text
stringlengths 2.85k
2.55M
| label
class label 11
classes |
|---|---|
A Recurrent Neural Model with Attention for
the Recognition of Chinese Implicit Discourse Relations
Samuel Rönnqvist1,2,∗ , Niko Schenk2,∗ and Christian Chiarcos2
1
Turku Centre for Computer Science – TUCS, Åbo Akademi University, Turku, Finland
2
Applied Computational Linguistics Lab, Goethe University, Frankfurt am Main, Germany
[email protected]
{schenk,chiarcos}@informatik.uni-frankfurt.de
arXiv:1704.08092v1 [cs.CL] 26 Apr 2017
Abstract
We introduce an attention-based Bi-LSTM
for Chinese implicit discourse relations
and demonstrate that modeling argument
pairs as a joint sequence can outperform
word order-agnostic approaches. Our
model benefits from a partial sampling
scheme and is conceptually simple, yet
achieves state-of-the-art performance on
the Chinese Discourse Treebank. We also
visualize its attention activity to illustrate
the model’s ability to selectively focus on
the relevant parts of an input sequence.
1
Introduction
True text understanding is one of the key goals in
Natural Language Processing and requires capabilities beyond the lexical semantics of individual
words or phrases. Natural language descriptions
are typically driven by an inter-sentential coherent structure, exhibiting specific discourse properties, which in turn contribute significantly to the
global meaning of a text. Automatically detecting
how meaning units are organized benefits practical downstream applications, such as question answering (Sun and Chai, 2007), recognizing textual entailment (Hickl, 2008), sentiment analysis
(Trivedi and Eisenstein, 2013), or text summarization (Hirao et al., 2013).
Various formalisms in terms of semantic coherence frameworks have been proposed to account for these contextual assumptions (Mann
and Thompson, 1988; Lascarides and Asher,
1993; Webber, 2004). The annotation schemata
of the Penn Discourse Treebank (Prasad et al.,
2008, PDTB) and the Chinese Discourse Treebank
(Zhou and Xue, 2012, CDTB), for instance, define
∗
Both first authors contributed equally to this work.
discourse units as syntactically motivated character spans in the text, augmented with relations
pointing from the second argument (Arg2, prototypically, a discourse unit associated with an explicit discourse marker) to its antecedent, i.e., the
discourse unit Arg1. Relations are labeled with
a relation type (its sense) and the associated discourse marker. Both, PDTB and CDTB, distinguish explicit from implicit relations depending on
the presence of such a marker (e.g., because/ 因).1
Sense classification for implicit relations is by far
more challenging because the argument pairs lack
the marker as an important feature. Consider, for
instance, the following example from the CDTB
as implicit C ONJUNCTION:
Arg1: 会谈就一些原则和具体问题进行了
深入讨论，达成了一些谅解 In the talks, they
discussed some principles and specific questions
in depth, and reached some understandings
Arg2: 双 方 一 致 认 为 会 谈 具 有 积 极 成 果
Both sides agree that the talks have positive results
Motivation: Previous work on implicit sense labeling is heavily feature-rich and requires domainspecific, semantic lexicons (Pitler et al., 2009;
Feng and Hirst, 2012; Huang and Chen, 2011).
Only recently, resource-lean architectures have
been proposed. These promising neural methods attempt to infer latent representations appropriate for implicit relation classification (Zhang
et al., 2015; Ji et al., 2016; Chen et al., 2016). So
far, unfortunately, these models have been evaluated only on four top-level senses—sometimes
even with inconsistent evaluation setups.2 Furthermore, most systems have initially been designed
for the English PDTB and involve complex, task1
The set of relation types and senses is completed by alternative lexicalizations (A LT L EX/discourse marker rephrased),
and entity relations (E NT R EL/anaphoric coherence).
2
E.g., four binary classifiers vs. four-way classification.
Our Contribution: In this work, we release the
first attention-based recurrent neural sense classifier, specifically developed for Chinese implicit
discourse relations. Inspired by Zhou et al. (2016),
our system is a practical adaptation of the recent
advances in relation modeling extended by a novel
sampling scheme.
Contrary to previous assertions by Rutherford
et al. (2016), our model demonstrates superior
performance over traditional bag-of-words approaches with feedfoward networks by treating
discourse arguments as a joint sequence. We evaluate our method within an independent framework and show that it performs very well beyond
standard class-level predictions, achieving stateof-the-art accuracy on the CDTB test set.
We illustrate how our model’s attention mechanism provides means to highlight those parts of
an input sequence that are relevant for the classification decision, and thus, it may enable a better understanding of the implicit discourse parsing problem. Our proposed network architecture
is flexible and largely language-independent as it
operates only on word embeddings. It stands out
due to its structural simplicity and builds a solid
ground for further development towards other textual domains.
2
Approach
We propose the use of an attention-based bidirectional Long Short-Term Memory (Hochreiter
ay
er
Sense labels
O
ut
pu
tl
y
r
⊕
A'1
h'2
h''2
A'2
A''1
A''2
e1
e2
t1
t2
h'k
..
.
h''1
hk
..
.
h2
⊕
⊕
h''k
A'k
A''k
..
.
Re
c
la urre
ye n
rs t
b
la edd
ay
ye in
er
r g
Em
tl
pu
h1
ek
..
.
A
tte
la ntio
ye n
r
·α
h'1
In
specific architectures (Liu and Li, 2016), while
discourse modeling techniques for Chinese have
received very little attention in the literature and
are still seriously underrepresented in terms of
publicly available systems. What is more, over
80% of all words in Chinese discourse relations
are implicit—compared to only 52% in English
(Zhou and Xue, 2012).
Recently, in the context of the CoNLL 2016
shared task (Xue et al., 2016), a first independent
evaluation platform beyond class level has been
established. Surprisingly, the best performing neural architectures to date are standard feedforward
networks, cf. Wang and Lan (2016); Schenk et al.
(2016); Qin et al. (2016). Even though these specific models completely ignore word order within
arguments, such feedforward architectures have
been claimed by Rutherford et al. (2016) to generally outperform any thoroughly-tuned recurrent
architecture.
tk
(<ARG1>, 会 谈, ..., </ARG1>, <ARG2>, 双 方, ..., </ARG2>)
Token input sequence
Figure 1: The attention-based bidirectional LSTM
network for the task of modeling argument pairs
for Chinese implicit discourse relations.
and Schmidhuber, 1997, LSTM) network to predict senses of discourse relations. The model
draws upon previous work on LSTM, in particular its bidirectional mode of operation (Graves and
Schmidhuber, 2005), attention mechanisms for recurrent models (Bahdanau et al., 2014; Hermann
et al., 2015), and the combined use of these techniques for entity relation recognition in annotated
sequences (Zhou et al., 2016). More specifically,
our model is a flexible recurrent neural network
with capabilities to sequentially inspect tokens and
to highlight which parts of the input sequence are
most informative for the discourse relation recognition task, using the weighting provided by the attention mechanism. Furthermore, the model benefits from a novel sampling scheme for arguments,
as elaborated below. The system is learned in an
end-to-end manner and consists of multiple layers,
which are illustrated in Figure 1.
First, token sequences are taken as input and
special markers (<ARG1>, </ARG1>, etc.) are
inserted into the corresponding positions to inform
the model on the start and end points of argument
spans. This way, we can ensure a general flexibility in modeling discourse units and could easily extend them with additional context, for instance. In our experiments on implicit arguments,
only the tokens in the respective spans are considered. Note that, unlike previous works, our approach models Arg1-Arg2 pairs as a joint sequence
and does not first compute intermediate representations of arguments separately.
Second, an input layer encodes tokens using
one-hot vector representations (ti for tokens at positions i ∈ [1, k]), and a subsequent embedding
layer provides a dense representation (ei ) to serve
as input for the recurrent layers. The embedding
layer is initialized using pre-trained word vectors,
in our case 300-dimensional Chinese Gigaword
vectors (Graff and Chen, 2005).3 These embeddings are further tuned as the network is trained
towards the prediction task. Embeddings for unknown tokens, e.g., markers, are trained by backpropagation only. Note that, tokens, markers and
the pre-trained vectors represent the only source of
information for the prediction task.
For the recurrent setup, we use a layer of LSTM
networks in a bidirectional manner, in order to better capture dependencies between parts of the input sequence by inspection of both left and righthand-side contexts at each time step. The LSTM
holds a state representation as a continuous vector
passed to the subsequent time step, and it is capable of modeling long-range dependencies due to
its gated memory. The forward (A0 ) and backward
(A00 ) LSTMs traverse the sequence ei , producing
sequences of vectors h0i and h00i respectively, which
are then summed together (indicated by ⊕ in Figure 1).
The resulting sequence of vectors hi is reduced
into a single vector and fed to the final softmax
output layer in order to classify the sense label y
of the discourse relation. This vector may be obtained either as the final vector h produced by an
LSTM, or through pooling of all hi , or by using
attention, i.e., as a weighted sum over hi . While
the model may be somewhat more difficult to optimize using attention, it provides the added benefit
of interpretability, as the weights highlight to what
extent the classifier considers the LSTM state vectors at each token during modeling. This is particularly interesting for discourse parsing, as most
previous approaches have provided little support
for pinpointing the driving features in each argument span.
Finally, the attention layer contains the trainable
3
http://www.cs.brandeis.edu/˜clp/
conll16st/dataset.html
vector w (of the same dimensionality as vectors
hi ) which is used to dynamically produce a weight
vector α over time steps i by:
α = sof tmax(wT tanh(H))
where H is a matrix consisting of vectors hi . The
output layer r is the weighted sum of vectors in H:
r = HαT
Partial Argument Sampling: For the purpose
of enlarging the instance space of training items
in the CDTB, and thus, in order to improve the
predictive performance of the model, we propose
a novel partial sampling scheme of arguments,
whereby the model is trained and validated on sequences containing both arguments, as well as single arguments. A data point (a1 , a2 , y), with ai being the token sequence of argument i, is expanded
into {(a1 , a2 , y), (a1 , a2 , y), (a1 , y), (a2 , y)}. We
duplicate bi-argument samples (a1 , a2 , y) (in
training and development data only) to balance
their frequencies against single-argument samples.
Two lines of motivation support the inclusion
of single argument training examples, grounded
in linguistics and machine learning, respectively.
First, it has been shown that single arguments in
isolation can evoke a strong expectation towards
a certain implicit discourse relation, cf. Asr and
Demberg (2015) and, in particular, Rohde and
Horton (2010) in their psycholinguistic study on
implicit causality verbs. Second, the procedure
may encourage the model to learn better representations of individual argument spans in support of
modeling of arguments in composition, cf. LeCun
et al. (2015). Due to these aspects, we believe this
data augmentation technique to be effective in reinforcing the overall robustness of our model.
Implementational Details: We train the model
using fixed-length sequences of 256 tokens with
zero padding at the beginning of shorter sequences
and truncate longer ones. Each LSTM has a vector
dimensionality of 300, matching the embedding
size. The model is regularized by 0.5 dropout rate
between the layers and weight decay (2.5e−6 ) on
the LSTM inputs. We employ Adam optimization
(Kingma and Ba, 2014) using the cross-entropy
loss function with mini batch size of 80.4
4
The model is implemented in Keras https://
keras.io/.
Rank
1
2
3
4
5
6
7
CDTB Development Set
System
% accuracy
Wang and Lan (2016)
73.53
Qin et al. (2016)
71.57
Schenk et al. (2016)
70.59
Rutherford and Xue (2016)
68.30
Weiss and Bajec (2016)
66.67
Weiss and Bajec (2016)
61.44
Jian et al. (2016)
21.90
This Paper:
93.52∗
Rank
1
2
3
4
5
6
7
CDTB Test Set
System
Wang and Lan (2016)
Schenk et al. (2016)
Rutherford and Xue (2016)
Qin et al. (2016)
Weiss and Bajec (2016)
Weiss and Bajec (2016)
Jian et al. (2016)
This Paper:
% accuracy
72.42
71.87
70.47
67.41
64.07
63.51
21.73
73.01
Table 1: Non-explicit parser scores on the official CoNLL 2016 CDTB development and test sets.
(∗ Scores on development set are obtained through partial sampling and are not directly comparable.)
Sense Label
C ONJUNCTION
Training
Dev’t
Test
5,174
189
228
majority class
(66.3%)
(62.8%)
(64.8%)
1,188
1,099
187
66
56
26
26
7
7,804
48
50
10
3
1
0
0
0
301
40
71
8
1
3
1
0
0
352
E XPANSION
E NT R EL
C AUSATION
C ONTRAST
P URPOSE
C ONDITIONAL
T EMPORAL
P ROGRESSION
# impl. rels
Table 2: Implicit sense labels in the CDTB.
3
Evaluation
We evaluate our recurrent model on the CoNLL
2016 shared task data5 which include the official
training, development and test sets of the CDTB;
cf. Table 2 for an overview of the implicit sense
distribution.6
In accordance with previous setups (Rutherford
et al., 2016), we treat entity relations (E NT R EL)
as implicit and exclude A LT L EX relations. In the
evaluation, we focus on the sense-only track, the
subtask for which gold arguments are provided
and a system is supposed to label a given argument pair with the correct sense. The results are
shown in Table 1.
With our proposed architecture it is possible to
correctly label 257/352 (73.01%) of implicit rela5
http://www.cs.brandeis.edu/˜clp/
conll16st/
6
Note that, in the CDTB, implicit relations appear almost
three times more often than explicit relations. Out of these,
65% appear within the same sentence. Finally, 25 relations in
the training set have two labels.
tions on the test set, outperforming the best feedforward system of Wang and Lan (2016) and all
other word order-agnostic approaches. Development and test set performances suggest the robustness of our approach and its ability to generalize
to unseen data.
Ablation Study: We perform an ablation study to
quantitatively assess the contribution of two of the
characteristic aspects of our model. First, we compare the use of the attention mechanism against
the simpler alternative of feeding the final LSTM
hidden vectors (h0k and h001 ) directly to the output
layer. When attention is turned off, this yields
an absolute decrease in performance of 2.70% on
the test set, which is substantial and significant according to a Welch two-sample t-test (p < .001).
Second, we independently compare the use of the
partial sampling scheme against training on the
standard argument pairs in the CDTB. Here, the
absence of the partial sampling scheme yields an
absolute decrease in accuracy of 5.74% (p < .001),
which demonstrates its importance for achieving
competitive performance on the task.
Performance on the PDTB: As a side experiment, we investigate the model’s language independence by applying it to the implicit argument
pairs of the English PDTB. Due to computational
time constraints we do not optimize hyperparameters, but instead train the model using identical
settings as for Chinese, which is expected to lead
to suboptimal performance on the evaluation data.
Nevertheless, we measure 27.09% accuracy on the
PDTB test set (surpassing the majority class baseline of 22.01%), which shows that the model has
potential to generalize across implicit discourse
relations in a different language.
C ONJUNCTION :
<Arg1> 会 谈 就 ⼀ 些 原则 和 具体 问题 进⾏ 了 深 ⼊ 讨 论 ， 达 成 了 ⼀ 些 谅 解 </Arg1>
In the talks, they discussed some principles and specific questions in depth, and reached some understandings
<Arg2> 双 ⽅ ⼀ 致 认 为 会谈 具有 积极 成果 </Arg2>
Both sides agree that the talks have positive results
E NT R EL :
<Arg1> 他
说
：
我们
希望
澳门
政府
对于
这
三
个
问题
继续
给予
关注
，
He said: We hope that the Macao government will continue to pay attention to these three issues,
以
求 得 最 后 的 妥善 解决 </Arg1>
in order to find a final proper solution
<Arg2> 李 鹏 说 ， ⻙ 奇 ⽴ 总督 为 澳门 问题 的 顺 利 解 决 做 了 许 多 有 益 的 ⼯ 作 ，
Peng Li said, Governor Liqi Wei has done a lot of useful work for the smooth settlement of the Macao question,
对 此 我 们 表 ⽰ 赞赏 </Arg2>
we appreciate that
Figure 2: Visualization of attention weights for Chinese characters with high (dark blue) and low (light
blue) intensities. The underlined English phrases are semantically structure-shared by the two arguments.
Visualizing Attention Weights: Finally, in Figure 2, we illustrate the learned attention weights
which pinpoint important subcomponents within
a given implicit discourse relation. For the implicit C ONJUNCTION relation the weights indicate
a peak on the transition between the argument
boundary, establishing a connection between the
semantically related terms understandings–agree.
Most E NT R ELs show an opposite trend: here second arguments exhibit larger intensities than Arg1,
as most entity relations follow the characteristic
writing style of newspapers by adding additional
information by reference to the same entity.
4
Summary & Outlook
In this work, we have presented the first attentionbased recurrent neural sense labeler specifically
developed for Chinese implicit discourse relations.
Its ability to model discourse units sequentially
and jointly has been shown to be highly beneficial, both in terms of state-of-the-art performance
on the CDTB (outperforming word order-agnostic
feedforward approaches), and also in terms of
insightful observations into the inner workings
of the model through its attention mechanism.
The architecture is structurally simple, benefits
from partial argument sampling, and can be eas-
ily adapted to similar relation recognition tasks. In
future work, we intend to extend our approach to
different languages and domains, e.g., to the recent
data sets on narrative story understanding or question answering (Mostafazadeh et al., 2016; Feng
et al., 2015). We believe that recurrent modeling
of implicit discourse information can be a driving
force in successfully handling such complex semantic processing tasks.7
Acknowledgments
The authors would like to thank Ayah Zirikly,
Philip Schulz and Wei Ding for their very helpful suggestions on an early draft version of the paper, and also thank the anonymous reviewers for
their valuable feedback and insightful comments.
We are grateful to Farrokh Mehryary for technical support with the attention layer implementation. Computational resources were provided by
CSC – IT Centre for Science, Finland, and Arcada
University of Applied Sciences, Helsinki, Finland.
Our research at Goethe University Frankfurt was
supported by the project ‘Linked Open Dictionaries (LiODi, 2015-2020)’, funded by the German
Ministry for Education and Research (BMBF).
7
The code involved in this study is publicly
available
at
http://www.acoli.informatik.
uni-frankfurt.de/resources/.
References
Fatemeh Torabi Asr and Vera Demberg. 2015. Uniform Information Density at the Level of Discourse Relations: Negation Markers and Discourse Connective Omission. In 11th International
Conference on Computational Semantics (IWCS).
page 118. http://www.coli.uni-saarland.de/ fatemeh/iwcs2015.pdf.
Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua
Bengio. 2014. Neural Machine Translation by
Jointly Learning to Align and Translate. CoRR
abs/1409.0473. http://arxiv.org/abs/1409.0473.
Tsutomu Hirao, Yasuhisa Yoshida, Masaaki Nishino,
Norihito Yasuda, and Masaaki Nagata. 2013.
Single-Document Summarization as a Tree Knapsack Problem. In Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, EMNLP 2013, 18-21 October 2013, Grand Hyatt Seattle, Seattle, Washington, USA, A meeting of SIGDAT, a Special
Interest Group of the ACL. pages 1515–1520.
http://aclweb.org/anthology/D/D13/D13-1158.pdf.
Sepp Hochreiter and Jürgen Schmidhuber. 1997. Long
Short-Term Memory. Neural Comput. 9(8):1735–
1780. https://doi.org/10.1162/neco.1997.9.8.1735.
Jifan Chen, Qi Zhang, Pengfei Liu, Xipeng Qiu,
and Xuanjing Huang. 2016. Implicit Discourse
Relation Detection via a Deep Architecture with
Gated Relevance Network. In Proceedings of the
54th Annual Meeting of the Association for Computational Linguistics, ACL 2016, August 7-12,
2016, Berlin, Germany, Volume 1: Long Papers.
http://aclweb.org/anthology/P/P16/P16-1163.pdf.
Hen-Hsen Huang and Hsin-Hsi Chen. 2011. Chinese
Discourse Relation Recognition. In Proceedings
of 5th International Joint Conference on Natural
Language Processing. Asian Federation of Natural
Language Processing, Chiang Mai, Thailand, pages
1442–1446. http://www.aclweb.org/anthology/I111170.
Minwei Feng, Bing Xiang, Michael R. Glass,
Lidan Wang, and Bowen Zhou. 2015.
Applying deep learning to answer selection: A
study and an open task. In 2015 IEEE Workshop on Automatic Speech Recognition and
Understanding, ASRU 2015, Scottsdale, AZ,
USA, December 13-17, 2015. pages 813–820.
https://doi.org/10.1109/ASRU.2015.7404872.
Yangfeng Ji, Gholamreza Haffari, and Jacob Eisenstein. 2016. A Latent Variable Recurrent Neural Network for Discourse-Driven Language Models. In Proceedings of the 2016 Conference of
the North American Chapter of the Association
for Computational Linguistics: Human Language
Technologies. Association for Computational Linguistics, San Diego, California, pages 332–342.
http://www.aclweb.org/anthology/N16-1037.
Vanessa Wei Feng and Graeme Hirst. 2012. Text-level
Discourse Parsing with Rich Linguistic Features. In
Proceedings of the 50th Annual Meeting of the Association for Computational Linguistics: Long Papers
- Volume 1. Association for Computational Linguistics, Stroudsburg, PA, USA, ACL ’12, pages 60–68.
http://www.aclweb.org/anthology/P12-1007.
Ping Jian, Xiaohan She, Chenwei Zhang, Pengcheng
Zhang, and Jian Feng. 2016.
Discourse
Relation Sense Classification Systems for
CoNLL-2016 Shared Task.
In Proceedings
of the CoNLL-16 shared task. Association for
Computational Linguistics,
pages 158–163.
https://doi.org/10.18653/v1/K16-2022.
David Graff and Ke Chen. 2005. Chinese Gigaword.
LDC Catalog No.: LDC2003T09, ISBN, 1:58563–
58230.
Diederik P. Kingma and Jimmy Ba. 2014. Adam:
A method for stochastic optimization.
CoRR
abs/1412.6980. http://arxiv.org/abs/1412.6980.
Alex Graves and Jürgen Schmidhuber. 2005.
Framewise phoneme classification with bidirectional LSTM and other neural network architectures. Neural Networks 18(5-6):602–610.
https://doi.org/10.1016/j.neunet.2005.06.042.
Alex Lascarides and Nicholas Asher. 1993. Temporal Interpretation, Discourse Relations and Commonsense entailment. Linguistics and Philosophy
16(5):437–493.
Karl Moritz Hermann, Tomas Kocisky, Edward
Grefenstette, Lasse Espeholt, Will Kay, Mustafa Suleyman, and Phil Blunsom. 2015. Teaching machines to read and comprehend. In Advances in Neural Information Processing Systems. pages 1693–
1701.
Andrew Hickl. 2008.
Using Discourse Commitments to Recognize Textual Entailment.
In
Proceedings of the 22nd International Conference on Computational Linguistics - Volume 1.
Association for Computational Linguistics, Stroudsburg, PA, USA, COLING ’08, pages 337–344.
http://dl.acm.org/citation.cfm?id=1599081.1599124.
Yann LeCun, Yoshua Bengio, and Geoffrey Hinton.
2015. Deep learning. Nature 521(7553):436–444.
Yang Liu and Sujian Li. 2016. Recognizing Implicit Discourse Relations via Repeated Reading: Neural Networks with Multi-Level Attention.
In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, EMNLP 2016, Austin, Texas,
USA, November 1-4, 2016. pages 1224–1233.
http://aclweb.org/anthology/D/D16/D16-1130.pdf.
William C. Mann and Sandra A. Thompson. 1988.
Rhetorical structure theory: Toward a functional theory of text organization. Text 8(3):243–281.
Nasrin Mostafazadeh, Nathanael Chambers, Xiaodong
He, Devi Parikh, Dhruv Batra, Lucy Vanderwende,
Pushmeet Kohli, and James Allen. 2016. A Corpus and Cloze Evaluation for Deeper Understanding of Commonsense Stories. In Proceedings of the
2016 Conference of the North American Chapter of
the Association for Computational Linguistics: Human Language Technologies. Association for Computational Linguistics, San Diego, California, pages
839–849. http://www.aclweb.org/anthology/N161098.
Emily Pitler, Annie Louis, and Ani Nenkova. 2009.
Automatic Sense Prediction for Implicit Discourse
Relations in Text. In Proceedings of the Joint Conference of the 47th Annual Meeting of the ACL and
the 4th International Joint Conference on Natural
Language Processing of the AFNLP: Volume 2 Volume 2. Association for Computational Linguistics, Stroudsburg, PA, USA, ACL 2009, pages 683–
691. http://www.aclweb.org/anthology/P/P09/P091077.pdf.
Rashmi Prasad, Nikhil Dinesh, Alan Lee, Eleni Miltsakaki, Livio Robaldo, Aravind Joshi, and Bonnie Webber. 2008. The Penn Discourse TreeBank
2.0. In Proceedings, 6th International Conference
on Language Resources and Evaluation. Marrakech,
Morocco, pages 2961–2968.
Lianhui Qin, Zhisong Zhang, and Hai Zhao. 2016.
Shallow Discourse Parsing Using Convolutional
Neural Network. In Proceedings of the CoNLL-16
shared task. Association for Computational Linguistics, pages 70–77. https://doi.org/10.18653/v1/K162010.
Hannah Rohde and William Horton. 2010. Why or
what next? Eye movements reveal expectations
about discourse direction. Talk at the 23rd Annual
CUNY Conference on Human Sentence Processing.
New York, NY.
Attapol Rutherford and Nianwen Xue. 2016. Robust Non-Explicit Neural Discourse Parser in English and Chinese. In Proceedings of the CoNLL-16
shared task. Association for Computational Linguistics, pages 55–59. https://doi.org/10.18653/v1/K162007.
Attapol T. Rutherford, Vera Demberg, and Nianwen Xue. 2016. Neural Network Models for Implicit Discourse Relation Classification in English
and Chinese without Surface Features.
CoRR
abs/1606.01990. http://arxiv.org/abs/1606.01990.
Niko Schenk, Christian Chiarcos, Kathrin Donandt,
Samuel Rönnqvist, Evgeny Stepanov, and Giuseppe
Riccardi. 2016. Do We Really Need All Those
Rich Linguistic Features? A Neural Network-Based
Approach to Implicit Sense Labeling. In Proceedings of the CoNLL-16 shared task. Association for Computational Linguistics, pages 41–49.
https://doi.org/10.18653/v1/K16-2005.
Mingyu Sun and Joyce Y Chai. 2007. Discourse processing for context question answering based on
linguistic knowledge. Knowledge-Based Systems
20(6):511–526.
Rakshit S. Trivedi and Jacob Eisenstein. 2013. Discourse connectors for latent subjectivity in sentiment analysis. In Human Language Technologies: Conference of the North American Chapter of the Association of Computational Linguistics, Proceedings, June 9-14, 2013, Westin Peachtree
Plaza Hotel, Atlanta, Georgia, USA. pages 808–813.
http://aclweb.org/anthology/N/N13/N13-1100.pdf.
Jianxiang Wang and Man Lan. 2016. Two Endto-end Shallow Discourse Parsers for English and
Chinese in CoNLL-2016 Shared Task. In Proceedings of the CoNLL-16 shared task. Association for Computational Linguistics, pages 33–40.
https://doi.org/10.18653/v1/K16-2004.
Bonnie L. Webber. 2004.
D-LTAG: extending lexicalized TAG to discourse.
Cognitive Science 28(5):751–779.
http://dblp.unitrier.de/db/journals/cogsci/cogsci28.html.
Gregor Weiss and Marko Bajec. 2016. Discourse Sense
Classification from Scratch using Focused RNNs. In
Proceedings of the CoNLL-16 shared task. Association for Computational Linguistics, pages 50–54.
https://doi.org/10.18653/v1/K16-2006.
Nianwen Xue, Hwee Tou Ng, Sameer Pradhan, Bonnie Webber, Attapol Rutherford, Chuan Wang, and
Hongmin Wang. 2016. The CoNLL-2016 Shared
Task on Shallow Discourse Parsing. In Proceedings
of the Twentieth Conference on Computational Natural Language Learning - Shared Task. Association
for Computational Linguistics, Berlin, Germany.
Biao Zhang, Jinsong Su, Deyi Xiong, Yaojie Lu, Hong
Duan, and Junfeng Yao. 2015. Shallow Convolutional Neural Network for Implicit Discourse Relation Recognition. In Proceedings of the 2015
Conference on Empirical Methods in Natural Language Processing, EMNLP 2015, Lisbon, Portugal, September 17-21, 2015. pages 2230–2235.
http://aclweb.org/anthology/D/D15/D15-1266.pdf.
Peng Zhou, Wei Shi, Jun Tian, Zhenyu Qi, Bingchen
Li, Hongwei Hao, and Bo Xu. 2016. AttentionBased Bidirectional Long Short-Term Memory Networks for Relation Classification. In Proceedings
of the 54th Annual Meeting of the Association for
Computational Linguistics, ACL 2016, August 7-12,
2016, Berlin, Germany, Volume 2: Short Papers.
http://aclweb.org/anthology/P/P16/P16-2034.pdf.
Yuping Zhou and Nianwen Xue. 2012.
PDTBstyle Discourse Annotation of Chinese Text.
In Proceedings of the 50th Annual Meeting
of the Association for Computational Linguistics (Volume 1: Long Papers). Association for
Computational Linguistics, Jeju Island, Korea,
pages 69–77. http://www.aclweb.org/anthologynew/P/P12/P12-1008.bib.
| 2cs.AI
|
arXiv:1708.06407v1 [math.GN] 21 Aug 2017
The minimizing vector theorem
in symmetrized max-plus algebra
Cenap Özel
Artur Piękosz
Eliza Wajch
Hanifa Zekraoui
Department of Mathematics, King Abdulaziz University,
Jeddah, Kingdom of Saudi Arabia
Institute of Mathematics, Cracow University of Technology,
Cracow, Poland
Department of Mathematics and Physics,
University of Natural Sciences and Humanities in Siedlce,
Siedlce, Poland
Department of Mathematics, Larbi Ben M’hidi University,
Oum El Bouaghi, Algeria
August 23, 2017
Abstract
Assuming ZF and its consistency, we study some topological and geometrical properties of the symmetrized max-plus algebra in the absence
of the axiom of choice in order to discuss the minimizing vector theorem
for finite products of copies of the symmetrized max-plus algebra. Several relevant statements that follow from the axiom of countable choice
restricted to sequences of subsets of the real line are shown. Among them,
it is proved that if all simultaneously complete and connected subspaces
of the plane are closed, then the real line is sequential. A brief discussion
about semidenrites is included. Older known proofs in ZFC of several
basic facts relevant to proximinal and Chebyshev sets in metric spaces are
replaced by new proofs in ZF. It is proved that a nonempty subset C of the
symmetrized max-plus algebra is Chebyshev in this algebra if and only if
C is simultaneously closed and connected. An application of it to a version
of the minimizing vector theorem for finite products of the symmetrized
max-plus algebra is shown. Open problems concerning some statements
independent of ZF and other statements relevant to Chebyshev sets are
posed.
Key words: symmetrized max-plus algebra, metric, complete metric, Cantor complete metric, proximinal set, Chebyshev set, convexity,
geometric convexity, minimizing vector theorem, semidendrite, ZF, axiom of countable choice for the real line, independence results.
MSC[2010]: Primary: 15A80, 16Y60. Secondary: 03E25, 54F15.
1
1
Introduction
The Max-Plus Algebra (shortly: MPA), known also as the max-plus dioid ([10])
or the tropical semiring ([22]), is the idempotent semiring (see Definition 1 in
[17]) over the union of the set of all real numbers and one-point set consisting
of ε = −∞, equipped with the maximum (as the addition) and the addition (as
the multiplication). The zero and the unit of this semiring are respectively ε
and 0.
Despite the remarkable analogy between these two basic operations in MPA
and the conventional algebra, the principal difference reflects in missing of inverse elements related to the first operation: max. A partial remedy is the
Symmetrized Max-Plus Algebra (shortly: SMPA) which is an extension of MPA
built by a method similar to that used in extending a given set of all natural
numbers satisfying Peano’s postulates to a set that can be called the set of all
integers.
Compared to conventional linear algebra and linear system theory, the linear max-plus algebra is far from fully developed, even though many papers and
textbooks were written on the topics of matrix calculations, among them, relevant to different types of decomposition, spectral theory, linear systems and
linear optimization. Some important papers dealing with these subjects also
were written in SMPA; for instance, let us refer to [21].
In this paper, we give some topological and geometrical properties of SMPA
in order to start a deeper investigation of the minimizing vector theorem in
products of SMPA and point out new open problems relevant to Chebyshev
sets.
Although most mathematicians seem to assume ZFC as the axiomatic foundation, we realize that the axiom of choice (AC) is not needed to obtain the
main results of our work. Therefore, the basic set-theoretic assumption of this
study is ZF, formulated in [16], without any form of the axiom of choice. If it
occurs necessary, we also use AC or another axiom which is a weaker form of
AC; however, we clearly inform which statements in our article are independent
of ZF. We use the notation of [12] for ZF and axioms relevant to the axiom
of choice. All topological notions not defined here can be found in [5], [12] and
[18]. The book [14] is the most useful source of information about validity and
falsity in models of ZF.
Our paper is organized as follows. Section 2 concerns basic properties of
d-proximinal and d-Chebyshev sets in an arbitrary metric space (X, d). New
proofs in ZF to several facts on d-proximinality are shown. The proofs are
more subtle than the already known ones in ZFC. We give a brief introduction
about calculation in MPA and the construction of max-plus algebra of pairs
(MPAP) in subsection 3.1. In subsection 3.2, we give a short presentation of
the construction of SMPA from MPAP and some remarks about computation in
SMPA. In section 4, we define equivalent Euclidean and inner metrics in the set S
of all points of SMPA. Equipped with the natural topology induced by the inner
metric, the space S is a semidendrite which is embeddable in the complex plane
C. Some properties of semidendrites in ZF are shown. Section 5 concerns finite
2
products Sn of S. Natural metrics in Sn are considered. A notion of a geometric
segment in Sn is introduced and it is shown that geometric segments can be
identified with suitable broken lines in Cn . Section 6 is about completeness
and Cantor-completeness of the metrics in Sn defined in section 5. A few settheoretic open problems are posed. New equivalents of the independent of ZF
statement that R is sequential are found. In section 7, we introduce the structure
of a semimodule in Sn . Section 8 is devoted to distinct concepts of convex sets
in Sn . The final section 9 is about the minimizing vector theorem in Sn . We give
here a complete, satisfactory characterization of Chebyshev sets in S. We also
investigate finite products of Chebyshev sets in S. We pose open problems on the
minimizing vector theorem. Our paper is a preparation to deeper investigations
of Chebyshev sets in Sn in the future.
We shall use the set ω of all finite ordinal numbers of von Neumann as the
set of all non-negative integers (see [16]).
2
Chebyshev sets in metric spaces-preliminaries
The aim of this section is to establish terminology and basic properties of Chebyshev sets in an arbitrary metric space. An extensive bibliography on Chebyshev
sets in normed spaces which includes also survey articles is given in [6], so it is
not necessary to repeat this bibliography here.
2.1
d-proximinal and d-Chebyshev sets
Throughout this section, we assume that (X, d) is a metric space. As usual, for
x ∈ X and a positive real number r, we denote Bd (x, r) = {y ∈ X : d(x, y) < r}
and B̄d (x, r) = {y ∈ X : d(x, y) ≤ r}. The symbol τ (d) stands for the topology
induced by d. For simplicity, we denote by X the topological space (X, τ (d)).
If Y is a subset of X, we shall consider Y as a metric subspace of (X, d) and
as a topological subspace of (X, τ (d)). The following definition can be regarded
as an adaptation of Definition 2.3 of [6] to arbitrary metric spaces. If it is not
stated otherwise, Rn will stand for the n-dimensional Euclidean space equipped
with the Euclidean metric induced by the standard inner product in Rn .
Definition 1. Let K ⊆ X.
(i) We define a set-valued mapping Pd,K : X → P(X) by Pd,K (x) = {y ∈ K :
d(x, y) = d(x, K)} if K 6= ∅ and by Pd,K (x) = ∅ if K = ∅, where x ∈ X.
(ii) We say that a d-nearest point to x ∈ X in K or, equivalently, the best
d-approximation of x in K is a point which belongs to the set Pd,K (x).
(iii) The set K is called d-proximinal in X if Pd,K (x) 6= ∅ for all x ∈ X.
(iv) The set K is called a d-Chebyshev set in X if Pd,K (x) is a singleton for
each x ∈ X.
3
(v) If x ∈ X is such that Pd,K (x) is a singleton, we denote by pd,K (x) the
unique point of Pd,K (x) and we call pd,K (x) the d-projection of the point
x onto K.
(vi) If K is a d-Chebyshev set in X, then the mapping pd,K : X → K is called
the d-projection of X onto K. If this does not lead to misunderstanding,
d-projections will be called metric projections.
Definition 2. Let Y be a subset of the metric space (X, d) and let K ⊆ Y .
We say that K is d-proximinal in Y if, for each y ∈ Y , the set Pd,K (y) is
nonempty. If, for each y ∈ Y , the set Pd,K (y) is a singleton, we say that K is
d-Chebyshev in Y .
A set which is d-Chebyshev in a subspace Y of the metric space (X, d) need
not be d-Chebyshev in X. (Take a non-Chebyshev subset Y in R2 . Then Y is
Chebyshev in itself.)
Although it has been written a lot about Chebyshev sets in normed vector
spaces, in general, it is not known much about the nature of d-Chebyshev sets for
an arbitrary metric d; however, the concepts of proximinal and Chebyshev sets
in metric spaces, especially in strictly convex metric spaces, were investigated in
the past, for instance, in [1], [19] and [20]. Unfortunately, even in the case when
d is induced by the inner product of a Hilbert space over the field R, versions
of the axiom of choice have been involved in proofs of some properties of dChebyshev sets, while it is more desirable to investigate whether a statement
is provable in ZF or it can fail in a model of ZF. For instance, the statement
that Rn is a Fréchet space was used in the proof on page 363 of [23] that every
Chebyshev set in the n-dimensional Euclidean space Rn is closed. However, in
view of Theorem 4.54 of [12], that R is Fréchet is an equivalent of CC(R) which
asserts that every nonempty countable collection of nonempty subsets of R has
a choice function (see Definition 2.9 (1) of CC(R) in [12] and Form 94 in [14]).
In Cohen’s original model of ZF (see model M1 in [14]), R is not Fréchet (see
Form 73 of [14]). Therefore, the proof in [23] that Chebyshev sets in Rn are
closed is not a proof in ZF but it is a too complicated proof in ZF + CC(R).
If we replace the norm by the metric d in Proposition 2.9 of [6] and in its proof
in [6], then we can see that the following more general fact than Proposition
2.9 of [6] has a trivially simple proof in ZF which is included below to show its
simplicity:
Proposition 3. Each d-proximinal set in a nonempty metric space (X, d) is
nonempty and closed in (X, d). In particular, each d-Chebyshev set in the metric
space (X, d) is closed in (X, d).
Proof. Let A be a d-proximinal set in X. It is obvious that A 6= ∅ because
X 6= ∅. Let x ∈ cld (A) and a ∈ Pd,A (x). Then d(x, a) = d(x, A) = 0, so x ∈ A
because x = a. Therefore, A = cld (A).
Definition 4. The d-graph of a set A ⊆ X is the set
[
Grd (A) =
[{x} × Pd,A (x)].
x∈X
4
Proposition 5. The d-graph of a closed subset of X is closed in X × X.
Proof. Let K be a closed subset of X. If K = ∅, then Grd (K) = ∅. Assume
that K 6= ∅. Let (x, y) ∈ (X × X) \ Grd (K). Then y ∈
/ Pd,K (x). If y ∈
/ K,
then U = X × (X \ K) is a neighbourhood of (x, y) such that U ∩ Grd (K) = ∅.
Now, suppose that y ∈ K. Then d(x, y) > d(x, K) because y ∈
/ Pd,K (x).
Let r be a positive real number such that 3r < d(x, y) − d(x, K). Put V =
Bd (x, r) × Bd (y, r). Suppose that (z, t) ∈ V ∩ Grd (K). Then t ∈ Pd,K (z), so
d(z, t) = d(z, K). Moreover, d(x, y) ≤ d(x, z) + d(z, t) + d(t, y) < d(z, K) + 2r ≤
d(z, x) + d(x, K) + 2r ≤ d(x, K) + 3r.This gives that d(x, y) − d(x, K) < 3r
which is impossible. Hence V ∩ Grd (K) = ∅. In consequence, (X × X) \ Grd (K)
is open in X × X.
Propositions 3 and 5 taken together imply the following theorem:
Theorem 6. If K is a d-proximinal subset of X then the d-graph of K is closed
in X × X. In particular, if K is a d-Chebyshev set in X, then the d-projection
of X onto K has a closed graph.
Theorem 6 is an extension and generalization of Corollary 2.19 from [6] and,
for instance, of Theorem 4 from [19]. However, the proofs to Corollary 2.19 in
[6] and to Theorem 4 in [19] are not proofs in ZF because, in ZF, only sequential
closedness of Chebyshev sets was shown in [6] and [19], while it is known that
a sequentially closed subset of a metric space may fail to be closed in a model
of ZF. Namely, as in Theorem 4.55 of [12], we denote by CC(cR) the following
statement: Every sequence of nonempty complete metric subspaces of R has
a choice function. In view of Theorem 4.55 of [12], that R is sequential is an
equivalent of CC(cR).
Remark 7. In Cohen’s original model (see model M1 in [14]), for each natural
number n ≥ 1, the space Rn is not sequential because it contains non-separable
subspaces (see Form 92 of [14]).
It is known that CC(R) is strictly stronger than CC(cR) in ZF (see pages
74-77 of [12]).
Definition 8. Let K be a nonempty subset of X. We say that the mapping
Pd,K : X → P (X) is d-continuous at x ∈ X if Pd,K (x) is a singleton and the
following condition is satisfied:
∀ε∈(0;+∞) ∃δ∈(0;+∞) ∀y,z∈X [(d(x, y) < δ ∧ z ∈ Pd,K (y)) ⇒ d(z, pd,K (x)) < ε].
Definition 8 is a modification of Definition 2.23 of [6]. In the case when
K is a nonempty d-Chebyshev set in X, then the metric projection pd,K of X
onto K is continuous if and only if Pd,K is d-continuous at each x ∈ X in the
sense of our Definition 8, while pd,K is sequentially continuous if and only if
Pd,K is continuous at each x ∈ X in the sense of Definition 2.23 of [6] under
the additional assumption that X is a normed vector space and d is induced by
the norm of X. In general, given metric spaces (X, d) and (Y, ρ), the sequential
5
continuity of a mapping f : X → Y is not equivalent to the continuity of f in
ZF (see Theorem 4.54 of [12] and Form 94 in [14] which fails in model M1 of
[14]). It can de deduced from the proof in ZF to Theorem 3.15 in [12] that the
following theorem holds true in ZF:
Theorem 9. If X is a separable metrizable space and Y is a metrizable space,
then a mapping f : X → Y is continuous if and only if f is sequentially continuous.
Remark 10. Let K be a d-Chebyshev set in a normed vector space X where
d is induced by the norm of X. If X is finitely dimensional, the classical proofs
of the continuity of the metric projection pd,K of X onto K given, for instance,
in [6] and [23] have a gap in ZF because only sequential continuity of pd,K was
shown there. To deduce from the sequential continuity of pd,K its continuity, we
need Theorem 9. If X is infinitely dimensional, pd,K may fail to be continuous
(see Example 2.54 in [6] and page 246 in [9]).
We are going to give a direct proof in ZF of the continuity of pd,K if K
is a d-Chebyshev set in an arbitrary nonempty metric space (X, d) having the
Heine-Borel property.
Definition 11. Let Y be a subset of the metric space (X, d). We say that:
(i) Y has the Heine-Borel property with respect to d or, equivalently,
Y is d-totally complete if, for each K ⊆ Y , it is true that K is compact
if K is simultaneously closed in Y and d-bounded;
(ii) Y is d-boundedly compact if, for each positive real number r and for
each x ∈ X, the set Y ∩ B̄d (x, r) is compact.
The Heine-Borel property, called also total compactness by some authors
(see [19]), is a familiar notion in topology and in the theory of metric spaces.
Definition 11(ii) is an adaptation of Definition 2.25 of [6] to metric spaces, given
here in order to notice that a subset Y of the metric space (X, d) is d-boundedly
compact if and only if Y has the Heine-Borel property with respect to d.
Theorem 12. Let K be a d-boundedly compact subset of the metric space (X, d)
and let x ∈ X be such that Pd,K (x) is a singleton. Then the mapping Pd,K is
d-continuous at x.
Proof. Suppose that Pd,K is not d-continuous at x. There exists a positive real
number ε such that, for each n ∈ ω \ {0}, the set
Fn = [
[
{Pd,K (y) : y ∈ Bd (x,
1
)}] \ Bd (pd,K (x), ε)
n
is nonempty. Let n ∈ ω \ {0} and y ∈ Bd (x, n1 ). For any z ∈ Pd,K (y), we have:
d(x, z) ≤ d(x, y) + d(y, z) ≤ n1 + d(y, K) ≤ n1 + d(y, x) + d(x, K) ≤ n2 + d(x, K).
This implies that the set Fn is d-bounded. Hence the set Cn = clX (Fn ) is dbounded. Then Cn ∩ K is compact because K is d-boundedly compact. Since
6
Fn ⊆ Cn ∩ K, we obtain that Cn ⊆ K. Of course, the collection {Cn : n ∈
ω \ {0}} is T
centred. Thus, it follows from the compactness of Cn that there
exists z0 ∈ n∈ω\{0} Cn . Then z0 ∈ K. Let g(t) = d(t, K) for t ∈ X. Let us
fix a positive real number η. Since the function g is continuous, there exists
n0 ∈ ω \ {0} such that n10 < ε and if t ∈ Bd (x, n10 ), then |g(t) − g(x)| < η.
Since z0 ∈ clX (Fn ), there exists z1 ∈ Fn such that d(z0 , z1 ) < n1 . There
exists y1 ∈ Bd (x, n1 ) such that z1 ∈ Pd,K (y1 ). Since z0 ∈ K, we have that
d(x, K) ≤ d(x, z0 ). On the other hand, d(x, z0 ) ≤ d(x, y1 )+d(y1 , z1 )+d(z1 , z0 ) ≤
2
2
n + d(y1 , K) ≤ n + d(x, K) + η. This, together with the arbitrariness of the
numbers η > 0 and n ≥ n0 , implies that d(x, z0 ) ≤ d(x, K). In consequence,
we obtain that d(x, z0 ) = d(x, K) = d(x, pd,K (x)), so z0 = pd,K (x). However,
ε ≤ d(pd,K (x), z1 ) ≤ d(pd,K (x), z0 ) + d(z0 , z1 ) ≤ d(pd,K (x), z0 ) + n1 and thus
d(pd,K (x), z0 ) > 0. The contradiction obtained completes the proof.
An application of our proof to Theorem 12 is that the following immediate
consequence of Theorem 12 holds true in ZF:
Corollary 13. If K is a d-boundedly compact d-Chebyshev set in X, then the
d-projection of X onto K is continuous.
2.2
Operations on d-Chebyshev sets
As in subsection 2.1, let us assume that (X, d) is a nonempty metric space.
Proposition
14. Let C be a nonempty collection of d-Chebyshev sets in X and
S
let K = C. Suppose that K is d-boundedly compact and has the property that,
for each pair x, y of points of K, there exists C ∈ C such that x, y ∈ C. Then
K is a d-Chebyshev set in X.
Proof. Let x0 ∈ X. Take any c ∈ K. Put r = d(x0 , c) and E = B̄d (x0 , r) ∩ K.
Then d(x0 , K) = d(x0 , E). Since E is nonempty and compact, there exists
c0 ∈ E such that d(x0 , c0 ) = d(x0 , E), so K is d-proximinal. Suppose that
c1 , c2 ∈ K and d(x0 , c1 ) = d(x0 , c2 ) = d(x0 , K). There exists C ∈ C such that
c1 , c2 ∈ C. Let i ∈ {1, 2} and let a = pd,C (x0 ). Then d(x0 , ci ) ≤ d(x0 , a) because
a ∈ K. Since ci ∈ C, we have d(x0 , a) ≤ d(x0 , ci ). Hence d(x0 , ci ) = d(x0 , a).
This implies that ci = a becauce Pd,C (x0 ) = {a}.
The following notions have appeared useful for investigations of Chebyshev
sets, especially in strictly convex metric spaces (see, for instance, [1], [19] and
[20]):
Definition 15. Let Y ⊆ X and let x, y ∈ Y . Then:
(i) (cf. Definition 2.1 of [20]) the d-segment in Y between points x and y is
the set [x, y]d,Y defined by:
[x, y]d,Y = {z ∈ Y : d(x, z) + d(z, y) = d(x, y)};
(ii) if Y = X, then [x, y]d = [x, y]d,Y ;
7
(iii) (cf. Definition 2.6 of [20]) a set A ⊆ Y is called d-convex in Y if, for each
pair of points x, y ∈ A, the inclusion [x, y]d,Y ⊆ A holds;
(iv) (cf. Definition 2.1 of [20]) a point z ∈ X is d-between x and y in Y if
z ∈ [x, y]d,Y .
The following proposition, relevant to Corollary 1 of [1], follows immediately
from Proposition 14:
Proposition 16. If all d-segments in X are d-Chebyshev sets, then every
nonempty d-boundedly compact d-convex set in X is d-Chebyshev in X.
Proposition 17. Suppose that A1 , A2 and A = A1 ∪ A2 are all d-Chebyshev
sets in X. Then, for each x ∈ X, we have:
pd,A1 (x) if d(x, pd,A1 (x)) ≤ d(x, pd,A2 (x))
pd,A (x) =
.
(1)
otherwise
pd,A2 (x)
Proof. Let x ∈ X. Denote a = pd,A (x) and ai = pd,Ai (x) for i ∈ {1, 2}. Of
course, d(x, a) ≤ d(x, ai ) for i ∈ {1, 2}. Suppose that a ∈ A1 . Then d(x, a1 ) ≤
d(x, a), so d(x, a) = d(x, a1 ). This implies that a = a1 and d(x, a1 ) ≤ d(x, a2 ).
If a ∈ A2 \ A1 , then a = a2 and d(x, a1 ) > d(x, a).
Using similar elementary arguments as in the proof to Proposition 17, one
can show that the following proposition holds:
Proposition 18. Let n ∈ ω \ {0}. Assume that,Sfor each i ∈ n, we are given
a d-Chebyshev set Ai in X such that the set A = i∈n Ai is d-Chebyshev in X.
Let x ∈ X and let a = pd,A (x). Then a = pd,Ai (x) for each i ∈ n such that
a ∈ Ai .
Now, we assume that n
Q∈ ω \ {0} and (Xi , di ) is a nonempty metric space
for each i ∈ n, while X = i∈n Xi . Usually, the equivalent metrics ρk for k ∈ 3
are considered in X where, for each x, y ∈ X, we have:
ρ0 (x, y) =
ρ1 (x, y) =
max{di (x(i), y(i)) : i ∈ n},
sX
di (x(i), y(i))2 ,
i∈n
ρ2 (x, y) =
X
di (x(i), y(i)).
i∈n
Since the theorem below will be of essential importance, we include its simple
proof for completeness:
Q
Theorem 19. Let Ai ⊆ Xi for each i ∈ n and let A = i∈n Ai , while k ∈ {1, 2}.
Then the following conditions are satisfied:
Q
(i) For each x ∈ X, the equality Pρk ,A (x) = i∈n Pdi ,Ai (x(i)) holds.
8
(ii) A is a ρk -Chebyshev set in X if and only if Ai is a di -Chebyshev set in
Xi for each i ∈ n.
Proof. To prove (i), assume first that a ∈ Pρk ,A (x). Let j ∈ n. Then a(j) ∈ Aj .
Consider any yj ∈ Aj and define y ∈ X as follows: y(i) = a(i) for each i ∈ n\{j},
while y(j) = yj . Since y ∈ A, we have ρk (x, a) ≤ ρk (x, y). This implies
thatQdj (x(j), a(j)) ≤ dj (x(j), yj ), so a(j) ∈ Pdj ,Aj (x(j)) and, in consequence,
a ∈ i∈n Pdi ,Ai (x(i)).
Q
Now, assume that b ∈ i∈n Pdi ,Ai (x(i)). Then b ∈ A. Moreover, if z ∈ A,
then, since di (x(i), b(i)) ≤ di (x(i), z(i)) for each i ∈ n, we have that ρk (x, b) ≤
ρk (x, z). Hence b ∈ Pρk ,A (x). This completes the proof to (i).
That (ii) also holds is an immediate consequence of (i).
The following example shows that Theorem 19 cannot be extended to the
metric ρ0 .
Example 20. Let A0 = A1 = [−1, 1]. For x, y ∈ R, let d0 (x, y) = d1 (x, y) =
|x − y|. Then, for x, y ∈ R2 , we have ρ0 (x, y) = max{|x(0) − y(0)|, |x(1) − y(1)|}.
Now, let x∗ = (0, 2). Then {0} = Pd0 ,A0 (0) and {1} = Pd1 ,A1 (2), while, for
A = A0 × A1 , the set Pρ0 ,A (x∗ ) is infinite because Pρ0 ,A (x∗ ) = [−1, 1] × {1}.
3
Calculations in dioids of interest
In this section, we recall the construction and calculations in algebraic structures
of our interest. Readers can compare this material with [21].
3.1
Calculations in MPA and MPAP
The max as the addition operation and the addition as the multiplication operation are denoted respectively by ⊕ and ⊗. The union R∪{ε} (where ε = −∞ ∈
/R
is a fixed element) is denoted by Rε and the semiring (Rε , ⊕, ⊗) is denoted by
Rmax . So, for any a, b ∈ Rε ,
a ⊕ b = max (a, b) and a ⊗ b = a + b.
(2)
The structure Rmax is a dioid (=an idempotent semiring) and a semifield (see
Definition 1 in [17]). Notice that Rnmax is naturally a semimodule over Rmax .
Analogously to conventional algebra, scalars in linear combinations are written before vectors. The rules for the order of evaluation of the tropical operations
are similar to those of conventional algebra, i.e. max-plus algebraic power has
the highest priority, and the tropical multiplication has a higher priority than
the tropical addition. All rules of calculation in Rmax can be illustrated in the
following example:
9
Example 21.
2 ⊕ 35 ⊕ 2−1 ⊗ 1 ⊕ ε2
=
=
=
=
=
=
2 ⊕ (5 × 3 ⊕ (−2)) ⊗ 1 ⊕ ε
2 ⊕ (15 ⊕ (−2)) ⊗ 1
2 ⊕ (15 ⊗ 1 ⊕ (−2) ⊗ 1)
2 ⊕ ((15 + 1) ⊕ (−2 + 1))
2 ⊕ (16 ⊕ (−1))
max (2, 16, −1) = 16
There is no inverse related to max operation, which makes a simple equation
a ⊕ x = b often insolvable. For this reason, we first consider the algebra of pairs
as a tool in the construction of the Symmetrized Max-Plus Algebra.
Let Pε = Rε × Rε with the operations ⊕ and ⊗ defined by operations in (1)
as follows:
(a, b) ⊕ (c, d) =
(a, b) ⊗ (c, d) =
(a ⊕ c, b ⊕ d)
(a ⊗ c ⊕ b ⊗ d, a ⊗ d ⊕ b ⊗ c)
(3)
(4)
It is easy to verify that (Pε , ⊕, ⊗) is a dioid with (ε, ε) and (0, ε) as its zero
and unit, respectively. Notice that (ε, ε) is the absorbing element for ⊗. This
structure is called the Max-Plus Algebra of Pairs (abbreviated to MPAP)
and denoted by Pmax . As Rmax is isomorphic to the subdioid of pairs (a, ε),
a ∈ Rε , then Rmax itself can be considered as a subdioid of Pmax . If u =
(x, y) ∈ Pε , then one defines the max-plus norm (or absolute value) of u
as |u|⊕ = x ⊕ y and we have two unary operators: ⊖ (the max-plus algebraic
minus operator) and (·)• (the balance operator) such that ⊖u = (y, x) and
u• = u⊕(⊖u) = (|u|⊕ , |u|⊕ ). The properties of these operators can be expressed
in the following proposition:
Proposition 22 (cf. [21]). For all u and v in Pε , we have:
•
•
(i) u• = (⊖u) = (u• ) ;
•
(ii) u ⊗ v • = (u ⊗ v) = u• ⊗ v = u• ⊗ v • ;
(iii) ⊖ (⊖u) = u, ⊖ (u ⊕ v) = (⊖u) ⊕ (⊖v) and
(⊖u) ⊗ v = ⊖ (u ⊗ v) = u ⊗ (⊖v).
Remark 23. The above can be interpreted: (·)• is a projection onto another
subdioid of Pε consisting of fixed points of ⊖ and ⊖ is an involution commuting
with ⊕.
We shall write u ⊖ v instead of u ⊕ (⊖v).
3.2
Construction of SMPA
For readers’ convenience, let us recall in brief the construction of SMPA described in [21].
10
Since u ⊖ u = (|u|⊕ , |u|⊕ ) = (ε, ε) ⇒ u = (ε, ε), then Pmax is, as Rmax ,
without inverses with respect to ⊕. For this reason, the balance relation ▽ is
introduced: For all u = (a, b) ∈ Pε and v = (c, d) ∈ Pε , we have
u ▽ v ⇔ a ⊕ d = b ⊕ c.
(5)
This relation is reflexive and symmetric, but not transitive because, for real
numbers x, y, z such that x > z > y, we have that
x = x ⊕ x = x ⊕ y 6= y ⊕ z,
which gives (x, y) ▽ (x, x) and (x, x) ▽ (z, x), but (x, y) is not balanced with
(z, x). So we need to modify the relation ▽ somewhat to have an associated
with it equivalence relation ∼ in Pε such that it is possible, in a standard
way, to extend the operations ⊕ and ⊗ of Pmax to the quotient set S = Pε /∼
to obtain the semiring Pmax /∼ which contains not inverses but symmetrized
elements which have some properties of inverses. Let the equivalence relation
∼ be defined as follows: for all u, v ∈ Pε with u = (a, b) and v = (c, d) where
a, b, c, d ∈ Rε , we have
u ▽ v if a 6= b and c 6= d
u∼v⇔
.
(6)
u = v
otherwise
The structure (S, ⊕, ⊗) = Pmax /∼ is called the Symmetrized Max-Plus Algebra (shortly: SMPA) and denoted by Smax . The semifield Rmax can be
identified with a subdioid of Smax , and Smax can be considered as a semimodule
over Rmax .
We have three types of equivalence classes of ∼ which are called max-plus
positive, max-plus negative and balanced elements, respectively. Namely, for
u ∈ Pε , let u = {v ∈ Pε : u ∼ v}. Then, for a ∈ R, we get
(a, ε) =
(ε, a) =
{(a, b) ∈ Pmax : a > b} ,
{(a, b) ∈ Pmax : a < b} .
(7)
(8)
Moreover, for each a ∈ Rε , we get
(a, a) = {(a, a)} .
(9)
If a ∈ R, the element ⊕a = (a, ε) of S is called max-plus positive, while the
element ⊖a = (ε, a) of S is called max-plus negative. If a ∈ Rε , the element
a• = (a, a) of S is called balanced. The element ε• is the zero element of
Smax . Let S⊕ be the set of all elements of S which are either max-plus positive
or equal to ε• . Let S⊖ be the set of all elements of S which are either max-plus
negative or equal to ε• . Finally, let S• be the set of all balanced elements of S.
The set S can be written as the union S⊕ ∪ S⊖ ∪ S• . The sets S⊕ , S⊖ and S•
have only one common point ε• . Let us put ⊕ε = ⊖ε = ε• . We can identify
S⊕ with Rmax by assigning ⊕a to a ∈ Rε . Similarly, S⊖ can be identified with
⊖Rmax = {⊖a : a ∈ Rε }, and S• can be identified with (Rmax )• = {a• : a ∈ Rε }.
11
Definition 24 (cf. [21]). Let a ∈ S. The max-plus positive part a⊕ and the
max-plus negative part a⊖ of a are defined as follows:
(i) if a ∈ S⊕ , then, a⊕ = a and a⊖ = ε;
(ii) if a ∈ S⊖ then a⊕ = ε and a⊖ = ⊖a;
(iii) if a ∈ S• , then there exists x ∈ Rε such that a = (x, x) and then a⊕ = a⊖
= x.
It is clear from the previous definition that, for each a ∈ S, the following
equality holds:
a = a⊕ ⊖ a⊖ .
(10)
This decomposition is unique if we allow only a = x ⊖ y (x, y ∈ S⊕ ) with either
x = ε or y = ε or x = y.
4
Topology and geometry in SMPA
4.1
Metrics on SMPA
In this paragraph, we introduce two distances on S.
Proposition 25 (cf. [21]). The formula
|a| = a⊕ ⊕ a⊖
(11)
defines an absolute value on Smax as a dioid (or a norm with Smax understood
as a semimodule over the semifield Rmax ).
Proof. Step 1). If |a| = ε, then
max a⊕ , a⊖ = −∞ ⇔ a⊕ = a⊖ = −∞ ⇔ a = ε.
.
Step 2). Naturally, we have
⊕
⊖
|a ⊗ b| = (a ⊗ b) ⊕ (a ⊗ b) .
We discuss three cases:
If a, b ∈ S⊕ , then
⊕
⊖
a = a⊕ , b = b⊕ ⇒ (a ⊗ b) = a⊕ ⊗ b⊕ , (a ⊗ b) = ε
and
|a| = a⊕ , |b| = b⊕ ,
which gives
|a ⊗ b| = a⊕ ⊗ b⊕ = |a| ⊗ |b| .
12
If a, b ∈ S⊖ , then a⊖ = ⊖a and a⊕ = ε , b⊖ = ⊖b and b⊕ = ε, which yields
⊕
(⊖a) ⊗ (⊖b) = (a ⊗ b) ,
then,
⊖
(a ⊗ b) = ε, |a| = ⊖a, |b| = ⊖b,
which gives
⊕
|a ⊗ b| = (a ⊗ b) = (⊖a) ⊗ (⊖b) = |a| ⊗ |b| .
If a ∈ S⊕ , b ∈ S⊖ , then a = a⊕ , b⊖ = ⊖b and b⊕ = ε, which yields
⊖
a ⊗ (⊖b) = ⊖ (a ⊗ b) = (a ⊗ b) ,
then,
which gives
(a ⊗ b)⊕ = ε, |a| = a, |b| = ⊖b,
|a ⊗ b| = (a ⊗ b)⊖ = a ⊗ (⊖b) = |a| ⊗ |b| .
So in all cases, the |·| is multiplicative.
Step 3). Similarly, we can prove the triangle inequality (equality, in fact).
Remark 26. The max-plus absolute value |·| does not produce a distance on S
in the similar way as the usual absolute value produces a distance on R because,
if we define ρ(x, y) = |x ⊖ y| for x, y ∈ S, then, for a ∈ S with a 6= ε• , we get
ρ (a, a) = |a ⊖ a| = |a• | = |a| 6= ε.
Corollary 27. Each a ∈ S can be written as
a = sgn(a) |a|
(12)
where sgn(a) ∈ {⊕, ⊖,• }. If a 6= ε• , then this formula is unique.
Corollary 28. One can write the formulas for actions in S:
a ⊕ b = sgn(a ⊕ b)( |a| ⊕ |b|),
a ⊗ b = sgn(a) sgn(b) (|a| ⊗ |b|)
sgn(a), |a| > |b|,
sgn(b), |b| > |a|,
where sgn(a ⊕ b) =
sgn(a), |a| = |b|, sgn(a) = sgn(b),
•,
|a| = |b|, sgn(a) 6= sgn(b),
and the multiplication table for signs is obvious.
(13)
(14)
Definition 29. The canonical embedding of S into C √is the mapping Φ :
S → C defined in the following way: for e−∞ = 0, θ = −1+2 3i and any r ∈ Rε ,
we put
Φ(⊕r) = θer , Φ(⊖r) = θ2 er , Φ(r• ) = er .
13
We denote the Euclidean metric in Rn by de,n or by de when it clear what n
is. In particular, for the plane C = R2 , we have de = de,2 . As we have already
mentioned, Rn will be usually considered with its Euclidean metric de and with
the natural topology induced by de .
Definition 30. The natural (or usual) topology of S is the smallest topology
in S which makes the mapping Φ : S → C continuous.
We shall consider two natural metrics in S that are defined below:
Definition 31. Let d1 : S × S−→R be defined as follows:
d1 (a, b) = de (Φ(a), Φ(b))
for all a, b ∈ S. The metric d1 in S will be called the Euclidean distance in S.
Definition 32. Let d2 : S × S−→R be defined as follows:
|exp |a| − exp |b|| if sgn(a) = sgn(b)
d2 (a, b) =
exp |a| + exp |b| if sgn(a) 6= sgn(b)
for all a, b ∈ S. The metric d2 in S will be called the inner distance in S.
Remark 33. It is clear that both Euclidean and inner distances in S are equivalent and they induce the natural topology of S; moreover, the metrics d1 and
d2 are uniformly equivalent and equal on each of the lines S⊕ , S⊖ , S• . As a
topological space, we shall consider S only with its natural topology. The space
S is homeomorphic with the subspace Φ(S) of C = R2 . The mapping Φ is an
isometry of the metric space (S, d1 ) onto the metric subspace Φ(S) of (C, de ).
The metric space (S, d2 ) is isometric with:
1) the metric subspace (R×{0})∪({0}×[0, +∞)) of the metric space (R2 , ρ2 )
where ρ2 (x, y) = |x(0) − y(0)| + |x(1) − y(1)| for all x, y ∈ R2 ;
2) the metric subspace {z ∈ R2 : z(1) = |z(0)| ∨ (z(0) = 0 ∧ z(1) < 0} of the
metric space (R2 , ρ0 ) where ρ0 (x, y) = max(|x(0) − y(0)|, |x(1) − y(1)|) for all
x, y ∈ R2 .
Let X be a topological space. We recall that it is said that a point c ∈ X
disconnects X between points a, b ∈ X if there exists a pair U, V of disjoint
open sets in X such that X \ {c} = U ∪ V and a ∈ U , while b ∈ V . The space X
is a dendrite if it is a continuum such that, for each pair a, b of distinct points
of X, there exists a point c ∈ X which disconnects X between points a and b. It
is said that X is a semicontinuum if, for each pair a, b of points of X, there is
a continuum C in X such that a, b ∈ C. Let us say that X is a semidendrite if
it is a semicontinuum such that, for each pair a, b of distinct points of X, there
exists a point c ∈ X which disconnects X between a and b.
Proposition 34. The space S is a semidendrite.
Proof. It suffices to notice that S is homeomorphic with the semidendrite Φ(S)
on the plane C.
14
Basic properties of dendrites are nicely described in [2] and [18]. Some
of the properties can be easily proved in ZF and they are also properties of
semidendrites. For example, that every semidendrite is a Hausdorff space and
that every semicontinuum contained in a semidendrite is a semidendrite are
statements obviously true in ZF. However, the axiom of choice is heavily used,
for instance, in an adaptation to metrizable semidendrites of the known proof
for metrizable dendrites of the following theorem:
Theorem 35. It holds true in ZFC that, for each pair a, b of distinct points of
a metrizable semidendrite X, there exists in X exactly one arc with end-points
a and b.
Problem 36. Can ZFC be replaced by ZF in Theorem 35?
4.2
Continuity of the inner operations
In this paragraph, we study the continuity and the discontinuity of the inner
operations in Smax with respect to the usual topology in S. The following
proposition is obvious:
Proposition 37. The operation ⊗ is continuous in the usual topology of S.
Proposition 38. The operation ⊕ is discontinuous in the usual topology of S.
Proof. For r ∈ R, we have discontinuity of ⊕ : S × S → S at (⊕r, ⊖r), (⊖r, ⊕r),
(⊕r, r• ), (r• , ⊕r), (⊖r, r• ), and (r• , ⊖r). For example:
lims→r− ⊕(⊕s, r• ) = r• ,
lims→r+ ⊕(⊕s, r• ) = ⊕r.
Remark 39. Let τ be the quotient topology induced by the mapping | · | : S →
Rε where Rε is equipped with its standard topology. The topology τ is not T0 ;
however, both operations ⊕ and ⊗ from (S, τ ) × (S, τ ) to (S, τ ) are continuous.
5
Topology and geometry of finite products of
SMPA
Let n ∈ ω \ {0}. Then the set Sn of all functions x : n → S is identified with
the n-Cartesian product of S. In particular, S1 is identified with S. The usual
or natural topology of Sn is the product topology of the space Sn where S is
equipped with its usual topology (see Definition 30).
5.1
The main metrics in Sn
We consider the Euclidean metric d1 and the inner metric d2 in S. For j ∈ {1, 2},
it is most natural to associate with the distance function dj the following metrics
15
ρ0,j , ρ1,j and ρ2,j in Sn :
ρ0,j (x, y)
ρ1,j (x, y)
= max{dj (x(i), y(i)) : i ∈ n},
sX
dj (x(i), y(i))2 ,
=
i∈n
ρ2,j (x, y)
=
X
dj (x(i), y(i))
i∈n
where x, y ∈ Sn . Of course, the metrics ρk,j are all equivalent and they induce
the natural topology of Sn . We shall pay a special attention to the metrics ρ1,1
and ρ1,2 .
Definition 40. The metric ρ1,1 will be denoted by D1 and called the Euclidean
metric in Sn , while the metric ρ1,2 will be denoted by D2 and called the inner
metric in Sn .
Definition 41. For the
Q canonical embedding Φ of S into C, let Φi = Φ where i ∈
n. Then the product i∈n Φi will be denoted by [n Φ] and called the canonical
embedding of Sn into Cn .
Remark 42. The mapping [n Φ] is an isometry of the metric space (Sn , D1 )
onto the metric subspace [n Φ](Sn ) of (Cn , de,2n ) where the set Cn is identified
with R2n in an obvious way.
Remark 43. For x, y ∈ Sn , let Lx,y be the shortest broken line in Cn which
connects [n Φ](x) with [n Φ](y) and is contained in [n Φ](Sn ). Then D2 (x, y) is
the length in Cn of the curve Lx,y , while
[x, y]D2 = [n Φ]−1 (Lx,y ).
5.2
Geometric segments as broken lines
For x, y ∈ Rn (or x, y ∈ Cn , resp.), let [x, y] = {(1 − t)x + ty : 0 ≤ t ≤ 1} be the
traditional segment with end-points x, y.
Definition 44. Let x, y ∈ Sn . Then:
(i) the D2 -segment in Sn between x and y will be denoted by [x, y]g and called
the geometric segment between x and y;
(ii) if the traditional segment [[n Φ](x), [n Φ](y)] in Cn is a subset of [n Φ](Sn ),
then [n Φ]−1 ([[n Φ](x), [n Φ](y)]) will be called the traditional segment in
Sn between x and y.
For an arbitrary geometric segment [a, b]g , let us demonstrate a method of
finding aSpossibly small finite collection C of traditional segments such that
[a, b]g = C.
16
Definition 45. For distinct u, v ∈ {⊕, ⊖, •}, we define an embedding Ψ(u,v) :
Su ∪ Sv → R as follows: for each r ∈ R, we put •r = r• , Ψ(u,v) (ur) = er , while
Ψ(u,v) (vr) = −er . We put Ψ(u,v) (ε• ) = 0. Moreover, let Ψ(u,u) : Su → [0; +∞)
be defined by the equality Ψ(u,u) (x) = Ψ(u,v) (x) for each x ∈ Su and each
v ∈ {⊕, ⊖, •} \ {u}.
Now, let us fix n ∈ ω \ {0} and let a, b ∈ Sn be given. Assume that a 6= b.
For each j ∈ n we choose an ordered pair (u(j), Q
v(j)) ∈ {⊕, ⊖, •} × {⊕, ⊖, •}
such that a(j) ∈ S u(j) and b(j) ∈ S v(j) . Let Ψ = j∈n Ψ(u(j),v(j)) . To simplify
notation, put Ψj = Ψ(u(j),v(j)) for j ∈ n. Consider the traditional segment
[Ψ(a), Ψ(b)] in Rn . Let J be the set of all j ∈ n which have the property that
there exists a unique tj ∈ [0, 1] such that (1 − tj )Ψj (a(j)) + tj Ψj (b(j)) = 0
and let K ⊆ J be a maximal subset of J such that, for all l, p ∈ K, tl 6= tp if
l 6= p. Suppose that the set K is nonempty and it has exactly k elements. Let
j0 , . . . , jk−1 be the injective sequence of all elements of K such that tjl ≤ tjp if
l, p ∈ k and l ≤ p. Put xp =S(1 − tjp )Ψ(a) + tjp Ψ(b) for each p ∈ K. We get the
broken line L = [Ψ(a), x0 ]∪ p∈k−1 [xp , xp+1 ]∪[xk−1 , Ψ(b)]. Then Ψ−1 (L) is the
geometric segment [a, b]g , while the length |L| of L is equal to D2 (a, b). Now,
it is easily seen that every geometric segment in Sn is a union of at most n + 1
traditional segments. The following example shows how to use this method in
practice.
Example 46. Let us consider points a, b ∈ S3 where a = (⊕0, ⊖ln3, [ln2]• )
and b = (⊖0, 0• , ⊕0). To follow the method described above,Qwe put u(0) =
⊕, v(0) = ⊖, u(1) = ⊖, v(1) = •, u(2) = •, v(2) = ⊕ and Ψ = j∈3 Ψ(u(j),v(j)) .
For t ∈ [0, 1], we have (1 − t)Ψ(a) + tΨ(b) = (1 − 2t, 3 − 4t, 2 − 3t). Then
K = {0, 1, 2}, while t0 = 12 , t1 = 34 , t2 = 23 . In consequence, x0 = (0, 1, 21 ), x1 =
(− 13 , 31 , 0), x2 = (− 21 , 0, − 41 ), For the broken line L = [Ψ(a), x0 ] ∪ [x0 , x1 ] ∪
√
[x1 , x2 ] ∪ [x2 , Ψ(b)], we have |L| = D2 (a, b) = 29. In this case, [a, b]g is a union
of four traditional segments in S3 .
Remark 47. Let u, v ∈ {⊕, ⊖, •} be distinct. Suppose that a ∈ Su \ {ε• }, while
b ∈ Sv \ {ε• }. Then [a, b]d1 = {a, b}.
6
Completeness in Sn
According to [11] and [15], to consider completeness in metric spaces in the
absence of the axiom of choice, it is necessary to make a distinction between the
notions of metric completeness and Cantor completeness.
6.1
Completeness and Cantor completeness of metrics
To make our terminology clear, we recall the following definitions with an additional notation which is more suitable to our aims:
Definition 48. Let (X, d) be a metric space and let Y ⊆ X. It is said that:
17
(i) the set Y is complete in (X, d) or, equivalently, d-complete in X if each
d-Cauchy sequence of points of Y converges in (X, d) to a point from Y ;
(ii) the set Y is Cantor complete in (X, d) or, equivalently, Y is Cantor dcomplete in X if, for each sequence (Fn ) of nonempty d-closed in Y sets
such thatTFn+1 ⊆ Fn for each n ∈ ω \ {0} and limn→+∞ diamd (Fn ) = 0,
we have n∈ω\{0} Fn 6= ∅ where diamd (Fn ) denotes the diameter of Fn in
the metric space (X, d).
All Cantor complete subspaces of the metric space (X, d) are closed in (X, d)
but, in some models of ZF, complete subsets of metric spaces need not be closed
(see [11], Proposition 6 of [15] and Theorem 4.55 of [12]).
Let us recall that CC is the axiom of countable choice which states that
every nonempty countable collection of nonempty sets has a choice function
(see Definition 2.5 of [12] and Form 8 in [14]). In the light of Theorem 7 of [15],
that every complete metric space is Cantor complete is an equivalent of CC.
6.2
Completeness in S
Theorem 49. Let i ∈ {1, 2}. Then (S, di ) is a Cantor complete metric space.
Therefore, all closed subspaces of S are di -complete. That all di -complete subspaces of S are closed in S is an equivalent of CC(cR).
Proof. The subspace Φ(S) of the metric space (C, de ) is Cantor complete. This,
together with Remark 33, implies that the metric spaces (S, d1 ) and (S, d2 ) are
Cantor complete. Now, suppose that CC(cR) fails. Then, by Theorem 4.55 of
[12], R is not sequential. Let X be a subset of R+ such that X is sequentially
closed but not closed in R. Let us consider X as a metric subspace of the
complete metric space (R, d) where d(x, y) = |x − y| for x, y ∈ R. Then the
metric d is complete on X. Since the metric space X is isometric with a metric
subspace of the metric subspace Φ(S⊕ ) of the semidendrite Φ(S), it follows that,
for i ∈ {1, 2}, not all complete subspaces of (S, di ) are closed in S.
Finally, let us assume that CC(cR) holds. Then, by Theorem 4.55 of [12],
all complete subspaces of R are closed in R. This implies that all complete
subspaces of (S, di ) are closed in S where i ∈ {1, 2}.
Remark 50. For i ∈ {1, 2}, it is independent of ZF that all di -complete subspaces of S are closed in S. In the light of Remark 7, for instance, in model M1
of [14], there exist metric subspaces of (S, di ) that are simultaneously complete
and not closed.
Remark 51. There is a mistake in condition 7 of Theorem 4.54 of [12]. Namely,
condition 7 of Theorem 4.54 on page 74 of [12] is the following statement:
(*) Each second countable topological space is separable.
It is worth to notice that (*) is an equivalent of CC. To show that (*) implies
CC, one can use the method of the proof to Theorem 4.64 in [12]. It is well
known by topologists and set theorists that CC implies (*). Condition 7 of
Theorem 4.54 of [12] should be replaced by the following statement:
18
Every second countable T0 -space is separable.
We know from a private communication of S. da Silva with E. Wajch that Horst
Herrlich was aware of his mistake after [12] had appeared in print.
6.3
Completeness of the main metrics of Sn
One can deduce from Theorem 49 and from the definitions of ρk,i (see subsection
5.1) that the following theorem holds:
Theorem 52. Let i ∈ {1, 2} and k ∈ 3. Then, for each n ∈ ω \ {0}, the
metric space (Sn , ρk,i ) is Cantor complete, so all closed subspaces of Sn are
ρk,i -complete.
Problem 53. Let i ∈ {1, 2}, k ∈ 3. Consider any n ∈ ω \ {0} with n > 1. Does
CC(cR) imply in ZF that all ρk,i -complete subspaces of Sn are closed in Sn ?
A satisfactory solution to Problem 53 is unknown to us and it may be complicated. However, the following proposition follows easily from known facts:
Proposition 54. Let i ∈ {1, 2}, k ∈ 3 and let n ∈ ω \ {0}. If CC(R) holds,
then each ρk,i - complete subspace of Sn is closed in Sn .
Proof. Suppose that CC(R) holds. Let X ⊆ Sn . In view of Theorem 4.54 of
[12], the subspace X of Sn is separable. If X is not closed in Sn , it follows from
the separability of X and from the first-countability of Sn that there exists a
sequence of points of X which converges in Sn to a point from Sn \ X. Then X
is not ρk,i -complete.
Definition 55. For a positive integer n, let us say that a set A ⊆ Rn is well
placed if, for each open subset V of Rn such that V ∩ A 6= ∅, there exist
p, q ∈ V ∩ Qn such that [p, q] ∩ A 6= ∅ where [p, q] is the standard segment in Rn
with end-points p and q.
The following theorem leads to a partial solution to Problem 53:
Theorem 56. For each positive integer n > 1, the following conditions are all
equivalent in ZF:
(i) CC(cR);
(ii) each well placed complete subset of Rn is closed in Rn ;
(iii) each well placed, simultaneously connected and complete subset of Rn is
closed in Rn ;
(iv) each well placed, simultaneously connected and complete subspace of the
plane R2 is closed in R2 .
19
Proof. It is clear that the implications (ii) ⇒ (iii) ⇒ (iv) hold. Let us assume
(i) and use some ideas of the proof that (7) implies (5) in Theorem 4.55 in [12].
Consider a well placed, nonempty complete subset X of Rn . Let J = {(p, q) ∈
Qn × Qn : [p, q] ∩ X 6= ∅}. For each pair (p, q) ∈ J, let Y (p, q) = {t ∈ [0, 1] :
(1 − t)p + tq ∈ X}. Then Y (p, q) is a completeQsubspace of R. Under the
assumption that CC(cR) holds, there exists f ∈ (p,q)∈J Y (p, q). Since X is
well placed, the set {f (p, q) : (p, q) ∈ J} is dense in X. Hence X is separable.
Since every separable complete subset of a metric space is closed, the set X is
closed in Rn . In consequence, (i) ⇒ (ii).
Now, suppose that CC(cR) fails. In view of Theorem 4.55 of [12], there
exists a sequentially closed subset A of R such that A is not closed in R. Let
I = [0, 1] be the unit interval of R and let X = (A × I) ∪ (R × {1}). Of course
X is well placed and connected in R2 . Let (zn )n∈ω\{0} be a Cauchy sequence in
X and let xn , yn be real numbers such that zn is the ordered pair (xn , yn ) for
n ∈ ω \ {0}. Put C = {a ∈ A : ({a} × I) ∩ {zn : n ∈ ω \ {0}} =
6 ∅}. Suppose
that C is infinite. Then there exists an injective subsequence (znk ) of (zn ) such
that xnk ∈ A. The sequence (xnk ) is a Cauchy sequence of points of A. Let
x0 be the limit of (xnk ) in R. Since A is sequentially closed, we have x0 ∈ A.
Suppose that the segment {x0 } × I does not contain accumulation points of the
sequence (znk ). Then, by the compactness of {x0 } × I, there exists an open set
V in R2 such that ({x0 } × I) ⊆ V and V ∩ {znk : k ∈ ω \ {0}} = ∅. This is
impossible because x0 is the limit of (xnk ). The contradiction obtained shows
that the sequence (znk ) has an accumulation point z0 ∈ {x0 } × I. Then z0
is the limit of (zn )Sin X. Now, suppose that C is finite. Then the subspace
Y = (R × {1}) ∪ a∈C ({a} × I) is closed in R2 and, moreover, there exists
n0 ∈ ω \ {0} such that zn ∈ Y for each natural number n > n0 . Since Y is
complete and closed in R2 , the limit of (zn ) in R2 belongs to Y . In consequence,
X is complete. Of course, X is not closed in R2 .
Corollary 57. Let n > 1 be a positive integer. Let i ∈ {1, 2} and k ∈ 3. If all
simultaneously ρk,i - complete and connected subspaces of Sn are closed in Sn ,
then CC(cR) holds.
Problem 53 can be regarded as equivalent to the following one:
Problem 58. Let n ∈ ω \ {0} and n > 1. If M is a model of ZF such that
CC(cR) holds in M, is it true in M that every complete subspace of Rn is closed
in Rn ?
7
Product semimodules from SMPA
Let n ∈ ω \ {0}. For all x, y ∈ Sn , one can define x ⊕ y ∈ Sn and x ⊗ y ∈ Sn in
an obvious way:
(x ⊕ y)(i) =
x(i) ⊕ y(i)
(15)
(x ⊗ y)(i) =
x(i) ⊗ y(i)
(16)
20
where, for each i ∈ n, x(i) ⊕ y(i) and x(i) ⊗ y(i) are defined as in Corollary 28.
Then (Sn , ⊕, ⊗) is a semiring.
7.1
S as a semimodule
By a scalar we mean an element of Rε . For a ∈ S, λ ∈ Rε , we define
λ ⊗ a = sgn(a)(λ + |a|).
(17)
The operation ⊗ : Rε × S → S is well defined, and we can consider it as an outer
operation from Rε × S to S or the restriction to Rε × S of the inner operation
⊗ in Smax because Rmax ⊂ Smax . Considering ⊗ as the outer operation, we can
look at Smax as at a semimodule (“semi-vector-space”) over the semifield Rmax .
7.2
Sn as a semimodule
Let n ∈ ω \ {0}. We consider Sn as the product semimodule with the inner
operation ⊕ defined by equation (15) and the outer operation ⊗ : Rε × Sn → Sn
defined as follows:
(λ ⊗ a)(i) = λ ⊗ a(i),
(18)
where λ ∈ Rε , a ∈ Sn and i ∈ n, while λ ⊗ a(i) is as in (17).
We deduce the following two corollaries from Propositions 37 and 38, respectively:
Corollary 59. The operation ⊗ : Sn × Sn → Sn is continuous in the usual
topology of Sn .
Corollary 60. The operation ⊕ : Sn × Sn → Sn is discontinuous in the usual
topology of Sn .
8
Convexities in Sn
If, for each pair of points x, y of a space X, a segment [x, y] with end-points x and
y is defined, then a set A ⊆ X is called convex if, for all x, y ∈ A, we have [x, y] ⊆
A (see, for instance, [13]). In particular, for x, y ∈ Sn , we can consider distinct
kinds of segments with end-points x and y, among them, traditional segments,
geometric segments and semimodule segments. These kinds of segments in Sn
lead to distinct kinds of convexity in Sn . Let us turn our attention to traditional,
semimodule and geometric convexity in Sn . One can also investigate ρk,j -convex
sets in Sn for each k ∈ 3 and j ∈ {1, 2}.
21
8.1
Traditionally convex sets in Sn
Let n ∈ ω \ {0}. A convex set in Cn is a set A ⊆ Cn such that, for each pair
x, y ∈ A, the traditional segment [x, y] in the affine space Cn is contained in A.
Definition 61. A set C ⊆ Sn will be called traditionally convex if [n Φ](C)
is convex in Cn .
Remark 62. Let us notice that a nonempty set C ⊆ S is simultaneously compact and traditionally convex in S if and only if C is a traditional segment in
S.
All traditionally convex subsets of Sn are connected.
8.2
Semimodule convex sets in Sn
The semimodule structure of Sn allows us to think about semimodule convexity
of subsets of Sn . In analogy to conventional algebra, we define a semimodule
segment [a, b]sm with end-points a ∈ Sn and b ∈ Sn as follows:
[a, b]sm = {(λ ⊗ a) ⊕ (γ ⊗ b) : λ, γ ∈ Rε , with λ ⊕ γ = 0}.
(19)
n
Definition 63. A set A ⊆ S is said to be semimodule convex if, for each
pair a, b of points of A, the semimodule segment [a, b]sm is contained in A.
Convexity in arbitrary idempotent semimodules was introduced by Cohen,
Gaubert and Quadrat [3]. Some analogues of famous theorems of functional
analysis like Minkowski’s theorem, separation theorem, and some notions of geometry like simplices and convex polytopes were considered in Max-Plus Algebra
(for instance, one can refer to [4], [7], [8]).
Remark 64. We have the following semimodule segments in S for r, s ∈ R+
with r < s:
[⊕r, ⊖r]sm = {⊕r, ⊖r, r• },
(20)
[⊕r, r• ]sm = {⊕r, r• },
(21)
[⊖r, ⊕s]sm = {⊖r, r• } ∪ {⊕t : r < t ≤ s}
(22)
[r• , ⊕s]sm = {r• } ∪ {⊕t : r < t ≤ s}
(23)
[⊖r, s• ]sm = {⊖r} ∪ {t• : r ≤ t ≤ s}
(24)
Other cases of semimodule segments in S are similar or obvious.
Notice that the intersection of two semimodule segments may not be a semimodule segment. Also, if [a, b]sm and [c, d]sm are semimodule segments in S
such that [a, b]sm ∩ [c, d]sm 6= ∅, then, in general, the union [a, b]sm ∪ [c, d]sm
need not be a semimodule segment. However, if c ∈ [a, b]sm , then [a, b]sm =
[a, c]sm ∪ [c, b]sm . All disconnected semimodule segments in S have either two
or three connected components.
22
Example 65. Consider the semimodule segment in S between a = ⊕1 and
b = ⊖0. Then [a, b]sm = {b} ∪ {0• } ∪ {⊕t : t ∈ (0, 1]}, so the segment [a, b]sm
has three connected components. Moreover, this segment is not closed in S. For
i ∈ {1, 2}, both ⊖0 and 0• are di -nearest points to ε• in [a, b]sm . More precisely,
Pdi ,[a,b]sm (ε• ) = {⊖0, 0• } for i ∈ {1, 2}.
Example 66. A semimodule segment in S2 can have five connected components. For example, consider
C = [(⊕0, ⊖1), (⊖1, ⊕0)]sm = {(⊕0, ⊖1), (⊖1, ⊕0), (⊖1, 0•), (0• , ⊖1)} ∪
∪{(⊖t, ⊖s) : max(t, s) = 1, t, s > 0.}
Then C has exactly four isolated points and exactly one connected component
of C is of the form [2 Φ]−1 (L) where L ⊆ C is a broken line with its end-points
deleted.
8.3
Geometric convexity in Sn
Definition 67. A set A ⊆ Sn is said to be geometrically convex if, for each
pair a, b of points of A, the geometric segment [a, b]g is contained in A.
The notions of a geometrically convex set and a D2 -convex set in Sn coincide.
Of course, each traditionally convex subset of Sn is geometrically convex but
not every geometrically convex set in Sn is traditionally convex. Since geometric
segments in Sn are connected sets, each geometrically convex subset of Sn is
connected. In the case of S, we can state the following propositions:
Proposition 68. The geometric segment [a, b]g between points a, b ∈ S is the
intersection of all connected subsets of S that contain both a and b.
Proposition 69. A set A ⊆ S is geometrically convex if and only if A is
connected.
The proofs to Propositions 68 and 69 are so easy that we omit them. Of
course, if n > 1 then, for distinct a, b ∈ Sn , the intersection of all connected
subsets of Sn that contain both a and b is the two-point set {a, b}.
Let us recall that, if X is a topological space, a, b ∈ X, while C is a continuum
in X such that a, b ∈ C, then C is called irreducible between a and b if, for each
continuum D ⊆ C such that a, b ∈ D, the equality D = C holds. The axiom of
choice is used in the standard proof of the theorem of ZFC asserting that, for
each pair of distinct points of a Hausdorff continuum X there exists a continuum
C in X such that C is irreducible between a and b.
Problem 70. Is there a model M of ZF such that there exists a Hausdorff
continuum X in M which has the property that, for a pair a, b of distinct points
of X, there does not exist in M a subcontinuum C of X such that C is irreducible
between a and b?
23
Remark 71. Suppose that we work inside a model M of ZFC. Let X be a
semidendrite in M and let a, b ∈ X. We can generalize properties of dendrites
given on page 125 in [18] to semidendrites. Namely, we can prove that there
exists in M a unique continuum C ⊆ X such that a, b ∈ C and C is irreducible
between a and b. Let us denote this continuum C by [a, b]c−g and call it the
c-geometric segment of X between a and b. Now, assume, in addition, that
X is metrizable in M and that a 6= b. Then [a, b]c−g is an arc with end-points
a and b (see Lecture VIII of [18]). Notice that [a, b]c−g is the intersection of all
connected subspaces of X that contain both a and b, so [a, b]c−g is the smallest
(with respect to inclusion) among all connected subsets of X that contain both
a and b. Let us call a set A ⊆ X c-geometrically convex in X if, for each
pair a, b of distinct points of A, the inclusion [a, b]c−g ⊆ A holds. Then a set
A ⊆ X is connected if and only if it is c-geometrically convex in X. If X = S,
then M can be any model of ZF.
Remark 72. For a pair a, b of points of R2 , let C(a, b) be the collection of
all connected subsets of R2 which contain both a and b. Then each continuum
C ∈ C(a, b) such that C is irreducible between
a and b is a minimal (with respect
T
to inclusion) element of C(a, b). Since {C : C ∈ C(a, b)} = {a,
T b}, to define a cgeometric segment in R2 between a and b as the intersection {C : C ∈ C(a, b)}
does not lead to anything reasonable. However, it may be still interesting to
investigate in deep all topological spaces X which have the property that, for
each pair a, b of points of X, the intersection of all connected subsets of X that
contain both a and b is connected in X.
Remark 73. Let [a, b]g and [c, d]g be geometric segments in S. Then the
intersection [a, b]g ∩ [c, d]g is a geometric segment in S. Also, if c ∈ [a, b]g ,
then [a, b]g = [a, c]g ∪ [c, b]g . However, if [a, b]g ∩ [c, d]g 6= ∅, then [a, b]g ∪ [c, b]g
is a dendrite which need not be a geometric segment in S.
9
The minimizing vector theorem in Sn
In what follows, all Hilbert spaces are assumed to be over the field R. Given
a Hilbert space H, we consider H as a metric space equipped with the metric
ρH of H which is induced by the inner product of H, i. e. ρH (x, y) = kx − yk
where k·k is the norm defined by the inner product of H. The minimizing vector
theorem, also called the Hilbert projection theorem, is a famous result in convex
analysis saying that it holds true in ZF + CC that, for every point x ∈ H in a
Hilbert space H and every nonempty closed convex subset C of H, there exists a
unique point y ∈ C for which kx − yk is the distance of x from C. In particular,
this is true in every model M of ZF + CC for any closed affine subspace C of
H provided H is in M and we work on H inside M. In the case when C is a
closed affine subspace of the H, a necessary and sufficient condition for y ∈ C
to be the nearest point to x in C is that the vector x − y be orthogonal to C.
A result of ZF is the following version of the vector minimizing theorem:
24
Theorem 74. For every Hilbert space H, for each nonempty, simultaneously
convex and Cantor complete subset C of H and for each x ∈ H, there exists a
unique y ∈ C such that kx − yk is the distance of x from C.
As we have already mentioned in subsection 6.1, according to Theorem 7 of
[15], in a model of ZF, a complete subset of a metric space need not be Cantor
complete. Therefore, one may ask whether it holds true in every model M of
ZF that all simultaneously complete and convex subsets of a Hilbert space H
in M are closed in H. A partial answer to this question is given in the following
proposition:
Proposition 75. Let H be a finitely dimensional Hilbert space and let C be a
nonempty convex complete subset of H. Then C is closed in H.
Proof. Let Y be the intersection of all affine subspaces of H that contain C.
Since H is finitely dimensional, it is known that intY (C) 6= ∅. Let a ∈ intY (C)
and b ∈ clY (C). Then the interval I = [a, b] = {(1 − t)a + tb : t ∈ R, 0 ≤ t ≤ 1}
is contained in clY (C), while I \ {b} is a subset of intY (C). Of course, there
exists a sequence (xn ) of points of I \ {b} which converges in H to b. Since
(xn ) is a Cauchy sequence of points of C, we have b ∈ C because C is complete.
Hence C = clY (C) = clH (C).
At this moment, we do not know a satisfactory solution to the following
problem:
Problem 76. Is is true in ZF that, for an arbitrary infinitely dimensional
Hilbert space H, all simultaneously convex and complete subsets of H are Cantor
complete in H?
We recommend [6] as one of the newest survey articles on the minimizing
vector theorem. We are going to state some simple versions of this theorem for
Sn .
9.1
A characterization of Chebyshev sets in S
For i ∈ {1, 2}, the following theorem gives a complete characterization of di Chebyshev sets in S:
Theorem 77. For each nonempty closed subset C of S, the following conditions
are equivalent:
(i) C is connected;
(ii) C is a d1 -Chebyshev set;
(iii) C is a d2 -Chebyshev set.
Proof. Assume C is connected and x ∈ S. Consider the following two cases (a)
and (b):
25
(a) ε• ∈
/ C. Then C is contained in one of the lines S⊕ , S⊖ , S• , where
connnectedness is equivalent with traditional convexity, so Pd1 ,C (x) = Pd2 ,C (x)
and Pd1 ,C (x) is a singleton.
(b) ε• ∈ C. Then C is star-like; hence, it is easily seen that Pd1 ,C (x) =
Pd2 ,C (x) and Pd1 ,C (x) is a singleton which lies on the same line as x.
In consequence, (i) implies both (ii) and (iii). On the other hand, if C were
disconnected, there would exist a connected component K of S \ C, distinct
points a1 , a2 ∈ C ∩ clS (K) and points y1 , y2 ∈ K such that a1 , a2 ∈ Pdi ,C (yi ) for
i ∈ {1, 2}. Therefore, each of (ii) and (iii) implies (i).
Corollary 78. If C is a nonempty, simultaneously connected and closed subset
of S, then pd1 ,C = pd2 ,C .
Proposition 79. For each nonempty closed set C ⊆ S, that C is geometrically
convex in S is equivalent to each one of conditions (i) − (iii) of Theorem 76.
Let us have a brief look at semimodule convex closed subsets of S.
Example 80. The semimodule segment A = [⊕0, ⊖0]sm = {⊕0, ⊖0, 0•} is a
closed, semimodule convex set in S with three connected components. For each
i ∈ {1, 2}, we have Pdi ,A (ε• ) = A.
Proposition 81. Let C be a nonempty, simultaneously semimodule convex and
closed subset of S. Then C is di -proximinal and, for each x ∈ S, the set Pdi ,C (x)
consists of at most three points.
Proof. This is clear since, on each of the lines S⊕ , S⊖ , S• , semimodule convexity
agrees with the traditional convexity.
9.2
Chebyshev sets in Sn
n
Definition 82. A set
Q A ⊆ S is said to be box semimodule convex if it is
a Cartesian product i∈n Ai of semimodule convex subsets Ai of S.
We can generalize Proposition 81 as follows:
Proposition 83. Let j, k ∈ {1, 2}. Suppose that A is a nonempty, closed, box
semimodule convex subset of Sn . Then A is ρk,j -proximinal and, moreover, for
each x ∈ Sn , the set Pρk,j ,A (x) consists of at most 3n points.
Proof. Since closed balls in (Sn , ρk,j ) are compact, each nonempty closed set
in Sn is ρk,j -proximinal. That Pρk,j ,A (x) consists of at most 3n points follows
directly from Theorem 19 (i) and from Proposition 81.
Example 84. In view of Example 80, the set A = {⊕0, ⊖0, 0•}n in Sn is box
semimodule convex. Let x0 ∈ Sn be defined by: x0 (i) = ε• for each i ∈ n.
Then, if k ∈ 3 and j ∈ {1, 2}, the set Pρk,j ,A (x0 ) consists of exactly 3n points.
Q
Theorem 85. Let j, k ∈ {1, 2}. Suppose that A = i∈n Ai where each Ai is
nonempty and closed in S. Then A is ρk,j -Chebyshev in Sn if and only if A is
connected.
26
Proof. It suffices to apply Theorems 19(ii) and 77.
The following problem is still unsolved:
Problem 86. Let i ∈ {1, 2} and A be a nonempty closed subset of Sn where
n > 1. Is it true that A is Di -Chebyshev if and only if A is geometrically convex
in Sn ?
Let us leave Problem 86 for future investigations in another work.
Acknowledgement. This research was supported by King Abdulaziz University. The article is a part of research project initiated by Hanifa Zekraoui.
References
[1] G. C. Ahuja, T. D. Narang, S. Trehan: Best approximation on convex sets
in a metric space, J. Approx. Theory 12 (1974) 94–97.
[2] J. J. Charatonik, W. Charatonik: Dendrites, Aportaciones Matemáticas,
Serie Comunicaciones 22 (1998) 227–253.
[3] G. Cohen, S. Gaubert, J. P. Quadrat: Duality and Separation Theorems in
Idempotent Semimodules, Linear Algebra Appl. 379 (2004) 395–422.
[4] M. Develin, B. Sturmfels: Tropical Convexity, Documenta Math. 9 (2004)
1–27.
[5] R. Engelking, K. Sieklucki: Topology. A Geometric Approach, Heldermann
Verlag, Berlin (1992).
[6] J. Fletcher, W. B. Moors: Chebyshev Sets, J. Aust. Math. Soc. 98 (2015)
161–231.
[7] S. Gaubert, R. D. Katz: The Minkowski theorem for max-plus convex sets,
Linear Algebra Appl. 421(2-3) (2007) 356–369.
[8] S. Gaubert, R. D. Katz: Max-Plus Convex Geometry, in: Proceedings of
the 9th International Conference on Relational Methods in Computer Science and 4th International Workshop on Applications of Kleene Algebra
(RelMiCS/AKA 2006), R. A. Schmidt (ed.), Lecture Notes in Comput. Sci.
4136, Springer-Verlag, Berlin-Heidelberg (2006) 192–206.
[9] J. R. Giles: Convex analysis with application in the differentiation of convex
functions, Research Notes in Mathematics 58, Pitman Publishing, Boston
(1982).
[10] M. Gondran, M. Minoux: Graphs, Dioids and Semirings. New Models and
Algorithms, Operations Research/Computer Science Interfaces 41, Springer
Science & Business Media (2008).
27
[11] G. Gutierres: Sequential Topological Conditions in R in the Absence of the
Axiom of Choice, Math. Log. Quart. 49 (3) (2003) 293–298.
[12] H. Herrlich: Axiom of Choice, Springer-Verlag, Berlin-Heidelberg (2006).
[13] L. Hörmander: Notions of Convexity, Progress in Mathematics 127,
Birkhäuser, Boston-Basel-Berlin (1994).
[14] P. Howard, J. E. Rubin: Consequences of the Axiom of Choice, Mathematical Surveys and Monographs 59, Amer. Math. Soc., Providence (1998).
[15] K. Keremedis: On Sequentially Closed Subsets of the Real Line in ZF,
Math. Log. Quart. 61(1-2) (2015) 24–31.
[16] K. Kunen: The Foundations of Mathematics, Individual Authors and College Publications, London (2009).
[17] Max-Plus Working Group: Max-Plus Algebra and Applications to System
Theory and Optimal Control, in: Proceedings of the International Congress
of Mathematicians August 3–11, 1994 Zürich, Switzerland, S. D. Chatterji
(ed.), Birkhäuser, Basel (1995) 1511–1522.
[18] J. Mioduszewski: Lectures on Topology. Connected Sets and Continua (in
Polish), Silesian University Press, Katowice (2011).
[19] T. D. Narang: Best approximation in metric spaces, Pub. Mat. UAB 27(2)
(1983) 71–80.
[20] T. D. Narang, Sangeeta: A Note on the Convexity of Chebyshev Sets, Bol.
Soc. Paran. Mat. 27(1) (2009) 59–63.
[21] B. De Schutter, B. De Moor: The QR Decomposition and the Singular Value Decomposition in the Symmetrized Max-Plus Algebra Revisited,
SIAM Rev. 44(3) (2002) 417–454.
[22] I. Simon: On Semigroups of Matrices Over the Tropical Semiring, Informatique Théorique et Appl. 28(3–4) (1994) 277–294.
[23] R. Webster: Convexity, Oxford University Press, New York 1994.
28
| 0math.AC
|
MR image reconstruction using deep density priors
Kerem C. Tezcan1,† , Christian F. Baumgartner1 , Ender Konukoglu1 ,
for the Alzheimers Disease Neuroimaging Initiative∗
arXiv:1711.11386v3 [cs.CV] 17 Jan 2018
1
Computer Vision Laboratory, ETH Zürich
†
[email protected]
December, 2017
Abstract
Purpose: MR image reconstruction exploits regularization to compensate for missing k-space
data. In this work, we propose to learn the probability distribution of MR image patches with neural
networks and use this distribution as prior information constraining images during reconstruction,
effectively employing it as regularization.
Methods: We use variational autoencoders (VAE) to learn the distribution of MR image patches,
which models the high-dimensional distribution by a latent parameter model of lower dimensions
in a non-linear fashion. The proposed algorithm uses the learned prior in a Maximum-A-Posteriori
estimation formulation. We evaluate the proposed reconstruction method with T1 weighted images
and also apply our method on images with white matter lesions.
Results: Visual evaluation of the samples showed that the VAE algorithm can approximate the
distribution of MR patches well. The proposed reconstruction algorithm using the VAE prior produced high quality reconstructions. The algorithm achieved normalized RMSE, CNR and CN values
of 2.77%, 0.43, 0.11; 4.29%, 0.43, 0.11, 6.36%, 0.47, 0.11 and 10.00%, 0.42, 0.10 for undersampling
ratios of 2, 3, 4 and 5, respectively, where it outperformed most of the alternative methods. In the
experiments on images with white matter lesions, the method faithfully reconstructed the lesions.
Conclusion: We introduced a novel method for MR reconstruction, which takes a new perspective
on regularization by using priors learned by neural networks. Results suggest the method compares
favorably against the other evaluated methods and can reconstruct lesions as well.
Keywords: Reconstruction, MRI, prior probability, MAP estimation, machine learning, variational inference, deep learning
1
Introduction
Acquisition time in magnetic resonance (MR) imaging is directly related to the number of samples
acquired in k-space. For high quality images, a large number of samples, and therefore long acquisition
times are necessary. Reducing acquisition time in a reliable manner is an important question in MR
imaging and many methods exploiting different properties of the k-space [1] and hardware design [2]
have found wide use in clinical practice.
On the other hand, a substantial amount of effort went into the investigation of reconstruction
methods from randomly or regularly undersampled k-space acquisitions. The random undersampling
∗
Two of the images used in preparation of this article were obtained from the Alzheimers Disease Neuroimaging Initiative (ADNI) database (adni.loni.usc.edu). As such, the investigators within the ADNI contributed
to the design and implementation of ADNI and/or provided data but did not participate in analysis or writing of this report.
A complete listing of ADNI investigators can be found at: http://adni.loni.usc.edu/wpcontent/uploads/how to apply/ADNI Acknowledgement List.pdf
1
approach was primarily motivated by the compressed sensing framework where the incoherence between k-space sampling and some sparsifying transform was exploited to achieve theoretically exact
reconstruction [3, 4, 5, 6, 7]. Similarly, regular undersampling schemes and corresponding reconstruction algorithms were extensively investigated [8, 9, 10, 11]. The common aspect of these two approaches
is that they both are interested in inverting an underdetermined system of equations by finding good
regularizers to overcome the ill-posedness. A key observation is that the use of regularizers in both
approaches can essentially be viewed as introduction of prior information to the reconstruction problem.
Recently, researchers started to employ deep neural networks (DNN) [12] for MR reconstruction
inspired by its success in computer vision and medical image analysis [13]. In the related literature
there are two main approaches for applying DNNs to the reconstruction problem. The first is to teach
the networks a mapping going from the undersampled images to their fully sampled versions [14]. After
training, an undersampled image is fed into the network to obtain its reconstruction at the output.
This approach does not have any explicit data consistency term and cannot guarantee this property at
test time. To overcome this limitation, in [15] the authors introduce a deep cascaded network. Here a
block of convolutional neural network (CNN) is followed by a data consistency term and this structure
is repeated multiple times. This guarantees consistency with measured data also at test time while
the CNNs perform de-aliasing.
The second approach is to use the efficient trainability of the networks to improve existing algorithms. In [16] the authors show that the iterations of the alternating direction method of multipliers
(ADMM) algorithm solving a dictionary based `-1 regularized reconstruction problem can be expressed
as a multi-layer network. Kernels and non-linear functions are parameterized and learned by the network which improve on the fixed version in the original method. Both deep learning approaches suffer
from specificity of training. Learned mappings, kernels and non-linearities are specific to a certain
undersampling pattern and possibly field-of-view (FOV). Hence, the networks need to be retrained for
each different configuration of FOV, undersampling factor and pattern, which limits wide use of these
methods.
In computer vision, DNNs have also been used for approximating high dimensional probability
distributions from samples through unsupervised learning. Such approximations allow to estimate
the likelihood of a previously unseen data point. One such approach is the variational auto encoder
(VAE) algorithm [17, 18]. Using VAEs, it is possible to approximate the distribution of patches of MR
images and compute likelihood of a previously unseen image. Furthermore, the likelihood function is
differentiable since it is modelled as a neural network.
In this work we develop a probabilistic reconstruction model that uses priors learned via VAEs,
and deploys them as regularizers, which we term deep density prior (DDP) based reconstruction. The
approach can be used to learn the prior distribution using fully sampled examples once and reconstruct
for different sampling parameters without retraining. To achieve this we first formulate a Bayesian
model of the imaging process and express the reconstruction problem as a Maximum-A-Posteriori
(MAP) estimation problem. We show how some of the previous methods fit into this formulation and
then present the proposed model using a VAE-learned prior. In the experiments, we first demonstrate
that VAEs can approximate distribution of patches of MR images by showing examples of patches
sampled from the prior. We then show reconstruction results from the proposed model and compare
with both more conventional approaches as well as recent DNN based methods.
2
Methods
In the first two parts of this section, we provide a brief background on Bayesian formulation of the
MR reconstruction problem and the VAE algorithm. Then, starting from Section 2.3, we present our
main technical contribution, that is learning a prior for MR image patches and integrating it in the
reconstruction problem.
2
2.1
Bayesian formulation of the MR reconstruction problem
The MR image is denoted as m ∈ CN , where N is number of voxels1 . The imaging operation is given
by the undersampling encoding operation E = U F S, where S : CN → CN ×γ is the sensitivity encoding
operator. Here, γ is number of coils, F : CN ×γ → CN ×γ is the Fourier operator and U : CN ×γ → CM ×γ
is the undersampling operator, with M < N .
Assuming complex valued zero mean normal distributed noise, denoted as η, the acquired data
y ∈ CM ×γ can be written as y = Em + η. Under this noise model the data likelihood becomes
p(y|m) = N (y|Em, ση ) =
1
1
(Em−y)H (Em−y)
2ση
e
,
(2πση2 )M/2
(1)
where H denotes the Hermitian transpose and ση is the standard deviation of the noise. In reconstruction, we are interested in the posterior distribution p(m|y), i.e. the probability of the image being m
given the k-space measurements, which can be written using the Bayes’ theorem as
p(m|y) =
p(y|m)p(m)
.
p(y)
(2)
The common approach to model the reconstruction problem, which we will also use, is to use the MAP
estimation
arg max p(m|y) = arg max p(y|m)p(m),
(3)
m
m
where the equality is due to p(y) not depending on m. p(m) is called the prior term and represents
the information one has about the fully sampled image before the data acquisition. Taking the log of
both sides and plugging in the definition of the likelihood term yields
arg max log p(m|y) = arg max [log p(y|m) + log p(m)]
(4)
m
m
1
2 M/2
H
= arg max − log (2πση )
−
(Em − y) (Em − y) + log p(m) . (5)
2ση
m
Rewriting the likelihood and ignoring terms that do not vary with m yields
1
2
arg max log p(m|y) = arg max − kEm − yk2 + log p(m)
λ
m
m
(6)
where we define λ , 2ση . Taking the maximum, or equivalently taking the minimum of the negative of
the expression and multiplying all sides with λ recovers the conventional formulation of a reconstruction
problem with a data and a regularization term
m̂ = arg minkEm − yk22 − λ log p(m).
(7)
m
From the point of view of Equation 7, regularization-based reconstruction algorithms differ in the
log prior term they use. A generic prior expression can be written as log p(m) = kΨ(m − µ)kp , where
Ψ is a sparsifying operator, µ is the expected image and k · kp is the `-p norm.
For instance, forcing the image to be sparse in some domain can be expressed as assuming the data
is i.i.d. according to the Laplace distribution in the domain [19, 20]. Mathematically, this is
p(m) = | det(Ψ)|p(Ψm) = | det(Ψ)| Lap(Ψm|0, b) = | det(Ψ)|
Y 1
|(Ψm)i |
exp(−
),
2b
b
(8)
i
1
In this work we focus on 2D imaging, however, the same techniques can be applied to 3D imaging and this extension
will be a part of our future work.
3
where we set µ = 0, denote the matrix determinant with det2 , use | · | as the magnitude and i indexes
voxels in the image. Now, taking − log of the expression and ignoring the terms that do not vary with
m we get
1X
1
− log p(m) ∝
|(Ψm)i | = kΨmk1 .
(9)
b
b
i
Setting b = 1/λ yields the prior term in Equation 7. Furthermore, setting Ψ as the gradient operator
with first order finite differences, denoted by ∇, corresponds to enforcing sparsity in the gradient
domain recovering the basic total variation compressed sensing reconstruction
m̂ = minkEm − yk22 + λk∇mk1 .
m
Alternatively, both Ψ and µ can be estimated from data. One example for this approach is to assume a
Normal prior on m, p(x) = N (µm , Σm ) and estimate the mean µm and the covariance Σm from a lowresolution image of the same object [10, 22]. The posterior p(y|m) is then also a Normal distribution
due to conjugacy relations and the MAP estimate is given with the posterior mean [23]
−1
1 H
1 H
−1
−1
=
E E + Σm
E y + Σm µm
ση
ση
−1
1 H
1 H
−1
= µm +
E E + Σm
E (y − Eµm ) .
ση
ση
µpost
A similar idea has been applied in k-t SENSE and k-t PCA methods in the dynamic imaging setting
to exploit spatio-temporal correlations [22, 10].
In this work, we follow the same approach of estimating the prior term from example images and
propose to approximate − log p(m) with a neural network model. We train a VAE on patches of fully
sampled MR images to capture the distribution of patches and use this prior for reconstruction.
2.2
Learning the data distribution with VAEs
VAE is an unsupervised learning algorithm used to approximate high-dimensional data distributions [17, 18]. We introduce VAEs very briefly and refer the reader to [17] for further discussion.
The main goal of the VAE algorithm is to approximate the data distribution using a latent model
given as
Z
Z
p(x) =
p(x, z)dz =
Z
p(x|z)p(z)dz,
(10)
Z
where z ∈ RL denote the latent variables, x ∈ RP the data and p(z) the prior over the z’s. Direct computation of p(x), log(p(x)) or p(z|x) requires integrating over z, which is not feasible even
with relatively small L. Variational approximation uses an approximate distribution for the posterior
q(z|x) ≈ p(z|x) to address this problem. Using the approximation, log p(x) can be decomposed into
two terms [24]
p(x, z)
log p(x) = ELBO + KLD = Eq(z|x) log
+ KL [q(z|x)||p(z|x)] .
(11)
q(z|x)
The first term is referred to as the evidence lower bound (ELBO) and the second term is the KullbackLeibler divergence (KLD) between the approximate and true posteriors.
VAE parameterizes the data likelihood and the approximate posterior using networks, i.e. q(z|x) =
qθ (z|x) and p(x|z) = pφ (x|z), and during training maximizes ELBO for a given set of data samples
2
Where we used the change of variables rule to relate the distributions [21].
4
with respect to the parameters of the networks θ and φ. Maximizing ELBO will minimize the KLD
and maximize log p(x). The optimization during training is written as
N
1 X
max Ex∼p(data) [ELBO(x, θ, φ)] ≈ max
Eqθ (z|xn ) [log pφ (xn |z)] − KL [qθ (z|xn )||p(z)] ,
θ,φ
θ,φ N
(12)
n=1
where xn is the nth training sample.
The networks qθ (z|x) and pφ (x|z) are called the encoder and the decoder, respectively. The former
takes the data sample x and encodes it into a posterior distribution in the latent space with network
parameters θ. If the posterior distribution is modelled as a Gaussian, then the encoder outputs a mean
and a covariance matrix for z depending on x. The decoder network on the other hand, takes a latent
vector z and maps it to a conditional distribution of the data given z.
In this work, we use the vanilla VAE design [17] except for the data likelihood, where we use a
multi-modal Gaussian pφ (x|z) = N (x|µφ (z), Σφ (z)), similar to [25].
In our model, we use the VAE model to approximate the prior distribution of magnitude of 2D MR
patches of fully sampled images. In other √
words,
√ we approximate p(|x|), where x is a patch extracted
from a complex valued MR image of size P x P . We provide the exact parameters of the network
design in Section 2.5.
2.3
Deep density prior reconstruction model
Once the network is trained and optimal values θ∗ and φ∗ are found, we can integrate the prior within
a Bayesian formulation of the reconstruction problem. We make two key observations to achieve this.
First, an approximate log likelihood of an image patch x can be obtained by evaluating ELBO(|x|)
with the optimal parameters.
p(z)
ELBO(|x|) = Eqθ∗ (z||x|) log pφ∗ (|x||z) + log
,
(13)
qθ∗ (z||x|)
where we use the magnitude of the patch |x| within the ELBO. This allows us to formulate the
proposed model as the following MAP estimation problem
X
arg minkEm − yk22 −
ELBO (|xr |) ,
(14)
m
xr ∈Ω(m)
where Ω(m) denotes a set of (overlapping) patches covering the image m and |xr | is the magnitude
of the rth image patch. Note that this approach assumes independence between different patches and
therefore, ignores statistical dependencies between them. However, it also makes the computations
much easier.
Since an exact computation of the ELBO term requires evaluating the expectation with respect to
q(z||x|), which is computationally not feasible, we use a Monte Carlo sampling approach to calculate
the ELBO as follows
J
p(zj )
1X
log p(|x||zj ) + log
, with zj ∼ q(z||x|).
ELBO(|x|) ≈
J
q(zj ||x|)
(15)
j=1
Here J represents the number of Monte-Carlo samples.
Our second key observation is that the approximation in Equation 15 is differentiable since each
term is defined through networks that are themselves differentiable. This is the critical aspect that
allows integrating the trained VAE as a prior into an optimization based reconstruction algorithm.
5
Plugging the ELBO approximation into Equation 14, we obtain the formulation of the DDP reconstruction problem
J
j
X
X
p(z )
1
arg minkEm − yk22 −
log p(|xr ||zj ) + log
, where zj ∼ q(z||xr |),
(16)
j ||x |)
J
q(z
m
r
xr ∈Ω(m)
j=1
where the first term is the usual data term and the second term within the summation is the regularization term that arises from the learned prior.
We can compute the total derivative of the prior term with respect to each image patch as follows
p(zj )
R(|x|, zj ) , log p(|x||zj ) + log
q(zj ||x|)
J
J
d 1 X
1X d
R(|x|, zj )
R(|x|, zj ) =
dx J
J
dx
j=1
j=1
J
j
X
x 1
∂
dz
∂
=
R(|x|, zj ) + j R(|x|, zj )
,
|x| J
∂|x|
∂z
d|x|
(17)
(18)
(19)
j=1
where we defined R(x, zj ) for notational simplicity. The second term in the last line is due to the
dependency of the samples zj to x. The term x/|x| is due to taking the derivative of the magnitude
with respect to the image patch.
Algorithm 1 Deep density prior (DDP) reconstruction using POCS. See text for a more detailed
explanation.
1: y: undersampled k-space data
2: E: undersampling encoding operator
3: VAE: the trained VAE
4: procedure DDPrecon(y, E, VAE)
5:
m0 ← E H y
. initialize with the zero-filled image
6:
for t = 0 : T − 1 do
. main loop: POCS iterations
7:
mt,0 ← mt
8:
for k = 0 : K − 1 do
. inner loop: iterations for the prior projection Pprior
t,k
t,k
9:
{xr } ← image2patches(m )
. creates a set of patches covering the image
10:
for r = 1 : no
. loop over all the patches in {xtr }
h ofPpatches do i
J
d
1
j
11:
gr ← dx
. calculate the derivative acc. to Eq. 19
j=1 R(|x|, z )
t,k
J
12:
13:
14:
15:
16:
17:
18:
19:
20:
2.4
end for
g ← patches2image({gr })
mt,k+1 ← mt,k + α · g
end for
mt ← |mt,K |
mt+1 ← mt − E H (Emt − y)
end for
return mT
end procedure
x=xr
. Pprior mt
. Pabs mt
. PDC mt
. Resulting reconstruction
Optimization using projection onto convex sets
We solve the DDP optimization problem given in Equation 16 using the projection onto convex sets
(POCS) algorithm [26], specifically using the formulation in [27]. POCS is an iterative minimization
6
process, where the solution variable is projected sequentially onto different convex sets, each defined
by one of the constraints in the problem.
The projection for the data consistency term is given simply as PDC m = m − E H (Em − y) [26].
Since we do not have a projection operator for the prior term, we approximate it by several gradient
ascent steps with a small step size. We use the final image at the end of the ascent steps as the
projected image patch. We define the operation as follows: i) we create a set of patches {xtr } = Ω(mt )
from the image mt at iteration t, ii) obtain the derivatives for each of these patches using Equation 19,
which have the same size as the patches themselves, iii) create a derivative image from these derivative
patches by averaging the values where the patches overlap, iv) update the image using the derivative
image, v) repeat this K times. To reduce edge effects resulting from patchwise projections, we use four
sets of overlapping patches.
As the likelihood term is only trained on magnitude patches and cannot convey any information
regarding the phase, we also use a phase projection following the prior projection, where we set the
phase of the image to zero, given as Pabs m = |m|. Notice we do this because we expect the images
to have zero phase in this specific application. It is however possible to change this to any other
constraint on the phase of the image, such as a low resolution estimation of the phase [3] or zerodivergence constraint for phase contrast flow imaging reconstruction [28].
With the data consistency, prior and phase projections defined as above, one step of reconstruction
within the POCS framework becomes
mt+1 = PDC Pabs Pprior mt .
(20)
We apply T POCS steps to complete the reconstruction. Algorithm 1 provides a summary of the
reconstruction procedure.
2.5
Details on VAE network architecture and training
We use batches of 50 patches of size 28x28 and a 60 dimensional latent space. Both networks are
depicted in Figure 1. For the encoding network, the input is an image patch and the output is the
mean and the covariance of the corresponding posterior. The decoding network takes a vector z as
input and outputs the mean and the covariance of the corresponding likelihood for the patch. Both
networks are mostly convolutional with single fully connected layers. All convolutional layers for both
networks use 3x3 kernels and all layers also have additive bias terms throughout the network.
Due to stability issues we use the log of the variance values throughout the network. We initialize
the network weights with a truncated normal initializer with standard deviation 0.05. We use Adam [29]
for optimization (learning rate of 5e-4, default momentum values in Tensorflow).
3
3.1
Experimental setup
MR image data
We evaluated our method using structural images from the Human Connectome Project (HCP)3 data
set [30]. High quality and large number of HCP images are ideal for learning priors with the VAE
model.
We took 2D slices from the T1 weighted 3D MPRAGE images acquired with 2400 ms, 2.14 ms and
1000 ms for TR, TE and TI, respectively, flip angle of 8 degrees, a standard field of view for all subjects
of 224x224x224 mm3 with 0.7 mm isotropic resolution using a bandwidth of 210 Hz/voxel. Images were
3
Further details of the HCP data set can be found in HCP-S500+MEG2-Release-Reference-Manual and HCPS1200-Release-Appendix-I (obtainable from https://www.humanconnectome.org/study/hcp-young-adult/document/500subjects-data-release)
7
Figure 1: Architecture of the encoding (top) and decoding (bottom) networks of our VAE. Arrows
indicated with C# are convolutional layers followed by ReLU non-linear activation except C4 of
decoding network, which is not followed by non-linear activation. Arrows indicated by FC are fully
connected layers. FC of the decoding network is followed by a ReLU activation but not of the encoding
network.
acquired with a 3T Siemens device. Fat suppression was used during acquisition. In our experiments,
we used the minimally preprocessed images in order to also have the corresponding FreeSurfer [31]
segmentations, which we use in our evaluation. Preprocessing steps were resampling to 1mm isotropic
resolution with FOV matrix 260x311x260 and rigid alignment. Despite the rigid alignment, there
were substantial local orientation differences between the images from different subjects. We used
five central slices with skipping four slices between each. We cropped the images to a size of 252x308,
removing only background, to reduce computational load. We normalized the images by mapping their
99th percentile to 1.
We used images from 158 subjects (790 images in total) to train the prior VAE model. The VAE
model trained for 200k iterations. As test data, we took central slices from images of 17 separate
subjects.
To test if the proposed reconstruction method can be used on a domain that is different from the
one the prior is trained on, we experimented with two slices from the Alzheimer’s Disease Neuroimaging
Initiative (ADNI) data set4 . The images were selected from subjects with Alzheimer’s disease and who
4
Two of the images used in the preparation of this article were obtained from the Alzheimers Disease Neuroimaging
Initiative (ADNI) database (adni.loni.usc.edu). The ADNI was launched in 2003 as a public-private partnership, led by
Principal Investigator Michael W. Weiner, MD. The primary goal of ADNI has been to test whether serial magnetic
resonance imaging (MRI), positron emission tomography (PET), other biological markers, and clinical and neuropsychological assessment can be combined to measure the progression of mild cognitive impairment (MCI) and early Alzheimers
8
have visible white matter lesions to also test whether the proposed method will be able to faithfully
reconstruct lesions. The acquisition parameters of the ADNI images were different than HCP: TR,
TE, TI values were 7.34 ms, 3.03 ms, 400 ms, flip angle was 11 degrees, FOV matrix was 196x256x256
with resolution 1.2 mm x 1 mm x 1 mm. The images were acquired with a 3T GE scanner with no fat
suppression and were bias-field corrected with the N3 algorithm [32]. We extracted the central slices
that showed the largest lesions from these images and further cropped the FOV to 168x224 to get rid
of the empty regions in the images to accelerate computations. We normalized the image intensities
the same way as the HCP data.
3.2
Setup and evaluation
In our assessment, we artificially undersampled the test images in k-space, reconstructed them back
and compared the results with the original images. We experimented with varying undersampling
ratios, which we denote with R when presenting results.
We used Cartesian undersampling in one dimension while fully sampling in the other dimension,
corresponding to phase-encoding and readout directions, respectively. We present some examples of
undersampling patterns in Figures 3 and 4. We generated the patterns by randomly sampling a 1D
Gaussian distribution along the phase-encoding dimension. We randomly drew many sampling patterns
from the Gaussian distribution and selected the ones with the best peak-to-side ratio. In addition, we
added the central 15 profiles to these selected patterns to fully sample the low-frequency components.
We used 2, 3, 4 and 5 for net undersampling ratios (including the fully sampled center). We assumed
a single coil imaging with uniform sensitivity throughout the experiments for proof of concept, as the
proposed method works independent of the number of coils.
We tried different configurations of latent space size and patch size to see how sensitive the method
is to changes in these parameters. We used patch sizes of 12x12, 20x20, 28x28, and 36x36 and latent
space dimensions of 40, 60 and 80. For each configuration we retrained the prior model. We also
analyzed the proposed method’s robustness to noise. To this end, we added synthetic noise to the fully
sampled images before undersampling. The original images had a SNRA between 50 to 70, depending
on the subject. We added 4 different levels of noise with SNR values 40, 30, 20 and 10. Notice the
actual noise on the images is higher (SNR nearly half) due to the base noise on the original images.
We studied the different configurations and robustness to noise using 5 of the test images.
While assessing the DDP method, we generated a new undersampling pattern for each test to make
sure the reconstruction results covered the variability of the undersampling patterns. We reconstructed
the test images using 30 POCS iterations (T=30) for R=2 and 3 and 60 for R=4 and 5, 10 iterations
for the prior projection (K=10) and the step size was set to α=1e-4.
We compared the reconstruction with the fully sampled ”ground truth” images using Root-MeanSquared-Error (RMSE), Contrast-to-Noise-Ratio (CNR) and Contrast difference (CN) computed at the
gray and white matter boundary in the brain. In order to obtain white matter boundaries we applied
binary erosion (with a square structuring element of size 7x7) to the white matter segmentations, as
computed by FreeSurfer, and took the difference of the original and eroded segmentations.
3.3
Compared methods
As a basic benchmark, we evaluated zero-filling reconstructions. As the first baseline, we used total
variation (TV) reconstruction as described in [3]. We used the BART toolbox implementation that is
publicly available [33]. We used the ”pics” tool with TV regularization (regularization strength 0.075)
and the ADMM parameter ρ as 1. We used 20 conjugate gradient steps and 4500 total iterations
(”bart pics -R T:3:0:0.0075 -u1 -C20 -i4500”). We ran a grid search for the best parameter setting in
the RMSE sense, which we chose as above.
disease (AD). For up-to-date information, see www.adni-info.org
9
As the second method, we used reconstruction using dictionary learning (DL) [34]5 . We used 200
iterations, a patch size of 36 voxels and 36 dictionary atoms. Furthermore, we set number of signals
used for training to 7200 and the overlap stride to 2. We used K-SVD learning with both sparsity and
error threshold. The sparsity level was set to 7. The error threshold was set to 0.046 for the first four
iterations, then to 0.0322 for the rest of the iterations. We used 15 K-SVD iterations. We choe the
parameters as sugested by the authors in code, but increased the number of iterations.
We employed the recently proposed ADMM-Net [16] 6 . We modified the code very slightly to work
with non-square images and Cartesian undersampling patterns. To train the model, we used the same
790 images, which we used to train the VAE model. To train and test the ADMM-Net, we fixed the
undersampling pattern for all undersampling ratios, as required by the method. We used 15 stages
with 8 filters (filter size 3x3), padding as 1. We did not use weight decay during training. We set
the maximum iteration number to 25 for the L-BFGS algorithm. It trained to maximum iterations
for R=2 (45 hours), 13 iterations before convergence for R=3 (42 hours), 18 iterations before running
out of time (120 hours) for R=4 and maximum iterations for R=5 on a GPU (GeForce GTX TITAN
X). The normalized mean squared errors were 0.078 and 0.035 for R=2, 0.11 and 0.071 for R=3, 0.15
and 0.11 for R=4, and 0.19 and 0.13 for R=5, before and after training, respectively. We took the
parameter setting for which the best results were reported in the paper.
We also compared with the methods shift-invariant discrete wavelet reconstruction (SIDWT) [35],
fast multiclass dictionary learning (FDLCP) [36] and patch based directional wavelets (PBDW) [37].
For these methods we cropped the images to 256x256 and generated new undersampling patterns for
this size since the authors implementations worked only with that size7 . We did not modify the code
for these methods and took the parameters them as set by the authors in the code. For comparison,
we ran our proposed method on these images as well.
Finally we compared to an algorithm based on BM3D denoising, namely BM3D-MRI [38]8 and
used the parameters as set by the author in the code.
4
Results
We start by showing patches sampled from the prior model trained for patch-size of 28x28 and latent
dimension of 60 in Figure 2. These patches were generated by simply feeding 16 random z vectors
drawn from unit Gaussian to the decoder network. The decoder network outputs the mean images,
i.e. µφ (z), and the corresponding variance images, i.e. Σφ (z). We would like to note that we have
not cherry-picked these images. The sampled mean patches, shown on the left, look realistic where
gray matter, white matter and gyri/sulci structures are clearly defined and properly positioned. These
generated samples suggest that the VAE algorithm is able to approximate the underlying distribution
of MR patches. Variance images images on the right show that VAE places high variance in the CSF
areas as well as boundaries between structures. We observe that isolated gray matter or CSF islands
receive high variance.
Next, we show visual reconstruction results. Figure 3 shows results for one of the test images from
the HCP data for R=2. The sampling pattern is also shown in the figure. All the methods achieve
good reconstruction quality in terms of reproducing structural and texture information, except the TV
images which are cartoon-like. The error images show higher error in regions where gray matter and
white matter structures are intertwined.
Aliasing artifacts become more prominent at R=3 as seen in Figure 4. SIDWT cannot complete
dealiasing, TV and PBDW are have structural and texture problems. ADMM-Net cannot reconstruct
5
Code available at http://www.ifp.illinois.edu/∼yoram/DLMRI-Lab/Documentation.html
https://github.com/yangyan92/Deep-ADMM-Net
7
Implementations can be found on http://csrc.xmu.edu.cn/csg publications en.html
8
Code available at http://web.itu.edu.tr/eksioglue/pubs/BM3D MRI.htm
6
10
(a) Mean
(b) Variance
Figure 2: 16 image patches sampled from the prior learned by the VAE algorithm. Images on the
left are the mean predictions and the images on the right are the corresponding variance maps. The
patches are 28x28.
the small GM island in the zoomed image. DDP, DLMRI and FDLCP perform well.
We show four more randomly selected images from the test set in supplementary Figures S4 and S5.
We also show results for higher undersampling ratios in supplementary Figures S6 and S7.
We present the quantitative results for reconstruction accuracy in Table 1. In the first part, i.e.
for the full FOV reconstructions BM3D-MRI performs the best for R=2 and 3, followed by the DDP.
After R=4 the DDP method works best. For the cropped FOV part, The FDLCP reconstruction
outperforms all the other methods in all cases in terms of RMSE with the proposed method following
as the 2nd. In terms of CNR, the proposed method performs better than all the others.
A plot showing convergence of the POCS algorithm can be seen in Figure S1. Results for patch
size and latent dimension can be seen in supplementary Figure S2.
Lastly, we show DDP reconstructions in Figure 5 for the ADNI images for R=2. We used the VAE
model was trained on HCP images, which is composed of healthy subjects, to reconstruct the images
here. The reconstructed images recover gray matter-white matter structures and edges faithfully. The
white matter lesions are also well reconstructed. The error maps do not indicate a specific increase of
error in the lesion regions.
5
Discussion
MR patches sampled from the learned distribution shown in Figure 2 support our hypothesis that the
VAE model can learn to approximate the distribution of MR patches. The variance maps show that
the sulci-like generated structures filled with CSF have higher variance, which is in accordance with
the literature [39].
The reconstruction examples in Figure 3 and Figure 4 show that our proposed VAE based deep
density prior reconstruction method produces promising results. This suggests that the learned distribution of MR patches can be used as prior for reconstruction. Reconstructions with R=3 shows
11
12
(b) US
(c) DDP
(d) TV
(e) DLMRI
(f) ADMM
(g) BM3D
(h) DDP
(i) SIDWT
(j) FDLCP
(k) PBDW
Figure 3: Reconstruction results for R=2. First row shows the fully sampled image (FS), the undersampled image
(US) and the results of different reconstruction methods. Second row shows the used undersampling pattern and the
error maps (intensities clipped to (-0.3, 0.3)). Third row shows a zoomed in part of the images. US, DDP, TV,
DLMRI, ADMM-Net and BM3D-MRI images are produced with the undersampling pattern that was used to train the
ADMM-Net for comparability. In the second part DDP, SIDWT, FDLCP and PBDW are reconstructed for a smaller
FOV.
(a) FS
13
(b) US
(c) DDP
(d) TV
(e) DLMRI
Figure 4: Similar display as in Figure 3 with R=3.
(a) FS
(f) ADMM
(g) BM3D
(h) DDP
(i) SIDWT
(j) FDLCP
(k) PBDW
(a) FS
(b) US
(c) DDP
(d) Error
Figure 5: DDP reconstruction results for two images with white matter lesions due to Alzheimer’s
disease from the ADNI data set for R=2. Images show the original (FS), undersampled (US), reconstructed (DDP) and the error maps from left to right. Lesions are clearly visible in the reconstructed
images as well. Error map values are clipped to (-0.3,0.3)
14
limitations of the methods.
In the full FOV case, the proposed method performs better than most other methods in terms of
quality metrics except for the BM3D-MRI and better or equally in visual quality assessment. The
BM3D based method performs worse than the DDP in higher R’s since denoising becomes more
difficult.
For the cropped FOV FDLCP performs consistently better, which is a highly optimized method.
The results we report here are for proof of concept and are open to improvement. Compared to most
other methods learning the data distribution and using this as a prior works better than enforcing a
specific distribution in a specific domain, such as the Laplace distribution in the gradient domain as in
TV or more generalized distributions in a kernel domain as in DL. These methods are inherently limited
by their formulations compared to the more general formulation of the DDP approach presented here.
Hence the DDP approach can be expected to improve over these methods further.
Two of the limitations of our method are the variational assumption and the unit Gaussian prior
for the latent space. A lot of work went into investigating different latent space prior structures
such as Gaussian mixtures or graphical models which could in principal improve the encoding and
representative capacity of the latent variables to address the latter limitation [40, 41]. Similarly a lot
of research is going into different density estimation methods, for instance using generative adversarial
networks [42]. These have the potential to improve on the approximation of the true likelihood and
hence to improve reconstruction quality as well.
Another limitation of the proposed method is the independence assumption over patches. This
assumption can be removed by taking into account the statistical dependence between different patches
in an image. However, this is likely to increase the computational complexity of the reconstruction
method.
Experiments with different configurations show that the DDP reconstruction is not sensitive to
patch size and latent space dimensions for a reasonable range. Furthermore, as expected the performance degrades with decreasing SNR.
We emphasize the generality of our method in contrast to most DNN based methods in the literature. Most methods require a specific mapping to be learned for every undersampling pattern and/or
lack an explicit data consistency term, whereas our method does not suffer from these limitations. Our
method learns the prior directly on images and can be used to reconstruct any sampling scheme faithful
to the measured data without need of retraining as long as the images are from the same domain.
The ADNI reconstruction results are also encouraging in two terms. First, they show that the
learned model does not blur out the lesion regions during the prior projection, which could be expected
since the training images do not have any examples of lesions. However, for a proper treatment of
images with lesions, we believe the training data set should include images with lesions.
Secondly, the proposed method performs reasonably despite the domain difference between the
training and test sets. Although the two data sets are acquired at the same field strength, they still
differ in the acquisition protocol and imaging parameters. Furthermore, our method is invariant to
FOV changes but not to scale changes. HCP and ADNI voxel sizes are different (1mm x 1mm vs. 1.2
mm x 1 mm), however the reconstructions still look plausible. Notice that the HCP has fat suppression
which reduces the signal from the skull significantly, however ADNI images do not have the suppression,
making the dealiasing more challenging. The results indicate that the learned model can generalize to
different scales and imaging protocols without retraining. More challenging problems which might be
attacked by domain adaptation methods such as training and testing on different imaging modalities
still require further research.
15
FS
Zero-fill
DDP
TV
DLMRI
ADMMNet
BM3D-MRI
FS
DDP
SIDWT
FDLCP
PBDW
R=2
RMSE
13.03(1.13)
2.64(0.39)
3.87(0.47)
4.48(0.52)
3.65(0.39)
1.92(0.36)
2.73(0.38)
4.49(0.98)
2.63(0.35)
3.24(0.38)
CNR
0.48(0.10)
0.40(0.09)
0.48(0.11)
0.46(0.11)
0.46(0.11)
0.48(0.11)
0.48(0.10)
0.48(0.10)
0.48(0.10)
0.45(0.11)
0.48(0.10)
0.47(0.11)
CN
0.12(0.02)
0.12(0.02)
0.12(0.02)
0.12(0.02)
0.12(0.02)
0.12(0.02)
0.12(0.02)
0.12(0.02)
0.12(0.02)
0.12(0.02)
0.12(0.02)
0.12(0.02)
R=3
RMSE
21.15(1.36)
4.14(0.50)
7.56(0.83)
7.25(0.83)
7.41(0.46)
4.23(1.05)
4.67(0.92)
9.42(1.62)
4.35(0.87)
5.59(0.94)
CNR
0.48(0.10)
0.33(0.07)
0.48(0.11)
0.40(0.10)
0.40(0.10)
0.44(0.11)
0.46(0.10)
0.48(0.10)
0.48(0.10)
0.39(0.09)
0.45(0.10)
0.44(0.10)
CN
0.12(0.02)
0.09(0.02)
0.12(0.02)
0.09(0.02)
0.10(0.02)
0.11(0.02)
0.11(0.02)
0.12(0.02)
0.12(0.02)
0.12(0.02)
0.14(0.03)
0.13(0.02)
R=4
RMSE
24.92(1.91)
6.36(1.69)
11.40(1.39)
10.72(1.31)
11.80(0.69)
8.08(2.48)
7.28(1.27)
14.57(1.96)
6.72(0.89)
8.51(0.98)
CNR
0.48(0.10)
0.31(0.06)
0.47(0.11)
0.35(0.09)
0.33(0.09)
0.36(0.09)
0.43(0.10)
0.48(0.10)
0.46(0.10)
0.33(0.08)
0.41(0.10)
0.38(0.09)
CN
0.12(0.02)
0.08(0.02)
0.11(0.02)
0.08(0.02)
0.08(0.02)
0.09(0.02)
0.10(0.02)
0.12(0.02)
0.11(0.02)
0.10(0.02)
0.13(0.03)
0.12(0.02)
R=5
RMSE
27.36(1.79)
10.00(2.39)
14.56(1.23)
13.87(1.25)
13.54(0.70)
11.70(2.76)
12.72(2.66)
18.76(2.80)
9.62(1.48)
11.38(1.39)
CNR
0.48(0.10)
0.30(0.06)
0.42(0.11)
0.31(0.08)
0.30(0.08)
0.32(0.08)
0.38(0.09)
0.48(0.10)
0.39(0.08)
0.32(0.07)
0.35(0.08)
0.34(0.08)
CN
0.12(0.02)
0.08(0.02)
0.10(0.02)
0.07(0.02)
0.07(0.02)
0.08(0.02)
0.09(0.02)
0.12(0.02)
0.10(0.02)
0.10(0.02)
0.11(0.02)
0.10(0.02)
R - net undersampling ratio, RMSE - root mean squared error in percentage, CNR - contrast to noise
ratio, CN - contrast.
Table 1: Table summarizing results for different reconstruction quality metrics. Numbers indicate the
mean (and standard deviation) of the error metric for N=17 test images. Second part of the table
reports results for the cropped FOV.
6
Conclusion
In this paper we proposed a novel method, termed deep density prior for MR reconstruction from
undersampled k-space acquisitions. The method uses the VAE algorithm to learn the distribution of
MR patches from a training set composed of fully sampled images. The model then uses this learned
distribution as a probabilistic prior in a Bayesian reconstruction framework. We have shown that
the VAE can approximate the distribution and generate realistic looking MR patches. Furthermore,
reconstruction with the DDP approach yielded promising results for HCP and ADNI data sets in terms
of visual quality and quantitative measures.
Acknowledgements
Authors would like to acknowledge Prof. Klaas Pruessmann and Prof. Sebastian Kozerke for valuable
discussions. We thank NVIDIA for their GPU donation. The presented work is funded by Swiss
National Science Foundation grant number: 205321 173016. For the ADNI data usage statement see
the footnote9 .
References
[1] D. A. Feinberg, J. D. Hale, J. C. Watts, L. Kaufman, and A. Mark, “Halving mr imaging time
by conjugation: demonstration at 3.5 kg.,” Radiology, vol. 161, no. 2, pp. 527–531, 1986. PMID:
9
Data collection and sharing for this project was funded by the Alzheimer’s Disease Neuroimaging Initiative (ADNI)
(National Institutes of Health Grant U01 AG024904) and DOD ADNI (Department of Defense award number W81XWH12-2-0012). ADNI is funded by the National Institute on Aging, the National Institute of Biomedical Imaging and
Bioengineering, and through generous contributions from the following: AbbVie, Alzheimers Association; Alzheimers
Drug Discovery Foundation; Araclon Biotech; BioClinica, Inc.; Biogen; Bristol-Myers Squibb Company; CereSpir, Inc.;
Cogstate; Eisai Inc.; Elan Pharmaceuticals, Inc.; Eli Lilly and Company; EuroImmun; F. Hoffmann-La Roche Ltd and its
affiliated company Genentech, Inc.; Fujirebio; GE Healthcare; IXICO Ltd.; Janssen Alzheimer Immunotherapy Research
& Development, LLC.; Johnson & Johnson Pharmaceutical Research & Development LLC.; Lumosity; Lundbeck; Merck
& Co., Inc.; Meso Scale Diagnostics, LLC.; NeuroRx Research; Neurotrack Technologies; Novartis Pharmaceuticals
Corporation; Pfizer Inc.; Piramal Imaging; Servier; Takeda Pharmaceutical Company; and Transition Therapeutics.
The Canadian Institutes of Health Research is providing funds to support ADNI clinical sites in Canada. Private
sector contributions are facilitated by the Foundation for the National Institutes of Health (www.fnih.org). The grantee
organization is the Northern California Institute for Research and Education, and the study is coordinated by the
Alzheimers Therapeutic Research Institute at the University of Southern California. ADNI data are disseminated by the
Laboratory for Neuro Imaging at the University of Southern California.
16
3763926.
[2] K. P. Pruessmann, M. Weiger, M. B. Scheidegger, and P. Boesiger, “Sense: Sensitivity encoding
for fast mri,” Magnetic Resonance in Medicine, vol. 42, no. 5, pp. 952–962, 1999.
[3] M. Lustig, D. Donoho, and J. M. Pauly, “Sparse MRI: The application of compressed sensing for
rapid MR imaging,” Magnetic Resonance in Medicine, vol. 58, no. 6, pp. 1182–1195, 2007.
[4] M. Guerquin-kern, M. Hberlin, K. P. Pruessmann, and M. Unser, “A fast wavelet-based reconstruction method for magnetic resonance imaging,” IEEE Transactions on Medical Imaging,
pp. 1649–1660, 2011.
[5] S. G. Lingala, Y. Hu, E. Dibella, and M. Jacob, “Accelerated dynamic MRI exploiting sparsity and
low-rank structure: K-t SLR,” IEEE Transactions on Medical Imaging, vol. 30, no. 5, pp. 1042–
1054, 2011.
[6] D. L. Donoho, “Compressed sensing,” IEEE Transactions in Information Theory, vol. 52,
pp. 1289–1306, Apr. 2006.
[7] E. J. Cands, J. K. Romberg, and T. Tao, “Stable signal recovery from incomplete and inaccurate
measurements,” Communications on Pure and Applied Mathematics, vol. 59, no. 8, pp. 1207–1223,
2006.
[8] L. Wissmann, A. Gotschy, C. Santelli, K. C. Tezcan, S. Hamada, R. Manka, and S. Kozerke,
“Analysis of spatiotemporal fidelity in quantitative 3d first-pass perfusion cardiovascular magnetic
resonance,” Journal of Cardiovascular Magnetic Resonance, vol. 19, p. 11, Jan 2017.
[9] K. P. Pruessmann, M. Weiger, P. Brnert, and P. Boesiger, “Advances in sensitivity encoding with
arbitrary k-space trajectories,” Magnetic Resonance in Medicine, vol. 46, no. 4, pp. 638–651, 2001.
[10] H. Pedersen, S. Kozerke, S. Ringgaard, K. Nehrke, and Y. K. Won, “K-t PCA: Temporally
constrained k-t BLAST reconstruction using principal component analysis,” Magnetic Resonance
in Medicine, vol. 62, no. 3, pp. 706–716, 2009.
[11] J. F. M. Schmidt, L. Wissmann, R. Manka, and S. Kozerke, “Iterative k-t principal component
analysis with nonrigid motion correction for dynamic three-dimensional cardiac perfusion imaging,” Magnetic Resonance in Medicine, vol. 72, no. 1, pp. 68–79, 2014.
[12] Y. LeCun, Y. Bengio, and G. Hinton, “Deep learning.,” Nature, vol. 521, pp. 436–444, May 2015.
[13] G. Litjens, T. Kooi, B. E. Bejnordi, A. A. A. Setio, F. Ciompi, M. Ghafoorian, J. A. van der
Laak, B. van Ginneken, and C. I. Snchez, “A survey on deep learning in medical image analysis,”
Medical Image Analysis, vol. 42, no. Supplement C, pp. 60 – 88, 2017.
[14] B. Zhu, J. Z. Liu, B. R. Rosen, and M. S. Rosen, “Neural network mr image reconstruction with
automap: Automated transform by manifold approximation,” in Proceeding of the 25th Annual
Meeting of ISMRM, ISMRM, 2017.
[15] J. Schlemper, J. Caballero, J. V. Hajnal, A. Price, and D. Rueckert, A Deep Cascade of Convolutional Neural Networks for MR Image Reconstruction, pp. 647–658. Cham: Springer International
Publishing, 2017.
[16] Y. Yang, J. Sun, H. Li, and Z. Xu, “Deep admm-net for compressive sensing mri,” in Advances in
Neural Information Processing Systems 29 (D. D. Lee, M. Sugiyama, U. V. Luxburg, I. Guyon,
and R. Garnett, eds.), pp. 10–18, Curran Associates, Inc., 2016.
17
[17] D. P. Kingma and M. Welling, “Auto-encoding variational bayes.,” CoRR, vol. abs/1312.6114,
2013.
[18] D. J. Rezende, S. Mohamed, and D. Wierstra, “Stochastic backpropagation and approximate
inference in deep generative models,” in Proceedings of the 31st International Conference on
Machine Learning (E. P. Xing and T. Jebara, eds.), vol. 32 of Proceedings of Machine Learning
Research, (Bejing, China), pp. 1278–1286, PMLR, 22–24 Jun 2014.
[19] S. D. Babacan, R. Molina, and A. K. Katsaggelos, “Bayesian compressive sensing using laplace
priors,” Transactions in Imgage Processing, vol. 19, pp. 53–63, Jan. 2010.
[20] D. P. Wipf and B. D. Rao, “Sparse bayesian learning for basis selection,” IEEE Transactions on
Signal Processing, vol. 52, pp. 2153–2164, Aug 2004.
[21] K. P. Murphy, Machine Learning: A Probabilistic Perspective. The MIT Press, 2012.
[22] J. Tsao, P. Boesiger, and K. P. Pruessmann, “k-t blast and k-t sense: Dynamic mri with high
frame rate exploiting spatiotemporal correlations,” Magnetic Resonance in Medicine, vol. 50, no. 5,
pp. 1031–1042, 2003.
[23] A. Gelman, J. B. Carlin, H. S. Stern, and D. B. Rubin, Bayesian Data Analysis, Second Edition
(Chapman & Hall/CRC Texts in Statistical Science). Chapman and Hall/CRC, 2 ed., July 2003.
[24] C. M. Bishop, Pattern Recognition and Machine Learning. Springer, 2006.
[25] S. Yeung, A. Kannan, Y. Dauphin, and L. Fei-Fei, “Tackling over-pruning in variational autoencoders,” CoRR, vol. abs/1706.03643, 2017.
[26] A. A. Samsonov, E. G. Kholmovski, D. L. Parker, and C. R. Johnson, “Pocsense: Pocs-based
reconstruction for sensitivity encoded magnetic resonance imaging,” Magnetic Resonance in
Medicine, vol. 52, no. 6, pp. 1397–1406, 2004.
[27] A. R. De Pierro and E. S. Helou Neto, “From convex feasibility to convex constrained optimization using block action projection methods and underrelaxation,” International Transactions in
Operational Research, vol. 16, no. 4, pp. 495–504, 2009.
[28] C. Santelli, M. Loecher, J. Busch, O. Wieben, T. Schaeffter, and S. Kozerke, “Accelerating 4d
flow mri by exploiting vector field divergence regularization,” Magnetic Resonance in Medicine,
vol. 75, no. 1, pp. 115–125, 2016.
[29] D. P. Kingma and J. Ba, “Adam:
vol. abs/1412.6980, 2014.
A method for stochastic optimization,” CoRR,
[30] M. F. Glasser, S. N. Sotiropoulos, J. A. Wilson, T. S. Coalson, B. Fischl, J. L. Andersson,
J. Xu, S. Jbabdi, M. Webster, J. R. Polimeni, D. C. V. Essen, and M. Jenkinson, “The minimal
preprocessing pipelines for the human connectome project,” NeuroImage, vol. 80, no. Supplement
C, pp. 105 – 124, 2013. Mapping the Connectome.
[31] M. Reuter, N. J. Schmansky, H. D. Rosas, and B. Fischl, “Within-subject template estimation
for unbiased longitudinal image analysis,” NeuroImage, vol. 61, no. 4, pp. 1402 – 1418, 2012.
[32] J. Sled, A. Zijdenbos, and A. Evans, “A nonparametric method for automatic correction of intensity nonuniformity in mri data,” IEEE Transactions in Medical Imaging, vol. 17, pp. 87–97, 01
2002.
18
[33] M. Uecker, J. I. Tamir, F. Ong, S. Iyer, J. Y. Cheng, and M. Lustig, “Bart: version 0.4.00,” Dec.
2016.
[34] S. Ravishankar and Y. Bresler, “Mr image reconstruction from highly undersampled k-space data
by dictionary learning,” IEEE transactions on medical imaging, vol. 30, pp. 1028–41, 11 2010.
[35] B. Ning, X. Qu, D. Guo, C. Hu, and Z. Chen, “Magnetic resonance image reconstruction using
trained geometric directions in 2d redundant wavelets domain and non-convex optimization,”
Magnetic Resonance Imaging, vol. 31, no. 9, pp. 1611 – 1622, 2013.
[36] Z. Zhan, J. F. Cai, D. Guo, Y. Liu, Z. Chen, and X. Qu, “Fast multiclass dictionaries learning
with geometrical directions in mri reconstruction,” IEEE Transactions on Biomedical Engineering,
vol. 63, pp. 1850–1861, Sept 2016.
[37] X. Qu, D. Guo, B. Ning, Y. Hou, Y. Lin, S. Cai, and Z. Chen, “Undersampled mri reconstruction
with patch-based directional wavelets,” Magnetic Resonance Imaging, vol. 30, no. 7, pp. 964 –
977, 2012.
[38] E. M. Eksioglu, “Decoupled algorithm for mri reconstruction using nonlocal block matching model:
Bm3d-mri,” Journal of Mathematical Imaging and Vision, vol. 56, pp. 430–440, Nov 2016.
[39] R. Tanno, D. E. Worrall, A. Gosh, E. Kaden, S. N. Sotiropoulos, A. Criminisi, and D. Alexander,
“Bayesian image quality transfer with cnns: Exploring uncertainty in dmri super-resolution,” in
Intl Conf. Medical Image Computing and Computer Assisted Intervention (MICCAI), Springer,
September 2017.
[40] N. Dilokthanakul, P. A. M. Mediano, M. Garnelo, M. C. H. Lee, H. Salimbeni, K. Arulkumaran, and M. Shanahan, “Deep unsupervised clustering with gaussian mixture variational autoencoders,” CoRR, vol. abs/1611.02648, 2016.
[41] M. Johnson, D. K. Duvenaud, A. Wiltschko, R. P. Adams, and S. R. Datta, “Composing graphical
models with neural networks for structured representations and fast inference,” in Advances in
Neural Information Processing Systems 29 (D. D. Lee, M. Sugiyama, U. V. Luxburg, I. Guyon,
and R. Garnett, eds.), pp. 2946–2954, Curran Associates, Inc., 2016.
[42] M. Arjovsky, S. Chintala, and L. Bottou, “Wasserstein generative adversarial networks,” in Proceedings of the 34th International Conference on Machine Learning, ICML 2017, Sydney, NSW,
Australia, 6-11 August 2017, pp. 214–223, 2017.
19
8
1800
1780
7
1740
6
1720
5
1700
RMSE (%)
log prior
1760
1680
4
1660
1640
3
0
5
10
15
20
25
30
POCS iter.
Figure S1: log prior (red) and RMSE (blue) values between the original and reconstructed image during
POCS reconstruction from an undersampled image with R=2. Convergence is achieved approximately
after 15 iterations for R=2. The RMSE starts with high value as the iterations are initialized with the
zero-filled image. Notice that the likelihood given as the log prior also increases stably.
20
6.5
RMSE (%)
6.0
5.5
5.0
4.5
4.0
3.5
,
0
8
=
ld
,
0
8
=
ld
,
0
8
=
ld
,
0
8
=
ld
,
0
6
=
ld
,
0
6
=
ld
,
0
6
=
ld
,
0
6
=
ld
,
0
4
=
ld
,
0
4
=
ld
,
,
0
4
=
ld
0
4
=
ld
6
8
3
=
s
p
0
2
=
s
p
2
2
=
s
p
6
1
=
s
p
8
3
=
s
p
0
2
=
s
p
2
2
=
s
p
6
1
=
s
p
8
3
=
s
p
0
2
=
s
p
2
2
=
s
p
1
=
s
p
Figure S2: RMSE values for varying the latent space dimension (ld) and patch size (ps) (5 subjects,
R=3). The method works relatively similar for different configurations indicating robustness to parameter selection to some extend.
21
18
16
RMSE (%)
14
12
10
8
R=2
R=3
R=4
R=5
6
4
40
35
30
25
20
SNR of added noise
15
10
Figure S3: RMSE vs SNR of added noise (5 subjects). As expected increasing noise increases the error
in reconstruction. Notice the net SNR on the images is around half of the added value, due to the
noise in the original images.
22
Figure S4: Reconstruction results from 4 different subjects with R=2. Rows show the original and
undersampled images and the DDP, TV, DLMRI and ADMM-Net, BM3D-MR reconstructions and
reconstructions with the cropped FOV for DDP, SIDWT, FDLCP, PBDW in the up-down direction.
23
Figure S5: Same display as in as Figure S4 with R=3
24
25
(b) US
(c) DDP
(d) TV
(e) DLMRI
(f) ADMM
Figure S6: Similar display as in Figures 3 and 4 with R=4.
(a) FS
(g) BM3D
(h) DDP
(i) SIDWT
(j) FDLCP
(k) PBDW
26
(b) US
(c) DDP
(d) TV
(e) DLMRI
(f) ADMM
Figure S7: Similar display as in Figures 3 and 4 with R=5.
(a) FS
(g) BM3D
(h) DDP
(i) SIDWT
(j) FDLCP
(k) PBDW
| 1cs.CV
|
Using a hierarchy of Domain Specific Languages in
complex software systems design
V. S. Lugovsky <[email protected]>
arXiv:cs/0409016v1 [cs.PL] 9 Sep 2004
February 1, 2008
Abstract
A new design methodology is introduced, with some examples on building Domain
Specific Languages hierarchy on top of Scheme.
1
Introduction
nized as “true” macros as they hide an access
to the host language.
This situation looks like a paradox. On the
one hand, industry uses the metaprogramming
ideas and tools, and it is easy to imagine how
it would suffer without them. On the other
hand, industry does not want to hear anything
related to the metaprogramming. It does not
want people inventing new programming languages — plenty of industry coders barely use
only one language and IT managers believe
without any reason that they can not be taught
to use more [6].
Industry prefers to “re–invent a wheel” and
to express any sort of complexity in the form
of libraries for static, steady languages. For
some strange reason learning complicated libraries for a language which barely fits problem
domain needs is preferred to learning a small
new language specifically designed for it.
In this paper I am trying to advocate the
metaprogramming approach as the major design methodology for complex systems. Sounds
like another one “silver bullet” invention?
There were many methodologies claiming to
solve all possible problems of the mankind —
RUP, eXtreme programming, etc. Why do we
need another one? Simply because the previous approaches did not succeed. They were
too tied to particular programming technologies (mostly — OOP varieties), which are definitely not “silver bullets”. Metaprogramming
Programs that write programs that write programs (...) Too complicated? A hackers technique which can not be applied to the “real
world” problems? This is exactly how IT industry specialists think about metaprogramming. And this is a completely wrong notion!
Metaprogramming is the only known way
to reduce the complexity significantly. In some
areas “programs that write programs” are accepted by the industry due to the enormous
level of complexity of the corresponding handwritten code — regular expressions, lexers and
parsers generators to name a few, and code wizards and templates in popular “integrated development environments” are also widely used.
But this does not help at all in the overall methodology recognition. The industry’s
most beloved and widely buzzworded language,
Java, does not have even such a rudimentary
preprocessor as C does. Very few C++ programmers have an idea on how to use the templates, they just utilize STL without any understanding of the true source of the power.
Even in the enlightened world of Lisp programming the misunderstanding is surprisingly
wide: almost all of Lisp dialects and Scheme
implementations have problems with macros
(not so many people are using them), and even
the current Scheme standard R5RS contains
only hygienic macros that can hardly be recog1
methodology is different, it strongly encourages of domain specific languages is not very tricky.
the use of all possible programming technolo- An approach I will describe here is based on
metaprogramming techniques. It requires a so
gies and to invent the “impossible” ones.
called Core Language, on top of which we will
build a hierarchy of our domain specific lan2 Domain specific lan- guages. The Core Language should possess the
following properties:
guages
Below I am providing an outline of the proposed methodology.
Any problem domain can be best expressed
using a language (mathematical, programming,
natural, ...) specially designed for it. In most
cases there should be one entity in a language
for every entity in a problem domain. For example, if a problem domain is the recognition
of syntax constructions in a characters stream,
the Domain Specific Language should contain
characters and characters sets as a primary entity and automata constructions for expressing
syntax. That is enough — regular expressions
language is designed. It is hard to believe that
somebody will ever invent anything better for
this purpose than this “most optimal” DSL.
If a problem domain is already specified as
an algebra, we even do not have to design the
DSL: it will be this algebra itself, galvanised
with any underlying computational semantics
— this is the way SQL was born. If a problem domain is 3D graphics, linear algebra and
stereometry should be used. All the languages
and data formats dedicated to 3D contain subsets of this formal theories.
As it is stated in [1],
“The object of a DSL-based software architecture is to minimise the
semantic distance between the system’s specification and its implementation.”
3
Core language
For any problem it is convenient to have a language that best fits it. There already exist specialized languages for some common problems.
But what to do if none is available? The answer is trivial: implement it. Implementation
• True macros. That is, we must have an
access to a complete programming language (preferably the same as a host
language, or a different one) inside the
macro definitions. Macros should be real
programs which can do anything that the
programs written in the host language
can do. Macros are producing the code
in the host language, in the form of text
or directly as an abstract syntax tree.
• True runtime eval. Programs that are
generated in the runtime should be evaluated. This can be a different language
than the host language, or, better, the
same one.
• Turing-completeness. This should be a
real programming language, equivalent in
its expressive power to the “general purpose” languages.
• Simplicity. It is an extensible core and
should not contain any unnecessary complexity that can be later added by a user
who really needs it.
• Comprehensive and easy to use data
types system. If a type system is well
suited for expressing any possible abstract syntax trees, the language fits this
requirement.
On top of the Core Language we have to
build functionality that will be needed to implement programming languages. It is lexing,
parsing, intermediate languages that fit well
computational models different from the model
of the Core Language (e.g., if the core language
is imperative or an eager functional, we will
need a graph reduction engine to implement
lazy functional DSLs, or a term unification engine to implement logical languages and a stack
machine if we have to go to lower levels). The
2
Core Language enriched with this “Swiss army
knife” for programming languages development
then becomes a major tool for any project.
4
New methodology
The development process must fit in the following chain:
• divide the problem into sub–problems,
possibly using some object oriented design techniques, or whatever fits better.
• formalize each sub–problem.
5
Scheme example
A good example of a practical Core Language is
Scheme (with addition of Common Lisp–style
macros). It uses S–expressions as an AST,
and S–expressions composition is very natural. S–expressions are good enough to represent any possible AST (for example, XML is
naturally represented as SXML). It provides a
true runtime eval hosting the same language
as in compile time. There exist some practical
and efficient Scheme implementations which
provide performance acceptable for most tasks,
good FFI, and, thus, integration with legacy libraries.
• implement the Domain Specific Language after this formalization, using the
Let us start with adding the functionality
Core Language and other DSL with the
described
above to Scheme. First of all we will
same semantics.
need parsing — not all of our team members
• solve the problem using the best possible are fond of parentheses, so we have to implelanguage.
ment many complicated syntaxes. The most
This way any project will grow into a tree (hier- natural way for a functional programming lanarchy) of domain specific languages. Any lan- guage is to implement a set of parsing combinaguage is a subset or a superset of another lan- tors for building recursive descendant parsers
guage in the hierarchy (or, may be, combina- (mostly LL(1), but it is not such a fixed limit as
tion of several languages), and the amount of LALR(1) for Yacc–like automata generators).
coding for a new language if we already have
a deep and comprehensive hierarchy is quite
small.
A development team working within this
methodology should consist of at least one specialist who maintains this hierarchy, an architect who formalizes problems, and a number of
coders who specialize in particular problem domains, they even may not be programmers at
all — they just have to know well their domains
and operate them in terms that are as close as
possible to the native problem domain terminology. For example, HTML designer will be
happy operating HTML–like tags for his templates (that is why JSP custom tags are so popular); mathematician will find a language modelled after the standard mathematical notation
intuitive — for this reason Wolfram Mathematica is so popular among non-programmers;
game script writer will operate a language expressing characters, their properties and action
rules — stating, not programming. This list can
be continued infinitely.
Of course we will use metaprogramming
wherever possible. All the parsers should be
functions which consume a list of tokens (e.g.
characters) as an input and return the result in
the following form:
((RESULT anyresult) unparsed-input-rest)
or
((FAIL reason) input)
To access the parsing result we will provide
the following macros:
(define-macro (success? r )
‘(not (eq? (caar ,r ) ’FAIL)))
And if we are sure that we have some result, we will use the following macro to extract
it (otherwise, this will return a fail message):
(define-macro (result r )
‘(cdar ,r ))
3
The last definition looks surprisingly comIn any case, we can access the rest of the
pact, thanks to the pselect macro. From this
stream after the parsing pass:
stage the power of metaprogramming becomes
(define-macro (rest r )
more and more obvious.
‘(cdr ,r ))
Just as a reference, we will show here the
These macros could also be implemented as definition of a choice combinator:
functions. But all the macros are available in (define-macro (pOR0 p1 p2 )
the context of macro definitions while functions ‘(λ (l)
are not.
(let ((r1 (,p1 l)))
Almost all of the parsers should fail on the
(if (success? r1 )
end of the input, so the following safeguard
r1
macro will be extremely useful:
(,p2 l)))))
(define-macro (parser p)
‘(λ (l)
(if (∅? l) ’((FAIL "EMPTY"))
(,p l))))
And its nested version is obvious:
(define-macro (pOR p1 . p~o )
0
‘(pselect pOR ,p1 ,@p~o ))
We will skip the rest of the combinators
Now this game becomes more interesting. definitions and just show what we gained after
Here is a very handy macro that nests a se- all. For example, now to define a floating point
quence of applications into the form of (m p1 number recognizer, we can use this definition:
(m p2 . . . (m px pn ))):
(define parse-num
(p+
(define-macro (pselect m p1 . p~o )
(pOR (pcsx-or (#\− #\+))
(if (∅? p~o ) p1
parse-any)
(let ((p2 (car p~o ))
(pMANY pdigit)
(px (cdr p~o )))
OR
(p
(p+ (pcharx #\.)
‘(,m ,p1 (pselect ,m ,p2 ,@px )))))
(pMANY pdigit))
Sequence parsing combinator with two arparse-any)))
guments can be declared as follows:
It looks like BNF, but still too Schemish.
(define-macro (p+0 p1 p2 )
This is already a Domain Specific Language on
‘(λ (l)
top of Scheme, but it does not conform to the
(let ((r1 (,p1 l)))
perfectionist requirement. However, we can use
(if (success? r1 )
this still not perfect parsing engine to imple(let ((r2 (,p2 (rest r1 ))))
ment an intermediate regular expressions lan(if (success? r2 )
guage as a macro. Omitting the definitions, we
(cons (cons ’RESULT
will show the previous recognizer implemented
(append
in a new way:
(result r1 )
(define parse-num
(result r2 )))
(regexp
(rest r2 ))
((#\− / #\+) / parse-any) +
(cons (list ’FAIL "p+" (car r2 )) l)))
(pdigit ∗) +
r1 ))))
(("." + (pdigit ∗)) /
And it will be immediately turned into the
parse-any)))
sequence parsing combinator with an arbitrary
This new Domain Specific Language can be
number of arguments:
used in many ways. For example, we can build
(define-macro (p+ p1 . p~o )
‘(pselect p+0 ,p1 ,@p~o ))
a simple infix pre–calculator for constants:
(define-macro (exp1 . v )
4
(defparsers
(letrec
((epr
(let
((body
(regexp
(num :−> $ 0) /
(lst -> (aprs epr )))))
(regexp
((body + (SCM psym +) + epr )
:−> (list (+ $ 0 $ 2))) /
((body + (SCM psym −) + epr )
:−> (list (− $ 0 $ 2))) /
((body + (SCM psym ∗) + epr )
:−> (list (∗ $ 0 $ 2))) /
((body + (SCM psym / ) + epr )
:−> (list (/ $ 0 $ 2))) / body
))))
(car (result (epr v ))))))
only languages with a computational model
which is close to the model of Scheme (eager dynamically typed functional languages
with imperative features), but any possible
languages, providing small intermediate DSLs
which simulate alternative computational models. For those who need very lowlevel power it
is possible to produce an intermediate code in
C language (for example, the Bigloo Scheme [3]
implementation allows to include C code when
compiling through C backend). For implementing complicated runtime models it is easy to
produce an intermediate Forth–like DSL on top
of Scheme and then use both Scheme and Forth
metaprogramming powers.
6
Alternatives
To make the picture complete, it is necesAnd then, wherever we want to calculate a
sary to mention other possible choices for the
numerical constant in the compilation time, we
Core Language. The popular programming
may use the exp1 macro:
language, C++, could become such a Core
(exp1 5 + ((10 / 2)−(1 / 5)))
Language relatively easily. It has a TuringThis language does not look like Scheme complete macro system, unfortunately, featurany more. And we can go even further, im- ing the language different from the host lanplementing a Pascal (or Rlisp)–like language guage (so only one stage preprocessing is poson top of Scheme, using just the same regexp sible). It lacks a good type system, but it could
macro to describe both a lexer and a parser, be simulated on top of the existing lowlevel feaand then to compile the resulting code to the tures. There exist some implementations of
the recursive descendant parsing combinators
underlying Scheme.
for C++ (e.g., Boost Spirit library [2]), implementation of the functional programming (e.g.,
(pasqualish
Boost Lambda [2]), and even Lisp compilers on
"
top of the C++ template system. The runtime
function fac(x)
evaluation is available in different ways: using
begin
pluggable scripting languages other than C++,
if (x > 0) then
using the C++ interpreter [14]. An interesting
x*fac(x - 1)
approach is described in [13].
else 1;
Another choice is Forth. It is a powerful
end
metalanguage,
but the core language remains
")
too lowlevel and unsafe. Forth is often the only
No more parenthesis that frighten non–Lisp choice available for the embedded systems with
programmers so much! Now even Pascal pro- limited resources.
It is worth mentioning modern experimengrammers can use Scheme.
The code samples above demonstrate some tal extensions for strictly typed functional lanof the techniques available in this approach. guages: Template Haskell [12] and MetaOCaml
The complete implementation can be down- [11]. Both of them conform well to all of
loaded from [4]. It is possible to produce not the Core Language requirements. Objective
5
Caml also provides one–stage metaprogram- metaprogramming too.
ming using a sophisticated preprocessing engine CamlP4. And OCaml is quite good for
Conclusion
implementing interpreters using the closure– 7
based technique. Some examples can be found
The idea of metaprogramming is not something
in [7].
esoteric. Metaprogramming is used widely by
No doubt that Common Lisp would also commercial programmers, they just do not rebe a very good platform since it shares almost alize it. The methodology proposed in this paall the features with Scheme with exception of per is an attempt of uncovering all the hidsimplicity. The killing feature of Common Lisp den power of the metaprogramming techniques
is advanced runtime compilation in some of the available.
major implementations (CMU CL [8] and its
The Scheme example presented above is
descendant SBCL [9] are good examples), and part of the working project, which already
the defmacro is guaranteed to be working in all proved the supremacy of this approach. A subthe implementations available, which is a great set of the Domain Specific Languages hierarchy
advantage over Scheme.
designed for the WWW data acquiring project
For relatively small projects Tcl [10] would is shown on the Fig. 1.
be a good choice. Its computational model
The subject discussed requires future reis based on rewrites (and primary data struc- search and practical approbation, whose final
tures are just the strings of text), which ren- result may be a completely formalized, matheders an extremely powerful metaprogramming matically strict methodology description and a
tool. JavaScript language is also based on Core Language which will best fit this methodthe rewrites semantics, so it could be used for ology.
6
References
[1] Diomidis Spinellis. Reliable software implementation using domain specific languages. In
G. I. Schuëller and P. Kafka, editors, Proceedings ESREL ’99 — The Tenth European
Conference on Safety and Reliability, pages 627–631, Rotterdam, September 1999. ESRA,
VDI, TUM, A. A. Balkema //
[draft] http://www.dmst.aueb.gr/dds/pubs/conf/1999-ESREL-SoftRel/html/dsl.html
[2] The Boost project // http://www.boost.org/
[3] The Bigloo Practical Scheme implementation // http://www.bigloo.org/
[4] V. S. Lugovsky, DSLEngine project home // http://dslengine.sourceforge.net/
[5] P. Graham, The Hundred-Year Language // http://www.paulgraham.com/hundred.html
[6] P. Graham, The Python Paradox // http://www.paulgraham.com/pypar.html
[7] V. S. Lugovsky, publications list // http://ontil.ihep.su/˜vsl
[8] CMU Common Lisp // http://www.cons.org/cmucl/
[9] Steel Bank Common Lisp // http://sbcl.sourceforge.net/
[10] Tcl programming language resource // http://tcl.activestate.com/
[11] MetaOCaml project home // http://www.metaocaml.org/
[12] T. Sheard, S. P. Jones, Template metaprogramming for Haskell //
http://research.microsoft.com/˜simonpj/papers/meta-haskell/
[13] Tempo project home // http://compose.labri.fr/prototypes/tempo/
[14] CINT project home // http://root.cern.ch/root/Cint.html
7
Core language
Parsing combinators
Stack machine
Lexer
Graph machine
Unification machine
Templates language
Parser generator
Forth-like
Regular expressins
Data aquision regexps
Pascal-like
Rule engine
SQL templates
Figure 1: A sample DSLs hierarchy subset for the Web crawler project.
8
| 6cs.PL
|
Ristretto: Hardware-Oriented Approximation of Convolutional Neural
Networks
By
arXiv:1605.06402v1 [cs.CV] 20 May 2016
Philipp Matthias Gysel
B.S. (Bern University of Applied Sciences, Switzerland) 2012
Thesis
Submitted in partial satisfaction of the requirements for the degree of
Master of Science
in
Electrical and Computer Engineering
in the
Office of Graduate Studies
of the
University of California
Davis
Approved:
Professor S. Ghiasi, Chair
Professor J. D. Owens
Professor V. Akella
Professor Y. J. Lee
Committee in Charge
2016
-i-
Copyright c 2016 by
Philipp Matthias Gysel
All rights reserved.
To my family ...
-ii-
Contents
List of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
v
List of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
vii
Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
viii
Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
ix
1 Introduction
1
2 Convolutional Neural Networks
3
2.1
Training and Inference . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3
2.2
Layer Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6
2.3
Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
10
2.4
Computational Complexity and Memory Requirements . . . . . . . . . .
11
2.5
ImageNet Competition . . . . . . . . . . . . . . . . . . . . . . . . . . . .
12
2.6
Neural Networks With Limited Numerical Precision . . . . . . . . . . . .
14
3 Related Work
19
3.1
Network Approximation . . . . . . . . . . . . . . . . . . . . . . . . . . .
19
3.2
Accelerators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
21
4 Fixed Point Approximation
25
4.1
Baseline Convolutional Neural Networks . . . . . . . . . . . . . . . . . .
25
4.2
Fixed Point Format . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
26
4.3
Dynamic Range of Parameters and Layer Outputs . . . . . . . . . . . . .
27
4.4
Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
29
5 Dynamic Fixed Point Approximation
32
5.1
Mixed Precision Fixed Point . . . . . . . . . . . . . . . . . . . . . . . . .
32
5.2
Dynamic Fixed Point . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
33
5.3
Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
34
-iii-
6 Minifloat Approximation
38
6.1
Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
38
6.2
IEEE-754 Single Precision Standard . . . . . . . . . . . . . . . . . . . . .
38
6.3
Minifloat Number Format . . . . . . . . . . . . . . . . . . . . . . . . . .
38
6.4
Data Path for Accelerator . . . . . . . . . . . . . . . . . . . . . . . . . .
39
6.5
Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
40
6.6
Comparison to Previous Work . . . . . . . . . . . . . . . . . . . . . . . .
42
7 Turning Multiplications Into Bit Shifts
43
7.1
Multiplier-free Arithmetic . . . . . . . . . . . . . . . . . . . . . . . . . .
43
7.2
Maximal Number of Shifts . . . . . . . . . . . . . . . . . . . . . . . . . .
44
7.3
Data Path for Accelerator . . . . . . . . . . . . . . . . . . . . . . . . . .
45
7.4
Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
45
8 Comparison of Different Approximations
47
8.1
Fixed Point Approximation . . . . . . . . . . . . . . . . . . . . . . . . .
48
8.2
Dynamic Fixed Point Approximation . . . . . . . . . . . . . . . . . . . .
49
8.3
Minifloat Approximation . . . . . . . . . . . . . . . . . . . . . . . . . . .
49
8.4
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
49
9 Ristretto: An Approximation Framework for Deep CNNs
51
9.1
From Caffe to Ristretto . . . . . . . . . . . . . . . . . . . . . . . . . . . .
51
9.2
Quantization Flow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
51
9.3
Fine-tuning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
52
9.4
Fast Forward and Backward Propagation . . . . . . . . . . . . . . . . . .
53
9.5
Ristretto From a User Perspective . . . . . . . . . . . . . . . . . . . . . .
54
9.6
Release of Ristretto . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
55
9.7
Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
56
-iv-
List of Figures
2.1
Network architecture of AlexNet . . . . . . . . . . . . . . . . . . . . . . .
4
2.2
Convolution between input feature maps and filters . . . . . . . . . . . .
6
2.3
Pseudo-code for convolutional layer . . . . . . . . . . . . . . . . . . . . .
7
2.4
Fully connected layer with activation . . . . . . . . . . . . . . . . . . . .
8
2.5
Pooling layer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
9
2.6
Parameter size and arithmetic operations in CaffeNet and VGG-16
. . .
11
2.7
ImageNet networks: accuracy vs size . . . . . . . . . . . . . . . . . . . .
13
2.8
Inception architecture
. . . . . . . . . . . . . . . . . . . . . . . . . . . .
14
2.9
Data path with limited numerical precision . . . . . . . . . . . . . . . . .
15
2.10 Quantized layer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
16
3.1
ASIC vs FPGA vs GPU . . . . . . . . . . . . . . . . . . . . . . . . . . .
23
4.1
Dynamic range of values in LeNet . . . . . . . . . . . . . . . . . . . . . .
27
4.2
Dynamic range of values in CaffeNet . . . . . . . . . . . . . . . . . . . .
28
4.3
Fixed point results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
30
5.1
Fixed point data path . . . . . . . . . . . . . . . . . . . . . . . . . . . .
32
5.2
Dynamic fixed point representation . . . . . . . . . . . . . . . . . . . . .
34
5.3
Static vs dynamic fixed point . . . . . . . . . . . . . . . . . . . . . . . .
35
6.1
Minifloat number representation . . . . . . . . . . . . . . . . . . . . . . .
39
6.2
Minifloat data path . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
40
6.3
Minifloat results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
41
7.1
Representation for integer-power-of-two parameter . . . . . . . . . . . . .
44
7.2
Multiplier-free data path . . . . . . . . . . . . . . . . . . . . . . . . . . .
45
8.1
Approximation of LeNet . . . . . . . . . . . . . . . . . . . . . . . . . . .
47
8.2
Approximation of CIFAR-10 . . . . . . . . . . . . . . . . . . . . . . . . .
48
-v-
8.3
Approximation of CaffeNet . . . . . . . . . . . . . . . . . . . . . . . . . .
48
9.1
Quantization flow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
52
9.2
Fine-tuning with full precision weights . . . . . . . . . . . . . . . . . . .
53
9.3
Network brewing with Caffe . . . . . . . . . . . . . . . . . . . . . . . . .
54
9.4
Network quantization with Ristretto . . . . . . . . . . . . . . . . . . . .
55
-vi-
List of Tables
3.1
ASIC vs FPGA vs GPU . . . . . . . . . . . . . . . . . . . . . . . . . . .
23
4.1
Fixed point results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
30
5.1
Dynamic fixed point quantization . . . . . . . . . . . . . . . . . . . . . .
36
5.2
Dynamic fixed point results . . . . . . . . . . . . . . . . . . . . . . . . .
37
6.1
Minifloat results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
41
7.1
Multiplier-free arithmetic results . . . . . . . . . . . . . . . . . . . . . . .
45
-vii-
Abstract
Ristretto: Hardware-Oriented Approximation of Convolutional Neural
Networks
Convolutional neural networks (CNN) have achieved major breakthroughs in recent years.
Their performance in computer vision have matched and in some areas even surpassed human capabilities. Deep neural networks can capture complex non-linear features; however
this ability comes at the cost of high computational and memory requirements. State-ofart networks require billions of arithmetic operations and millions of parameters.
To enable embedded devices such as smart phones, Google glasses and monitoring
cameras with the astonishing power of deep learning, dedicated hardware accelerators can
be used to decrease both execution time and power consumption. In applications where
fast connection to the cloud is not guaranteed or where privacy is important, computation
needs to be done locally. Many hardware accelerators for deep neural networks have
been proposed recently. A first important step of accelerator design is hardware-oriented
approximation of deep networks, which enables energy-efficient inference.
We present Ristretto, a fast and automated framework for CNN approximation. Ristretto simulates the hardware arithmetic of a custom hardware accelerator. The framework
reduces the bit-width of network parameters and outputs of resource-intense layers, which
reduces the chip area for multiplication units significantly. Alternatively, Ristretto can
remove the need for multipliers altogether, resulting in an adder-only arithmetic. The
tool fine-tunes trimmed networks to achieve high classification accuracy.
Since training of deep neural networks can be time-consuming, Ristretto uses highly
optimized routines which run on the GPU. This enables fast compression of any given
network.
Given a maximum tolerance of 1%, Ristretto can successfully condense CaffeNet and
SqueezeNet to 8-bit. The code for Ristretto is available.
-viii-
Acknowledgments
First and foremost, I want to thank my major advisor Professor Soheil Ghiasi for his
guidance, inspiration and encouragement that he gave me during my graduate studies.
Thanks to him, I had the privilege to do research in a dynamic research group with
excellent students. He provided my with all the ideas, equipment and mentorship I needed
for writing this thesis.
Second I would like to thank the graduate students at UC Davis who contributed to
my research. I was fortunate to work together with members of the LEPS Group, the
Architecture Group as well as the VLSI Computation Lab. Most notably, Mohammad
Motamedi, Terry O’Neill, Dan Fong and Joh Pimentel helped me with advice, technical
knowledge and paper editing. I consider myself extremely lucky that I had their support
during my graduate studies, and I look forward to continuing our friendship in the years
to come. I’m humbled by the opportunity to do research with Mohammad Motamedi,
a truly bright PhD student. Our early joint research projects motivated me to solve
challenging problems and to strive for extraordinary research results.
Third, I am grateful to all members of my thesis committee: Professor John Owens,
Venkatesh Akella, and Yong J. Lee. Professor Owens spurred me on to conduct an in-depth
analysis of related work; additionally he gave me valuable improvement suggestions for
my thesis. Early in this research project, Professor Akella guided me in reading papers
on hardware acceleration of neural networks. Professor Lee helped me significantly to
improve the final version of this document.
Finally I’d like to thank my family for supporting my studies abroad, and especially
my girlfriend Thirza. I am grateful to my family and friends for always motivating me to
pursue my academic goals; without them I would not have come this far.
-ix-
Chapter 1
Introduction
One of the major competitions in AI and computer vision is the ImageNet Large Scale
Visual Recognition Competition (Russakovsky et al., 2015). This annually held competition has seen state-of-the-art image classification accuracies by deep networks such as
AlexNet by Krizhevsky et al. (2012), VGG (Simonyan and Zisserman, 2015), GoogleNet
(Szegedy et al., 2015) and ResNet (He et al., 2015). All winners since 2012 have used
deep convolutional neural networks. These networks contain millions of parameters and
require billions of arithmetic operations.
Training of large networks like AlexNet is a very time-consuming process and can take
multiple days or even weeks. The training procedure of these networks is only possible
thanks to recent advancements in graphics processing units. High-end GPUs enable fast
deep learning, thanks to their large throughput and memory capacity. When training
AlexNet with Berkeley’s deep learning framework Caffe (Jia et al., 2014) and Nvidia’s
cuDNN (Chetlur et al., 2014), a Tesla K-40 GPU can process an image in just 4ms.
While GPUs are an excellent accelerator for deep learning in the cloud, mobile systems
are much more sensitive to energy consumption. In order to deploy deep learning algorithm in energy-constraint mobile systems, various approaches have been offered to reduce
the computational and memory requirements of convolutional neural networks (CNNs).
Various FPGA-based accelerators (Suda et al., 2016; Qiu et al., 2016) have proven that
it is possible to use reconfigurable hardware for end-to-end inference of large CNNs like
AlexNet and VGG. Moreover, we see an increasing amount of ASIC designs for deep
1
CNNs (Chen et al., 2016; Sim et al., 2016; Han et al., 2016a).
Before implementing hardware accelerators, a first crucial step consists of condensing
the neural network in question. Various work has been conducted recently to reduce
the computational and memory requirements of neural networks. However, no opensource project exists which would help a hardware developer to quickly and automatically
determine the best way to reduce the complexity of a trained neural network. Moreover,
too aggressive compression of neural networks leads to reduction in classification accuracy.
In this thesis we present Ristretto, a framework for automated neural network approximation. The framework is open source and we hope it will speed up the development
process of energy efficient hardware accelerators. Our framework focuses on condensing
neural networks without adding any computational complexity such as decompression or
sparse matrix multiplication. Ristretto is a Caffe-based approximation framework and was
developed for fast, automated, efficient and flexible CNN compression for later deployment
in hardware accelerators. Ristretto aims at lowering the area requirements for processing
elements, and lowering the memory footprint which in turn reduces or eliminates off-chip
memory communication.
This thesis analyses the resource-requirements of convolutional neural networks. Based
on these findings, different approximation strategies are proposed to reduce the resourceintense parts of CNNs. For all different approximation strategies, we present an in-depth
analysis of the compression vs accuracy trade-off.
Parts of this thesis are based on previous publications with two other authors: Mohammad Motamedi from the University of California, Davis and Professor S. Ghiasi from
the same university.
• Hardware-oriented Approximation of Convolutional Neural Networks,
Philipp Gysel, Mohammad Motamedi, and Soheil Ghiasi, arXiv preprint arXiv:
1604.03168 (2016). Gysel et al. (2016)
• PLACID: a Platform for Accelerator Creation for DCNNs, Mohammad
Motamedi, Philipp Gysel and Soheil Ghiasi, under review. Motamedi et al. (2016)
2
Chapter 2
Convolutional Neural Networks
Over time, different feature extraction algorithms have been used for image processing
tasks. SIFT (Lowe, 2004) and HOG (histogram of oriented gradients by Dalal and Triggs
(2005)) were state-of-art for feature extraction, but they relied on handcrafted features.
Neural networks in contrast can automatically create both high-level and low-level features. For a long time, deep neural networks were hindered by their computational complexity. However, advances in both personal computers and general purpose computing
have enable the training of larger networks with more parameters. In 2012, the first deep
convolutional neural network with 8 parameter layers was proposed by Krizhevsky et al.
(2012). State-of-the art deep CNNs use a series of convolutional layers which enables
them to extract very high-level features from images. Convolutional neural networks have
proven to overshadow conventional neural networks in complex vision tasks.
2.1
Training and Inference
Convolutional neural networks have two computation phases. The forward propagation is
used for classification, and the backward propagation for training. Like other algorithms
in machine learning, CNNs use a supervised training procedure to find network parameters which yield good classification accuracy. Throughout this thesis, we use the terms
parameters and weights interchangeably.
3
2.1.1
Forward Propagation
Input to the forward propagation is the image to classify. The forward path consists of
different layers which process the input data in a chained manner. Deep CNNs use many
such layers, the last of which is used to predict the image class.
A typical CNN architecture termed AlexNet is shown in Figure 2.1. As input to the
network, we use an RGB image with dimensions 224×224. The first five layers in AlexNet
are convolutional layers, and the last three are fully connected layers. Convolutional
layers use 2d filters to extract features from the input. The first convolutional layer
generates 2×48 feature maps, each of which represents the presence or absence of a lowlevel feature in the input image. To reduce the spatial dimension of feature maps and to
add translation-invariance, pooling layers are used which do sub-sampling. Moreover, a
non-linear layer is added which enables the network to learn non-linear features.
The last convolutional layer creates 2×128 feature maps with spatial dimension 13×13.
This layer is followed by three dense layers, which produce the weighted sum of inputs.
The last layer has 1000 output nodes, which are the predictions for the 1000 image classes.
Figure 2.1. Network architecture of AlexNet by Krizhevsky et al. (2012).
Forward propagation depends on network parameters in two ways. First, the convolutional layers rely on feature filters. Second, the fully connected layers contain many
parameters, each of which serves as weighted connection between a specific input and
output node. These parameters are learned during the training procedure.
4
2.1.2
Backward Propagation
Training of a deep CNN requires thousands or even millions of labeled images. The
network is exposed to these images, and the network parameters are gradually updated
to make up for prediction errors. The purpose of backward propagation is to find the
error gradient with respect to each network parameter. In a later step this error gradient
is used for a parameter update.
Training is done in batches of images. Several images are run through the network
in forward path. Denoting x as the network input, w as the network parameters and
f as the overall CNN function, the network output is given by z 0 = f (x, w). Since all
images are labeled, the desired network output z is known. Given many pairs of images
and ground-truth (x1 , z1 ), ..., (xn , zn ), we define a loss function l(z, z 0 ) which denotes the
penalty for predicting z 0 instead of z.
We average the loss for a batch of images, and update the parameters according to
the formula below:
wt+1 = wt − α ·
δl t
(w )
δw
(2.1)
This formula requires us to calculate the network output error w.r.t. each parameter.
The above parameter update is called stochastic gradient descent, which relies on a learning rate α. There actually exist many optimizations for this parameter update such as
Nesterov momentum as explained by Bengio et al. (2013) or Adam rule (Kingma and Ba,
2015). All these learning rules require the error gradient w.r.t each network parameter.
Moreover these optimization procedures all work in batch mode, i.e., a batch of images
(e.g. 256 images) is run through the network and the parameter update is based on the
average error gradient. The error surface of neural networks is non-convex. Nevertheless, batch-based learning rules can avoid local minima by using many different training
examples.
For computation of the error gradient w.r.t the network parameters, we first compute
the error gradient with respect to each layer output, starting with the second last layer
and back propagation to the second layer. In a second step, the gradients w.r.t. the layer
5
outputs can be used to compute the gradients w.r.t. the network parameters, using the
chain rule for derivatives.
2.2
Layer Types
Deep convolutional neural networks process the input data layer by layer. Each layer has
a specific purpose as explained in the following paragraphs.
Convolutional layer: This layer type is used for feature extraction. Deep CNNs have
many convolutional layers; AlexNet for example has five layers of this type. The feature
extraction is done by a series of L convolutional layers. Each layer uses 2d kernel filters
which extract features from input feature maps (IFM). The result of multiple feature
extractions is summed up to form one output feature map (OFM). This process is shown
in Figure 2.2, where two filter banks, each with three kernels, are used to generate 2
output feature maps. The number of input and output feature maps is denoted by N
and M , and the size of one output feature map is R×C. One kernel has dimension
K×K, and a layer has N ×M of these kernels. The feature extraction consists of a series
of multiplication-accumulation (MAC) operations, as shown in Figure 2.3. Each output
pixel is the sum of the 2d convolutions between the input feature maps and the respective
kernels. To generate the neighboring output pixels, the kernel stack is slided across the
spacial dimension by stride S.
The time and space complexity of convolutional layers is given in Equations 2.2 and
2.3. Assuming a stride S of 1, the computational complexity is R · C times larger than
the number of parameters. It is for this reason that the computational bottleneck of deep
CNNs comes from convolutional layers.
Figure 2.2. Convolution is the result of extracting 2d features from multiple feature
maps. Image credit: Motamedi et al. (2016).
6
Figure 2.3. Pseudo-code for convolutional layer. Image credit: Motamedi et al. (2016).
runtime = O(RCM N K 2 )
(2.2)
parameter size = O(M N K 2 )
(2.3)
Fully connected layer: Fully connected layers serve the same purpose as convolutional
layers, namely feature extraction. Fully connected layers, alternatively termed dense
layers, build the weighted sums of their input. Thus all input nodes are connected to
all output nodes, which requires a relatively large amount of network parameters. Fully
connected layers are the basic building block of classic neural networks, which are normally
a concatenation of dense layers. In convolutional neural networks, the first layers are
normally convolutional layers, and only one or two layer at the very end are dense.
The mathematical function of a fully connected layer is a simple matrix-vector product.
The layer output nodes z depend on the input vector x, the parameter-matrix W and the
bias b (Equation 2.4). Denoting N = |x| as the number of input nodes and M = |z| as
the number of outputs, the time and space complexity is given in Equations 2.5 and 2.6.
For layers with many input nodes, the parameter size usually makes up for a large part
of the overall network size.
Most fully connected layers are followed by a non-linear activation function f . Figure
2.4 depicts a fully connected layer followed by an activation layer.
z=W·x+b
7
(2.4)
runtime = O(N M )
(2.5)
parameter size = O(N M )
(2.6)
Figure 2.4. Fully connected layer followed by an activation function. Image credit:
Karpathy.
Rectified Linear Unit (ReLU): To enable neural networks to capture non-linear
relations, several non-linear layers are inserted into the network architecture. Traditionally
we used the Sigmoid or hyperbolic tangent function for this purpose. However, those
classic non-linear functions have several disadvantages. First, their gradient becomes very
small for large values, which means the error gradient during training vanishes. Moreover
these two non-linear functions are relatively costly in terms of computing power.
As alternative, nearly all state-of-art CNNs use Rectified Linear Units (ReLU). This
activation function was first proposed by Nair and Hinton (2010) for Restricted Boltzmann
Machines. The work by Krizhevsky et al. (2012) was the first to apply this simplified
activation to a deep neural networks. Deep networks trained with this activation function
converge much faster.
The function of a ReLU layer maps negative values to zero: f (x) = max(0, x).
Normalization layers: Local Response Normalization (LRN) layers serve the purpose
of normalization across feature maps or across the spatial dimension. There are generally
two type of LRN layers: LRN across feature maps and LRN within feature maps. The
8
first type of normalization serves the purpose of normalizing neurons in the same spatial
position, but different feature maps. This creates competition between neurons generated
by different kernels and improves the top-1 accuracy of AlexNet (Krizhevsky et al., 2012)
by 1.4%. The exact mathematical formulation of LRN across channels can be found in
the same paper.
Many state-of-art CNNs use LRN layers to increase accuracy (Krizhevsky et al., 2012;
Szegedy et al., 2015). One notable exception is the network by Simonyan and Zisserman
(2015) which performs very well without any kind of normalization.
Most recently, a new normalization strategy termed Batch Normalization (BN) was
proposed (Ioffe and Szegedy, 2015). This strategy was adopted by the most recent winner
of the ILSVRC competition (Russakovsky et al., 2015).
While both normalization strategies help for faster convergence and better prediction
accuracy, they also add computational overhead, especially batch normalization. Unfortunately these normalization layers require a very large dynamic range for intermediate
values. In AlexNet for example, the intermediate values of LRN layers are 214 times larger
than any intermediate value from another layer. For this reason this thesis assumes LRN
and BN layers are to be implemented in floating point, and we concentrate on the approximation of other layer types. Notice that previous work by Suda et al. (2016) chose
32-bit floating point for FPGA-based LRN layers.
Figure 2.5. Max pooling layer with 2×2 filter size. Image credit: Karpathy.
Pooling: Pooling layers are normally used between successive convolutional layers in
CNNs. They can be considered as sub-sampling functions. The purpose of pooling layers
9
is to reduce the spatial dimension of feature maps and encode translation invariance. By
reducing the feature map dimensions, we also reduce the parameter size and computational
complexity, and thus lower chances for over-fitting.
The most commonly used pooling type in CNNs is MAX pooling, although other
types such as average pooling and L2-norm pooling exist. MAX pooling does not required
any arithmetic operation except comparisons. Since this layer type is cheap in terms of
computation and parameter size, this thesis leaves this layer type as is and we perform
no approximation.
An example of a MAX pooling operation is shown in Figure 2.5, where 2x2 kernels
are used to extract the maximum value. The result is stored as output on the right side
of the Figure. Pooling layers are associated with a kernel size and a stride. The stride
indicates how many pixels are skipped between pooling operations. Notice that in the
above example, the data size is reduced by 75%.
2.3
Applications
Deep convolutional neural networks have pushed the limits of artificial intelligence in
a wide range of applications. Recent winners (Krizhevsky et al., 2012; Simonyan and
Zisserman, 2015; Szegedy et al., 2015; He et al., 2015) of ImageNet (Russakovsky et al.,
2015) competition have continuously improved machine’s abilities in image classification.
The most recent winners of this competition even outperform human vision. Besides image
classification, deep networks show state-of-art performance in object detection (Girshick,
2015) as well as speech recognition (Hinton et al., 2012). Other applications include
playing games (Silver et al., 2016), as well as art (Gatys et al., 2015). Notice that artificial
neural networks as well as recurrent neural networks could potentially be approximated
in a similar way as we do in this thesis.
10
2.4
Computational Complexity and Memory Requirements
The complexity of deep CNNs can be split into two parts. First, the convolutional layers
contain more than 90% of the required arithmetic operations. The second resource-intense
layer type are fully connected layers, which contain over 90% of the network parameters.
An energy-efficient accelerator for CNNs needs to 1) offer a large enough computational
throughput and 2) offer a memory-bandwidth that can keep the processing elements busy.
Figure 2.6. Parameters and arithmetic operations in CNNs. Data from CaffeNet is
colored green, data from VGG-16 network is colored violet.
CaffeNet is the Caffe-version of AlexNet by Krizhevsky et al. (2012). CaffeNet was
developed for the ImageNet data set, which has 1000 image classes. Figure 2.6 shows the
required arithmetic operations and parameters size of AlexNet by layer type. The major
part of arithmetic operations comes from convolutional layers: this layer type requires a
total of 2.15 G operations. The arithmetic operations in all other layers sum up to 117 M
operations. The parameter size of CaffeNet is 250 MB, of which 235 MB comes from fully
connected layers.
The same trend can be observed for the 16-layer version of VGG by Simonyan and Zisserman (2015): Extracting features in convolutional layers is computation-intense, while
fully connected layers are memory-intense. Since most computation and memory require-
11
ments come from fully connected and convolutional layers, this thesis concentrates on
approximating these two layer types only.
2.5
ImageNet Competition
The ILSVRC competition (Russakovsky et al., 2015) is a large scale image classification
and detection challenge which has been held annually since 2010. More than fifty institutions have participated in the challenge, among them companies like Google, Microsoft,
Qualcomm, as well as various universities. In its first year the competition consisted of
an image classification challenge only. In the meantime, different new challenges were introduced such as object detection and localization in images, object detection from video
as well as scene classification.
For training of the classifier, ILSVRC provides the ImageNet data set, which consisted
originally of 1000 image categories and a total of 1.2 million images. In the meantime,
this data base has been significantly expanded, and currently holds over 14 million labeled
images.
A common performance measure for deep CNNs is their classification accuracy on
the ILSVRC 2012 data set. For this challenge, the classifier is trained on a training
data set, and tested by the researcher on a validation data set. For the official score in
the ILSVRC competition, a private test data set is used which is not available to the
public. Most researchers give their performance numbers in top-1 and top-5 measure.
The top-1 measure gives the percentage of images that were classified correctly on either
the validation or test set. Since the ImageNet data set has very fine-grained image classes
which are sometimes even hard for humans to distinguish, most researchers prefer to use
the top-5 measure. Here, an image is considered as classified correctly if one of the top5 predictions is correct. In the remainder of this thesis, we will concentrate on top-1
accuracy, since this captures better how CNN approximation affects network accuracy.
2.5.1
Network Size vs Accuracy
On one hand, recent network architectures indicate that deeper networks perform better.
The winner of 2014’s localization challenge (Simonyan and Zisserman, 2015) experimented
12
with different depths for their network. Going from 11 parameter layers to 19 improved
their top-1 classification for over 4%. Another experiment by the winner of 2015 (He
et al., 2015) used very deep networks. Their network architecture improves by over 2%
when expanding the net from 34 to 152 layers.
On the other hand, some research shows that even relatively small networks can
achieve good classification performance. In the classification challenge of 2014, GoogLeNet
(Szegedy et al., 2015) outperformed VGG (Simonyan and Zisserman, 2015) with a network
capacity that was over 19X smaller. The GoogLeNet network is based on the inception
idea described in section 2.5.2. A newer network architecture by Iandola et al. (2016)
uses an adapted inception concept with smaller convolutional kernels. Their small, yet
quite accurate network termed SqueezeNet was developed specifically for mobile devices.
SqueezeNet has the accuracy of AlexNet (Krizhevsky et al., 2012), but contains 50X
fewer parameters. SqueezeNet avoids fully connected layers, which drastically reduces the
parameter size to below 5 MB.
In summary, there is a delicate trade-off between network size and accuracy (see
Figure 2.7). GoogLeNet finds a good balance for reducing both classification error and
network size. This network by Google outperforms AlexNet and VGG in both aspects.
SqueezeNet is the smallest architecture, but its accuracy is not outstanding. ResNet has
the best accuracy, but it is also the largest network both in terms of parameter layers and
parameter size.
Figure 2.7. ImageNet networks: accuracy vs size.
13
2.5.2
Inception Idea
The inception idea is a concept proposed by Szegedy et al. (2015) which was used to
build the GoogLeNet architecture. GoogLeNet was developed with the goal of achieving
high classification accuracy on the ImageNet data set with a budget of 1.5 billion MAC
operations for inference. The authors avoid sparse representations and chose to use readily
available dense components. The GoogLeNet architecture contains several replicas of the
inception module shown in Figure 2.8. An inception module contains 1×1, 3×3 and
5×5 convolution kernels. In order to reduce the number of feature maps and thus the
computational complexity, 1×1 convolutions are added in front of the expensive 3×3 and
5×5 convolutions. The inception module also contains an alternative max pooling path.
Figure 2.8. Inception architecture used by GoogLeNet. Image credit: Szegedy et al.
(2015).
2.6
Neural Networks With Limited Numerical Precision
This section focuses on limited precision arithmetic for convolutional neural networks.
Most deep learning frameworks (Jia et al., 2014; Theano Development Team, 2016; Abadi
et al., 2015) use 32-bit or 64-bit floating point for CNN training and inference. However, it
has be shown (Du et al., 2015) that CNNs have a relatively high error resilience; moreover
CNNs can be trained in a discrete parameter space. In the following subsections, we
14
describe the process of quantizing a full precision network to limited precision numbers.
Additionally we introduce different rounding schemes for the quantization step, and we
describe different options for optimizing the classification accuracy of a limited precision
network.
2.6.1
Quantization
Figure 2.9. Simulated data path for convolutional and fully connected layers. The
layer inputs, layer outputs and layer parameters are discrete-valued numbers.
As discussed previously, the goal of this thesis is to provide a framework for approximating the forward path of any given CNN. For this purpose we compress the number
format in convolutional and fully connected layers. These two layer types, which are the
most resource-demanding part of a deep network, require the same arithmetic operations,
namely a series of multiplication-and-accumulation (MAC). In this thesis we simulate the
arithmetic of a hardware accelerator. The simulated data path is shown in Figure 2.9.
The difference between this simulated data path and the original full precision data
path is the quantization step of weights, layer inputs, and layer outputs. Therefore the
condensed networks will suffer from quantization errors, which can affect the network
accuracy.
In this thesis we propose a framework which can approximate 32-bit floating point
networks by condensed ones which use quantized values. In order to simulate a condensed
15
layer, given a full precision reference network, the following three steps are required:
• Quantization of the layer input and weights to reduced precision format (using m
and n bits for number representation, respectively)
• Perform the MAC operations using the quantized values
• The final result is again quantized
These steps are summarized in Figure 2.10.
Figure 2.10. Our quantization framework performs these steps to simulate the hardware arithmetic.
2.6.1.1
Data path in hardware
In Figure 2.9, the layer input values and layer weights serve as input to the multiplication
units. To leverage the inherent sources of parallelism, a hardware accelerator will use
many of these units in parallel. Each multiplication unit gets one layer input value and
one network weight per computation round. The different results are accumulated in an
adder tree, and the final sum is the layer output. Notice that some of the layer input
and weight values are actually reused for different multipliers, depending on the exact
characteristics of the layer in question.
To reduce the number of bits required for number representation, our approximation
framework quantizes both the layer inputs and weights. The resulting values can be
represented using significantly fewer bits. As a result, the multiplication units require
less area. In order to simplify simulation of hardware, our framework uses 32-bit floating
point for accumulation. To achieve the same result in a hardware accelerator, the adder
16
tree should use 32-bit. Adders are much cheaper in terms of area and power, compared
to multipliers. Thus it is acceptable to use more bits for number representation in the
adder tree.
As a last step, the layer output is quantized to limited-precision format. This reduces
the memory requirements, especially if results need to be written back to off-chip memory.
2.6.2
Rounding Schemes
Different rounding schemes can be used for value quantization.
Round nearest even: Round-nearest-even is an unbiased scheme which rounds to the
nearest discrete value. Denoting as the quantization step size and bxc as the largest
quantization value less or equal to x, Gupta et al. (2015) define round nearest as follows:
round(x) =
bxc,
if bxc ≤ x ≤ x +
bxc + , if bxc +
2
2
(2.7)
<x≤x+
As round-nearest-even is deterministic, we chose this rounding scheme for inference, i.e.,
at test time all the parameters are rounded deterministically, and so are the layer outputs.
Round stochastic: Another rounding scheme termed stochastic rounding was used by
Gupta et al. (2015) for the weight updates of 16-bit neural networks. Gupta et al. (2015)
define stochastic rounding as follows:
round(x) =
bxc,
w.p. 1 −
bxc + , w.p.
x−bxc
(2.8)
x−bxc
Stochastic rounding adds randomness to the quantization procedure, which can have
an averaging effect during training. We chose to use this rounding scheme when quantizing network parameters during fine-tuning, as explained later in the next subsection.
Stochastic rounding has an expected rounding error of zero, i.e. E(round(x)) = 0.
2.6.3
Optimization in Discrete Parameter Space
Training of neural networks can be seen as an optimization problem, where the goal is to
find the optimal set of parameters which minimizes the classification error for a given set
17
of images. A practical solution to this problem is to use stochastic gradient descent, as
explained in subsection 2.1.1. In the traditional setting of 64-bit floating point training,
this optimization is a continuous problem with a smooth error surface. The error surface of
neural networks depends on its input and its current parameters. For quantized networks,
this error surface becomes discrete. This new optimization problem – where the goal is to
find an optimal set of discrete valued parameters – is an NP-hard problem. One approach
to find a good set of discrete parameters is to train in limited precision ‘from scratch’.
In this approach, we would train the network with quantized parameters right from the
start. All the weight updates would be discrete.
We chose to use another approach: Our framework first trains a network in the continuous
domain, then quantizes the parameters, and finally fine-tunes in discrete parameter space.
This way we can fully leverage pre-trained networks which saves considerable amount of
time.
During this retraining procedure, the network learns how to classify images with limited precision parameters. Since the network weights can only have discrete values, the
main challenge consists in the weight update. We adopt the idea of previous work by
Courbariaux et al. (2015) which uses full precision shadow weights. Small weight updates ∆w are applied to the full precision weights w, whereas the discrete weights w0 are
sampled from the full precision weights. The sampling during fine-tuning is done with
stochastic rounding. For more details on the fine-tuning procedure of quantized networks,
please refer to section 9.3.
18
Chapter 3
Related Work
In the first section of this chapter, we review different network approximation techniques.
In the second part, we describe related work in hardware accelerator design.
3.1
Network Approximation
3.1.1
Fixed Point Approximation
Various solutions have been offered to reduce the resource-requirement of CNNs. Traditionally neural networks are trained in 32-bit floating point. However fixed point arithmetic is less resource hungry than floating point arithmetic. Moreover, it has been shown
that fixed point arithmetic is adequate for neural network computation. This observation has been leveraged recently to condense deep CNNs. Gupta et al. (2015) show that
networks on datasets like CIFAR-10 (10 images classes) can be trained in 16-bit. Further
trimming of the same network uses as low as 7-bit multipliers (Courbariaux et al., 2014).
Another approach by Courbariaux et al. (2015) uses only binary weights, again on the
same network. A similar proposal represents the weights of the same network with +1,
0 and -1 values (Sung et al., 2015). While these proposed fixed point schemes work well
with small networks, only limited work has been done on large CNNs like AlexNet.
3.1.2
Network Pruning and Shared Weights
Off-chip memory access makes up for a significant part of the total energy budget of any
data-intense application. As deep CNNs have typically more than 10 MB of parameters,
19
an important step is to compress the size of the network parameters. The deep compression pipeline proposed by Han et al. (2016b) addresses this problem. The authors achieve
a network parameter compression rate of up to 49X for deep CNNs using a three-step
pipeline. In a first step, the ’unimportant’ connections of a trained network are removed.
The resulting sparse network is then retrained to regain its classification accuracy, and the
pruning step is repeated. After some iterations of pruning and fine tuning, the remaining
parameters are clustered together to form shared weights. These shared weights are again
fine-tuned to find optimal centroids. In a last step, a lossless data compression scheme
(Huffman Coding) is applied to the final weights.
3.1.3
Binary Networks
Since memory access has a relatively high energy cost, it is desirable to reduce the network
parameter size. This motivated BinaryConnect (Courbariaux et al., 2015), a work which
represents weights in binary format, rather than in traditional 32-bit floating point. This
approach reduces parameter size by factor 32X and removes the need of multiplications
in the forward path. BinaryConnect achieves near-state-of-art performance on 10-class
datasets (MNIST, CIFAR-10, SVHN).
A later work by Lin et al. (2016) takes this idea a step further by turning multiplications in the backward propagation into bit shifts. Layer activations are approximated
by integer power of 2 numbers, while error gradients are retained in full precision. This
proposal significantly reduces the hardware requirements for accelerators.
Combining the two previous ideas, ‘Binarized Neural Network’ (Courbariaux et al.,
2016) uses binary weights and layer activations. These numbers are constraint to +1 and
-1 for both forward and backward propagation. Convolutional neural networks mainly
consist of multiply-accumulate operations. For a binarized network, these operations are
replaced by binary XNOR and binary count. To improve training results, the proposed
method uses a bit-shift-based batch normalization as well as a shift-based parameter
update.
Finally the work of Rastegari et al. (2016) applies the idea of binary networks to
ImageNet data set. The three previously mentioned approaches work well with small
20
networks; however they show limited performance on large networks for the ImageNet data
set. The work by Rastegari et al. (2016) proposes two network architectures. Both are
based on AlexNet (Krizhevsky et al., 2012) and use different degrees of binarization. First,
the proposed Binary-Weights-Network shows a speedup of 2X for CPU implementation
and achieves an accuracy within 3% of AlexNet. Second, XNOR- Net has binary weights
and layer outputs. XNOR-Net turns most MAC operations into binary XNOR and bit
count, however at a relatively high accuracy drop (12.4%).
3.2
Accelerators
Different accelerator platforms have been used to accelerate CNN inference. In what
follows we review proposed accelerators on GPUs, FPGAs and ASICs.
3.2.1
GPU Acceleration
Given the high throughput and memory bandwidth of today’s GPUs, different research
has focused on accelerating GPU-based inference of neural networks. A proposal by
Denton et al. (2014) uses clustered filters and low rank approximation. They achieve a
speedup of 2X for convolutional layers of AlexNet, compared to a non-optimized GPU implementation. Another work by Mathieu et al. (2013) achieves better results by replacing
convolution through FFT. Finally the neural network compression pipeline proposed by
Han et al. (2016b) uses pruning and weight-sharing. When this compression is applied to
dense layers of AlexNet, forward propagation is 4X faster and 3X more energy efficient.
In their later paper (Han et al., 2016a), they show a Titan X based GPU implementation
with a throughput of 3.23 TOPs.
High-end GPUs require a lot of energy. As a case in point, Nvidia’s Tesla K-40 has an
average power consumption of 235 W when running DGEMM. This motivated researchers
to consider accelerator platforms with smaller power budgets.
3.2.2
FPGA-based Accelerators
Field programmable gate arrays (FPGA) can offer high throughput per power. FPGAbased accelerators have a shorter development time than ASICs, however they can’t match
the throughput of GPUs. Different FPGA-based accelerators for neural networks have
21
been proposed. An approach by Zhang et al. (2015) uses Vivado HLS to accelerate the
convolutional layers of AlexNet. Their floating point implementation achieves a throughput of over 60 GFLOPs at a power budget of 20 W.
A subsequent proposal by Suda et al. (2016) uses OpenCL to implement whole VGG
(Simonyan and Zisserman, 2015) net on an Altera Stratix V board. Their throughputoptimized design achieves an overall throughput of 117.8 GOPs. Finally a recent Xilinxbased implementation (Qiu et al., 2016) achieves the start-of-art throughput of 137 GOPs.
Their 16-bit fixed point implementation requires less than 10 W.
3.2.3
Custom Accelerators (ASIC)
Custom architectures have the highest throughput and energy efficiency, however their
design time is significant. DaDianNao by Chen et al. (2014) is a super-computer for
machine learning at 28 nm technology. Their chip relies on large on-chip memory (which
takes up nearly half of the area) and achieves significant speedups and power savings
compared to the GPU. A later implementation termed Eyeriss (Chen et al., 2016) can
run the convolutional layers of AlexNet in forward path at 34 frames per second (74.6
GOPs), using only 0.278 W. The chip is about 2X slower than a throughput optimized
embedded GPU, but 13X more energy efficient. Eyeriss uses 16-bit fixed point. Finally
EIE (Han et al., 2016a) is an ASIC which leverages the deep compression pipeline (Han
et al., 2016b). EIE infers significantly compressed networks 13X faster than a GeForce
GTX Titan X. Since the neural network is pruned and has shared weights, the whole
network fits to on-chip memory which allows to infer images using just 0.6 W.
3.2.4
Comparison Of Accelerator Platforms
In this section we compare different accelerator platforms in terms of throughput and
throughput per power. We take the performance numbers from recently published papers,
the source of our numbers can be found in Table 3.1.
22
Table 3.1. Throughput and power consumption of different accelerator platforms. All
works implement an ImageNet network.
Platform
Throughput
Power
Throughput
per power
Source
ASIC
74.6 GOP/s
278 mW
268 GOP/s/W
Chen et al. (2016)
Xilinx Zynq ZC706
137 GOP/s
9.63 W
14.2 GOP/s/W
Qiu et al. (2016)
NVIDIA TK1
155 GOP/s
10.2 W
15.2 GOP/s/W
Chen et al. (2016)
Titan X
3.23 TOP/s
250 W
12.9 GOP/s/W
Han et al. (2016a)
The ASIC design by Chen et al. (2016) is optimized for large networks and low power
consumption. Their work concentrates on convolutional layers of AlexNet. Other works
which concentrate on fully connected layers only show similar throughput (Han et al.,
2016a). Predictably, the ASIC design shows the highest throughput per power (see Figure
3.1).
Figure 3.1. Throughput and throughput per power for different accelerator platforms.
Numbers are taken from Table 3.1.
For GPU performance, we consider an implementation which uses cuBLAS for fully
connected layers of AlexNet. Convolutional layers would yield lower throughput, since
this layer type requires rearrangement of data before the matrix-matrix multiplication.
23
The acceleration on the GPU achieves the highest throughput among all accelerators (see
Figure 3.1). The high GPU throughput of 3.23 TOP/s comes at a relatively high power
cost (250 W). Embedded GPUs require less power, but their throughput is proportionally
lower. As a case in point, we consider the mobile GPU implementation of AlexNet
by Chen et al. (2016). When comparing the two GPU implementations, the mobile
GPU’s throughput per power is only slightly better than that of the high-end GPU (15.2
GOP/s/W vs 12.9 GOP/s/W).
The FPGA implementation from Qiu et al. (2016) is an end-to-end implementation of
the 16-layer version of VGG. The FPGA implementation uses 16-bit fixed point arithmetic
to reduce memory and computation requirements. Moreover the authors use pruning
in fully connected layers to reduce parameter size. Another work by Suda et al. (2016)
achieves nearly the same throughput without weight pruning. The FPGA implementation
is head-to-head with the embedded GPU implementation. The latter has 11% more
throughput and 6% more throughput per power.
24
Chapter 4
Fixed Point Approximation
This chapter covers the approximation of convolutional neural networks with fixed point
numbers. While normal inference is done in 32-bit floating point, using bit-width reduced
format for intermediate results can increase both throughput and energy efficiency of
hardware accelerators.
4.1
Baseline Convolutional Neural Networks
In the remainder of this document, we will discuss different approaches for approximating
CNNs in a hardware friendly manner. In each section, we approximate the following
CNNs:
1. LeNet1 was proposed by LeCun et al. (1998). This network consists of two convolutional and two fully connected layers and can be used to classify handwritten
digits (MNIST dataset).
2. The CIFAR-10 data set (Krizhevsky, 2009) has 10 image classes such as airplanes,
bird, and truck. The CIFAR-10 Full model2 was developed by Caffe for the
CIFAR-10 data set. The network has three convolutional layers followed by one
fully connected layer. Moreover, the model has two local response normalization
(LRN) layers.
1 https://github.com/BVLC/caffe/blob/master/examples/mnist/lenet_train_test.prototxt
2 https://github.com/BVLC/caffe/blob/master/examples/cifar10/cifar10_full_train_test.prototxt
25
3. CaffeNet3 is the Caffe version of AlexNet (Krizhevsky et al., 2012) which is the
winner of the 2012 ILSVRC competition. This network can classify images into
the 1000 ImageNet categories, which vary from animal and plant species to various
human-made objects. CaffeNet has five convolutional layers, three fully connected
layers and two LRN layers. CaffeNet has 60 million parameters and 650,000 neurons.
4. GoogLeNet4 was proposed by Szegedy et al. (2015) and won the 2014 ILSVRC
competition. This network is based on the inception idea, which uses convolutional
and pooling layers with small kernel sizes. GoogLeNet has 12X fewer parameters
than AlexNet but manages to improve the accuracy significantly.
5. SqueezeNet5 by Iandola et al. (2016) was developed with the goal of a small
network with the accuracy of AlexNet (Krizhevsky et al., 2012). SqueezeNet relies
on convolutional layers with 1x1 and 3x3 kernels. No fully connected layers or
normalization layers are needed.
4.2
Fixed Point Format
Standard inference of deep neural networks uses 32-bit floating point. We replace the
parameter and layer outputs with the following fixed point number format: [IL.FL], where
IL and FL denote the integer and fractional length of the numbers, respectively. The
number of bits used to represent each value is therefor IL+FL. To quantize floating point
numbers to fixed point, we use round-nearest. We use 2s-complement numbers, thus the
largest positive value we can represent is:
xmax = 2IL−1 − 2−FL
(4.1)
Note that in the following experiments, all truncated numbers use a shared fixed point
format, i.e., they share the same integer and fractional length. For a representation using
dynamic adaption of integer and fractional part, please refer to chapter 5.
3 https://github.com/BVLC/caffe/blob/master/models/bvlc_reference_caffenet/train_val.prototxt
4 https://github.com/BVLC/caffe/wiki/Model-Zoo
5 https://github.com/DeepScale/SqueezeNet/blob/master/SqueezeNet_v1.0/train_val.prototxt
26
4.3
Dynamic Range of Parameters and Layer Outputs
In this subsection we analyze the dynamic range of numbers in two neural networks. This
analysis will help to understand the optimal choice for integer and fractional bits in fixed
point representation.
4.3.1
Dynamic Range in Small CNN
Figure 4.1. Dynamic range of values in LeNet.
In a first step we do this analysis for LeNet. We performed the forward propagation of 100
images to compute intermediate values in the network. The value distribution is shown in
Figure 4.1. Note that this histogram data is the result of truncating all values to integer
power of two. We can see that on average, parameters are smaller than layer outputs.
99% of the trained network parameters are between 20 and 2−10 . For fully connected
layers however, 99% of the layer outputs are in the range 25 ...2−4 .
In order to quantize both the layer outputs and network parameters to 8-bit fixed
point, a part of the values needs to be saturated. We achieved the best quantization
results with the Q.4.4 format. This indicates that large layer outputs are more important
than small network parameters.
27
4.3.2
Dynamic Range in Large CNN
This subsection contains the analysis for the relatively large CaffeNet network. We performed the forward propagation of 50 images on a trained CaffeNet network. The resulting
dynamic range is shown in Figure 4.2. Similarly to a small network, parameters tend to
be smaller than layer outputs. However, for this large network, the average difference
between these two number categories is much larger. This is to be expected, since layer
outputs are the result of a multiplication accumulation process, which yields a much larger
result for big layers. As a case in point, we can compare the relatively large second parameter layer (447.9M MAC operations) with the relatively small last parameter layer (4.1M
MAC operations). While the second layer’s largest value is larger than 29 , all values are
below 26 in the last layer.
Figure 4.2. Dynamic range of values in CaffeNet.
Since the dynamic range of values is much larger than in LeNet, more bits are required
for a fixed point representations. Our experiments show the best 16-bit fixed point results
when using the Q9.7 format. Notice that a significant part of the network parameters
gets saturated in this quantization, since there are not enough fractional bits. Only very
few layer outputs (0.46% in convolutional layers) are too large to be represented, while a
large part of the parameters (21.23%) is truncated to zero. Similarly to the analysis with
LeNet, large layer outputs are more important than small parameters.
28
4.4
Results
This subsection covers the results of quantizing trained 32-bit floating point networks to
fixed point.
4.4.1
Optimal Integer and Fractional Length
For a given network and target bit-width, the layer outputs and network parameters of
convolutional and fully connected layers all share the same fixed point format. The bitwidth is the sum of integer and fractional length. The choice of fractional length is crucial
and will decide which values need to be saturated. Our quantization procedure tries
different partitionings of the bit-width into integer and fractional part. The best setting
is retained and the resulting fixed point network is fine-tuned. Notice that different
choices for integer and fractional length are conceivable. For example, only the layer
output quantization could be considered to find a good partitioning, since the network
parameters can be adapted in the fine-tuning step. However, our experiments on three
different networks show that a joint optimization of layer outputs and parameters yields
the best results after fine-tuning.
4.4.2
Quantization to Fixed Point
We quantized three of our baseline networks to fixed point: LeNet, CIFAR-10 and CaffeNet. To calculate relative accuracy of a bit-width reduced network, we divide the fixed
point accuracy by the 32-bit floating point accuracy. First we consider the relatively small
LeNet network for handwritten digit recognition. The quantization from 32-bit floating
point to 8-bit fixed point incurs a relative accuracy loss of 10.3% (see Figure 4.3). After
fine-tuning, the absolute accuracy loss shrinks to 0.27% (Table 4.1), indicating LeNet
works well in 8-bit fixed point.
29
Figure 4.3. Normalized classification accuracy of fixed point networks.
The second baseline network we consider is CIFAR-10. This network classifies images
into classes such as truck, ship, dog, bird. As this is a more challenging task which requires
a larger network, the layer outputs are larger too, and the network is more sensitive to
quantization errors. When the network is quantized to 8-bit, the network output is random
and the accuracy drops to 10%. Since our quantization framework is unable to achieve
good results at 8-bit, we double the bit-width. The network works fine in 16-bit, with
a relative accuracy loss below 1%. The best results were achieved using 8 integer bits,
whereas LeNet only required 4 integer bits (see Table 4.1).
Table 4.1. Fine-tuned fixed point networks. Numbers in brackets indicate accuracy
without fine-tuning.
Network
LeNet
CIFAR-10
CaffeNet top-1
Baseline Fixed point
accuracy bit-width
99.15%
8-bit
81.69%
16-bit
56.90%
16-bit
Fixed point Fixed point
accuracy
format
Q4.4
98.88% (88.90%)
Q8.8
81.38% (80.94%)
Q9.7
52.48% (52.13%)
Finally we quantize CaffeNet, a network for ImageNet classification. As before, 8-bit
quantization yields poor results, which is the reason we choose 16-bit fixed point. The
30
relative accuracy loss after quantization is 8.4% and the fine-tuned network achieves an
accuracy within 4.5% of the baseline (compared in absolute values).
In order to increase the accuracy of the quantized networks, we introduce dynamic
fixed point in the next section.
31
Chapter 5
Dynamic Fixed Point Approximation
In this chapter we discuss quantization of a floating point CNN to a dynamic fixed point
version. We extend the fixed point format to dynamic fixed point, and show how it can
be used to further decrease parameter size while maintaining a high prediction accuracy.
5.1
Mixed Precision Fixed Point
Figure 5.1. Data path of a fixed point convolutional or fully connected layer.
The data path of fully connected and convolutional layers consists of a series of MAC operations (multiplication and accumulation), as shown in Figure 5.1. The layer activations
are multiplied with the network weights, and these multiplication results are accumulated
to form the output.
32
As shown by Lin et al. (2015); Qiu et al. (2016), it is a good approach to use mixed
precision, i.e., different parts of a CNN use different bit-widths. In Figure 5.1, m and n
refer to the number of bits used to represent layer outputs and layer weights, respectively.
Multiplication results are accumulated using an adder tree which gets thicker towards the
end. The adder outputs in the first level are m + n + 1 bits wide, and the bit-width grows
by 1 bit in each level. In the last level, the bit-width is m + n + lg2 (x), where x is the
number of multiplication operations per output value. In the last stage, the bias is added
to form the layer output. For each network layer, we need to find a good balance between
reducing the bit-widths (m and n) and maintaining a good classification accuracy.
5.2
Dynamic Fixed Point
The different parts of a CNN have a significant dynamic range. In large layers, the
outputs are the result of thousands of accumulations, thus the network parameters are
much smaller than the layer outputs. Fixed point has only limited capability to cover
a wide dynamic range. Dynamic fixed point can be a good solution to overcome this
problem, as shown by Courbariaux et al. (2014). In dynamic fixed point, each number is
represented as follows:
s
−F L
(−1) · 2
B−2
X
2i · xi
(5.1)
i=0
Here B denotes the bit-width, s the sign bit, F L is the fractional length, and x the
mantissa bits. Since the intermediate values in a network have different ranges, it is
desirable to group fixed point numbers into groups with constant FL. So the number of
bits allocated to the fractional part is constant within that group, but different compared
to other groups. Each network layer is split into two groups: one for the layer outputs,
one for the layer weights. This allows to better cover the dynamic range of both layer
outputs and weights, as weights are normally significantly smaller. On the hardware side,
it is possible to realize dynamic fixed point arithmetic using bit shifters.
33
Figure 5.2. Dynamic fixed point with variable length of fractional part.
The concept of dynamic fixed point is depicted in Figure 5.2, where two numbers
are both represented in 8 bits, but belong to a different group (i.e., they have different
fractional length).
5.2.1
Choice of Number Format
When we approximate a neural network with dynamic fixed point numbers, we need to
choose a number format for each number group. Each layer has two such groups: the
layer parameters and the layer outputs. Within each group, all numbers are represented
using the same integer and fractional length.
To find the optimal set of number formats, we could perform an exhaustive search,
however this is not efficient for large neural networks. Instead, we follow a specific rule
that automatically determines the required number of integer bits. More specifically, we
choose enough bits to avoid saturation. So for a given set of numbers S, the required
integer length IL is given by Equation 5.2.
IL = dlg2 (max x + 1)e
S
(5.2)
This relation defines the integer length of layer parameters. For layer outputs, we
reduce the integer length by one, since our experiments show slightly better results this
way.
5.3
Results
In this section we present the results of approximating 32-bit floating point networks
by condensed dynamic fixed point models. All classification accuracies were obtained
34
running the respective network on the whole validation dataset. We follow the general
approximation procedure explained in section 2.6.
5.3.1
Impact of Dynamic Fixed Point
We used our Ristretto framework to quantize CaffeNet (AlexNet) into fixed point, and
compare traditional fixed point with dynamic fixed point. To allow a simpler comparison,
all layer outputs and network parameters share the same bit-width. Results show a good
performance of static fixed point for as low as 18-bit (Figure 5.3). However, when reducing
the bit-width further, the accuracy starts to drop significantly, while dynamic fixed point
has a stable accuracy. We can conclude that dynamic fixed point performs significantly
better for such a large network. The reason is that dynamic fixed point allows us to adapt
the number of bits allocated to integer and fractional part, according to the dynamic range
of different parts of the network.
Figure 5.3. Top-1 accuracy for CaffeNet on ILSVRC 2014 validation dataset. Integer
length refers to the number of bits assigned to the integer part of fixed point numbers.
5.3.2
Quantization of Individual Network Parts
In this section, we present the results for approximating different parts of a network. For
each experiment, only one category is quantized to dynamic fixed point, and the rest
remains in full precision. Table 5.1 shows the quantization impact for three different
networks. For each network, we quantize the layer outputs, the convolutional kernels
(CONV), and the parameters of fully connected layers (FC) independently. In all three
35
nets, the convolution kernels and layer activations can be trimmed to 8-bit with an absolute accuracy change of only 0.3%. Fully connected layers are more affected from trimming
weights to 8-bit, the absolute change is maximally 0.9%. Interestingly, LeNet weights can
be trimmed to as low as 2 bits, with absolute accuracy change below 0.4%.
Table 5.1. Dynamic fixed point quantization results for different parts of network.
Only one number category is cast to fixed point, and the remaining numbers are in
floating point format.
Fixed point bit-width
16-bit
8-bit
4-bit
2-bit
LeNet, 32-bit floating point accuracy: 99.1%
Layer output
CONV parameters
FC parameters
99.1% 99.1% 98.9% 85.9%
99.1% 99.1% 99.1% 98.9%
99.1% 99.1% 98.9% 98.7%
Full CIFAR-10, 32-bit floating point accuracy: 81.7%
Layer output
CONV parameters
FC parameters
81.6% 81.6% 79.6% 48.0%
81.7% 81.4% 75.9% 19.1%
81.7% 80.8% 79.9% 77.5%
CaffeNet Top-1, 32-bit floating point accuracy: 56.9%
Layer output
CONV parameters
FC parameters
5.3.3
56.8% 56.7% 06.0% 00.1%
56.9% 56.7% 00.1% 00.1%
56.9% 56.3% 00.1% 00.1%
Fine-tuned Dynamic Fixed Point Networks
Here we report the accuracy of five networks that were condensed and fine-tuned with
Ristretto. All networks use dynamic fixed point parameters as well as dynamic fixed point
layer outputs for convolutional and fully connected layers. LeNet performs well in 2/4bit, while CIFAR-10 and the three ImageNet CNNs can be trimmed to 8-bit (see Table
5.2). Surprisingly, these compressed networks still perform nearly as well as their floating
point baseline. The relative accuracy drops of LeNet, CIFAR-10 and SqueezeNet are very
small (<0.6%), whereas the approximation of the larger CaffeNet and GoogLeNet incurs
a slightly higher cost (0.9% and 2.3% respectively).
36
Table 5.2. Fine-tuned, dynamic fixed point networks. Numbers in brackets indicate
accuracy without fine-tuning.
Network
Layer
outputs
CONV
FC
parameters parameters
LeNet (Exp 1)
LeNet (Exp 2)
Full CIFAR-10
CaffeNet
SqueezeNet
GoogLeNet
4-bit
4-bit
8-bit
8-bit
8-bit
8-bit
4-bit
2-bit
8-bit
8-bit
8-bit
8-bit
4-bit
2-bit
8-bit
8-bit
8-bit
8-bit
32-bit
baseline
99.15%
99.15%
81.69%
56.90%
57.68%
68.92%
Fixed point
accuracy
98.95%
98.81%
81.44%
56.00%
57.09%
66.57%
(98.72%)
(98.03%)
(80.64%)
(55.77%)
(55.25%)
(66.07%)
The SqueezeNet (Iandola et al., 2016) architecture was developed with the goal of a
small CNN that performs well on the ImageNet data set. Ristretto can make the already
small network even smaller, so that its parameter size is less than 2 MB. This condensed
network is well- suited for deployment in smart mobile systems.
All five 32-bit floating point networks can be approximated well in 8-bit and 4-bit
fixed point. For a hardware implementation, this reduces the size of multiplication units
by about one order of magnitude. Moreover, the required memory bandwidth is reduced
by 4–8X. Finally, it helps to hold 4–8X more parameters in on-chip buffers.
Some previous work (Courbariaux et al., 2014) concentrated on training with fixed
point arithmetic from the start and shows little performance decline for as short as 7bit fixed point numbers on LeNet. Our approach is different in that we train with high
numerical precision, then quantize to fixed point, and finally fine-tune the fixed point
network. Our condensed model achieves superior accuracy with as low as 4-bit fixed
point, on the same data set. While more sophisticated data compression schemes could
be used to achieve higher network size reduction, our approach is very hardware friendly
and imposes no additional overhead such as decompression.
37
Chapter 6
Minifloat Approximation
6.1
Motivation
Chapters 4 and 5 concentrated on fixed point approximation of deep CNNs. Since the
training of neural networks is normally done in floating point, it is an intuitive approach
to condense these models to smaller floating point numbers. This section analyses the
network approximation through minifloat, i.e., floating point numbers with 16 bits or
smaller.
6.2
IEEE-754 Single Precision Standard
According to IEEE-754 standard, single precision numbers have 1 sign bit, 8 exponent
bits and 23 mantissa bits. The mantissa’s first bit (always ’1’) is added implicitly, and
the stored exponent is biased by 127. Numbers with all zeros or ones in the exponent
have a special meaning. An exponent with all zeros either represents the number 0 or
a denormalized number, depending on the mantissa bits. For the case of all ones in the
exponent, the number represented is either +/-INF or NaN.
6.3
Minifloat Number Format
In order to condense networks and reduce their computational and memory requirements,
we will represent floating point numbers with much fewer bits than the IEEE-754 standard.
We follow the standard to a large degree when going to 12-bit, 8-bit, or 6-bit numbers,
but our format differs in some details. Namely, the exponent bias is lowered according to
38
the number of bits assigned to the exponent:
bias = 2bits−1 − 1
(6.1)
Here bits denotes the number of bits assigned to the exponent. Another difference to
the IEEE standard is that we don’t support denormalized numbers, INF and NaN. INF is
replaced by saturated numbers, and denormalized numbers are replace by 0. Finally, the
number of bits assigned to the exponent and mantissa part don’t follow a specific rule.
To be more precise, our Ristretto framework automatically searches for the best balance
between exponent and mantissa bits. As a case in point, a 16-bit minifloat number (Figure
6.1) could be represented with 1 sign bit, 5 exponent bits and 10 mantissa bits.
Figure 6.1. Example of a 16-bit minifloat number with 5 exponent bits and 10 mantissa
bits.
6.3.1
Network-specific Choice of Number Format
Similar to dynamic fixed point, we need to choose a specific number format per bit-width,
i.e., partition the available bits into exponent and mantissa. In order to approximate a
neural network with minifloat numbers, we need to find a suitable number of exponent
and mantissa bits. We use enough exponent bits to avoid saturation:
bits = dlg2 (lg2 (max x) − 1) + 1e
S
(6.2)
S is the set of numbers which we approximate. This choice of exponent bits assures
no saturation happens, under the assumption that we use infinitely many mantissa bits.
6.4
Data Path for Accelerator
The data path of convolutional and fully connected layers is depicted in Figure 6.2. For
simplicity, we only consider fixed precision arithmetic, i.e., all number categories shared
the same minifloat format. Similar to the fixed point data path, network parameters and
39
layer inputs are multiplied and accumulated. Input to each multiplier is a pair of numbers,
each in minifloat format. The output of each multiplier is 3 bits wider than the input
numbers. In a next step, the multiplication results are accumulated in full precision. In a
last step the bias is added in minifloat format, and the final result is trimmed to minifloat.
When implemented in a hardware accelerator, the data path’s input and output are
minifloat numbers. In case the neural network in question is too large to fit into on-chip
memory, the layer outputs and parameters need to be stored in off-chip memory. Since
both these number categories are represented in minifloat, we can achieve significant
energy savings thanks to reduced data transfer.
Figure 6.2. Data path of a minifloat convolutional or fully connected layer.
6.5
Results
In this section, we analyze the impact of lowering the bit-width of floating point numbers.
We used our approximation framework to query different CNNs which use minifloat numbers in convolutional and fully connected layers. To find the accuracy of the condensed
networks, we follow the quantization flow described in section 2.6.
We quantized three CNNs to 12, 8 and 6-bit minifloat. The quantization is done
for layer outputs and parameters of fully connected and convolutional layers. For each
network, we show the classification accuracy of both the 32-bit baseline, followed by
minifloat versions (Figure 6.3). We calculate the normalized accuracy by dividing the
40
minifloat network’s performance by the 32-bit floating point network accuracy.
The results indicate that LeNet has no classification loss when shrinking layer outputs
and layer parameters to 8-bit. CIFAR-10 and CaffeNet can be used in 12-bit, again with
no loss in accuracy.
Figure 6.3. Normalized classification accuracy of minifloat networks.
We fine-tuned the 8-bit versions of the three networks. Table 6.1 shows the accuracy of
the minifloat networks. CIFAR-10 has an absolute accuracy drop below 1%, and CaffeNet
incurs an absolute drop of 4.6%. For LeNet, minifloat actually increases accuracy. Minifloat adds more regularization to LeNet and increases the accuracy by 0.05%, compared
to the 32-bit network.
Table 6.1. Fine-tuned minifloat networks. Numbers in brackets indicate accuracy
without fine-tuning.
Network
32-bit
accuracy
LeNet
CIFAR-10
CaffeNet top-1
99.15%
81.69%
56.90%
Minifloat Minifloat
Exponent bits,
bit-width accuracy
mantissa bits
8-bit
99.20% (99.20%) 4-bit, 3-bit
8-bit
80.85% (80.47%) 5-bit, 2-bit
8-bit
52.52% (52.30%) 5-bit, 2-bit
41
Compared to dynamic fixed point results in section 5.3.3, minifloat requires more bits.
For 8-bit CaffeNet, the absolute accuracy drop of dynamic fixed point is small (below
1%), whereas minifloat incurs a relatively large drop (4.38%).
6.6
Comparison to Previous Work
Previous work (Deng et al.) approximated network parameters of AlexNet (CaffeNet)
with 8-bit minifloat. They analyze the impact of quantizing a varying percentage of the
network parameters. Our results achieve significantly better accuracy, thanks to a careful
choice of minifloat format and a fine-tuning step.
42
Chapter 7
Turning Multiplications Into Bit
Shifts
7.1
Multiplier-free Arithmetic
Hardware accelerators for convolutional neural networks need to be energy-efficient to
allow for deployment in mobile devices. Fully connected layers and convolutional layers
consist of additions and multiplications, of which the latter requires a much larger chip
area. This motivated previous research to eliminate all multipliers by using integer power
of two weights (Tang and Kwan, 1993; Mahoney and Elhanany, 2008). These weights can
be considered as minifloat numbers with zero mantissa bits. By using such weights, all
multiplications turn into bit shifts, which can save a significant amount of energy on a
hardware accelerator.
We now detail the approximation of convolutional and fully connected layers with
multiplier-free arithmetic. Although we assume strictly positive weights in this discussion, it is straight forward the expand the approximation procedure to both positive and
negative weights. The computation of convolutional and fully connected layers consists
of multiplication-and-accumulation operations. Equation 7.1 shows the necessary operations in a full precision network. The layer inputs xj are multiplied with layer parameters
wj and the accumulation yields the result zi . To simplify this discussion, we assume the
input data has been rearranged such that output zi = wT · x. In order to switch to
multiplier-free arithmetic, we first approximate parameters by the closest integer-power-
43
of-two number (Equation 7.2). Now output zi can be approximated by equation 7.3 which
is multiplier-free. Notice that the last equation relies on the power-of-two exponents ej ,
not the original parameters.
zi =
X
xj · w j
(7.1)
j
ej = round(lg2 (wj ))
zi ≈
X
(7.2)
xj <<ej
(7.3)
j
7.2
Maximal Number of Shifts
Nearly all network weights of a trained network are between +1 and −1, but most of them
are close to zero. The quantization of these parameters to power-of-two has the highest
impact (in terms of absolute value change) to the weights close to +1 and −1. These
weights will only be able to take on the values 1, 12 ,
1
4
and so on.
We encode the number of shifts in 4 bits (see Figure 7.1). This implies the parameter
exponents can have 8 different values. We choose to represent the exponent values such
that ei ∈ [−8, ..., −1] and use this format for the subsequent experiments.
The motivation for this format is two-fold. First of all, using only 4 bits for parameters
reduces the memory requirements tremendously. Second, the smallest possible value in
this format is 2−8 . Parameters smaller than that have only a minor effect on the network
output. Moreover only few parameters are below this smallest value. Our analysis from
section 4.3.2 shows only 10.97% of parameters in CaffeNet are lower than the smallest
possible value. For LeNet, this percentage is even smaller (5.83%).
Figure 7.1. Example of a 4-bit parameter.
44
7.3
Data Path for Accelerator
The adder-only arithmetic of convolutional and fully connected layers is shown Figure
7.2. The 4-bit parameters indicate how many bit-shifts are required for the layer inputs.
To enable shifts by multiple bits in one clock cycle, barrel shifters should be used. Notice
that this data path has no multipliers at all, which can potentially save significant chip
area. To simplify this analysis, we only focus on the impact of removing multiplications.
The layer inputs and outputs are kept in full precision format.
Figure 7.2. Data path of a convolutional or fully connected layer.
7.4
Results
Table 7.1. Classification accuracy of nets with power-of-two weights. Numbers in
brackets indicate accuracy without fine-tuning.
32-bit floating Multiplier-free
point weights network
LeNet
99.15%
99.16% (99.09%)
CIFAR-10
81.69%
77.48% (74.65%)
CaffeNet top-1 56.90%
53.25% (50.58%)
Network
We used our Ristretto framework to simulate the effect of removing all multiplications
from convolutional and fully connected layers. Our framework quantizes all network
45
parameters to the nearest integer-power-of-two number. Table 7.1 compares networks
with power-of-two weights and networks with single precision weights.
For LeNet, the absolute classification accuracy drop for the quantized weights is 0.1%.
CIFAR-10 and CaffeNet are more affected by the weight quantization (4.21% and 3.65%
absolute accuracy drop). At first glance, the results for the larger two networks might be
discouraging. However, it is surprising that the nets with weight quantization still have
a decent classification accuracy. The power-of-two weights can be stored in just 4 bits
(the exponents range from -1 to -8). This allows for tremendous energy savings: First
the traffic to off-chip memory is reduced, as the weights are not 32-bit but 4-bit. Second,
multipliers are replaced with simple bit shifters.
46
Chapter 8
Comparison of Different
Approximations
In this chapter, we compare the different approximation strategies for convolutional neural
networks. For this purpose, we consider three networks: LeNet, CIFAR-10 and CaffeNet.
We analyze how well the approximation schemes can lower the bit-width without hurting
accuracy. In all experiments, the parameters and layer outputs of convolutional and fully
connected layers are condensed to smaller bit-width. The approximation results without
fine-tuning are shown in Figures 8.1, 8.2, 8.3. For all three neural networks, dynamic fixed
point has the best performance, followed by minifloat approximation. All approximation
schemes perform well at 16-bit, but as we lower the bit-width the accuracy drops.
Figure 8.1. Approximation of LeNet.
47
Figure 8.2. Approximation of CIFAR-10.
Figure 8.3. Approximation of CaffeNet.
8.1
Fixed Point Approximation
Fixed point is the approximation scheme that requires the least energy and development
time for a hardware accelerator. However, it is also the approximation with the poorest
performance for small bit-widths. For CaffeNet for example, the dynamic range of values
is significant, as shown in subsection 4.3.2. In 15-bit convolutional layers, 0.45% of layer
outputs are too large and need to be represented in saturated format. When moving to
14- bit fixed point, 2.82% of the layer outputs are saturated. Since large layer outputs are
very important for the network’s accuracy, this leads to a significant accuracy drop (see
48
Figure 8.1).
8.2
Dynamic Fixed Point Approximation
Dynamic fixed point shows the best performance among the three approximation schemes.
Dynamic fixed point combines the advantages of both fixed point and minifloat: On one
hand this format allows to use all bits for the mantissa part which helps for a good
accuracy. On the other hand dynamic fixed point can cover a large dynamic range,
just like floating point, thanks to the exponent that is stored implicitly. LeNet can be
approximated with just 5-bit numbers, achieving the same accuracy as the 32-bit floating
point model. The same holds true for an 8-bit CIFAR-10 network. Finally the CaffeNet
architecture can be approximated with 8-bit dynamic fixed point, at an absolute accuracy
drop of 0.3%.
8.3
Minifloat Approximation
Floating point numbers can cover a large dynamic range, thanks to their exponent. Minifloat performs significantly better than static fixed point. However minifloat approximation shows a sharp accuracy drop when going to very low bit-widths. This sharp drop is
at the point where there are not enough bits for the exponent. For LeNet for example,
the accuracy of 5-bit arithmetic is 97.96%. In this setting, we use 4 bits for the exponent,
no mantissa bits and one sign bit. When we lower the bit-width further, the exponent is
unable to cover the dynamic range of values, and the accuracy drops sharply to 10.09%.
For the other two networks, we can see a similar effect. Both CIFAR-10 and CaffeNet
need 5 exponent bits, according to Equation 6.2. Since we need one more bit for the
sign, those two networks need at least 6-bit minifloat numbers in order to achieve good
classification performance.
8.4
Summary
Dynamic Fixed point is very well suited for approximation of neural networks. This
approximation shows the best accuracy at low bit-widths. Although dynamic fixed point
requires some more chip area than pure fixed point arithmetic, this approximation is very
49
well suited for hardware acceleration of neural networks. The bit-width can be reduced
to 4-bit or 8-bit for LeNet, CIFAR-10, and CaffeNet. This reduces the required memory
bandwidth and footprint significantly, which is expected to yield significant energy savings
for FPGA and ASIC designs.
50
Chapter 9
Ristretto: An Approximation
Framework for Deep CNNs
9.1
From Caffe to Ristretto
According to Wikipedia, Ristretto is ‘a short shot of espresso coffee made with the normal
amount of ground coffee but extracted with about half the amount of water’. Similarly,
our compressor removes the unnecessary parts of a CNN, while making sure the essence the ability to predict classes from images - is preserved. With its strong community and
fast training for deep CNNs, Caffe created by Jia et al. (2014) is an excellent framework to
build on. Ristretto takes a trained model as input, and automatically brews a condensed
network version. Input and output of Ristretto are a network description file (prototxt)
and the network parameters. The condensed model in Caffe-format can then be used for
a hardware accelerator.
9.2
Quantization Flow
Ristretto can condense any 32-bit floating point network to either fixed point, minifloat
or integer power of two parameters. Ristretto’s quantization flow has five stages (Figure
9.1). In the first step, the dynamic range of the weights is analyzed to find a compressed
number representation. For dynamic fixed point, Ristretto allocates enough bits to the
integer part to avoids saturation of large values. Similarly, for minifloat approximation,
the framework makes sure enough bits are allocated to the exponent. To quantize full
51
precision numbers into a smaller number format, Ristretto uses round nearest even.
The second step runs several thousand images in forward path. The generated layer
activations are analyzed to generate statistical parameters. Ristretto allocates enough
bits to the new number format to avoid saturation of layer activations.
Next Ristretto performs a binary search to find the optimal number of bits for convolutional weights, fully connected weights, and layer outputs. For dynamic fixed point, a
certain network part is quantized, while the rest remains in floating point. Since there are
three network parts that should use independent bit-widths, iteratively quantizing one
network part allows us to find the optimal bit-width for each part. Once a good trade-off
between small number representation and classification accuracy is found, the resulting
network can be fine-tuned.
Figure 9.1. Network approximation flow with Ristretto.
9.3
Fine-tuning
In order to make up for the accuracy drop incurred by quantization, the quantized network
is fine-tuned in Ristretto. During this retraining procedure, the network learns how to
classify images with discrete-valued parameters w0 . During fine-tuning, we will calculate
small weight updates ∆w. Since these small weight updates may be below the quantization
step size of the discrete parameters, we also keep a set of full precision weights w.
Ristretto uses the fine-tuning procedure shown in Figure 9.2. For each batch, the full
precision weights are quantized to reduced-precision format. During forward propagation,
these discrete weights are used to compute the layer outputs zl . Each layer l turns its
input batch xl into output zl , according to its function fl : (xl , w0 ) → zl . Assuming the
last layer computes the loss, we denote f as the overall CNN function. The goal of back
52
propagation is to compute the error gradient
δf
δw
with respect to each quantized parameter.
For parameter updates we use the Adam rule by Kingma and Ba (2015). As an important
observation, we do not quantize layer outputs during fine-tuning. We use floating point
layer outputs instead, which enables Ristretto to analytically compute the error gradient
with respect to each parameter. In contrast, scoring of the network is done with reduced
precision layer outputs.
Figure 9.2. Fine-tuning with shadow weights. The left side shows the training process with full-precision shadow weights. On the right side the fine-tuned network is
benchmarked on the validation data set. Quantized values are represented in orange.
To achieve the best fine-tuning results, we used a learning rate that is an order of
magnitude lower than the last full precision training iteration. Since the choice of hyper parameters for retraining is crucial (Bergstra and Bengio, 2012), Ristretto relies on
minimal human intervention in this step.
9.4
Fast Forward and Backward Propagation
Ristretto brews a condensed network with reduced precision weights and layer activations.
For simulation of the forward propagation in hardware, Ristretto uses full floating point
for accumulation. This follows the thought of Gupta et al. (2015) and is conform with
our description of the forward data path in subsection 2.6.1.1. During fine-tuning, the
full precision weights need to be quantized for each batch, but after that all computation
can be done in floating point (Figure 9.2). Therefore Ristretto can fully leverage opti-
53
mized matrix- matrix multiplication routines for both forward and backward propagation.
Thanks to its fast implementation on the GPU, a fixed point CaffeNet can be tested on
the ILSVRC 2014 validation dataset (50k images) in less than 2 minutes (using one Tesla
K-40 GPU).
9.5
Ristretto From a User Perspective
Ristretto is based on the highly optimized Caffe-framework and follows its principles. A
Caffe user will appreciate the smooth and easy-to-understand integration of Ristretto.
Figure 9.3. Network training with Caffe, the base framework of Ristretto.
Caffe: Development of an image processing algorithm in Caffe starts with a network
description file (see Figure 9.3). This file is written by the user and contains the hyper
parameters of the neural network architecture. The Caffe framework uses the Google
Protocol Buffer format to encode the network information. Networks are represented as
directed acyclic graphs. The vertices are layers which do computation based on the input
data and layer parameters. Data flows from one layer to another in so called ‘blobs’.
As second item, the Caffe user needs a labeled training data set. Finally Caffe requires
a solver file, which contains the hyper parameters for training, such as initial learning rate
and training duration.
Once the network description file, the data set and the solver are prepared, the Caffe
tool can be used for training. The result of training is a file containing the trained
network parameters. This parameter file – together with the network description file –
can be deployed for classification of arbitrary images.
54
Figure 9.4. Network quantization and fine-tuning with Ristretto.
Ristretto: Our approximation framework can be seen as a feature extension to Caffe.
In fact, for the network compression pipeline described in Figure 9.4, some steps require
the traditional Caffe tool (which was left as-is). Ristretto starts where Caffe ends: a
trained network serves as input to the quantization pipeline. The user has several options
to choose from for quantization, for instance he can set the error margin and an approximation strategy. The Ristretto tool quantizes the trained 32-bit floating point network to
the smallest possible bit-width representation. Ristretto produces the network description
of the condensed model, which follows the format of Caffe. The condensed model contains
limited-precision layers, which are a Ristretto-specific feature. Moreover each layer has
quantization parameters, such as the number of integer and fractional bits in the case of
fixed point approximation.
At this point, the quantized model description could be used to score the network on
a data set. However, in order to increase accuracy, the user is advised to fine-tune the
new model. The user writes a new solver file which will be used for fine-tuning. The Caffe
tool will fine-tune the condensed model to achieve the highest possible accuracy.
9.6
Release of Ristretto
Ristretto is released as open source project and has the following strengths:
• Automation: Ristretto performs automatic trimming of any given CNN.
• Flexibility: Various trimming schemes are supported.
55
• Accuracy: Ristretto fine-tunes trimmed networks.
• Speed: Ristretto runs on the GPU and leverages optimized CUDA routines.
Ristretto has a homepage1 and the source code is available2 .
9.7
Future Work
Ristretto follows the modular source code architecture of Caffe. New features such as new
limited precision layer types can be added to Ristretto easily. In this section we discuss
different possible future steps.
9.7.1
Network Pruning
The most energy-costly operation for CNN accelerators is off-chip memory access. Since
large networks don’t fit into on-chip memory, it is imperative to compress the network.
Most network weights come from fully connected layers. It has been shown that a significant part of the connections in fully connected layers can be removed. Previous work
(Han et al., 2016b) achieves high network compression rates with no loss in classification
accuracy.
9.7.2
Binary Networks
The first published work to represent ImageNet networks with binary weights was by
Rastegari et al. (2016). Their results show that very deep networks can be approximated
with binary weights, although at an accuracy drop of around 3% for CaffeNet and 6% for
GoogLeNet. Substituting 32-bit floating point parameters with just one bit of information
necessarily reduces the network’s accuracy to extract features from images. The challenge
with binary parameters is to achieve high prediction accuracy without increasing the
parameter size or adding additional computational overhead.
9.7.3
C-Code Generation
Many academic projects use high-level synthesis tools for FPGA and ASIC based accelerators. The standard development tool-chain starts with a C-implementation of the
1 http://ristretto.lepsucd.com/
2 https://github.com/pmgysel/caffe
56
algorithm, which then undergoes many unit tests to verify correct functionality. In a
next step the C-code is manually converted to System-C which serves as input to high
level synthesis (HLS). Highly optimized HLS tools can produce very efficient Verilog code
within a fraction of the time which would be needed for manual Verilog coding. We plan
to add a feature to Ristretto which allows for automatic generation of the C-code of a
condensed network. This feature will produce the necessary code files as well as a dump
of the extracted low-precision parameters.
We hope that Ristretto will enable researchers to speed up their development time
for CNN accelerators. Hopefully Ristretto will gain traction in the community. As it is
an open-source project, the community can help adding new features to the framework,
which will make the tool even more powerful.
57
References
Abadi, M., Agarwal, A., Barham, P., Brevdo, E., Chen, Z., Citro, C., Corrado, G. S.,
Davis, A., Dean, J., Devin, M., Ghemawat, S., Goodfellow, I., Harp, A., Irving, G.,
Isard, M., Jia, Y., Jozefowicz, R., Kaiser, L., Kudlur, M., Levenberg, J., Mané, D.,
Monga, R., Moore, S., Murray, D., Olah, C., Schuster, M., Shlens, J., Steiner, B.,
Sutskever, I., Talwar, K., Tucker, P., Vanhoucke, V., Vasudevan, V., Viégas, F., Vinyals,
O., Warden, P., Wattenberg, M., Wicke, M., Yu, Y., and Zheng, X. TensorFlow: LargeScale Machine Learning on Heterogeneous Systems, 2015. URL http://tensorflow.
org/.
Bengio, Y., Boulanger-Lewandowski, N., and Pascanu, R. Advances in Optimizing Recurrent Networks. In 2013 IEEE International Conference on Acoustics, Speech and
Signal Processing (ICASSP), pages 8624–8628. IEEE, 2013.
Bergstra, J. and Bengio, Y. Random Search for Hyper-Parameter Optimization. The
Journal of Machine Learning Research, 13(1):281–305, 2012.
Chen, Y.-H., Krishna, T., Emer, J., and Sze, V. Eyeriss: An Energy-Efficient Reconfigurable Accelerator for Deep Convolutional Neural Networks. In IEEE International
Solid-State Circuits Conference, ISSCC 2016, Digest of Technical Papers, pages 262–
263, 2016.
Chen, Y., Luo, T., Liu, S., Zhang, S., He, L., Wang, J., Li, L., Chen, T., Xu, Z., Sun,
N., and Temam, O. DaDianNao: A Machine-Learning Supercomputer. In Proceedings
of the 47th Annual IEEE/ACM International Symposium on Microarchitecture, pages
609–622. IEEE Computer Society, 2014.
Chetlur, S., Woolley, C., Vandermersch, P., Cohen, J., Tran, J., Catanzaro, B., and
Shelhamer, E.
cuDNN: Efficient Primitives for Deep Learning.
arXiv:1410.0759, 2014.
58
arXiv preprint
Courbariaux, M., David, J.-P., and Bengio, Y. Training Deep Neural Networks with Low
Precision Multiplications. arXiv preprint arXiv:1412.7024, 2014.
Courbariaux, M., Bengio, Y., and David, J.-P. BinaryConnect: Training Deep Neural
Networks with binary weights during propagations. In Advances in Neural Information
Processing Systems, pages 3105–3113, 2015.
Courbariaux, M., Hubara, I., Soudry, D., El-Yaniv, R., and Bengio, Y. Binarized Neural
Networks: Training Deep Neural Networks with Weights and Activations Constrained
to +1 or -1. arXiv preprint arXiv:1602.02830, 2016.
Dalal, N. and Triggs, B. Histograms of oriented gradients for human detection. In IEEE
Computer Society Conference on Computer Vision and Pattern Recognition, 2005, volume 1, pages 886–893. IEEE, 2005.
Deng, Z., Xu, C., Cai, Q., and Faraboschi, P.
Reduced-Precision Memory Value
Approximation for Deep Learning. http://www.labs.hpe.com/techreports/2015/
HPL-2015-100.html. Accessed: 2016-05-09.
Denton, E. L., Zaremba, W., Bruna, J., LeCun, Y., and Fergus, R. Exploiting Linear
Structure Within Convolutional Networks for Efficient Evaluation. In Advances in
Neural Information Processing Systems, pages 1269–1277, 2014.
Du, Z., Lingamneni, A., Chen, Y., Palem, K. V., Temam, O., and Wu, C. Leveraging the
Error Resilience of Neural Networks for Designing Highly Energy Efficient Accelerators.
Computer-Aided Design of Integrated Circuits and Systems, IEEE Transactions on, 34
(8):1223–1235, 2015.
Gatys, L. A., Ecker, A. S., and Bethge, M. A Neural Algorithm of Artistic Style. arXiv
preprint arXiv:1508.06576, 2015.
Girshick, R. Fast R-CNN. In Proceedings of the IEEE International Conference on
Computer Vision, pages 1440–1448, 2015.
59
Gupta, S., Agrawal, A., Gopalakrishnan, K., and Narayanan, P. Deep Learning with
Limited Numerical Precision. In Proceedings of the 32nd International Conference on
Machine Learning (ICML-15), pages 1737–1746, 2015.
Gysel, P., Motamedi, M., and Ghiasi, S. Hardware-oriented Approximation of Convolutional Neural Networks. arXiv preprint arXiv:1604.03168, 2016.
Han, S., Liu, X., Mao, H., Pu, J., Pedram, A., Horowitz, M. A., and Dally, W. J.
EIE: Efficient Inference Engine on Compressed Deep Neural Network. arXiv preprint
arXiv:1602.01528, 2016a.
Han, S., Mao, H., and Dally, W. J. Deep Compression: Compressing Deep Neural Networks with Pruning, Trained Quantization and Huffman Coding. In International
Conference on Learning Representations, 2016b.
He, K., Zhang, X., Ren, S., and Sun, J. Deep Residual Learning for Image Recognition.
arXiv preprint arXiv:1512.03385, 2015.
Hinton, G., Deng, L., Yu, D., Dahl, G. E., Mohamed, A.-R., Jaitly, N., Senior, A.,
Vanhoucke, V., Nguyen, P., Sainath, T. N., et al. Deep Neural Networks for Acoustic
Modeling in Speech Recognition: The Shared Views of Four Research Groups. Signal
Processing Magazine, IEEE, 29(6):82–97, 2012.
Iandola, F. N., Moskewicz, M. W., Ashraf, K., Han, S., Dally, W. J., and Keutzer, K.
SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and <0.5MB model
size. arXiv:1602.07360, 2016.
Ioffe, S. and Szegedy, C. Batch Normalization: Accelerating Deep Network Training by
Reducing Internal Covariate Shift. In Proceedings of The 32nd International Conference
on Machine Learning, pages 448–456, 2015.
Jia, Y., Shelhamer, E., Donahue, J., Karayev, S., Long, J., Girshick, R., Guadarrama,
S., and Darrell, T. Caffe: Convolutional Architecture for Fast Feature Embedding.
60
In Proceedings of the ACM International Conference on Multimedia, pages 675–678.
ACM, 2014.
Karpathy, A. Stanford CS231n course: Convolutional Neural Networks for Visual Recognition. http://cs231n.github.io/neural-networks-1/. Accessed: 2016-05-09.
Kingma, D. and Ba, J. Adam: A Method for Stochastic Optimization. In International
Conference on Learning Representations, 2015.
Krizhevsky, A. Learning Multiple Layers of Features from Tiny Images. Master’s thesis,
University of Toronto, 2009.
Krizhevsky, A., Sutskever, I., and Hinton, G. E. ImageNet Classification with Deep
Convolutional Neural Networks. In Advances in Neural Information Processing Systems,
pages 1097–1105, 2012.
LeCun, Y., Bottou, L., Bengio, Y., and Haffner, P. Gradient-Based Learning Applied to
Document Recognition. Proceedings of the IEEE, 86(11):2278–2324, 1998.
Lin, D., Talathi, S., and Annapureddy, S. Fixed Point Quantization of Deep Convolutional
Networks. arXiv preprint arXiv:1511.06393, 2015.
Lin, Z., Courbariaux, M., Memisevic, R., and Bengio, Y. Neural Networks with Few
Multiplications. In International Conference on Learning Representations, 2016.
Lowe, D. G. Distinctive Image Features from Scale-Invariant Keypoints. International
Journal of Computer Vision, 60(2):91–110, 2004.
Mahoney, V. and Elhanany, I. A backpropagation neural network design using adder-only
arithmetic. In 51st Midwest Symposium on Circuits and Systems, 2008, pages 894–897.
IEEE, 2008.
Mathieu, M., Henaff, M., and LeCun, Y. Fast Training of Convolutional Networks through
FFTs. arXiv preprint arXiv:1312.5851, 2013.
61
Motamedi, M., Gysel, P., and Ghiasi, S. PLACID: A Platform for Accelerator Creation
for DCNNs. Under review, 2016.
Nair, V. and Hinton, G. E. Rectified Linear Units Improve Restricted Boltzmann Machines.
In Proceedings of the 27th International Conference on Machine Learning
(ICML-10), pages 807–814, 2010.
Qiu, J., Wang, J., Yao, S., Guo, K., Li, B., Zhou, E., Yu, J., Tang, T., Xu, N., Song,
S., Wang, Y., and Yang, H. Going Deeper with Embedded FPGA Platform for Convolutional Neural Network. In Proceedings of the 2016 ACM/SIGDA International
Symposium on Field-Programmable Gate Arrays, pages 26–35, 2016.
Rastegari, M., Ordonez, V., Redmon, J., and Farhadi, A. XNOR-Net: ImageNet Classification Using Binary Convolutional Neural Networks. arXiv preprint arXiv:1603.05279,
2016.
Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy,
A., Khosla, A., Bernstein, M., Berg, A. C., and Fei-Fei, L. ImageNet Large Scale
Visual Recognition Challenge. International Journal of Computer Vision, 115(3):211–
252, 2015.
Silver, D., Huang, A., Maddison, C. J., Guez, A., Sifre, L., Van Den Driessche, G.,
Schrittwieser, J., Antonoglou, I., Panneershelvam, V., Lanctot, M., et al. Mastering
the game of Go with deep neural networks and tree search. Nature, 529(7587):484–489,
2016.
Sim, J., Park, J.-S., Kim, M., Bae, D., Choi, Y., and Kim, L.-S. A 1.42 TOPS/W deep
convolutional neural network recognition processor for intelligent IoE systems. In 2016
IEEE International Solid-State Circuits Conference (ISSCC), pages 264–265. IEEE,
2016.
Simonyan, K. and Zisserman, A. Very Deep Convolutional Networks for Large-Scale
Image Recognition. In International Conference on Learning Representations, 2015.
62
Suda, N., Chandra, V., Dasika, G., Mohanty, A., Ma, Y., Vrudhula, S., Seo, J.-s., and
Cao, Y. Throughput-Optimized OpenCL-based FPGA Accelerator for Large-Scale Convolutional Neural Networks. In Proceedings of the 2016 ACM/SIGDA International
Symposium on Field-Programmable Gate Arrays, pages 16–25. ACM, 2016.
Sung, W., Shin, S., and Hwang, K. Resiliency of Deep Neural Networks under Quantization. arXiv preprint arXiv:1511.06488, 2015.
Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S., Anguelov, D., Erhan, D., Vanhoucke,
V., and Rabinovich, A. Going Deeper with Convolutions. In Proceedings of the IEEE
Conference on Computer Vision and Pattern Recognition, pages 1–9, 2015.
Tang, C. Z. and Kwan, H. K. Multilayer Feedforward Neural Networks with Single
Powers-of-Two Weights. Signal Processing, IEEE Transactions on, 41(8):2724–2727,
1993.
Theano Development Team. Theano: A Python framework for fast computation of
mathematical expressions. arXiv e-prints, abs/1605.02688, May 2016. URL http:
//arxiv.org/abs/1605.02688.
Zhang, C., Li, P., Sun, G., Guan, Y., Xiao, B., and Cong, J. Optimizing FPGA-based
Accelerator Design for Deep Convolutional Neural Networks. In Proceedings of the 2015
ACM/SIGDA International Symposium on Field-Programmable Gate Arrays, pages
161–170. ACM, 2015.
63
| 9cs.NE
|
Computing Maximum Flow with Augmenting Electrical Flows
arXiv:1608.06016v1 [cs.DS] 21 Aug 2016
Aleksander Mądry∗
MIT
[email protected]
Abstract
10 1
e m 7 U 7 -time algorithm for the maximum s-t flow problem (and the
We present an O
minimum s-t cut problem) in directed
with m arcs and largest integer capacity U . This
graphs
10
e
7
-time algorithm of Mądry [43] in the unit-capacity
matches the running time of the O (mU )
e (m√n log U )-time algorithm of Lee and Sidford
case, and improves over it, as well as over the O
[38], whenever U is moderately large and the graphis sufficiently
sparse.
1
10
e
7
7
By well-known reductions, this also gives us an O m B -time algorithm for the maximumcardinality bipartite
b-matching
problem in which the largest integer demand is B. This,
10
e
again, matches the O (mB) 7 -time algorithm of Mądry [43], when B = 1, which corresponds
to the
maximum-cardinality bipartite matching problem, and outperforms it, as well as the
e (m√n log B)-time algorithm of Lee and Sidford [38], for moderate values of B and sufficiently
O
sparse graphs.
One of the advantages of our algorithm is that it is significantly simpler than the ones
presented in [43] and [38]. In particular, these algorithms employ a sophisticated interior-point
method framework, while our algorithm is cast directly in the classic augmenting path setting
that almost all the combinatorial maximum flow algorithms use. At a high level, the presented
algorithm takes a primal dual approach in which each iteration uses electrical flows computations
both to find an augmenting s-t flow in the current residual graph and to update the dual solution.
We show that by maintain certain careful coupling of these primal and dual solutions we are
always guaranteed to make significant progress.
∗
Supported by NSF grant CCF-1553428, and a Sloan Research Fellowship.
1
Introduction
The maximum s-t flow problem and its dual, the minimum s-t cut problem, are two of the most
fundamental and extensively studied graph problems in combinatorial optimization [51, 1, 50].
They have a wide range of applications (see [2]), are often used as subroutines in other algorithms
(see, e.g., [4, 53]), and a number of other important problems – e.g., bipartite matching problem [10]
– can be reduced to them. Furthermore, these two problems were often a testbed for development
of fundamental algorithmic tools and concepts. Most prominently, the Max-Flow Min-Cut theorem
[14, 17] constitutes the prototypical primal-dual relation.
Several decades of extensive work resulted in a number of developments on these problems (see
Goldberg and Rao [22] for an overview) and many of their generalizations and special cases. Still,
despite all this effort, the basic problem of computing maximum s-t flow and minimum s-t cut in
general graphs resisted progress for a long time. In particular, for a number of years, the best
2
1
running time bound for the problem was an O(m min{m 2 , n 3 } log(n2 /m) log U ) (with U denoting
the largest integer arc capacity) bound established in a a breakthrough paper by Goldberg and Rao
1
2
[22] and this bound, in turn, matched the O(m min{m 2 , n 3 }) bound for unit-capacity graphs due
to Even and Tarjan [15] – and, independently, Karzanov [25] – that were put forth more than 40
years ago.
The above bounds were improved only fairly recently. Specifically, in 2013, Mądry
[43]
presented
10
e
an interior-point method based framework for flow computations that gave an O m 7 -time1 algorithm for the unit-capacity caseof the
maximum s-t flow and minimum s-t cut problems. This
3
e
finally broke the long-standing O n 2 running time barrier for sparse graphs, i.e., for m = O(n).
Later on, Lee and Sidford [38] developed a variant of interior-point method that enabled them to
obtain improvement for the regime of dense graphs. In particular, their algorithm is able to compute
e (m√n log U ) time and thus improve over
the (general) maximum s-t flow and minimum s-t cut in O
the Goldberg-Rao bound whenever the input graph is sufficiently dense.
It is also worth mentioning that, as a precursor to the above developments, substantial progress
was made in the context of (1 − ε)-approximate variant of the maximum s-t flow problem in undirected graphs. In 2011, Christiano et al. [7] developed an algorithm that allows one to compute a
(1 + ε)-approximation
to the undirected maximum s-t flow (and the minimum s-t cut) problem in
1
−11/3
e
3
O mn ε
time. Their result relies on devising a new approach to the problem that combines
electrical flow computations with multiplicative weights update method (see [4]). Later, Lee et
– algorithm that employs purely
al. [37] presented a quite different – but still electrical-flow-based
e mn1/3 ε−2/3 -time (1 + ε)-approximation for the case of
gradient-descent-type view to obtain an O
unit capacities. Very recently, this line of work was culminated by Sherman [54] and Kelner et al.
[26] who independently showed how to integrate non-Euclidean gradient-descent methods with fast
poly-logarithmic-approximation algorithms for cut problems of Mądry [41] to get an O(m1+o(1) ε−2 )time (1 + ε)-approximation to the undirected maximum flow problem. Then, Peng [48] built on
e mε−2 , running time.
these works to obtain a truly nearly-linear, i.e., O
Finally, we note that, in parallel to the above work that is focused on designing weaklypolynomial algorithms for the maximum s-t flow and minimum s-t cut problems, there is also
a considerable interest in obtaining running time bounds that are strongly-polynomial, i.e., that do
1
e (f ) denotes O(f logc f ), for some constant c.
We recall that O
1
not depend on the values of arc capacities. The current best such bound is O(mn) and it follows
by combining the algorithms of King et al. [28] and Orlin [47].
Bipartite Matching Problem. Another problem that is related to the maximum s-t problem –
and, in fact, can be reduced to it – is the (maximum-cardinality) bipartite matching problem. This
problem is a fundamental assignment task with numerous applications (see, e.g., [1, 40]) and long
history that has its roots in the works of Frobenius [18, 19] and König [29, 30, 31] from the early
20th century (see [52]). Already in 1931, König [32] and Egerváry [13] provided first constructive
characterization of maximum matchings in bipartite graphs. This characterization can be turned
into a polynomial-time algorithm. Then, in 1973, Hopcroft and Karp [24] and, independently,
√
Karzanov [25], devised the celebrated O(m n)-time algorithm. For 40 years this bound remained
the best one known in the regime of relatively sparse graphs. Only recently Mądry [43] obtained an
e m10/7 running time. It turns out, however, that whenever the input graph is dense,
improved, O
i.e., when m is close to n2 even better bounds can be obtain. In this setting, one can combine the
algebraic approach of Rabin and Vazirani [49] – that itself builds on the work of Tutte [57] and
Lovász [39] – with matrix-inversion techniques of Bunch and Hopcroft [6] to get an algorithm that
runs in O(nω ) time (see [45]), where ω ≤ 2.3727 is the exponent of matrix multiplication [9, 59].
Also, later on, Alt et al. [3], and Feder and Motwani [16] developed combinatorial algorithms that
√
2
offer a slight improvement – by a factor of, roughly, log n/ log nm – over the O(m n) bound of
Hopcroft and Karp whenever the graph is sufficiently dense.
Finally, a lot of developments has been done in the context of the (maximum-cardinality) matching problem in general, i.e., not necessarily bipartite, graphs. Starting with the pioneering work of
Edmonds [11], these developments led to bounds that essentially match the running time guarantees
that were previously known only for bipartite case. More specifically, the running time bound of
√
O(m n) for the general-graph case was obtained by Micali and Vazirani [44, 60] (see also [20] and
[21]). Then, Mucha and Sankowski [46] gave an O(nω )-time algorithms for general graphs that
builds on the algebraic characterization of the problem due to Rabin and Vazirani [49]. This result
was later significantly simplified by Harvey [23].
1.1
Our Contribution
In this paper, we put forth a new algorithm for solving the maximum s-t flow and the minimum
s-t cut problems in directed graphs. More precisely,
10 we
develop an algorithm that computes the
1
e
maximum s-t flow of an input graph in time O m 7 U 7 , where m denotes the number of arcs of
that graph and U its largest integer capacity. Known reductions imply similar running time bounds
for the minimum s-t cut problem as well as for the maximum-cardinality bipartite b-matching
problem, a natural generalization of the maximum bipartite matching problem
each vertex
10 in1 which
e
7
7
v has a degree demand bv . For that problem, our algorithm yields an O m B -time algorithm,
with B being the largest (integer) vertex demand.
In the light of the above, for the unit-capacity/demand cases, the resulting algorithms
match
e mU 107
the performance of the algorithm of Mądry [43]. The latter algorithm, however, runs in O
e (mB) 107 running time for the bipartite b-matching problem) in
time (which translates into an O
the case of arbitrary capacities/demands. Consequently, the significantly better dependence of the
running time of our algorithm on the largest capacity U / largest demand B makes it much more
2
favorable in that setting. In fact, even though that dependence on U /B is still polynomial it enables
our algorithm to remain competitive, for a non-trivial range of parameters, with the best existing
e (m√n log U )-time algorithm
algorithms that run in time that is logarithmic in U /B, such as the O
of Lee and Sidford [38].
Even more crucially, the key advantage of our algorithm is that it is significantly simpler than
both the algorithm of Mądry [43] and that of Lee and Sidford [38]). Both these algorithms rely
heavily on the interior-point method framework. Specifically, [43] designed a certain new variant of
path-following interior-point method algorithm for the near-perfect bipartite b-matching problem
that encoded the input maximum s-t flow instance. It then used electrical flow computations
1to
e m2
converge to the near-optimal solution for that problem. In order to break the bottlenecking O
iteration bound, however, Mądry [43] needed to, first, develop an extensive toolkit for perturbing
and preconditioning the underlying electrical flow computation and, then, to combine this machinery
with a very careful and delicate analysis of the resulting dynamics.
Our algorithm also relies on electrical flow computations but it abandons the above methodology
and works instead fully within the classic augmenting path framework that almost all the previous
combinatorial maximum s-t flow algorithms used. In this framework, developed by Ford and Fulkerson [17] (see also [14]), the flow is built in stages. Each stage corresponds to finding a so-called
augmenting flow in the current residual graph, which is a directed graph that encodes the solution
found so far. The algorithm terminates when the residual graph admits no more augmenting flows,
i.e., there is no path from s to t in it, since in this case the solution found so far has to be already
optimal.
The chief bottleneck in the running time analysis of augmenting path based algorithms is ensuring that each flow augmentation stage makes sufficient progress. Specifically, one wants to obtain a
good trade off between the amount of flow pushed in each augmentation step and the time needed
to implement each such flow push. One simple approach is to just use here s-t path computations.
This is a nearly-linear time procedure but it only guarantees pushing one unit of flow each time.
A much more sophisticated primitive developed in this context are blocking flow computations.
Combining this primitive with a simple duality argument enabled Goldberg and Rao [22], who built
1
2
on the work of Even and Tarjan [15], to obtain an O(m min{m 2 , n 3 } log U )-time maximum flow
that remained the best known algorithm for nearly two decades. Unfortunately, trying to improve
such blocking flow-based approaches turned out to be extremely difficult and no progress was made
here so far, even in the unit-capacity case for which the best known bounds were established over
40 years ago.
One of the key contributions of this paper is bringing a new type of primitive: electrical flows to
the augmenting path framework; and showing how to successfully use it to outperform the blocking
flow-based methods. Specifically, our algorithm finds augmenting flows by computing electrical
flows in certain symmetrization of the current residual graph – see Section 3.3 for more details.
(Note that performing such a symmetrization is necessary as residual graphs are inherently directed
while electrical flows are inherently undirected.) The key difficulty that arises in this context,
however, is ensuring that this symmetrized residual graph can still support a significant fraction
of the s-t capacity of the original residual graph. It is not hard to see that, in general, this might
not be the case. To address this problem we introduce a certain careful coupling of the primal
and dual solutions, which is inspired by the so-called centrality condition arising in interior-point
method based maximum flow algorithms (see [43]). We then show that maintaining this coupling
and applying a simple preconditioning technique let us guarantee that looking only for the flows in
3
the symmetrized version of the residual graph stillprovides sufficient progress in each iteration and,
e m 32 log U -time algorithm.
in particular, immediately delivers a O
We then (see Section 4) build on that basic algorithm and develop an ℓp -geometric understanding of its running time analysis. This understanding guides us towards a simple electrical flow
perturbation technique – akin to the perturbation techniques used in [7] and [43] – that enables us
√
to break the Ω( m) iterations bottleneck that all the blocking flow-based algorithms were suffering
from, and thus get the final, improved result.
We believe that further study of this new augmenting flow based framework will deliver even
faster and simpler algorithms.
1.2
Organization of the Paper
We begin the technical part of the paper in Section 2 by presenting some preliminaries on the
maximum flow problem and the notion of electrical
3flows. Then, in Section 3, we present our
e m 2 log U -time maximum s-t flow algorithm.
framework and demonstrate how it yields an O
Finally, inSection 4, we show how to refine our basic framework to obtain the improved running
e m 107 U 17 .
time of O
2
Preliminaries
Throughout this paper, we will be viewing graphs as having both lower and upper capacities.
Specifically, we will denote by G = (V, E, u ) a directed graph with a vertex set V , an arc set E (we
+
allow parallel arcs), and two (non-negative) integer capacities u−
e and ue , for each arc e ∈ E. (We
will explain the role of these capacities below.) Usually, m will denote the number |E| of arcs of
the graph in question and n = |V | will be the number of its vertices. We view each arc e of G as
an ordered pair (u, v), where u is its tail and v is its head.
Observe that this perspective enables us to view undirected graphs as directed ones in which
the ordered pair (u, v) ∈ E is an (undirected) edge (u, v) and the order just specifies the orientation
of that edge (from u to v).
Maximum Flow Problem. The basic notion of this paper is the notion of a flow. Given a graph
G, we view a flow in G as a vector f ∈ Rm that assigns a value fe to each arc e of G. If this
value is negative we interpret it as having a flow of |fe | flowing in the direction opposite to the arc
orientation. (This convention is especially useful when discussing flows in undirected graphs.)
We say that a flow f is an σ-flow, for some demands σ ∈ Rn iff it satisfies flow conservation
constraints with respect to that demands. That is, we have that
X
X
fe −
fe = σv , for each vertex v ∈ V .
(1)
e∈E + (v)
e∈E − (v)
Here, E + (v) (resp. E − (v)) is the set of arcs of G that are entering (resp. leaving) vertex v.
Intuitively, these constraints enforce that the net balance of the total in-flow into vertex v and the
total out-flow out
P of that vertex is equal to σv , for every v ∈ V . (Observe that this implies, in
particular, that v σv = 0.)
4
Furthermore, we say that a σ-flow f is feasible in G iff f obeys the the capacity constraints:
+
−u−
e ≤ fe ≤ ue ,
for each arc e ∈ E.
(2)
In other words, we want each arc e to have a flow that is at most u+
e if it flows in the direction
−
of e’s orientation (i.e., fe ≥ 0), and at most ue , if it flows in the opposite direction (i.e., fe < 0).
Note that setting all u−
e s be equal to zero recovers the standard notion of flow feasibility in directed
graphs.
One type of flows that will be of special interest to us are s-t flows, where s (the source) and t
(the sink) are two distinguish vertices of G. Formally, an s-t flow is a σ-flow whose demand vector
σ is equal to F · χs,t, where F ≥ 0 is called the value of f and χs,t is a demand vector that has −1
(resp. 1) at the coordinate corresponding to s (resp. t) and zeros everywhere else.
Now, the maximum flow problem corresponds to a task in which we are given a (directed) graph
G = (V, E, u ) with integer capacities as well as a source vertex s and a sink vertex t and want to
find a feasible (in the sense of (2)) s-t flow of maximum value. We will denote this maximum value
as F ∗ .
Residual Graphs. A fundamental object in many maximum flow algorithms (including ours) is
the notion of a residual graph. Given a graph G = (V, E, u ) and a feasible σ-flow f in that graph
(it is useful to think σ = F · χs,t ), we define the residual graph Gf (of G with respect to f ) as a
b (f )) over the same vertex and arc set as G and such that, for each arc e = (u, v)
graph Gf = (V, E, u
of G, its lower and upper capacities are defined as
+
u
b+
e (f ) := ue − fe and
−
u
b−
e (f ) := ue + fe .
(3)
b−
We will refer to u
b+
e (f )) as forward residual capacity (resp. backward residual capacity)
e (f ) (resp. u
of e and also define the residual capacity u
be (f ) of e as the minimum of these two, i.e., u
be (f ) :=
+ (f )}. Note that the value of residual capacity depends on the flow f but we will
(f
),
u
b
min{b
u−
e
e
ensure that it is always clear from the context with respect to which flow the residual capacity is
measured. Also, observe that feasibility of f implies that all residual capacities are always nonnegative (cf. (2)).
The main reason why residual graphs are useful in computing maximum flows is that they
constitute a very convenient representation of the progress made so far. Specifically, we have the
following important fact. (Again, it is useful to think here of the maximum s-t flow setting, in which
σ = F ∗ χs,t .)
Fact 2.1. Let σ be some demand and G = (V, E, u ) be a graph in which a demand of σ can be
routed, i.e., there exists a σ-flow f ∗ that is feasible in G. Also, for any 0 ≤ α ≤ 1, let f be a feasible
b (f )) be the residual graph of G with respect to f . We have that
ασ-flow in G, and Gf = (V, E, u
(a) one can route a demand of (1 − α)σ in Gf ;
(b) if f ′ is a feasible α′ σ-flow in Gf , for some α′ , then f + f ′ is a feasible (α + α′ )σ-flow in G.
Intuitively, the above fact enables us to reduce the task of routing a demand σ in G to a sequence
of computations of augmenting α′ σ-flows in the residual graph Gf . We know that as long as we
have not yet computed a feasible σ-flow in G, Gf can route a demand of (1 − α)σ-flow, where
(1 − α) > 0 is the fraction of routed demand that we are still “missing”, and each new augmenting
α′ σ-flow found in Gf brings us closer to routing σ in full in G. (Note that initially Gf is equal to
G and Gf is changing after each new augmenting α′ σ-flow is found.)
5
Electrical Flows and Vertex Potentials. Another notion that will play a fundamental role in
this paper is the notion of electrical flows. Here, we just briefly review some of the key properties
that we will need later. For an in-depth treatment we refer the reader to [5].
Consider a graph G and a vector of resistances r ∈ Rm that assigns to each edge e its resistance
re > 0. For a given σ-flow f in G, let us define its energy (with respect to resistances r ) Er (f ) to
be
X
Er (f ) :=
re fe2 .
(4)
e
For a given demand vector σ and a vector of resistances r , we define the electrical σ-flow in
G (that is determined by resistances r ) to be the flow that minimizes the energy Er (f ) among all
flows with demand σ in G. As energy is a strictly convex function, one can easily see that such a
flow is unique. (It is important to keep in mind that such flow is not required to be feasible with
respect to capacities of G, in the sense of (2).)
A very useful property of electrical flows is that they can be characterized in terms of vertex
potentials inducing them. Namely, one can show that a flow f with demands σ in G is an electrical
σ-flow determined by resistances r iff there exist vertex potentials φv (that we collect into a vector
φ ∈ Rn ) such that, for any edge e = (u, v) in G,
fe =
φv − φu
.
re
(5)
In other words, a f with demands σ is an electrical σ-flow iff it is induced via (5) by some vertex
potential φ. (Note that the orientation of edges matters in this definition.) The above equation
corresponds to the Ohm’s law known from physics.
Note that we are able to express the energy Er (f ) (see (4)) of an electrical σ-flow f in terms of
the potentials φ inducing it as
X (φv − φu )2
.
(6)
Er (f ) =
re
e=(u,v)
One of the consequences of the above is that one can develop a dual characterization of the
energy of an electrical σ-flow in terms of optimization over vertex potentials. Namely, we have the
following lemma whose proof can be found, e.g., in [43] Lemma 2.1.
Lemma 2.2. For any graph G = (V, E), any vector of resistances r , and any demand vector σ,
1
= min
Er (f ∗ ) φ|σT φ=1
X
e=(u,v)∈E
(φv − φu )2
,
re
where f ∗ is the electrical σ-flow determined by r in G. Furthermore, if φ∗ are the vertex potentials
b := φ∗ /Er (f ∗ ).
corresponding to f ∗ then the minimum is attained by taking φ to be equal to φ
Note that the above lemma provides a convenient way of lowerbounding the energy of an electrical σ-flow. One just needs to expose any vertex potentials φ such that σ T φ = 1 and this will
immediately constitute an energy lowerbound.
6
Laplacian Solvers. The fact that the electrical σ-flow determined by resistances r is the only flow
with demands σ that can be induced by vertex potentials (cf. (5)) has an important consequence.
It enables us to reduce electrical σ-flow computations to solving a linear system. In fact, the task
of finding vertex potentials that induce that flow can be cast as a Laplacian linear system. That is,
a linear system in which the constraint matrix corresponds to a Laplacian of the underlying graph
with weights given by the (inverses of) the resistances r .
Now, from the algorithmic point of view, the crucial property of Laplacian systems is that we
can solve them, up to a very good approximation, very efficiently. Namely, there is a long line
of work [56, 33, 34, 27, 8, 35, 36] that builds on an earlier work of Vaidya [58] and Spielman and
Teng [55] and gives us a number of Laplacian system solvers that run in only nearly-linear time
and, in case of more recent variants, are conceptually fairly simple. In particular, this line of work
establishes the following theorem.
Theorem 2.3. For any ε > 0, any graph G with n vertices
and m edges, any demand vector
−1
e
σ, and any resistances r , one can compute in O m log ε
time vertex potentials φ̃ such that
∗
kφ̃ − φ∗ kL ≤ εkφ∗ kL , where L is the Laplacian
of
G,
φ
are
potentials
inducing the electrical σ-flow
p
T
determined by resistances r , and kφkL := φ Lφ.
Even though the solutions delivered by the above Laplacian solvers are only approximate, the
quality of approximation that it delivers is more than sufficient for our purposes. Therefore, in the
rest of this paper we assume that these solutions are exact. (See, e.g., [43] for discussion how to
deal with inexactness of the electrical flows computed.)
Vector Norms. We will find it useful to work with various ℓp -norms of vectors. To this end, for
any p > 0, we define the ℓp -norm kh kp of a vector h as
khkp :=
X
i
!1
p
|hi |p
.
(7)
In particular, the ℓ∞ -norm is given by kh k∞ := maxi |hi |.
3
Augmenting Residual Graphs with Electrical Flows
In this section, we put forth the general framework we will use to solve the maximum s-t problem.2
In particular,
we
3
demonstrate how this framework enables us to solve the maximum s-t flow problem
e
2
in O m log U time, where m = |E| is the number of arcs in of the input graph and U is its largest
(integer) capacity.
More precisely, for any
s-t flow instance G = (V, E, u ) and any value F ≥ 0, our
maximum
3
e
algorithm will work in O m 2 log U time and either: compute a feasible s-t flow of value F in G;
or conclude that the maximum s-t flow value F ∗ of G is strictly smaller than F .
2
Observe that our treatment can be immediately extended to routing arbitrary demands since this problem can
always be reduced to the directed maximum flow problem by adding a super-source and super-sink and connecting
all the surpluses in the original graph to the former and all the deficits in that graph to the latter via arcs with
appropriate capacities.
7
Note that such procedure can be turned into a “classic” maximum s-t flow by applying binary
search over values of F and incurring a multiplicative running time overhead of only O(log U n). (In
fact, a standard use of capacity scaling technique [12] enables one to keep the overall running time
of the resulting algorithm be only linear, instead of quadratic, in log U .)
As mentioned earlier, our algorithm follows the primal dual augmenting paths based framework.
At a high level, in each iteration (see Section 3.2), it uses electrical flow computations to compute
an augmenting flow as well as an update to the dual solution. To ensure that each augmenting
iteration makes enough progress, it maintains a careful coupling of the primal and dual solution.
We describe it below.
3.1
Primal Dual Coupling
Let us fix our target flow value F and, for notational convenience, , for any 0 ≤ α ≤ 1, let us denote
by χα the demand vector αF χs,t , i.e., the demand corresponding to routing α-fraction of the target
flow value F of the s-t flow. Also, let us define χ to be the demand equal to χ1 .
Again, our algorithm will be inherently primal dual in nature. That is, in addition to maintaining
a primal solution: a χα -flow f , for some 0 ≤ α ≤ 1, that is feasible in G, it will also maintain a
dual solution y ∈ Rn , which should be viewed as an embedding of all the vertices of the graph G
into a line.
Consequently, our goal will be to either to make f be a feasible flow with demands χα and
α = 1, which corresponds to routing the target s-t flow value F in full, or to make the dual solution
y certify that the target demand χ1 = χ cannot be fully routed in G and thus F > F ∗ .
Well-coupled Solutions. Our primal dual scheme will be enforcing a very specific coupling of
the primal solution f and the dual solution y. More precisely, let us define for each arc e = (u, v)
∆e (y ) := yv − yu ,
(8)
to be the “stretch“ of the arc e in the embedding given by y. Also, let Gf be the residual graph of
G with respect to f and let us define, for a given arc e, a potential function
Φe (f ) :=
1
1
,
u
b−
e (f )
(9)
for each arc e.
(10)
u
b+
e (f )
−
−
+
b−
where we recall that u
b+
e (f ) := ue + fe ) are forward (resp. backward)
e (f ) := ue − fe (resp. u
residual capacities of the arc e. (See preliminaries, i.e., Section 2, for details.)
Then, our intention is to maintain the following relation between f and y :
∆e (y ) = Φe (f )
Intuitively, this condition ensures that the line embedding y stretches each arc e in the direction
of the smaller of the residual capacities, and that this stretch is inversely proportional to the value of
−
that capacity. (Note that if G was undirected and thus the initial capacities u+
e and ue were equal,
the direction of smaller residual capacity is also the direction in which the flow f flows through the
arc e.) It is worth pointing out that this coupling condition is directly inspired by (and, in fact,
can be directly derived from) a certain variant of centrality condition used by interior-point method
based maximum flow algorithms (see [43]).
8
Even though condition (10) expresses our intended coupling, it will be more convenient to work
with a slightly relaxed condition that allows us to have small violations of that ideal coupling.
Specifically, we say that a primal dual solution (f , y ) is γ-coupled iff
|∆e (y ) − Φe (f )| ≤
γe
u
be (f )
for all arcs e = (u, v).
(11)
m
Here, u
be (f ) = min{b
u+
b−
e (f ), u
e (f )} is the (symmetrized) residual capacity of the arc e, and γ ∈ R
is the violation vector that we intend to keep very small. In particular, we say that a primal dual
1
, i.e.,
solution (f , y ) is well-coupled iff its violation vector γ has its ℓ2 -norm (see (7)) be at most 100
1
kγk2 ≤ 100 .
One of the key consequences of maintaining a well-coupled primal dual pair of solutions (f , y )
is that it enables us to use y as a dual certificate for inability to route certain demands in G. The
following lemma gives us a concrete criterion for doing that.
Lemma 3.1. Let (f , y ) be a well-coupled primal dual solution with f being a χα -flow, for some
0 ≤ α < 1. If
2m
χT y >
(1 − α)
then the demand χ cannot be routed in G, i.e., F > F ∗ .
Note that the choice of the constant 2 in the above lemma is fairly arbitrary. In principle, any
constant strictly larger than 1 would suffice.
2m
but the demand χ can still be routed
Proof. Assume for the sake of contradiction that χT y > (1−α)
in G. As f is an χα -flow that is feasible in G, by Fact 2.1, there has to exist a χ(1−α) -flow f ′ that
is feasible in the residual graph Gf . Observe that, by (1) and (8), we have in that case that
X
X
X
X
X
(f ′ )T ∆(y ) =
fe′ ∆e (y ) =
fe′ (yv − yu ) =
yv
fe′ −
fe′ (12)
e
v
e=(u,v)
e∈E + (v)
e∈E − (v)
T
= (1 − α)F (yt − ys ) = (1 − α)χ y > 2m.
We want now to use the feasibility of f ′ to derive an upper bound on the inner produce (f ′ )T ∆(y )
that will violate the above lower bound and thus deliver the desired contradiction.
To this end, let us consider some particular arc e. Assume wlog that fe′ ≥ 0 (handling of the
case of fe′ < 0 is completely analogous). We have that the contribution of that arc to the inner
product (f ′ )T ∆(y ) is, by (11), at most
1
1
γe fe′
γe fe′
′
′
′
≤ fe
,
−
+
fe ∆e (y ) ≤ fe Φe (f ) +
u
be (f )
u
be (f )
u
b+
u
b−
e (f )
e (f )
where we recall that γ is the violation vector of (f , y ) (cf. (11)).
As f ′ is feasible in Gf , we need to have that fe′ ≤ u
b+
b+
b−
b+
e (f ). So, if u
e (f ) ≤ u
e (f ), i.e., u
e (f ) =
u
be (f ), then
f ′ (1 + γe )
1
1
γe fe′
′
′
≤ e
≤ (1 + γe ).
fe ∆e (y ) ≤ fe
− −
+
+
u
be (f )
u
be (f )
u
be (f ) u
be (f )
9
Otherwise, we have that
1
(1 − γe )fe′
1
1 − γe
γe fe′
1
′
′
′
≤ fe
≤ 1,
≤1−
− −
−
fe ∆e (y ) ≤ fe
+
+
+
u
be (f )
u
be (f )
u
be (f )
u
be (f ) u
be (f )
u
be (f )
as γe ≤ kγk∞ ≤ kγk2 ≤ 12 . So, in either case, this contribution is at most (1 + γe ). Summing these
contributions over all arcs we get that
X
X
√
(f ′ )T ∆(y ) ≤
fe′ ∆e (y ) ≤
(1 + γe ) ≤ m + mkγk2 ≤ 2m,
e
e
which indeed contradicts the lower bound (12). The lemma follows.
In the light of the above discussion, for a given well-coupled primal dual solution (f , y ), we
should view the value of α as a measure of our primal progress, while the value of the inner product
σ T y can be seen as a measure of our dual progress.
Initialization. The coupling condition (11) ties the primal and dual solutions f and y fairly
tightly. In fact, coming up with some primal dual solutions that are well-coupled, which we need
to initialize our framework, might be difficult.
Fortunately, finding such a pair of initial well-coupled solutions turns out to be easy, if our input
graph G is undirected. In that case, just taking f to be a trivial zero flow and y to be a trivial allzeros embedding makes condition (11) satisfied (with γe s being all zero). After all, the residual graph
+
−
Gf with respect to such zero flow f is just the graph G itself and thus u
b+
b−
e (f ) = ue = ue = u
e (f ).
Furthermore, even though we are interested in solving directed instances too, every such instance
can be reduced to an undirected one. Specifically, we have the following lemma.
Lemma 3.2. Let G be an instance of the maximum s-t flow problem with m arcs and the maximum
e (m) time, one can construct an
capacity U , and let F be the corresponding target flow value. In O
′
instance G of undirected maximum s-t flow problem that has O(m) arcs and the maximum capacity
U , as well as target flow value F ′ such that:
(a) if there exists a feasible s-t flow of value F in G then a feasible s-t flow of value F ′ exists in G′ ;
e (m) time a feasible s-t flow of
(b) given a feasible s-t flow of value F ′ in G′ one can construct in O
value F in G.
The proof of the above lemma boils down to a known reduction of the directed maximum flow
problem to its undirected version – see, e.g., Theorem 3.6.1 in [42] for details. Consequently, from
now on we can assume without loss of generality that we always have a well-coupled primal dual
pair to initialize our framework.
3.2
Progress Steps
Once we described our basic framework and how to initialize it, we are ready to put forth its main
ingredient: progress steps that enable us to gradually improve the primal dual solutions that we
maintain. To this end, let us fix a well-coupled primal dual solutions (f , y ) with f being a χα -flow,
for some 0 ≤ α < 1, that is feasible in G. Our goal in this section will be to use (f , y ) to compute, in
nearly-linear time, another pair of well-coupled primal dual solutions (f + , y + ) that bring us closer
10
to the optimal solutions. The flow f + we obtain will be a χα′ -flow feasible in G, for α′ > α. So,
the resulting flow update f + − f is an augmenting flow that is feasible in our current residual graph
Gf and pushes (α′ − α)-fraction of the target s-t flow.
We will compute (f + , y + ) in two stages. First, in the augmentation step, we obtain a pair of
b ), with fb being a χα′ -flow, for α′ > α, that is feasible in G. These solutions make
solutions (fb , y
progress toward the optimal solutions but might end up being not well-coupled. Then, in the fixing
step, we correct (fb, yb ) slightly by adding a carefully chosen flow circulation, i.e., a flow with allb so as to make the resulting solutions (f + , y + ) be
zeros demands, to fb and an dual update to y
well-coupled, as desired.
The key primitive in both these steps will be electrical flow computation. As we will see, the
crucial property of electrical flows we will rely on here is their“self-duality“. That is, the fact that
each electrical flow computation gives us both the flow and the corresponding vertex potentials
that are coupled to it via Ohm’s law (5). This enables us not only to update our primal and dual
solutions with that flow and vertex potentials, respectively, but also, much more crucially, this
coupling introduced by Ohm’s law will be exactly what will allow us to (approximately) maintain
our desired primal dual coupling property (11).
Augmentation Step. To perform an augmentation step we compute first an electrical χ-flow fe
in G with the resistances r defined as
1
1
+
,
(13)
re := +
−
(b
ue (f ))2 (b
ue (f ))2
for each arc e. Note that the resistance re is proportional, roughly, to the inverse of the square of
the residual capacity u
be (f ) of that arc. So, in particular, it becomes very large whenever residual
capacity of the arc e is small, and vice versa. As we will see shortly, this correspondence will allow
us to control the amount of flow that fe sends over each edge and thus ensure that the respective
residual capacities are not violated.
Let φ̃ be the vertex potentials inducing fe (via the Ohm’s law (6)). Then, we obtain the new
primal and dual solution (fb, yb ) as follows:
fbe := fe + δfee
ybv := yv + δφev
for each arc e
(14)
for each vertex v,
where δ is the desired step size. Observe that this update is exactly an augmentation of our current
flow f with the (scaled) electrical flow δfe and adding to our dual solution the (scaled) vertex
potentials δφ̃. This, in particular, means that the new flow fb we obtain here is a χα′ -flow with
α′ = α + δ.
(15)
The step size δ, however, will have to be carefully chosen. On one hand, as we see in (15), the
larger it is the more progress we make. On the other hand, though, it has to be small enough so as
to keep the flow δfe feasible in Gf (and thus, by Fact 2.1, to make the flow fb + f feasible in G).
Note that, a priori, we have no direct control over neither the directions in which the electrical
χ-flow fe is flowing thorough each arc nor the amount of that flow. So, in order to establish a grasp
on what is the right setting of δ, it is useful to define a congestion vector ρ given by
ρe :=
fee
,
u
be (f )
11
(16)
for each arc e. One can view ρe as a normalized measure of how much the electrical flow fe overflows
the residual capacity u
be (f ) of the arc e and in what direction. In other words, the sign of ρe encodes
the direction of the flow fee .
It is now not hard to see that to ensure that δfe is feasible in Gf , i.e., that no residual capacity
is violated by the update (14), it suffices that δ|ρe | ≤ 41 , for all arcs e, or, equivalently, that
δ≤
1
,
4kρk∞
(17)
where k · k∞ is the standard ℓ∞ -norm.
It is also worth pointing out that the congestion vector ρ turns out to capture (up to a small
multiplicative factor) the contribution of each arc e to the energy Er (fe ) of fe. In particular, we have
the following simple but important observation.
Lemma 3.3. For any arc e, ρ2e ≤ re fee2 ≤ 2ρ2e and
where k · k2 is the standard ℓ2 -norm.
kρk22 ≤ Er (fe) ≤ 2kρk22 ,
Proof. Note that, by definition of the resistance re (13), (16), and the definition of the residual
capacity u
be , we have that
ρ2e ≤
fee2
2fee2
(ρe u
be (f ))2
+
−2
−
−2 e2
2
e
≤
≤
(b
u
(f
))
+
(b
u
(f
))
= 2ρ2e .
f
=
r
f
≤
e e
e
e
e
u
be (f )2
u
be (f )2
u
be (f )2
We thus also have that
kρk22 =
as desired.
X
e
ρ2e ≤
X
e
re fee2 = Er (fe) ≤
X
2ρ2e = 2kρk22 ,
e
This link between the energy-minimizing nature of the electrical σ-flow fe and the ℓ2 -norm of
the congestion vector ρ will end up being very important. One reason for that is the fact that
ℓ∞ -norm is always bounded by the ℓ2 -norm. Consequently, we can use this connection to control
the ℓ∞ -norm of the vector ρ and thus the value of δ needed to satisfy the feasibility condition (17).
It turns out, however, that just ensuring that our augmenting flow is feasible is not enough for
our purposes. Specifically, we also need to control the coupling of our primal dual solutions, and
the feasibility bound (17) might be not sufficiently strong for this purpose. We thus have to develop
analyze the impact of the update (14) on the coupling condition (11) more closely.
To this end, let us first notice the following fact that stems from a standard application of the
Taylor’s theorem. Its proof appears in Appendix A.1
Fact 3.4. For any u1 , u2 > 0 and x such as |x| ≤ u4 , where u = min{u1 , u2 }, we have that
1
1
1
1
1
1
−
−
+
=
+
x + x2 ζ,
u1 − x u2 + x
u1 u2
u21 u22
where |ζ| ≤
5
.
u3
12
Now, the above approximation bound enables us to get an accurate estimate of how the coupling
condition evolves during the augmentation step (14). That is, for any arc e, applying Fact 3.4 with
e
b−
u1 = u
b+
e (f ) and x = δ fe , gives us that
e (f ), u2 = u
1
1
1
1
b
Φe (f ) = +
+ −
δfee + (δfee )2 ζe , (18)
− −
= Φe (f ) +
2
(b
u+
(b
ue (f ))2
u
be (f ) − δfee u
be (f ) + δfee
e (f ))
with |ζe | ≤ ube (f5 )3 .
Observe that the above expression tells us that the first order approximation of the change in
the primal contribution of the arc e to the coupling condition (11) caused by the update (14) is
exactly
1
1
b
+ −
δfee = re δfee ,
Φe (f ) − Φe (f ) ≈
2
2
(b
u+
(f
))
(b
u
(f
))
e
e
where we also used (13). (In fact, the choice of the resistances r was made exactly to make the
above statement true.)
Furthermore, by Ohm’s law (5) and the definition of our augmentation step (14), we have that
re δfee = δ φev − φeu = ∆e (b
y ) − ∆e (y ),
which is exactly the change in the dual contribution of the arc e = (u, v) to the coupling condition
(11) caused by the augmentation step update (14).
So, up to first order approximation, these two contributions cancel out, leaving the coupling
(11) intact. Consequently, any increase in the violation of the coupling condition must come from
the second-order terms in the approximation (18). The following lemma makes this precise and is
proved in Appendix A.2.
Lemma 3.5. Let 0 < δ ≤ (4kρk∞ )−1 and the primal dual solution (f , y ) be γ-coupled. Then, we
have that, for any arc e = (u, v),
∆e (b
y ) − Φe (fb ) ≤
4
3 γe
+ 7(δρe )2
.
u
be (fb)
Fixing Step. Although Lemma 3.5 enables us to bound the deterioration of the primal dual
coupling during the augmentation step, we cannot prevent this effect altogether. Therefore, we
need to introduce a fixing step that deals with this problem. More precisely, we develop a procedure
that uses a single electrical flow computation to significantly reduce that violation, provided it was
not too large to begin with. This is formalized by the following lemma, whose proof appears in
Appendix A.3.
Lemma 3.6. Let (g , z ) be a ς-coupled primal dual solution, with g being a feasible χα′ -flow and
1
e (m) time, we can compute a primal dual solution (ḡ , z̄ ) that is well-coupled and in
kςk2 ≤ 50
. In O
which ḡ is still a χα′ -flow.
Now, after putting Lemmas 3.5 and 3.6 together, we are finally able to state the condition that
δ in the update (14) has to satisfy in order to ensure that the solutions (f + , y + ) we obtain after
performing the augmentation and fixing step is still well-coupled.
13
Lemma 3.7. (f + , y + ) is a well-coupled primal dual solution with f + being a χα′ -flow that is feasible
in G whenever
δ ≤ (33kρk4 )−1 ,
The above lemma tells us that the step size δ of our augmentation step (14) should be governed
by the ℓ4 -norm (see (7)) of the congestion vector (16). Observe that the ℓ4 -norm of a vector is
always upper bounding its ℓ∞ -norm. So, the condition (17) is subsumed by this ℓ4 -norm bound.
Proof. Note that we always have that kρk4 ≥ kρk∞ . So, the condition (17) is automatically satisfied
and the flow f + is indeed a χα′ -flow that is feasible in G.
Now, to argue about well-coupling of (f + , y + ), notice that, in the light of Lemma 3.6, it suffices
b ) obtained after executing the augmentation step (14)
to argue that the primal dual solution (fb , y
1
b -coupled with kb
is γ
γ k2 ≤ 50 .
To this end, observe that by Lemma 3.5 and the fact that (f , y ) are well-coupled, i.e., γ-coupled
1
, we have that
with kγk2 ≤ 100
kb
γ k2 =
≤
v
u
2
uX 4
2
2
t
γe + 7(δρe )
γ
be ≤
3
e
e
s
X
4
2
7
1
4
2
kγk2 + 7δ
+ 7δ2 kρk24 ≤
+
<
.
ρ4e ≤
3
300
150
1089
50
e
sX
The lemma follows.
3.3
Analysis of the Algorithm
We want now to analyze the overall running time of our algorithm. Recall that given our target
demand χ that corresponds to sending F units of flow from the source s to the sink t, our overarching
goal is to either route this demand fully in G or provide a dual certificate that it is impossible to
route χ in G.
We aim to achieve this goal by maintaining and gradually improving a primal dual solution
(f , y ). In this solution, f is a χα -flow (which corresponds to routing an α fraction of the desired
demand χ) that is feasible in G and f and y are well-coupled, i.e., tied to each other via condition
b having sufficiently small ℓ2 -norm. As described in Section 3.2, each
(11) with the violation vector γ
iteration runs in nearly-linear time and boils down to employing electrical flow computations to
find an augmenting flow in the current residual graph Gf (as well as to update the dual solution to
maintain well-coupling).
Consequently, all we need to do now is to lower bound the amount of progress that each of these
iteration makes. Ideally, we would like to prove that in each iteration in which f already routed
α-fraction of the desired flow, i.e., f is a feasible χα -flow, the step size δ (see (14)) can be taken to
be at least
δ ≥ (1 − α)δ̂,
(19)
for some fixed δ̂ > 0. Observe that if such a lower bound was established then, by (15), it would
imply that each iteration finds an augmenting flow that routes at least δ̂-fraction of the amount of
flow still to be routed. As a result, after executing at most O(δ̂−1 log mU ) iterations, the remaining
value of flow to be routed would be at most 1 and thus a simple flow rounding and augmenting path
14
finding
would yield
the final answer (see, e.g., [43]), making the overall running time be at most
−1
e δ̂ m log U .
O
Unfortunately, a priori, it is difficult to provide any such non-trivial unconditional lower bound
on the amount of primal progress we make in each iteration. After all, it could be the case that
the target flow cannot be even routed in G. More importantly though, even if the target flow could
be routed in G, and thus the residual graph always admitted augmenting flows of sufficiently large
value, it is still not clear that our flow augmenting procedure would be able to find them. (It is
worth noting that this problem is by no means specific to our algorithm. In fact, in all the maximum
flow algorithms that rely on the augmenting paths framework ensuring that each iteration makes a
sufficient primal progress is a chief bottleneck in the analysis.)
The root of the problem here is that our flow augmenting procedure is based on electrical flows
and these are undirected in nature. Consequently, the augmenting flows that it finds have to come
from a fairly restricted class: s-t flows that are feasible in a certain “symmetrized" version of the
residual graph.
bf to be an
To make it precise, given a residual graph Gf , let us define its symmetrization G
undirected graph in which each arc e has its forward and backward capacity equal to u
be (f ), i.e., to
+
−
the minimum of the forward u
be (f ) and backward u
be (f ) residual capacities in Gf . Observe now that
each (electrical) augmenting flow δfe found in the augmentation step (cf. (14)) is not only feasible
bf – this is exactly what the condition (17)
in the residual graph Gf but also in its symmetrization G
enforces.
bf too. In
However, not all augmenting s-t flows that are feasible in Gf have to be feasible in G
fact, it can happen that a large maximum s-t flow value that the residual graph Gf supports mostly
bf , and thus prevents our algorithm from making a sufficient good
vanishes in its symmetrization G
primal progress. (Again, a difficulty of a exactly the same nature arises in the analysis of the classic
flow augmenting algorithms such as [15, 25, 22].)
Preconditioning Arcs. It turns out, however, that there is a fairly simple way to circumvent
the above difficulty and ensure that the kind of direct, “primal-only“ analysis we hoped for above
can indeed be performed. Namely, we just need to “precondition“ our input graph by adding to it
a large number of s-t arcs of sufficiently large capacities.
More precisely, we modify our input graph G by adding to it m undirected arcs between the
source s and sink t with a forward and backward capacities equal to 2U and their orientation being
from s to t. We will call these arcs preconditioning arcs. Observe that after this modification the
number of arcs of our graph as well as its maximum capacity at most doubled, and the maximum
s-t flow value changed additively by exactly 2mU . In particular, the preconditioning arcs constitute
exactly half of all the arcs and the amount of s-t flow that they alone can support is at least twice
the s-t throughput of the rest of the graph. (Also, as these arcs are undirected they do not interfere
with our initialization procedure – cf. Lemma 3.2.3 ) Consequently, we can just focus on analyzing
the running time of our algorithm on this preconditioned instance and the bounds we establish will
immediately translate over to the original instance.4
3
More precisely, we can just first initialize our framework for the original graph, as before, and only precondition
the undirected graph resulting from that initialization.
4
Note that the preconditioning arcs have to be fully saturated in any maximum s-t flow. So, simply dropping
these arcs and the flow on them will yield the maximum s-t flow in the original graph.
15
As already mentioned, somewhat surprisingly, once such preconditioning arcs are in place and
bf
our primal dual solution (f , y ) is well-coupled, it is always the case that the symmetrization G
of our current residual graph Gf retains a constant fraction of the s-t throughput. Intuitively
speaking, well-coupling prevents the “shortcutting” preconditioning arc from getting “clogged” too
quickly. Instead, their residual capacity is consumed at the same rate as that of the rest of the
graph. Consequently, these arcs alone are always able to provide enough of s-t throughput in the
bf of the residual graph Gf . This is made precise in the following lemma, whose
symmetrization G
proof appears in Appendix A.4
Lemma 3.8. Let (f , y ) be a well-coupled primal dual solution in the (preconditioned) graph G and
let f be a χα -flow, for some 0 ≤ α < 1, that is feasible in G. We have either that:
bf of the residual graph Gf ;
(a) there exists a χ (1−α) -flow f ′ that is feasible in the symmetrization G
10
(b) or
χT y
>
2m
(1−α)
implying that our target demand χ cannot be routed in G (cf. Lemma 3.1).
Note that if our target demand χ is exactly the demand F ∗ χs,t of the maximum s-t flow, the
second condition cannot ever trigger and thus indeed it is the case that the symmetrization of the
(preconditioned) residual graph retains a constant fraction of the original s-t throughput.
bf of the residual graph Gf retains
Lower Bounding δ̂. Once we proved that the symmetrization G
most of its s-t flow throughput (see Lemma 3.8), we are finally able to provide an absolute lower
bound δ̂ (cf. (19)) on the amount of primal progress each iteration of our algorithm makes. To
this end, we upper bound first the energy, or, (almost) equivalently, the ℓ2 -norm of the congestion
vector (see Lemma 3.3) of the electrical flow that we use in our augmentation step (see (14)).
Lemma 3.9. Let (f , y ) be a well-coupled primal dual solution, with f being a χα -flow that is feasible
in Gf , for some 0 ≤ α < 1. Let fe be an electrical χ-flow determined by the resistances r given by
(13). We have that either:
(a) kρk22 ≤ Er (fe) ≤
explicit constant;
(b) or, χT y >
CE m
,
(1−α)2
2m
(1−α) ,
where ρ is the congestion vector defined in (16), and CE > 0 is an
i.e., our target demand χ cannot be routed in G.
Proof. By Lemma 3.8, we either have that χT y >
2m
(1−α) ,
i.e., our condition (b) triggers, or there
bf of the residual graph Gf .
exists a χ (1−α) -flow f ′ that is feasible in the symmetrization G
10
Let us first bound the energy Er (f ′ ) of that flow f ′ . To this end, observe that, by definition
(13), we have that
′
Er (f ) =
X
e
re (fe′ )2
≤
X
e
1
1
+ −
+
2
(b
ue (f ))
(b
ue (f ))2
(fe′ )2
X f ′ 2
e
≤ 2m,
≤2
u
b
(f
)
e
e
bf . (Recall that the
where the last inequality follows by the fact that the flow f ′e is feasible in G
bf are equal to u
forward and backward capacities of an arc e in G
be (f ).)
16
Consequently, if we scale f ′ by
10
(1−α)
then it will become a χ-flow and its energy will be at most
200
m. However, by definition, the energy Er (fe) of the electrical χ-flow fe cannot be larger than
(1−α)2
the energy of any other χ-flow. We thus can conclude that
kρk22 ≤ Er (fe ) ≤
200
100
Er (f ′ ) ≤
m,
(1 − α)2
(1 − α)2
where we also used Lemma 3.3. So, taking CE := 200 concludes the proof.
Now, we should notice that by Lemma 3.7 it suffices that we always have that
δ≤
1
(1 − α)
1
≤
≤ √
,
33kρk4
33kρk2
33 CE m
(20)
where we used Lemma 3.9 and the simple fact that it is always the case that kρk4 ≤ kρk
2 .3 Conse
√
e m 2 log U
quently, by (19), we see that we can take δ̂ := (33 CE m)−1 , which gives us the desired O
time algorithm.
Finally, we want to emphasize again that even though our above analysis was based solely on
analyzing our primal progress5 , maintaining the dual solution and the primal dual coupling (11)
was absolutely critical for its success.
10 1
e
An Improved O m 7 U 7 -Time Algorithm
4
e m 107 U 17 ,
In this section, we present an improved maximum s-t flow algorithm that runs in time O
where U is the value of the largest (integer) capacity of the graph. In the case of the unit capacities,
10
e (m) 7 -time algorithm of Mądry
i.e., U = 1, this running time matches the running time of the O
e (mU ) 107 , O
e (m√n log U )} that stems for
[43], and improves over the best known bound of min{O
the work of Mądry [43] and Lee and Sidford [38] whenever U is moderately large (and the graph is
sufficiently sparse).
e m 23 log U -time algorithm presented in Section 3.
To achieve this goal, we will modify the O
These modifications will enable us to lower bound the amount of primal progress δ̂ (cf. (19)) this
algorithm makes in each iteration. Specifically, we aim to show that
1 −1
,
δ̂ ≥ m 2 −η
(21)
1
− 71 logm U − O(log log mU ), whenever that modified algorithm executes progress
for certain η := 14
steps (see Section 3.2). Clearly,
in Section 3.3, establish that lower bound on δ̂ ensures
1 as discussed
10
1
−η
e
e
that we need at most O m 2 log U = O m 7 U 7 progress steps to converge to the optimum
solution.
5
In fact, one could perform it even without resorting explicitly to the dual infeasibility certificates χT y .
17
4.1
Modified Algorithm
e m 32 log U -time algorithm we presented in Section 3, let
Before we explain how we modify the O
us understand first what is its main running time bottleneck.
To this end, recall that the key quantity that captures the progress that each augmentation
iteration makes is the ℓ4 -norm kρk4 of the congestion vector ρ of the electrical χ-flow fe determined
by the resistances r (see (13)). More precisely, by Lemma 3.7, we need to always have that
(1 − α)δ̂ ≤ δ ≤ (33kρk4 )−1 .
In Section 3.3, we bounded that ℓ4 -norm by simply noticing that it always has to be upper
bounded by the ℓ2 -norm of that vector and that, by Lemma 3.3, this ℓ2 -norm is directly tied to
the energy Er (fe ) ofthe electrical
flow fe . Furthermore, since we were able to also prove that this
√
m
(see Lemma 3.9) the final O( m) bound on δ̂−1 followed. (See (20)
energy is always O (1−α)
2
for details.)
Now, the key observation is that such bounding of the ℓ4 -norm of the vector ρ with its ℓ2 -norm
might be wasteful. That is, even though it is not hard to construct examples in which the congestion
vector ρ has these two norms be fairly close, such examples are inherently fragile. More precisely,
as first pointed out by Christano et al. [7] in the context of an analogous ℓ∞ - vs ℓ2 -norm trade off,
whenever kρk4 ≈ kρk2 , it must be necessarily be the case that most of the energy of the electrical
flow fe is contributed by a very small set of arcs. Moreover, if one perturbs such arcs by increasing
their resistance, this will result in a great increase of the energy of the corresponding electrical flow.
Christano et al. [7], and later Mądry [43] have demonstrated that a careful exploitation of this
phenomena can lead to ensuring that such fragile worst-case instances do not appear too often and
thus a better bound on the ℓ4 -norm can be achieved.
Arc Boosting. Our improved algorithm will also follow the perturbation paradigm we outlined
above. To this end, we introduce an operation of boosting that we will use to perturb high-energy
contributing arcs whenever needed.
Formally, let us fix a primal dual well-coupled solution (f , y ) and a particular arc e = (u, v).
be (f ) is completely symmetric). We
be (f ) (the case of u
b−
Let us assume wlog that u
b+
e (f ) = u
e (f ) = u
define a boost of e to be an operation that modifies the input graph and the solution (f , y ) in the
following way:
(1) The arc e is replaced by an u-v path consisting of β(e) := 2 + ⌈ ub2U
⌉ arcs e1 , . . . , eβ(e) , all
e (f )
oriented towards v.
−
−
−
+
+
(2) The first two arcs e1 , e2 are just copies of the arc e, with u+
e1 = ue2 := ue and ue1 = ue2 := ue .
−
(3) For all the remaining (β(e) − 2) arcs, we have that u+
e, for 2 < i ≤ β(e),
ei := +∞ and uei := u
where u
e ≥ U is a value that we will set below.
(4) The flow f pushes the original flow fe over the created u-v path.
(5) Finally, if v0 = u, v1 , . . . , vβ(e) = v are the consecutive vertices on the created path then:
• yv0 := yu and yvβ(e) := yv ;
18
• yv1 := yv , and yv2 := yv1 + Φe (f ) = yv + Φe (f );
(y )
∆
Φe (f )
e2
= − β(e)−2
, for each 2 < i ≤ β(e).
• yv3 , . . . , yvβ(e)−1 are set so as ∆ei (y ) = − β(e)−2
Observe that after the above modifications, we have that
∆e1 (y ) − Φe1 (f ) = ∆e (y) − Φe (f )
∆e2 (y ) = Φe2 (f ) = Φe (f )
1
Φei (f ) = −
= ∆ei (y ),
(e
u + fe )
for each 2 < i ≤ β(e), as long as we set
u
e :=
β(e) − 2
− fe .
Φe (f )
Therefore, the solution (f , y ) remains feasible in the (modified) residual graph Gf and is wellcoupled too. Also, note that since Φe (f ) ≤ ube1(f ) and |fe | ≤ U , the fact that (β(e) − 2) ≥ ub2U
e (f )
ensures that u−
e ≥ U , for all 2 < i ≤ β(e). Consequently, the effective forward and backward
ei = u
capacities of the created path are the same as the forward and backward capacities of the original
arc e, and thus the s-t throughput of our graph remains unchanged.
We thus see that, from our point of view, the boosting operation had only two real consequences:
one of them is the intended effect and the other one is an undesired side effect. The intended effect
is that the effective resistance with respect to the resistances r (cf. (13)) of the created path is at
least twice as large as the original resistance re of the arc e. (Note that the arcs e1 and e2 have
exactly the same residual capacities, and thus the resistances re1 and re2 are both equal to the
resistance re of the original arc e.) The undesired consequence, however, is that the number of arcs
in our graph increased by β(e) − 1.
The key impact of that (at least) doubling of the resistance of the arc e is that, intuitively, it
makes the energy of our electrical χ-flow fe increase proportionally to the initial contribution of
that arc to the (squared) ℓ2 -norm of the congestion vector ρ. This is made precise in the following
lemma. (This lemma is stated in a slightly more general form for future reference. For now, one
can think of the set S consisting of a single arc e.)
Lemma 4.1. Let fe be an electrical χ-flow determined by the resistance r in our graph G and let ρ
′
be the corresponding congestion vector defined by (16). If fe is the electrical χ-flow determined by
the resistances r ′ in the graph G′ that was obtained by boosting a set of arcs S then
!!
X ρ2
′
1
e
Er (fe).
Er ′ (fe ) ≥ 1 +
8
kρk22
e∈S
Proof. Let φ̃ be the vertex potentials inducing fe, by Lemma 2.2 and (16), we have that
and χT φ̃ = Er (fe).
1
=
Er (fe )
X
e′ =(u′ ,v′ )
19
(φev′ − φeu′ )2
re′ Er (fe)2
′
We want now to use the same voltages φ̃ to lowerbound the energy Er ′ (fe ) of the electrical χ-flow
′
′
fe after boosting the arcs in S. Observe that the effective resistance ru,v
between the endpoints u
and v of some boosted arc e = (u, v) ∈ S after that boost is at least 2re . So, when we analyze the
′
energy of fe we can treat the u-v path that was created as just an arc e but with the resistance re′
′
equal to ru,v
≥ 2re .
Therefore, by using Lemma 2.2 again, we can conclude that
2
X
X (φev − φeu )2
e
e
(φv′ − φu′ )
1
+
≤
′
re′ Er (fe )2
re′ Er (fe)2
Er ′ (fe )
e′ =(u′ ,v′ )∈S
/
e=(u,v)∈S
2
X (φev′ − φeu′ )2
X
X (φev − φeu )2
e
e
1
(φv − φu )
−
≤
1−
=
2
2
e
e
e
re′ Er (f )
2re Er (f )
Er (f )
2re Er (fe)
e′ =(u′ ,v′ )
e=(u,v)∈S
e=(u,v)∈S
!
!
X re fe2
X ρ2
1
1
e
e
1−
≤
1−
≤
,
4kρk22
Er (fe )
2Er (fe)
Er (fe )
e∈S
e∈S
where we also used Lemma 3.3 and Ohm’s law (5). The lemma follows by noticing that
(1 + 2ε ) whenever 0 ≤ ε ≤ 14 .
1
(1−ε)
≥
Boosting Use and Early Termination. Once we introduced the boosting operation, we can
state the two modifications of the basic algorithm from Section 3 that we will make.
The first one is relatively minor: we terminate our algorithm whenever
1
(1 − α)F ≤ m 2 −η .
(22)
1
That is, once we know that we are within an additive factor of m 2 −η of the target flow, we do not
execute progress steps anymore. Observe that in that case we can just round our current flow and
1
compute the optimal solution using at most m 2 −η augmenting paths computation (see, e.g., [43] for
details). So, if we perform these operations, our algorithm will still run within the desired running
time bound.
The second modification specifies how and when we apply boosting operations. To this end,
given an electrical χ-flow fe and its congestion vector ρ, let us define an arc e to be high-energy if
1
|ρe | ≥ ρ∗ =
m 2 −3η
,
Cρ∗ (1 − α)
(23)
for certain constant Cρ∗ > 0 that we will set later (see (28) below).
Now, recall that our goal is to ensure that (21) holds, which corresponds to ensuring that, by
1 −η
m2
. However, for reasons that will be clear shortly, we will actually want
Lemma 3.7, kρk4 ≤ 33(1−α)
to have a stronger, ℓ3 -norm bound on ρ. Specifically, we want to make a progress step only when
1
m 2 −η
kρk3 ≤
,
33(1 − α)
and then use the step size
δ=
1
δ̂
=
(1 − α)
33(1 − α)kρk3
20
(24)
when making the augmentation step described in Section 3.2. Observe that as kρk3 ≥ kρk4 , this
new condition still satisfies the ℓ4 -norm requirement set forth in Lemma 3.7.
Unfortunately, as mentioned earlier, it can sometime happen that the bound (24) is violated.
Therefore, we want to use our boosting technique to enforce that such violations do not happen too
often. Specifically, whenever condition (24) does not hold we boost all the arcs in the set S ∗ , which
is defined to be the set of up to k∗ := m4η high-energy arcs e with highest values of |ρe |. In other
words, we order all the high-energy arcs in a non-increasing order with respect to |ρe | and take S ∗
to consist of the first up to k∗ of them. (Note that, in principle, in some iterations there can be
only very few, or even no, high-energy arcs.)
Once we boost the set S ∗ , we simply proceed to the next iteration (in the new version of the
graph that our boosting operation created). Note that, at first glance, there is nothing preventing
us from needing to execute such a boosting step in each iteration, and thus to never be able to
execute progress steps. However, as we argue in the next section, we are actually guaranteed to be
able to execute progress steps often enough and thus make sufficiently good progress overall.
4.2
Analysis of the Improved Algorithm
We proceed now to analyzing our modified algorithm and establishing the improved running time
bound. This will require tackling two issues.
First one is controlling the increase in the number of arcs resulting from performing of all our
boosting operations. Specifically, we want to maintain the following invariant.
Invariant 4.2. The total increase in the number of arcs is at most
m
10 .
As we will see, ensuring that this invariant is maintained will be what determines our final bound
on the value of η (cf. (21)).
We need to maintain this invariant in order to be able to use the machinery we developed in
Section 3. (It is not hard to check that, as long as that invariant is not violated, the whole analysis
performed there is still valid, subject to slight adjustment of the corresponding constants.)
The other issue we need to tackle corresponds to bounding the total number of boosting steps
we execute. After all, we want to ensure that we execute progress steps often enough. In order
to obtain such a bound we will perform a potential based argument with our potential function
being the energy Er (fe ) of the electrical χ-flow fe we compute in each iteration. (Note that this
electrical flow fe is always determined by the resistances r (cf. (13)) in the current version of the
graph G.) One can easily convince oneself that this energy has to be at least U12 at the beginning
of the algorithm and, by Lemma 3.9 (and provided Invariant 4.2 holds), it is never larger than
Er (fe) ≤
CE m
≤ O(m3 U 2 ).
(1 − α)2
(Here, the second inequality follows due to the fact that 1 ≤ F ≤ 2mU and we have our termination
condition (22).) We want now to establish two claims:
(I) Whenever a boosting
step is executed, the energy Er (fe ) increases by a factor of at least
−2η
1 + CB m
, for some constant CB > 0. (To prove this statement we will use Lemma 4.1.)
(II) Whenever a progress step is performed,
with the condition (24) satisfied, the energy Er (fe )
−2η
decreases by at most 1 + CD m
, for some other constant CD > 1. (Proving this statement
is where we leverage the fact that we enforce the stronger, ℓ3 -norm condition (24) on ρ.)
21
Observe that once the above two claims are established, we can amortize the number of boosting
steps against the number of good progress steps. Specifically, a simple calculation shows that unless
the number of boosting step is within a O(log m log U ) factor of the number of the progress steps,
the resulting increase of the energy Er (fe ) would violate the absolute upper bound O(m3 U 2 ) we
established. Furthermore, since we execute a progress step only when condition (24) is satisfied,
this means that our desired lower bound (21) on δ̂ holds. Therefore,
by our
reasoning
1 in Section
−1
e δ̂ log mU = O
e m 2 −η log mU .
3.3, we know that the number of progress steps is at most O
As each boosting and progress step can be executed in nearly-linear time, our desired improved
running time bound will then follow.
In the light of the above, it remains to only argue that claims (I) and (II) are indeed correct
and that the Invariant 4.2 is never violated.
Preservation of the Invariant 4.2. To prove that the Invariant 4.2 is indeed preserved, let us
, where
recall that each arc boosting operation increases the number of arcs by β(e) − 1 ≤ ub2U
e (f )
u
be (f ) is the residual capacity of the arc e at the time the boosting occurred. As it turns out, since
we only boost arcs that are high-energy, i.e., arcs e with |ρe | ≥ ρ∗ at the time of the boosting,
one can use a simple vertex potential based argument to argue that β(e) has to be always at most
O(m4η U ).
Lemma 4.3. Assume that Invariant 4.2 holds, let fe be an electrical χ-flow determined by the
resistances r and let e = (u, v) be a high-energy arc that was boosted. Then, the total arc number
increase β(e) − 1 is at most O(m4η U ).
2U
u
be (f ) .
So, it suffices that we prove that u
be (f ) ≥ Ω(m−4η ). To this
end, let φ̃ be vertex potentials inducing the electrical flow fe and let Rs,t be the effective resistance
between the source s and sink t in our graph. As fe is an s-t flow of value F , we need to have that
Proof. Recall that β(e) − 1 ≤
F 2 Rs,t = Er (fe) ≤
CE m
,
(1 − α)2
where we used Lemma 3.9 and the fact that Invariant 4.2 holds.
On the other hand, we know that the vertex potential drop φet −φes between s and t is non-negative
(as the electrical flow has to flow from s to t) and at most
1
φet − φes = F Rs,t ≤
CE m 2 +η
CE m
≤
,
(1 − α)2 F
(1 − α)
1
where the last inequality follows as (1 − α)F ≥ m 2 −η due to our termination condition (22).
However, as fe is an s-t flow it must be the case that the vertex potential difference between the
endpoints of our arc e = (u, v) cannot be larger than such difference between the vertex potentials
of s and t. That is, we have that
1
CE m 2 +η
.
|φev − φeu | ≤ |φet − φes | ≤
(1 − α)
(25)
But, by Ohm’s law (5) and definition (13) of the resistance r , we have that
1
|φev − φeu | = |fee |re ≥
|ρe |
ρ∗
m 2 −3η
|fee |
=
≥
≥
,
u
be (f )2
u
be (f )
u
be (f )
Cρ∗ (1 − α)b
ue (f )
22
where we used the definition (23) of the high-energy arcs.
So, putting this inequality together with the bound (25), gives us the desired lower bound on
u
be (f ). The lemma follows.
∗
4η
Observe now that as our boosting
operation never boosts more than k = m arcs at a time
1
−η
e m 2 log U boosting steps, by Lemma 4.3, the total arc number increase
and we have at most O
throughout the whole algorithm is at most
e m 21 −η log U · O(m4η U ) = O
e m 12 +7η U < m ,
m4η · O
10
1
− 17 logm U − O(log log mU ) with appropriately chosen constant in the
provide we set η := 14
O(log log mU ) term. Thus Invariant 4.2 is indeed never violated.
Establishing Claim (I). We want now to prove that each boosting step results in an increase
of the energy Er (fe) by a factor of at least 1 + Ω(m−2η ) . In the light of Lemma 4.1, it suffices we
show that if S ∗ is the set of high-energy arcs that got boosted due to the condition (24) not holding
then these arcs had to have a large contribution to the (square) of the ℓ2 -norm of the congestion
vector ρ. That is, we want to argue that
X ρ2
e
−2η
).
2 ≥ Ω(m
kρk
2
∗
(26)
e∈S
Clearly, once we establish that then the desired energy increase will follow from Lemma 4.1 applied
to S = S ∗ .
To prove (26), let us first consider the case that S ∗ has maximum size, i.e., |S ∗ | = k∗ = m4η .
Note that all the arcs in S ∗ are high-energy arcs. Therefore, (23) gives us in this case that
∗ 2
X ρ2
m1−2η
e
4η (ρ )
≥ Ω(m−2η ),
≥
m
·
≥
kρk22
kρk22
Cρ2∗ (1 − α)2 kρk22
∗
e∈S
CE m
where we used the fact that Lemma 3.9 implies that kρk22 ≤ (1−α)
2 . So, (26) holds in this case
∗
4η
and we can focus on the situation when |S | < m . Observe that this means that there is no
high-energy arcs that are not in S ∗ .
We want to argue that in such situation the arcs of S ∗ contribute at least a half of the total
(cubed) ℓ3 -norm of the vector ρ. That is, that we have that
X
e∈S ∗
1
|ρe |3 ≥
m 2 −η
66(1 − α)
!3
(27)
.
To this end, observe that the contribution of arcs that are not high energy (and thus not in S ∗ )
to that (cubed) ℓ3 -norm is, by the Cauchy-Schwartz inequality, (23) and Lemma 3.9, at most
X
e∈S
/ ∗
|ρe |3
!
≤
max |ρe |
e∈S
/ ∗
X
e∈S
/ ∗
ρ2e
!
3
CE m 2 −3η
≤ ρ∗ kρk22 ≤
≤
Cρ∗ (1 − α)3
23
1
m 2 −η
66(1 − α)
!3
,
(28)
provided the constant Cρ∗ is set to be large enough. Therefore, since the condition (24) does not
hold, we must have that the inequality (27) is indeed valid.
Now, if there exists an arc e ∈ S ∗ such that |ρe | ≥
1
m 2 −η
(1−α)
then the contribution of this arc
alone suffices to make (26) true. We can thus assume that maxe∈S ∗ |ρe | <
Cauchy-Schwartz inequality to (27) to obtain that
X
e∈S
P
|ρe |3
1
ρ2e ≥
≥
maxe∈S ∗ |ρe |
66
∗
e∈S ∗
1
m 2 −η
66(1 − α)
!2
1
m 2 −η
(1−α)
and apply the
,
which, by Lemma 3.9, establishes (26). Claim (I) is thus proved.
Establishing Claim (II). To prove Claim (II) we will tie the potential decrease of energy during
progress steps to the quantity that we already control: the ℓ3 -norm of the congestion vector ρ. (In
fact, the desire to control the energy decrease is precisely the reason why we enforce the stronger
ℓ3 -norm condition (24) instead of the ℓ4 -norm condition that Lemma 3.7 suggests.)
We make this connection precise in the following lemma, whose proof appears in Appendix A.5.
Lemma 4.4. Let (f , y ) be a well-coupled feasible primal dual solution and let fe be the corresponding
electrical χ-flow determined by the resistances r and congestion vector ρ that satisfies the ℓ3 -norm
condition (24). Then, after execution of the progress steps, as described in Section 3.2, the energy
Er (fe) decreases by a factor of at most
kρk23
1+O
.
kρk22
Observe that, by (24) and Lemma 3.9, we have that
m1−2η (1 − α)2
kρk23
≤
= O(m−2η ),
(33)2 CE m(1 − α)2
kρk22
which immediately establishes Claim (II) and thus concludes the analysis of the improved algorithm.
Acknowledgments. We are grateful to Michael Cohen, Slobodan Mitrović, Dimitris Tsipras, and
Adrian Vladu for a number of helpful discussions on this topic.
24
A
A.1
Appendix
Proof of Fact 3.4
Let us define
g(x) :=
By Taylor’s theorem, we have that
1
1
−
u1 − x u2 + x
1
1
x2
=
−
+
g(x) = g(0) + g (0)x + g (z)
2
u1 u2
′
′′
1
1
+ 2
2
u1 u2
.
x+x
2
1
1
−
3
(u1 − z)
(u2 + z)3
where z is some value with |z| ≤ |x|. As a result, we can conclude that
64
1
64
5
1
≤
−
+
≤ 3,
|ζ| =
3
3
3
3
(u1 − z)
(u2 + z)
u
27u1 27u2
(29)
(30)
as we wanted to show.
A.2
Proof of Lemma 3.5
By the definition of our update (14) and the approximation bound (18) as well as the coupling
condition (11), we have that
1
1
b
e
e
+ − 2 δfee − (δfee )2 ζe
∆e (b
y ) − Φe (f ) = ∆e (y ) + δ(φv − φu ) − Φ(f ) −
2
(b
u+
)
(b
ue )
e
γ
e
≤ δ (φev − φeu ) − re fee − (δfee )2 ζe +
u
be (f )
e2
=
7δ2 fube2 + 34 γe
7(δρe )2 + 43 γe
γe
e
(δfee ) ζe +
=
,
≤
u
be (f )
u
be (fb)
u
be (fb )
2
where the second to last equality follows by Ohm’s law (6), the last equality follows by (16) and we
be (f ) since δ ≤ (4kρk∞ )−1 . The lemma follows.
also used the fact that u
be (fb) ≥ 43 u
A.3
Proof of Lemma 3.6
For each arc e, let us define
θe :=
1
1
+ −
2
(b
u+
(g
))
(b
u
(g
))2
e
e
−1
(31)
(∆e (z ) − Φe (g )) .
Intuitively, θe is the first-order correction to the coupling of the primal dual solution (g , z ) with
respect to the arc e.
Note that by the fact that (g , z ) is ς-coupled we have that
|θe | ≤
1
1
+ −
+
2
(b
ue (g ))
(b
ue (g ))2
−1
2
|∆e (z ) − Φe (g )| ≤ u
be (g )
25
ςe
u
be (g )
= ςe u
be (g ).
(32)
Let us now define a flow g ′ to be the flow g with the “correcting“ flow θ added. That is,
ge′ := ge + θe ,
1
and, by (32), |θe | ≤ ςe u
for each arc e. Observe that, as kςk∞ ≤ kςk2 ≤ 50
be (g ), this flow g ′ is
51
u
be (g ) ≥ u
be (g ′ ) ≥ 49
be (g ).
feasible in Gg and, in fact, 50
50 u
Furthermore, if we analyze the coupling of (g ′ , z ), by Fact 3.4, for each arc e, we get that
1
1
′
− −
∆e (z ) − Φe (g ) = ∆e (z ) −
u
b+
u
be (g ) + θe
e (g ) − θe
1
1
= ∆e (z ) − Φe (g ) −
+ −
θe − θe2 ζe′
2
2
(b
u+
(g
))
(b
u
(g
))
e
e
2ς 2
2
5b
u
(g
)
5ς
51ςe2
e
e
e
= −θe2 ζe′ ≤
=
≤
,
u
be (g )3
u
be (g )
10b
ue (g ′ )
where the third equality follows by the definition of θe (31) and the second to last inequality follows
by (32).
e -coupled with
We thus see that the solutions (g ′ , z ) are well-coupled. In fact, they are γ
v
u
s
X
uX 11ςe2 2 51
51
2
t
ke
γ k2 =
kςk22 ≤
.
(33)
≤
γe =
e
2
10
25000
e
e
b )-flow, where
So, these solutions (g ′ , z ) have all the desired properties except g ′ is an (χα′ + σ
b are the demands of the “correcting“ flow θ we added – and not a χα′ -flow we need.
σ
To remedy this deficiency, we compute an electrical (−b
σ)-flow θb determined by the to resistances
rbe :=
1
′ 2
(b
u+
e (g ))
+
1
,
′ 2
(b
u−
e (g ))
(34)
and obtain our desired solutions (ḡ , z̄ ) by adding this electrical flow to the flow g ′ , and simultab that induce θb to the dual solution z . In other words, we
neously adding the vertex potentials φ
set
ḡ
z̄
:= g ′ + θb
b
:= z + φ.
(35)
Clearly, ḡ is a χα′ -flow, as desired. It thus only remains to ensure that the solutions (ḡ , z̄ ) are still
well-coupled.
To analyze the coupling of (ḡ , z̄ ) one should note first that the procedure we used to obtain
b)
these solutions from the solutions (g ′ , z ) is analogous to the one we used obtain the solutions (fb , y
from the solutions (f , y ) in our augmentation step. More precisely, the update (35) is an exact
analogue of the update (14) with δ = 1.
So, to quantify the change in coupling we can follow the approach we used when analyzing the
augmentation step. Specifically, let us define a congestion vector ρe as
ρbe :=
θbe
,
u
be (g ′ )
26
for each arc e. By Lemma 3.3, we know that
b ≤ Erb (−θ),
e 22 ≤ Erb (θ)
kρk
where the second inequality follows by the energy minimizing property of electrical flows. (Note
that −θ is a (−b
σ )-flow.) We thus have that, by (34),
X
X
1
1
2
2
rbe (−θe ) =
Erb (−θ) =
+ ′ 2 +
− ′ 2 θe
(b
u
(g
))
(b
u
(g
))
e
e
e
e
X 51ςe 2 21
X
1
2
2
,
(ςe u
be (g )) ≤
2
≤ kςk22 ≤
≤
′
2
u
be (g )
50
20
2000
e
e
where the first inequality follows by (32). As a result, we can conclude that
e 22 ≤ Erb (−θ) ≤
kρk
1
,
2000
(36)
which ensures, in particular, that our analogue of condition (17) (for δ = 1) is satisfied.
b -coupled with
Consequently, we can use Lemma 3.5 to conclude that the solutions (ḡ , z̄ ) is γ
v
u
s
2
X
uX 4
204
7
1
4
2
2
t
e 22 ≤
γ
ee + 7b
ρe
γ k2 + 7kρk
+
<
,
≤ ke
γ
be ≤
kb
γ k2 =
3
3
75000
2000
100
e
e
where we used the estimates (33) and (36). So, (ḡ , z̄ ) is well-coupled, as desired.
A.4
Proof of Lemma 3.8
2m
Let us assume that χT y ≤ (1−α)
. Otherwise, by Lemma 3.1, we have a dual certificate that the
target demand χ cannot be routed in G, i.e., the condition (b) triggers.
As all the preconditioning arcs have s and t as their endpoints and are oriented from s to t, we
have that
χT y
2m
∆e (y ) = yt − ys =
≤
,
(37)
F
(1 − α)F
for each such arc e. Also, all these preconditioning arcs are completely indistinguishable from the
point of view of our algorithm. They have the same endpoints, orientation, and initial (residual)
capacities and stretch. Consequently, their resistances are always equal when we compute our
electrical flows (see, e.g., (13)) and thus, in turn, their forward and backward residual capacities
evolve in exactly the same way during each of the augmentation and fixing steps (see Section 3.2).
We want to argue that, for each such arc e,
u
be (f ) ≥
(1 − α)F
.
5m
(38)
bf of the residual
As the residual capacity u
be (f ) is exactly the capacity of e in the symmetrization G
graph Gf , and preconditioning arc constitute exactly half of all the m arcs, a flow f ′ that just routes
(1−α)F
bf , and
units of flow from s to t on each such arc will be the desired χ (1−α) -flow feasible in G
5m
10
our proof will follow.
27
To establish (38), let us fix some representative preconditioning arc e and assume for the sake
of contradiction that (38) does not hold. Since the total preconditioning arc capacity constitutes at
least 23 -fraction of the total s-t capacity F ∗ of the graph G, we must have that
−
u+
e = ue ≥
2 2 ∗ 4F ∗
· F =
,
m 3
3m
(39)
−
where u+
e and ue are the initial capacities of the preconditioning arcs. As a result, if (38) did not
hold, by (3), we know that
−
|fe | ≥ max{u+
b+
b−
e −u
e (f ), ue − u
e (f )} ≥
5 ∗
F + αF
4F ∗ (1 − α)F
2F ∗
−
≥ 2
>
,
3m
5m
3m
3m
where we used the fact that F ≤ 23 F ∗ as otherwise we would be able to immediately tell that it is
impossible to route at least 32 -fraction of the s-t flow only via preconditioning arcs.
In fact, we also know that fe > 0. This is so, as e is oriented from s to t and thus having
∗
fe < 0 would mean that there is a flow of strictly more than F3 from t towards s over all the
preconditioning arcs. It would, however, be impossible to balance this “backward” t-s flow via an
∗
s-t flow over the remaining arcs as they can support an s-t flow of at most F3 . So, indeed fe > 0.
Furthermore, if (38) would indeed not hold then fe > 0, (3) and (39) would imply that
u
b+
be (f ) <
e =u
(1 − α)F
F∗
1
1 −
u
b− (f )
≤
≤ u−
ue (f ) − fe ) ≤ e
,
e = (b
3m
2m
2
2
2
where we again used the fact that F ≤ 32 F ∗ .
Now, combining the above bound with the fact that (f , y ) is well-coupled (11), we can conclude
that, for any preconditioning arc e,
γe
1
(1 − γe )
1
1
γe
= +
≥
−
− −
−
u
be (f )
be (f )
u
be (f )
2b
ue (f )
u
be (f ) u
be (f ) u
2
5m(1 − √m )
(1 − 2γe )
2m
≥
>
,
2b
ue (f )
2(1 − α)F
(1 − α)F
∆e (y ) ≥ Φe (f ) −
=
where we again used the assumption that (38) does not hold as well as the fact that γe ≤ √1m since
otherwise the contribution of all the m
2 preconditioning arcs to the ℓ2 -norm of the violation vector
γ alone would make this norm violate the well-coupling bound. However, the above lower bound on
∆e (y ) directly violates our bound (37). So, we reached a contradiction that proves (38) and thus
our lemma follows.
A.5
Proof of Lemma 4.4
By examining the augmentation and fixing steps that are performed during the progress steps (cf.
(14) and (35) in the proof of Lemma 3.6) and applying standard Taylor approximations to the
resulting changes in the resistances r (cf. (13)), one obtains that, for each arc e, the resulting arc
resistance re′ change is such that
re′ ≤ (1 + O(δ|ρe | + κe ))−1 re ,
28
(40)
where ρ is the congestion vector (16), δ = (33kρk3 )−1 is the step size, and κ is a vector such that
kκk2 ≤ 1. (Roughly speaking, the δρe term corresponds to augmentation step and the κe term
corresponds to the fixing step.)
Now, let φe be the vertex potentials inducing the electrical χ-flow fe. By Lemma 2.2, we know
that
X (φev − φeu )2
1
=
.
Er (fe)
re Er (fe )2
e=(u,v)
′
′
Furthermore, if Er ′ (fe ) is the energy of an electrical χ-flow fe determined by the new resistances
r ′ , applying Lemma 2.2 again, we obtain that
1
′
Er ′ (fe )
X (φev − φeu )2
X (1 + O(δ|ρe | + κe )) (φev − φeu )2
≤
re′ Er (fe )2
re Er (fe )2
e=(u,v)
e=(u,v)
!!
X (δρe + κe )(re fe2 )
1
e
1+O
=
e
e
Er (f )
Er (f )
e
!!
3
X (δρ + κe ρ2 )
1
e
e
≤
1+O
,
kρk22
Er (fe )
e
≤
where we used (40), Ohm’s law (5) and Lemmas 3.3 and 3.9.
Finally, observe that since δ = (33kρk3 )−1 and kκk2 ≤ 1, an application of the Cauchy-Schwartz
inequality gives us that
X
(δρ3e + κe ρ2e ) ≤ δkρk33 + kκk2 kρk24 ≤ δkρk33 + kκk2 kρk23 ≤ O(kρk23 ),
e
as desired. The lemma thus follows.
References
[1] R. K. Ahuja, T. L. Magnanti, and J. B. Orlin. Network flows: theory, algorithms, and applications. Prentice-Hall, 1993.
[2] R. K. Ahuja, T. L. Magnanti, J. B. Orlin, and M. R. Reddy. Applications of Network Optimization, volume 7 of Handbooks in Operations Research and Management Science. North-Holland,
1995.
[3] H. Alt, N. Blum, K. Mehlhorn, and
p M. Paul. Computing a maximum cardinality matching in
1.5
m/ log n). Inf. Process. Lett., 37(4):237–240, 1991.
a bipartite graph in time O(n
[4] S. Arora, E. Hazan, and S. Kale. The multiplicative weights update method: a meta-algorithm
and applications. Theory of Computing, 8(1):121–164, 2012.
[5] B. Bollobas. Modern Graph Theory. Springer, 1998.
[6] J. R. Bunch and J. E. Hopcroft. Triangular factorization and inversion by fast matrix multiplication. Mathematics of Computation, 28(125):231–236, 1974.
29
[7] P. Christiano, J. Kelner, A. Mądry, D. Spielman, and S.-H. Teng. Electrical flows, Laplacian systems, and faster approximation of maximum flow in undirected graphs. In STOC’11:
Proceedings of the 43rd Annual ACM Symposium on Theory of Computing, pages 273–281,
2011.
[8] M. B. Cohen, R. Kyng, G. L. Miller, J. W. Pachocki, R. Peng, A. B. Rao, and S. C. Xu. Solving
SDD linear systems in nearly m log1/2 n time. In STOC’14: Proceedings of the 46th Annual
ACM Symposium on Theory of Computing, pages 343–352, 2014.
[9] D. Coppersmith and S. Winograd. Matrix multiplication via arithmetic progressions. Journal
of Symbolic Computation, 9:251–280, 1990.
[10] T. H. Cormen, C. E. Leiserson, R. L. Rivest, and C. Stein. Introduction to Algorithms. The
MIT Press, 3rd edition, 2009.
[11] J. Edmonds. Paths, trees, and flowers. Canadian Journal of Mathematics, 17:449–467, 1965.
[12] J. Edmonds and R. M. Karp. Theoretical improvements in algorithmic efficiency for network
flow problems. Journal of the ACM, 19(2):248–264, 1972.
[13] J. Egerváry. Matrixok kombinatorius tulajdonságairól. Matematikai és Fizikai Lapok, 38:16–28,
1931.
[14] P. Elias, A. Feinstein, and C. E. Shannon. A note on the maximum flow through a network.
IRE Transactions on Information Theory, 2, 1956.
[15] S. Even and R. E. Tarjan. Network flow and testing graph connectivity. SIAM Journal on
Computing, 4(4):507–518, 1975.
[16] T. Feder and R. Motwani. Clique partitions, graph compression and speeding-up algorithms.
Journal of Computer and System Sciences, 51(2):261Ű–272, 1995.
[17] L. R. Ford and D. R. Fulkerson. Maximal flow through a network. Canadian Journal of
Mathematics, 8:399–404, 1956.
[18] F. G. Frobenius. Über matrizen aus nicht negativen elementen. Sitzungsberichte der Königlich
Preussischen Akademie der Wissenschaften zu Berlin, pages 456–477, 1912.
[19] F. G. Frobenius. Über zerlegbare determinanten. Sitzungsberichte der Königlich Preussischen
Akademie der Wissenschaften zu Berlin, pages 274–277, 1917.
[20] H. N. Gabow and R. E. Tarjan. Faster scaling algorithms for general graph matching problems.
Journal of the ACM, 38(4):815–853, 1991.
[21] A. V. Goldberg and A. V. Karzanov. Maximum skew-symmetric flows and matchings. Mathematical Programming, 100(3):537–568, 2004.
[22] A. V. Goldberg and S. Rao. Beyond the flow decomposition barrier. Journal of the ACM,
45(5):783–797, 1998.
[23] N. J. A. Harvey. Algebraic algorithms for matching and matroid problems. SIAM Journal on
Computing, 39(2):679–702, 2009.
30
[24] J. Hopcroft and R. Karp. An n5/2 algorithm for maximum matchings in bipartite graphs.
SIAM Journal on Computing, 2(4):225–231, 1973.
[25] A. V. Karzanov. O nakhozhdenii maksimal’nogo potoka v setyakh spetsial’nogo vida i nekotorykh prilozheniyakh. Matematicheskie Voprosy Upravleniya Proizvodstvom, 5:81–94, 1973. (in
Russian; title translation: On finding maximum flows in networks with special structure and
some applications).
[26] J. A. Kelner, Y. T. Lee, L. Orecchia, and A. Sidford. An almost-linear-time algorithm for approximate max flow in undirected graphs, and its multicommodity generalizations. In SODA’14:
Proceedings of the 25th Annual ACM-SIAM Symposium on Discrete Algorithms, pages 217–226,
2014.
[27] J. A. Kelner, L. Orecchia, A. Sidford, and Z. A. Zhu. A simple, combinatorial algorithm for
solving SDD systems in nearly-linear time. In STOC’13: Proceedings of the 45th Annual ACM
Symposium on the Theory of Computing, pages 911–920, 2013.
[28] V. King, S. Rao, and R. Tarjan. A faster deterministic maximum flow algorithm. Journal of
Algorithms, 17(3):447–474, 1994.
[29] D. König. Vonalrendszerek és determinánsok. Mathematikai és Természettudományi Értesitö,
33:221–229, 1915.
[30] D. König. Über graphen und ihre anwendung auf determinantentheorie und mengenlehre.
Mathematische Annalen, 77:453–465, 1916.
[31] D. König. Sur un probléme de la théorie générale des ensembles et la théorie des graphes.
Revue de Métaphysique et de Morale, 30:443–449, 1923.
[32] D. König. Graphok és matrixok. Matematikai és Fizikai Lapok, 38:116–119, 1931.
[33] I. Koutis, G. L. Miller, and R. Peng. Approaching optimality for solving SDD systems. In
FOCS’10: Proceedings of the 51st Annual IEEE Symposium on Foundations of Computer Science, pages 235–244, 2010.
[34] I. Koutis, G. L. Miller, and R. Peng. A nearly m log n-time solver for SDD linear systems.
In FOCS’11: Proceedings of the 52nd Annual IEEE Symposium on Foundations of Computer
Science, pages 590–598, 2011.
[35] R. Kyng, Y. T. Lee, R. Peng, S. Sachdeva, and D. A. Spielman. Sparsified cholesky and
multigrid solvers for connection laplacians. In STOC’16: Proceedings of the 48th Annual ACM
Symposium on Theory of Computing, 2016.
[36] R. Kyng and S. Sachdeva. Approximate Gaussian elimination for Laplacians: Fast, sparse,
and simple. In FOCS’16: Proceedings of the 57th Annual IEEE Symposium on Foundations of
Computer Science, 2016.
[37] Y. T. Lee, S. Rao, and N. Srivastava. A new approach to computing maximum flows using
electrical flows. In STOC’13: Proceedings of the 45th Annual ACM Symposium on the Theory
of Computing, pages 755–764, 2013.
31
[38] Y. T. Lee and
√ A. Sidford. Path finding methods for linear programming: Solving linear programs in Õ( rank) iterations and faster algorithms for maximum flows. In FOCS’14: Proceedings of the 55th Annual IEEE Symposium on Foundations of Computer Science, pages 424–433,
2014.
[39] L. Lovász. On determinants, matchings and random algorithms. Fundamentals of Computation
Theory, 565–574, 1979.
[40] L. Lovász and D. M. Plummer. Matching Theory. Elsevier Science, 1986.
[41] A. Mądry. Fast approximation algorithms for cut-based problems in undirected graphs. In
FOCS’10: Proceedings of the 51st Annual IEEE Symposium on Foundations of Computer Science, pages 245–254, 2010.
[42] A. Mądry. From Graphs to Matrices, and Back: New Techniques for Graph Algorithms. PhD
thesis, Massachusetts Institute of Technology, 2011.
[43] A. Mądry. Navigating central path with electrical flows: from flows to matchings, and back.
In FOCS’13: Proceedings of the 54th Annual IEEE Symposium on Foundations of Computer
Science, pages 253–262, 2013.
p
[44] S. Micali and V. V. Vazirani. An O( |V | · |E|) algoithm for finding maximum matching in
general graphs. In FOCS’80: Proceedings of the 21st Annual IEEE Symposium on Foundations
of Computer Science, pages 17–27, 1980.
[45] M. Mucha. Finding maximum matchings via Gaussian elimination. PhD thesis, University of
Warsaw, 2005.
[46] M. Mucha and P. Sankowski. Maximum matchings via Gaussian elimination. In FOCS’04:
Proceedings of the 45th Annual IEEE Symposium on Foundations of Computer Science, pages
248–255, 2004.
[47] J. B. Orlin. Max flows in O(nm) time, or better. In STOC’13: Proceedings of the 45th Annual
ACM Symposium on the Theory of Computing, pages 765–774, 2013.
[48] R. Peng. Approximate undirected maximum flows in O(mpolylog(n)) time. In SODA’16:
Proceedings of the 27th Annual ACM-SIAM Symposium on Discrete Algorithms, pages 1862–
1867, 2016.
[49] M. O. Rabin and V. V. Vazirani. Maximum matchings in general graphs through randomization.
J. Algorithms, 10(4):557–567, Dec. 1989.
[50] A. Schrijver. On the history of the transportation and maximum flow problems. Mathematical
Programming, 91:437–445, 2002.
[51] A. Schrijver. Combinatorial Optimization: Polyhedra and Efficiency. Springer, 2003.
[52] A. Schrijver. On the history of combinatorial optimization (till 1960). In R. W. K. Aardal,
G.L. Nemhauser, editor, Handbook of Discrete Optimization, pages 1–68. Elsevier, 2005.
32
√
[53] J. Sherman. Breaking the multicommodity flow barrier for O( log n)-approximations to sparsest cuts. In FOCS’09: Proceedings of the 50th Annual IEEE Symposium on Foundations of
Computer Science, pages 363–372, 2009.
[54] J. Sherman. Nearly maximum flows in nearly linear time. In FOCS’13: Proceedings of the 54th
Annual IEEE Symposium on Foundations of Computer Science, pages 263–269, 2013.
[55] D. A. Spielman and S.-H. Teng. Solving sparse, symmetric, diagonally-dominant linear systems in time O(m1.31 ). In FOCS’03: Proceedings of the 44th Annual IEEE Symposium on
Foundations of Computer Science, pages 416–427, 2003.
[56] D. A. Spielman and S.-H. Teng. Nearly-linear time algorithms for graph partitioning, graph
sparsification, and solving linear systems. In STOC’04: Proceedings of the 36th Annual ACM
Symposium on the Theory of Computing, pages 81–90, 2004.
[57] W. T. Tutte. The factorization of linear graphs. Journal of the London Mathematical Society,
22:107–111, 1947.
[58] P. M. Vaidya. Solving linear equations with symmetric diagonally dominant matrices by constructing good preconditioners. Unpublished manuscript, UIUC 1990. A talk based on the
manuscript was presented at the IMA Workshop on Graph Theory and Sparse Matrix Computation, October 1991, Mineapolis.
[59] V. Vassilevska Williams. Multiplying matrices faster than Coppersmith-Winograd. In
STOC’12: Proceedings of the 44th Annual ACM Symposium on the Theory of Computing,
pages 887–898, 2012.
[60] V. p
V. Vazirani. A theory of alternating paths and blossoms for proving correctness of the
O( |V ||E|) general graph matching algorithms. Combinatorica, 14 (1):71–109, 1994.
33
| 8cs.DS
|
Function-Specific Mixing Times and
Concentration Away from Equilibrium
Maxim Rabinovich, Aaditya Ramdas
Michael I. Jordan, Martin J. Wainwright
{rabinovich,aramdas,jordan,wainwrig}@berkeley.edu
University of California, Berkeley
arXiv:1605.02077v2 [math.ST] 30 Sep 2016
October 4, 2016
Abstract
Slow mixing is the central hurdle when working with Markov chains, especially those used for
Monte Carlo approximations (MCMC). In many applications, it is only of interest to estimate the
stationary expectations of a small set of functions, and so the usual definition of mixing based on
total variation convergence may be too conservative. Accordingly, we introduce function-specific
analogs of mixing times and spectral gaps, and use them to prove Hoeffding-like function-specific
concentration inequalities. These results show that it is possible for empirical expectations of
functions to concentrate long before the underlying chain has mixed in the classical sense, and we
show that the concentration rates we achieve are optimal up to constants. We use our techniques
to derive confidence intervals that are sharper than those implied by both classical Markov chain
Hoeffding bounds and Berry-Esseen-corrected CLT bounds. For applications that require testing,
rather than point estimation, we show similar improvements over recent sequential testing results
for MCMC. We conclude by applying our framework to real data examples of MCMC, providing
evidence that our theory is both accurate and relevant to practice.
1
Introduction
Methods based on Markov chains play a critical role in statistical inference, where they form the
basis of Markov chain Monte Carlo (MCMC) procedures for estimating intractable expectations [see,
e.g., 9, 28]. In MCMC procedures, it is the stationary distribution of the Markov chain that typically
encodes the information of interest. Thus, MCMC estimates are asymptotically exact, but their
accuracy at finite times is limited by the convergence rate of the chain.
The usual measures of convergence rates of Markov chains—namely, the total variation mixing
time or the absolute spectral gap of the transition matrix [21]—correspond to very strong notions of
convergence and depend on global properties of the chain. Indeed, convergence of a Markov chain in
total variation corresponds to uniform convergence of the expectations of all unit-bounded function
to their equilibrium values. The resulting uniform bounds on the accuracy of expectations [4, 11,
18, 19, 20, 22, 27, 29] may be overly pessimistic—not indicative of the mixing times of specific
expectations such as means and variances that are likely to be of interest in an inferential setting.
Another limitation of the uniform bounds is that they typically assume that the chain has arrived
at the equilibrium distribution, at least approximately. Consequently, applying such bounds requires
either assuming that the chain is started in equilibrium—impossible in practical applications of
MCMC—or that the burn-in period is proportional to the mixing time of the chain, which is also
unrealistic, if not impossible, in practical settings.
Given that the goal of MCMC is often to estimate specific expectations, as opposed to obtaining
the stationary distribution, in the current paper we develop a function-specific notion of convergence with application to problems in Bayesian inference. We define a notion of “function-specific
1
mixing time,” and we develop function-specific concentration bounds for Markov chains, as well
as spectrum-based bounds on function-specific mixing times. We demonstrate the utility of both
our overall framework and our particular concentration bounds by applying them to examples of
MCMC-based data analysis from the literature and by using them to derive sharper confidence
intervals and faster sequential testing procedures for MCMC.
1.1
Preliminaries
We focus on discrete time Markov chains on d states given by a d × d transition matrix P that
satisfies the conditions of irreducibility, aperiodicity, and reversibility. These conditions guarantee
the existence of a unique stationary distribution π. The issue is then to understand how quickly
empirical averages of functions of the Markov chain, of the form f : [d] → [0, 1], approach the
stationary average, denoted by
µ : = EX∼π [f (X)].
The classical analysis of mixing defines convergence rate in terms of the total variation distance:
dTV p, q =
sup
EX∼p f (X) − EY ∼q f (Y ) ,
(1)
f : Ω→[0, 1]
where the supremum ranges over all unit-bounded functions. The mixing time is then defined as
the number of steps required to ensure that the chain is within total-variation distance δ of the
stationary distribution—that is
n
o
T (δ) : = min n ∈ N | max dTV πn(i) , π ≤ δ ,
(2)
i∈[d]
(i)
where N = {1, 2, . . .} denotes the natural numbers, and πn is the distribution of the chain state
Xn given the starting state X0 = i.
Total variation is a worst-case measure of distance, and the resulting notion of mixing time can
therefore be overly conservative when the Markov chain is being used to approximate the expectation
of a fixed function, or expectations over some relatively limited class of functions. Accordingly, it
is of interest to consider the following function-specific discrepancy measure:
Definition 1 (f -discrepancy). For a given function f , the f -discrepancy is
df p, q = EX∼p f X − EY ∼q f Y .
(3)
The f -discrepancy leads naturally to a function-specific notion of mixing time:
Definition 2 (f -mixing time). For a given function f , the f -mixing time is
n
o
Tf δ = min n ∈ N | max df πn(i) , π ≤ δ .
(4)
i∈[d]
In the sequel, we also define function-specific notions of the spectral gap of a Markov chain, which
can be used to bound the f -mixing time and to obtain function-specific concentration inequalities.
2
1.2
Related work
Mixing times are a classical topic of study in Markov chain theory, and there is a large collection
of techniques for their analysis [see, e.g., 1, 6, 21, 23, 26, 30]. These tools and the results based on
them, however, generally apply only to worst-case mixing times. Outside of specific examples [5, 7],
relatively little is known about mixing with respect to individual functions or limited classes of
functions. Similar limitations exist in studies of concentration of measure and studies of confidence
intervals and other statistical functionals that depend on tail probability bounds. Existing bounds
are generally uniform, or non-adaptive, and the rates that are reported include a factor that encodes
the global mixing properties of the chain and does not adapt to the function [4, 11, 18, 19, 20, 22,
27, 29]. These factors, which do not appear in classic bounds for independent random variables, are
generally either some variant of the spectral gap γ of the transition matrix, or else a mixing time of
the chain T δ0 for some absolute constant δ0 > 0. For example, the main theorem from [20] shows
that for a function f : [d] → [0, 1] and a sample X0 ∼ π from the stationary distribution, we have
P
N
o
n
1 X
γ0
· 2 N ,
f Xn − µ ≥ ≤ 2 exp −
N
2 2 − γ0
n=1
(5)
where the eigenvalues of P are given in decreasing order as 1 > λ2 (P ) ≥ · · · ≥ λd (P ), and we denote
the spectral gap of P by
γ0 : = min 1 − λ2 (P ), 1 .
The requirement that the chain start in equilibrium can be relaxed by adding a correction for
the burn-in time [27]. Extensions of this and related bounds, including bounded-differences-type
inequalities and generalizations to continuous Markov chains and non-Markov mixing processes have
also appeared in the literature (e.g., [19, 29]).
The concentration result has an alternative formulation in terms of the mixing time instead of
the spectral gap [4]. This version and its variants are weaker, since the mixing time can be lower
bounded as
T δ ≥
1
1
1
1
− 1 log
≥
− 1 log
,
γ∗
2δ
γ0
2δ
(6)
where we denote the absolute spectral gap [21] by
γ∗ : = min 1 − λ2 , 1 − λd
≤ γ0 .
In terms of the minimum
probability πmin : = mini πi , the corresponding upper bound is an
1
extra factor of log πmin
larger, which potentially leads to a significant gap between γ10 and T δ0 ,
even for a moderate constant such as δ0 = 18 . Similar distinctions arise in our analysis, and we
elaborate on them at the appropriate junctures.
1.3
Organization of the paper
In the remainder of the paper, we elaborate on these ideas and apply them to MCMC. In Section 2,
we state some concentration guarantees based on function-specific mixing times, as well as some
spectrum-based bounds on f -mixing times, and the spectrum-based Hoeffding bounds they imply.
Section 3 is devoted to further development of these results in the context of several statistical
models. More specifically, in Section 3.1, we show how our concentration guarantees can be used
3
to derive confidence intervals that are superior to those based on uniform Hoeffding bounds and
CLT-type bounds, whereas in Section 3.2, we analyze the consequences for sequential testing. In
Section 4, we show that our mixing time and concentration bounds improve over the non-adaptive
bounds in real examples of MCMC from the literature. Finally, the bulk of our proofs are given in
Section 5, with some more technical aspects of the arguments deferred to the appendices.
2
Main results
We now present our main technical contributions, starting with a set of “master” Hoeffding bounds
with exponents given in terms of f -mixing times. As we explain in Section 2.3, these mixing time
bounds can be converted to spectral bounds bounding the f -mixing time in terms of the spectrum.
(We give some techniques for the latter in Section 2.2).
Recall that we use µ : = Eπ [f ] to denote the mean. Moreover, we follow standard conventions
in setting
λ∗ : = max λ2 (P ), λd (P ) , and λ0 : = max λ2 (P ), 0 .
so that the absolute spectral gap and the (truncated) spectral gap introduced earlier are given by
γ∗ : = 1 − λ∗ , and γ0 : = 1 − λ0 . In Section 2.2, we define and analyze corresponding functionspecific quantities, which we introduce as necessary.
2.1
Master Hoeffding bound
In this section, we present a master Hoeffding bound that provides concentration rates that depend
on the mixing properties of the chain only through the f -mixing time Tf . The only hypotheses on
burn-in time needed for the bounds to hold are that the chain has been run for at least N ≥ Tf
steps—basically, so that thinning is possible—and that the chain was started from a distribution
π0 whose f -discrepancy distance from π is small—so that the expectation of each f Xn iterate is
close to µ—even if its total-variation discrepancy from π is large. Note that the latter requirement
imposes only a very mild restriction, since it can always be satisfied by first running the chain for
a burn-in period of Tf steps and then beginning to record samples.
Theorem 1. Given any fixed > 0 such that df π0 , π ≤ 2 and N ≥ Tf 2 , we have
(
)
N
h1 X
i
2 N
.
P
f Xn ≥ µ + ≤ exp −
(7)
N
8T
f
2
n=1
Compared to the bounds in earlier work [e.g., 20], the bound (7) has several distinguishing
features. The primary difference is that the “effective” sample size
Neff : =
N
,
Tf (/2)
(8a)
is a function of f , which can lead to significantly sharper bounds on the deviations of the empirical
means than the earlier uniform bounds can deliver. Further, unlike the uniform results, we do not
require that the chain has reached equilibrium, or even approximate equilibrium, in a total variation
sense. Instead, the result applies provided that the chain has equilibrated only approximately, and
only with respect to f .
The reader might note that if one actually has access to a distribution π0 that is /2-close to π in
f -discrepancy, then an estimator of µ with tail bounds similar to those guaranteed by Theorem 1 can
4
be obtained as follows: first, draw N i.i.d. samples from π0 , and second, apply the usual Hoeffding
inequality for i.i.d. variables. However, it is essential to realize that Theorem 1 does not require
that such a π0 be available to the practitioner. Instead, the theorem statement is meant to apply
in the following way: suppose that—starting from any initial distribution—we run an algorithm for
N ≥ Tf (/2) steps, and then use the last of N − Tf (/2) samples to form an empirical average. Our
concentration result then holds with an effective sample size of
burnin
Neff
:=
N − Tf (/2)
N
=
− 1.
Tf (/2)
Tf (/2)
(8b)
In other words, the result can be applied with an arbitrary initial π0 , and accounting for burnin merely reduces the effective sample size by one. By contrast, such an interpretation does not
actually hold for the original result of [20]: it requires an initial sample X1 ∼ π, but such an exact
sample is not attainable after any finite burn-in period.
The appearance of the function-specific mixing time Tf in the bounds comes with both advantages and disadvantages. A notable disadvantage, shared with the mixing time versions of the
uniform bounds, is that
spectrum-based bounds on the mixing time (including our f -specific ones)
1
introduce a log πmin
term that can be a significant source of looseness. On the other hand, obtaining rates in terms of mixing times comes with the advantage that any bound on the mixing
time translates directly into a version of the concentration bound (with the mixing time replaced
−1
by its upper bound). Moreover, since the πmin
term is likely to be an artifact of the spectrum-based
approach, and possibly even just of the proof method, it may be possible to turn the mixing time
based bound into a stronger spectrum-based bound with a more sophisticated analysis. We go part
−1
term.
of the way toward doing this, albeit without completely removing the πmin
An analysis based on mixing time also has the virtue of better capturing the non-asymptotic
behavior of the rate. Indeed, as a consequence of the link (6) between mixing and spectral graph
(as well as matching upper bounds [21]), for any fixed function f , there exists a function-specific
spectral-gap γf > 0 such that
Tf
1
1
≈
log
+O 1 ,
2
γf
for 1.
(8c)
These asymptotics can be used to turn our aforementioned theorem into a variant of the results
of Léon and Perron [20], in which γ0 is replaced by a value γf that (under mild conditions) is at
least as large as γ0 . However, as we explore in Section 4, such an asymptotic spectrum-based view
loses a great deal of information needed to deal with practical cases, where often γf = γ0 and yet
Tf (δ) T (δ) even for very small values of δ > 0. For this reason, part of our work is devoted to
deriving more fine-grained concentration inequalities that capture this non-asymptotic behavior.
By combining our definition (8a) of the effective sample size Neff with the asymptotic expansion (8c), we arrive at an intuitive interpretation of Theorem 1: it dictates that the effective sample
γf N
size scales as Neff ≈ log(1/)
in terms of the function-specific gap γf and tolerance . This interpretation is backed by the Hoeffding bound derived in Corollary 1 and it is useful as a simple mental
model of these bounds. On the other hand, interpreting the theorem this way effectively plugs in
the asymptotic behavior of Tf and does not account for the non-asymptotic properties of the mixing
time; the latter may actually be more favorable and lead to substantially smaller effective sample
sizes than the naive asymptotic interpretation predicts. From this perspective, the master bound
has the advantage that any bound on Tf that takes advantage of favorable non-asymptotics translates directly into a stronger version of the Hoeffding bound. We investigate these issues empirically
in Section 4.
5
Based on the worst-case Markov Hoeffding bound (5), we might hope that the Tf ( 2 ) term in
Theorem 1 is spurious and removable using improved techniques. Unfortunately, it is fundamental.
This conclusion becomes less surprising if one notes that even if we start the chain in its stationary
distribution and run it for N < Tf () steps, it may still be the case that there is a large set Ω0 such
that for i ∈ Ω0 and 1 ≤ n ≤ N ,
|f (Xn ) − µ| a.s. if X0 = i.
(9)
This behavior is made possible by the fact that large positive and negative deviations associated
with different values in Ω0 can cancel out to ensure that E [f (Xn )] = µ marginally. However, the
lower bound (9) guarantees that
!
!
N
N
X
1 X
1 X
P
f (Xn ) ≥ µ + ≥
πi · P
f (Xn ) ≥ µ + | X0 = i
N
N
n=1
n=1
i∈Ω0
≥ π (Ω0 ) ,
so that if π (Ω0 ) 0, we have no hope of controlling the large deviation probability unless N &
Tf (). We make this intuitive argument precise in Section 2.5.
2.2
Bounds on f -mixing times
We generally do not have direct access either to the mixing time T δ or the f -mixing time Tf δ .
Fortunately, any bound on Tf translates directly into a variant of the tail bound (7). Accordingly,
this section is devoted to methods for bounding these quantities. Since mixing time bounds are
equivalent to bounds on dTV and df , we frame the results in terms of distances rather than times.
These results can then be inverted in order to obtain mixing-time bounds in applications.
The simplest bound is simply a uniform bound on total variation distance, which also yields a
bound on the f -discrepancy. In particular, if the chain is started with distribution π0 , then we have
1
dTV πn , π ≤ √
· λn∗ · dTV π0 , π .
πmin
(10)
In order to improve upon this bound, we need to develop function-specific notions of spectrum
and spectral gaps. The simplest way to do this is simply to consider the (left) eigenvectors to
which the function is not orthogonal and define a spectral gap restricted only to the corresponding
eigenvectors.
Definition 3 (f -eigenvalues and spectral gaps). For a function f : [d] → R, we define
n
o
Jf : = j ∈ [d] | λj 6= 1 and qjT f 6= 0 ,
(11a)
where qj denotes a left eigenvector associated with λj . Similarly, we define
λf = max λj ,
j∈Jf
and γf = 1 − λf .
(11b)
Using this notation, it is straightforward to show that if the chain is started with the distribution
π0 , then
s
Eπ f 2
· λnf · df π0 , π .
(12)
df πn , π ≤
πmin
6
This bound, though useful in many cases, is also rather brittle: it requires f to be exactly orthogonal
to the eigenfunctions of the transition matrix. For example, a function f0 with a good value of λf
can be perturbed by an arbitrarily small amount in a way that makes the resulting perturbed
function f1 have λf = λ∗ . More broadly, the bound is of little value for functions with a small but
nonzero inner product with the eigenfunctions corresponding to large eigenvalues (which is likely to
occur in practice; cf. Section 4), or in scenarios where f lacks symmetry (cf. the random function
example in Section 2.4).
In order to address these issues, we now derive a more fine-grained bound on df . The basic idea
is to split the lower f -spectrum Jf into a “bad” piece J, whose eigenvalues are close to 1 but whose
eigenvectors are approximately orthogonal to f , and a “good” piece Jf \ J, whose eigenvalues are
far from 1 and which therefore do not require control on the inner products of their eigenvectors
with f . More precisely, for a given set J ⊂ Jf , let us define
∆∗J : = 2 J × max khj k∞ × max qjT f ,
j∈J
j∈J
n
o
λ−J : = max λj | j ∈ Jf \ J .
λJ : = max
n
o
λj | j ∈ J ,
and
We obtain the following bound, expressed in terms of λ−J and λJ , which we generally expect to
obey the relation 1 − λ−J 1 − λJ .
Lemma 1 (Sharper f -discrepancy bound). Given f : [d] → [0, 1] and a subset J ⊂ Jf , we have
s
df πn , π ≤
∆∗J
λnJ
· dTV (π0 , π) +
Eπ f 2
· λn−J df (π0 , π).
πmin
(13)
The above bound, while easy to apply and comparatively easy to estimate, can be loose when
the first term is a poor estimate of the part of the discrepancy that comes from the J part of the
spectrum. We can get a still sharper estimate by instead making use of the following vector quantity
that more precisely summarizes the interactions between f and J:
X
hJ n : =
qjT f · λnj hj .
j∈J
This quantity leads to what we refer to as an oracle adaptive bound, because it uses the exact value
of the part of the discrepancy coming from the J eigenspaces, while using the same bound as above
for the part of the discrepancy coming from Jf \ J.
Lemma 2 (Oracle f -discrepancy bound). Given f : [d] → [0, 1] and a subset J ⊂ Jf , we have
s
df πn , π ≤
π0 − π
T
hJ n
+
Eπ f 2
· λn−J · df π0 , π .
πmin
(14)
We emphasize that, although Lemma 2 is stated in terms of the initial distribution π0 , when we
apply the bound in the real examples we consider, we replace all quantities that depend on π0 by
their worst cases values, in order to avoid dependence on initialization; this results in a khJ n k∞
term instead of the dot product in the lemma.
7
2.3
Concentration bounds
The mixing time bounds from Section 2.2 allow us to translate the master Hoeffding bound into
a weaker but more interpretable—and in some instances, more directly applicable—concentration
bound. The first result we prove along these lines applies meaningfully only to functions f whose
absolute f -spectral gap γf is larger than the absolute spectral gap γ∗ . It is a direct consequence
of the master Hoeffding bound and the simple spectral mixing bound (12), and it delivers the
asymptotics in N and promised in Section 2.1.
Corollary 1. Given any > 0 such that df π0 , π ≤ 2 and N ≥ Tf 2 , we have
N
exp − 2
1 X
8 log
f (Xn ) ≥ µ + ≤
P
2
N
n=1
exp − N
8
!
γf N
#
"
2
√
πmin
if ≤
2λ
√ f ,
πmin
otherwise.
Deriving a Hoeffding bound using the sharper f -mixing bound given in Lemma 1 requires more
care, both because of the added complexity of managing two terms in the bound and because one
of those terms does not decay, meaning that the bound only holds for sufficiently large deviations
> 0.
The following result represents one way of articulating the bound implied by Lemma 1; it leads
to improvements over the previous two results when the contribution from the bad part of the
spectrum J—that is, the part of the spectrum that brings γf closer to 1 than we would like—is
negligible at the scale of interest. Recall that Lemma 1 expresses the contribution of J via the
quantity ∆∗J .
Corollary
2. Given a triple of positive numbers (∆, ∆J , ∆∗J ) such that ∆J ≥ ∆∗J and N ≥ Tf ∆J +
∆ , we have
"
1
P
N
N
X
n=1
#
f Xn ≥ µ + 2 (∆J + ∆) ≤
exp −
exp −
2
∆J +∆
2
2
∆J +∆
2
1−λ−J N
log
N
1
√
∆ πmin
if ∆ >
if ∆ ≤
λ
√ −J ,
πmin
(15)
λ
√ −J .
πmin
Similar arguments can be applied to combine the master Hoeffding bounds with the oracle f mixing bound Lemma 2, but we omit the corresponding result for the sake of brevity. The proofs
for both aforementioned corollaries are in Section 5.2.
2.4
Example: Lazy random walk on C2d
In order to illustrate the mixing time and Hoeffding bounds from Section 2.2, we analyze their
predictions for various classes of functions on the 2d-cycle C2d , identified with the integers modulo
2d. In particular, consider the Markov chain corresponding to a lazy random walk on C2d ; it has
transition matrix
1
if v = u,
2
1 if v = u + 1 mod 2d,
Puv = 41
(16)
if
v
=
u
−
1
mod
2d,
4
0 otherwise.
8
It is easy to see that the chain is irreducible, aperiodic, and reversible, and its stationary
distribution is uniform. It can be shown [21] that its mixing time scales proportionally to d2 .
However, as we now show, several interesting classes of functions mix much faster, and in fact, a
“typical” function, meaning a randomly chosen one, mixes much faster than the naive mixing bound
would predict.
Parity function. The epitome of a rapidly mixing function is the parity function:
(
1 if u is odd,
fparity (u)) : =
(17)
0 otherwise.
It is easy to see that no matter what the choice of initial distribution π0 is, we have E fparity (X1 ) =
1
2 , and thus fparity mixes in a single step.
Periodic functions. A more
general class
of examples arises from considering the eigenfunctions
of P , which are given by gj u = cos πju
d ; [see, e.g., 21]. We define a class of functions of varying
regularity by setting
1 + gj
, for each j = 0, 1, . . . , d.
2
Here we have limited j to 0 ≤ j ≤ d because fj and f2d−j behave analogously. Note that the parity
function fparity corresponds to fd .
Intuitively, one might expect that some of these functions mix well before d2 steps have elapsed—
both because the vectors {fj , j 6= 1} are orthogonal to the non-top eigenvectors with eigenvalues
close to 1 and because as j gets larger, the periods of fj become smaller and smaller, meaning that
their global behavior can increasingly be well determined by looking at local snapshots, which can
be seen in few steps.
Our mixing bounds allow us to make this intuition precise, and our Hoeffding bounds
allow us
to prove correspondingly improved concentration bounds for the estimation of µ = Eπ fj = 1/2.
Indeed, we have
( π2 j 2
1 − cos πj
if j ≤ d2 ,
2
d
γ fj =
≥ 124d
(18)
2
if d2 < j ≤ d.
2
fj =
Consequently, equation (12) predicts that
(
d2
24 1
1
if j ≤ d2 ,
· j2
2 2 log 2d + log δ
π
Tfj δ ≤ T̃fj δ =
(19)
log 2d + 2 log 1δ
if d2 < j ≤ d,
where we have used the trivial bound Eπ f 2 ≤ 1 to simplify the inequalities. Note that this yields
an improvement over d2 for j & log d. Moreover, the bound (19) can itself be improved, since
each fj is orthogonal to all eigenfunctions other than 1 and gj , so that the log d factors can all be
removed by a more carefully
argued form of Lemma 1. It thus follows directly from the bound (18)
that if we draw N + T̃fj 2 samples, we obtain the tail bound
3d2
2 N
+Nb
exp
−
·
if j ≤ d2 ,
√
h 1 NX
i
2
2
π j
1
log 2 2d/
P
fj Xn ≥ + ≤
(20)
2 N
d
N0
2
√
exp −
2 < j ≤ d,
n=N
16 log 2 2d/
b
where the burn-in time is given by Nb = T̃fj /2 . Note again that the sharper analysis mentioned
above would allow us to remove the log 2d factors.
9
Random functions. A more interesting example comes from considering a randomly chosen
function f : C2d → [0, 1]. Indeed, suppose that the function values are sampled iid from some
distribution ν on [0, 1] whose mean µ∗ is 1/2:
iid
{f (u), u ∈ C2d } ∼ ν.
(21)
We can then show that for any fixed δ ∗ > 0, with high probability over the randomness of f , have
d log d log d + log 1δ
Tf (δ) .
,
for all δ ∈ (0, δ ∗ ].
(22)
δ2
√ d , this scaling is an improvement over the global mixing time of order d2 log(1/δ).
For δ log
d
The core idea behind the proof of equation (22) is to apply Lemma 1 with
s
s
(
)
d
d
Jδ : = j ∈ N ∩ [1, 2d − 1] | j ≤ 4δ
.
(23)
or j ≥ 2d − 4δ
log d
log d
It can be
q shown that khj k∞ = 1 for all 0 ≤ j < 2d and that with high probability over f ,
|qjT f | . logd d simultaneously for all j ∈ Jδ , which suffices to reduce the first part of the sharper
f -discrepancy bound to order δ.
In order to estimate the rate of concentration, we proceed as follows. Taking δ = c0 for a
suitably chosen universal constant c0 > 0, we show that ∆J : = 4 ≥ ∆∗J . We can then set ∆ = 4
and observe that with high probability over f , the deviation in Corollary 2 satisfies the bound
2
1
2 (∆J + ∆) ≤ . With δ as above, we have 1 − λ−J ≥ dclog
d for another universal constant c1 > 0.
Thus, if we are given N + Tf /2 samples for some N ≥ Tf 2 , then we have
(
)
N +Tf (/2)
4N
X
1
c
2
,
P
f (Xn ) ≥ µ + ≤ exp −
(24)
N
d log d log 4 + log 2d
n=Tf (/2)
for some c2 > 0. Consequently, it suffices for the sample size to be lower bounded by
d log d log 1/ + log d
,
N&
4
in order to achieve an estimation accuracy of . Notice that this requirement is an improvement
2
2
over the d2 from the uniform Hoeffding bound provided that ( logd d )1/2 . Proofs of all these
claims can be found in Appendix B.
2.5
Lower bounds
Let us now make precise the intuitive argument set forth at the end of Section 2.1. The
basic
idea is to start with an arbitrary candidate function δ : (0, 1) → (0, 1) such that Tf 2 in the
denominator of the function-specific Hoeffding bound (7) can be replaced by Tf (δ ()) and show
that if δ () ≥ , the replacement is not actually possible. We prove this fact by constructing a
Markov chain (which is independent of ) and a function (which depends on both and δ) such that
the Hoeffding bound is violated for the Markov chain-function pair for some value of N (which in
general depends on the chain and ).
As the following precise result shows, our lower bound continues to hold for an arbitrary constant
in the exponent of the Hoeffding bound, meaning that Theorem 1 is optimal up to constants. We
give the proof in Section 5.3.
10
Proposition 1. For every constant c1 > 0 and ∈ (0, 1), there exists a Markov chain Pc1 , a number
of steps N = N (c1 , ) and a function f = f such that
Pπ
!
N
1 X
1
c1 N 2
.
f (Xn ) −
≥ > 2 · exp −
N
2
Tf (δ())
(25)
n=1
3
Statistical applications
We now consider how our results apply to Markov chain Monte Carlo (MCMC) in various statistical
settings. Our investigation proceeds along three connected avenues. We begin by showing, in
Section 3.1, how our concentration bounds can be used to provide confidence intervals for stationary
expectations that avoid the over-optimism of pure CLT predictions without incurring the prohibitive
penalty of the Berry-Esseen correction—or the global mixing rate penalty associated with spectralgap-based confidence intervals. Then, in Section 3.2, we show how our results allow us to improve
on recent sequential hypothesis testing methodologies for MCMC, again replacing the dependence
on the spectral gap by a dependence on the f -mixing time. Later, in Section 4, we illustrate the
practical significance of function-specific mixing properties by using our framework to analyze three
real-world instances of MCMC, basing both the models and datasets chosen on real examples from
the literature.
3.1
Confidence intervals for posterior expectations
In many applications, a point estimate of Eπ f does not suffice; the uncertainty in the estimate must
be quantified, for instance by providing (1 − α) confidence intervals for some pre-specified constant
α. In this section, we discuss how improved concentration bounds can be used to obtain sharper
confidence intervals. In all cases, we assume the Markov chain is started from some distribution π0
that need not be the stationary distribution, meaning that the confidence intervals must account
for the burn-in time required to get close to equilibrium.
We first consider a bound that is an immediate consequence of the uniform Hoeffding bound
given by [20]. As one would expect, it gives contraction at the usual Hoeffding rate but with an
effective sample size of Neff ≈ γ0 (N − T0 ), where T0 is the tuneable burn-in parameter. Note that
this means that no matter how small Tf is compared to the global mixing time T , the effective size
incurs the penalty for a global burn-in and the effective sample size is determined by the global
spectral parameter γ0 . In order to make this precise, for a fixed burn-in level α0 ∈ (0, α), define
q
N (α, α0 ) : = 2 2 − γ0 ·
s
log 2/ α − α0
.
γ0 N − T α0
Then the uniform Markov Hoeffding bound [20, Thm. 1] implies that the set
N
X
1
unif
f Xn ± N α, α0
IN
α, α0 =
N − T α0 /2
(26a)
(26b)
n=T α0 /2 +1
is a 1 − α confidence interval. Full details of the proof are given in Appendix C.1.
Moreover, given that we have a family of confidence intervals—one for each choice
of α0 ∈
∗
(0, α)—we can obtain the sharpest confidence interval by computing the infimum N α : = inf N α, α0 .
0<α0 <α
11
Equation (26b) then implies that
h
unif
IN
α =
1
N − T α0
N
X
f (Xn ) ± ∗N (α)
i
n=T (α0 /2)+1
is a 1 − α confidence interval for µ.
We now consider one particular application of our Hoeffding bounds to confidence intervals, and
find that the resulting interval adapts to the function, both in terms of burn-in time required, which
now falls from a global mixing time to an f -specific mixing time, and in terms of rate, which falls
from γ10 to Tf (δ) for an appropriately chosen δ > 0. We first note that the one-sided tail bound of
Theorem 1 can be written as e−rN ()/8 , where
"
#
N
−1 .
rN () : = 2
(27)
Tf 2
If we wish for each tail to have probability mass that is at most α/2, we need to choose
> 0 so
2
that rN ≥ 8 log α , and conversely any such corresponds to a valid two-sided 1 − α confidence
interval. Let us summarize our conclusions:
−1
Theorem 2. For any width N ∈ rN
8 log 2/α , ∞ , the set
func
IN
:=
1
N − Tf
N
X
2
f Xn ± N
2
n=Tf
is a 1 − α confidence interval for the mean µ = Eπ f .
In order to make the result more amenable to interpretation, first note that for any 0 < η < 1,
we have
#
"
N
−1
valid for all ≥ η.
(28)
rN ≥ 2
Tf η2
|
{z
}
rN,η ()
Consequently, whenever rN,η (N ) ≥ 8 log α2 and N ≥ η, we are guaranteed that a symmetric interval
of half-width N is a valid 1 − α -confidence interval. Summarizing more precisely, we have:
Corollary 3. Fix η > 0 and let
N =
If N ≥ Tf
η
2 ,
−1
rN,η
√
2
8 log
=2 2
α
s
Tf
η
2
· log 2/α
.
N − Tf η2
func is a 1 − α confidence interval for µ = E f .
then IN
π
Often, we do not have direct access to Tf δ , but we can often obtain an upper bound T̃f δ
that is valid for all δ > 0. In Section 5.4, therefore, which contains the proofs for this section, we
prove a strengthened form of Theorem 2 and its corollary in that setting.
12
A popular alternative strategy for building confidence intervals using MCMC depends on the
Markov central limit theorem (e.g., [8, 17, 12, 28]). If the Markov CLT held exactly, it would lead
to appealingly simple confidence intervals of width
r
˜N = σf,asym
log(2/α)
,
N
PN
2
is the asymptotic variance of f .
where σf,asym
: = limN →∞ N1 VarX0 ∼π
n=1 f Xn
Unfortunately, the CLT does not hold exactly, even after the burn-in period. The amount by
which it fails to hold can be quantified using a Berry-Esseen
bound
for Markov chains, as we now
PN
discuss. Let us adopt the compact notation S̃N = n=1 f Xn − µ . We then have the bound [22]
P
S̃N
σf,asym
√
e−γ0 N
13
1
≤s −Φ s ≤ √
+
· √ ,
√
3 πmin σf,asym πmin γ0 N
N
(29)
where Φ is the standard normal CDF. Note that this bound accounts for both the non-stationarity
error and for non-normality error at stationarity. The former decays rapidly at the rate e−γ0 N ,
1
while the latter decays far more slowly, at the rate γ √
.
N
0
While the bound (29) makes it possible to prove a corrected CLT confidence interval, the resulting bound has two significant drawbacks. The first is that it only holds for extremely large
log 1/πmin
sample sizes, on the order of π 1 γ 2 , compared to the order
required by the uniform
γ0
min 0
Hoeffding bound. The second, shared by the uniform Hoeffding bound, is that it is non-adaptive
and therefore bottlenecked by the global mixing properties of the chain. For instance, if the sample
size is bounded below as
N ≥ max
1
1
6084
2
, 2 2
,
log √
2
γ0
πmin α
γ0 σf,asym πmin α
then both terms of equation (26b) are bounded by 1/6, and the confidence intervals take the form
BE
IN
N
1 X
=
f Xn ± σf,asym
N
n=1
s
2 log 6/α
.
N
(30)
See Appendix C.2 for the justification of this claim.
It is important to note that the width of this confidence interval involves a hidden form of
σ2
mixing penalty. Indeed, defining the variance σf2 = Varπ f X and ρf : = σ2 f , we can rewrite
f,asym
the width as
s
N = σf
2 log 6/α
.
ρf N
Thus, for this bound, the quantity ρf captures the penalty due to non-independence, playing the
role of γ0 and γf in the other bounds. In this sense, the CLT bound adapts to the function f , but
only when it applies, which is at a sample-size scale dictated by the global mixing properties of the
chain (i.e., γ0 ).
13
3.2
Sequential testing for MCMC
For some applications, full confidence intervals may be unnecessary; instead, a practitioner may
merely want to know whether µ = Eπ [f ] lies above or below some threshold 0 < r < 1. In these
cases, we would like to develop a procedure for distinguishing between the two possibilities, at a
given tolerable level 0 < α < 1 of combined Type I and II error. The simplest approach is, of
course, to choose N so large that the 1 − α confidence interval built from N MCMC samples lies
entirely on one side of r, but it may be possible to do better by using a sequential test. This latter
idea was recently investigated in Gyori and Paulin [15], and we consider the same problem settings
that they did:
(a) Testing with (known) indifference region, involving a choice between
H0 : µ ≥ r + δ
H1 : µ ≤ r − δ;
(b) Testing with no indifference region—that is, the same as above but with δ = 0.
For the first setting (a), we always assume 0 < δ < ν : = min(µ, 1 − µ), and the algorithm is
evaluated on its ability to correctly choose between H0 and H1 when one of them holds,
but it
incurs no penalty for either choice when µ falls in the indifference region r − δ, r + δ . The error
of a procedure A can thus be defined as
P A X1:∞ = H1
if µ ∈ H0 ,
err A, f = P A X1:∞ = H0
if µ ∈ H1 ,
0
otherwise.
The rest of this subsection is organized as follows. For the first setting (a), we analyze a
procedure Afixed that makes a decision after a fixed number N := N (α) of samples. We also
analyze a sequential procedure Aseq that chooses whether to reject at a sequence N0 , . . . , Nk , . . . of
decision times. For the second, more challenging, setting (b), we analyze Ahard , which also rejects
at a sequence of decision times. For both Aseq and Ahard , we calculate the expected stopping times
of the procedures.
As mentioned above, the simplest procedure Afixed would choose a fixed number N of samples to
be collected
P based onthe target level α. After collecting N samples, it forms the empirical average
µ̂N = N1 N
n=1 f Xn and outputs H0 if µ̂N ≥ r + δ, H1 if µ̂N ≤ r − δ, and outputs a special
indifference symbol, say I, otherwise.
The sequential algorithm Aseq makes decisions as to whether to output one of the hypotheses or
continue testing at a fixed sequence of decision times, say Nk . These times are defined recursively
by
1
1
N0 = M · min ,
,
r 1−r
k
Nk = N0 1 + ξ ,
(31)
(32)
where M > 0 and 0 < ξ < 2/5 are parameters of the algorithm. At each time Nk for k ≥ 1, the
algorithm Aseq checks if
µ̂Nk ∈ r −
M
M
,r +
.
Nk
Nk
14
(33)
If the empirical average lies in this interval, then the algorithm continues sampling; otherwise, it
outputs H0 or H1 accordingly in the natural way.
For the sequential algorithm Ahard , let N0 > 0 be chosen arbitrarily,1 and let Nk be defined in
terms of N0 as in (32). It once again decides at each Nk for k ≥ 1 whether to output an answer or
to continue sampling, depending on whether
µ̂Nk ∈ r − k α , r + k α .
When this inclusion holds, the algorithm continues; when it doesn’t hold, the algorithm outputs
H0 or H1 in the natural way. The following result is restricted to the stationary case; later in the
section, we turn to the question of burn-in.
Theorem 3. Assume that α ≤ 52 . For Afixed , Aseq , Ahard to all satisfy err A, f ≤ α, it suffices
to (respectively) choose
2Tf δ log α1
N=
,
(34)
δ2
8Tf 2δ log √2αξ
M=
, and
(35)
δ
n
log 1/α + 1 + 2 log k o
2
,
(36)
k α = inf > 0 :
≥
Nk
8Tf 2
where we let inf ∅ = ∞.
Our results differ from those of [15] because the latter implicitly control the worst-case error of
the algorithm
err A =
sup err A, f ,
f : Ω→[0, 1]
while our analysis controls err A, f directly. The corresponding choices made in [15] are
s
log( √2αξ )
log(1/α)
log(1/α) + 1 + 2 log k
N=
,M =
, and k (α) =
.
2
γ0 δ
γ0 δ
γ0 Nk
Hence, the Tf parameter in our bounds plays the same role that γ10 plays in their uniform bounds.
As a result of this close correspondence, we easily see that our results improve on the uniform result
for a fixed function f whenever it
to its stationary expectation faster than the chain itself
converges
1
converges— i.e., whenever Tf δ ≤ 2γ0 .
The value of the above tests depends substantially on their sample size requirements. In setting
(a), algorithm Aseq is only valuable if it reduces the number of samples needed compared to Afixed .
In setting (b), algorithm Ahard is valuable because of its ability to test between hypotheses separated
only by a point, but its utility is limited if it takes too long to run. Therefore, we now turn to the
question of bounding expected stopping times.
In order to o carry out the stopping time analysis, we introduce the true margin ∆ = |r − µ|.
First, let us introduce some useful notation. Let N (A) be the number of sampled collected by A.
Given a margin schedule k , let
∆
, and N0∗ 1:∞ : = N ∗
k0∗ 1:∞ : = min k ≥ 1 : k ≤
k0
2
1:∞
.
We can bound the expected stopping times of Aseq , Ahard in terms of ∆ as follows:
1
In Gyori and Paulin [15], the authors set N0 =
100
γ0
, but this is inessential.
15
Theorem 4. Assume either H0 or H1 holds. Then,
s
M
4 2Tf δ/2 M
E N (Aseq ) ≤ 1 + ξ
+
+ 8Tf δ/2 + 1 ;
(37)
∆
∆
∆
∗
32αTf ∆/4
E N (Ahard ) ≤ 1 + ξ N0 + 1 +
.
(38)
∆2
With minor modifications to the proofs in [15], we can bound the expected stopping times of
their procedures as
s
o
nM
2
M
4
+1 ;
E[N (Aseq )] ≤ (1 + ξ)
+
+
∆
∆ γ0 ∆ γ0
4α
E[N (Ahard )] ≤ (1 + ξ) N0∗ + 1 +
.
γ0 ∆2
In order to see how the uniform and adaptive bounds compare, it is helpful to first note that,
under either H0 or H1 , we have the lower bound ∆ ≥ δ. Thus, the dominant term in the expectations
in both cases is (1 + ξ)M/∆. Consequently, the ratio between the expected stopping times is
approximately equal to the ratio between the M values—viz.,
Madapt
≈ γ0 Tf δ/2 .
Munif
(39)
As a result, we should expect a significant improvement in terms of number
of samples when the
relaxation time γ10 is significantly larger than the f -mixing time Tf δ/2 . Framed in absolute terms,
we can write
√
√
Tf δ/2 log 2/ αξ
log 2/ αξ
N̄unif (Aseq ) ≈
and N̄adapt (Aseq ) ≈
.
γ0 δ∆
δ∆
Up to an additive term, the bound for Ahard is also qualitatively similar to earlier ones, with
replaced by ∆12 .
4
1
δ∆
Analyzing mixing in practice
We analyze several examples of MCMC-based Bayesian analysis from our theoretical perspective.
These examples demonstrate that convergence in discrepancy can in practice occur much faster
than suggested by naive mixing time bounds and that our bounds help narrow the gap between
theoretical predictions and observed behavior.
4.1
Bayesian logistic regression
Our first example is a Bayesian logistic regression problem introduced by Robert and Casella [28].
The data consists of 23 observations of temperatures (in Fahrenheit, but normalized by dividing by
100) and a corresponding binary outcome—failure (y = 1) or not (y = 0) of a certain component;
the aim is to fit a logistic regressor, with parameters α, β ∈ R2 , to the data, incorporating a prior
and integrating over the model uncertainty to obtain future predictions. More explicitly, following
the analysis in Gyori and Paulin [14], we consider the following model:
1
p α, β | b = · eα exp − eα /b
b
p y | α, β, x ∝ exp α + βx ,
16
which corresponds to an exponential prior on eα , an improper uniform prior on β and a logit link for
prediction. As in Gyori and Paulin [14], we target the posterior byrunning
a Metropolis-Hastings
4 0
algorithm with a Gaussian proposal with covariance matrix Σ =
. Unlike in their paper,
0 10
however, we discretize the state space to facilitate exact analysis of the transition matrix and to
make our theory directly applicable. The resulting state space is given by
Ω=
n
o
α̂ ± i · ∆, β̂ ± j · ∆ | 0 ≤ i, j ≤ 8 ,
where ∆ = 0.1 and (α̂, β̂) is the MLE. This space has d = 172 = 289 elements, resulting in a
289 × 289 transition matrix that can easily be diagonalized.
Robert and Casella [28] analyze the probability of failure when the temperature x is 65◦ F; it is
specified by the function
f65
exp α + 0.65β
.
α, β =
1 + exp α + 0.65β
Note that this function fluctuates significantly under the posterior, as shown in Figure 2.
We find that this function also happens to exhibit rapid mixing. The discrepancy df65 , before
entering an asymptotic regime in which it decays exponentially at a rate 1 − γ ∗ ≈ 0.386, first drops
from about 0.3 to about
to the predicted 10 iterations from
0.01 in just 2 iterations, compared
n
the naive bound df n ≤ dTV n ≤ √π1min · 1 − γ ∗ . Figure 3 demonstrates this on a log scale,
comparing the naive bound to a version of the bound in Lemmas 1 and 2. Note that the oracle
f -discrepancy bound improves significantly over the uniform
baseline, even though the non-oracle
version does not. In this calculation, we took J = 2, . . . , 140 to include the top half of the
spectrum excluding 1 and computed khj k∞ directly from P for j ∈ J and likewise for qjT f65 . The
oracle bound is given by Lemma 2. As shown in panel (b) of Figure 3, this decay is also faster than
that of the total variation distance.
An important point is that the quality of the f -discrepancy bound depends significantly on the
choice of J. In the limiting case where J includes the whole spectrum below the top eigenvalue,
the oracle bound becomes exact. Between that and J = ∅, the oracle bound becomes tighter and
tighter, with the rate of tightening depending on how much power the function has in the higher
versus lower eigenspaces. Figure 4 illustrates this for a few settings of J, showing that although
for this function and this chain, a comparatively large J is needed to get a tight bound, the oracle
bound is substantially tighter than the uniform and non-oracle f -discrepancy bounds even for small
J.
4.2
Bayesian analysis of clinical trials
The problem of missing data often necessitates Bayesian analysis, particularly in settings where
uncertainty quantification is important, as in clinical trials. We illustrate how our framework would
apply in this context by considering a clinical trials dataset [3, 14].
The dataset consists of n = 50 patients, some of whom participated in a trial for a drug and
exhibited early indicators (Yi ) of success/failure and final indicators (Xi ) of success/failure. Among
the 50 patients, both indicator values are available for nX = 20 patients; early indicators are
available for nY = 20 patients; and no indicators are available for n0 = 10 patients. The analysis
17
depends on the following parameterization:
P Xi = 1 | Yi = 0 = γ0 ,
P Xi = 1 | Yi = 1 = γ1 ,
P Xi = 1 | Yi missing = p.
Note that, in contrast to what one might expect, pis to be interpreted as the marginal probability
that Xi = 1, so that in actuality p = P Xi = 1 unconditionally; we keep the other notation,
however, for the sake of consistency with past work [3, 14]. Conjugate uniform (i.e., Be 1, 1 )
priors are placed on all the model parameters.
The unknown variables include the parameter triple γ0 , γ1 , p and the unobserved Xi values
30
for nY + n0 = 30 patients, and the full sample space is therefore Ω̃ = [0, 1]3 × 0, 1 . We cannot
estimate the transition matrix for this chain, even with a discretization with as coarse a mesh as
∆ = 0.1, since the number of states would be d = 103 ×230 ∼ 1012 . We therefore make two changes to
the original MCMC procedure. First, we collapse out the Xi variables to bring the state space down
to [0, 1]3 ; while analytically collapsing out the discrete variables is impossible, we can estimate the
transition probabilities for the collapsed chain analytically by sampling the Xi variables conditional
on the parameter values and forming a Monte Carlo estimate of the collapsed transition
probabilities.
Second, since the function of interest in the original work—namely, f γ0 , γ1 , p = 1 p > 0.5 —
depends only on p, we fix γ0 and γ1 to their MLE values and sample only p, restricted to the unit
interval discretized with mesh ∆ = 0.01.
As Figure 1 shows, eigenvalue decay occurs rapidly for this sampler, with γ ∗ ≈ 0.86. Mixing
thus occurs so quickly that none of the bounds—uniform or function-specific—get close to the truth,
due to the presence of the constant terms (and specifically the large term √π1min ≈ 2.14 × 1033 ).
Nonetheless, this example still illustrates how in actual fact, the choice of target function can make
a big difference in the number of iterations required for accurate estimation; indeed, if we consider
the two functions
f1 (p) : = 1 p > 0.5 ,
and f2 (p) : = p,
we see in Figure 5 that the mixing behavior differs significantly between them: whereas the discrepancy for the second decays at the asymptotic exponential rate from the outset, the discrepancy
for the first decreases faster (by about an order of magnitude) for the first few iterations, before
reaching the asymptotic rate dictated by the spectral gap.
4.3
Collapsed Gibbs sampling for mixture models
Due to the ubiquity of clustering problems in applied statistics and machine learning, Bayesian
inference for mixture models (and their generalizations) is a widespread application of MCMC
[10, 13, 16, 24, 25]. We consider the mixture-of-Gaussians model, applying it to a subset of the
schizophrenic reaction time data analyzed in Belin and Rubin [2]. The subset of the data we
consider consists of 10 measurements, with 5 coming from healthy subjects and 5 from subjects
diagnosed with schizophrenia. Since our interest is in contexts where uncertainty is high, we chose
the 5 subjects from the healthy group whose reaction times were greatest and the 5 subjects from
the schizophrenic group whose reaction times were smallest. We considered a mixture with K = 2
18
components, viz.:
µb ∼ N 0, ρ2 , b = 0, 1,
ω ∼ Be α0 , α1
Zi | ω ∼ Bern ω
Xi | Zi = b, µ ∼ N µb , σ 2 .
We chose relatively uninformative priors, setting α0 = α1 = 1 and ρ = 237. Increasing the value
chosen in the original analysis [2], we set σ ≈ 70; we found that this was necessary to prevent the
posterior from being too highly concentrated, which would be an unrealistic setting for MCMC. We
ran collapsed Gibbs on the indicator variables Zi by analytically integrating out ω and µ0:1 .
As Figure 1 illustrates, the spectral gap for this chain is small—namely, γ∗ ≈ 3.83 × 10−4 —yet
the eigenvalues fall off comparatively quickly after λ2 , opening up the possibility for improvement
over the uniform γ∗ -based bounds. In more detail, define
zb∗ : = b b b b b 1 − b 1 − b 1 − b 1 − b 1 − b ,
corresponding to the cluster assignments in which the patient and control groups are perfectly
separated (with the control group being assigned label b). We can then define the indicator for
exact recovery of the ground truth by
f (z) = 1 z ∈ z0∗ , z1∗ .
As Figure 6 illustrates, convergence in terms of f -discrepancy occurs much faster than convergence in total variation, meaning that predictions of required burn-in times and sample size based
on global metrics of convergence drastically overestimate the computational and statistical effort
required to estimate the expectation of f accurately using the collapsed Gibbs sampler. This behavior can be explained in terms of the interaction between the function f and the eigenspaces of P .
Although the pessimistic constants in the bounds from the uniform bound (10) and the non-oracle
function-specific bound (Lemma 1) make their predictions overly conservative, the oracle version
of the function-specific bound (Lemma
2) begins to make exact predictions after just a hundred
iterations when applied with J = 1, . . . , 25 ; this corresponds to making exact predictions of Tf δ
for δ ≤ δ0 ≈ 0.01, which is a realistic tolerance for estimation of µ. Panel (b) of Figure 6 documents
this by plotting the f -discrepancy oracle bound against the actual value of df on a log scale.
Bound type
Tf 0.01
Tf 10−6
Uniform
31,253
55,312
Function-Specific
25,374
49,434
Function-Specific (Oracle)
98
409
Actual
96
409
Table 1: Comparison
on Tf δ for different values of δ. The uniform bound corresponds
of bounds
to the bound Tf δ ≤ T δ , the latter of which can be bounded by the total variation bound. The
function-specific bounds correspond to Lemmas 1 and 2, respectively. Whereas the uniform and
non-oracle f -discrepancy bounds make highly conservative predictions, the oracle f -discrepancy
bound is nearly sharp even for δ as large as 0.01.
The mixture setting also provides a good illustration of how the function-specific Hoeffding
bounds can substantially improve on the uniform Hoeffding bound. In particular, let us compare
19
the Tf -based Hoeffding bound (Theorem 1) to the uniform Hoeffding bound established by Léon
and Perron [20]. At equilibrium, the penalty for non-independence in our bounds is (2Tf (/2))−1
compared to roughly γ∗−1 in the uniform bound. Importantly, however, our concentration bound
applies unchanged even when the chain has not equilibrated, provided it has approximately equilibrated with respect to f . As a consequence, our bound only requires a burn-in of Tf (/2), whereas
the uniform Hoeffding bound does not directly apply for any finite burn-in. Table 1 illustrates the
size of these burn-in times in practice. This issue can be addressed using the method of Paulin [27],
but at the cost of a burn-in dependent penalty dTV (T0 ) = supπ0 dTV (πn , π):
N
i
n
X
o
1
γ0
· 2 N − T0 ,
P
f (Xn ) ≥ µ + ≤ dTV T0 + exp −
N − T0
2 1 − γ0
n=T
h
(40)
0
where we have let T0 denote the burn-in time. Note that a matching bound holds for the lower
tail. For our
we computed the tightest version of the bound (40), optimizing T0 in
experiments,
the range 0, 105 for each value of the deviation . Even given this generosity toward the uniform
bound, the function-specific bound still outperforms it substantially, as Figure 7 shows.
For
the function-specific bound, we used the function-specific oracle bound (Lemma 2) to bound
Tf 2 ; this nearly coincides with the true value when ≈ 0.01 but deviates slightly for larger values
of .
5
Proofs of main results
This section is devoted to the proofs of the main results of this paper.
5.1
Proof of Theorem 1
We begin with the proof of the master Hoeffding bound from Theorem 1. At the heart of the
proof is the following bound on the moment-generating function (MGF) for the sum of an appropriately thinned subsequence of the function values {f (Xn )}∞
n=1 . In particular, let us introduce the
e
shorthand notation Xm,t : = X(m−1)Tf (/2)+t and N0 : = N/Tf ( 2 ). With this notation, we have the
following auxiliary result:
Lemma 3 (Master MGF bound). For any scalars β ∈ R, ∈ (0, 1), and integer t ∈ 0, Tf ( 2 ) , we
have
"
#
N0
X
1
1 2
e
E exp β
≤ exp
β + βµ + β · N0 .
(41)
f Xm,t
2
2
m=1
em,t , we have
See Section 5.1.1 for the proof of this claim. Recalling the definition of X
(/2) N0
h PN
i
n TfX
o
X
α n=1 f (Xn )
e
E e
= E exp α
f (Xm,t )
t=1
m=1
n
= E exp αTf (/2)
Tf (/2) N0
X X
o
1
em,t )
f (X
Tf (/2)
t=1
1
≤
Tf /2
Tf
m=1
#
/2 −1 "
N0
n
X
X
o
em,t
E exp αTf /2
f X
,
t=0
m=1
20
where the last inequality follows from Jensen’s inequality, as applied to the exponential function.
Applying Lemma 3 with β = αTf /2 , we conclude
n 1
o
PN
1
E eα n=1 f (Xn ) ≤ exp
+ αTf
µ + α2 Tf2
· N0 ,
αTf
2
2
2
2
2
valid for α > 0. By exponentiating and applying Markov’s inequality, it follows that
#
"
N
PN
1 X
f Xn ≥ µ + ≤ e−α(µ+) E eα n=1 f (Xn )
P
N
n=1
h
i
1
2 2
≤ exp
· − α Tf ( ) + α Tf ( ) N0 ( ) .
2
2
2
2
The proof of Theorem 1 follows by taking α =
2Tf ( 2 )
since
#
N
1 X
1 2 2
P
· −
+
f Xn ≥ µ + ≤ exp
· N0
N
2
2
4
n=1
2
N0
≤ exp −
8
(
)
2 N
.
= exp −
8Tf 2
"
5.1.1
Proof of Lemma 3
For the purposes of the proof, fix t and let Wm = X̃m,t . For convenience, also define a dummy
constant random variable W0 : = 0. Now, by assumption, we have
|E [f (W1 )] − µ| ≤
and |E [f (Wm+1 ) | Wm ] − µ| ≤ .
2
2
We therefore have the bound
"N
0
i
h P
Y
α m f (Wm )
eα[f (Wm )−E[f (Wm )
≤E
E e
#
| Wm−1 ]]
· eαµN0 +
αN0
2
.
(42)
m=1
But now observe that the random variables ∆m = f (Wm )−E [f (Wm ) | Wm−1 ] are deterministically
bounded in [−1, 1] and zero mean conditional on Wm−1 . Moreover, by the Markovian property,
this implies that the same is true conditional on W<m : = W0:(m−1) . It follows by standard MGF
bounds that
α2
E eα∆m | W<m ≤ e 2 .
Combining this bound with inequality (42), we conclude that
h P
i
αN0
α2
E eα m f (Wm ) ≤ e 2 ·N0 · eαµN0 + 2 ,
as claimed.
5.2
Proofs of Corollaries 1 and 2
In this section, we prove the derived Hoeffing bounds stated in Corollaries 1 and 2.
21
5.2.1
Proof of Corollary 1
The proof is a direct application of Theorem 1. Indeed, it suffices to note that if ≤
Tf
2
log √πmin
≤
2
log λ1f
=
log
2
+ 21 log
log λ1f
1
πmin
then
,
which yields the first bound. Turning to the second bound, note that if >
implies that Tf 2 = 1, which establishes the claim.
5.2.2
2λ
√ f ,
πmin
2λ
√ f ,
πmin
then equation (12)
Proof of Corollary 2
The proof involves combining Theorem 1 with Lemma 1, using the setting = 2 (∆
+
∆J ). We begin
2
by combining the bounds λJ ≤ 1, dTV (π0 , πn ) ≤ 1, df (π0 , πn ) ≤ 1, and Eπ f ≤ 1 with the
claim of Lemma 1 so as to find that
λn
λn
df (π0 , πn ) ≤ ∆∗J + √ −J ≤ ∆J + √ −J .
πmin
πmin
It follows that
Tf ( ) = Tf (∆J + ∆) ≤
2
log
1
∆
+ 12 log
1
log( λ−J )
1
πmin
whenever ∆ ≤
λ
√ −J .
πmin
Plugging into Theorem 1 now yields the first part of the bound. On the other hand, if ∆ >
then Lemma 1 implies that Tf (∆J + ∆) = 1, which proves the bound in the second case.
5.3
λ
√ −J ,
πmin
Proof of Proposition 1
In order to prove the lower bound in Proposition 1, we first require an auxiliary lemma:
Lemma 4. Fix a function δ : (0, 1) → (0, 1) with δ () > . For every constant c0 ≥ 1, there exists
a Markov chain Pc0 and a function f on it such that µ = 12 , yet, for N = c0 Tf (δ ()), and starting
the chain from stationarity,
!
N
1 X
1
1
Pπ
f (Xn ) −
≥ ≥ .
N
2
3
n=1
Using this lemma, let us now prove Proposition 1. Suppose that we make the choices
log 7
c0 : =
≥ 1, and Nc1 , : = c0 Tf () ,
c1 2
in Lemma 4. Letting Pc0 be the corresponding Markov chain and f the function guaranteed by the
lemma, we then have
Nc1 ,
X
c1 Nc1 , 2
1
1
1
2
Pπ
f (Xn ) −
≥ ≥
>
≥ 2 · exp −
.
Nc1 ,
2
3
7
Tf (δ())
n=1
22
Proof of Lemma 4: It only remains to prove Lemma 4, which we do by constructing pathological
function on a chain graph, and letting our Markov chain be the lazy random walk on this graph.
For the proof, fix > 0, let δ = δ () and let Tf = Tf (δ). Now choose an integer d > 0 such that
d > 2c0 and let the state space be Ω be the line graph with 2d elements with the standard lazy
random walk defining P . We then set
(
1
− δ 1 ≤ i ≤ d,
f (i) = 21
d + 1 ≤ i ≤ 2d.
2 +δ
It is then clear that Tf = 1.
Define the bad event
E=
d
3d
X1 ∈ 0,
∪
, 2d .
2
2
When this occurs, we have
0
N
1
1 X
f (Xn ) −
≥δ>
0
N
2
with probability one,
n=1
for all N 0 < d2 . Since N = c0 < d2 , we can set N 0 = N .
On the other hand, under π, the probability of E is ≥ 31 . (It is actually about 12 , but we want
to ignore edge cases.) The claim follows immediately.
5.4
Proofs of confidence interval results
Here we provide the proof of the confidence interval corresponding to our bound (Theorem 2).
Proofs of the claims (26b) and (30) can be found in Appendix C.
As discussed in Section 3.1, we actually prove a somewhat stronger form of Theorem 2, in order
to guarantee that the confidence interval can be straightforwardly built using an upper bound T̃f
on the f -mixing time rather than the true value. Setting T̃f = Tf recovers the original theorem.
Specifically, suppose T̃f : N → R+ is an upper bound on Tf and note that the corresponding tail
bound becomes e−r̃N ()/8 , where
"
#
N
r̃N () = 2
−1 .
T̃f 2
This means that, just as before we wanted to make the rate rN in equation (27) at least
as large
as 8 log α2 , we now wish to do the same with r̃N , which means choosing N with r̃N N ≥ 8 log α2 .
We therefore have the following result.
−1
Proposition 2. For any width N ∈ r̃N
8 log 2/α , ∞ , the set
func
IN
=
1
N − Tf
N
X
2
f Xn ± N
n=T̃f
is a 1 − α confidence interval for µ = Eπ f .
23
2
Proof. For notational economy, let us introduce the shorthands τf () = Tf 2 and τ̃f = T̃f
Theorem 1 then implies
N
X
N − τf 2
1
P
f Xn ≥ µ + ≤ exp −
·
N − τ̃f
4τf
2
.
n=τ̃f
N − τ̃f 2
·
4τ̃f
r̃N
.
= exp −
4
≤ exp −
Setting = N yields
P
1
N − τ̃f
N
X
f Xn ≥ µ + N ≤
n=τ̃f
α
.
2
The corresponding lower bound leads to an analogous bound on the lower tail.
As we did with Corollary 3, we can derive a more concrete, though slightly weaker, form of this
result that is more amenable to interpretation. We derive the corollary from the specialized bound
by setting T̃f = Tf .
To obtain this bound, define the following lower bound, in parallel with equation (28):
N
r̃N ≥ r̃N,η : = 2
η − 1 , ≥ η.
T̃f 2
Since this is a lower bound,
we see that whenever N ≥ η and r̃N,η N ≥ 8 log α2 , N is a valid
half-width for a 1 − α -confidence interval for the stationary mean centered at the empirical mean.
More formally, we have the following:
Proposition 3. Fix η > 0 and let
−1
N = r̃N,η
If N ≥ T̃f
η
2 ,
v
u
√ u T̃f η2 · log 2/α
2
t
8 log
=2 2
.
α
N − T̃f η2
func is a 1 − α confidence interval for µ = E f .
then IN
π
Proof. By assumption, we have
η ≤ N
This implies T̃f
N
2
≥ T̃f
η
2 ,
v
u
u T̃f
η = 2t
η
2
· log 2/α
.
N − T̃f η2
which yields
r̃N N = 2N
N
T̃f 2N
N
− 1 ≥ 2N
η − 1 = 8 log 2/α .
T̃f 2
But now Proposition 2 applies, so that we are done.
24
5.5
Proofs of sequential testing results
In this section, we collect various proofs associated with our analysis of the sequential testing
problem.
5.5.1
Proof of Theorem 3 for Afixed
We provide a detailed proof when H1 is true, in which case we have µ ≤ r − δ; the proof for the
other case is analogous. When H1 is true, we need to control the probability P Afixed X1:N = H0 .
In order to do so, note that Theorem 1 implies that
N
1 X
P Afixed X1:N = H0 = P
f Xn ≥ r + δ
N
≤P
1
N
n=1
N
X
n=1
≤ exp −
Setting N =
5.5.2
2Tf (δ) log
δ2
1
α
f Xn ≥ µ + 2δ
δ2N
.
2Tf δ
yields the bound P Afixed X1:N = H0 ≤ α, as claimed.
Proof of Theorem 3 for Aseq
The proof is nearly identical to that given by [15], with Tf δ/2 replacing
that H1 holds, so µ ≤ r − δ. In this case, it is certainly true that
1
γ0 .
We again assume
err Aseq , f = P ∃k : Aseq X1:Nk = H0
= P ∃k :
Nk
M
1 X
f Xn ≥ r +
Nk
Nk
n=1
≤
∞
X
k=1
It follows by Theorem 1, with k = δ +
M
Nk ,
Nk
1 X
M
P
f Xn ≥ r +
.
Nk
Nk
n=1
that
Nk
Nk
1 X
M
1 X
M
≤P
P
f Xn ≥ r +
f Xn ≥ µ + δ +
Nk
Nk
Nk
Nk
n=1
n=1
2k Nk
8Tf 2k
2 Nk
≤ exp − k δ .
8Tf 2
≤ exp −
√
In order to simplify notation, for the remainder of the proof, we define τ : = 4Tf (δ/2), β : =
25
αξ
2 ,
and ζk : =
δ 2 Nk
2τ log(1/β) .
In terms of this notation, we have M =
exp −
2τ log(1/β)
,
δ
and hence that
2k Nk
1
M 2
= exp −
· δ 2 Nk + 2δM +
2τ
2τ
Nk
δ 2 Nk
2τ log2 1/β
= exp −
+ log 1/β +
2τ
δ 2 Nk
−1
= exp − log 1/β 1 + ζk + ζk
= β · exp − log 1/β ζk + ζk−1 .
It follows that the error probability is at most
β
∞
X
exp − log 1/β ζk + ζk−1 .
k=1
We now finish the proof using two small technical lemmas, whose proofs we defer to Appendix D.
Lemma 5. In the above notation, we have
∞
∞
n
h
n
X
X
`
−` io
o
exp − log(1/β) 1 + ξ + 1 + ξ
.
exp − log(1/β) ζk + ζk−1 ≤ 4
`=0
k=1
Lemma 6. For any integer c ≥ 0, we have
(1 + ξ)` + (1 + ξ)−` ≥ 2(c + 1)
for all ` ∈
h
9c 9(c+1)
5ξ , 5ξ
.
9
, we find that the error is at
Using this bound, and grouping together terms in blocks of size 5ξ
most
∞
∞
X
`
−` 36 X
2 c+1
4
exp − log 1/β 1 + ξ + 1 + ξ
≤
·
β
.
5ξ
c=0
`=0
√
Since both α and ξ are at most 25 , we have β = 2αξ ≤ 15 , and hence the error probability is bounded
as
∞
36β 3
36β X 2 c+1
36β 2
3β 2
3α
≤
≤
≤
β
=
< α.
2
2
5ξ
2ξ
4
5ξ 1 − β
25ξ 1 − β
c=0
5.5.3
Proof of Theorem 3 for Ahard
We may assume that H1 holds, as the other case is analogous. Under H1 , letting k0 be the smallest
k such that k < ∞, we have
err(Ahard , f ) ≤
∞
X
P (µ̂Nk ≥ r + k ) ≤
k=k0
∞
X
P(µ̂Nk ≥ µ + 2k ).
k=k0
By Theorem 1, and the definition of k , we thus have
err (Ahard , f ) ≤
∞
X
k=k0
Nk 2k
exp −
8Tf 2k
!
≤
=
as claimed.
26
∞
α X 1
2
k2
k=k0
2
π
12
α < α,
5.5.4
Proof of Theorem 4 for Aseq
We may assume H1 holds; the other case is analogous. Note that
E[N ] ≤ N1 +
∞
X
Nk+1 − Nk P N > Nk
k=1
≤ N1 +
∞
X
k=1
≤ N1 +
∞
X
k=1
Nk
1 X
M
M
Nk+1 − Nk P
f Xn ∈ r −
, r+
Nk
Nk
Nk
n=1
Nk
M
1 X
f Xn > r −
Nk+1 − Nk P
Nk
Nk
∞
X
n=1
Nk
M
1 X
= N1 +
Nk+1 − Nk P
f Xn > µ + ∆ −
Nk
Nk
n=1
k=1
)
(
2
∞
X
∆Nk − M +
.
≤ N1 +
Nk+1 − Nk exp −
8T
δ/2
N
f
k
k=1
Our proof depends on the following simple technical lemma, whose proof we defer to Appendix D.3.
Lemma 7. Under the conditions of Theorem 4, we have
(
2 )
Z ∞
∞
X
∆Nk − M +
≤ 1+ξ 1+
h(s)ds ,
Nk+1 − Nk exp −
8Tf δ/2 Nk
N1
k=1
o
n
(∆s−M )2
where h(s) : = exp − 8Tf (δ/2)s+ .
(43)
Given this lemma, we then follow the argument of Gyori and Paulin [15] in order to bound the
integral. We have
s
Z ∞
4 2Tf δ/2 M
h s ds ≤
+ 8Tf δ/2 .
∆
∆
N1
1
1
To conclude, note that either r ≥ ∆ or 1 − r ≥ ∆, since 0 < µ < 1, so that min 1r , 1−r
≤ ∆
. It
follows that
(1 + ξ)M
N1 ≤ 1 + ξ N0 ≤
.
∆
Combining the bounds yields the desired result.
5.5.5
Proof of Theorem 4 for Ahard
For concreteness, we may assume H1 holds, as the H0 case is symmetric. We now have that
#
"
#
"
Nk
Nk
1 X
1 X
P [N ≥ Nk ] ≤ P
f Xn − r ≤ k ≤ P
f Xn ≥ µ + ∆ − k .
Nk
Nk
n=1
n=1
For convenience, let us introduce the shorthand
(
k
Tf ∆−
+
2
Tf,k : =
1
27
if k ≤ ∆,
otherwise.
Applying the Hoeffding bound from Theorem 1, we then find that
(
)
2
Nk
P [N ≥ Nk ] ≤ exp − + · ∆ − k + .
8Tf,k
Observe further that
E[N ] = N1 +
∞
X
Nk+1 − Nk P N > Nk
k=1
≤ Nk0∗ +1 +
∞
X
Nk+1 − Nk P N > Nk
k=k0∗ +1
≤ 1+ξ
N0∗
+1 +
∞
X
Nk+1 − Nk P N > Nk .
k=k0∗ +1
Combining the pieces yields
E[N ] ≤ 1 + ξ
N0∗
∞
X
+1 +
k=k0∗ +1
2
Nk
Nk+1 − Nk exp −
+ · ∆ − k + .
8Tf,k
(44)
The crux of the proof is a bound on the infinite sum, which we pull out as a lemma for clarity.
Lemma 8. The infinite sum (44) is upper bounded by
∞
X
k=k0∗ +1
∞
X
2
∆2
Nk
.
·
∆
−
≤
α
·
exp
−
m
·
Nk+1 − Nk exp −
k +
+
8Tf,k
32Tf ∆
4
m=1
See Appendix D.4 for the proof of this claim.
Lemma 8 then implies that
∞
X
k=k0∗ +1
Nk+1 − Nk exp −
Nk
Tf
·
∆
4
∞
X
∆2
∆2
≤α·
exp − m ·
∆
32
32T
f
4
m=1
∆2
α exp −
=
32Tf
1 − exp
32αTf
≤
∆2
∆
4
∆
4
∆2
32Tf ∆
4
.
The claim now follows from equation (44).
6
Discussion
A significant obstacle to successful application of statistical procedures based on Markov chains—
especially MCMC—is the possibility of slow mixing. The usual formulation of mixing is in terms
of convergence in a distribution-level metric, such as the total variation or Wasserstein distance.
On the other hand, algorithms like MCMC are often used to estimate equilibrium expectations
over a limited class of functions. For such uses, it is desirable to build a theory of mixing times
28
with respect to these limited classes of functions and to provide convergence and concentration
guarantees analogous to those available in the classical setting, and our paper has made some steps
in this direction.
In particular, we introduced the f -mixing time of a function, and showed that it can be characterized by the interaction between the function and the eigenspaces of the transition operator. Using
these tools, we proved that the empirical averages of a function f concentrate around their equilibrium values at a rate characterized by the f -mixing time; in so doing, we replaced the worst-case
dependence on the spectral gap of the chain, characteristic of previous Markov chain concentration
bounds, by an adaptive dependence on the properties of the actual target function. Our methodology yields sharper confidence intervals, as well as better rates for sequential hypothesis tests in
MCMC, and we have provided evidence that the predictions made by our theory are accurate in
some real examples of MCMC and thus potentially of significant practical relevance.
Our investigation also suggests a number of further questions. Two important ones concern the
continuous and non-reversible cases. Both arise frequently in statistical applications—for example,
when sampling continuous parameters or when performing Gibbs sampling with systematic scan—
and are therefore of considerable interest. As uniform Hoeffding bounds do exist for the continuous
case and, more recently, have been established for the non-reversible case, we believe many of our
conclusions should carry over to these settings, although somewhat different methods of analysis
may be required.
Furthermore, in practical applications, it would be desirable to have methods for estimating or
bounding the f -mixing time based on samples. It would also be interesting to study the f -mixing
times of Markov chains that arise in probability theory, statistical physics, and applied statistics
itself. While we have shown what can be done with spectral methods, the classical theory provides
a much larger arsenal of techniques, some of which may generalize to yield sharper f -mixing time
bounds. We leave these and other problems to future work.
Acknowledgments
The authors thank Alan Sly and Roberto Oliveira for helpful discussions about the lower bounds
and the sharp function-specific Hoeffding bounds (respectively). This work was partially supported
by NSF grant CIF-31712-23800, ONR-MURI grant DOD 002888, and AFOSR grant FA9550-14-10016. In addition, MR is supported by an NSF Graduate Research Fellowship and a Fannie and
John Hertz Foundation Google Fellowship.
References
[1] David Aldous and Persi Diaconis. Shuffling cards and stopping times. American Mathematical
Monthly, 93(5):333–348, 1986.
[2] Thomas R Belin and Donald B Rubin. The analysis of repeated-measures data on schizophrenic
reaction times using mixture models. Statistics in Medicine, 14(8):747–768, 1995.
[3] Scott M Berry, Bradley P Carlin, J Jack Lee, and Peter Muller. Bayesian Adaptive Methods
for Clinical Trials. CRC press, 2010.
[4] Kai-Min Chung, Henry Lam, Zhenming Liu, and Michael Mitzenmacher. Chernoff-hoeffding
bounds for markov chains: Generalized and simplified. In 29th International Symposium on
Theoretical Aspects of Computer Science, STACS 2012, February 29th - March 3rd, 2012,
29
Paris, France, pages 124–135, 2012. doi: 10.4230/LIPIcs.STACS.2012.124. URL http://dx.
doi.org/10.4230/LIPIcs.STACS.2012.124.
[5] Mark Conger and Divakar Viswanath. Riffle shuffles of decks with repeated cards. The Annals
of Probability, pages 804–819, 2006.
[6] Persi Diaconis and James Allen Fill. Strong stationary times via a new form of duality. Ann.
Probab., 18(4):1483–1522, 10 1990. doi: 10.1214/aop/1176990628. URL http://dx.doi.org/
10.1214/aop/1176990628.
[7] Persi Diaconis and Bob Hough. Random walk on unipotent matrix groups. arXiv preprint
arXiv:1512.06304, 2015.
[8] James M Flegal, Murali Haran, and Galin L Jones. Markov chain monte carlo: Can we trust
the third significant figure? Statistical Science, 23(2):250–260, 2008.
[9] Andrew Gelman, John B. Carlin, Hal S. Stern, David B. Dunson, Aki Vehtari, and Donald B.
Rubin. Bayesian Data Analysis, Third Edition. Chapman and Hall/CRC, 2013.
[10] Zoubin Ghahramani and Thomas L Griffiths. Infinite latent feature models and the indian
buffet process. In Advances in Neural Information Processing Systems, pages 475–482, 2005.
[11] David Gillman. A chernoff bound for random walks on expander graphs. SIAM Journal on
Computing, 27(4):1203–1220, 1998.
[12] Peter W Glynn and Eunji Lim. Asymptotic validity of batch means steady-state confidence
intervals. In Advancing the Frontiers of Simulation, pages 87–104. Springer, 2009.
[13] Thomas L Griffiths and Mark Steyvers. Finding scientific topics. Proceedings of the National
Academy of Sciences, 101(suppl 1):5228–5235, 2004.
[14] Benjamin M Gyori and Daniel Paulin. Non-asymptotic confidence intervals for mcmc in practice. arXiv preprint arXiv:1212.2016, 2012.
[15] Benjamin M. Gyori and Daniel Paulin. Hypothesis testing for markov chain monte carlo. Statistics and Computing, pages 1–12, July 2015. ISSN 0960-3174. doi: 10.1007/s11222-015-9594-1.
URL http://dx.doi.org/10.1007/s11222-015-9594-1.
[16] Sonia Jain, Radford M Neal, et al. Splitting and merging components of a nonconjugate
dirichlet process mixture model. Bayesian Analysis, 2(3):445–472, 2007.
[17] Galin L Jones and James P Hobert. Honest exploration of intractable probability distributions
via markov chain monte carlo. Statistical Science, 16(4):312–334, 2001.
[18] Aldéric Joulin, Yann Ollivier, et al. Curvature, concentration and error estimates for markov
chain monte carlo. The Annals of Probability, 38(6):2418–2442, 2010.
[19] Aryeh Kontorovich, Roi Weiss, et al. Uniform chernoff and dvoretzky-kiefer-wolfowitz-type
inequalities for markov chains and related processes. Journal of Applied Probability, 51(4):
1100–1113, 2014.
[20] Carlos A. Léon and François Perron. Optimal hoeffding bounds for discrete reversible markov
chains. Ann. Appl. Probab., 14(2):958–970, 05 2004. doi: 10.1214/105051604000000170.
30
[21] David A. Levin, Yuval Peres, and Elizabeth L. Wilmer. Markov Chains and Mixing Times.
American Mathematical Society, 2008.
[22] Pascal Lezaud. Chernoff and berry–esséen inequalities for markov processes. ESAIM: Probability and Statistics, 5:183–201, 2001.
[23] Sean P Meyn and Richard L Tweedie. Markov Chains and Stochastic Stability. Springer Science
& Business Media, 2012.
[24] David M. Mimno, Matthew D. Hoffman, and David M. Blei. Sparse stochastic inference for
latent dirichlet allocation. In Proceedings of the 29th International Conference on Machine
Learning, ICML 2012, Edinburgh, Scotland, UK, June 26 - July 1, 2012, 2012. URL http:
//icml.cc/discuss/2012/784.html.
[25] Radford M Neal. Markov chain sampling methods for dirichlet process mixture models. Journal
of computational and graphical statistics, 9(2):249–265, 2000.
[26] Yann Ollivier. Ricci curvature of markov chains on metric spaces. Journal of Functional
Analysis, 256(3):810–864, 2009.
[27] Daniel Paulin. Concentration inequalities for markov chains by marton couplings and spectral
methods. arXiv preprint arXiv:1212.2015, 2012.
[28] Christian P. Robert and George Casella. Monte Carlo Statistical Methods (Springer Texts in
Statistics). Springer-Verlag New York, Inc., Secaucus, NJ, USA, 2005. ISBN 0387212396.
[29] Paul-Marie Samson et al. Concentration of measure inequalities for Markov chains and φ-mixing
processes. The Annals of Probability, 28(1):416–461, 2000.
[30] Alistair Sinclair. Improved bounds for mixing rates of markov chains and multicommodity
flow. Combinatorics, Probability, and Computing, 1(04):351–370, 1992.
A
Proofs for Section 2.2
In this section, we gather the proofs of the mixing time bounds from Section 2.2, namely equations (10) and (12) and Lemmas 1 and 2.
A.1
Proof of the bound (10)
Recall that
df (p, q) =
Ep f X − Eq f (Y ) .
sup
f : [d]→[0, 1]
It follows from equation (12) that
dTV πn , π =
sup
df πn , π
f : [d]→[0, 1]
λnf
· df π0 , π
√
πmin
f : [d]→[0, 1]
1
=√
· λn∗ · dTV π0 , π ,
πmin
≤
sup
as claimed.
31
A.2
Proof of equation (12)
√
Let D = diag π . Then the matrix A = DP D−1 is symmetric and so has an eigendecomposition
P
of the form A = γ1 γ1T + dj=2 λj γj γjT . Using this decomposition, we have
T
P = 1π +
d
X
λj hj qjT ,
j=2
where hj : = D−1 γj and qj : = Dγj . Note that the vectors {qj }dj=2 correspond to the left eigenvectors
associated with the eigenvalues {λj }dj=2 .
Now, if we let π0 be an arbitrary distribution over [d], we have
df (πn , π) = π0T P n f − π T P n f ≤ (π0 − π)T P n f .
P
P
Defining Pf : = 1π T + j∈Jf λj hj qjT , we have P n f = Pfn f . Moreover, if we define P̃f : = j∈Jf λj hj qjT ,
T
T
and correspondingly Ãf : = DP̃f D−1 , we then have the relation π0 − π P̃f = π0 − π Pf . Consequently, by the definition of the operator norm and sub-multiplicativity, we have
T
df πn , π ≤ π0 − π P̃fn f
≤ |||Ãf |||nop kDf k2 D−1 π0 − π 2
v
u
u X π0,i − πi 2 n
· λf df π0 , π .
= tEπ f 2 ·
πi
i∈[d]
In order to complete the proof, let Z ∈ {0, 1}d denote the indicator vector Zj = 1 X0 = j .
Observe that the function
X zi − πi 2
r(z) : =
πi
i∈[d]
is convex in terms of z. Thus, Jensen’s inequality implies that
X π0,i − πi
Eπ0 r(Z) ≥ r Eπ0 Z = r π0 =
πi
2
.
i∈[d]
On the other hand, for any fixed value X0 = j, corresponding to Z = ej , we have
2
X
1 − πj
1 − πj
1
+
πi =
≤
.
r z = r ej =
πj
πj
πmin
i6=j
q
[f 2 ]
We deduce that df (πn , π) ≤ Eππmin
· λnf · df (π0 , π), as claimed.
A.3
Proof of Lemma 1
We observe that
(π0 − π)T hJ (n) ≤ kπ0 − πk1 khJ (n)k∞
= 2dTV π0 , π khJ (n)k∞
o
nX
n
≤ 2dTV π0 , π
λj · qjT f · khj k∞
j∈J
≤ 2dTV
o
n
π0 , π 2 J · max qjT f · max khj k∞ ,
j∈J
32
j∈J
as claimed.
A.4
Proof of Lemma 2
We proceed in a similar fashion as in the proof of equation (12). Begin with the identity proved
there, viz.
df πn , π = (π0 − π)T P̃fn f ,
P
where P̃f = j∈Jf λj hj qjT . Now decompose P̃f further into
X
X
λj hj qjT .
PJ =
λj hj qjT and P−J =
j∈J
j∈Jf \J
n . We thus find that
Note also that P̃fn = PJn + P−J
n
f .
df (πn , π) ≤ (π0 − π)T PJn f + (π0 − π)T P−J
Now observe that PJn f = hJ (n), so (π0 − π)T PJn f = (π0 − π)T hJ (n)|. On the other hand, the
second term can be bounded using the argument from the proof of equation (12) to obtain
s
Eπ f 2
n
(π0 − π)T P−J
f ≤
· λn−Jδ · df (π0 , π),
πmin
as claimed.
B
Proofs for Section 2.4
In this section, we provide detailed proofs of the bound (24), as well as the other claims about the
random function example on C2d .
Proposition 4. Let f : [d] → [0, 1] with f (i) ∼ τ iid from some distribution on [0, 1]. There exists
∗
a universal constant c0 > 0 such that with probability ≥ 1 − 128√δd log d over the randomness f , we
have
c0 d log d log 128d
δ
Tf δ ≤
for all 0 < δ ≤ δ ∗ .
δ2
Proof. We proceed by defining a “good event” Eδ , and then showing that the stated bound on Tf (δ)
holds conditioned on this event. The final step is to show that P[Eδ ] is suitably close to one, as
claimed.
The event Eδ is defined in terms of the interaction between f and the eigenspaces of P corresponding to eigenvalues close to 1. More precisely, denote the indices of these eigenvalues by
s
s
n
d
d o
Jδ : = j ∈ {1, . . . , 2d − 1} | j ≤ 4δ
or j ≥ 2d − 4δ
.
log d
log d
The good event Eδ occurs when f has small inner product with all the corresponding eigenfunctions—
that is
r
n
10 log d o
Eδ : = max |qjT f | ≤ 2
.
j∈Jδ
d
Viewed as family of events indexed by δ, these events form a decreasing sequence. (In particular,
the associated sequence of sets Jδ is increasing in δ, in that whenever δ ≤ δ ∗ , we are guaranteed
that Jδ ⊂ Jδ∗ .)
33
Establishing the bound conditionally on Eδ : We now exploit the spectral properties of the
transition matrix to bound Tf conditionally
on the event Eδ . Recall that the lazy random walk on
πj
1
C2d has eigenvalues λj = 2 1 + cos( d ) for j ∈ [d], with corresponding unit eigenvectors
1
1 ωj
vjT = √
2d
πij
· · · ωj2d−1 , ωj : = e d .
P
T
(See [21] for details.) We note that this diagonalization allows us to write P = 1π T + 2d−1
j=1 λj hj qj ,
√
√
vj
where hj = 2d · vj and qj = √2d
, where we have used the fact that diag π = √12d · I. Note that
khj k∞ = 1.
q
Combining Lemma 1 with the bounds λJδ ≤ 1, khj k∞ ≤ 1, and Jδ ≤ 8δ logd d , we find that
s
df (πn , π) ≤ 16δ
√
d
· max qjT f + d · λn−Jδ .
log d j∈J
Therefore, when the event Eδ holds, we have
√
√
df (πn , π) ≤ 32 10 · δ + d · λn−Jδ .
(45)
In order to conclude the argument, we use the fact that
λ−Jδ =
where j0 = 4δ
q
d
log d .
1 + maxj∈Jf \Jδ cos
πj
d
2
1 + cos
≤
2
πj0
d
,
On the other hand, we also have
π 2 x2
π 2 x2 π 4 x4
+
≤1−
,
cos πx ≤ 1 −
2
24
12
for all |x| ≤ 1,
which implies that
λ−Jδ ≤ 1 −
2π 2 δ 2
2π 2 δ 2
≤ exp −
.
3d log d
3d log d
Together with equation (45), this bound implies that for n ≥
whence
3d log d log
2π 2 δ 2
d
δ
, we have
√
df (πn , π) ≤ (32 10 + 1) δ ≤ 128δ.
Replacing δ by
δ
128
throughout, we conclude that for
3 (128)2 d log d log 128d
3 · 213 d log d log 128d
δ
δ
=
·
,
2π 2 δ 2
π2
δ2
we have df (πn , π) ≤ δ with probability at least P Eδ/128 .
n≥
34
√
dλn−Jδ ≤ δ,
Controlling the probability of Eδ : It now suffices to prove P(Eδ ) ≥ 1− √d δlog d , since this implies
that P Eδ/128 ≥ 1 − 128√δd log d , as required. In order to do so, observe that the vectors {qj }dj=1 are
rescaled versions of an orthonormal collection of eigenvectors, and hence
E qjT f = Eν [µ] · qjT 1 = 0.
We can write the inner product as qjT f = Aj + iBj , where (Aj , Bj ) are a pair of real numbers. The
triangle inequality then guarantees that |qjT f | ≤ |Aj | + |Bj |, so that it suffices to control these two
absolute values.
By definition, we have
Aj =
2d−1
πj`
1 X
f ` · cos
,
2d
d
`=0
2 = cos2 πj` ≤ 1.
showing that it is the sum of sub-Gaussian random variables with parameters σ`,j
d
1
. A parallel argument applies
Thus, the variable Aj is sub-Gaussian with parameter at most σj2 ≤ 2d
to the scalar Bj , showing that it is also sub-Gaussian with parameter at most σj2 .
By the triangle inequality, we have |qjT f | ≤ |Aj | + |Bj |, so it suffices to bound Aj and |Bj |
separately. In order to do so, we use sub-Gaussianity to obtain
s
r2
dr 2
d
P max |Aj | ≥ r ≤ |J| · e− 2 ≤ 8δ
· e− 2 .
j∈J
log d
With r : =
q
2 log 16d
,
d
we have
r
P max Aj ≥
j∈Jδ
2 log 16d
δ
≤ √
.
d
2 d log d
Applying a similar argument to Bj and taking a union bound, we find that
r
P max
j∈Jδ
Since 2
q
2 log 16d
d
qjT f
≥2
2 log 16d
δ
≤√
.
d
d log d
q
d
≤ 2 10 log
for d ≥ 2, we deduce that
d
r
1 − P (Eδ ) = P max qjT f ≥ 2
j∈Jδ
10 log d
d
!
≤√
δ
,
d log d
as required.
The concentration result now follows.
Proposition 5. The random function f on C2d defined in equation (21) satisfies the mixing time
and tail bounds
c0 d log d log d + log 12
Tf
≤
,
2
2
35
and
1
P
N
N +Tf (/2)
X
f Xn ≥ µ + ≤ exp −
n=Tf /2
c 4 N
1 1
.
d log d log( ) + log d
2
over the randomness of f provided ≥ c3
with probability at least 1 − √dc2log
d
c0 , c1 , c2 , c3 > 0 are universal constants.
log d 1/2
,
d
where
2
4δ
Proof. We first note that from the proof of Proposition 4, we have the lower bound 1 − λ−Jδ ≥ dclog
d,
√
∗
valid for all δ ∈ (0, 1). The proof of the previous proposition guarantees that ∆J ≤ 32 10δ, so setting
δ = 128√10 yields
√
= 32 10δ ≥ ∆∗J ,
4
and
1 − λ−Jδ ≥
c04 2
.
d log d
Now, by Proposition 4, there is a universal constant c5 > 0 such that, with probability at least
1 − 128√δd log d , we have
c5 d log d log d/δ 0
Tf δ 0 ≤
2
δ0
for all δ 0 ≥ δ.
In particular, we have
Tf
c02 d log d log d/
≤
2
2
with this same probability. Thus, we have this bound on Tf with the high probability claimed in
the statement of the proposition.
c04 2
We now finish by taking ∆ = 4 in Corollary 2. Noting that ∆J + ∆ = 2 and 1 − λ−J ≥ d log
d
completes the proof.
C
Proofs for Section 3.1
We now prove correctness of the confidence intervals based on the uniform Hoeffding bound (5),
and the Berry-Esseen bound (29).
C.1
Proof of claim (26b)
This claim follows directly from a modified uniform Hoeffding bound, due to [27]. In particular, for
any integer T0 ≥ 0, let dTV (T0 ) = supπ0 dTV (π0 P T0 , π) be the worst-case total variation distance
from stationarity after T0 steps. Using this notation, [27] shows that for any starting distribution
π0 and any bounded function f : [d] → [0, 1], we have
P
1
N − T0
N
X
f Xn − µ ≥ ≤ 2 exp −
n=T0 +1
36
γ0
· 2 N + 2dTV T0 .
2 2 − γ0
(46)
We now use the bound (46) to prove our claim (26b). Recall that we have chosen T0 so that
dTV T0 ≤ α0 /2. Therefore, the bound (46) implies that
N
n
o
X
1
γ0
· 2N N + α0
P
f (Xn ) − µ ≥ N ≤ 2 exp −
N − T0
2 2 − γ0
n=T0 +1
≤2·
α − α0
+ α0 = α,
2
as required.
C.2
Proof of the claim (30)
We now use the result (29) to prove the claim (30).
By the lower bound on N , we have
α
13
1
e−γ0 N
α
≤
and
· √ ≤ .
√
√
3 πmin
6
σf,asym πmin γ0 N
6
It follows from equation (29) that
"
#
N
X
√ α
1
P
f Xn ≥ µ + N ≤ Φ N N +
σf,asym N
3
n=1
N 2 α
≤ exp −
· +
2 N
3
α
= ,
2
and since a matching bound holds for the lower tail, we get the desired result.
D
Proofs for Section 5.5
In this section, we gather the proofs of Lemmas 5–8.
D.1
Proof of Lemma 5
Observe that the function
g(ζ) : = exp
n
o
− log(1/β) (ζ + ζ −1 )
is increasing on (0, 1] and decreasing on [1, ∞). Therefore, bringing ζ closer to 1 can only increase
the value of the function.
Now, for fixed k ≥ 1, define
min ` : 1 + ξ ` ≥ ζk
if ζk ≤ 1,
`
`k : =
max ` : 1 + ξ ≤ ζk
otherwise.
`
In words, the quantity `k is either the smallest integer such that 1 + ξ is bigger than ζk (if ζk ≤ 1)
`
or the largest integer such that 1 + ξ is smaller than ζk (if ζk ≥ 1).
37
With this definition, we see that 1+ξ
`
that g 1 + ξ k ≥ g ζk , and hence
∞
X
`k
always lies between ζk and 1, so that we are guaranteed
∞
X
g ζk ≤
g (1 + ξ)`k .
k=1
k=1
Thus, it suffices to show that at most two distinct values of k map to a single `k . Indeed, when this
mapping condition holds, we have
∞
X
∞
∞
X
X
`
`
g ζk ≤ 2
g 1+ξ
≤4
g 1+ξ .
k=1
`=−∞
`=0
In order to prove the stated mapping condition, note first that `k is clearly nondecreasing in k,
so that we need to prove that `k+2 > `k for all k ≥ 1. It is sufficent to show that ζk+2 ≥ 1 + ξ ζk ,
since this inequality implies that `k+2 ≥ `k + 1.
k
We now exploit the fact that ζk = ank for some absolute constant a, where nk = bn0 1+ξ c. For
k
this, let b = n0 1 + ξ , so that nk = bbc. Since nk+1 > nk , we have 1 + ξ b ≥ b 1 + ξ bc ≥ bbc + 1,
and hence
2
2
b 1 + ξ bc
1+ξ b−1
nk+2
=
≥
nk
bbc
bbc
1 + ξ bbc + 1 − 1
≥
bbc
≥ 1 + ξ,
as required.2
D.2
Proof of Lemma 6
When c = 0 and ` = 0, we note that the claim obviously holds with equality. On the other hand,
the left hand side is increasing in `, so that the c = 0 case follows immediately.
Turning to the case c > 0, we first note that it is equivalent to show that
(1 + ξ)2` − 2(c + 1)(1 + ξ)` + 1 ≥ 0
9c 9(c+1)
for all ` ∈ ( 5ξ
, 5ξ ).
It suffices toshow that (1 + ξ)` is at least as large as the largest root of the the quadratic equation
z 2 − 2 c + 1 z + 1 = 0. This largest root is given by
z∗ = c + 1 +
Consequently, it is enough to show that ` ≥
verify that
p
c (c + 2) ≤ 2(c + 1).
log 2(c+1)
log(1+ξ) .
Since
9c
5ξ
is a lower bound on `, we need to
9c
log 2 (c + 1)
≥
.
5ξ
log(1 + ξ)
2
We thank Daniel Paulin for suggesting this argument as an elaboration on the shorter proof in Gyori and Paulin
[15].
38
In order to verify this claim, note first that since ξ ≤ 25 , we have log(1 + ξ) ≥ ξ − 21 ξ 2 ≥ 45 ξ, whence
log 2 (c + 1)
5 log 2 (c + 1)
≤
.
log(1 + ξ)
4ξ
Differentiating the upper bound in c, we find that its derivative is
5
5
9
≤
≤ ,
4 (c + 1) ξ
8ξ
5ξ
so it actually suffices to verify the claim for c = 1, which can be done by checking numerically that
5 log 4
≤ 95 .
4
D.3
Proof of Lemma 7
Our strategy is to split the infinite sum into two parts: one corresponding to the range of s where
h is constant and equal to 1 and the other to the range of s where h is decreasing. In terms of the
Nk , these two parts are obtained by splitting the sum into terms with k < k0 and k ≥ k0 , where
k0 ≥ 1 is minimal such that M ≤ ∆Nk for k ≥ k0 .
For convenience in what follows, let us introduce the convenient shorthand
2
∆Nk − M +
Tk : = exp −
.
2τf δ/2 Nk
Now, if k0 = 1, we note that h must then be decreasing for s ≥ N1 , so that
∞
X
∞
Z
h s ds.
Nk+1 − Nk Tk ≤
N1
k=1
Otherwise, if k0 > 1, we have
∞
X
Z
∞
Nk+1 − Nk Tk ≤
h s ds.
Nk0
k=k0
For k < k0 , we have Tk = 1, so that when k < k0 − 1, we have
Nk+1 − Nk
2
∆Nk − M + Z Nk+1
=
h s ds.
exp −
2τf δ/2 Nk
Nk
Thus
kX
0 −1
Nk+1 − Nk
k=1
2
∆Nk − M + Z Nk0 −1
exp −
=
h s ds.
2τf δ/2 Nk
N1
Note that this implies
Z
∞
N1
2
∆s − M +
exp −
ds ≥ Nk0 −1 .
2τf δ/2 s
39
Finally, we observe that Nk+1 ≤ 1 + ξ Nk + 1 + ξ, so that Nk0 − Nk0 −1 ≤ ξNk0 −1 + 1 + ξ. Putting
together the pieces, we have
Nk0 − Nk0 −1
2
2
Z ∞
∆Nk0 −1 − M +
∆s − M +
exp −
exp −
≤1+ξ+ξ
ds,
2τf δ/2 Nk0 −1
2τf δ/2 s
N1
and hence
∞
X
Nk+1 − Nk
k=1
D.4
2
Z
∆Nk − M +
∞
≤1+ξ+ 1+ξ
h s ds.
exp −
2τf δ/2 Nk
N1
Proof of Lemma 8
+
∗
Observe that for k > k0∗ , we have ∆ − k ≥ ∆
2 . It follows that for k > k0 , we have Tf,k ≤ Tf
Thus, we can bound each term in the sum by
∆
4
.
2
Nk
∆2
Nk
Nk+1 − Nk exp −
·
.
·
∆
−
≤
N
−
N
exp
−
k
k+1
k
+
+
32
8Tf,k
Tf ∆
4
|
{z
}
Fk
Furthermore, the exponential in the definition of Fk is a decreasing function of Nk , so we further
bound the overall sum as
∞
X
k=k0∗ +1
Fk ≤
∞
X
exp − n ·
n=N0∗ +1
= exp −
∆2
32Tf ∆
4
∞
∆2 X
∆2
·
×
exp
−
m
·
32Tf ∆
32Tf ∆
4
4
m=1
N0∗
∞
∆2
∆ 2 X
·
.
×
exp − m ·
2
∆/2
32Tf ∆
4
m=1
N0∗
= exp −
8Tf
2
On the other hand, by the definition of N0∗ , k0∗ ≤
Tf
∆
2,
so
k0∗
∆/2
≤ Tf
.
2
2
By the definition of k , however, we know that
log 1/α
log 1/α + 1 + 2 log k
2k
≥
≥
,
Nk
Nk
8Tf 2k
which implies that (∆/2)2 N0∗ ≥ log 1/α 8Tf
(∆/2)
.
2
40
Re-arranging yields the claim.
1.0
0.6
Eigenvalue
Eigenvalue
0.8
0.4
0.2
0.0
0
50
100 150 200 250 300
Eigenvalue rank
(a)
1.0
0.8
0.6
0.4
0.2
0.0
0.2
0
20
40
60
80
Eigenvalue rank
100
(b)
1.0
Eigenvalue
0.8
0.6
0.4
0.2
0.0
0
200
400 600 800 1000
Eigenvalue rank
(c)
Figure 1: Spectra for three example chains: (a) Metropolis-Hastings for Bayesian logistic regression;
(b) collapsed Gibbs sampler for missing data imputation; and (c) collapsed Gibbs sampler for a
mixture model.
41
Posterior probability
0.035
0.025
0.015
0.005
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
P(y = 1|x = 0.65)
0
1.0
2
0.8
Discrepancy value
Log discrepancy
Figure 2: Distribution of f65 values under the posterior. Despite the discretization and truncation
to a square, it generally matches the one displayed in Figure 1.2 in Robert and Casella [28].
4
6
Uniform
Adaptive
Adaptive (Oracle)
Actual
8
10
0
5
10
15
Iterations
20
(a)
Function
Total variation
0.6
0.4
0.2
0.0
0 1 2 3 4 5 6 7 8 9
Iteration
(b)
Figure 3: (a) Discrepancies (plotted on log-scale) for f65 as a function of iteration number. The
prediction of the naive bound is highly pessimistic; the f -discrepancy bound goes part of the way
toward closing the gap and the oracle version of the f -discrepancy bound nearly completely closes
the gap in the limit and also gets much closer to the right answer for small iteration numbers. (b)
Comparison of the function discrepancy df65 and the total variation discrepancy dTV . They both
decay fairly quickly due to the large spectral gap, but the function discrepancy still falls much faster.
42
2
2
4
6
Uniform
Adaptive
Adaptive (Oracle)
Actual
8
10
0
5
10
15
Iterations
4
Uniform
Adaptive
Adaptive (Oracle)
Actual
8
10
0
5
10
15
Iterations
6
Uniform
Adaptive
Adaptive (Oracle)
Actual
8
0
5
20
(c)
10
15
Iterations
2
4
6
Uniform
Adaptive
Adaptive (Oracle)
Actual
8
10
0
5
10
15
Iterations
(d)
Figure 4: Comparisons of the uniform, non-oracle function-specific, and oracle function-specific
bounds for various choices of J. In each case, J = {2, . . . , Jmax }, with Jmax = 50 in panel (a),
Jmax = 100 in panel (b), Jmax = 200 in panel (c), and Jmax = 288 in panel (d). The oracle bound
becomes tight in the limit as Jmax goes to d = 289, but it offers an improvement over the uniform
bound across the board.
43
20
(b)
0
2
6
4
10
20
(a)
0
Log discrepancy
Log discrepancy
0
Log discrepancy
Log discrepancy
0
20
Change in log discrepancy
0.8
1.0
1.2
1.4
1.6
1.8
f(p) = 1[p > 0.5]
f(p) = p
1 2 3 4 5 6 7 8 9 10
Iteration
Figure 5: Change in log discrepancy for the two functions f (p) = 1 p ≥ 0.5 and f (p) = p considered
above. Whereas f (p) = p always changes at the constant rate dictated by the spectral gap, the
indicator discrepancy decays more quickly in the first few iterations.
1.0
Function
Total variation
Log discrepancy
Discrepancy
0.8
0.6
0.4
0.2
0.0
0
20
40 60
Iteration
80
100
(a)
0.0
0.5
1.0
1.5
2.0
2.5
3.0
3.5
Adaptive (Oracle)
Actual
0
50
100
150
Iteration
200
(b)
Figure 6: (a) Comparison of the f -discrepancy df and the total variation discrepancy dTV over
the first 100 iterations of MCMC. Clearly the function mixes much faster than the overall chain.
(b) The predicted value of log df (according to the f -discrepancy oracle bound—Lemma 2) plotted
against the true value. The predictions are close to sharp throughout and become sharp at around
100 iterations.
44
Log tail probability bound
0
2
4
6
8
10
12
14
16
18
Adaptive
Uniform
0.1
0.0
Deviation
0.1
Figure 7: Comparison of the (log) tail probability bounds provided by the uniform Hoeffding bound
due to [20] with one version of our function-specific Hoeffding bound (Theorem 1). Plots are based
on N = 106 iterations,
and choosing the optimal burn-in for the uniform bound and a fixed burn-in
−6
of 409 ≥ Tf 10
iterations for the function-specific bound. The function-specific bound improves
over the uniform bound by orders of magnitude.
45
| 10math.ST
|
ApproxDBN: Approximate Computing for
Discriminative Deep Belief Networks
Xiaojing Xu, Srinjoy Das, and Ken Kreutz-Delgado
arXiv:1704.03993v3 [cs.NE] 6 May 2017
Electrical and Computer Engineering, UC San Diego
Xiaojing Xu and Srinjoy Das have contributed equally to this work.
Abstract. Probabilistic generative neural networks are useful for many
applications, such as image classification, speech recognition and occlusion removal. However, the power budget for hardware implementations
of neural networks can be extremely tight. To address this challenge we
describe a design methodology for using approximate computing methods to implement Approximate Deep Belief Networks (ApproxDBNs) by
systematically exploring the use of (1) limited precision of variables; (2)
criticality analysis to identify the nodes in the network which can operate with such limited precision while allowing the network to maintain
target accuracy levels; and (3) a greedy search methodology with incremental retraining to determine the optimal reduction in precision to enable maximize power savings under user-specified accuracy constraints.
Experimental results show that significant bit-length reduction can be
achieved by our ApproxDBN with constrained accuracy loss.
Keywords: Generative Neural Networks; Approximate Computing; Deep
Belief Networks
1
Introduction
Neural networks provide high performance for applications such as image recognition, text retrieval and pattern completion [1][2]. Enabling such algorithms
to operate on low-power, realtime platforms such as mobile phones and Internet of Things (IoT) devices is an area of critical interest [3][4][5]. Fortunately,
because neural networks are inherently error-resilient [6], high accuracy of arithmetic representations and operations is not necessary to generate sufficiently
accurate performance of such algorithms. Researchers have explored the errorresilience properties of neural networks to achieve power efficiency, such as the
use of limited variable precision[7][8]. The majority of hardware neural network
implementations have been done for feedforward deterministic networks [4][7].
On the other hand, stochastic Deep Belief Networks (DBNs) have been used for
generating and recognizing images [9], video sequences [10], and motion-capture
data [11]. Here, we develop a design methodology that uses limited variable precision to implement a Discriminative DBN which performs classification tasks.
We are motivated, in part, by the work of Venkataramani et al. [7] who
rank neurons based on their impact on the overall performance of the network
and apply limited precision to “less critical” (or “resilient”) nodes, followed by a
retraining process to “heal the error” caused by approximation so that the power
2
Xiaojing Xu, Srinjoy Das, and Ken Kreutz-Delgado
benefit of approximation can be maximized with tolerable accuracy drop. In this
paper, we carry criticality analysis and retraining to the domain of stochastic,
generative neural networks. When applying approximation, two things to be
decided are the number of approximated neurons and the approximation degree,
which for us is the variable precision. Previous work [7][12] usually chooses a
fixed number of approximated neurons, however this choice is based on some
manually chosen threshold and could reduce opportunities for approximation.
We propose to use an iteratively changing number of approximated neurons and
an iteratively changing degree of approximation to fully exploit the potential of
approximation.
In this paper, we develop a methodology for designing energy efficient implementations of a Discriminative DBN (DDBN) that meets a required level of
discrimination performance. We particularly focus on the reduction of variable
precision (VP), which is synonymous with variable bit-length. The problem is to
find the bit-length distribution which results in lowest power within specified performance constraints. To find the distribution we propose a context dependent
(i.e., application domain data-dependent) design framework for the principled
approximation of the variables of a DDBN used for inference purposes. This is
done by first performing criticality analysis on a full-precision DDBN to provide a guideline for choosing the neurons to reduce VP and then, based on the
criticality analysis, we search in a greedy way for the lowest VPs satisfying the
given network accuracy constraint. With these VPs, we retrain the DDBN to
improve its performance, leaving room for further VP reduction. We repeat this
approximation-retraining process until the accuracy constraint is violated with
any further bit-length reduction. Summarizing, our contributions are:
• The use of criticality analysis on a stochastic, generative model such as DBN
to decide which neurons are less critical and are therefore more amenable to
approximation.
• The use of retraining on an approximated DBN to optimize network parameters with limited precision.
• The development of a greedy search methodology to determine the optimal bit-length distribution indicating minimum power designs within userspecified accuracy constraints.
2
2.1
Discriminative Deep Belief Networks
Restricted Boltzmann Machines
A Restricted Boltzmann Machine (RBM) [13] is a generative stochastic neural
network that can learn a probability distribution over a set of inputs (Fig. 1).
The conditional distributions are [1]:
P (hj |v) = σ(bj +
X
i
vi Wij )
and
P (vi |h) = σ(bvi +
X
hj Wij ) ,
(1)
j
where σ(x) is the sigmoid function. The training method is described in [14].
ApproxDBN: Approximate Computing for Discriminative DBNs
2.2
Discriminative Deep Belief Networks
v1
3
bv, W, b
h
1
v2
Deep Belief Networks (DBNs) are probabilistic generh2
ative models which learn a deep hierarchical representation of the training data. Reference [1] shows that
DBNs are just stacked RBMs and can be learned in a
greedy manner by sequentially learning RBMs. To use
hn
vm
a DBN as a classifier, one simple way is to replace the
hidden layer h
last RBM with a Discriminative RBM [9], created by visible layer v
concatenating h(L−1) and the “one-hot” class vector
c, as shown in Fig. 2. It is trained in the same way Fig. 1. RBM is a bipartite
as a DBN without class units, but with the additional graph, consisting of a visible layer vector v and a
constraints on c such that [2]
P (ci |h(L) ) = softmax(bci +
X
(L)
hj Wijc ) .
(2)
hidden layer vector h. bv
and b are biases for v and
h respectively, and W is
the weight between them.
P j
with softmax(xi ) = exp(xi )/ k exp(xk ). This architecture results in a Discriminative Deep Belief Network (DDBN). When used for inference, hidden units and classification units are
successively sampled using previous-layer-conditional probabilities. Sampling is
repeated and classification units averaged to get the final classification result.
classification
layer c
bx, W(1), b(1)
input layer x
hidden layer h(1)
b(L-2), W(L-1), b(L-1)
bc, Wc, b(L)
b(L-1), W(L), b(L)
hidden layer h(L-1)
hidden layer h(L)
Fig. 2. Discriminative DBN used for classification. It has a classification layer concatenated with the second last hidden layer to form a Discriminative RBM instead of the
last RBM. The dotted arrows show the direction of signal propagation in inference,
from input to the last hidden layer h(L) to the classification layer c
3
Criticality Analysis
Previous work implementing approximate computing frameworks for neural networks includes ApproxANN [7] and AxNN [12] where neuron criticality analysis
is implemented to identify neurons which are less critical to the overall accuracy
of the network and where approximate computing elements can be instantiated.
In these feedforward neural networks, the Euclidean distance between the output
of classification and the true label is used as the loss function during training
of such networks. During criticality analysis, the derivative of this loss function
is calculated with respect to the value of every neuron as that neuron’s error
contribution to classification error on the current sample. The magnitude of the
average error contributions over all training samples is then used to characterize
the criticality of the particular neuron. In our work we are concerned with the
4
Xiaojing Xu, Srinjoy Das, and Ken Kreutz-Delgado
stochastic generative model of the DDBN and only consider the criticality of
hidden neurons because in a DDBN these are the major source of computation
and power consumption.
To generalize the criticality analysis of feedforward deterministic neural network [7][12] to generative stochastic model DDBNs, there are two key steps: (i)
choose a relevant loss function expressible as a function of the values of hidden
neurons, and (ii) create a surrogate stochastic derivative since we don’t have the
derivative of a continuous and deterministic loss function as in feedforward deterministic networks. This “derivative” is used to determine the error contribution
of the hidden neurons for the specific sample.
x
W(1), b(1) z(1)=W(1)*x+b(1) a(1)=σ(z(1)) h(1)=１(a(1)>rand)
hidden
layer 1
Input
layer
h(L)
zc=Wc’*h(L)+bc ac=softmax(zc) c=１(ac>rand)
W c, b c
classification
layer
hidden
layer L
Fig. 3. DDBN inference process using L hidden layers
Our choice of loss function is Loss = |t − P (c)|2 , where t is the one-hot true
label vector of the sample. It should be noted that this loss function is not the
same one used in the DDBN training, but it properly interprets classification accuracy which is used as the metric evaluating the performance of Discriminative
DBN inference. To calculate the derivatives, we need the values of hidden neurons to be continuous, therefore we use the probabilities a(l) instead of binary
states h(l) in Fig. 3 for the values of hidden neurons in criticality analysis. With
the notations illustrated in Fig. 3, P (c) is the same as ac , and the derivative is
∂ Loss
(L)
∂hi
=
∂ Loss ∂acj ∂zlc
∂acj ∂zlc ∂h(L)
i
XX
j
l
!
=
X c
X c
(aj − tj )
(al=j − acj acl )Wilc .
j
(3)
l
(L)
This is the contribution of hi to the accuracy. To obtain the contribution
(l)
of hi with l < L, we apply the chain rule:
∂ Loss
(l)
=
X ∂ Loss ∂h(l+1)
j
∂hi
(l+1)
j
∂hj
(l)
,
(4)
∂hi
where the second factor can be derived as follows:
(l+1)
∂hj
(l)
∂hi
(l)
=
∂σ(hj )
(l)
∂hi
= 1i=j (hj
(l+1)
(l+1) 2
− (hj
) ).
(5)
To determine the criticality of a hidden unit hi from the above equations, we
take the magnitude of the average contribution evaluated on all training images.
ApproxDBN: Approximate Computing for Discriminative DBNs
4
5
Approximate Deep Belief Networks Design Approach
4.1 Limited Variable Precision (VP)
Fixed-point is preferred to floating-point representation because the latter requires extensive area and power for digital hardware implementations [15]. We
use Qm.n to denote the fixed-point number format with the most significant bit
being the mth bit to the left of the decimal point(including sign bit), and n bits
fractional part. The sum m + n is the total bit-length. Our experiments show
that replacing floating-point 64-bit by fixed-point 64-bit representations (Q0.64
for activation function outputs and Q8.56 for others) doesn’t affect the network
performance, so we use 64-bit fixed-point VP as our implementation baseline.
When working on further bit-length reduction, it is usual to just change the
fractional bit length in limited bit length designs and leave the integer bit length
intact [3], we do the same in our limited VP design by reducing n in Qm.n;
4.2 Accuracy Improvement with Retraining
The original DDBN is trained with floating-point 64-bit full precision and therefore its parameters such as weights and biases may not be optimal after the
approximations based on limited precision are introduced. Thus it is necessary
to perform retraining of the Approximated DDBN using limited precision to
reduce the impact of approximation on classification performance. The initial
values of parameters in retraining are set to the previous trained network parameter values after approximation and after each round of updates in retraining,
the parameters are represented by limited precision.
START
Train full-precision DDBN
A (Accuracy) > Amin (Accuracy Constraint)
Represent h, c by 64-bit fixed-point
A>Amin?
P1
Y
Uniformly reduce bit-length of h, c by ∆b
N
Uniformly increase bit-length of h, c by ∆b
Calculate criticalities of hidden units; let the number of
resilient neurons n = the number of hidden neurons
Level 1
Reduce bit-length of resilient h by ∆b
Level 3
A>Amin?
Y
N
Increase bit-length of resilient h by ∆b
P2
Level 2
n=n-∆n
Y
n>0?
N
Y
Bit-length change
since last (re)training?
Retrain with limit-precision h, c
N
END
Resilient
W (64)
W (64/60)
Fig. 4. ApproxDBN
Design Approach.Wh,(64/60)
c here stand
for variables feeding
into hidden
neuron
and classification layers
W (64/60/56)
Resilien
neuron
Resilien
neuron
97%
Bit length reduction
based on criticality
92%
Incremental
retrain
94%
Bit length reduction
based on criticality
Bit length reduction(64->60->56->...); power saving; accuracy stays above the 90% constraint
90%
6
4.3
Xiaojing Xu, Srinjoy Das, and Ken Kreutz-Delgado
Approximate DBN Design Approach
Our approach to designing an ApproxDBN with multiple limited VPs based
on a full-precision trained DDBN is illustrated in Fig. 4. It finds the optimal
VP configuration under some user-specified accuracy constraints in a greedy
manner. There are two phases of variable bit-length reduction. In the first phase
(P1 in Fig. 4), we apply uniform bit-length reduction for all variables until
the accuracy drops below the constraint. In the second phase (P2), we try to
further reduce the precisions of some variables in the hidden layers. The basic
idea is to first use criticality analysis to decide the criticality order of all hidden
neurons, and then reduce the precisions of the less critical neurons in a greedy
manner until the accuracy constraint is violated. Then retraining is performed
with limited VPs in-place to improve the accuracy. The approximation-retraining
process is repeated until no bit-length reduction can be done with the accuracy
constraint anymore. Specifically, maximum bit-length reduction is obtained with
three loops. The outermost loop (Level 1 in Fig. 4) calculates the criticality
order of hidden neurons with the input DDBN, and after approximation, retrains
the DDBN before the next iteration. The middle loop (Level 2) decides which
neurons to apply bit-length reduction based on a iteratively changing threshold
to allow less and less neurons reduce precision. The innermost loop (Level 3)
is where approximation is done. It iterates to reduce bit-lengths of the neurons
chosen by Level 2 until the bit-length drops to 0 or the accuracy drops below the
constraint. The program terminates when no further bit-length change occurs.
5
5.1
Experimental Results
Experimental Setup
We use a Discriminative DBN to perform classification on the MNIST dataset,
which contains 60,000 training samples and 10,000 test samples of 28 x 28 grayscale handwritten digits. The image data are binarized using a threshold of 0.5.
Therefore the DDBN structure in our analysis is composed of 784 input units
and 10 class units. Baseline DDBN has a floating-point 64-bit variable precision.
5.2
Benefits of Criticality Analysis
Fig. 5. Accuracy relative to baseline vs. the number of approximated neurons with
random approximating order and criticality analysis based approximating order for 4
different DDBN structures having one layer through four layers. a-b-c means there are
three hidden layers with respectively a, b and c neurons in each layer, etc. The baseline
accuracies for the 4 structures are 93.41%, 93.51%, 94.24% and 95.07% respectively
We show the benefits of criticality analysis by exploring four different DDBN
structures with the MNIST dataset (see Fig. 5). For each DDBN, we set the
ApproxDBN: Approximate Computing for Discriminative DBNs
7
baseline bit-length to Q8.8 and approximate different numbers (0 to all) of hidden
neurons by setting their bit-length to 0. As seen in Fig. 5 approximating less
critical neurons results in higher accuracies compared to approximating the same
numbers of neurons randomly for DDBNs with more than 1 layer.
5.3
ApproxDBN Design Flow
P1
P2
Bit-length reduction
Retraining
Fig. 6. The relative accuracy and sum of bit-length change while obtaining ApproxDBN with a 300-200-100 DDBN and a maximum accuracy loss of 10%.
We demonstrate the change in accuracy and the sum of hidden neurons
bit-length during the ApproxDBN design flow as the network goes through successive iterations of criticality determination and retraining (Fig. 6). For each
iteration in Phase 2, approximation is first applied to existing relatively high
precision neurons as a result of which classification accuracy is reduced. This is
followed by retraining which results in adjustment of network parameters to the
updated precision and this consequently results in performance improvement.
This process is repeated until at the end of the 6th iteration the accuracy can’t
stay above 90%, the target accuracy anymore with any approximation. The sum
of bit-length keeps decreasing during the process, and the design methodology
operates in a greedy way to make sure no variable will have bit-length increase
in the process so that the overall power consumption will be reduced.
5.4 Benefits of Using Criticality Analysis and Retraining
Using a 300-200-100 DDBN, Fig. 7 shows that both criticality analysis and
retraining help ApproxDBN obtain a distribution of more lower bit-length variables. Criticality based selection of approximated neurons performs much better
than random based selection in reduction of VPs. Retraining is important if
more accuracy loss has been allowed in the bit-length reduction stages of Fig. 6
as then there will be more room for accuracy improvement with retraining.
6
Conclusion
An approximate computing framework with constrained accuracy loss has been
developed for Discriminative Deep Belief Networks (DDBNs). This framework
makes use of limited variable precisions, and systematically identifies and approximates neurons that least affect network performance via application of
criticality analysis and retraining. Experimental results on DDBNs using the
8
Xiaojing Xu, Srinjoy Das, and Ken Kreutz-Delgado
Fig. 7. Bit-length distribution for (1) ApproxDBN, the approximate DDBN using
our design methodology, (2) ApproxDBN without criticality analysis, (3) ApproxDBN
without retraining, and (4) ApproxDBN without both criticality analysis and retraining, under different relative accuracy loss constraints
MNIST dataset show a significant reduction of bit-length with little accuracy
loss. Future work involves extending such approximate computing techniques to
more general network architectures utilizing approximate multipliers and performing generative tasks such as pattern completion and denoising.
References
1. G. Hinton and R. Salakhutdinov. Reducing the dimensionality of data with neural
networks. Science, 313(5786):504–507, 2006.
2. H. Larochelle, M. Mandel, R. Pascanu, and Y. Bengio. Learning algorithms for
the classification restricted boltzmann machine. JMLR, 13:643–669, 2012.
3. A. Yasoubi, R. Hojabr, and M. Modarressi. Power-efficient accelerator design for
neural networks using computation reuse. IEEE Comp. Arch. Letters, 2016.
4. Y.-H. Chen, T. Krishna, J. Emer, and V. Sze. Eyeriss: An energy-efficient reconfigurable accelerator for deep convolutional neural networks. IEEE JSSC, 2016.
5. S. Das, B. Pedroni, P. Merolla, et al. Gibbs sampling with low-power spiking digital
neurons. In IEEE ISCAS, pages 2704–2707, 2015.
6. Christopher M Bishop. Neural networks for pattern recognition. 1995.
7. S. Venkataramani, A. Ranjan, K. Roy, and A. Raghunathan. Axnn: energy-efficient
neuromorphic systems using approximate computing. In Proc. Int. Symposium on
Low power electronics & design, pages 27–32, 2014.
8. S. Gupta, A. Agrawal, K. Gopalakrishnan, and P. Narayanan. Deep learning with
limited numerical precision. arXiv preprint arXiv:1502.02551, 2015.
9. G. Hinton, S. Osindero, and Y.-W. Teh. A fast learning algorithm for deep belief
nets. Neural computation, 18(7):1527–1554, 2006.
10. I. Sutskever and G. Hinton. Learning multilevel distributed representations for
high-dimensional sequences. In AISTATS, volume 2, pages 548–555, 2007.
11. G. Taylor, G. Hinton, and S. Roweis. Modeling human motion using binary latent
variables. Advances in neural information processing systems, 19:1345, 2007.
12. Q. Zhang, T. Wang, Tian, et al. ApproxANN: an approximate computing framework for artificial neural network. In DATECE, pages 701–706, 2015.
13. J. Hopfield. Neural networks and physical systems with emergent collective computational abilities. PNAS, 79(8):2554–2558, 1982.
14. G. Hinton. A practical guide to training restricted boltzmann machines. Momentum, 9:926, 2010.
15. D. Ly and P. Chow. A high-performance fpga architecture for restricted boltzmann
machines. In ACM/SIGDA ISFPGA, pages 73–82, 2009.
| 9cs.NE
|
arXiv:1608.03090v1 [cs.SY] 10 Aug 2016
Regression Models for Output Prediction of
Thermal Dynamics in Buildings∗
Georgios C. Chasparis†
Software Competence Center Hagenberg GmbH
Department of Data Analysis Systems
Softwarepark 21
4232 Hagenberg, Austria
Email: [email protected]
Thomas Natschlaeger
Software Competence Center Hagenberg GmbH
Department of Data Analysis Systems
Softwarepark 21
4232 Hagenberg, Austria
Email: [email protected]
Standard (black-box) regression models may not necessarily
suffice for accurate identification and prediction of thermal
dynamics in buildings. This is particularly apparent when
either the flow rate or the inlet temperature of the thermal
medium varies significantly with time. To this end, this paper analytically derives, using physical insight, and investigates linear regression models with nonlinear regressors for
system identification and prediction of thermal dynamics in
buildings. Comparison is performed with standard linear regression models with respect to both a) identification error,
and b) prediction performance within a model-predictivecontrol implementation for climate control in a residential
building. The implementation is performed through the EnergyPlus building simulator and demonstrates that a careful
consideration of the nonlinear effects may provide significant
benefits with respect to the power consumption.
1
Introduction
The increased demand for electricity power and/or fuel
consumption in residential buildings requires an efficient
control design for all heating/cooling equipment. To this end,
recently, there have been several efforts towards a more efficient climate control in residential buildings [2, 3, 4, 5]. Ef-
∗ This work is part of the mpcEnergy project which is supported within
the program Regionale Wettbewerbsfähigkeit OÖ 2007-2013 by the European Fund for Regional Development as well as the State of Upper Austria.
The research reported in this article has been (partly) supported by the Austrian Ministry for Transport, Innovation and Technology, the Federal Ministry of Science, Research and Economy, and the Province of Upper Austria
in the frame of the COMET center SCCH. An earlier version of part of this
paper appeared in [1].
† Address all correspondence related to this paper to this author.
ficiency of power/fuel consumption is closely related to the
ability of the heating/cooling strategy to incorporate predictions of the heat-mass transfer dynamics and exogenous thermal inputs (e.g., outdoor temperature, solar radiation, user
activities, etc.). Naturally this observation leads to modelpredictive control (MPC) implementations (see, e.g., [2, 3]).
The performance though of any such implementation will be
closely related to the performance of the prediction models
for the evolution of both the indoor temperature and the exogenous thermal inputs.
A common approach in the derivation of prediction
models for the indoor temperature evolution relies on the
assumption that the heat-mass transfer dynamics can be
approximated well by linear models. For example, in
the optimal supervisory control formulation introduced in
[2] for controlling residential buildings with an heatingventilation-air-conditioning (HVAC) system, a linear (statespace) model is considered. As pointed out in [2], including
the actual dynamics might be complicated for the derivation
of optimal control strategies. Similar is the assumption in the
dynamic game formulation of HVAC thermally controlled
buildings in [6], where again a linear model is assumed for
the overall system. Also, in references [7, 8], detailed nonlinear representations of the heat-mass transfer dynamics (of
a vapor-compression system in [7] and for a building in [8])
are linearized about an operating point to allow for a more
straightforward controller design.
Given this difficulty in incorporating nonlinear representations in the controller design, several efforts for identification of heat-mass transfer dynamics have adopted linear
transfer functions, such as the standard ARX or ARMAX
black-box model structures (cf., [9]). Examples of such im-
plementations include the MIMO ARMAX model in [10],
the MISO ARMAX model for HVAC systems in [11] and
the ARX model structure considered in [12].
On the other hand, the nonlinear nature of the heatmass transfer dynamics in buildings has been pointed out by
several papers, including the switched linear dynamics for
modeling the intermittent operation of RH systems in [4],
the multiple-time scale analysis considered in [12, 5] and
the physics-based representation of the air-conditioning systems in [7]. In reference [12], a comparison is performed
between standard linear ARX models with two time-scale
transfer models, which according to the authors better represent thermal models. Indeed, the operation of either an
radiant-heating (RH) system or an HVAC system usually
takes place in a faster time-scale than the room dynamics,
an observation that was also utilized for the derivation of a
(nonlinear) model predictive controller in [5] using singularperturbation techniques.
The following questions naturally emerge: What is the
effect of the nonlinearities of an RH or an HVAC system in
the identification error of the indoor temperature? Can more
detailed model representations be utilized to reduce identification error and improve energy efficiency? This paper begins with the observation that thermal dynamics in buildings
are nonlinear in nature due to the operation of the RH and
HVAC systems. We avoid introducing any multiple timescale approximation as in [5] and we adopt a detailed representation of the heat-mass transfer dynamics (using standard
Bond-graph models). In principle, linear transfer models
incorporating linear regressors, such as Output-Error (OE),
ARX or ARMAX structures, may be sufficient for system
identification in buildings. However, we wish to investigate
whether this is a reasonable assumption and whether more
accurate representations of the dynamics are necessary.
To this end, linear regression models with nonlinear regressors are derived analytically using physical insight for
prediction of the indoor temperature in residential buildings
with RH and/or HVAC systems. The derivation accounts for
alternative information structures depending on the sensors
available, which is the main contribution of this paper. The
derived identification models are compared against standard
(black-box) linear regression models trained with simulated
data generated in EnergyPlus (V7-2-0) building simulator
tool developed by the U.S. Department of Energy [13]. Comparison is also performed with respect to the prediction performance of the derived models when used within a standard
MPC for climate control in a residential building. The results indicate that a more detailed design of the thermal-mass
transfer prediction models can be beneficial with respect to
the energy consumption. This paper extends prior work of
the same authors [1], since it extends the derivation of the regression models into a larger family of information structures
(sensors available), while it provides a detailed comparison
with respect to the energy consumption.
In the remainder of the paper, Section 2 provides a description of the overall framework and the objective of this
paper. Section 3 provides a short introduction to a class of
linear regression models (briefly LRM) for system identifica-
tion of heat-mass transfer dynamics. Section 4 analytically
derives using physical insight linear regression models with
nonlinear regressors (briefly NRM). In Section 5, we compare the performance of the derived NRM with LRM with
respect to both a) identification error, and b) prediction performance within a standard MPC for climate control. Finally,
Section 6 presents concluding remarks.
Notation:
• col{x1 , ..., xn }, for some real numbers x1 , ..., xn ∈ R,
denotes the column vector (x1 , ..., xn ) in Rn . Also,
row{x1 , ..., xn } denotes the corresponding row vector.
• diag{x1 , ..., xn } denotes the diagonal matrix in Rn×n
with diagonal entries x1 , ..., xn .
• for any finite set A, |A| denotes the cardinality of A.
.
• = denotes definition.
2
Framework & Objective
When addressing system identification and control of
heat-mass transfer dynamics in buildings, the assumed model
of indoor temperature prediction should be a) accurate
enough to capture the phenomena involved, and b) simple
enough to accommodate a better control design. Given that
prior literature has primarily focused on linear state-space or
transfer models, we wish to address the following:
1. investigate the performance of standard (black-box)
linear (regression) models (LRM) in the identification/prediction of heat-mass transfer dynamics in residential buildings;
2. derive analytically, using physical insight and under alternative information structures, an accurate representation of the dynamics using linear regression models with
nonlinear regressors (NRM);
3. provide a detailed comparison between the derived
NRM with standard LRM and assess the potential energy saving.
To this end, we consider residential buildings divided
into a set of zones I . Each zone is controlled independently
in terms of heating using either a RH and/or an HVAC system. We will use the notation i, j to indicate representative
elements of the set of zones I . The temperature of a zone i is
affected by the temperature of the neighboring zones and/or
the outdoor environment, denoted Ni , i.e., those zones which
are adjacent to i, separated by some form of separator.
In the remainder of this section, we provide some standard background on heat-mass transfer dynamics (based on
which the derivation of the prediction models will be presented). We further discuss the considered assumptions with
respect to the available sensors, and finally some generic
properties of the dynamics.
2.1
Background: Heat-mass transfer dynamics
We will use standard modeling techniques of heat-mass
transfer dynamics for each thermal zone i ∈ I , following
a standard Bond-graph approach [14] and the simplified
Fourier’s law for heat conduction and Newton’s law for heat
Fig. 1: Bond-graph approximation of heat-mass transfer dynamics of a thermal zone i ∈ I .
convection (cf., [15]). Such bond-graph approximations of
thermal dynamics are equivalent to a series of electrical
resistance-capacitor (RC) connections. Figure 1 provides a
detailed bond-graph approximation for a thermal zone i ∈ I
with both RH and HVAC heating equipment.
heat transfer dynamics for a separator i j can be written as
Q̇−
s,i j
We use C to denote the thermal capacitance of a separator (e.g., wall) or a thermal zone (e.g., room), and the symbol
R to denote the heat resistance, according to an electricity
equivalent terminology. We also use the symbol Q̇ to denote
heat flow and the symbol T to denote temperature.
The subscript s denotes separator-related parameters;
the subscript r denotes zone-related parameters, and usually
corresponds to a room; the subscript w denotes parameters
related to the RH system; and the subscript a denotes HVACrelated parameters. Finally, a pair i j denotes the interconnection between two neighboring zones i, j ∈ I .
Separators have been modeled as two thermal resistances separated by a capacitance, with the superscript “+”
denoting the “inside” part of the separator, and the superscript “−” denoting the “outside” part of the separator. For
example, R+
s,i j denotes the thermal resistance that it is adjacent to i. Furthermore, Tr,i , i ∈ I , denotes the temperature
of zone i and Ts,i j , j ∈ Ni , denotes the (internal) temperature of the separator i j. The separator exchanges heat with
the neighboring zones i and j due to heat convection, whose
−
resistance has been incorporated within R+
s,i j and Rs,i j . The
1
(Ts,i j − Tr,i )
R+
s,i j
1
= − (Tr, j − Ts,i j )
Rs,i j
+
= Q̇−
s,i j − Q̇s,i j ,
Q̇+
s,i j =
Cs,i j Ṫs,i j
(1)
+
where Q̇−
s,i j and Q̇s,i j denote the heat input and output for
separator i j, respectively.
Regarding the RH system, it is modeled by two thermal
restrictors Rw,i (V̇w,i ) separated by a thermal capacitance Cw,i
(that determines the heat stored in the thermal medium). In
.
particular, Rw,i = Rw,i (V̇w,i ) = 1/ρw cwV̇w,i and Cw,i = cw ρwVw ,
where ρw , cw and Vw are the density, specific heat capacity and volume of the water in the RH system. Due to the
capacitance of the thermal medium, a state variable Tw,i is
introduced that represents the temperature of the water at the
point of heat exchange. The radiator exchanges heat with
the thermal zone due to heat convection of resistance Rc,i .
Furthermore, the inlet water temperature in the RH system is
+
denoted by Tw,i
, while the outlet water temperature, denoted
−
Tw,i , is considered approximately equal to Tw,i (if we assume
uniform temperature distribution within the radiator device).
Thus, the heat transfer dynamics of the RH system are
Q̇w,i =
1
(Tw,i − Tr,i )
Rc,i
Cw,i Ṫw,i =
1
(T + − Tw,i ) − Q̇w,i ,
Rw,i (V̇w,i ) w,i
(2)
vector, respectively. Furthermore, we define
.
A
i (ui ) =
ar,i (V̇a,i ) row{a+
arw,i
rs,i j } j∈Ni
col{a+
0
s,i j } j∈Ni diag{as,i j } j∈Ni
aw,i
0
awc,i (V̇w,i )
where Q̇w,i is the heat input to zone i through the radiator.
Regarding the HVAC system, we consider the standard
assumption that the fluid is incompressible, thus we are only
interested in the thermal part of the system [16]. A thermal
resistance, Ra,i (V̇a,i ), is introduced to represent a restrictor of
heat flow. It is given by Ra,i = Ra,i (V̇a,i ) = 1/ρa caV̇a,i where ρa
denotes the air density, ca denotes the air specific heat capacity at constant pressure, and V̇a,i is the air volume rate. The
outlet temperature of air in the HVAC system is considered
approximately equal to the zone temperature, while the in+
let air temperature is denoted by Ta,i
. Thus, the heat input
attributed to the HVAC system is
Q̇HVAC,i =
1
(T + − Tr,i ).
Ra,i (V̇a,i ) a,i
Finally, the heat-mass transfer dynamics of the thermal
zone i can be expressed as follows
∑ Q̇+s,i j + Q̇w,i + Q̇HVAC,i + Q̇ext,i .
(4)
j∈Ni
From equations (1), (2), (3) and (4), we can derive in a
straightforward manner the state-space representation of the
overall system, which can be written in a generalized form
as follows
ẋi (t) = Ai (ui (t))xi (t) + Ei (ui (t))di (t),
i ∈ I,
(5)
where
Tr,i
.
xi = col{Ts,i j } j∈Ni ,
Tw,i
+
Tw,i
+
Ta,i
.
di =
col{Tr, j } j∈N ,
Q̇ext,i
.
ui =
where
.
ar,i (V̇a,i ) = − ∑ j∈Ni 1/Cr,i R+s,i j − 1/Cr,i Rc,i − 1/Cr,i Ra,i (V̇a,i ),
. 1
.
+
a+
ara,i (V̇a,i ) = 1/Cr,i Ra,i (V̇a,i ),
rs,i j = /Cr,i Rs,i j ,
.
. 1
−
a+ = 1/Cs,i j R+s,i j , a−
s,i j = /Cs,i j Rs,i j ,
(6)
. s,i j+
. 1
−
as,i j = −as,i j − as,i j , aww,i (V̇w,i ) = /Cw,i Rw,i (V̇w,i ),
. 1
. 1
. 1
aw,i = /Cw,i Rc,i , aext,i = /Cr,i , arw,i = /Cr,i Rc,i ,
.
awc,i (V̇w,i ) = −aw,i − aww,i (V̇w,i ).
(3)
The disturbance Q̇ext,i denotes heat inputs from external
sources. It will correspond to a vector of external heat flow
rates, attributed to solar radiation, human presence, etc.
Cr,i Ṫr,i =
.
E
i (ui ) =
0
ara,i (V̇a,i )
0
aext,i
−
0
0
diag{as,i j } j∈Ni 0
aww,i (V̇w,i )
0
0
0
V̇w,i
V̇a,i
,
i
are the state vector, control-input vector and disturbance
2.2
Information structures
Different information structures will be considered depending on the sensors available. More specifically, for the
remainder of the paper, we consider the following cases:
• Full information (FI): All state, control and disturbance
variables can be measured except for the separator state
variables {Ts,i j } j .
• Medium Information (MI): All state, control and disturbance variables can be measured except for the separator state variables, {Ts,i j } j , and the water state variable
at the point of heat exchange, Tw,i .
• Limited information (LI): All state, control and disturbance variables can be measured except for the separator state variables {Ts,i j } j , the water state variable Tw,i
−
+
and the disturbance variables Tw,i
and Tw,i
.
The specifics can also be visualized in Table 1. The introduced alternative structures intend on demonstrating the
flexibility of the proposed derivation of nonlinear regressors,
since different residential buildings may not necessarily incorporate all sensors of the FI structure.
The case of limited information (LI) corresponds to the
case where only the flow rates are known in the RH and/or
the HVAC system, while the corresponding inlet/outlet temperatures of the thermal medium cannot be measured. Such
assumption may be reasonable for most today’s residential
buildings.
Alternative information structures may be considered,
such as the ones where Q̇ext,i (due to radiation or peoples’
presence) is not known, however such cases can be handled
through the above cases and with small modifications.
Tr,i
{Ts,i j } j∈Ni
Tw,i
V̇w,i
V̇a,i
+
Tw,i
+
Ta,i
{Tr, j } j∈Ni
Q̇ext,i
(FI)
X
X
X
X
X
X
X
X
(MI)
×
X
×
X
X
X
X
X
X
(LI)
×
X
×
×
X
X
×
×
X
X
Table 1: Information structures with measured variables.
2.3
The RH dynamics can be written as follows
System separation under FI
Under the full information structure (FI), a natural separation of the system dynamics can be introduced between
the zone and RH dynamics. Such separation will be used at
several points throughout the paper, and it is schematically
presented in Figure 2.
ẋw,i (t) = Aw,i (uw,i (t))xw,i (t) + Ew,i (uw,i (t))dw,i (t),
where
.
xw,i = Tw,i ,
dw,i
uw,i
dr,i
Tw,i
RH
.
Aw,i (ui ) = awc,i (V̇w,i ) ,
ur,i
Fig. 2: System separation architecture under FI.
ẋr,i (t) = Ar,i (ur,i (t))xr,i (t) + Er,i (ur,i (t))dr,i (t)
(7)
.
where ur,i = V̇a,i is the control-input vector and
Tw,i
+
Ta,i
.
dr,i =
col{Tr, j } j∈N ,
i
Q̇ext,i
.
xr,i =
Tr,i
col{Ts,i j } j∈Ni
,
are the state vector and disturbance vector of the zone dynamics, respectively. Furthermore, we define
.
Ar,i (ui ) =
row{a+
rs,i j } j∈Ni
ar,i (V̇a,i )
,
col{a+
s,i j } j∈Ni diag{as,i j } j∈Ni
0
aext,i
. arw,i ara,i (V̇a,i )
Er,i =
.
0
0
diag{a−
s,i j } j∈Ni 0
.
dw,i =
+
Tw,i
Tr,i
.
Ew,i = aww,i (V̇w,i ) aw,i .
3
In particular, the overall (heat-mass transfer) dynamics
for a single zone can be written as follows
.
uw,i = V̇w,i ,
are the state vector, control-input vector and disturbance vector of the RH system, respectively. Furthermore, we define
the following system matrices
Tr,i
Zone
(8)
Linear Regression Models (LRM)
We would like to focus on models for system identification which (to great extent) retain the physical interpretation of the parameters. To this end, an output-error model
structure (cf., [9]) is considered, where a one-to-one correspondence between the identification parameters and the
discrete-time model parameters can be established. In the
remainder of this section, we provide some background on
the Output-Error (OE) model structure, while we also discuss the implicit assumptions made when a linear structure
of the dynamics is assumed.
3.1
Background: Output-Error (OE) model
If we assume that the relation between input u and the
undisturbed output of the system w can be written as a linear difference equation, and that the disturbance consists of
white measurement noise v, then we obtain:
w(k) + f1 w(k − 1) + ... + fn f w(k − n f )
= b1 u(k − 1) + ... + bnb u(k − nb )
(9)
and the output is y(k) = w(k) + v(k), for some positive integers n f ≥ nb . The parameter vector to be determined is:
.
θ = (b1 , . . . , bnb , f1 , . . . , fn f ). Since w(k) is not observed, it
should be constructed from previous inputs and it should
carry an index θ.
The natural predictor (resulting from a maximum a pos-
teriori predictor1 ) is: ŷ(k|θ) = w(k, θ), and it is constructed
from past inputs only. If we define the vector:
.
ϕ(k, θ) = u(k − 1), . . . , u(k − nb ),
−w(k − 1, θ), . . . , −w(k − n f , θ) ,
may lead to significant identification errors. To see this, let
us consider the full information structure (FI). In a forthcoming section, we will show that the following approximation
of the RH part of the dynamics holds for sufficiently small
sampling period ε > 0:
(10)
then, the predictor can be written more compactly as:
ŷ(k|θ) = ϕ(k, θ)T θ, leading to a linear regression model
(LRM), where ϕ(k, θ) is called the regression vector. To simplify notation, in several cases we will write ϕ(k) instead of
ϕ(k, θ). Note that in the above expression, the w(k − j, θ),
j = 1, 2, ..., n f , are not observed, but using the above maximum a-posteriori predictor, they can be computed using
previous inputs as follows: w(k − j, θ) = ŷ(k − j|θ), j =
1, 2, ..., n f .
Such output-error model predictors will be used
throughout the paper. However, note that, depending on
the physics of the identified system, the regression vector
ϕ defined in (10) may be nonlinear with respect to the input/output lagged variables. In such cases, the resulting predictor will be referred to as a linear regression model with
nonlinear regressors (NRM).
3.2
Tw,i (k + 1) ≈
+
(1 + εawc,i (V̇w,i (k)))Tw,i (k) + εaww,i (V̇w,i (k))Tw,i
(k)+
εaw,i Tr,i (k),
plus higher-order terms of ε.
According to this (finite-response) approximation, a linear regression vector ϕ may describe well the evolution of
Tw,i as long as either the flow rate V̇w,i or the temperatures
+
Tw,i and Tw,i
are not varying significantly with time. However, variations in the flow rate V̇w,i and in the water temperature Tw,i may be large with time. Similar are the conclusions
when investigating the effect of the air flow rate V̇a,i in the
evolution of the zone temperature. The exact effect of these
nonlinear effects cannot be a-priori determined.
4
Linear Regression Models with Nonlinear Regressors
(NRM)
In this section, we will explore the formulation of OE
model structures when the regressors in ϕ may be nonlinear
functions of lagged input/output variables. Such investigation will be performed under the FI, MI and LI structure,
extending prior work of the same authors [1] to a larger set
of possible information structures (beyond the FI structure).
For the derivation, we will be using the following notation: Let q−1 {·} denote the one-step delay operator, i.e.,
q−1 {x(k)} = x(k − 1). Note that the delay operator is linear.
Let us also define the following operators:
.
Ps,i j = [(1 + εas,i j )q−1 ]
.
Pwc,i (V̇w,i ) = [(1 + εawc,i (V̇w,i ))q−1 ]
.
Pr,i (V̇a,i ) = [(1 + εar,i (V̇a,i ))q−1 ]
Discussion
Output-error model structures do not consider any disturbance terms in the process dynamics, instead they only
consider measurement noise. Hence, such model structures
are rather appropriate for providing a clear picture of the impact of the assumed (physical) model into the prediction error. On the other hand, when considering structures with process noise, such as ARMAX model structures (cf., [9, Section 4.2]), the process dynamics are perturbed by artificial
terms, which are not easily motivated by the physical insight.
Since the goal of this paper is to evaluate the impact of the
assumed (physical) dynamics, we consider OE models more
appropriate for evaluating predictions. This, however, does
not imply that the performance of an OE model structure is
necessarily better compared to an ARMAX model. The goal
of this paper is not to provide such a comparison.
Furthermore, identifiability (cf. [9, Section 4.6]) of the
considered OE model structures will be guaranteed by the
(inherent) controllability of the heat-mass transfer dynamics.
where ε > 0 defines the sampling period. Define also
.
Qs,i j = [1 − Ps,i j ],
.
Qwc,i (V̇w,i ) = [1 − Pwc,i (V̇w,i )],
.
Qr,i (V̇a,i ) = [1 − Pr,i (V̇a,i )].
3.3
For any operator P{·}, P−1 {·} will denote its inverse operator, i.e.,
Linear approximation & implicit assumptions
We wish to explore the utility of an OE model structure
using a linear regression vector ϕ into identifying the nonlinear system dynamics of Equation (5). Note that the original
dynamics (5) are bilinear in nature due to multiplications of
the flow rates (V̇w,i , V̇a,i ) with state or disturbance variables.
An investigation of the original dynamics (5) reveals that
when an output-error model is considered that admits a linear regression vector ϕ, we implicitly admit assumptions that
1 i.e., the one that maximizes the probability density function of the output given observations up to the previous time instant.
P−1 {P{x(k)}} = x(k).
Property 4.1. For each zone i ∈ I , the family of operators
{Qs,i j } j∈Ni are pairwise commutative, i.e.,
Qs,i j Qs,i j0 {x(k)} = Qs,i j0 Qs,i j {x(k)}
for any j, j0 ∈ I with j 6= j0 .
Proof. For each zone i ∈ I , and for any two neighboring
zones j, j0 ∈ Ni , we have the following:
T̂w,i (k|θw,i ) ≈ ϕw,i (k)T θw,i plus higher-order terms of ε, where
θw,i is a vector of unknown parameters and
Qs,i j Qs,i j0 {x(k)}
= [1 − Ps,i j ] 1 − Ps,i j0 {x(k)}
= [1 − Ps,i j ] x(k) − (1 + εas,i j0 )x(k − 1)
= x(k) − (1 + εas,i j0 )x(k − 1)−
(1 + εas,i j )x(k − 1) + (1 + εas,i j )(1 + εas,i j0 )x(k − 2).
T̂w,i (k − 1|θw,i )
. V̇w,i (k − 1)T̂w,i (k − 1|θw,i )
.
ϕw,i (k) =
+
V̇w,i (k − 1)Tw,i
(k − 1)
Tr,i (k − 1)
It is straightforward to check that the same expression also
occurs if we expand Qs,i j0 Qs,i j , due to the fact that (1+εas,i j )
commutes with (1 + εas,i j0 ). Thus, the conclusion follows. •
Proof. By a Taylor-series expansion of the RH dynamics, the
finite-step response of the water temperature Tw,i (k) is
Tw,i (k) ≈
(1 + εawc,i (V̇w,i (k − 1)))Tw,i (k − 1)+
+
εaww,i (V̇w,i (k − 1))Tw,i
(k − 1) + εaw,i Tr,i (k − 1),
Another two properties that will be handy in several
cases are the following:
Property 4.2. For each zone i ∈ I , we have
plus higher-order terms of ε. Equation (12) directly results
from the definition of the delay operator Pwc,i (V̇w,i ). From
Equation (12), a regression vector may be derived of the desired form. •
[Qs,i j Qwc,i (V̇w,i (k))] = [Qwc,i (V̇w,i (k))Qs,i j ]−
(1 + εas,i j )ε (1 − q−1 )awc,i (V̇w,i (k)) q−2 .
Proof. See Appendix A. •
Property 4.3. For some finite index set A and signal x(k),
h
∏ Qs,i j
j∈A
i
{x(k)} = x(k) +
|A|
∑ αm x(k − m),
(13)
(11)
m=1
for some constants α1 , ..., α|A| .
Proof. The conclusion follows in a straightforward manner
by induction and the definition of the delay operator Qs,i j . •
4.1.2 Zone dynamics
A similar approach to Proposition 4.1 for the derivation
of a nonlinear regression vector can also be employed for the
zone dynamics.
Proposition 4.2 (NRM for Zone under FI). For
sufficiently small sampling period ε > 0, the maximum
a-posteriori predictor of Tr,i (k) can be approximated by
T̂r,i (k|θr,i ) ≈ ϕr,i (k)T θr,i plus higher-order terms of ε, where
θr,i is a vector of unknown parameters and
|Ni |+1
col{T̂r,i (k − m|θr,i )}m=1
n
o
|Ni |+1
col col{Tr, j (k − m)}m=2
j∈Ni
Ni |+1
. col{V̇a,i (k − m)T̂r,i (k − m|θr,i )}|m=1
.
ϕr,i (k) =
col{V̇a,i (k − m)T + (k − m)}|Ni |+1
a,i
m=1
|Ni |+1
col{T̂w,i (k − m|θw,i )}m=1
|Ni |+1
col{Q̇ext,i (k − m)}m=1
4.1
Full-Information (FI) structure
Using the natural decomposition of the dynamics under
the FI structure described in Section 2.3, in the following
subsections we provide a derivation (using physical insight)
of nonlinear regression vectors under an OE model structure
for the two subsystems of the overall dynamics.
4.1.1 RH dynamics
By isolating the RH dynamics, we first formulate a
(finite-response) prediction of the water temperature Tw,i , as
the following proposition describes.
Proposition 4.1 (NRM for RH under FI). For
sufficiently small sampling period ε > 0, the RH dynamics (8)
can be approximated by:
Tw,i (k) ≈ Pwc,i (V̇w,i (k − 1)){Tw,i (k)}+
+
εaww,i (V̇w,i (k − 1))Tw,i
(k − 1) + εaw,i Tr,i (k − 1), (12)
plus higher-order terms of ε. Furthermore, the maximum
a-posteriori predictor of Tw,i (k) can be approximated by
Proof. See Appendix B. •
4.2
Medium-Information (MI) structure
The case of MI structure is slightly different from the
FI structure, since the water temperature, Tw,i , at the point
of heat exchange cannot be directly measured, therefore the
dynamics of the RH and the zone cannot be separated as was
demonstrated in Figure 2. The following proposition provides the corresponding NRM predictor for the MI structure.
Proposition 4.3 (NRM for Zone under MI). For sufficiently small sampling period ε > 0, the maximum aposteriori predictor of Tr,i (k) can be approximated by
T̂r,i (k|θr,i ) ≈ ϕr,i (k)T θr,i plus higher-order terms of ε, where
θr,i is a vector of unknown parameters and
.
ϕr,i (k) =
|Ni |+2
col{T̂r,i (k − m|θr,i )}m=1
n
o
|Ni |+2
col col{Tr, j (k − m)}m=2
j∈Ni
N
+2
|
|
i
col{V̇a,i (k − m)T̂r,i (k − m|θr,i )}m=1
|Ni |+2
+
col{
V̇
(k
−
m)T
(k
−
m)}
a,i
a,i
m=1
N
+2
|
|
i
+
col{V̇a,i (k − m)V̇w,i (k − m)Ta,i
(k − m)}m=2
(14)
N
+1
| i|
.
col{
V̇
(k
−
m
−
1)
T̂
(k
−
m|θ
)}
w,i
r,i
r,i
m=1
|Ni |+2
col{V̇w,i (k − m)T̂r,i (k − m|θr,i )}m=2
|Ni |+2
col{V̇w,i (k − m)V̇a,i (k − m)T̂r,i (k − m|θr,i )}m=2
|Ni |+2
+
col{V̇w,i (k − m)Tw,i
(k − m)}m=2
|Ni |+2
col{Q̇ext,i (k − m)}m=1
|Ni |+2
col{V̇w,i (k − m)Q̇ext,i (k − m)}m=2
Proof. The proof follows similar reasoning with the proof of
Proposition 4.2 and it is presented in Appendix C. •
4.3
Limited-Information (LI) structure
Similarly to the derivation of the NRM for the FI and
MI structures, we may also derive the corresponding NRM
for the case of the LI structure. Note that the difference of
such information structure (compared to MI) is the fact that
+
+
the inlet temperatures of the thermal mediums (Tw,i
and Ta,i
)
cannot be measured. In such case, the only valid assumption
is the fact that the inlet temperature of the thermal mediums
remain constant with time. In this case, Proposition 4.3 con+
+
tinue to hold, when Tw,i
and Ta,i
are replaced with unity in
(14) (i.e., they become part of the parameters in θr,i ).
5
Performance Evaluation
In this section, we provide a comparison between standard linear regression models and the nonlinear ones derived
in Section 4. The comparison will be performed both with
respect to a) identification error, and b) prediction performance within a standard MPC implementation for climate
control in a residential building. The second part is considered the most representative of the performance of an identification model since even small prediction errors might lead
to significant performance degradation.
5.1
Simulation platform
In order to compare the proposed nonlinear regression
vectors with standard linear regressors, we used the EnergyPlus (V7-2-0) building simulator developed by the U.S.
Department of Energy [13]. The BCVTB simulation tool
has also been used for allowing data collection and also climate control developed in MATLAB to be implemented during run-time. A typical (three-storey) residential building in
Linz, Austria, was accurately modeled and simulated with
the EnergyPlus environment to allow for collecting data from
a realistic residential environment.
5.2
Data generation
The data collection for system identification was performed under normal operating conditions of the heating system during the winter months (October - April) under the
weather conditions of Linz, Austria. To replicate normal operating conditions, a standard hysteresis controller was employed, according to which the water flow V̇w,i (k) is updated
regularly at time instances tk = kTsam where Tsam = 1/12h.
The control law is as follows:
.
V̇w,i (k) =
V̇w,max
0
if pi (k) > 0 and
(Tr,i (k) < Tset − ∆T or
(Tr,i (k) ≥ Tset − ∆T and Tr,i (k − 1) ≤ Tr,i (k))),
else
where ∆T determines a small temperature range about the
desired (set) temperature, Tset , in which the control flow
maintains its value and V̇w,max denotes the maximum water
flow set to 0.0787kg/sec. Furthermore, pi (k) ∈ {0, 1} indicates whether people are present in thermal zone i. In other
words, the hysteresis controller applies only if someone is
present in the thermal zone i, which can be easily determined
through a motion detector. Furthermore, the inlet water tem+
perature Tw,i
is determined by the following heating curve:
+
Tw,i
.
=
(
ρ0 + ρ1 · (Tset − Tout )ζ , if Tset > Tout
ρ0 ,
else.
(15)
where ρ0 , ρ1 and ζ are positive constants. For the simulation,
we set ∆T = 0.1, ρ0 = 29.30, ρ1 = 0.80 and ζ = 0.97. The
set temperature was set equal to Tset = 21oC.
There exists a natural ventilation system that operates
autonomously with an intermittent flow pattern V̇a,i . Thus,
the only control parameter is V̇w,i . All parameters mentioned
in Table 1 can be measured, except for {Ts,i j } j , which allows
for evaluating all considered structures (FI, MI and LI).
5.3
Recursive system identification
To utilize the linear and nonlinear regression vectors for
system identification of heat-mass transfer dynamics of a residential building, we use an OE model structure as described
in Section 4, while we employ a regularized recursive leastsquares implementation (cf., [17, Section 12.3]) for training its parameters. The reason for implementing a recursive identification procedure is primarily due to the fact that
the size of the available data is quite large, which makes the
use of standard least-squares approaches practically infeasible. Besides, predictions for the zone temperature will be
needed continuously during run-time, demanding for more
efficient computational schemes. Furthermore, a recursive
least squares implementation allows for an adaptive response
to more recent data, thus capturing more recent effects.
the form (17) the following condition is satisfied:2
#2
"
E ∆Gi0 (q)yi (k) + ∑ ∆Gil (q)zl (k)
l∈L
5.4
Identification experiments
The experiments performed in this study will involve a)
a standard linear regression model, with b) the derived NRM
of Proposition 4.3 (which corresponds to the MI structure).
Our intention is to evaluate the benefit of the more accurate
physical representation of the models derived in Section 4.
In both regression vectors (linear or nonlinear) we consider
|Ni | = 1, i.e., we implicitly assume that there is a single
neighboring zone. This can be justified by the fact that the
building can be considered as a single thermal zone, since
the same heating schedule is employed in all rooms.
In particular, the linear regression model (LRM) implemented corresponds to the following structure:
|Ni |+2
col{
T̂
(k
−
m|θ
)}
r,i
r,i m=1
n
o
|Ni |+2
col col{Tr, j (k − m)}m=1
j∈Ni
|Ni |+2
col{V̇a,i (k − m)}m=1
.
|Ni |+2
+
ϕr,i (k) =
,
col{Ta,i (k − m)}m=1
|Ni |+2
col{V̇w,i (k − m)}m=1
N
+2
|
|
col{T + (k − m|θr,i )} i
w,i
m=1
|Ni |+2
col{Q̇ext,i (k − m)}m=1
(16)
i.e., it only considers delayed samples of the state and the
available input and disturbance parameters. The NRM model
implemented corresponds to the one of Equation (14).
5.5
Persistence of Excitation
Following the analysis of Section 4, for any thermal
.
zone i ∈ I , the output of the zone yi (k) = Tr,i (k) can briefly
be expressed in the following form:
(17)
l∈L
for some linear transfer functions Gi0 (q), Gil (q), l ∈ L and
some index set L ⊂ N. The terms zil (k) represent time sequences of input, disturbance or output signals, or products
of such terms. For example, if we consider the MI model
presented in (14), the terms zil may correspond to a disturbance, such as Tr, j , j ∈ Ni , or to a product of a control input
and the output, such as V̇w,i (k)Tr,i (k).
A model of the generic form of equation (17) is uniquely
determined by its corresponding transfer functions Gi0 (q),
Gil (q), l ∈ L. Thus, in order for the data to be informative,
the data should be such that, for any two different models of
(18)
where ∆Gi0 (q), ∆Gil (q) denote the corresponding differences between the transfer functions of the two different
.
models. Moreover, E = limN→∞ 1/N ∑Nk=1 E[·] and E[·] denotes the expectation operator. Since the models are assumed
different, either ∆Gi0 (q) or ∆Gil (q), for some l ∈ L, should
be nonzero. Thus, if ∆Gil (q) 6= 0, for some l ∈ L, the persistence of excitation 3 of zil (k) will guarantee (18).
Since data are collected through a closed-loop experiment, special attention is needed for terms zil (k) which involve the product of the water flow rate V̇w,i (k) with the output yi (k) of the system, such as the term V̇w,i (k)yi (k) in (14).
Note, however, that this term is nonlinear in both yi (k) and
the presence/occupancy indicator, since the hysteresis controller a) applies only when someone is present in the zone
and b) is a nonlinear function of the output yi (k). Thus, it
is sufficient for either the input signal or the occupancy indicator to be persistently exciting in order for the experiment
to be informative. In Figure 3, we provide the positive spectrum generated through the discrete Fast Fourier Transform
(DFFT) of the disturbance and control signals.
For either the simplified model of (16) or the more detailed model of (14), note that any transfer function (from
an input/disturbance to the zone temperature) requires the
identification of at most 2( Ni + 2) parameters. Thus, in
order for an experiment to be informative enough, it is sufficient that the considered inputs/disturbances are persistently exciting of order 2( Ni + 2) or higher. In the case of
Ni = 1 (considered in this experiment), it suffices that any
input/disturbance signal is persistently exciting of order 6 (or
5 in case one of the frequencies is at zero, as it is the case for
all signals in Figure 3). As demonstrated in Figure 3, this
condition is satisfied by the input/disturbance signals, since
for all of them the positive spectrum is non-zero in at least 3
distinct frequencies.
5.6
yi (k) = Gi0 (q)yi (k) + ∑ Gil (q)zil (k),
6= 0,
Identification error comparison
In Figure 4, we demonstrate the resulting identification error under the LRM of (16) and the NRM of (14) for
Ni = 1. Note that the NRM achieves a smaller identification error of about 10%. This reduced error is observed later
in time (due to the larger training time required by the larger
number of terms used in the nonlinear regression vector).
The data used for training correspond to data collected between October and April from the simulated building, however the same data have been reused several times for better
fitting. Thus, the time axis in Figure 4 does not correspond
to the actual simulation time, but it corresponds to the accumulated time index of the reused data.
2 cf.,
3 cf.,
discussion in [9, Section 13.4].
[9, Definition 13.2]
Presence
Running Average Squared Error (oC2 )
0.4
Power
400
0.3
200
0
LRM
NRM
0
0.1
0.2
0.3
0.4
Water Flow
0.5
0.6
0.2
0.1
0
1
0.5
Power
400
Power
Power
1
0.8
0.6
0.4
0.2
0
1
0.8
0.6
0.4
0.2
0
Power
2
0
·105
0
·104
0
0.1
0.2
0.3
Water Inlet Temperature
1
0.5
1.5
2
2.5
Outdoor Temperature
5 · 10−2
·107
0.1
0.4
0.15
3
·10−2
0.2
min
·105
Nhor n
α∑
o
2
p̂i (k) T̂r,i (k) − Tset (k) /Nhor +
k=0
Nhor −1 n
Solar Radiation
∑
k=0
s.t.
1
0.5
0
5 · 10−2
0.1
0.15
0.2
Frequency (1/h)
Fig. 3: Positive Spectrum of Input and Disturbance Signals.
5.7
3.5
weather conditions may dramatically change within a few
days). This may potentially result in a significant degradation of the prediction performance when the prediction has
not taken into account the effect of the nonlinearities present
in the dynamics. The effect of these nonlinearities may not
be known a-priori and a thorough investigation is necessary
to understand their potential economic impact.
To this end, we designed a standard MPC for the RH
system of the main living area of the residential building.
The goal is to evaluate the performance of the previously designed (nonlinear) regression vectors for the zone temperature prediction compared to standard (black-box) linear regression vectors. The structure of the MPC is rather simple
and addresses the following optimization problem.
o
−
+
βTsam Tw,i
(k) − T̂w,i
(k) + γTsamV̇w,i (k)
(19a)
1.5
0
3
Fig. 4: Identification performance for the zone temperature.
200
0
2
1.5
2.5
time (hours)
Prediction performance comparison
Small identification error may not necessarily guarantee
a good prediction performance, especially when predictions
are required for several hours ahead. From the one hand,
training for system identification is based upon the one-step
ahead innovations, which does not necessarily guarantee a
similar accurate prediction for several hours ahead. On the
other hand, training might have been performed under operating conditions that are (possibly) different from the operating conditions under which predictions are requested (e.g.,
var.
T
(19b)
−
T̂w,i
(k) ≈ T̂w,i (k) ≈ ϕw,i (k) · θw,i
+
Tw,i (k) ∈ {40oC, 45oC},
(19c)
T̂r,i (k) ≈ ϕr,i (k) θr,i
V̇w,i (k) ∈ {0, V̇w,max },
(19d)
(19e)
k = 0, 1, 2, ..., Nhor − 1,
Note that the first part of the objective function (19a)
corresponds to a comfort cost. It measures the average
squared difference of the zone temperature from the desired
(or set) temperature entered by the user at time k. The set
temperature was set equal to 21oC throughout the optimization horizon. The variable p̂i (k) ∈ {0, 1} holds our estimates
on whether people are present in zone i at time instance k.
The second part of the objective function (19a) corresponds to the heating cost, while the third part corresponds
to the pump-electricity cost. The nonnegative parameters
β, γ were previously identified for the heating system of
the simulated building and take values: β = 0.3333kW /oCh,
γ = 0.5278 · 103 kW sec/hm3 . The non-negative constant α is
introduced to allow for adjusting the importance of the comfort cost compared to the energy cost. A large value has been
assigned equal to 106 to enforce high comfort.
The p̂i (k), k = 1, 2, ...,, as well as the outdoor temperature, Tout , and the solar gain, Q̇ext (k), are assumed given
(i.e., predicted with perfect accuracy). This assumption is
essential in order to evaluate precisely the impact of our temperature predictions (19b) in the performance of the derived
optimal controller.
The sampling period was set to Tsam = 1/12h, the optimization period was set to Topt = 1h, and the optimization
horizon was set to Thor = 5h. This implies that Nhor = 5 ·12 =
60. Furthermore, the control variables are the inlet water
temperature which assumes two values (40oC and 45oC) and
the water flow which assumes only two values, the minimum
0kg/sec and the maximum V̇w,max = 0.0787kg/sec.
For the prediction model of (19b), we used either the
LRM of (16) or the NRM of (14), both trained offline using data collected during normal operation of a hysteresis
controller (as demonstrated in detail in Section 5.2). For
the prediction model of the outlet water temperature used in
the computation of the cost function, we used the prediction
model derived in (13). As in the case of Section 5.6 where the
LRM and the NRM were compared, we evaluated the value
of the cost function under these two alternatives. The performances of the two models with respect to the comfort cost
(which corresponds to the 1st term of the objective (19a))
are presented in Figure 5. The performances of the two models with respect to the energy spent (i.e., heating and pump
electricity cost), which correspond to the 2nd and 3rd term
of the objective (19a), are presented in Figure 6. Note that
the NRM achieves lower comfort cost using less amount of
energy which is an indication of small prediction errors.
3
2.5
2
1.5
1
0.5
0
Comfort Cost |Tr,i (k) − Tset | (oC)
LRM
NRM
0
200
400
600 800
Time (h)
1,000 1,200
Fig. 5: Running-average comfort cost under LRM and NRM.
5.8
Discussion
Note that the NRM achieves an improvement to both
prediction accuracy and cost performance. The fact that the
LRM may not provide a similarly good performance as the
Heating and Pump Electricity Cost (kW h/Tsam )
0.2
0.1
0
LRM
NRM
0
200
400
600 800
Time (h)
1,000 1,200
Fig. 6: Running-average heating and pump electricity cost
under LRM and NRM.
NRM could be attributed to the fact that the operating conditions for testing the MPC controller might not be identical with the training conditions. Note, for example, that
the training phase was performed under the heating-curve
pattern of (15), while the MPC employs only two alternative water inlet temperatures. This variation in the inlet water temperature may have exhibited different patterns in the
nonlinear parts of the dynamics, thus affecting the prediction
performance.
The simulation study presented in this paper considered
a regression vector generated with Ni = 1, which corresponds to the case of a single neighboring zone (e.g., the
outdoor temperature). The proposed regression models are
also applicable for higher-order models (i.e., when more than
one neighboring zones are present), however it might be necessary that some of the input signals are designed to include
larger number of frequencies. Note, however, that the input
signals that can be designed (e.g., medium flow rates and inlet temperatures) appear as products with other disturbance
variables in the derived regression models (e.g., equation
(14)). Thus, persistence of excitation for such products of
non-correlated signals is an easier task. On the other hand,
some of the disturbance signals (e.g., the solar radiation or
the outdoor temperature) cannot be designed. Of course, the
prediction model will operate at conditions that are similar to
which it was identified. However, the impact of such disturbance signals when they are not informative enough needs to
be further investigated.
6
Conclusions
We derived analytically regression models admitting
nonlinear regressors specifically tailored for system identification and prediction for thermal dynamics in buildings. The
proposed models were compared with standard (black-box)
linear regression models derived through the OE structure,
and an improvement was observed with respect to both the
resulting prediction performance as well as the energy cost.
The greatest advantage of the proposed identification scheme
relies on the fact that it provides a richer and more accurate
representation of the underlying physical phenomena, contrary to standard black-box identification schemes.
References
[1] Chasparis, G., and Natschläger, T., 2014. “Nonlinear system identification of thermal dynamics in buildings”. In Control Conference (ECC), 2014 European,
pp. 1649–1654.
[2] Nghiem, T., and Pappas, G., 2011. “Receding-horizon
supervisory control of green buildings”. In Proc. of the
2011 American Control Conference, pp. 4416–4421.
[3] Oldewurtel, F., Parisio, A., Jones, C., Morari, M., Gyalistras, D., Gwerder, M., Stauch, V., Lehmann, B., and
Morari, M., 2012. “Use of model predictive control and
weather forecasts for energy efficient building climate
control”. Energy and Buildings, 45, pp. 15–27.
[4] Nghiem, T., Pappas, G., and Mangharam, R., 2013.
“Event-based green scheduling of radiant systems in
buildings”. In Proc. of 2013 American Control Conference (ACC), pp. 455–460.
[5] Touretzky, C., and Baldea, M., 2013. “Model reduction
and nonlinear MPC for energy management in buildings”. In Proc. of 2013 American Control Conference
(ACC), pp. 455–460.
[6] Coogan, S., Ratliff, L., Calderone, D., Tomlin, C., and
Sastry, S., 2013. “Energy management via pricing in
LQ dynamic games”. In Proc. of 2013 American Control Conference (ACC), pp. 443–448.
[7] Rasmussen, B., Alleyne, A., and Musser, A., 2005.
“Model-driven system identification of transcritical vapor compression systems”. IEEE Transactions on Control Systems Technology, 13(3), pp. 444–451.
[8] Maasoumy, M., Razmara, M., Shahbakhti, M., and Vincentelli, A. S., 2014. “Handling model uncertainty in
model predictive control for energy efficient buildings”.
Energy and Buildings, 77, pp. 377 – 392.
[9] Ljung, L., 1999. System Identification: Theory for the
User, 2nd ed. Prentice Hall Ptr, Upper Saddle River,
NJ.
[10] Yiu, J.-M., and Wang, S., 2007. “Multiple ARMAX
modeling scheme for forecasting air conditioning system performance”. Energy Conversion and Management, 48, pp. 2276–2285.
[11] Scotton, F., Huang, L., Ahmadi, S., and Wahlberg, B.,
2013. “Physics-based modeling and identification for
HVAC systems”. In Proc. of 2013 European Control
Conference (ECC), pp. 1404–1409.
[12] Malisani, P., Chaplais, F., Petit, N., and Feldmann, D.,
2010. “Thermal building model identification using
time-scaled identification models”. In Proc. of 49th
IEEE Conference on Decision and Control, pp. 308–
315.
[13] EnergyPlus. EnergyPlus Energy Simulation software
Version: 7-2-0.
[14] Karnopp, D., Margolis, D., and Rosenberg, R., 2012.
System Dynamics: Modeling, Simulation and Control
of Mechatronic Systems, 5th ed. John Wiley & Sons,
Inc, Hoboken, NJ.
[15] Thirumaleshwar, M., 2009. Fundamentals of Heat &
Mass Transfer. Dorling Kindersley (India) Pvt. Ltd,
New Dehli, India.
[16] Karnopp, D., 1978. “Pseudo bond graphs for thermal
energy transport”. ASME Journal of Dynamic Systems,
Measurement and Control, 100, pp. 165–169.
[17] Sayed, A., 2003. Fundamentals of Adaptive Filtering.
John Wiley & Sons, Inc., New Jersey.
A
Proof of Property 4.2
We can write:
[Qs,i j Qwc,i
(V̇w,i (k))] {x(k)}
= Qs,i j Qwc,i (V̇w,i (k)) {x(k)}
= Qs,i j x(k) − (1 + εawc,i (V̇w,i (k)))x(k − 1)
= x(k) − (1 + εawc,i (V̇w,i (k)))x(k − 1)−
(1 + εas,i j )x(k − 1) + (1 + εas,i j )x(k − 2)+
(1 + εas,i j )εawc,i (V̇w,i (k − 1))x(k − 2).
Similarly, we have:
[Qwc,i (V̇w,i (k))Qs,i j ] {x(k)}
= x(k) − (1 + εawc,i (V̇w,i (k)))x(k − 1)−
(1 + εas,i j )x(k − 1) + (1 + εas,i j )x(k − 2)+
(1 + εas,i j )εawc,i (V̇w,i (k))x(k − 2).
Thus, we have:
[Qs,i j Qwc,i (V̇w,i (k))] {x(k)} =
[Qwc,i (V̇w,i (k))Qs,i j ] {x(k)} −
(1 + εas,i j )ε awc,i (V̇w,i (k)) − awc,i (V̇w,i (k − 1)) x(k − 2).
Finally, we may write:
[Qs,i j Qwc,i (V̇w,i (k))] =
[Qwc,i (V̇w,i (k))Qs,i j ]−
(1 + εas,i j )ε (1 − q−1 )awc,i (V̇w,i (k)) q−2 .
(20)
B
Proof of Proposition 4.2
A Taylor-series expansion of the continuous-time zone
dynamics leads to the following approximation:
Tr,i (k) ≈
Pr,i (V̇a,i (k − 1)){Tr,i (k)} + ε
∑ a+rs,i j Ts,i j (k − 1)+
j∈Ni
+
εarw,i Tw,i (k − 1) + εara,i (V̇a,i (k − 1))Ta,i
(k − 1)+
εaext,i Q̇ext,i (k − 1),
(21)
plus higher-order terms of ε. Furthermore, note that the finite
response of the separator-dynamics can be approximated by:
n
o
−
+
Ts,i j (k) ≈ Q−1
(22)
s,i j εas,i j Tr,i (k − 1) + εas,i j Tr, j (k − 1)
plus higher-order terms of ε. Given that the family of operators Qs,i j , j ∈ Ni are pairwise commutative by Property 4.1,
if we replace Ts,i j into (21) and we multiply both sides of the
above expression by the composite operator Qs,i j1 · · · Qs,i j|N | ,
i
briefly denoted by ∏ j∈Ni Qs,i j {·}, we have:
Tr,i
h (k) ≈
1−
∏ Qs,i j
j∈Ni
h
Qs,i j
∏
j∈Ni
2
ε
ε
i
Pr,i (V̇a,i (k − 1)){Tr,i (k)} +
∑
a+
rs,i j
∑
a+
rs,i j
j∈Ni
2
j∈Ni
εarw,i
h
∏
j∈Ni
h
ε
in
o
Tr,i (k) +
∏
j∈Ni
εaext,i
h
in
o
∏ Qs,i` a+s,i j Tr,i (k − 2) +
h `6= j
in
o
−
Q
a
T
(k
−
2)
+
r,
j
s,i`
∏
s,i j
`6= j
in
o
Qs,i j Tw,i (k − 1) +
in
o
+
Qs,i j ara,i (V̇a,i (k − 1))Ta,i
(k − 1) +
h
∏ Qs,i j
j∈Ni
i
Q̇ext,i (k − 1) ,
plus higher-order terms of ε. The conclusion follows directly
by using Property 4.3 and expanding the terms of the above
expression.
εaext,i Qwc,i (V̇w,i (k − 2)) Q̇ext,i (k − 1)
plus higher-order terms of ε. Applying to both sides of the
above expression the composite operator Qs,i j1 · · · Qs,i j|N | ,
i
briefly denoted by ∏ j∈Ni Qs,i j {·}, and making use of Property 4.1, we have:
n
o
Tr,i (k) ≈
∏ Qs,i j Qwc,i (V̇w,i (k − 2))
j∈Ni
∏ Qs,i j Qwc,i (V̇w,i (k − 2))
j∈Ni
2
ε
Pr,i (V̇a,i (k − 1)){Tr,i (k)} +
∑ a+rs,i j ∏
Qs,i` [Qs,i j Qwc,i (V̇w,i (k − 2))]Q−1
s,i j
∑ a+rs,i j ∏
−1
Qs,i` [Qs,i j Qwc,i (V̇w,i (k − 2))]Qs,i
j
`∈Ni \ j
n j∈Ni
o
+
as,i j Tr,i (k − 2) +
ε2
`∈Ni \ j
n j∈Ni
o
−
as,i j Tr, j (k − 2) +
n
o
+
ε2 arw,i ∏ Qs,i j aww,i (V̇w,i (k − 2))Tw,i
(k − 2) +
j∈Ni
2
ε arw,i aw,i
∏ Qs,i j {Tr,i (k − 2)} +
j∈Ni
ε
∏ Qs,i j Qwc,i (V̇w,i (k − 2))
j∈Ni
n
o
+
ara,i (V̇a,i (k − 1))Ta,i
(k − 1) +
εaext,i
∏ Qs,i j Qwc,i (V̇w,i (k − 2))
j∈Ni
C
Proof of Proposition 4.3
Under the MI structure, the finite-response approximation of (21) continues to hold for sufficiently small step-size
sampling interval ε. Note that the separator dynamics can
still be approximated according to Equation (22). Furthermore, a Taylor-series expansion of the water dynamics leads
to the following finite-response approximation:
Tw,i (k) ≈ Q−1
wc,i (V̇w,i (k − 1))
n
o
+
εaww,i (V̇w,i (k − 1))Tw,i
(k − 1) + εaw,i Tr,i (k − 1)
plus higher-order terms of ε. According to Property 4.2, we
may perform the substitution:
[Qs,i j Qwc,i (V̇w,i (k − 2))] =
[Qwc,i (V̇w,i (k − 2))Qs,i j ]−
(1 + εas,i j )ε (1 − q−1 )awc,i (V̇w,i (k − 2)) q−2 .
Thus, we may write:
∏
plus higher-order terms of ε. If we replace Ts,i j and Tw,i into
(21) and we apply to both sides of the resulting expression
the operator Qwc,i (V̇w,i (k − 2)), we get
n
o
Qwc,i (V̇w,i (k − 2)) Tr,i (k) ≈
Qwc,i (V̇w,i (k − 2)) Pr,i (V̇a,i (k − 1)){T
n r,i (k)} + o
−1
+
2
ε ∑ ars,i j Qwc,i (V̇w,i (k − 2))Qs,i j a+
s,i j Tr,i (k − 2) +
j∈Ni
ε
2
∑
j∈Ni
n
o
−1
−
a+
Q
(
V̇
(k
−
2))Q
a
T
(k
−
2)
+
r,
j
wc,i
w,i
rs,i j
s,i j
s,i j
+
ε2 arw,i aww,i (V̇w,i (k − 2))Tw,i
(k − 2)+
2
ε arw,i aw,i Tr,i (k − 2)+
n
o
+
εQwc,i (V̇w,i (k − 2)) ara,i (V̇a,i (k − 1))Ta,i
(k − 1) +
Q̇ext,i (k − 1)
j∈Ni
2
ε
n
o
Qs,i j Qwc,i (V̇w,i (k − 2))Qr,i (V̇a,i (k − 1)) Tr,i (k) ≈
∑ a+rs,i j ∏
`∈Ni \ j
j∈Ni
Qs,i` Qwc,i (V̇w,i (k − 2))
n
o
a+
s,i j Tr,i (k − 2) +
ε2
∑ a+rs,i j ∏
`∈Ni \ j
j∈Ni
Qs,i` Qwc,i (V̇w,i (k − 2))
n
o
a−
s,i j Tr, j (k − 2) +
n
o
+
ε2 arw,i ∏ Qs,i j aww,i (V̇w,i (k − 2))Tw,i
(k − 2) +
j∈Ni
ε2 arw,i aw,i
∏ Qs,i j {Tr,i (k − 2)} +
j∈Ni
ε
∏ Qs,i j Qwc,i (V̇w,i (k − 2))
j∈Ni
n
o
+
ara,i (V̇a,i (k − 1))Ta,i
(k − 1) +
εaext,i
∏
j∈Ni
|Ni |+2
Qs,i j Qwc,i (V̇w,i (k − 2)) Q̇ext,i (k − 1) , (23)
Qwc,i (V̇w,i (k − 2))Qr,i (V̇a,i (k − 1)){Tr,i (k)} =
Tr,i (k) − 2Tr,i (k − 1) + Tr,i (k − 2)−
εar,i (V̇a,i (k − 1))Tr,i (k − 1)−
εawc,i (V̇w,i (k − 2))Tr,i (k − 1)+
εawc,i (V̇w,i (k − 2))Tr,i (k − 2)+
εar,i (V̇a,i (k − 2))Tr,i (k − 2)+
ε2 awc,i (V̇w,i (k − 2))ar,i (V̇a,i (k − 2))Tr,i (k − 2).
∏
n
o
Qs,i` Qwc,i (V̇w,i (k − 2)) a+
T
(k
−
2)
r,i
s,i j
`∈Ni \ j
≈
Note that
α5,mV̇w,i (k − m)V̇a,i (k − m)Tr,i (k − m),
m=2
plus higher-order terms of ε. In the following, we will approximate the terms of the above expression, by neglecting
terms of order of ε3 or higher.
|Ni |+2
∑
m=2
∏
j∈Ni
=
|Ni |+2
∑
(4)
+
α1,mV̇w,i (k − m)Tw,i
(k − m),
∏ Qs,i j
j∈Ni
The following also hold:
=
|Ni |+2
∑
m=2
n
o
Qwc,i (V̇w,i (k − 2)) a+
T
(k
−
2)
r,i
s,i j
+
=a+
s,i j Tr,i (k − 2) − (1 + εawc,i (V̇w,i (k − 2)))as,i j Tr,i (k − 3)
n
o
Tr,i+ (k − 2)
(5)
α1,m Tr,i+ (k − m),
∏ Qs,i j Qwc,i (V̇w,i (k − 2))
+
≈a+
s,i j Tr,i (k − 2) − as,i j Tr,i (k − 3)
n
o
+
Qwc,i (V̇w,i (k − 2)) ara,i (V̇a,i (k − 1)Ta,i
(k − 1)
(2)
α1,m Tr,i (k − m),
n
o
+
Qs,i j aww,i (V̇w,i (k − 2))Tw,i
(k − 2)
m=2
(24)
(1)
∑
j∈Ni
=
|Ni |+2
∑
m=1
|Ni |+2
+
=ara,i (V̇a,i (k − 1))Ta,i
(k − 1)−
∑
(5)
+
α1,mV̇a,i (k − m)Ta,i
(k − m)+
(5)
+
α2,mV̇w,i (k − m)V̇a,i (k − m)Ta,i
(k − m),
+
(1 + εawc,i (V̇w,i (k − 2)))ara,i (V̇a,i (k − 2))Ta,i
(k − 2)
m=2
Qwc,i (V̇w,i (k − 2)) Q̇ext,i (k − 1)
∏ Qs,i j Qwc,i (V̇w,i (k − 2))
j∈Ni
=Q̇ext,i (k − 1) − (1 + εawc,i (V̇w,i (k − 2)))Q̇ext,i (k − 2)
where the first expression ignores terms of ε since, in (23),
it multiplies an expression of order of ε2 . Using the above
approximations and Property 4.3, the terms of (23) can be
approximated as follows:
n
o
Tr,i (k)
∏ Qs,i j Qwc,i (V̇w,i (k − 2))Qr,i (V̇a,i (k − 1))
j∈Ni
=Tr,i (k) +
|Ni |+2
∑
m=1
|Ni |+1
∑
m=1
|Ni |+2
∑
m=2
|Ni |+2
∑
m=1
(1)
α1,m Tr,i (k − m)+
(1)
α2,mV̇a,i (k − m)Tr,i (k − m)+
(1)
α3,mV̇w,i (k − 1 − m)Tr,i (k − m)+
(1)
α4,mV̇w,i (k − m)Tr,i (k − m)+
=
|Ni |+2
∑
m=1
|Ni |+2
∑
m=2
n
o
+
ara,i (V̇a,i (k − 1))Ta,i
(k − 1)
Q̇ext,i (k − 1)
(6)
α1,m Q̇ext,i (k − m)+
(6)
α2,mV̇w,i (k − m)Q̇ext,i (k − m).
(×)
for some constant parameters α×,m ∈ R.
It is straightforward to check that using the above approximations, Equation (23) can be written as a linear regression with a nonlinear regression vector of the form depicted
in Equation (14).
| 3cs.SY
|
arXiv:1605.09181v2 [q-fin.PM] 25 Aug 2016
The use of the multi-cumulant tensor analysis
for the algorithmic optimisation of investment
portfolios.
†
Krzysztof Domino [email protected]
†
Institute of Theoretical and Applied Informatics, Polish
Academy of Sciences, Gliwice, Poland
August 26, 2016
Abstract
The cumulant analysis plays an important role in non Gaussian distributed
data analysis. The shares’ prices returns are good example of such data.
The purpose of this research is to develop the cumulant based algorithm and
use it to determine eigenvectors that represent investment portfolios with
low variability. Such algorithm is based on the Alternating Least Square
method and involves the simultaneous minimisation 2’nd – 6’th cumulants of
the multidimensional random variable (percentage shares’ returns of many
companies). Then the algorithm was tested during the recent crash on the
Warsaw Stock Exchange. To determine incoming crash and provide enter
and exit signal for the investment strategy the Hurst exponent was calculated
using the local DFA. It was shown that introduced algorithm is on average
better that benchmark and other portfolio determination methods, but only
within examination window determined by low values of the Hurst exponent.
Remark that the algorithm of is based on cumulant tensors up to the 6’th
order calculated for a multidimensional random variable, what is the novel
idea. It can be expected that the algorithm would be useful in the financial
data analysis on the world wide scale as well as in the analysis of other types
of non Gaussian distributed data.
Keywords cumulant tensors, ALS–class algorithm, Hurst exponent, financial data analysis, stock exchange.
1
1
Introduction
Let us consider the multidimensional frequency distribution of shares’ prices’
percentage returns. The optimization (minimization) of higher cumulants of
this distribution is used to determine investment portfolios, to test if they
are better on average than the benchmark, during the crash. The proposed
procedure is based on [1] and implies the investigation of cumulants tensors
– the n’th cumulant of the multidimensional random variable is represented
by the n–dimensional tensor [1, 2]. For this purpose, I introduce the generalisation of the classical Value at Risk (VaR) procedure [3], where the left
Eigenvector Decomposition (EVD) of the second cumulant (the covariance)
matrix is performed, and the multidimensional normal distribution of financial data is assumed. In classical EVD approach, the portfolio with minimal
variance corresponds to the last eigenvector. However, the classical EVD
method fails to anticipate the risk of investment portfolios since the second
cumulant fails to represent the extreme events, where drops of shares’ prices
values are high and cross–correlated. This happens mainly due to the break
down of the central limit theorem resulting from the time–varying variance of
financial data. The Autoregressive Conditional Heteroskedasticity (ARCH),
that violates both independence and identical distribution assumptions of
the central limit theorem, was recorded for many types of financial data
[4, 5, 6, 7, 8]. Recall also the impact of long range auto–correlations of shares’
returns [9, 10, 11, 12, 13, 14]. It is worth to mention the work [15, 16], where
authors shows that moments or cumulants (of order 6 or 8) may be necessary to account for the severe price fluctuations, that are usually observed
at short time scales (e.g. portfolio rebalanced at a weekly, time scales). In
my research I would examine portfolio rebalanced at the 20 trading days
(approximately monthly scale) as it is often performed in practice in assets
management. To search for the severe price fluctuations I used the Hurst
exponent indicator.
Following this arguments, high cumulants analysis should anticipate extreme events, improving the search for portfolios with low variability. There
are some works implying the use of 2’nd, 3’rd and 4’th cumulant of multivariate shares’ returns [17, 18]. In this research I use the 5’th and the 6’th
cumulant as well, what is a new approach for multivariate shares’ returns.
In general the proposed algorithm is based on the High Order Singular Value
Decomposition (HOSVD) and Alternating Least Square (ALS) procedure [2].
To compare the proposed method with others (such as EVD), the author, for
each method, creates the family of investment portfolios which are supposed
to be safer than a benchmark. Then portfolios are compared using the result function that is an average percentage change of portfolios’ values – an
2
average portfolio results. Other result functions are also discussed:
1. a mode of percentage change of portfolios’ values,
2. a maximal loss / minimal gain – the result of the “worst portfolio”,
3. a minimal loss/ maximal gain – the result of the “best portfolio”.
The major motivation for this research is to introduce the automation
method of analysis of data that are not Gaussian distributed. Good example of such data are financial data, especially during the rupture and crash
period. It is why, I focus in this work, on the financial data analysis. To
determine the rupture and crash period and introduce the enter and exit signal of an investment strategy, I use the Hurst exponent indicator calculated
for the WIG20 index, using the local DFA. This paper give some additional
incentive for the development of cumulants tensors calculation method at
low computational complexity. Afterwards the multi–cumulant analysis may
be applied for large financial data sets and tested against many crashes on
many markets. Additionally the method may be used to analyse other (non–
financial) data that are not Gaussian distributed.
2
The classical approach, the covariance matrix EVD
Let us take the M –dimensional random variable of size T , X ∈ R(T ×M ) ,
being the percentage returns of M shares. Its marginal variables are Xi , and
values are xt,i :
xt=1,1 · · · xt=1,M
..
..
X = [X1 , . . . , Xi , . . . , XM ] = ...
(1)
.
.
.
xt=T,1 · · · xt=T,M
An unbiased estimator of variance of the i’th marginal random variable (Xi )
is:
T
1 X
(xt,i − Xi )2 ,
(2)
σi2 =
T − 1 t=1
and an unbiased estimator of covariance between (Xi ) and (Xj ) is:
T
covi,j
1 X
=
(xt,i − Xi )(xt,j − Xj ).
T − 1 t=1
3
(3)
The variance and the covariance can be represented by the M ×M symmetric
covariance matrix, called also the second cumulant matrix – C2 (notice σi2 =
covi,i ):
σ12
cov1,2 · · · cov1,L
cov2,1
· · · cov2,L
σ22
C2 =
(4)
.
..
..
..
.
.
.
.
.
.
covL,1 covL,2 · · ·
σL2
Definition 2.1. The Eigenvalue Decomposition – EVD. Consider the covariance (second cumulant) symmetric matrix. The matrix can be diagonalized
in the following way:
C2 = V ΣV | ,
(5)
where Σ = V | C2 V is the diagonal matrix with diagonal values σi02 =
(V | C2 V )ii and V is unitary M × M factors matrix, such that σi02 are sorted
in descending order:
σ102 0 · · · 0
0 σ 02 · · · 0
2
(6)
Σ = ..
.. . .
.. .
.
.
.
.
02
0
0 · · · σM
The i’th column of V is the eigenvector that corresponds with the eigenvalue
σi02 . Rows in the i’th column of V are factors that give the linear combination
of marginal random variables with the combination’s variance σi02 . The last
eigenvector would give the linear combination of marginal random variables
02
.
with the smallest combination’s variance – σM
The classical EVD procedure has been often used in the portfolio risk
determination. However, it requires the multidimensional Gaussian distribution of shares’ returns, where all information about the variability of the
frequency distribution is stored in the covariance matrix. As mentioned before the financial data (shares’ returns) are not Gaussian distributed and the
classical EVD procedure has often failed in the investment portfolio’s risk
determination [19]. It is why the author proposes to extend the classical
EVD procedure by taking into consideration also cumulants of order higher
than 2 – the higher cumulants.
3
Cumulants
Let us consider the M dimensional random variable X = [X1 , . . . , XM ]. The
n’th cumulant Cn of such variable is the n–mode tensor [2], with elements
4
κα1 ,...,αn (X) [20, 21]:
κα1 ,...,αn (X) =
∂n
log (E (exp(τ · X| )))
∂τα1 , ∂τα2 , . . . , ∂ταn
τ =0
.
(7)
where τ is the argument vector τ = [τ1 , · · · , τi , · · · τM ], and E() is the expected
|
{z
}
M
value operator. Formulas used to calculate cumulants up to 4’th order are
well known [20, 21]. The author has calculated 5’th and 6’th cumulants by
the direct use of (7). Here analysed data were substituted for the random
variable X ∈ R(T ×M ) , and computer differentiations were performed at point
τ = [τ1 , · · · τM ] = 0, using ForwardDiff and DualNumbers library in Julia
| {z }
M
programming [22].
3.1
The multi–cumulant decomposition.
To investigate the financial data the author takes many cumulant tensors
C2 , . . . , Cn , where n = 4 or n = 6. The calculation of cumulants of order
n > 6 might require larger data series, but non–stationary of financial data
[10] makes the investigation of long time series less adequate than shorter
data series. To achieve the factor matrix V , the author proposes the following
ALS–class algorithm, where the search for the local maximum of the function
Φ(V ) is performed [23, 24]. Following the maximisation procedure which can
not be solved precisely, the author will find the local maximum using the
iteration procedure [24] and show that the results are meaningful.
Definition 3.1. The Φ(V ) function. Consider the i’th core–tensor Ti that
is the contraction of Ci tensor and i factor matrices V :
X
(8)
(Ci )j1 ,···,ji Vj1 l1 · · · Vji li .
(Ti )l1 ,···,li =
j1 ,···,ji
The ALS procedure proposed in [1, 24] refers to the search for the common
factor matrix V that maximise Φ4 (V ).
4
X1
1
Φ4 (V ) = ||V | C2 V ||2 +
||Ti ||2 .
2!
i!
i=3
(9)
The author proposes to extend the analysis up to the 6’th cumulant which
are more sensitive to extreme “tail events”. Hence the author defines Φ6 (V ):
6
Φ6 (V ) =
X1
1
||V | C2 V ||2 +
||Ti ||2 .
2!
i!
i=3
5
(10)
To find the common factor matrix V , the ALS–based algorithm is proposed by author and presented at subsection (3.2). The idea of the algorithm
is based on the algorithm proposed in [23] where the iteration procedure was
used for the search for the local maximum of the following function:
Φ0 (V ) = ||V | C2 V ||2 + αn ||Tn ||2 .
(11)
The proposed algorithm works for the general case (any Φn (V )), but computations were performed for n = 4 and n = 6. Racall that ALS algorithms
move information into the upper left corner of the core–tensor and order the
information in the sense of the Frobenius Norm. Take the linear transformation of analysed data X:
Y = XV,
(12)
where Y = [Y1 , . . . , Yj , . . . , YM ]. Here Yj represents percentage returns of the
j’th portfolio. Elements of Y are:
yt,j =
M
X
xt,i Vi,j
(13)
i=1
The rear columns of the factor matrix would give the investment portfolio
with little variability.
3.2
The algorithm.
The algorithm used to determine the factor matrix V given cumulant symmetric tensors C2 · · · Cn , it is a general algorithm and work for each n ≥ 3.
Let Ci(1) be the unfold of the tensor Ci in the first mode [2]. The first factor
matrix anzatz is computed as a matrix that columns are left eigenvectors of
the following matrix:
Ci(1)
Cn(1)
C2
···
···
(14)
2!
i!
n!
At k’th interaction, we have the (Vk−1 ) factor matrix. Now the following
procedure is performed. The contraction of the Ci tensor (matrix) and i − 1
factor matrices (Vk−1 ) is performed:
X
(Si )j1 ,l2 ···,li =
(Ci )j1 ,j2 ···,ji (Vk−1 )j2 l2 · · · (Vk−1 )ji li .
(15)
j2 ,···,ji
To compute Vk we takes left eigenvectors of the following matrix:
Si(1)
Sn(1)
S2
···
···
2!
i!
n!
The procedure is repeated to satisfaction the stop condition.
6
(16)
4
The investigation of financial data.
The cumulant analysis was performed in the optimal portfolios searching
problem. Let us consider the price of a i’th share at time t – pt,i . Its
percentage return is
pt,i − p(t−1),i
xt,i =
· 100%.
(17)
p(t−1),i
In our case t numerates trading days (the analyse of daily returns was performed) and pt,i the closing price of i’th share the given trading day numbered
by t. Next the multidimensional random variable X of percentage returns
is constructed. To construct investment portfolios we use the factor matrix
V . The j’th portfolio returns are one dimensional random variable Yj with
elements yt,j .
The naive method of factor matrix determination uses the Eigenvalue
Decomposition (EVD) of the covariance matrix [3]. This procedure is not
fully adequate since shares returns are not Gaussian distributed, especially
the rupture and crisis period [9, 10, 11, 12] – importantly such period can be
predicted by the use of the Hurst exponent. To anticipate higher cumulants
of shares returns as well, the author proposes to determine the factor matrix
V by searching for the local maximum of the Φ4 (V ) function as well as
Φ6 (V ) function – using cumulant tensors up to the 6’th order, what is a
new approach. The proposed Φ6 (V ) method is used to chose portfolios with
returns that have low absolute values of high cumulants. Hence the method
is supposed to work well where the portfolio’s variability is a disadvantage. It
happens during the crash of the financial market, hence the author tests the
method during the last rupture and crisis on the Warsaw Stock Exchange.
4.1
The data analysis.
The author has examined M = 10 dimensional random variable Tab. (1),
being daily percentage returns of the shares of 10 most liquid companies
from the WIG20 index at the time 12.05.2010 – 04.08.2016 (the WIG20 index
includes 20 most liquid companies traded on the Warsaw Stock Exchange).
Recent composition of the WIG20 index is presented in Fig. (1) The WIG20
index reached maximum at 14.05.2015 and then has fallen rapidly – the crash
has occurred. To introduce the signal of incoming crash, the Hurst exponent
was calculated for the WIG20 index using the local Detrended Fluctuation
Analysis (DFA) [10, 13].
7
WIG20 index
2800
2600
2400
2200
2000
1800
Hurst Exponent
1600
Aug 20
12
1
Feb 20
3
Aug 20
13
1
Feb 20
4
Aug 20
14
1
Feb 20
5
Aug 20
15
1
Feb 20
6
Aug 20
1
Feb 20
6
16
Aug 20
16
0.50
Exit signal
0.45
0.40
Enter signal
12
Aug 20
1
Feb 20
3
13
Aug 20
1
Feb 20
4
14
Aug 20
1
Feb 20
5
15
Aug 20
Figure 1: WIG20 index and the Hurst exponent.
i
company
1
PKOBP
2
PZU
3
PEKAO
4 PKNORLEN
5
PGE
6
KGHM
7
BZWBK
8
LPP
9
PGNIG
10
MBANK
contribution
to WIG20 %
14.64
14.04
11.65
8.45
7.52
7.14
5.21
4.77
3.55
3.00
contribution to
benchmark % (BPi )
18.31
17.55
14.57
10.57
9.40
8.93
6.51
5.96
4.43
3.75
Table 1: the 10 most liquid companies of the WIG20 index, their value contribution to the WIG20 index (at 20.03.2015) and as their value contribution
to proposed benchmark portfolio.
8
4.1.1
The Hurst exponent.
To determine the rupture and crisis period of the stock exchange, where the
examined investment strategy was tested the Hurst exponent was calculated
using the local DFA. The parameters for DFA were the same as in [25]: 500
days long observation window was used to examine past closing value of the
WIG20 index. Having the Hurst exponent, I introduce the signals of entry
and exit for proposed investment strategy. Recall that in [10] the Hurst
exponent was calculated using the local DFA for the index of Polish Stock
Exchange, and it was shown that before a crash (near a rupture point), the
Hurst Exponent has minima . 0.4. Hence the entry threshold value was
chosen as Hentry = 0.4. The exit threshold was chosen as Hexit = 0.425 –
data with high negative auto–correlation was chosen for a test.
Regarding the recent crash the entry signal occurred at 19.12.2014 and
the exit signal at 10.09.2015. To examine the algorithm, I introduce the
20 trading days (approx. 1 months) long investment windows – as it is
performed in practice assets management. First window starts a day after
the enter signal – 22.12.2014, and there are 9 windows within a test period
22.12.2014 – 27.01.2015, 27.01.2015 – 24.02.2015, 24.02.2015 – 24.03.2015,
24.03.2015 – 23.04.2015, 23.04.2015 – 22.05.2015, 22.05.2015 – 22.06.2015,
22.06.2015 – 20.07.2015, 20.07.2015 – 17.08.2015, 17.08.2015 – 14.09.2015
(the last window ends just after exit point). For each window, cumulants
are calculated using a test series of length T = 1100, that ends just before
the examination window. Next investment returns are analysed for data in
given window – the testing set. In next subsections the analysis is discussed
in details for the 7’th window of 22.06.2015 – 20.07.2015. Then the analysis
results are presented for other windows.
4.1.2
Optimal portfolios determination – training.
Let us discuss in details the procedure for the exemplary window of 22.06.2015
– 20.07.2015. Given the training set, the factor matrix is determined using
different methods, such as EVD, Φ4 (V ) and Φ6 (V ). Here also the Independent Component Analysis (ICA) was used for more general comparison. The
Φ4 (V ) method requires the calculation of 3’rd and 4’th cumulants. For Φ6 (V )
also 5’th and 6’th cumulant tensors are required, which were calculated by
the direct use of Eq. (7). Given Φ4 (V ) and Φ6 (V ) the algorithm introduced
in subsection (3.2) was used for the factor matrix V determination.
In Fig. (2) some cumulant value of the one dimensional random
variable,
PM
that is the j’th investment portfolio Yj (with elements yt,i = i=1 xt,i Vi,j ) are
presented for different methods of the factor matrix determination. Generally
9
large cumulants values were stored in first portfolios where j = 1, 2, · · ·. For
further investigation I took the 5 rear portfolios, where j ∈ [5, 10] as those
that have low cumulants’ absolute values.
4.1.3
Testing optimal portfolios.
After the training (the determination of V ) has been completed, the testing
of portfolios is performed. The factor matrices (V ) columns contain both
positive and negative values, the later corresponds to the negative value of
shares in the portfolio – the short sale. To diminish the use of the sort
sale, the test portfolios were compared with the benchmark portfolio. Shares
values contributions in benchmark portfolio – BPi are given in Tab. (1). In
proposed test portfolios the value contribution of the i’th share in the j’th
portfolio would be:
αBPi + Vi,j
,
(18)
T Vi,j = P10
(αBP
+
V
)
i
i,j
i=1
the α = 7 was taken, to make cases of the short sale rare. For testing, shares
prices of companies, see Tab. (1) were taken. Testing set is represented by:
pt0 ,i , where t0 is time in the testing window. The percentage return of j’th
portfolio after L trading days is:
P
p
P10
(t0 =L+1,i)
10
− 10
X
i=1 T Vi,j
i=1 T Vi,j
p(t0 =L+1,i)
p(t0 =1,i)
=
T Vi,j
.
P rj (L) =
P10
p(t0 =1,i)
i=1 T Vi,j
i=1
(19)
In Fig. (3), returns after 10 and 20 trading days are presented. Remark, in
this research transaction costs were not taken into account. The benchmark
portfolio contributions can be reproduced by simply substituting ∀i,j Vi,j = 0
to Eq. (18).
4.2
Discussion.
Analysing Fig. (3), one can see that the Φ6 (V ) method gives in both cases
3 portfolios that are better than the benchmark and 2 that are as good as
benchmark. In the reminding part of the paper, I discuss the statistics of
returns of such 5 portfolios.
One can also conclude, that each method of factor matrix determination
(Φ6 (V ), Φ4 (V ), EVD, ICA) produces the worst portfolio
min P rj (L),
j∈[6,10]
10
(20)
cumulant’s value
14
EV D
Φ4 (V )
12
Φ6 (V )
10
8
4
−1.0
2
cumulant’s value
8
EV D
Φ4 (V )
Φ6 (V )
1 2 3 4 5 6 7 8 9 10
number of portfolio − j
(a) second cumulant
10
0.0
−0.5
6
0
0.5
−1.5
1 2 3 4 5 6 7 8 9 10
number of portfolio − j
(b) third cumulant (normalized)
EV D
Φ4 (V )
Φ6 (V )
500
400
EV D
Φ4 (V )
Φ6 (V )
300
6
200
4
100
2
0
0
1 2 3 4 5 6 7 8 9 10
number of portfolio − j
(c) fourth cumulant (normalized)
−100
1 2 3 4 5 6 7 8 9 10
number of portfolio − j
(d) sixth cumulant (normalized)
Figure 2: Values of cumulants for portfolios, the larger j the portfolio is
supposed to be less “variable”.
11
percentage returns
−0.005
−0.010
−0.015
−0.020
−0.025
−0.030
−0.035
−0.040
−0.045
−0.050
−0.01
−0.02
−0.03
−0.04
−0.05
−0.06
−0.07
6
7
8
9
10
number of portfolio − j
−0.08
(a) After 10 trading days
Bench.
Φ6 (V )
ICA
EV D
Φ4 (V )
6
7
8
9
10
number of portfolio − j
(b) After 20 trading days
Figure 3: Returns of 5 portfolios, investments window 22.06.2015 –
20.07.2015.
which return is minimal and often smaller than benchmark’s return. Those
minimum of portfolios’ returns are presented in Fig. (4a). Analysing minimum of portfolios’ returns one can conclude that out of all methods (Φ6 (V ),
Φ4 (V ), EVD, ICA) the Φ6 (V ) method gives smallest loss – its worst portfolio
is almost as good as benchmark. The worst results gives the ICA method,
this is due to large variability of returns – see Fig. (3), such method is not
desirable during a crisis. Similarly best portfolios can found:
max P rj (L),
(21)
j∈[6,10]
Their results are presented in Fig. (4b). Best results presents the ICA, however this method is not safe. Next best are Φ6 (V ) and Φ4 (V ).
It is worth checking now, which portfolio determination method is best
on average. In Fig. (4c), mean values of portfolios’ returns are presented:
10
1X
P rj (L).
5 j=6
(22)
Remark that all methods but Φ6 (V ) give an average return similar to or
worse than the benchmark. It is a worthy result, since it is hard to beat
the benchmark on average. To examine a typical portfolio, the mode of
portfolios’ returns can be mentioned as well – see Fig. (4d), here Φ6 (V ) gives
results, better than other methods, and slightly better the the benchmark.
12
portfolios′ returns
0.01
0.00
−0.01
−0.02
−0.03
−0.04
−0.05
−0.06
−0.07
−0.08
0.01
0.00
−0.01
−0.02
−0.03
ICA
EV D
Φ4 (V )
−0.04
−0.05
Φ6 (V )
Bench.
2 4 6 8 10 12 14 16 18 20
lenght of test series − L
(a) minimum of portfolio’s returns
portfolios′ returns
2 4 6 8 10 12 14 16 18 20
lenght of test series − L
(b) maximum of portfolio’s returns
0.01
0.00
−0.01
−0.02
−0.03
−0.04
−0.05
−0.06
−0.06
2 4 6 8 10 12 14 16 18 20
lenght of test series − L
(c) mean of portfolio’s returns
0.01
0.00
−0.01
−0.02
−0.03
−0.04
−0.05
−0.06
−0.07
−0.08
2 4 6 8 10 12 14 16 18 20
lenght of test series − L
(d) mode of portfolio’s returns
Figure 4: Statistics of investment returns, 7’th window of 22.06.2015 –
20.07.2015.
13
Concluding, statistics of the φ6 (V ) method are better than other methods
and the benchmark. Results of other 20 days observation windows within
the observation period determined by the Hurst exponent and outside it are
discussed in next subsection.
4.3
All observation windows.
The analysis was performed for following observation windows: 22.12.2014
– 27.01.2015, 27.01.2015 – 24.02.2015, 24.02.2015 – 24.03.2015, 24.03.2015
– 23.04.2015, 23.04.2015 – 22.05.2015, 22.05.2015 – 22.06.2015, 22.06.2015
– 20.07.2015, 20.07.2015 – 17.08.2015, 17.08.2015 – 14.09.2015. The first
window starts a trading day after the enter signal recorded at 19.12.2014.
The WIG20 index increased in first for windows, the maximum appeared in
the 5’th window where the crisis started, the last (9’th window) ends just
after the exit signal recorded at 10.09.2015. Windows 5 – 9 are crisis windows.
Since I an interested in the investment strategy outcome between the enter
and the exit signal, I present the cumulative results of investment that starts
a trading day after the enter signal and ends just after exit signal. For
each window factor matrices are calculated separately, investment is made
at the first point in a window, at the last point of the window shares are
sold and the mean of returns of 5 portfolios is calculated. Cumulative of
such mean returns are presented in Fig. (5a). In Fig. (5b) the cumulative
results are presented for crisis portfolios 5 - 9, here investment starts at
23.04.2015. In Fig. (6) similar results are presented, but now mode of returns
of 5 portfolios is calculated in each window and the cumulative results are
presented. Analysing Fig. (5, 6) one can conclude that the Φ6 (V ) method
on average gives best results at the exit point and during the crisis.
In Fig. (7a) cumulative results are presented, if in each window the worst
portfolio was chosen (unlucky choice) – there Φ6 (V ) method is worse than a
benchmark, but slightly better than other methods. In Fig. (7b) cumulative
results are presented, if in each window the best portfolio was chosen (lucky
choice) – there Φ6 (V ) method is better than all other methods apart from
ICA. However the ICA produces also very bad portfolios (worst minimum),
and hence is not adequate for a crisis.
To test a method a bit more, I introduced observation windows after
the exit signal, and number them as 10’th to 15’th, the cumulative results
of means and modes of portfolio returns are presented in (8). Investment
starts at 14.09.2015 and investment windows are 14.09.2015 – 12.10.2015,
12.10.2015 – 09.11.2015, 09.11.2015 – 08.12.2015, 08.12.2015 – 12.01.2016
and 12.01.2016 – 09.02.2016. It can be concluded that beyond the exit signal
the Φ6 (V ) method gives results similar to other methods and the benchmark.
14
cumulative returns
0.15
0.10
0.05
0.00
−0.05
−0.10
−0.15
ICA
EV D
Φ4 (V )
Φ6 (V )
Bench.
1 2 3 4 5 6 7 8 9
number of invetment window
0.02
0.00
−0.02
−0.04
−0.06
−0.08
−0.10
−0.12
−0.14
−0.16
5
6
7
8
9
number of invetment window
(a) whole investment 22.12.2014 – 14.09.2015 (b) crisis windows 23.04.2015 – 14.09.2015
cumulative returns
Figure 5: Cumulative of means of portfolios returns.
0.15
0.05
0.10
0.00
0.05
−0.05
0.00
−0.05
−0.10
−0.15
−0.10
ICA
EV D
Φ4 (V )
−0.15
Φ6 (V )
Bench.
1 2 3 4 5 6 7 8 9
number of invetment window
−0.20
5
6
7
8
9
number of invetment window
(a) whole investment 22.12.2014 – 14.09.2015 (b) crisis windows 23.04.2015 – 14.09.2015
Figure 6: Cumulative of modes of portfolios returns.
15
0.05
cumulative returns
0.05
0.00
0.00
−0.05
−0.05
−0.10
−0.15
−0.20
−0.25
ICA
EV D
Φ4 (V )
−0.10
Φ6 (V )
Bench.
5
6
7
8
9
number of invetment window
(a) Cumulatives of minima of returns
−0.15
5
6
7
8
9
number of invetment window
(b) Cumulatives of maxima of returns
Figure 7: Best and worst cumulative result at crisis windows 23.04.2015 –
14.09.2015.
Hence the use of the Hurst exponent to determine the proper enter and exit
signal appears to be crucial.
5
Conclusions
The author has used the multi–cumulant tensor analysis to analyse financial
data and determine optimal investment portfolio with low absolute values
of cumulants of their percentage returns. For this purpose, the author has
analysed daily returns of shares traded on the Warsaw Stock Exchange to
determine the factor matrix that represents such portfolios and test them
during the recent rupture and crash period on the Warsaw Stock Exchange.
The main result of this work is the introduction of the algorithm that
uses 2’nd – 6’th cumulant tensors to analyse multivariate financial data and
determine the investment portfolios that have low variability (low cumulants’
absolute values). The Hurst exponent, calculated by the local DFA for the
WIG20 index, indicates the auto–correlation phase on the stock market (the
rupture period and the early stage of the crisis). At this phase, the introduced method is on average better than the benchmark and other tested
methods. Importantly the Hurst exponent condition appears to be necessary
to achieve this result. The examination of the method can be extended in
further research, e.g. the algorithm can be tested on many stock exchanges.
The algorithm can also be used to analyse other (non–financial) data that
16
cumulative returns
0.00
−0.05
0.00
EV D
Φ4 (V )
−0.05
Φ6 (V )
Bench.
−0.10
−0.10
−0.15
−0.15
−0.20
−0.20
−0.25
10
11
12
13
14
number of invetment window
−0.25
10
11
12
13
14
number of invetment window
(a) Cumulatives of means of returns
(b) Cumulatives of modes of returns
Figure 8: Statistics for investment after exit signal, 14.09.2015 – 09.02.2016
are non–Gaussian distributed.
Acknowledgements
The research was partially financed by the National Science Centre, Poland
- project number 2014/15/B/ST6/05204
References
[1] J. Morton, “Algebraic models for multilinear dependence,” 2009.
[2] T. G. Kolda and B. W. Bader, “Tensor decompositions and applications,”
SIAM review, vol. 51, no. 3, pp. 455–500, 2009.
[3] P. Best, Implementing value at risk. John Wiley & Sons, 2000.
[4] V. Akgiray, “Conditional heteroscedasticity in time series of stock returns: Evidence and forecasts,” Journal of business, pp. 55–80, 1989.
[5] T. Bollerslev, “A conditionally heteroskedastic time series model for
speculative prices and rates of return,” The review of economics and
statistics, pp. 542–547, 1987.
17
[6] T. Bollerslev, R. Y. Chou, and K. F. Kroner, “Arch modeling in finance:
A review of the theory and empirical evidence,” Journal of econometrics,
vol. 52, no. 1-2, pp. 5–59, 1992.
[7] R. F. Engle and K. F. Kroner, “Multivariate simultaneous generalized
arch,” Econometric theory, vol. 11, no. 01, pp. 122–150, 1995.
[8] G. W. Schwert and P. J. Seguin, “Heteroskedasticity in stock returns,”
the Journal of Finance, vol. 45, no. 4, pp. 1129–1155, 1990.
[9] B. B. Mandelbrot, The variation of certain speculative prices. Springer,
1997.
[10] D. Grech and G. Pamuła, “The local hurst exponent of the financial time
series in the vicinity of crashes on the polish stock exchange market,”
Physica A: Statistical Mechanics and its Applications, vol. 387, no. 16,
pp. 4299–4308, 2008.
[11] Ł. Czarnecki, D. Grech, and G. Pamuła, “Comparison study of global
and local approaches describing critical phenomena on the polish stock
exchange market,” Physica A: Statistical Mechanics and its Applications,
vol. 387, no. 27, pp. 6801–6811, 2008.
[12] G. L. Vasconcelos, “A guided walk down wall street: an introduction to
econophysics,” Brazilian Journal of Physics, vol. 34, no. 3B, pp. 1039–
1065, 2004.
[13] K. Domino, “The use of the hurst exponent to predict changes in trends
on the warsaw stock exchange,” Physica A: Statistical Mechanics and
its Applications, vol. 390, no. 1, pp. 98–109, 2011.
[14] K. Domino, “The use of the hurst exponent to investigate the global
maximum of the warsaw stock exchange wig20 index,” Physica A: Statistical Mechanics and its Applications, vol. 391, no. 1, pp. 156–169,
2012.
[15] Y. Malevergne and D. Sornette, “Multi-moments method for portfolio
management: Generalized capital asset pricing model in homogeneous
and heterogeneous markets,” Available at SSRN 319544, 2002.
[16] M. Rubinstein, E. Jurczenko, and B. Maillet, Multi-moment asset allocation and pricing models, vol. 399. John Wiley & Sons, 2006.
[17] J. C. Arismendi and H. Kimura, “Monte carlo approximate tensor moment simulations,” Available at SSRN 2491639, 2014.
18
[18] E. Jondeau, E. Jurczenko, and M. Rockinger, “Moment component analysis: An illustration with international stock markets,” Swiss Finance
Institute Research Paper, no. 10-43, 2015.
[19] U. Cherubini, E. Luciano, and W. Vecchiato, Copula methods in finance.
John Wiley & Sons, 2004.
[20] M. G. Kendall et al., “The advanced theory of statistics.,” The advanced
theory of statistics., no. 2nd Ed, 1946.
[21] E. Lukacs, “Characteristics functions,” Griffin, London, 1970.
[22] J. Revels, T. Papamarkou, and M. Lubin, “Forwarddiff.jl,” JuliaDiff/ForwardDiff.jl, 2015.
[23] L. De Lathauwer and J. Vandewalle, “Dimensionality reduction in
higher-order signal processing and rank-(r 1, r 2,. . . , r n) reduction
in multilinear algebra,” Linear Algebra and its Applications, vol. 391,
pp. 31–55, 2004.
[24] B. Savas and L.-H. Lim, “Quasi-newton methods on grassmannians
and multilinear approximations of tensors,” SIAM Journal on Scientific
Computing, vol. 32, no. 6, pp. 3352–3393, 2010.
[25] K. Domino and T. Błachowicz, “The use of copula functions for modeling
the risk of investment in shares traded on the warsaw stock exchange,”
Physica A: Statistical Mechanics and its Applications, vol. 413, pp. 77–
85, 2014.
19
| 5cs.CE
|
arXiv:1606.08970v3 [math.GR] 21 Feb 2018
THE ROTATING NORMAL FORM OF BRAIDS IS
REGULAR
JEAN FROMENTIN
Abstract. Defined on Birman–Ko–Lee monoids, the rotating normal
form has strong connections with the Dehornoy’s braid ordering. It can
be seen as a process for selecting between all the representative words
of a Birman–Ko–Lee braid a particular one, called rotating word. In
this paper we construct, for all n > 2, a finite-state automaton which
recognizes rotating words on n strands, proving that the rotating normal
form is regular. As a consequence we obtain the regularity of a σ-definite
normal form defined on the whole braid group.
1. Introduction
Originally, the group Bn of n-strand braids was defined as the group of
isotopy classes of n-strand geometric braids. An algebraic presentation of Bn
was given by E. Artin in [1]:
σi σj = σj σi
for |i − j| > 2
σ1 , ... , σn−1
.
(1)
σi σj σi = σj σi σj for |i − j| = 1
An n-strand braid is an equivalence class consisting of (infinitely many)
words in the letters σi±1 . The standard correspondence between elements of
the presented group Bn and geometric braids consists in using σi as a code
for the geometric braid where only the ith and the (i + 1)st strands cross,
with the strand originally at position (i + 1) in front of the other.
i+1
i+1
i
i
σi
σi−1
σ3−1 σ2−1 σ1−1 σ2−1 σ3
Figure 1. Interpretation of a word in the letters σi±1 as a
geometric braid diagram.
2010 Mathematics Subject Classification. 20F36, 20M35, 20F10.
Key words and phrases. dual braid monoid, rotating normal form, regular language,
automata.
1
2
JEAN FROMENTIN
In 1998, J.S. Birman, K.H. Ko, and S.J. Lee [3] introduced and investi∗
gated for each n a submonoid B+
n of Bn , which is known as the Birman–
Ko–Lee monoid. The name dual braid monoid was subsequently proposed
because several numerical parameters obtain symmetric values when they
+∗
are evaluated on the positive braid monoid B+
n and on Bn , a correspondence that was extended to the more general context of Artin–Tits groups
∗
by D. Bessis [2] in 2003. The dual braid monoid B+
n is the submonoid of Bn
generated by the braids ai,j with 1 6 i < j 6 n, where ai,j is defined by
ai,j = σi · · · σj −1 σj σj−−11 · · · σi−1 . In geometrical terms, the braid ai,j corresponds to a crossing of the ith and jth strands, both passing behind the
(possible) intermediate strands.
4
≈
1
Figure 2. In the geometric braid a1,4 , the strands 1 and 4
cross under the strands 2 and 3.
Remark. In [3], the braid ai,j is defined to be σj −1 · · · σi+1 σi σi−+11 · · · σj−−11 ,
corresponding to a crossing of the ith and jth strands, both passing in
front of the (possible) intermediate strands. The two definitions lead to
isomorphic monoids. Our choice is this of [13] and has connections with
+∗
∗
Dehornoy’s braid ordering: B+
n−1 is an initial segment of Bn .
By definition, σi equals ai,i+1 and, therefore, the positive braid monoid B+
n
∗
is included in the monoid B+
n , a proper inclusion for n > 3 since the braid
a1,3 does not belong to the monoid B+
3.
For n > 2, we denote by An the set {ap,q | 1 6 p < q 6 n}. If p and q are
two integers of N satisfying p 6 q, we denote by [p, q] the interval {p, ..., q}
of N. The interval [p, q] is said to be nested in the interval [r, s] if the
∗
relation r < p < q < s holds. The following presentation of the monoid B+
n
is given in [3].
∗
Proposition 1.1. The monoid B+
n is presented by generators An and relations:
ap,q ar,s = ar,s ap,q
for [p, q] and [r, s] disjoint or nested,
ap,q aq,r = aq,r ap,r = ap,r ap,q
for 1 6 p < q < r 6 n.
(2)
(3)
∗
Since [2] and [3] it is known that the dual braid monoid B+
n admits
a Garside structure whose simple elements are in bijection with the noncrossing partitions of n. In particular, there exists a normal form associated
with this Garside structure, the so-called greedy normal form.
∗
The rotating normal form is another normal form on B+
n , and was in∗
troduced in [12] and [13]. Roughly speaking, for every braid β ∈ B+
n the
THE ROTATING NORMAL FORM OF BRAIDS IS REGULAR
3
rotating normal form picks up a unique representative word on the letters An
among all of these representing β. It can be seen as a map rn from the dual
∗
∗
braid monoid B+
n to the set of words An . The language of all n-rotating
∗
words, denoted by Rn , is then the image of B+
n under the map rn . We recall
that the rotating normal form is a dual version of the alternating normal
form introduced by P. Dehornoy in [8] and itself building on S. Burckel’s
normal form defined in [4]
The aim of this paper is to construct for all n > 2 an explicit finite-state
automaton which recognizes the language Rn , implying that the language of
n-rotating words is regular. Following [13] and [14] we can define, from the
rotating normal form, a σ-definite normal form defined on the whole braid
group. The corresponding language is then proven to be regular.
The paper is divided as follows. In section 2 we recall briefly the construction of the rotating normal form and its useful already known properties.
In the third section we describe the left reversing process on dual braid
monoids. In section 4 we give a syntactical characterization of n-rotating
normal words. In the fifth section we construct, for each n > 2, a finite-state
automaton which recognizes the language Rn of n-rotating normal words.
Section 5 is devoted to establish the regularity of the σ-definite normal form.
In last section we prove that the rotating normal form is not right-automatic.
2. The rotating normal form
The main ingredient used to define the rotating normal form is the Gar−1
∗
side automorphism φn of B+
n defined by φn (β) = δn β δn where δn =
+∗
a1,2 a2,3 ...an−1,n is the Garside braid of Bn . In terms of Birman–Ko–Lee
generators, the map φn can be defined by:
(
ap+1,q+1 for q 6 n − 1,
φn (ap,q ) =
(4)
a1,p+1
for q = n.
Geometrically, φn should be viewed as a rotation, which makes sense provided braid diagrams are drawn on a cylinder rather than on a plane rectangle.
6
1
3
2
3
4
4
3
4
φ6
−
7
5 → 2
2
5
1
1
6
1
Figure 3. Rolling up the usual braid diagram helps us to
visualize the symmetries of the braids ap,q . On the resulting
cylinder, ap,q naturally corresponds to the chord connecting
vertices p and q. With this representation, φn acts as a clockwise rotation of the marked circles by 2π/n.
6
4
JEAN FROMENTIN
∗
For β and γ in B+
n , we say that γ is a right divisor of β, if there exists a
+∗
′
dual braid β of Bn satisfying β = β ′ γ.
∗
Definition 2.1. For n > 3 and β a braid of B+
n , the maximal braid β1
+∗
∗
lying in Bn−1 that right-divides the braid β is called the B+
n−1 -tail of β.
∗
Using basic Garside properties of the monoid B+
n we obtain the following
∗
result (Proposition 2.5 of [13]) which allows us to express each braid of B+
n
+∗
as a unique finite sequence of braids lying in Bn−1 .
∗
Proposition 2.2. Assume n > 3. For each non-trivial braid β of B+
n there
+∗
exists a unique sequence (βb , ..., β1 ) of braids of Bn−1 satisfying βb 6= 1 and
β = φnb−1 (βb ) · ... · φn (β2 ) · β1 ,
for each k > 1, the B
+∗
n −1
-tail of
φnb−k (βb ) ·
(5)
... · φn (βk+1 ) is trivial.
(6)
Under the above hypotheses, the sequence (βb , ..., β1 ) is called the φn splitting of the braid β. It is shown in [13] that Condition (6) can be replaced
by:
∗
b−k
for each k > 1, βk is the B+
n−1 -tail of φn (βb ) · ... · φn (βk −1 ) · βk .
φ62 (β3 )
φ6 (β2 )
β1
φ63 (β4 )
3
2
(7)
4
5
6
1
∗
Figure 4. The φ6 -splitting of a braid of B+
6 . Starting from
the right, we extract the maximal right divisor that keeps
the sixth strand unbraided, then extract the maximal right
divisor that keeps the first strand unbraided, etc.
∗
Example 2.3. Consider the braid β = a1,2 a2,3 a1,2 a2,3 of B+
3 . Using relations (3) on the underlined factors we obtain:
β = a1,2 a2,3 a1,2 a2,3 = a1,2 a2,3 a1,3 a1,2 = a1,2 a1,3 a1,2 a1,2 .
We decompose β as φ3 (γ1 ) · β1 with γ1 = φ3−1 (a1,2 a1,3 ) = a1,3 a2,3 and β1 =
a1,2 a1,2 . As the word a1,2 a1,3 is alone in its equivalence class, the braid
∗
φ3 (γ1 ) = a1,2 a1,3 is not right-divisible by a1,2 and so its B+
2 -tail is trivial.
The braid φ3 (γ1 ) is exactly the one of (6) for n = 3 and k = 1. Considering
γ1 instead of β we obtain γ1 = φ3 (γ2 ) · β2 with γ2 = φ3−1 (a1,3 a2,3 ) = a2,3 a1,2
and β2 = 1. As the word a1,3 a2,3 is also alone in its equivalence class,
∗
the braid φ3 (γ2 ) = a1,3 a2,3 is not right-divisible by a1,2 and so its B+
2 tail is trivial. The braid φ3 (γ2 ) is the one of (6) for n = 3 and k = 2.
THE ROTATING NORMAL FORM OF BRAIDS IS REGULAR
5
We decompose the braid γ2 as φ3 (γ3 ) · β3 with γ3 = φ3−1 (a2,3 ) = a1,2 and
∗
β3 = a1,2 . As the braid φ3 (γ3 ) = a2,3 is not right-divisible by a1,2 , its B+
2 -tail
is trivial. The braid φ3 (γ3 ) it the one of (6) for n = 3 and k = 3. Eventually
we express γ3 as φ3 (γ4 ) · β4 with γ4 = 1 and β4 = a1,2 . The process ends as
the remaining braid γ4 is trivial. We conclude that the φ3 -splitting of β is
(β4 , β3 , β2 , β1 ) = (a1,2 , a1,2 , 1, a21,2 ).
Before giving the definition of the rotating normal we fix some definitions
about words.
Definition 2.4. Assume n > 2. A word on the alphabet An is an An −1
±
word. A word on the alphabet A±
n = An ⊔ An is an An -word. The braid
±
represented by the An -word w is denoted by w. For w, w′ two A±
n -words,
we say that w is equivalent to w′ , denoted by w ≡ w′ if w = w′ holds. The
empty word is denoted by ε.
∗
The n-rotating normal form is an injective map rn from B+
n to the set of
An -words defined inductively using the φn -splitting.
∗
k
Definition 2.5. For β ∈ B+
2 , we define r2 (β) to be the unique A2 -word a1,2
∗
representing β. The rotating normal form of a braid β ∈ B+
n with n > 3 is:
rn (β) = φnb−1 (rn−1 (βb )) · ... · φn (rn−1 (β2 )) · rn−1 (β1 ),
where (βb , ... , β1 ) is the φn -splitting of β. A word w is said to be n-rotating
∗
if it is the n-rotating normal form of a braid of B+
n .
∗
As the n-rotating normal form of a braid of B+
n−1 is equal to its (n − 1)rotating normal form we can talk, without ambiguities, about the rotating
normal form of a dual braid.
Example 2.6. We reconsider the braid β of Example 2.3. We know that
the φ3 -splitting of β is (a1,2 , a1,2 , 1, a21,2 ). Since r2 (1) = ε, r2 (a1,2 ) = a1,2
and r2 (a21,2 ) = a21,2 we obtain:
r3 (β) = φ33 (a1,2 ) · φ32 (a1,2 ) · φ3 (ε) · a21,2 = a1,2 a1,3 a1,2 a1,2 .
Some properties of the rotating normal form have been established in [13].
Connections, established in [12] and [13], between the rotating normal form
and the braid ordering introduced by P. Dehornoy in [6] are based on these
properties.
We finish this section with some already known or immediate properties
about φn -splittings and n-rotating words.
Definition 2.7. For every non-empty word w, the last letter of w is denoted
∗
by w# . Assume n > 2. For each non-trivial braid β in B+
n , we define the
last letter of β, denoted β # , to be the last letter in the rotating normal form
of β.
Lemma 2.8 (Lemma 3.2 of [13]). Assume n > 3 and let (βb , ..., β1 ) be the
∗
φn -splitting of a braid of B+
n .
6
JEAN FROMENTIN
(i) For k > 2, the letter βk# is of type a..,n−1 unless βk = 1.
(ii) For k > 3 and for k = b, we have βk 6= 1.
The fact that βb is not trivial is a direct consequence of the definition
of φn -splitting. As, for k > 2, the braid β ′ = φn (βk#+1 )βk is a right divisor
of φnb−k (βb ) · ... · βk , it must satisfy some properties. In particular, if βk#+1 =
∗
ap−1,n−1 holds then the B+
n−1 -tail of φn (ap,n βk ) is trivial by (6).
Definition 2.9. For n > 3 and p, r, s in [1, n − 1], we say that a letter ar,s
is an ap,n -barrier if the relation r < p < s holds.
There exists no ap,n -barrier with n 6 3 and the only ap,4 -barrier is a1,3 ,
which is an a2,4 -barrier. By definition, if the letter x is an ap,n -barrier, then
∗
in the presentation of B+
n there exists no relation of the form ap,n · x =
y · ap,n allowing one to push the letter ap,n to the right through the letter x:
so, in some sense, x acts as a barrier.
The following result guarantees that the rotating normal form of a braid
satisfying some properties must contain an ap,n -barrier.
Lemma 2.10 (Lemma 3.4 of [13]). Assume n > 4, p ∈ [2, n − 2] and let β
+∗
∗
be a braid of B+
n−1 such that the Bn−1 -tail of φn (ap,n β) is trivial. Then
the rotating normal form of β is not the empty word and it contains an
ap,n -barrier.
It turns out that the above mentioned braid property are easily checked
for entries of a φn -splitting.
Lemma 2.11 (Lemma 3.5 of [13]). Assume n > 3 and let (βb , ..., β1 ) be
∗
the φn -splitting of some braid of B+
n . Then, for each k in [2, b − 1] such
#
that βk+1 is not an−2,n−1 (if any), the rotating normal form of βk contains
a φn (βk#+1 )-barrier.
3. Left reversing for dual braid monoid
The left-reversing process was introduced by P. Dehornoy in [7]. It is a
powerful tool for the investigation of division properties in some monoids as
stated by Proposition 3.6.
Definition 3.1. A monoid M defined by a presentation hS | Ri+ is leftcomplemented if there exists a map f : S × S → S ∗ satisfying:
R = f (x, y)x = f (y, x)y | (x, y) ∈ S 2
together with f (x, x) = ε for all x ∈ S.
As the relation x = x is always true for x ∈ S we say that M is leftcomplemented even if x = x does not occur in R for x ∈ S.
THE ROTATING NORMAL FORM OF BRAIDS IS REGULAR
7
∗
The monoid B+
3 with presentation of Proposition 1.1 is left-complemented
with respect to the map f given by
f (a1,2 , a2,3 ) = f (a1,2 , a1,3 ) = a1,3 ,
f (a2,3 , a1,2 ) = f (a2,3 , a1,3 ) = a1,2 ,
f (a1,3 , a1,2 ) = f (a1,3 , a2,3 ) = a2,3 .
∗
However the monoid B+
4 with presentation of Proposition 1.1 is not leftcomplemented. Indeed there is no relation of type ... a1,3 = ... a2,4 . It follows
that the words f (a1,3 , a2,4 ) and f (a2,4 , a1,3 ) are not well-defined.
In general for 1 6 p < r < q < s 6 n, the words f (ap,q , ar,s ) and
∗
f (ar,s , ap,q ) are not defined for the presentation of B+
n given in Proposi∗
tion 1.1. In order to obtain a left-complemented presentation of B+
n we
must exhibit some extra relations.
For example, the relation a2,3 a1,4 a1,3 ≡ a3,4 a1,2 a2,4 holds and so we can
consider f (a1,3 , a2,4 ) to be a2,3 a1,4 . However the relation a1,4 a2,3 a1,3 ≡
a3,4 a1,2 a2,4 is also satisfied and so f (a1,3 , a2,4 ) = a1,4 a2,3 is an other valid
choice.
Lemma 3.2. For n > 2, the map fn : An × An → A∗n defined by:
ε
for ap,q = ar,s ,
ap,s
for q = r,
for p = r and q > s,
as,q
fn (ap,q , ar,s ) = ar,p
for q = s and p > r,
ar,q ap,s for p < r < q < s,
as,q ar,p for r < p < s < q,
ar,s
otherwise.
∗
provides a structure of left-complemented monoid to B+
n .
Proof. Direct computations using Proposition 1.1 establish fn (x, y) · x ≡
fn (y, x) · y for all (x, y) ∈ A2n .
Assume n > 2. As illustrated above, the characterization of the map fn
∗
from the presentation of B+
n is not well-defined: many choices are possible.
Our map fn admits the following remarkable property: for every letters ap,q
and ar,s of An , the last letter of the word fn (ap,q , ar,s ) is of the form at,s ,
with t 6 r, whenever q < s. This property will be use in the sequel, for
example in the proof of Lemma 4.6.
Definition 3.3. Assume n > 2. For w and w′ two A±
n -words, we say that
w left-reverses in one step to w′ , denoted by w y1 w′ , if we can obtain w′
from w substituting a factor xy −1 (with x, y ∈ An ) by fn (x, y)−1 fn (y, x). We
say that w left-reverses to w′ , denoted by w y w′ , if there exists a sequence
1
w = w1 , ..., wℓ = w′ of A±
n -words satisfying wk y wk +1 for k ∈ [1, ℓ − 1].
8
JEAN FROMENTIN
1
Example 3.4. The word u = a1,2 a2,3 a1,2 a−
1,3 left-reverses to a2,3 a2,3 as the
following left reversing sequence shows (reversed factor are underlined)
−1
−1
1
1
1
1
a1,2 a2,3 a1,2 a−
1,3 y a1,2 a2,3 a1,3 a2,3 y a1,2 a1,2 a2,3 a2,3 y a2,3 a2,3 ,
1
which is denoted by a1,2 a2,3 a1,2 a−
1,3 y a2,3 a2,3 .
A consequence of Lemma 1.1 of [7] is the following definition of left denominator and left numerator of an A±
n -word.
Definition 3.5. Assume n > 2. For w an A±
n -word, we denote by D(w) and
N (w) the An -words, if there exist, such that w y D(w)−1 N (w) holds. If
such words exist, then they are unique, hence the notations D(w) and N (w)
are unambiguous. The word N (w) is the left numerator of w while the
word D(w) is its left denominator.
Reconsidering Example 3.4, we obtain that the left denominator of u
is D(u) = ε and that its left numerator is N (u) = a2,3 a2,3 .
∗
Assume n > 2. By Example 8 of [10] based on [3], the monoid B+
n is a
Garside monoid and fn is a left lcm selector. A consequence of Lemma 4.3
of [10] is that for every w ∈ A±
n , the words N (w) and D(w) exist. We obtain
also the following result:
Proposition 3.6. Assume n > 2. For w an An -word and ap,q in An , the
1
braid w is right-divisible by ap,q if and only if D(w · a−
p,q ) is empty.
1
Since the left denominator of u = a1,2 a2,3 a1,2 a−
1,3 is empty, the braid a1,3
right-divides the braid a1,2 a2,3 a1,2 .
4. Characterization of rotating normal words
The aim of this section is to give a syntactical characterization of nrotating words among An -words.
∗
Lemma 4.1. Assume n > 3. For β in B+
n−1 there is equivalence between
(i) an An -word representing β contains an ap,n -barrier,
(ii) every An -word representing β contains an ap,n -barrier.
Proof. Clearly (ii) implies (i). For (i) ⇒ (ii) it is sufficient to prove that
relations of Proposition 1.1 preserve ap,n -barriers. For relation (2) this is
immediate since it conserves involved letters. If one of the three words
ar,s as,t , as,t ar,s and ar,t ar,s of relations (3) contains an ap,n -barrier then the
two others also, as illustrated by the following chord diagram.
p
s
r
t
n
THE ROTATING NORMAL FORM OF BRAIDS IS REGULAR
9
A letter ar,s is an ap,n -barrier if and only if its chord intersects properly that
of the letter ap,n .
As a consequence, containing an ap,n -barrier is not a word property but
a braid one.
∗
Lemma 4.2. Assume that n > 3. A non-empty braid β ∈ B+
n has triv+∗
ial Bn−1 -tail if and only if there exists a unique letter in An that rightdivides β, and this letter is of the form a..,n .
∗
Proof. If ap,n is the only letter right-divising β, then the B+
n−1 -tail of β is
+∗
certainly trivial. Conversely, assume that the Bn−1 -tail of β is trivial. Hence,
no letter ap,q with q < n right-divides β. Assume, for a contradiction, that
two distinct letters ap,n and aq,n with p < q < n right-divide β. The braid
β is then right-divisible by their lcm
ap,n ∨ aq,n = ap,q aq,n = aq,n ap,n = ap,n , ap,q ,
∗
hence by ap,q , which is impossible since the B+
n−1 -tail of β is trivial.
We conclude that, under some hypotheses, the last letter of a word is a
braid invariant.
Definition 4.3. For n > 3 and p, q ∈ [2, n − 1], we say that an n-rotating
word w is an ap,n -ladder lent on aq−1,n−1 if there exist a decomposition
w = v0 x1 v1 . . . vh−1 xh vh ,
a strictly increasing sequence j(0), ..., j(h), with j(0) = p and j(h) = n − 1,
and a sequence i(1), ..., i(h) such that:
(i) for each k 6 h, the letter xk is ai(k),j(k) with i(k) < j(k − 1) < j(k),
(ii) for each k < h, the word vk contains no aj(k),n -barrier,
(iii) the last letter of vh is aq−1,n−1 .
6
5
4
3
2
1
Figure 5. An a2,6 -ladder lent on a3,5 (the last letter). The
gray line starts at position 1 and goes up to position 5 using
the bar of the ladder. The empty spaces between bars in
the ladder are represented by a framed box. In such boxes
the vertical line representing the letter ai,j does not cross the
gray line. The bar of the ladder are represented by black
thick vertical lines.
Condition (i) is equivalent to saying that for each k 6 h, the letter xk is
an aj(k−1),n -barrier of type a..,j(k).
10
JEAN FROMENTIN
An immediate adaptation of Proposition 3.9 of [13] is:
∗
Lemma 4.4. Assume n > 4, p ∈ [2, n − 2] and let β be a braid of B+
n−1 such
+∗
that the Bn−1 -tail of φn (β) is trivial and such that β contains an ap,n -barrier.
Then the rotating normal form of β is an ap,n -ladder lent of β # .
In order to obtain a syntactical characterization of n-rotating words we
want a local version of condition (6) characterizing a φn -splitting. The
following result is the first one in this way.
∗
Proposition 4.5. Assume n > 4. For β ∈ B+
n−1 and p ∈ [2, n − 2] there is
equivalence between:
∗
(i) the B+
n−1 -tail of φn (ap,n β) is trivial,
+∗
(ii) the Bn−1 -tail of φn (β) is trivial and β contains an ap,n -barrier,
Our proof of Proposition 4.5 rests on the following Lemma.
Lemma 4.6. Assume n > 3. For u an An−1 -word and p ∈ [1, n − 1], the left
1
denominator D(ua−
p,n ) is not empty. More precisely, their exists q ∈ [1, p]
1 −1 begins with a−1 .
such that D(ua−
q,n
p,n )
Proof. Assume that w1 , ..., wℓ is a reversing sequence from the word w1 =
−1
−1
1
ua−
p,n to the word D(w1 ) N (w1 ). For k ∈ [1, ℓ] we denote by yk the leftmost
negative letter in wk , if such a negative letter exists in wk . We will show by
induction on k ∈ [1, ℓ] the existence of a non-increasing sequence r(k) such
that yk = ar(k),n .
By construction, we have r(1) = p and y1 = ar(1),n . Each reversing
step consists in replacing a factor xy −1 of wk by fn (x, y)−1 fn (y, x). For
k ∈ [1, ℓ − 1], if the reversed factor of wk does not contain the leftmost
occurence of yk−1 in wk then, yk+1 equals yk and so r(k + 1) = r(k). Assume
now that the reversed factor contains the leftmost negative letter yk of wk
and that yk is equal to ar(k),n . In particular, their exits a positive letter x
1
such that the reversed factor is of the form xyk−1 = xa−
r(k),n . As all letters
of wk on the left of the leftmost negative letter of wk must belong to An−1 ,
the letter x is of the form ai,j with 1 6 i < j < n. Lemma 3.2 implies
for j = r(k),
ai,n
fn (x, yk ) = fn (ai,j , ar(k),n ) = ar(k),j ai,n for i < r(k) < j,
ar(k),n
otherwise,
which gives in particular
−1
−1
xyk = ai,j ar(k),n
(
1
a−
i,n · · ·
y
1
a−
r(k),n · · ·
for i < r(k) 6 j,
otherwise.
(8)
It follows that yk+1 is equal to ar(k+1),n with r(k + 1) 6 r(k). Eventually we
−1
1
obtain that ua−
p,n left reverses to ar(ℓ),n · · · with r(ℓ) 6 r(1) = p and so the
desired property on D(uap,n ) holds.
THE ROTATING NORMAL FORM OF BRAIDS IS REGULAR
11
∗
Proof of Proposition 4.5. Let us first assume (i). As the the B+
n−1 -tail of φn (β)
is a right-divisor of φn (ap,n β) the first statement of (ii) holds and the second
statement of (ii) is Lemma 2.10.
We now prove (ii) ⇒ (i). By Lemma 4.2 and Condition (ii), there exists
a unique letter in An that right-divises φn (β) and it is of type a..,n . It follows
that the last letter β # of β is of type a..,n−1 . We denote by w the rotating
normal form of β. Let ar,s be an An -letter different from β # . It remains to
show that ar,s cannot be a right divisor of ap,n β.
Assume first s 6 n − 1. Then the letter ar,s belongs to An−1 . As β lies
∗
in B+
n−1 , Lemma 4.2 implies that the braid ar,s could not be a right-divisor
1
of β. Hence, by Proposition 3.6 the word D(w a−
r,s ) must be non-empty. As
±
±
the reversing of an An−1 -word is also an An−1 -word, there exists a letter at,t′
with t′ < n such that:
−1
1
ap,n w a−
r,s y ap,n at,t′ · · · ,
holds. Clearly, the braid at,t′ is not a right divisor of ap,n (since we have
1
t′ < n). Therefore, by Proposition 3.6, the left denominator of ap,n w a−
r,s is
not empty, and we conclude that ar,s is not a right divisor of ap,n β.
Assume now s = n. Hypotheses on β plus Lemma 4.4 imply that w is an
ap,n -ladder lent on β # . Following Definition 4.3, we write
w = v0 x 1 v1 · · · vh − 1 x h vh .
By Lemma 4.6, there exist two maps η and µ from N to itself such that
−1
−1
′ −1
′ −1
1
wa−
r,n = wh aη(h),n y wh aµ(h),n · · · y ... y w0 aη(0),n · · · y w0 aµ(0),n · · · ,
where for all k ∈ [0, h],
wk = v0 x1 v1 · · · vk−1 xk vk ,
wk′ = v0 x1 v1 · · · vk−1 xk .
By construction w0 is v0 while w0′ is the empty word and η(h) is r. For
k ∈ [0, h], Lemma 4.6 with u = vk and p = η(k) imply that the word
−1
′
1
wk a−
η(k),n · · · = wk vk aη(k),n · · ·
1
left-reverses to wk′ a−
µ(k),n · · · with µ(k) 6 η(k). Then, for k 6= 0, Lemma 4.6
(with u = vk−1 and p = µ(k)) implies that the word
−1
1
wk′ a−
µ(k),n · · · = wk −1 xk aµ(k),n · · ·
1
left-reverses to wk−1 a−
η(k −1),n · · · with η(k − 1) 6 µ(k). Using an inductive
argument on k = h, ..., 0 we then obtain:
µ(0) 6 η(0) 6 µ(1) 6 ... 6 µ(h) 6 η(h) = r.
(9)
Following Definition 4.3 we write xk = ai(k),j(k) .
We will now prove for all k ∈ [0, h − 1],
µ(k + 1) 6 j(k + 1) ⇒ η(k) < j(k).
(10)
12
JEAN FROMENTIN
Let k ∈ [0, h − 1] and assume µ(k + 1) 6 j(k + 1). Definition 4.3 (i)
guarantees the relation i(k + 1) < j(k) < j(k + 1). For the case where
µ(k + 1) 6 i(k + 1) we have:
η(k) 6 µ(k + 1) 6 i(k + 1) < j(k),
and we are done in this case. The remaining case is µ(k + 1) > i(k + 1). By
relation (8), with i = i(k + 1), j = j(k + 1) and r = µ(k + 1) (satisfying the
relation i < r 6 j) we obtain
−1
−1
1
xk+1 a−
µ(k +1),n = ai(k +1),j(k +1) aµ(k +1),n y ai(k +1),n v,
for some A±
n -word v. In particular, we have η(k) = i(k + 1) < j(k) and (10)
is established.
For k = h − 1 the left-hand member of (10) is satisfied since j(h) is equal
to n − 1 and
µ(h) 6 η(h) = r 6 n − 1
holds by definition of r. Properties (9) and (10) imply µ(k) < j(k) for
all k ∈ [0, h − 2]. In particular we have µ(0) < j(0) = p together with
−1
1
wa−
r,n y aµ(0),n · · · . As aµ(0),n cannot be a right divisor of ap,n it follows
1
that the left denominator of ap,n w a−
r,n is also non-empty and so that ar,n is
not a right divisor of ap,n β.
As the reader can see, the case p = n − 1 is excluded from Proposition 4.5.
This case is treated by the following result.
∗
Proposition 4.7. Assume n > 3. For β a non-trivial braid of B+
n−1 there
is equivalence between:
∗
(i) the B+
n−1 -tail of φn (an−1,n β) is trivial,
+∗
(ii) the Bn−1 -tail of φn (β) is trivial.
Proof. As a right divisor of φn (β) is also a right of φn (an−1,n β), the implication (i) ⇒ (ii) is clear. Let us now prove (ii) ⇒ (i). For x ∈ An−1
∗
and y, z in An , the only relations in B+
n of the form an−1,n x = y z with
(y, z) 6= (an−1,n , x) are commutations relations, for which x is of type ap,q
with p < q < n − 1. Hence, if w is a representative of β, the relations in
∗
the dual braid monoid B+
n−1 allow only transforming the word an−1,n w into
words of the form u an−1,n v. The An−2 -word u represents a dual braid that
commute with an−1,n and u v represents β. By Lemma 4.2 the only letter in
An that right-divides φn (β) is of type a..,n , hence any representative word
of β must end with a letter of type a..,n−1 . In particular v is not an An−2 word and so v is not empty and it admits only β # as last letter. We have
then established that every representative word of an−1,n β ends with the
∗
letter β # which is of type a..,n−1 . Therefore, the B+
n−1 -tail of φn (an−1,n β) is
trivial.
Theorem 4.8. Assume n > 3. A finite sequence (βb , ..., β1 ) of braids
+∗
∗
in B+
n−1 is the φn -splitting of a braid of Bn if and only if:
(i) for k > 3 and k = b, the braid βk is not trivial,
THE ROTATING NORMAL FORM OF BRAIDS IS REGULAR
13
∗
(ii) for k > 2, the B+
n−1 -tail of φn (βk ) is trivial,
(iii) if, for k > 3, we have βk# 6= an−2,n−1 , then βk−1 contains an φn (βk# )barrier.
∗
Proof. Let (βb , ..., β1 ) be the φn -splitting of some braid of B+
n . Condition (i)
∗
is Lemma 2.8.(ii). Condition (6) implies that the B+
n−1 -tail of
φnb−k (βb ) · ... · φn (βk+1 )
∗
is trivial for k > 1. In particular the B+
n−1 -tail of φn (βk +1 ) must be trivial
for k > 1, which implies (ii). Condition (iii) is Lemma 2.11.
∗
Conversely, let us prove that a sequence (βb , ..., β1 ) of braids of B+
n−1 satis+∗
fying (i), (ii) and (iii) is the φn -splitting of some braid of Bn . Condition (i)
implies that βb is not trivial. For k > 2 we denote by γk the braid
γk = φnb−k (βb ) · ... · φn (βk+1 ) · βk .
For k > 3, and for k > 2 whenever β2 6= 1, we first prove
βk# is the only An -letter that right-divides γk .
#
(11)
We note that Condition (i) guarantees the existence of βk for k > 3. For
∗
k = b, Condition (ii) implies that the B+
n−1 -tail of φn (βb ) is trivial. Hence, by
Lemma 4.2 the only An−1 -letter that right-divides βb is βb# . Since any right
+∗
∗
divisors of a braid of B+
n−1 lies in Bn−1 , we have established (11) for k = b.
Assume (11) holds for k > 4, or k > 3 whenever β2 6= 1, and let us
prove it for k − 1. By Condition (ii), there exists p such that βk# is ap−1,n−1 .
∗
Conditions (ii) and (iii) imply that the B+
n−1 -tail of φn (βk −1 ) is trivial and
that βk−1 contains an ap,n -barrier whenever p < n − 1 holds. Hence, by
Proposition 4.5 (for p < n − 1) and Proposition 4.7 (for p = n − 1), the
∗
B+
n−1 -tail of the braid φn (ap,n βk −1 ) is trivial. Using Lemma 4.2, we obtain
that φn (ap,n βk−1 ) is right-divisible by a unique An -letter. It follows that the
unique An -letter right-divising ap,n βk−1 is βk#−1 .
We denote by uap−1,n−1 and v two An -words representing γk and βk−1
respectively. The braid γk−1 is then represented by φn (u)ap,n v. Let y be
an An -letter different from βk#−1 . As y is not a right divisor of ap,n βk−1 ,
Proposition 3.6 implies the existence of an An -letter x different from ap,n
satisfying
φn (u)ap,n vy −1 y φn (u)ap,n x−1 · · · .
The word φn (u)ap,n represents φn (γk ). By induction hypothesis x is not a
right divisor of φn (γk ). Then Proposition 3.6 implies that D(φn (u)ap,n x−1 )
is not empty. It follows D(φn (u)ap,n vy −1 ) 6= ε and so, always by Proposition 3.6, the letter y is not a right divisor of γk−1 . Eventually we have
established (11) for k > 3.
A direct consequence of (11) and Condition (ii) is that the only An -letter
∗
right-dividing φn (γk ) is of type a..,n and so, by Lemma 4.2 the B+
n−1 -tail
of the braid γk is trivial for k > 3 and for k = 2 whenever βk 6= 1. It
∗
remains to establish that the B+
n−1 -tail of φn (γ2 ) is also trivial whenever β2
is trivial. Assume β2 = 1. Condition (iii) implies β3# = an−2,n−1 . By (11),
14
JEAN FROMENTIN
an−2,n−1 is the only An -letter that right-divides γ3 . Since γ2 = φn (γ3 ) holds,
the letter φn2 (an−2,n−1 ) = a1,n is the only letter right-dividing φn (γ2 ). In
∗
particular the B+
n−1 -tail of φn (γ2 ) is trivial.
Conditions (i), (ii) and (iii) are easy to check if the braids β1 , ..., βb are
given by their rotating normal forms.
Corollary 4.9. Assume n > 3 and let (wb , ..., w1 ) be a finite sequence
of An−1 -words. The An -word
φnb−1 (wb ) · ... · φn (w2 ) · w1 ,
(12)
is n-rotating if the following conditions are satisfied:
(i) for k > 1, the word wk is (n − 1)-rotating,
(ii) for k > 3, the word wk ends with ap−1,n−1 for some p,
(iii) the word w2 is either empty (except for b = 2) or ends with ap−1,n−1
for some p,
(iv) if, for k > 3, the word wk ends with ap−1,n−1 with p 6= n − 1 then the
word wk−1 contains an ap,n -barrier.
Proof. Assume that (wb , ..., w1 ) satisfies Conditions (i)-(iv) and let us prove
that the word w defined in (12) is rotating.
We denote by βi (resp. β) the braid represented by wi (resp. w). By Condition (i) and Definition 2.5, the word w is rotating if and only if (βb , ..., β1 ) is
a φn -splitting. Conditions (ii) and (iii) imply Condition (i) of Theorem 4.8.
Theorem 4.8.(iii) is a consequence of (ii) and (iv).
We now prove Condition (ii) of Theorem 4.8. Let k be in [2, b]. If
∗
the B+
n−1 -tail of φn (βk ) is not trivial, then there exists ap,q , with 1 6 p <
∗
q < n, that right-divides φn (βk ). As βk lies in B+
n−1 , we must have p 6= 1 and
∗
therefore βk is right-divisible by ap−1,q−1 with q − 1 6 n − 2. The B+
n−1 -tail
of wk is then not trivial. Since the word wk is (n − 1)-rotating, its last
∗
letter must come from its B+
n−1 -tail. Hence wk must end with a letter ai,j
satisfying j 6 n − 2, which is in contradiction with Conditions (ii) and (iii).
We conclude applying Theorem 4.8.
It is not true that any decomposition of an n-rotating word as in (12)
satisfies Conditions (i)−(iv) of Corollary 12. However we have the following
result.
Proposition 4.10. For n > 3 and every n-rotating word w, there exists
a unique sequence (wb , ..., w1 ) of (n − 1)-rotating words such that w decomposes as in (12) and Conditions (ii) − (iv) of 4.9 hold.
Proof. By definition of a rotating normal word and by Lemma 2.8 such a
sequence exists. Let us prove now the unicity. Assume w is a n-rotating normal word and that (wb , ..., w1 ) and (wc′ , ..., w1′ ) are two different sequences of
(n − 1)-rotating normal words satisfying Conditions (ii) and (iii) of Corollary 4.9. Let k be the minimal integer satisfying wk 6= wk′ . Since the sum of
the word lengths of the two sequences are the same, we have k 6 min{b, c}.
THE ROTATING NORMAL FORM OF BRAIDS IS REGULAR
15
Without loss of generality, we may assume that wk′ is a proper suffix of
wk , i.e., wk = u·wk′ . Let x be the last letter of wk′ +1 or the last letter of wk′ +2
if wk+1 is empty. By Conditions (ii) and (iii) of Corollary 4.9, the letter x is
equal to ap−1,n−1 for some p and wk admits either φn (ap−1,n−1 )wk′ = ap,n wk′
or φn2 (ap−1,n−1 )wk′ = a1,p+1 wk′ as suffix. The first case is impossible since wk
is an An−1 -word. The second case may occur only for k = 1 and w2′ = ε.
In this case w2′ is empty and so Condition (iv) of Corollary 4.9 implies that
the last letter of w3′ , which is x, is equal to an−2,n−1 . This implies that wk
admits a1,n u as suffix, which is also impossible since wk is an An−1 -word.
A direct consequence of Corollary 4.9 and Proposition 4.10 is
Theorem 4.11. For n > 3, an An -word w is rotating if and only if it can
be expressed as in (12) subject to Conditions (i) − (iv) of Corollary 4.9.
5. Regularity
In this section we will show that the language of n-rotating words, denoted
by Rn , is regular, i.e., there exists a finite-state automaton recognizing the
language of n-rotating words. As the rotating normal form is defined using
right division it is more natural for an automaton to read words from the
right. For w = x0 · ... · xk an An -word we will denote by Π(w) the mirror
word xk · ... · x0 . By Theorem 1.2.8 of [11] the language Rn is regular if and
only if the language Π(Rn ) is. In this section we will construct an automaton
recognizing Π(Rn ).
For us a finite-state automaton is a 6-tuple (S ∪ {⊗}, A, µ, Y, i, ⊗) where
S ∪ {⊗} is the finite set of states, A is a finite alphabet,
µ : (S ∪ {⊗}) × A → S ∪ {⊗}
is the transition function, Y ⊆ S is the set of accepting states and i is the
initial state. In this paper each automaton is equipped with an undrawn
dead state ⊗ and each state except the dead one is accepting, i.e., Y = S
always hold.
To characterize A it is enough to describe µ on (s, x) ∈ S × A instead of (S ∪ {⊗}) × A. For example an automaton recognizing the language R2 = Π(R2 ) is A2 = ({1, ⊗}, {a1,2 }, µ2 , {1}, 1, ⊗) with µ2 (1, a1,2 ) = 1.
The corresponding automaton diagram is:
a1,2
1
The horizontal arrow points to the initial state.
We will now construct an automaton A3 recognizing the language Π(R3 ).
16
JEAN FROMENTIN
Proposition 5.1. An A3 -word xeb b
a1,2
xk = a2,3
a1,3
3
2
1
· ... · ae1,3
ae2,3
ae1,2
where
if k ≡ 1 mod 3,
if k ≡ 2 mod 3,
if k ≡ 3 mod 3.
is rotating if and only if ek 6= 0 for all k > 3.
Proof. The 2-rotating words are powers of a1,2 . Let w be the word of the
k
statement. Defining wk to be ae1,2
, we obtain
w = φnb−1 (wb ) · ... · φn (w2 ) · w1 .
As each non-empty word wk ends with a1,2 , Condition (iv) of Corollary 4.9
is vacuously true. Hence the word w is rotating if and only if it satisfies
Conditions (ii) and (iii) of Corollary 4.9, i.e., ek is not 0 for k > 3.
As a consequence the following automaton recognizes the language Π(R3 ):
a1,3
a1,2
a1,3
0
2
a1,3
a1,2
3
a1,2
a2,3
1
a2,3
a2,3
We now explain the principle of our construction of an automaton An
recognizing the language Π (Rn ) for n > 4. Following the Condition (i) of
Corollary 4.9, the construction will be done by induction on n > 3. Quite
natural it will be more complicated for n > 4 than for n = 2 and n = 3
since Condition (iv) of Corollary 4.9 is not empty for n > 4 and must be
checked. As suggested by Condition (ii) the language Rn∗ (see Definition 5.2)
of n-rotating words ending with a letter a..,n is in the core of the process.
Hence, instead of directly construct the automaton An we will construct an
automaton A∗n recognizing the language Π(Rn∗ ). As the language Π(Rn ) is
the concatenation of Π(Rn−1 ) and Π(Rn∗ ), the automaton An is naturally
obtained from An−1 and A∗n .
Assume that the automaton A∗n−1 recognizing Π(Rn∗ −1 ) is given. In order to check Condition (iv) of Corollary 4.9, we must modify A∗n−1 to store
information about the existence of a..,n-barriers in the involved word. A
duplication of states contained in A∗n−1 together with some standard modifications on the transition function allow us to store one bit of information.
Since there is exactly n − 3 types of a..,n -barriers we must multiply the number of states of A∗n−1 by at most 2n−3 in order to obtain an automaton Bn0 −1
recognizing Π(Rn∗ −1 ) and detecting barriers.
THE ROTATING NORMAL FORM OF BRAIDS IS REGULAR
17
Then, for k ∈ [1, n − 1] we construct an automaton Bnk −1 recognizing
the language Π(φnk (Rn∗ −1 )) by applying φnk to Bn0 −1 . Eventually we obtain
A∗n by plugging cyclically automata Bn0 −1 , ..., Bnn−−11 . Transitions between au1
tomata Bnk −1 and Bnk+
−1 will be done in respect with Condition (iv) of Corollary 4.9 thanks to informations about encountered barriers stored in Bnk −1 .
Definition 5.2. For n > 2, we denote by Rn∗ the language of n-rotating
words that are empty or end with a letter of type ap,n for some p.
As explained above, we will first construct by induction on n > 3 an
automaton A∗n for the language Π (Rn∗ ).
Definition 5.3. A partial automaton is a 6-tuple P = (S∪{⊗}, A, µ, S, I, ⊗)
where S, A and µ are defined as for an automaton and I : A → S ∪ {⊗} is
the initial map. The closure of a partial automaton P is the automaton
A(P ) = (S ∪ {◦, ⊗}, A, µc , S ∪ {◦}, ◦, ⊗)
whose transition function is given by
(
I(x)
µc (s, x) =
µ(s, x)
if s = ◦,
otherwise.
In the sequel, we will plug partial automata using their initial maps. A
partial automaton is represented as an automaton excepted for the initial
map I: for each x ∈ A we draw an arrow attached to the state I(x) and
labelled x. We say that a partial automaton recognizes a given language if
its closure does.
a1,3
a1,3
a1,3
2
a1,3
a1,2
a1,3
3
a1,2
◦
2
a1,3
a1,2
3
a1,2
a2,3
a2,3
1
a2,3
a2,3
1
a2,3
a2,3
Figure 6. The partial automaton P3 and the corresponding
closure, which recognizes the language Π(R3∗ ).
We will now show how to construct by induction a partial automaton Pn∗
recognizing Π (Rn∗ ) for n > 3. For n = 3 this is already done by Figure 6.
For the sequel we assume that n > 4 and that
Pn∗−1 = (Sn∗ −1 ∪ {⊗}, An−1 , µ∗n−1 , Sn∗ −1 , In∗−1 , ⊗)
is a given partial automaton which recognizes the language Π (Rn∗ −1 ).
18
JEAN FROMENTIN
We construct a partial automaton Pn0 that recognize Π(Rn∗ −1 ) and store
informations about the encountered barriers. We define Sn0 to be the set
Sn0 = {0} × Sn∗ −1 × P({a2,n , ..., an−2,n }).
A state in Sn0 is then written (0, s, m). The integer 0 is used to identify this
particular partial automaton among those that will eventually constitute Pn∗ .
The set m stores informations on encountered barriers.
For ai,j ∈ An−1 we denote by bar(ai,j ) the set of letters ap,n such that ai,j
is an ap,n -barrier, i.e.,
bar(ai,j ) = {ap,n | i < p < j}.
Definition 5.4. We define Pn0 = (Sn0 ∪ {⊗}, An−1 , µ0n , Sn0 , In0 , ⊗) to be the
partial automaton defined by: for all x ∈ An−1 ,
(
(0, In∗−1 (x), bar(x)) if In∗−1 (x) 6= ⊗,
0
In (x) =
⊗
if In∗−1 (x) = ⊗.
and for all (0, s, m) ∈ Sn0 and for all x ∈ An−1 ,
(
(0, µ∗n−1 (s, x), m ∪ bar(x))
µ0n ((0, s, m), x) =
⊗
if µ∗n−1 (s, x) 6= ⊗,
if µ∗n−1 (s, x) = ⊗.
Proposition 5.5. The partial automaton Pn0 recognizes the language Π (Rn∗ −1 ).
Moreover an accepted An -word Π(w) contains an ap,n -barrier if and only if
Pn0 has state (0, s, m) with ap,n ∈ m after reading Π(w).
Proof. We denote by A and A′ the closure of Pn∗−1 and Pn0 respectively.
Assume w = w1 · . . . · wℓ is an An -word of length ℓ. For all k ∈ [1, ℓ], we
denote by sk (resp. s′k ) the state of the automaton A (resp. A′ ) after reading
the k-th letter of Π(w), i.e., the letter wℓ−k+1 . By construction of µ0n , for
all k ∈ [1, ℓ], we have
(
⊗
if sk = ⊗,
s′k =
(0, sk , mk ) otherwise (for a certain set mk ).
In particular s′ℓ 6= ⊗ if and only if sℓ 6= ⊗. Hence both automata accept the
word Π(w) or not and so Pn0 recognizes Π(Rn∗ −1 ) since Pn∗−1 does.
Let us prove the result on mℓ whence Π(w) is accepted. By an immediate
induction of k ∈ [1, ℓ], the set mk contains bar(wℓ ) ∪ . . . ∪ bar(wℓ−k+1 ). The
set mℓ is then bar(w1 ) ∪ ... ∪ bar(wℓ ). We conclude by definition of bar(),
that w contains an ap,n -barrier if and only if ap,n belongs to mℓ .
As the only ap,4 -barrier in A4 is a1,3 , the partial automaton P40 is obtained
from P3∗ by connecting edges labelled a1,3 to a copy of P3 , as illustrated on
Figure 7.
We now construct n − 1 twisted copies of Pn0 using the word homomorphism φn . For t = (0, s, m) ∈ Sn0 and k ∈ [1, n − 1], we define Φnk (t) to be
THE ROTATING NORMAL FORM OF BRAIDS IS REGULAR
a1,3
a1,3
0,2
∅ a1,2
a1,3
0,1
∅
a2,3
a2,3
a1,3
a1,3
0,2
a2,4
0,3
∅
a1,2
a1,3
0,1
a2,4
a2,3
a1,3
19
a1,2
0,3
a2,4
a2,3
a2,3
a1,2
0,2
a2,4
a1,3
a1,2
a1,3
0,1
∅
a2,3
a1,2
0,3
a2,4
a1,3
a2,3
0,1
a2,4
a2,3
a2,3
Figure 7. The partial automaton P40 . Obsolete transitions
from P3∗ are in gray. New added transitions are dashed. The
right partial automaton is P40 without inaccessible states.
the state (k, s, m). We also define Snk to be Φnk Sn0 and
Pnk = Snk ∪ {⊗}, φnk (An−1 ) , µkn , Snk , Ink , ⊗ ,
to be the partial automaton given by Ink (φnk (x)) = Φnk In0 (x) and
µkn ((k, s, m), φnk (x)) = Φnk µ0n ((0, s, m), x) ,
with the convention Φnk (⊗) = ⊗. In other words, Pnk is obtained from Pn0 by
replacing the letter x by φnk (x) and the state (0, s, m) by (k, s, m). By Proposition 5.5, we obtain immediately that Pnk recognizes the language φnk (Π(Rn∗ ))
and store informations about encountered barriers.
We can now construct the partial automaton Pn∗ by plugging cyclically n
partial automata Pnk for k ∈ [0, n − 1]. Transition between two adjacent
partial automata is done using initial maps accordingly to Condition (iv) of
Corollary 4.9.
Definition 5.6. We define Pn∗ = (Sn∗ ∪ {⊗}, An , µ∗n , Sn , In∗ , ⊗), with Sn∗ =
Sn0 ⊔ ... ⊔ Snk to be the partial automaton given by
1
In (x) if x = ap,n with p 6= 1,
In∗ (x) = In2 (x) if x = a1,n ,
⊗
otherwise,
and with transition function
µkn ((k, s, m), φnk (x))
k +1 k
In (φn (x))
µ∗n ((k, s, m), φnk (x)) = Ink+1 (φnk (x))
⊗
with the convention Inn = In0 .
if x ∈ An−1 ,
if x = an−1,n
if x = ap,n with 2 6 p 6 n − 2
and ap,n ∈ m,
otherwise,
20
JEAN FROMENTIN
We summarize the construction of the partial automaton Pn∗ on the following diagram.
φ2n (An−1 )
φn (An−1 )
T1
ap,n
Pn1
p6=1
a1,n
Pn2
T0
An−1
Pn∗
Pn0
Pnk
Tn−1
Tk
Pnn−1
φn−1
(An−1 )
n
φkn (An−1 )
Pnk+1
φk+1
(An−1 )
n
An arrow labelled Tk represents the set of transitions µ∗n ((k, s, m), φnk (ap,n )).
Lemma 5.7. The partial automaton Pn∗ recognizes the language Π (Rn∗ ).
Proof. Let A∗ be the closure of Pn∗ and w be a non-empty An -word. There
exists a unique sequence (wb , ..., w1 ) of An−1 -words such that wb 6= ε, w is
equal to
φnb−1 (wb ) · ... · φn (w2 ) · w1
and for all i, the word φni (wi ) is the maximal suffix of φnb−1 (wb ) · ... · φni (wi )
belonging to φni (A∗n−1 ). By definition of In∗ , the word Π(w) is accepted by Pn∗
only if w ends with a letter ap,n for some p.
We assume now that w is such a word. Thus the first integer j such that
wj is non-empty is 2 or 3. More precisely, we have j = 2 if p > 1 and
j = 3 if p = 1. In both cases, after reading the first letter of Π(w), the
automaton A∗ is in state s ∈ Snj . The automaton reaches a state outside of
Snj if it goes to the state ⊗ or if it reads a letter outside of φnj −1 (An−1 ), i.e.,
a letter of φnj (Π(wj +1 )). This is a general principle: after reading a letter
of φni−1 (Π(wi )) the automaton A is in state (t, s, m) with t = i mod n. By
construction of Pnt , the word φni−1 (Π(wi )) provides an accepted state if and
only if wi is a word of Rn−1 .
At this point we have shown that Π(w) is accepted by A only if w is
empty or if w satisfies w# = ap,n together with Conditions (i), (ii) and (iii)
of Corollary 4.9. Let i be in [j, k − 1] and assume that A is in an acceptable
state (t, s, m) with t = i mod n after reading the word
Π(φni−1 (wi ) · ... · φn (w2 ) · w1 ).
We denote by x the letter wi#+1 . By construction of wi+1 , the letter x does
not belong to φni−1 (An−1 ) and so x = φni (ap,n ) for some p. By definition of µ∗n
we have µ∗n ((t, s, m), φni (ap,n )) 6= ⊗ if and only if p = n − 1 or if p ∈ [2, n − 2]
together with ap,n ∈ m. As, by construction of Pnt , we have ap,n ∈ m if and
only if wi contains an ap,n -barrier, Condition (iv) of Corollary 4.9 is satisfied.
THE ROTATING NORMAL FORM OF BRAIDS IS REGULAR
21
Eventually, by Corollary 4.9, the word Π(w) is accepted by A if and only
if w belongs to Rn∗ .
a2,4
a2,4
a2,3
1,2
a2,4
a2,4
a2,4
a3,4
a2,3
a2,4
a3,4
1,1
a2,4
1,1
∅
a1,4
a3,4
a1,2
0,2
a2,4
a2,3
a1,4
a3,4
a3,4
a1,2 a0,3
2,4
a1,3
1,3
a2,4
0,1
a2,4 a2,3
a1,3
2,1
∅
a1,4
a1,3
a1,3
a1,3
a1,3
a1,4 2,1
a2,4
0,1
∅
a1,3
a1,4
2,2
a2,4
a3,4
2,3 a
a2,4 3,4
a2,3
a1,2
a1,2
a2,3
3,1
a2,4
a1,2
a1,3
3,3
a2,4
a1,4
3,1
∅
a2,4
a1,4
a1,2
a2,4
a2,4
3,2
a2,4
a2,4
Figure 8. Partial automaton recognizing the language Π (R4∗ ).
Assume that an automaton An−1 = (Sn−1 ∪ {⊗}, An−1 , µn−1 , Sn−1 , i, ⊗)
recognizing the language Π(Rn−1 ) for n > 4 is given. Plugging An−1 to the
partial automaton Pn∗ = (Sn∗ ∪ {⊗}, An , µ∗n , Sn∗ , In∗ ) we construct an automaton An = (Sn ∪ {⊗}, An , µn , Sn , i, ⊗) defined by Sn = Sn−1 ⊔ Sn∗ and
µn−1 (s, x) if s ∈ Sn−1 and x ∈ An−1 ,
µn (s, x) = In∗ (x)
if s ∈ Sn−1 and x ∈ An \ An−1 ,
∗
µn (s, x)
if s ∈ Sn∗ .
Proposition 5.8. If An−1 recognizes Π(Rn−1 ), the automaton An recognizes
the language Π(Rn ).
Proof. Let w be an An -word, w1 be the maximal suffix of w which is an An−1 word and w′ be the corresponding prefix. By Corollary 4.9, the word w is
rotating if and only if w1 and w′ are. By construction, the automaton An
is in acceptable state after reading Π(w1 ) if and only if w1 is an (n − 1)rotating word. Hence w is accepted only if w1 is rotating. Assume this is
22
JEAN FROMENTIN
a1,2
0
a2,3 a2,3
a1,3
a1,3
1
2
a1,3
a2,3
a1,2
3
a2,4
a3,4
a2,4
a2,4
a3,4
a2,3
a2,4
a1,2
0,2
a2,4
a1,3
1,3
a2,4
a3,4
1,1
a2,4
1,1
∅
a1,4
a3,4
a3,4
a1,2 a0,3
2,4
a1,4
a2,3
1,2
a2,4
a2,4
a1,2
a1,4
0,1
a2,4 a2,3
a2,3
a1,3
2,1
∅
a1,3
a1,3
a1,3
a1,3
a1,4 2,1
a2,4
0,1
∅
a1,3
a1,4
2,2
a2,4
a3,4
2,3 a
a2,4 3,4
a2,3
a1,2
a1,2
a2,3
3,1
a2,4
a1,2
a1,3
3,3
a2,4
a1,4
3,1
∅
a2,4
a1,4
a1,2
a2,4
a2,4
3,2
a2,4
a2,4
Figure 9. Automaton A4 for the language Π(R4 ).
the case. By Lemma 5.7 the automaton An is in an acceptable state after
reading Π(w′ ) if and only if the word w′ is rotating. Eventually the word
Π(w) is accepted by A if and only if w1 and w′ are both rotating, which is
equivalent to w is rotating.
By Proposition 5.8, the language Π(Rn ) is regular and so we obtain:
Theorem 5.9. The language of n-rotating words Rn is regular.
THE ROTATING NORMAL FORM OF BRAIDS IS REGULAR
23
6. Application to σ-definite representative
It it well known since [6] that Artin’s braid groups are orderable, by
an ordering that enjoys many remarkable properties [9]. The existence of
this ordering rests on the property that every non-trivial braid admits a
σ-definite representative, i.e., a braid word w in standard Artin generator
σ1 , σ2 , ... in which the generator σi with highest index i occurs only positively (no σi−1 ), in which case w is called σi -positive, or only negatively (no
σi ), in which case w is called σi -negative. Thanks to the word homomorphism defined by ap,q 7→ σp ... σq−1 σq σq−−11 ... σp−1 , the notions of σ-definite,
σi -positive and σi -negative words extend naturally to A±
n -words. In [13] a
particular σ-definite representative is obtained for every non-trivial braid
from its rotating normal form.
We say that a braid has index k > 2 if it belongs to Bk and not to Bk−1
with the convention B1 = {1}. Each representative, in standard Artin
generator, of a braid of index k > 2 must contain a letter σk−1 or a letter
σk−−1 1 . Hence such a braid is either σk−1 -positive or σk−1 -negative.
First we recall how uniquely express a braid as a quotient.
Proposition 6.1 (Proposition 6.1 of [13]). Assume n > 3. Each braid β
of Bn admits a unique expression δn−t w where t is a non-negative integer,
w is an n-rotating normal word, and the braid w is not left-divisible by δn
unless t is zero.
A direct adaptation to dual braid context of Proposition 4.4 of [14] gives:
∗
Proposition 6.2. Assume n > 3 and β is a braid of B+
n . Let t be a
positive integer and (wb , ..., w1 ) the φn -splitting of β. If t > b − 1 holds then
the quotient δn−t β is represented by the σn−1 -negative A±
n -word
δn−t+b+1 wb δn−1 wb−1 δn−1 ... w2 δn−1 w1 ,
(13)
where δn is written as the word a1,2 a2,3 ... an−1,n . Moreover if t < b − 1 then
the quotient δn−t β is not σn−1 -negative.
Proposition 6.2 gives specific σ-definite representative to each σn−1 -negative
braid of Bn . Unfortunately this representative word is not unique.
−1
Definition 6.3. We denote by Rσn the words of the form (13) such that
the braid φnb−1 (wb ) · ... · φn (w2 ) · w1 is not left-divisible by δn .
Assume (wb , ..., w1 ) is a sequence of An−1 -words satisfying Conditions (i)–
(iv) of Corollary 4.9. In section 5 we have explicitly constructed by induction
a finite-state automaton recognizing the mirror of the word φnb−1 (wb ) · · · w1 .
This construction rests on verification of conditions of Corollary 4.9 and on
the detection of transition between words φni−1 (wi ) and φni (wi+1 ). Thus, if we
add a letter $ to the alphabet An we can construct a finite-state automaton
recognizing the mirrors of words wb $wb−1 $...$w2 $w1 .
Since An contains no negative letters, the word δn−1 can play the role of the
letter $. Hence their exists a finite-state automaton recognizing the reverse
24
JEAN FROMENTIN
of the word
wb δn−1 wb−1 δn−1 ... w2 δn−1 w1 .
We then obtain that mirror of words of (13) constitute a regular language.
−1
Proposition 6.4. For n > 2, the language Rσn
is regular.
Proof. The result is immediate for n = 2. Assume n > 3. We denote by
Wn the set of words as in (13). As explained previously the language Wn is
regular. Let Bn be a finite-state automaton recognizing Wn and let w be a
word as in (13). Following Section 9.2 of [11] we can modify Bn such that it
∗
memorize the maximal simple element head(w) of B+
n that left-divides the
−1
b
dual braid φn (wb )·...·φn (w2 )·w1 . Accepting only words such that head(w)
is different from δn we obtain an finite-sate automaton which recognize the
−1
language Rσn .
Since the inverse of σn−1 -positive braid is a σn−1 -negative braid, each
−1
braid of index n admits a representative either in Rσn or in its inverse
+
−1
−1
+
Rσn = {w−1 | w ∈ Rσn }. As Rσn is regular the language Rσn is also
+
−1
regular. Defining two regular languages Rσ2 = {ak1,2 | k < 0} and Rσ2 =
{ak1,2 | k > 0} we immediately obtain:
Proposition 6.5. Assume n > 2. Then each braid of Bn admits a unique
σ-definite representative lying in
Sn = {ε} ⊔
n
G
−1
+
Rσk ⊔ Rσk .
(14)
k=2
Since the union of regular languages is also regular, we obtain:
Theorem 6.6. For n > 2 the language Sn of σ-definite representative of
braids of Bn is regular.
7. Automaticity
Using syntactical characterization of rotating words we have proved that
the language of n-rotating words is regular. For W a finite-state automaton,
we denote by L(W ) the language recognized by W .
From [5] and [11] we introduce the following definition:
Definition 7.1. Let M be a monoid. A right automatic structure, resp.
left automatic structure, on M consists of a set A of generators of M , a
finite-state automaton W over A, and finite-state automata Mx over (A, A),
for x ∈ A ∪ {ε}, satisfying the following conditions:
(i) the map π : L(W ) → M is surjective,
(ii) for x ∈ A ∪ {ε}, we have (u, v) ∈ L(Mx ) if and only if ux = y, resp.
xu = y, and both u and v are elements of L(W ).
THE ROTATING NORMAL FORM OF BRAIDS IS REGULAR
25
A consequence of Theorem 2.3.5 of [11] is that automaticity for groups
is equivalent to fellow traveller property. As pointed in [5] the situation is
more complicated for monoids. But with Definition 3.11 of [5] for fellow
traveller property on monoids we have, Proposition 3.12 of [5], that right
(resp. left) automaticity of monoids implies right (resp. left) fellow traveller
property of monoids.
As pointed by the anonymous referee the rotating normal form is not
right automatic (or even asynchronously right automatic) for n > 4. For all
integers k > 0, we define two words
w = (a2,3 a1,2 a1,3 )k a3k
1,2 ,
and w′ = (a1,3 a2,3 a1,2 )k a1,4 a3k
2,3 .
The φ4 -splitting of w is (w) and w is 3-rotating by Proposition 5.1 and so 4rotating. The φ4 -splitting of w′ is (a2,3 , ..., a2,3 , 1, a3k
2,3 ). Using Corollary 4.9
{z
}
|
3k+1
we have that w′ is 4-rotating.
We observe that a2,3 a1,2 a1,3 a31,2 ≡ a1,3 a2,3 a1,2 a32,3 = δ33 . Since by [3] the
∗
3k
braid δ33 is in the center of B+
3 we have w ≡ δ3 and
w′ ≡ (a1,3 a2,3 a1,2 )k ak2,3 a1,4 ≡ δ33k a1,4 ≡ wa1,4 .
Finally, w and w′ end with 3k copies of two different letters of A4 , hence
the (asynchronous) fellow traveller property is falsified. As a consequence
we obtain that the n-rotating normal form is not automatic for n = 4 (and
for n > 4 as well).
At this time we don’t know if the rotating normal form is left automatic.
We think that contrary to the previous example, if w is an n-rotating word
and x is an An -letter then the existence of barrier in x prevents migration
to far on the right of x during the computation of rotating normal form of
the braid represented by xw.
Acknowledgments. The author wishes to thank the anonymous referee
for his/her very sharp and constructive comments.
References
[1] E. Artin. Theorie der Zöpfe. Abh. Math. Sem. Univ. Hanburg, 4:47–72, 1925.
[2] D. Bessis. The dual braid monoid. Ann. Sci. École Norm. Sup. (4), 36(5):647–683,
2003.
[3] J. S. Birman, K. H. Ko, and S. J. Lee. A new approach to the word and conjugacy
problems in the braid groups. Adv. Math., 139(2):322–353, 1998.
[4] S. Burckel. The wellordering on positive braids. J. Pure Appl. Algebra, 120(1):1–17,
1997.
[5] C. M. Campbell, E. F. Robertson, N. Ruškuc, and R. M. Thomas. Automatic semigroups. Theoretical Computer Science, 250(1-2):365–391, 2001.
[6] P. Dehornoy. Braid groups and left distributive operations. Trans. Amer. Math. Soc.,
345(1):115–150, 1994.
26
JEAN FROMENTIN
[7] P. Dehornoy. Groups with a complemented presentation. J. Pure Appl. Algebra,
116:115–137, 1997. Special volume on the occasion of the 60th birthday of Professor Peter J. Freyd.
[8] P. Dehornoy. Alternating normal forms for braids and locally Garside monoids. J.
Pure Appl. Algebra, 212(11):2413–2439, 2008.
[9] P. Dehornoy, I. Dynnikov, D. Rolfsen, and B. Wiest. Ordering braids, volume 148 of
Mathematical Surveys and Monographs. American Mathematical Society, Providence,
RI, 2008.
[10] P. Dehornoy and L. Paris. Gaussian groups and Garside groups, two generalisations
of Artin groups. Proc. London Math. Soc. (3), 79(3):569–604, 1999.
[11] D. B. A. Epstein, J. W. Cannon, D. F. Holt, S. V. F. Levy, M. S. Paterson, and
W. P. Thurston. Word processing in groups. Jones and Bartlett Publishers, Boston,
MA, 1992.
[12] J. Fromentin. A well-ordering of dual braid monoids. C. R. Math. Acad. Sci. Paris,
346:729–734, 2008.
[13] J. Fromentin. Every braid admits a short sigma-definite expression. Journal of the
European Mathematical Society (JEMS), 13(6):1591–1631, 2011.
[14] J. Fromentin and L. Paris. A simple algorithm for finding short sigma-definite representatives. J. Algebra, 350:405–415, 2012.
| 4math.GR
|
SalNet360: Saliency Maps for omni-directional images
with CNN
Rafael Monroy∗ , Sebastian Lutz∗ , Tejo Chalasani∗ , Aljosa Smolic
arXiv:1709.06505v1 [cs.CV] 19 Sep 2017
{monroyrr,lutzs,chalasat,smolica}@scss.tcd.ie
Trinity College Dublin, Ireland
Abstract
The prediction of Visual Attention data from any kind of media is of valuable
use to content creators and used to efficiently drive encoding algorithms. With
the current trend in the Virtual Reality (VR) field, adapting known techniques
to this new kind of media is starting to gain momentum. In this paper, we
present an architectural extension to any Convolutional Neural Network (CNN)
to fine-tune traditional 2D saliency prediction to Omnidirectional Images (ODIs)
in an end-to-end manner. We show that each step in the proposed pipeline
works towards making the generated saliency map more accurate with respect
to ground truth data.
Keywords: saliency, omnidirectional image (ODI), convolutional neural
network (CNN), virtual reality (VR)
1. Introduction
The field of Virtual Reality (VR) has experienced a resurgence in the last
couple of years. Major corporations are now investing considerable efforts and
resources to deliver new Head-Mounted Displays (HMDs) and content in a field
that is starting to become mainstream.
One of the applications for VR headsets is the displaying of so-called Omnidirectional Images (ODIs). These images portray an entire scene as seen from
∗ These
authors contributed equally to this work.
Preprint submitted to Elsevier Signal Processing Journal: Image Communications September 19, 2017
a static point of view, and when viewed through a VR headset, allow for an
immersive user experience. The most common method for storing ODIs is by
applying equirectangular, cylindrical or cubic projections and saving them as
standard two-dimensional images [1].
One of the many directions of research in the VR field is Visual Attention,
the main goal of which is to predict the most probable areas in a picture the
average person will look at, by analysing an image. This kind of information is
collected by performing experiments with subjects looking at pictures while an
eye-tracker, together with the Inertial Measurement Unit of the headset, records
the location in the image the user is looking at [2]. By collecting this data from
several subjects, it is possible to create a saliency map that highlights the regions
where most people looked at. Knowing which portions in an image are the
most observed can be used, for example, to drive compression and segmentation
algorithms [3][4].
Earlier works on image saliency made use of manually-designed feature maps
that relate to salient regions and when combined in some form produce a final
saliency map [5][6][7]. With the advent of deep neural networks, GPU-optimised
code and availability of vast amounts of annotated data, research efforts in
computer vision tasks such as recognition, segmentation and detection have been
focused on creating deep learning network architectures to learn and combine
features maps automatically driven by data [8][9][10]. As evident from the MIT
Saliency Benchmarks [11], variants of Convolutional Neural Networks (CNNs)
are the top-performing algorithms for generating saliency maps in traditional
2D images. Good techniques for transfer learning[12] made sure that CNNs can
be used for predicting user gaze even with a relatively small amount of data
[13][14][15][16].
Saliency prediction techniques originally designed for traditional 2D images
cannot be directly used on ODIs due to the heavy distortions present in most
projections and a different nature in observed biases. De Abreu et al. [17]
demonstrated that removing the centre bias present on most techniques for
traditional 2D images significantly improves the performance when estimating
2
saliency maps for ODIs.
A recent effort to attract attention to the problem of creating saliency maps
for ODIs was presented in the Salient360! Grand Challenge at the ICME 2017
Conference[18]. Participants were provided with a series of ODIs (40 in total)
for which head- and eye-tracking ground truth data was given in the form of
saliency maps. The details on how the dataset was built are described by Rai
et al. [2]. Twenty-five paired images and ground truth were later provided to
properly evaluate the submissions to the challenge. In the presented paper we
follow the experimental conditions of the challenge, i.e. we used the initial 40
images to train the proposed CNN and the results we present were calculated
using the 25 test images.
The approach we present is similar to that of De Abreu et al. [17] in that
the core CNN to create the saliency maps can be switched once better networks
become available. We start with the premise that the heavy distortions near the
poles in an equirectangular ODI negatively affect the final saliency map and the
nature of the biases near the equator and poles differs from those in a traditional
2D image. To address these issues, we make the following contributions:
• Subdividing the ODI into undistorted patches.
• Providing the CNN with the spherical coordinates for each pixel in the
patches.
This paper is organised as follows. Section 2 enlists work related to saliency
maps for traditional 2D images and ODIs. Then, Section 3 describes the pipeline
and the required pre- and post-processing steps. The end-to-end trainable CNN
architecture we use here is then described in Section 4. Results are presented
in Section 5. Finally, in Section 6 ideas for future research opportunities and
conclusions are provided.
2. Previous work
Until the advent of CNNs, most saliency models relied on extracting feature
maps, which depict the result of decomposing the image, highlighting a particu3
lar characteristic. These feature maps are then linearly combined by computing
a series of weights to be applied to each map. Features like those presented
by Itti and Koch [5] were initially based on knowledge of the low-level human
vision, e.g. color, orientation and frequency responses.
The catalogue of features used to calculate saliency maps has increased over
the years, including for example, global features [6], face detection features [19],
location bias features [20] and others. Judd et al.[7] created a framework and
model that combined many of these features to compute saliency maps.
The MIT Saliency Benchmark was later developed by Bylinskii et al., allowing researchers to submit and compare their saliency computation algorithms
[11].
With the increasing availability of VR headsets, saliency computation methods specifically tailored for ODIs have started to surface. Due to the fact that
ODIs describe, in actuality, a sphere, saliency models designed for traditional 2D
images cannot be directly used. Bogdanova et al. analyse the sphere depicted
in an ODI, creating low-level vision feature maps based on intensity, colour and
orientation features [21].
In recent years, the interest in CNNs has grown and most of the topperforming submissions in the MIT Saliency Benchmark system correspond to
CNN-based approaches. One of the first methods to use CNNs for saliency was
presented by Vig et. al. [22], which used feature maps from convolution layers
as the final features of a linear classifier. Kümmerer et. al. [12] published an
extension of this approach to use CNNs as feature extractors. They extended
their approach in [23], where they use the VGG [24] features to predict saliency
with no fine-tuning. Instead they only train a few readout layers on top of the
VGG layers. The method developed by Liu et. al. [13] used a Multi-Resolution
CNN, where the network was trained with image regions centred on fixation
and non-fixation eye locations at several scales. Another approach using CNNs
to extract features was shown by Li et. al. [25], where they obtained features
from convolutional layers at several scales and combined them with handcrafted
features which were trained on a Random Forest. One of the first end-to-end
4
Figure 1: ODI Saliency Detection Pipeline.
trained CNNs for saliency prediction was introduced by Pan et. al. [16]. Similar
work from Kruthiventi et al. [15] added a novel Location Biased Convolutional
Layer, which allowed them to model location dependent patterns.
As CNNs are becoming more complex and sophisticated in other Computer
Vision domains, these advances are being transferred to saliency prediction applications, for example, the method proposed by Pan et. al. [26], introducing
adversarial examples to train their network. The work of Liu et. al. [27] introduces a deep spatial contextual LSTM to incorporate global features in their
saliency prediction.
The amount of published research focusing on CNNs applied to ODIs is
at the moment very small. De Abreu et. al. [17] uses translated versions of
the ODI as input to a CNN trained for traditional 2D images to generate an
omnidirectional saliency map.
3. Method
In this section we detail the pipeline used to obtain the saliency map for a
given ODI. Fig. 1 shows a diagram with the complete pipeline. Our method
takes an ODI as input and splits it into six patches using the pre-processing
steps described in 3.1. Each of these six patches is sent through the CNN, the
details of which we delve into in Section 4. The output of the CNN for all the
5
(a) Full ODI
(b) Nadir view
Figure 2: Example of distortions present in equirectangular ODIs.
patches are then combined using the post-processing technique mentioned in
Section 3.2
3.1. Pre-processing
Mapping a sphere onto a plane requires introducing heavy distortions to the
image, which are most evident at the poles. When looking at those distorted
areas in the ODI, it becomes very hard to identify what they are supposed
to be depicting. Fig. 2 gives an example of the distortions observed on an
equirectangular ODI. From Fig. (2a) alone, it is not possible to recognise the
object on the table. Once the nadir view is undistorted as seen in Fig. (2b), it
is clear that the object in question is a cake.
In order to reduce the effect of these distortions on saliency estimation, we
divide the ODI into equally-sized patches by rendering viewing frustums with
a field of view of (FOV) approximately 90 degrees each. This field of view was
selected to keep distortions low, cover the entire sphere with six patches and
have a similar FOV to that of the Oculus Rift, which was used to create the
dataset (approximately 100 degrees).
By specifying the field of view per patch and its resolution, it is possible
to calculate the spherical coordinates of each pixel in the patch. These are
then used to find the corresponding pixels in the ODI by applying the following
equations:
6
y
y
ϕ
x
x
θ
z
z
(a) Spherical Coordinates
(b) Sliding frustum
Figure 3: Spherical coordinates definition and sliding frustum used to create the patches.
π
2
φ+
y = sh 1 −
π
π
2
x = sw
θ +
2π
(1)
(2)
Where θ and φ are the spherical coordinates of each pixel, see Fig. 3a. The
variables sw and sh correspond to the ODI’s width and height respectively. A
graphical representation of how patches look like when sampling the sphere can
be seen in Fig. 3b.
The process of generating patches is also applied during training, which will
be discussed in Section 4.2. During the saliency map computation six patches
with fixed views are generated. Two of these views are oriented towards the
nadir and zenith, the other four are pointed towards the horizon but rotated
horizontally to cover the entire band at the sphere’s equator. Fig. 4 illustrates
such partitions when applied to an equirectangular ODI.
3.2. Post-processing
As previously mentioned, the CNN takes the six patches and their spherical
coordinates as inputs. As result, the CNN generates a saliency map for each of
7
Figure 4: Patches extracted from the ODI.
these patches. Consequently, they have to be combined to a single saliency map
as output. For that, we project each pixel of each patch to the equirectangular
ODI, using their per-pixel spherical coordinates. We apply forward-projection
with nearest-neighbour interpolation for simplicity. In order to fill holes and
smooth the result, we apply a Gaussian filter with a kernel size of 64 pixels. This
kernel size was selected because it was found to give the best results in terms
of correlation with the ground truth. Such processing is efficient and sufficient,
as we do not compute output images for viewing, but estimate saliency maps.
4. SalNet360
Since ODIs depict an entire 360-degree view, their size tend to be considerably large. This factor, together with the current hardware limitations, prohibits
using them directly as inputs in a CNN without heavily down-scaling the ODI.
Deep CNNs need large amounts of data to avoid over-fitting. The deeper the
network is, the better its accuracy in general [16], but that also means the parameter space increases accordingly. Since training data consisted of only 40
images, the resulting amount of data would not be enough for properly training
a CNN.
8
Figure 5: Network Architecture.
To address this problem we generated one hundred patches (paired colour
image and ground truth saliency map) per ODI by placing the viewing frustum
to random locations. The saliency maps per patch are used as labels in the
CNN. Fig. 5 illustrates the architecture of the proposed network.
Our network consists of two parts, the Base CNN and the refinement architecture. The Base CNN is trained to detect saliency maps for traditional 2D
images. It has been pre-trained using the SALICON dataset [28]. The second
part is a refinement architecture that is added after the Base CNN. It takes
a 3-channel feature map as input: the output saliency map of the Base CNN
and the spherical coordinates per pixel as two channels. This combination of
the Base CNN and the Saliency Refinement has been trained as will be discussed in Section 4.2. Omnidirectional saliency maps differ from traditional 2D
saliency maps in that they are affected by the view or position of the users head.
Combining the spherical coordinates of each pixel as extra input provides the
network in the second stage the information to highlight or lower the already
computed salient regions depending on their placement in the ODI. We delve
into the details of network architecture and training in the next subsections.
9
Table 1: Network parameters.
Layer
Input depth
Output depth
Kernel size
Stride
Padding
Activation
conv1
3
96
7×7
1
3
ReLU
pool1
3
3
3×3
2
0
-
conv2
96
256
5×5
1
2
ReLU
pool2
256
256
3×3
2
0
-
conv3
256
512
3×3
1
1
ReLU
conv4
512
256
5×5
1
2
ReLU
conv5
256
128
7×7
1
3
ReLU
conv6
128
32
11×11
1
5
ReLU
conv7
32
1
13×13
1
6
ReLU
deconv1
1
1
8×8
4
2
-
merge
-
-
-
-
-
-
conv8
3
32
5×5
1
2
ReLU
pool3
32
32
3×3
2
0
-
conv9
32
64
3×3
1
2
ReLU
conv10
64
32
5×5
1
2
ReLU
conv11
32
1
7×7
1
3
ReLU
deconv2
1
1
4×4
2
1
-
4.1. Network Architecture
The architecture of our Base CNN was inspired by the deep network introduced by Pan et. al. [16] and shares the same layout as the VGG CNN M
architecture from [29] in the first three layers. This allowed us to initialise the
weights of these layers with the weights that have been learned on the ImageNet classification task. After the first three layers, four more convolution
layers follow, each of them followed by a Rectified Linear Unit (ReLU) activation function. Two max-pooling layers after the first and second convolution
layers reduce the size of the feature maps for subsequent convolutions and add
some translation invariance. To upscale the final feature map to a saliency map
with the same dimensions as the input image, a deconvolution layer is added at
the end. As will be discussed in Section 4.2, we train our whole network in two
stages. In the first stage only this Base CNN is trained on traditional 2D images with the help of an Euclidean Loss function directly after the deconvolution
layer. In our final network, however, this loss function is moved to the end of
our refinement architecture and the output of the deconvolution layer is merged
with the 2-channel per-pixel spherical coordinates as input to the second part
of our architecture, the Saliency Refinement.
As has been stated above, our architecture does not use the entire ODI at the
10
same time. Instead, the ODI is split into patches and for each of these patches
the saliency map is calculated individually. The idea of the second part of our
architecture, the Saliency Refinement, is to take the saliency map generated
from the Base CNN and refine it with the information on where in the ODI this
patch is located. For example, this allows us to model the strong bias towards
the horizon that the ground truth saliency maps for ODIs show. The whole
Saliency Refinement stage consists of four convolution layers, one max-pooling
layer after the first convolution and one deconvolution layer at the end. All
activation functions are ReLUs and the loss function at the end is Euclidean
Loss. Our full architecture is visualised in Fig. 5 and the hyperparameters for
all layers are presented in Table 1.
We experimented with different activation layers and different loss functions
in the network, however, we found that using ReLUs and Euclidean Loss yielded
the best metrics on the dataset.
4.2. Training
We performed training in two stages. In the first stage we only train the
first part of the network, the Base CNN. The first three layers are initialised
using pre-trained weights from VGG CNN M network from [29]. We then use
SALICON data [28] to train the first part of the network. The data (both images
and their saliencies) is normalised by removing the mean of pixel intensity values
and re-scaling the data to a [-1,1] interval. We also scale all the images and
saliency maps to 360x240 resolution and split the dataset into two sets of 8000
images for training and 2000 images for testing. The network is then trained for
20,000 iterations using the stochastic gradient descent method and with batch
size of four.
For the second stage of the training we add the Saliency Refinement part to
the network and use the dataset we created with ODIs from [2]. As described in
Section 3.1, 100 patches are randomly sampled per ODI to generate a dataset
of 4000 images, their ground truth saliency and spherical coordinates. We preprocess all the images using the same techniques we used for the first stage
11
of training. The weights for the Base CNN are initialised using the weights
we obtained from the first stage training. We started with a base learning
rate of 1.3e-7 and reduced it by 0.7 after every 500 iterations. The network
is trained for 22,000 iterations with a 10% split for test data and the rest of
the images are used for training. The batch size is set to be five while the test
error is monitored every 100 iterations to look for divergence. We present our
experiments and results in the next section.
5. Results
In this section we describe the experiments and results that indicate that
our method enhances a CNN trained with traditional 2D images and allows
it to be applied to ODIs. Before we discuss the actual numbers, we provide a
short introduction to each of the metrics used to evaluate the generated saliency
maps.
5.1. Performance measures
Bylinskii et al. [30] differentiate between two types of performance metrics
based on their required ground truth format. Location-based metrics need a
ground truth with values at discrete fixation locations, while distribution-based
metrics consider the ground truth and saliency maps as continuous distributions.
Four metrics were used to evaluate the results of our system: the KullbackLeibler divergence (KL), the Pearson’s Correlation Coefficient (CC), the Normalized Scanpath Saliency (NSS) and the Area under ROC curve (AUC). Both
the KL and the CC are distribution-based metrics, whereas the NSS and AUC
are location-based. In the case of the former two metrics, the predicted saliency
map of our approach is normalised to generate a valid probability distribution.
For further details on these metrics, we refer to Bylinskii et al. [30].
5.2. Experiments
In order to test the improvements obtained from our proposed method, we
defined three scenarios to be compared. The first one consists of applying our
12
Table 2: Comparison of the three experimental scenarios.
∗ indicates a significant improvement in performance compared to the Base CNN (t-test,
p < 0.01).
KL
CC
NSS
AUC
Base CNN
1.597
0.416
0.630
0.648
Above + Patches
0.625
0.474
0.566
0.659
Above + Spherical Coords.
0.458*
0.548*
0.755*
0.701*
Base CNN, see Fig. 5 while taking the entire ODI as input by downscaling it to a
resolution of 800×400 pixels. The predicted saliency map is then upscaled to the
resolution of the original image. In our second scenario, we divide the ODI in six
undistorted patches as described in Section 3.1 and run these patches separately
through the Base CNN. Afterwards we recombine the predicted saliencies of
these patches to form the final saliency map. Finally, the third scenario consists
of the whole pipeline as described in Section 3, where the spherical coordinates
of the patches are also considered by the CNN during inference.
The results of our experiments can be seen in Table 2, where we show the
average KL, CC, NSS and AUC for the 25 test images. As can be seen from the
table, only running the entire ODI through the Base CNN gives relatively poor
results compared to the later scenarios. By dividing the ODI into patches and
using them individually before being recombined, our results are improved in
most of the metrics. After adding in the spherical coordinates to the inference
of each patch, the results clearly improve considerably in all the metrics.
A visual example of the results of the three scenarios can be seen in figure
6. In the top row from left to right, the input ODI and a blended version of
the ODI and ground truth are shown. In the bottom row again from left to
right the predicted saliency maps blended into the input ODI for each scenario
are presented. The first scenario on the far left (Base CNN only) predicted two
big salient centres, which are not found in the ground truth map. The second
scenario improved on these predictions by finding salient areas that also cover
salient areas in the ground truth. However, it still predicts two highly salient
areas in wrong locations and introduced some salient areas of low impact at
13
Figure 6: Comparison of the three experimental scenarios. Top row: On the left the input
ODI, on the right the ground truth saliency map blended with the image.
Bottom row: From left to right, the result of the three experimental scenarios: Base CNN,
Base CNN + Patches, Base CNN + Patches + Spherical Coords.
the bottom part of the image that are incorrectly labelled as salient. Finally,
the third scenario clearly covers all the salient areas in the ground truth and
removes the incorrectly labelled salient regions at the bottom of the image that
the second scenario introduced. It is, however, too generous in the amount of
area that the salient regions cover, which seems to be the main issue in this
scenario.
In Fig. 8, some of the best-performing results are shown. As a comparison,
two of the lowest-performing results are shown in Fig. 9. A summary of the results of all images can be seen in Table 3. Fig. 7 provides a visual representation
of the values obtained for each metric.
5.3. Issues
As can be seen on the predicted saliency maps in Figs. 8 and 9, the merging
of the predicted patches leaves a distinct pattern in the final saliency map,
leaving in some cases a lattice-like pattern. We tried to mitigate this issue by
applying a Gaussian Blur, but it is still noticeable in some of the results and
consequently has a negative effect on the KL and CC scores. In the saliency
prediction of the individual patches in both figures, another set of artefacts can
also be observed. We speculate that these line patterns stem from the addition
of the spherical coordinates, because these predictions correspond to the patches
14
Table 3: Individual results for each ODI. The two best and worst results have been highlighted
in green and red respectively.
Image index
KL
CC
NSS
AUC
1
0.589
0.507
0.749
0.702
8
0.484
0.655
1.055
0.772
16
0.249
0.610
0.396
0.629
19
0.384
0.512
0.739
0.693
20
0.618
0.497
0.787
0.729
26
0.276
0.596
0.686
0.684
48
0.622
0.266
0.616
0.702
50
0.479
0.590
0.983
0.746
60
0.273
0.671
0.459
0.663
65
0.475
0.531
0.817
0.737
69
0.433
0.686
1.064
0.769
71
0.404
0.388
0.522
0.662
72
0.459
0.271
0.353
0.609
73
0.364
0.536
0.690
0.695
74
0.385
0.656
0.589
0.663
78
0.514
0.581
0.903
0.716
79
0.360
0.818
1.303
0.779
85
0.351
0.714
0.975
0.751
91
0.498
0.628
0.892
0.700
93
0.496
0.488
0.761
0.694
94
0.583
0.468
0.871
0.763
95
0.455
0.624
0.894
0.703
96
0.582
0.722
1.270
0.808
97
0.471
0.369
0.371
0.626
98
0.652
0.309
0.126
0.529
Mean
0.458
0.548
0.755
0.701
Std. Dev.
0.112
0.144
0.286
0.062
15
KL
CC
0.7
0.9
0.6
0.8
0.7
0.5
0.6
0.4
0.5
0.3
0.4
0.3
0.2
0.2
0.1
0
0.1
0
1 8 16 19 20 26 48 50 60 65 69 71 72 73 74 78 79 85 91 93 94 95 96 97 98
Score
Average
Score
NSS
Average
AUC
1.4
0.9
1.2
0.8
0.7
1
0.6
0.8
0.5
0.6
0.4
0.3
0.4
0.2
0.2
0
1 8 16 19 20 26 48 50 60 65 69 71 72 73 74 78 79 85 91 93 94 95 96 97 98
0.1
0
1 8 16 19 20 26 48 50 60 65 69 71 72 73 74 78 79 85 91 93 94 95 96 97 98
Score
Average
1 8 16 19 20 26 48 50 60 65 69 71 72 73 74 78 79 85 91 93 94 95 96 97 98
Score
Average
Figure 7: Plots for each of the metrics applied to the test ODIs.
that show the poles. However, when recombining the patches and applying the
post-processing steps discussed above, these artefacts are no longer noticeable.
We expect that increasing the amount of training data could help alleviate some
of the issues.
6. Conclusions
We showed in Section 5, that dividing an omnidirectional image into patches
and adding a Saliency Refinement architecture that takes into consideration
spherical coordinates to an existing Base CNN can considerably improve the
results in omnidirectional saliency prediction. We envision several potential
improvements that could be made to increase performance. Our network was
trained using the Euclidean Loss function, which is a relatively simple loss function that is applicable to a wide variety of cases, such as saliency prediction. It
tries to minimise the pixel-wise difference between the calculated saliency map
and the ground truth. However, to optimise based on any of the performance
metrics mentioned in 5.1, custom loss functions could be used. In this way, the
network would be trained to specifically minimise in regards to the KL, CC or
16
(a) Image index: 69, KL: 0.433, CC: 0.686, NSS: 1.064, AUC: 0.769
(b) Image index: 79, KL: 0.360, CC: 0.818, NSS: 1.303, AUC: 0.779
Figure 8: Two of the best-performing examples. Top row: From left to right: Input image, six
extracted patches, ground truth blended with input image. Bottom row: From left to right:
Predicted saliency map, predicted saliency maps from each patch, ground truth saliency map.
NSS.
As mentioned in Section 5.3, one of the bigger issues that affect our results
are the artefacts that are created when recombining the patches to create the
final saliency map. Instead of just using Gaussian Blur to remove these artefacts,
a more sophisticated process could be implemented since these artefacts always
appear in the same way and location.
In most Deep Learning applications, improvements are often made by making the networks deeper. Our network is based on the Deep Convolutional
Network by Pan et. al. [16], but compared to the current state-of-the-art in
other Computer Vision tasks e.g. classification and segmentation, this network
17
(a) Image index: 72, KL: 0.459, CC: 0.271, NSS: 0.353, AUC: 0.609
(b) Image index: 98, KL: 0.652, CC: 0.309, NSS: 0.126, AUC: 0.529
Figure 9: Two of the lowest-performing examples. Top row: From left to right: Input image,
six extracted patches, ground truth blended with input image. Bottom row: From left to right:
Predicted saliency map, predicted saliency maps from each patch, ground truth saliency map.
is still relatively simple. We are confident that updating the Base CNN to recent
advances will improve the results in omnidirectional saliency.
Finally, as with all Deep Learning tasks, a large amount of data is needed
to achieve good results. Unfortunately there is currently only a relatively small
amount of ground truth saliency maps available, and even less data for omnidirectional saliency. It is our hope that more data will become available in the
future which can be used to get better results.
In this work, we present an end-to-end CNN that is specifically tailored for
estimating saliency maps for ODIs. We provide evidence which indicates that
part of the discrepancies found when using CNNs trained on traditional 2D im18
ages is due to the heavy distortions found on the projected ODIs. These issues
can be addressed by dividing the ODI in undistorted patches before calculating the saliency map. Furthermore, in addition to the patches, biases due to
the location of objects on the sphere can be considered by using the spherical
coordinates of the pixels in the patches before computing the final saliency map.
Acknowledgement
The present work was supported by the Science Foundation Ireland under
the Project ID: 15/RP/2776 and with the title V-SENSE: Extending Visual
Sensation through Image-Based Visual Computing.
19
References
[1] M. Yu, H. Lakshman, B. Girod, A framework to evaluate omnidirectional
video coding schemes, in: IEEE Int. Symposium on Mixed and Augmented
Reality, 2015.
[2] Y. Rai, J. Gutirrez, P. Le Callet, A dataset of head and eye movements for
omni-directional images, in: Proceedings of the 8th International Conference on Multimedia Systems, ACM, 2017.
[3] L. Itti, Automatic foveation for video compression using a neurobiological
model of visual attention, in: IEEE Transactions on Image Processing,
Vol. 13, 2004.
[4] B. C. Ko, J.-Y. Nam, Object-of-interest image segmentation based on human attention and semantic region clustering, in: Journal of Optical Society of America A, Vol. 23, 2006.
[5] L. Itti, C. Koch, A saliency-based mechanism for overt and covert shift of
visual attention, in: Vision Research, Vol. 40, 2000.
[6] A. Torralba, A. Oliva, M. Castelhano, J. Henderson, Contextual guidance
of eye movements and attention in real-world scenes: The role of global
features in object search, in: Psychological Review, Vol. 113, 2006.
[7] T. Judd, K. Ehinger, F. Durand, A. Torralba, Learning to predict where
humans look, in: ICCV, 2009.
[8] A. Krizhevsky, I. Sutskever, G. E. Hinton, Imagenet classification with
deep convolutional neural network, in: Advances in neural information
processing systems, 2012, pp. 1097–1105.
[9] R. Girshick, J. Donahue, T. Darrell, J. Malik, Rich feature hierarchies for
accurate object detection and semantic segmentation, in: CVPR, 2014, pp.
580–587.
20
[10] J. Long, E. Shelhamer, T. Darrell, Fully convolutional networks for semantic segmentation, in: CVPR, 2015, pp. 3431–3440.
[11] Z. Bylinskii, T. Judd, A. Borji, L. Itti, A. O. F. Durand, A. Torralba, Mit
saliency benchmark, http://saliency.mit.edu/results_mit300.html.
[12] M. Kümmerer, L. Theis, M. Bethge, Deep gaze i: Boosting saliency prediction with feature maps trained on imagenet, in: International Conference
on Learning Representations, 2015.
[13] N. Liu, J. Han, D. Zhang, S. Wen, T. Liu, Predicting eye fixations using
convolutional neural networks, in: CVPR, 2015.
[14] M. Cornia, L. Baraldi, G. Serra, R. Cucchiara, A deep multi-level network
for saliency prediction, in: CoRR, 2016.
[15] S. Kruthiventi, K. Ayush, R. Babu, Deepfix: A fully convolutional neural network for predicting human eye fixations, in: IEEE Transactions on
Image Processing, 2017.
[16] J. Pan, E. Sayrol, X. G. i Nieto, K. McGuinness, N. O’Connor, Shallow
and deep convolutional networks for saliency prediction, in: CVPR, 2016.
[17] A. D. Abreu, C. Ozcinar, A. Smolic, Look around you: Saliency maps for
omnidirectional images in vr applications, in: QoMEX, 2017.
[18] Salient360!: Visual attention modeling for 360◦ images grand challenge,
http://www.icme2017.org/grand-challenges/, accessed: 2017-06-06.
[19] M. Cerf, J. Harel, W. Einhauser, C. Koch, Predicting human gaze using
low-level saliency combined with face detection, in: Advances in Neural
Information Processing Systems, Vol. 20, 2008.
[20] B. W. Tatler, The central fixation bias in scene viewing: Selecting an optimal viewing position independently of motor biased and image feature
distribution, in: Journal of Vision, Vol. 7, 2007.
21
[21] I. Bogdanova, A. Bur, H. Hugli, Visual attention on the sphere, in: IEEE
Transactions on Image Processing, Vol. 17, 2008.
[22] E. Vig, M. Dorr, D. Cox, Large-scale optimization of hierarchical features
for saliency prediction in natural images, in: CVPR, 2014.
[23] M. Kümmerer, T. S. Wallis, M. Bethge, Deepgaze ii: Reading fixations from deep features trained on object recognition, in: arXiv preprint
arXiv:1610.01563, 2016.
[24] K. Simonyan, A. Zisserman, Very deep convolutional networks for largescale image recognition, in: CoRR, Vol. abs/1409.1556, 2014.
[25] G. Li, Y. Yu, Visual saliency detection based on multiscale deep cnn features, in: IEEE Transactions on Image Processing, Vol. 25, 2016.
[26] J. Pan, C. Canton, K. McGuinness, N. O’Connor, J. Torres, E. Sayrol,
X. G. i Nieto, Salgan: Visual saliency prediction with generative adversarial
networks, in: ArXiv, 2017.
[27] N. Liu, J. Han, A deep spatial contextual long-term recurrent convolutional
network for saliency detection, in: arXiv preprint arXiv:1610.01708, 2016.
[28] M. Jiang, S. Huang, J. Duan, Q. Zhao, Salicon: Saliency in context, in:
CVPR, 2015.
[29] K. Chatfield, K. Simonyan, A. Vedaldi, A. Zisserman, Return of the devil
in the details: Delving deep into convolutional nets, in: British Machine
Vision Conference, 2014.
[30] Z. Bylinskii, T. Judd, A. Oliva, A. Torralba, F. Durand, What do different
evaluation metrics tell us about saliency models?, in: ArXiv, 2016.
22
| 1cs.CV
|
D-MODULES, BERNSTEIN-SATO POLYNOMIALS AND
F -INVARIANTS OF DIRECT SUMMANDS
arXiv:1611.04412v1 [math.AC] 14 Nov 2016
JOSEP ÀLVAREZ MONTANER1 , CRAIG HUNEKE2 , AND LUIS NÚÑEZ-BETANCOURT3
Abstract. We study the structure of D-modules over a ring R which is a direct summand of a polynomial or a power series ring S with coefficients over a field. We relate
properties of D-modules over R to D-modules over S. We show that the localization
Rf and the local cohomology module HIi (R) have finite length as D-modules over R.
Furthermore, we show the existence of the Bernstein-Sato polynomial for elements in R.
In positive characteristic, we use this relation between D-modules over R and S to show
that the set of F -jumping numbers of an ideal I ⊆ R is contained in the set of F -jumping
numbers of its extension in S. As a consequence, the F -jumping numbers of I in R form
a discrete set of rational numbers. We also relate the Bernstein-Sato polynomial in R
with the F -thresholds and the F -jumping numbers in R.
Contents
1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2. Background on D-modules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3. D-modules over direct summands . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.1. Differentiable direct summands . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.2. Finite length of D-modules over direct summands . . . . . . . . . . . . . . . . . . . . . .
3.3. Bernstein-Sato polynomial over direct summands. . . . . . . . . . . . . . . . . . . . . . .
4. F -invariants of direct summands . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.1. Test ideals of direct summands . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.2. F -thresholds and Bernstein-Sato polynomials . . . . . . . . . . . . . . . . . . . . . . . . . .
Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1
4
5
6
9
10
13
14
18
21
21
1. Introduction
Direct summands of polynomial rings play an important role in representation theory,
combinatorial algebra, commutative algebra, and algebraic geometry. For instance, rings
of invariants under a linearly reductive group actions, and affine toric rings are direct
summands of polynomial rings. Rings corresponding to Grassmannian, Veronese and
2010 Mathematics Subject Classification. Primary 14F10, 13N10, 13A35, 16S32 ; Secondary 13D45,
14B05, 14M25, 13A50 .
Key words and phrases. D-modules; Bernstein-Sato polynomial; direct summands; ring of invariants;
local cohomology; F -jumping numbers; test ideals; F -thresholds.
1
The first author was partially supported by Generalitat de Catalunya 2014SGR-634 project and
Spanish Ministerio de Economía y Competitividad MTM2015-69135-P.
2
The second author was partially supported by the NSF Grant 1460638.
3
The third author was partially supported by the NSF Grant 1502282 .
1
2
J. ÀLVAREZ MONTANER, C. HUNEKE, AND L. NÚÑEZ-BETANCOURT
Segre varieties also belong to this family of rings. If one assumes that the field has
characteristic zero, then the ring associated to the n × m matrices of rank at most t is a
direct summand of a polynomial ring. Since the Hochster-Roberts Theorem [HR74], which
states that direct summands of regular rings are Cohen-Macaulay, it has been expected
that direct summands behave similarly to regular rings in certain aspects. For instance,
direct summands are normal rings [HH89] with rational singularities in characteristic zero
[Bou87] or strongly F -regular singularities in prime characteristic [HH89]. The main goal
of this manuscript is to find structural properties of important D-modules over direct
summands of regular rings, which resemble the situation over regular rings.
We first focus on modules over the ring of K-linear differential operators, DR|K , where
K is a field, and R is a direct summand of a regular K-algebra S. Significant research
has been focused on the finite generation of DR|K for rings of invariants [Kan77, LS95a,
Lev81, Tra06, Sch95, MVdB98, LS95b, PR05], Grassmann varieties [Tra10], and toric
affine algebras [Mus87, Jon94, Mus94, ST01, ST04, ST09]. In addition, the simplicity
of DR|K as a ring and the simplicity of R as DR|K -module have also been a topic of
interest [SVdB97, Smi95, Tra00]. Unfortunately, the rings of differential operators are not
always finitely generated for singular rings, even in characteristic zero [BGG72]. Instead
of focusing on the ring DR|K , we switch gears and study the structure of an important
class of DR|K -modules.
We study the structure of localization and local cohomology modules over direct summands. For this we recall their finiteness properties over regular rings. Suppose that
S is either K[x1 , . . . , xn ] or K[[x1 , . . . , xn ]]. If K has characteristic zero, then the ring
of K-linear differential
operators,
DS|K , equals the S-algebra generated by the partial
E
D
derivatives, S
∂
, . . . , ∂x∂
∂x1
n
. In particular, DS|K is a left and right Noetherian ring. Fur-
thermore, every localization Sf , and therefore every local cohomology module HIi (S), has
finite length as DS|K -module. If K has prime characteristic, DS|K is no longer Noetherian. However, Sf and HIi (S) still have finite length as DS|K -modules. These finiteness
properties have a significant impact on the study of local cohomology over regular rings
[Lyu93, Lyu97]. Our first result recovers these properties for direct summands.
Theorem A (see Theorem 3.8). Let K be a field and S be either K[x1 , . . . , xn ] or
K[[x1 , . . . , xn ]], and R be a K-subalgebra. Suppose that R is a direct summand of S.
Then Rf and HIi (R) have finite length as DR|K -modules for every f ∈ R and any ideal
I ⊆ R. Furthermore,
i
λDR|K (Rf ) ≤ λDS|K (Sf ) and λDR|K (HIi (R)) ≤ λDS|K (HIS
(S)).
The previous theorem holds for a larger class of modules (see Theorem 3.10), which includes HIi11 (· · · (HIiℓℓ (R)) · · · ). Lyubeznik introduced a category of DR|K -modules [Lyu00a,
Lyu00b], denoted by C(R, K), to compensate for the lack of the notion of holonomic Dmodules in prime and mixed characteristic at the time. The category C(R, K) resembles
the class of holonomic modules in the sense that both consist of D-modules of finite length
in equal characteristic.
Since its introduction [Ber72, SS72], the Bernstein-Sato polynomial bf (s) has been an
important invariant of a hypersurface in characteristic zero. In fact, bf (s) is an important
object in the study of singularities. For instance, it relates to the theory of vanishing cycles
[Del71], V -filtrations and monodromies [Mal83, Kas83], jumping numbers of multiplier
D-MODULES, BERNSTEIN-SATO POLYNOMIALS & F -INVARIANTS
3
ideals [Kol97, ELSV04, BS05], and zeta functions [DL92]. In our second main result, we
develop the theory of the Bernstein-Sato polynomial over a direct summand. To the best
of our knowledge, this is one of the first efforts to extend the theory of the Bernstein-Sato
polynomial for singular rings (see [HM16] for an study of Bernstein-Sato polynomials of
ideals in a normal toric ring).
Theorem B (see Theorem 3.14). Let K be a field of characteristic zero, let S be either
K[x1 , . . . , xn ] or K[[x1 , . . . , xn ]], and let R be a K-subalgebra. Suppose that R is a direct
summand of S. Then for every element f ∈ R \ {0}, there exists δ(s) ∈ DR|K [s] and
b(s) ∈ Q[s] such that
(1.0.1)
δ(t) · f t+1 = b(t)f t
for every t ∈ Z. As a consequence, Rf is a cyclic DR|K -module.
Under the hypothesis of the previous theorem, we call the Bernstein-Sato polynomial
of f in R, bR
f (s), the monic polynomial of smallest degree satisfying the Equation 1.0.1
for some δ(s). As a consequence of the proof of Theorem B, the roots of bR
f (s) consists of
negative rational numbers when R is a direct summand of a polynomial ring. We hope
that bR
f (s) relates to other invariants that measure singularities of hypersurfaces in direct
summands.
In Remark 3.18, we show how different versions of the Bernstein-Sato theory can be
extended to direct summands, which include the one given by Sabbah [Sab87] and the
one given by Budur, Mustaţǎ, and Saito [BMS06a]. Very recently, Hsiao and Matusevish
extended and studied Bernstein-Sato polynomials associated to ideals in a normal toric
ring [HM16]. We note that not every K-algebra has Bernstein-Sato polynomials (see
Example 3.19).
We point out that it was previously known that if R is an affine toric ring and I ⊆ R
is a monomial ideal, then HIi (R) has finite length as DR|K module [Hsi12]. In this case,
it was also shown that if f is a monomial, then Rf is a cyclic DR|K -module [Hsi12]. Even
for toric rings, Theorems A and B recover and extend these results, as the monomial
hypothesis is not longer needed.
In the second part of the paper we shift to positive characteristic to study the ring of
differential operators, DR . We show that Rf is generated by f1 as DR -module for direct
summands of regular rings (see Proposition 4.4). We use this fact and the ideas behind the
proof of Theorems A and B to study the F -jumping numbers of R. These invariants are
used to measure singularities in prime characteristic. Specifically, the F -jumping numbers
are the values where the generalized test ideals change [HY03]. The test ideals can be
seen as an analogue, in prime characteristic, of multiplier ideals [Smi97, Smi00, HY03].
In particular, the F -pure threshold, which is the first F -jumping number, serves as the
analogue of the log-canonical threshold [HY03, MTW05]. We recall that multiplier ideals
are defined using resolution of singularities (see [Laz04]). In contrast, test ideals are
defined in terms of tight closure theory [HH90, HH94a, HH94b, HY03] so, a priori, it
is not clear that F -jumping numbers form a discrete set of rational numbers. Since
the introduction of the test ideals [HY03], intense research has been devoted around
this question [BMS08, BMS09, Sch11, BSTZ10, ST14b, CEMS]. Efforts have also been
dedicated to compute these invariants, mainly over regular rings [Her14, ST14b, Her15,
HNBWZ16]. Our main result in this direction is that the set of the F -jumping numbers
of R is a subset of F -jumping numbers of S.
4
J. ÀLVAREZ MONTANER, C. HUNEKE, AND L. NÚÑEZ-BETANCOURT
Theorem C (see Theorem 4.12 and Corollary 4.13). Let S be a regular F -finite domain,
and R be an F -finite ring. Suppose that R ⊆ S and that R is a direct summand of S.
Let I ⊆ R denote an ideal. Then, the set F -jumping numbers of I in R is a subset of
the set F -jumping numbers of IS in S. In particular, the set F -jumping numbers of I in
R is formed by rational numbers and has no accumulation points. As a consequence, the
F -pure threshold of I is a rational number.
The previous result is relevant when the extension R → S is not finite, since the finite
case follows from the work of Schwede and Tucker [ST14a].
Theorem C gives a set of candidates for the F -jumping numbers in R, which could
potentially give algorithms to compute the invariants for direct summands (see Remark
4.14).
Motivated by the work of Mustaţǎ, Takagi and Watanabe [MTW05] for smooth complex varieties, we relate our notion of Bernstein-Sato polynomial to invariants in positive
characteristic, namely the F -thresholds and the F -jumping numbers. These invariants
are equal for regular rings [BMS08]; however, they may differ for singular rings. Suppose
that R is a direct summand of a polynomial ring over Q, and f ∈ R. In Theorems 4.20
and 4.21 we show that the truncated base p-expansion of any F -jumping number and any
F -threshold of the reduction of f modulo p is a root of the Bernstein-Sato polynomial
bf (s) modulo p. These type of results have been used to recover roots of the Bernstein-Sato
polynomials via methods in prime characteristic in the regular setting (see [BMS06b]).
2. Background on D-modules
In this section we briefly recall the basics on the theory of rings of differential operators
as introduced by Grothendieck [Gro67, §16.8].
Let R be a Noetherian ring. The ring of differential operators of R is the subring
DR ⊆ HomZ (R, R) whose elements are defined inductively as follows: the differential
0 ∼
operators of order zero are defined by the multiplication by elements of R, i.e. DR
= R. We
say that δ ∈ HomZ (R, R) is an operator of order less than or equal to m if [δ, r] = δr − rδ
0
1
is an operator of order less than or equal to m − 1. We have a filtration DR
⊆ DR
⊆ ···
given by the order and the ring of differential operators is defined as
[
m
DR =
DR
.
m∈N
If A ⊆ R is a subring, we denote DR|A ⊆ DR the subring of differential operators that
are A-linear. In particular DR = DR|Z . If R contains a field K of characteristic p > 0,
e
every additive map is Z/pZ-linear and thus DR = DR|Z/pZ . Moreover, let Rp ⊆ R
(e)
be the subring consisting of pe powers of elements of R and set DR := HomRpe (R, R).
S
(e)
In this case, DR|K ⊆ DR ⊆ e∈N DR . If K is a perfect field, then DR|K = DR and
S
(e)
DR = e∈N DR when R is F -finite (see [Smi87, Theorem 2.7] and [Yek92, Theorem
1.4.9]).
Example 2.1. Let R be either the polynomial ring A[x1 , . . . , xn ] or the formal power
series ring A[[x1 , . . . , xn ]] with coefficients in a ring A. The ring of A-linear differential
operators is:
1 dt
| i = 1, . . . , n; t ∈ N ,
DR|A = R
t! dxti
D-MODULES, BERNSTEIN-SATO POLYNOMIALS & F -INVARIANTS
that is, the free R-module generated by the differential operators
A = K is a field of characteristic zero, we have
d
d
DR|K = R
.
,...,
dx1
dxn
1 dt
.
t! dxti
5
Furthermore, if
Example 2.2. Let S = K[x1 , . . . , xn ] be a polynomial ring over a field K of characteristic
zero and, given an ideal I ⊆ S, set R = S/I. Then, using the results about differential
operator of quotient rings of polynomials [MR01, Theorem 5.13], the ring of K-linear
differential operators of R is given by
{δ ∈ DS|K | δ(I) ⊆ I}
DR|K :=
IDS|K
It is clear that R is a DR|A module. Several of the results presented in this manuscript
concern to the DR|A -module structure of the localization, Rf at an element f ∈ R, and
the local cohomology modules HIi (R) associated to an ideal I ⊆ R. For this reason, we
recall a few properties and definitions regarding these objects.
• Localization: Let M be a DR|A - module and f ∈ R. Then Mf is also a DR|A -module,
where the action of a differential operator δ ∈ DR|A on fvt ∈ Mf is defined inductively as
0
follows: If δ ∈ DR|A
has order zero then δ · fvt = δ·v
. Now, suppose that the action of every
ft
n+1
n
element in DR|A
has been defined. Let δ ∈ DR|A . Then,
δ · v − [δ, f t ] ·
v
δ· t =
f
ft
v
ft
,
which follows from the fact that [δ, f t ] · fvt is defined. With this DR|A -module structure
on Mf , the localization map M → Mf is a morphism of DR|A -modules.
• Local cohomology modules: Let I ⊆ R be an ideal generated by a sequence of elements
f = f1 , . . . , fℓ ∈ R, and let M be any R-module. The Čech complex of M with respect
to f is defined by
M
M
Mfi →
Mfi fj → · · · → Mf1 ···fℓ → 0,
Č• (f ; M) : 0 → M →
i
i,j
where the maps on every summand are localization maps up to a sign. The local cohomology of M with support on I is defined by
HIi (M) = H i (Č• (f ; M)).
One may check that it is independent of the set of generators of I. It follows from this
construction that every local cohomology module over a DR|A -module is again a DR|A module.
3. D-modules over direct summands
Let A ⊆ R ⊆ S be an extension of Noetherian rings such that R is a direct summand
of S with splitting morphism β : S → R. Our first aim in this section is to relate differential operators on S to differential operators on R using the splitting β. Then we
introduce DR|A -modules that are direct summands of DS|A -modules, for which the differential structure is compatible. This idea is used to obtain properties of the localizations
6
J. ÀLVAREZ MONTANER, C. HUNEKE, AND L. NÚÑEZ-BETANCOURT
Rf and the local cohomology modules HIi (R) as a DR|A -module from the properties of Sf
i
and HIS
(S) as a DS|A -module. The same techniques are used in order to develop a theory
of Bernstein-Sato polynomials, even if R is not regular.
3.1. Differentiable direct summands. Let A ⊆ R ⊆ S be Noetherian rings such that
R is a direct summand of S with splitting morphism β : S → R. We start relating
differential operators in DS|A to differential operators in DR|A using the splitting β. We
first show that the order of a differential operator cannot increase when we compose it
with the splitting β. This idea has been used for rings of invariants over a group action
G
[Sch95]. Namely, there exists a map1 DS|K
→ DS G |K . However, this map may not be
surjective [Sch95, Example 5.7]. It is injective for many classes of groups, in particular for
finite groups, but, to the best of our knowledge, the injectivity of this map is not known
in general. For the study of local cohomology modules of S G in characteristic zero as
G
DS|K
-module, we refer to [Put16].
We point that the following lemma was implicit in Smith’s proof of the statement: S is
a simple DS|Fp -module, then R is also DR|Fp is also simple [Smi95, Proposition 3.1]. We
include this proof for the sake of completeness.
Lemma 3.1. Let A ⊆ R ⊆ S be three Noetherian rings. Let ι : R → S denote the
n
inclusion and let β : S → R be any R-linear morphism. Then, for every δ ∈ DS|A
, we
n
have that δ̃ := β ◦ δ|R ∈ DR|A .
Proof. To avoid heavy notation, we also write δ for δ|R We first note that δ̃ ∈ HomA (R, R).
We proceed by induction on the order n of the differential operator. If n = 0, there exists
g ∈ S such that δ(w) = gw for every w ∈ S. Then,
δ̃(v) = β ◦ δ(v) = β(δ(v)) = β(gv) = vβ(g) = β(g)v.
n+1
for every v ∈ R. We now assume the claim for n. Let δ ∈ DS|A
, and f ∈ R. Then,
[δ̃, f ](v) = δ̃(f v) − f δ̃(v)
= β ◦ δ(f v) − f (β ◦ δ(v))
= β(δ((f v)) − β(f δ(v))
= β(δ((f v) − f δ(v))
= β([δ, f ](v))
]
= [δ,
f ](v)
n+1
n
]
for all v ∈ R. Then, [δ̃, f ] = [δ,
f ] ∈ DR|A
by induction hypothesis. Then, δ̃ ∈ DS|A
.
We now introduce DR|A -modules that are direct summands of DS|A -modules, for which
the differential structure is compatible.
Definition 3.2. Let A ⊆ R ⊆ S let be three Noetherian rings. Suppose that the inclusion
R ⊆ S has a splitting β : S → R. We say that a DR|A -module M, is a differential direct
summand of the DS|A -module N compatible with β, if M ⊆ N and we have a splitting
θ : N → M of R-modules such that
θ(δ · v) = (β ◦ δ|R ) · v.
1D G
S|K
denotes the ring of equivariant differential operators over the field K.
D-MODULES, BERNSTEIN-SATO POLYNOMIALS & F -INVARIANTS
7
for every v ∈ M and δ ∈ DS|A . In this case, we say that θ is a differentiable splitting map
compatible with β.
Lemma 3.3. Let A ⊆ R ⊆ S be three Noetherian rings. Suppose that the inclusion R ⊆ S
has a splitting β : S → R. Let M ⊆ N be differentiable direct summand compatible with
β. Let V, W ⊆ M be DR|A -submodules of M. If DS|A V = DS|A W , then V = W.
e Let v ∈ V .
Proof. We denote by θ : N → M a differential splitting, and β ◦ δ|R by δ.
Since DS|A V = DS|A W, there exists δ1 , . . . , δℓ ∈ DS|A and w1 , . . . , wℓ ∈ W such that
v = δ1 w1 + . . . + δℓ wℓ . We apply θ to both sides to obtain
v = θ(v) = θ (δ1 w1 + . . . + δℓ wℓ ) = θ(δ1 w1 ) + . . . + θ(δℓ wℓ ) = δe1 w1 + . . . + δeℓ wℓ ,
where the last step follows from the fact that θ is a differential splitting compatible with
β. Then, v ∈ DR|A W = W. We conclude that V ⊆ W. Likewise, W ⊆ V, which concludes
the proof.
The main result of this subsection is the following upper bound for the length, as
D-module, of a differential direct summand.
Proposition 3.4. Let A ⊆ R ⊆ S let be three Noetherian rings. Suppose that the
inclusion R ⊆ S has a splitting β : S → R. Let M ⊆ N be differentiable direct summands
compatible with β. Then, λDR|A (M) ≤ λDS|A (N).
Proof. We note that if N does not have finite length as DS|A -modules the claim is clear.
We assume that λDS|A (N) is finite, and proceed by contradiction. Let ℓ = λDS|A (N). Let
0 ( V1 ( . . . ( Vℓ+1 be a strictly increasing chain of DR|A -submodules of M. By Lemma
3.3 0 ( DS|A V1 ( . . . ( DS|A Vℓ+1 is a strictly increasing chain of DS|A -submodules of N.
Then, ℓ + 1 ≤ λDS|A (N), which is a contradiction.
Definition 3.5. Let A ⊆ R ⊆ S let be three Noetherian rings. Suppose that the inclusion
R ⊆ S has a splitting β : S → R. Given two differentiable direct summands M1 ⊆ N1
and M2 ⊆ N2 with differentiable splittings θ1 : N1 → M1 and θ2 : N2 → M2 , we say that a
map φ : M1 → M2 is a morphism of differential direct summands if φ ∈ HomDS|A (N1 , N2 ),
φ(M1 ) ⊆ M2 , φ|M1 ∈ HomDR|A (M1 , M2 ), and the following diagram
M1
φ| M
/
N1
θ1
/
M1
φ| M
φ
1
θ2
1
/ N2
/ M2
M2
commutes. For the sake of notation, we often write φ for φ|N1 .
Proposition 3.6. Let A ⊆ R ⊆ S let be three Noetherian rings. Suppose that the
inclusion R ⊆ S has a splitting β : S → R. Let M be a DR|A -module and N be a
DS|A -module. If M is a differential direct summand of N, then Mf is a differential direct
summand of Nf for every element of f ∈ R. Furthermore, the localization maps give a
morphism of differential direct summands.
Proof. Let θ : N → M be a differential splitting compatible with β. For the sake of
notation, we use the same symbols for induced maps in the localization. We now show
that the map Mf → Nf induced by θ is a differential splitting compatible with β.
8
J. ÀLVAREZ MONTANER, C. HUNEKE, AND L. NÚÑEZ-BETANCOURT
Let δe denote the differential operator β ◦ δ|R . We start showing that θ(δ · fvn ) = δe · fvn
for every element v ∈ M and δ ∈ DS|A , by induction on the order of δ. If δ has order zero
we have, using the action of δ as given in Section 2, that
v
δ·v
1
1
1
θ δ· n =θ
= n θ(δ · v) = n θ(δ · v) = n (δe · v),
n
f
f
f
f
f
where the last equality follows because θ : M → N is a differentiable splitting. Then,
v
1
δe · v
v
θ δ · n = n (δe · v) = n = δe · n .
f
f
f
f
We now assume our claim for differential operators of order equal to or less than n, and
suppose that δ has order n + 1.
!
δ · v − [δ, f n ] · fvt
v
θ δ· t =θ
f
ft
!
[δ, f n ] · fvt
δ·v
=θ
−θ
ft
ft
!
[δ, f n ] · fvt
1
= t θ (δ · v) − θ
f
ft
!
[δ, f n ] · fvt
1e
because θ : M → N is a differentiable spliting
= tδ · v − θ
f
ft
=
=
=
=
^
[δ,
f n] ·
1e
δ
·
v
−
ft
ft
^
δe · v − [δ,
f n ] · vt
v
ft
by induction hypothesis as [δ, f n ] ∈ DAn (S);
f
ft
^
δe · v − [δ,
f n] ·
ft
e f n] ·
δe · v − [δ,
ft
v
ft
v
ft
as a consequence of the proof of Lemma 3.1;
v
ft
To verify that the localization map M → Mf induces a map of differential direct summands, we note that all the conditions in Definition 3.5 are satisfied by the usual properties
of localization maps, and the fact that localization maps are D-module morphisms.
= δe ·
Lemma 3.7. Let A ⊆ R ⊆ S let be three Noetherian rings. Suppose that the inclusion R ⊆ S has a splitting β : S → R. Let M1 ⊆ N1 and M2 ⊆ N2 be differentiable
direct summands with differential splittings θ1 and θ2 . Let φ : N1 → N2 be a map of
differentiable direct summands. Then, Ker(φ|M1 ) ⊆ Ker(φ), Im(φ|M1 ) ⊆ Im(φ|M1 ) and
Coker(φ|M1 ) ⊆ Coker(φ) are also differentiable direct summands. Furthermore, the inclusion and projection maps are morphism of differentiable direct summands.
D-MODULES, BERNSTEIN-SATO POLYNOMIALS & F -INVARIANTS
9
Proof. Since φ is a map of DS|A modules and φ|M1 is also a map of DR|A -modules, we
note that the induced maps on kernel, images and cokernel are also maps of DS|A and
DR|A -modules respectively. The rest follows from the fact that the diagram
M1
φ|M
/
θ1
/
M1
φ| M
φ
1
M2
N1
/
N2
θ2
.
1
/
M2
commutes, and that θ1 , θ2 are differential splittings.
3.2. Finite length of D-modules over direct summands. Let K be a field and S
be either K[x1 , . . . , xn ] or K[[x1 , . . . , xn ]]. One of the main results in the theory of Dmodules is that both the localizations Sf and the local cohomology modules HIi (S) have
finite length as DS|K -modules. If K is a field of characteristic zero, this result follows
from the fact that localizations and local cohomology modules belong to the class of holonomic DS|K -modules (see [Bjö79, Cou95]). These results can be extended to differentiably
admissible K-algebras, which is a larger class or regular rings containing the rational numbers (see [MNM91, NB13]). For intance, if (R, m, K) is a ramified regular ring of mixed
characteristic p > 0, then R[1/p] is a differentiably admissible algebra [NB13].
In positive characteristic, these results were proved by Lyubeznik using his theory of
F -finite F -modules [Lyu97]. In fact, he showed that Sf and HIi (S) have finite length as
DS|K -modules for a larger class of Noetherian regular K-algebras of prime characteristic.
Indeed, F -finite F -modules behave quite likely as holonomic modules in characteristic
zero.
Despite the fact that we do not have a theory of holonomic or F -finite F -modules for
direct summands, we are still able to prove the finite length of localization and local
cohomology modules in this setting.
Theorem 3.8. Let K be a field and S be either K[x1 , . . . , xn ] or K[[x1 , . . . , xn ]]. Let
K ⊆ R ⊆ S be a subring with a splitting β : S → R. Then Rf and HIi (R) have finite
length as DR|K -modules for every f ∈ R and any ideal I ⊆ R. Furthermore, λDR|K (Rf ) ≤
i
λDS|K (Sf ) and λDR|K (HIi (R)) ≤ λDS|K (HIS
(S)).
Proof. Since R ⊆ S is a differential splitting, we have that Rf ⊆ Sf is a differential splitting for every f ∈ R by Proposition 3.6. Then we have λDR|K (Rf ) ≤ λDS|K (Sf ) by Proposition 3.4. Let I = (f1 , . . . , fℓ ) ⊆ R, and f = f1 , . . . , fℓ . We note that Č• (f ; R) ⊆ Č• (f ; S)
is a complex of differential direct summands, and so, the local cohomology, HIi (R) ⊆
i
HIS
(S), is again a differential direct summand by Lemma 3.7. Then, λDR|K (HIi (R)) ≤
i
λDS|K (HIS
(S)) by Proposition 3.4.
We may extend the property of finite length to other DR|K -modules if we consider the
category C(R, A) introduced by Lyubeznik [Lyu00a, Lyu00b].
Definition 3.9. We denote by C(R, A) the smallest subcategory of DR|A -modules that
contains Rf for all f ∈ R and that is closed under subobjects, extensions and quotients.
The category C(R, A) is also closed under localization and, in some sense, behaves as
well as holonomic or F-finite F-modules. In particular, this category contains iterated local
10
J. ÀLVAREZ MONTANER, C. HUNEKE, AND L. NÚÑEZ-BETANCOURT
cohomology modules HIi11 · · · HIiℓℓ (R), and more generally, Lyubeznik functors T defined in
[Lyu93].
Theorem 3.10. Let K be a field and S be either K[x1 , . . . , xn ] or K[[x1 , . . . , xn ]]. Let
K ⊆ R ⊆ S let be a subring with a splitting β : S → R. Then, every DR|K -module in
C(R, K) has finite length as DR|K -module.
Proof. This statements follows from Theorem 3.8, since Rf has finite length as DR|K module, and every object in C(R, K) is build from localization modules with operations
that preserve the property of having finite length.
From the previous theorem, we obtain as a corollary a result by the third author
regarding associated primes of local cohomology.
Corollary 3.11 ([NB12]). Let K be a field and S be either K[x1 , . . . , xn ] or K[[x1 , . . . , xn ]].
Let K ⊆ R ⊆ S let be a subring with a splitting β : S → R. Then, AssR HIi (R) is a finite
set for every ideal I ⊆ R and every integer i.
Proof. We note that a simple DR|K -module, M 6= 0, has only one associated prime. In fact,
if Q is a maximal element in the set of the associated primes of M, then AssR HQ0 (M) = Q
and HQ0 (M) is a nonzero DR|K -submodule of M. Since HIi (R) has finite length, there exists
an increasing sequence of DR|K -modules 0 = M0 ⊆ M1 ⊆ . . . ⊆ Mℓ = HIi (R) such that
S
Mj+1 /Mj is a simple DR|K -module. Then, AssR HIi (R) ⊆ ℓj=0 AssR Mj+1 /Mj .
Remark 3.12. There are other structural properties that local cohomology modules
satisfy over regular rings containing a field, which are proved using D-modules. For
instance, local cohomology modules have finite Bass numbers and their injective dimension
is bounded above by the dimension of their support. Both of these properties fail for direct
summands. For instance, R = K[ac, ad, bc, bd] is a direct summand of S = K[a, b, c, d],
2
H(ab,ad)(R)
is a local cohomology module with an infinite socle [Har68], so not all Bass
2
numbers are finite. In addition, H(ab,ad)(R)
is a simple DR|K -module [Hsi12] supported at
the maximal ideal (ab, ad, cd, bc)R, which is not an injective R-module, so the injective
dimension can be bigger than the dimension of the support.
3.3. Bernstein-Sato polynomial over direct summands. Let S be either the polynomial ring K[x1 , . . . , xn ] or the formal power series ring K[[x1 , . . . , xn ]], where K is a field
of characteristic zero and R ⊆ S a direct summand. We aim to extend the well-known
theory of Bernstein-Sato over S to the direct summand R. We start reviewing the basic
facts.
D
E
d
d
Recall that DS|K = S dx1 , . . . , dxn is the ring of K-linear differential operators of S.
Let DS|K [s] be the polynomial ring in the indeterminate s and coefficients in DS|K . For
any nonzero element f ∈ S, there exists a polynomial b(s) 6= 0 and a differential operator
δ(s) ∈ DS|K [s] satisfying the following functional equation:
(3.3.1)
δ(s) · f f s = b(s)f s .
This is an identity in M[f s ] := Sf [s]f s which is the rank one free Sf [s]-module with
generator f s viewed as a formal symbol. After taking the specializations M[f s ] → Sf ,
given by f s 7→ f t for t ∈ Z, we obtain the following equation
(3.3.2)
δ(t) · f t+1 = b(t)f t .
D-MODULES, BERNSTEIN-SATO POLYNOMIALS & F -INVARIANTS
11
The set of all polynomials b(s) for which there exists an operator δ(s) such that Equation
3.3.1 holds forms an ideal of K[s]. The monic polynomial that generates this ideal is called
the Bernstein-Sato polynomial [Ber72, SS72] and is denoted as bf (s).
A fundamental result states that the roots of bf (s) are negative rational numbers in the
polynomial case2 [Mal75, Kas77]. As a consequence, we may view bf (s) ∈ Q[s].
A key point in this theory is that we can give a DS|K [s]-module structure to M[f s ] as
follows: The action of the derivatives on a general element gf s is given by
d
d
s
−1 d
· gf =
· g + sgf
(f ) f s ,
dxi
dxi
dxi
and the multiplication by s is DS|K -linear. Unfortunately, when we turn our attention to
a non-regular ring R, we are not able to define a DR|K [s]-module structure on Rf [s]f s .
However, it still makes sense to study the existence of an element δ(s) ∈ DR|K [s] and a
polynomial b(s) ∈ Q[s] that satisfy Equation 3.3.2 for every t ∈ Z.
In order to make precise the setup that we introduce the following general definition
that encodes all the Bernstein-Sato data.
Definition 3.13. Let A ⊆ R be two Noetherian rings, and f ∈ R. We consider the
A[s]-module,
B = (δ(s), b(s)) ∈ DR|A [s] × A[s] | δ(t) · f t+1 = b(t)f t ∀t ∈ Z in Rf .
Let Bf be the image of B under the projection map on the second entry. If Bf = 0, we
say that f has no Bernstein-Sato polynomial. If Bf 6= 0, we call its monic generator the
A
Bernstein-Sato polynomial of f , which is denoted by bR
f (s). If the rings R and A are
clear from the context, we just write bf (s).
Our next result shows that Bernstein-Sato polynomials exist for a direct summand,
R ⊆ S, as long as they exist for S.
Theorem 3.14. Let A ⊆ R ⊆ S be three Noetherian rings. Suppose that the inclusion
A
R ⊆ S has a splitting β : S → R. Let f ∈ R. If bSf A (s) exists, then bR
f (s) also exists.
SA
A
Furthermore, bR
f (s) divides bf (s).
Proof. Let δ(s) ∈ DS|A [s] be such that δ(t) · f t+1 = b(t)f t for all t ∈ Z, where the equality
is taken in Sf . Since Rf is a differential direct summand of Sf , we have that
(3.3.3)
β ◦ δ(t)|R · f t+1 = b(t)f t
A
for every t ∈ Z in Rf . Then, bSf A (s) is a multiple of bR
f (s), since this is the monic
polynomial satisfying 3.3.3.
Corollary 3.15. Let K be a field and S be either K[x1 , . . . , xn ] or K[[x1 , . . . , xn ]]. Let
K ⊆ R ⊆ S be a direct summand with an splitting β : S → R. Then the Bernstein-Sato
polynomial exists for every element f ∈ R. Furthermore, in the polynomial case all the
K
roots of bR
f (s) are negative rational numbers if S is polynomial.
Proof. The Bernstein-Sato polynomial bSf K (s) exists [Ber72, SS72]. If S is polynomial,
bSf K (s) only has negative rational roots [Mal75, Kas77].
2
Indeed, this result is also stated in the analytic case (S = C{x1 , . . . , xn }).
12
J. ÀLVAREZ MONTANER, C. HUNEKE, AND L. NÚÑEZ-BETANCOURT
Remark 3.16. Theorems 3.8, 3.10, and 3.14 can be extended to direct summands of
differentiably admissible K-algebras, where K is a field of characteristic zero (see [MNM91,
NB13]). In particular, we may also consider direct summands of S = C{x1 , . . . , xn }, and
the rationality of the roots of the Bernstein-Sato polynomials also holds [Mal75, Kas77].
S
We now present an example that show that bR
f (s) could differ from bf (s). This example
was kindly given to us by Jack Jeffries.
Example 3.17. Let R = Q[xu, yv] and S = Q[x, y, u, v] We note that R = S G , where G
is a torus with coordinates λ1 , λ2 acting by
−1 −1
−1
x 7→ λ1 λ2 x, y 7→ λ1 λ−1
2 y, u 7→ λ1 λ2 u, v 7→ λ1 λ2 v.
Consider f = xu − yv. There exists an isomorphism from R to a polynomial ring in two
1
variables, which sends f to a variable. Thus, bR
f (s) = s + 1 and Rf is generated by f
as a DR|K -module. In contrast, bSf (s) = (s + 1)(s + 2), and so, Sf 6= DS|K · f1 [Wal05,
Lemma 1.3]. Furthermore, Hfi (R) is a simple DR|K -module, but Hf1(S) is not a simple
DS|K -module.
Remark 3.18. There are several variants and extensions where the theory of BernsteinSato polynomial applies. The results of Theorem 3.14 and Corollary 3.15 can be extended
to these cases:
• Bernstein-Sato polynomial associated to an algebraic variety: Let I ⊆ S be an ideal
generated by a sequence of nonzero polynomials f = f1 , . . . , fℓ . Budur, Mustaţă and Saito
[BMS06a] defined the Bernstein-Sato polynomial, bI (s), associated to the ideal I. After
a convenient shifting, bI (s) is also an invariant of the variety V (I). They proved the
existence of a nonzero polynomial b(s) ∈ Q[s], and differential operators δc (s1 , . . . , sℓ ) ∈
DR|K [s1 , . . . , sℓ ] satisfying, in our setting, a functional equation
Y
ℓ
ℓ
Y
X Y
ti
ti +ci
fi
= b(t1 + . . . + tr )
fiti ,
δc (t1 , . . . , tℓ ) ·
(3.3.4)
c
i
ℓ
i=1
i=1
c∈Z
i=1,...,ℓ
ci <0
for every ti ∈ Z and where the sum varies over finitely many c = (c1 , . . . , cℓ ) ∈ Zℓ . We
think of this as a functional equation in Sf1 ···fℓ .
Let R ⊆ S be a direct summand with a splitting β : S → R and assume that fi ∈ R,
i = 1, . . . , ℓ. Then we obtain the analogue of Equation 3.3.4 for R by composing the
differential operators δc (t1 , . . . , tℓ ) with the splitting β.
• Bernstein-Sato ideal associated to a family of hypersurfaces: Let f = f1 , . . . , fℓ be
a sequence of elements in S. Then, there exists a nonzero polynomial b(s1 , . . . , sℓ ) ∈
Q[s1 , . . . , sℓ ], and differential operators δ(s1 , . . . , sℓ ) ∈ DR|K [s1 , . . . , sℓ ] satisfying, in our
setting, a functional equation
(3.3.5)
δ(t1 , . . . , tℓ ) ·
ℓ
Y
i=1
fiti +1
= b(t1 , . . . , tℓ )
ℓ
Y
fiti .
i=1
for every ti ∈ Z. The ideal in Q[s1 , . . . , sℓ ] generated by these polynomials b(s1 , . . . , sℓ ) is
the Bernstein-Sato ideal associated to the sequence f . If R ⊆ S is a direct summand and
fi ∈ R, i = 1, . . . , ℓ, we use the same compose by β again to obtain the equation in R.
D-MODULES, BERNSTEIN-SATO POLYNOMIALS & F -INVARIANTS
13
We end this subsection with a example, generously suggested by Jack Jeffries, which
shows that Bernstein-Sato polynomials do not exist in general.
Example 3.19. Let R = C[x, y, z]/(x3 + y 3 + z 3 ). We consider R as a graded ring where
every variable has degree 1, and we denote the C-subspace of homogeneous elements of
degree ℓ by [R]ℓ . We note that DR|C is also a graded ring, where a differential operator
has degree ℓ ∈ Z if δg = 0 or deg(δg) = deg(g) + ℓ for every homogeneous element g ∈ R.
We note that this identity extends to any localization of R at a homogeneous element.
The C-subspace
of differential operators of degree ℓ is denoted by DR|C (ℓ). In this case,
L
DR|C = ℓ∈Z DR|C (ℓ). Furthermore, DR|C (ℓ) = 0 for every ℓ < 0 [BGG72, Proposition
1]. Let f ∈ R be any element of positive degree u. We now show by contradiction that
f has no Bernstein-Sato polynomial (as in Definition 3.13). Suppose that there exists
δ[s] ∈ DR|C [s] and b(s) ∈ C[s] such that b(s) 6= 0 and δ(t)f t+1 = b(t)f t for every t ∈ Z.
We fix t ∈ Z such that t > 0 and b(t) 6= 0. We set δ = δ(t), and we take δℓ ∈ DR|C (ℓ) such
that δ = δ0 + · · · + δk , which is possible because there are no operators of negative degree
[BGG72, Proposition 1]. Then,
M
b(t)f t = δf t+1 = δ0 f t+1 + · · · + δk f t+1 ∈
[R]ℓ .
ℓ>(t+1)u
As a consequence, tu > (t + 1)u for every ℓ appearing in the equation above, because
b(t)f t is a nonzero homogeneous element of degree tu. This is a contradiction because
u > 0.
1
Remark 3.20. We note that even in the case where DR|K
has no elements of negative degree, one can find elements of negative degree in DR|K . For instance, if R =
C[a, b, c, d]/(ad − bc) ∼
= C[xy, xz, yu, zu], which is a direct summand of S = C[x, y, z, u]
with splitting β : S → R defined on the monomials by β(xα1 y α2 z α3 uα4 ) = xα1 y α2 z α3 uα4
if α1 + α4 = α2 + α3 and β(xα1 y α2 z α3 uα4 ) = 0 otherwise. Since R is a graded ring with
an isolated singularity, there are no elements of order 1 and negative degree [Kan77].
∂ ∂
However, β ◦ ∂x
is a differential operator on R of order 2 and negative degree.
∂y
Theorem 3.14 and Example 3.19 motivate the following question.
Question 3.21. What conditions are necessary and sufficient for a K-algebra for the
existence of the Bernstein-Sato polynomial?
4. F -invariants of direct summands
In this section we study different invariants in positive characteristic: F -jumping numbers and F -thresholds. We point out that in a regular ring these two families of invariants
are the same. However, over singular rings they usually differ, even for direct summands.
For this reason, we need to study them using different approaches.
We recall that R acquires an structure of R-module by restriction of scalars via the e-th
iteration of the Frobenius map. F e . We denote this module action on R by F∗e R. Namely,
e
we have r · F∗e x = F∗e (r p x) ∈ F∗e R for r ∈ R and F∗e x ∈ F∗e R. If R is a reduced ring,
e
it is a common practice to identify F∗e R with R1/p , via the isomorphism of R-modules
e
F∗e r 7→ r 1/p .
Throughout this section we consider rings containing a field of positive characteristic
that are F -finite, that is, rings such that F∗e R is a finitely generated R-module.
14
J. ÀLVAREZ MONTANER, C. HUNEKE, AND L. NÚÑEZ-BETANCOURT
4.1. Test ideals of direct summands. In this subsection we prove that the F -jumping
numbers of an ideal I of a direct summand R of a regular ring S are contained in the
F -jumping numbers of the extended ideal IS. This result follows from comparison of rings
of differential operators and Cartier morphisms between R and S. Differential operators
have been used previously to show discreteness and rationality of F -jumping numbers for
regular rings. However, they have not been used in the singular case as far as we know.
We point out that Chiecchio, Enescu, Miller, and Schwede [CEMS] used the fact that
R is a direct summand of the symbolic Rees algebra R = ⊕∞
t=0 OX (−tKX ), to prove
discreteness and rationality of F -jumping numbers in certain cases, given that R is finitely
generated as R-algebra. From recent results in the minimal model program [BCHM10,
HM10, HX15] (see also [Kol10, Theorem 92]), one could expect that R is finitely generated
for every strongly F -regular ring, in particular, for direct summands of regular rings.
However, to the best of our knowledge, this question is still open. Furthermore, this
question relates to the longstanding problem of weakly F -regular versus strongly F -regular
[CEMS]. We point out that our strategy differs from theirs by using differential operators
over a singular ring.
Definition 4.1. Let R be an F -finite Noetherian ring.
• We say that R if F -pure if the inclusion map R → F∗e R splits for all e > 0.
• Assume that R is a domain. We say that R is strongly F -regular if for every r ∈ R
there exists e ∈ N such that the ϕ : R → F∗e R defined by 1 7→ F∗e r splits.
Regular rings are strongly F -regular. Moreover, direct summands of strongly F -regular
rings are strongly F -regular by a result of [HH89].
We now introduce the definition and preliminary results regarding Cartier operators
and its relation with differential operators in positive characteristic.
Definition 4.2. Let R be an F -finite ring.
e
• An additive map ψ : R → R is a pe -linear map if ψ(rf ) = r p ψ(f ). Let FRe be the set
of all the pe -linear maps. Then, we have FRe ∼
= HomR (R, F∗e R).
e
• An additive map φ : R → R is a p−e -linear map if φ(r p f ) = rφ(f ). Let CRe be the set
of all the p−e -linear maps. Then, we have CRe ∼
= HomR (F∗e R, R).
The main examples of pe and p−e -linear maps are the iterated Frobenius and Cartier
morphisms. For this reason, these morphisms are usually referred as Frobenius and Cartier
operators. They have a close relation with differential operators. Recall from Section 2
that for F -finite rings we have
[ (e)
DR =
DR .
e
∼
Since R is reduced, we have
= HomR (F∗e R, F∗e R). It is natural to consider the
(e)
pairing CRe ⊗R FRe −→ DR sending φ ⊗ ψ to its composition ψ ◦ φ. If R is a regular ring,
(e)
the previous map is an isomorphism of Abelian groups, that is CRe ⊗R FRe ∼
= DR . One
concludes, roughly speaking, that it is equivalent to consider differential operators or just
Cartier operators.
In general, for singular rings, we do not have such an isomorphism. However, in the
F -pure case we make an easy observation that relates differential operators with Cartier
maps. This relation is the heart of Lemma 4.10, which is a key part of our strategy
towards Theorem C.
(e)
DR
D-MODULES, BERNSTEIN-SATO POLYNOMIALS & F -INVARIANTS
15
Remark 4.3. Let R be an F -pure F -finite ring. If γ : F∗e R → R is a splitting, the map
Ψ : HomR (F∗e R, F∗e R) → HomR (F∗e R, R) defined by δ 7→ γ ◦ δ is surjective. Then, the
(e)
corresponding map DR → CRe is also surjective.
Let S be a regular ring. The so-called Frobenius descent, that follows from the isomor(e)
phism CSe ⊗S FSe ∼
= DS and Morita equivalence, is a key ingredient to show that Sf = DS · f1
[ÀMBL05, Theorem 1.1]. This is an important result regarding the D-module structure
of the localization Sf at an element f ∈ S. This result was extended by Takagi and Takahashi [TT08, Corollary 2.10] to the case of graded rings with finite F-representation type.
This class of rings include graded direct summands. In our first result in this section, we
prove that this occurs for all direct summands of F -finite regular domains.
Theorem 4.4. Let S be a regular F -finite domain. Let R ⊆ S be an extension of Noetherian rings such that R a direct summand of S. Then Rf is generated by f1 as DR -module.
(e′ )
Proof. Since Sf is generated as DS -module by f1 , for every e there exists δ ∈ DS , with
e′ ≥ e, such that δ( f1 ) =
e′
e′
1
.
f pe
e′
Then, by multiplying this equation by f p , we obtain
e
δ(f p −1 ) = f p −p .
(e′ )
Let β : S → R denote a splitting. Let δe = β ◦ δ. We note that δe ∈ DR . In addition,
′
′
e pe −1 ) = f pe −pe . Hence, δ(
e 1 ) = 1pe . We conclude that Rf is generated as DR -module
δ(f
f
f
1
by f .
Test ideals have been a fundamental tool in the theory of tight closure developed by
Hochster and the second author [HH90, HH94a, HH94a]. Hara and Yoshida [HY03] extended the theory to include test ideals τR (I λ ) associated to pairs (R, I λ ) where I ⊆ R is
an ideal and λ ∈ R is a parameter. A new approach to test ideals3 by means of Cartier
operators was given by Blickle, Mustaţă and Smith [BMS08, BMS09] in the case that R is
a regular ring. Their approach has been extremely useful in order to extended the theory
of test ideals to non-regular rings. We refer to [Sch11, BB11, Bli13] for a more general
setting.
Definition 4.5. Suppose that R is a strongly F -regular ring. Let I ⊆ R be an ideal, and
λ ∈ R>0 . The test ideal4 of the pair (I, λ) is defined by
[
e
τR (I λ ) =
CRe I ⌈p λ⌉ .
e∈N
We note that the chain of ideals
e
so, τR (I λ ) = CRe I ⌈p λ⌉ for e ≫ 0.
e
{CRe I ⌈p λ⌉ }
is increasing [TT08, Proposition 4.4], and
We now summarize basic well-known properties of test ideals. In the case of strongly
F -regular rings we refer to [TT08, Lemma 4.5]. For the general case of F -finite rings, we
refer [Bli13, Proposition ].
Proposition 4.6. Let R be an F -finite ring, I, J ⊆ R ideals, and λ, λ′ ∈ R>0 . Then,
(1) If I ⊆ J, then τR (I λ ) ⊆ τR (J λ ).
3Indeed,
they describe the generalization to pairs of the big test ideal considered in [HH94a].
this are the big test ideals. This ideals are equal to the original test ideals for graded rings
[LS99] and Gorenstein rings [LS01].
4Precisely,
16
J. ÀLVAREZ MONTANER, C. HUNEKE, AND L. NÚÑEZ-BETANCOURT
′
(2) If λ < λ′ , then τR (I λ ) ⊆ τR (I λ ).
′
(3) There exists ǫ > 0, such that τR (I λ ) = τR (I λ ), if λ′ ∈ [λ, λ + ǫ).
Therefore, every ideal I ⊆ R is associated to a family of test ideals τR (I λ ) parameterized
by real numbers λ ∈ R>0 . Indeed, they form a nested chain of ideals. The real numbers
where the test ideals change are called F -jumping numbers. We now make this precise,
Definition 4.7. Let R be an F -finite ring and let I ⊆ R be an ideal. A real number λ is
an F -jumping number of I if
τR (I λ ) 6= τR (I λ−ǫ )
for every ǫ > 0.
Remark 4.8. Let R be a strongly F -regular F -finite ring of characteristic p > 0, and
I ⊆ R. If R is either a local ring or a graded K algebra with I homogeneous, then the
F -pure threshold of I is defined in terms of maps R → F∗e R that split [TW04, Definition
2.1]. Specifically,
fpt(I) := sup{λ | ∀e ≫ 0 ∃f ∈ I ⌊(p
e −1)λ⌋
Then, fpt(I) is the first F -jumping number of I.
such that 1 7→ F∗e f splits}.
From now on we consider extensions of Noetherian rings R ⊆ S such that R is a direct
summand of S. Our aim is to relate test ideals of an ideal in R with the test ideal of its
extension to S. The first result in this direction shows that the F -pure threshold of an
ideal of R is greater or equal than the F -pure threshold of its extension in S.
Proposition 4.9. Let S be a regular F -finite domain. Let R ⊆ S be an extension of
Noetherian rings such that R is a direct summand of S. Then, τS ((IS)λ ) ∩ R ⊆ τR (I λ )
for every ideal I ⊆ R and λ ∈ R>0 . In particular, fptS (IS) ≤ fptR (I).
e
Proof. Let f ∈ τS ((IS)λ ) ∩ R. Then, there exists e ∈ N such that f ∈ CSe (IS)⌈p λ⌉ . Then,
e
there exist g1 , . . . , gℓ ∈ I ⌈p λ⌉ and φ1 , . . . , φℓ ∈ CeS such that φ1 (g1 ) + . . . + φℓ (gℓ ) = f .
e
Let β : S → R denote a splitting. Then, β ◦ φi |R ∈ CRe . Thus, β(φi (gi )) ∈ CRe I ⌈p λ⌉ ⊆
τR (I λ ). Then,
f = φ1 (g1 ) + . . . + φℓ (gℓ );
= β(φ1 (g1 ) + . . . + φℓ (gℓ ));
= β(φ1 (g1 )) + . . . + β(φℓ (gℓ )).
Hence, f ∈ CRe (I
⌈pe λ⌉
) ⊆ τR (I λ ).
The following results are a key part of our strategy towards Theorem C.
(e)
(e)
Lemma 4.10. Let R be an F -pure F -finite ring, and I, J ⊆ R ideals. If DR I = DR J,
then CRe I = CRe J.
Proof. We have that
CRe I = HomR (F∗e R, R) · F∗e I by Definition 4.2
= γ(HomR (F∗e R, F∗e R) · F∗e I) by Remark 4.3
(e)
(e)
= γ(HomR (F∗e R, F∗e R) · F∗e J) by Definition 4.2 and DR I = DR J;
= HomR (F∗e R, R) · F∗e J by Remark 4.3
= CRe J by Definition 4.2
D-MODULES, BERNSTEIN-SATO POLYNOMIALS & F -INVARIANTS
17
Lemma 4.11. Let R ⊆ S be an extension of Noetherian rings such that R is a direct
summand of S. Let I ⊆ R denote an ideal. If CSe (IS)r = CSe (IS)t for r, t ∈ N, then
CRe I r = CRe I t .
Proof. We assume without loss of generality that r ≤ t. We have CRe I t ⊆ CRe I r , thus we
focus on the other containment. Since CSe (IS)r = CSe (IS)t , we have that
[pe ]
e
(e)
(e)
DS (IS)r = (CSe (IS)r )[p ] = CSe (IS)t
= DS (IS)t .
(e)
(e)
Then, I r ⊆ DS (IS)t . Let f ∈ I r . Then, there exists φ1 , . . . , φℓ ∈ DS and g1 , . . . , gℓ ∈ I t
(e)
such that f = φ1 (g1 ) + . . . + φℓ (gℓ ). Let β : S → R denote a splitting. Then, β ◦ φi |R ∈ DR
(e)
(e)
and β(φi (gi )) ∈ DR I t . It follows that f ∈ DR (I t ) since
f = φ1 (g1 ) + . . . + φℓ (gℓ )
= β(φ1 (g1 ) + . . . + φℓ (gℓ ))
= β(φ1 (g1 ) + . . . + β(φℓ (gℓ )
Therefore, I r
(e)
(e)
= β(φ1 (g1 )) + . . . + β(φℓ (gℓ )).
(e)
(e) t
(e) t
(e) r
(e)
= DR I t . We conclude that
⊆ DR I . Hence, DR I ⊆ DR DR I
DR I r = DR I t , and so, CRe I r = CRe I t by Lemma 4.10.
We are now ready to prove the main result in this section.
Theorem 4.12. Let S be a regular F -finite domain. Let R ⊆ S be an extension of
Noetherian rings such that R is a direct summand of S. Let I ⊆ R denote an ideal, and
λ1 , λ2 ∈ R>0 . If τS ((IS)λ1 ) = τS ((IS)λ2 ), then τR (I λ1 ) = τR (I λ2 ).
e
Proof. There exists t ∈ N such that for e > N, τS ((IS)λ1 ) = CSe (IS)⌈p λ1 ⌉ and τS ((IS)λ2 ) =
e
e
e
CSe (IS)⌈p λ2 ⌉ . Then, CSe (IS)⌈p λ1 ⌉ = CSe (IS)⌈p λ2 ⌉ for e ≥ t. By Lemma 4.11, we have that
e
e
CRe I ⌈p λ1 ⌉ = CRe I ⌈p λ2 ⌉ for e ≥ t. Then,
[
[
[
[
e
e
e
e
τR (I λ1 ) =
CRe I ⌈p λ1 ⌉ =
CRe I ⌈p λ1 ⌉ =
CRe I ⌈p λ2 ⌉ =
CRe I ⌈p λ2 ⌉ = τR (I λ2 ).
e∈N
e≥t
e≥t
e∈N
Corollary 4.13. Let S be a regular F -finite domain. Let R ⊆ S be an extension of
F -finite Noetherian rings such that R is a direct summand of S. Let I ⊆ R denote an
ideal. Then, the set of F -jumping numbers of I in R is a subset of the set of F -jumping
numbers of IS in S. In particular, the set of F -jumping numbers of I in R is formed by
rational numbers and has no accumulation points.
Proof. The first claim follows immediately from the definition of F -jumping numbers and
Theorem 4.12. The second claim follows from the fact that the set of F -jumping numbers
consits of rational numbers and it has no accumulation points for F -finite regular domains
[ST14b, Theorem B].
Remark 4.14. The previous result gives candidates to compute the F -jumping numbers
of I. Furthermore, we obtain an algorithm to find the test ideals and F -jumping numbers
of I given that we know the corresponding objects for IS.
18
J. ÀLVAREZ MONTANER, C. HUNEKE, AND L. NÚÑEZ-BETANCOURT
e
e
If τS (IS λ ) = CSe (IS)⌈p λ⌉ for some e ≫ 0, then τR (I λ ) = CRe I ⌈p λ⌉ . As a consequence,
we can know whether a F -jumping number of IS is also a F -jumping number of I. To
do this, we need to compute first the next F -jumping number for IS, say λ′ . We now
′
compute where the test ideal of τS (IS λ ) and τS (IS λ ) stabilize. We use this to compute
′
the test ideals τR (I λ ) and τR (I λ ). Then, these ideals differ if and only if λ is an F -jumping
number for I in R.
It is natural to ask if the conclusion of Corollary 4.13 holds in characteristic zero.
Question 4.15. Let S = K[x1 , . . . , xn ] be a polynomial ring over a field of characteristic
zero. Let R ⊆ S be such that R a direct summand of S. Let I ⊆ R denote an ideal. Is
the set of jumping numbers for the multiplier ideals of I in R a subset of the set jumping
numbers for the multiplier ideals of IS in S?
4.2. F -thresholds and Bernstein-Sato polynomials. Mustaţă, Takagi and Watanabe
introduced some asymptotic invariants associated to an ideal a in an F -finite regular
local ring R [MTW05]. Roughly speaking, given any ideal J ⊆ R, the F -threshold cJ (a)
measures the containment of powers of a in the Frobenius powers of J. The set of F thresholds when we vary J coincides with the set of F -jumping numbers of a [BMS08,
Corollary 2.30]. Together with the second author, they extended the notion of F -threshold
to all Noetherian rings in prime characteristic [HMTW08]. This definition is given by a
limit, which they showed to exists for F -pure rings. This limit was recently showed to
exist in full generality [DSNBP16, Theorem A]. In this subsection we extend previous
relations between F -threshold, F -jumping numbers, and the Bernstein-Sato polynomial,
previously known only for polynomial rings [MTW05]. For the sake of exposition, we
restrict ourselves to principal ideals and to finitely generated Q-algebras. We start by
recalling this relation in the regular case. Let S = Q[x1 , . . . , xn ]. Let f denote the class of
f in Fp [x1 , . . . , xn ], where p is greater any denominator appearing in f . Then, for p ≫ 0,
bSf (νfa (pe )) ≡ 0 mod p
√
for every ideal a ⊆ Fp [x1 , . . . , xn ] such that f ∈ a [MTW05, Proposition 3.11].
We start by recalling the definition of F -thresholds.
√
Definition 4.16. Let R be a Noetherian rings, and J, a ⊆ R be ideals such that J ⊆ a.
e
Let νJa (pe ) = max{t | J t 6⊆ a[p ] }. We denote the F -threshold of J with respect to a by
νJa (pe )
.
e→∞
pe
ca (J) = lim
It is worth mentioning that, in the singular case, the F -thresholds may differ from the
F -jumping numbers (see [HWY14, Section 6] for a few examples). For this reason, it is
not clear, a priori, that F -thresholds are rational numbers that form a discrete set. We
note that this is the case if R is a direct summand of a regular ring S.
Proposition 4.17. Let S be a regular F -finite domain. Let R ⊆ S be a direct summand
of S. Then, ca (J) is a rational number for every ideal J ⊆ R. Furthermore, the set
√
{ca (J) | a ⊆ R & J ⊆ a}
has no accumulation points.
D-MODULES, BERNSTEIN-SATO POLYNOMIALS & F -INVARIANTS
19
Proof. We have that ca (J) = caS (JS) [HMTW08, Proposition 2.2(v)]. Every F -threshold,
caS (JS), is an F -jumping number for test ideals of a [BMS08]. Both statements, rationality and discreteness, follows from the fact that the set of F -jumping numbers consists in
rational numbers and it has no accumulation points for F -finite regular domains [ST14b,
Theorem B].
We now start relating the Bernstein-Sato polynomial with F -thresholds for direct summands. For this we need to establish a lemma that allow us to restrict Q-linear differential
operators to Z-linear operators.
Lemma 4.18. Let S = Q[x1 , . . . , xn ], I ⊆ S be an ideal, and R = S/I. Let SZ =
Z[x1 , . . . , xn ], J = I ∩ SZ , and RZ = SZ /J. Then, for every δ ∈ DR|Q , there exists c ∈ N
such that cδ(RZ ) ⊆ RZ . Furthermore, cδ ∈ DR|Z .
Proof. Recall that
DR|Q =
{∂ ∈ DS|Q | ∂(I) ⊆ I}
.
IDS|Q
Given δ ∈ DR|Q , let ∂ ∈ DS|Q be such that δf = ∂f for every f ∈ R. Since
d
d
DS|Q = S
,...,
,
dx1
dxn
there exists an integer c such that
d
d
.
,...,
c∂ ∈ SZ
dx1
dxn
Furthermore, c∂(SZ ) ⊆ SZ and c∂(J) ⊆ J. It follows that cδf = c∂f ∈ RZ for every
f ∈ RZ . Moreover, cδ ∈ DR|Z as c∂ ∈ DSZ |Z .
Remark 4.19. Let R be a finitely generated Z-algebra. Then, there is a ring morphism
n
DR|Z → DR⊗Z Fp |Fp , defined by δ 7→ δ, where δ·f = δ · f . Furthermore, this map sends DR|Z
n
to DR⊗Z Fp |Fp . We point out that this map may not be surjective [Smi95, Pages 384-385 ].
We now relate the Bernstein-Sato polynomial to F -thresholds.
Theorem 4.20. Let S = Q[x1 , . . . , xn ], I ⊆ S be an ideal, and R = S/I. Let SZ =
Z[x1 , . . . , xn ], J = I ∩ SZ , and RZ = SZ /J. Suppose that bR
f (s) exists. Then, there exists
m ∈ N, depending only on RZ , such that
a e
bR
f (νf (p )) ≡ 0
mod p
for every prime number p > m, and every ideal a ⊆ RZ /pRZ such that f ∈
particular, this holds when R is a direct summand of a polynomial ring over Q.
P
j
Proof. We pick a differential operator δ(s) = m
j=1 δj s ∈ DR|Q [s] such that
(4.2.1)
√
a. In
δ(t)f t+1 = bf (t)f t
for every t ∈ Z. By Lemma 4.18 we may clean up the coefficients of δ(s) so, without any
loss of generality, we may assume that δ(s) ∈ DR|Z [s] as long as we pick p > c, where c is
as in Lemma 4.18. Moreover, me may multiply both sides of Equation 4.2.1 by the least
common multiple of the denominators of every coefficient of bR
f (s). We pick p bigger than
any of this denominators. We may also assume that it is an equation in RZ .
20
J. ÀLVAREZ MONTANER, C. HUNEKE, AND L. NÚÑEZ-BETANCOURT
Now we reduce this equation mod p. By our choice of p, bR
f (s) is not the zero polynomial
in Fp [s]. We have the equation
δ(t)f
t+1
= bf (t)f
t
for every t ∈ N. We pick p > max nj , where nj denotes the order of the differential operator
(1)
δj . Then, the order of δ(t) is bounded above by p − 1, so δ(t) ∈ DR = HomR (F∗ R, F∗ R)
√
for every t ∈ N. Let a ⊆ RZ /pRZ such that f ∈ a. We set ν = νfa (pe ), thus we have
that f
ν+1
e
e
e
∈ a[p ] and δ(ν + 1)a[p ] ⊆ a[p ] . Then,
ν
bf (ν)f = δ(ν)f
ν+1
e
∈ a[p ] .
e
Since f 6∈ I [p ] , we deduce that bf (ν) ≡ 0 mod p.
We also point out that νfa (pe ) is the e-truncation of the base p-expansion of ca (f ). For
a similar result for F -pure thresholds, we refer to [Her12, Key Lemma].
Recall that the set of F -thresholds may differ from the set of F -jumping numbers for
non-regular rings. Our next result shows that we still have an analogous relation between
Bernstein-Sato polynomials and F -jumping numbers.
Theorem 4.21. Let S = Q[x1 , . . . , xn ], I ⊆ S be an ideal, and R = S/I. Let SZ =
Z[x1 , . . . , xn ], J = I ∩ SZ , and RZ = SZ /J. Suppose that bR
f (s) exists. Then, there exists
m ∈ N, depending only on RZ ,
bR
f (ν) ≡ 0 mod p
ν
if p > m and CRe Z ⊗Z Fp f 6= CRe Z ⊗Z Fp f
summand of a polynomial ring over Q.
ν+1
. In particular, this holds when R is a direct
Proof. As in the proof of Theorem 4.20, for p ≫ 0, we can assume that δ(s) ∈ DR|Z [s].
R
We can assume that no denominator in any coefficient in bR
f (s) becomes zero. Then, bf (s)
is not the zero polynomial in Fp [s]. We have the equation
δ(t)f
t+1
= bf (t)f
t
(1)
for every t ∈ N. We can also assume that δ(t) ∈ DR = HomR (F∗ R, F∗ R) for every t ∈ N.
ν
ν+1
If bf (ν) 6≡ 0 mod p, f ∈ D (1) f . Then,
ν
D (e) f ⊆ D (e) D (1) f
Since D (e) f
a+1
ν+1
CRe Z ⊗Z Fp f
p.
ν
ν+1
= D (e) f
⊆ D (e) f , we conclude that D (e) f
ν
ν+1
.
= D (e) f
ν+1
. Then, CRe Z ⊗Z Fp f
ν
=
by Remark 4.10, which contradicts our hypothesis. Hence, bf (ν) ≡ 0 mod
Remark 4.22. Theorem 4.20 and 4.21 also hold for ideals that are not principal if we
consider the Bernstein-Sato polynomial for arbitrary ideals (see Remark 3.18). These
claims can be also extended for any field of characteristic zero. We focused on principal
ideals over finitely generated Q-algebras for the sake of accessibility. The experts would
easily adapt the arguments in the proof to the more general setting.
D-MODULES, BERNSTEIN-SATO POLYNOMIALS & F -INVARIANTS
21
Acknowledgments
We thank Jack Jeffries for suggesting Examples 3.17 and 3.19. We also thank Jen-Chieh
Hsiao, Luis Narváez-Macarro and Karl Schwede for several helpful comments. The third
author is grateful to Karen E. Smith for inspiring conversations.
References
[ÀMBL05]
Josep Àlvarez Montaner, Manuel Blickle, and Gennady Lyubeznik. Generators of D-modules
in positive characteristic. Math. Res. Lett., 12(4):459–473, 2005. 15
[BB11]
Manuel Blickle and Gebhard Böckle. Cartier modules: finiteness results. J. Reine Angew.
Math., 661:85–123, 2011. 15
[BCHM10] Caucher Birkar, Paolo Cascini, Christopher D. Hacon, and James McKernan. Existence of
minimal models for varieties of log general type. J. Amer. Math. Soc., 23(2):405–468, 2010.
14
[Ber72]
I. N. Bernšteı̆n. Analytic continuation of generalized functions with respect to a parameter.
Funkcional. Anal. i Priložen., 6(4):26–40, 1972. 2, 11
[BGG72]
I. N. Bernšteı̆n, I. M. Gel′ fand, and S. I. Gel′ fand. Differential operators on a cubic cone.
Uspehi Mat. Nauk, 27(1(163)):185–190, 1972. 2, 13
[Bjö79]
Jan-Erik Björk. Rings of differential operators, volume 21 of North-Holland Mathematical
Library. North-Holland Publishing Co., Amsterdam, 1979. 9
[Bli13]
Manuel Blickle. Test ideals via algebras of p−e -linear maps. J. Algebraic Geom., 22(1):49–83,
2013. 15
[BMS06a]
Nero Budur, Mircea Mustaţǎ, and Morihiko Saito. Bernstein-Sato polynomials of arbitrary
varieties. Compos. Math., 142(3):779–797, 2006. 3, 12
[BMS06b]
Nero Budur, Mircea Mustaţă, and Morihiko Saito. Roots of Bernstein-Sato polynomials for
monomial ideals: a positive characteristic approach. Math. Res. Lett., 13(1):125–142, 2006.
4
[BMS08]
Manuel Blickle, Mircea Mustaţǎ, and Karen E. Smith. Discreteness and rationality of F thresholds. Michigan Math. J., 57:43–61, 2008. Special volume in honor of Melvin Hochster.
3, 4, 15, 18, 19
[BMS09]
Manuel Blickle, Mircea Mustaţă, and Karen E. Smith. F -thresholds of hypersurfaces. Trans.
Amer. Math. Soc., 361(12):6549–6565, 2009. 3, 15
[Bou87]
Jean-François Boutot. Singularités rationnelles et quotients par les groupes réductifs. Invent.
Math., 88(1):65–68, 1987. 2
[BS05]
Nero Budur and Morihiko Saito. Multiplier ideals, V -filtration, and spectrum. J. Algebraic
Geom., 14(2):269–282, 2005. 3
[BSTZ10]
Manuel Blickle, Karl Schwede, Shunsuke Takagi, and Wenliang Zhang. Discreteness and
rationality of F -jumping numbers on singular varieties. Math. Ann., 347(4):917–949, 2010.
3
[CEMS]
Alberto Chiecchio, Florian Enescu, Lance Miller, and Karl Schwede. Test ideals in rings
with finitely generated anti-canonical algebras. Journal of the Institute of Mathematics of
Jussieu , to appear. 3, 14
[Cou95]
S. C. Coutinho. A primer of algebraic D-modules, volume 33 of London Mathematical Society
Student Texts. Cambridge University Press, Cambridge, 1995. 9
[Del71]
Pierre Deligne. Théorie de Hodge. I. In Actes du Congrès International des Mathématiciens
(Nice, 1970), Tome 1, pages 425–430. Gauthier-Villars, Paris, 1971. 2
[DL92]
J. Denef and F. Loeser. Caractéristiques d’Euler-Poincaré, fonctions zêta locales et modifications analytiques. J. Amer. Math. Soc., 5(4):705–720, 1992. 3
[DSNBP16] Alessando De Stefani, Luis Núñez-Betancourt, and Felipe Pérez. On the existence of Fthresholds and related limits. arXiv:1605.03264, 2016. 18
[ELSV04]
Lawrence Ein, Robert Lazarsfeld, Karen E. Smith, and Dror Varolin. Jumping coefficients
of multiplier ideals. Duke Math. J., 123(3):469–506, 2004. 3
22
[Gro67]
J. ÀLVAREZ MONTANER, C. HUNEKE, AND L. NÚÑEZ-BETANCOURT
A. Grothendieck. Éléments de géométrie algébrique. IV. Étude locale des schémas et des
morphismes de schémas IV. Inst. Hautes Études Sci. Publ. Math., (32):361, 1967. 4
[Har68]
Robin Hartshorne. Cohomological dimension of algebraic varieties. Ann. of Math. (2),
88:403–450, 1968. 10
[Her12]
Daniel J. Hernández. F -purity of hypersurfaces. Math. Res. Lett., 19(2):389–401, 2012. 20
[Her14]
Daniel J. Hernández. F -pure thresholds of binomial hypersurfaces. Proc. Amer. Math. Soc.,
142(7):2227–2242, 2014. 3
[Her15]
Daniel J. Hernández. F -invariants of diagonal hypersurfaces. Proc. Amer. Math. Soc.,
143(1):87–104, 2015. 3
[HH89]
Melvin Hochster and Craig Huneke. Tight closure and strong F -regularity. Mém. Soc. Math.
France (N.S.), (38):119–133, 1989. Colloque en l’honneur de Pierre Samuel (Orsay, 1987).
2, 14
[HH90]
Melvin Hochster and Craig Huneke. Tight closure, invariant theory, and the Briançon-Skoda
theorem. J. Amer. Math. Soc., 3(1):31–116, 1990. 3, 15
[HH94a]
Melvin Hochster and Craig Huneke. F -regularity, test elements, and smooth base change.
Trans. Amer. Math. Soc., 346(1):1–62, 1994. 3, 15
[HH94b]
Melvin Hochster and Craig Huneke. Tight closure of parameter ideals and splitting in
module-finite extensions. J. Algebraic Geom., 3(4):599–670, 1994. 3
[HM10]
Christopher D. Hacon and James McKernan. Existence of minimal models for varieties of
log general type. II. J. Amer. Math. Soc., 23(2):469–490, 2010. 14
[HM16]
Jen-Chieh Hsiao and Laura Felicia Matusevich. Bernstein-Sato polynomials on normal toric
varieties. arXiv:1608.03646, 2016. 3
[HMTW08] Craig Huneke, Mircea Mustaţă, Shunsuke Takagi, and Kei-ichi Watanabe. F-thresholds,
tight closure, integral closure, and multiplicity bounds. Michigan Math. J., 57:463–483,
2008. Special volume in honor of Melvin Hochster. 18, 19
[HNBWZ16] Daniel J. Hernández, Luis Núñez-Betancourt, Emily E. Witt, and Wenliang Zhang. F-Pure
thresholds of homogeneous polynomials. Michigan Math. J., 65(1):57–87, 2016. 3
[HR74]
Melvin Hochster and Joel L. Roberts. Rings of invariants of reductive groups acting on
regular rings are Cohen-Macaulay. Advances in Math., 13:115–175, 1974. 2
[Hsi12]
Jen-Chieh Hsiao. D-module structure of local cohomology modules of toric algebras. Trans.
Amer. Math. Soc., 364(5):2461–2478, 2012. 3, 10
[HWY14]
Daisuke Hirose, Kei-ichi Watanabe, and Ken-ichi Yoshida. F -thresholds versus a-invariants
for standard graded toric rings. Comm. Algebra, 42(6):2704–2720, 2014. 18
[HX15]
Christopher D. Hacon and Chenyang Xu. On the three dimensional minimal model program
in positive characteristic. J. Amer. Math. Soc., 28(3):711–744, 2015. 14
[HY03]
Nobuo Hara and Ken-Ichi Yoshida. A generalization of tight closure and multiplier ideals.
Trans. Amer. Math. Soc., 355(8):3143–3174 (electronic), 2003. 3, 15
[Jon94]
A. G. Jones. Rings of differential operators on toric varieties. Proc. Edinburgh Math. Soc.
(2), 37(1):143–160, 1994. 2
[Kan77]
Jean-Michel Kantor. Formes et opérateurs différentiels sur les espaces analytiques complexes. Bull. Soc. Math. France Mém., (53):5–80, 1977. 2, 13
[Kas83]
M. Kashiwara. Vanishing cycle sheaves and holonomic systems of differential equations. In
Algebraic geometry (Tokyo/Kyoto, 1982), volume 1016 of Lecture Notes in Math., pages
134–142. Springer, Berlin, 1983. 2
[Kas77]
Masaki Kashiwara. B-functions and holonomic systems. Rationality of roots of B-functions.
Invent. Math., 38(1):33–53, 1976/77. 11, 12
[Kol97]
János Kollár. Singularities of pairs. In Algebraic geometry—Santa Cruz 1995, volume 62 of
Proc. Sympos. Pure Math., pages 221–287. Amer. Math. Soc., Providence, RI, 1997. 3
[Kol10]
János Kollár. Exercises in the birational geometry of algebraic varieties. In Analytic and
algebraic geometry, volume 17 of IAS/Park City Math. Ser., pages 495–524. Amer. Math.
Soc., Providence, RI, 2010. 14
[Laz04]
Robert Lazarsfeld. Positivity in algebraic geometry. II, volume 49 of Ergebnisse der Mathematik und ihrer Grenzgebiete. 3. Folge. A Series of Modern Surveys in Mathematics [Results
D-MODULES, BERNSTEIN-SATO POLYNOMIALS & F -INVARIANTS
[Lev81]
[LS95a]
[LS95b]
[LS99]
[LS01]
[Lyu93]
[Lyu97]
[Lyu00a]
[Lyu00b]
[Mal75]
[Mal83]
[MNM91]
[MR01]
[MTW05]
[Mus87]
[Mus94]
[MVdB98]
[NB12]
[NB13]
[PR05]
[Put16]
[Sab87]
[Sch95]
23
in Mathematics and Related Areas. 3rd Series. A Series of Modern Surveys in Mathematics].
Springer-Verlag, Berlin, 2004. Positivity for vector bundles, and multiplier ideals. 3
Thierry Levasseur. Grade des modules sur certains anneaux filtrés. Comm. Algebra,
9(15):1519–1532, 1981. 2
T. Levasseur and J. T. Stafford. Invariant differential operators and an homomorphism of
Harish-Chandra. J. Amer. Math. Soc., 8(2):365–372, 1995. 2
T. Levasseur and J. T. Stafford. Invariant differential operators and an homomorphism of
Harish-Chandra. J. Amer. Math. Soc., 8(2):365–372, 1995. 2
Gennady Lyubeznik and Karen E. Smith. Strong and weak F -regularity are equivalent for
graded rings. Amer. J. Math., 121(6):1279–1290, 1999. 15
Gennady Lyubeznik and Karen E. Smith. On the commutation of the test ideal with localization and completion. Trans. Amer. Math. Soc., 353(8):3149–3180 (electronic), 2001.
15
Gennady Lyubeznik. Finiteness properties of local cohomology modules (an application of
D-modules to commutative algebra). Invent. Math., 113(1):41–55, 1993. 2, 10
Gennady Lyubeznik. F -modules: applications to local cohomology and D-modules in characteristic p > 0. J. Reine Angew. Math., 491:65–130, 1997. 2, 9
Gennady Lyubeznik. Finiteness properties of local cohomology modules: a characteristicfree approach. J. Pure Appl. Algebra, 151(1):43–50, 2000. 2, 9
Gennady Lyubeznik. Finiteness properties of local cohomology modules for regular local
rings of mixed characteristic: the unramified case. Comm. Algebra, 28(12):5867–5882, 2000.
Special issue in honor of Robin Hartshorne. 2, 9
B. Malgrange. Le polynôme de Bernstein d’une singularité isolée. In Fourier integral operators and partial differential equations (Colloq. Internat., Univ. Nice, Nice, 1974), pages
98–119. Lecture Notes in Math., Vol. 459. Springer, Berlin, 1975. 11, 12
B. Malgrange. Polynômes de Bernstein-Sato et cohomologie évanescente. In Analysis and
topology on singular spaces, II, III (Luminy, 1981), volume 101 of Astérisque, pages 243–
267. Soc. Math. France, Paris, 1983. 2
Z. Mebkhout and L. Narváez-Macarro. La thèorie du polynôme de Bernstein-Sato pour
les algèbres de Tate et de Dwork-Monsky-Washnitzer. Ann. Sci. École Norm. Sup. (4),
24(2):227–256, 1991. 9, 12
J. C. McConnell and J. C. Robson. Noncommutative Noetherian rings, volume 30 of Graduate Studies in Mathematics. American Mathematical Society, Providence, RI, revised edition, 2001. With the cooperation of L. W. Small. 5
Mircea Mustaţǎ, Shunsuke Takagi, and Kei-ichi Watanabe. F-thresholds and Bernstein-Sato
polynomials. pages 341–364, 2005. 3, 4, 18
Ian M. Musson. Rings of differential operators on invariant rings of tori. Trans. Amer. Math.
Soc., 303(2):805–827, 1987. 2
Ian M. Musson. Differential operators on toric varieties. J. Pure Appl. Algebra, 95(3):303–
315, 1994. 2
Ian M. Musson and Michel Van den Bergh. Invariants under tori of rings of differential
operators and related topics. Mem. Amer. Math. Soc., 136(650):viii+85, 1998. 2
Luis Núñez-Betancourt. Local cohomology properties of direct summands. J. Pure Appl.
Algebra, 216(10):2137–2140, 2012. 10
Luis Núñez-Betancourt. On certain rings of differentiable type and finiteness properties of
local cohomology. J. Algebra, 379:1–10, 2013. 9, 12
W. Plesken and D. Robertz. Constructing invariants for finite groups. Experiment. Math.,
14(2):175–188, 2005. 2
Tony J. Puthenpurakal. Local cohomology modules of invariant rings. Math. Proc. Cambridge Philos. Soc., 160(2):299–314, 2016. 6
C. Sabbah. Proximité évanescente. II. Équations fonctionnelles pour plusieurs fonctions
analytiques. Compositio Math., 64(2):213–241, 1987. 3
Gerald W. Schwarz. Lifting differential operators from orbit spaces. Ann. Sci. École Norm.
Sup. (4), 28(3):253–305, 1995. 2, 6
24
J. ÀLVAREZ MONTANER, C. HUNEKE, AND L. NÚÑEZ-BETANCOURT
[Sch11]
[Smi87]
[Smi95]
[Smi97]
[Smi00]
[SS72]
[ST01]
[ST04]
[ST09]
[ST14a]
[ST14b]
[SVdB97]
[Tra00]
[Tra06]
[Tra10]
[TT08]
[TW04]
[Wal05]
[Yek92]
Karl Schwede. Test ideals in non-Q-Gorenstein rings. Trans. Amer. Math. Soc.,
363(11):5925–5941, 2011. 3, 15
S. P. Smith. The global homological dimension of the ring of differential operators on a
nonsingular variety over a field of positive characteristic. J. Algebra, 107(1):98–105, 1987. 4
Karen E. Smith. The D-module structure of F -split rings. Math. Res. Lett., 2(4):377–386,
1995. 2, 6, 19
Karen E. Smith. F -rational rings have rational singularities. Amer. J. Math., 119(1):159–
180, 1997. 3
Karen E. Smith. The multiplier ideal is a universal test ideal. Comm. Algebra, 28(12):5915–
5929, 2000. Special issue in honor of Robin Hartshorne. 3
Mikio Sato and Takuro Shintani. On zeta functions associated with prehomogeneous vector
spaces. Proc. Nat. Acad. Sci. U.S.A., 69:1081–1082, 1972. 2, 11
Mutsumi Saito and William N. Traves. Differential algebras on semigroup algebras. In Symbolic computation: solving equations in algebra, geometry, and engineering (South Hadley,
MA, 2000), volume 286 of Contemp. Math., pages 207–226. Amer. Math. Soc., Providence,
RI, 2001. 2
Mutsumi Saito and William N. Traves. Finite generation of rings of differential operators
of semigroup algebras. J. Algebra, 278(1):76–103, 2004. 2
Mutsumi Saito and Ken Takahashi. Noetherian properties of rings of differential operators
of affine semigroup algebras. Osaka J. Math., 46(2):529–556, 2009. 2
Karl Schwede and Kevin Tucker. On the behavior of test ideals under finite morphisms. J.
Algebraic Geom., 23(3):399–443, 2014. 4
Karl Schwede and Kevin Tucker. Test ideals of non-principal ideals: computations, jumping
numbers, alterations and division theorems. J. Math. Pures Appl. (9), 102(5):891–929, 2014.
3, 17, 19
Karen E. Smith and Michel Van den Bergh. Simplicity of rings of differential operators in
prime characteristic. Proc. London Math. Soc. (3), 75(1):32–62, 1997. 2
William N. Traves. Tight closure and differential simplicity. J. Algebra, 228(2):457–476,
2000. 2
William N. Traves. Differential operators on orbifolds. J. Symbolic Comput., 41(12):1295–
1308, 2006. 2
Will Traves. Differential operators on Grassmann varieties. In Symmetry and spaces, volume
278 of Progr. Math., pages 197–207. Birkhäuser Boston, Inc., Boston, MA, 2010. 2
Shunsuke Takagi and Ryo Takahashi. D-modules over rings with finite F -representation
type. Math. Res. Lett., 15(3):563–581, 2008. 15
Shunsuke Takagi and Kei-ichi Watanabe. On F-pure thresholds. J. Algebra, 282(1):278–297,
2004. 16
Uli Walther. Bernstein-Sato polynomial versus cohomology of the Milnor fiber for generic
hyperplane arrangements. Compos. Math., 141(1):121–145, 2005. 12
Amnon Yekutieli. An explicit construction of the Grothendieck residue complex. Astérisque,
(208):127, 1992. With an appendix by Pramathanath Sastry. 4
Deptartament de Matemàtiques, Universitat Politècnica de Catalunya, Av. Diagonal
647, Barcelona 08028, Spain
E-mail address: [email protected]
Department of Mathematics, University of Virginia, Charlottesville, VA 22904-4135,
USA
E-mail address: [email protected]
Centro de Investigación en Matemáticas, Guanajuato, Gto., México
E-mail address: [email protected]
| 0math.AC
|
arXiv:1611.10093v1 [math.GR] 30 Nov 2016
On products of groups with abelian subgroups of
small index
Bernhard Amberg and Yaroslav Sysak∗
Abstract
It is proved that every group of the form G = AB with two subgroups
A and B each of which is either abelian or has a quasicyclic subgroup
of index 2 is soluble of derived length at most 3 . In particular, if A is
abelian and B is a locally quaternion group, this gives a positive answer
to Question 18.95 of ”Kourovka notebook” posed by A.I.Sozutov.
1
Introduction
Let the group G = AB be the product of two subgroups A and B , i.e.
G = {ab | a ∈ A, b ∈ B} . It was proved by N. Itô that the group G is
metabelian if the subgroups A and B are abelian (see [1, Theorem 2.1.1]).
In connection with Itô’s theorem a natural question is whether every group
G = AB with abelian-by-finite subgroups A and B is metabelian-by-finite [1,
Question 3] or at least soluble-by-finite. However, this seemingly simple question
is very difficult to attack and only partial results in this direction are known.
A positive answer was given for linear groups G by the second author in [7]
(see also [8]) and for residually finite groups G by J. Wilson [1, Theorem 2.3.4].
Furthermore, N.S. Chernikov proved that every group G = AB with centralby-finite subgroups A and B is soluble-by-finite (see [1, Theorem 2.2.5]).
It is natural to consider first groups G = AB where the two factors A and
B have abelian subgroups with small index. There are a few known results
in the case when both factors A and B have an abelian subgroup of index
at most 2 . It was shown in [3] that G is soluble and metacyclic-by-finite if
A and B have cyclic subgroups of index at most 2 , and it is proved in [2]
that G is soluble if A and B are periodic locally dihedral subgroups. A more
general result that G = AB is soluble if each of the factors A and B is either
abelian or generalized dihedral was obtained in [4] by another approach. Here a
group is called generalized dihedral if it contains an abelian subgroup of index 2
and an involution which inverts the elements of this subgroup. Clearly dihedral
groups and locally dihedral groups, i.e. groups with a local system of dihedral
subgroups, are generalized dihedral.
1991 Mathematics Subject Classification. Primary 20D40
Key words and phrases. Products of groups, soluble group, generalized dihedral group,
locally quaternion group
∗ The second author likes to thank the Deutsche Forschungsgemeinschaft for financial support and the Institute of Mathematics of the University of Mainz for its excellent hospitality
during the preparation of this paper.
1
We recall that a group is called quasicyclic (or a Prüfer group) if it is an infinite locally cyclic p -group for some prime p . It is well-known that quasicyclic
subgroups of abelian groups are their direct factors. Furthermore, it seems to
be known and will be shown below that every non-abelian group having a quasicyclic subgroup of index 2 is either an infinite locally dihedral or a locally
quaternion group. It should be noted that for each prime p , up to isomorphism, there exists a unique locally dihedral group whose quasicyclic subgroup
is a p -group, and there is only one locally quaternion group. These and other
details about such groups can be found in [6], p. 45 - 50.
Theorem 1.1. Let the group G = AB be the product of two subgroups A and
B each of which is either abelian or has a quasicyclic subgroup of index 2 .
Then G is soluble with derived length at most 3 . Moreover, if the subgroup
B is non-abelian and X is its quasicyclic subgroup, then AX = XA is a
metabelian subgroup of index 2 in G .
As a direct consequence of this theorem, we have an affirmative answer to
Question 18.95 of the ”Kourovka notebook” posed by A.I.Sozutov.
Corollary 1.2. If a group G = AB is the product of an abelian subgroup A
and a locally quaternion subgroup B , then G is soluble.
It is also easy to see that if each of the factors A and B in Theorem 1.1
has a quasicyclic subgroup of index 2 , then their quasicyclic subgroups are
permutable. As a result of this the following holds.
Corollary 1.3. Let the group G = A1 A2 · · · An be the product of pairwise
permutable subgroups A1 , ..., An each of which contains a quasicyclic subgroup
of index 2 . Then the derived subgroup G′ is a direct product of the quasicyclic
subgroups and the factor group G/G′ is elementary abelian of order 2m for
some positive integer m ≤ n .
The notation is standard. If H is a subgroup of a group G and g ∈ G ,
then the normal closure of H in G is the normal subgroup of G generated by
all conjugates of H in G , and g G is the conjugacy class of G containing g ,
respectively.
2
Preliminary lemmas
Our first lemma lists some simple facts concerning groups with quasicyclic subgroups of index 2 which will be used without further explanation.
Lemma 2.1. Let G be a non-abelian group containing a quasicyclic p -subgroup
X of index 2 and y ∈ G \ X . Then y 2 ∈ X and the following statements
hold:
1) every subgroup of X is characteristic in G ;
2) the group G is either locally dihedral or locally quaternion;
3) the derived subgroup G′ coincides with X ;
4) every proper normal subgroup of G is contained in X ;
2
5) if G is locally quaternion, then p = 2 , y 4 = 1 , xy = x−1 for all
x ∈ X , the center Z(G) coincides with hy 2 i and is contained in every
non-trivial subgroup of G , the coset yX coincides with the conjugacy
class y G = y X ;
6) if G is locally dihedral, then y 2 = 1 , xy = x−1 for all x ∈ X , Z(G) = 1
and the coset yX coincides with the conjugacy class y G = y X for p > 2
and Z(G) is the subgroup of order 2 in X for p = 2 ;
7) the factor group G/Z(G) is locally dihedral.
Proof. In fact, only statement 2) needs an explanation. Clearly G = Xhyi for
some y ∈ G with y 2 ∈ X and each cyclic subgroup hxi of X is normal in G .
Therefore for p > 2 we have y 2 = 1 and either xy = x or xy = x−1 . Since
X contains a unique cyclic subgroup of order pn for each n ≥ 1 , the equality
xy = x for some x 6= 1 holds for all x ∈ X , contrary to the hypothesis that
G is non-abelian. Therefore xy = x−1 for all x ∈ X and hence the group
G is locally dihedral. In the case p = 2 each subgroup hxi of X properly
containing the subgroup hy 2 i has index 2 in the subgroup hx, yi . If x is of
order 2n for some n > 3 , then the element y can be chosen such that either
y 4 = 1 and hx, yi is a generalized quaternion group with xy = x−1 or y 2 = 1
and hx, yi is one of the following groups: dihedral with xy = x−1 , semidihedral
n−2
n−2
with xy = x−1+2
or a group with xy = x1+2
(see [5], Theorem 5.4.3).
It is easy to see that from this list only generalized quaternion and dihedral
subgroups can form an infinite ascending series of subgroups, so that the 2 group G can be either locally quaternion or locally dihedral, as claimed.
Lemma 2.2. Let G be a group and M an abelian minimal normal p -subgroup
of G for some prime p . Then the factor group G/CG (M ) has no non-trivial
finite normal p -subgroup.
Proof. Indeed, if N/CG (M ) is a finite normal p -subgroup of G/CG (M ) and
x is an element of order p in M , then the subgroup hN, xi is finite and so the
centralizer CM (N ) is a non-trivial normal subgroup of G properly contained
in M , contrary to the minimality of M .
We will say that a subset S of G is normal in G if S g = S for each g ∈ G
which means that sg ∈ S for every s ∈ S .
Lemma 2.3. Let G be a group and let A, B be subgroups of G . If a normal
subset S of G is contained in the set AB and S −1 = S , then the normal
subgroup of G generated by S is also contained in AB . In particular, if i is
an involution with iG ⊆ AB and N is the normal closure of the subgroup hii
in G , then AN ∩ BN = A1 B1 with A1 = A ∩ BN and B1 = AN ∩ B .
Proof. Indeed, if s, t ∈ S , then t = ab and (s−1 )a = cd for some elements
a, c ∈ A and b, d ∈ B . Therefore s−1 t = s−1 ab = a(s−1 )a b = (ac)(db) ∈ AB
and hence the subgroup hs | s ∈ Si is contained in AB and normal in G .
Moreover, if N is a normal subgroup of G and N ⊆ AB , then it is easy to
see that AN ∩ BN = (AN ∩ B)N = (AN ∩ B)N = (A ∩ BN )(AN ∩ B) (for
details see [1], Lemma 1.1.3)
The following slight generalization of Itô’s theorem was proved in [7] (see
also [8], Lemma 9).
3
Lemma 2.4. Let G be a group and let A, B be abelian subgroups of G . If H
is a subgroup of G contained in the set AB , then H is metabelian.
3
The product of an abelian group and a group
containing a quasicyclic subgroup of index 2
In this section we consider groups of the form G = AB with an abelian subgroup
A and a subgroup B = X < y > in which X is a quasicyclic subgroup of index
2 and y ∈ B \ X .
Lemma 3.1. Let the group G = AB be the product of an abelian subgroup A
and a non-abelian subgroup B with a quasicyclic subgroup X of index 2 . If
G has non-trivial abelian normal subgroups, then one of these is contained in
the set AX .
Proof. Suppose the contrary and let N be the set of all non-trivial normal
subgroups of G contained in the derived subgroup G′ . Then AG = 1 and
AN X 6= AX for each N ∈ N . Since G = AB = AX ∪ AXy and AX ∩
AXy = ∅ , for every N ∈ N the intersection N X ∩ AXy is non-empty and
so G = AN X . Moreover, as X = B ′ ≤ G′ by Lemma 2.1, it follows that
G′ = DN X with D = A ∩ G′ . It is also clear that G = hA, Xi , because
otherwise hA, Xi = AX is a normal subgroup of index 2 in G . In particular,
A∩X = 1.
For each N ∈ N we put AN = A ∩ BN and BN = AN ∩ B . Then
AN N = BN N = AN BN by [1], Lemma 1.1.4, and the subgroup BN is not
contained in X , because otherwise N is contained in the set AX , contrary to
the assumption. Let XN = BN ∩ X and CN = AN ∩ N XN . Then XN is a
subgroup of index 2 in BN and CN N = N XN is a normal T
subgroup of G ,
because (N XN )X = N XN and (N CN )A = N CN . Put M = N ∈N N .
Since G = AN X for each N ∈ N , the factor group G/N is metabelian
by Lemma 2.4. Therefore also the factor group G/M is metabelian and so its
′
derived subgroup G′ /M
T is abelian. Clearly if M = 1 , then D = A ∩ G ≤
′
AG = 1 and so G = N ∈N N X = X , contrary to the assumption. Thus M
is an abelian minimal normal subgroup of G . If M is finite, then its centralizer
CG (M ) in G contains X and so the group G = A(M X) is metabelian. Then
G′ = DM X is abelian and hence D = A ∩ G′ ≤ AG = 1 . Therefore G′ = M X
and so X is a normal subgroup of G , contrary to the assumption. Thus M
is infinite and then M XM = CM M is an abelian normal subgroup of G with
finite cyclic subgroup XM whose order is a prime power pk ≥ 1 .
pk
If M contains no elements of order p , then M ∩ CM = 1 and CM
= 1 , so
that CM is of order pk . But then the subgroups AM and BM are of order
2pk and hence the subgroup AM M = BM M = AM BM is finite, contrary to
the choice of M . Thus M is an elementary abelian p -subgroup and so the
factor group Ḡ = G/CG (M ) has no non-trivial finite normal p -subgroups by
Lemma 2.2. In particular, the center of Ḡ has no elements of order p . On
the other hand, G = AM X and G′ = DM X with D = A ∩ G′ , so that
Ḡ = ĀX̄ is a metabelian group and its derived subgroup Ḡ′ = D̄X̄ is abelian.
Thus D̄ ≤ Ā ∩ Ḡ′ is a central subgroup of Ḡ which contains no elements of
4
order p . But then X̄ and so each of its subgroup are normal in Ḡ . This final
contradiction completes the proof.
It should be noted that if in Lemma 3.1 the subgroup B is locally dihedral,
then the group G = AB is soluble by [4], Theorem 1.1. Therefore the following
assertion is an easy consequence of this lemma.
Corollary 3.2. If the group G = AB is the product of an abelian subgroup
A and a locally dihedral subgroup B containing a quasicyclic subgroup X of
index 2 , then AX = XA is a metabelian subgroup of index 2 in G .
Proof. Indeed, let H be a maximal normal subgroup of G with respect to the
condition H ⊆ AX . If X ≤ H , then AH = AX is a metabelian subgroup
of index 2 in G by Itô’s theorem. In the other case the intersection H ∩ X
is finite and hence HX/H is the quasicyclic subgroup of index 2 in BH/H .
Since G/H = (AH/H)(BH/H) is the product of the abelian subgroup AH/H
and the locally dihedral subgroup BH/H , the set (AH/H)(HX/H) contains
a non-trivial normal subgroup F/H of G/H by Lemma 3.1. But then F is a
normal subgroup of G which is contained in the set AX and properly contains
H . This contradiction completes the proof.
In the following lemma G = AB is a group with an abelian subgroup A
and a locally quaternion subgroup B = X < y > in which X is the quasicyclic
2 -subgroup of index 2 and y is an element of order 4 , so that xy = x−1 for
each x ∈ X and z = y 2 is the unique involution of B . It turns out that in
this case the conjugacy class z G of z in G is contained in the set AX .
Lemma 3.3. If G = AB and A ∩ B = 1 , then the intersection z A ∩ AXy is
empty.
Proof. Suppose the contrary and let z a = bxy for some elements a, b ∈ A
−1
and x ∈ X . Then b−1 z = (xy)a
and from the equality (xy)2 = z it
−1
follows that (b−1 z)4 = 1 and b−1 zb−1 z = z a . Therefore b−1 z a b−1 = zz a
and hence bz a b = z a z . As z a = bxy , we have b(bxy)b = (bxy)z and so
bxyb = xyz . Thus (xy)−1 b(xy) = zb−1 . Furthermore, bxyb−1 = (zb−1 )a , so
that bzb−1 = ((zb−1 )2 )a = (xy)−a b2 (xy)a , i.e. the elements z and b2 are
conjugate in G by the element g = b−1 (xy)−a . Since g = cd for some c ∈ B
−1
and d ∈ A , we have b2 = z g = z d and so z = (b2 )d = b2 , contrary to the
hypothesis of the lemma. Thus z A ∩ AXy = ∅ , as desired.
Theorem 3.4. Let the group G = AB be the product of an abelian subgroup
A and a locally quaternion subgroup B . If X is the quasicyclic subgroup of
B , then AX = XA is a metabelian subgroup of index 2 in G . In particular,
G is soluble of derived length at most 3 .
Proof. Let Z be the center of B , N the normal closure of Z in G and
X = B ′ , so that X is the quasicyclic subgroup of index 2 in B . If A∩B 6= 1 ,
then Z is contained in A ∩ B by statement 4) of Lemma 2.1 and so N = Z .
Otherwise it follows from Lemma 2.3 that N = Z G = Z A is contained in
the set AX . Then N is a metabelian normal subgroup of G by Lemma 2.4
and the factor group BN/N is locally dihedral by statement 7) of Lemma 2.1.
Since the factor group G/N = (AN/N )(BN/N ) is the product of an abelian
5
subgroup AN/N and the locally dihedral subgroup BN/N , it is soluble by [4],
Theorem 1.1, and so the group G is soluble.
Now if X ≤ N , then AN = AX is a metabelian subgroup of index 2 in
G and so the derived length of G does not exceed 3 . In the other case the
intersection N ∩ X is finite and hence N X/N is the quasicyclic subgroup of
index 2 in BN/N . Therefore AX = XA by Corollary 3.2 and this completes
the proof.
4
The product of a locally quaternion and a
generalized dihedral subgroup
In this section we consider groups of the form G = AB with a locally quaternion
subgroup A and a generalized dihedral subgroup B . The main part is devoted
to the proof that every group G of this form has a non-trivial abelian normal
subgroup. In what follows G = AB is a group in which A = Qhci with a
quasicyclic 2 -subgroup Q of index 2 and an element c of order 4 such that
ac = a−1 for each a ∈ Q and B = X⋊ < y > with an abelian subgroup X
and an involution y such that xy = x−1 for each x ∈ X .
Let d = c2 denote the involution of A . The following assertion is concerned
with the structure of the centralizer CG (d) of d in G . It follows from statement
4) of Lemma 2.1 that the normalizer of every non-trivial normal subgroup of A
is contained in CG (d) .
Lemma 4.1. The centralizer CG (d) is soluble.
Proof. Indeed, if Z = hdi , then the factor group CG (d)/Z = (A/Z)(CB (d)Z/Z)
is a product of the generalized dihedral subgroup A/Z and the subgroup
CB (d)Z/Z which is either abelian or generalized dihedral. Therefore CG (d)/Z
and thus CG (d) is a soluble group by [4], Theorem 1.1, as claimed.
The following lemma shows that if G has no non-trivial abelian normal
subgroup, then the index of A in CG (d) does not exceed 2 .
Lemma 4.2. If CB (d) 6= 1 , then either CX (d) = 1 or G contains a nontrivial abelian normal subgroup.
Proof. If X1 = CX (d) , then X1 is a normal subgroup of B and CG (d) =
ACB (d) . Therefore the normal closure N = X1G is contained in CG (d) , because X1G = X1BA = X1A . Since CG (d) and so N is a soluble subgroup by
Lemma 4.1, this completes the proof.
Consider now the normalizers in A of non-trivial normal subgroups of B .
Lemma 4.3. Let G have no non-trivial abelian normal subgroup. If U is a
non-trivial normal subgroup of B , then NA (U ) = 1 . In particular, A ∩ B = 1 .
Proof. Indeed, if NA (U ) 6= 1 , then d ∈ NA (U ) and so the normal closure
G
B
hdi = hdi is contained in the normalizer NG (U ) = NA (U )B . Since NA (U ) 6=
A , the subgroup NA (U ) is either finite or quasicyclic, so that NG (U ) and thus
G
hdi is soluble. This contradiction completes the proof.
6
Lemma 4.4. If CX (d) = 1 , then G contains a non-trivial abelian normal
subgroup.
Proof. Since G = AB , for each x ∈ B there exist elements a ∈ A and b ∈ B
such that dx = ab . If b ∈
/ X , then b = a−1 dx is an element of order 2 and
x
x
−1
2k
so d ad = a . As a = d for some k ≥ 0 , it follows that dx ddx = d and
hence ab = dx = (dx )d = (ab)d = abd . Therefore bd = b and so b ∈ CB (d) . In
particular, if CB (d) = 1 , then b ∈ X , so that in this case the conjugacy class
dG = dB is contained in the set AX .
Assume that CB (d) 6= 1 and the group G has no non-trivial normal
subgroup. Then CX (d) = 1 by Lemma 4.2 and without loss of generality
CB (d) = hyi . Then G = (Ahyi)X and so the quasicyclic subgroup Q of A is
normalized by y . In particular, dy = d and the subgroup Qhyi can be either
abelian or locally dihedral. We consider first the case when y centralizes Q
and show that in this case the conjugacy class dG is also contained in the set
AX .
Indeed, otherwise there exist elements a ∈ A and b, x ∈ B such that
dx = ab and b ∈
/ X . Then b ∈ CB (d) = hyi by what was proved above, so that
b = y and dx = ay . As dB = dhyiX = dX , we may suppose that x ∈ X . But
−1
2
then dx = (dx )y = ay = dx and hence dx = d . Therefore x2 ∈ hyi and so
x2 = 1 . In particular, if X has no involution, then dG = dX ⊆ AX . We show
next that the case with an involution x ∈ X cannot appear.
Clearly in this case x is a central involution in B and so the subgroup
D = hd, xi generated by the involutions d and x is dihedral. It is easy to
see that d and x cannot be conjugate in G and the center of D is trivial,
because otherwise the centralizer CG (x) properly contains B , contradicting
Lemma 4.3. Thus dx is an element of infinite order and so D = hdxi ⋊ hxi has
no automorphism of finite order more than 2 . On the other hand, if u ∈ A
and v ∈ B , then Duv = D if and only if Du = D and Dv = D , so that
NG (D) = NA (D)NB (D) . Therefore NA (D) = hdi and hence z = (dx)2 is an
element of infinite order in NB (D) . But then z ∈ X and so hzi is a normal
subgroup of B normalized by d , again contradicting Lemma 4.3. Thus X has
no involution, as claimed.
Finally, if N is the normal closure of the subgroup hdi in G , then AN =
N X = A1 X1 with A1 = A∩N X and X1 = AN ∩X by Lemma 2.3. Therefore
the subgroup A1 X1 is soluble by Theorem 3.4, so that N and hence G has a
non-trivial abelian normal subgroup, contrary to our assumption.
Thus the subgroup Qhyi is locally dihedral and so y inverts the elements of
Q . Since A = Qhci with ac = a−1 for all a ∈ A , the element cy centralizes Q
and hence the subgroup Qhcyi is abelian. But then the group G = (Qhcyi)B
as the product of an abelian and a generalized dihedral subgroup is soluble by
[4], Theorem 1.1. This final contradiction completes the proof.
Theorem 4.5. Let the group G = AB be the product of a locally quaternion
subgroup A and a generalized dihedral subgroup B . Then G is soluble. Moreover, if B has a quasicyclic subgroup of index 2 , then G is metabelian.
Proof. If A ∩ X 6= 1 , then the centralizer CG (d) is of index at most 2 in
G and so G is soluble by Lemma 4.1. Let N be a normal subgroup of G
maximal with respect to the condition A ∩ N X = 1 . Then BN = (A ∩ BN )B
and the subgroup A ∩ BN is of order at most 2 . Therefore the subgroup N
7
is soluble and the factor group G/N = (AN/N )(BN/N ) is the product of the
locally quaternion subgroup AN/N and the subgroup BN/N which is either
abelian or generalized dihedral. Hence it follows from Theorem 3.4 and Lemmas
4.2 and 4.4 that G/N has a non-trivial abelian normal subgroup M/N . Put
L = M Q ∩ M X , Q1 = Q ∩ M X and X1 = M Q ∩ X . Then L = M Q1 = M X1
and Q1 6= 1 , because A ∩ M X 6= 1 by the choice of M . It is also clear
that L is a soluble normal subgroup of G , because (M Q1 )A = M Q1 and
(M X1 )B = M X1 . Therefore the factor group G/L and so the group G is
soluble if AL/L is of order 2 . In the other case AL/L is locally dihedral
and BL/L is abelian or generalized dihedral. Since G/L = (AL/L)(BL/L) ,
it follows that G/L and so G is soluble by [4], Theorem 1.1. Moreover, if the
subgroup X is quasicyclic, then the subgroups Q and X centralize each other
by [1], Corollary 3.2.8, so that QX is an abelian normal subgroup of index 4
in G and thus G is metabelian.
The Proof of Theorem 1.1 is completed by a direct application of Corollary
3.2, Theorem 3.4 and Theorem 4.5.
References
[1] Amberg, B., Franciosi, S., de Giovanni, F., Products of Groups. The Clarendon Press, Oxford University Press, Oxford, 1992.
[2] Amberg, B., Fransman, A., Kazarin, L., Products of locally dihedral subgroups, J. Algebra, 350 (2012), 308–317.
[3] Amberg, B., Sysak, Ya., Products of two groups containing cyclic subgroups
of index at most 2, Arch. Math. 90 (2008), 101–111.
[4] Amberg, B., Sysak, Ya., On products of groups which contain abelian subgroups of index at most 2, J. Group Theory 16 (2013), 299 –318.
[5] Gorenstein, D., Finite groups, Harper and Row, New York, 1968.
[6] Kegel, O.H., Wehrfritz, B. A. F., Locally finite groups, North-Holland,
Amsterdam, 1973.
[7] Sysak, Ya., Products of almost abelian groups. In: Investigations of groups
with restrictions for subgroups (In Russian), Akad. Nauk Ukrain. SSR, Inst.
Mat., Kiev (1988), 81–85.
[8] Sysak, Ya., Products of groups and local nearrings, Note Mat. 28 (2008),
181–216.
Address of the authors:
Bernhard Amberg
Institut für Mathematik
der Universität Mainz
D-55099 Mainz
Germany
Yaroslav P. Sysak
Institute of Mathematics
Ukrainian National Academy of Sciences
01601 Kiev
Ukraine
8
| 4math.GR
|
A Deep Q-Learning Agent for the L-Game with
Variable Batch Training
Petros Giannakopoulos and Yannis Cotronis
National and Kapodistrian University of Athens - Dept of Informatics and Telecommunications
Ilisia, 15784 - Greece
Abstract. We employ the Deep Q-Learning algorithm with Experience Replay to
train an agent capable of achieving a high-level of play in the L-Game while selflearning from low-dimensional states. We also employ variable batch size for
training in order to mitigate the loss of the rare reward signal and significantly
accelerate training. Despite the large action space due to the number of possible
moves, the low-dimensional state space and the rarity of rewards, which only come
at the end of a game, DQL is successful in training an agent capable of strong play
without the use of any search methods or domain knowledge.
1
1.1
Introduction
Related Work
The seminal Deep-Mind’s paper [1] demonstrated how to make deep reinforcement
learning obtain human-level performance on a large set of Atari 2600 games. Their
method, called Deep Q-Learning involved Q-learning, a deep convolutional neural
network followed by one fully connected hidden layer and one fully connected output
layer, and a set of additional techniques. The input to the learning system involved
high-dimensional visual pixel data.
Deep-Mind later introduced AlphaGo [2], a Go playing program that combines
Monte Carlo tree search with convolutional neural networks for searching (policy
network) and evaluating (value network) positions, with Deep Reinforcement
Learning used for training both networks and supervised mentoring, used in the initial
stages of training, with games pulled from a database of games between human
players.
An agent for the game of Hex, called NeuroHex was recently introduced [3].
The system for training the agent employed Deep Q-Learning and consisted of a
convolutional network followed by one fully connected output layer, differing from
AlphaGo's architecture which was fully convolutional. They also used supervised
mentoring for network initialization.
[4] showed that DQL can also be effective for reinforcement learning problems
with low-dimensional states. They achieved state-of-the-art performance in Keepaway
soccer using a shallow network made only from fully connected layers and trained
directly from the low-dimensional input state.
One of the earliest and most successful examples of applying Reinforcement
Learning combined with a neural network to board games is TD-gammon [5]. There,
the self-trained agent achieved superhuman levels of play by approximating stateaction values using the Temporal Difference learning method.
1.2
The L-Game
The L-Game is an abstract strategy board game created by Edward de Bono and was
first presented in his book [6]. His goal was to create a game that, contrary to games
like Chess which achieved difficulty through complexity, would be simple yet would
still require a high degree of skill.
The game is played by two players. The board consists of sixteen squares (4x4).
Each player has a flat 3x2 L-shaped piece that exactly covers four squares. In
addition, there are two circular neutral pieces, each occupying a 1x1 square. The
starting position is shown in Figure 1.
Fig. 1: The L-Game’s board
and starting positions.
Fig. 2: One of the final winning arrangements in
the L-Game.
On each turn, a player picks up his L-piece and replaces it on the board to cover
four empty squares, at least one of which must be different from the four squares just
vacated. The piece may be rotated or even flipped over if desired. After the L-piece
has been placed, the player may optionally choose one (but not both) of the neutral
pieces and move it to any vacant square. The objective of the game is to leave your
opponent without a legal move for his L-piece. In Figure 2(a) it is Black’s turn to
play. He first repositions his L-piece and then a neutral piece as shown in 2(b). Now
White is unable to move his L-piece and thus loses.
The L-Game is a perfect information, zero-sum game. Neither player has an
advantage in any way (e.g. by starting first). In a game with two perfect players,
neither will ever win nor lose. The state size of the L-Game allows for precise
analysis. There are known to be 2296 different possible valid ways the pieces can be
arranged, not counting a rotation or mirror of an arrangement as a new arrangement
(which would bring the total arrangements up to 18368) and considering the two
neutral pieces to be identical. Any arrangement can be reached during the game, with
it being any player's turn. There are 15 basic winning positions, where one of the Lpieces is blocked. Another 14 positions are known to lead to a win after a maximum
of 5 moves. Building a winning strategy requires memory of the known winning
positions, spatial acuity, and the ability to plan ahead.
The domain can be modelled as a fully observable Markov Decision Process
(MDP) where, for each player, a state s is represented by the position of the player
and neutral pieces on the game board. Action a is a legal move chosen by the player
which leads to the new state of the game board s’. If a ends the game, s’ is a terminal
state and each player receives a reward r, positive or negative, depending if he won or
lost the game.
2
2.1
Overview of This Work
Challenges
In this work we explore the application of Reinforcement Learning [7] with Deep QLearning to the L-Game. There are some additional challenges involved in applying
this method, so successful with Atari, to the L-Game and to other board games with
similar general domain characteristics.
One challenge is the large number of actions: up to 128 possible moves for each
state of the L-Game. Since Q-learning performs a maximization over all available
actions, this large number might cause the noise in estimation to overwhelm the
useful signal, resulting in catastrophic maximization bias. Despite this, we found that
the linear network was still able to achieve good learning results from the lowdimensional input states.
Another challenge is that the reward signal occurs only at the end of a game, so
it is infrequent, coming after a sequence of several turns. This means that most
updates are based only on network evaluations without immediate win/loss feedback.
The question is whether the learning process will allow this end-of-game reward
information to propagate back to the middle and early game. To address this
challenge, we save and then assemble all the states for a single played game, from the
first state to the last state (were the reward signal appears), into a batch. We then use
this batch to update the gradient. Essentially, we perform batch updates but the batch
size is not fixed but variable and equal to the number of states contained in a single
played game. This helps to minimize, as much as possible, the variance of stochastic
gradient updates and allows reaching good results with fewer training epochs
(games).
2.2
Problem Definition
We use a feedback scheme where a win is set to be worth a reward of +1 and a loss a
reward of -1. The reward given for all intermediate moves between the start and end
of game is 0. Therefore, the ground truth for the Q-value of every possible stateaction pair q(S,A) assumes a binary form with its value being either 1 or -1 for every
possible state-action pair. We want the agent to maximize his reward during a single
game (or episode) by winning. The network approximates the Q-value of every action
for a given state, or simply the subjective probability that a particular move will result
in a win minus the probability that it will result in a loss. The accurate prediction of
the true value of q(S,A) for all states encountered during play is the measure of how
strong a player the agent will be by following the policy π which takes the action a
with the highest estimated Q-value (or win probability) argmax[Q(s)[a]] for each
state s.
2.3
Problem Modelling
We use the Torch library [8] to build and train out network. The input state size is the
size of the board: 4x4 = 16 data points. This dictates the use of an input layer to the
network consisting of 16 nodes. There up to 128 possible moves (or actions) which
points to the use of an output layer consisting of 128 nodes. We experimented with
various sizes and number of hidden layers and we found that the best results were
obtained with just 2 hidden layers of 512 nodes width each. Due to the low
dimensionality of the input state, deep architectures do not give an advantage in the
learning process since there’s less need to model complex non-linear relationships in
problems with simple input such in this case. Activation function for every layer
except the output was selected to be Rectified Linear Units (ReLU). The output of the
network is a vector containing the value of every possible move corresponding to the
input state. Illegal moves (occupied positions) are still evaluated by the network but
are subsequently pruned.
2.4
Learning Process and Results
We do not use any mentoring to initialize the network. All the learning is performed
through self-play. We found that random exploration during training was enough for
the network to experience and successfully learn from a wide variety of game
positions.
We implement Deep Q-learning with Experience Replay as first introduced in
[1]. Our implementation differentiates in the fact that an episode (=game) consists of
a sequence of state-action pairs with a singular reward being given at the end of the
game based on win or loss and no rewards being given for actions in-between the start
and the end. For this reason, storing and randomly sampling state-action pairs from
the replay memory either individually or in small batches may cause the rare reward
signal to be lost in the estimation noise. In an attempt to mitigate this, we temporarily
keep all the state-action pairs encountered during a game in a table and, once the
game ends, we push this game and all the experiences acquired with it in the
experience replay memory as shown in Figure 3. Sampling from the replay memory is
done in similar fashion: a full previously played game is randomly sampled and used
for batch-updating the weights of the network.
Fig. 3: Structure of the Replay Memory. Each game consists of an arbitrary
amount of state-action pairs leading to the final state and the reward.
We save a large set of the most recently played games (10000) in the Replay
Memory and sample N games from that set on each recall from the memory. A value
of N = 0 means that Replay Memory is not used and it serves as the baseline. We
experimented with the impact the number of games sampled per recall has on learning
performance for the first 25000 epochs (Figure 4). It is evident that the use of
experience replay improves performance considerably over baseline up to N = 10
games sampled per recall. Higher sample sizes do not appear to provide additional
benefits and very high sample sizes may even slow down learning. We ended up
using a sample size of N = 10 games per recall.
Since we want to place as much weight as possible to the outcome of a game
versus intermediate moves, we need to use a high discount factor γ. We tried values
for γ between 0.7 and 1 and we found a value of γ = 0.9 to deliver the best training
results. Lower values yielded slower convergence while higher did not provide any
further acceleration of learning and values approaching 1 caused action values to
diverge.
We used an annealed ϵ schedule for the ϵ-greedy policy. We decrease it linearly
from 0.05 to 0.01, meaning the agent makes a random move from 5% of the time at
the beginning of training to 1% near the end. ϵ is fixed at 0.01 during validating.
For the backpropagation algorithm, we experimented with ADADELTA [9],
RMSProp and SGD with Nesterov momentum [10]. We found SGD with Nesterov
momentum and an annealed learning rate schedule to deliver the best results,
achieving faster approach to convergence and the lowest final error. The performance
of each backpropagation algorithm for the first 25000 epochs is shown in Figure 5.
Fig. 4: Effect of Replay Memory
sample size on learning performance.
Fig. 5: Performance of the
backpropagation algorithms tested.
Figure 6 shows the learning performance for the first 25000 epochs when using
a fixed mini-batch size of 32 episodes [(s, a, r, s’) pairs] per learning step to update
the gradients, vs. using an entire game consisting of an arbitrary number of episodes.
For training, the agent assumes the roles of both players and begins learning through
self-play for 100000 games. To gauge the agent’s progress we periodically (every
1000 training episodes) validate by playing 1000 games against an agent playing
randomly and against a perfect agent based on Minimax who can never lose. Training
took 9 hours on an Intel® Core™ i7-3770K. The final trained agent achieves a 98%
winning rate, playing as either Player 1 or Player 2, over 10000 games versus the
random player, while making a random move every 100 moves. Furthermore, it
achieves a 95% draw rate versus a perfect (undefeatable) minimax player. We
consider a game to be drawn in this case, if it goes on for over 100 turns without a
winner.
Fig. 6: Learning rate using a fixed
mini-batch size vs. using variable.
Fig. 7: Agent performance versus a
random player as training progresses.
2.5
Conclusion
In this paper we developed a game playing agent based on Deep Q-Learning for a
challenging board game. We evaluated the performance of Deep Q-Learning on a task
which involves, contrary to the original Atari video game playing application of Deep
Q-Learning, learning from low-dimensional states, a large action space and a rare
singular reward coming at the end of an episode. We also did not use mentoring for
guiding the network at the initial stages of training and all exploration is performed by
the agent. The results of the experiments show that Deep Q-Learning is able to
produce well-performing agents, in spite of these challenges. Our agent achieves
playing performance close to a perfect player, without any domain knowledge and
mentoring, purely through self-play. Deep Learning techniques do increase
performance, with experience replay significantly improving the results. For the
backpropagation algorithm, we found SGD with Nesterov momentum and a decaying
learning rate to perform better in this task compared to adaptive methods such as
RMSProp and ADADELTA. We found that the use of variable batch training can
provide substantial benefits to the performance of Reinforcement Learning
applications on the domain of board games, where the reward signal is inherently rare,
allowing an agent to learn more efficiently from it.
References
[1]
Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei a Rusu, Joel Veness, Marc G
Bellemare, Alex Graves, Martin Riedmiller, Andreas K Fidjeland, Georg Ostrovski, Stig Petersen,
Charles Beattie, Amir Sadik, Ioannis Antonoglou, Helen King, Dharshan Kumaran, Daan Wierstra,
Shane Legg, and Demis Hassabis. Human-level control through deep reinforcement learning.
Nature, 518(7540):529–533, 2015.
[2]
David Silver, Aja Huang, Chris J. Maddison, Arthur Guez, Laurent Sifre, George van den Driessche,
Julian Schrittwieser, Ioannis Antonoglou, Veda Panneershelvam, Marc Lanctot, Sander Dieleman,
Dominik Grewe, John Nham, Nal Kalchbrenner, Ilya Sutskever, Timothy Lillicrap, Madeleine
Leach, Koray Kavukcuoglu, Thore Graepel, and Demis Hassabis. Mastering the game of Go with
deep neural networks and tree search. Nature, 529(7587):484–489, 2016.
[3]
Kenny Young, Gautham Vasan, Ryan Hayward. NeuroHex: A Deep Q-learning Hex Agent.
arXiv:1604.07097 [cs.AI], 2016
[4]
Mateusz Kurek, Wojciech Jaskowski. Heterogeneous Team Deep Q-Learning in Low-Dimensional
Multi-Agent Environments. Proceedings of IEEE 2016 Conference on Computational Intelligence
and Games, 201-208, 2016.
[5]
Gerald Tesauro. Temporal difference learning and td-gammon. Communications of the ACM,
38(3):58–68, 1995
[6]
Edward de Bono. The Five-Day Course in Thinking. Penguin Books, 1992.
[7]
R.S. Sutton and A.G. Barto. Reinforcement learning, volume 9. MIT Press, 1998.
[8]
R. Collobert and K. Kavukcuoglu and C. Farabet. Torch7: A Matlab-like Environment for Machine
Learning. BigLearn, NIPS Workshop, 2011.
[9]
Matthew D. Zeiler. ADADELTA: An Adaptive Learning Rate Method. arXiv:1212.5701 [cs.LG],
2012.
[10]
Ilya Sutskever, James Martens, George Dahl, Geoffrey Hinton. On the importance of initialization
and momentum in deep learning. JMLR W&CP 28 (3) :1139-1147, 2013
| 2cs.AI
|
A study on the combined interplay between
stochastic fluctuations and the number of flagella
in bacterial chemotaxis
Daniela Besozzia Paolo Cazzanigab Matteo Dugoa
Dario Pescinib Giancarlo Maurib
a Università degli Studi di Milano
Dipartimento di Informatica e Comunicazione
Via Comelico 39, 20135 Milano, Italy
[email protected], [email protected]
b Università degli Studi di Milano-Bicocca
Dipartimento di Informatica, Sistemistica e Comunicazione
Viale Sarca 336, 20126 Milano, Italy
cazzaniga/pescini/[email protected]
The chemotactic pathway allows bacteria to respond and adapt to environmental changes, by tuning
the tumbling and running motions that are due to clockwise and counterclockwise rotations of their
flagella. The pathway is tightly regulated by feedback mechanisms governed by the phosphorylation
and methylation of several proteins. In this paper, we present a detailed mechanistic model for
chemotaxis, that considers all of its transmembrane and cytoplasmic components, and their mutual
interactions. Stochastic simulations of the dynamics of a pivotal protein, CheYp, are performed by
means of tau leaping algorithm. This approach is then used to investigate the interplay between the
stochastic fluctuations of CheYp amount and the number of cellular flagella. Our results suggest that
the combination of these factors might represent a relevant component for chemotaxis. Moreover, we
study the pathway under various conditions, such as different methylation levels and ligand amounts,
in order to test its adaptation response. Some issues for future work are finally discussed.
1
Introduction
Recent experimental investigations at the single-cell level [9] have evidenced the presence of biological
noise, due to the inherently stochastic interactions between those molecular species that occur in low
amounts inside the cell. Standard modeling approaches based on ordinary differential equations cannot
effectively capture the effects of biological random processes, which are able to lead the cell to different
states (e.g., bistability). In the last years, indeed, many algorithms that perform stochastic simulations
of biochemical reaction systems have proved their intrinsic suitability for reproducing the dynamics of
many cellular processes (see, e.g., [24] and references therein). For instance, the stochastic simulation
algorithm (SSA) [10] is able to provide an exact realization of the system dynamics and to account for
random fluctuations in the temporal evolution of molecules; however, it can be very slow for those systems where the number of reactions – or even the amount of a few molecular species – is large, as it is
frequently the case for complex cellular processes involving many species and many interactions among
them. In order to overcome this drawback, a faster and reliable stochastic simulation algorithm, called
tau leaping [4], has been recently proposed. Tau leaping represents an efficient tool for the modeling of
R.J.Back, I.Petre, E. de Vink (Eds.): Computational
Models for Cell Processes (CompMod 2009)
EPTCS 6, 2009, pp. 47–62, doi:10.4204/EPTCS.6.4
48
Interplay of stochastic fluctuations and number of flagella in bacterial chemotaxis
stochastic processes in individual cells (see, e.g., [6]), as it can easily handle very detailed and mechanistic descriptions of all possible molecular interactions or molecule modifications, which can bring about
a huge increase in the number of reactions and molecular species (e.g., phosphorylation or methylation
states of proteins). In this work we exploit the efficiency of tau leaping to simulate the dynamics of
bacterial chemotaxis, in order to generate stochastic time series of a pivotal chemotactic protein, which
will be then analyzed with respect to the number of flagella in bacterial cells.
Chemotaxis is an efficient signal transduction pathway which allows bacterial cells to move directionally, in response to specific attractants or repellents occurring in their surroundings. The pathway consists
of several transmembrane and cytoplasmic proteins acting as signal sensors and response regulators [11],
which rule the reversal of the flagellar motor (governed by the phosphorylation and dephosphorylation
of a key protein, CheY). This process induces a switch between running and tumbling movements, with
a frequency that allows a temporal sampling (through random walks) of homogeneous environments.
Anyway, in the presence of a gradient of attractants or repellents, the bacteria are able to respond quickly
by reducing the frequency of flagellar reversal between clockwise and counterclockwise rotations, which
cause a longer running motion in a biased direction. The frequency of switching is then reset to the
random walk level if the concentration of the external ligands remains constant in time. At the molecular scale, this adaptation property is implemented by the coordinated action of methyltransferase and
methylesterase proteins acting on the transmembrane receptors.
The genetic regulation and biochemical functions of the proteins involved in chemotaxis are well
known, and several models have already been proposed to study their complex interplay as well as the
robustness of this system [1, 23, 17, 13, 15, 19]. In the model we present hereby, we consider very
detailed protein-protein interactions for the chemotactic pathway in E. coli, in response to attractant
molecules, which sum up to 62 biochemical reactions and 32 molecular species. The temporal evolution
of the phosphorylated form of CheY (CheYp) is investigated under different conditions, such as the
deletion of other proteins involved in the pathway, the addition of distinct amounts of external ligand,
and the effect of different methylation states.
The results obtained through stochastic simulations of this model are then used to propose an analysis on the interplay between the stochastic fluctuations of CheYp and the number of cellular flagella,
which occur in a few units in the individual bacterium (around half a dozen in E. coli). The aim of this
analysis is to devise the mean time periods during which the cell either performs a running or a tumbling
motion, considering both the coordination of flagella and the randomness that is intrinsic in the chemotactic pathway. Indeed, experimental observations show that the running motion requires all flagella to
be simultaneously synchronized in the counterclockwise rotation, which occurs when CheYp is not interacting with the proteins regulating the flagellar motor; on the contrary, when at least one flagellum is not
coordinated with the others, then the bacterium performs a tumbling movement. To distinguish between
these two states, we will assume that the cell is sensitive to a threshold level of CheYp, that is evaluated
as the mean value of CheYp at steady state. Because of stochastic fluctuations, the amount of CheYp
will randomly switch from below to above this value, thus reversing the rotation from counterclockwise
to clockwise rotations of each flagellum. Therefore, the original contribution of our work consists in
linking the synchronization of all flagella to the stochastic fluctuations of CheYp, as the core component that stands at the basis of chemotactic motions. To this aim, we define a procedure to identify the
synchronization of rotations of all flagella, and we use it to compare the mean time intervals of running
and tumbling motions – as well as of the adaptation times after ligand addition – according to a varying
number of flagella.
The paper is structured as follows. In Section 2 we give a description of the bacterial chemotactic
pathway, and present the mechanistic model of this system. In Section 3, after a brief explanation of
D. Besozzi et al.
49
the functioning of tau leaping, we show the results of simulations for the temporal evolution of the
pivotal protein in chemotaxis (CheYp), that have been obtained by using the model previously defined.
Then, in Section 4 we exploit the outcome of stochastic simulations to study the relations between the
fluctuations of CheYp and the number of flagella occurring in a bacterial cell. We conclude the paper
with some topics for future extensions of our work.
2
The modeling of bacterial chemotaxis
In this section we present the chemotaxis signaling pathway and define the mechanistic model that describes the molecular interactions therein involved.
2.1 Bacterial chemotaxis
Chemotaxis is a signal transduction pathway that allows swimming bacteria to perform biased movements in ever-changing environments, by efficiently sensing concentration gradients of beneficial or
toxic chemicals in their immediate proximity. The binding of ligand molecules triggers an event cascade
involving several transmembrane and cytoplasmic proteins, which eventually affects the concentration
of a pivotal response regulator, CheY. This protein rapidly diffuses inside the cell and interacts with the
proteins of the flagellar motors, thus inducing clockwise (CW) and counterclockwise (CCW) rotation
of each flagellum. When flagella are turning CW, they are uncoordinated and the bacterium performs
a tumbling movement, while if they are all turning CCW, they form a bundle and get coordinated, thus
allowing the cell to swim directionally (the so-called running movement). In a homogeneous environment, bacteria perform a temporal sampling of their surroundings by moving with a random walk, that
is caused by a high switch frequency of the flagellar motors rotations, that alternate rapid tumblings
with short runnings. In the presence of a ligand concentration gradient, instead, bacteria carry out directional swimming toward/against the attractants/repellents, by reducing the switch frequency of flagella
rotations, that results in longer running movements. If the ligand concentration remains constant in time,
then the switch frequency is reset to the prestimulus level, therefore realizing an adaptation of the chemotactic response to the change in ligand concentration. In what follows, we consider the chemosensory
system of E. coli bacteria, in response to attractant chemicals.
The chemotactic pathway, depicted in Figure 1, has been well characterized from a molecular point
of view [3, 11, 25]. External signals are detected by transmembrane methyl-accepting proteins (MCPs),
which are linked to cytoplasmic histidine protein kinases (CheA) by means of scaffold proteins (CheW).
These three proteins constitute the sensor module (i.e. the receptor complexes) of the whole pathway;
each protein occurs as a dimer in every receptor complex. The role of CheA is to transduce the presence
of an external ligand toward the inside of the cell, by phosphorylating two cytoplasmic proteins, called
CheY and CheB. The transfer of the phosphoryl group to these proteins is more probable – that is, the
activity of CheA is stronger – in absence of external ligands. CheY and CheB compete for the binding
to CheA, but the phosphotransfer to CheY is faster than to CheB [25]; this fact assures that the proper
chemotactic response can be generated before the process of adaptation occurs, as explained hereafter.
CheY is the response regulator protein which, after being phosphorylated, interacts with the proteins
FliM of the flagellar motors, inducing the CW rotation of the flagellum and the tumbling movements
(FliM is a key component of the processes that stand downstream of the chemotaxis signaling, and
therefore will not be explicitly included in our model; anyway, some considerations about its role within
the model are discussed in Section 5). In presence of external ligands, the activity of CheA is reduced: the
50
Interplay of stochastic fluctuations and number of flagella in bacterial chemotaxis
Figure 1: Signal transduction pathway in bacterial chemotaxis: solid arrows indicate enzyme-catalyzed
reactions, dashed arrows indicate autocatalysis; CH3 denotes the methyl group, P the phosphoryl group
(the dimensions of components are not scaled).
concentrations of phosphorylated CheY diminishes, its interaction with the flagellar motors is reduced,
the CCW rotation is switched on, and bacteria can perform longer running movements. The termination
of this signal transduction process is mediated by another cytoplasmic protein, CheZ, which acts as
an allosteric activator of CheY dephosphorylation. Concurrently to the processes involving CheY, the
chemosensory system possesses an adaptation response which depends on the methylation level of the
receptors. Methylation reactions are modulated by the coordinated interplay between proteins CheR
and CheB. Up to 4-6 methyl groups are constantly transferred to the cytoplasmic domain of MCPs by
the constitutively active methyltransferases CheR. On the other side, the demethylation of MCPs occurs
by means of the phosphorylated form of the methylesterase CheB. The methylation state of MCPs also
intervene on the regulation of CheA: when MCPs are highly methylated, CheA is more active; when
MCPs are unmethylated, the activity of CheA is reduced. In the latter case, also the concentrations
of phosphorylated CheB diminishes, and this in turn lets the methylation state of MCPs increase, with
a consequent renewed activity of CheA, and so on through a continuous feedback control. Therefore,
the cell is able to adapt to environmental changes and return to the random walk sampling when the
concentration gradient of the attractant remains constant in time. This feedback mechanism also allows
bacteria to widen the range of ligand concentration to which they can respond, making them very sensible
to low environmental variations.
2.2 A mechanistic model
For the modeling of the chemotaxis pathway, we have considered detailed protein-protein interactions
which sum up to a total of 62 reactions and 32 molecular species [7]. The initial amounts – given as
number of molecules, as reported in [21] – of the 7 elementary chemotactic proteins are the following:
4000 dimers of MCPs; 4000 dimers of CheW; 4000 dimers of CheA; 17000 monomers of CheY; 12000
monomers of CheZ; 200 monomers of CheR; 1700 monomers of CheB (plus a constant amount of 1.2
·106 ATP molecules that are needed for phosphorylation reactions). The initial amounts of all other
molecular species appearing in the model are equal to zero, as they are produced by mimicking the formation and dissociation of protein complexes, and by describing the phosphorylation/dephosphorylation
of cytoplasmic proteins and the methylation/demethylation of MCPs, in both the conditions of presence
and absence of external ligands.
D. Besozzi et al.
51
Each reaction in the model is given in the form “reagents → products”, where the notation X + Y
is used to represent a molecular interaction between species X and Y, while X::Y denotes that X and
Y are chemically bound in the formation of a complex (see Table 1). Note that only monomolecular
or bimolecular reactions are here considered; the formation of complexes consisting of more than two
species is performed stepwise. The phosphorylated form of species X, with X ∈ {CheA, CheB, CheY},
is denoted by Xp, while the methylation state of receptor MCP is denoted by MCPm , for m = 0, . . . , 4
(that is, five methylation states are considered).
The reactions describe the following molecular interactions:
• association of the three dimers (2MCP, 2CheW and 2CheA) constituting each ternary receptor
complex (reactions 1-4);
• binding and unbinding of ligand molecules to the receptor complex in the five methylation states
(reactions 28-32 and 33-37, respectively);
• methylation and demethylation of MCPs, in absence and in presence of ligand molecules (reactions
5-8 and 9-12, 38-41 and 42-45, respectively);
• autophosphorylation of CheA in the five methylation states of MCPs, in absence and in presence
of ligand molecules (reactions 13-17 and 46-50, respectively);
• phosphotransfer to CheY in the five methylation states of MCPs, in absence and in presence of
ligand molecules (reactions 18-22 and 51-55, respectively);
• phosphotransfer to CheB in the five methylation states of MCPs, in absence and in presence of
ligand molecules (reactions 23-27 and 56-60, respectively);
• dephosphorylation of CheYp and CheBp (reactions 61-62).
According to literature, the ternary receptor complex 2MCPm ::2CheW::2CheA is assumed to be
stable for the duration of the signal transduction process [22]; moreover, the synthesis and degradation
rates of all chemotactic proteins are assumed to occur at a much slower scale than the chemotactic
response (hence, the reactions corresponding to these processes have not been included in the model).
A stochastic constant is associated to each reaction: it is needed to evaluate the probability of that
reaction to occur when performing stochastic simulations, as explained in Section 3.1. The stochastic
constants used for the simulations shown in Section 3.2 are reported in the caption of Table 1 (all values
are expressed in sec−1 ). These values have been partly derived from literature [18], and partly tuned to
account for the following biological features [14, 16]: (1) the binding affinity of the ligand is directly
proportional to the methylation state of MCPs; (2) the ligand-receptor binding reactions occur at a faster
rate with respect to phosphorylation and methylation/demethylation reactions; (3) the methylation and
demethylation activities of CheR and CheBp are, respectively, inversely and directly proportional to the
methylation state of MCPs; (4) the rate of phosphotransfer from CheA to CheY and CheB depends on the
rate of autophosphorylation of CheA. According to these constraints, which set the relative magnitude of
some constants with respect to others, the estimation of the unavailable constants has been performed by
testing the effect of a range of values for each constant within every module of the model (that is, a group
of reactions corresponding to a specific process, such as, e.g., reactions 1-4 that model the formation of
the receptor complexes). Within this range, the finally chosen value for each constant is the one that gave
a good reproduction of the expected behaviour of the biological subsystem described by that module.
Then, every other module has been sequentially added to the previous ones, following the same iterative
process, to perform a comprehensive and biologically correct simulation of the whole pathway.
Interplay of stochastic fluctuations and number of flagella in bacterial chemotaxis
52
Table 1: The 62 reactions of the model of bacterial chemotaxis. The values of the corresponding
stochastic constants are: c1 = 0.1, c2 = 0.01, c3 = 0.1, c4 = 0.02, c5 = 0.325, c6 = 0.29, c7 = 0.165, c8 =
0.05, c9 = 0.0044, c10 = 0.0175, c11 = 0.0306, c12 = 0.035, c13 = 5.0 · 10−7 , c14 = 7.0 · 10−6 , c15 = 2.8 ·
10−5 , c16 = 5.0 · 10−5 , c17 = 6.8 · 10−5 , c18 = 5.0 · 10−4 , c19 = 0.0035, c20 = 0.014, c21 = 0.025, c22 =
0.0336, c23 = 2.0 · 10−4 , c24 = 0.0014, c25 = 0.0056, c26 = 0.01, c27 = 0.0135, c28 = 0.6, c29 = 0.8, c30 =
1.0, c31 = 1.2, c32 = 1.4, c33 = 15.0, c34 = 15.0, c35 = 15.0, c36 = 15.0, c37 = 15.0, c38 = 4.0 · 10−4 , c39 =
3.75 · 10−4 , c40 = 3.5 · 10−4 , c41 = 2.125 · 10−4 , c42 = 6.0 · 10−4 , c43 = 0.0044, c44 = 0.0175, c45 =
0.0343, c46 = 1.0 · 10−8 , c47 = 5.0 · 10−7 , c48 = 7.0 · 10−6 , c49 = 2.8 · 10−5 , c50 = 5.0 · 10−5 , c51 = 1.0 ·
10−5 , c52 = 5.0 · 10−4 , c53 = 0.0035, c54 = 0.014, c55 = 0.03, c56 = 1.0 · 10−5 , c57 = 2.0 · 10−4 , c58 =
0.0014, c59 = 0.0056, c60 = 0.0112, c61 = 0.0080, c62 = 1.0
1
2
3
4
5-8
9-12
13-17
18-22
23-27
28-32
33-37
38-41
42-45
46-50
51-55
56-60
61
62
3
Reagents
2MCPm + 2CheW
2MCPm ::2CheW
2MCPm ::2CheW + 2CheA
2MCPm ::2CheW::2CheA
2MCPm ::2CheW::2CheA + CheR
2MCPm ::2CheW::2CheA + CheBp
2MCPm ::2CheW::2CheA + ATP
2MCPm ::2CheW::2CheAp + CheY
2MCPm ::2CheW::2CheAp + CheB
lig + 2MCPm ::2CheW::2CheA
lig::2MCPm ::2CheW::2CheA
lig::2MCPm ::2CheW::2CheA + CheR
lig::2MCPm ::2CheW::2CheA + CheBp
lig::2MCPm ::2CheW::2CheA + ATP
lig::2MCPm ::2CheW::2CheAp + CheY
lig::2MCPm ::2CheW::2CheAp + CheB
CheYp + CheZ
CheBp
Products
2MCPm ::2CheW
2MCPm + 2CheW
2MCPm ::2CheW::2CheA
2MCPm ::2CheW + 2CheA
2MCPm+1 ::2CheW::2CheA + CheR
2MCPm−1 ::2CheW::2CheA + CheBp
2MCPm ::2CheW::2CheAp
2MCPm ::2CheW::2CheA + CheYp
2MCPm ::2CheW::2CheA + CheBp
lig::2MCPm ::2CheW::2CheA
lig + 2MCPm ::2CheW::2CheA
lig::2MCPm+1 ::2CheW::2CheA + CheR
lig::2MCPm−1 ::2CheW::2CheA + CheBp
lig::2MCPm ::2CheW::2CheAp
lig::2MCPm ::2CheW::2CheA + CheYp
lig::2MCPm ::2CheW::2CheA + CheBp
CheY + CheZ
CheB
Methyl. state
m=0
m=0
m=0
m=0
m = 0, . . . , 3
m = 1, . . . , 4
m = 0, . . . , 4
m = 0, . . . , 4
m = 0, . . . , 4
m = 0, . . . , 4
m = 0, . . . , 4
m = 0, . . . , 3
m = 1, . . . , 4
m = 0, · · · , 4
m = 0, . . . , 4
m = 0, . . . , 4
Stochastic simulations of the chemotactic response regulator
In this section, we briefly present the stochastic algorithm used to perform the simulation of the model
defined in Section 2.2. Then, we show the results of the dynamics of the chemotactic response regulator
protein (the phosphorylated form of CheY) under different conditions of the chemotactic system.
3.1 Tau leaping
The algorithm called tau leaping [4] is an approximated and faster version of the seminal Gillespie’s
stochastic simulation algorithm (SSA) [10]. Both algorithms allow to generate the temporal evolution of
chemicals contained inside a well stirred volume, in given and fixed experimental conditions. Chemicals
interact with each other by means of given reactions, whose physical and chemical properties are encompassed in a specified stochastic constant associated to each reaction. Reactions are applied according to a
probability distribution, that is determined – at each computation step – by the current state of the system
(given by the number of molecules of each chemical) and by the value of all reaction constants. SSA and
tau leaping share the characteristic that repeated (independent) executions will produce different tempo-
D. Besozzi et al.
53
ral dynamics, even starting from the same initial configuration, thus reflecting the inherent noise of the
system. The two algorithms differ with respect to the way reactions are applied. In SSA, only one reaction can be applied at each step; the reaction that will be actually simulated, and the waiting time before
the next reaction will take place, depend on two independent random numbers drawn from the uniform
unit interval [0,1]. In tau leaping, instead, several reactions can be chosen and executed simultaneously,
by the sampling of Poissonian distributions and by choosing an opportune time increment (we refer to
[4] for further details). So doing, the computational burden typical of SSA for large systems consisting of many reactions and many molecular species, can be largely reduced. Indeed, with tau leaping it is
possible to handle detailed descriptions of many molecular interactions and chemical modifications (e.g.,
phosphorylation or methylation states of proteins), providing fast and reliable stochastic simulations of
mechanistic models of complex biological systems [6]. On the other side, tau leaping is not guaranteed
to reproduce the exact behavior of the system, but the accuracy of the simulation can be controlled.
Here we report a sketch of the functioning of tau leaping. We denote by X a well stirred system in
thermal equilibrium, consisting of N molecular species S1 , . . . , SN , which can interact through M chemical
reactions r1 , . . . , rM . Let Xi (t) be the number of molecules of chemical Si at time t, and x = X(t) ≡
(X1 (t), . . . , XN (t)) the state of the system at time t. The aim of the procedure is to fire several reactions
for each time interval [t,t + τ ). In order to find out which reactions will be executed, we have to calculate
the probability that a reaction r j will occur in the next infinitesimal time interval [t,t + dt), starting from
the system state x. This probability is given by a j (x)dt, which is the propensity function of reaction r j
and is defined as a j (x) = h j (x)·c j , where h j (x) is the number of distinct reactant molecules combinations
in r j , and c j is the stochastic constant associated to r j .
Given a state x of the system X , we denote by K j (τ , x,t) the exact number of times that a reaction
r j will be fired in the time interval [t,t + τ ), so that K(τ , x,t) is the exact probability distribution vector
(having K j (τ , x,t) as elements). For arbitrary values of τ , the computation of the values of K j (τ , x,t) is as
difficult as resolving the corresponding Chemical Master Equation for that system. On the contrary, if τ is
small enough so that the change in the state x during [t,t + τ ) is so slight that no propensity function will
suffer an appreciable change in its value (this is called the leap condition), then it is possible to evaluate a
good approximation of K j (τ , x,t) by using the Poisson random variables with mean and variance a j (x)· τ .
Hence, starting from the state x and choosing a τ value that satisfies the leap condition, we can update the
state of the system at time t + τ according to X(t + τ ) = x + ∑ j=i,...,M v j Pj (a j (x), τ ), where Pj (a j (x), τ )
denotes an independent sample of the Poisson random variable with mean and variance a j (x) · τ , and
v j ≡ (v1 j , . . . , vN j ) is the state change vector whose element vi j represents the stoichiometric change
of the species Si due to reaction r j . Summarizing, each iterative step of the algorithm consists of four
stages: (1) generate the maximum changes of each species that satisfy the leap condition; (2) compute
the mean and variance of the changes of the propensity functions; (3) evaluate the leap value τ exploiting
the auxiliary quantities previously computed; (4) toss the reactions to apply; (5) update the system state
(see [4] for further details).
The accuracy of this algorithm can be fixed a priori by means of an error control parameter ε (0 <
ε ≤ 1); for the following simulations, we have used the value ε = 0.03 as suggested in [4]. The algorithm
has been implemented in our laboratory [5]; the program is written in C language, compiled under Linux
using the GCC compiler. All the simulations have been performed using a Personal Computer with an
Intel Core2 CPU (2.66 GHz) running Linux. The mean duration time for one run, for the simulation of
the dynamics of CheYp over 3000 sec, is about 4-5 seconds (with the initial values of chemical amounts
given in Section 2.2 and the constants reported in Table 1). All the figures reported in Section 3.2, unless
otherwise stated, represent the mean value over 50 independent runs of tau leaping, each one executed
with the same initial conditions.
Interplay of stochastic fluctuations and number of flagella in bacterial chemotaxis
54
3.2 Results of simulations for the dynamics of CheYp
The dynamics of CheYp has been analyzed by considering various conditions, such as the addition and
removal of different ligand amounts, distinct methylation states of MCPs and deletion of other chemotactic proteins.
We start by reporting in Figure 2, left side, the response of the system to the addition of two consecutive amounts of external ligand: the first stimulus corresponds to a ligand amount of 100 molecules,
added to the system at time t = 3000 sec and removed at time t = 6000 sec, while the second stimulus corresponds to a ligand amount of 500 molecules, added at time t = 9000 sec and removed at time
t = 12000 sec. Note that, since the amount of CheYp is equal to 0 at the beginning of the simulation,
its dynamics shows a marked increase which then reaches – due to the counteraction of CheZ, CheR and
CheB – a steady state level. Starting from this level, the addition of the ligands has been simulated by
changing its amount from 0 to 100 (500, respectively) molecules, thus mimicking the environmental situation where the bacterium encounters a different concentration of attractant molecules. Vice versa, the
removal of the ligands has been simulated by putting the value of the ligand back to 0. In the time interval
between the addition and the removal of each ligand stimulus, the amount of ligand molecules has been
kept at the constant value of 100 and 500, respectively, thus mimicking the presence of an environmental
homogenous concentration. This has been done in order to test the adaptation capabilities of the system.
In both cases, we can see that the system is able to respond to a step-increase of the ligands by achieving
a sharp and fast decrease in CheYp (that is, the negative peaks at time instants t = 3000 and t = 9000
sec). Immediately after this transient, the amount of CheYp returns to a steady state value, which differs
from the prestimulus level only for a few tens of molecules, at most, according to the amount of added
ligand. In this phase, the bacterium is returning to the prestimulus switching and thus to the random walk
sampling of its surroundings. When the ligand is removed, CheYp shows another transient behavior, corresponding to a sharp and fast increase of its amount, that is in line with experimental observations (see
[2, 19]). After this second transient, the amount of CheYp correctly returns to the prestimulus steady
state level.
2000
1550
1500
1800
1450
1400
CheYp Molecules
CheYp Molecules
1600
1400
1350
1300
1250
1200
1200
1150
1000
50 ligands
100 ligands
500 ligands
1000
1100
800
1050
0
2000
4000
6000
8000
Time [sec]
10000
12000
14000
0
100
200
300
400
500
Time [sec]
Figure 2: Dynamics of CheYp. Left: adaptation response to two consecutive stimuli. Right: comparison
of transient and steady state response to different ligand amounts.
In Figure 2, right side, we compare the transients and steady states reached by CheYp after the
addition of distinct ligand amounts. This figure shows that the response magnitude at steady state and
the adaptation time of CheYp is only slightly sensitive to the ligand amount, being the relative differences
less than a few tens of molecules and less than a few seconds, respectively. The mean values of the steady
state of CheYp before the stimulus (SS1), after the ligand addition (SS2) and after the ligand removal
(SS3) are reported in Table 2, together with the values of its minimal and maximal values immediately
D. Besozzi et al.
55
Table 2: Steady state values and minimal/maximal transient values of CheYp after addition and removal
of distinct ligand amounts.
Ligand amount
50 molecules
100 molecules
500 molecules
1000 molecules
SS1
1486.7
1486.7
1486.7
1486.7
Min
1245.4
1160.7
1078.4
1058.6
SS2
1500.9
1495.1
1481.4
1478.2
Max
1626.0
1645.4
1653.2
1665.8
SS3
1474.7
1474.3
1469.4
1474.7
after the ligand addition and removal (Min and Max, respectively), for the four ligand amounts (50, 100,
500, 1000 molecules) considered in the right side of Figure 2.
In Figure 3 we show how the dynamics of CheYp changes when CheB is deleted from the system
at time t = 3000 sec, in both conditions of absence of external ligands (left side) and of presence of 100
molecules of ligand (right side) added at time t = 3000 sec. CheB is the methylesterase that, once being
phosphorylated by CheA, increases the methylation state of MCPs, thus keeping CheA more active.
This, in turn, causes an increase in the amount of CheYp, which is evident from its new steady state level
reached after CheB deletion, and also from its less negative transient decrease after ligand addition.
3000
2400
2800
2200
2600
2000
CheYp Molecules
CheYp Molecules
2400
2200
2000
1800
1600
1800
1600
1400
1400
1200
1200
∆CheB
Wild type
∆CheB
Wild type
1000
1000
0
1000
2000
3000
Time [sec]
4000
5000
6000
0
1000
2000
3000
Time [sec]
4000
5000
6000
Figure 3: Comparison of dynamics of CheYp in normal condition and after deletion of CheB at t = 3000
sec, without ligand (left) and with simultaneous addition of 100 ligand molecules (right).
2000
18000
∆CheR
Wild type
1800
∆CheZ
Wild type
16000
1600
14000
1400
CheYp Molecules
CheYp Molecules
12000
1200
1000
800
10000
8000
6000
600
4000
400
2000
200
0
0
0
1000
2000
3000
Time [sec]
4000
5000
6000
0
1000
2000
3000
Time [sec]
4000
5000
6000
Figure 4: Comparison of dynamics of CheYp in normal condition and after deletion of CheR (left) and
CheZ (right) at t = 3000 sec, both simulated with a simultaneous addition of 100 ligand molecules.
Similarly, in Figure 4 we show the dynamics of CheYp when either CheR (left side) or CheZ (right
Interplay of stochastic fluctuations and number of flagella in bacterial chemotaxis
1800
1200
1600
1000
1400
800
CheYp Molecules
CheYp Molecules
56
1200
600
1000
400
800
200
600
0
0
1000
2000
3000
Time [sec]
4000
5000
6000
0
1000
2000
3000
Time [sec]
4000
5000
6000
Figure 5: Dynamics of CheYp when only 3 (left) and 2 (right) methylation states are active.
side) are deleted from the system at time t = 3000 sec, simultaneously to the addition of 100 ligand
molecules (the temporal evolution of CheYp when no ligand is added is basically equivalent). When
CheR is deleted, its methyltransferase activity is silenced, the MCPs are no more methylated, and hence
the amount of CheYp tends to zero. On the contrary, when CheZ is deleted, all CheY molecules always
remain phosphorylated. For the sake of completeness, we have also simulated the dynamics of CheYp
when either CheB, CheR or CheZ are deleted from the system since the time instant t = 0, in order to
have a comparison about the initial temporal evolution of CheYp and the steady state levels it can reach.
In these conditions, the model correctly simulates [1, 12, 20] a very low production of CheYp when
CheR is deleted, and an increased production (albeit with different magnitudes) when either CheB or
CheZ are deleted (data not shown).
Finally, in Figure 5 we compare the dynamics of CheYp in response to the addition of 100 ligand
molecules at t = 3000 sec, when only 3 (left side) or 2 (right side) methylation states of the receptors are
allowed. In practice, this is achieved by initially putting to zero the values of the stochastic constants of
methylation and demethylation reactions for levels m = 4 and m = 3, respectively. In both cases, we see
that the system is not able to adapt, as the steady state level of CheYp reached after the addition of the
ligand is substantially lower than the steady state when all methylation levels are activated.
The outcome of the stochastic simulations reported in this section validates the model presented
in Section 2.2, as the dynamics of CheYp under different conditions of the chemotactic pathway well
compare with experimental evidences.
4
The interplay between stochastic fluctuations and the number of bacterial flagella
In this section we make use of the simulations based on our model of chemotaxis to analyze the interplay
between stochastic fluctuations of CheYp and the number of flagella occurring on the cell, in order
to outline the influence of synchronization of flagellar motors on the swimming behavior, and on the
adaptation mechanism of the bacterium to the environmental changes. To this aim, we consider the
dynamics of CheYp at steady state, as well as its transient step-decrease that takes place immediately
after the chemotactic stimulus. In both cases, we are interested in devising the time periods during which
the cell performs either a running or a tumbling motion. In particular we will assume that: (1) the
time spent in alternating CW and CCW rotations during the steady state corresponds to the random walk
sampling of the environment – where we expect more time spent in tumbling than in running motions; (2)
D. Besozzi et al.
57
the time required to return to the prestimulus level of CheYp (that is, the transient response immediately
after the ligand addition) corresponds to the chemotactic adaptation time – where we expect a much
longer and uninterrupted time interval of running motion with respect to the steady state condition.
As explained in Section 2.1, a running motion requires that all flagella are simultaneously synchronized in a CCW rotation – which occurs when CheYp is not interacting with the proteins FliM of the
flagellar motors, that is, when its intracellular concentration diminishes with respect to a reference value.
To distinguish between the CW and CCW rotations of a single flagellum, we assume that the flagellar
motor switch is sensitive to a threshold level of CheYp, that is hereby evaluated as the mean value of
CheYp at steady state (see also [17], where a similar approach of threshold-crossing mechanism for motor switching was tested, albeit that work considered only a single flagellum and did not propose any
investigation on the simultaneous coordination of many flagella). When the amount of CheYp is below
this threshold, each flagellum is rotating CCW, while when the amount of CheYp is above the threshold, each flagellum is rotating CW. In what follows, we make a one-to-one correspondence between
the behavior of a single flagellum and a temporal evolution of CheYp generated by one run of the tau
leaping algorithm, that is, we consider a different and independent stochastic simulation for each and
every flagellum (albeit starting from the same initial conditions for the whole system). In other words,
we assume that flagella are independent one another – as no molecular interactions between them have
been evidenced in bacterial cells. Nonetheless, they all overlook on the same intracellular ambient, that
is, they are all subject to the same temporal evolution of CheYp, apart from the stochastic noise existing
among independent simulations. In order to determine the synchronization of all flagella that will induce
a running motion of the bacterium, we therefore need to identify the time instants when all flagella are
rotating CCW, that is, to select the time intervals when all the temporal evolutions of CheYp are below
the fixed threshold.
Formally, we proceed as follows. Let n = 1, . . . , 10 be the number of flagella f1 , . . . , fn whose influence we want to test, and let si , i = 1, . . . , n, be the time series of CheYp (generated by a single tau
leaping run) associated to each fi . For any fixed value of n, the total time of the simulation considered
to generate the dynamics of CheYp is the same for all si . This simulation time, hereby denoted by ∆tsim ,
is chosen long enough to have a meaningful evaluation of the mean intervals of running and tumbling in
the analysis performed below (e.g., ∆tsim = 40000, 60000, 120000 sec for n = 1, 5, 10, respectively). The
threshold for CheYp is evaluated in the following way: we choose an initial time instant at the steady
state level – distant enough from the step decrease of CheYp after ligand addition, i.e. 1000 sec afterward
– and then, starting from this instant and till the end of ∆tsim , we calculate the mean value µi =< si > for
each si . Then, we define a common threshold µ for all flagella, that is, µ = n1 ∑i=1,...,n µi . This threshold
is then considered as the reference value also for the portion of the CheYp dynamics corresponding to
the transient decrease after ligand addition. In Figure 6, top panel on the left side, we show a part of ∆tsim
over a single simulation of CheYp, where both the initial transient response and the stochastic fluctuations around the threshold are evidenced. For all the results discussed below, the different values of µ
have been found to be approximatively equal to 1480 molecules.
The next step consists in detecting, for each fi , the time intervals during which the amount of CheYp
remains below µ , each one of these intervals corresponding to a CCW rotation time interval of that flagellum. Namely, for each si we identify the time intervals ∆ttrue ⊆ ∆tsim such that ∆ttrue = {t ∈ ∆tsim |
si (t) − µ ≤ 0}. Note that this simple mechanism of single threshold-crossing could be extended to consider more complex situations – e.g., a double threshold-crossing mode can be assumed – whereby one
simply asks for analogous conditions to be satisfied. Similarly, for each si we can locate the complementary time intervals ∆t f alse ⊆ ∆tsim such that ∆t f alse = {t ∈ ∆tsim | si (t) − µ > 0}; these intervals correspond
to the time that each flagellum fi spends in a CW rotation. Stated in other terms, we can associate to
58
Interplay of stochastic fluctuations and number of flagella in bacterial chemotaxis
Figure 6: Threshold-crossing intervals in stochastic fluctuations of CheYp (left) and synchronization of
running motion between 2 flagella (right).
each si a function CCWsi : ∆tsim → {true, f alse} defined as:
true if si (t) − µ ≤ 0
CCWsi (t) =
f alse otherwise
In Figure 6, bottom panel on the left side, we show the values of this function for the CheYp simulation given in the upper panel. As it can be seen at a glance, the transient response after ligand addition
(when the amount of CheYp is initially below µ ) corresponds to a longer and uninterrupted interval of
CCW rotation of that flagellum.
Once that the set of all ∆ttrue intervals – or, equivalently, of all functions CCWsi – have been found out
for each flagellum, we start the process of synchronization for any given number n of flagella. To this aim,
n
n = {t ∈ ∆t
let us define Tsync
sim | CCWsi (t) = true for all i = 1, . . . , n}. Tsync is the set of all times during
which all time series si are below the threshold µ , that is, the time intervals during which all flagella are
rotating CCW. More precisely, we identify these intervals as the running motion of the bacterium, i.e.
n corresponds to the time of directional swimming – when all flagella are coordinated in a bundle.
Tsync
As an example, in Figure 6, right side, we represent the functions CCWsi (t) = true for i = 1, 2, and the
n , n = 2. The complementary set, T n
n
corresponding set Tsync
unsync = ∆tsim \ Tsync , corresponds instead to
tumbling motion – when at least one flagellum (over the set of n flagella considered time by time) is
n
rotating CW. Namely, Tunsync
= {t ∈ ∆tsim | there exists i = 1, . . . , n such that CCWsi (t) = f alse}.
n are influWe are now interested in understanding if and how the time intervals within the set Tsync
enced by the increase of n. We have performed this analysis over a set of 10 distinct in silico experiments
(each one corresponding to a cell with n flagella, with n = 1, . . . , 10), and then we have evaluated the
mean values of the following three parameters:
1. the time intervals corresponding to a running motion of the bacterium, h∆trun i, when all flagella
are rotating CCW (that is, when all time series si are below µ );
2. the time intervals corresponding to a tumbling motion of the bacterium, h∆ttumb i, when at least one
flagellum over the n flagella is rotating CW (that is, when at least one time series si is above µ );
3. the time intervals corresponding to the transient decrease of CheYp after ligand addition, h∆tadapt i,
that is, the adaptation time during which the bacterium is performing a longer running motion.
The results for h∆trun i are reported in Figure 7, left side, where we can see that the mean time
intervals of running motion are very short, and their values decrease in a (qualitative) exponential way as
D. Besozzi et al.
59
3.5
300
< ∆ trun >
3
105
< ∆ ttumb >
< ∆ tadapt >
100
250
2.5
95
1.5
Time [sec]
Time [sec]
Time [sec]
200
2
150
90
85
100
1
80
50
0.5
0
75
0
0
2
4
6
8
10
70
0
2
4
Flagella
6
8
10
0
2
Flagella
4
6
8
10
Flagella
Figure 7: Variation of mean time values of running (left) and tumbling motions (middle), and of adaptation time (right), with respect to the number of flagella.
Table 3: Values of mean time intervals for running, tumbling and adaptation.
n
1
5
10
h∆trun i (sec)
3.102
0.606
0.310
h∆ttumb i (sec)
3.062
18.73
297.4
h∆trun i/h∆ttumb i
1.013
0.032
0.001
h∆tadapt i (sec)
104.0
73.48
72.22
the number n of flagella increases, as expected. Similarly, the results for h∆ttumb i evidence a (qualitative)
exponential increase with respect to n, as reported in Figure 7, middle part. As reference, the precise
values of the mean running and tumbling time intervals are given in Table 3, together with their ratio, for
three values of n. The running-to-tumbling ratio, which decreases as n increases, highlights the relevance
of the number of flagella and the necessity of their synchronization with respect to the chemotactic
behavior of the bacterium. That is, we see that for n = 1 the time spent in running or tumbling motions
is approximatively equivalent, but if coordination among many flagella (n = 10) has to take place, then
the running motions are highly reduced with respect to tumbling motions, which is in agreement with
biological expectations.
The results for h∆tadapt i are reported in Figure 7, right side, and in Table 3. In this case, it is not
possible to recognize a simple function for the curve progress, and we see that the variation of the time
intervals is within a range of a few tens of seconds. Once more, this result seems to be in agreement with
biological expectations, as the response of the bacterium to an environmental change (i.e. the addition
or removal of ligands) should not be strictly dependent on the number of flagella that are present on its
surface, otherwise the chemotactic pathway would not guarantee an appropriate adaptation mechanism,
independently from the variation of the number of flagella among distinct individuals in a population of
cells.
5
Discussion
In this paper we have investigated the possible influence of stochastic fluctuations of the chemotactic
protein CheYp on the running motion of bacterial cells, with respect to an increasing number of flagella
in the individual bacterium. To this aim, we have defined a procedure to identify the synchronization
of CCW rotations of each and every flagella, and then we have compared the mean time intervals of
running and tumbling motions of the cell, as well as of adaptation times to ligand addition, according
to the different numbers of flagella. We have shown that, on the one hand, the running-to-tumbling ra-
60
Interplay of stochastic fluctuations and number of flagella in bacterial chemotaxis
tio highlights the relevance of the number of flagella, and the necessity of their synchronization with
respect to the chemotactic behavior of the bacterium. On the other hand, the adaptation time does not
seem to be strongly influenced by the varying number of flagella in distinct individual cells. These results have been obtained by performing stochastic simulations of a very detailed mechanistic model of
the bacterial chemotaxis pathway, that takes into account all proteins, and their respective interactions,
involved in both signaling and response. All post-translational modifications of proteins, such as methylation and phosphorylation, have been considered because of their relevant roles in the feedback control
mechanisms governing this pathway. In particular, by exploiting the efficiency of tau leaping algorithm,
we have investigated the dynamics of the pivotal protein involved in chemotaxis, CheYp, under different
conditions, such as the deletion of other chemotactic proteins, the addition of distinct amounts of external
ligand, the effect of different methylation states of the receptors.
Concerning the analysis of the interplay between CheYp fluctuations and the number of flagella,
other relevant biological aspects of chemotaxis, that stand downstream of the signaling process, might
represent valuable points to be considered for future research. For instance, it is known that each flagellar
motor switch-complex is constituted by two principal components: a group of proteins, called FliM, FliN,
FliG (assembled in a ring), and the torque-generating complexes, called MotA and MotB. In E. coli, a
typical flagellar ring contains 34 copies of FliM, each of which can bind one copy of CheYp. In [8] it
is suggested that binding of CheYp to FliM modifies the displacement of protein FliG, which directly
interacts with the Mot complexes and therefore modulates the switch state of the flagellum. Moreover,
flagellar motor switching has been found to be highly sensitive to the concentration of CheYp (having
a Hill coefficient ≈ 10), though the binding of CheYp to FliM has a low level of cooperativity (Hill
coefficient ≈ 1). In [8], the hypothesis that CheYp can interact more favorably with the FliM displaced in
the CW orientation, than those in the CCW orientation, is put forward. In [20], in addition, the following
mechanism is considered for the control of flagellar motor by means of CheYp: the number of CheYp
molecules bound to FliM determines the probability of CW or CCW rotation, while the switch is thrown
by thermal fluctuations. In other words, CheYp only changes the stabilities of the two rotational states,
by shifting the energy level of CCW-state up and of CW-state down, by a magnitude that is directly
proportional to the number of bound molecules. Therefore, interesting questions related to stochastic
fluctuations of CheYp, that might be coupled with the investigation on the number of cellular flagella,
are: how many FliM proteins at each flagellar switch have to be occupied by CheYp in order to generate
the CW rotation of each flagellum and of the bacterium [3]? What is the corresponding probability
of throwing the reversal switch? Can a double-threshold crossing mechanism [3] be more suitable to
effectively detect the CW and CCW rotational states?
Nonetheless, other related features might be relevant points for a further extension of this work. For
instance, the gradient of CheYp that can be present inside the cytoplasm – due to the diffusion from the
area of its phosphorylation (close to chemotactic receptors) to the area of its activity (around the flagellar
motors) – can be a significant aspect in chemotaxis, together with the localization of CheZ (that controls
the dephosphorylation of CheYp) and of the flagella around the cell [15]. With respect to this matter,
how are diffusion processes related to the interactions between CheYp and the flagellar motors? And
how does diffusion intervene on the chemotactic response?
We believe that the definition of detailed mechanistic models, like the one proposed in this paper
for chemotaxis, coupled to the use of efficient procedures for the analysis of stochastic processes in
individual cells, can be a good benchmark to investigate the combined roles of many biological factors
interplaying within a common system. With this perspective, the development of formal methods specifically devised for the analysis of properties (e.g., synchronization) of stochastic systems represents indeed
a hot topic research in biological modeling.
D. Besozzi et al.
61
References
[1] U. Alon, M. G. Surette, N. Barkai & S. Leibler (1999): Robustness in bacterial chemotaxis. Nature
397(6715), pp. 168–171. Available at http://dx.doi.org/10.1038/16483.
[2] N. Barkai & S. Leibler (1997): Robustness in simple biochemical networks. Nature 387(6636), pp. 913–917.
Available at http://dx.doi.org/10.1038/43199.
[3] A. Bren & M. Eisenbach (2001): Changing the direction of flagellar rotation in bacteria by modulating the
ratio between the rotational states of the switch protein FliM. J. Mol. Biol. 312(4), pp. 699–709. Available
at http://dx.doi.org/10.1006/jmbi.2001.4992.
[4] Y. Cao, D. T. Gillespie & L. R. Petzold (2006): Efficient step size selection for the tau-leaping simulation
method. J. Chem. Phys. 124(4), p. 044109. Available at http://dx.doi.org/10.1063/1.2159468.
[5] P. Cazzaniga, D. Pescini, D. Besozzi & G. Mauri (2006): Tau leaping stochastic simulation method in P
systems. In: H. J. Hoogeboom, G. Păun, G. Rozenberg & A. Salomaa, editors: 7th International Workshop
on Membrane Computing WMC 2006, 4361. LNCS, pp. 298–313. Available at http://dx.doi.org/10.
1007/11963516_19.
[6] P. Cazzaniga, D. Pescini, D. Besozzi, G. Mauri, S. Colombo & E. Martegani (2008): Modeling and stochastic
simulation of the Ras/cAMP/PKA pathway in the yeast Saccharomyces cerevisiae evidences a key regulatory function for intracellular guanine nucleotides pools. J. Biotechnol. 133(3), pp. 377–385. Available at
http://dx.doi.org/10.1016/j.jbiotec.2007.09.019.
[7] M. Dugo (2009): Modellazione stocastica della chemiotassi batterica. Bachelor thesis, Università degli Studi
di Milano.
[8] C. M. Dyer, A. S. Vartanian, H. Zhou & F. W. Dahlquist (2009): A Molecular Mechanism of Bacterial
Flagellar Motor Switching. J. Mol. Biol. 388, pp. 71–84. Available at http://dx.doi.org/10.1016/j.
jmb.2009.02.004.
[9] M. B. Elowitz, A. J. Levine, E. D. Siggia & P. S. Swain (2002): Stochastic gene expression in a single cell.
Science 297, pp. 1183–1186. Available at http://dx.doi.org/10.1126/science.1070919.
[10] D. T. Gillespie (1977): Exact stochastic simulation of coupled chemical reactions. Journ. Phys. Chem.
81(25), pp. 2340–2361. Available at http://dx.doi.org/10.1021/j100540a008.
[11] M. S. Jurica & B. L. Stoddard (1998): Mind your B’s and R’s: bacterial chemotaxis, signal transduction and
protein recognition. Structure 6(7), pp. 809–813. Available at http://www.ncbi.nlm.nih.gov/pubmed/
9687374.
[12] C. Kim, M. Jackson, R. Lux & S. Khan (2001): Determinants of chemotactic signal amplification in Escherichia coli. J. Mol. Biol. 307, pp. 119–135. Available at http://dx.doi.org/10.1006/jmbi.2000.
4389.
[13] M. D. Levin, C. J. Morton-Firth, W. N. Abouhamad, R. B. Bourret & D. Bray (1998): Origins of individual
swimming behavior in bacteria. Biophys. J. 74(1), pp. 175–181. Available at http://dx.doi.org/10.
1016/S0006-3495(98)77777-X.
[14] G. Li & R. M. Weis (2000): Covalent modification regulates ligand binding to receptor complexes in the
chemosensory system of Escherichia coli. Cell 100(3), pp. 357–365. Available at http://view.ncbi.nlm.
nih.gov/pubmed/10676817.
[15] K. Lipkow, S. S. Andrews & D. Bray (2005): Simulated diffusion of phosphorylated CheY through the
cytoplasm of Escherichia coli. J. Bacteriol. 187(1), pp. 45–53. Available at http://dx.doi.org/10.
1128/JB.187.1.45-53.2005.
[16] B. A. Mello & Y. Tu (2003): Perfect and near-perfect adaptation in a model of bacterial chemotaxis. Biophys.
J. 84(5), pp. 2943–2956. Available at http://dx.doi.org/10.1016/S0006-3495(03)70021-6.
[17] C. J. Morton-Firth & D. Bray (1998): Predicting temporal fluctuations in an intracellular signalling pathway.
J. Theor. Biol. 192(1), pp. 117–128. Available at http://dx.doi.org/10.1006/jtbi.1997.0651.
[18] C. J. Morton-Firth, T. S. Shimizu & D. Bray (1999): A free-energy-based stochastic simulation of the Tar
receptor complex. J. Mol. Biol. 286(4), pp. 1059–1074. Available at http://dx.doi.org/10.1006/jmbi.
62
Interplay of stochastic fluctuations and number of flagella in bacterial chemotaxis
1999.2535.
[19] C. V. Rao, J. R. Kirby & A. P. Arkin (2004): Design and diversity in bacterial chemotaxis: a comparative
study in Escherichia coli and Bacillus subtilis. PLoS Biol. 2(2), p. E49. Available at http://dx.doi.org/
10.1371/journal.pbio.0020049.
[20] B. E. Scharf, K. A. Fahrner, L. Turner & H. C. Berg (1998): Control of direction of flagellar rotation in
bacterial chemotaxis. Proc. Natl. Acad. Sci. U S A 95(1), pp. 201–206. Available at http://www.pnas.
org/content/95/1/201.abstract.
[21] T. S. Shimizu & D. Bray (2002): Modelling the bacterial chemotaxis receptor complex. Novartis Found.
Symp. 247, pp. 162–77; discussion 177–81, 198–206, 244–52. Available at http://www.pdn.cam.ac.uk/
groups/comp-cell/Papers/Shimizu02b.pdf.
[22] V. Sourjik (2004): Receptor clustering and signal processing in E. coli chemotaxis. Trends Microbiol. 12(12),
pp. 569–576. Available at http://dx.doi.org/10.1016/j.tim.2004.10.003.
[23] P. A. Spiro, J. S. Parkinson & H. G. Othmer (1997): A model of excitation and adaptation in bacterial
chemotaxis. Proc. Natl. Acad. Sci. U S A 94(14), pp. 7263–7268. Available at http://www.pnas.org/
content/94/14/7263.abstract.
[24] T. E. Turner, S. Schnell & K. Burrage (2004): Stochastic approaches for modelling in vivo reactions.
Computational Biology and Chemistry 28, pp. 165–178. Available at http://dx.doi.org/10.1016/j.
compbiolchem.2004.05.001.
[25] G. H. Wadhams & J. P. Armitage (2004): Making sense of it all: bacterial chemotaxis. Nat. Rev. Mol. Cell.
Biol. 5(12), pp. 1024–1037. Available at http://dx.doi.org/10.1038/nrm1524.
| 5cs.CE
|
Empirical Evaluation of Abstract Argumentation: Supporting the
Need for Bipolar and Probabilistic Approaches
Sylwia Polberg and Anthony Hunter∗
arXiv:1707.09324v2 [cs.AI] 8 Dec 2017
University College London, Department of Computer Science
66-72 Gower Street, London WC1E, United Kingdom
Abstract
In dialogical argumentation, it is often assumed that the involved parties will always correctly identify the intended statements posited by each other and realize all of the associated relations, conform to the three acceptability
states (accepted, rejected, undecided), adjust their views whenever new and correct information comes in, and that a
framework handling only attack relations is sufficient to represent their opinions. Although it is natural to make these
assumptions as a starting point for further research, dropping some of them has become quite challenging.
Probabilistic argumentation is one of the approaches that can be harnessed for more accurate user modelling. The
epistemic approach allows us to represent how much a given argument is believed or disbelieved by a given person,
offering us the possibility to express more than just three agreement states. It comes equipped with a wide range
of postulates, including those that do not make any restrictions concerning how initial arguments should be viewed.
Thus, this approach is potentially more suitable for handling beliefs of the people that have not fully disclosed their
opinions or counterarguments with respect to standard Dung’s semantics. The constellation approach can be used
to represent the views of different people concerning the structure of the framework we are dealing with, including
situations in which not all relations are acknowledged or when they are seen differently than intended. Finally, bipolar
argumentation frameworks can be used to express both positive and negative relations between arguments.
In this paper we will describe the results of an experiment in which participants were asked to judge dialogues in
terms of agreement and structure. We will compare our findings with the aforementioned assumptions as well as with
the constellation and epistemic approaches to probabilistic argumentation and bipolar argumentation.
Keywords— Dialogical argumentation, probabilistic argumentation, abstract argumentation
∗ This research is funded by EPSRC Project EP/N008294/1 “Framework for Computational Persuasion”.We thank the reviewers for their valuable
comments that helped us to improve this paper.
i
Contents
1
Introduction
3
2
Dung’s Argumentation Framework
5
3
Bipolar Argumentation
7
4
Probabilistic Argumentation
4.1 Epistemic Probabilistic Argumentation . . . .
4.1.1 Preferential Postulates . . . . . . . .
4.1.2 Explanatory Postulates . . . . . . . .
4.1.3 Value Postulates . . . . . . . . . . .
4.1.4 Multi–type Postulates . . . . . . . .
4.1.5 Properties of Epistemic Postulates . .
4.1.6 Relation to Labeling–based Semantics
4.2 Constellation Probabilistic Argumentation . .
.
.
.
.
.
.
.
.
9
9
10
10
11
11
12
12
14
5
Experiment Description
5.1 Dialogues & Tasks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.2 Recruitment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
14
15
16
6
Results
6.1 Argument Graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.1.1 Dialogue 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.1.1.1 Intended Graph Analysis . . . . . . . . . . . . . . .
6.1.1.2 Participant–Sourced Graphs Analysis . . . . . . . . .
6.1.2 Dialogue 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.1.2.1 Intended Graph Analysis . . . . . . . . . . . . . . .
6.1.2.2 Participant–Sourced Graphs Analysis . . . . . . . . .
6.2 Postulate Satisfaction . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.2.1 Dialogue 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.2.2 Dialogue 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.3 Statements vs Relations . . . . . . . . . . . . . . . . . . . . . . . . . .
6.3.1 Dialogue 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.3.1.1 Effect of a Relation on its Source . . . . . . . . . . .
6.3.1.2 Effect of an Argument on the Relations it Carries Out
6.3.2 Dialogue 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.3.2.1 Effect of a Relation on its Source . . . . . . . . . . .
6.3.2.2 Effect of an Argument on the Relations it Carries Out
6.4 Changes in Beliefs . . . . . . . . . . . . . . . . . . . . . . . . . . . .
17
19
20
20
22
28
28
29
36
36
41
45
45
45
49
50
50
54
57
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
7
Related Work
59
8
Conclusions
61
9
Statistics Appendix
9.1 Dialogue 1 . . . . . . . . . . . .
9.1.1 Postulate Satisfaction . .
9.1.2 Statements vs. Relations:
9.1.3 Statements vs. Relations:
9.2 Dialogue 2 . . . . . . . . . . . .
9.2.1 Postulate Satisfaction . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Effect of a Relation on its Source . . . . . . . . . .
Effect of an Argument on the Relations it Carries Out
. . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
9.3
9.2.2 Statements vs. Relations: Effect of a Relation on its Source . . . . . . . . . . . . . . . . . .
9.2.3 Statements vs. Relations: Effect of an Argument on the Relations it Carries Out . . . . . . . .
Dialogues 1& 2: Changes in Beliefs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2
1
Introduction
At the heart of abstract argumentation lies Dung’s framework developed in [27], which treats arguments as abstract
atomic entities that can be connected through an attack relation. Since its introduction, this framework has been endowed with numerous new semantics as well as generalized in various ways. These generalizations include structures
that can represent new types of relations, such as support, as well as those that handle properties such as preferences or
probabilities of arguments or relations [9]. Despite the available approaches, dialogical argumentation appears to rely
heavily on frameworks handling only attacks between arguments, such as Dung’s framework. It also makes certain
assumptions that are supported by the argumentation theory, but not always by empirical results. Argumentation, in
many ways, simplifies human reasoning and appears to consider people to be in principle rational, good reasoners, that
may be simply uninformed. Unfortunately, this approach might not always be adequate.
One of the core concepts of defeasible reasoning and therefore abstract argumentation is the fallibility of human
perception. Thus, we need to be able to reason even with incomplete information and be prepared to retract our
conclusions in the face of new data. From a certain perspective, most of the abstract argumentation approaches can
be quite conservative in their handling of these issues. Although the defeasibility of arguments and notions such as
undercutting attack are widely acknowledged, at the same time there is an assumption that a universal attack relation
exists. In other words, it is believed that every person participating in a given dialogue will interpret arguments and the
relations between them in exactly the same way, and every time new information is presented, it will be understood
and linked to the existing arguments in the “intended and correct” way. However, this does not seem realistic in various
scenarios. For example, the transcript of a TV debate “as it is” can be perceived differently from its processed version,
i.e. one in which arguments are identified, organized and clearly presented w.r.t. the chosen argument structure. Not
all relevant pieces of information have to be explicitly stated and it can happen that people taking part in the debate
may be purposefully ambiguous. Consequently, we may have to deal with arguments that have implicit premises or
conclusions and thus run the risk of participants interpreting them differently. Therefore, we need to acknowledge
that people can view the nature of the framework associated with a given dialogue differently and that we should be
able to represent such differences and uncertainties. Furthermore, these issues can themselves become a part of the
discussion, which due to the aforementioned assumptions is rarely considered.
The fact that dialogical argumentation relies so heavily on attack–based frameworks for modelling purposes has
also affected the way we perceive dialogues on a higher level. This approach became rather heavily conflict–centered,
by which we understand that people participating in a dialogue view each other as opponents whose arguments should
be defeated and knowledge “rectified”. Although this view may very well work in a court room, it has negative effects
in the context of, for example, physical or mental health, as seen in [41]. For these applications, it may be more
fruitful to consider dialogue parties as partners rather than opponents and use a more support–oriented approach in the
discussion. Thus, rather than giving arguments against the current views of our dialogue partners, we try to motivate
them or give them arguments for changing their behaviour and opinions. A given piece of information can also be
framed in a positive or negative way, which affects the way people react to it [59, 57]. Hence, not only negative, but
also positive relations between arguments can play a role in dialogical argumentation, and the use of various types of
support should be considered.
Finally, the applicability of argumentation semantics in dialogical argumentation is not sufficiently verified [19].
The use of two or three values, as promoted by Dung’s semantics, oversimplifies the varying degrees to which we
may agree or disagree with a given argument. For example, while a mother of an 8 year old girl might disagree
with her staying the night at her best friend’s house, this is probably nowhere near to her disagreement with the child
going alone to a heavy metal concert. Thus, there is a need for expressing varying degrees of agreement with a given
argument. Another issue lies in the nature of the semantics themselves. For example, successful persuasion is often
seen as synonymous with winning a dialogue game w.r.t. a given semantics, such as grounded or admissible. However,
it is easy to see that in the cases such as a doctor persuading a patient to stop smoking or to go on a diet, holding a
dialectically winning position and actually convincing a person to do or not to do something can be two completely
different things. A person taking part in a dialogue might also exhibit a number of perception biases or reasoning
paradoxes and, due to inability or lack of cooperation, withhold information from the doctor. Hence, he or she does
not have to reason in a way that adheres to the current argumentation semantics [45, 44].
To summarize, there is a need for empirical evaluation of argumentation approaches, which so far have received
limited attention [55, 19, 56]. Although this task can be quite challenging, as what we are dealing with is human
3
perception and judgment which are not always perfect, it is necessary. One has to realize that due to the nature of
abstract argumentation, the validity of its approaches is typically inherited from other methods or obtained through
novelty and technical correctness. For example, in order to defend the introduction of a new framework, we can
explain how it can be instantiated with a given formalism, and show that the answers the framework produces exhibit
certain desirable properties connected to the formalism we instantiated it with. The validity of the new framework thus
depends on the validity of the formalisms it is linked to. Another possible way to argue in favour of a new approach
is through the abstract argumentation itself, where the new semantics or framework is shown to be considerably
different or more expressive than the existing ones. This method is often paired with presenting a realistic motivating
example that argues in favor of the new framework or semantics. Although both of these strategies promote new
ideas and innovation, there comes a time when they need to be verified in real life. Without empirical evidence, we
can accidentally increase a gap between applying argumentation and successfully applying argumentation in real life
situations.
In this paper we describe the results of an empirical study in which participants are presented with dialogues
separated into five stages. At every stage, they are asked to declare why and how much they agree or disagree with
the presented statements and how they view the relations between them. The purpose of this study was to investigate
certain aspects of abstract argumentation as used by laypeople, rather than to make a number of initial assumptions
that we prove or disprove based purely on the behaviour of experts. We have been able to draw a number of important
observations and gained some evidence concerning various formalisms available in abstract argumentation. However,
we would like to note that due to the exploratory nature of our experiments, our results should be treated as indicative
and as a basis for further studies, rather than as an indisputable proof for or against a given argumentation approach.
Observation 1 The data supports the use of the constellation approach to probabilistic argumentation – people
may interpret statements and relations between them differently, and not necessarily in the intended manner.
The constellation approach can represent our uncertainty about the argument graphs describing our opponents
views.
Observation 2 People may explicitly declare that two statements are connected, however, they might not be sure of
the exact nature of the relation between them. We therefore also need to express the uncertainty that a person
has about his or her own views, which can potentially be addressed with the constellation approach or with the
introduction of a suitable framework.
Observation 3 The data supports the use of epistemic approach to probabilistic argumentation:
– people may assign levels of agreement to statements going beyond the 3–valued Dung’s approach,
– the epistemic postulates, in contrast to the standard semantics, can be highly adhered to and due to their
nature, allow us to analyze why classical semantics may fail to explain the participants behaviour,
– the extended epistemic postulates allow us to model situations where the perceived “strength” of a relation
might not necessarily be tightly connected with the level of agreement assigned to its source.
Observation 4 The data supports the use of bipolar argumentation frameworks – the notion of defence does not
necessarily account for all of the positive relations between the statements viewed by the participants.
Observation 5 The data supports the use of bipolar argumentation in combination with the prudent/careful approaches – many additional attacks perceived by the participants can be explained by the existing notions of
indirect conflicts in these settings.
Observation 6 The data shows that people use their own personal knowledge in order to make judgments and might
not necessarily disclose it.
Observation 7 The data shows that presenting a new and correct piece of information that a given person was not
aware of does not necessarily lead to changing that person’s beliefs.
This paper is organized as follows. In Sections 2 to 4 we review the necessary background on Dung’s argumentation frameworks, bipolar argumentation frameworks and probabilistic argumentation. In Section 5 we explain the set
4
up of our experiment and analyze its results in Section 6. This includes the analysis of 1) the argument graphs created
for the experiment as well as those provided by the participants, 2) the satisfaction rates of various epistemic postulates
on the provided frameworks, 3) the connection between the level of agreement assigned to a given statement and the
way the relations it carries out are perceived, and 4) changes in participants’ opinions throughout the experiment. We
close this paper with the discussion on the works related to our study and pointers for future work.
2
Dung’s Argumentation Framework
We start with a review of abstract argumentation as proposed by Dung in [27]. He introduced the following, straightforward framework, that can be easily depicted using a graph where nodes play the role of arguments and edges
represent conflicts:
Definition 2.1. A Dung’s abstract argumentation framework (AF for short) is a pair F = (A, R), where A is a set
of arguments and R ⊆ A × A represents an attack relation.
An argument A ∈ A is an attacker of B ∈ A iff (A, B) ∈ R. By abuse of notation, we say that a set of elements
attacks another element if it contains an appropriate attacker. The way we decide which arguments can be accepted or
rejected (or neither) is called a semantics. Depending on whether a set of arguments or a labeling is returned, we deal
with the extension–based and labeling–based semantics [11, 2].
An extension is a set of arguments E ⊆ A that satisfies the requirements imposed by a given semantics. The
classical semantics [27] are built on the notion of defence:
Definition 2.2. Let F = (A, R) be a Dung’s framework. An argument A ∈ A is defended by a set E ⊆ A in F 1 if
for every B ∈ A s.t. B attacks A, there exists C ∈ E that attacks B. A set E ⊆ A is:
(cf) conflict–free in F iff for no A, B ∈ E, A is an attacker of B.
(ad) admissible in F iff it is conflict–free in F and defends all of its members.
(pr) preferred in F iff it is maximal w.r.t. set inclusion admissible in F .
(co) complete in F iff it is admissible in F and all arguments defended by E are contained in E.
(st) stable in F iff it is conflict–free in F and for every A ∈ A \ E there exists an argument B ∈ E that attacks A.
(gr) grounded in F iff it is the least complete extension of F .
A
B
C
D
E
Figure 1: Sample argument graph
Example 1. Consider the Dung’s framework F = (A, R) with A = {A, B, C, D, E} and the attack relation
R = {(A, B), (C, B), (C, D), (D, C), (D, E), (E, E)}, as depicted in Figure 1. It has eight conflict–free extensions
in total: {A, C},{A, D}, {B, D}, {A}, {B}, {C}, {D} and ∅. As B is attacked by an unattacked argument, it cannot
be defended against it. Consequently, {A, C},{A, D}, {A}, {C}, {D} and ∅ are our admissible sets. From this
{A, C}, {A, D} and {A} are complete. We end up with two preferred extensions, {A, C} and {A, D}. However,
only {A, D} is stable. Finally, {A} is the grounded extension.
More types of extension–based semantics have been proposed in the recent years [2] and even though we will not
consider all of them here, we would like to recall certain notions that will be useful in the next sections. As observed
in the prudent and careful semantics [22, 23], based on the interplay of attack and defence, there might be additional
positive and negative indirect interactions between the arguments that go beyond direct attack and defence:
1 Defence
is often also referred to as acceptability: we say that A is acceptable w.r.t. E iff E defends A.
5
Definition 2.3. Let F = (A, R) be a Dung’s framework and let A, B ∈ A be two arguments:
• A indirectly attacks B iff there exists an odd–length path from A to B in F .
• A indirectly defends B iff there exists an even–length path from A to B in F . The length of this path is not
zero.
• A is controversial w.r.t. B iff A indirectly attacks and indirectly defends B.
Additionally, for arguments A, B, C ∈ A, the pair (A, B) is super–controversial w.r.t. C iff A indirectly attacks C
and B indirectly defends C.
Originally, these notions were used to impose stronger restrictions on the conflict–free sets. Although we will not
recall the semantics themselves, the aforementioned notions will be useful to us in the context of this work. In particular, we will test whether certain elements of the argument graphs sourced from the participants of our experiments
could potentially be the manifestations of the above definitions.
Example 1 (Continued). In the considered framework, we can observe that each of C, D and E function as indirect
attackers and defenders of E. In addition to that, we can consider C and D as indirectly attacking each other. They
also serve as their own defenders. We can observe that C and D also respectively indirectly attack and defend B.
Let us now focus on the labeling–based semantics, in particular those corresponding to the classical extension–
based ones. An argument labeling is a total function L : A → {in, out, und} [11, 2]. By in(L), out(L) and und(L)
we denote the arguments mapped respectively to in, out and und(ecided) by L. We will often write a labeling as a
triple (I, O, U ), where I, O and U are sets of arguments mapped to in, out and und. We can now introduce the notion
of legality, on which many semantics are based.
Definition 2.4. Let F = (A, R) be a Dung’s framework and let L : A → {in, out, und} be a labeling:
• X ∈ in(L) is legally in iff all its attackers are in out(L).
• X ∈ out(L) is legally out iff it has an attacker in in(L).
• X ∈ und(L) is legally und iff not all of its attackers are in out(L) and it does not have an attacker in in(L).
Definition 2.5. Let L : A → {in, out, und} be a labeling:
(lab-cf) L is conflict–free iff every A ∈ out(L) is legally out and there are no arguments A, B ∈ in(L) s.t. A is an
attacker of B.
(lab-ad) L is admissible iff every A ∈ in(L) is legally in and every A ∈ out(L) is legally out.
(lab-co) L is complete iff it is admissible and every A ∈ und(L) is legally und.
The preferred, stable and grounded labelings are obtained from the complete ones by using the constraints from
Table 1.
Restriction on a complete labeling L
No argument A ∈ A s.t. L(A) = und
Maximal no. of A ∈ A s.t. L(A) = in
Maximal no. of A ∈ A s.t. L(A) = out
Maximal no. of A ∈ A s.t. L(A) = und
Minimal no. of A ∈ A s.t. L(A) = in
Minimal no. of A ∈ A s.t. L(A) = out
Semantics
stable (ST)
preferred (PR)
preferred (PR)
grounded (GR)
grounded (GR)
grounded (GR)
Table 1: Relation between different labelings
The properties of the labeling–based semantics and their correspondence to the classical extension–based family
have already been studied in [11, 2]:
6
Theorem 2.6 ([2]). Let F = (A, R) be a Dung’s framework and E ⊆ A be a σ–extension of F , where σ ∈
{cf, ad, co, gr, pr, st}. Let E + = {B | there exists A ∈ E s.t. A is an attacker of B}. Then (E, E + , A \ (E ∪ E + )) is a
σ–labeling of F . Let L be a σ–labeling of F , where σ ∈ {cf, ad, co, gr, pr, st}. Then in(L) is a σ–extension of F .
Example 1 (Continued). Let us come back to the previously analyzed framework. Its admissible labelings are visible
in Table 2. We can observe that one admissible extension can be associated with more than one labeling. However, out
of the possible interpretations, only L3 , L9 and L13 are complete. They are now also in one–to–one relation with the
complete extensions. L3 is the grounded labeling, while L9 and L13 are preferred. Only L13 is stable; we can observe
that this is the only labeling in which no argument is assigned und.
Labeling
Extension
#
A
B
C
D
E
∅
L1
und
und
und
und
und
{A}
L2
in
und
und
und
und
{A}
L3
in
out
und
und
und
{C}
L4
und
out
in
out
und
{C}
L5
und
und
in
out
und
{D}
L6
und
und
out
in
und
{D}
L7
und
und
out
in
out
{A, C}
L8
in
und
in
out
und
{A, C}
L9
in
out
in
out
und
{A, D}
L10
in
und
out
in
und
{A, D}
L11
in
und
out
in
out
{A, D}
L12
in
out
out
in
und
{A, D}
L13
in
out
out
in
out
Table 2: Admissible labelings of the framework ({A, B, C, D, E}, {(A, B), (C, B), (C, D), (D, C), (D, E),
(E, E)})
3
Bipolar Argumentation
In Dung’s framework, from direct attacks we can derive defence, which can be seen as a type of a positive indirect
relation between arguments. However, defence does not account for all the possible forms of support between arguments, and a structure going beyond attack was required. Consequently, the notion of abstract support and the bipolar
argumentation framework [12, 13, 15] were introduced, followed by the deductive [5], necessary [43, 42] and evidential supports [46, 47, 53], with the latter two developed in their own dedicated frameworks. Although there are
significant differences between these supports and the way that their dedicated frameworks model them, particularly
in the context of support cycles, they have also been partially recreated in bipolar argumentation frameworks in order
to perform a comparative study [15]. Further results on this topic can be found in [53, 49]. The bipolar argumentation
framework itself is defined as follows:
Definition 3.1. The bipolar argumentation framework (BAF for short) is a tuple (A, R, S), where A is a set of
arguments, R ⊆ A × A represents the attack relation and S ⊆ A × A the support.
We say that there is a sequence of supports between arguments A, B ∈ A if there is a sequence of arguments
(A, C1 , ...Cn , B) s.t. ASC1 , C1 SC2 , . . . , Cn SB.
One of the ways bipolar argumentation frameworks model support on the semantics level is by transforming it into
attack [15, 17]. The base framework is extended with indirect conflicts, stemming from the interplay of existing attacks
and supports. The resulting structure can then be evaluated like a Dung’s framework, particularly when semantics that
are at least complete are concerned. The type of support we try to model affects what sort of additional indirect attacks
are created. Although this approach does not account for all possible support semantics [49], it is sufficient for the
purpose of this work. Typically, the following indirect conflicts are distinguished:
Definition 3.2. Let BF = (A, R, S) be a BAF. The indirect attacks of BF are as follows:
• there is a supported attack from A to B iff there exists an argument C s.t. there is a sequence of supports from
A to C and (C, B) ∈ R.
• there is a secondary attack2 from A to B iff there exists an argument C s.t. there is a sequence of supports
from C to B and (A, C) ∈ R.
2 This
attack was also referred to as diverted in [13].
7
C
A
B
(a) Supported attack
C
A
A
B
C
B
(b) Secondary attack
C
A
(c) Mediated attack
D
C
A
B
(d) Extended attack
B
A
C
D
B
(e) Super–mediated attacks (squiggly edges) and existing sup- (f) Super–extended attacks (squiggly edges) and existing secported and mediated attacks (dotted edges)
ondary and extended attacks (dotted edges)
Figure 2: Indirect attacks in BAFs. Solid edges represent direct attacks, dashed edges represent support, and squiggly
or dotted edges represent indirect attacks.
• there is an extended attack3 from A to B iff there exists an argument C s.t. there is a sequence of supports
from C to A and (C, B) ∈ R.
• there is a mediated attack from A to B iff there exists an argument C s.t. there is a sequence of supports from
B to C and (A, C) ∈ R.
• there is a super–mediated attack from A to B iff there exists an argument C s.t. there is a sequence of supports
from B to C and a direct or supported attack from A to C.
• there is a super–extended attack4 from A to B iff there exists an argument C s.t. there is a sequence of supports
from C to A and a direct or secondary attack from C to B.
Additionally, the interactions between these auxiliary conflicts can also lead to the creation of new attacks. In
particular, the super–mediated attack is a mixture of supported and mediated attacks (more details can be found in
[49]). Nevertheless, the above notions are sufficient for the remainder of this report. We can see them exemplified in
Figure 2.
We would like to stress that even though there are many types of conflicts available, it does not mean that all
of them need to be used – the choice depends on what we intend to use a given BAF for. Usually, only some of the
conflicts are studied at a time, particularly if we consider specialized forms of support rather than just the abstract type.
What needs to be stated explicitly is that BAFs were meant as research frameworks for analyzing the different types
of support and the consequences of their interplay with attack. Therefore, there is no “absolute” way to choose what
sort of indirect attacks need to be taken into account and different interpretations of support might call for different
attacks. In our experiment, we will focus on finding the existing notions that would allow us to reproduce the relations
identified by the participants, rather than on fixing the interpretation of the used supports and stating that given indirect
conflicts should have been used.
Example 2. Consider the bipolar argumentation framework BF = (A, R, S) with A = {A, B, C, D, E}, R =
{(C, B), (C, D), (D, C), (E, E)} and S = {(A, B), (D, E)}, depicted in Figure 3. We can create the following
indirect conflicts for this framework. Since D supports E, which is a self–attacker, we can create a supported attack
3 We
4 In
recall only one form of the extended attack, as other ones are already subsumed by the direct and secondary attacks.
[17] this attack is referred as an n+-attack.
8
(D, E) and a (super) mediated one (E, D). These conflicts also lead to the super–mediated attack (D, D). Due to the
same support, we can create a secondary attack (C, E) and a (super) extended one (E, C). As C attacks B, which is
supported by A, we can create a (super) mediated attack (C, A).
If we decide to use all of the aforementioned attacks, then the Dung’s framework associated with BF is F =
(A, R′ ), where R′ = R ∪ {(C, A), (C, E), (D, D), (D, E), (E, C), (E, D)}. The sets ∅ and {C} are its admissible
and complete extensions. {C} is the stable and preferred extension, while ∅ is grounded. Following the approach from
[15], we can treat these extensions as the extensions of BF .
If we consider only the secondary and (super) extended attacks, we obtain a Dung’s graph with the set of attacks
R′′ = R ∪ {(C, E), (E, C)}. The admissible extensions of this framework are ∅, {A}, {C}, {D}, {A, C}, {A, D},
{B, D} and {A, B, D}. Out of this, {A}, {A, C} and {A, B, D} are complete, with the first set being grounded and
the other two preferred. Only {A, C} is stable.
A
B
C
D
E
Figure 3: Sample bipolar argument graph. Solid edges represent attack, dashed edges represent support and squiggly
edges represent indirect attacks.
4
Probabilistic Argumentation
The proposals for probabilistic argumentation extend Dung’s framework to address various aspects of uncertainty
arising in argumentation. The two main approaches to probabilistic argumentation are the constellations and the
epistemic approaches [32]. In the constellations approach, the uncertainty is in the topology of the graph. This
approach is useful when one agent is not sure what arguments and attacks another agent is aware of. This can be
captured by a probability distribution over the space of possible argument graphs, where each graph has a chance of
being the real model of the agent. In the epistemic approach, the topology of the argument graph is fixed, but there is
uncertainty as to the degree to which each argument is believed. It thus provides us with a more fine–grained approach
towards modelling the acceptability of arguments. In this section we provide a brief revision of these two probabilistic
formalisms and for further reading concerning the differences between them we refer the readers to [32, 52].
4.1 Epistemic Probabilistic Argumentation
We now turn to the epistemic approach to probabilistic argumentation [58, 32, 33, 4, 36]. In this section we review
and extend the results from [34].
P
Definition 4.1. A mass distribution over arguments
A is a function P : 2A → [0, 1] s.t.
E⊆A P (E) = 1. The
P
probability of an argument A ∈ A is P (A) = E⊆A s.t. A∈E P (E).
The probability of a single argument is understood as the belief that an agent has in it, i.e. we say that an agent
believes an argument A to some degree when P (A) > 0.5, disbelieves an argument to some degree when P (A) < 0.5,
and neither believes nor disbelieves an argument when P (A) = 0.5. This belief can be interpreted in various ways
[32], for example, if we assume that an argument has a standard premise–conclusion structure, then the degree of
belief in an argument can be seen as the degree to which the agent believes the premises and the conclusion drawn
from those premises.
Just like argument graphs come equipped with different types of semantics that are meant to capture various intuitions as to what is a “good” collection of arguments (and/or attacks), the probabilistic frameworks are accompanied by
9
a number of postulates that capture the properties of a probability distribution. While classical semantics tend to represent a number of properties at the same time, a single postulate focuses on a single aspect at a time. Thus, while there
are many epistemic postulates, they tend to be quite simple. They also allow a more detailed view on the participant
behaviour and can allow us to analyze the cases in which classic semantics may fail to explain it. Consequently, they
can provide more feedback to argumentation systems, such as for computational persuasion, than normal semantics
do. For example, the low performance of the argumentation semantics such as complete does not really inform the
system what aspect of the participant reasoning does not meet the semantics requirements. In contrast, through the use
of the postulates we could analyze whether the issue lies within conflict–freeness, defense or not accepting/rejecting
arguments that should be accepted/rejected.
We will recall some of the existing postulates for the epistemic semantics and introduce certain argument–centered
alternatives of the properties from [52], namely the preferential postulate PRE, strict STC, protective PRO, restrained
RES, discharging DIS, guarded GRD, trusting TRU, anticipating ANT, demanding DEM, binary BIN and n–valued
VALn . The available postulates can be grouped in various ways. In what follows, we separate them into the following,
not necessarily disjoint, types: preferential (rational), explanatory and value families.
4.1.1 Preferential Postulates
The first family of postulates focuses on resolving a conflict between two arguments and deciding how believed or
disbelieved the attacker and attackee should be. They are primarily of the form “for every A, B ∈ A, if (A, B) ∈ R,
then X holds”, where X specifies the conditions on the beliefs assigned to the attacker and the attackee. The intuitions
behind the introduced postulates are as follows. Let us assume that we believe an argument attacking another argument.
The question now is what should be our belief in the attackee. The least restrictive, PRE postulate, allows us to believe
the attackee as long as it is believed more than the attacker. If we follow RAT, then we do not believe the attackee,
and if STC, then we disbelieve it. In a dual manner, the PRO postulate ensures that if we believe the attackee, we
disbelieve its attackers. The RES postulate strengthens PRO and STC and is equivalent to saying that if two arguments
are in conflict, at least one of them should be disbelieved. Hence, we no longer have the option to be undecided about
an argument.
Definition 4.2. A probability mass distribution P for F is:
(PRE) preferential if for every A, B ∈ A s.t. (A, B) ∈ R, if P (A) > 0.5 and P (B) > 0.5, then P (A) < P (B).
(RAT) rational if for every A, B ∈ A s.t. (A, B) ∈ R, P (A) > 0.5 implies P (B) ≤ 0.5.
(STC) strict if for every A, B ∈ A s.t. (A, B) ∈ R, P (A) > 0.5 implies P (B) < 0.5.
(PRO) protective if for every A, B ∈ A s.t. (A, B) ∈ R, P (B) > 0.5 implies P (A) < 0.5.
(RES) restrained if for every A, B ∈ A, if (A, B) ∈ R then P (A) ≥ 0.5 implies P (B) < 0.5.
4.1.2 Explanatory Postulates
We now come to the explanatory family. The purpose of the postulates of this type is to demand that the degree of
belief assigned to an argument is justified by the degrees of belief associated with the arguments related to it. They
are roughly of the form “for every argument B ∈ A, if P (B) meets a condition X then it has an attacker A s.t.
P (A) meets condition Y ” or “for every argument B ∈ A, if for all of its attackers A, P (A) meets condition Y ,
then P (B) meets a condition X”. For instance, DIS demands that if an argument is disbelieved, then it possesses
a believed attacker. GRD postulate relaxes this requirement by allowing the use of undecided attackers as well. On
the other hand, the TRU property requires us to believe an argument when we have no reason against it, i.e. when
all of its attackers are disbelieved. ANT modifies TRU by saying that lack of believed attackers is a good reason to
believe the attackee. Finally, the DEM property requires that a completely disbelieved argument has to be paired with
a completely believed attacker, and a completely believed argument can only be attacked by completely disbelieved
arguments.
Definition 4.3. A probability mass distribution P for F is:
10
(DIS) discharging if for every B ∈ A, if P (B) < 0.5 then there exists an argument A ∈ A s.t. (A, B) ∈ R and
P (A) > 0.5.
(GRD) guarded if for every B ∈ A, if P (B) < 0.5 then there exists an argument A ∈ A s.t. (A, B) ∈ R and
P (A) ≥ 0.5.
(TRU) trusting if for every B ∈ A, if P (A) < 0.5 for all A ∈ A s.t. (A, B) ∈ R, then P (B) > 0.5.
(ANT) anticipating if for every B ∈ A, if P (A) ≤ 0.5 for all A ∈ A s.t. (A, B) ∈ R, then P (B) > 0.5.
(DEM) demanding if for every A ∈ A, if P (A) = 1, then for every B ∈ A s.t. (B, A) ∈ R, P (B) = 0, and if
P (A) = 0, then ∃B ∈ A s.t. (B, A) ∈ R and P (B) = 1.
4.1.3 Value Postulates
All of the previously listed postulates are relatively general and tell us whether to believe or disbelieve an argument,
but not to what degree. This is where the value family of postulates comes in. The FOU and SFOU postulates tell us
how much we should believe initial arguments. OPT and SOPT provide us with lower bounds for the degrees of belief
we have in an element based on the beliefs we have in its attackers. The BIN postulate prohibits any indecisiveness we
may have about the arguments. MIN, NEU, MAX permit the use of only 0, 0.5 and 1 degrees of belief respectively,
while TER allows for all three. Finally, with VALn postulate we can distinguish those distributions that assign to their
arguments no more than n distinct values in total.
Definition 4.4. A probability mass distribution P for F is:
(SFOU) semi–founded if P (A) ≥ 0.5 for every initial A ∈ A.
(FOU) founded if P (A) = 1 for every initial A ∈ A.
P
(SOPT) semi–optimistic if P (A) ≥ 1 − B∈{A}− P (B) for every A ∈ A that is not initial.
P
(OPT) optimistic if P (A) ≥ 1 − B|(B,A)∈R P (B) for every A ∈ A.
(BIN) binary if for no A ∈ A, P (A) = 0.5.
(TER) ternary if P (A) ∈ {0, 0.5, 1} for every A ∈ A.
(NEU) neutral if P (A) = 0.5 for every A ∈ A.
(MAX) maximal if P (A) = 1 for every A ∈ A.
(MIN) minimal if P (A) = 0 for every A ∈ A.
(VALn ) n–valued if |{x | ∃A ∈ A s.t. P (A) = x}| ≤ n.
4.1.4 Multi–type Postulates
Finally, we have the postulates that can be seen as shared between the families. In particular, on the intersection of
the rational (preferential) and value types are the COH, INV and JUS postulates. COH gives us the upper bound of
the belief we might have in an argument based on its (strongest) attacker. By combining this upper bound with the
lower bound from the OPT postulate, we receive the JUS property. The INV postulate requires that the belief in the
attackee is dual to the belief in the attacker. Please note that the DEM, SFOU and FOU postulates can be seen as
shared between the value and explanatory families, rather than belonging to just one of them. We leave it to the reader
to classify them as he or she sees fit.
Definition 4.5. A probability mass distribution P for F is:
(COH) coherent if for every A, B ∈ A s.t. (A, B) ∈ R, P (A) ≤ 1 − P (B).
(INV) involutary if for every A, B ∈ A, if (A, B) ∈ R then P (A) = 1 − P (B).
(JUS) justifiable if it is coherent and optimistic.
11
4.1.5 Properties of Epistemic Postulates
In Proposition 4.6 and Figure 4 we show some of the relations between the presented postulates, extending the results
available in [35]. These properties make it more explicit which postulates are more restrictive than the others and show
additional connections between the different families of postulates we have presented.
Proposition 4.6. Let F be an argument graph and let PX be the collection of all distributions on F satisfying postulate
X. The following holds:
POP T (F ) = PSOP T (F ) ∩ PF OU (F )
PT ER (F ) ∩ PT RU (F ) ⊆ PF OU (F )
(PCOH (F ) ∩ PT ER (F )) = (PP RO (F ) ∩ PST C (F ) ∩ PT ER (F ))
PIN V (F ) ∩ PF OU ⊆ PDEM (F )
PP RE
P
PSF OU
PRAT
PT RU
PGRD
PF OU
PSOP T
PST C
PP RO
PANT
PDIS
POP T
PINV
PCOH
PRES
PM AX
PNEU
PJ U S
PM IN
∅
∅
PNEU
PDEM
PM IN
PT ER
PM AX
PBIN
P
Figure 4: Classes of probability functions where Pµ1 → Pµ2 denotes Pµ1 ⊆ Pµ2
Example 1 (Continued). Let us come back to the argument graph F = ({A, B, C, D, E}, {(A, B), (C, B), (C, D),
(D, C), (D, E), (E, E)}) from Figure 1. In Table 3 we have listed a number of probability distributions and marked
which postulates are or are not satisfied by them under F . For the sake of simplicity, we have only focused on the
resulting argument probabilities, not on the full description of the distributions.
4.1.6 Relation to Labeling–based Semantics
In the previous sections we could have observed that the epistemic approach offers a wide variety of postulates describing certain properties that a given probability function may or may not possess. In this section, we would like to
show how some of the combinations of these postulates can capture the intuitions behind the standard labeling based
semantics. This information will become useful once we analyze the postulate adherence in our experiment.
12
COH
JUS
INV
DIS
GRD
TRU
ANT
DEM
SFOU
FOU
SOPT
OPT
E
0.8
0.5
0.5
0.4
0.5
0.5
0.5
0.5
0.5
0.5
RES
D
0.4
0.2
0.4
0
0
0.4
0
0.5
0.7
0.5
PRO
C
0.2
0.6
0.6
1
1
0.5
1
0.5
0.4
0.5
STC
B
0.5
0.5
0.4
0.5
0
0
0
1
1
0.5
RAT
A
0.6
0.3
0.7
1
1
1
0.5
0.5
0.6
0.5
PRE
#
P1
P2
P3
P4
P5
P6
P7
P8
P9
P10
×
X
X
X
X
X
X
X
X
X
×
X
X
X
X
X
X
X
×
X
×
×
X
×
X
X
X
X
×
X
×
X
X
X
X
X
X
×
×
X
×
×
×
X
×
×
×
×
×
×
×
×
×
×
X
X
X
×
×
X
×
×
×
×
X
×
×
×
×
×
×
×
×
×
×
×
×
×
×
X
×
×
X
×
X
×
X
X
X
X
×
×
X
×
X
X
X
X
X
X
×
×
X
×
X
×
×
×
X
×
×
×
×
×
×
×
×
×
X
×
X
X
X
X
X
X
X
×
×
X
X
×
X
X
X
X
X
X
X
X
×
×
×
X
X
X
×
×
×
×
×
×
X
×
X
×
X
X
X
X
×
×
×
×
X
×
×
×
×
×
Table 3: Satisfaction of epistemic postulates on Dung’s framework from Figure 1.
Definition 4.7. Let F = (A, R) be an argument graph and let P be a probability distribution. The epistemic labeling
associated with P is LP = (I, O, U ), where I = {A ∈ A | P (A) > 0.5}, O = {A ∈ A | P (A) < 0.5}, and
U = {A ∈ A | P (A) = 0.5}. A labeling L and a probability function P are congruent, denoted L ∼ P , if for all
A ∈ A we have that L(A) = in ⇔ P (A) = 1, L(A) = out ⇔ P (A) = 0, and L(A) = und ⇔ P (A) = 0.5.
Please note that if L ∼ P , then L = LP , i.e. if a labeling and a probability function are congruent then this
labeling is also an epistemic labeling of this function.
Proposition 4.8. Let P be an epistemic distribution and let F = (A, R) be a Dung’s framework.
• P ∈ PRAT (F ) ∩ PDIS (F ) iff LP ∈ cf(F ).
• P ∈ PP RO (F ) ∩ PDIS (F ) iff LP ∈ ad(F ).
• P ∈ PP RO (F ) ∩ PST C (F ) ∩ PDIS (F ) ∩ PT RU (F ). iff LP ∈ co(F ).
If we were to consider retrieving the extension–based semantics through epistemic postulates, the above results
should be rephrased in the following manner: if a probability distribution satisfies the given postulates, then the set of
in arguments of the associated labeling is an extension of a given type, and if a set of arguments is an extension of a
given type, then there exists a probability distribution satisfying the given postulates s.t. the set of believed arguments
coincides with the extension. The only relaxation we are aware of is in the case of conflict–free semantics, where a set
of arguments is conflict–free iff the associated probability distribution is rational.
Example 1 (Continued). For every labeling listed in Table 2, we can create a congruent probability function and show
that it satisfies the required postulates. For example, we can consider the admissible labeling L6 and its congruent
distribution s.t. P (A) = P (B) = P (E) = 0.5, P (C) = 0 and P (D) = 1. The protective and discharging postulates
are easily satisfied, however, the trusting and strict ones are not – in particular, A should have been believed and B, E
disbelieved. Addressing these issues would produce a distribution that is associated with the complete labeling L13 .
Let us now look at the distributions in Table 3. All P3 , P5 , P7 , P8 and P10 are both rational and discharging.
It is easy to verify that the associated labelings LP3 = LP5 = {A : in, B : out, C : in, D : out, E : und},
LP7 = {A : und, B : out, C : in, D : out, E : und}, LP8 = {A : und, B : in, C : und, D : und, E : und} and
LP10 = {A : und, B : und, C : und, D : und, E : und} are conflict–free. All of them, with the exception of LP8 , are
also admissible – at the same time, we can observe that P8 is not protective. Finally, only LP3 and LP5 are complete,
which is reflected by P3 and P5 being the only of the listed functions that are additionally strict and trusting.
Therefore, as we can observe, epistemic probability, understood as the degree of belief, can also express various
levels of agreement in the Dung’s sense, i.e. accepting, rejecting and being undecided about an argument.
13
4.2 Constellation Probabilistic Argumentation
Here we review the constellation approach to probabilistic argumentation from [31], which extends the methods from
[28] and [38]. The constellation approach allows us to represent the uncertainty over the topology of the graph.
Each subgraph of the original graph is assigned a probability which is understood as the chances of it being the
actual argument graph of the agent. It can be used to model what arguments and attacks an agent is aware of. If our
uncertainty is about which arguments appear in the graph, then only the full (induced) subgraphs of the argument graph
have a non–zero probability. If we are only uncertain about which attacks appear, then it is the spanning subgraphs of
the argument graph that can have a non–zero probability.
Definition 4.9. Let F = (A, R) and F ′ = (A′ , R′ ) be two argument graphs. F ′ is a subgraph of F , denoted F ′ ⊑ F ,
iff A′ ⊆ A and R′ ⊆ (A′ × A′ ) ∩ R. The set of subgraphs of F is Sub(F ) = {F ′ | F ′ ⊑ F }. A subgraph (A′ , R′ )
is full iff A′ ⊆ A and R′ = (A′ × A′ ) ∩ R. A subgraph (A′ , R′ ) is spanning iff A′ = A and R′ ⊆ R.
P
Definition 4.10. A subgraph distribution is a function P c : Sub(F ) → [0, 1] with F ′ ∈Sub(F ) P c (F ′ ) = 1. A
subgraph distribution P c is a full subgraph distribution iff if (A′ , R′ ) is not a full subgraph, then P c ((A′ , R′ )) =
0. A subgraph distribution P c is a spanning subgraph distribution iff if (A′ , R′ ) is not a spanning subgraph,
P c ((A′ , R′ )) = 0.
The constellation semantics can be seen as a two–level construct. Each graph in the subgraph distribution is
evaluated with a given base semantics, the results of which are later paired with the probabilities of the frameworks
producing them, thus leading to constellation semantics. Although we will not be using this particular method for
evaluating argument graphs in this paper, we briefly recall and exemplify it so that it is more clear that it is different
from the epistemic approach. Further details can be found in [32, 52].
Determining the probability that a set of arguments or a labeling follows the semantics of a particular type (e. g.
grounded, preferred, etc.) is done by collecting the probabilities of the subgraphs producing the desired extensions or
labelings. In a similar fashion, we can derive the probability of an argument being accepted in a labeling of a given
type.
Definition 4.11. For E ⊆ A and σ ∈ {cf, ad, co, pr, gr, st}, the probability that L : E → {in, out, und} is a σ–labeling
is:
X
Pσ (L) =
P c (F ′ )
F ′ ∈Sub(F ) s.t. L∈σ(F ′ )
Definition 4.12. Given a semantics σ ∈ {ad, co, pr, gr, st}, the probability that A ∈ A is assigned an in status in a
σ–labeling is
X
Pσ (A) =
P (F ′ )
F ′ ∈Sub(F ) s.t. L∈σ(F ′ ) and A∈in(L)
Example 3. Consider the graph F = ({A, B}, {(A, B)}. Its subgraphs are F1 = ({A, B}, {(A, B)}, F2 =
({A, B}, ∅), F3 = ({A}, ∅), F4 = ({B}, ∅) and F5 = (∅, ∅). Out of them, F1 , F3 , F4 and F5 are full, and F1 and
F2 are spanning. Consider the following subgraph distribution P c : P c (F1 ) = 0.09, P c (F2 ) = 0.81, P c (F3 ) = 0.01
and P c (F4 ) = 0.09 and P c (F5 ) = 0. The probability of a given set being a grounded extension is as follows:
Pgr ({A, B}) = P c (F2 ) = 0.81; Pgr ({A}) = P c (F1 ) + P c (F3 ) = 0.1; Pgr ({B}) = P c (F4 ) = 0.09; and Pgr ({}) =
P c (F5 ) = 0. Therefore Pgr (A) = 0.91 and Pgr (B) = 0.9.
For further reading concerning the constellation approach to probabilistic argumentation we refer the readers to
[31, 28, 38, 25]. Computational results can be found in [26, 29, 30], and in [39, 50] we can find approaches combining
the constellation probabilities with generalizations of Dung’s framework. Finally, in [24] we can find a characterization
of one of the versions of the constellation approach in terms of probabilistic logic.
5
Experiment Description
In this section we explain how the experiment was carried out. The purpose of our study was to gather the opinions
on dialogues concerning the topic of flu vaccines. To this end, we prepared two separate dialogues and asked the
14
participants we recruited online to judge at least one of them. In the following sections we will show the dialogues we
have used, explain the tasks given to the participants and describe the recruitment process.
5.1 Dialogues & Tasks
The dialogues presented to the participants can be seen in Tables 4 and 5. The first dialogue discusses the use of the
flu vaccine by hospital staff members, while the second concerns the safety of the children flu vaccine. Every dialogue
is split into five steps in which two fictional discussants, person 1 (P1) and person 2 (P2), take turns in presenting their
opinions. A given statement, once uttered, is visible throughout the rest of the dialogue. Although the general topic is
the same, i.e. concerns flu vaccinations, we can observe that the statements used in the dialogues do not overlap. The
dialogues have been created based on the information found on the Centers for Disease Control and Prevention (CDC
USA) and National Health Services (NHS UK) websites.
Table 4: Dialogue 1 between people P1 and P2. This dialogue starts with P1 claiming that hospital staff do not need
to receive flu shots, to which P2 objects. The two counterarguments of P1 are then defeated by P2. The table presents
at which steps a given statement was visible, who uttered it and what was its content.
Steps
Person
Statement
1 to 5
P1
A
1 to 5
P2
B
2 to 5
P1
C
3 to 5
P2
D
4 to 5
P1
E
5
P2
F
Content
Hospital staff members do not need to receive flu shots.
Hospital staff members are exposed to the flu virus a lot. Therefore, it would be
good for them to receive flu shots in order to stay healthy.
The virus is only airborne and it is sufficient to wear a mask in order to protect
yourself. Therefore, a vaccination is not necessary.
The flu virus is not just airborne, it can be transmitted through touch as well.
Hence, a mask is insufficient to protect yourself against the virus.
The flu vaccine causes flu in order to gain immunity. Making people sick, who
otherwise might have stayed healthy, is unreasonable.
The flu vaccine does not cause flu. It only has some side effects, such as
headaches, that can be mistaken for flu symptoms.
Table 5: Dialogue 2 between people P1 and P2. This dialogue starts with P1 claiming that the vaccine is not safe, to
which P2 objects, and the discussion proceeds to revolve around whether it contains mercury–based compounds or
not. The table presents at which steps a given statement was visible, who uttered it and what was its content.
Steps
Person
Statement
1 to 5
1 to 5
P1
P2
2 to 5
P1
A
B
C
D
3 to 5
P2
E
F
4 to 5
P1
G
5
P2
H
Content
The flu vaccine is not safe to use by children.
The flu vaccine does not contain poisonous components and is safe to use.
The vaccine contains some mercury compounds.
The mercury compounds are poisonous and therefore the vaccine is not safe to
use.
The child vaccine does not contain any mercury compounds.
The virus is only accompanied by stabilizers and possibly trace amounts of antibiotics used in its production.
The vaccine contains a preservative called thimerosal which is a mercury-based
compound.
Children receive the nasal spray vaccine and thimerosal has been removed from it
over 15 years ago.
During the experiment, at every stage of the dialogue the participants were presented with three tasks - Agreement,
15
Explanation and Relation - with an additional Awareness question at the end of the dialogue:
• Agreement: the participants were asked to state how much they agree or disagree with a given statement. They
were allowed to choose one of the seven options (Strongly Agree, Agree, Somewhat Agree, Neither Agree nor
Disagree, Somewhat Disagree, Disagree, Strongly Disagree) or select the answer Don’t Know.
• Explanation: the participants were asked to explain the chosen level of agreement for every statement. In
particular, we requested them to provide us with any reason they may have for disagreement that was not
mentioned in the dialogue and to explain reasons for changing their opinions compared to the previous step in
the discussion (if applicable).
• Relation: the participants were asked to state how they viewed the relation between the statements. For every
listed pair, they could say whether one statement was A good reason against, A somewhat good reason against,
Somewhat related, but can’t say how, A somewhat good reason for, A good reason for the other statement or
select the answer N/A (i.e. that the statements were unrelated). The questions were kept with the flow of the
dialogue, i.e. the source of a given relation is preceded by the target in the presented discussion. For example,
we would ask how statement F is related to A, but not the other way around. Thus, the temporal aspects of the
dialogues are taken into account by the graph, similarly as is done in the analysis of online discussions [8, 10].
This also reduced the number of questions we had to ask and therefore simplified the task for the participants.
• Awareness: the participants were asked which of the presented statements they were familiar with prior to the
experiment. By this we understand that, for example, someone they know may have expressed the opinion
contained in a given statement, they might have read it in some source, or generally heard about it before. To
put it simply, we wanted to know what statements the participants were aware of, independently of whether they
agreed or disagreed with them.
5.2 Recruitment
The recruitment of the participants was done using the Amazon Mechanical Turk (AMT for short) and the survey itself
was carried out using the SurveyMonkey, which are both common platforms for experiments of this type. In addition
to the tasks explained in the previous section, the participants were subjected to an additional language exercise, two
attention checks and a comprehension test in order to ensure that they had sufficient skills to complete the test and
that they worked honestly. The language exercise was on an intermediate level in terms of difficulty. The attention
checks presented the participants with two sentences requesting them to choose particular answers. They were meant
to disqualify the participants who were too distracted during the experiment or simply resorted to random clicking
in order to complete the survey. Finally, in the comprehension exercise, they were asked to select the answers that
described what the statements presented in the dialogue concerned. Prior to the survey, the participants were warned
of such tests. We also provided them with explanations and examples of the tasks they would be asked to complete,
including what should be understood as a reason for or against or what we meant by being aware of a given statement.
We also requested the participants not to use Google or Wikipedia in order to verify the statements in the dialogues.
At the very end of these instructions the participants obtained the code used to unlock the actual questionnaire without
which they could not proceed. Moreover, it was also necessary for them to accept the terms and conditions of our
experiment, which included familiarizing themselves with these explanations. Hence, we took all reasonable steps in
informing the participants what to expect.
We ran the survey until we had collected 80 survey responses (40 per dialogue) in which the participants had sufficiently high scores in the language, attention and comprehension tests. As we have noted previously, the participants
- if they wished to do so - were allowed to judge both of the dialogues. We have found 11 people that have completed
both tests, which means we have recruited a total of 69 unique participants. The total number of entries (not participants) was equal to 156, this brings us to an acceptability rate of around 51%. The analysis of the demographics data
on our participants can be found in the auxiliary appendix at [51]. This research project has been approved by the
designated ethics officer in the Computer Science Department at University College London. The participants could
withdraw from the experiment at any point in time and could refuse to provide any piece of information they deemed
too private.
16
6
Results
In this section we will review the results of our study. We will first describe the argument graphs obtained from our
participants as well as the ones we had in mind when creating the dialogues. Given these structures, we will discuss if
and how the declared levels of agreement satisfy the epistemic postulates we have recalled in Section 4.1. Finally, we
will discuss the changes in participants’ beliefs throughout the dialogues.
Before we start, we would like to introduce some necessary notions. As we have explained previously, in addition
to declaring that one statement is a good reason for or against another statement, a possible answer in the relation task
was Somewhat related, but can’t say how. In order to be able to represent this dependency that a participant could not
classify, we propose the notion of a tripolar argument graph:
Definition 6.1. The tripolar argument graph is a tuple T F = (A, R, S, D) where A is a set of arguments, R ⊆
A × A is the set of attacks, S ⊆ A × A is the set of supports and D ⊆ A × A is the set of dependencies.
Please note that we introduce the tripolar graph only for representation purposes and we do not define any semantics
for it. When required, we can simply extract a Dung’s or bipolar graph from it. Its primary purpose is simply to
mark relations which, potentially due to confusion or the difficulty of the experiment, the participants had problems
classifying. Consequently, unlike in [48, 18, 21], it does not necessarily mean that the relation is really neither an
attack nor a support, or that the relation might not exist.
In the following sections the ability to compare two graphs will be important for us. To this end, we introduce
various types of subgraphs. To put it simply, if all arguments and links contained in one framework are present in
another framework, we say that the first framework is a subgraph of another. If the links are perceived in the same
manner (e.g. if it is an attack in one graph, it is also attack in the other), we are dealing with a correct subgraph, which
is the most commonly considered type. We can then relax this notion in a number of ways. For example, if we allow
links such as attacking or supporting to become classified as dependent, we can use the notion of a confusion subgraph.
In a dual fashion, by allowing the dependent links to be classified as attacking or supporting, we create a precision
subgraph. Finally, if we only concern ourselves with verifying whether the arguments connected in one framework are
connected in another, independently of what the nature of that connection might be, we can use the notion of a lenient
subgraph:
Definition 6.2. Let T F = (A, R, S, D) and T F ′ = (A′ , R′ , S ′ , D′ ) be two tripolar argument graphs. We say that:
• T F ′ is a correct subgraph of T F iff A′ ⊆ A, R′ ⊆ R, S ′ ⊆ S and D′ ⊆ D.
• T F ′ is a confusion subgraph of T F iff A′ ⊆ A, R′ ⊆ R ∪ D and S ′ ⊆ S ∪ D and D′ ⊆ D.
• T F ′ is a precision subgraph of T F iff A′ ⊆ A, R′ ⊆ R and S ′ ⊆ S and D′ ⊆ D ∪ R ∪ S.
• T F ′ is a lenient subgraph of T F iff A′ ⊆ A and (R′ ∪ S ′ ∪ D′ ) ⊆ (R ∪ S ∪ D).
We say that a tripolar framework T F is clarified if D = ∅.
Given the fact that the new types of subgraphs were introduced as a way to relax the classical correct subgraph
notion, there certain relations between them. We can observe that every correct subgraph is also a confusion and a
precision subgraph, and every confusion subgraph or precision subgraph is a lenient subgraph. However, not every
confusion subgraph meets precision requirements and not every precision subgraph meets confusion requirements.
Example 4. Let us consider the graphs T F1 = ({A, B, C, D}, {(A, B), (B, C)}, {(A, C)}, ∅), T F2 = ({A, B, C, D},
{(A, C)}, {(A, B), (B, C)}, ∅), T F3 = ({A, B, C, D}, {(A, B), (B, C)}, {(A, C), (A, D)}, ∅) and T F4 =
({A, B, C, D}, {(A, B)}, {(A, C), (A, D)}, {(B, C)}). We can observe that any framework is its own subgraph
of any type. In addition to that, T F1 is a correct subgraph of T F3 . It is a confusion subgraph of both T F3 and T F4 , a
precision subgraph of T F3 and a lenient subgraph of all of the listed frameworks. On the other hand, T F2 is not a correct, confusion or precision subgraph of any other framework apart from itself. It is however a lenient subgraph of all
of the listed structures. T F3 is not a correct subgraph of any other framework apart from itself and is a confusion and
lenient subgraph of T F4 . T F4 is only its own correct and confusion subgraph, and a precision and lenient subgraph
of T F3 and T F4 .
17
In addition to being able to represent the answers provided by the participants, we would also like to be able to
measure the disagreement between the various frameworks we obtain at a given stage of the dialogue. For this purpose,
we create a notion distance between frameworks conceptually similar to the ones from [21, 7, 14]. The distance is
measured in terms of links by which two given frameworks differ. In particular, if the given arguments are seen as
unrelated in one structure, but as attacking/supporting or dependent in the other, we set the difference to 2 and 1
respectively. If a link has opposing polarities (i.e. attacking vs supporting) in the frameworks, the distance is set again
to 2. If it is attacking or supporting in one and dependent in another, the distance is set to 1. Finally, if the nature of a
given relation is seen in the same way, then the difference is naturally 0. The distance between the frameworks is then
simply a sum of all such differences:
Definition 6.3. Let T F = (A, R, S, D) and T F ′ = (A, R′ , S ′ , D′ ) be two tripolar frameworks defined over the same
′
set of arguments s.t. R ∩ S ∩ D = ∅ and R′ ∩ S ′ ∩ D′ = ∅. Let RelT F = R ∪ S ∪ D and RelT F = R′ ∪ S ′ ∪ D′ be
′
the sets of all relations in both structures. The difference between T F and T F on edge α is defined in the following
manner:
′
• If α ∈
/ RelT F ∩ RelT F :
0
′
diff(T F, T F , α) = 1
2
if α ∈
/ RelT F ∪ RelT F
if α ∈ D ∪ D′
otherwise
′
′
• If α ∈ RelT F ∩ RelT F :
1
1
diff(T F, T F ′ , α) = 2
2
0
if α ∈ D ∧ α ∈
/ D′
if α ∈
/ D ∧ α ∈ D′
if α ∈ S ∧ α ∈ R′
if α ∈ R ∧ α ∈ S ′
otherwise
The distance between T F and T F ′ is then defined as:
X
d(T F, T F ′ ) =
diff(T F, T F ′ , α)
α∈RelT F ∪RelT F ′
N/A
1
2
2
D
R
1
1
2
S
Figure 5: Difference measure between attacking, supporting, dependent and nonexistent relations
The introduced notion of a distance between the frameworks meets the requirements of a metric, i.e. the following
holds:
Lemma 6.4. Let T F1 = (A, R1 , S1 , D1 ), T F2 = (A, R2 , S2 , D2 ) and T F3 = (A, R3 , S3 , D3 ) be tripolar frameworks defined over the same sets of arguments and s.t. for i ∈ {1, 2, 3}, Ri ∩ Si ∩ Di = ∅. The following holds:
1. d(T F1 , T F2 ) ≥ 0.
18
2. d(T F1 , T F2 ) = 0 ⇐⇒ T F1 = T F2 5 .
3. d(T F1 , T F2 ) = d(T F2 , T F1 ).
4. d(T F1 , T F3 ) ≤ d(T F1 , T F2 ) + d(T F2 , T F3 ).
Finally, in order to be able to analyze which postulates may or may not be satisfied by the participants, we need to
map the declared levels of agreement to actual numerical values. By this we understand that if an argument X is assigned an answer such as Strongly Disagree by a participant, then there is a numerical value y associated with Strongly
Disagree s.t. for the purpose of this analysis we can say that P (X) = y. From such assignments, a full probability
distribution over the powerset of arguments can be reproduced [36]. However, given the way the probabilities are used
by the epistemic postulates that we have recalled, knowing that P (X) = y is sufficient for our purposes. We propose
the following assignments:
• Strongly Agree – 6/6.
• Agree – 5/6.
• Somewhat Agree – 4/6.
• Neither Agree nor Disagree – 3/6.
• Somewhat Disagree – 2/6.
• Disagree – 1/6.
• Strongly Disagree – 0/6.
Additionally, the Don’t know answer is available in the questionnaire. Unless stated otherwise, for calculation
purposes we assign it the same value as Neither Agree nor Disagree (i.e. 3/6). It is a possible way of shifting from
a four–valued setting (in, out, undecided and unknown) to a three–valued one (in, out, undecided), where inability to
make a decision is treated similarly as indecisiveness.
Finally, we create the graphs based on the answers in relation task in the following manner. The Reason for answer
we will mark as support, Reason against as attack, and the Somewhat related, but can’t say how as dependent. The
declared strength of the attack or support relation will only become relevant once we analyze its connection to the
agreement level assigned to its source in Section 6.3.
6.1 Argument Graphs
In this section we will analyze the argument graphs we have received from the participants based on their responses to
the relation tasks. In what follows we will distinguish the following types of graphs:
• the intended graph is meant to depict the minimal set of relations we considered reasonable for a given set of
arguments.
• the augmented graph is obtained by adding to the intended graph the indirect relations from the prudent/careful
or bipolar argumentation approaches (Definitions 2.3 and 3.2).
• the participant–sourced graphs:
– the declared graph is constructed from the answers given to us by the participants in the Relation tasks,
where the Reason for answer we will mark as support, Reason against as attack, and the Somewhat related,
but can’t say how as dependent. The arguments correspond to the statements appearing at a given step of
the dialogue.
– the expanded graph is constructed from the declared graph, extended with the statements extracted from
the answers that the participants have provided in the Explanation task.
5 In
this context we say that T F1 = T F2 iff R1 = R2 , S1 = S2 and D1 = D2 .
19
– the common graph is the declared graph created by the highest number of participants at a given step in
the dialogue.
The intended graph simply depicts certain minimal constraints we had in mind for this experiment. For example,
we had intended the statement “Hospital staff members are exposed to the flu virus a lot. Therefore, it would be good
for them to receive flu shots in order to stay healthy.” to attack “Hospital staff members do not need to receive flu
shots.”. Please note that this graph does not necessarily contain all of the possible interactions between the statements.
For example, we decided against putting any supporting links in. The purpose of this graph was merely to encompass
the minimal constraints we had intended the participants to recognize. Consequently, this does not necessarily mean
that it contains all the possible relations that could be extracted from the dialogue.
The augmented graph is built from the intended graph through the addition of additional relations, similarly as
indirect edges can be added to bipolar graphs for evaluation (see Section 3 and [49] for further details). The purpose
of this graph is to check whether the relations declared by the participants in addition to the ones stated in the intended
graph can in fact be reproduced through the use of notions from Definitions 2.3 and 3.2. In particular, we will augment
the intended graph with attacks using the definitions of indirect conflicts from the aforementioned definitions and with
supports using the definition of indirect defense. We will then observe that these notions do not necessarily give us
all the possible relations defined by the participants and analyze how the ones that are not accounted for could be
explained.
At this point we would like to stress that in this report, we will follow the abstract interpretation of supports, i.e.
one in which it is understood simply as a positive relation between arguments without any further requirements [12].
As stated in Section 3, we will not impose any restrictions as to which indirect conflicts should be used. The reason
for this is two–fold. First of all, there is no consensus as to which indirect conflicts should be associated with a given
specialized form of support, such as deductive or necessary [49, 17, 54]. Second of all, we are not aware of any
empirical study verifying that people indeed associate with a given type of support the indirect attacks ascribed to it in
theory. There also appears to be no study testing whether a given supporting edge is assigned the same interpretation
both by theory and by actual people. All of these are extremely interesting questions, however, they are beyond the
scope of this particular study.
We would also like to explain our approach to reproducing supports obtained from the participants through defense
in the augmented graph. First of all, both defenses and supports are forms of positive relations between arguments.
Hence, interactions between them are unavoidable, particularly given the results in [49, 17]. Taking into account that
our participants are not argumentation experts and their task was to mark positive relations in the dialogue, what they
have identified could fall into either support or defense category. This will always be the case in any experiment that
takes into account people from all walks of life, not just specialists in our area, and we cannot discard parts of their
answers and extract as supports only those relations that in our opinion are “real”. What we therefore show in Sections
6.1.1.2 and 6.1.2.2 is that even though many of the positive links identified by the participants can be reproduced
through defense, there are support edges that cannot be explained in this way.
The expanded graph contains an auxiliary statement P that embodies the additional reasons for and against the
statements in the dialogues that the participants may have given in their explanations and that have not appeared in the
dialogue. For example, P could contain a statement such as “Masks are insufficient to prevent the flu, gloves are also
needed.” or “The vaccine does cause flu, everyone I know gets very sick after the shot.”.
In this section we will perform two types of analyses. We first consider the intended graphs created for both of the
dialogues (Sections 6.1.1.1 and 6.1.2.1) and see how they are reflected by the graphs obtained from the participants.
Based on this information we create the notions of the total and core samples, used throughout the rest of this paper.
We then focus on looking at the declared and common graphs in Sections 6.1.1.2 and 6.1.2.2. In particular, we explain
up to what extent the common graphs can be reproduced from the intended graphs by the use of indirect defenses and
attacks from the prudent/careful and bipolar setting (see Definitions 2.3 and 3.2).
6.1.1 Dialogue 1
6.1.1.1 Intended Graph Analysis
The minimal graphs we had wanted the participants to acknowledge at every stage of Dialogue 1 are presented in
Figure 6. The first thing we would like to analyze is how the intended graphs are related to the ones declared by the
20
participants.
In the first chart in Figure 7 we can see for how many participants the intended graph was a subgraph of a given
type of their declared graph. The portion of the participants who have correctly recognized the intended relations at a
given step of the dialogue does not exceed 50%, with the worst performance of 30% at the final step of the dialogue.
The results look somewhat more interesting once we look at the confusion subgraph, which - as the name suggests
- assumes the participants may become confused and specify the link as simply dependent rather than attacking or
supporting. The biggest difference in the number of participants between the correct and confusion subgraphs can be
seen in the last step of the dialogue, which might point to participants’ exhaustion or to the complexity of the pieces
of information presented in the dialogue. Finally, we can consider looking at the lenient subgraphs, in which it only
matters that the statements we have seen as related are also seen as such by the participants, independently of the
nature of this relation. The high number of participants whose declared graphs contained the intended graph in this
lenient approach tells us that the participants had no major problems in recognizing the related statements, however,
marking the type of this connection has caused some issues. All of the declared graphs that have not met the leniency
requirements missed at least one of the (C, B), (E, B) or (F, E) attacks. Nevertheless, in all of the approaches we
can see that the lowest number of participants satisfying a given subgraph relation can be seen in the last two steps of
the dialogue in which statements E and F were presented.
A
A
A
A
A
B
B
B
B
B
C
C
C
C
D
D
D
E
E
F
(a) Step 1
(b) Step 2
(c) Step 3
(d) Step 4
(e) Step 5
Figure 6: The intended argument graphs for Dialogue 1. Solid edges represent the attack relation.
Although the first chart in Figure 7 presents how many participants recognized the intended graphs at a given step
in the dialogue, it does not show how many participants recognized the intended graphs at a given number of steps.
For example, even though 18 (resp. 20) participants recognized the intended graph correctly at the first step (resp.
second step) of the dialogue, it does not necessarily mean that 18 participants recognized the intended graph correctly
at both of these steps. Consequently, in the second chart in Figure 7 we present at how many steps the participants
have recognized the intended graph in a particular way (correct, confusion, lenient). The lenient requirements are quite
easily satisfied, just like in the previous case. However, we can observe that a relatively small number of participants
(7 w.r.t. the correct and 12 w.r.t. the confusion subgraph) recognized the intended graph in all dialogue stages. We
believe this indicates that the relation task might have been more confusing than we had anticipated. Consequently, we
have decided to report the results from all of the participants – we will refer to them as the total sample from now on
– as well as from those that have recognized the intended framework in at least four out of five dialogue stages w.r.t.
the confusion subgraph approach – we will refer to them as the core sample. The samples contain respectively 40 and
21
16 participants.
Correct subgraph
Confusion subgraph
Lenient subgraph
Number of Participants
Number of participants whose declared
graphs contain the intended graphs
at a given step
Number of participants whose declared
graphs contain the intended graphs
in a given number of steps
40
30
20
10
36
40
22
39
39
25
23
20
18
36
36
19
19
15
12
9
20
14
5 5
12
1
0
0
1
2
3
4
Step Number
8 7
5
0
1
3
3
0
5 4
0
2
3
Number of Steps
7
0
4
5
Figure 7: Analysis of intended graph containment in Dialogue 1
6.1.1.2 Participant–Sourced Graphs Analysis
In this section we would like to take a closer look at the graphs declared by the participants throughout the first
dialogue. Although we have already analyzed whether the participants have recognized the intended graph, we have
not yet seen what relations - in addition to the ones in Figure 6 – they have also observed. We can observe that at
every step of the dialogue, the number of unique declared graphs we have obtained is smaller than the number of
participants (auxiliary data appendix is available at [51]). This is a natural effect of the fact that different participants
could answer the relation tasks in the same manner at some of the dialogue stages. Nevertheless, the further we are in
the dialogue, the more common the graphs that have only been declared by one participant become. As seen in Figure
8, also the bigger the chances that the framework declared by a participant will not be clarified (i.e. it will contain
dependent links, see Definition 6.2), which was to be expected given the increasing number of edges at every stage.
Although at every step we can find a graph that has been declared by the largest number of participants, the obtained
structure rarely accounts for more than 50% of the (total) sample. Thus, not every participant perceives the dialogue
in the same manner, in part because they do not appear to be always aware of the same facts and arguments (we will
discuss this issue further in Section 6.4). This disparity in knowledge and perception lends support to the use of the
constellation approach in opponent modelling. In particular, if we treat the portion of participants declaring a given
graph as a probability, what we obtain at every stage of the dialogue is actually a subgraph distribution.
As we have stated before, at every stage of the dialogue we can find a framework that has been declared by the
largest number of participants. We present these structures in Figure 9. Only in the first step of the dialogue we obtain
2 equally common graphs, one in which (B, A) edge is seen as attacking and one in which it is supporting (both are
declared by 18 participants). These proportions change notably once we look at the core sample, where the (B, A) link
is primarily attacking. We believe this behaviour can be put down to the initial confusion caused by the experiment,
particularly that afterwards the edge is seen mostly as attacking by the participants from both samples.
We can observe that w.r.t. the intended graphs, the participants have declared significantly more relations in the
common graphs. However, as we will see, many of these edges can in fact be explained by the existing notions of
indirect relations. We will first look at the common graphs from the perspective of the prudent/careful semantics
(Definition 2.3). In other words, we perform a conflict–centered analysis, where the positive edges between the
statements are seen only as a result of the interplay between the attacks. By extending the intended graphs with
additional conflicts stemming from the indirect attacks and additional supports corresponding to the indirect defenses,
22
80
% of Declared Graphs
67.5%
62.5%
60
Total
sample
Core
sample
42.5%
40
20
31.25%
18.75%
31.25%
25%
20% 18.75%
10%
0
1
2
3
Step Number
4
5
Figure 8: Portion of unclarified graphs (i.e. with nonempty set of dependencies) obtained in Dialogue 1
A
A
A
A
A
B
B
B
B
B
C
C
C
C
D
D
D
E
E
A
B
(a) Step 1
F
(b) Step 2
(c) Step 3
(d) Step 4
(e) Step 5
Figure 9: The common argument graphs for Dialogue 1 based on the total sample. With the exception of the bottom
framework in Step 1, they are also common in the core sample. The thicker edges represent the relations appearing in
the intended graph. Solid edges stand for attack and dashed for support links.
we obtain the augmented graphs visible in Figure 10. The augmented graphs obtained in the first three steps are
identical with the common frameworks declared by the participants. The differences between them start appearing
once the statement E is introduced in the fourth step of the dialogue. In particular, in the augmented framework, the
(E, D) attack and (E, C) support are unaccounted for. Once F is presented in the fifth step of the dialogue, also the
(F, D) support and (F, C) attack are missing.
This situation could be interpreted as a sign that the intended framework is too conservative and that the missing
edges should have been included in it in the first place. However, there are also other possible interpretations that do not
resort to modifying the intended graph. According to Definition 2.3, E is controversial w.r.t. D as it is an attacker of B,
which is defended by D. Similarly, F is controversial w.r.t. C, as F attacks A which is defended by C. Consequently,
it might be the case that the (E, D) and (F, C) attacks are manifestations of these issues 6 . The missing (E, C) and
6 Please
note that controversy could also lead the D attacking E and C attacking F , but such relations could not have been defined by the
23
(F, D) supports could then be explained in two ways. If we were to use the aforementioned controversy as a basis
for additional conflicts and add them to the (intended or augmented) graph, the absent supports could potentially be
created by allowing indirect defense to also take into account these new conflicts in their definition. This could perhaps
give rise to “controversial defense”, the existence of which we are not aware of in the literature. Another possible,
and perhaps more natural approach, is to consider the statements that indirectly attack (defend) the same statement as
positively related. In this respect, we can see E and C as working towards the same goal, which is defending A and
attacking B. Nevertheless, we are not aware of this aspect being explored in the context of attack–based graphs.
Let us now look again at the common graphs, but from the perspective of bipolar argumentation. Like in the previous case, we extend the intended graph with support relations corresponding to the indirect defenses from Definition
2.3. However, instead of extending the graph with the indirect conflicts in the prudent/careful sense, we will use the
ones offered by bipolar argumentation. The obtained augmented graphs are visible in Figure 11. By adding the secondary or supported attacks (in this particular case, they overlap) to the intended graphs, we recreate the augmented
structures visible in Figure 10. By considering the (super) mediated attacks that adhere to the flow of the dialogue (i.e.
the source of the attack has to appear later than the target), we can create the (E, D) and (F, C) attacks, which were
missing in the previous approach. Consequently, we are now only missing the (E, C) and (F, D) supports. This issue
can be addressed in two ways. One method is to repeat the process of adding defense–based support to the graph. This
turns E into a supporter of C (it provides defense against D) and F into a supporter of D (it provides defense against
E). Another approach is to consider (E, C) as a supporting link that should have been included in the intended graph
in the first place. In such a situation, the (F, C) conflict can be reproduced by a secondary attack, while the (F, D)
support would still require a repetition of the step in which we add the defense–based support.
We can observe that both approaches offer a way of reconstructing the common graphs (with the exception of
the ({A}, ∅, {(B, A)}, ∅) at step 1) from the intended one. The purely conflict–based approach would require the
introduction of certain auxiliary notions. The bipolar approach uses methods already available in the literature, but in
turn requires some of the procedures to be repeated.
A
A
A
A
A
B
B
B
B
B
C
C
C
C
D
D
D
E
E
F
(a) Step 1
(b) Step 2
(c) Step 3
(d) Step 4
(e) Step 5
Figure 10: The intended argument graphs for Dialogue 1 extended with indirect conflicts and defenses from Definition
2.3. Solid edges represent the attack relation and dashed edges represent the defense–generated support relation. Thick
edges represent the relations appearing in the intended graphs.
In addition to the presented graphs, in Table 6 we can see how often a given relation was declared as attacking,
participants due to the experiment set up.
24
A
A
A
A
A
B
B
B
B
B
C
C
C
C
D
D
D
E
E
F
(a) Step 1
(b) Step 2
(c) Step 3
(d) Step 4
(e) Step 5
Figure 11: The intended argument graphs for Dialogue 1 extended with support coming from indirect defenses from
Definition 2.3, secondary/supported attacks and super–mediated attacks following the dialogue flow. Dash dotted gray
edges represent the additional supports obtainable if we again add defense–based support. Solid edges represent the
attack relation and dashed edges represent the support relation. Thick edges represent the relations appearing in the
intended graphs.
supporting, dependent or nonexistent (in [51] we can find the same data w.r.t the dialogue steps). There are various
differences between how a given relation was perceived by the participants belonging to the core and total samples.
However, we can observe that for a given link, the most commonly associated type by one sample is the same as in
the other. The only exceptions are the (F, A) and (F, C) links, which were considered primarily attacking in the core
sample and dependent in the total sample. There appear to be more differences in what is the second and the third
most common type. For example, a relation primarily seen as attacking and secondly as supporting in the total sample
is more likely to be seen as primarily attacking and secondly dependent in the core one. However, we believe these
differences are an effect of the method we have used to select the core sample and are most likely not indicative of
any particular behaviour patterns. Finally, we can note that if we extracted graphs based on the data found in Table 6
for the core sample, we would recreate the common graphs for that sample. In the total case, we would not obtain the
additional support graph from step 1, and due to the issue with the (F, A) and (F, C) links, the common graphs would
be precision subgraphs of the structures extracted from the table.
In [51] we have also calculated the average distance from one framework to other declared frameworks obtained
from the participants, understood as the average of all distances from a given structure to the remaining 39 ones in the
total sample and 15 in the core one, i.e.:
Definition 6.5. Let n be the number of participants and let {G1 , . . . , Gn } be the frameworks they have declared. The
average distance from a framework Gi , where 1 ≤ i ≤ n, to other frameworks is defined as:
Pn
j=1 d(Gj , Gi )
(1)
avg dist(Gi ) =
n−1
The results for the total sample can be seen in Figure 12, while the overall summary of our findings is presented
in Table 7. In this table we present the average distance from the common framework at a given step to all of the
other frameworks as well as the general analysis of all of the obtained average distances. We provide the minimum,
maximum and median of the obtained values, and the overall average distance (i.e. the average of the averages).
25
Relation
(B, A)
(C, A)
(C, B)
(D, A)
(D, B)
(D, C)
(E, A)
(E, B)
(E, C)
(E, D)
(F, A)
(F, B)
(F, C)
(F, D)
(F, E)
in R
85
12.50
89.06
81.25
2.08
91.67
21.88
81.25
12.50
68.75
50
12.50
62.50
0
81.25
Core Sample
in S
in D
1.25 13.75
84.38 3.13
3.13
7.81
6.25 12.50
97.92
0
0
8.33
62.50 15.63
3.13 15.63
71.88 15.63
3.13 28.13
6.25 43.75
68.75 18.75
6.25 31.25
68.75 31.25
0
18.75
N/A
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
in R
60.50
29.38
68.75
53.33
10
66.67
30
53.75
28.75
53.75
27.50
10
30
0
52.50
Total Sample
in S
in D
31.50
8
63.75 5.63
20
7.50
30.83
15
84.17 5.83
25
8.33
45
16.25
16.25 21.25
43.75
20
10
27.50
30
40
62.50
25
27.50 37.50
55
37.50
22.50
20
N/A
0
1.25
3.75
0.83
0
0
8.75
8.75
7.5
8.75
2.5
2.5
5
7.5
5
Table 6: Occurrences of the declared relations in Dialogue 1 (values are expressed as %)
We can observe that the further we are in the dialogue, the higher these values become, which is natural given the
increasing size of the frameworks. The average distance calculated for the common framework tends to be smaller
than the overall average and less or equal to the obtained median. In the core sample, all of the averages for the
common frameworks are also the minimum ones from all of the obtained values. We can therefore observe that the
fact that this framework is defined by the largest number of participants and is close to or identical with the framework
induced by Table 6 is reflected in the obtained distances between the structures declared by the participants.
Minimum
Maximum
Average
Median
Common
Framework
Minimum
Maximum
Average
Median
Step 1
Step 2
Step 3
Step 4
Step 5
Core Sample
Common
Framework
Total Sample
1.03
1.67
3.44
6.74
11.33
0.92
1.67
3.44
6.74
10.62
1.03
4.54
7.33
13.36
18.77
1.02
2.40
4.75
8.78
13.40
1.03
1.92
4.23
8.39
12.62
0.33
0.40
1.33
2.07
5.13
0.33
0.40
1.33
2.07
5.13
1.80
3.87
7.60
11.93
13.13
0.52
0.72
2.22
3.44
7.17
0.33
0.40
1.33
2.07
6.87
Table 7: Analysis of the average distance values for the declared frameworks for Dialogue 1, where average distance
is formulated in Definition 6.5. We include the average distance from the common framework to other frameworks,
minimum and maximum average distances amongst all the averages, median, and overall average.
26
1
0.5
Average Distance
Average Distance
Average Distance
Average Distance
Average Distance
Step 1
0
0
5
10
15
20
Step 2
25
30
35
40
0
5
10
15
20
Step 3
25
30
35
40
0
5
10
15
20
Step 4
25
30
35
40
0
5
10
15
20
Step 5
25
30
35
40
0
5
10
15
25
30
35
40
4
2
0
8
6
4
2
0
15
10
5
0
20
10
0
20
Figure 12: The list of average distances from the graph declared by a given participant to the graphs of the remaining
participants at a given step in Dialogue 1.
27
6.1.2 Dialogue 2
6.1.2.1 Intended Graph Analysis
The graphs depicting the minimal set of relations we had intended the participants to recognize in the second
dialogue are presented in Figure 13. We will now discuss the reason why the (G, F ) edge is grayed out. Argument
G was meant to be a counterargument for E and F . In the first case the contradiction is more obvious. In the
second instance, the fact that the virus is accompanied only by stabilizers and antibiotics means it is not accompanied
by thimerosal, which is only a preservative. Thus, the contradiction depends on the distinction between stabilizers
and preservatives. Based on the explanations provided by some of the participants we can however observe that
these two notions were occasionally confused, which additional research has shown to be a common situation in
reality. Consequently, thimerosal was seen as an example of a stabilizer and as a result, G could have been in fact
understood as supporting F rather than attacking it. Hence, declaring this relation differently was a conscious and
somewhat justified decision, not a result of misunderstanding the exercise or an unintentional choice (a “misclick”).
Nevertheless, removing this link from the intended graph does not significantly affect the core sample and would allow
one more person in w.r.t. the original approach. This indicates that the majority of people marking (G, F ) differently
than intended have also had other issues that prevented them from entering the core sample. Thus, this issue deserves
attention on its own, and for now we will evaluate the answers without modifying the initially set methodology.
A
A
A
A
A
B
B
B
B
B
C
C
C
C
D
D
D
D
E
E
E
F
F
F
G
G
H
(a) Step 1
(b) Step 2
(c) Step 3
(d) Step 4
(e) Step 5
Figure 13: The intended argument graphs for Dialogue 2. Solid edges represent the attack relation. The crossed gray
edges represent attacks that some participants have interpreted as supports in a way that can be considered justified.
Similarly as in the case of Dialogue 1, in Figure 14 we present how many participants recognized the intended
graph at a given step of the dialogue, and how many participants recognized the intended graph in a given number
of steps. We create the core sample for Dialogue 2 by gathering participants who have acknowledged the intended
graph w.r.t. the confusion subgraph relation in at least 4 stages. Hence, from now on we will be working with the total
28
sample of size 40 and the core sample of size 15.
Correct subgraph
Confusion subgraph
Lenient subgraph
Number of Participants
Number of participants whose declared
graphs contain the intended graphs
at a given step
Number of participants whose declared
graphs contain the intended graphs
in a given number of steps
40
30
36
40
40
26
27
22
24
20
10
38
38
19
20
14
14
36
11
16
0
0
1
2
3
4
Step Number
9 7
8
10
5
0
4 2
0
1
6 8
2
8
6 7
2
0
2
3
Number of Steps
4
4
5
Figure 14: Analysis of intended graph containment in Dialogue 2
6.1.2.2 Participant–Sourced Graphs Analysis
Let us now look at the graphs given to us by the participants of Dialogue 2 (full data can be found in the data
appendix in [51]). Just like in the previous dialogue, the further we are in the discussion, the more single–participant
graphs we obtain. However, in this case, the maximal dispersion occurs during the fourth rather than fifth step, in
which the participants tend to agree a bit more again. As the dialogue progresses, the number of clarified graphs
decreases, as seen in Figure 15. Although there are some differences between the decrease in this dialogue and in
Dialogue 1, particularly concerning the second stage, the end states are not that different. It therefore appears that
participants in both samples had similar issues in deciding on the nature of some of the edges.
In Figure 16 we have depicted the common graphs at every stage of the dialogue, where thick edges represent the
relations contained in the intended graph. Just like in Dialogue 1, at the very first step the participants are asked only
about the (B, A) relation. Although more people have seen the (B, A) relation as attacking rather than supporting (22
vs 14 in the total sample), both groups are important enough to be reported. Many people who have marked (B, A) as
supporting did not make it to the core sample, in which the attacking interpretation is dominant.
Similarly as we had done in the case of the first dialogue, we will now analyze whether the relations contained in
the common graphs, but not in the intended ones, could possibly be explained with additional notions such as indirect
attacks and defenses. Let us first consider creating the augmented graphs with the use of the indirect attacks and
defenses from the prudent/careful setting (see Definition 2.3). The results are visible in Figure 17. Starting from step
2, there are differences between the common and the augmented frameworks. They are primarily caused by C and F
being initially disconnected from other statements. In total, we are missing eight edges – four attacks (C, B), (F, A),
(F, C) and (F, D), and four supports – (C, A), (D, C), (F, B) and (F, E). We will now try to explain them following
the conflict–centered perspective.
Let us start with the (C, B) conflict and the (D, C) and (C, A) supports. Statement C simply presents a “fact” that
vaccine contains a particular substance, which in this case is a mercury compound. As such, it does not contradict any
previous information. However, it becomes conflicting once it is paired with the information that mercury compounds
are poisonous, contained in statement D. This supporting (D, C) link possibly reflects this pairing and makes the
(C, B) attack perfectly reasonable. One can therefore argue that either (C, B) or both (C, B) and (D, C) should have
been included in the intended graph. If only (C, B) was added to the graph, we could potentially reproduce the (D, C)
support by treating statements working towards the same goal as positively related, similarly as proposed in the case
29
% of Declared Graphs
80
70%
60
46.67%
37.5%
40
20
66.67%
60%
10%
46.67%
50%
Total
sample
Core
sample
37.5%
13.33%
0
1
2
3
Step Number
4
5
Figure 15: Portion of unclarified graphs (i.e. with nonempty set of dependencies) obtained in Dialogue 2
of (E, C) and (F, D) supports in Dialogue 1. Adding only (C, B) or both (C, B) and (D, C) would also lead to the
introduction of the (C, A) support through indirect defense.
A similar analysis could be carried out for the relations associated with F . Statement F serves as a further backing
for E, which was recognized as support by the participants. Consequently, including the (F, E) support and (F, C) and
(F, D) attacks in the intended graph (or, based on the aforementioned explanation, just the (F, C) and (F, D) attacks)
would lead to the reproduction of the remaining missing relations through the use of indirect attack or defense.
Let us now look at the intended graphs augmented with supports corresponding to the indirect defenses from the
prudent/careful setting and with indirect attacks associated with the bipolar setting. By using secondary/supported
attacks and defense–induced support, we obtain the augmented frameworks identical to the ones present in Figure
17. Thus, again we are missing the same set of relations as in the conflict–centered analysis: four attacks – (C, B),
(F, A), (F, C) and (F, D) – and four supports – (C, A), (D, C), (F, B) and (F, E). In this case, the (super) mediated
attacks do not offer any additional insight. By considering (super) extended attacks, we can retrieve the (C, B) attack,
though only starting from the fourth step of the dialogue. In the same fashion we can also obtain the (F, A), (F, C)
and (F, D) conflicts at the last stage of the discussion. Nevertheless, these changes occur later than the arguments
are introduced and our main issue is not resolved. In the conflict–centered analysis, we had proposed adding certain
attacks (or both attacks and supports) to the intended graph in order to bridge the gap between the common graph and
the augmented graph. In this, more support–centered, analysis, we can consider adding the (D, C) and (F, E) supports
to the intended graph. Adding the (D, C) support would allow us to recreate the (C, B) conflict as an extended attack.
The inclusion of the (F, E) support would lead to the recreation of the missing (F, C) and (F, D) attacks as supported
attacks in the augmented graph.
We are thus left with the recreation of the (F, A) attack and the (C, A) and (F, B) supports. In order to do so, we
can repeat the procedure of adding indirect defenses and conflicts, i.e. we extend the augmented graph in the way we
extended the intended one. Then F becomes a supported and secondary attacker of A through the secondary/supported
attack (E, A) or the supported attack (F, D). Additionally, F is now an indirect defender of B through the supported
attack (F, D). Finally, (C, A) can be recreated through indirect defense generated by the the extended conflict (C, B).
We can observe that although certain supports can be recreated through indirect defense, this is not the only possible
reason why the participants have declared these relations. For example, we can also observe the similarity between
statement A and the conclusion of D and the fact that G is a more detailed reiteration of C. If we view (F, E) support
as a relation that should have been in the intended graph, then given the fact that E is a defense–mimicking supporter
of B, the (F, B) support could be seen as a particular mixture stemming from the support transitivity. Consequently,
it is possible to be recreate some of the missing supports using a different methodology.
To summarize, we can observe that recreation of the common graph from the intended one in Dialogue 2 requires
the addition of at least 2 relations carried out by C and F . However, just like in Dialogue 1, we would still need to
introduce auxiliary notions when recreating the common graphs via the prudent/careful approach only and perform a
two–step recreation via the bipolar approach.
30
A
A
A
B
B
B
C
C
D
D
A
E
B
F
(a) Step 1
(b) Step 2
(c) Step 3
A
A
B
B
C
C
D
D
E
E
F
F
G
G
H
(d) Step 4
(e) Step 5
Figure 16: The common argument graphs for Dialogue 2 based on the total sample. With the exception of the bottom
framework in Step 1, they are also common in the core sample. The thicker edges represent the relations appearing in
the intended graph. Solid edges stand for attack and dashed for support links.
31
A
A
A
B
B
B
C
C
D
D
E
F
(a) Step 1
(b) Step 2
(c) Step 3
A
A
B
B
C
C
D
D
E
E
F
F
G
G
H
(d) Step 4
(e) Step 5
Figure 17: The intended argument graphs for Dialogue 2 extended with indirect conflicts and defenses from Definition
2.3. The thicker edges represent the relations appearing in the intended graph. Solid edges stand for attack and dashed
ones for defense–generated support.
32
In addition to the graph reproduction analysis, in Table 8 we list how often a given relation was perceived as attacking, supporting, dependent or nonexistent during the dialogue (more detailed tables can be found in [51]). Similarly
as in the case of Dialogue 1, we can observe that the dominating assignment for a given edge is the same in both core
and total samples, though the second and third most common choices often differ. By focusing on the most common
edges, we can reconstruct the graphs from Figure 16.
Relation
(B, A)
(C, A)
(C, B)
(D, A)
(D, B)
(D, C)
(E, A)
(E, B)
(E, C)
(E, D)
(F, A)
(F, B)
(F, C)
(F, D)
(F, E)
(G, A)
(G, B)
(G, C)
(G, D)
(G, E)
(G, F)
(H, A)
(H, B)
(H, C)
(H, D)
(H, E)
(H, F)
(H, G)
in R
85.33
13.33
75
11.67
85
11.67
73.33
8.89
91.11
84.44
73.33
8.89
82.22
82.22
2.22
16.67
60
10
16.67
86.67
80
73.33
13.33
86.67
73.33
0
0
80
Core Sample
in S
in D
2.67
12
73.33 13.33
3.33 21.67
81.67 6.67
3.33 11.67
60
25
8.89 17.78
86.67 4.44
0
8.89
0
15.56
11.11 15.56
84.44 6.67
0
17.78
0
17.78
88.89 8.89
63.33
20
10
30
83.33 6.67
70
13.33
3.33
10
6.67 13.33
0
26.67
73.33 13.33
0
13.33
6.67
20
93.33 6.67
73.33 26.67
6.67 13.33
N/A
0
0
0
0
0
3.33
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
in R
66.50
27.50
68.13
27.50
76.25
27.50
58.33
18.33
65
61.67
58.33
13.33
53.33
55.83
11.67
23.75
61.25
25
21.25
70
63.75
45
12.50
60
52.50
5
7.50
50
Total Sample
in S
in D
26.50
7
58.75 13.13
12.50 16.25
65.63 6.88
12.50 9.38
51.88 16.25
26.67 14.17
73.33 8.33
17.50 12.50
21.67 13.33
29.17 11.67
75.83
10
30.83 12.50
29.17 12.50
75.83 9.17
58.75 16.25
16.25 21.25
63.75 8.75
65
11.25
16.25
10
18.75
15
25
27.50
67.50 17.50
20
12.50
22.50 22.50
77.50 12.50
72.50
15
32.50
10
N/A
0
0.63
3.13
0
1.88
4.38
0.83
0
5
3.33
0.83
0.83
3.33
2.50
3.33
1.25
1.25
2.50
2.50
3.75
2.50
2.50
2.50
7.50
2.50
5
5
7.50
Table 8: Occurrences of the declared relations in Dialogue 2 (expressed as %)
In [51] we have also calculated the average distance from one framework to other declared frameworks obtained
from the participants, understood as the average of all distances from a given structure to the remaining 39 ones in the
total sample and 14 in the core one (see also Definition 6.5). The results in the case of the total sample are visible in
Figure 18, with the overall summary of our findings visible in Table 9. In this table we present the average distance
from the common framework at a given step to all of the other frameworks as well as the general analysis of all of the
obtained average distances. We provide the minimum, maximum and median of the obtained values, and the overall
average distance (i.e. the average of the averages).
The same observations as in the case of Dialogue 1 can be made for Dialogue 2. The average distance calculated
for the common framework tends to be smaller than the overall average and less or equal to the obtained median. In
the case of the core sample, all of the averages for the common frameworks are also the minimum ones from all of the
obtained values. This is also true in the case of the total sample, with the exception of the third step of the dialogue. The
fact that the common framework is defined by the largest number of participants and is identical with the framework
induced by Table 8 is reflected in the obtained distances between the structures declared by the participants.
33
Minimum
Maximum
Average
Median
Common
Framework
Minimum
Maximum
Average
Median
Step 1
Step 2
Step 3
Step 4
Step 5
Core Sample
Common
Framework
Total Sample
0.82
3.38
9.82
13.28
16.26
0.82
3.38
9.77
13.28
16.26
1.23
7.74
18.08
25.49
37.13
0.97
4.74
13.07
17.91
22.59
0.82
4.21
13.41
16.87
22.29
0.29
1.93
5.00
6.64
7.21
0.29
1.93
5.00
6.64
7.21
1.86
8.14
15.57
21.64
25.64
0.48
2.95
7.70
10.15
11.43
0.29
2.00
5.79
8.29
8.36
Table 9: Analysis of the average distance values for the declared frameworks for Dialogue 2, where average distance
is formulated in Definition 6.5. We include the average distance from the common framework to other frameworks,
minimum and maximum average distances amongst all the averages, median, and overall average.
34
1
0.5
Average Distance
Average Distance
Average Distance
Average Distance
Average Distance
Step 1
0
0
5
10
15
20
Step 2
25
30
35
40
0
5
10
15
20
Step 3
25
30
35
40
0
5
10
15
20
Step 4
25
30
35
40
0
5
10
15
20
Step 5
25
30
35
40
0
5
10
15
25
30
35
40
8
6
4
2
0
15
10
5
0
20
10
0
40
30
20
10
0
20
Figure 18: The list of average distances from the graph declared by a given participant to the graphs of the remaining
participants at a given step in Dialogue 2.
35
6.2 Postulate Satisfaction
In this section we analyze if and how the postulates we recalled in Section 4.1 are reflected in the behaviours of
the participants. In order to do so, we will evaluate the probability distributions extracted from the agreement tasks
against the intended, expanded, declared and common graphs (see the end of the introduction of Section 6 on page
19 concerning how these distributions were obtained). Given the fact that the postulates are defined only for Dung’s
frameworks, we need to consider the attack subgraphs of the aforementioned graphs in order to perform the analysis.
Although this means that the frameworks contain information not yet harnessed by the postulates, in Sections 6.1.1.2
and 6.1.2.2 we could have observed that many of the supporting links could be explained with indirect defence or lead
to additional indirect conflicts. Thus, we believe that the effect of the positive relations in the obtained graphs to be
reasonably, even if not completely, reflected by the negative relations. Therefore, the analysis performed only on the
attack graph is still valuable.
The results will be presented in the following manner. For every participant, we calculate the proportion of steps
on which a given postulate was satisfied according to a given graph. For example, if we have five dialogue steps and
the obtained statement probabilities conformed to the rational postulate on four of them, we obtain the adherence rate
of 80% (or 0.8). The values we report are the averages of these scores on the total and core samples. We then analyze
the differences between the results obtained between the graphs and samples. The additional statistical analysis can
be found in Section 9.
6.2.1 Dialogue 1
The results concerning the average adherence rates of the postulates we have recalled in Section 4.1 on the intended,
declared, expanded and common graphs are presented in Figures 20 (total sample) and 21 (core sample). Let us first
look at the total sample. In the majority of cases, the results obtained on the intended and common graphs do not differ
significantly, which can be seen as supporting our graph augmentation approach from Section 6.1.1.2 (see also Table 10
in Section 9.1.1). All of the results appear to be closely related on postulates that, in general, have a high satisfaction
rate or belong to the preferential family. Thus, the adherence rates between the graphs are closely connected for
postulates such as preferential, strict, protective and restrained, coherent, involutary and semi–optimistic. Since the
ternary and binary postulates do not depend on the structure of the graph in question, they do not contribute to this
particular observation. Nevertheless, once the satisfaction rate goes down and we look at the valued or explanatory
postulates, the results begin to diverge and it is mostly the intended and common graphs that still behave similarly (see
Table 10 in Section 9.1.1).
We can thus observe good adherence to the postulates and a similarity between the results on all of the graphs
in the case the preferential branch of our epistemic postulates. The obtained satisfaction rates slowly decrease as
the more specialized versions are considered, reaching the minimum on the justifiable postulate. Low adherence to
this property is most likely caused by the optimistic property, which is also rarely satisfied. In principle, postulates
considered shared between preferential and value based families do not perform as well as the ones classified as
only preferential. We would also like to observe that the preferential postulate is highly satisfied not just due to its
“leniency”, but also because it reflects the reasoning of various participants. For example, some of the participants
satisfying the preferential, but not the rational postulate in the first step of the dialogue, justified their agreement with
both of the statements by saying that e.g. “A: I agree that Hospital staff members don’t need to receive flu shots unless
they want to. They should not be required to. ” and “B: It is also true that hospital workers are exposed to the flu a lot
so if they choose to it would be beneficial for them to receive the shot.”, or “It’s been shown, I think, that mysteriously,
nurses and doctors are less prone to catching the diseases they treat in general.” and “Extra precautions couldn’t hurt
during a particularly nasty flu season.”. To summarize, these results mean that participants are in general capable of
reasoning well when it comes to considering the conflicts between two statements.
Lower satisfaction rate of the postulates and/or bigger dissimilarities between the results on different graphs (see
Table 10 in Section 9.1.1) can be observed particularly in the case of postulates that are highly affected by the personal
knowledge of the participants and by accuracy of modelling the perceived attacks. For example, these are the cases in
which the results for the expanded graph – which is precisely the declared graph augmented with statements extracted
from explanations - differ from the results for the declared graph. What was considered an initial argument may not
have been seen in the same way by the participants, thus reducing the adherence to the optimistic postulate. This, along
36
with the reluctance of the participants to choose the extreme values (see the ternary postulate), may explain the lower
performance of the founded property. In particular, by relaxing the founded restriction of complete belief to the semi–
founded one, which simply requests non–disbelief, we obtain much better results. All of the guarded, discharging,
trusting and anticipating postulates are intended to grasp the idea of believing an argument when there is no reason
against it and disbelieving an argument only if there is a reason for it. Consequently, they are highly dependent on
the accuracy of the graphs we create for the participants in terms of their knowledge, and are as such more difficult
to measure and estimate than the postulates of the preferential family. The fact that the demanding postulate appears
to be the least affected by it might be connected to the ternary postulate – again, if the extreme values are not chosen,
this property is satisfied more easily.
We can observe that the semi–optimistic postulate is commonly satisfied, even twice as often as the coherent
one. This has certain important implications. The purpose of these properties is to set respectively lower and upper
bounds on the belief we have in a statement in the face of its attackers. For example, given the performance of the
protective and strict postulates, it appears that while the participants are capable of choosing what to believe and what
to disbelieve, they can be more “generous” with their belief than anticipated by some of the postulates.
We can now consider the results obtained for the core sample. We can observe that the average satisfaction rate
of the postulates that the participants appeared to adhere to in the total sample increased in the core sample. In other
words, the preferential, rational, strict, protective and restrained postulates are, on average, satisfied more often. The
results obtained between the graphs appear to be much more closely connected than in the total sample (see Table 10 in
Section 9.1.1). Significant differences between the adherence rates can be observed primarily in the case of declared
and expanded graph. They are also bigger than in the case of the total sample, i.e. the average of the differences
between adherence rates of participants between the declared and expanded graphs is bigger in the total sample only
in the case of the involutary and demanding postulates. This may be caused by the participants being more forthcoming
with their personal knowledge and thus the expanded graphs being more precise. It would also explain the somewhat
better results on the explanatory and optimistic, founded and justifiable postulates, which are more dependent on
such factors. Finally, we can consider the binary and ternary postulates. In both samples, the binary postulate has
a high adherence rate, which means that the participants were capable of deciding whether to believe or disbelieve
the presented statements using their own knowledge and the information found in the dialogue. Thus, the text itself
could not have been overly complicated for the participants. The low adherence to the ternary postulate in both of the
samples, combined with the low indecisiveness of the participants, shows that people often exhibit various degrees of
belief and disbelief and are reluctant to completely accept or completely reject a given statement.
In addition to the postulates available in Figures 20 and 21, we have also considered a slight modification of
the VALn property. This postulate tells us how many distinct probabilities have been used in a given probability
distribution. We can recall that from every user we obtain five distributions, one per every step of the dialogue.
However, in some dialogue steps the number of used arguments is not sufficient to make any important observations
w.r.t. the VALn postulate. Consequently, rather than counting the distinct values in a single distribution, we count the
distinct values used in all five distributions defined by a participant 7 . The results are presented in Figure 19 (see also
the auxiliary data appendix at [51] for a more detailed table). We can observe that the vast majority of the participants
did not use 3 or less values during the dialogue, independently of whether we consider the total or the core sample. The
preferred options appear to be 4 and 6 (total sample) or 4 and 5 (core sample) values. These results are significantly
different from what we would expect if the participants were providing us with random answers and show that what
we obtained were in fact conscious choices8 .
The above observations indicate that the two (or three) valued perception of acceptance exhibited by the classical
Dung’s semantics oversimplifies human modelling and might be insufficient to model the beliefs of the majority of
participants. This also confirms the findings of [55]. We may also recall that rational, strict, protective, discharging and
trusting postulates were related to classical conflict–free, admissible and complete semantics. While the performance
of the preferential branch indicates that in general the statements accepted by the participants are not conflicting, the
discharging and trusting results show that admissibility and completeness may be more problematic to satisfy. Thus,
7 In
this analysis, we treat Don’t know the same as Neither Agree nor Disagree due to the fact that they both evaluate to 3/6.
the assumption that every value is equally likely to be selected, if the participants were answering randomly throughout Dialogue 1 then
the probability of them using exactly 7 values throughout the dialogue is 0.704, probability of them using exactly 6 values is 0.272, of using 5
values is 0.024, of using 4 is 0.00048, and of 3 or less is 0.0000015. Exact multinomial test shows there are significant differences between the
obtained and the expected probabilities (obtained p–value was small enough that it was rounded up to 0 by the used software).
8 Under
37
% of Participants
breaking down the classical semantics into smaller, separate properties, might give us a more detailed insight into the
reasoning of the participants and provide more feedback in situations in which classical semantics are not satisfied.
Consequently, more fine–grained approaches than Dung’s semantics should be considered in order to provide a more
accurate human modelling.
43.75%
40
35%
30
20
10
17.5%18.75%
17.5%
12.5%
Total sample
22.5%
Core sample
12.5%
6.25%
5%
2
3
4
5
Number of Values
6
6.5%
2.5%
7
Figure 19: Usage of different levels of agreement in Dialogue 1
38
Preferential
Rational
Strict
Protective
Restrained
Coherent
Involutary
Justifiable
Binary
SemiOptimistic
SemiFounded
Ternary
Founded
Optimistic
Demanding
Guarded
Discharging
Trusting
Anticipating
0%
10%
20%
30%
40%
50%
60%
70%
Beliefs on declared graphs Beliefs on expanded graphs
Beliefs on common graphs Beliefs on intended graphs
Figure 20: Postulate satisfaction on the total sample in Dialogue 1
39
80%
90%
Preferential
Rational
Strict
Protective
Restrained
Coherent
Involutary
Justifiable
Binary
SemiOptimistic
SemiFounded
Ternary
Founded
Optimistic
Demanding
Guarded
Discharging
Trusting
Anticipating
0%
10%
20%
30%
40%
50%
60%
70%
Beliefs on declared graphs Beliefs on expanded graphs
Beliefs on common graphs Beliefs on intended graphs
Figure 21: Postulate satisfaction on core sample Dialogue 1
40
80%
90%
6.2.2 Dialogue 2
In Figures 22 and 23 we can observe the average adherence rates to the epistemic postulates in the total and core
samples in Dialogue 2. Let us focus on the total sample first. Some of the observations we have made in the case
of Dialogue 1 carry over to Dialogue 2. Similarly as in the first dialogue, the results obtained for the declared,
expanded and common graphs appear to be similar in the case of preferential family (see Table 19). However, unlike
in the previous case, the behaviour of the intended graph on this family appears to be more distinct from the rest.
Fortunately, with the exception of the preferential postulates, the similarities between the results on the common and
intended graphs, present in first dialogue, hold here as well. This behaviour may indicate that the intended graphs
were too conservative in terms of the considered attacks w.r.t. the common graphs (see also Section 6.1.2.2). We can
also note that similarly as in the first dialogue, the adherence to the rational family of postulates decreases the more
specialized the property is. Again, the justifiable postulate is barely adhered to.
The differences between the results on the declared graph and other graphs in the case of explanatory and value–
based families are more visible than in the first dialogue (see also Table 10 in Section 9.1.1 and Table 19 in Section
9.2.1). In principle, the largest number of differences can be observed between the intended and declared graphs, and
exceeds our findings in the first dialogue. Additionally, many of the postulates we have considered do not perform as
well as in the first dialogue in terms of the declared graphs. Nevertheless, the rational, semi–optimistic, semi–founded,
guarded, discharging and particularly demanding properties appear to achieve better results than in the previous dialogue. We believe some of this behaviour (including also the lower performance of the strict and restrained postulates)
can be explained with the behaviour of the binary property. The low satisfaction rate of this particular postulate means
that the participants had problems in deciding whether to believe or disbelieve the presented statements using their own
knowledge and the information found in the dialogue. This dialogue presented more specialized information than the
first dialogue and without access to, for example, Wikipedia, the participants had problems verifying the statements.
The presence of the undecided assignments appears to have benefited some of the postulates that are easily satisfied
with neutral distributions. We can observe that all of the properties that have experienced an increase in adherence
belong to this group. At the same time the undecided assignments harmed those postulates that try to force more
decisiveness, such as strict or protective properties, or force belief or disbelief in the face of undecided attackers (e.g.
restrained or anticipating postulates). Consequently, in addition to the postulates vulnerable to participants who do
not share all of their knowledge (see the analysis of Dialogue 1), we can also identify those that are vulnerable or
benefiting from the indecisive participants.
We can now consider the results obtained for the core sample. The average adherence rates of the postulates in the
core sample are quite close to the ones from the total sample. The results obtained for different graphs also appear to
be more closely connected (see Table 19 in Section 9.2.1). The most visible changes can be observed in the case of the
explanatory family of postulates (with the exception of the demanding property), which have all increased their values.
The results w.r.t. the expanded graphs on the guarded and discharging postulates become even more distinct from the
declared graphs in terms of the average difference of participant responses between these graphs. It therefore appears
that the participants were more forthcoming with their personal knowledge, in particular with the additional reasons
against believing certain statements. We therefore again identify the properties more vulnerable to such information.
In Figure 24 we can observe the number of participants who have chosen a given number of distinct probabilities
throughout the dialogue 9 (see also the auxiliary data appendix at [51] for a more detailed table). Similarly to Dialogue
1, the majority of participants choose 4 or more values, though in this case 5 values appear more often in the total
sample than 4 values, and they are picked equally often in the core sample. This increase may be related to a wider
use of the Neither Agree nor Disagree and Don’t know values, which is visible in the difference of the adherence to
the binary postulate between the dialogues. Also in this case it is easy to see that the selection of the values by the
participants is clearly not random10 and thus the tendency of the participants to choose more than 3 values is their
conscious choice.
9 The
presented results treated Don’t know as the same as Neither Agree nor Disagree.
the assumption that every value is equally likely to be selected, if the participants were answering randomly throughout Dialogue 2 then
the probability of them using exactly 7 values throughout the dialogue is 0.876, probability of them using exactly 6 values is 0.121, of using 5
values is 0.003, of using 4 is 0.00002, and of 3 or less is 0.0000000095. Exact multinomial test shows there are significant differences between
the obtained and the expected probabilities (obtained p–value was small enough that it was rounded up to 0 by the used software).
10 Under
41
Preferential
Rational
Strict
Protective
Restrained
Coherent
Involutary
Justifiable
Binary
SemiOptimistic
SemiFounded
Ternary
Founded
Optimistic
Demanding
Guarded
Discharging
Trusting
Anticipating
0%
10%
20%
30%
40%
50%
60%
70%
Beliefs on declared graphs Beliefs on expanded graphs
Beliefs on common graphs Beliefs on intended graphs
Figure 22: Postulate satisfaction on the total sample in Dialogue 2
42
80%
90%
Preferential
Rational
Strict
Protective
Restrained
Coherent
Involutary
Justifiable
Binary
SemiOptimistic
SemiFounded
Ternary
Founded
Optimistic
Demanding
Guarded
Discharging
Trusting
Anticipating
0%
10%
20%
30%
40%
50%
60%
70%
Beliefs on declared graphs Beliefs on expanded graphs
Beliefs on common graphs Beliefs on intended graphs
Figure 23: Postulate satisfaction on core sample in Dialogue 2
43
80%
90%
% of Participants
42.5%
40
33.33%
30
10
Total sample
22.5%
20%
20
33.33%
Core sample
13.33%
5%
6.67%
2
5%
3
4
5
Number of Values
6.67%
6
5%
6.67%
7
Figure 24: Usage of different levels of agreement in Dialogue 2
44
6.3 Statements vs Relations
Relations in argumentation frameworks are often defined in terms of the argument structure we are dealing with.
Consequently, they are seen as secondary to arguments, and as such their acceptability is rarely considered. One
of the exceptions is the attack–based semantics for argumentation frameworks, in which the status assigned to a
relation corresponds to the status assigned to its source [60] and vice versa. The same behaviour is replicated by other
frameworks allowing the attacks to appear in extensions [3, 20] if we limit ourselves to non–recursive graphs. In this
section we would like to verify this view by analyzing the answers to the agreement tasks and relation tasks.
We will perform our analysis in the following manner. For every relation of a given type, by which we understand A
good reason against, A somewhat good reason against, A good reason for, A somewhat good reason for and Somehow
related, but can’t say how, we create a distribution based on the levels of agreement assigned to the statements carrying
them out. Moreover, for every possible level of agreement, we gather the types of relations that the statements assigned
a given level carry out. This allows us to observe whether by knowing the value ascribed to a relation we can guess
the value assigned to its source and vice versa. In particular, this allows us to verify whether the connection between
the values assigned to arguments and relations is as strong as we might think based on some of the research in abstract
argumentation. Therefore, what we would expect to obtain is either 1) relations considered “somewhat good” to be
carried out primarily by the statements that the participants disagree with and “good” relations to be sourced primarily
by the statements that the participants agree with, or 2) relations considered “somewhat good” to be carried out
primarily by the statements that have low levels of agreement or disagreement and relations considered “good” to be
sourced primarily by the statements that have high levels of agreement or disagreement. In both cases we expect the
Somewhat related, but can’t say how edges to be primarily associated with statements of an undecided or unknown
status. Our analysis will be carried out on the original as well as pooled data, i.e. one in which certain categories can
be merged. Both relation types and agreement levels can be grouped by strength or polarity:
• Agreement grouping:
– By strength: we group the agreement levels according to how strongly believed or disbelieved they are. We
distinguish between Strong Belief (created from Strongly Agree and Strongly Disagree), Moderate Belief
(combining Agree and Disagree), Weak Belief (grouping Somewhat Agree and Somewhat Disagree), and
Neither (created from Neither Agree nor Disagree and Don’t Know) categories.
– By polarity: we group the agreement levels according to whether they represent belief or disbelief. We distinguish between Believed (grouping the Somewhat Agree, Agree and Strongly Agree levels), Disbelieved
(created from Somewhat Disagree, Disagree and Strongly Disagree), and Neither (formed from Neither
Agree nor Disagree and Don’t Know) categories.
• Relation grouping:
– By strength: we group relations according to whether they are good or somewhat good reasons. We
distinguish between Strong Relations (formed from A good reason for and A good reason against), Normal
Relations (grouping A somewhat good reason for and A somewhat good reason against), and Dependencies
(representing Somehow related, but can’t say how).
– By polarity: we group relations according to whether they are positive or negative. We distinguish between
Attacks (formed from A good reason against and and A somewhat good reason against), Support (grouping
A good for and A somewhat good reason for), and Dependencies (representing Somehow related, but can’t
say how).
We consider pooling data in order to be able to more clearly investigate certain trends that may become more
apparent with fewer categories.
6.3.1 Dialogue 1
6.3.1.1 Effect of a Relation on its Source
45
In Figure 25 we can observe how the levels of agreements of the sources of relations of a given kind are distributed.
The results show that this agreement distribution is in principle dependent on the nature of the relation in question 11 .
Nevertheless, the distributions visible in Figure 25 do show certain similarities and we carry out additional pairwise–
analysis.
We can observe that the distributions of relations marked as being good reasons are relatively similar, as also
supported by the results in Table 11 in Section 9.1.2. In both cases, independently of whether we choose the core or
total sample, the dominating belief assigned to the source is Strongly Agree, with Agree and Strongly Disagree next in
line.
Although some similarities can also be observed between the relations considered as somewhat good reasons for
or against a given statement, they are more visible in case of the core sample (Table 11 in Section 9.1.2). In the total
sample, the proportions of source agreement levels are in principle distinct. It appears that more participants have
chosen disagreement rather than agreement types of sources, though it does not appear that one particular belief is
substantially more common than the other in the total sample. In the core sample, the Somewhat Disagree option
appears to be chosen slightly more often. No particular pattern appears to hold for the Somewhat related, but can’t say
how relation.
Although none of the agreement distributions is random (see Table 12 in Section 9.1.2), we can observe that the
“weaker” the relation, the more blurred the picture becomes. Consequently, if we were given a relation, informed of
its type and then asked to guess what should the belief in its source be, our best chances of answering correctly are in
the case of edges marked as good reasons.
Let us now look at the results obtained for the pooled data. Given the similarity in the behaviour of the relations
marked as good reasons for and against in both samples and the somewhat good reasons for and against in the core
sample, we will now pool our data according to its strength. We first consider grouping both arguments and relations
according to this criterion. We obtain the distributions visible in Figure 26a. The results show that they depend on
the chosen relation type, both in the overall 12 and pairwise analysis (see Table 13 in Section 9.1.2), and that they are
not random (see Table 14 in Section 9.1.2). We can observe that strong relations are primarily carried out by strong
arguments, i.e. those that the participants strongly agree or disagree with. The slightly weaker relations appear to be
carried out more by moderately believed arguments, however, the obtained result accounts for less than 50% of the
cases. The sources of relations marked as dependencies do not seem to follow any particular pattern. We can only
observe that weakly believed arguments are the least common sources.
Let us now consider pooling arguments by polarity and relations by strength. The obtained distributions are depicted in Figure 26b. We can observe that in this case, strong relations are primarily carried out by believed arguments.
Normal relations tend to be carried out by disbelieved arguments, though only in the core sample such sources account
for more than 50% of the answers. Finally, relations marked as dependencies appear to be carried out by all possible
types of arguments, to the point that the results obtained in the case of the core sample are similar to random (see
Table 14 in Section 9.1.2). Nevertheless, results in Table 13 in Section 9.1.2 show that despite pooling, the nature of
the source is dependent on the type of relation we consider13 .
Let us now briefly consider pooling relations and arguments according to polarity. Given the similarities in the
behaviour of relations marked as good reasons in the original data, it is not surprising that the results obtained for
attacks and supports are similar in both core and total samples (see also Table 13 in Section 9.1.2), despite the fact that
the overall analysis shows that source distributions depend on the chosen relation14 . In both cases, the pooled belief
marked as Neither is the least common option, with Believed being the most common choice with the exception of the
Support relations on the core sample. Again, relations marked as Dependencies do not follow any particular pattern,
even to the point that the answers can be considered random (see Table 14 in Section 9.1.2). This similarity between
attacks and supports persists even if we consider pooling agreement levels according to strength (see Tables 13 and 14
11 G–test for independence yields G–value 317.77 with 28 degrees of freedom and p–value less than 2.2 × 10−16 . This result was obtained using
the library Deducer (likelihood.test function) in R.
12 G–test for independence yields G–value 208.01 with 6 degrees of freedom and p–value less than 2.2 × 10−16 . This result was obtained using
the library Deducer (likelihood.test function) in R.
13 G–test for independence yields G–value 175.33 with 4 degrees of freedom and p–value less than 2.2 × 10−16 . This result was obtained using
the library Deducer (likelihood.test function) in R.
14 G–test for independence yields G–value 62.846 with 4 degrees of freedom and p–value of 7.312 × 10−13 . This result was obtained using the
library Deducer (likelihood.test function) in R.
46
Total sample
Core sample
A good reason against
A good reason for
A somewhat good
reason against
D
SD
DK
So
D
A
So
A
NA
D
SA
SD
D
So
D
DK
A
So
A
NA
D
SA
60%
50%
40%
30%
20%
10%
0%
Somewhat related,
but can’t say how
A somewhat good
reason for
D
SD
So
D
DK
So
A
NA
D
A
SA
SD
D
So
D
DK
So
A
NA
D
A
SA
SD
D
So
D
DK
A
So
A
NA
D
SA
60%
50%
40%
30%
20%
10%
0%
Figure 25: The levels of agreements assigned to the sources of relations of a given type in Dialogue 1. We use
the following abbreviations: Strongly Agree (SA), Agree (A), Somewhat Agree (SoA), Neither Agree nor Disagree
(NAD), Somewhat Disagree (SoD), Disagree (D), Strongly Disagree (SD), Don’t Know (DK)
in Section 9.1.2), again despite overall result indicating that all of the relations are distinct 15 .
To conclude, we can observe that the dependent relations do not appear to favour sources marked as Neither in any
particular way, which does not agree with the models we have recalled at the start of the section. Nevertheless, certain
interesting patterns can be observed for strong and normal relations. In particular, strong relations tend to be carried
out by arguments that are believed (polarity pooling) and strongly believed (strength pooling). Grouping relations by
polarity leads to similar behaviour of attack and support relations, which may mean that future experimental data on
one of them might be generalized to the other.
15 G–test for independence yields G–value 69.849 with 6 degrees of freedom and p–value of 4.391 × 10−13 . This result was obtained using the
library Deducer (likelihood.test function) in R.
47
Total sample
Strong Relation
Core sample
Dependency
Normal Relation
N
WB
MB
SB
N
WB
MB
SB
N
WB
MB
SB
70%
60%
50%
40%
30%
20%
10%
0%
(a) Arguments and relations pooled by strength
Strong Relation
Dependency
Normal Relation
B
N
D
B
N
D
B
N
D
70%
60%
50%
40%
30%
20%
10%
0%
(b) Arguments pooled by polarity and relations pooled by strength
Figure 26: The levels of agreements assigned to the sources of relations of a given type in Dialogue 1 according to
a given type of pooling. We use the following abbreviations: Strong Belief (SB), Normal Belief (NB), Weak Belief
(WB), Neither (N), Disbelief (D), Belief (B).
48
6.3.1.2 Effect of an Argument on the Relations it Carries Out
Let us now consider the dual problem i.e. what relations are carried out by an argument assigned a given agreement
level? In Figure 27 we can find the data extracted from the answers of our participants. Although in principle the
relation distribution does depend on the strength of the argument carrying it out 16 , we perform additional pairwise
analysis. We can observe that while in the previous case the charts of relations with the same status but opposite
polarity shared certain similarities, the grouping in this case is somewhat different. The graphs of Strongly Agree,
Agree and potentially Strongly Disagree do tend to put more focus on the good reason for and against edges, however,
they in general lead to different results (see Table 15 in Section 9.1.3). The Somewhat Agree, Somewhat Disagree
and Disagree statements on the other hand appear to lead more to relations marked as somewhat good reasons for or
against. The similarity in the types of relations these three agreement levels carry out is also reflected by the results
in Table 15 in Section 9.1.3. The Somewhat good reason against and Somehow related, but can’t say how are the
two most common choices when it comes to the statements that the participants neither agreed nor disagreed with, the
latter being the dominant relation resulting from the elements marked as Don’t know. These two levels of agreement
also lead to similar distributions (see Table 15 in Section 9.1.3).
Based on the results in Table 16 in Section 9.1.3, we can also observe that our results are in principle not random.
The results for agreement levels marked as Somewhat in the core sample are not significant, however, given the
distributions in Figure 27, this might be related to the sample size.
Total sample
Strongly Agree
Core sample
Strongly Disagree
Agree
Disagree
Somewhat Agree
Neither Agree
nor Disagree
Somewhat Disagree
++
+
?
-
--
++
+
?
-
--
++
+
?
-
--
++
+
?
-
--
60%
50%
40%
30%
20%
10%
0%
Don’t know
++
+
?
-
--
++
+
?
-
--
++
+
?
-
--
++
+
?
-
--
60%
50%
40%
30%
20%
10%
0%
Figure 27: The types of relations carried out by statements of a given level of agreement in Dialogue 1. We use the
following abbreviations for the relation types: A good reason against (- -), A somewhat good reason against (-), A
good reason for (++), A somewhat good reason for (+), Somewhat related, but can’t say how (?)
In many cases, this initial analysis points to the statements with a particular level of agreement favouring relations
of a corresponding “power”. Thus, we will now consider pooling our results by strength, as visible in Figure 28a.
16 G–test
for independence yields G–value 317.77 with 28 degrees of freedom and p–value less than 2.2 × 10−16 . This result was obtained using
the library Deducer (likelihood.test function) in R.
49
All of the belief strengths affect the associated relation distributions in the total sample17 , although the results for
the moderate and weak belief appear to be more closely related on the core sample (see Table 17 in Section 9.1.3).
Nevertheless, the obtained distributions are not random (see Table 18 in Section 9.1.3) and we can observe certain
interesting patterns emerging. Strong belief leads primarily to strong relations and weak belief leads primarily to
normal relations (obtained results account for 55% to 72% of the answers, depending on the sample). Relations
caused by the arguments with moderate belief are almost evenly split between strong and normal relations. Arguments
that are classified as Neither often carry out dependencies, however, they do not account for 50% of all of the possible
relations.
Let us now consider pooling agreement levels according to belief and relations according to strength. By doing
so, we obtain the distributions visible in Figure 28b. We can observe that believed arguments lead primarily to strong
relations, disbelieved to normal relations, and neither to dependencies. However, clear dominance is visible only in
the case of believed arguments. In all cases, the chosen agreement level leads to a different relation distribution18 that
is not random (see Tables 17 and 18 in Section 9.1.3).
Finally, let us briefly consider agreement levels paired with relations pooled by polarity. Although tests show that
the obtained results are in principle distinct, many similarities become apparent in the pairwise analysis19 . If we group
levels by polarity, we obtain the result that the proportions of the relation types carried out by believed and disbelieved
arguments are not significantly different (see Tables 17 and 18 in Section 9.1.3). This similarity may in part be due
to the connections between the Somewhat Agree, Somewhat Disagree and Disagree agreement levels. It might also
explain the fact that when we consider merging arguments by strength, the similarities between the weak and moderate
or strong levels of agreement become visible (see Tables 17 and 18 in Section 9.1.3).
To conclude, we can observe that the tendencies observable in Figures 28a and 28b seem to agree with what
the models we have previously recalled would assume. However, the domination of the desired relations is not as
significant in all situations as we would expect. Similarly as in the previous relation analysis, the tendencies for
strongly believed (strength pooling) and believed (polarity pooling) arguments are the most visible.
6.3.2 Dialogue 2
6.3.2.1 Effect of a Relation on its Source
In Figure 29 we can observe how the levels of agreements on the sources of relations are distributed. The tests
show that the agreement distributions depend on the type of the relation we consider, both in the overall 20 and pairwise
analysis w.r.t. the total sample and in almost all cases w.r.t. the core sample (see Table 20 in Section 9.2.2). In all
cases, the obtained distributions are not random (Table 21 in Section 9.2.2).
As in the previous dialogue, we can observe certain similarities in the distributions of relations marked as good
reasons for and against, particularly in the case of the core sample (Table 20 in Section 9.2.2). However, it is easy to
see that these results are substantially different from the ones observed in Dialogue 1 (see Figure 25). In this case, the
Agree and Neither Agree nor Disagree options appear to be the two most common choices. Nevertheless, none of them
exceeds 25% of the given answers. The options representing agreement appear to be favoured over those representing
disagreement, which we will take into consideration when analyzing pooled results.
The Agree and Neither Agree nor Disagree answers are also common when it comes to relations marked as A
somewhat good reason against. However, depending on whether we are dealing with the total or core sample, the
Somewhat Agree or Somewhat Disagree value is more frequent. It is worth noticing is that the strongest possible
values – Strongly Agree and Strongly Disagree – are the least common. These values are also rarely chosen in the case
of the A somewhat good reason for relation. For this edge, the Neither Agree nor Disagree level of agreement is the
most common value, both in the total and in the core sample.
17 G–test for independence yields G–value 208.01 with 6 degrees of freedom and p–value less than 2.2 × 10−16 . This result was obtained using
the library Deducer (likelihood.test function) in R.
18 G–test for independence yields G–value 175.33 with 6 degrees of freedom and p–value less than 2.2 × 10−16 . This result was obtained using
the library Deducer (likelihood.test function) in R.
19 When pooling arguments by strength, G–test for independence yields G–value 69.849 with 6 degrees of freedom and p–value of 4.391×10−13 .
When pooling arguments by polarity, G–test for independence yields G–value 62.846 with 4 degrees of freedom and p–value of 7.312 × 10−13 .
This result was obtained using the library Deducer (likelihood.test function) in R.
20 G–test for independence yields G–value 390.02 with 28 degrees of freedom and p–value less than 2.2 × 10−16 . This result was obtained using
the library Deducer (likelihood.test function) in R.
50
Total sample
Strong Belief
Core sample
Moderate Belief
Weak Belief
Neither
D
NR
SR
D
NR
SR
D
NR
SR
D
NR
SR
70%
60%
50%
40%
30%
20%
10%
0%
(a) Arguments and relations pooled by strength
Belief
Neither
Disbelief
D
NR
SR
D
NR
SR
D
NR
SR
70%
60%
50%
40%
30%
20%
10%
0%
(b) Arguments pooled by polarity and relations pooled by strength
Figure 28: The types of relations carried out by statements of a given level of agreement in Dialogue 1 according to
a given pooling. We use the following abbreviations for relation types: Strong Relation (SR),Normal Relation (NR),
Dependency (D).
Finally, we reach the relations marked as Somewhat related, but can’t say how. This is the only case in which
the domination of a single value – Neither Agree nor Disagree – is clearly visible, particularly in the case of the core
sample. Additionally, in the total sample we can observe that the Don’t know option may also be of some importance.
Thus, we can observe that independently of the chosen relation type, the Neither Agree nor Disagree agreement level
of the source appears to be quite common, often with Agree, Somewhat Agree or Somewhat Disagree not far behind.
The observed behaviour of the relations is also notably different from what we have seen in Dialogue 1.
Let us now analyze pooled data. We start by grouping relations by strength, and pool agreement levels first by
strength and afterwards by polarity. We obtain the distributions visible in Figures 30a and 30b. The results show that
the way the sources are distributed depends on the chosen relation type, both in the overall 21 and pairwise analysis
(see Table 22 in Section 9.2.2). Furthermore, the results are also not random (Table 23 in Section 9.2.2). Let us
now look more closely at the results with strength pooling on both agreement levels and relations. We can observe
that unlike in the first dialogue, the distribution associated with the strong relations does not appear to follow any
particular pattern. We can only note that weak belief is the least common type of a source. In a similar fashion, normal
relations are carried out by strongly believed arguments the least often, with the remaining types of sources being
spread relatively evenly. Finally, while in Dialogue 1 relations marked as Dependencies did not appear to follow any
patterns, in Dialogue 2 the dominance of sources grouped as Neither is clear.
The dominance of Neither sources in Dependencies is also quite clear if we look at the results of arguments pooled
by polarity and relations by strength. We can observe that in this analysis, strong relations tend to be carried out by
21 In case both arguments and relations have been pooled by strength, the G–test for independence yields G–value 232.25 with 6 degrees of
freedom and p–value less than 2.2 × 10−16 . In pooling arguments by polarity and relations by strength, we obtain G–value 167.86, 4 degrees of
freedom and again p–value less than 2.2 × 10−16 . This result was obtained using the library Deducer (likelihood.test function) in R.
51
Total sample
Core sample
A good reason against
A good reason for
A somewhat good
reason against
D
SD
DK
So
D
A
So
A
NA
D
SA
SD
D
So
D
DK
A
So
A
NA
D
SA
60%
50%
40%
30%
20%
10%
0%
Somewhat related,
but can’t say how
A somewhat good
reason for
D
SD
So
D
DK
So
A
NA
D
A
SA
SD
D
So
D
DK
So
A
NA
D
A
SA
SD
D
So
D
DK
A
So
A
NA
D
SA
60%
50%
40%
30%
20%
10%
0%
Figure 29: The levels of agreements assigned to sources of relations of a given type in Dialogue 2. We use the following abbreviations: Strongly Agree (SA), Agree (A), Somewhat Agree (SoA), Neither Agree nor Disagree (NAD),
Somewhat Disagree (SoD), Disagree (D), Strongly Disagree (SD), Don’t Know (DK)
believed arguments. Although this repeats the results of Dialogue 1, in this case the dominance is not entirely clear
and believed sources do not account for 50% of the relations. Additionally, unlike in Dialogue 1, the disbelieved
arguments appear to be the least common sources. No particular pattern emerges when we look at normal relations –
we can only observe that the sources marked as Neither appear to be somewhat less common than other types.
Let us now briefly consider pooling relations and arguments according to polarity. Despite the fact that the overall
analysis shows that source distributions depend on the chosen relation22 , pairwise analysis gives us a different result.
Although the similarities in the behaviour of relations marked as good reasons in the original data were only apparent
in the core sample, the pooled results obtained for attacks and supports are similar in both core and total samples
(see also Table 22 in Section 9.2.2). Similarly as in the first dialogue, the Believed level is the most common choice.
However, in this dialogue, the relations marked as Dependencies do follow a pattern and are carried out primarily by
arguments marked as Neither. In all cases, our results are not random (see Table 23 in Section 9.2.2). Let us now
pool agreement levels according to strength. The overall test shows that the source distribution depends on the chosen
relation 23 . However, a detailed analysis shows that agreement distributions associated with attacks and supports still
22 G–test for independence yields G–value 132.84 with 4 degrees of freedom and p–value less than 2.2 × 10−16 . This result was obtained using
the library Deducer (likelihood.test function) in R.
23 G–test for independence yields G–value 142.15 with 6 degrees of freedom and p–value less than 2.2 × 10−16 . This result was obtained using
the library Deducer (likelihood.test function) in R.
52
share certain similarities (see Tables 22 and 23 in Section 9.2.2), even though it becomes apparent mostly on the core
sample.
To conclude, although there are certain characteristics present in both dialogues, not all of the findings from
Dialogue 2 coincide with those from Dialogue 1. This is particularly the case when we consider pooling relations
by strength. Results associated with agreements pooled by strength are quite distinct from the results in Dialogue 1.
It is worth observing that relations marked as Dependency, which previously did not follow any particular pattern,
now favour agreement levels marked as Neither. While strong relations in both dialogues do tend to favour believed
arguments more than disbelieved or neither in polarity pooling of agreement levels, the domination of this value in
Dialogue 1 is not replicated in Dialogue 2.
Total sample
Strong Relation
Core sample
Dependency
Normal Relation
N
WB
MB
N
SB
WB
MB
N
SB
WB
MB
SB
70%
60%
50%
40%
30%
20%
10%
0%
(a) Arguments and relations pooled by strength
Strong Relation
Dependency
Normal Relation
B
N
D
B
N
D
B
N
D
70%
60%
50%
40%
30%
20%
10%
0%
(b) Arguments pooled by polarity and relations pooled by strength
Figure 30: The levels of agreements assigned to the sources of relations of a given type in Dialogue 2 according to
a given type of pooling. We use the following abbreviations: Strong Belief (SB), Normal Belief (NB), Weak Belief
(WB), Neither (N), Disbelief (D), Belief (B).
53
6.3.2.2 Effect of an Argument on the Relations it Carries Out
In Figure 31 we can find the data showing what sort of relations are carried out by the statements with a particular
agreement level. In general, the choice of agreement level affects the resulting relation types24 . However, in the
pairwise analysis, certain similarities become more visible (see Table 24 in Section 9.2.3).
We may notice that some of the observations made in the case of Dialogue 1 carry over to Dialogue 2. In particular,
the charts created for Strongly Agree, Agree and Strongly Disagree tend to put more focus on the good reason for and
against edges. Nevertheless, the distributions are technically speaking distinct, and the similarity between the strong
agreement and disagreement is detected only in the core sample. The Disagree level, depending on the sample we
are dealing with, favours A good reason against and Somewhat good reason for edges, with A good reason for and A
somewhat good reason against next in line.
The Somewhat Agree and Somewhat Disagree charts are quite similar, as also supported by results in Table 24
in Section 9.2.3. In this case, the A somewhat good reason against agreement level is perhaps more common than
other options. The Neither Agree nor Disagree agreement level appears to favour the Somewhat related, but can’t
say how option the most and the Somewhat good reason for/against answers the least. Finally, we can observe that
the distributions obtained from the total and core samples in the case of the Don’t know chart are quite different; one
favours the A good reason for answer, while the other the Somewhat good reason for/against options. We can observe
that the distributions associated with the agreement levels are in principle not random (see Table 25 in Section 9.2.3),
though there appear to be certain difficulties with the Disagree and Somewhat Disagree values in the core sample.
Total sample
Strongly Agree
Core sample
Strongly Disagree
Agree
Disagree
Somewhat Agree
Neither Agree
nor Disagree
Somewhat Disagree
++
+
?
-
--
++
+
?
-
--
++
+
?
-
--
++
+
?
-
--
60%
50%
40%
30%
20%
10%
0%
Don’t know
++
+
?
-
--
++
+
?
-
--
++
+
?
-
--
++
+
?
-
--
60%
50%
40%
30%
20%
10%
0%
Figure 31: The types of relations carried out by statements of a given level of agreement in Dialogue 2. We use the
following abbreviations for the relation types: A good reason against (- -), A somewhat good reason against (-), A
good reason for (++), A somewhat good reason for (+), Somewhat related, but can’t say how (?)
Let us now consider pooling our results. We start by grouping arguments and relations by strength, as visible in
24 G–test
for independence yields G–value 390.02 with 28 degrees of freedom and p–value less than 2.2 × 10−16 . This result was obtained using
the library Deducer (likelihood.test function) in R.
54
Figure 32a. All of the belief strengths affect the associated relation distributions, both in the overall25 and pairwise
analysis (see Table 26 in Section 9.2.3). Additionally, the answers extracted from the users are not random (see Table
27 in Section 9.2.3). We can observe that the strong relations account for at least 50% of connections carried out by
arguments that are strongly or moderately believed. Weak belief appears to be split evenly between the strong and
normal relations. Arguments that are neither believed nor disbelieved can lead to all three types of relations, though
strong relations are more favoured than others. Consequently, not all of the observations made in the case of Dialogue
1 are repeated in Dialogue 2.
Let us now consider grouping arguments by polarity and relations by strength (see Figure 32b). Again, all of the
belief strengths affect the associated relation distributions, both in the overall26 and pairwise analysis (see Table 26 in
Section 9.2.3). The answers extracted from the users are also not random (see Table 27 in Section 9.2.3). Nevertheless,
there are certain similarities between the obtained distributions. In particular, we can observe that strong relations
appear to be the most common relations carried out by arguments marked as Believed or Neither, independently of
the sample, and by Disbelieved in the core sample. In particular, in the first and the last case these relations account
for over 60% of all edges. What is worth observing is that dependent links are the least common when it comes to
arguments that are believed or disbelieved. Nevertheless, it is easy to see that the results obtained for agreement levels
marked as Neither and Disbelieved in the second dialogue are different from the first dialogue.
We will now now briefly consider agreement levels paired with relations pooled by polarity. Although tests show
that the obtained results are in principle distinct, certain similarities become apparent in the pairwise analysis27 . If we
group agreement levels by polarity, we obtain the result that the proportions of the relation types carried out by the
believed and disbelieved arguments are not significantly different (see Table 26 in Section 9.2.3), particularly in the
case of the total sample. When we consider merging arguments by strength, certain similarities between the weak,
moderate and strong levels of agreement become visible (see Tables 26 and 27 in Section 9.2.3).
To conclude, despite various differences between Dialogue 1 and Dialogue 2, it again appears that analysis in
which relations are pooled by polarity leads to situations where relation distributions might not in all cases depend on
the agreement level of the source. When pooling relations by strength, we can also observe that similarly to Dialogue 1,
the strongly believed (strength pooling) and believed (polarity pooling) arguments lead primarily to strong relations.
However, the remaining pooled agreement levels do not appear to follow exactly the same patterns. Even though
certain shifts towards weaker relations can be observed in Figure 32a, their progress can be seen as “slower” than in
Dialogue 1.
25 G–test for independence yields G–value 232.25 with 6 degrees of freedom and p–value less than 2.2 × 10−16 . This result was obtained using
the library Deducer (likelihood.test function) in R.
26 G–test for independence yields G–value 167.86 with 4 degrees of freedom and p–value less than 2.2 × 10−16 . This result was obtained using
the library Deducer (likelihood.test function) in R.
27 When pooling arguments by strength, G–test for independence yields G–value 142.15 with 6 degrees of freedom and p–value less than
2.2 × 10−16 . When pooling arguments by polarity, G–test for independence yields G–value 132.84 with 4 degrees of freedom and p–value less
than 2.2 × 10−16 . This result was obtained using the library Deducer (likelihood.test function) in R.
55
Total sample
Strong Belief
Core sample
Moderate Belief
Weak Belief
Neither
D
NR
SR
D
NR
SR
D
NR
SR
D
NR
SR
90%
80%
70%
60%
50%
40%
30%
20%
10%
0%
(a) Arguments and relations pooled by strength
Belief
Neither
Disbelief
D
NR
SR
D
NR
SR
D
NR
SR
70%
60%
50%
40%
30%
20%
10%
0%
(b) Arguments pooled by polarity and relations pooled by strength
Figure 32: The types of relations carried out by statements of a given level of agreement in Dialogue 2 according to a
given pooling. We use the following abbreviations for the relation types: Strong Relation (SR),Normal Relation (NR),
Dependency (D).
56
6.4 Changes in Beliefs
In this section we discuss the statement awareness declared by the participants in both of the dialogues and the changes
– or rather, the lack of them – in participants’ opinions. In Figure 33 we can observe that the statements we have used
in the dialogues were, in general, not common knowledge.
Total sample
Core sample
Dalogue 1
Dialogue 2
100
% of Participants
80
60
40
20
0
A
B
C
D
E
F
A
B
C
D
E
F
G
H
Figure 33: The percentage of participants in the core and total samples in both dialogues who have stated that they
were aware of a given argument prior to the experiment
Nevertheless, this lack of awareness of the presented pieces of information did not result in participants changing
their opinions easily. For example, in the first dialogue, the majority of participants – resp. 90% and 87.5% in the total
and core samples – agreed with B on all stages of the dialogue, with the remainder disagreeing with B on all stages of
the dialogue. The majority of participants – 75% and 68.75% in the total and core samples – disagreed with A on all
stages of the dialogue, with resp. 17.5% and 25% agreeing with A on all stages of the dialogue. This leaves us only
with a couple of participants “swapping sides” in the dialogue. More such people can be found in the second dialogue,
where we have 17.5% and 26.67% participants changing between agreeing, disagreeing and being undecided at least
once in the case of statement B in the total and core samples, with 10% and 20% changing their opinions in the case
of A. With all the opportunities that the participants were given to change their opinion, the number of people who
did not at least once alternate between belief and disbelief appears extremely high. Even in a more detailed analysis,
where we differentiate between all the declarable levels of agreement rather than just belief–disbelief, the significant
changes appear to primarily concern argument B in the first dialogue (see Table 28). Hence, even though the dialogues
appear to have certain effect on the participants, it is more visible on the fine–grained level.
Despite these results, there are few notable changes that have occurred and which we would like to discuss. In
the first dialogue, we have two participants who have changed their position concerning the statement A: one who
initially disbelieved A, but believed it at the end, and one who initially believed A and disbelieved it at the end (this
particular person was also present in the core sample for in this dialogue). In both cases, the change appeared in the
second step of the dialogue. The first person initially disagreed with both A and B and apparently interpreted the
dialogue as discussing whether the hospital staff should be demanded to take the vaccine, not whether it would be best
for them to take it. Once statement C was presented, which in the opinion of this participant reinforced the fact that
the doctors don’t need the shots, he started agreeing with A and no further counterarguments fixed this situation. The
second participant initially agreed with both A and B and provided no real explanation for his opinions. However, the
appearance of C, with which the participant disagreed, led to further disagreement with A throughout the rest of the
dialogue.
In the second dialogue, we have five participants that we would like to discuss: 1) two for whom initial agreement
with B turned to disagreement without affecting A, 2) one for whom agreement (disagreement) with B (with A)
57
turned to disagreement (agreement), 3) one for whom disagreement with A turned into indecisiveness, and 4) one for
whom indecisiveness with B turned into disagreement. The last three participants belong to the core sample. In the
majority of these cases, the participants, either due to some initial skepticism or lack of knowledge, became swayed
by the incorrect statements, namely C, D and G. This was often paired with the inability to verify the correctness of
H, or with interpreting it as an admission that the contained substances were so harmful that they had to be removed
for the sake of the children28.
In Figure 34 we present the average change in beliefs of the participants, which is calculated in the following
manner. For every participant, we have summed the differences in their beliefs throughout all the stages in the dialogue
on statements they were aware of and those there were not w.r.t. the value assignment explained at the beginning of
Section 6. By dividing this sum by the number of statements declared as aware or unaware by a given participant
(assuming this number is non–zero), we have obtained an average change in statements of a given type per person.
Based on this we calculate the average change in statements of a given type per sample29 . As we may observe, the
changes are quite modest, and relatively close in both aware and unaware statements. This is particularly visible in
Dialogue 1, where the average change is smaller than the single smallest change possible (e.g. the difference between
Strongly Agree to Agree, which are mapped to 6/6 and 5/6 respectively, is approximately 0.167). The variability
in Dialogue 2 is somewhat larger, though one also has to bear in mind that more statements were presented to the
participants and therefore they were given more opportunities to change their mind per statement. Nevertheless, it
is possible that some of the observed changes were noise rather than actual modifications in beliefs caused by the
dialogue. This is particularly supported by the fact that various possible explanations for these differences, such as the
appearance of an attacker, could have been paired with an increase as well as decrease in the beliefs declared by the
participants.
Average Change
in Beliefs
Dialogue 1
Dialogue 2
0.3
0.2
Total sample
0.1
Core sample
0
Aware Unaware Aware Unaware
Figure 34: Average change in beliefs in dialogues per sample, calculated using the value assignment from page 19.
It appears that in many cases, the participants were able to evaluate the statements they were not familiar with
by using their own knowledge and other facts that they were aware of. Unfortunately, this also happened in the
cases where the knowledge of the participants was incorrect, such as the vaccines causing autism or flu shots causing
flu. However, the complete inability to verify the pieces of information contained in the statements with their own
knowledge had made some participants of Dialogue 2 choose to neither agree nor disagree with the initial statements
presented in it. This might mean that depending on the difficulty of the dialogue, too much as well as too little personal
knowledge might affect how open to change people are.
We would also like to note that we have found participants who, on at least one occasion, gave answers such as
“I agree/disagree with this statement”, “This information is (in)accurate” or “I think/believe this is true/false” in the
explanation tasks and refrained from providing a reason for their opinions. By considering only those cases in which
the participants claimed they were not aware of the presented pieces of information and the stage of the dialogue they
were at should not have given them sufficient reasons for evaluating these statements, we have found respectively 8
and 5 participants for the first and second dialogues resorting to such explanations. These participants also exhibited
insignificant, if any, changes in their opinions. Therefore, it might be the case that for various reasons, they could not
or did not want to provide us with further details.
28 The thimerosal has been removed from the vaccine primarily due to social pressure and overwhelming distrust, not any scientific evidence of
its harmfulness.
29 Participants who did not have any unaware (resp. aware) statements were excluded from the appropriate calculation.
58
7
Related Work
In this section we would like to review the results of other works dealing with empirical evaluation of argumentation
approaches and discuss other works dealing with more fine–grained models to argument acceptability. We begin by
analyzing the work by Rahwan et al [55], which focuses on the issue of reinstatement. The authors have prepared a
total of 10 short dialogues in order to study this phenomenon. A sample dialogue for the case of simple reinstatement
is as follows:
A Louis applied the brake and the brake was not faulty. Therefore, the car slowed down.
B The brake fluid was empty. Therefore, the brake was faulty.
C The car had just undergone maintenance service. Therefore, the brake fluid was not empty.
The dialogue was split into three stages, where first A, then A and B, and finally all A, B and C were presented.
For each problem, the participants had to choose a value from a 7–point scale ranging from Certainly false to Certainly
true that best describes their confidence in the conclusion of A. The obtained results support the notions of defeat and
reinstatement, that is, that the belief in A decreases once it is defeated and increases when it is defended, though still
remaining significantly lower than prior to the defeat.
This study, similarly to ours, lends support to the use of more fine–grained approaches towards describing the
beliefs of the participants, such as the epistemic approaches. However, as we can observe, the dialogues used in this
study were much simpler and shorter than ours, and unlikely to be affected by any subjective views of the participants.
The results show that reinstatement as such does occur and thus support the idea behind semantics such as grounded
or preferred. However, at this point it cannot be claimed that they are applicable in a more complex setting and that
they are a general proof for the use of argumentation semantics in computational persuasion.
The next study we would like to consider is by Cerutti et al [19], which again studies various forms of reinstatement.
The authors have prepared two texts – a base one and an extended one – on four topics (weather forecast, political
debate, used car sale and romantic relationship). For example, the texts prepared for the political debate scenario are
as follows:
Base Case: In a TV debate, the politician AAA argues that if Region X becomes independent then Xs citizens will be
poorer than now. Subsequently, financial expert Dr. BBB presents a document, which scientifically shows that
Region X will not be worse off financially if it becomes independent.
Extended Case: In a TV debate, the politician AAA argues that if Region X becomes independent then Xs citizens
will be poorer than now. Subsequently, financial expert Dr. BBB presents a document, which scientifically
shows that Region X will not be worse off financially if it becomes independent. After that, the moderator of the
debate reminds BBB of more recent research by several important economists that disputes the claims in that
document.
The participants are given one of the eight scenarios and then asked to determine which of the following positions
they think is accurate:
• PA : I think that AAAs position is correct (e.g. “Xs citizens will be poorer than now”).
• PB : I think that BBBs position is correct (e.g. “Xs citizens will not be worse off financially”).
• PU : I cannot determine if either AAAs or BBBs position is correct (e.g. “I cannot conclude anything about
Region Xs finances”).
Next, the participants were asked certain questions concerning the text. In particular, concerning the agreement,
the participants are asked “How much do you agree with the following statements?” and respond on a 7–point scale
from Disagree to Agree for each statement.
The results show that the majority of participants in the base case scenarios agree with position PB and with position PU in the extended scenarios. They suggest a correspondence between the formal theory used by the authors and
59
its representation in natural language. However, additional analyses show that there are some significant deviations,
apparently caused by the personal knowledge of the participants. For example, while PB was the most common choice
in the base weather forecast scenario, there were participants who chose PU and explained their decision with “All
weather forecasts are notoriously inaccurate”.
Our study, similarly to this one, points to the personal knowledge of the participants affecting their decisions. However, due to a different methodology, it is difficult to compare the other findings. We have presented the participants
with an ongoing dialogue and tried to monitor the changes in their beliefs at every step of the way, similarly as in [55].
In this study, a given participant receives only one text to evaluate, i.e. the base and extended scenarios are judged by
different people. We therefore cannot claim that if a person was first shown the base case, and then the extended one,
the PU choice would still be more common than PB , which is what our results would point to.
Another interesting study worth considering is [56], which focuses on exploring the abilities of argumentation
theory, relevance heuristics, machine and transfer learning for predicting the argument choices of participants, with
a particular focus on the machine learning. Argumentation theory is verified using three experiments, in which the
dialogues are used to construct bipolar argumentation frameworks and the sets of arguments selected by the participants
are contrasted with grounded, preferred and stable extensions. In the first experiment consisting of 6 scenarios, the
authors create bipolar argumentation frameworks which are not known to the participants, present two standpoints
from two parties and ask the participants to choose which of the additional four arguments they would use next if
they were one of the deliberants in the discussion. In the second experiment, real life data is annotated and structured
into a bipolar argumentation framework that is later analyzed. In the third experiment, a chat bot is created, aimed at
discussing flu vaccination with the participants. In this case, both the chat and the participant can only use arguments
from a predefined list. Finally, in an additional experiment, an argument theory agent was implemented in order to
provide suggestions to the participants during a two–person chat.
The authors report that a substantial part of the results (or in some cases, even the majority) do not conform to the
results predicted by the semantics. It is worth mentioning that the stated adherence to the conflict–free extension–based
semantics is 78% and is similar to our results concerning the rational postulate, which corresponds to this semantics.
The created agent was also considered helpful only by a small percentage of the participants. Nevertheless, the causes
for such behaviour of the semantics are not investigated, and the participants were not allowed to explain their decisions
(the first and the third experiment) or there was no possibility to ask them for further input (the second experiment).
Moreover, unlike in our study, the participants were evaluated against the graphs constructed by the authors or annotators. As shown by our results, they do not necessarily reflect how the participants view the relations between the
arguments. Finally, there is no discussion concerning whether these particular bipolar argument framework semantics
used in this experiment [1] are applicable. The stable and conflict–free semantics in this work are based on direct and
supported attacks, and only the direct ones need to be defended from and can be used for defence. Taking into account
the fact that this approach has been superseded by a number of different methods since it has been introduced, the
presented results indicate that these particular semantics are not useful in modelling of the user behaviour, rather than
there exists a deeper issue within the argumentation itself.
The aforementioned studies focused on the empirical evaluation of certain phenomenona in argumentation. However, there are also other studies, which focus more on the behavioural methods or computational linguistics perspective rather than the argumentation perspective and can be seen as complementary to ours. These works often use
arguments sourced from e.g. social media, and analyze the relation between the persuasiveness of an argument and its
traits, the persuasiveness of an argument or relations between the arguments and the personality or the emotions of the
participant, and more. Some of them, such as [40], also point to the importance of the prior knowledge in changing
one’s beliefs. These studies are an important line of work which can be harnessed in creating computational persuasion methods that are tailored to the participants and which can provide guidance in transforming logical arguments to
natural language arguments. Therefore, they should be considered in the next steps of our work. We refer the readers
to [61, 40] for further details.
Although in this work we have mostly focused on the analysis of epistemic probabilities, we would like to note that
this is not the only approach allowing a more fine–grained perspective on argument acceptability. There also are other
interesting methods that allow score assignments, such as certain forms of preference or weight–based argumentation
(see [9, 6] for an overview). However, many of these works share one common problem, namely that the values
associated with the arguments are quite abstract and do not have a meaning of their own. Even though using them we
60
can, for example, state that a participant agrees with one argument more than with another, we cannot state whether
he agrees with it in the first place. Let us consider a framework containing a single argument and no attacks. Giving
this argument a preference 0.3 has no effect, as it will always be accepted in any extension. Giving it the epistemic
probability of 0.3, which is interpreted as disbelief, will lead it to it being always rejected. Consequently, despite
certain similarities on the structural level, the semantics of the preference and weight–based frameworks are quite
different from epistemic probabilities.
8
Conclusions
Our work, through empirically verifying certain prevailing assumptions in argumentation, is about informing the
design of formalisms that can model how participants think about arguments arising in a dialogue. We observe that
handling the uncertainty concerning both the participants’ opinions about arguments and the structure of the graph is of
critical importance, and that a Dung framework equipped only with the standard semantics is insufficient to represent
the views of the participants. Our results can be summarized in the following manner:
Observation 1 The data supports the use of the constellation approach to probabilistic argumentation for modelling
the argument graphs representing the views of dialogue participants. In particular, in Section 6.1 we have
seen that people may interpret statements and relations between them differently and without adhering to the
intended relations. Furthermore, their personal knowledge can affect their perception and evaluation of the
dialogue. Thus, the constellation approach can represent our uncertainty about the argument graphs describing
our opponents views.
Observation 2 People may explicitly declare that two given statements are connected, however, they might not be
sure of the exact nature of the relation between them. In Section 6.1 we have observed that the portion of
the graphs declared by the participants that were not clarified (i.e. contained at least one relation marked as
dependent) reached even 70%. We therefore also need to express the uncertainty that a person has about his or
her own views. Although for the purpose of the analysis we have introduced the notion of a tripolar framework
in order to be represent such situations, this issue can potentially be addressed with the constellation approach
to argumentation. By this we understand that associated with a given unclarified framework we can have a
number of graphs which interpret the dependent links as attacking or supporting, each with a given probability
that could, for example, be obtained from the graph distribution of all participants.
Observation 3 The data supports the use of the epistemic approach to probabilistic argumentation. In Section 6.2
we calculated how many values the participants were using throughout the dialogues, and we could have observed that in most of the cases three values were insufficient to represent the participants’ opinions. Another of
our important observations concerns the adherence to the epistemic postulates, analyzed in the aforementioned
section. While classical semantics tend to represent a number of properties at the same time, a single postulate
tends to focus on a single aspect, as seen in Section 4.1. They therefore allow a more detailed view on the
participant behaviour and can allow us to analyze the cases in which classic semantics may fail to explain it.
The epistemic postulates can also provide more feedback to argumentation systems, such as for computational
persuasion, than normal semantics do. For example, the low performance of the argumentation semantics such
as complete does not really inform the system what aspect of the participant reasoning does not meet the semantics requirements. In contrast, the high performance of the preferential postulates and low performance of
the explanatory ones, can inform the system that it has insufficient information about the participant’s personal
knowledge and that it should proceed with querying him or her for further data. This is particularly important
as our data also shows that people use their own personal knowledge in order to make judgments and might not
necessarily disclose it.
The extended epistemic approach [52] would also allow us to model situations where the way a given relation
is perceived is not necessarily tightly related to the way how much we believe or disbelieve its source, which as
seen in Section 6.3, does tend to occur. It is likely that there are more properties of an argument or statement,
such as how detailed and informative it is, which affect how the relations carried out by it are seen by people.
While our analysis concerned the impact of the source of a relation on the relation and vice versa, it is possible
61
that both source and target can affect the relation and thus approaches from [16, 37] could be verified. Thus,
these issues need further investigation.
Observation 4 The data supports the use of bipolar argumentation frameworks. In Section 6.1 we have observed
that the participants explicitly view certain relations as supporting and that the notion of defence does not
account for all of the positive relations that the participants have identified between the presented statements.
In particular, we could have observed that there are new support relations arising in the context of the dialogue,
such as support coming from statements working towards the same goal.
Observation 5 The data supports the use of bipolar argumentation in combination with the prudent/careful approaches. In Section 6.1 we have observed that many additional attacks perceived by the participants can be
explained by the existing notions of indirect conflicts in these settings. They can therefore be used to model
auxiliary conflicts arising in the context of a dialogue, but not necessarily created on the logical level.
Observation 6 The data shows that people use their own personal knowledge in order to make judgments and might
not necessarily disclose it. In Section 6.2 we have shown how the differences between the declared and expanded
graphs become visible on the postulates highly sensitive to personal knowledge, such as the ones belonging to
the explanatory family. Additionally, in Section 6.4 we have noted that a number of participants were able to
evaluate the presented statements despite the lack of awareness of certain information. This, combined with
them not providing any explanations for their opinions, may mean that argumentation systems need to handle
participants not willing to share their knowledge.
Observation 7 The data shows that presenting a new and correct piece of information that a given person was not
aware of does not necessarily lead to changing that person’s beliefs. In Section 6.4 we have observed that
throughout the dialogues, not many participants have changed their opinions in a significant manner. Moreover,
quite often those who have, have done so under the influence of incorrect information and presenting them with
the correct data has not managed to rectify their opinions.
Our exploratory study shows that the most common approaches to argumentation might be too simplistic in order to
adequately grasp human reasoning. However, we do not believe that the argumentation theory as a field is insufficient
altogether. In particular, we have highlighted the correspondence of the obtained results to various, less common
formalisms, such as probabilistic and bipolar frameworks, and prudent and careful approaches. Consequently, these
methods could be merged in the future in order to develop abstract argumentation tools that can be used in dialogical
argumentation with more success. Nevertheless, further and more specialized studies concerning our observations
should be carried out. In particular, our experiment could be seen as dynamic, as it concerned two dialogues between
different parties. It would be interesting to observe whether our findings could be replicated in a more static setting,
where arguments are presented randomly and not in the context of a dialogue. For example, this shift could affect the
perception of the indirect attacks, defenses and supports between different statements. We believe further studies will
provide more insight into this matter.
62
References
[1] L. Amgoud, C. Cayrol, M. C. Lagasquie-Schiex, and P. Livet. On bipolarity in argumentation frameworks.
International Journal of Intelligent Systems, 23(10):1062–1093, 2008.
[2] Pietro Baroni, Martin Caminada, and Massimiliano Giacomin. An introduction to argumentation semantics. The
Knowledge Engineering Review, 26(4):365–410, 2011.
[3] Pietro Baroni, Federico Cerutti, Massimiliano Giacomin, and Giovanni Guida. AFRA: Argumentation framework with recursive attacks. International Journal of Approximate Reasoning, 52(1):19–37, 2011.
[4] Pietro Baroni, Massimiliano Giacomin, and Paolo Vicig. On rationality conditions for epistemic probabilities
in abstract argumentation. In Simon Parsons, Nir Oren, Chris Reed, and Federico Cerutti, editors, Proceedings
of the 5th International Conference on Computational Models of Argument (COMMA 2014), volume 266 of
Frontiers in Artificial Intelligence and Applications, pages 121–132. IOS Press, 2014.
[5] Guido Boella, Dov Gabbay, Leendert van der Torre, and Serena Villata. Support in abstract argumentation. In
Pietro Baroni, Federico Cerutti, Massimiliano Giacomin, and Guillermo R. Simari, editors, Proceedings of the
3rd International Conference on Computational Models of Argument (COMMA 2010), volume 216 of Frontiers
in Artificial Intelligence and Applications, pages 111–122, Amsterdam, The Netherlands, The Netherlands, 2010.
IOS Press.
[6] Elise Bonzon, Jérôme Delobelle, Sébastien Konieczny, and Nicolas Maudet. A comparative study of rankingbased semantics for abstract argumentation. In Dale Schuurmans and Michael Wellman, editors, Proceedings of
the 30th AAAI Conference on Artificial Intelligence (AAAI 2016), pages 914–920. AAAI Press, 2016.
[7] Richard Booth, Martin Caminada, Mikolaj Podlaszewski, and Iyad Rahwan. Quantifying disagreement in
argument-based reasoning. In Wiebe van der Hoek, Lin Padgham, Vincent Conitzer, and Michael Winikoff,
editors, Proceedings of the 11th International Conference on Autonomous Agents and Multiagent Systems (AAMAS 2012), pages 493–500. IFAAMAS, 2012.
[8] Tom Bosc, Elena Cabrio, and Serena Villata. Tweeties squabbling: Positive and negative results in applying
argument mining on social media. In Pietro Baroni, Thomas F. Gordon, Tatjana Scheffler, and Manfred Stede,
editors, Proceedings of the 6th International Conference on Computational Models of Argument (COMMA 2016),
volume 287 of Frontiers in Artificial Intelligence and Applications, pages 21–32. IOS Press, 2016.
[9] Gerhard Brewka, Sylwia Polberg, and Stefan Woltran. Generalizations of Dung frameworks and their role in
formal argumentation. IEEE Intelligent Systems, 29(1):30–38, January 2014.
[10] Elena Cabrio and Serena Villata. Node: A benchmark of natural language arguments. In Simon Parsons, Nir
Oren, Chris Reed, and Federico Cerutti, editors, Proceedings of the 5th International Conference on Computational Models of Argument (COMMA 2014), volume 266 of Frontiers in Artificial Intelligence and Applications,
pages 449–450. IOS Press, 2014.
[11] Martin Caminada and Dov M. Gabbay. A logical account of formal argumentation. Studia Logica, 93(2):109–
145, 2009.
[12] Claudette Cayrol and Marie-Christine Lagasquie-Schiex. Bipolar abstract argumentation systems. In Iyad Rahwan and Guillermo R. Simari, editors, Argumentation in Artificial Intelligence, pages 65–84. Springer, 2009.
[13] Claudette Cayrol and Marie-Christine Lagasquie-Schiex. Coalitions of arguments: A tool for handling bipolar
argumentation frameworks. International Journal of Intelligent Systems, 25(1):83–109, 2010.
[14] Claudette Cayrol and Marie-Christine Lagasquie-Schiex. Weighted argumentation systems: A tool for merging
argumentation systems. In Taghi M. Khoshgoftaar and Xingquan Zhu, editors, Proceedings of the 23rd IEEE
International Conference on Tools with Artificial Intelligence (ICTAI 2011), pages 629–632. IEEE Computer
Society, 2011.
63
[15] Claudette Cayrol and Marie-Christine Lagasquie-Schiex. Bipolarity in argumentation graphs: Towards a better
understanding. International Journal of Approximate Reasoning, 54(7):876–899, 2013.
[16] Claudette Cayrol and Marie-Christine Lagasquie-Schiex. From preferences over arguments to preferences over
attacks in abstract argumentation: A comparative study. In Nikolaos Bourbakis, editor, Proceedings of the
25th IEEE International Conference on Tools with Artificial Intelligence (ICTAI 2013), pages 588–595. IEEE
Computer Society, 2013.
[17] Claudette Cayrol and Marie-Christine Lagasquie-Schiex. An axiomatic approach to support in argumentation.
In Elizabeth Black, Sanjay Modgil, and Nir Oren, editors, Proceedings of the 3rd International Workshop on
Theory and Applications of Formal Argumentation (TAFA 2015), volume 9524 of Lecture Notes in Computer
Science, pages 74–91. Springer International Publishing, 2015.
[18] Federico Cerutti, Alexis Palmer, Ariel Rosenfeld, Jan Snajder, and Francesca Toni. A pilot study in using
argumentation frameworks for online debates. In Proceedings of the 1st International Workshop on Systems and
Algorithms for Formal Argumentation (SAFA) 2016), pages 63–74, 2016.
[19] Federico Cerutti, Nava Tintarev, and Nir Oren. Formal arguments, preferences, and natural language interfaces
to humans: an empirical evaluation. In Torsten Schaub, Gerhard Friedrich, and Barry O’Sullivan, editors, Proceedings of the 21st European Conference on Artificial Intelligence (ECAI 2014), volume 263 of Frontiers in
Artificial Intelligence and Applications, pages 1033–1034. IOS Press, August 2014.
[20] Andrea Cohen, Sebastian Gottifredi, Alejandro J. Garcı́a, and Guillermo R. Simari. On the acceptability semantics of argumentation frameworks with recursive attack and support. In Pietro Baroni, Thomas F. Gordon,
Tatjana Scheffler, and Manfred Stede, editors, Proceedings of the 6th International Conference on Computational
Models of Argument (COMMA 2016), volume 287 of Frontiers in Artificial Intelligence and Applications, pages
231–242. IOS Press, 2016.
[21] Sylvie Coste-Marquis, Caroline Devred, Sébastien Konieczny, Marie-Christine Lagasquie-Schiex, and Pierre
Marquis. On the merging of Dung’s argumentation systems. Artificial Intelligence, 171(10):730–753, 2007.
[22] Sylvie Coste-Marquis, Caroline Devred, and Pierre Marquis. Inference from controversial arguments. In Geoff
Sutcliffe and Andrei Voronkov, editors, Proceedings of the 12th International Conference on Logic for Programming, Artificial Intelligence, and Reasoning (LPAR 2005), volume 3835 of Lecture Notes in Computer Science,
pages 606–620. Springer-Verlag Berlin Heidelberg, 2005.
[23] Sylvie Coste-Marquis, Caroline Devred, and Pierre Marquis. Prudent semantics for argumentation frameworks.
In Andrew Lim, editor, Proceedings of the 17th IEEE International Conference on Tools with Artificial Intelligence (ICTAI 2005), pages 568–572, Washington, DC, USA, 2005. IEEE Computer Society.
[24] Dragan Doder and Stefan Woltran. Probabilistic argumentation frameworks – a logical approach. In Umberto
Straccia and Andrea Calı̀, editors, Proceedings of the 8th International Conference on Scalable Uncertainty
Management (SUM 2014), volume 8720 of Lecture Notes in Computer Science, pages 134–147. Springer International Publishing, 2014.
[25] Pierpaolo Dondio. Multi-valued and probabilistic argumentation frameworks. In Simon Parsons, Nir Oren,
Chris Reed, and Federico Cerutti, editors, Proceedings of the 5th International Conference on Computational
Models of Argument (COMMA 2014), volume 266 of Frontiers in Artificial Intelligence and Applications, pages
253–260. IOS Press, 2014.
[26] Pierpaolo Dondio. Toward a computational analysis of probabilistic argumentation frameworks. Cybernetics
and Systems, 45(3):254–278, 2014.
[27] Phan Minh Dung. On the acceptability of arguments and its fundamental role in nonmonotonic reasoning, logic
programming and n-person games. Artificial Intelligence, 77:321–357, 1995.
64
[28] Phan Minh Dung and Phan Minh Thang. Towards (probabilistic) argumentation for jury-based dispute resolution.
In Pietro Baroni, Federico Cerutti, Massimiliano Giacomin, and Guillermo R. Simari, editors, Proceedings of the
3rd International Conference on Computational Models of Argument (COMMA 2010), volume 216 of Frontiers
in Artificial Intelligence and Applications, pages 171–182, Amsterdam, The Netherlands, The Netherlands, 2010.
IOS Press.
[29] Bettina Fazzinga, Sergio Flesca, and Francesco Parisi. On the complexity of probabilistic abstract argumentation.
In Francesca Rossi, editor, Proceedings of the 23rd International Joint Conference on Artificial Intelligence
(IJCAI 2013), pages 898–904. AAAI Press, 2013.
[30] Bettina Fazzinga, Sergio Flesca, and Francesco Parisi. On the complexity of probabilistic abstract argumentation
frameworks. ACM Transactions on Computational Logic, 16(3):1–39, 2015.
[31] Anthony Hunter. Some foundations for probabilistic abstract argumentation. In Bart Verheij, Stefan Szeider, and
Stefan Woltran, editors, Proceedings of the 4th International Conference on Computational Models of Argument
(COMMA 2012), volume 245 of Frontiers in Artificial Intelligence and Applications, pages 117–128. IOS Press,
2012.
[32] Anthony Hunter. A probabilistic approach to modelling uncertain logical arguments. International Journal of
Approximate Reasoning, 54(1):47–81, 2013.
[33] Anthony Hunter and Matthias Thimm. Probabilistic argument graphs for argumentation lotteries. In Simon
Parsons, Nir Oren, Chris Reed, and Federico Cerutti, editors, Proceedings of the 5th International Conference
on Computational Models of Argument (COMMA 2014), volume 266 of Frontiers in Artificial Intelligence and
Applications, pages 313–324. IOS Press, September 2014.
[34] Anthony Hunter and Matthias Thimm. Probabilistic argumentation with epistemic extensions and incomplete
information. Technical report, ArXiv, May 2014.
[35] Anthony Hunter and Matthias Thimm. On partial information and contradictions in probabilistic abstract argumentation. In Chitta Baral, James Delgrande, and Frank Wolter, editors, Proceedings of the 15th International
Conference on Principles of Knowledge Representation and Reasoning (KR 2016), pages 53–62. AAAI Press,
April 2016.
[36] Anthony Hunter and Matthias Thimm. Probabilistic reasoning with abstract argumentation frameworks. Journal
of Artificial Intelligence Research, 2017. Forthcoming.
[37] Souhila Kaci and Christophe Labreuche. Arguing with valued preference relations. In Weiru Liu, editor, Proceedings of the 11th European Conference on Symbolic and Quantitative Approaches to Reasoning with Uncertainty (ECSQARU 2011), volume 6717 of Lecture Notes in Computer Science, pages 62–73. Springer Berlin
Heidelberg, 2011.
[38] Hengfei Li, Nir Oren, and Timothy J. Norman. Probabilistic argumentation frameworks. In Sanjay Modgil, Nir
Oren, and Francesca Toni, editors, Proceedings of the 1st International Workshop on the Theory and Applications of Formal Argumentation (TAFA 2011), volume 7132 of Lecture Notes in Computer Science, pages 1–16.
Springer-Verlag Berlin Heidelberg, 2011.
[39] Hengfei Li, Nir Oren, and Timothy J. Norman. Relaxing independence assumptions in probabilistic argumentation. In Proceedings of the 10th International Workshop on Argumentation in Multi-Agent Systems (ArgMAS
2013), 2013.
[40] Stephanie Lukin, Pranav Anand, Marilyn Walker, and Steve Whittaker. Argument strength is in the eye of the
beholder: Audience effects in persuasion. In Mirella Lapata, Phil Blunsom, and Alexander Koller, editors,
Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics
(EACL 2017): Volume 1, Long Papers, pages 742–753. Association for Computational Linguistics, 2017.
65
[41] Hien Nguyen and Judith Masthoff. Designing persuasive dialogue systems: Using argumentation with care. In
Harri Oinas-Kukkonen, Per Hasle, Marja Harjumaa, Katarina Segerståhl, and Peter Øhrstrøm, editors, Proceedings of the 3rd International Conference on Persuasive Technology (PERSUASIVE 2008), pages 201–212, Berlin,
Heidelberg, 2008. Springer Berlin Heidelberg.
[42] Farid Nouioua. AFs with necessities: Further semantics and labelling characterization. In Weiru Liu, V.S.
Subrahmanian, and Jef Wijsen, editors, Proceedings of the 7th International Conference on Scalable Uncertainty
Management (SUM 2013), volume 8078 of Lecture Notes in Computer Science, pages 120–133. Springer-Verlag
Berlin Heidelberg, 2013.
[43] Farid Nouioua and Vincent Risch. Argumentation frameworks with necessities. In Salem Benferhat and John
Grant, editors, Proceedings of the 5th International Conference on Scalable Uncertainty Management (SUM
2011), volume 6929 of Lecture Notes in Artificial Intelligence, pages 163–176. Springer-Verlag Berlin Heidelberg, 2011.
[44] Jane Ogden. Health Psychology: A Textbook. Open University Press, 5 edition, 2012.
[45] James M. Olson. Psychological barriers to behavior change:how to indentify the barriers that inhibit change.
Canadian Family Physician, 38:309–319, February 1992.
[46] Nir Oren and Timothy J. Norman. Semantics for evidence-based argumentation. In Philippe Besnard, Sylvie
Doutre, and Anthony Hunter, editors, Proceedings of the 2nd International Conference on Computational Models
of Argument (COMMA 2008), volume 172 of Frontiers in Artificial Intelligence and Applications, pages 276–
284. IOS Press, 2008.
[47] Nir Oren, Chris Reed, and Michael Luck. Moving between argumentation frameworks. In Pietro Baroni, Federico Cerutti, Massimiliano Giacomin, and Guillermo R. Simari, editors, Proceedings of the 3rd International
Conference on Computational Models of Argument (COMMA 2010), volume 216 of Frontiers in Artificial Intelligence and Applications, pages 379–390, Amsterdam, The Netherlands, The Netherlands, 2010. IOS Press.
[48] Sylwia Polberg. Understanding the Abstract Dialectical Framework. In Loizos Michael and Antonis Kakas,
editors, Proceedings of the 15th European Conference on Logics in Artificial Intelligence (JELIA 2016), volume
10021 of Lecture Notes in Artificial Intelligence, pages 430–446. Springer International Publishing, 2016.
[49] Sylwia Polberg. Intertranslatability of abstract argumentation frameworks. Technical Report DBAI-TR-2017104, Institute for Information Systems, Technical University of Vienna, 2017.
[50] Sylwia Polberg and Dragan Doder. Probabilistic abstract dialectical frameworks. In Eduardo Fermé and João
Leite, editors, Proceedings of the 14th European Conference on Logics in Artificial Intelligence (JELIA 2014),
volume 8761 of Lecture Notes in Computer Science, pages 591–599, Cham, 2014. Springer International Publishing.
[51] Sylwia Polberg and Anthony Hunter. Data Appendix. http://www0.cs.ucl.ac.uk/staff/a.hunter/papers/dia
2017.
[52] Sylwia Polberg, Anthony Hunter, and Matthias Thimm. Belief in attacks in epistemic probabilistic argumentation. In Proceedings of the 11th International Conference on Scalable Uncertainty Management (SUM 2017),
Lecture Notes in Artificial Intelligence. Springer-Verlag Berlin Heidelberg, 2017. Forthcoming.
[53] Sylwia Polberg and Nir Oren. Revisiting support in abstract argumentation systems. In Simon Parsons, Nir Oren,
Chris Reed, and Federico Cerutti, editors, Proceedings of the 5th International Conference on Computational
Models of Argument (COMMA 2014), volume 266 of Frontiers in Artificial Intelligence and Applications, pages
369–376. IOS Press, 2014.
66
[54] Henry Prakken. On support relations in abstract argumentation as abstractions of inferential relations. In Torsten
Schaub, Gerhard Friedrich, and Barry O’Sullivan, editors, Proceedings of the 21st European Conference on
Artificial Intelligence (ECAI 2014), volume 263 of Frontiers in Artificial Intelligence and Applications, pages
735–740. IOS Press, 2014.
[55] I. Rahwan, M Madakkatel, J. Bonnefon, R Awan, and S. Abdallah. Behavioural experiments for assessing the
abstract argumentation semantics of reinstatement. Cognitive Science, 34(8):1483–1502, 2010.
[56] Ariel Rosenfeld and Sarit Kraus. Providing arguments in discussions on the basis of the prediction of human
argumentative behavior. ACM Transactions on Interactive Intelligent Systems, 6(4):30:1–30:33, December 2016.
[57] Alexander J Rothman and Peter Salovey. Shaping perceptions to motivate healthy behavior: the role of message
framing. Psychological Bulletin, 121(1):3–19, 1997.
[58] Matthias Thimm. A probabilistic semantics for abstract argumentation. In Luc De Raedt, Christian Bessiere,
Didier Dubois, Patrick Doherty, Paolo Frasconi, Fredrik Heintz, and Peter Lucas, editors, Proceedings of the 20th
European Conference on Artificial Intelligence (ECAI 2012), volume 242 of Frontiers in Artificial Intelligence,
pages 750–755. IOS Press, August 2012.
[59] Amos Tversky and Daniel Kahneman. The Framing of Decisions and the Psychology of Choice, volume 4 of
NATO ASI Series, pages 107–129. Springer Berlin Heidelberg, 1985.
[60] Serena Villata, Guido Boella, and Leendert Van Der Torre. Attack semantics for abstract argumentation. In Toby
Walsh, editor, Proceedings of the 22nd International Joint Conference on Artificial Intelligence (IJCAI 2011),
pages 406–413. AAAI Press, 2011.
[61] Serena Villata, Elena Cabrio, Sahbi Jraidi, Imène; Benlamine, Maher Chaouachi, Claude Frasson, and Fabien
Gandon. Emotions and personality traits in argumentation: An empirical evaluation. Argument & Computation,
8(1):61–87, 2017.
9
Statistics Appendix
9.1 Dialogue 1
9.1.1 Postulate Satisfaction
Total Sample
Preferential
Rational
Strict
Protective
Restrained
Coherent
Involutary
Justifiable
Semi Optimistic
Semi Founded
Founded
Optimistic
Demanding
Guarded
Discharging
Trusting
Anticipating
Core Sample
Preferential
Rational
Strict
Protective
Restrained
Coherent
Involutary
Justifiable
SemiOptimistic
SemiFounded
Founded
Optimistic
Demanding
Guarded
Discharging
Trusting
Anticipating
Intended vs
Declared
Z
p–value
1.428
0.149
2.417
0.019
1.609
0.115
1.102
0.289
0.880
0.411
-1.783
0.075
-1.344
0.183
2.917
0.004
1.799
0.064
4.107
0.000
3.353
0.001
3.213
0.001
-1.083
0.354
4.220
0.000
4.220
0.000
4.224
0.000
4.224
0.000
Intended vs
Declared
Z
p–value
1.732
0.250
1.730
0.250
1.414
0.500
0.318
1.000
-0.138
0.750
-0.138
0.750
0.615
0.750
1.730
0.250
0.615
0.750
0.841
0.500
0.382
1.000
0.382
1.000
1.000
1.000
0.841
0.500
0.841
0.500
0.841
0.500
0.841
0.500
Intended vs
Expanded
Z
p–value
1.472
0.143
2.426
0.018
1.795
0.079
1.102
0.289
0.880
0.411
-0.252
0.804
-0.461
0.650
2.917
0.004
1.420
0.144
1.847
0.066
0.956
0.351
1.150
0.260
-2.383
0.016
2.109
0.034
2.133
0.032
2.203
0.026
2.203
0.026
Intended vs
Expanded
Z
p–value
1.730
0.250
1.730
0.250
1.732
0.250
0.318
1.000
-0.138
0.750
1.208
0.375
1.033
0.500
1.730
0.250
0.034
1.000
-1.587
0.124
-2.169
0.031
-1.932
0.055
-0.046
1.000
-1.390
0.176
-1.389
0.176
-1.442
0.152
-1.442
0.152
Intended vs
Common
Z
p–value
1.999
0.125
1.732
0.250
1.732
0.250
1.732
0.250
1.414
0.500
NA
NA
NA
NA
NA
NA
-0.296
1.000
-1.732
0.250
-2.449
0.031
-2.000
0.125
1.999
0.125
-1.414
0.500
-1.414
0.500
-1.732
0.250
-1.732
0.250
Intended vs
Common
Z
p–value
1.000
1.000
1.000
1.000
1.000
1.000
1.000
1.000
NA
NA
NA
NA
NA
NA
NA
NA
-1.000
1.000
-1.413
0.500
-1.732
0.250
-1.732
0.250
NA
NA
-1.413
0.500
-1.413
0.500
-1.730
0.250
-1.730
0.250
Declared vs
Expanded
Z
p–value
1.000
1.000
1.000
1.000
1.000
1.000
NA
NA
NA
NA
2.447
0.031
2.235
0.063
NA
NA
-1.414
0.500
-3.304
0.001
-3.153
0.002
-2.823
0.008
-2.447
0.031
-3.153
0.002
-3.153
0.002
-3.153
0.002
-3.153
0.002
Declared vs
Expanded
Z
p–value
1.000
1.000
1.000
1.000
1.000
1.000
NA
NA
NA
NA
1.413
0.500
1.000
1.000
NA
NA
-1.000
1.000
-2.617
0.016
-2.617
0.016
-2.431
0.031
-1.000
1.000
-2.431
0.031
-2.431
0.031
-2.431
0.031
-2.431
0.031
Declared vs
Common
Z
p-value
-0.221
0.819
-2.017
0.061
-1.079
0.306
-0.608
0.577
-0.608
0.577
1.783
0.075
1.344
0.183
-2.917
0.004
-1.462
0.146
-4.682
0.000
-4.412
0.000
-4.084
0.000
2.299
0.026
-4.682
0.000
-4.682
0.000
-4.772
0.000
-4.772
0.000
Declared vs
Common
Z
p-value
-0.500
1.000
-1.414
0.500
-0.577
1.000
0.138
0.750
0.138
0.750
0.138
0.750
-0.615
0.750
-1.730
0.250
-1.033
0.500
-1.995
0.125
-1.996
0.125
-1.996
0.125
-1.000
1.000
-1.995
0.125
-1.995
0.125
-2.226
0.063
-2.226
0.063
Expanded vs
Common
Z
p–value
-0.544
0.601
-2.199
0.036
-1.364
0.203
-0.608
0.577
-0.608
0.577
0.252
0.804
0.461
0.650
-2.917
0.004
-0.990
0.331
-2.225
0.025
-1.870
0.062
-1.735
0.083
3.025
0.002
-2.356
0.017
-2.363
0.017
-2.559
0.009
-2.559
0.009
Expanded vs
Common
Z
p–value
-1.414
0.500
-1.732
0.250
-1.414
0.500
0.138
0.750
0.138
0.750
-1.208
0.375
-1.033
0.500
-1.730
0.250
-0.034
1.000
1.313
0.227
1.671
0.094
1.361
0.191
0.046
1.000
1.089
0.313
1.115
0.266
0.855
0.426
0.855
0.426
Table 10: The results of the Wilcoxon signed rank test with Pratt adjustment for a given postulate evaluated on two
separate graphs on the total and core samples in Dialogue 1. The results have been obtained using R library coin. We
have highlighted the fields with p–value greater than 0.05, i.e. those for which we cannot reject the null hypothesis.
In case the adherence rate for a given postulate was identical for two graphs for all participants, the algorithm has
returned NA.
9.1.2 Statements vs. Relations: Effect of a Relation on its Source
Total Sample
-++
+
Core Sample
-++
+
G–value
153.455
DF
7
p–value
0.000
G–value
5.076
145.156
++
DF
7
7
p–value
0.651
0.000
G–value
89.014
15.290
82.477
+
DF
7
7
7
p–value
0.000
0.032
0.000
G–value
94.698
52.819
84.490
37.405
G–value
85.098
DF
7
p–value
0.000
G–value
10.231
38.538
++
DF
7
7
p–value
0.176
0.000
G–value
77.735
6.837
45.756
+
DF
7
7
7
p–value
0.000
0.446
0.000
G–value
69.700
26.354
30.842
35.980
?
DF
7
7
7
7
?
DF
7
7
7
7
p–value
0.000
0.000
0.000
0.000
p–value
0.000
0.000
0.000
0.000
Table 11: Results of G–test for independence between relations of given type on total and core samples. We use the
following abbreviations for the relation types: A good reason against (- -), A somewhat good reason against (-), A
good reason for (++), A somewhat good reason for (+), Somewhat related, but can’t say how (?). DF stands for degrees
of freedom. We have highlighted the fields with p–value greater than 0.05, i.e. those for which we cannot reject the
null hypothesis. These results were obtained using library Deducer (likelihood.test function) in R.
χ2
DF
p–value
-401.219
7
0.000
Total Sample
++
+
66.134 393.783 42.745
7
7
7
0.000
0.000
0.000
?
40.784
7
0.000
-340.297
7
0.000
Core Sample
++
+
44.036 108.651 57.959
7
7
7
0.000
0.000
0.000
?
17.267
7
0.016
Table 12: Results of chi–squared goodness of fit test on levels of agreement of sources of a relation of a given type in
Dialogue 1. We use the following abbreviations for the relation types: A good reason against (- -), A somewhat good
reason against (-), A good reason for (++), A somewhat good reason for (+), Somewhat related, but can’t say how (?).
DF stands for degrees of freedom. We have highlighted the fields with p–value greater than 0.05, i.e. those for which
we cannot reject the null hypothesis. These results were obtained using R.
Agreement pooled
according to strength
Strong Relation
Normal Relation
Agreement pooled
according to polarity
Strong Relation
Normal Relation
Agreement pooled
according to polarity
Attack
Support
Agreement pooled
according to strength
Attack
Support
Relations pooled according to strength
Total Sample
Core Sample
Normal Relation
Dependency
Normal Relation
Dependency
G–value DF p–value G–value DF p–value G–value DF p–value G–value DF p–value
145.715
3
0.000
89.373
3
0.000
66.202
3
0.000
61.424
3
0.000
47.221
3
0.000
19.351
3
0.000
Total Sample
Core Sample
Normal Relation
Dependency
Normal Relation
Dependency
G–value DF p–value G–value DF p–value G–value DF p–value G–value DF p–value
113.155
2
0.000
103.253
2
0.000
86.140
2
0.000
53.169
2
0.000
25.713
2
0.000
31.549
2
0.000
Relations pooled according to polarity
Total Sample
Core Sample
Support
Dependency
Support
Dependency
G–value DF p–value G–value DF p–value G–value DF p–value G–value DF p–value
0.673
2
0.714
48.684
2
0.000
4.906
2
0.086
35.441
2
0.000
54.445
2
0.000
29.528
2
0.000
Total Sample
Core Sample
Support
Dependency
Support
Dependency
G–value DF p–value G–value DF p–value G–value DF p–value G–value DF p–value
7.553
3
0.056
48.315
3
0.000
2.302
3
0.512
39.198
3
0.000
56.617
3
0.000
29.107
3
0.000
Table 13: Results of G–test for independence between relations of given type in Dialogue 1 on total and core samples
according to a given pooling. DF stands for degrees of freedom. We have highlighted the fields with p–value greater
than 0.05, i.e. those for which we cannot reject the null hypothesis. These results were obtained using library Deducer
(likelihood.test function) in R.
Agreement pooled
according to strength
χ2
DF
p–value
Agreement pooled
according to polarity
χ2
DF
p–value
Agreement pooled
according to polarity
χ2
DF
p–value
Agreement pooled
according to strength
χ2
DF
p–value
Strong Relation
472.913
3
0.000
Strong Relation
519.513
2
0.000
Attack
272.556
2
0.000
Attack
184.929
3
0.000
Relations pooled according to strength
Total Sample
Normal Relation Dependency Strong Relation
78.459
30.627
297.046
3
3
3
0.000
0.000
0.000
Total Sample
Normal Relation Dependency Strong Relation
136.832
11.412
185.490
2
2
2
0.000
0.003
0.000
Relations pooled according to polarity
Total Sample
Support
Dependency
Attack
251.086
11.412
120.544
2
2
2
0.000
0.003
0.000
Total Sample
Support
Dependency
Attack
229.567
30.627
185.019
3
3
3
0.000
0.000
0.000
Core Sample
Normal Relation
46.399
3
0.000
Core Sample
Normal Relation
97.475
2
0.000
Core Sample
Support
64.511
2
0.000
Core Sample
Support
79.000
3
0.000
Dependency
9.853
3
0.020
Dependency
0.960
2
0.619
Dependency
0.960
2
0.619
Dependency
9.853
3
0.020
Table 14: Results of chi–squared goodness of fit test on levels of agreement of sources of a relation of a given type
in Dialogue 1 on total and core samples according to a given pooling. DF stands for degrees of freedom. We have
highlighted the fields with p–value greater than 0.05, i.e. those for which we cannot reject the null hypothesis. These
results were obtained using R.
9.1.3 Statements vs. Relations: Effect of an Argument on the Relations it Carries Out
Total
sample
SD
D
SoD
NAD
DK
SoA
A
Core
sample
SD
D
SoD
NAD
DK
SoA
A
D
G–value
34.873
p–value
0.000
SoD
G–value p–value
35.761
0.000
3.452
0.485
NAD
G–value p–value
44.077
0.000
19.668
0.001
25.352
0.000
DK
G–value p–value
17.742
0.001
13.085
0.011
19.199
0.001
4.571
0.334
SoA
G–value p–value
33.778
0.000
8.647
0.071
6.513
0.164
38.749
0.000
25.826
0.000
p–value
0.022
SoD
G–value p–value
5.418
0.247
2.413
0.660
NAD
G–value p–value
23.947
0.000
19.445
0.001
13.304
0.010
DK
G–value p–value
28.026
0.000
29.463
0.000
20.222
0.000
6.920
0.140
SoA
G–value p–value
2.570
0.632
0.477
0.976
0.351
0.986
4.976
0.290
6.329
0.176
D
G–value
11.457
A
G–value
11.239
61.659
63.386
59.589
25.004
47.747
p–value
0.024
0.000
0.000
0.000
0.000
0.000
A
G–value
5.042
25.615
13.042
22.987
19.897
4.931
p–value
0.283
0.000
0.011
0.000
0.001
0.294
SA
G–value p–value
23.965
0.000
118.932
0.000
105.564
0.000
97.628
0.000
40.261
0.000
80.135
0.000
13.623
0.009
SA
G–value p–value
36.282
0.000
89.884
0.000
51.616
0.000
50.419
0.000
50.310
0.000
14.571
0.006
20.806
0.000
Table 15: Results of G–test for independence between different argument acceptance levels in Dialogue 1 on total and
core samples. In all cases we have obtained 4 degrees of freedom. We use the following abbreviations: Strongly Agree
(SA), Agree (A), Somewhat Agree (SoA), Neither Agree nor Disagree (NAD), Somewhat Disagree (SoD), Disagree
(D), Strongly Disagree (SD), Don’t Know (DK). DF stands for degrees of freedom. We have highlighted the fields
with p–value greater than 0.05, i.e. those for which we cannot reject the null hypothesis. These results were obtained
using library Deducer (likelihood.test function) in R.
χ2
DF
p–value
SD
11.314
4
0.023
D
32.225
4
0.000
SoD
33.057
4
0.000
Total Sample
NAD
DK
35.972 14.727
4
4
0.000
0.005
SoA
35.980
4
0.000
A
57.000
4
0.000
SA
171.296
4
0.000
SD
12.239
4
0.016
D
23.340
4
0.000
SoD
7.296
4
0.121
Core Sample
NAD
DK
13.909 25.600
4
4
0.008
0.000
SoA
2.000
4
0.736
A
18.706
4
0.001
SA
190.611
4
0.000
Table 16: Results of chi–squared goodness of fit test on relations carried out by arguments of a given acceptance level
in Dialogue 1. We use the following abbreviations: Strongly Agree (SA), Agree (A), Somewhat Agree (SoA), Neither
Agree nor Disagree (NAD), Somewhat Disagree (SoD), Disagree (D), Strongly Disagree (SD), Don’t Know (DK). DF
stands for degrees of freedom. We have highlighted the fields with p–value greater than 0.05, i.e. those for which we
cannot reject the null hypothesis. These results were obtained using R.
Relations pooled
according to strength
Strong Belief
Moderate Belief
Weak Belief
Relations pooled
according to polarity
Strong Belief
Moderate Belief
Weak Belief
Moderate Belief
G–value p–value
36.110
0.000
Moderate Belief
G–value p–value
7.965
0.019
Relations pooled
according to strength
Disbelieved
Believed
Relations pooled
according to polarity
Disbelieved
Believed
Agreement pooled according to strength
Total Sample
Core Sample
Weak Belief
Neither
Moderate Belief
Weak Belief
G–value p–value G–value p–value G–value p–value G–value p–value
122.613
0.000
100.750
0.000
40.240
0.000
33.854
0.000
44.053
0.000
52.513
0.000
2.165
0.339
49.153
0.000
Total Sample
Core Sample
Weak Belief
Neither
Moderate Belief
Weak Belief
G–value p–value G–value p–value G–value p–value G–value p–value
5.437
0.066
56.972
0.000
4.124
0.127
2.060
0.357
4.981
0.083
37.544
0.000
0.229
0.892
49.160
0.000
Agreement pooled according to polarity
Total Sample
Core Sample
Believed
Neither
Believed
Neither
G–value p–value G–value p–value G–value p–value G–value p–value
98.123
0.000
40.289
0.000
74.909
0.000
37.864
0.000
100.516
0.000
62.282
0.000
Total Sample
Core Sample
Believed
Neither
Believed
Neither
G–value p–value G–value p–value G–value p–value G–value p–value
5.811
0.055
37.922
0.000
5.042
0.080
36.045
0.000
62.659
0.000
34.388
0.000
Neither
G–value p–value
65.784
0.000
29.473
0.000
20.024
0.000
Neither
G–value p–value
41.822
0.000
25.974
0.000
19.549
0.000
Table 17: Results of G–test for independence between different argument acceptance levels in Dialogue 1 on total and
core samples according to a given pooling. In all cases we have obtained 2 degrees of freedom. We have highlighted
the fields with p–value greater than 0.05, i.e. those for which we cannot reject the null hypothesis. These results were
obtained using library Deducer (likelihood.test function) in R.
Relations pooled
according to strength
χ2
DF
p–value
Relations pooled
according to polarity
χ2
DF
p–value
Agreement pooled according to strength
Total Sample
Core Sample
Strong Belief Moderate Belief Weak Belief Neither Strong Belief Moderate Belief Weak Belief
280.351
75.107
102.118
13.865
192.360
33.832
17.645
2
2
2
2
2
2
2
0.000
0.000
0.000
0.001
0.000
0.000
0.000
Total Sample
Core Sample
Strong Belief Moderate Belief Weak Belief Neither Strong Belief Moderate Belief Weak Belief
121.032
74.476
63.430
7.000
116.919
43.486
15.516
2
2
2
2
2
2
2
0.000
0.000
0.000
0.030
0.000
0.000
0.000
Agreement pooled according to polarity
Relations pooled
Total Sample
Core Sample
according to strength Disbelieved Believed Neither Disbelieved Believed Neither
χ2
84.105
333.680 13.865
65.475
197.077
8.340
DF
2
2
2
2
2
2
p–value
0.000
0.000
0.001
0.000
0.000
0.015
Relations pooled
Total Sample
Core Sample
according to polarity
Disbelieved Believed Neither Disbelieved Believed Neither
χ2
69.584
180.390
7.000
69.475
107.868
6.298
DF
2
2
2
2
2
2
p–value
0.000
0.000
0.030
0.000
0.000
0.043
Neither
8.340
2
0.015
Neither
6.298
2
0.043
Table 18: Results of chi–squared goodness of fit test on relations carried out by arguments of a given acceptance level
in Dialogue 1 on total and core samples according to a given pooling. DF stands for degrees of freedom. We have
highlighted the fields with p–value greater than 0.05, i.e. those for which we cannot reject the null hypothesis. These
results were obtained using R.
9.2 Dialogue 2
9.2.1 Postulate Satisfaction
Total Sample
Preferential
Rational
Strict
Protective
Restrained
Coherent
Involutary
Justifiable
SemiOptimistic
SemiFounded
Founded
Optimistic
Demanding
Guarded
Discharging
Trusting
Anticipating
Core Sample
Preferential
Rational
Strict
Protective
Restrained
Coherent
Involutary
Justifiable
SemiOptimistic
SemiFounded
Founded
Optimistic
Demanding
Guarded
Discharging
Trusting
Anticipating
Intended vs
Declared
Z
p–value
3.231
0.001
3.021
0.002
1.957
0.050
2.410
0.015
2.019
0.043
2.019
0.043
0.465
0.641
2.236
0.063
-3.196
0.001
4.291
0.000
2.825
0.008
2.827
0.008
1.363
0.250
4.072
0.000
3.365
0.000
4.208
0.000
3.636
0.000
Intended vs
Declared
Z
p–value
2.226
0.063
2.224
0.063
2.025
0.063
2.124
0.035
1.983
0.051
2.215
0.027
0.961
0.625
NA
NA
-2.647
0.008
2.224
0.063
1.000
1.000
NA
NA
1.729
0.250
1.993
0.125
0.745
0.594
1.730
0.250
0.659
0.500
Intended vs
Expanded
Z
p–value
3.256
0.001
3.035
0.002
2.550
0.010
2.485
0.012
2.026
0.042
3.794
0.000
2.425
0.015
2.000
0.125
-3.834
0.000
0.389
0.704
-0.971
0.315
0.284
0.918
-0.508
0.684
0.194
0.850
-0.157
0.880
0.916
0.363
0.346
0.735
Intended vs
Expanded
Z
p–value
2.226
0.063
2.224
0.063
2.025
0.063
2.124
0.035
1.983
0.051
2.647
0.006
1.594
0.219
NA
NA
-3.056
0.002
-1.061
0.359
-0.659
0.500
-1.000
1.000
0.618
0.750
-1.393
0.188
-1.687
0.102
-0.130
0.906
-0.809
0.531
Intended vs
Common
Z
p–value
2.995
0.004
2.642
0.016
3.585
0.000
4.295
0.000
3.718
0.000
4.190
0.000
2.235
0.063
NA
NA
-4.498
0.000
NA
NA
NA
NA
NA
NA
1.999
0.125
-1.732
0.250
-2.824
0.008
-1.414
0.500
-1.999
0.125
Intended vs
Common
Z
p–value
1.994
0.125
1.729
0.250
1.993
0.125
2.924
0.004
2.229
0.063
2.777
0.008
1.414
0.500
NA
NA
-3.172
0.001
NA
NA
NA
NA
NA
NA
1.730
0.250
NA
NA
-1.413
0.500
NA
NA
-1.000
1.000
Declared vs
Expanded
Z
p–value
1.414
0.500
1.000
1.000
1.999
0.125
1.000
1.000
1.000
1.000
3.845
0.000
3.154
0.002
-1.000
1.000
-2.643
0.016
-3.963
0.000
-3.156
0.002
-2.644
0.016
-1.867
0.094
-3.842
0.000
-3.841
0.000
-3.588
0.000
-3.305
0.001
Declared vs
Expanded
Z
p–value
NA
NA
NA
NA
NA
NA
NA
NA
NA
NA
2.229
0.063
1.414
0.500
NA
NA
-1.994
0.125
-2.433
0.031
-1.413
0.500
-1.000
1.000
-1.000
1.000
-2.429
0.031
-2.427
0.031
-1.729
0.250
-1.729
0.250
Declared vs
Common
Z
p-value
-1.632
0.132
-2.029
0.050
-0.143
0.892
0.378
0.712
-0.131
0.905
0.472
0.640
0.497
0.613
-2.236
0.063
-1.062
0.308
-4.291
0.000
-2.825
0.008
-2.827
0.008
-0.255
1.000
-4.298
0.000
-4.500
0.000
-4.330
0.000
-4.104
0.000
Declared vs
Common
Z
p-value
-1.000
1.000
-1.413
0.500
-0.961
0.625
0.435
0.703
-0.411
0.813
-0.494
1.000
0.049
1.000
NA
NA
-0.960
0.500
-2.224
0.063
-1.000
1.000
NA
NA
0.049
1.000
-1.993
0.125
-1.994
0.125
-1.730
0.250
-1.413
0.500
Expanded vs
Common
Z
p–value
-1.849
0.080
-2.044
0.048
-0.941
0.351
0.307
0.766
-0.182
0.863
-2.421
0.015
-1.914
0.063
-2.000
0.125
0.416
0.757
-0.389
0.704
0.971
0.315
-0.284
0.918
1.275
0.173
-0.631
0.535
-0.888
0.381
-1.259
0.213
-0.986
0.331
Expanded vs
Common
Z
p–value
-1.000
1.000
-1.413
0.500
-0.961
0.625
0.435
0.703
-0.411
0.813
-1.679
0.125
-0.961
0.625
NA
NA
0.535
1.000
1.061
0.359
0.659
0.500
1.000
1.000
0.659
0.500
1.393
0.188
1.359
0.191
0.130
0.906
0.513
0.688
Table 19: The results of the Wilcoxon signed rank test with Pratt adjustment for a given postulate evaluated on two
separate graphs on the total and core samples in Dialogue 2. The results have been obtained using R library coin. We
have highlighted the fields with p–value greater than 0.05, i.e. those for which we cannot reject the null hypothesis.
In case the adherence rate for a given postulate was identical for two graphs for all participants, the algorithm has
returned NA.
9.2.2 Statements vs. Relations: Effect of a Relation on its Source
Total Sample
-++
+
Core Sample
-++
+
G–value
146.771
DF
7
p–value
0.000
G–value
44.558
70.229
++
DF
7
7
p–value
0.000
0.000
G–value
112.838
49.829
82.165
+
DF
7
7
7
p–value
0.000
0.000
0.000
G–value
122.577
125.574
110.601
90.918
G–value
86.269
DF
7
p–value
0.000
G–value
7.506
70.549
++
DF
7
7
p–value
0.378
0.000
G–value
83.920
18.330
74.209
+
DF
7
7
7
p–value
0.000
0.011
0.000
G–value
123.81499
89.734
110.2263
42.580
?
DF
7
7
7
7
?
DF
7
7
7
7
p–value
0.000
0.000
0.000
0.000
p–value
0.000
0.000
0.000
0.000
Table 20: Results of G–test for independence between relations of given type in Dialogue 2 on total and core samples.
We use the following abbreviations for the relation types: A good reason against (- -), A somewhat good reason against
(-), A good reason for (++), A somewhat good reason for (+), Somewhat related, but can’t say how (?). DF stands
for degrees of freedom. We have highlighted the fields with p–value greater than 0.05, i.e. those for which we cannot
reject the null hypothesis. These results were obtained using library Deducer (likelihood.test function) in R.
χ2
DF
p–value
-220.015
7
0.000
Total Sample
++
+
183.075 146.260 87.918
7
7
7
0.000
0.000
0.000
?
378.960
7
0.000
-134.442
7
0.000
Core Sample
++
+
59.454 79.750 46.842
7
7
7
0.000
0.000
0.000
?
396.177
7
0.000
Table 21: Results of chi–squared goodness of fit test on levels of agreement of sources of a relation of a given type in
Dialogue 2. We use the following abbreviations for the relation types: A good reason against (- -), A somewhat good
reason against (-), A good reason for (++), A somewhat good reason for (+), Somewhat related, but can’t say how (?).
DF stands for degrees of freedom. We have highlighted the fields with p–value greater than 0.05, i.e. those for which
we cannot reject the null hypothesis. These results were obtained using R.
Agreement pooled
according to strength
Strong Relation
Normal Relation
Agreement pooled
according to polarity
Strong Relation
Normal Relation
Agreement pooled
according to polarity
Attack
Support
Agreement pooled
according to strength
Attack
Support
Relations pooled according to strength
Total Sample
Core Sample
Normal Relation
Dependency
Normal Relation
Dependency
G–value DF p–value G–value DF p–value G–value DF p–value G–value DF p–value
99.442
3
0.000
120.905
3
0.000
98.797
3
0.000
127.352
3
0.000
127.914
3
0.000
54.266
3
0.000
Total Sample
Core Sample
Normal Relation
Dependency
Normal Relation
Dependency
G–value DF p–value G–value DF p–value G–value DF p–value G–value DF p–value
38.622
2
0.000
108.870
2
0.000
14.090
2
0.001
120.824
2
0.000
127.904
2
0.000
55.808
2
0.000
Relations pooled according to polarity
Total Sample
Core Sample
Support
Dependency
Support
Dependency
G–value DF p–value G–value DF p–value G–value DF p–value G–value DF p–value
3.595
2
0.166
122.975
2
0.000
9.883
2
0.007
107.440
2
0.000
104.779
2
0.000
78.490
2
0.000
Total Sample
Core Sample
Support
Dependency
Support
Dependency
G–value DF p–value G–value DF p–value G–value DF p–value G–value DF p–value
9.338
3
0.025
128.599
3
0.000
0.965
3
0.810
104.209
3
0.000
106.480
3
0.000
82.158
3
0.000
Table 22: Results of G–test for independence between relations of given type in Dialogue 2 on total and core samples
according to a given pooling. DF stands for degrees of freedom. We have highlighted the fields with p–value greater
than 0.05, i.e. those for which we cannot reject the null hypothesis. These results were obtained using library Deducer
(likelihood.test function) in R.
Agreement pooled
according to strength
χ2
DF
p–value
Agreement pooled
according to polarity
χ2
DF
p–value
Agreement pooled
according to polarity
χ2
DF
p–value
Agreement pooled
according to strength
χ2
DF
p–value
Strong Relation
118.474
3
0.000
Strong Relation
108.641
2
0.000
Attack
63.392
2
0.000
Attack
112.872
3
0.000
Relations pooled according to strength
Total Sample
Normal Relation Dependency Strong Relation
145.419
278.048
38.630
3
3
3
0.000
0.000
0.000
Total Sample
Normal Relation Dependency Strong Relation
10.771
141.128
61.437
2
2
2
0.005
0.000
0.000
Relations pooled according to polarity
Total Sample
Support
Dependency
Attack
24.454
141.128
59.744
2
2
2
0.000
0.000
0.000
Total Sample
Support
Dependency
Attack
83.463
278.048
22.326
3
3
3
0.000
0.000
0.000
Core Sample
Normal Relation
73.749
3
0.000
Core Sample
Normal Relation
7.732
2
0.021
Core Sample
Support
4.805
2
0.090
Core Sample
Support
20.270
3
0.000
Dependency
188.726
3
0.000
Dependency
106.235
2
0.000
Dependency
106.235
2
0.000
Dependency
188.726
3
0.000
Table 23: Results of chi–squared goodness of fit test on levels of agreement of sources of a relation of a given type
in Dialogue 2 on total and core samples according to a given type of pooling. DF stands for degrees of freedom. We
have highlighted the fields with p–value greater than 0.05, i.e. those for which we cannot reject the null hypothesis.
These results were obtained using R.
9.2.3 Statements vs. Relations: Effect of an Argument on the Relations it Carries Out
Total
sample
SD
D
SoD
NAD
DK
SoA
A
Core
sample
SD
D
SoD
NAD
DK
SoA
A
D
G–value
88.802
p–value
0.000
SoD
G–value p–value
41.240
0.000
46.449
0.000
NAD
G–value p–value
57.479
0.000
116.328
0.000
40.291
0.000
DK
G–value p–value
39.781
0.000
72.405
0.000
45.535
0.000
10.459
0.033
SoA
G–value p–value
39.381
0.000
30.161
0.000
5.529
0.237
50.145
0.000
41.020
0.000
p–value
0.000
SoD
G–value p–value
37.146
0.000
4.049
0.399
NAD
G–value p–value
64.903
0.000
26.022
0.000
21.170
0.000
DK
G–value p–value
83.274
0.000
10.662
0.031
11.723
0.020
43.599
0.000
SoA
G–value p–value
48.583
0.000
6.652
0.155
11.731
0.019
56.900
0.000
18.957
0.001
D
G–value
48.190
A
G–value
18.469
72.721
46.450
77.095
41.392
37.557
p–value
0.001
0.000
0.000
0.000
0.000
0.000
A
G–value
19.065
22.830
17.529
72.986
53.371
17.817
p–value
0.001
0.000
0.002
0.000
0.000
0.001
SA
G–value p–value
15.769
0.003
129.046
0.000
60.854
0.000
74.098
0.000
55.609
0.000
68.731
0.000
21.771
0.000
SA
G–value p–value
6.361
0.174
61.635
0.000
46.062
0.000
87.561
0.000
94.664
0.000
55.203
0.000
20.983
0.000
Table 24: Results of G–test for independence between different argument acceptance levels in Dialogue 2 on total and
core samples. In all cases we have obtained 4 degrees of freedom. We use the following abbreviations: Strongly Agree
(SA), Agree (A), Somewhat Agree (SoA), Neither Agree nor Disagree (NAD), Somewhat Disagree (SoD), Disagree
(D), Strongly Disagree (SD), Don’t Know (DK). DF stands for degrees of freedom. We have highlighted the fields
with p–value greater than 0.05, i.e. those for which we cannot reject the null hypothesis. These results were obtained
using library Deducer (likelihood.test function) in R.
χ2
DF
p–value
SD
87.128
4
0.000
D
117.032
4
0.000
SoD
22.202
4
0.000
Total Sample
NAD
DK
36.956 37.706
4
4
0.000
0.000
SoA
26.618
4
0.000
A
137.961
4
0.000
SA
156.983
4
0.000
SD
90.706
4
0.000
D
5.596
4
0.231
SoD
7.067
4
0.132
Core Sample
NAD
DK
59.016 16.133
4
4
0.000
0.003
SoA
32.727
4
0.000
A
98.417
4
0.000
SA
140.519
4
0.000
Table 25: Results of chi–squared goodness of fit test on relations carried out by arguments of a given acceptance level
in Dialogue 2. We use the following abbreviations: Strongly Agree (SA), Agree (A), Somewhat Agree (SoA), Neither
Agree nor Disagree (NAD), Somewhat Disagree (SoD), Disagree (D), Strongly Disagree (SD), Don’t Know (DK). DF
stands for degrees of freedom. We have highlighted the fields with p–value greater than 0.05, i.e. those for which we
cannot reject the null hypothesis. These results were obtained using R.
Relations pooled
according to strength
Strong Belief
Moderate Belief
Weak Belief
Relations pooled
according to polarity
Strong Belief
Moderate Belief
Weak Belief
Moderate Belief
G–value p–value
64.174
0.000
Moderate Belief
G–value p–value
1.847
0.397
Relations pooled
according to strength
Disbelieved
Believed
Relations pooled
according to polarity
Disbelieved
Believed
Agreement pooled according to strength
Total Sample
Core Sample
Weak Belief
Neither
Moderate Belief
Weak Belief
G–value p–value G–value p–value G–value p–value G–value p–value
89.900
0.000
101.892
0.000
55.723
0.000
88.655
0.000
7.532
0.023
97.014
0.000
10.005
0.007
76.925
0.000
Total Sample
Core Sample
Weak Belief
Neither
Moderate Belief
Weak Belief
G–value p–value G–value p–value G–value p–value G–value p–value
9.971
0.007
78.619
0.000
6.068
0.048
5.081
0.079
7.261
0.026
88.552
0.000
0.000
1.000
49.026
0.000
Agreement pooled according to polarity
Total Sample
Core Sample
Believed
Neither
Believed
Neither
G–value p–value G–value p–value G–value p–value G–value p–value
34.173
0.000
100.464
0.000
8.874
0.012
43.919
0.000
104.099
0.000
114.249
0.000
Total Sample
Core Sample
Believed
Neither
Believed
Neither
G–value p–value G–value p–value G–value p–value G–value p–value
2.630
0.268
74.999
0.000
15.702
0.000
40.888
0.000
106.419
0.000
104.415
0.000
Neither
G–value p–value
146.970
0.000
59.026
0.000
40.821
0.000
Neither
G–value p–value
72.506
0.000
55.628
0.000
39.409
0.000
Table 26: Results of G–test for independence between different argument acceptance levels in Dialogue 2 on total
and core samples according to a given type of pooling. In all cases we have obtained 2 degrees of freedom. We have
highlighted the fields with p–value greater than 0.05, i.e. those for which we cannot reject the null hypothesis. These
results were obtained using library Deducer (likelihood.test function) in R.
Relations pooled
according to strength
χ2
DF
p–value
Relations pooled
according to polarity
χ2
DF
p–value
Agreement pooled according to strength
Total Sample
Core Sample
Strong Belief Moderate Belief Weak Belief Neither Strong Belief Moderate Belief Weak Belief
348.713
273.129
134.568
116.882
286.052
136.630
55.665
2
2
2
2
2
2
2
0.000
0.000
0.000
0.000
0.000
0.000
0.000
Total Sample
Core Sample
Strong Belief Moderate Belief Weak Belief Neither Strong Belief Moderate Belief Weak Belief
168.530
271.362
134.827
50.852
94.518
107.849
65.200
2
2
2
2
2
2
2
0.000
0.000
0.000
0.000
0.000
0.000
0.000
Agreement pooled according to polarity
Relations pooled
Total Sample
Core Sample
according to strength Disbelieved Believed Neither Disbelieved Believed Neither
χ2
213.405
463.158 116.882
105.506
281.212 11.626
DF
2
2
2
2
2
2
p–value
0.000
0.000
0.000
0.000
0.000
0.003
Relations pooled
Total Sample
Core Sample
according to polarity
Disbelieved Believed Neither Disbelieved Believed Neither
χ2
213.633
351.965
50.852
68.486
212.382
8.311
DF
2
2
2
2
2
2
p–value
0.000
0.000
0.000
0.000
0.000
0.016
Neither
11.626
2
0.003
Neither
8.311
2
0.016
Table 27: Results of chi–squared goodness of fit test on relations carried out by arguments of a given acceptance level
in Dialogue 2 on total and core samples according to a given pooling. DF stands for degrees of freedom. We have
highlighted the fields with p–value greater than 0.05, i.e. those for which we cannot reject the null hypothesis. These
results were obtained using R.
Total
Core
Total
Core
Dialogue 2
Dialogue 1
9.3 Dialogues 1& 2: Changes in Beliefs
Argument
Between
steps
1 and 2
2 and 3
3 and 4
4 and 5
1 and 2
2 and 3
3 and 4
4 and 5
1 and 2
2 and 3
3 and 4
4 and 5
1 and 2
2 and 3
3 and 4
4 and 5
A
B
C
Z
pvalue
Z
pvalue
-0.551
1.890
-1.986
2.157
0.069
1.732
-0.577
1.000
-0.378
-0.794
-2.448
-0.013
0.000
1.413
-1.995
0.494
0.716
0.125
0.063
0.039
1.000
0.250
1.000
1.000
1.000
0.656
0.031
1.000
1.000
0.500
0.125
1.000
-5.997
5.846
-5.826
5.773
-3.726
3.792
-3.873
3.873
-4.126
-1.102
0.740
0.764
0.577
-3.011
3.372
1.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.406
0.688
0.445
1.000
0.003
0.000
1.000
D
E
F
Z
pvalue
Z
pvalue
Z
pvalue
1.783
0.780
1.964
0.085
0.544
0.051
-5.351
-1.134
0.000
0.453
2.875
0.005
0.000
2.000
0.046
1.000
0.125
1.000
-0.069
-3.553
1.000
0.000
1.513
0.188
5.449
-4.781
1.297
0.000
0.000
0.227
0.622
-0.839
1.154
0.627
0.440
0.359
1.635
-0.952
3.316
-1.717
-0.647
0.000
0.125
0.564
1.413
-1.466
1.996
0.500
0.234
0.125
-2.653
-0.960
G
Z
pvalue
Z
pvalue
0.156
0.434
0.349
-1.393
0.699
0.227
0.421
0.591
0.003
0.500
-2.232
-0.576
0.033
0.750
1.308
0.219
Table 28: The results of the Wilcoxon signed rank test with Pratt adjustment between the acceptance levels declared
by the participants at steps i and i + 1 for a given argument, in a given dialogue and sample. The results have been
obtained using R library coin. We have highlighted the fields with p–value greater than 0.05, i.e. those for which we
cannot reject the null hypothesis. Due to the fact that arguments F and H appear only at the last stages of Dialogues
1 and 2 respectively, they possess only a single distribution and the change analysis cannot be performed for them.
| 2cs.AI
|
arXiv:1507.05272v1 [cs.CE] 19 Jul 2015
Under consideration for publication in Theory and Practice of Logic Programming
1
Optimizing Phylogenetic Supertrees Using
Answer Set Programming
LAURA KOPONEN and EMILIA OIKARINEN and TOMI JANHUNEN
HIIT and Department of Computer Science
Aalto University
P.O. Box 15400, FI-00076 AALTO, Finland
(e-mail: {Laura.J.Koponen, Emilia.Oikarinen, Tomi.Janhunen}@aalto.fi)
LAURA SÄILÄ
Department of Geosciences and Geography
University of Helsinki
P.O. Box 64, FI-00014 University of Helsinki, Finland
(e-mail: [email protected])
submitted 1 January 2003; revised 1 January 2003; accepted 1 January 2003
Abstract
The supertree construction problem is about combining several phylogenetic trees with
possibly conflicting information into a single tree that has all the leaves of the source
trees as its leaves and the relationships between the leaves are as consistent with the
source trees as possible. This leads to an optimization problem that is computationally
challenging and typically heuristic methods, such as matrix representation with parsimony
(MRP), are used. In this paper we consider the use of answer set programming to solve the
supertree construction problem in terms of two alternative encodings. The first is based
on an existing encoding of trees using substructures known as quartets, while the other
novel encoding captures the relationships present in trees through direct projections. We
use these encodings to compute a genus-level supertree for the family of cats (Felidae).
Furthermore, we compare our results to recent supertrees obtained by the MRP method.
KEYWORDS: answer set programming, phylogenetic supertree, quartets, projections, Felidae
1 Introduction
In the supertree construction problem, one is given a set of phylogenetic trees
(source trees) with overlapping sets of leaf nodes (representing taxa) and the goal
is to construct a single tree that respects the relationships in individual source trees
as much as possible (Bininda-Emonds 2004). The concept of respecting the relationships in the source trees varies depending on the particular supertree method
at hand. If the source trees are compatible, i.e., there is no conflicting information regarding the relationships of taxa in the source trees, then supertree construction is easy (Aho et al. 1981). However, this is rarely the case. It is typical that source trees obtained from different studies contain conflicting informa-
2
L. Koponen et al.
tion, which makes supertree optimization a computationally challenging problem
(Foulds and Graham 1982; Day et al. 1986; Byrka et al. 2010).
One of the most widely used supertree methods is matrix representation with parsimony (MRP) (Baum 1992; Ragan 1992) in which source trees are encoded into a
binary matrix, and maximum parsimony analysis is then used to construct a tree.
Other popular methods include matrix representation with flipping (Chen et al. 2003)
and MinCut supertrees (Semple and Steel 2000). There is some criticism towards
the accuracy and performance of MRP, indicating input tree size and shape biases
and varying results depending on the chosen matrix representation (Purvis 1995;
Wilkinson et al. 2005; Goloboff and Pol 2002). An alternative approach is to directly consider the topologies induced by the source trees, for instance, using quartets (Piaggio-Talice et al. 2004) or triplets (Bryant 1997), and try to maximize the
satisfaction of these topologies resulting in maximum quartet (resp. rooted triplet)
consistency problem. The quartet-based methods have received increasing interest
over the last few years (Snir and Rao 2012) and the quality of supertrees produced
have been shown to be on a par with MRP trees (Swenson et al. 2011).
There are a number of constraint-based approaches tailored for the phylogeny
reconstruction problem (Kavanagh et al. 2006; Brooks et al. 2007; Wu et al. 2007;
Sridhar et al. 2008; Morgado and Marques-Silva 2010). In phylogeny reconstruction, one is given a set of sequences (for instance gene data) or topologies (for instance quartets) as input and the task is to build a phylogenetic tree that represents
the evolutionary history of the species represented by the input. In (Brooks et al. 2007),
answer set programming (ASP) is used to find cladistics-based phylogenies, and in
(Kavanagh et al. 2006; Sridhar et al. 2008) maximum parsimony criteria are applied, using ASP and mixed integer programming (MIP), respectively. The most
closely related approach to our work is the one in (Wu et al. 2007) where an ASP
encoding for solving the maximum quartet consistency problem for phylogeny reconstruction is presented. The difference to supertree optimization is that in phylogeny reconstruction, typically almost all possible quartets over all sets of four
taxa are available, with possibly some errors. In supertree optimization the overlap of source trees is limited and the number of quartets obtained from source
trees is much smaller than the number of possible quartets for the supertree.
For example, the supertree shown in Figure 2 (right), with 34 leaf nodes, displays 46 038 different quartets, while the source trees used to construct it only
contributed 11 319 distinct quartets, some of which were mutually incompatible.
In (Morgado and Marques-Silva 2010) a constraint programming solution is introduced for the maximum quartet consistency problem. There are also related studies
of supertree optimization based on constraint reasoning. In (Chimani et al. 2010)
a MIP solution for minimum flip supertrees is presented, and in (Gent et al. 2003)
constraint programming is used to produce min-ultrametric trees using triplets.
However, in both cases the underlying problem is polynomially solvable. Furthermore, ASP has also been used to formalize phylogeny-related queries in (Le et al. 2012).
In this paper we solve the supertree optimization problem in terms of two alternative ASP encodings. The first encoding is based on quartets and is similar
to the one in (Wu et al. 2007), though instead of using an ultrametric matrix, we
Optimizing Phylogenetic Supertrees Using Answer Set Programming
outgroup
Felis
Lynx
Panthera
Puma
3
outgroup
Leopardus
Panthera
Acinonyx
Lynx
Felis
Puma
Fig. 1. Phylogenetic trees from (Fulton and Strobeck 2006) on the left and from
(Flynn et al. 2005) on the right, abstracted to genus-level (for more details, see
Section 4).
use a direct encoding to obtain the tree topology. However, the performance of the
quartet-based encoding does not scale up. Our second encoding uses a novel approach capturing the relationships present in trees through projections, formalized
in terms of the maximum projection consistency problem. We use these encodings
to compute a genus-level supertree for the family of cats (Felidae) and compare our
results to recent supertrees obtained from the MRP method.
The rest of this paper is organized as follows. We present the supertree problem
in Section 2, and introduce our encodings for supertree optimization in Section 3.
In Section 4, we first compare the efficiency of the encodings, and then use the
projection-based encoding to compute a genus-level supertree for the family of cats
(Felidae). We compare our supertrees to recent supertrees obtained using the MRP
method. Finally, we present our conclusions in Section 5.
2 Supertree problem
A phylogenetic tree of n taxa has exactly n leaf nodes, each corresponding to one
taxon. The tree may be rooted or unrooted. In this work we consider rooted trees
and assume that the root has a special taxon called outgroup as its child. An inner
node is resolved if it has exactly two children, otherwise it is unresolved. If a tree
contains any unresolved nodes, it is unresolved; otherwise, it is resolved. Resolution
is the ratio of resolved inner nodes in a phylogenetic tree. A higher resolution is
preferred, as this means that more is known about the relationships of the taxa.
The problem of combining a set of phylogenetic trees with (partially) overlapping
sets of taxa into a single tree is known as the supertree construction problem. In
the special case where each source tree contains exactly the same set of species, it is
also called the consensus tree problem (Steel et al. 2000). In order to combine trees
with different taxa, one needs a way to split the source trees into smaller structures
which describe the relationships in the trees at the same time. There are several
ways to achieve this, for instance by using triplets (rooted substructures with three
leaf nodes) or quartets (unrooted substructures with four leaf nodes).
A quartet (topology) is an unrooted topological substructure of a tree. The quartet ((I , J ), (K , L)) is in its canonical representation if I < J , I < K , and K < L,
where “<” refers to the alphabetical ordering of the names of the taxa. From now
on, we will consider canonical representations of quartets. We say that a tree T
4
L. Koponen et al.
displays a quartet ((I , J ), (K , L)), if there is an edge in the tree T that separates T
into two subtrees so that one subtree contains the pair I and J as its leaves and the
other subtree contains the pair K and L as its leaves. For any set of four taxa appearing in a resolved phylogenetic tree T , there is exactly one quartet displayed by
T . Furthermore, we say that two phylogenetic trees T and T ′ are not compatible,
if there is a set of four taxa for which T and T ′ display a different quartet.
Example 1
Consider the two phylogenetic trees in Figure 1. It is easy to see that these trees are
not compatible. For the taxa Felis, Lynx, Panthera, and Puma, the tree on the left
displays the quartet ((Felis,Lynx),(Panthera,Puma)), while the tree on the right
displays the quartet ((Felis,Puma),(Lynx,Panthera)).
Let tx(T ) denote the set of taxa in the leaves of a tree T and qt(T ) the set
of all quartets that are displayed by T . For a collection S of phylogenetic trees,
S
S
we define qt(S ) as the multiset1 T ∈S qt(T ) and tx(S ) = T ∈S tx(T ). Given any
phylogenetic tree T , the set qt(T ) uniquely determines it (Erdős et al. 1999).
The quartet compatibility problem is about finding out whether a set of quartet
topologies qt(S ) for a collection of phylogenetic trees S is compatible, i.e., if there
is a phylogeny T on the taxa in tx(S ) that displays all the quartet topologies in
qt(S ). The maximum quartet consistency problem for a supertree takes as input a
set of quartet topologies qt(S ) for a collection of phylogenetic trees S , and the goal
is to find a phylogeny T on the taxa tx(S ) that displays the maximum number of
quartet topologies in qt(S ) (Piaggio-Talice et al. 2004).
The topology of a tree T can be captured more directly using projections of T .
Given a set S ⊆ tx(T ), the projection of T with respect to S , denoted by TS , is
obtained from T by removing all structure related to the taxa in tx(T ) \ S . This
may imply that entire subtrees are removed and non-branching nodes are deleted.
We say that T displays another tree T ′ if tx(T ′ ) ⊆ tx(T ) and Ttx(T ′ ) = T ′ .
Example 2
If the left tree in Figure 1 is projected with respect to {Puma, Lynx, Felis}, the following tree results: ((Puma,Lynx),Felis). The right tree yields a different projection
((Puma,Felis),Lynx) illustrating the topological difference of the trees.
When comparing a phylogeny T with other phylogenies, an obvious question is
which projections should be used. Rather than using arbitrary sets S ⊆ tx(T ) for
projections TS , we suggest to use the subtrees of T . We denote this set by sub(T ).
It is clear that T displays T ′ for every T ′ ∈ sub(T ). Moreover, if T displays T ′′
for every T ′′ ∈ sub(T ′ ) and tx(T ) = tx(T ′ ), then T = T ′ . More generally, the
more subtrees of T ′ are displayed by T , the more alike T and T ′ are as trees. This
observation suggests defining the maximum projection consistency problem for a
supertree in analogy to the maximum quartet consistency problem. The input for
this problem consists of the multiset sub(T1 ) ∪ . . . ∪ sub(Tn ) induced by a given
collection T1 , . . . , Tn of phylogenetic trees. The goal is to find a supertree T such
1
We use multisets in order to give more weight to structures appearing in several source trees.
Optimizing Phylogenetic Supertrees Using Answer Set Programming
5
that tx(T ) = tx({T1 , . . . , Tn }) and T displays as many subtrees from the input as
possible—disregarding orientation. This objective is aligned with the quartet-based
approach: if T displays a particular subtree T ′ , then it also displays qt(T ′ ).
Example 3
Consider again the trees in Figure 1. The non-trivial subtrees of the left tree are:
(outgroup,(Felis,(Lynx,(Panthera,Puma)))), (Felis,(Lynx,(Panthera,Puma))),
(Lynx,(Panthera,Puma)), (Panthera,Puma)
The right tree displays only the subtree (Panthera,Puma) as its projection.
3 Encodings for supertree optimization
We assume that the reader is familiar with basic ASP terminology and definitions,
and we refer the reader to (Baral 2003; Gebser et al. 2012) for details. Our encodings are based on the input language of the gringo 3.0.4 grounder (Gebser et al. 2009)
used to instantiate logic programs. In this section, two alternative encodings for the
supertree construction problem are presented. Both encodings rely on the same formalization of the underlying tree structure, but have different objective functions
as well as different representations for the input data. We begin by developing a
canonical representation for phylogenies based on ordered trees in Section 3.1. The
first encoding based on quartet information is then presented in Section 3.2. The
second one exploiting projections of trees is developed in Section 3.3.
3.1 Canonical phylogenies
Our encodings formalize phylogenies as ordered trees whose leaf nodes correspond
to taxa (species or genera) of interest. The simplest possible (atomic) tree consists
of a single node. Thus we call the leaves of the tree atoms and formalize them in
terms of the predicate atom/1. We assume that the number of atoms is available
through the predicate atomcnt/1, and furthermore that atoms have been ordered
alphabetically so that the first atom is accessible through the predicate fstatom/1,
while the predicate nxtatom/2 provides the successor of an atom. These predicates
can be straightforwardly expressed in the input language of gringo and we skip
their actual definitions. Full encodings are published with tools (see Section 4).
To formalize the structure of an ordered tree with N leaves, we index the leaf
nodes using numbers from 1 to N . Any subsequent numbers up to 2N − 1 will
be assigned to inner nodes as formalized by lines 2–4 of Listing 1. Depending on
the topology of the tree, the number of inner nodes can vary from 1 to N − 1.
In the former case, the tree has an edge from the root to every leaf but a full
binary tree results in the latter case. If viewed as phylogenies, the former leaves
all relationships unresolved whereas the latter gives a fully resolved phylogeny. The
predicate pair/2 defined in line 5 declares that the potential edges of the tree
always proceed in the descending order of node numbers. This scheme makes loops
impossible and prohibits edges starting from leaf nodes. The rule in line 8 chooses
at most 2N − 2 edges for the tree up to 2N − 1 nodes. The constraint in line 9
6
L. Koponen et al.
Listing 1. An ASP Encoding of Directed Trees/Forests
1
2
3
4
5
6
7
8
9
10
11
12
13
14
% Domains
node (1..2* N -1) : - atomcnt ( N ) .
leaf ( X ) : - node ( X ) , X <= N , atomcnt ( N ) .
inner ( X ) : - node ( X ) , X >N , atomcnt ( N ) .
pair ( X , Y ) : - inner ( X ) , node ( Y ) , X > Y .
% Choose edges
{ edge (X , Y ) : pair (X , Y ) } 2* N -2 : - atomcnt ( N ) .
: - edge (X , Z ) , edge (Y , Z ) , pair ( X ;Y , Z ) , X < Y .
: - edge (X , Y ) , pair (X , Y ) , inner ( Y ) , not edge (Y , Z ) : pair (Y , Z ) .
% Assign atoms to leaves
asgn (1 , A ) : - node (1) , fstatom ( A ) .
asgn ( N +1 , B ) : - node ( N ) , asgn (N , A ) , nxtatom ( A , B ) .
ensures that a directed tree/forest rather than a directed acyclic graph is obtained.
The purpose of the constraint in line 10 is to deny branches ending at inner nodes.
The fixed assignment of atoms to leaf nodes 1 . . . N according to their alphabetical
order takes place in lines 13–14 using predicates fstatom/1 and nxtatom/2. This
is justified by a symmetry reduction, since N ! different assignments to leaf nodes
would be considered otherwise and no tree topology is essentially ruled out.
However, as regards tree topologies themselves, further symmetry reductions are
desirable because the number of optimal phylogenies can increase substantially otherwise. Listing 2 provides conditions for a canonical ordering for the inner nodes.
The order/2 predicate defined in lines 2–3 captures pairs of inner nodes that must
be topologically ordered in a tree being constructed. The ireach/2 predicate defined by rules in lines 4 and 5 gives the irreflexive reachability relation for nodes, i.e.,
a node is not considered reachable from itself. The constraint in line 6 effectively
states that the numbering of inner nodes must follow the depth-first descending
order, i.e., any inner nodes X below Y must have higher numbers than Z. The remaining degree of freedom concerns the placement of leaves to subtrees. To address
this, we need to find out the minimum2 leaf (node) for each subtree. The min/2
predicate defined in lines 9–10 captures the actual minimum leaf Y beneath an inner
node X. The orientation constraint in line 11 concerns inner nodes Y and Z subject
to topological ordering, identifies the minimum leaf W in the subtree rooted at Z,
and ensures that this leaf is smaller than any leaf V in the subtree rooted at Y. This
also covers the case that V is the respective minimum leaf under Y. The orientation
constraint above generalizes that of (Brooks et al. 2007) for non-binary trees and
we expect that canonical trees will have further applications beyond this work.
Finally, there are some further requirements specific to phylogenies. We assume
that certain subsidiary predicates have already been defined. The predicate root/1
is used to identify root nodes. Inner nodes that remain completely disconnected are
marked as unused by the predicate unused/1. Otherwise, the node is in use as cap-
2
Recall that the numbering of leaf nodes corresponds to the alphabetical ordering of the taxa.
Optimizing Phylogenetic Supertrees Using Answer Set Programming
7
Listing 2. Encoding for Canonical Phylogenies
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
% Depth - first ordering on internal nodes
order (Y , Z ) : - edge (X , Y ) , edge (X , Z ) , pair (X , Y ; Z ) , inner ( Y ; Z ) ,
Y >Z , not edge ( X , W ) : Y > W : W > Z : pair (X , W ) .
ireach (X , Y ) : - edge (X , Y ) , pair ( X , Y ) .
ireach (X , Y ) : - ireach (X , Z ) , edge (Z , Y ) , pair (Z , Y ) .
: - order (Y , Z ) , pair (Y , Z ) , ireach (Y , X ) , inner ( X ) , X < Y .
% Determine the orientatio n of leaf nodes
min (X , Y ) : - ireach (X , Y ) , inner ( X ) , leaf ( Y ) ,
not ireach (X , Z ) : Z < Y : leaf ( Z ) .
: - order (Y , Z ) , pair (Y , Z ) , ireach (Y , V ) , min ( Z , W ) , leaf ( V ; W ) , V < W .
% Constrain ts for phylogeni e s
: - unused ( X ) , used ( Y ) , inner ( X ; Y ) , X < Y .
: - root ( X ) , root ( Y ) , inner ( X ; Y ) , X < Y .
: - not root ( X ) : inner ( X ) .
: - leaf ( X ) , not edge (Y , X ) : pair (Y , X ) .
: - inner ( X ) , root ( X ) , not outgroup ( X ) .
: - inner ( X ) , not root ( X ) , outgroup ( X ) .
: - edge (X , Y ) , pair (X , Y ) , not edge ( X , Z ) : pair (X , Z ) : Z != Y .
tured by used/1. Moreover, a node is an outgroup node, formalized by outgroup/1,
if it is assigned to the special outgroup taxon or one of its child nodes is so assigned
(cf. Figure 1). Lines 14–20 list the additional constraints for a phylogeny. Only the
highest numbers are allowed for unused nodes (line 14). The root must be a unique
inner node (lines 15 and 16). Every leaf must be connected (line 17). The special
outgroup leaf must be associated with the root node (lines 18 and 19). Every inner
node that is actually used must have at least two children (line 20): the denial of
unary nodes is justified because they are not meaningful for phylogenies.
3.2 Quartet-based approach
The first encoding is quartet-based. Each source tree is represented as the set of
all quartets that it displays. The predicate quartet/4 represents one input quartet
in canonical form. Listing 3 shows the objective function for the quartet encoding.
For each quartet appearing in the input, we check if it is satisfied by the current
output tree candidate. The auxiliary predicate reach/2 marks reachability from
inner nodes to atoms (species) assigned to leaves. The output tree is rooted, so
given any inner node X in the tree, there is a uniquely defined subtree rooted at X,
and reach(X,A) is true for any atom A corresponding to a leaf node of the subtree.
A quartet consisting of two pairs is satisfied by the output tree, if for one pair there
exists at least one inner node X such that the members of the pair are descendants
of X, while the members of the other pair do not appear in that subtree.
The predicate quartetwt/5 assigns a weight to each quartet structure. In the
unweighted case, this weight is equal to the number of source trees that display the
quartet. In the weighted case, source trees stemming from computational studies
based on molecular input data were weighted up by a factor of four. For example,
8
L. Koponen et al.
Listing 3. Optimization function for the quartet encoding
1
2
3
4
5
6
7
8
9
reach (X , A ) : - inner ( X ) , ireach ( X , Y ) , asgn (Y , A ) , atom ( A ) .
% Maximize number of satisfied quartets
satisfied ( A1 , A2 , A3 , A4 ) : - quartet ( A1 , A2 , A3 , A4 ) , inner ( X ) ,
reach (X , A1 ) , reach (X , A2 ) , not reach (X , A3 ) , not reach (X , A4 ) .
satisfied ( A1 , A2 , A3 , A4 ) : - quartet ( A1 , A2 , A3 , A4 ) , inner ( X ) ,
reach (X , A3 ) , reach (X , A4 ) , not reach (X , A1 ) , not reach (X , A2 ) .
# maximize [ satisfied ( A1 , A2 , A3 , A4 ) = W : quartetwt ( A1 , A2 , A3 , A4 , W ) ].
if a particular quartet was present in three source trees, two of which were from
molecular studies while the third one was not, the total weight would be 4 + 4 + 1.
3.3 Projection-based approach
The second encoding is based on direct projections of trees and the idea is to
identify which inner nodes in the selected phylogeny correspond to subtrees present
in the input trees. Input trees are represented using a function symbol t as a tree
constructor. For instance, the leftmost tree in Figure 1 is represented by a term
t(outgroup,t(felis,t(lynx,t(panthera,puma)))) ·
(1)
For simplicity, it is assumed here that t always takes two arguments although in
practice, some of the input trees are non-binary, and a more general list representation is used instead. In the encoding, projections of interest are declared in terms
of the predicate proj/1. The predicate comp/1, defined in line 2 of Listing 4, identifies compound trees as those having at least one instance of the constructor t.
The set of projections is made downward closed by the rule in line 3. For instance,
outgroup and t(felis,t(lynx,t(panthera,puma))) are projections derived from
(1) by a single application of this rule. In line 4, atoms are recognized as trivial tree
projections with no occurrences of t such as outgroup above.
The reach/2 predicate, defined in lines 7 and 8 of Listing 4, generalizes the respective predicate from Listing 3 for arbitrary projections T and includes a new
base case for immediate assignments (line 7). A compound tree T is assigned to
an inner node X by default (line 11) and the predicate denied/2 is used to specify exceptions in this respect. It is important to note that if edge(X,Y) is true,
then X is an inner node and used(X) is true, too. The first exception (line 12) is
that T is already assigned below X in the phylogeny. The second case (lines 13–14)
avoids mapping distinct subtrees of t(T1,T2) on the same subtree in the phylogeny.
Thirdly, if t(T1,T2) is to be assigned at inner node X, then T1 and T2 must have
been assigned beneath X in the phylogeny (lines 15–18). Finally, the constraint in
line 20 insists that each inner node is assigned at least one projection because the
node could be removed from the phylogeny otherwise. The net effect of the constraints introduced so far is that if T1 and T2 have been assigned to nodes X and Y,
respectively, then t(T1,T2) is assigned to the least common ancestor of X and Y.
Optimizing Phylogenetic Supertrees Using Answer Set Programming
9
Listing 4. Projection-Based Optimization of the Phylogeny
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
% Projectio ns of the phylogeny
comp ( t ( T1 , T2 ) ) : - proj ( t ( T1 , T2 ) ) .
proj ( T1 ; T2 ) : - comp ( t ( T1 , T2 ) ) .
atom ( X ) : - proj ( X ) , not comp ( X ) .
% R e a c h a b i l i t y from a node to a projection
reach (X , T ) : - node ( X ) , asgn ( X , T ) , proj ( T ) .
reach (X , T ) : - ireach (X , Y ) , node ( X ; Y ) , reach (Y , T ) , proj ( T ) .
% Assign compound trees to inner nodes
asgn ( X , T ) : - inner ( X ) , used ( X ) , not denied ( X , T ) , comp ( T ) .
denied (X , T ) : - edge (X , Y ) , pair ( X , Y ) , comp ( T ) , reach (Y , T ) .
denied (X , t ( T1 , T2 ) ) : - edge (X , Y ) , pair (X , Y ) , comp ( t ( T1 , T2 ) ) ,
T1 < T2 , reach (Y , T1 ) , reach (Y , T2 ) .
denied (X , t ( T1 , T2 ) ) : - inner ( X ) , used ( X ) , comp ( t ( T1 , T2 ) ) ,
not reachvia (X , Z , T1 ) : pair (X , Z ) .
denied (X , t ( T1 , T2 ) ) : - inner ( X ) , used ( X ) , comp ( t ( T1 , T2 ) ) ,
not reachvia (X , Z , T2 ) : pair (X , Z ) .
reachvia (X ,Y , T ) : - edge (X , Y ) , pair (X , Y ) , reach (Y , T ) , proj ( T ) .
: - inner ( X ) , used ( X ) , not asgn ( X , T ) : comp ( T ) .
% Optimize the assignment of compound trees
unassigned ( T ) : - comp ( T ) , not asgn (X , T ) : node ( X ) .
next ( X , T ) : - edge (X , Y ) , pair (X , Y ) , asgn ( Y , T ) , proj ( T ) .
separated ( t ( T1 , T2 ) ) : - edge ( X , Y ) , pair (X , Y ) , asgn (X , t ( T1 , T2 ) ) ,
not next (X , T1 ) .
separated ( t ( T1 , T2 ) ) : - edge ( X , Y ) , pair (X , Y ) , asgn (X , t ( T1 , T2 ) ) ,
not next (X , T2 ) .
# minimize [ unassigned ( T ) = AC * W : acnt ( T , AC ) : projwt (T , W ) : comp ( T ) ,
separated ( T ) = W : projwt (T , W ) : comp ( T ) ].
The rest of Listing 4 concerns the objective function we propose for phylogeny
optimization. The predicate unassigned/1 captures compound trees T which could
not be assigned to any inner node by the rules above. This is highly likely if mutually
inconsistent projections are provided as input. It is also possible that a compound
projection t(T1,T2) is assigned further away from the subtrees T1 and T2, i.e., they
are not placed next to t(T1,T2). The predicate separated/1 holds for t(T1,T2) in
this case (lines 24–28). The purpose of the objective function (line 30) is to minimize
penalties resulting from these aspects of assignments. For unassigned compound
trees T, this is calculated as the product of the number of atoms in T and the weight3
of T. These numbers are accessible via auxiliary predicates acnt/2 and projwt/2 in
the encoding. Separated compound trees are further penalized by their weight (line
29). Since the rules in lines 2–3, 13–18, 25–28 only cover binary trees they would
have to be generalized for any fixed arity which is not feasible. To avoid repeating
the rules for different arities, we represent trees as lists (of lists) in practice.
3
As before, the weight is 4 for projections originating from molecular studies and 1 otherwise.
10
L. Koponen et al.
4 Experiments
Data. We use a collection of 38 phylogenetic trees from (Säilä et al. 2011; Säilä et al. 2012)
covering 105 species of Felidae as our source trees.4 There are both resolved and unresolved trees, all rooted with outgroup, in the collection and the number of species
varies from 4 to 52. The total number of species in the source trees makes supertree
analysis even with heuristic methods challenging, and computing the full supertree
for all species at once is not feasible with our encodings. Thus, we consider the following simplifications of the data. In Section 4.1 we use genus-specific projections
of source trees to compare the efficiency of our two encodings. In Section 4.2 we
reduce the size of the instance by considering the genus-level supertree as a first
step towards solving the supertree problem for the Felidae data.
Experimental setting. We used two identical 2.7-GHz CPUs with 256 GB of RAM
to compute optimal answer sets for programs grounded by gringo 3.0.4. The stateof-the-art solver5 clasp 3.1.2 (Gebser et al. 2011) was compared with a runner-up
solver wasp6 (Alviano et al. 2015) as of 2015-06-28. Moreover, we studied the performance of MAXSAT solvers as back-ends using translators lp2acyc 1.29 and
lp2sat 1.25 (Gebser et al. 2014), and a normalizer lp2normal 2.18 (Bomanson et al. 2014)
from the asptools 7 collection. As MAXSAT solvers, we tried clasp 3.1.2 in its
MAXSAT mode (clasp-s in Table 1), an openwbo-based extension8 (Martins et al. 2014)
of acycglucose R739 (labeled acyc in Table 1) also available in the asptools collection, and sat4j9 (Le Berre and Parrain 2010) dated 2013-05-25.
4.1 Genus-specific supertrees
To produce genus-specific source trees for a genus G, we project all source trees
to the species in G (and the outgroup). Genera with fewer than five species are
excluded as too trivial. Thus, the instances of Felidae data have between 6 and
11 species each, and the number of source trees varies between 2 and 22. In order
to be able to compare the performance of different solvers for our encodings, we
compute one optimum here and use a timeout of one hour. In Table 1 we report the
run times for the best-performing configuration of each solver for both encodings.10
Moreover, the methods based on unsatisfiable cores turned out to be ineffective in
general. Hence, branch-and-bound style heuristics were used.
The performance of the projection encoding scales up better than that of the
quartet encoding when the complexity of the instance grows. Our understanding
is that in the quartet encoding the search space is more symmetric than in the
projection encoding: in principle any subset of the quartets could do and this has to
4
5
6
7
8
9
10
Source trees in Newick format are provided in the online appendix (Appendix D).
http://potassco.sourceforge.net
http://github.com/alviano/wasp.git
Subdirectories download/ and encodings/ at http://research.ics.aalto.fi/software/asp/
http://sat.inesc-id.pt/open-wbo/
http://www.sat4j.org/
We exclude sat4j, which had the longest run times, from comparison due to space limitations.
Optimizing Phylogenetic Supertrees Using Answer Set Programming
Genus
Hyperailurictis
Lynx
Leopardus
Dinofelis
Homotherium
Felis
Panthera
a
b
c
d
Taxa Trees
6
7
8
9
9
11
11
2
8
6
2
3
12
22
claspa
qtet proj
0.0
0.0
0.6
0.1
0.7
39.6
1395.8
0.0
0.0
0.1
0.0
0.0
21.9
45.6
11
waspb
qtet proj
acycc
qtet proj
clasp-sd
qtet proj
0.0
0.0
0.0
0.1
1.7
0.2
0.0
0.1
0.1
0.1
290.8 120.6
– 456.3
0.0
0.0
0.0
0.0
1.1
0.4
0.1
0.1
0.1
0.0
122.7 59.6
– 174.6
0.0
0.0
0.6
0.0
0.0
27.7
944.2
0.0
0.0
0.1
0.1
0.0
20.8
67.1
Options: --config=frumpy (proj) and --config=trendy (qtet)
Options: --weakconstraints-algorithm=basic
Options: -algorithm=1 and -incremental=3
Options --config=frumpy (proj) and --config=tweety (qtet)
Table 1. Time (s) to find one optimum for genus-specific data using different solvers
using quartet (qtet) and projection (proj) encoding (– marks timeout).
be excluded in the optimality proof. On the other hand, the mutual incompatibilities
of projections can help the solver to cut down the search space more effectively.
4.2 Genus-level abstraction
We generate 28 trees abstracted to the genus level from the 38 species-level trees.
The abstraction is done by placing each genus G under the node N furthest
away from the root such that all occurrences of the species of genus G are in
the subtree below N . Finally, redundant (unary) inner nodes are removed from
the trees. The trees that included fewer than four genera were excluded. Following (Säilä et al. 2011; Säilä et al. 2012), Puma pardoides was treated as its own
genus Pardoides, and Dinobastis was excluded as an invalid taxon. As further preprocessing, we removed the occurrences of genera Pristifelis, Miomachairodus, and
Pratifelis appearing in only one source tree each. These so-called rogue taxa have
unstable placements in the supertree, due to little information about their placements in relation to the rest of the taxa. The rogue taxa can be a posteriori placed
in the supertree in the position implied by their single source tree. After all the
preprocessing steps, our genus-level source trees have 34 genera in total and the
size of the trees varies from 4 to 22 genera.
We consider the following schemes from (Säilä et al. 2011; Säilä et al. 2012):
All-FM-bb-wgt Analysis with a constraint tree separating the representatives of
Felinae and Machairodontinae into subfamilies, with weight 4 given to source
trees from molecular studies.
F-Mol Analysis using molecular studies only and extinct species pruned out (leaving 20 source trees and 15 genera, which are all representatives of Felinae).
Noticeably, the first setting allows us to split the search space and to compute
the supertree for Felinae and Machairodontinae separately. The best resolved tree
in (Säilä et al. 2011; Säilä et al. 2012) was obtained using the MRP supertree for
12
L. Koponen et al.
outgroup
Proailurus
Pseudaelurus
Hyperailurictis
Stenailurus
Metailurus
Dinofelis
Adelphailurus
Promegantereon
Paramachaerodus
Smilodon
Megantereon
Nimravides
Machairodus
Amphimachairodus
Xenosmilus
Homotherium
Styriofelis
Neofelis
Panthera
Pardoides
Catopuma
Pardofelis
Leptailurus
Profelis
Caracal
Leopardus
Lynx
Felis
Otocolobus
Prionailurus
Miracinonyx
Puma
Acinonyx
outgroup
Proailurus
Pseudaelurus
Hyperailurictis
Stenailurus
Metailurus
Dinofelis
Adelphailurus
Promegantereon
Paramachaerodus
Smilodon
Megantereon
Nimravides
Machairodus
Amphimachairodus
Xenosmilus
Homotherium
Styriofelis
Neofelis
Panthera
Pardoides
Catopuma
Pardofelis
Leptailurus
Profelis
Caracal
Leopardus
Lynx
Felis
Otocolobus
Prionailurus
Miracinonyx
Puma
Acinonyx
Fig. 2. Left: Best-resolution 50% majority consensus MRP genus-level supertree
modified from (Säilä et al. 2011; Säilä et al. 2012) using scheme All-F-Mol-bbwgt; Right: The optimal genus-level supertree using projection encoding and
scheme All-FM-bb-wgt.
F-Mol abstracted to the genus level as a constraint tree (scheme All-F-Mol-bbwgt). We include the best resolved tree by Säilä et al. to the comparison as well.
We use clasp for the computation of all optimal models. The considered schemes
turned out to be unfeasible for the quartet-based encoding (no optimum was reached
by a timeout of 48 hours), and only results from the projection encoding are included. It turns out that there exists a unique optimum for the projection encoding
for both schemes. In the All-FM-bb-wgt scheme, the global optimum was identified in 4 hours and 56 minutes, while it was located in 52 minutes for F-Mol
using --config=trendy which performed best on these instances. The respective
run times are 1.5 hours and 20 minutes using parallel clasp 3.1.2 with 16 threads.
The MRP supertrees in (Säilä et al. 2011; Säilä et al. 2012) are computed using
the full species-level data with the Parsimony Ratchet method (Nixon 1999). For
the resulting shortest trees, 50% majority consensus trees were computed and the
best supported supertree according to (Wilkinson et al. 2005) out of different runs
(with various MRP settings) originates from scheme All-FM-bb-wgt, while the
best resolved tree was obtained using scheme All-F-Mol-bb-wgt. Finally, the
species-level supertree is collapsed to the genus level. The optimal supertree for
the projection encoding and the MRP supertrees from Säilä et al. described above
(projected to the set of genera considered in our experiments) are presented in
Figure 2 and the online appendix (Appendices A–C).
As the true supertree is not known for this real-life dataset, the goodness of the
output tree can only be measured based on how it reflects the source trees. To assess
the quality of the output trees and to compare them with the MRP trees, we considered the number of satisfied quartets of source trees, the resolution of the supertree,
Optimizing Phylogenetic Supertrees Using Answer Set Programming
Method
Resolution
QSa
%QSb
Vc
All-FM-bb-wgt
All-FM-bb-wgt
All-F-Mol-bb-wgt
proj
MRP
MRP
0.90
0.85
0.93
14 076
12 979
13 910
0.84
0.77
0.83
0.43
0.45
0.42
F-Mol
F-Mol
proj
MRP
1.00
1.00
4 395
4 389
0.86
0.86
0.25
0.27
Scheme
a
b
c
13
Number of satisfied quartets from source trees
Percentage of satisfied quartets from source trees
Support according to (Wilkinson et al. 2005)
Table 2. Comparison between the optimal supertree for the projection encoding
(proj) and the best MRP supertrees.
and support values (Wilkinson et al. 2005). Support varies between 1 and −1, indicating good and poor support, respectively, of the relationships in source trees. The
results are given in Table 2, showing that the optimum of the projection encoding
satisfies more quartets of the input data than the MRP supertrees.
Finally, the differences of the objective functions of our two encodings can be illustrated by computing the supertree of 5 highly conflicting source trees of 8 species
of hammerhead sharks from (Cavalcanti 2007). The optimum for the projection encoding is exactly the same as source tree (b) in (Cavalcanti 2007), whereas the
optimum for quartet encoding is exactly the same as source tree (a). Thus, the two
objective functions are not equivalent in the case of conflicting source trees.
5 Conclusion
In this paper we propose two ASP encodings for phylogenetic supertree optimization. The first, solving the maximum quartet consistency problem, is similar to
the encoding in (Wu et al. 2007) and does not perform too well in terms of run
time when the size of the input (source trees and number of taxa therein) grows.
The other novel encoding is based on projections of trees and the respective optimization problem is formalized as the maximum projection consistency problem. We use real data, namely a collection of phylogenetic trees for the family of
cats (Felidae) and first evaluate the performance of our encodings by computing
genus-specific supertrees. We then compute a genus-level supertree for the data and
compare our supertree against a recent supertree computed using MRP approach
(Säilä et al. 2011; Säilä et al. 2012). The projection-based encoding performs better than the quartet-based one and produces a unique optimum for the two cases
we consider (with rogue taxa removed). Obviously, this is not the case in general
and in the case of several optima, consensus and majority consensus supertrees can
be computed. Furthermore, our approach produces supertrees comparable to ones
obtained using MRP method. For the current projection-based encoding, the problem of optimizing a species-level supertree using the Felidae data is not feasible as
a single batch. Further investigations how to tackle the larger species-level data are
14
L. Koponen et al.
needed. Possible directions are for instance using an incremental approach and/or
parallel search.
6 Acknowledgments
This work has been funded by the Academy of Finland, grants 251170 (Finnish
Centre of Excellence in Computational Inference Research COIN), 132995 (LS),
275551 (LS), and 250518 (EO). We thank Martin Gebser, Ian Corfe, and anonymous
reviewers for discussion and comments that helped to improve the paper.
References
Aho, A. V., Sagiv, Y., Szymanski, T. G., and Ullman, J. D. 1981. Inferring a tree
from lowest common ancestors with an application to the optimization of relational
expressions. SIAM Journal on Computing 10, 3, 405–421.
Alviano, M., Dodaro, C., Leone, N., and Ricca, F. 2015. Advances in WASP. In Proceedings of the 13th International Conference on Logic Programming and Nonmonotonic
Reasoning, LPNMR 2015. Lecture Notes in Computer Science, vol. 9345. Springer.
Baral, C. 2003. Knowledge Representation, Reasoning, and Declarative Problem Solving.
Cambridge University Press, New York, NY, USA.
Baum, B. R. 1992. Combining trees as a way of combining data sets for phylogenetic
inference, and the desirability of combining gene trees. Taxon 41, 1, 3–10.
Bininda-Emonds, O. R. 2004. Phylogenetic Supertrees: Combining Information to Reveal
the Tree of Life. Computational Biology. Springer.
Bomanson, J., Gebser, M., and Janhunen, T. 2014. Improving the normalization of
weight rules in answer set programs. In Proceedings of the 14th European Conference
on Logics in Artificial Intelligence, JELIA 2014. Lecture Notes in Computer Science,
vol. 8761. Springer, 166–180.
Brooks, D. R., Erdem, E., Erdoğan, S. T., Minett, J. W., and Ringe, D. 2007.
Inferring phylogenetic trees using answer set programming. Journal of Automated Reasoning 39, 4, 471–511.
Bryant, D. 1997. Building trees, hunting for trees, and comparing trees. Ph.D. thesis,
University of Canterbury.
Byrka, J., Guillemot, S., and Jansson, J. 2010. New results on optimizing rooted
triplets consistency. Discrete Applied Mathematics 158, 11, 1136–1147.
Cavalcanti, M. J. 2007. A phylogenetic supertree of the hammerhead sharks (Carcharhiniformes, Sphyrnidae). Zoological Studies 46, 1, 6–11.
Chen, D., Diao, L., Eulenstein, O., Fernández-Baca, D., and Sanderson, M. 2003.
Flipping: a supertree construction method. DIMACS series in discrete mathematics and
theoretical computer science 61, 135–162.
Chimani, M., Rahmann, S., and Böcker, S. 2010. Exact ILP solutions for phylogenetic
minimum flip problems. In Proceedings of the First ACM International Conference on
Bioinformatics and Computational Biology, BCB 2010. ACM, 147–153.
Day, W. H., Johnson, D. S., and Sankoff, D. 1986. The computational complexity of
inferring rooted phylogenies by parsimony. Mathematical biosciences 81, 1, 33–42.
Erdős, P. L., Steel, M. A., Székely, L. A., and Warnow, T. 1999. A few logs suffice
to build (almost) all trees (i). Random Structures and Algorithms 14, 2, 153–184.
Optimizing Phylogenetic Supertrees Using Answer Set Programming
15
Flynn, J. J., Finarelli, J. A., Zehr, S., Hsu, J., and Nedbal, M. A. 2005. Molecular
phylogeny of the Carnivora (Mammalia): assessing the impact of increased sampling on
resolving enigmatic relationships. Systematic Biology 54, 2, 317–337.
Foulds, L. R. and Graham, R. L. 1982. The Steiner problem in phylogeny is NPcomplete. Advances in Applied Mathematics 3, 1, 43–49.
Fulton, T. L. and Strobeck, C. 2006. Molecular phylogeny of the Arctoidea (Carnivora): effect of missing data on supertree and supermatrix analyses of multiple gene
data sets. Molecular phylogenetics and evolution 41, 1, 165–181.
Gebser, M., Janhunen, T., and Rintanen, J. 2014. Answer set programming as SAT
modulo acyclicity. In Proceedings of the 21st European Conference on Artificial Intelligence, ECAI 2014. IOS Press, 351–356.
Gebser, M., Kaminski, R., Kaufmann, B., and Schaub, T. 2012. Answer Set Solving
in Practice. Synthesis Lectures on Artificial Intelligence and Machine Learning. Morgan
& Claypool Publishers.
Gebser, M., Kaminski, R., Ostrowski, M., Schaub, T., and Thiele, S. 2009. On the
input language of ASP grounder Gringo. In Proceedings of the 10th International Conference on Logic Programming and Nonmonotonic Reasoning, LPNMR 2009. Lecture
Notes in Computer Science, vol. 5753. Springer, 502–508.
Gebser, M., Kaufmann, B., Kaminski, R., Ostrowski, M., Schaub, T., and Schneider, M. T. 2011. Potassco: The Potsdam answer set solving collection. AI Commun. 24, 2, 107–124.
Gent, I. P., Prosser, P., Smith, B. M., and Wei, W. 2003. Supertree construction
with constraint programming. In Proceedings of the 9th International Conference on
Principles and Practice of Constraint Programming, CP 2003. Lecture Notes in Computer Science, vol. 2833. Springer, 837–841.
Goloboff, P. A. and Pol, D. 2002. Semi-strict supertrees. Cladistics 18, 5, 514–525.
Kavanagh, J., Mitchell, D. G., Ternovska, E., Manuch, J., Zhao, X., and Gupta,
A. 2006. Constructing Camin-Sokal phylogenies via answer set programming. In Proceedings of the 13th International Conference on Logic for Programming, Artificial
Intelligence, and Reasoning, LPAR 2006. Lecture Notes in Computer Science, vol. 4246.
Springer, 452–466.
Le, T., Nguyen, H., Pontelli, E., and Son, T. C. 2012. ASP at work: An ASP
implementation of PhyloWS. In Technical Communications of the 28th International
Conference on Logic Programming, ICLP 2012. LIPIcs, vol. 17. 359–369.
Le Berre, D. and Parrain, A. 2010. The Sat4j library, release 2.2. Journal on Satisfiability, Boolean Modeling and Computation 7, 59–64.
Martins, R., Manquinho, V., and Lynce, I. 2014. Open-WBO: a modular MaxSAT
solver. In Theory and Applications of Satisfiability Testing, SAT 2014. Lecture Notes
in Computer Science, vol. 8561. Springer, 438–445.
Morgado, A. and Marques-Silva, J. 2010. Combinatorial optimization solutions for
the maximum quartet consistency problem. Fundam. Inform. 102, 3-4, 363–389.
Nixon, K. C. 1999. The parsimony ratchet, a new method for rapid parsimony analysis.
Cladistics 15, 4, 407–414.
Piaggio-Talice, R., Burleigh, J. G., and Eulenstein, O. 2004. Quartet supertrees.
In Phylogenetic Supertrees. Springer, 173–191.
Purvis, A. 1995. A modification to Baum and Ragan’s method for combining phylogenetic
trees. Systematic Biology 44, 2, 251–255.
Ragan, M. A. 1992. Phylogenetic inference based on matrix representation of trees.
Molecular phylogenetics and evolution 1, 1, 53–58.
16
L. Koponen et al.
Säilä, L. K., Fortelius, M., Oikarinen, E., Werdelin, L., and Corfe, I. 2012.
Fossil mammals, phylogenies and climate: the effects of phylogenetic relatedness on
range sizes and replacement patterns in changing environments. In Proceedings of 60th
Annual Symposium of Vertebrate Palaeontology and Comparative anatomy, SVPCA
2012. Poster.
Säilä, L. K., Fortelius, M., Oikarinen, E., Werdelin, L., Corfe, I., and Tuomola,
A. 2011. Taxon replacement: Invasion or speciation? First results for a supertree of
Neogene mammals. Journal of Vertebrate Paleontology 31, 3, suppl., 184A.
Semple, C. and Steel, M. 2000. A supertree method for rooted trees. Discrete Applied
Mathematics 105, 1, 147–158.
Snir, S. and Rao, S. 2012. Quartet MaxCut: a fast algorithm for amalgamating quartet
trees. Molecular phylogenetics and evolution 62, 1, 1–8.
Sridhar, S., Lam, F., Blelloch, G. E., Ravi, R., and Schwartz, R. 2008. Mixed
integer linear programming for maximum-parsimony phylogeny inference. IEEE/ACM
Transactions on Computational Biology and Bioinformatics 5, 3, 323–331.
Steel, M., Dress, A. W., and Bocker, S. 2000. Simple but fundamental limitations
on supertree and consensus tree methods. Systematic Biology 49, 2, 363–368.
Swenson, M. S., Suri, R., Linder, C. R., and Warnow, T. 2011. An experimental
study of Quartets MaxCut and other supertree methods. Algorithms for Molecular
Biology 6, 1, 7.
Wilkinson, M., Cotton, J. A., Creevey, C., Eulenstein, O., Harris, S. R., Lapointe, F.-J., Levasseur, C., Mcinerney, J. O., Pisani, D., and Thorley, J. L.
2005. The shape of supertrees to come: tree shape related properties of fourteen supertree methods. Systematic biology 54, 3, 419–431.
Wilkinson, M., Pisani, D., Cotton, J. A., and Corfe, I. 2005. Measuring support
and finding unsupported relationships in supertrees. Systematic Biology 54, 5, 823–831.
Wu, G., You, J.-H., and Lin, G. 2007. Quartet-based phylogeny reconstruction with
answer set programming. IEEE/ACM Transactions on Computational Biology and
Bioinformatics 4, 1, 139–152.
| 5cs.CE
|
Planning with Information-Processing Constraints and
Model Uncertainty in Markov Decision Processes
Jordi Grau-Moya1,2,3 , Felix Leibfried1,2,3 , Tim Genewein1,2,3 , and Daniel A. Braun1,2
1
arXiv:1604.02080v1 [cs.AI] 7 Apr 2016
2
Max Planck Institute for Intelligent Systems, Tübingen, Germany,
[email protected],
Max Planck Institute for Biological Cybernetics, Tübingen, Germany,
3
Graduate Training Centre for Neuroscience, Tübingen, Germany
Abstract. Information-theoretic principles for learning and acting have been proposed to solve particular classes of Markov Decision Problems. Mathematically,
such approaches are governed by a variational free energy principle and allow
solving MDP planning problems with information-processing constraints expressed
in terms of a Kullback-Leibler divergence with respect to a reference distribution. Here we consider a generalization of such MDP planners by taking model
uncertainty into account. As model uncertainty can also be formalized as an
information-processing constraint, we can derive a unified solution from a single generalized variational principle. We provide a generalized value iteration
scheme together with a convergence proof. As limit cases, this generalized scheme
includes standard value iteration with a known model, Bayesian MDP planning,
and robust planning. We demonstrate the benefits of this approach in a grid world
simulation.
Keywords: bounded rationality, model uncertainty, robustness, planning, Markov
Decision Processes
1
Introduction
The problem of planning in Markov Decision Processes was famously addressed by
Bellman who developed the eponymous principle in 1957 [1]. Since then numerous
variants of this principle have flourished in the literature. Here we are particularly interested in a generalization of the Bellman principle that takes information-theoretic constraints into account. In the recent past there has been a special interest in the KullbackLeibler divergence as a constraint to limit deviations of the action policy from a prior.
This can be interesting in a number of ways. Todorov [2,3], for example, has transformed the general MDP problem into a restricted problem class without explicit action
variables, where control directly changes the dynamics of the environment and control
costs are measured by the Kullback-Leibler divergence between controlled and uncontrolled dynamics. This simplification allows mapping the Bellman recursion to a linear
algebra problem. This approach can also be be generalized to continuous state spaces
leading to path integral control [4,5]. The same equations can also be interpreted in
terms of bounded rational decision-making where the decision-maker has limited computational resources that allow only limited deviations from a prior decision strategy
2
Grau-Moya et al.
(measured by the Kullback-Leiber divergence in bits) [6]. Such a decision-maker can
also be instantiated by a sampling process that has restrictions in the number of samples it can afford [7]. Disregarding the possibility of a sampling-based interpretation,
the Kullback-Leibler divergence introduces a control information cost that is interesting
in its own right when formalizing the perception action cycle [8].
While the above frameworks have led to interesting computational advances, so
far they have neglected the possibility of model misspecification in the MDP setting.
Model misspecification or model uncertainty does not refer to the uncertainty arising
due to the stochastic nature of the environment (usually called risk-uncertainty in the
economic literature), but refers to the uncertainty with respect to the latent variables that
specify the MDP. In Bayes-Adaptive MDPs [9], for example, the uncertainty over the
latent parameters of the MDP is explicitly represented, such that new information can
be incorporated with Bayesian inference. However, Bayes-Adaptive MDPs are not robust with respect to model misspecification and have no performance guarantees when
planning with wrong models [10]. Accordingly, there has been substantial interest in
developing robust MDP planners [11,12,13]. One way to take model uncertainty into
account is to bias an agent’s belief model from a reference Bayesian model towards
worst-case scenarios; thus avoiding disastrous outcomes by not visiting states where
the transition probabilities are not known. Conversely, the belief model can also be biased towards best-case scenarios as a measure to drive exploration—also referred in the
literature as optimism in face of uncertainty [14,15].
When comparing the literature on information-theoretic control and model uncertainty, it is interesting to see that some notions of model uncertainty follow exactly the
same mathematical principles as the principles of relative entropy control [3]. In this
paper we therefore formulate a unified and combined optimization problem for MDP
planning that takes both, model uncertainty and bounded rationality into account. This
new optimization problem can be solved by a generalized value iteration algorithm. We
provide a theoretical analysis of its convergence properties and simulations in a grid
world.
2
Background and Notation
In the MDP setting the agent at time t interacts with the environment by taking action
at ∈ A while in state st ∈ S. Then the environment updates the state of the agent to
st+1 ∈ S according to the transition probabilities T (st+1 |at , st ). After each transition
s
the agent receives a reward Rstt+1
,at ∈ R that is bounded. For our purposes we will
consider A and S to be finite. The aim of the agent is to choose its policy π(a|s) in
order to maximize the total discounted expected reward or value function for any s ∈ S
"T −1
#
X
∗
t st+1
V (s) = max lim E
γ Rst ,at
π
T →∞
t=0
with discount factor 0 ≤ γ < 1. The expectation is over all possible trajectories
ξ =
QT −1
s0 , a0 , s1 . . . of state and action pairs distributed according to p(ξ) = t=0 π(at |st )
T (st+1 |at , st ). It can be shown that the optimal value function satisfies the following
Information-Processing Constraints and Model Uncertainty in MDPs
recursion
V ∗ (s) = max
π
X
h 0
i
s
π(a|s)T (s0 |a, s) Rs,a
+ γV ∗ (s0 ) .
3
(1)
a,s0
At this point there are two important implicit assumptions. The first is that the policy π can be chosen arbitrarily without any constraints which, for example, might not
be true for a bounded rational agent with limited information-processing capabilities.
The second is that the agent needs to know the transition-model T (s0 |a, s), but this
model is in practice unknown or even misspecified with respect to the environment’s
true transition-probabilities, specially at initial stages of learning. In the following, we
explain how to incorporate both bounded rationality and model uncertainty into agents.
2.1
Information-Theoretic Constraints for Acting
Consider a one-step decision-making problem where the agent is in state s and has
s0
to choose a single action a from the set A to maximize the reward Rs,a
, where s0
is the next
A perfectly rational agent selects the optimal action a∗ (s) =
Pthe state.
0
s0
argmaxa s0 T (s |a, s)Rs,a
. However, a bounded rational agent has only limited reP
s0
sources to find the maximum of the function s0 T (s0 |a, s)Rs,a
. One way to model
such an agent is to assume that the agent has a prior choice strategy ρ(a|s) in state
s before a deliberation process sets in that refines the choice strategy to a posterior
distribution π(a|s) that reflects the strategy after deliberation. Intuitively, because the
deliberation resources are limited, the agent can only afford to deviate from the prior
strategy by a certain amount of information bits. This can be quantified by the relative
P
entropy DKL (π||ρ) = a π(a|s) log π(a|s)
ρ(a|s) that measures the average information cost
of the policy π(a|s) using the source distribution ρ(a|s). For a bounded rational agent
this relative entropy is bounded by some upper limit K. Thus, a bounded rational agent
has to solve a constrained optimization problem that can be written as
max
π
X
π(a|s)
X
0
s
T (s0 |a, s)Rs,a
s.t. DKL (π||ρ) ≤ K
s0
a
This problem can be rewritten as an unconstrained optimization problem
F ∗ (s) = max
π
=
X
π(a|s)
a
X
0
s
T (s0 |a, s)Rs,a
−
s0
1
DKL (π||ρ)
α
X
P
0
s0
1
log
ρ(a|s)eα s0 T (s |a,s)Rs,a .
α
a
(2)
(3)
where F ∗ is a free energy that quantifies the value of the policy π by trading off the
average reward against the information cost. The optimal strategy can be expressed
analytically in closed-form as
∗
π (a|s) =
ρ(a|s)eα
P
0
s0
s
T (s0 |a,s)Rs,a
Zα (s)
4
Grau-Moya et al.
P
P
0
s0
with partition sum Zα (s) =
a ρ(a|s) exp α
s0 T (s |a, s)Rs,a . Therefore, the
maximum operator in (2) can be eliminated and the free energy can be rewritten as in
(3). The Lagrange multiplier α quantifies the boundedness of the agent. By setting α →
∞ we recover a perfectly rational agent with optimal policy π ∗ (a|s) = δ(a − a∗ (s)).
For α = 0 the agent has no computational resources and the agent’s optimal policy is to
act according to the prior π ∗ (a|s) = ρ(a|s). Intermediate values of α lead to a spectrum
of bounded rational agents.
2.2
Information-Theoretic Constraints for Model Uncertainty
In the following we assume that the agent has a model of the environment Tθ (s0 |a, s)
that depends on some latent variables θ ∈ Θ. In the MDP setting, the agent holds a
belief µ(θ|a, s) regarding the environmental dynamics where θ is a unit vector of transition probabilities into all possible states s0 . While interacting with the environment the
agent can incorporate new data by forming the Bayesian posterior µ(θ|a, s, D), where
D is the observed data. When the agent has observed an infinite amount of data (and
assuming θ∗ (a, s) ∈ Θ) the belief will converge to the delta distribution µ(θ|s, a, D) =
δ(θ −θ∗ (a, s)) and the agent will act optimally according to the true transition probabilities, exactly as in ordinary optimal choice strategies with known models. When acting
under a limited amount of data the agent
the value of an action a with
Pcannot determine
s0
the true transition model according to s0 T (s0 |a, s)Rs,a
, but it can only determine an
R
P
s0
.
expected value according to its beliefs θ µ(θ|a, s) s0 Tθ (s0 |a, s)Rs,a
The Bayesian model µ can be subject to model misspecification (e.g. by having a
wrong likelihood or a bad prior) and thus the agent might want to allow deviations from
its model towards best-case (optimistic agent) or worst-case (pessimistic agent) scenarios up to a certain extent, in order to act more robustly or to enhance its performance in
a friendly environment [16]. Such deviations can be measured by the relative entropy
DKL (ψ|µ) between the Bayesian posterior µ and a new biased model ψ. Effectively,
this allows for mathematically formalizing model uncertainty, by not only considering
the specified model but all models within a neighborhood of the specified model that
deviate no more than a restricted number of bits. Then, the effective expected value of
an action a while having limited trust in the Bayesian posterior µ can be determined for
the case of optimistic deviations as
Z
X
1
∗
s0
(4)
F (a, s) = max ψ(θ|a, s)
Tθ (s0 |a, s)Rs,a
− DKL (ψ||µ)
ψ
β
θ
0
s
for β > 0, and for the case of pessimistic deviations as
Z
X
1
s0
F ∗ (a, s) = min ψ(θ|a, s)
Tθ (s0 |a, s)Rs,a
− DKL (ψ||µ)
ψ
β
θ
0
s
for β < 0. Conveniently, both equations can be expressed as a single equation
F ∗ (a, s) =
1
log Zβ (a, s)
β
(5)
Information-Processing Constraints and Model Uncertainty in MDPs
R
with β ∈ R and Zβ (s, a) =
the optimal biased belief
θ
5
P
s0
µ(θ|a, s) exp β s0 Tθ (s0 |a, s)Rs,a
when inserting
X
1
s0
µ(θ|a, s) exp β
Tθ (s0 |a, s)Rs,a
ψ (θ|a, s) =
Zβ (a, s)
0
!
∗
s
into either equation (4) or (5). By adopting this formulation we can model any degree
of trust in the belief µ allowing deviation towards worst-case or best-case with −∞ ≤
β ≤ ∞. For the case of β → −∞ we recover an infinitely pessimistic agent that
considers only worst-case scenarios, for β → ∞ an agent that is infinitely optimistic
and for β → 0 the Bayesian agent that fully trusts its model.
3
Model Uncertainty and Bounded Rationality in MDPs
In this section, we consider a bounded rational agent with model uncertainty in the infinite horizon setting of an MDP. In this case the agent must take into account all future
rewards and information costs, thereby optimizing the following free energy objective
∗
F (s) = max ext lim E
π
ψ T →∞
T
−1
X
t=0
γ
t
Rsstt+1
,at −
ψ(θt |at , st ) 1
π(at |st )
1
log
− log
β
µ(θt |at , st ) α
ρ(at |st )
!
(6)
where the extremum operator ext can be either max for β > 0 or min for β < 0,
0 < γ < 1 is the discount factor and the expectation E is Q
over all trajectories ξ =
T −1
s0 , a0 , θ0 , s1 , a1 , . . . aT −1 , θT −1 , sT with distribution p(ξ) = t=0 π(at |st ) ψ(θt |at , st )
Tθt (st+1 |at , st ). Importantly, this free energy objective satisfies a recursive relation and
thereby generalizes Bellman’s optimality principle to the case of model uncertainty and
bounded rationality. In particular, equation (6) fulfills the recursion
"
∗
1
π(a|s)
log
+
α
ρ(a|s)
1
ψ(θ|a, s)
Eψ(θ|a,s) − log
+
β
µ(θ|a, s)
F (s) = max ext Eπ(a|s) −
π
ψ
#
h 0
i
s
∗ 0
ETθ (s0 |a,s) Rs,a + γF (s )
. (7)
Applying variational calculus and following the same rationale as in the previous
sections [6], the extremum operators can be eliminated and equation (7) can be reexpressed as
h
h 0
ii αβ
1
s
+ γF ∗ (s0 )
F ∗ (s) = log Eρ(a|s) Eµ(θ|a,s) exp βETθ (s0 |a,s) Rs,a
(8)
α
6
Grau-Moya et al.
because
1
1
π(a|s)
F (s) = max Eπ(a|s)
log Zβ (a, s) − log
π
β
α
ρ(a|s)
1
α
= log Eρ(a|s) exp
log Zβ (a, s) ,
α
β
∗
(9)
(10)
where
h 0
i 1
ψ(θ|a, s)
s
∗ 0
Zβ (a, s) = ext Eψ(θ|a,s) ETθ (s0 |a,s) Rs,a + γF (s ) − log
ψ
β
µ(θ|a, s)
h 0
i
s
= Eµ(θ|a,s) exp βETθ (s0 |a,s) Rs,a
+ γF ∗ (s0 )
(11)
with the optimizing arguments
ψ ∗ (θ|a, s) =
h 0
i
1
s
µ(θ|a, s) exp βETθ (s0 |a,s) Rs,a
+ γF (s0 )
Zβ (a, s)
1
ρ(a|s) exp
π (a|s) =
Zα (s)
∗
α
log Zβ (a, s)
β
(12)
and partition sum
α
Zα (s) = Eρ(a|s) exp
log Zβ (a, s) .
β
With this free energy we can model a range of different agents for different α and
β. For example, by setting α → ∞ and β → 0 we can recover a Bayesian MDP planner
and by setting α → ∞ and β → −∞ we recover a robust planner. Additionally, for
α → ∞ and when µ(θ|a, s) = δ(θ − θ∗ (a, s)) we recover an agent with standard value
function with known state transition model from equation (1).
3.1
Free Energy Iteration Algorithm
Solving the self-consistency equation (8) can be achieved by a generalized version of
value iteration. Accordingly, the optimal solution can be obtained by initializing the
free energy at some arbitrary value F and applying a value iteration scheme B i+1 F =
BB i F where we define the operator
"
1
π(a|s)
log
+
α
ρ(a|s)
1
ψ(θ|a, s)
Eψ(θ|a,s) − log
+
β
µ(θ|a, s)
BF (s) = max ext Eπ(a|s) −
π
ψ
#
h 0
i
s
0
ETθ (s0 |a,s) Rs,a + γF (s )
(13)
Information-Processing Constraints and Model Uncertainty in MDPs
7
with B 1 F = BF , which can be simplified to
h
h 0
ii αβ
1
s
0
0
BF (s) = log Eρ(a|s) Eµ(θ|a,s) exp βETθ (s |a,s) Rs,a + γF (s )
α
In Algorithm (1) we show the pseudo-code of this generalized value iteration scheme.
Given state-dependent prior policies ρ(a|s) and the Bayesian posterior beliefs µ(θ|a, s)
and the values of α and β, the algorithm outputs the equilibrium distributions for the
action probabilities π(a|s), the biased beliefs ψ(θ|a, s) and estimates of the free energy
value function F ∗ (s). The iteration is run until a convergence criterion is met. The
convergence proof is shown in the next section.
Algorithm 1: Iterative algorithm solving the self-consistency equation (8)
Input: ρ(a|s), µ(θ|a, s), α, β
Initialize: F ← 0, Fold ← 0
while not converged do
forall the s ∈ S do
ii α
h
h 0
β
s
+ γFold (s0 )
F (s) ← α1 log Eρ(a|s) Eµ(θ|a,s) exp βETθ (s0 |a,s) Rs,a
end
Fold ← F
end
log Zβ (a, s)
i
h 0
s
1
+ γF (s0 )
µ(θ|a, s) exp βETθ (s0 |a,s) Rs,a
ψ(θ|a, s) ← Zβ (a,s)
π(a|s) ←
1
ρ(a|s) exp
Zα (s)
α
β
return π(a|s), ψ(θ|a, s), F (s)
4
Convergence
Here, we show that the value iteration scheme described through Algorithm 1 converges
to a unique fixed point satisfying Equation (8). To this end, we first prove the existence
of a unique fixed point (Theorem 1) following [17,18], and subsequently prove the
convergence of the value iteration scheme presupposing that a unique fixed point exists
(Theorem 2) following [19].
0
s
Theorem 1. Assuming a bounded reward function Rs,a
, the optimal free-energy vector
∗
F (s) is a unique fixed point of Bellman’s equation F ∗ = BF ∗ , where the mapping
B : R|S| → R|S| is defined as in equation (13)
Proof. Theorem 1 is proven through Proposition 1 and 2 in the following.
8
Grau-Moya et al.
Proposition 1. The mapping Tπ,ψ : R|S| → R|S|
"
Tπ,ψ F (s) = Eπ(a|s) −
1
π(a|s)
log
+
α
ρ(a|s)
1
ψ(θ|a, s)
Eψ(θ|a,s) − log
+
β
µ(θ|a, s)
h
s0
ETθ (s0 |a,s) Rs,a
#
i
+ γF (s )
. (14)
0
converges to a unique solution for every policy-belief-pair (π, ψ) independent of the
initial free-energy vector F (s).
Proof. By introducing the matrix Pπ,ψ (s, s0 ) and the vector gπ,ψ (s) as
Pπ,ψ (s, s0 ) := Eπ(a|s) Eψ(θ|a,s) [Tθ (s0 |a, s)] ,
"
h
0
s
gπ,ψ (s) := Eπ(a|s) Eψ(θ|a,s) ETθ (s0 |a,s) Rs,a
#
ψ(θ|a, s)
1
π(a|s)
1
− log
,
− log
β
µ(θ|a, s)
α
ρ(a|s)
i
Equation (14) may be expressed in compact form: Tπ,ψ F = gπ,ψ + γPπ,ψ F . By applying the mapping Tπ,ψ an infinite number of times on an initial free-energy vector F ,
the free-energy vector Fπ,ψ of the policy-belief-pair (π, ψ) is obtained:
i
Fπ,ψ := lim Tπ,ψ
F = lim
i→∞
i−1
X
i→∞
t
i
γ t Pπ,ψ
gπ,ψ + lim γ i Pπ,ψ
F,
i→∞
|
{z
}
t=0
→0
which does no longer depend on the initial F . It is straightforward to show that the
quantity Fπ,ψ is a fixed point of the operator Tπ,ψ :
Tπ,ψ Fπ,ψ = gπ,ψ + γPπ,ψ lim
i→∞
0
= γ 0 Pπ,ψ
gπ,ψ + lim
i−1
X
t=0
i→∞
= lim
i→∞
t
γ t Pπ,ψ
gπ,ψ
i
X
t
γ t Pπ,ψ
gπ,ψ
t=1
i−1
X
t
i
γ t Pπ,ψ
gπ,ψ + lim γ i Pπ,ψ
gπ,ψ = Fπ,ψ .
i→∞
|
{z
}
t=0
→0
Furthermore, Fπ,ψ is unique. Assume for this purpose an arbitrary fixed point F 0 such
i
that Tπ,ψ F 0 = F 0 , then F 0 = limi→∞ Tπ,ψ
F 0 = Fπ,ψ .
Proposition 2. The optimal free-energy vector F ∗ = maxπ extψ Fπ,ψ is a unique fixed
point of Bellman’s equation F ∗ = BF ∗ .
Information-Processing Constraints and Model Uncertainty in MDPs
9
Proof. The proof consists of two parts where we assume ext = max in the first part and
ext = min in the second part respectively. Let ext = max and F ∗ = Fπ∗ ,ψ∗ , where
(π ∗ , ψ ∗ ) denotes the optimal policy-belief-pair. Then
F ∗ = Tπ∗ ,ψ∗ F ∗ ≤ max max Tπ,ψ F ∗ =: Tπ0 ,ψ0 F ∗
π
ψ
{z
}
|
Induction
≤
Fπ0 ,ψ0 ,
=BF ∗
where the last inequality can be straightforwardly proven by induction and exploiting
the fact that Pπ,ψ (s, s0 ) ∈ [0; 1]. But by definition F ∗ = maxπ maxψ Fπ,ψ ≥ Fπ0 ,ψ0 ,
hence F ∗ = Fπ0 ,ψ0 and therefore F ∗ = BF ∗ . Furthermore, F ∗ is unique. Assume
for this purpose an arbitrary fixed point F 0 = Fπ0 ,ψ0 such that F 0 = BF 0 with the
corresponding policy-belief-pair (π 0 , ψ 0 ). Then
F ∗ = Tπ∗ ,ψ∗ F ∗ ≥ Tπ0 ,ψ0 F ∗
Induction
≥
Fπ0 ,ψ0 = F 0 ,
and similarly F 0 ≥ F ∗ , hence F 0 = F ∗ .
Let ext = min and F ∗ = Fπ∗ ,ψ∗ . By taking a closer look at Equation (13), it can
be seen that the optimization over ψ does not depend on π. Then
F ∗ = Tπ∗ ,ψ∗ F ∗ ≥ min Tπ∗ ,ψ F ∗ =: Tπ∗ ,ψ0 F ∗
Induction
≥
ψ
Fπ∗ ,ψ0 .
But by definition F ∗ = minψ Fπ∗ ,ψ ≤ Fπ∗ ,ψ0 , hence F ∗ = Fπ∗ ,ψ0 . Therefore it holds
that BF ∗ = maxπ minψ Tπ,ψ F ∗ = maxπ Tπ,ψ∗ F ∗ and similar to the first part of the
proof we obtain
F ∗ = Tπ∗ ,ψ∗ F ∗ ≤ max Tπ,ψ∗ F ∗ =: Tπ0 ,ψ∗ F ∗
π
{z
}
|
Induction
≤
Fπ0 ,ψ∗ .
=BF ∗
But by definition F ∗ = maxπ Fπ,ψ∗ ≥ Fπ0 ,ψ∗ , hence F ∗ = Fπ0 ,ψ∗ and therefore
F ∗ = BF ∗ . Furthermore, Fπ∗ ,ψ∗ is unique. Assume for this purpose an arbitrary fixed
point F 0 = Fπ0 ,ψ0 such that F 0 = BF 0 . Then
F 0 = Tπ0 ,ψ0 F 0 ≤ Tπ0 ,ψ∗ F 0
Induction
≤
Fπ0 ,ψ∗
Induction
≤
Tπ0 ,ψ∗ F ∗ ≤ Tπ∗ ,ψ∗ F ∗ = F ∗ ,
and similarly F ∗ ≤ F 0 , hence F ∗ = F 0 .
Theorem 2. Let be a positive number satisfying <
η
1−γ
where γ ∈ (0; 1) is the
0
s
discount factor and where u and l are the bounds of the reward function Rs,a
such that
0
s
l ≤ Rs,a ≤ u and η = max{|u|, |l|}. Suppose that the value iteration scheme from
Algorithm 1 is run for i = dlogγ (1−γ)
e iterations with an initial free-energy vector
η
F (s) = 0 for all s. Then, it holds that maxs |F ∗ (s) − B i F (s)| ≤ , where F ∗ refers to
the unique fixed point from Theorem 1.
10
Grau-Moya et al.
Proof. We start the proof by showing that the L∞ -norm of the difference vector between the optimal free-energy F ∗ and B i F exponentially decreases with the number of
iterations i:
max F ∗ (s) − B i F (s) =: F ∗ (s∗ ) − B i F (s∗ )
s
1
1
π(a|s∗ )
Eq. (9)
= max Eπ(a|s∗ )
log Zβ (a, s∗ ) − log
π
β
α
ρ(a|s∗ )
1
1
π(a|s∗ )
− max Eπ(a|s∗ )
log Zβi (a, s∗ ) − log
π
β
α
ρ(a|s∗ )
1
1
∗
i
∗
≤ max Eπ(a|s∗ )
log Zβ (a, s ) − log Zβ (a, s )
π
β
β
1
1
≤ max
log Zβ (a, s∗ ) − log Zβi (a, s∗ )
a
β
β
1
1
=:
log Zβ (a∗ , s∗ ) − log Zβi (a∗ , s∗ )
β
β
s0
1
ψ(θ|a∗ , s∗ )
Eq. (11)
∗ 0
= ext Eψ(θ|a∗ ,s∗ ) ETθ (s0 |a∗ ,s∗ ) Rs,a + γF (s ) − log
ψ
β
µ(θ|a∗ , s∗ )
s0
1
ψ(θ|a∗ , s∗ )
i−1
0
− ext Eψ(θ|a∗ ,s∗ ) ETθ (s0 |a∗ ,s∗ ) Rs,a + γB F (s ) − log
ψ
β
µ(θ|a∗ , s∗ )
≤ max Eψ(θ|a∗ ,s∗ ) ETθ (s0 |a∗ ,s∗ ) γF ∗ (s0 ) − γB i−1 F (s0 )
ψ
≤ γ max F ∗ (s) − B i−1 F (s)
Recur.
s
≤ γ i max |F ∗ (s) − F (s)| ≤ γ i
s
η
,
1−γ
where we exploit the fact that |extx f (x) − extx g(x)| ≤ maxx |f (x) − g(x)| and that
the free-energy is bounded through the reward bounds l and u with η = max{|u|, |l|}.
η
For a convergence criterion > 0 such that ≥ γ i 1−γ
, it then holds that i ≥
logγ
5
(1−γ)
η
presupposing that <
η
1−γ .
Experiments: Grid World
This section illustrates the proposed value iteration scheme with an intuitive example
where an agent has to navigate through a grid-world. The agent starts at position S ∈ S
with the objective to reach the goal state G ∈ S and can choose one out of maximally
four possible actions a ∈ {↑, →, ↓, ←} in each time-step. Along the way, the agent can
encounter regular tiles (actions move the agent deterministically one step in the desired
direction), walls that are represented as gray tiles (actions that move the agent towards
the wall are not possible), holes that are represented as black tiles (moving into the
hole causes a negative reward) and chance tiles that are illustrated as white tiles with a
question mark (the transition probabilities of the chance tiles are unknown to the agent).
Reaching the goal G yields a reward R = +1 whereas stepping into a hole results in
a negative reward R = −1. In both cases the agent is subsequently teleported back
Information-Processing Constraints and Model Uncertainty in MDPs
11
to the starting position S. Transitions to regular tiles have a small negative reward of
R = −0.01. When stepping onto a chance tile, the agent is pushed stochastically to an
adjacent tile giving a reward as mentioned above. The true state-transition probabilities
of the chance tiles are not known by the agent, but the agent holds the Bayesian belief
N (s)
Y s0 s0i
s0N (s)
s01
i
µ(θ s,a |a, s) = Dirichlet Φs,a
, . . . , Φs,a
=
(θs,a
)Φs,a −1
i=1
s0N (s)
s01
s0
where transition model is denoted as Tθs,a (s0 |s, a) = θs,a
and θ s,a = θs,a
. . . θs,a
and N (s) is the number of possible actions in state s. The data is incorporated into the
s0N (s)
0
s01
model as a count vector Φs,a
, . . . , Φs,a
where Φss,a represents the number of times
that the transition (s, a, s0 ) has occurred. The prior ρ(a|s) for the actions at every state
is set to be uniform. An important aspect of the model is that in the case of unlimited
observational data, the agent will plan with the correct transition probabilities.
We conducted two experiments with discount factor γ = 0.9 and uniform priors ρ(a|s) for the action variables. In the first experiment, we explore and illustrate
the agent’s planning behavior under different degrees of computational limitations (by
varying α) and under different model uncertainty attitudes (by varying β) with fixed
uniform beliefs µ(θ|a, s). In the second experiment, the agent is allowed to update its
beliefs µ(θ|a, s) and use the updated model to re-plan its strategy.
5.1
The Role of the Parameters α and β on Planning
Figure 1 shows the solution to the variational free energy problem that is obtained by
iteration until convergence according to Algorithm 1 under different values of α and β.
In particular, the first row shows the free energy function F ∗ (s) (Eq. (8)). The second,
third and fourth row show heat maps of the position of an agent that follows the optimal
policy (Eq. (12)) according to the agent’s biased beliefs (plan) and to the actual transition probabilities in a friendly and unfriendly environment, respectively. In chance tiles,
the most likely transitions in these two environments are indicated by arrows where the
agent is teleported with a probability of 0.999 into the tile indicated by the arrow and
with a probability of 0.001 to a random other adjacent tile.
In the first column of Fig. 1 it can be seen that a stochastic agent (α = 3.0) with
high model uncertainty and optimistic attitude (β = 400) has a strong preference for the
broad corridor in the bottom by assuming favorable transitions for the unknown chance
tiles. This way the agent also avoids the narrow corridors that are unsafe due to the
stochasticity of the low-α policy. In the second column of Fig. 1 with low α = 3 and
high model uncertainty with pessimistic attitude β = −400, the agent strongly prefers
the upper broad corridor because unfavorable transitions are assumed for the chance
tiles. The third column of Fig. 1 shows a very pessimistic agent (β = −400) with high
precision (α = 11) that allow the agent to safely choose the shortest distance by selecting the upper narrow corridor without risking any tiles with unknown transitions. The
fourth column of Fig. 1 shows a very optimistic agent (β = 400) with high precision. In
this case the agent chooses the shortest distance by selecting the bottom narrow corridor
that includes two chance tiles with unknown transition.
Grau-Moya et al.
Free Energy
α = 3 β = 400
S
?
?
?
?
?
?
α = 3 β = − 400
G
S
?
?
?
?
?
?
α = 11 β = − 400
G
S
?
?
?
?
?
?
G
α = 11 β = 400
S
?
?
?
?
?
?
1
G
0
Free Energy
12
1
0
1
2
3
4
5
6
7
8
S
?
?
?
G
G
S
S
?
?
?
?
?
?
G
G
S
S
?
?
?
?
?
?
G
G
S
S
?
?
?
?
?
?
G
Density
?
?
?
0.0
1.0
G
Density
S
0.0
1.0
S
G
0 1 2 3 4 5 6 7 8
S
G
S
G
S
G
Density
Unfriendly Environment
Friendly Environment
Plan
1.0
0.0
Fig. 1. The four different rows show free energy values and heat-maps of planned trajectories according to the agent’s beliefs over state-transitions in chance tiles, heat-maps of real trajectories
in a friendly environment and in an unfriendly environment respectively. The Start-position is
indicated by S and the goal state is indicated by G. Black tiles represent holes with negative reward, gray tiles represent walls and chance tiles with a question mark have transition probabilities
unknown to the agent. The white tiles with an arrow represent the most probable state-transition
in chance tiles (as specified by the environment). Very small arrows in each cell encode the policy π(a|s) (the length of each arrow encodes the probability of the corresponding action under
the policy, highest probability action is indicated as a red arrow). The heat map is constructed
by normalizing the number of visits for each state over 20000 steps, where actions are sampled
from the agent’s policy and state-transitions are sampled according to one of three ways: the second row according to the agent’s belief over state-transitions ψ(θ|a, s), in the third and fourth
row according to the actual transition probabilities of a friendly and an unfriendly environment
respectively. Different columns show different α and β cases.
Information-Processing Constraints and Model Uncertainty in MDPs
5.2
13
Updating the Bayesian Posterior µ with Observations from the Environment
2
S
G
2
3
3
4
4
0
1
2
3
4
S
0
Environment
?
0
0
1
2
3
4
Optimal
0.10
?
0.05
1
1
2
G
S
?
?
G
2
S
?
?
G
3
3
?
4
0
1
2
3
Pessimistic
?
0
1
2
3
Optimistic
4
β fixed
25
20
15
10
5
00
50 100 150 200 250 300
Step
0.10
50 100 150 200 250 300
Step
0.05
0.05
0.150
35
30
0.00
0.10
4
4
α fixed
Data points acquired
1
40
35
30
25
20
15
10
5
00
α = 12. 0 β = 0. 2
α = 12. 0 β = 5. 0
α = 12. 0 β = 20. 0
50 100 150 200 250 300
Step
Average reward
0
1
Average reward
0
Data points acquired
Similar to model identification adaptive controllers that perform system identification
while the system is running [20], we can use the proposed planning algorithm also in
a reinforcement learning setup by updating the Bayesian beliefs about the MDP while
executing always the first action and replanning in the next time step. During the learning phase, the exploration is governed by both factors α and β, but each factor has a
different influence. In particular, lower α-values will cause more exploration due to the
inherent stochasticity in the agent’s action selection, similar to an -greedy policy. If α
is kept fixed through time, this will of course also imply a “suboptimal” (i.e. bounded
optimal) policy in the long run. In contrast, the parameter β governs exploration of
states with unknown transition-probabilities more directly and will not have an impact
on the agent’s performance in the limit, where sufficient data has eliminated model
uncertainty. We illustrate this with simulations in a grid-world environment where the
agent is allowed to update its beliefs µ(θ|a, s) over the state-transitions every time it
enters a chance tile and receives observation data acquired through interaction with the
environment—compare left panels in Figure 2. In each step, the agent can then use the
updated belief-models for planning the next action.
0.00
0.05
0.10
0.150
α = 5. 0 β = 0. 2
α = 8. 0 β = 0. 2
α = 12. 0 β = 0. 2
50 100 150 200 250 300
Step
Fig. 2. The effect of α and β when updating beliefs over 300 interaction steps with the environment. The four panels on the left show the grid-world environment and the pertaining optimal
policy if the environment is known. The lower left panels show paths that the agent could take
depending on its attitude towards model uncertainty. The panels on the right show the number of
acquired data points, that is the number of times a chance tile is entered, and the average reward
(bottom panels) for fixed α (varying β) or fixed β (varying α). The average reward at each step
is computed as follows. Each time the agent observes a state-transition in a chance tile and updates its belief model, 10 runs of length 2000 steps are sampled (using the agent’s current belief
model). The average reward (bold lines) and standard-deviation (shaded areas) across these 10
runs are shown in the figure.
14
Grau-Moya et al.
Figure 2 (right panels) shows the number of data points acquired (each time a chance
tile is visited) and the average reward depending on the number of steps that the agent
has interacted with the environment. The panels show several different cases: while
keeping α = 12.0 fixed we test β = (0.2, 5.0, 20.0) and while keeping β = 0.2 fixed
we test α = (5.0, 8.0, 12.0). It can be seen that lower α leads to better exploration, but it
can also lead to lower performance in the long run—see for example rightmost bottom
panel. In contrast, optimistic β values can also induce high levels of exploration with
the added advantage that in the limit no performance detriment is introduced. However,
high β values can in general also lead to a detrimental persistence with bad policies, as
can be seen for example in the superiority of the low-β agent at the very beginning of
the learning process.
6
Discussion and Conclusions
In this paper we are bringing two strands of research together, namely research on
information-theoretic principles of control and decision-making and robustness principles for planning under model uncertainty. We have devised a unified recursion principle that extends previous generalizations of Bellman’s optimality equation and we have
shown how to solve this recursion with an iterative scheme that is guaranteed to converge to a unique optimum. In simulations we could demonstrate how such a combination of information-theoretic policy and belief constraints that reflect model uncertainty
can be beneficial for agents that act in partially unknown environments.
Most of the research on robust MDPs does not consider information-processing constraints on the policy, but only considers the uncertainty in the transition probabilities by
specifying a set of permissible models such that worst-case scenarios can be computed
in order to obtain a robust policy [11,12]. Recent extensions of these approaches include more general assumptions regarding the set properties of the permissible models
and assumptions regarding the data generation process [13]. Our approach falls inside
this class of robustness methods that use a restricted set of permissible models, because
we extremize the biased belief ψ(θ|a, s) under the constraint that it has to be within
some information bounds measured by the Kullback-Leibler divergence from a reference Bayesian posterior. Contrary to these previous methods, our approach additionally
considers robustness arising from the stochasticity in the policy.
Information-processing constraints on the policy in MDPs have been previously
considered in a number of studies [3,21,22,17], however not in the context of model
uncertainty. In these studies a free energy value recursion is derived when restricting
the class of policies through the Kullback-Leibler divergence and when disregarding
separate information-processing constraints on observations. However, a small number
of studies has considered information-processing constraints both for actions and observations. For example, Polani and Tishby [8] and Ortega and Braun [6] combine both
kinds of information costs. The first cost formalizes an information-processing cost in
the policy and the second cost constrains uncertainty arising from the state transitions
directly (but crucially not the uncertainty in the latent variables). In both informationprocessing constraints the cost is determined as a Kullback-Leibler divergence with
respect to a reference distribution. Specifically, the reference distribution in [8] is given
Information-Processing Constraints and Model Uncertainty in MDPs
15
by the marginal distributions (which is equivalent to a rate distortion problem) and in
[6] is given by fixed priors. The Kullback-Leibler divergence costs for the observations
in these cases essentially correspond to a risk-sensitive objective. While there is a relation between risk-sensitive and robust MDPs [23,24,25], the innovation in our approach
is at least twofold. First, it allows combining information-processing constraints on the
policy with model uncertainty (as formalized by a latent variable). Second, it provides
a natural setup to study learning.
The algorithm presented here and Bayesian models in general [9] are computationally expensive as they have to compute possibly high-dimensional integrals depending
on the number of allowed transitions for action-state pairs. However, there have been
tremendous efforts in solving unknown MDPs efficiently, especially by sampling methods [26,27,28]. An interesting future direction to extend our methodology would therefore be to develop a sampling-based version of Algorithm 1 to increase the range of
applicability and scalability [29]. Moreover, such sampling methods might allow for
reinforcement learning applications, for example by estimating free energies through
TD-learning [30], or by Thompson sampling approaches [31,32] or other stochastic
methods for adaptive control [20].
Acknowledgments This study was supported by the DFG, Emmy Noether grant BR4164/11. The code was developed on top of the RLPy library [33].
References
1. Richard Bellman. Dynamic Programming. Princeton University Press, Princeton, NJ, USA,
1 edition, 1957.
2. Emanuel Todorov. Linearly-solvable markov decision problems. In Advances in neural
information processing systems, pages 1369–1376, 2006.
3. Emanuel Todorov. Efficient computation of optimal actions. Proceedings of the national
academy of sciences, 106(28):11478–11483, 2009.
4. Daniel A Braun, Pedro A Ortega, Evangelos Theodorou, and Stefan Schaal. Path integral
control and bounded rationality. In Adaptive Dynamic Programming And Reinforcement
Learning (ADPRL), 2011 IEEE Symposium on, pages 202–209. IEEE, 2011.
5. Bart van den Broek, Wim Wiegerinck, and Hilbert J. Kappen. Risk sensitive path integral
control. In UAI, 2010.
6. Pedro A Ortega and Daniel A Braun. Thermodynamics as a theory of decision-making with
information-processing costs. In Proc. R. Soc. A, volume 469, page 20120683. The Royal
Society, 2013.
7. Pedro A Ortega and Daniel A Braun. Generalized thompson sampling for sequential
decision-making and causal inference. Complex Adaptive Systems Modeling, 2(1):2, 2014.
8. Naftali Tishby and Daniel Polani. Information theory of decisions and actions. In Perceptionaction cycle, pages 601–636. Springer, 2011.
9. Michael O’Gordon Duff. Optimal Learning: Computational procedures for Bayes-adaptive
Markov decision processes. PhD thesis, University of Massachusetts Amherst, 2002.
10. Shie Mannor, Duncan Simester, Peng Sun, and John N Tsitsiklis. Bias and variance approximation in value function estimates. Management Science, 53(2):308–322, 2007.
11. Arnab Nilim and Laurent El Ghaoui. Robust control of markov decision processes with
uncertain transition matrices. Operations Research, 53(5):780–798, 2005.
16
Grau-Moya et al.
12. Garud N Iyengar. Robust dynamic programming. Mathematics of Operations Research,
30(2):257–280, 2005.
13. Wolfram Wiesemann, Daniel Kuhn, and Berç Rustem. Robust markov decision processes.
Mathematics of Operations Research, 38(1):153–183, 2013.
14. István Szita and András Lőrincz. The many faces of optimism: a unifying approach. In
Proceedings of the 25th international conference on Machine learning, pages 1048–1055.
ACM, 2008.
15. István Szita and Csaba Szepesvári. Model-based reinforcement learning with nearly tight
exploration complexity bounds. In Proceedings of the 27th International Conference on
Machine Learning (ICML-10), pages 1031–1038, 2010.
16. Lars Peter Hansen and Thomas J Sargent. Robustness. Princeton university press, 2008.
17. Jonathan Rubin, Ohad Shamir, and Naftali Tishby. Trading value and information in mdps.
In Decision Making with Imperfect Decision Makers, pages 57–74. Springer, 2012.
18. DP Bertsekas and JN Tsitsiklis. Neuro-dynamic programming. 1996.
19. Alexander L Strehl, Lihong Li, and Michael L Littman. Reinforcement learning in finite
mdps: Pac analysis. The Journal of Machine Learning Research, 10:2413–2444, 2009.
20. Karl J Åström and Björn Wittenmark. Adaptive control. Courier Corporation, 2013.
21. Hilbert J Kappen. Linear theory for control of nonlinear stochastic systems. Physical review
letters, 95(20):200201, 2005.
22. J Peters, K Mülling, Y Altun, Fox D Poole, et al. Relative entropy policy search. In TwentyFourth National Conference on Artificial Intelligence (AAAI-10), pages 1607–1612. AAAI
Press, 2010.
23. Yun Shen, Michael J Tobia, Tobias Sommer, and Klaus Obermayer. Risk-sensitive reinforcement learning. Neural computation, 26(7):1298–1328, 2014.
24. Takayuki Osogami. Robustness and risk-sensitivity in markov decision processes. In Advances in Neural Information Processing Systems, pages 233–241, 2012.
25. Yinlam Chow, Aviv Tamar, Shie Mannor, and Marco Pavone. Risk-sensitive and robust
decision-making: a cvar optimization approach. In Advances in Neural Information Processing Systems, pages 1522–1530, 2015.
26. Stéphane Ross, Joelle Pineau, Brahim Chaib-draa, and Pierre Kreitmann. A bayesian approach for learning and planning in partially observable markov decision processes. The
Journal of Machine Learning Research, 12:1729–1770, 2011.
27. Arthur Guez, David Silver, and Peter Dayan. Efficient bayes-adaptive reinforcement learning
using sample-based search. In Advances in Neural Information Processing Systems, pages
1025–1033, 2012.
28. Arthur Guez, David Silver, and Peter Dayan. Scalable and efficient bayes-adaptive reinforcement learning based on monte-carlo tree search. Journal of Artificial Intelligence Research,
pages 841–883, 2013.
29. Pedro A Ortega, Daniel A Braun, and Naftali Tishby. Monte carlo methods for exact & efficient solution of the generalized optimality equations. In Robotics and Automation (ICRA),
2014 IEEE International Conference on, pages 4322–4327. IEEE, 2014.
30. Roy Fox, Ari Pakman, and Naftali Tishby. G-learning: Taming the noise in reinforcement
learning via soft updates. arXiv preprint arXiv:1512.08562, 2015.
31. Pedro A Ortega and Daniel A Braun. A minimum relative entropy principle for learning and
acting. Journal of Artificial Intelligence Research, pages 475–511, 2010.
32. Pedro A Ortega and Daniel A Braun. A bayesian rule for adaptive control based on causal
interventions. In 3d Conference on Artificial General Intelligence (AGI-2010). Atlantis Press,
2010.
33. Alborz Geramifard, Christoph Dann, Robert H Klein, William Dabney, and Jonathan P How.
Rlpy: A value-function-based reinforcement learning framework for education and research.
Journal of Machine Learning Research, 16:1573–1578, 2015.
| 3cs.SY
|
arXiv:1702.05641v1 [math.ST] 18 Feb 2017
On discrimination between two close distribution
tails.
Rodionov I. V.∗
March 17, 2018
1
Introduction. Main result.
Statistics deals often with discrimination of close distributions based on censored or truncated data, in particular, for high-risk insurances and reliability
problems. The situation when one observes data exceeding a pre-determined
threshold is well-studied, see [1], [2], [3] and references therein. On the other
hand statistics of extremes says that only higher order statistics should be used
for discrimination of close distribution tails, wherein moderate sample values
can be modeled with standard statistical tools. In particular, such approach for
distributions from Gumbel maximum domain of attraction (for the definitions
see [4]) is considered in [5], [6], [7]. As well, any estimators of the extreme value
indices γ and ρ (see [8]) can be used also to discriminate the distribution tails.
Notice that we do not assume belonging the corresponding distribution function
to a maximum domain of attraction.
Definition 1 The distribution functions F and G are said to be satisfied the
condition B(F, G) if for some ε > 0 and x0
1 − G(x)
is nondecreasing with x > x0 .
(1 − F (x))1−ε
(1)
Denote by Θ(F0 ) the class of continuous distribution functions F1 satisfying
either B(F1 , F0 ) or B(F0 , F1 ). Consider the simple hypothesis H0 : F = F0 and
the alternative hypothesis H1 : F ∈ Θ(F0 ), where F0 is continuous. Notice that
if distribution functions F, G satisfy either B(F, G) or B(G, F ) for some ε > 0
then it holds for all ε1 , 0 < ε1 < ε. So denote
ε(F, G) = max{ε : F, G satisfy either B(F, G) or B(G, F ) for ε}.
∗ Moscow State University, Faculty of Mathematics and Mechanics and Moscow Institute
of Physics and Technology, Faculty of Innovations and High Technologies. E-mail: [email protected]
1
Denote by Θε (F0 ) the class of continuous distribution functions F1 satisfying
either B(F1 , F0 ) or B(F0 , F1 ) with ε(F0 , F1 ) ≥ ε and consider another alternative hypothesis H1,ε : F ∈ Θε (F0 ). Let X1 , . . . , Xn be i.i.d. random variables
with a common distribution function F . Denote by X(1) ≤ . . . ≤ X(n) the order
statistics for them. Introduce the Hill-like statistics
Rk,n = ln(1 − F0 (X(n−k) )) −
1
k
n
X
i=n−k+1
ln(1 − F0 (X(i) )).
which we are going to use for the problem of discrimination between the two
introduced above hypotheses when k higher order statistics are known. Remark
that if F0 is Pareto distribution function with parameter γ, then
d
Rk,n = γH /γ,
where γH is the Hill estimator of γ. If furthermore F0 belongs to Fréchet maxdomain of attraction, then Rk,n behaves asymptotically as γH /γ, that is, theirs
ratio tends to one as n → ∞. We will show that the distributions of Rk,n if either
H0 or H1 fulfilled are different which can give a statistical for discrimination the
hypotheses. The following two results describe the behavior of Rk,n as k, n → ∞
with k < n provided H0 or H1 is fulfilled.
Theorem 1 If H0 holds then
√
d
k(Rk,n − 1) −→ ξ as k, n → ∞,
where ξ is standard normal random variable, i.e. ξ ∼ N (0, 1).
This theorem gives obvious goodness-of-fit test for the tail of F. Besides, the
following result provides some information about the consistency of this test.
Assume that H0 does not hold and F is equal to F1 which is different from F0 .
Denote x∗ , the right endpoint of F1 , that is, x∗ = sup{x : F1 (x) < 1}. Assume
that F0 and any F1 ∈ Θ(F0 ) have the same right endpoint (how to discriminate
distributions with different endpoints, see [10], [4]). Further consider x∗ = +∞,
otherwise change variables y = 1/(x∗ − x) gives the assumption. The following
theorem shows consistency of the proposed test.
Theorem 2
(i) If H1 holds then
p
d
kn |Rkn ,n − 1| −→ +∞
provided kn → ∞, kn /n → 0 as n → ∞.
(ii) If H1,ε holds then under the same conditions
inf
F1 ∈Θε (F0 )
p
d
kn |Rkn ,n − 1| −→ +∞.
2
The considered test makes it possible to discriminate, for example, two normal distributions with different variances, but we should weaken the condition
(1) to discriminate two normal distributions with the same variance and different means. But weakening the condition (1) imposes some conditions on
behavior of the sequence kn .
Definition 2 The distribution functions F and G are said to satisfy the condition C(F, G) if for some ε > 0 and x0
1 − G(x)
is nondecreasing,
(1 − F (x))(− ln(1 − F (x)))ε
x > x0 .
(2)
Denote by Θ′ (F0 ) the class of continuous distribution functions F1 satisfying
either C(F1 , F0 ) or C(F0 , F1 ) and the following condition: for some δ ∈ (0, 1)
1 − F1 (x) ≤ (1 − F0 (x))δ , x > x0 .
(3)
See, if distribution functions F, G satisfy either C(F, G) or C(G, F ) for some
ε > 0 then it holds for all ε1 , 0 < ε1 < ε. Denote
ε′ (F, G) = max{ε : F, G satisfy either C(F, G) or C(G, F ) for ε}.
Denote by Θ′ε (F0 ) the class of continuous distribution functions F1 satisfying (3)
and either C(F1 , F0 ) or C(F0 , F1 ) with ε′ (F0 , F1 ) ≥ ε. As before, consider the
simple hypothesis H0 : F = F0 and two alternative hypotheses H1′ : F ∈ Θ′ (F0 ),
′
H1,ε
: F ∈ Θ′ε (F0 ) with continuous F0 .
Theorem 3
(i) If H1′ holds then
p
d
kn |Rkn ,n − 1| −→ +∞
1/2−α
provided kn /n → 0, kn
/ ln n → +∞, for some α ∈ (0, 1/2), as n → ∞.
′
(ii) If H1,ε
holds then under the same conditions
p
d
inf′
kn |Rkn ,n − 1| −→ +∞.
F1 ∈Θε (F0 )
2
2.1
Auxiliary results and proofs.
Auxiliary results.
Since Rn depends on the higher order statistics we cannot immediately use independence of the random variables (X1 , . . . , Xn ). Therefore consider the conditional distribution of Rn given X(n−k) = q applying the following lemma.
Lemma 1 ([4]) Let X, X1 , . . . , Xn be i.i.d. random variables with common
distribution function F, and let X(1) ≤ . . . ≤ X(n) be the nth order statistics.
For any k = 1 . . . n− 1, the conditional joint distribution of {X(i) }ni=n−k+1 given
X(n−k) = q is equal to the (unconditional) joint distribution of the corresponding
∗ k
set {X(i)
}i=1 of order statistics for i.i.d. random variables {Xi∗ }ki=1 having the
distribution function
3
Fq (x) = P (X ≤ x|X > q) =
F (x) − F (q)
,
1 − F (q)
x > q.
We call Fq (x), x > q, the tail distribution function linked with the distribution
function F. Consider two continuous distribution functions F and G and a random variable ξq with distribution function Gq , where q ∈ R is some parameter.
Let
1 − F (q)
.
ηq = ln
1 − F (ξq )
Clear, ηq ≥ 0 for all q ∈ R.
The crucial point in the proof of Theorem 2 is studying of asymptotical behavior
of ηq .
Proposition 1 Let Fq and Gq are tail distribution functions of F and G respectively. Then
(i) If for some x0 , q > x0 , and any x > q, Fq (x) = Gq (x), then ηq is standard
exponential.
(ii) Gq (x) ≥ Fq (x) for any x > q if and only if ηq is stochastically smaller than
a standard exponential random variable.
Gq (x) ≤ Fq (x) for any x > q if and only if ηq is stochastically larger than
a standard exponential random variable.
(iii) Gq (x) ≥ Fq (x) for any x > q ≥ x0 and some x0 if and only if (1 −
G(x))/(1 − F (x)) is nonincreasing function as x > x0 .
2.2
Proof of Proposition 1.
(i) Let Fq (x) = Gq (x) for all x > q, then we have for the distribution function
of ηq ,
1 − F (q)
1 − F (q)
≤y =P
P (ηq ≤ y) = P ln
≤ ey =
1 − F (ξq )
1 − F (ξq )
1 − F (q)
←
−y
1−
= P ξq ≤ F
.
(4)
= P F (ξq ) ≤ 1 − (1 − F (q))e
ey
Furthermore, for the same x,
F F ← 1 − 1−F (q)
− F (q)
y
e
1 − F (q)
P ξq ≤ F ← 1 −
=
= 1 − e−y .
ey
1 − F (q)
(ii) Now assume that for all x > q and some q ∈ R, Gq (x) ≥ Fq (x). Then from
(4), since 1 − (1 − F (q))e−y ≥ F (q) for all y ≥ 0 it follows that
(q)
G F ← 1 − 1−F
− G(q)
y
e
P (ηq ≤ y) =
≥
1 − G(q)
4
F F← 1 −
1−F (q)
ey
1 − F (q)
− F (q)
= 1 − e−y .
(5)
Conversely, assume that ηq is stochastically smaller than a standard exponential
random variable, that is, P (ηq ≤ y) ≥ 1 − e−y for all y ≥ 0. With (4) we get
that
(q)
(q)
G F ← 1 − 1−F
− G(q)
1 − G F ← 1 − 1−F
y
ey
e
≥ 1−e−y ⇐⇒
≤ e−y
1 − G(q)
1 − G(q)
1 − G(q)
1 − F (q)
←
≤1−
⇐⇒
1−
⇐⇒ G F
y
e
ey
1 − F (q)
1 − G(q)
←
←
1−
F
1−
≤G
.
ey
ey
Denote zF = F ← (1 − e−y (1 − F (q))) and zG = G← (1 − e−y (1 − G(q))) . Since
F (zF ) = 1 − e−y (1 − F (q)) and G(zG ) = 1 − e−y (1 − G(q)), we have,
e−y =
1 − F (zF )
1 − G(zG )
=
.
1 − G(q)
1 − F (q)
Further, since zF ≤ zG then
1 − F (zF )
1 − G(zG )
1 − G(zF )
=
≤
.
1 − F (q)
1 − G(q)
1 − G(q)
This observation completes the proof since zF ∈ [q, ∞). The proof of the second
assertion is similar.
(iii) We have,
G(x) − G(q)
F (x) − F (q)
1 − G(x)
1 − F (x)
≥
∀x > q ≥ x0 ⇐⇒
≤
∀x > q ≥ x0 ⇐⇒
1 − G(q)
1 − F (q)
1 − G(q)
1 − F (q)
1 − G(q)
1 − G(x)
1 − G(x)
≤
∀x > q ≥ x0 ⇐⇒
is nonincreasing for all x > x0 .
1 − F (x)
1 − F (q)
1 − F (x)
2.3
Proof of Theorem 1.
Under the conditions of Theorem 1, F0 (X1 ) is uniformly distributed on [0, 1],
that is, F0 (X1 ) ∼ U [0, 1], hence − ln(1−F0 (X)) is standard exponential random
variable. It follows from Rényi’s representation (see [4]), that
− ln(1 − F0 (X(n−i) )) + ln(1 − F0 (X(n−k) ))
5
k−1
i=0
k−1
k
X
En−j+1
d
,
=
j
j=i+1
i=0
where E1 , E2 . . . are independent standard exponential variables. Therefore the
distribution of the left-hand side does not depend on n and
− ln(1 − F0 (X(n−i) )) + ln(1 − F0 (X(n−k) ))
k−1 d
=
i=0
E(k−i)
k−1
i=0
,
where E(1) ≤ . . . ≤ E(k) are the nth order statistics of the sample {Ei }ki=1 .
Finally we have,
!
k−1
k
X
X
√
√
√
1
1
d
k(Rk,n − 1) = k
E(k−i) − 1 = k
Ej − 1 ,
k i=0
k j=1
and the assertion follows from the Central Limit Theorem.
2.4
Proof of Theorem 2.
We first prove (i). The steps of the proof are similar to corresponding steps in
[6] and [7]. Consider asymptotic behavior of Rkn ,n as n → ∞. Denote
Yi = ln(1 − F0 (q)) − ln(1 − F0 (Xi∗ )),
n
are i.i.d. random variables introduced in Lemma 1 with the
where {Xi∗ }ki=1
distribution function
Fq (x) =
F1 (x) − F1 (q)
, q < x.
1 − F1 (q)
d
Taking F = F0 and G = F1 we have, Yi = ηq , i ∈ {1, . . . , kn }. Notice that,
n
of the
in view of Lemma 1, the joint distribution of order statistics {Y(i) }ki=1
kn
sample {Yj }i=1 is equal to the joint conditional distribution of order statistics
n
n
given X(n−kn ) = q, where
of {Zj }kj=1
{Z(j) }kj=1
Zj = ln(1 − F0 (X(n−kn ) )) − ln(1 − F0 (X(n−j+1) )), j = 1, ..., kn .
Clear,
Rn,kn =
kn
1 X
Zi .
kn i=1
So, the conditional distribution of Rkn ,n given X(n−k) = q is equal to the distriPkn
bution of k1n i=1
Yi . Further, distribution functions F1 and F0 satisfy B(F0 , F1 )
or B(F1 , F0 ). First suppose that the condition B(F0 , F1 ) holds for some ε > 0
and x0 . Since x∗ = +∞, X(n−kn ) → +∞ a.s., we may consider the case q > x0
only. Proposition 1 (iii) implies, that
(1 − F0 (x))1−ε
1 − F1 (x)
, x > x0 .
≥
1 − F1 (x0 )
(1 − F0 (x0 ))1−ε
6
With (5), we get that,
P (Y1 ≤ x) = 1 −
1−
1 − F1 F0← 1 −
1 − F0 F0← 1 −
1 − F1 (q)
1−ε
1−F0 (q)
ex
(1 − F0 (q))1−ε
1−F0 (q)
ex
≤
= 1 − e−(1−ε)x ,
hence Y1 is stochastically larger than a random variable E ∼ Exp(1 − ε), write
Y1 ≫ E. Further, let E1 , . . . , Ekn are i.i.d. random variables with distribution
function H(x) = 1 − e−(1−ε)x , then
!
!
kn
kn
p
p
1 X
1 X
Yi − 1 ≫ kn
Ei − 1 .
(6)
kn
kn i=1
kn i=1
Since (6) holds for all q > x0 , and X(n−kn ) → +∞ a.s. as n → ∞, we have
under the conditions of Theorem 2, that
!
kn
p
p
1 X
kn (Rkn ,n − 1) ≫ kn
Ei − 1 .
(7)
kn i=1
It follows from Lindeberg-Feller theorem, that
!
kn
p
1
1 X
d
Ei −
−
→ ξ ∼ N (0, 1), n → ∞,
(1 − ε) kn
kn i=1
1−ε
therefore
p
kn
kn
1 X
Ei − 1
kn i=1
!
P
−
→ +∞, n → ∞.
(8)
Finally, with (7), we have,
p
P
kn (Rkn ,n − 1) −
→ +∞, n → ∞.
If the condition B(F0 , F1 ) holds, then
p
P
kn (Rkn ,n − 1) −
→ −∞, n → ∞,
and the proof is similar. The second assertion easily follows from (7) and (8).
2.5
Proof of Theorem 3.
Firstly we prove (i). Denote F (x) = 1 − F (x). In notation of the proof of
Theorem 2, find the distribution of Y1 . First assume that C(F0 , F1 ) holds. With
(5) and Proposition 1 (iii) we have,
←
F1 F0 (F0 (q)e−x )
P (Y1 ≤ x) = 1 −
≤
F1 (q)
7
1−
ε
F0 (q)e−x − ln(F0 (q)e−x )
F (q)(− ln F0 (q))ε
For ε, c ∈ (0, 1),
=1−e
−x
1+
x
− ln F 0 (q)
ε
.
(1 + cx)ε ≥ 1 + cε − cεe−x , x ≥ 0,
and G(x) = 1 − e−x (1 + cε − cεe−x ) is the distribution function. Hence,
P (Y1 ≤ x) ≥ 1 − e−x −
ε
(1 − e−x ).
− ln F0 (q)
Further, let ζ, ζ1 , . . . , ζkn be i.i.d. random variables with this distribution function. Therefore, like the proof of Theorem 2,
!
!
kn
kn
p
p
1 X
1 X
kn
(9)
Yi − 1 ≫ kn
ζi − 1 .
kn i=1
kn i=1
Clear,
so we have,
p
kn
ε
, V ar ζ = 1 −
Eζ = 1 +
−2 ln F0 (q)
kn
1 X
ζi − 1
kn i=1
!
p
= kn
kn
1 X
ζi − Eζ
kn i=1
ε
−2 ln F0 (q)
!
+
p
kn
2
,
ε
.
−2 ln F0 (q)
(10)
√
Consider now the statistic kn / ln F0 (X(n−kn ) ). Denote Ri = F1 (Xi ), i =
1, . . . , n. Since F1 is continuous, R1 , . . . , Rn are i.i.d. standard uniform random variables and R(kn ) = F1 (X(n−kn ) ). Theorem 2.2.1 [4] implies, that
kn d
n
√
R(kn ) −
−
→ N (0, 1), n → ∞.
(11)
n
kn
Using the delta method (see [11]) for the function f (x) = −x/ ln x, we have
R(kn )
kn /n
n
P
√
−
→ 0, n → ∞,
−
− ln(kn /n)
kn − ln R(kn )
since under the conditions of theorem
kn
1
1
f′
=−
→ 0, n → ∞.
+
n
ln(n/kn ) (ln(n/kn ))2
Further,
R(kn )
kn /n
=
−
ln R(kn )
ln(kn /n)
p
1
kn /n
1
+ kn
,
−
−
ln(R(kn ) )
ln R(kn )
ln(kn /n)
n
√
kn
n
√
kn
R(kn )
ln R(kn )
8
and (11) implies that the first summand in the right hand side tends to 0 in
probability. Therefore,
p
1
1
P
kn
−
→ 0, n → ∞,
−
ln R(kn )
ln(kn /n)
and under the conditions of Theorem 3,
√
√
p
kn
kn
1
1
P
+
−
→ +∞, n → ∞.
= kn
−
− ln R(kn )
− ln R(kn )
− ln(kn /n)
− ln(kn /n)
On the other hand, from (3) it follows that
√
√
√
kn
kn
kn
P
≥
=
−
→ +∞
←
−1 ln R
−δ
− ln F0 (X(n−kn ) )
− ln F0 F1 (R(kn ) )
(kn )
(12)
as n → ∞. Further, it follows from the Law of large numbers for triangular
arrays (see [9]), that for any ǫ > 0
!
kn
p
1 X
kn
ζi − 1 = oP (knǫ ), n → ∞.
kn i=1
It means that the term in the left hand side is asymptotically smaller in probability than knǫ . Hence for any q, given X(n−kn ) = q
p
kn
kn
1 X
Yi − 1
kn i=1
!
P
−
→ +∞, n → ∞,
and finally,
p
P
kn (Rkn ,n − 1) −
→ +∞, n → ∞.
If the condition C(F1 , F0 ) holds, then
p
P
kn (Rkn ,n − 1) −
→ −∞, n → ∞,
and the proof is the same. The second assertion clearly follows from (9), (10)
and (12).
References
[1] Dufour R., Maag U.R. Distribution Results for Modified KolmogorovSmirnov Statistics for Truncated or Censored Samples. — Technometrics,
1978, v. 20, p. 29–32.
[2] Guilbaud O. Exact Kolmogorov-Type Test for Left-Truncated and/or
Right-Censored Data. — Journal of American Statistical Association, 1998,
v. 83, p. 213–221.
9
[3] Chernobai A., Menn C., Rachev S. T., Truck S. Estimation of operational
value-at-risk in the presence of minimum collection thresholds. — Tech.
Rep., University of California, Santa Barbara, Calif, USA, 2005.
[4] Fereira A., Haan L. de. Extreme value theory. An introduction. N. Y.:
Springer, Springer Series in Operations Research and Financial Engineering, 2006.
[5] Gardes L., Girard S., Guillou A. Weibull tail-distributions revisited: a
new look at some tail estimators. — Journal of Statistical Planning and
Inference, 2009, v. 141, p. 429–444.
[6] Rodionov I. V. A discrimination test for tails of Weibull-like distributions. — to appear in Probability Theory and its Applications.
[7] Rodionov I. V. Discrimination of close hypotheses on distribution tails using
higher order statistics. — to appear in Extremes.
[8] Haan L. de, Resnick S. Second-order regular variation and rates of convergence in extreme value theory. — The Annals of Probability, 1996, v. 24,
i. 1, p. 97–124.
[9] Mikusheva A. E. The law of large numbers and the logarithmic law for
arrays. — Fundamental and applied mathematics, 2000, v. 6, i. 1, p. 195–
206 (in Russian).
[10] Falk M. Some best estimators for distributions with finite endpoint. —
Statistics, 1995, v. 27, p. 115–125.
[11] Resnick S. I. A probability path. Birkhauser, 1999.
10
| 10math.ST
|
SideEye: A Generative Neural Network Based
Simulator of Human Peripheral Vision
arXiv:1706.04568v2 [cs.NE] 23 Oct 2017
Lex Fridman
MIT
Bryan Reimer
MIT
Benedikt Jenik
MIT
Christoph Zetzsche
University of Bremen
(a) Looking at the GPS navigation screen in the center stack.
Shaiyan Keshvari
MIT
Ruth Rosenholtz
MIT
(b) Looking at the pedestrian.
Figure 1: Peripheral vision simulation using a generative neural network trained on a behaviorally-validated model of human
peripheral vision. Our SideEye tool uses this model to simulate what an observer sees when foveating the location of the mouse
pointer. Note: In the figure on the left, when looking at the center stack, the pedestrian is not visible in the periphery.
ABSTRACT
INTRODUCTION AND RELATED WORK
Foveal vision makes up less than 1% of the visual field. The
other 99% is peripheral vision. Precisely what human beings
see in the periphery is both obvious and mysterious in that we
see it with our own eyes but can’t visualize what we see, except in controlled lab experiments. Degradation of information in the periphery is far more complex than what might be
mimicked with a radial blur. Rather, behaviorally-validated
models hypothesize that peripheral vision measures a large
number of local texture statistics in pooling regions that overlap and grow with eccentricity. In this work, we develop a
new method for peripheral vision simulation by training a
generative neural network on a behaviorally-validated fullfield synthesis model. By achieving a 21,000 fold reduction
in running time, our approach is the first to combine realism and speed of peripheral vision simulation to a degree that
provides a whole new way to approach visual design: through
peripheral visualization.
In the fovea (the central rod-free area of the retina, approximately 1.7° in diameter), recognition is relatively robust and
effortless. However, more than 99% of visual field lies outside the fovea, here referred to as the periphery. Peripheral vision has considerable loss of information relative to the fovea.
This begins at the retina, which employs variable spatial resolution to get past the bottleneck of the optic nerve. However,
it does not end there, but continues with neural operations
in visual cortex. Reduced peripheral acuity has only a tiny
effect, compared with peripheral vision’s sensitivity to clutter, known as visual crowding (discussed in more detail in
the “Crowding in Peripheral Vision” section). Unlike acuity
losses, which impact only tasks relying on quite high spatial frequencies, crowding occurs with a broad range of stimuli [20]. It is ever-present in real-world vision, in which the
visual system is faced with cluttered scenes full of objects and
diverse textures. Crowding constrains what we can perceive
at a glance.
(a) Original 512 × 512 pixel image.
(b) Foveated image with fixation on (256, 256).
Figure 2: Example “foveated” image. Given an original image (a), the foveated image provides a visualization of the information
available from a single glance (b) assuming the two subfigures are 14.2 inches away from the observer on a printed page (2.6
inches in height and width) and the observer’s eyes are fixated at the center of the original image. By selecting a new model
fixation, one can similarly get predictions for that fixation. Any part of the image that appears clearly in (b) is predicted to be
easy to perceive when fixating at the center of (a). The stripes in the flag are clear. The cat is clearly identifiable as such, and
clearly sits next to a book. The pictures within the book are not predicted to have a clear organization, given the modeled fixation.
The image in this example comes from the dataset used in the paper.
The field of human vision has recently made significant advances in understanding and modeling peripheral vision. A
successful model has been shown to predict performance at
a number of peripheral and full-field vision tasks [2, 8, 12,
27, 34]. Critically, from the point of view of converting this
understanding to design intuitions, researchers have visualized both reduced acuity and visual crowding using foveated
texture synthesis techniques that generate new image samples
that have the same texture statistics in a large number of overlapping “pooling” regions [8, 26]. Such visualizations facilitate intuitions about peripheral vision, e.g. for design [23],
and also enable testing models of peripheral vision. However, generating each synthesis can take a number of hours,
limiting the utility of this technique.
In this work, we develop and release a Foveated Generative
Network (FGN) architecture for end-to-end learning of the
foveation task and an online tool (SideEye) for real-time simulation of peripheral vision on user-submitted designs. The
primary goal of this approach is to reduce the running time
of generating a human-realistic visualization of peripheral vision from hours to milliseconds while maintaining reasonable
consistency with the behaviorally validated models. Fig. 2
shows an example visualizing the degradation of spatial information in the periphery. Being able to perform a visualization
like this in under a second has several significant applications
(see list below). As discussed in the “Running Time Performance” section, the average running time of 700 ms could be
further significantly reduced. As it approaches 33 ms (or 30
fps), the following applications become even more feasible:
• Interface Design: Explore various graphic user interface
design options with the SideEye tool on the fly by adjusting
the fixation point and visualizing the full-field appearance
of the design given the fixation point in near real-time (see
Fig. 1). One example of this application is the A/B testing
of website designs [10]. An illustrative case study of this
testing-based design methodology is presented in the the
“Application Case Study: A/B Testing of Design Layouts”
section. Communicate design intuitions and rationale to
members of the design or product team.
• Insights into Usability and Safety: Quickly gain intuitions about HCI issues such as whether, in an automotive
context, a user is likely to notice obstacles (i.e., pedestrians) while engaged with a cell phone, GPS system, or augmented reality headset. An example of this type of exploration using the SideEye tool is shown in Fig. 1.
• Behavioral Evaluation of Vision Model on HCIrelevant stimuli and tasks: The existing peripheral vision model has been extensively tested on a wide range
of stimuli and tasks, providing confidence that the model
extends to HCI-relevant visual tasks. Nonetheless, additional testing is always desireable. Previous model testing has utilized peripheral vision visualizations to generate
model predictions (see, e.g. [34], for the standard methodology). For HCI tasks, this requires generating hundreds
of foveated images dependent on subject fixation patterns.
FGN can generate the needed set of foveated images onthe-fly as the subject is performing the experiment.
field means that we cannot hope to understand much of vision
without understanding, controlling for, or otherwise accounting for the mechanisms of visual crowding.
• Video Foveation: Fast image foveation can be applied to
individual frames of a video. This is an important step
toward producing a model of peripheral vision in realworld viewing. However, there are further modeling challenges like accounting for peripheral encoding of motion
and maintaining temporal consistency would need to be
added to the architecture in order make video foveation a
powerful tool to explore human processing of spatiotemporal visual information.
A fair assessment of the current state of vision research is
that there exists a dominant theory of crowding. Crowding
has been equivalently attributed to “excessive or faulty feature integration”, “compulsory averaging”, or “forced texture
processing” “pooling”, resulting from of features over regions
that grow linearly with eccentricity [2,14,15,17,20]. Pooling
has typically been taken to mean averaging [17] or otherwise
computing summary statistics [2, 14] of features within the
local region.
The main contribution of this work is to use a deep learning approach to make a model of human vision fast enough
to provide a useful design tool. This will facilitate more effective communication of visual information, better usability, early insight into performance of a given design prior
to user testing, and better communication within the design
team [25]. To further use of this tool, we release the code and
an online in-browser version at https://****.***.***/
peripheral. We demonstrate that the resulting model successfully approximates the state-of-the-art behaviorally validated model, and yet is 21,000 times faster.
A Statistical Model
MODELING PERIPHERAL VISION
Crowding in Peripheral Vision
It is well known that the visual system has trouble recognizing peripheral objects in the presence of nearby flanking
stimuli, a phenomenon known as crowding (for reviews see:
[15, 20, 31]). Fig. 4 shows a classic demonstration. Fixating
the central cross, one can likely easily identify the isolated ‘A’
on the left but not the one on the right flanked by additional
letters. An observer might see these crowded letters in the
wrong order, e.g., ‘BORAD’. They might not see an ‘A’ at all,
or might see strange letter-like shapes made up of a mixture
of parts from several letters [14]. Move the flanking letters
farther from the target ‘A’, and at a certain critical spacing
recognition is restored. The critical spacing is approximately
0.4 to 0.5 times the eccentricity (the distance from the center
of fixation to the target) for a fairly wide range of stimuli and
tasks [3, 18, 19]. Authors in [20] have dubbed this Bouma’s
Law.
It should be clear that such “jumbling” in the periphery has
profound consequences for HCI design. In fact, crowding
is likely task-relevant for most real-world visual stimuli and
tasks. It has a far greater impact on vision than loss of acuity or color vision, and it is the dominant difference between
foveal and peripheral vision [24]. It impacts visual search,
object recognition, scene perception, perceptual grouping,
shape perception, and reading [20, 26, 27]. Crowding demonstrates a tradeoff in peripheral vision, in which significant information about the visual details is lost, and yet considerable information remains to support many real-world visual
tasks and to give us a rich percept of the world. The information that survives must suffice to guide eye movements and
give us a coherent view of the visual world [22]. Nonetheless, the pervasive loss of information throughout the visual
Authors in [2] operationalized earlier theories of statistical
processing in peripheral vision [14, 17] in terms of measurement of a rich set of texture statistics within local pooling
regions that grow linearly with eccentricity, in accord with
Bouma’s Law. They used as their candidate statistics those
identified by authors in [21], as that set has been successful at
describing texture perception (as judged by synthesis of textures that are difficult to discriminate from the original). Authors in [8] generalized this model to synthesize images from
local texture statistics computed across the field of view, often referred to as the “V2 model”, as they suggested that these
computations might occur in visual processing area V2 in the
brain. Authors in [26] have similarly developed full-field synthesis techniques, and refer to the full model as the Texture
Tiling Model (TTM).
Fig. 3 shows four examples of foveated images generated by
TTM. Note that because of the statistical nature of the model,
each input image corresponds to a large number of output images that share the same texture statistics. Critically, these
synthesized model images aid intuitions about what visual
tasks will be easy to perform at a glance, i.e. with the information available in the periphery plus the high resolution
fovea. Regions of the synthesized image that appear clear
are well represented by peripheral vision, according to the
model. Tasks that are easy to perform with the synthesized
images will be easy to perform at a glance. For instance, the
model predicts that it is obvious at a glance that the first image in the figure is of a cat, sitting amongst books and below
a flag. However, the layout and content of the open book may
be difficult to discern.
The V2 model and TTM [8, 26] share many features, including the local texture statistics, and overlapping pooling regions that overlap and grow linearly with eccentricity. This
paper utilizes the TTM synthesis procedure, so we adopt that
terminology.
Mounting evidence supports TTM as a good candidate model
for the peripheral encoding underlying crowding; it predicts
human performance at peripheral recognition tasks [2, 8, 12,
26], visual search [27, 34], and scene perception tasks [26],
and equating those local statistics creates visual metamers [8].
Both the V2 model and TTM are slow to converge, as they
must optimize to satisfy a large number of constraints arising
from the measured local texture statistics. Authors in [13], on
Original Image
TTM Foveation
FGN Foveation
Figure 3: Four examples (rows) from the evaluation dataset showing the original image (column 1), TTM-based foveation of the
image (column 2), and FGN-based foveation of the image (column 3). Unlike the deterministic radial blur function in the “Naive
Foveation” section, TTM is stochastic and can generate an arbitrary number of foveated images (based on a random number
generator seed) from a single source image. Therefore, one should not expect the TTM and FGN foveations to match pixel by
pixel, but rather their synthesis should have similar density and type of information degradation in the periphery.
{1,2,3}
Given an original undistorted color image xi, j
of dimen-
{4}
xi, j
Figure 4: A classic example demonstrating the “crowding”
effect.
the other hand, have taken a different approach to a related
problem. They apply simple image distortions, such as spatial warping, to an image, and have shown that it is surprisingly difficult to tell that anything is wrong away from the
fovea. Applying simple image distortions is fast to compute;
however, it is not well known what distortions best capture
the information available in peripheral vision; this distortion
work is not yet as well grounded in terms of being able to
predict task performance as TTM and the V2 model. Here
the aim is to use deep networks to produce distortions like
those introduced by TTM in a more computationally efficient
way.
A GENERATIVE MODEL FOR FOVEATED RENDERING
Fully Convolutional Networks as End-to-End Generators
Researchers have long desired to speed up successful computer vision and image processing models. In many cases
these successful models have taken an image as input, and
mapped that image to an output image through a slow optimization process. For example, mapping from a noisy image to a denoised image. Recent advances in neural networks
have provided a solution to speeding up some of these models, by learning the nonlinear mapping between the input and
output images for a given model. Once learned, one can map
from input to output using a relatively fast feed-forward network.
Fully convolution neural networks and other generative neural network models have been successfully used in computer vision literature for image segmentation [11, 16], deblurring [28, 29], denoising [4], inpainting [32], and superresolution [6], artifact removal [7], and general deconvolution [33]. While many of these examples learn a mapping
that removes image degradation, our work aims rather to add
degradation in a way that is representative of losses in human peripheral vision. As shown in the “Human-Realistic
Foveation” section, this is a highly nonlinear function in
that the mapping changes significantly with variation in both
global spatial context and in local texture statistics.
Neural networks have begun to be applied to the problem of
texture synthesis, i.e. synthesizing from an example texture
a new patch of perceptually similar texture. Authors in [9]
use the 16 convolutional and 5 pooling layers of the VGG19 network for texture synthesis. In the context of peripheral
vision, their work could be viewed as a method for synthesizing individual pooling regions as described in the “Modeling
Peripheral Vision” section. However, in addition to providing only a local texture synthesis, these new texture synthesis
techniques have not been behaviorally validated as models of
peripheral vision. Adding an behaviorally-validated, humanrealistic “attentional” (foveation) mechanism to a generative
network is a novel contributions of our work.
sion h × w × 3 and a spatial weight mask
of dimension
h×w×1, the task is to produce a foveated image, y, of dimension h × w × 3. The fourth channel of x captures the global
spatial component of the function to be learned as it relates to
the fixation point. The mask takes the form:
di, j =
q
xi, j,4 =
(i − fy )2 + ( j − fx )2
di, j , if di, j > dfovea
0,
otherwise
where di, j is the distance of each input pixel to the fixation
point ( fx , fy ), and dfovea is the radius (in pixels) of the foveal
region. For the results in the “Training and Evaluation of the
Generative Model” section, the fovea radius is 64 pixels.
The proposed foveated generative network (FGN) architecture is based on several components of CNN-based deconvolution approaches [4, 7, 33] and fully convolutional segmentation approaches [11,16]. A fully convolutional network
(FCN) can operate on large image sizes and produce output of
the same spatial dimensions. We extend the FCN architecture
with the foveation weight mask (see above) and propagate it
forward through the biases of each hidden layer in order for
the spatial relation with the fixation point to be accounted for
in computing the convolution and element-wise sigmoid for
each layer:
fk (x) = tanh (wk · fk−1 (x) + bk )
where wk and bk are the convolutions and biases at layer k,
respectively.
In our implementation of FGN, there are 4 convolutional layers w1,2,3,4 with w1 having 256 kernels of size 16 × 16 × 4,
w2 having 512 kernels of size 8 × 8 × 256, w3 having 512
kernels of size 1 × 1 × 512, and w4 having 3 kernels of size
8 × 8 × 512. The loss function is defined on the whole image
pair (x, Ti (x)) where Ti (x) is the output of the TTM model
on image x given a random seed of i. For purpose of FGN,
this forms a unique mapping between images, but it should be
noted that TTM can generate a very large number of images
Ti (x), ∀i ∈ N for a single input image x, since the number of
images that satisfy the statistical constraints imposed by the
optimization in TTM are upper-bounded by an exponential
function in the number image pixels.
Fig. 5 shows the fully convolutional architecture of FGN and
the TTM method used to generate the foveated image pairs.
The key aspect of the former is that the foveation is completed
with a single pass through the network.
TRAINING AND EVALUATION OF THE GENERATIVE
MODEL
We evaluate two models of foveation. The first is a naive
radial blur model known to be a poor visualization of peripheral perception as discussed in the “Introduction” section.
TTM: Texture Tiling Model
Original Image
Texture Synthesis Pooling Regions
Generated by TTM
“Ground truth”
used for training
FGN: Foveated Generative Network Model
Original Image
Convolutional Layers
Generated by FGN
Figure 5: The architecture of the foveated generative network (FGN) used to approximate the computationally-costly texture
tiling model (TTM). The output of the TTM is used as the ground truth for the end-to-end training of the FGN.
(a) Radial blur (ground truth).
(b) Radial blur (FGN generated).
(c) Pixel-based difference.
Figure 6: Example output of the radial blur computed directly (a) and learned through the FGN architecture (b). The pixel-bypixel difference between the two images (c) shows in black the pixels of the FGN-generated image that differs from the ground
truth.
However, it is a deterministic model for which there is a oneto-one mapping between source image and the ground truth.
Therefore, it is a good test of whether FGN can learn a function that is spatially dependent in a global sense on the fixation point, since the pixel-wise image difference is a more
reasonable metric of comparison for a deterministic model. In
addition, the greater visual interpretability of the naive radial
blur model allows us to gain intuition about the representation
learning power of the FGN model. It’s important to empha-
size, that the radial blur is a crude model of acuity loss that
cannot serve as a reasonable model of acuity loss in the periphery. In contrast to this, the second model of foveation we
consider is the TTM model that has been shown in behavioral
experiment to capture some of the more complex characteristics of perception in the periphery (i.e., crowding).
The original undistorted images in this paper are natural scene
images selected from the Places dataset [35]. 1,000 images were selected for training the FGN on both the radial
blur model (see the “Naive Foveation” section) and the TTM
model (see the “Human-Realistic Foveation” section). Another 1,000 images were used in evaluating FGN trained on
both models. All images were cropped and resized down to
512 × 512 pixels. For both training and quantitative evaluation, in this paper we assume a fixation at the center of the
input image, i.e. (w/2, h/2). For our application case studies
(the “Application Case Studies” section), we demonstrate the
ability to move the “fovea” of the trained model.
Naive Foveation: Training on Radial Blur Model Output
In order to evaluate the ability of FGN to learn a “foveating” function, we use a Gaussian blur with the standard deviation proportional to the distance away from the fixation. The
maximum standard deviation is set to 4 and decreases linearly with distance as both approach zero. Note that this blur
is made greater than that needed to mimic human peripheral
loss of acuity for the purpose of visualizing the effectiveness
of our training procedure. Fig. 6a shows the result of applying the radial blur on one of the images from test set. This
blurring function was applied to all 1,000 images in the training set and used as y in (x, y) image pairs for training an FGN
network to estimate the radial blur function. Fig. 6b shows
the result of running the image in Fig. 2a through the trained
network, and Fig. 6c shows the difference between this generated image and the ground truth.
Since radial blur is a deterministic function, we can estimate
the pixel error of the images generated by FGN. The trained
model was run on each of the 1,000 images in the test set
and achieved an average pixel difference of 2.3. Note that
the difference shown in Fig. 6c is inverted intensity-wise for
visualization clarity. This result is a quantifiable verification
that FGN can learn a simple radial blurring function, and thus
presumably can capture the loss of acuity in the periphery.
Human-Realistic Foveation: Training on TTM Output
The open question asked by this paper is whether a neural network can learn to degrade peripheral information in an image
in a way that is structurally similar to behaviorally validated
models like TTM. The results shown for 4 images in Fig. 5
and for 1,000 foveated test images made available online at
https://****.***.***/peripheral indicate that FGN is
able to capture many of the peripheral effects such as crowding and acuity loss. However, evaluating FGN’s ability to
capture the degree of this degradation not as straightforward
as evaluating a radial blur model. The TTM model produces
multiple output images for each input image, which can look
radically different while still maintaining consistent texture
statistics. One does not expect the FGN output to look exactly
like any given TTM output. Furthermore, peripheral vision
loses substantial local phase (location) information, an effect
well captured by TTM. These two factors make it impossible to evaluate FGN through pixel-based comparison with the
output of TTM. We cannot simply look at the difference between the TTM image and the FGN output, as we did when
evaluating radial blur. Here we show that FGN and TTM produce qualitatively similar distortions, and evaluate the degree
to which the TTM and FGN outputs match on the statistics
explicitly measured by TTM.
In Fig. 3, the first column has the original images, the second
column has the TTM foveated images, and the third column
has the FGN Foveation. Visual inspection of these images
reveals several key observations. First, the fovea region with
the 64 pixel radius is reproduced near-perfectly (the average
pixel intensity difference is below 1.9). Second, the results
capture a number of known effects of crowding, including the
“jumbling” loss of position information, coupled with preservation of many basic features such as orientation and contrast,
and dependence of encoding quality on local image contents,
including fairly good preservation of homogeneous textured
regions [14, 15, 31]. For example, the readability of text in
the periphery of the second images is degraded significantly
by its nonuniform positional displacement. Third, visual acuity decreases with distance from the fixation point for all 4
images.
Statistical Validation of FGN
The FGN model was trained and evaluated based on its ability
to mimic, in a meaningful statistic way, the foveated images
produced by the TTM model. Therefore, statistical validation of FGN was performed by comparing its output to TTM
output over the same exact pooling regions that were used
for the original TTM generation process. In other words,
this comparison evaluates the degree to which FGN is able
to mimic the texture feature vector on a region by region basis and thereby mimic the information degradation modeled
by TTM.
Fig. 7) shows the per-image difference in the feature vector
representing the texture statistics in each pooling region over
that image. Each bar represents a unique image. The error
in each case is computed for each of the 1,000 images in the
test dataset and sorted from highest (left) mean error to lowest (right). Only the highest 100 errors are shown in the figure for clarity. The mean and standard deviation of the error
are computed for each image by aggregating over each of the
values in the summary statistics vector in each of the pooling
regions. All mean errors are below 8% for the comparison
with the TTM output.
Running Time Performance
TTM hyper-parameters were chosen such that texture synthesis convergence was achieved. For these parameters, the
average running time per image was 4.2 hours. The model is
implemented in Matlab and given the structure of underlying
iterative optimization is not easily parallelizable.
The FGN architecture was implemented in TensorFlow [1]
and evaluated using NVIDIA GTX 980Ti GPU and a 2.6GHz
Intel Xeon E5-2670 processor. The average running time per
image was 0.7 seconds. That is an over 21,000-fold reduction in running time for foveating an image. There are several
aspect of this performance evaluation that indicate the possibility of significant further reductions in running time: (1) no
code optimization or architecture pruning was performed, (2)
the running time includes I/O read and write operations on
a SATA SSD drive, and (3) the GPU and CPU are 2-4 years
behind the top-of-the-line affordable consumer hardware.
APPLICATION CASE STUDIES
Percent Error in Summary Statistics (%)
20
15
10
5
0
Figure 7: Percent difference (error) between FGN output and TTM output. The error is computed for each of the 1,000 images
in the test dataset and sorted from highest (left) mean error to lowest (right). Only the highest 100 errors are shown in this figure
for clarity. The mean and standard deviation of the error are computed for each image by aggregating over each of the values in
the summary statistics vector in each of the pooling regions. All mean errors are below 8%.
Peripheral vision simulation as an approach to design is enabled by this work through the release of two things: (1) an algorithm implementation and (2) an online tool. As described
in the “Running Time Performance” section, the algorithm
implementation uses TensorFlow and Python, and provides a
pre-trained generative network that takes an image as input
and produces a foveated image as output. It also takes several parameters that specify the size and position of the fovea.
These parameters control the preprocessing of the image before it is passed through the network.
The online tool (named SideEye) provides an in-browser
JavaScript front-end that allows a user to load in an image
that is then passed to a Linux server backend where the FGN
network run inference on the image with the fovea position in
each of 144 different locations (12x12 grid on an image). The
result is 144 images, each foveated at grided locations. When
visualized together with the SideEye tool, these images automatically foveate to the position of a hovering mouse (or,
for a smartphone/tablet, the last position a finger touched the
display). Two sample mouse position for a resulting foveated
set are shown in Fig. 1.
been used to reveal the information available in a glance at
the website. We modeled the perception of the page when the
user points his eyes at the first word (“Serving”) of the most
relevant content. Based on the output of FGN on the first design (Fig. 8), the model predicts that a first-time visitor may
be able to recognize a truck, localize the logo, and tell that
the background image contains a natural scene of some sort.
Beyond that, the user might notice a black region, with blue
below it, and an orange circle. The black region, at least, appears to contain some text. Furthermore, the model predicts
that the user should be able to distinguish a light rectangle at
the top of the page, and a darker rectangle across the middle
of the page, both of which the user may guess are menu bars.
This is all useful information for a user to obtain at a glance.
In the second version of the page, the model predicts that a
user will easily determine that there is text against the blue
background. The location of the corporate logo is now very
clear, as are both of the menu bars. The blue region in the top
half of the page contains a red rectangle, possibly containing
some text. Higher level knowledge about web pages likely
suggests that this is probably a button.
Application Case Study: A/B Testing of Design Layouts
Given these FGN outputs, a designer could conclude that the
button in both designs is “salient” in the sense of being an
attention-getting color. However, the button is more obviously a button in the second design, even when fixating several degrees away. Given these predictions, it is perhaps no
surprise that users more frequently clicked on the button to
request a quote when interacting with the second design. Visualization of human perception, coupled with this kind of
reasoning process may help a designer to arriving at the more
effective design prior to testing with users, as described in [5].
A case study of a shipping website describes a significant increase in customers requesting quotes [5] based on a redesign
shown in Fig. 8. During the design process, FGN could have
Application Case Study: Logo Design Analysis
Both the SideEye tool and the pre-trained FGN network can
be used to perform fast peripheral vision simulation in helping understand how a user visually experiences the design
under question when first glancing at it. In the follow subsections, we provide two case studies where SideEye is used
to provide insight into how the design in question may appear in the periphery and what that may mean for the overall
experience that the design is intended to provide.
(a) Old design snapshot (original image).
(b) Old design snapshot (image foveated on “S” in “Serving”).
(c) New design snapshot (original image).
(d) New design snapshot (image foveated on “S” in “Serving”).
Figure 8: Given original images (a, c), the foveated images provides a visualization of human-realistic perception of the information available from a single glance (b, d). This particular visualization assumes that the two subfigures are 14.2 inches away
from the observer on a printed page (2.6 inches in height and width) and the observer’s eyes are fixated on the “S” in “Serving”
(middle left).
The ability to recognize a logo at a glance, even in the periphery, has significant potential upside for brand exposure
and recall [30]. We can use peripheral vision simulation to
evaluate how quickly the recognizability of a particular logo
design degrades in the periphery. Understanding how the logo
appears in the observer’s periphery may help understand the
potential effectiveness of product placement in the physical
world and in layout design on the web or on the printed page.
Fig. 9 shows the application of FGN to two versions of famous logos: (1) an earlier version of the company’s logo
and (2) the current version of their logo. The left two
columns show logos in their undistorted version. The right
two columns visualize these logos as they may appear in the
periphery when the fovea is centered at the black cross to the
left of the logo in question.
Using the metric of peripheral recognition, the new Spotify, Airbnb, and Google logos appear to be an improvement,
while both the new eBay and Starbucks logos appear to have a
decreased recognizability in the periphery. It should be noted,
that there may be other metrics under which the results are reversed, such as how memorable the logo is on repeated viewing. In that sense, peripheral vision simulation may be a useful process in a broader analysis of logo design.
CONCLUSION
We show that a generative network with a spatial foveating
component can learn in an end-to-end way to efficiently estimate the output of a human-realistic model of peripheral vision. We achieve a 4 orders-of-magnitude decrease in running
time, from 4.2 hours per image to 0.7 seconds per image. This
kind of jump in performance opens the door to a wide vari-
Full-Field View of Logo Designs
Peripheral View of Logo Designs
Undistorted 512 x 512 Image
Foveating the Left-Most Point (x=0, y=256)
Earlier Version
Later Version
Earlier Version
Later Version
Figure 9: Visualization of how two version of a famous brand’s logo appear in the periphery. The left two columns show the
undistorted full-field views of the logos. The right two columns show peripheral visualizations of those logos when the fovea is
centered on the black cross to the left of each respective logo. This kind of peripheral visualization process can be used as a tool
for understanding the initial visual exposure to a design when it first appears in the observer’s periphery.
ety of applications from interface design to virtual reality to
video foveation for studying human behavior and experience
in real-world interactions.
Code, Data, and Future Work
The SideEye tool and the FGN source code are made publicly
available at https://****.***.***/peripheral. In addition, the TTM-generated images used for training are made
available online. Future work will extend the size of the TTM
dataset from 1,000 to 100,000 images. This will allow other
groups to propose better-performing end-to-end architectures
trained on the TTM model.
REFERENCES
1. Martın Abadi, Ashish Agarwal, Paul Barham, Eugene
Brevdo, Zhifeng Chen, Craig Citro, Greg S Corrado,
Andy Davis, Jeffrey Dean, Matthieu Devin, and others.
2015. TensorFlow: Large-scale machine learning on
heterogeneous systems. Software available from
tensorflow.org (2015).
2. Benjamin Balas, Lisa Nakano, and Ruth Rosenholtz.
2009. A summary-statistic representation in peripheral
vision explains visual crowding. Journal of vision 9, 12
(2009), 13–13.
3. Herman Bouma. 1970. Interaction effects in parafoveal
letter recognition. Nature 226 (1970), 177–178.
4. Harold C Burger, Christian J Schuler, and Stefan
Harmeling. 2012. Image denoising: Can plain neural
networks compete with BM3D?. In Computer Vision
and Pattern Recognition (CVPR), 2012 IEEE
Conference on. IEEE, 2392–2399.
5. Smritti Chawla. 2013. Case Study - Lead Generation
Rate Shoots Up by 232% with Masthead Changes.
(2013). https://vwo.com/blog/
abtesting-increases-lead-generation-rate/
[Online; accessed 10-August-2016].
6. Chao Dong, Chen Change Loy, Kaiming He, and
Xiaoou Tang. 2014. Learning a deep convolutional
network for image super-resolution. In Computer
Vision–ECCV 2014. Springer, 184–199.
7. David Eigen, Dilip Krishnan, and Rob Fergus. 2013.
Restoring an image taken through a window covered
with dirt or rain. In Proceedings of the IEEE
International Conference on Computer Vision. 633–640.
8. Jeremy Freeman and Eero P Simoncelli. 2011.
Metamers of the ventral stream. Nature neuroscience 14,
9 (2011), 1195–1201.
9. Leon A Gatys, Alexander S Ecker, and Matthias Bethge.
2015. Texture synthesis and the controlled generation of
natural stimuli using convolutional neural networks.
arXiv preprint arXiv:1505.07376 (2015).
10. Bruce Hanington and Bella Martin. 2012. Universal
methods of design: 100 ways to research complex
problems, develop innovative ideas, and design effective
solutions. Rockport Publishers.
11. Kai Kang and Xiaogang Wang. 2014. Fully
convolutional neural networks for crowd segmentation.
arXiv preprint arXiv:1411.4464 (2014).
12. Shaiyan Keshvari and Ruth Rosenholtz. 2016. Pooling
of continuous features provides a unifying account of
crowding. Journal of vision 16, 3 (2016), 39–39.
13. Jan Koenderink, Whitman Richards, and Andrea J van
Doorn. 2012. Space-time disarray and visual awareness.
i-Perception 3, 3 (2012), 159–165.
14. Jerome Y Lettvin. 1976. On seeing sidelong. The
Sciences 16, 4 (1976), 10–20.
15. Dennis M Levi. 2008. Crowding - An essential
bottleneck for object recognition: A mini-review. Vision
research 48, 5 (2008), 635–654.
16. Jonathan Long, Evan Shelhamer, and Trevor Darrell.
2015. Fully convolutional networks for semantic
segmentation. In Proceedings of the IEEE Conference on
Computer Vision and Pattern Recognition. 3431–3440.
17. Laura Parkes, Jennifer Lund, Alessandra Angelucci,
Joshua A Solomon, and Michael Morgan. 2001.
Compulsory averaging of crowded orientation signals in
human vision. Nature neuroscience 4, 7 (2001),
739–744.
18. Denis G Pelli, Najib J Majaj, Noah Raizman,
Christopher J Christian, Edward Kim, and Melanie C
Palomares. 2009. Grouping in object recognition: The
role of a Gestalt law in letter identification. Cognitive
Neuropsychology 26, 1 (2009), 36–49.
19. Denis G Pelli, Melanie Palomares, and Najib J Majaj.
2004. Crowding is unlike ordinary masking:
Distinguishing feature integration from detection.
Journal of vision 4, 12 (2004), 12–12.
20. Denis G Pelli and Katharine A Tillman. 2008. The
uncrowded window of object recognition. Nature
neuroscience 11, 10 (2008), 1129–1135.
21. Javier Portilla and Eero P Simoncelli. 2000. A
parametric texture model based on joint statistics of
complex wavelet coefficients. International Journal of
Computer Vision 40, 1 (2000), 49–70.
22. Martin Rolfs, Donatas Jonikaitis, Heiner Deubel, and
Patrick Cavanagh. 2011. Predictive remapping of
attention across eye movements. Nature neuroscience
14, 2 (2011), 252–256.
23. Ruth Rosenholtz. 2011. What your visual system sees
where you are not looking.. In Human vision and
electronic imaging. 786510.
24. Ruth Rosenholtz. 2016. Capabilities and Limitations of
Peripheral Vision. Annual Review of Vision Science 2, 1
(2016).
25. Ruth Rosenholtz, Amal Dorai, and Rosalind Freeman.
2011. Do predictions of visual perception aid design?
ACM Transactions on Applied Perception (TAP) 8, 2
(2011), 12.
26. Ruth Rosenholtz, Jie Huang, and Krista A Ehinger.
2012a. Rethinking the role of top-down attention in
vision: effects attributable to a lossy representation in
peripheral vision. Frontiers in Psychology (2012).
27. Ruth Rosenholtz, Jie Huang, Alvin Raj, Benjamin J
Balas, and Livia Ilie. 2012b. A summary statistic
representation in peripheral vision explains visual
search. Journal of Vision 12, 4 (2012), 14–14.
28. Christian J Schuler, Michael Hirsch, Stefan Harmeling,
and Bernhard Schölkopf. 2014. Learning to deblur.
arXiv preprint arXiv:1406.7444 (2014).
29. Jian Sun, Wenfei Cao, Zongben Xu, and Jean Ponce.
2015. Learning a convolutional neural network for
non-uniform motion blur removal. In Computer Vision
and Pattern Recognition (CVPR), 2015 IEEE
Conference on. IEEE, 769–777.
30. Bo van Grinsven and Enny Das. 2016. Logo design in
marketing communications: Brand logo complexity
moderates exposure effects on brand recognition and
brand attitude. Journal of Marketing Communications
22, 3 (2016), 256–270.
31. David Whitney and Dennis M Levi. 2011. Visual
crowding: A fundamental limit on conscious perception
and object recognition. Trends in cognitive sciences 15,
4 (2011), 160–168.
32. Junyuan Xie, Linli Xu, and Enhong Chen. 2012. Image
denoising and inpainting with deep neural networks. In
Advances in Neural Information Processing Systems.
341–349.
33. Li Xu, Jimmy SJ Ren, Ce Liu, and Jiaya Jia. 2014. Deep
convolutional neural network for image deconvolution.
In Advances in Neural Information Processing Systems.
1790–1798.
34. Xuetao Zhang, Jie Huang, Serap Yigit-Elliott, and Ruth
Rosenholtz. 2015. Cube search, revisited. Journal of
vision 15, 3 (2015), 9–9.
35. Bolei Zhou, Agata Lapedriza, Jianxiong Xiao, Antonio
Torralba, and Aude Oliva. 2014. Learning deep features
for scene recognition using places database. In Advances
in neural information processing systems. 487–495.
| 1cs.CV
|
Finding Multiple New Optimal Locations in a Road Network
Ruifeng Liu
Ada Wai-Chee Fu
Zitong Chen
The Chinese University of
Hong Kong
The Chinese University of
Hong Kong
The Chinese University of
Hong Kong
arXiv:1606.01340v2 [cs.DB] 14 Jun 2016
[email protected]
[email protected] [email protected]
Silu Huang
Yubao Liu
Univeristy of Illinois
Urbana-Champagne
Sun Yat Sen University
[email protected]
[email protected]
ABSTRACT
new outlet on a set of eligible road segments so that the maximum
distance from any client to its nearest outlet is minimized. This is
an example of MinMax. Another example is in city planning for
new hospitals; we may set an objective to minimize the maximum
distance from any residence to its nearest hospital. Solutions to
MinMax are useful for the location selection for facilities in general, such as schools, libraries, sports facilities, fire stations, post
offices, police stations, etc.
Xiao el al. [23] first investigated the MinMax location query
problem in road networks. They considered the problem with a single new server. Their solution adopts a divide-and-conquer paradigm,
and has a time complexity of O(|V | + |S| + |C|)2 log(|V | + |S| +
|C|). Recently, an improved algorithm for Minmax location query
was proposed in [5]. Based on a new concept called the nearest
location component, their proposed algorithm can typically reduce
much of the search space. The time complexity of their algorithm
is O(m|V |log|V | + |V ||C|log|C|), where m is typically much
smaller than |C|. The run time is significantly less than that of the
algorithm for Minmax location query in [23].
However, the above solutions assume that only one new server
is to be added, and become inapplicable when more servers are to
be built. For real-life applications, such as a pizza delivery chain
store, there can be planning of multiple new outlets in a city for
a fiscal year. We may also have requirements of multiple new locations for public facilities. While the minmax problem for a single new server is tractable, it becomes intractable for two or more
new servers. The multi-server version of the MinMax problem is
shown to be NP-hard in [6]. A greedy algorithm is proposed which
applies the single server solution repeatedly until k servers are settled. However, results from the greedy algorithm can be arbitrarily
poor. Our empirical studies show that the greedy approach has poor
approximation ratio when compared to the optimal solutions.
We tackle the MinMax problem on three fronts. Firstly, when k
is relatively large, we propose an approximation algorithm which
we show to be 3-approximate, and which in our experiments produces good quality in the solutions. Secondly, when k is very small,
say k = 2 or k = 3, it would be much more convincing to provide
an optimal solution. This is due to the fact that in the intended applications, each new server involves a significant cost, and a higher
computation cost is well justified. We assume a small parameter
k. For intractable problems, there may exist a parameter such that
instances with small parameter values can be solved quickly. With
the assumption of a small k, MinMax is a candidate of parameterized problems. We show that MinMax is in XP, which is the class
of parameterized problems that can be solved in time f (k)ng(k)
for some computable functions f and g, where n is the input size
We study the problem of optimal location querying for locationbased services in road networks, which aims to find locations for
new servers or facilities. The existing optimal solutions on this
problem consider only the cases with one new server. When two
or more new servers are to be set up, the problem with minmax
cost criteria, MinMax, becomes NP-hard. In this work we identify
some useful properties about the potential locations for the new
servers, from which we derive a novel algorithm for MinMax, and
show that it is efficient when the number of new servers is small.
When the number of new servers is large, we propose an efficient
3-approximate algorithm. We verify with experiments on real road
networks that our solutions are effective and attains significantly
better result quality compared to the existing greedy algorithms.
1.
INTRODUCTION
We study the problem of optimal location querying (OLQ) in a
road network. In this problem we are given a set C of weighted
clients and a set S of servers in a road network G = (V, E), where
V is a set of vertices and E is a set of edges. We are also given a set
E 0 of eligible edges in the graph where new servers can be built.
The objective of this query is to identify a set of k (k ≥ 1) locations
(points) on E 0 such that k new servers set up at these locations
can optimize a certain cost function, which is defined based on the
distances between the servers and the clients. Here, we focus on the
minmax cost function. The problem is how to add new servers on
top of the existing servers so that the maximum weighted distance
from any client to its nearest server is minimized. We call such a
problem a MinMax problem, or simply MinMax.
Solutions to the MinMax location problem have been found to
be useful in various applications such as location planning and
location-based marketing, and have been studied in recent works
[23, 5, 6]. Consider the case of a pizza chain that plans to build a
new outlet in the city, aiming to minimize the worst case delivery
time to clients. Given the set S of existing pizza outlets and the
set C of all clients, the company wants to find a location for the
1
and k is a fixed parameter [12]. To show that MinMax is in XP,
we introduce the concepts of client cost lines and potential server
points. Based on their properties, we derive an optimal solution
for MinMax which is efficient for small k values, as illustrated by
experimental results on real datasets. Thirdly, when k is small but
too costly for the basic optimal algorithm, we derive optimization
strategies to further tame the computation costs. We illustrate with
experiments that our strategies work well on real city road networks
even for k up to 10.
This paper is organized as follows. Section 2 contains the problem definitions. Related works are discussed in Section 3. Section
4 presents our approximation algorithm. Sections 5, 6, and 7 describe our optimal solution. Optimization strategies are discussed
in section 8. Section 9 contains our experimental results. We conclude in Section 10.
2.
v1 c2
Notation
G = (V, E)
C
S
E0
w(c)
N NS (c)
c.dist
CostS (c)
d(p1 , p2 )
Cost(ci , x)
CCLℓ (ci )
p.pos
p.cost
P SP
kSP
cmax(P )
Y.maxcost
We consider a problem setting as follows. Let G = (V, E) be
an undirected graph with vertex set V and edge set E. C is a set
of clients and S is a set of servers. Each server or client is located
on some edge in graph G. Each client c in C is associated with a
positive weight w(c), which represents the importance of a client.
For example, if each client c represents a residential area, then w(c)
can be the population of c. The distance between two points p1 and
p2 on graph G is denoted by d(p1 , p2 ).
Assume that for each client c there is a unique server which is
nearest to c (this can be easily satisfied by shifting each client and
server by a very small unique distance). We denote the server that
is nearest to client c by N NS (c) and the distance between c and
its nearest server by c.dist, i.e., c.dist = d(c, N NS (c)). The cost
value of c, denoted by CostS (c), is equal to w(c) × c.dist.
We are given a set of eligible edges E 0 ⊆ E. The problem is to
place a set of k new servers, where each new server can be located
at any point on any edge in E 0 as long as there is no other server
at that same point. The introduction of E 0 is to model real-life
scenarios where feasible locations for the new servers are typically
restricted by factors such as city planning and zoning, financial considerations and property ownership.
We study the following optimal location query problem.
c∈C
c∈C
4
v5 3 v7
2
c6 2 v4
4
5
6
v8
Description
road network with vertex set V and edge set E
set of clients, |C| = n
set of servers
set of eligible edges where new servers can be built
weight of client c
server in set S closest to client c
the distance between c and its closest server N NS (c)
cost of client c, CostS (c) = w(c) × c.dist
the distance between two points p1 and p2 on graph G
w(ci ) × d(ci , x)
Client Cost Line of client ci on EI ℓ
p = (x, y) is a point on a CCL, and p.pos = x
p = (x, y) is a point on a CCL and p.cost = y
potential server point
a candidate set of k potential server points
maxc∈C {w(c) × d(c, N NS∪P (c))}
maxp∈Y p.cost
We classify the related work into three types: facility location
problem in a spatial setting, optimal location queries in a spatial
setting, and optimal location queries in road networks.
[Location problems in a spatial setting]
The facility location problem has been studied in past years [8].
The problem is about a finite set C of clients and a finite set P of
potential locations where new facilities can be built. The objective
is to find a subset P ′ in P that optimizes a predefined cost function. It is proved that the problem is NP-hard and many studies
focus on approximation algorithms [4]. Surveys of location facility
problems can be found in [8, 19, 10].
MinMax bears some similarity to the minimax facility location
problem (also called the k-center problem): given a point set X,
distance measure d, find a set S ⊆ X, |S| = k, so that the value of
maxp∈X (minq∈S (d(p, q)) is minimized. This problem is shown
to be NP-hard [13]. √For Euclidean space, an algorithm with time
complexity O(|P |O( k) ) is known [14]. An (1 + ǫ) approximation
algorithm is given in [16] that runs in O(k|X| ) time, and the authors
show empirically that it is efficient for small k values, say k ≤
5. The farthest point clustering method, which greedily picks the
farthest point from the selected points until k centers are found,
achieves 2-approximation [11]. An integer programming approach
is taken by [2] and their empirical studies consider graphs with up
to 3038 vertices.
The simple plant location problem takes a set I of potential sites
for plants, a set of clients, costs for setting up plants at sites i ∈ I ,
and transportation costs from i ∈ I to j ∈ J as inputs. It computes
a set P ⊆ I for plant locations so that the total cost of satisfying
all client demands is minimal [15]. The online variant of facility
location is studied in [18], which assumes that the clients join the
network one at a time, and the problem is to construct facilities
incrementally whilst minimizing the facility and service cost.
In general, the above problems differ from ours in that no existing servers (facilities) are given, and there is a finite set of potential
locations in a Lp space, whereas we assume that a road network is
given with existing servers.
(1)
The minmax cost resulting from P is given by
cmax(P ) = max{w(c) × d(c, N NS∪P (c))}
c1
3. RELATED WORK
D EFINITION 1 (M IN M AX P ROBLEM ). Given a graph G with
a set S of existing server locations, the MinMax k-location query
asks for a set of k locations on eligible edges E 0 in G to construct a
set of k new servers that minimizes the maximum cost of all clients.
A set P is returned where
P
4
6
Some of our notations used are listed in the following table.
PROBLEM DEFINITION
P = arg min (max{w(c) × d(c, N NS∪P (c))})
v6
2
4
5
6
c4 1
c3 1
v2/s2
s1
v3/c5
4
2
(2)
E XAMPLE 1. The following figure shows a road network G.
Each line segment corresponds to an edge. Each dot corresponds to
a vertex, server or client. There are 8 vertices, v1 , ..., v8 , 2 servers,
s1 , s2 , and 6 clients, c1 , ..., c6 , in this road network. ”v3 /c5 ”
means that client c5 is on vertex v3 . The number next to each line
segment is the distance between the two end points of the line segment. For example, d(v5 , v7 ) = 3, N NS (c4 ) = s2 , c4 .dist =
d(c4 , s2 ) = 2. Let w(c4 ) = 3, the cost value of c4 , CostS (c4 ), is
equal to w(c4 ) × c4 .dist = 3 × 2 = 6.
2
this can give an arbitrarily poor result. A simple example is where
2 new servers are to be built for 2 clients. The best locations are
at the 2 client sites while the greedy algorithm chooses one site in
between the 2 clients.
We propose a simple algorithm AppMinMax that collects new
server locations in a set A, which repeatedly selects the next server
location at the location of ci with the maximum CostS∪A (ci ), until
there are k servers in A. AppMinMax is shown in Algorithm 1.
Algorithm 1: AppM inM ax(G, C, S, k)
1
2
3
4
5
6
Input
: G, S, k, eligible edges E 0 , sorted C: c1 , ..., cn
Output : minmax cost: minmax(G, C, S, k),A
begin
A ← ∅;
for m = 1, ..., k do
A ← A ∪ {c′ } where c′ is the client with highest cost in
c1 , ..., cn .;
Compute CostS∪A (ci ) for each client ci ∈ c1 , ..., cn .;
Return maximum CostS∪A (ci ) and A, where 1 ≤ i ≤ n.;
T HEOREM 1. Algorithm 1 is a 3-approximation algorithm for
metric distances.
[OLQ in a spatial setting]
As a variation of the facility location problem, optimal location
query (OLQ) assumes that the potential facility set P is the whole
graph and there are some facilities built in advance. Most existing
works consider Lp distances. The algorithm in [3] minimizes the
maximum cost between a client to each facility when a new facility is built. There have also been studies on the problem of finding
k locations to optimize the total weight of clients that become attracted by the new servers, assuming that each client is attracted to
its nearest server. We refer to this problem as MaxSum. [9] solves
the MaxSum location query in spatial network with three different methods. [7] considers the MaxSum problem in a different
way. Given some weighted points in spatial databases, the algorithm finds the location of a rectangular region that maximizes the
sum of the weights of all the points covered by the rectangle. [1]
studies both the MaxSum and MinMax problems in the L2 space.
Yet another variation is the min-dist optimal location problem,
which is to find a new site to minimize the average distance from
each client to its closest server. A progressive algorithm is proposed in [25], which finds the exact answer given a finite number
of potential locations. More recent works can be found in [20, 21].
The MaxBRNN problem, which finds an optimal region to build a
new server that can attract the maximum number of clients in the
L2 space, is studied in [22]. An improved solution is proposed in
[17].
[OLQ in road networks]
The problems of proximity queries, including M inM ax, among
sets of moving objects in road networks are studied in [24]. The
MinMax location query problem in road networks with k = 1
is first investigated in [23]. Their solution adopts a divide-andconquer paradigm. The time complexity of their algorithm is O(|V |+
|S| + |C|)2 log(|V | + |S| + |C|) for MinMax location query.
An improved algorithm for MinMax with k = 1 is given in
[5]. The authors introduce a new concept called the nearest location component for each client c, denoted by N LC(c, d), which
is defined to be a set of all points on edges in G at distances no
more than d from c. The algorithm first computes the shortest distance between each client c and its closest server. Then, it sorts
all clients by their cost values. N LCs are built starting from the
client with the largest cost to look for a maximum integer m, such
that certain critical intersection of the N LC(cj , dj ), where dj =
Cost(cm )/w(cj ) for j ∈ [1, m], is non-empty. The time for this
step is O(m|V |log|V |). An optimal location p is proved to be at
such an intersection, and the time complexity to find such a location is O(|V ||C|log|C|). The total complexity is O(m|V |log|V |+
|V ||C|log|C|). The problem becomes NP-hard when k > 1, an
approximation algorithm is proposed in [5].
4.
PROOF: Assume that n clients c1 , c2 , ..., cn are ordered by nondecreasing costs, i.e., CostS (c1 ) ≥ CostS (c2 ) ≥ ... ≥ CostS (cn ).
If Q is an optimal solution, there must exist an integer m, k + 1 ≤
m ≤ n, such that CostS (cm+1 ) ≤ cmax(Q) ≤ CostS (cm ).
Suppose the result returned by the approximation algortihm is A.
In a possible case, A may build k servers on clients c1 , ..., ck , so
cmax(A) ≤ CostS (ck+1 ). Q contains k new servers s′1 , ..., s′k .
Let Ci′ be the set of clients in {c1 , ..., cm } whose nearest server in
S ∪ Q is s′i . The cost CostS∪Q (c) of each client c in Ci′ is less
than or equal to cmax(Q). C1′ ∪ ... ∪ Ck′ = {c1 , ...cm }. Let us
take the following subgraph for illustration:
c’/sa’
si '
c”
s”
In Algorithm 1, in each iteration, we build a new server s′a on the
client c′ , where c′ has the highest cost. c′ must belong to a client set
Ci′ because CostS (c′ ) cannot be smaller than CostS (cm ). Suppose c′′ is another client in Ci′ and s′i is the nearest server of both
c′ and c′′ in optimal solution Q. According to triangle inequality, Cost(c′′ , s′a ) = w(c′′ ) ∗ (d(c′′ , c′ )) ≤ w(c′′ ) ∗ (d(c′′ , s′i ) +
d(s′i , c′ )) = w(c′′ )∗d(c′′ , s′i )+w(c′′ )∗d(s′i , c′ ). w(c′′ )∗d(c′′ , s′i )
is the cost of c′′ in the optimal solution Q, so w(c′′ ) ∗ d(c′′ , s′i ) ≤
cmax(Q). Also, if w(c′′ ) ≤ w(c′ ), w(c′′ ) ∗ d(s′i , c′ ) ≤ w(c′ ) ∗
d(s′i , c′ ) ≤ cmax(Q), so Cost(c′′ , s′a ) ≤ 2cmax(Q).
If w(c′′ ) > w(c′ ), suppose w(c′′ ) = γw(c′ ) where γ > 1
and s′′ is the nearest server of c′′ in S ∪ A′ where A′ is the set
of locations for new servers before adding c′ . Since in Algorithm
1, c′ has the highest cost in the current iteration, CostS∪A′ (c′′ ) ≤
CostS∪A′ (c′ ). Cost(c′′ , s′′ ) = CostS∪A′ (c′′ ) ≤ CostS∪A′ (c′ ) ≤
Cost(c′ , s′′ ) ≤ w(c′ ) ∗ (d(c′ , s′i ) + d(s′i , c′′ ) + d(c′′ , s′′ )) ≤
′′ ′′
γ+1
cmax(Q)+ cmax(Q)
+ Cost(cγ ,s ) , Cost(c′′ , s′′ ) ≤ γ−1
cmax(Q).
γ
′′
′
′′
′′
′
′′
Since Cost(c , sa ) ≤ w(c ) ∗ d(c , si ) + w(c ) ∗ d(s′i , c′ ) =
w(c′′ ) ∗ d(c′′ , s′i ) + γw(c′ ) ∗ d(s′i , c′ ) ≤ cmax(Q) + γcmax(Q),
Cost(c′′ , s′a ) ≤ (1+γ)cmax(Q). The cost of c′′ after building s′a
γ+1
, 1+γ)cmax(Q).
is min(Cost(c′′ , s′′ ), Cost(c′′ , s′a )) ≤ min( γ−1
γ+1
When γ > 1, it is easy to verify that min( γ−1 , 1 + γ) ≤ 3. So the
cost of c′′ after building s′a is smaller than or equal to 3cmax(Q).
In both cases, the cost of all other clients in Ci′ after building
′
sa on c′ will be smaller than 3cmax(Q). After adding c′ to A
and recomputing the cost of all clients in C, if the client with the
highest cost is in set Ci′ , that means all clients have a cost smaller
than 3cmax(Q). Otherwise, the chosen client should belong to
another set Cj′ where 1 ≤ j ≤ k and the cost of all clients in Cj′
will be reduced to less than 3cmax(Q) with a new server. Since
C1′ ∪ ... ∪ Ck′ = {c1 , ...cm }, within k iterations, we can put at
least one server in each client set if the client set contains a client
with the highest cost. The maximum cost returned by Algorithm
1 must be smaller than 3cmax(Q) after k iterations, we conclude
that Algorithm 1 is a 3-approximation algorithm.
APPROXIMATION ALGORITHM
A greedy algorithm is proposed in [6], which is to select one
optimal location at a time until k locations are selected. However,
3
4
c6
2
4
5
v4
6
v8
(a)
c4 1
c3 1
v2/s2
c1
4
c6
2
v5 3 v7
2
v4
p7(6.7,2.2)
p9(7.6,4.8)
c2
4
6
6
2
v6
2
4
5
4
v5 3 v7
s1
v3/c5
4
2
6
c1
v1 c2
4
5
6
v8
p1(1,5)
c6
p0(0,4) p2(3,3)
p4(8,2)
c5 c1
v3
(b)
p3(6,0)
(c)
c4
c2
p5(0,6)
v4
c3
p9
p7
p6(0,6)
v1
p8(0,7)v2
(d)
c1
p12(6,6)
p11(4,2)
Cost
v6
2
4
5
6
c4 1
c3 1
v2/s2
s1
v3/c5
4
Cost
2
Cost
v1 c2
c2
v1
c5
p10(2,0)
v3
(e)
Figure 1: (a) Graph G, w(c1 )=w(c2 )=w(c5 )=w(c6 )=1, w(c3 )=8, w(c4 )=3, N LC(c1 ) in bold ; (b) N LC(c2 ) in bold; (c) CCLs for
G: plotting Cost(ci , p) on edge [v3 , v4 ]; p0 , p1 , p2 , p3 and p4 are P SP s ; (d) CCLs for G: plotting Cost(ci , p) on edge [v1 , v2 ];
p5 , p6 , p7 , p8 and p9 are P SP s ; (e) CCLs for G: plotting Cost(ci , p) on edge [v1 , v3 ]; p10 , p11 , and p12 are P SP s
After adding c′ to A and recomputing the cost of all clients in
C, if the client with the highest cost is in set Ci′ , that means each
client has a cost of no more than 2cmax(Q). Otherwise, the client
chosen should belong to another set Cj′ where 1 ≤ j ≤ k and the
cost of all clients in Cj′ will be reduced to no more than 2cmax(Q)
with a new server. After k iterations, the maximum cost returned
by Algorithm 1 must be no more than 2cmax(Q), so cmax(A) ≤
2cmax(Q), which proves that Algorithm 1 is a 2-approximation
algorithm when all clients have unit weight.
Consider the time complexity of Algorithm 1. Computing the
distance from each client to its nearest server takes O(|V | log |V |+
|C|) time, as shown in [6]. Sorting takes O(|C|log|C|) time. After
building each new server, it takes O(|V |log|V |) time to recompute all client costs. Thus, the time complexity is O(|C|log|C| +
k|V |log|V |). The storage space for running Dijkstras algorithm is
O(|V |) and the space for storing costs of all clients is O(|C|) . The
storage complexity is O(|V | + |C|).
L EMMA 1. Algorithm 1 is not (3 − ǫ)-approximate for ǫ > 0.
Figure 2 (a) shows a road network with 2 existing servers, s1 , s2 ,
and 2 clients, c1 , c2 . d(s1 , c1 ) = α2 + α + β, d(c1 , c2 ) = α + 1 and
d(c2 , s2 ) = α + 1 where α > 0 and β > 0. Assume that the client
weights are the following: w(c1 ) = 1, w(c2 ) = α. c1 and c2 have
their nearest servers as s1 and s2 , respectively. Since the nearest
server of c1 is s1 ,√d(s1 , c1 ) = α2 + α + β < 2α + 2 = d(c1 , s2 ).
We can get α < 9−4β+1
< 2. CostS (c1 ) = w(c1 ) × c1 .dist =
2
α2 + α + β and CostS (c2 ) = α2 + α. The client ordering is c1 , c2 .
α+1
α+1
α2+α+β
s2
c1
c2
s1
(a)
α+1
α+1
α2+α+β
s2
c1/s3
c2
s1
(b)
Figure 2: (a) A road network with w(c1 )=1, w(c2 ) = α, α > 0
and β > 0; (b) Solution of the greedy algorithm when k = 1
Figure 2 (b) shows the case when k = 1. Algorithm 1 builds a
new server s3 at c1 . The nearest server of c1 is s3 and CostS (c1 )
= 0. The maximum cost of all client is cmax(A) = CostS∪A (c2 )
= α2 + α. In an optimal solution, if we build the new server between c1 and c2 and the distance to c1 is α, then CostS (c1 ) =
CostS (c2 ) = α and the maximum cost among all clients is cmax(Q)
= α. So, Algorithm 1 is not (3 − ǫ)-approximate for ǫ > 0 when
clients have different weights, since cmax(A)/cmax(Q) = α + 1
and 0 < α < 2, and we can set α > 2 − ǫ.
We have a better guarantee when clients are unweighted.
5. EXACT SOLUTION FRAMEWORK
Our exact solution will make use of the Nearest Location Components as defined in [5]: N LC(c) = {p|d(c, p) ≤ c.dist and p
is a point on an edge in G}. For each client c, N LC(c) is a set
of all points on edges in G with a distance to c of at most c.dist.
In Figure 1 (a), the bold line segments correspond to N LC(c1 ),
where the distance between each point in N LC(c1 ) to c1 is at most
c1 .dist = d(c1 , s2 ) = 8.
For clarity, we shall overload the symbol P to stand for both a
set of locations and the set of servers built on these locations, and
ci to also stand for the location of ci .
The MinMax problem is to find a set P of k locations on the
road network for new servers so that with the new and existing
servers, the maximum among the costs CostS∪P (c) of all clients
c ∈ C is minimized. Let us name the n clients c1 , c2 , ..., cn , so that
the clients are ordered by non-increasing costs, i.e., CostS (c1 ) ≥
CostS (c2 ) ≥ ... ≥ CostS (cn ).
Our solution framework is based on 2 main observations: (O1)
From the definition of N LC, it is clear that in order to reduce the
cost of a client c, a new server must be located on N LC(c). (O2)
Since the clients are sorted by CostS (c), if Q is an optimal solution, there exists m, k ≤ m ≤ n, such that CostS (cm+1 ) ≤
cmax(Q) ≤ CostS (cm ).
From (O1), promising locations for new servers are on the N LCs
of clients with greater costs. From (O2), it suffices to consider
N LCs for client sets of Ci = {c1 , ..., ci } with increasing i, and
to stop at Cm when the best solution P for Cm introduces a cost
CostS∪P (cj ), for some 1 ≤ j ≤ m, higher than CostS (cm+1 ).
This suggests the following iterative process:
L EMMA 2. Algorithm 1 is 2-approximate when all clients have
unit weight.
PROOF: Assume that n clients c1 , c2 , ..., cn are ordered by nondecreasing costs, i.e., CostS (c1 ) ≥ CostS (c2 ) ≥ ... ≥ CostS (cn ).
Each client has unit weight. If Q is an optimal solution, there must
exist an integer m, k + 1 ≤ m ≤ n, such that CostS (cm+1 ) ≤
cmax(Q) ≤ CostS (cm ). Suppose the result returned by the approximation algorithm is A. In a possible case, A consists of k
servers at the clients c1 , ..., ck , so cmax(A) ≤ CostS (ck+1 ).
Optimal solution Q contains k new servers s′1 , ..., s′k . Each new
server s′i ∈ Q is the nearest server of a set of clients Ci′ where
Ci′ ⊂ {c1 , ...cm } for 1 ≤ i ≤ k and the cost of each client in Ci′
is less than or equal to cmax(Q). C1′ ∪ ... ∪ Ck′ = {c1 , ...cm }. In
Algorithm 1, for each iteration we build a new server s′a on a client
c′ , where c′ must belong to a client set Ci′ . Since each client in
Ci′ has unit weight and they share the same nearest server s′i with
maximum cost cmax(Q), according to triangle inequality, the cost
of all other clients in Ci′ after building s′a will be no more than
cmax(Q) + cmax(Q) = 2cmax(Q).
4
given a point p = (x, y) on CCLℓ (ci ), y = Cost(ci , x). We also
denote x by p.pos and y by p.cost.
Algorithm 2: Solution Framework
1
2
3
4
5
Input : G, S, k, E 0 , sorted client set C: c1 , ..., cn
Output : minmax cost and k optimal locations
begin
for client set Ci = {c1 , ..., ci }, i ∈ k..n do
Step 1: On each edge, find all locations for new servers
which have the potential to be in the optimal solution;
Step 2: Examine each combination P of k or less locations
derived in Step 1. Find the combination P with the minimum
cmax(P ) value assuming that the client set is Ci . Break if
this minmax value is greater than CostS (ci+1 );
E XAMPLE 2 (CCL). Figure 1(c) shows the CCLℓ (ci ) for
clients c1 , c2 , c5 , c6 on the edge ℓ = [v3 , v4 ]. The x-axis is for
positions on ℓ, and the y-axis is for the costs. Figures 1(d) and 1(e)
show CCLs on edges ℓ = [v1 , v2 ] and ℓ = [v1 , v3 ].
The union of CCLℓ (ci ) for all ℓ is denoted by CCL(ci ). The
values of p.pos of points on CCLℓ (c) correspond to the projection
of the CCL on ℓ. The union of the projections of CCLℓ (ci ) for all
edges ℓ spans the set of possible positions for a new server to attract
ci . Given a point p = (x, y) on CCLℓ (ci ), if x is the location of a
new server, then y is a possible new minmax cost. For convenience,
we also say that a new server is placed at p when it is placed at x.
CCLℓ (ci ) consists of one or two linear pieces, which may be
connected or disconnected. CCLℓ (ci ) ends at a point (x, z) if fi
is undefined in a non-empty interval (x, x′ ] or [x′ , x). For example,
in Figure 1(c) , ℓ = [v3 , v4 ], and CCLℓ (c6 ) ends at the point (6, 4).
If CCLℓ (ci ) ends at (x, z), then z = CostS (ci ), and z is the
highest cost value for CCL(ci ).
If a new server is placed at a point on CCLℓ (ci ) other than an
end point, the cost for ci is reduced as ci .dist becomes smaller.
The following definition helps to describe how a point is related to
reduction in client costs.
Return the best k potential locations found and the corresponding
minmax cost;
For Algorithm 2, a naive procedure for Step 1 is the following:
on each edge, for each subset x of Ci , find a potential location for
a new server to minimize the maximum cost of clients in x. Let
us consider the road network in Example 1. Suppose k = 2; we
need to build 2 new servers. The clients ci are sorted in the order of
CostS (ci ). c1 , c2 , c3 have the highest cost of 8, and we first try to
reduce the costs of C2 = {c1 , c2 } with the new servers. N LC(c1 )
and N LC(c2 ) are shown in Figures 1(a) and 1(b). The two N LCs
overlap on edges [v1 , v3 ], [v3 , v6 ], [v3 , v4 ] and [v3 , v5 ]. An optimal
location among these would be a point on edge [v3 , v4 ] with the
same distance from c1 and c2 . A new server built at such a location
in this segment will reduce the costs of both c1 and c2 to the same
value. However, since we have two servers, we may also consider
the best locations to simply reduce the cost of c1 by a new server,
and the best locations to reduce the cost of c2 by another new server.
Thus, all possible subsets of C2 are considered on each edge. Step
2 in the solution framework will then pick the best combination of
k locations to be the choice for Ci .
The two steps are repeated with C3 = {c1 , c2 , c3 }. For reducing
the cost of c3 , we may consider segments on [v1 , v2 ]. The next
highest cost client is c4 , with CostS (c4 ) = 6. If the best possible
reduced costs for c1 , c2 , c3 are smaller than 6, then we also try to
reduce the cost of c4 by examining N LC(c4 ). This continues until
the current minmax cost is higher than the cost of the next highest
cost client.
However, the above naive procedure for Step 1 is costly. If there
are γ client N LCs overlapping with the edge ℓ, the number of
subsets will be O(2γ ) ⊆ O(2n ). The difficulty lies with how to
avoid this exponential factor. Based on the concepts of client cost
lines and potential server points, to be introduced in Section 6, we
are able to reduce the complexity of Step 1 to O(γ 2 ) ⊆ O(n2 )
for each edge, as the number of potential locations is limited to
O((γ 2 )). Such a solution is presented in Section 7.
6.
D EFINITION 2 (ci COVERED BY p). We say that a client ci is
covered by a point p = (x, λ) on a CCL on edge ℓ if given a new
server s at x on ℓ, and S ′ = S ∪ {s}, CostS ′ (ci ) ≤ λ.
Note that p does not need to be in CCLℓ (ci ) in the above definition. For example, in Figure 1(c), p1 covers c6 , since CostS (c6 )
is less than the cost at p1 . p1 also covers c1 and c2 , being the intersection of CCLℓ (c1 ) and CCLℓ (c2 ). The idea of covering is to
identify subsets of clients that do not need to be covered by other
new servers. In the above, c6 is “covered” by p1 because if a server
is built at p1 .pos attaining a cost of p1 .cost for c1 and c2 , then we
need not try to cover c6 by another server, since reducing the cost of
c6 does not reduce the overall minmax cost. In Figure 1(c), to cover
all the clients c1 , c2 , c5 , c6 by a new server on ℓ, the best position
for the new server is that of p1 , and the maximum Cost(ci , x) of
the above clients will be minimized within ℓ when x = p1 .pos.
However, with multiple new servers, c1 may be covered by a new
server on another edge, and if we only need to cover c2 , c5 , c6 by a
new server on ℓ, the best position is that of p0 instead.
Next, we introduce a notion for the effect of choosing a location
p on the minmax cost concerning a client set C ′ .
D EFINITION 3 (zmax(C ′ , p)). Given point p on CCLℓ (ci )
for ci ∈ C ′ and edge ℓ in G. Let C ′′ ⊆ C ′ be the set of clients
in C ′ not covered by p. Define CostS (∅) = 0 and if C ′′ 6= ∅,
CostS (C ′′ ) = max(CostS (c)|c ∈ C ′′ ). zmax(C ′ , p) is defined
to be max(p.cost, CostS (C ′′ )).
CCL AND PSP
In this section we study the issues in Step 1 of the solution framework. We introduce the notions of client cost lines (CCL) and
potential server points (P SP ), and derive some useful properties
concerning CCLs and P SP s.
E XAMPLE 3 (zmax(C ′ , p)). To reduce the overall minmax
cost we need to reduce the costs of some clients. Consider Figure 1(c). Suppose we aim to reduce the costs of c1 and c5 by
placing a new server on [v3 , v4 ]. Compare the choices of p1 and
p2 , both of which cover c1 and c5 . Since p1 .cost > p2 .cost,
p2 is a better choice. We have zmax({c1 , c5 }, p1 ) = p1 .cost >
zmax({c1 , c5 }, p2 ) = p2 .cost.
6.1 Client Cost Lines (CCL)
Let ℓ = [a, b] be an edge covered by N LC(ci ). Call a and b
the boundaries of ℓ. Let x be a point in ℓ such that x ∈ N LC(ci ).
Define Cost(ci , x) = w(ci ) × d(ci , x). Given client ci , define
function fi (x) = Cost(ci , x) for x ∈ N LC(ci ) ∩ ℓ, which is a
piecewise linear function of x. Note that fi (x) is undefined for x 6∈
N LC(ci ). Call the plotted line for function fi on the positions on
ℓ the Client Cost Line for ci on ℓ, denoted by CCLℓ (ci ). Hence,
Next, we consider some useful properties concerning zmax(C ′ , p).
For a set P of locations in G, let us define the term cmax(P ) =
5
In Figure 1(c), ℓ = [v3 , v4 ], I(ℓ) = {p1 , p2 }, and B(ℓ) =
{p0 , p3 , p4 }. P SP (ℓ) is set to {p0 , p1 , p2 , p3 , p4 }. A proof of
the above lemma is given in the Appendix. We can further restrict
the selection of locations for new servers on an edge ℓ by replacing
Rule R1 with Rule R3 in the following.
maxc∈C {w(c) ∗ d(c, N NS∪P (c))}. A set P with |P | = k and
minimum cmax(P ) is an optimal solution.
We map locations of a candidate solution to points of the CCLs.
Consider a set P = {x1 , ..., xk } of k locations on E 0 . Each xi
intersects the CCL projections of a set of clients Ci . For a client
c in C ′ = C1 ∪ ... ∪ Ck , compare the costs, Cost(c, xi ), xi in
x1 , ..., xk , and pick xj with the smallest Cost(c, xj ). We say that
c is assigned to xj . In Figure 1(c), if k = 2, and x1 = v3 , x2 = v4 ,
then c2 , c5 will be assigned to v3 and c1 , c6 will be assigned to v4 .
For each xi in P , among all assigned clients, pick c′ whose CCL
has a highest cost at xi . The point of CCL(c′ ) at xi is called the
apex point of xi , denoted by ai . E.g., suppose xi = v3 in Figure
1(c), if clients c2 , c5 are assigned to v3 , then p0 is the apex point ai
of xi .
The candidate solution P can thus be also defined by the k apex
points a1 , a2 , ..., ak . Let A = {a1 , ..., ak }. We say that P is
mapped to A. Define amax(A) = cmax(P ). We prove the following in Appendix 1.
R3 choose an intersection point of 2 CCL linear pieces where
the slope of one piece is positive and the other is negative.
E XAMPLE 5 (RULE R3). In Figure 1(d), there are two intersection points for CCLℓ (c3 ) and CCLℓ (c4 ), namely, p7 and p9 .
However, p9 is the intersection of two linear pieces both having a
positive slope. In contrast, p7 satisfies Rule R3.
L EMMA 5. A MinMax solution exists where the k new servers
are located at points selected by R2 or R3.
PROOF: Based on Lemma 4, we only need to exclude the points
in I which are intersections of two linear pieces with both positive
or both negative slopes. Without loss of generality, consider a point
p = (x, y) in I on the intersection of CCLℓ (c) and CCLℓ (c′ ),
the slopes of which at p are both negative, and ℓ = [a, b], also
Cost(c, z) > Cost(c′ , z) for z > y. With similar arguments as in
the proof of Lemma 4, we can find a point q on CCLℓ (c) which
is in I(ℓ) ∪ B(ℓ) such that for the set C ′ of clients covered by p,
zmax(C ′ , p) ≥ zmax(C ′ , q).
Now, we have a definition for P SP s.
L EMMA 3. Let P be a set of k locations which is mapped to a
set A of k CCL points. Let p ∈ A and C ′ be the set of clients
covered by p. Let q 6∈ A be another point on a CCL, where
zmax(C ′ , p) ≥ zmax(C ′ , q). Let A′ = A − {p} ∪ {q}, then
amax(A′ ) ≤ amax(A).
Lemma 3 states that we can replace CCL point p by q in the candidate solution P without increasing the overall minmax cost. We
shall make use of this lemma in minimizing the number of potential
locations for new servers in the next subsection.
D EFINITION 4. A potential server point (P SP ) is a point selected by Rule R2 or Rule R3.
6.2 Potential Server Points (P SP )
L EMMA 6. There are O(m2 ) potential server points on an edge
which overlap with the N LCs of m clients.
In this subsection, we study how to determine potential points
on the CCLs for placing new servers. Let us call such critical
CCL points the potential server points (P SP s). Denote the set
of P SP s for ℓ by P SP (ℓ). Our task is to determine P SP (ℓ) for
each ℓ.
Consider an edge ℓ = [a, b]. We define the lowest boundary
(LB) point of CCLℓ (c) as follows: If client c lies on ℓ, thus,
(c, 0) ∈ CCLℓ (c), then (c, 0) is the LB point of CCLℓ (c); otherwise, there are two cases: (1) if both points (a, ya ) and (b, yb ) are
on CCLℓ (c), and ya < yb (ya > yb ), then (a, ya ) ((b, yb )) is the
LB point of CCLℓ (c); (2) else, if only point (a, ya ) ((b, yb )) is in
CCLℓ (c), then (a, ya ) ((b, yb )) is the LB point of CCLℓ (c). We
say that the LB point is on ℓ.
PROOF: Each client contributes at most two linear pieces of CCLs
on an edge ℓ. From Lemma 5, a P SP is either an intersection point
of CCLs or a lowest boundary point; the lemma thus follows.
7. MINMAX ALGORITHM
Here, we derive an algorithm for MinMax based on the findings
above. The pseudocode is shown in Algorithm 3. Let the resulting
minmax value be minmax(G, C, S, k). Let kSP be the set of k
P SP s returned. The positions of kSP , {p.pos|p ∈ kSP }, are
the locations for the k new servers. We examine combinations of
k P SP points as candidate solutions. We call each such combination a k-candidate. Given a k-candidate Y , denote the highest
cost among the P SP s in Y by Y.maxcost, i.e., Y.maxcost =
maxp∈Y p.cost. According to our solution framework in Section
5, we try to incrementally cover a set of clients {c1 , ..., ci }, from
i = 1, ..., |C|. Hence, we consider the clients c with decreasing
CostS (c). The key idea is to stop at some value of i smaller than
|C|. We make use of the following lemma.
E XAMPLE 4 (LB POINT ). In Figure 1(c), p0 is the LB point
of CCLℓ (c2 ), where ℓ = [v3 , v4 ], and p3 is the LB point of c1 . In
Figure 1(d), p5 is the LB point of CCLℓ (c2 ), where ℓ = [v1 , v5 ],
and p6 is the LB point of c4 .
We shall show that we can restrict the selection of locations for
new servers on an edge ℓ to the following two rules:
L EMMA 7 ( EARLY TERMINATION ). Let C ′ = {c1 , ..., cm }.
If CostS (cm+1 ) ≤ minmax(G, C ′ , S, k) < CostS (cm ), then
minmax(G, C, S, k) = minmax(G, C ′ , S, k).
R1 choose an intersection point of 2 or more CCLs.
R2 choose a lowest boundary point of a CCLℓ (c).
We begin the search with {c1 , .., ck }, since k servers can reduce
the costs of k clients {c1 , .., ck } to zero by choosing the client locations for placing the servers. The for loop at Lines 7-25 iterates
for clients from ck+1 to cn . The iteration for cm is incremental on
the previous iterations on c1 , ..., cm−1 avoiding repeated computation for the P SP s and k-candidates. Each iteration adds a new
CCL(cm ) due to the consideration of a client cm (Line 9). Lines
Let us denote the set of intersection points of CCLℓ s by I(ℓ)
and the set of LB points on ℓ by B(ℓ). Our first attempt is to set
the potential locations based on I(ℓ) ∪ B(ℓ). Let I = ∪ℓ I(ℓ) and
B = ∪ℓ B(ℓ).
L EMMA 4. A solution for MinMax exists where the k new servers
are located at the positions of k points in I ∪ B.
6
s2 , respectively. We know that c1 .dist = 8, c2 .dist = 8, c3 .dist
= 1, c4 .dist = 2, c5 .dist = 4, c6 .dist = 4. Thus, CostS (c1 ) =
w(c1 ) × c1 .dist = 1 × 8 = 8. Similarly, we have CostS (c2 ) =
8, CostS (c3 ) = 8, CostS (c4 ) = 6, CostS (c5 ) = 4, CostS (c6 )
= 4. Then, we have CostS (c1 ) ≥ CostS (c2 ) ≥ CostS (c3 ) ≥
CostS (c4 ) ≥ CostS (c5 ) ≥ CostS (c6 ). The client ordering is
c1 , c2 , c3 , c4 , c5 and c6 .
Since k = 2, we build CCL(c1 ) and CCL(c2 ) on [v3 , v4 ],
[v1 , v2 ], and [v1 , v3 ] (Line 3). Next, we get P SP s p0 , p1 , p3 on
[v3 , v4 ], p5 on [v1 , v2 ], p10 , p12 on [v1 , v3 ] (see Figures 1(c) to
(e)). Thus, P = {p0 , p1 , p3 , p5 , p10 , p12 } (Lines 5 and 6).
When m = 3, we know that the third client is c3 and thus curmax
= CostS (c3 ) = 8 (Line 8). We build CCL(c3 ) (Line 9). No P SP
or CCL is removed (Lines 10 and 11). No P SP is updated since
no P SP in P covers c3 (Line 12) and we get P SP p8 generated
by CCL(c3 ) (Line 14). In the while loop, 2-candidate Y must
contain p8 (Line 17). When Y = {p1 , p8 }, it covers c1 , c2 and c3
(Line 18). The fourth client is c4 . Since curmax is updated to
Y.maxcost = p1 .cost = 5 < 6 = CostS (c4 ) (Lines 21 and 22), we
exit the loop.
When m = 4, we build CCL(c4 ) (Line 9). CCL(c1 ) on [v1 , v3 ]
and p12 are removed (Lines 10 and 11) since p12 .cost = 6 ≥ 6 =
curmax. No P SP is updated (Line 12). We get P SP s p6 and
p7 generated by CCL(c4 ) (Line 14) (see Figure 1(d)). Now P =
{p0 , p1 , p3 , p5 , p6 , p7 , p8 , p10 } and Y should contain at least one
P SP in p6 , p7 . Only when Y = {p1 , p7 }, Y covers c1 , c2 , c3 and
c4 . Since curmax = CostS (c4 ) = 6 > 5 = p1 .cost = Y.maxcost
(Line 19), curmax is updated to Y.maxcost = 5 (Line 21). The
next client is c5 and curmax = 5 > 4 = CostS (c5 ). Thus, at Line
25, curmax > CostS (cm+1 ) holds, 5 and {p1 , p7 } are returned
as the solution by the algorithm.
Algorithm 3: M inM ax(G, C, S, k)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
Input
: G, S, k, eligible edges E 0 , sorted C: c1 , ..., cn
Output : minmax cost: minmax(G, C, S, k),kSP
begin
P ← ∅;
for m = 1, ..., k do
build CCL(cm ) on each edge in E 0 ;
Compute set Q of P SP s and clients covered by each P SP on E 0 ;
P ← P ∪ {Q};
for m = k + 1, ..., n do
curmax ← CostS (cm );
build CCL(cm ) on edges in E 0 ;
remove P SP s in P with cost ≥ curmax;
remove CCL segments on edges in E 0 with minimum cost of
segment ≥ curmax;
update clients covered by P SP s in P (to include cm );
Q1 ← {p : p ∈ P ∧ p covers cm };
Q2 ← set of P SP s generated by CCL(cm ) in E 0 ;
P ← P ∪ Q2 ;
/* next consider k-candidates involving points in Q1 ∪ Q2 */;
while ∃ new or updated k-candidate Y not explored in P do
if Y covers c1 , ..., cm then
if Y.maxcost < curmax then
kSP ← Y ;
curmax ← Y.maxcost;
if curmax ≤ CostS (cm+1 ) then
break;
if curmax > CostS (cm+1 ) then
Return curmax, kSP
10-11 apply our pruning strategies, which will be explained shortly.
At Line 13, we store in Q1 the existing P SP s which have been
updated at Line 12 to include cm . At Line 14, we generate new
P SP s from CCL(cm ) in E 0 according to Lemma 5, and store
them in Q2 . A set P is used to maintain the P SP s that have been
generated and not pruned thus far. At Line 17, the while loop looks
for new or updated k-candidates. New k-candidates are sets of k
points in P with at least one point from Q1 ∪ Q2 . By ‘updated’
k-candidates, we refer to new or old candidates which may have a
new set of clients that they cover; in particular, cm can be added as
a covered client.
Over the course of the computation, let kSP be the current best
k-candidate, and let kSP.maxcost = curmax. We can safely
dismiss P SP s with costs higher than curmax. This provides for
dynamic pruning of P SP s at Line 10 and implicitly at Line 14,
which is shown to be highly effective in our empirical studies.
Similarly, given a linear piece of CCLl (c), if no point p =
(x, y) on CCLl (c) has a cost y <= curmax, we can dismiss
CCLl (c) (Line 11).
At Line 22, if the maximum cost of the P SP s in the k-candidate
is less than CostS (cm+1 ), it means that CostS (cm+1 ) will be the
minmax cost if we do not try to reduce the cost of cm+1 . To reduce
the minmax cost, the solution must include at least a new P SP
from CCL(cm+1 ) or an existing P SP which will be updated at
Line 12 to cover cm+1 . Hence, we can exit the while loop at Line
22, and process cm+1 next.
At Line 24, if curmax > CostS (cm+1 ), it implies that we cannot improve the minmax cost by considering clients cm+1 , ...., cn .
Thus, the algorithm stops.
Suppose the iterative process of Algorithm 3 stops when m =
γ at Line 7. Computing the N LCs and CCLs for the clients
c1 , ..., cγ takes O(γ|V | log |V |) time. Let ρ be the number of eligible edges in these N LCs. Let α be the maximum number of
P SP s for an edge, α = O(γ 2 ). The time to compute the coverage for each P SP is O(γ). The total time to compute coverage
is O(ραγ). The running time of Algorithm 3 is O(γ|V | log |V | +
ραγ + k(ρα)k ). This shows that MinMax is in XP [12] and is
computable for small k values. For memory requirement, we need
to store the clients covered by the P SP s, which requires O(ραγ)
storage. k-candidates are computed on the fly. The memory complexity is thus O(ραγ).
8. OPTIMIZATION
In the previous subsection, we describe an algorithm solving the
MinMax problem which takes O(γ|V | log |V | + ραγ + k(ρα)k )
time, where O(ρα) is the number of P SP s processed, and O((ρα)k )
is the number of k-candidates considered, which is the dominating
factor. In this subsection, we introduce enhancement techniques to
optimize the algorithm by reducing the computation cost of P SP s
and k-candidates. The proposed strategies reduced the number of
P SP s to 20 or less for k ≤ 10 in our real datasets, which makes
the optimal solution more effective.
8.1 Early Termination of Iterations
In the iteration for client cm in Algorithm 3, we compute P SP s
generated by CCL(cm ) to include in P , update clients covered by
P SP s and look for k-candidates in P . If we can jump to the next
iteration for cm+1 before these steps, we may reduce the computation cost substantially. To achieve this, we introduce two strategies
that allow us to go to next iteration before computing P SP s for the
E XAMPLE 6. Let us take the road network in Figure 1 as an illustration. Suppose k = 2 and E 0 = [v3 , v4 ],[v1 , v2 ],[v1 , v3 ]. Assume that the client weights are the following: w(c1 )=1, w(c2 )=1,
w(c3 )=8, w(c4 )=3, w(c5 )=1 and w(c6 )=1. Note that c1 , c2 , c3 ,
c4 , c5 and c6 have their nearest servers as s2 , s1 , s2 , s2 , s1 and
7
L EMMA 8. In the iteration of cm , given a virtual k-candidate
in Pedge that covers c1 ,...,cm , there must exist a k-candidate Z with
k actual P SP s that cover c1 ,...,cm with Z.maxcost ≤ CostS (cm+1 ).
current iteration. The pseudocode that incorporates these strategies
is given in Algorithm 4.
8.1.1 Strategy 1: kSP based early termination
PROOF: Given a virtual k-candidate Zv in Pedge which covers
the client set Cm = {c1 , ..., cm }. For each pℓ ∈ Zv on edge ℓ,
let Zℓ be the set of clients covered by pℓ where Zℓ ⊆ Cm . Consider ci ∈ Zℓ , by the definition of pℓ , and since CostS (ci ) >
CostS (cm+1 ) by the sorted order of all clients, it follows that
CCLℓ (ci ) must be a single linear segment with two end points.
Thus, the point hℓ with the highest cost among all CCLℓ (ci ) for
ci ∈ Zℓ must cover all clients in Zℓ . The set of such points hℓ
forms a k-candidate Z that covers c1 , ..,cm with Z.maxcost ≤
CostS (cm+1 ). The lemma follows from Lemma 5 since an optimal k-candidate can be obtained from the actual P SP s.
Consider the iteration for cm . After building CCL(cm ) on eligible edges, we check whether any P SP in the current best kcandidate kSP can cover cm at Line 14 in Algorithm 4. If one of
the P SP s in kSP covers cm and kSP.maxcost < CostS (cm+1 ),
we can keep kSP and jump to the next iteration to process cm+1 .
This is because we can be sure that the minmax cost would be
either CostS (cm+1 ), or that a smaller cost can be found with a
solution set that also covers client cm+1 .
If we jump to the next iteration early in the iteration of cm ,
P SP s to be generated by CCL(cm ) are not computed and P SP s
in P are not updated to include coverage of cm . Thus, we use a set
N to keep track of clients whose iterations are terminated early. As
we continue with the for loop and come to an iteration where we
cannot jump early, say in the iteration of ci where i > m, we compute and update P SP s for clients in set N at Lines 23 and 24. The
value curmax typically becomes much smaller than CostS (c) for
c ∈ N if we jump multiple iterations after c. Many CCLs can
be removed and many P SP s can be disregarded. The number of
P SP s to be processed can thus be significantly reduced.
Based on Lemma 8, at Line 18, if we get a k-candidate from
Pedge that covers c1 ,...,cm , we can update curmax to Cost(cm+1 )
and jump to the next iteration without computing P SP s for cm .
The size of Pedge is |E 0 |, which is much smaller than the size of
P , hence the processing is much faster.
In our empirical study, we can find a k-candidate in Pedge that
meets the requirements in most cases. The reason is that in the iteration of cm , only clients c with CCLl (c)s containing some point
p with p.cost ≥ Cost(cm+1 ) are not covered by pℓ . Since CCLs
with a minimum cost greater than curmax are removed at Line 12,
and curmax ≤ CostS (cm ), we only consider CCLs containing
some point p such that p.cost ≤ curmax ≤ CostS (cm ). As the
clients are sorted, CostS (cm+1 ) is typically close to CostS (cm ),
thus, p.cost is also likely to be ≤ Cost(cm+1 ). It is therefore
likely for pℓ to cover all clients c such that CCLl (c) is on edge l.
As with Strategy 1, we keep cm in set N if we skip computations
of P SP s due to cm , to be handled in a future iteration.
8.1.2 Strategy 2: Virtual P SP s based early termination
In iteration m, for each edge ℓ, we create a virtual P SP pℓ
with pℓ .cost = CostS (cm+1 ). pℓ can be considered as a point
p = (x, y) with x undefined and y = CostS (cm+1 ). We say that
pℓ is a P SP of edge ℓ. Define the set of clients covered by pℓ
to be the set of all clients c such that the maximum cost of points
on CCLℓ (c) is smaller than CostS (cm+1 ). We call the normal
(non-virtual) P SP s the actual P SP s.
8.2 PSPs Pruning
c1
c2
c3
Cost
Cost
c1
pl
c4
c2
p2
c3
c4
(a)
p2
c2
Cost
Cost
We identify 2 properties of P SP s to eliminate redundant P SP s
from the computed P SP s.
(b)
p1
c3
l2
(c)
8.2.1 Strategy 3: Pruning by PSP comparison
l1
l1
c1
The complexity in computing k-candidates depends on the size
of Pedge at Line 17 and the size of Pedge ∪ P at Line 27. In this
subsection, we introduce two strategies for reducing the number of
P SP s, which can greatly speed up the algorithm.
pl
p1
P ROPERTY 1. If there exist two P SP s that cover the same set
of clients, we can safely disregard the one with a higher cost when
computing k-candidates.
p2
c1
p3
c2
c3
l2
RATIONALE FOR PROPERTY 1: Consider the example in Figure 3(c). p1 and p2 are two P SP s on edge l2 , and both cover c1 ,
c2 and c3 . p1 .cost < p2 .cost. Suppose we get a k-candidate
Z that contains P SP p2 , we can replace p2 by p1 directly. If
p2 .cost = Z.maxcost, then we can replace p2 by p1 to convert Z to Z ′ , and Z ′ .maxcost will be p1 .cost, which is less than
p2 .cost. If p2 .cost 6= Z.maxcost, then Z ′ .maxcost will be equal
to Z.maxcost after replacing p2 by p1 .
(d)
Figure 3: Illustrations for the enhancement strategies. (a) virtual P SP pl for edge ll covers c2 , c3 . (b) p2 is pruned by Pedge .
(c) p1 and p2 both cover c1 , c2 , c3 ; p2 is pruned. (d) p3 is pruned
when compared with p2 .
In Figure 3(a), the dotted line shows pl which is the virtual P SP
of edge l1 . The cost of pl equals CostS (cm+1 ). pl covers c2 , c3 ,
and c4 . There exist some points on CCL(c1 ) with costs greater
than pl .cost, so pl does not cover c1 .
We build a P SP set Pedge which consists of all virtual P SP s
of eligible edges at Lines 7 and 16 in Algorithm 4 and select kcandidates in Pedge at Line 17. Each selection is a combination
of k virtual P SP s in Pedge . The cost of each P SP in Pedge is
CostS (cm+1 ).
P ROPERTY 2. Suppose p1 and p2 are two P SP s on the set of
eligible edges, the set of clients covered by p1 is C1 and that covered by p2 is C2 . In the iteration of cm , if p1 .cost < p2 .cost <
CostS (cm+1 ) and C1 ⊂ C2 , we can disregard p1 when computing
k-candidates.
RATIONALE FOR PROPERTY 2: Consider Figure 3(d). p2 and p3
are two P SP s on edge l2 . p2 covers C2 = {c1 , c2 , c3 }. p3 covers
8
C3 = {c1 , c2 }. Suppose the dashed line shows CostS (cm+1 ), the
cost of the client in the next iteration. C3 ⊂ C2 and p3 .cost <
p2 .cost < CostS (cm+1 ). Consider a k-candidate Z that contains
P SP p3 , we can replace p3 by p2 , and get Z ′ = Z \ {p3 } ∪
{p2 }. Hence, clients covered by Z are also covered by Z ′ . If
Z.maxcost > CostS (cm+1 ), then Z ′ .maxcost = Z.maxcost.
Hence, if Z is an optimal solution, Z ′ is also an optimal solution. If
Z.maxcost ≤ CostS (cm+1 ), then Z ′ .maxcost is still less than or
equal to CostS (cm+1 ) since p3 .cost < p2 .cost < CostS (cm+1 ).
Thus, neither Z nor Z ′ will be returned as kSP at Lines 34 and
35. Therefore, we can disregard p3 when computing k-candidates.
Note that while p3 is disregarded in this iteration, it can be used for
computing k-candidate in a future iteration for some ci , i > m, if
p3 is updated at Line 23 to cover ci while p2 cannot cover ci .
Note that although the example above is based on only one edge,
in general we apply these 2 properties to P SP s in Pedge at Line
17 and Pedge ∪ P at Line 27 on all eligible edges. That is, the two
P SP s involved can be on two different edges.
Algorithm 4: QuickM inM ax(G, C, S, k)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
8.2.2
Strategy 4: Pedge based pruning
When we cannot jump to the next iteration by Strategy 1 or 2,
we will compute P SP s generated by CCL(c) for c ∈ N at Line
24. With the set Pedge , we can reduce the number of P SP s to
be computed in the following way. For each edge ℓ, a CCLℓ (c)
is said to be active if CCLℓ (c) contains a point p with p.cost >
CostS (cm+1 ) in the iteration of ci . Otherwise, CCLℓ (c) is inactive. While computing P SP s at Line 24, we check whether a
P SP p is generated by intersecting inactive CCLℓ (c)s, and if so,
p is dismissed. Clients covered by such a P SP must be also covered by pℓ , the virtual P SP of edge ℓ. Since pℓ will be considered
at Line 27, based on a similar rationale as that for Property 2, p can
be dismissed.
Consider Figure 3(b). pl is the virtual P SP on the edge l1 , and
it covers Cl1 = {c2 , c3 , c4 }. CCLl1 (c1 ) is active. CCLl1 (c2 )
and CCLl1 (c3 ) are inactive. p1 is an intersection of CCLl1 (c1 )
and CCLl1 (c3 ), which covers C1 = {c1 , c2 , c3 , c4 }. p2 is an
intersection of CCLl1 (c2 ) and CCLl1 (c3 ), which covers C2 =
{c2 , c3 , c4 }. C1 6⊆ Cl1 and C2 ⊆ Cl1 . We can thus dismiss p2 .
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
Compute set Q of P SP s and clients covered by each P SP on E 0 ;
P ← P ∪ Q;
For each ℓ in E 0 , compute clients covered by pℓ and Pedge ← {pℓ };
for m = k + 1, ..., n do
curmax ← CostS (cm ) ;
N ← N ∪ {cm };
build CCL(cm ) on edges in E 0 ;
remove P SP s in P with cost ≥ curmax;
remove CCLℓ (c) where c ∈ C on each edge ℓ in E 0 with a
minimum cost ≥ curmax;
if kSP.maxcost ≤ CostS (cm+1 ) and kSP covers cm then
continue
⊲ /* Strategy 1 */
update pℓ .cost and clients covered by pℓ for each pℓ in Pedge ;
while ∃ updated k-candidate Z not explored in Pedge do
if Z covers c1 , ..., cm then
curmax ← CostS (cm+1 ); kSP ← Z;
break;
if curmax = CostS (cm+1 ) then
continue
⊲ /* Strategy 2 */
update clients covered by P SP s in P (to include clients in N );
Q ← set of P SP s generated by CCL(c), c ∈ N ;
N ← ∅;
P ← P ∪ Q;
while ∃ new or updated k-candidate Z not explored in Pedge ∪ P
do
if Z covers c1 , ..., cm then
if Z.maxcost < curmax then
kSP ← Z;
curmax ← Z.maxcost;
if curmax ≤ CostS (cm+1 ) then
break;
if curmax > CostS (cm+1 ) then
Return curmax, kSP
a weight, generated randomly from a Zipf distribution with a skewness parameter of α > 1. The default setting is as follows: For SF:
|S|= 200, |C| = 100,000. For NYC: |S| = 500, |C| = 300,000. For
FLA: |S| = 1000, |C| = 600,000. α = ∞, meaning a unit weight
for each client. The default setting for |E 0 |/|E| is 10%.
We measure the quality and runtime by varying different parameters, including the number of new servers k, the number of clients
|C|, number of existing servers |S|, and the Zipf factor α in the
skewness of client weights.
8.3 Algorithm 4
We call the enhanced MinMax algorithm QuickMinMax. It is
shown in Algorithm 4. Strategy 1 is carried out at Lines 14-15.
Since |Pedge | is much less than |P |, we first select k-candidates
from Pedge at Line 17, and execute Strategy 2 at Lines 17-22. If
either Strategy 1 or 2 succeeds, note that cm is added to N at Line
10, and we jump to the next iteration to process ci+1 . Otherwise
we cannot terminate early, so N is processed at Lines 23 and 24
and reset to empty at Line 25. The remaining processing is similar
to Algorithm 3, except that Strategy 3 and Strategy 4 are activated
at Lines 17 and 27, and at Line 24, respectively.
9.
Input
: G,C,S,k, eligible edges E 0 , sorted c1 , ..., cn
Output : minmax cost: minmax(G, C, S, k), kSP
begin
P ← ∅; Pedge ← ∅; N ← ∅;
for m = 1, ..., k do
build CCL(cm ) on each edge in E 0 ;
9.1 Comparison of the Algorithms
We analyze the effects of different parameters on the MinMax
algorithms. We compare results of our approximation algorithm
(Approx) (Algorithm 1) and optimal algorithm (Opt) (Algorithm
4) with the results from the best-known greedy algorithms (Greedy)
in [5]. Note that the results of [5] and [23] are identical since they
compute the same optimal solution for a single new server, and the
corresponding approximation algorithms repeatedly select a single
new server until k new server locations are chosen. However, the
algorithm in [5] improves on the computation time compared to
[23], hence we only report the runtime of the greedy algorithms in
[5].
Let the minmax cost prior to adding any new server be M ax =
max(CostS (c)|c ∈ C), P be an optimal or approximate MinMax
solution with k locations. Let kM ax = max(CostS∪P (c)|c ∈
EMPIRICAL STUDIES
In this section, we evaluate the performance of our proposed algorithms. We run all experiments on a machine with a 3.4Ghz Intel
Core i7-4770 CPU and 16 GB RAM, running Ubuntu 12.04 LTS
Linux OS. All algorithms were implemented in C++ and compiled
with GNU c++ compiler.
We use three real world road network datasets: SF, NYC and
FLA, for San Francisco, New York City, and Florida, respectively.
SF contains 174,955 vertices and 223,000 edges. NYC contains
264,346 vertices and 733,846 edges and FLA contains 1,070,376
vertices and 2,712,798 edges. In our experiment, clients and servers
are generated randomly on all edges. Each client is associated with
9
k
C). Define Gain = M ax − kM ax, and Gain Ratio GR =
Gain/M ax. To quantify the advantage of the optimal solution
over the approximation algorithm, we measure the relative error of
Approx as Err = (O − A)/O, where A(O) is the Gain from
Approx (Opt).
k
Gain (Opt)
Gain (Approx)
Gain (Greedy)
GR (Opt) (%)
GR (Approx) (%)
GR (Greedy) (%)
Relative Error (Approx)
Relative Error (Greedy)
Time (Opt)(s)
Time (Approx)(s)
Time (Greedy)(s)
1
2
4
6
8
10
37.3
37.3
37.3
1.1
1.1
1.1
0.00
0.00
0.8
0.3
0.6
89.0
89.0
66.3
2.6
2.6
1.9
0.00
0.25
1.9
0.3
0.7
1277.8
1041.3
89.0
37.4
30.5
2.6
0.18
0.93
3.1
0.4
0.7
1722.4
1277.8
124.0
50.4
37.4
3.6
0.26
0.92
28.4
0.5
0.8
1800.5
1499.2
192.1
52.7
43.9
5.6
0.16
0.89
132.9
0.5
0.9
1986.5
1546.8
221.0
58.1
45.3
6.5
0.22
0.89
317.0
0.6
0.9
Gain (Opt)
Gain (Approx)
Gain (Greedy)
GR (Opt) (%)
GR (Approx) (%)
GR (Greedy) (%)
Relative Error (Approx)
Relative Error (Greedy)
Time (Opt)(s)
Time (Approx)(s)
Time (Greedy)(s)
1
2
4
6
8
10
5.9
5.9
5.9
0.4
0.4
0.4
0.00
0.00
1.4
0.5
1.2
40.1
40.1
6.0
2.8
2.8
0.4
0.00
0.85
1.6
0.6
1.3
77.1
75.7
28.8
5.2
5.3
2.0
0.02
0.63
2.04
0.9
1.5
217.2
106.9
41.0
15.0
7.4
2.8
0.50
0.81
22.2
1.1
1.6
313.3
217.2
59.1
21.6
15.0
4.1
0.30
0.81
301.9
1.3
1.8
324.1
281.8
68.8
22.4
19.4
4.7
0.13
0.78
401.8
1.5
2.0
Gain (Opt)
Gain (Approx)
Gain (Greedy)
GR (Opt) (%)
GR (Approx) (%)
GR (Greedy) (%)
Relative Error (Approx)
Relative Error (Greedy)
Time (Opt)(s)
Time (Approx)(s)
Time (Greedy)(s)
1
2
4
6
8
10
45.5
45.5
45.5
6.1
6.1
6.1
0.00
0.00
5.4
1.9
4.6
241.3
82.4
54.4
32.1
11.0
7.2
0.66
0.77
28.9
2.2
4.7
249.8
248.6
56.7
33.3
33.1
7.6
0.01
0.78
32.6
2.6
5.0
275.1
272.5
61.1
36.6
36.3
8.1
0.01
0.78
42.2
3.1
5.4
286.7
279.6
62.4
38.2
37.2
8.3
0.02
0.78
70.3
3.5
5.7
296.8
285.2
64.3
39.5
38.0
8.5
0.04
0.78
190.9
3.9
6.2
4
6
8
10
We analyze the effect of k and the Zipf parameter α on the Gain
in Table 5. The optimal solution consistently generated better solutions compared to the approximation method. Since the greedy
algorithm always takes more time and returns worse result compared to the approximation algorithm, we only show the Gain of
Opt and Approx.
α
Table 2: Gain, Gain Ratio (GR), Relative Error, and runtime
for MinMax with unit weights on NYC
k
2
974.0 3786.0 5845.9 8841.8 9272.9 10268.4
974.0 3786.0 5845.9 7454.2 8222.8 8701.9
974.0 1691.0 3786.0 5344.5 5845.9 6094.3
6.0
23.5
36.2
54.8
57.5
63.7
6.0
23.5
36.2
46.2
51.0
54.0
6.0
10.5
23.5
33.1
36.2
37.8
0.00
0.00
0.00
0.16
0.11
0.15
0.00
0.55
0.35
0.40
0.37
0.41
0.8
0.8
0.9
1.8
2.3
4.0
0.3
0.3
0.4
0.5
0.5
0.6
0.7
0.7
0.8
0.8
0.9
0.9
Table 4: Gain, Gain Ratio (GR), Relative Error, and runtime
for MinMax with skewed weights (α = 2) on SF
Table 1: Gain, Gain Ratio (GR), Relative Error, and runtime
for MinMax (Opt:Optimal solution; Approx: approximate solution; Greedy: greedy solution) with unit weights on SF
k
1
Gain (Opt)
Gain (Approx)
Gain (Greedy)
GR (Opt) (%)
GR (Approx) (%)
GR (Greedy) (%)
Relative Error(Approx)
Relative Error(Greedy)
Time (Opt)(s)
Time (Approx)(s)
Time (Greedy)(s)
Max
Opt Gain (SF)
k=2
k=4
k=8
Approx Gain (SF)
k=2
k=4 k=8
3786.0
5647.4
2133.8
89.0
3786.0
3449.1
1246.6
89.0
2
4
6
∞
16129
12124
6451
3417
5845.9
6578.1
2754.3
1277.8
9272.9
7597.8
4018.4
1800.5
α
Max
2
4
6
∞
7243
5133
3080
1449
α
Max
Opt Gain (FLA)
k=2
k=4
k=8
2
4
6
∞
3448
2413
1426
751
1141.2
252.3
128.8
241.2
Opt Gain (NYC)
k=2
k=4
k=8
688.6
798.9
209.6
40.1
843.1
1052.7
506.6
77.1
1287.6
773.3
318.0
249.8
2088.8
1270.9
762.5
313.4
1375.3
1001.4
457.7
286.7
5845.9
5647.4
2133.8
1041.3
8222.8
7187.0
3034.6
1499.2
Approx Gain (NYC)
k=2 k=4
k=8
688.6
349.4
209.6
40.1
789.2
820.1
458.3
75.7
1717.0
1224.3
734.6
217.2
Approx Gain (FLA)
k=2
k=4
k=8
616.8
252.2
128.8
82.4
1178.0
773.3
318.0
248.6
1363.9
968.9
450.6
279.6
Table 5: Gain values for MinMax on SF, NYC and FLA
9.2 Effects of Parameters on Opt
We analyze the effect of k on the runtime of the optimal algorithm in Figure 4. There are two different trends in the results. If
each client has a unit weight, the runtime increases exponentially
with k. When α 6= ∞, meaning that the weights of the clients are
not uniform, the runtime increases smoothly and it is easy to find
the new server locations. This is because the new servers should be
near to some clients with heavy weights.
We measure the effects of different parameters on the runtime of
the optimal algorithm. We study the effect of |C| in Figure 5(a).
The runtime increases with |C|. In the experiment, the number of
servers and their positions are fixed. The optimal algorithm result
and Gain do not change much as |C| increases. The number of
clients processed increases when |C| increases.
The effect of |S| is shown in Figure 5(b). The runtime increases
with |S| when k is small and decreases when k is large. The runtime depends on two factors: the time for building N LCs(CCLs)
and computing P SP s, and the time for processing P SP s. The first
factor dominates when k is small. When |S| is small, the cost difference between two consecutive clients in the sorted list is large,
so we can find the result quickly and the second factor is not significant. When k is large, the second factor dominates. When |S|
is small, the N LCs of clients are large and there are more P SP s.
Since the sizes of N LCs and number of P SP s decrease as |S|
increases, the runtime decreases when k increases.
Table 3: Gain, Gain Ratio (GR), Relative Error, and runtime
for MinMax with unit weights on FLA
For comparison of the optimal solution and approximate solution
with the greedy solution, we show the Gain and Gain Ratios GR
in Tables 1, 2, 3 and 4. The results clearly show the advantages
of the optimal solution and approximate solution compared to the
greedy algorithm. For example, to achieve the same level of gain
as 2 new servers in the optimal solution, 4, 6 and over 10 servers
will be needed from the greedy solution in SF, NYC, and FLA, respectively. With our target applications, it would incur a large and
unnecessary cost for the user to build more than 10 servers instead
of 2, thus the longer running time needed for the optimal solution
is well justified for the saving in this expense. The Gain and GR
of Approx are always better than Greedy and the relative error is
small. Approx has a poor gain with FLA for k = 2 because in FLA
the clients are more dispersed and it is harder to achieve near optimal result when the locations are limited to the client sites. Note
that for clients with skewed weights, Opt takes about the same time
as Greedy and Approx as shown in Table 4.
10
In Figure 5(c), we study the effect of |E 0 |. We vary |E 0 |/|E|.
The time increases steadily with the increase in |E 0 |. The pruning of N LC computation and P SP s is highly effective. In all our
experiments, the number of N LCs that are computed is a few hundred at most, and the number of P SP s that are not pruned is not
more than twenty. The increase in runtime is gentle due to the effective pruning. In Figure 5(d), we study the effect of Zipf parameter
α. Skewness in weights is beneficial to the runtime. Intuitively,
when there exist some heavy weight clients, it becomes easier to
select the new server locations, since they should be near to such
clients. This sharpens the pruning effects, and the running time of
Opt is similar to that of Approx when α 6= ∞ (see Table 4).
A similar set of results has been obtained for NYC. For FLA,
the results are shown in Figure 6. The trends are mostly similar
to SF, except that in Figure 6(b), the time decreases with |S| for
all values of k. This is because FLA is a state where locations are
more spread out compared to a city. Even with k = 2, many clients
are covered by the new servers, so the time for handling the P SP s
and k-candidates dominates, and this decreases as the N LC sizes
decrease with increasing |S|.
In summary, Opt can handle problems with small k values efficiently. Approx takes less time than Greedy and returns a solution with a much better quality. Thus, we improve on the status
quo where k = 1 is solved with an optimal solution and the cases
of k > 1 are solved by Greedy.
4
Total iterations
Strategy 1
Strategy 2
Strategies 1+2(%)
4
6
8
10
43
5
37
97.68
93
25
67
98.92
528
373
153
99.62
700
478
218
99.43
777
512
263
99.74
k (FLA)
Total iterations
Strategy 1
Strategy 2
Strategies 1+2(%)
2
4
6
8
10
242
179
62
99.59
278
214
63
99.64
349
282
66
99.71
409
331
76
99.51
444
347
91
98.65
103
time (seconds)
time (seconds)
1
10
10
2
4
6
k
(a)
8
10
100
4
6
k
time (seconds)
time (seconds)
10
10
0
10
-1
5
10
15
20
25
4
10
30
200
400
600
(b)
k=2
103
k=4
k=2
k=4
k=8
time (seconds)
time (seconds)
800 1000
|S|
102
102
101
101
100
100
10
15
0
|E |/|E| (%)
20
2
3
4
5
6
7
(c)
(d)
Figure 5: Effects of |C|, |S|, |E 0 |, and α for SF
103
100
k=2
k=8
2
10
101
2
4
6
5
|C| (10 )
8
k=2
k=4
80
time (seconds)
time (seconds)
k=4
k=8
60
40
20
10
1
2
3
(a)
4
5
(b)
102
k=4
k=2
k=4
k=8
k=8
time (seconds)
time (seconds)
3
|S| (10 )
k=2
160
120
80
40
101
5
10
15
0
|E |/|E| (%)
20
2
3
4
5
6
7
(c)
(d)
Figure 6: Effects of |C|, |S|, |E 0 |, and α for FLA
k (SF)
Total
Strategy 4
Strategies 4+3
PSPs pruned (%)
2
4
6
8
10
1247
37
3
99.75
3712
108
5
99.86
58458
543
8
99.98
78530
1512
17
99.98
81645
1597
20
99.98
k (FLA)
Total
Strategy 4
Strategies 4+3
PSPs pruned (%)
2
4
6
8
10
46580
228
3
99.99
77079
488
5
99.99
144833
644
7
99.99
229004
1710
9
99.99
229414
2660
10
99.99
Table 7: Number of PSPs after pruning by Strategies 3 and 4
k (SF)
2
4
6
8
10
SF
2
4
27
12316
141008
FLA
2
4
6
9
44
Table 8: Number of k-candidates computed
compute the new and updated k-candidates. In addition, P SP s
are pruned by different strategies, thus the number of k-candidates
computed is much smaller than the bound of O(|X|k ). These results show that our enhancement strategies lead to significant improvements in the overall performance.
unit weight
2
0
5
101
0
1
10
k=8
1
k=8
a=6
102
unit weight
10
(a)
a=4
a=6
102
2
103
a=2
a=4
2
10
|C| (10 )
Table 7 shows the total number of PSPs computed in the road
network, number of PSPs left after Strategy 4 and number of
PSPs left after both Strategies 4 and 3 in the last iteration. With
Strategy 4 and Strategy 3, we reduce the P SP number by 99.90%
on average. Only very few PSPs are used for computing k-candidates.
Table 8 shows the number of k-candidates computed in the optimal algorithm. Given a set X of P SP s in each iteration, we only
a=2
k=4
k=8
-1
Table 6: Number of iterations terminated early
103
k=2
k=4
10
10
We measure the effects of the enhancement strategies of optimal
algorithm QuickMinMax in Section 8. We show the number of iterations terminated early by Strategy 1 and Strategy 2 in Section
8.1 in Table 6. On average, we terminate early and jump to the next
iteration with Strategies 1 and 2 in over 99.25% of the iterations.
2
10
k=2
3
9.3 Effects of the Enhancement Strategies
k (SF)
3
10
8
10
10. CONCLUSION
(b)
We consider the problem of MinMax for finding multiple optimal minmax locations on a road network. We propose a novel
algorithm based on the concepts of client cost lines and potential
Figure 4: Effects of k on the runtime for (a) SF and (b) FLA for
MinMax (a stands for α in the legend.)
11
APPENDIX
server points, with search space pruning and early stopping strategies. Our empirical study shows that our proposed algorithm generates significantly better solutions compared to previous works on
three real road networks. Other OLQ problems with multiple new
servers will be interesting problems for further studies.
PROOF OF LEMMA 3: zmax(C ′ , p) = max(p.cost, CostS (C ′′ )),
where C ′′ is the set of clients in C ′ not covered by p, and CostS (C ′′ ) =
max(CostS (c)|c ∈ C ′′ ). When C ′ is the set of clients covered
by p, C ′′ = ∅. Hence, zmax(C ′ , p) = p.cost. Also, since
zmax(C ′ , p) ≥ zmax(C ′ , q) ≥ q.cost, thus, q.cost ≤ p.cost.
Since p ∈ A, and amax(A) is lower bounded by the costs of the
apex points in A, amax(A) ≥ p.cost. For a client c′ covered by
q, we have CostS∪A′ (c′ ) ≤ q.cost ≤ p.cost ≤ amax(A). For
a client c′ not covered by q, it is either covered by p or not covered by p. If c′ is not covered by p, since A′ − {q} = A − {p},
then CostS∪A′ (c′ ) = CostS∪A (c′ ). If c′ is covered by p, then
c′ ∈ C ′ . Given that zmax(C ′ , p) ≥ zmax(C ′ , q), zmax(C ′ , q) ≤
p.cost. Since c′ is not covered by q, it must hold that CostS (c′ ) >
q.cost. We deduce that zmax(C ′ , q) ≥ max(q.cost, CostS (c′ ))
= CostS (c′ ). Thus, CostS (c′ ) ≤ zmax(C ′ , q) ≤ p.cost. If c′
is covered by some point in S ∪ (A ∩ A′ ), then CostS∪A′ (c′ ) =
CostS∪A (c′ ), otherwise, c′ is not covered by any point in S ∪ A′ ,
and c′ is covered only by p in S ∪ A. We have CostS∪A′ (c′ ) =
CostS∪A (c′ ) = CostS (c′ ). Note that amax(A) ≥ CostS∪A (c)
for all c ∈ C. Thus, considering all cases for any given client
c′ , CostS∪A′ (c′ ) ≤ amax(A). We conclude that amax(A′ ) ≤
amax(A) and the lemma holds.
11. REFERENCES
[1] S. Cabello, J. M. Dı́az-Báñez, S. Langerman, C. Seara, and
I. Ventura. Reverse facility location problems. In CCCG, 2005.
[2] H. Calik and B. C. Tansel. Double bound method for solving the
p-center location problem. Computers and Operations Research,
40:2991–2999, 2013.
[3] J. Cardinal and S. Langerman. Min-max-min geometric facility
location problems. In Proc. European Workshop on Computational
Geometry), pages 149–152, 2006.
[4] M. Charikar and S. Guha. Improved combinatorial algorithms for
facility location and k-median problems. In Proc. IEEE FOCS, 1999.
[5] Z. Chen, Y. Liu, R. C.-W. Wong, J. Xiong, G. Mai, and C. Long.
Efficient algorithms for optimal location queries in road networks. In
SIGMOD, 2014.
[6] Z. Chen, Y. Liu, R. C.-W. Wong, J. Xiong, G. Mai, and C. Long.
Optimal location queries in road networks. In TODS, 2015.
[7] D.-W. Choi, C.-W. Chung, and Y. Tao. A scalable algorithm for
maximizing range sum in spatial databases. PVLDB,
5(11):1088–1099, 2012.
[8] Z. Drezner. Facility location: a survey of applications and methods.
Springer Verlag, 1995.
[9] Y. Du, D. Zhang, and T. Xia. The optimal-location query. In
Advances in Spatial and Temporal Databases, pages 163–180.
Springer, 2005.
[10] R. Farahani and M. Hekmatfar. facility location: Concepts, Models,
Algorithms and Case Studies. Physics-Verlag HD, 2009.
[11] T. Feder and D. Greene. Optimal algorithms for approximate
clustering. In STOC, pages 434–444, 1988.
[12] J. Flum and M. Grohe. Parameterized Complexity Theory. Springer
Verlag, 2006.
[13] D. Hochbaum and D. Shmoys. A best possible heuristic for the
k-center problem. Mathematics of Operations Research,
10(2):180–184, 1985.
[14] R. Hwang, R. Lee, and R. Chang. The slab dividing approach to
solve the euclidean p-center problem. Algorithmica, 9(1):1–22, 1993.
[15] J. Krarup and M. Pruzan. The simple plant location problem: Survey
and synthesis. European Journal of Operational Research,
12(1):36–57, 1983.
[16] P. Kumar and P. Kumar. Almost optimal solutions to k-clustering
problems. International Journal of Computational Geometry and
Applications, 20(4), 2010.
[17] Y. Liu, R. C.-W. Wong, K. Wang, Z. Li, C. Chen, and Z. Chen. A
new approach for maximizing bichromatic reverse nearest neighbor
search. Knowledge and information systems, 36(1):23–58, 2013.
[18] A. Meyerson. Online facility location. In Proc. IEEE FOCS, 2001.
[19] S. Nickel and J. Puerto. location theory: a unified approach.
Springer, 2005.
[20] J. Qi, R. Zhang, L. Kulik, D. Lin, and Y. Xue. The min-dist location
selection query. In ICDE, 2012.
[21] J. Qi, R. Zhang, Y. Xue, A. Y. Xue, and L. Kulik. The min-dist
location selection and facility replacement queries. World Wide Web,
17(6), 2014.
[22] R. C.-W. Wong, M. T. Özsu, P. S. Yu, A. W.-C. Fu, and L. Liu.
Efficient method for maximizing bichromatic reverse nearest
neighbor. Proceedings of the VLDB, 2(1):1126–1137, 2009.
[23] X. Xiao, B. Yao, and F. Li. Optimal location queries in road network
databases. In ICDE, pages 804–815. IEEE, 2011.
[24] Z. Xu. Processing proximity relations in road networks. In SIGMOD,
2010.
[25] D. Zhang, Y. Du, T. Xia, and Y. Tao. Progressive computation of the
min-dist optimal-location query. In VLDB, pages 643–654, 2006.
q’
p(x,y)
q(x’,y’)
p(x,y)
p(x,y)
q’
q’
q(x’,y’)
(x,z’)
CASE 1
q(x’,y’)
CASE 2
CASE 3
Figure 7: 3 cases in the proof of Lemma 4
PROOF OF LEMMA 4: Let ℓ = [a, b]. Assume I(ℓ) 6= ∅. If a point
p = (x, y) on CCLℓ (c) is not in I(ℓ), and p 6∈ B(ℓ), then there
must be a point q = (x′ , y ′ ) with q.cost < p.cost (i.e., y ′ < y)
on CCLℓ (c) closest to p which is either a lowest boundary point
or q ∈ I(ℓ). We show that replacing p by q does not increase
the overall minmax cost. Clearly, we only need to consider clients
covered by p. There are different cases for such a covered client c′ .
We show in each case that zmax({c′ }, p) ≥ zmax({c′ }, q).
[CASE 1]: Suppose p covers a client c′ where CCLℓ (c′ ) contains
no end point. We show that q also covers c′ . If this is true, then
since q.cost < p.cost, zmax({c′ }, p) ≥ zmax({c′ }, q). Assume
on the contrary that p covers such a client c′ and q does not. Recall that p = (x, y), q = (x′ , y ′ ). Without loss of generality, let
x′ < x. Since q cannot cover c′ , there exists a point q ′ = (x′ , z)
on CCLℓ (c′ ) with z > y ′ . This is illustrated in Figure 7 as CASE
1. Since p covers c′ , we have p′ = (x, z ′ ) ∈ CCLℓ (c′ ), where
z ′ < y. However, given the points p, q, p′ , q ′ , there must exist an
intersecting point (x̂, ŷ) of CCLℓ (c) and CCLℓ (c′ ), such that either x′ < x̂ < x, or x < x̂ < x′ , which contradicts the assumption
that q is the closest point to p in I. Thus, q also covers c′ .
[CASE 2]: Next, suppose an end point q ′ of CCLℓ (c′ ) is in (a, b)
in ℓ and CostS (c′ ) ≤ y ′ . I.e. q ′ is not a boundary of ℓ. Since
y ′ < y, clearly, zmax({c′ }, p) ≥ zmax({c′ }, q). See Figure 7
CASE 2 for a possible scenario.
[CASE 3]: Finally, consider the case where an end point q ′ of
CCLℓ (c′ ) is in (a, b) in ℓ and CostS (c′ ) > y ′ . Since p covers c′ , CostS (c′ ) ≤ y. Thus, zmax({c′ }, q) = CostS (c′ ) ≤
zmax({c′ }, p). See Figure 7 CASE 3 for an illustration.
From all the above cases, we conclude that for the set C ′ of
clients covered by p, zmax(C ′ , p) ≥ zmax(C ′ , q). The claim
thus follows from Lemma 3.
12
| 8cs.DS
|
arXiv:1501.05971v3 [cs.DS] 13 Dec 2015
Cooperative Greedy Pursuit Strategies for Sparse Signal
Representation by Partitioning
Laura Rebollo-Neira
Mathematics Department
Aston University
B3 7ET, Birmingham, UK
March 1, 2018
Abstract
Cooperative Greedy Pursuit Strategies are considered for approximating a signal partition
subjected to a global constraint on sparsity. The approach aims at producing a high quality
sparse approximation of the whole signal, using highly coherent redundant dictionaries. The cooperation takes place by ranking the partition units for their sequential stepwise approximation,
and is realized by means of i)forward steps for the upgrading of an approximation and/or ii)
backward steps for the corresponding downgrading. The advantage of the strategy is illustrated
by approximation of music signals using redundant trigonometric dictionaries. In addition to
rendering stunning improvements in sparsity with respect to the concomitant trigonometric
basis, these dictionaries enable a fast implementation of the approach via the Fast Fourier
Transform.
Keywords: Hierarchized Block Wise Pursuit Strategies, Optimized Orthogonal Matching Pursuit,
Sparse Representation of Music Signals.
1
Introduction
Compressible signals, such as audio and vision data, are characterized by samples containing a good
deal of redundancy. Consequently, if properly processed, the signal information content can be
accessed from a reduced data set. Transformations for data reduction are said to produce a sparse
representation of a signal if they can accurately reproduce the information the signal conveys, with
significantly less points that those by which the original signal is given. Most popular transformations
for signal processing (e.g. Discrete Cosine Transform and Discrete Wavelet Transform) do not modify
the signal size. A sparse representation is achieved, a posteriori, by disregarding the least relevant
1
points in the transformed domain. Nonetheless, much higher sparsity in a signal representation
may be achieved by allowing for the expansion of the transformed domain, thereby leaving room
for dedicated transformations adapted to the particular signal. Such a framework involves a large
redundant set called ‘dictionary’. The aim is to represent a signal as a superposition of, say K,
dictionary’s elements, which are called ‘atoms’, with K much less than the signal size.
Given a redundant dictionary, the problem of finding the sparsest approximation of a signal, up
to some predetermined error, is an NP-hard problem [1]. Consequently, the interest in practical
applications lies in the finding of those solutions which, being sparse in relation to other possible
representations, are also easy and fast to construct. Sparse practical solutions can be found by
what are known as Pursuit Strategies. Here the discrimination between two broad categories is
in order: i)The Basis Pursuit based approaches, which endeavor to obtain a sparse solution by
minimization of the 1-norm [2]. ii)Greedy algorithms which look for a sparse solution by stepwise
selection of dictionary’s atoms. Practical greedy algorithms, which originated as regression techniques
in statistics [3], have been popularized in signal processing applications as Matching Pursuit (MP) [4]
and Orthogonal Matching Pursuit (OMP) [5] methods. The approach, which in principle consider
the stepwise selection of single atoms, has been extended to multiple atom selection [6]. Dedicated
algorithms such as Stagewise Orthogonal Matching Pursuit [7], Compressive Sampling Matching
Pursuit [8], and Regularized Orthogonal Matching Pursuit [9] are known to be effective within the
context of the emerging theory of sampling, called compressive sensing/sampling. This theory asserts
that sparsity of a representation may also lead to more economical data collection [10–14]. In that
context the reconstruction problem is of a very particular nature, though: It is assumed that a signal
is sparse in an orthogonal basis and the goal is to reconstruct the signal from a reduced number
of measures. On the contrary, in this Communication we address the traditional representation
matter: the signal is assumed to be completely given by its samples. The aim is to produce a high
quality approximation of all those samples, as a K-term superposition of atoms belonging to a highly
2
coherent dictionary. In this case, minimization of the 1-norm is not effective and step wise greedy
selection of single atoms benefits sparsity results.
In practice, when trying to approximate real life signals using a redundant dictionary, there is
a need to approximate by partitioning. While this requirement normally comes from storage and
computational complexity demands, it does not represent a disadvantage. On the contrary, to capture
local properties of signals such as images, or music, non-local orthogonal transforms are also applied
on partitions to obtain superior results. The broadly used image compression standard JPEG and
the music compression standard MP3, for instance, operate by partitioning as a first step in the
compression process.
The central aim of this paper is to tailor pursuit algorithms for approximation by partitioning
without significant increment in computational complexity, in comparison with standard applications.
The examples presented here clearly show that a global constraint on sparsity, rather than quality
constraints on the individual partition’s units, may benefit enormously the quality of the signal
approximation. This is true even if the actual approximation of each unit is performed individually.
Because considerations are restricted to producing high quality approximations, there is no risk that
blocking artifacts could appear due to the signal division into non-overlapping pieces.
Notational convention
Throughout the paper R, C and N indicate the sets of real, complex, and natural numbers, respectively. Boldface letters are used to indicate Euclidean vectors, whilst standard mathematical fonts
indicate components, e.g., f ∈ RN , N ∈ N is a vector of components f (i), i = 1, . . . , N. A partition
of a signal f ∈ RN is represented as a set of disjoint pieces f{q} ∈ RNb , q = 1, . . . , Q, which for simplicity are assumed to be all of the same size and such that QNb = N. The signal is reconstructed
Q
from the partition through the operation f = Ĵq=1 f{q} where Ĵ represents the operator which concatenates the partition. The operation is defined as follows: given f{1} ∈ RNb and f{2} ∈ RNb ,
the vector f = f{1} Ĵ f{2} is a vector in R2Nb having components f (i) = f {1}(i) for i = 1, . . . , Nb ,
3
Q
and f (i) = f {2}(i − Nb ) for i = Nb + 1, . . . , 2Nb . Thus f = Ĵq=1 f{q} is a vector in RQNb having
components f (i) = f {q}(i − (q − 1)Nb ), i = (q − 1)Nb + 1, . . . , qNb , q = 1, . . . , Q. Consequently
hf, fi = kfk2 =
PQ
q=1
kf{q}k2 , were h·, ·i indicates the Euclidean inner product and k · k the induced
2-norm.
Paper contributions
Given a signal partition f{q} ∈ RNb , q = 1, . . . , Q and a dictionary D = dn ∈ RNb ; kdn k = 1
M
n=1
to
approximate the elements f{q} in the partition, the following outcome has been recently reported [15]:
A very significant gain in the sparsity of a signal approximation may be effectively obtained by a
greedy pursuit strategy, if the approximation of each piece f{q} (called ‘q-block’) is accomplished in a
hierarchized manner. Suppose that the k(q)-term approximation of f{q} is the atomic decomposition:
a
f {q} =
k(q)
X
c{q}n dℓ{q}n ,
q = 1, . . . , Q,
n=1
with the atoms dℓ{q}n , n = 1, . . . , k(q) selected from the dictionary D, via a stepwise greedy pursuit
strategy. Suppose also that the number of atoms to approximate the whole signal f is a fixed value
K, i.e., K =
PQ
q=1 k(q).
A possibility to handle this constraint is to consider a hierarchized selection
of the pieces f{q} to be approximated in each approximation step. Some remarkable results of this
strategy, which has been termed Hierarchized Block Wise (HBW) greedy strategy, are illustrated
in [15] by approximating images using the greedy algorithms MP and OMP. When these methods
are applied in the proposed HBW fashion are called HBW-MP and HBW-OMP, respectively.
While [15] focusses on highlighting the suitability of the HBW-OMP/MP method when the image
approximation is carried out in the wavelet domain, this Communication extends the method as well
as the range of applicability. The extended techniques are shown to produce hugely sparse high
quality approximation of music signals. This is accomplished with trigonometric dictionaries, which
are endowed with the additional advantage of enhancing the competitiveness of the approach in terms
of computational complexity.
4
The extension of the idea outlined in [15] comprises:
• A revised version of the approach, which optimizes the ranking of the pieces f{q}, q = 1, . . . , Q
for their step wise approximation without significant extra computational cost. Additionally,
the selection of atoms is also optimized by including the Optimized Orthogonal Matching
Pursuit (OOMP) criterion [16].
• A HBW backward approach for downgrading the approximation when required. This is realized
in a stepwise optimized manner, by removing terms in the approximations of the selected blocks.
• An alternative to the HBW strategy which consists of two stages. The first stage involves the
approximation of the pieces f{q}, q = 1, . . . , Q, up to a tolerance error. The second stage refines
the previous approximation by making possible the downgrading of the atomic decomposition of
some blocks, called ‘donors’, and the upgrading of the atomic decompositions of another blocks,
called ‘receivers’. Since this process is inspired in the Swapping-based-Refinement of greedy
strategies introduced in [17], we refer to it as HBW-Swapping-based-Refinement (HBW-SbR).
Further contributions of the paper are
• The illustration of the huge gain in the sparsity of the representation of music signals obtainable
by the proposed approach when using trigonometric dictionaries.
• The finding that a mixed trigonometric dictionary having both, discrete cosine and sine components, renders the best sparsity performance for melodic music signals, in comparison with
the single discrete cosine (or sine) dictionaries with the same redundancy. This result contrasts
with the fact that, the broadly used single discrete cosine basis is the basis yielding best sparsity results for music signals, in comparison to other trigonometric basis of the same nature.
Note: The discrete cosine/sine components of the mixed dictionary are not just the real and
imaginary parts of the complex exponentials in a Fourier series. The involved phase factors
make the whole difference in achieving high sparsity.
5
• The provision of implementation details in the form of algorithms for the proposed strategies
and the provision of dedicated algorithms to operate with trigonometric dictionaries via the
Fast Fourier Transform (FFT).
A library of MATLAB functions for implementing the proposed methods, and running the numerical
examples in this paper, are available for downloading on [18].
2
Hierarchized Blockwise OMP, revised
As already stated, a signal f ∈ RN will be considered to be the composition of Q identical and
Q
disjoint blocks f = Ĵq=1 f{q},, where f{q} ∈ RNb assuming that Nb Q = N. Each piece f{q} ∈ RNb
is approximated using a dictionary D =
dn ∈ RNb ; kdn k = 1
M
,
n=1
with M > Nb , by an atomic
decomposition of the form:
a
f {q} =
k(q)
X
c{q}n dℓ{q}n ,
q = 1 . . . , Q.
(1)
n=1
where the atoms dℓ{q}n , n = 1, . . . , k(q) are selected from the dictionary to approximate the block
k(q)
q. Thus, each set of selected indices Γ{q} = {ℓ{q}n }n=1 is a subset of the set of labels {n}M
n=1 which
identify the atoms in D. In general Γ{q} =
6 Γ{p} and k(q) 6= k(p) if p 6= q.
The HBW-OMP approach outlined in [15] selects atoms in (1) as indicated by the OMP approach,
i.e.: On setting initially k(q) = 0 and r0 {q} = f{q}, the algorithm picks the atoms in the atomic
decomposition of block q by selecting one by one the indices which satisfy:
ℓ{q}k(q)+1 = arg max dn , r
k(q)
n=1,...,M
{q} , with r
k(q)
{q} = f{q} −
k(q)
X
c{q}n dℓ{q}n if k(q) > 0.
(2)
n=1
The coefficients c{q}n , n = 1, . . . , k(q) in (2) are such that the norm rk(q) {q} is minimized. This
is ensured by requesting that rk(q) {q} = f{q} − P̂Vqk(q) f{q}, where P̂Vqk(q) is the orthogonal projection
operator onto Vqk(q) = span dℓ{q}n
k(q)
.
n=1
This condition also guarantees that the set dℓ{q}n
linearly independent.
6
k(q)
n=1
is
An implementation as in [16] provides us with two representations of P̂Vqk(q) . One of the represen
tations is achieved by orthonormalization of the set dℓ{q}n
k(q)
.
n=1
The other by biorthogonalization
of the same set. We implement the orthogonalization by Gram Schmidt method, including a rek(q)
e
orthogonalization step, as follows: The orthonormal set {w{q}
n }n=1 is inductively constructed from
e
w{q}
1 = w{q}1 = dℓ{q}1 through the process:
w{q}n
e
,
w{q}
n =
kw{q}n k
with w{q}n = dℓ{q}n −
n−1
X
i=1
e
e
w{q}
i w{q}
i , dℓ{q}n .
(3)
For numerical accuracy at least one re-orthogonalization step is usually needed. This implies to
recalculate the above vectors as
w{q}n ← w{q}n −
n−1
X
i=1
e
e
w{q}
i hw{q}
i , w{q}n i ,
(4)
k(q)
e
with the corresponding normalization giving rise to the orthonormal set {w{q}
n }n=1 . Notice that,
whilst this set can be used to calculate the orthogonal projection of f{q} onto Vqk(q) as
k(q)
X
P̂Vqk(q) f{q} =
n=1
e
e
w{q}
n hw{q}
n , f{q}i ,
(5)
this superposition is not the atomic decomposition in terms of the selected atoms. In order to
produce such a decomposition we use the other representation for an orthogonal projector given
in [16]. Namely,
P̂
Vqk(q)
f{q} =
k(q)
X
n=1
dℓ{q}n hb{q}n , f{q}i .
(6)
For a fixed q the vectors b{q}n , n = 1, . . . , k(q) in (6) are biorthogonal to the selected atoms,
i.e.
dℓ{q}n , b{q}m = 0 if n 6= m and 1 if n = m, and span the identical subspace i.e., Vqk(q) =
k(q)
k(q)
span{b{q}n }n=1 = span{dℓ{q}n }n=1 . In order to fulfill the last condition all the vectors need to be
updated when a new atom is introduced in the spanning set. Starting from b{q}1 = dℓ{q}1 both the
calculation and upgrading of the biorthogonal set is attained recursively as follows
b{q}k(q) =
w{q}k(q)
kw{q}k(q) k2
,
with w{q}k(q) as in (3)
E
D
b{q}n ← b{q}n − b{q}k(q) dℓ{q}k(q) , b{q}n ,
7
n = 1, . . . , (k(q) − 1).
(7)
These vectors produce the required atomic decomposition for the block q. The coefficients in (1) are
the inner products in (6), i.e. c{q}n = hb{q}n , f{q}i , n = 1, . . . , k(q).
The HBW variant of a greedy strategy establishes an order for upgrading the atomic decomposition of the blocks in the partition. Instead of completing the approximation of each block at once,
the atomic decomposition to be upgraded at each iteration corresponds to a block, say q ⋆ , selected
according to a greedy criterion. In other words: the HBW version of a pursuit method for approximating a signal by partitioning involves two selection instances: a) the selection of dictionary’s atoms
for approximating each of the blocks in the partition and b)the selection, at each iteration step, of
the block where the upgrading of the approximation is to be realized. In [15] the blocks are selected
through the process
q ⋆ = arg max
q=1,...,Q
D
E
dℓ{q}k(q)+1 , rk(q) {q} .
(8)
We notice, however, that condition (8) for the block’s selection can be optimized, in a stepwise sense,
without significant computational cost. The next proposition revise the condition (8) considered
in [15].
Proposition 1. Let ℓ{q}k(q)+1 be the index arising, for each value of q, from the maximization
process (2). In order to minimize the square norm of the total residual krk+1k2 at iteration k + 1 the
atomic decomposition to be upgraded should correspond to the block q ⋆ such that
D
E
k(q)
D
E
dℓ{q}k(q)+1 , r {q}
e
q ⋆ = arg max χ(q), where χ(q) = w{q}
,
f{q}
,
=
k(q)+1
q=1,...,Q
w{q}k(q)+1
(9)
e
with w{q}k(q)+1 and w{q}
k(q)+1 as given in (3).
Proof. Since at iteration k + 1 the atomic decomposition of only one block is upgraded by one atom,
the total residue at iteration k + 1 is constructed as
Q
rk+1 = Ĵp=1 r{p}k(p) Ĵ r{q}k(q)+1 .
p6=q
8
Then,
kr
k+1 2
k =
Q
X
p=1
p6=q
kr{p}k(p) k2 + kr{q}k(q)+1 k2 .
Moreover, since rk(q)+1 {q} = f{q} − P̂Vqk(q)+1 f{q} using the orthonormal vectors (3) to calculate
P̂Vqk(q)+1 f{q} we have:
kr
k(q)+1
2
{q}k = kr
k(q)
2
{q}k −
D
E2
e
w{q}k(q)+1 , f{q} ,
⋆
which is minimum for the q -value corresponding to the maximum value of
D
E
e
w{q}
k(q)+1 , f{q} .
Since w{q}k(q)+1 = dℓ{q}k(q)+1 − P̂Vqk(q) dℓ{q}k(q)+1 and P̂Vqk(q) is hermitian (9) follows.
e
Notice that, since the vectors w{q}
k(q)+1 will be used for subsequent approximations, the opti-
mized ranking of blocks (9) does not require significant extra computational effort. Only Q − 1 of
these vectors will not have been otherwise used when the algorithm stops. Apart from that, the
complexity of selecting the blocks remains being that of finding the maximum element of an array
of length Q, i.e. O(Q).
Before presenting the algorithm details we would like to consider also the optimization of the
selection of atoms (stage a) of a HBW pursuit strategy). Within our implementation the optimized
selection of atoms is readily achievable by the Optimized Orthogonal Matching Pursuit (OOMP)
approach [16], which selects the index ℓ{q}k(q)+1 through the maximization process:
ℓ{q}k(q)+1 = arg max
n=1,...,M
dn , rk(q) {q}
(1 − s{q}n )
1
2
,
with s{q}n =
k(q)
X
i=1
2
e
| hdn , w{q}
ii | .
(10)
The OOMP selection criterion (10) minimizes, in a stepwise sense, the norm of the local error when
approximating a single block [16]. The extra computational cost, in comparison to the OMP selection
step (c.f. (2)) is the calculation of the sum s{q}n in the denominator of (10), for every atom in the
dictionary, n = 1, . . . , M. As becomes clear in Algorithm 3, by saving the sum of previous iterations,
D
E
e
at iteration k(q) + 1 only the inner products dn , w{q}
k(q) , n = 1, . . . , M need to be calculated.
9
2.1
Algorithmic Implementation
The implementation details of the HBW-OOMP method is given in Algorithm 1, which is realized
through Algorithms 2 and 3. For a fixed number K the algorithm iterates until the condition
PQ
q=1
2.2
k(q) = K is met.
Trigonometric dictionaries and sparse representation of music signals
The viability of a sparse representation for a given signal depends in large part on the dictionary
choice. A possibility to secure a suitable dictionary for approximating a signal by partitioning
is through learning techniques [19–22]. Another possibility is to build the dictionary by merging
basis, or dictionaries, containing atoms of different nature. For instance, a Cosine-Dirac dictionary
enlarged by the incorporation of B-spline dictionaries has been shown to be suitable for producing
highly sparse high quality approximation of images [23]. Such dictionaries are very redundant. In
the case of melodic music, however, stunning sparsity levels may be achieved with relatively much
less redundant dictionaries. As will be illustrated here by numerical examples, it is the combination
of a Redundant Discrete Cosine Dictionary (RDCD) D c and a Redundant Discrete Sine Dictionary
(RDSD) D s , defined below, which yields highly sparse representation of music signals when processed
with the proposed techniques.
• D c = { wc1(n) cos π(2i−1)(n−1)
, i = 1, . . . , Nb }M
n=1 .
2M
, i = 1, . . . , Nb }M
• D s = { ws1(n) sin π(2i−1)(n)
n=1 ,
2M
where w c (n) and w s (n), n = 1, . . . , M are normalization factors as given by
√
if n = 1,
rNb
c
2π(n−1)N
π(n−1)
w (n) =
b)
M
Nb + sin( M ) sin(
if n 6= 1.
2π(n−1)
2
2(1−cos(
s
w (n) =
√
rNb
Nb
2
M
))
if n = 1,
−
2πnN
sin( πn
) sin( M b
M
))
2(1−cos( 2πn
M
10
)
if n 6= 1.
Algorithm 1 HBW-OOMP Procedure
Input: Signal partition f{q} ∈ RNb , q = 1, . . . , Q. Dictionary D = {dn ∈ RNb ; kdn k = 1}M
n=1 .
Number K of total atoms to approximate the whole signal.
Output: Sets Γ{q}, q = 1, . . . , Q containing the indices of the selected atom for each block.
k(q)
k(q)
e
Orthonormal and biorthogonal sets {w{q}
n }n=1 and {b{q}n }n=1 , for each block (c.f. (4) and (7)
k(q)
respectively). Coefficients of the atomic decompositions for each block {c{q}n }n=1 , q = 1, . . . , Q.
Approximation of the partition, f a {q}, q = 1, . . . , Q. Approximation of the whole signal f a .
{Initialization}
j=0
for q = 1 : Q do
k(q) = 1; r{q} = f{q}; s{q}n = 0, n = 1, . . . , M
{Select the potential first atom for each block, as below}
ℓ{q} = arg max |hdn , f{q}i|
n=1,...,M
{Store the maximum for each block}
e
e
Set w{q}1 = dℓ{q} ; w{q}
1 = w{q}1 ; b{q}1 = w{q}
1 ; Γ{q} = ∅
χ(q) =
dℓ{q} , f{q}
end for
while j < K do
{Select the block to be approximated (c.f.(9)) and store the index of the atom in the approxi-
mation of that block, as below}
q ⋆ = arg max χ(q)
q=1,...,Q
Γ{q ⋆ } ← Γ{q ⋆ } ∪ ℓ{q ⋆ }
{Update the residue corresponding to block q ⋆ , as below}
e ⋆ }k(q⋆ ) w{q
e ⋆ }k(q⋆ ) , f{q ⋆ }
r{q ⋆ } ← r{q ⋆ } − w{q
if k(q ⋆ ) > 1 then
k(q ⋆ )
{Upgrade the biorthogonal set {b{q ⋆ }n }n=1 to include the atom dℓ{q⋆ } (c.f. (7))}
end if
{Apply Algorithm 2 to select the index ℓ{q ⋆ } for a new potential atom for the block q ⋆ }
Increase k(q ⋆ ) ← k(q ⋆ ) + 1
e ⋆ }k(q⋆ ) from dℓ{q⋆ } , (c.f. (3) and (4)}
{Compute w{q ⋆}k(q⋆ ) and w{q
{Update the objective functional χ for the block q ⋆ }
χ(q ⋆ ) =
e ⋆ }k(q⋆ ) , f{q ⋆ }
w{q
Increase j ← j + 1
end while
{Calculation of coefficients and approximation}
for q = 1 : Q do
f a {q} = f{q} − r{q}; c{q}n = hb{q}n , f{q}i , n = 1, . . . , k(q)
end for
{Concatenation of the partition to produce the whole signal approximation}
Q
f a = Ĵq=1 f a {q}
11
Algorithm 2 Atom Selection Procedure
Input: Residue r{q ⋆}.
Dictionary D = {dn ∈ RNb ; kdn k = 1}M
n=1 .
Auxiliary sequence
e ⋆ }k(q⋆ ) for the upgrading. Set of already selected atoms Γ{q ⋆ }.
{s{q ⋆}n }M
n=1 and vector w{q
Output: Index ℓ{q ⋆ } corresponding to the selected atom dℓ{q⋆ } .
{Apply the Auxiliary Procedure, Algorithm 3, to upgrade the sequence {s{q ⋆ }n }M
n=1 with respect
e ⋆ }k(q⋆ ) }
to vector w{q
{Select the index, as below}
|hdn , r{q ⋆ }i|
ℓ{q ⋆ } = arg max
1
n=1,...,M (1 − s{q ⋆ }n ) 2
⋆
n∈Γ{q
/
}
Algorithm 3 Auxiliary Procedure
⋆
M
Input: Dictionary D = {dn ∈ RNb ; kdn k = 1}M
n=1 . Auxiliary sequence {s{q }n }n=1 and vector
e ⋆ }k(q⋆ ) for the upgrading.
w{q
e ⋆ }k(q⋆ ) .
Output: Upgraded sequence {s{q ⋆ }n }M
n=1 , with respect to w{q
for n = 1 : M do
s{q ⋆ }n = s{q ⋆ }n +
end for
e ⋆ }k(q⋆ )
dn , w{q
2
For M = Nb each of the above dictionaries is an orthonormal basis, the Orthogonal Discrete Cosine
Basis (ODCB) and the Orthogonal Discrete Sine Basis (ODSB), henceforth to be denoted Bc and
Bs respectively. The joint dictionary is an orthonormal basis for M =
Nb
,
2
the Orthogonal Discrete
Cosine-Sine Basis (ODCSB) to be indicated as Bcs . For M > Nb , D c is a RDCD and D s a RDSD.
If M >
Nb
2
D cs = D c ∪ D s becomes a Redundant Discrete Cosine-Sine Dictionary (RDCSD).
For the sake of discussing a fast calculation of inner products with trigonometric atoms, given a
vector y ∈ RM , let’s define
F (y, n, M) =
M
X
y(j)e− ı 2π
(n−1)(j−1)
M
,
n = 1, . . . , M.
(11)
j=1
When M = Nb (11) is the Discrete Fourier Transform of vector y ∈ RNb , which can be evaluated
using FFT. If M > Nb we can still calculate (11) via FFT by padding with zeros the vector y. Thus,
(11) can be used to calculate inner products with the atoms in dictionaries D c and D s . Indeed,
Nb
X
j=1
cos
π(n−1)
π(2j − 1)(n − 1)
y(j) = Re e− ı 2M F (y, n, 2M) , n = 1, . . . , M.
2M
12
(12)
and
Nb
X
sin
j=1
π(n−1)
π(2j − 1)(n − 1)
y(j) = − Im e− ı 2M F (y, n, 2M) , n = 2, . . . , M + 1,
2M
(13)
where Re(z) indicates the real part of z and Im(z) its imaginary part.
The computation of inner products with trigonometric dictionaries via FFT is outlined in Algorithm 4. For dictionaries D c and D s that procedure reduces the complexity for calculating the inner
products from O(Nb M), per block, to O(2M log2 2M), per block. For dictionary D cs the reduction
is larger: O(M log2 M), because both the real and imaginary parts of the FFT are used. When
Algorithm 1 is applied with trigonometric dictionaries the pieces implemented by Algorithms 2 and
3 should be replaced by Algorithms 6 and 7, given in Appendix A. In addition to speeding the calculations via FFT, Algorithms 6 and 7 avoid storing the dictionary. In the case of the dictionary D cs it
is assumed that both the Cosine and Sine components are of the same size and the first
M
2
elements
of the dictionary are Cosine atoms.
Algorithm 4 Computation of inner products with a trigonometric dictionary via FFT.
IPTrgFFT procedure: p =IPTrgFFT(r, M, Case)
Input: r ∈ RNb , M, number of elements in the dictionary, and Case (‘Cos’ or ‘Sin’).
Output: Vector p ∈ RM , with the inner products between r and ‘Cos’ or ‘Sin’ dictionaries.
{Computation of auxiliary vector t ∈ CM to compute p.}
t = FFT(r, M) {FFT with (M − Nb ) zero padding}.
Case ‘Cos’
p(n) =
1
w c (n)
Re(eı
π(n−1)
M
t(n)), n = 1, . . . , M{(c.f. (12))}
Case ‘Sin’
p(n) = − ws1(n) Im(eı
π(n−1)
M
t(n)), n = 1, . . . , M {(c.f.(13))}
Numerical Example I
The signal approximated in this example is a piece of piano melody shown in Fig 1. It consists of
N = 960512 samples (20 secs) divided into Q = 938 blocks with Nb = 1024 samples each. As sparsity
measure for a signal representation we consider the Sparsity Ratio (SR) defined as SR =
N
,
K
where
K is the total number of coefficients in the signal representation. As a measure of approximation
13
1
0.8
0.6
0.4
0.2
0
−0.2
−0.4
−0.6
−0.8
−1
0
1
2
3
4
5
6
7
8
9
10
5
x 10
Figure 1: Piano melody. Credit: Julius O. Smith, Center for Computer Research in Music and Acoustics
(CCRMA), Stanford University.
quality we use the standard Signal to Noise Ratio (SNR),
PNb ,Q
2
i=1 |f {q}(i)|
kfk
q=1
.
= 10 log10 PNb ,Q
a
2
kf − f a k2
i=1 |f {q}(i) − f {q}(i)|
2
SNR = 10 log10
q=1
This numerical example aims at illustrating the following outcomes in relation to the signal in hand.
1) The approximation power of all the orthogonal basis and dictionaries defined above remarkably
improve if, instead of applying OMP/OOMP independently to approximate each block f{q}
up to the same SNR, the HBW-OMP/OOMP approach is applied to match the sparsity of the
whole signal.
2) For approximating the signal with the OMP/OOMP greedy strategies, the redundant dictionaries D c and D cs perform significantly better than any of the orthogonal basis Bc , Bs and
Bcs .
In oder to demonstrate 1) and 2) each of the blocks f{q} is approximated independently with
OMP/OOMP, up to SNR=25 dB. Redundant dictionaries, with redundancy 2 and 4, are simply
creating by setting M = 2Nb and M = 4Nb in the definitions of D c and D s , with Nb = 1024. They
will be denoted as D c2 , D s2 and D c4 , D s4 , respectively. Approximations of the same quality are performed with each of the orthonormal basis Bc , Bs and Bcs , and the mixed dictionaries D cs2 = Bc ∪ Bs
14
Dict.
Bc
D c2
D c4
Bs
D s2
D s4
Bcs
D cs2
D cs4
OMP
SR SNR
14.38 25.0
17.75 25.0
19.39 25.0
7.65 25.0
12.13 25.0
13.17 25.0
10.77 25.0
20.45 25.0
23.83 25.0
HBW-OMP
SR SNR
14.38 35.19
17.75 34.91
19.39 34.64
7.65 28.31
12.13 29.39
13.17 29.34
10.77 30.09
20.45 35.81
23.83 35.56
OOMP
SR SNR
14.38 25.0
18.47 25.0
19.80 25.0
7.65 25.0
12.63 25.0
13.42 25.0
10.77 25.0
22.08 25.0
26.18 25.0
HBW-OOMP
SR
SNR
14.38 35.19
18.47 35.08
19.80 34.95
7.65 28.31
12.63 29.39
13.42 29.33
10.77 30.09
22.08 35.56
26.18 36.37
Table 1: Comparison of the approximation quality (SNR values) and sparsity (SR values) produced with
trigonometric basis B c , B s , B cs and redundant trigonometric dictionaries, D c2 , D c4 ,D s2 , D s4 , D cs2 , and D cs4 .
The second column shows the SR resulting when applying the block independent OMP approach to achieve
a SNR=25.0dB. The fifth column demonstrates the significant gain in SNR rendered by the HBW-OMP
strategy for the same sparsity. Further improvements of the same nature are shown in the last four columns
corresponding to the approaches OOMP and HBW-OOMP.
and D cs4 = D c2 ∪ D s2 . The results are presented in Table 1. As can be seen, the SR produced by
the redundant dictionaries D c2 , D c4 , D cs2 , and D cs4 is substantially larger than that corresponding to
any of the orthogonal basis. Moreover, in all the cases the HBW-OMP/OOMP strategies improve
notoriously (up to 11 dB) upon the OMP/OOMP approaches applied independently to produce a
uniform SNR in the approximation of each block.
It is noticed that the highest sparsity is attained by the dictionary D cs4 which, as already discussed,
involves the most effective implementation via FFT. It is also seen that HBW-OOMP over performs
HBW-OMP in terms of sparsity (the running times are similar). This will appear even more clearly
in Table 2, where the approximation of both methods are downgraded to produce a SNR of 25dB.
The next section discusses the HBW backward strategy which allows for removing atoms from the
signal approximation, in a stepwise optimized fashion.
2.3
Hierarchized Blockwise Backwards OOMP
We extend here the Backward Optimized Orthogonal Matching Pursuit (BOOMP) strategy [24] to
select also the blocks from which the atoms are to be removed for downgrading the approximation
15
of a whole signal. The BOOMP strategy is stepwise optimal because it minimizes, at each step,
the norm of the error resulting by downgrading, by one atom, an atomic decomposition. This
k(q)
result readily follows from the recursive equations for modifying vectors {b{q}n }n=1 to account
k(q)
for the elimination of, say the j-th atom, from the set {dℓ{q}n }n=1. For each q, the reduced set of
k(q)
k(q)
n6=j
n6=j
vectors {b{q}n }n=1 spanning the reduced subspace Vqk(q)/j = span{dℓ{q}n }n=1 can be quickly obtained
through the adaptive backward equations [24, 25]
b{q}n ← b{q}n − b{q}j
hb{q}n , b{q}j i
,
kb{q}j k2
n = 1, . . . , j − 1, j + 1, . . . , k(q).
(14)
Consequently, the coefficients of the atomic decomposition corresponding to the block q, from which
the atom dℓ{q}j is taken away, are modified as
c{q}n ← c{q}n − c{q}j
hb{q}j , b{q}n i
,
kb{q}j k2
n = 1, . . . , j − 1, j + 1, . . . , k(q).
(15)
For a fixed value of q, the BOOMP criterion removes the coefficient c{q}j ⋄ such that [24]
|c{q}j |
.
2
j=1,...,k(q) kb{q}j k
j ⋄ {q} = arg min
(16)
The equivalent criterion applies to the selection of the q ⋄ -block for the downgrading of the whole
approximation. The proof of the next proposition parallels the proof of (16) given in [24].
Q
Proposition 2. Assume that the approximation of a signal is given as f a = Ĵq=1 f a {q}, where f a {q} =
P̂Vqk(q) f{q}. Let j ⋄ {q}, q = 1, . . . , Q be the indices which satisfies (16) for each q. The block q ⋄ , from
where the atom dℓ{q⋄ } ⋄
j {q ⋄ }
is to be removed, in order to leave an approximation f↓a such that kf a − f↓a k
takes its minimum value, satisfies the condition:
q ⋄ = arg min
q=1,...,Q
|c{q}j ⋄ {q} |
.
kb{q}j ⋄ {q} k2
(17)
Proof. Since at each step only one atom is removed from a particular block, say the q-th one, it holds
that kf a − f↓a k = kf a {q} − f↓a {q}k. The identical steps as in [24] lead to the expression
D
E
b{q}j ⋄ {q} , f{q}
f a {q} − f↓a {q} = dℓ{q} ⋄
.
j {q}
kb{q}j ⋄ {q} k2
16
(18)
Thus kf a {q} − f↓a {q}k2 =
D
E
b{q}j ⋄ {q} , f{q}
kb{q}j ⋄ {q} k4
2
. Since
D
b{q}j ⋄ {q} , f{q}
E
= c{q}j ⋄ {q} it follows that
kf a −f↓a k is minimized by removing the atom j ⋄ {q ⋄ } corresponding to the block q ⋄ satisfying (17).
The HBW-BOOMP algorithm for downgrading a given atomic decomposition of a signal is presented in Algorithm 5.
Numerical Example II
In order to illustrate the backward approach, we change the information in the last four columns
of Table 1 to have the sparsity of all the approximations corresponding to SNR=25dB. For this we
downgrade the HBW-OMP/OOMP approximations to degrade the previous quality. As seen in Table
Dict.
Bc
D c2
D c4
Bs
D s2
D s4
Bcs
D cs2
D cs4
OMP
SR SNR
14.38 25.0
17.60 25.0
19.25 25.0
7.65 25.0
12.03 25.0
13.11 25.0
10.77 25.0
20.21 25.0
23.53 25.0
HBW-BOOMP
SR
SNR
25.56
25.0
34.05
25.0
37.01
25.0
13.67
25.0
25.74
25.0
28.18
25.0
19.94
25.0
42.43
25.0
50.19
25.0
OOMP
SR SNR
14.38 25.0
18.47 25.0
19.80 25.0
7.65 25.0
12.63 25.0
13.42 25.0
10.77 25.0
22.08 25.0
26.18 25.0
HBW-BOOMP
SR
SNR
25.56
25.0
36.34
25.0
39.06
25.0
13.67
25.0
27.56
25.0
29.59
25.0
19.94
25.0
48.48
25.0
59.68
25.0
Table 2: Comparison of sparsity (SR values) for the same SNR. The HBW-BOOMP results are obtained
by degrading the HBW-OMP/OOMP approximation in Table 1 to SNR=25dB.
2, the result is that the sparsity increases drastically. The SR improves up to 128% (for the D cs4
dictionary) in comparison with the standard application of the same approaches without ranking the
blocks. Notice also that the best SR result (for dictionary D cs4 ), is 133% higher than the best SR
result for an orthogonal basis (the Bc basis).
3
Refinement by Swaps
The suitability of the HBW strategy for approximating the class of signals we are considering is a
notable fact. However, the possibility of approximating every element of the partition completely in17
Algorithm 5 HBW-BOOMP
k(q)
Input: Biorthogonal sets {b{q}n }n=1 , q = 1, . . . , Q. Coefficients in the approximation of each
k(q)
block {c{q}n }n=1 , q = 1, . . . , Q. Sets of indices in the decomposition of each block Γ{q}, q =
1, . . . , Q. Number of total atoms for the downgraded approximation, Kd .
Output: Downgraded biorthogonal sets. Coefficients of the downgraded atomic decompositions
for each block. Indices of the remaining atoms.
{Initialization}
K = 0;
for q = 1 : Q do
k(q) = |Γ{q}| {Number of elements in Γ{q}}
K = K + k(q)
{Select the index of the potential first atom to be removed in each block, as below}
|c{q}n |
j ⋄ {q} = arg min
2
n=1,...,k(q) kb{q}n k
|c{q}j ⋄{q} |
χ↓ (q) =
{Store the minimum}
kb{q}j ⋄ {q} k2
end for
i=K
while i > Kd do
{Select the block to be downgraded, as below}
q ⋄ = arg min χ↓ (q)
q=1,...,Q
{Apply backward biorthogonalization (c.f. (14)) to downgrade the biorthogonal set corresponding to block q ⋄ , with respect to b{q ⋄ }j ⋄ {q⋄ } and shift indices}
{Update the coefficients corresponding to block q ⋄ (c.f. 15) and shift indices}
Downgrade the set Γ{q ⋄ } ← Γ{q ⋄ }/ℓ{q ⋄ }j ⋄ {q⋄ } and shift indices
Decrease k(q) ← k(q) − 1
{Select the new potential atom to be removed from block q ⋄ , as below}
|c{q ⋄ }n |
j ⋄ {q ⋄ } = arg min
⋄
2
n=1,...,k(q) kb{q }n k
{Update the objective functional χ↓ for the block q ⋄ , as below}
|c{q ⋄ }j ⋄ {q⋄ } |
χ↓ (q ⋄ ) =
kb{q ⋄ }j ⋄ {q⋄ } k2
Decrease i ← i − 1
end while
18
dependent of the others has the convenient feature of leaving room for straightforward parallelization.
This is particularly important for implementations on Graphics Processing Units (GPU), for instance.
Accordingly, an alternative worth exploring is to maintain the block independent approximation but
allowing for possible eventual HBW refinements.
Assuming that the approximation (1) of each block in a signal partition is known, the goal now
is to improve the signal approximation maintaining unaltered the total number of atoms K. The
proposed refinement consists of movement of atoms controlled by the following instructions:
i) Use criterion (17) to remove one atom from the block q ⋄ . Let us call this block a ‘donor’ block
and indicate it with the index qd .
ii) Use criterion (9) to incorporate one atom in the approximation of the block q ⋆ . Let us call this
block a ‘receiver’ and indicate it with the index qr .
iii) Denote by δd the square norm of the error introduced at step i)by downgrading the approximation of the donor block qd . Denote by δr the square norm of the gain in improving the
approximation of the receiver qr . Proceed according to the following rule:
If δr > δd accept the change and repeat steps i) and ii). Otherwise stop.
The above procedure is termed HBW-SbR-OMP/OOMP according to whether the selection of the
atom at stage ii) is realized through the OMP (c.f. (2)) or OOMP (c.f. (10)) criterion.
Notice that the HBW-SbR-OMP/OOMP methods, implemented as proposed above, possess the
desirable feature of acting only if the approximation can be improved by the proposed swap. The
possible situation for which qr = qd corresponds to a swap in the original SbR-OMP/OOMP approach
[17], which may take place here as a particular case.
Implementation Remark: For successive implementation of the steps i) and ii) one needs: to
downgrade the vectors b{q ⋄ }n , n = 1, . . . k(q) at step i) and to upgrade these vectors at step ii). In
order to allow for step ii) the orthogonal set {w{q ⋄ }n }k(q)
n=1 has to be downgraded as well. This can
19
be realized effectively by the plane rotation approach [26, 27].
Numerical Example III
We illustrate the HBW-SbR-OMP/OOMP approaches by applying them to the signal in Fig. 2,
which is a flute exercise consisting of N = 96256 samples divided into Q = 94 blocks with Nb = 1024
samples each.
Table 3 and 4 show the improvement in quality obtained by applying the HBW-SbR-OMP/OOMP
1
0.8
0.6
0.4
0.2
0
−0.2
−0.4
−0.6
−0.8
−1
0
1
2
3
4
5
6
7
8
9
10
4
x 10
Figure 2: Flute exercise: N = 96256 samples.
approaches on outputs of the OMP/OOMP approximations. Notice that in most cases the SNR
results are practically equivalent to those obtained by the HBW-OMP/OOMP approaches.
Dict.
Bc
D c2
D c4
Bs
D s2
D s4
Bcs
D cs2
D cs4
SR OMP Swaps HBW-SbR-OMP
10.13
25.0
1426
26.80
14.26
25.0
1013
26.69
15.58
25.0
945
26.63
7.27
25.0
2612
26.72
12.10
25.0
1397
27.24
13.16
25.0
1435
27.27
8.08
25.0
1822
26.65
29.00
25.0
636
27.38
35.17
25.0
518
27.36
HBW-OMP
26.79
26.67
26.61
26.73
27.27
27.25
26.65
27.37
27.36
Table 3: Comparison of quality (SRN values) for a fixed sparsity: that corresponding to SRN=25 with
the OMP approach. The forth column shows the number of swaps and the fifth columns the SRN achieved
by those swaps through the HBW-SbR-OMP refinement to the outputs yielding the second column. For
further comparison the last column shows the results corresponding to the HBW-OMP approach.
20
Dict.
Bc
D c2
D c4
Bs
D s2
D s4
Bcs
D cs2
D cs4
SR OOMP Swaps HBW-SbR-OOMP
10.13
25.0
1426
26.80
14.83
25.0
865
26.66
15.80
25.0
929
26.60
7.27
25.0
2612
26.72
12.58
25.0
1531
27.14
13.34
25.0
1445
27.27
8.08
25.0
1822
26.65
33.47
25.0
556
27.64
42.05
25.0
228
27.10
HBW-OOMP
26.79
26.68
26.61
26.73
27.27
27.26
26.65
27.63
27.69
Table 4: Same description as in Table 3 but for the OOMP, HBW-SbR-OOMP, and HBW-OOMP approaches.
4
Processing of large signals
As already discussed, the additional computational complexity introduced by the HBW raking of
Q blocks is only O(Q) (over the complexity of a standard pursuit algorithm implemented without
ranking the blocks). Storage does become more demanding though. The need for saving vectors (7)
and (3), for instance, implies to have to store 2K vectors of size Nb . This requirement restricts the
amount of blocks to be processed with a small system such as a standard laptop. Hence, we outline
here a processing scheme which makes it possible the application of HBW techniques on very large
signals. For this the whole signal needs to be divided into large segments of convenient size, say
Ns . More specifically, a number of, say P blocks, of size Nb are grouped together to form S larger
sP
segments g{s} = Ĵq=(s−1)P +1 f{q}, s = 1, . . . , S, where it is assumed that Nb P = Ns and SNs = Nb Q.
Each segment g{s} ∈ RNs is approximated with a HBW strategy as an entire individual signal. The
approximation of the original signal is assembled at the very end of the process by the operation
S
f a = Ĵs=1 ga {s}. In other words, the processing of large signals is realized by chopping the signal
into segments and processing each segment independently of the others. Nevertheless, the question
as to how to set the sparsity constraint for each segment needs further consideration. One could,
of course, require the same sparsity in every segment, unless information advising otherwise were
available. While uniform sparsity guarantees the same sparsity on the whole signal, in particular
21
cases, where the nature of the signal changes over its range of definition, it would not render the
best approximation. In order to avoid encountering this type of situation we introduce a preliminary
step: the randomization of all the small blocks in the signal partition. The implementation is carried
out as follows:
i) Given a signal f split it into Q blocks of size Nb . Apply an invertible random permutation Π
to scramble the block’s location (the bottom graph in Figure 3 provides a visual illustration of
how the signal in the top graph looks after this step). Let’s denote the re-arranged signal by
Q
f̃ = Ĵq=1 f̃{q}.
sP
ii) Group every P of the blocks f̃{q}, q = 1, . . . , Q to have S segments g̃{s} = Ĵq=(s−1)P +1 f̃ {q}, s =
1, . . . , S =
Q
.
P
iii) Approximate each of the segments g̃{s} independently of the others, as it each segment were
sP
an entire signal, i.e., obtain g̃a {s} = Ĵq=(s−1)P +1 f̃ a {q}, s = 1, . . . , S.
S
iv) Assemble the segments to have the approximation of the whole signal as f̃ a = Ĵs=1 g̃a {s}.
v) Reverse the permutation of the block’s location to obtain, from f̃ a , the approximation of the
original signal f a .
1
0
−1
0
5
10
15
5
x 10
1
0
−1
0
5
10
15
5
x 10
Figure 3: The top graph shows a piece (1433600 samples) of Piazzola music interpreted by Gidon Kremmer
and orchestra. The bottom graph shows the resulting signal after the randomization of the blocks.
22
Numerical Example IV
The signal to be approximated is a 32.5 secs (1433600 samples) piece of Piazzola music, shown
in the top graph of Figure 3. It is partitioned into Q = 1400 blocks of 1024 points each. After randomization (bottom graph of the same figure) the blocks are grouped into 50 segments,
each of which is independently approximated by the HBW-OMP/OOMP approach to produce a
SR=11.53. The resulting SNR of the approximated signal is 29.12 dB with HBW-OMP and 30.15 dB
with HBW-OOMP. Equivalent approximation quality is obtained with other realizations of the random process. For appreciation of this result the original signal was also approximated by HBWOMP/OOMP, but without segmentation. In that case the quality corresponding to SR=11.53 is
only 1.9% higher (SNR=29.66 dB with HBW-OMP and SNR=30.73 dB with HBW-OOMP). The
conclusion is that, even if the largest the segments the less the distortion for the same sparsity, applying HBW-OMP/OOMP on segments of a large signal may be still significantly more advantageous
than the independent approximation of the blocks. Indeed, in this example the latter yields a SNR
of 25 dB with OMP and 26.1 dB with OOMP.
5
Conclusions
Cooperative greedy pursuit strategies for approximating a signal partition subjected to a global
constraint on sparsity have been considered. The cooperation between partition units was realized in several ways: i)By ranking the partition units for their sequential stepwise approximation
(HBW-OMP/OOMP) ii)By ranking the partition units for stepwise downgrading of the whole approximation (HBW-BOOMP) iii)By allowing for downgrading of some partition units (donors) and
upgrading of another partitions units (receivers), to improve the approximation quality (HBW-SbROMP/OOMP). Comparisons of the OMP and OOMP criteria indicate that for the type of dictionaries
and signals considered here, the OOMP criterion renders improvements of sparsity up to 20% in relation to the equivalent strategy involving OMP.
23
The HBW algorithms maintain the complexity comparable with the totally independent approximation of each partition unit. The examples presented here, using trigonometric dictionaries for
achieving high quality approximation of musics signals, provide a clear illustration of the benefits
arising by the proposed strategies. Indeed, for a piano melody the increment in sparsity is 128%
higher than the sparsity produced by approximating each partition unit up to same quality. Comparison of trigonometric dictionaries results, against results obtained with trigonometric basis, give
an improvement of equivalent order: 133%. The mixed Cosine-Sine dictionary has yielded the best
sparsity results, not only for the test signals included in this paper but for a significant number of
other signals, all in the category of melodic music.
Because the proposed HBW implementation of pursuit strategies requires storage of order Nb K (Nb
being the length of each block in the partition and K the number of total atoms to approximate the
whole signals) a simple procedure for processing large signal was proposed: Previous to dividing the
signal into independent segments, each of which to be approximated with the identical constraint of
sparsity, the randomization of the partition units is advised.
24
Appendix A
Algorithm 6 Atom Selection Procedure via FFT
Input: Residue r{q ⋆ }. Dictionary Case (‘Cos’, ‘Sin’ or mixed ‘Cos-Sin’) and number of elements,
c
⋆
M
s
⋆
M
M. Sequence {s{q ⋆ }n }M
n=1 if Case is ‘Cos’ or ‘Sin’, and {s {q }n }n=1 , {s {q }n }n=1 if Case is ‘Cos-
e ⋆ }k(q⋆ ) to upgrade the sequences. Set of already selected atoms Γ{q ⋆ } (sets Γc {q ⋆ }
Sin’. Vector w{q
and Γs {q ⋆ } with the atoms in each component if Case is ‘Cos-Sin’).
Output: Selected index ℓ{q ⋆ } for a potential new atom for block {q ⋆ }.
Cases single ‘Cos’ or ‘Sin’
{Call IPTrgFFT procedure, Algorithm 4, to calculate inner products, and apply Algorithm 7 to
e ⋆ }k(q⋆ ) }
upgrade the sequence {s{q ⋆ }n }M
n=1 with respect to vector w{q
p=IPTrgFFT(r{q ⋆ }, 2M, Case),
{Select the index for a new potential atom for the block q ⋆ , as below.}
|p{q ⋆ }(n)|
ℓ{q ⋆ } = arg max
1
n=1,...,M (1 − s{q ⋆ }n ) 2
⋆
n∈Γ{q
/
}
Case ‘Cos-Sin’
{Same precedure as for the previous case but for both ‘Cos’ and ‘Sin’ cases},
pc {q ⋆ }=IPTrgFFT(r{q ⋆ }, M ‘Cos’); ps {q ⋆ }=IPTrgFFT(r{q ⋆ }, M, ‘Sin’)
s
⋆
M
{Apply Algorithm 7 to upgrade the sequences {sc {q ⋆ }n }M
n=1 and {s {q }n }n=1 and selec the index}
Set M ←
M
2
ℓc {q ⋆ } = arg max
n=1,...,M
n∈Γ
/ c {q ⋆ }
|pc {q ⋆ }(n)|
ℓs {q ⋆ } = arg max
1 ;
(1 − sc {q ⋆ }n ) 2
n=1,...,M
n∈Γ
/ s {q ⋆ }
|ps {q ⋆ }(n)|
{Evaluate the maximum value, as below}
|ps {q ⋆ }(ℓs {q ⋆ })|
|pc {q ⋆ }(ℓc {q ⋆ })|
s
⋆
;
µ
{q
}
=
µc {q ⋆ } =
1
1
(1 − sc {q ⋆ }ℓc {q⋆ } ) 2
(1 − ss {q ⋆ }ℓs {q⋆ } ) 2
µ{q ⋆} = max(µc {q ⋆ }, µs {q ⋆ })
if µ = µc then
ℓ{q ⋆ } = ℓc {q ⋆ }
else
ℓ{q ⋆ } = ℓs {q ⋆ } + M
end if
25
1
(1 − ss {q ⋆ }n ) 2
Algorithm 7 Auxiliary Procedure via FFT
e ⋆ }k(q⋆ ) . Dictionary Case: ‘Cos’, ‘Sin’ or ‘Cos-Sin’, and number of elements
Input: Vector w{q
c
⋆
M
s
⋆
M
M. Sequence {s{q ⋆ }n }M
n=1 , for Cases ‘Cos’ and ‘Sin’, or {s {q }n }n=1 , and {s {q }n }n=1 , for Case
‘Cos-Sin’.
Output:
c
⋆
M
Upgraded sequence, {s{q ⋆ }n }M
n=1 , for Cases ‘Cos’ and ‘Sin’, or {s {q }n }n=1 and
{ss {q ⋆ }n }M
n=1 for Case ‘Cos-Sin’.
Case ‘Cos’ or ‘Sin’
e ⋆ }k(q⋆ ) }
{Call IPTrgFFT procedure, Algorithm 4, to calculate inner products with vector w{q
e ⋆ }k(q⋆ ) , 2M, Case)
p{q ⋆ }=IPTrgFFT(w{q
{Upgrade the sequence, as below}
for n = 1 : M do
s{q ⋆ }n = s{q ⋆ }n + |p{q ⋆ }(n)|2
end for
Case ‘Cos-Sin’
e ⋆ }k(q⋆ )
{Call IPTrgFFT procedure, Algorithm 4, to calculate the inner products between vector w{q
and each of the dictionary components: Cases ‘Cos’ and ‘Sin’}
e ⋆ }k(q⋆ ) , M, ‘Cos’)
pc =IPTrgFFT(w{q
e ⋆ }k(q⋆ ) , M, ‘Sin’)
ps =IPTrgFFT(w{q
{Upgrade the sequences, as below}
for n = 1 :
M
2
do
sc {q ⋆ }n = sc {q ⋆ }n + |pc {q ⋆ }(n)|2
ss {q ⋆ }n = ss {q ⋆ }n + |ps {q ⋆ }(n)|2
end for
26
References
[1] B. K. Natarajan, “Sparse Approximate Solutions to Linear Systems”, SIAM Journal on Computing, 24, 227–234 (1995).
[2] S. S. Chen, D. L. Donoho, M. A Saunders, “Atomic Decomposition by Basis Pursuit”, SIAM
Journal on Scientific Computing, 20, 33–61 (1998).
[3] J. H. Friedman and W. Stuetzle, “Projection Pursuit Regression”, Journal of the American
Statistical Association, 76, 817–823 (1981).
[4] S. G. Mallat and Z. Zhang, “Matching Pursuits with Time-Frequency Dictionaries”, IEEE Trans.
Signal Process., 41, 3397–3415 (1993).
[5] Y.C. Pati, R. Rezaiifar, and P.S. Krishnaprasad, “Orthogonal matching pursuit: recursive
function approximation with applications to wavelet decomposition,” Proc. of the 27th ACSSC,1,
40–44 (1993).
[6] Y. Eldar, P. Kuppinger and H. Biölcskei, “Block-Sparse Signals: Uncertainty Relations and
Efficient Recovery”, IEEE Trans. Signal Process., 58, 3042–3054 (2010).
[7] D. L. Donoho , Y. Tsaig , I. Drori , J. Starck, “Stagewise Orthogonal Matching Pursuit”, IEEE
Transactions on Information Theory, 58, 1094–1121 (2006).
[8] D. Needell and J.A. Tropp, “CoSaMP: Iterative signal recovery from incomplete and inaccurate
samples”, Applied and Computational Harmonic Analysis, 26, 301–321 (2009).
[9] D. Needell and R. Vershynin, “Signal Recovery From Incomplete and Inaccurate Measurements
via Regularized Orthogonal Matching Pursuit”, IEEE Journal of Selected Topics in Signal Processing, 4, 310–316 (2010).
[10] D. L. Donoho, “Compressed sensing”, IEEE Trans. Inf. Theory,52, 1289–1306 (2006).
27
[11] E. Candès, J Romberg, and T. Tao, “Robust uncertainty principles: exact signal reconstruction
from highly incomplete frequency information,” IEEE Trans. Inf. Theory, 52, 489 –509 (2006).
[12] R. Baraniuk, “Compressive sensing”, IEEE Signal Processing Magazine, 24, 118–121, (2007).
[13] E. Candès and M. Wakin, “An introduction to compressive sampling”, IEEE Signal Processing
Magazin, 25, 21–30 (2008).
[14] R. Baraniuk, “More Is less: Signal processing and the data deluge”, Science, 331, 717–719
(2011).
[15] L. Rebollo-Neira, R. Matiol, S. Bibi, “Hierarchized block wise image approximation by greedy
pursuit strategies,” IEEE Signal Process. Letters, 20, 1175–1178 (2013).
[16] L. Rebollo-Neira and D. Lowe, “Optimized orthogonal matching pursuit approach”, IEEE Signal
Process. Letters, 9, 137–140 (2002).
[17] M. Andrle and L. Rebollo-Neira, “A swapping-based refinement of orthogonal matching pursuit
strategies”, Signal Processing, 86, 480–495 (2006).
[18] http://www.nonlinear-approx.info/examples/node01.html
[19] K. Kreutz-Delgado, J. F. Murray, B. D. Rao, K. Engan, T. Lee, and T. J. Sejnowski, “Dictionary
learning algorithms for sparse representation”, Neural Comp., 15, 349 – 396 (2003).
[20] M Aharon, M Elad, and A Bruckstein, “K-SVD: An Algorithm for Designing Overcomplete
Dictionaries for Sparse Representation”, IEEE Transactions on Signal Processing, 54, 4311–
4322 (2006).
[21] R. Rubinstein, M. Zibulevsky, M. Elad, “Double Sparsity: Learning Sparse Dictionaries for
Sparse Signal Approximation”, IEEE Trans. Signal Process. 58, 1553–1564 (2010).
28
[22] I. Tos̆ić and P. Frossard, “Dictionary Learning: What is the right representation for my signal?”,
IEEE Signal Process. Magazine, 28, 27–38 (2011).
[23] L. Rebollo-Neira, J.Bowley Sparse representation of astronomical images, Journal of The Optical
Society of America A, 30, 758–768 (2013).
[24] M. Andrle, L. Rebollo-Neira, E. Sagianos, “Backward-optimized orthogonal matching pursuit
approach”, IEEE Signal Proc. Let.,11,705–708 (2004).
[25] L. Rebollo-Neira, “Backward adaptive biorthogonalization”, Int. J. Math. Math. Sci. 1843–1853
(2004).
[26] Å. Björck, “Numerical Methods for Least Squares Problems”, SIAM, Philadelphia, 1996.
[27] G.W. Stewart, “On the stability of sequential updates and downdates”, IEEE Trans. Signal
Process., 43, 2642–2648 (1995).
29
| 8cs.DS
|
1
Wirelessly Powered Crowd Sensing: Joint Power Transfer,
Sensing, Compression, and Transmission
Xiaoyang Li, Changsheng You, Sergey Andreev, Yi Gong, and Kaibin Huang
arXiv:1711.02066v2 [cs.IT] 5 Mar 2018
System operation
Abstract—Leveraging massive numbers of sensors in user
equipment as well as opportunistic human mobility, mobile crowd
sensing (MCS) has emerged as a powerful paradigm, where
prolonging battery life of constrained devices and motivating
human involvement are two key design challenges. To address
these, we envision a novel framework, named wirelessly powered
crowd sensing (WPCS), which integrates MCS with wireless power
transfer (WPT) for supplying the involved devices with extra
energy and thus facilitating user incentivization. This paper
considers a multiuser WPCS system where an access point
(AP) transfers energy to multiple mobile sensors (MSs), each
of which performs data sensing, compression, and transmission.
Assuming lossless (data) compression, an optimization problem is
formulated to simultaneously maximize data utility and minimize
energy consumption at the operator side, by jointly controlling
wireless-power allocation at the AP as well as sensing-data
sizes, compression ratios, and sensor-transmission durations at
the MSs. Given fixed compression ratios, the proposed optimal
power allocation policy has the threshold-based structure with
respect to a defined crowd-sensing priority function for each
MS depending on both the operator configuration and the MS
information. Further, for fixed sensing-data sizes, the optimal
compression policy suggests that compression can reduce the
total energy consumption at each MS only if the sensing-data
size is sufficiently large. Our solution is also extended to the case
of lossy compression, while extensive simulations are offered to
confirm the efficiency of the contributed mechanisms.
I. I NTRODUCTION
The unprecedented growth of the Internet of Things (IoT)
applications for Smart Cities fuels the deployment of billions
of wireless IoT devices that help automate a wide range of services, such as temperature measurement, pollution assessment,
traffic monitoring, and public-safety surveillance. However,
traditional wireless sensor networking (WSN) solutions are
limited in their coverage and scalability, as well as suffer
from high maintenance costs [1]. Recently, leveraging massive
numbers of sensors in user handheld and wearable equipment
led to the emergence of mobile crowd sensing (MCS), which is
becoming a new paradigm that involves humans as part of the
sensing infrastructure [2]. The key MCS challenges are prolonging device battery life and facilitating human engagement.
These can be tackled by leveraging wireless power transfer
(WPT) techniques as a user incentive for charging mobile
sensors (MSs) within our proposed framework of wirelessly
powered crowd sensing (WPCS) [3].
This work was supported in part by Hong Kong Research Grants Council
under the Grants 17209917 and 17259416, and Natural Science Foundation
of Guangdong Province under Grant 2015A030313844. X. Li, C. You, and
K. Huang are with the Dept. of EEE at The University of Hong Kong, Hong
Kong (e-mail: [email protected], [email protected]; [email protected]).
X. Li is also with the Dept. of EEE at Southern University of Science and
Technology, Shenzhen, China. S. Andreev is with the Laboratory of Electronics and Communications Engineering, Tampere University of Technology,
Finland (e-mail: [email protected]). Y. Gong is with the Dept. of EEE at
Southern University of Science and Technology, Shenzhen, China (e-mail:
[email protected]).
Message
passing
Wireless
power
transfer
Mobile sensor 1 frame
Mobile sensor 2 frame
…
Mobile sensor N frame
Time
Crowd Sensing
Data fusion
center
Energy
beamforming
Data
transmission
Mobile sensor 1
Mobile sensor 2
Access point
Mobile sensor N
Figure 1: Our considered multiuser WPCS system.
In this work, we consider a multiuser WPCS system in
Fig. 1, which comprises an operator-deployed access point
(AP) that transfers energy to multiple MSs. To optimize
the operator’s reward, we design a set of efficient control
policies for maximizing data utility and minimizing energy
consumption simultaneously.
A. Mobile Crowd Sensing
The consideration of MCS for IoT in Smart Cities has
motivated research on advanced technologies that go beyond
conventional WSNs, such as personal location-based systems,
trusted sensing platforms, and people-centric sensing applications [2]. A key MCS design issue is how to involve people
into collaborate sensing, as users may not be willing to
contribute if the associated energy costs are high. There are
two approaches to address this issue, known as opportunistic
and participatory sensing [4], [5]. The former assumes that
applications may run in the background and opportunistically
collect data without user attention. The latter requires users to
consciously participate in sensing activities and timely transmit
the operator-requested data. However, this may consume significant device resources (i.e., battery and computing power),
and it is essential to design incentivization mechanisms that
compensate for the costs of participation [6]–[8].
Specifically, inspired by the location-based games, entertainment as an incentive can enrich the experience of participants who collect e.g., geographic data on user preferences and
interests. Service as an incentive is another widely-recognized
mechanism. A typical example here is traffic monitoring,
wherein users provide real-time traffic information (e.g., on
road congestion levels) and in return receive timely services,
2
such as accurate route planning. Finally, money as an incentive
is an intuitive method that interprets sensed data as goods
for gaining profit. Accordingly, an auction scheme for MCS
named reverse auction was proposed in [7], where the sellers
(i.e., participants) submit their bidding prices to the buyer (i.e.,
operator) to compete for the data sensing activities. Further, the
operator selects a subset of MSs for crowd sensing operation
and determines individual payments.
In [8], the authors designed an alternative framework by
utilizing Stackelberg game for monetary incentivization. The
operator and the MSs negotiate in a two-stage manner to
achieve the Stackelberg equilibrium: first the operator announces the payments and then each MS decides on whether
to participate. Here, incentive allocation in multiuser MCS
systems depends on data quality of each user, which has been
investigated in [9], [10]. In particular, the user data quality was
estimated in [9] by using unsupervised learning techniques for
characterizing long-term reputation. For privacy management,
inverse correlation between the service quality and the privacy
level was determined in [10] from the perspective of data
analytics. Despite these substantial efforts, the issue of energy
consumption at the MSs is not yet resolved. The proposed
work addresses it by relying on advanced WPT techniques
as well as contributes the first theoretical resource-allocation
design of trading energy for data, which we envisioned in [3].
B. Wirelessly Powered Communication and Sensing
Designed originally for point-to-point power transmission,
the WPT technology has already comprised advanced techniques from inductive coupling and electromagnetic radiation
[11], [12]. According to [13], the power to be wirelessly
transferred is on the level of micro-watt, thus WPT is mostly
applied for the ultra-dense small cell networks. Further integration of information and power transfer gave rise to
the emerging field of simultaneous wireless information and
power transmission (SWIPT) [14]. Multiple studies focused on
applying SWIPT to a variety of communication systems and
networks, including multiple-input-multiple-output (MIMO)
communications [14], [15], two-way transmissions [16], relaying [17]–[19], orthogonal frequency-division multiple access
(OFDMA) transmissions [20], and cognitive networking [21],
[22], among others. Further design efforts tailored it for practical setups by accounting for channel state information [23]
and quality of service [24]. Recently, WPT was considered
for unmanned aerial vehicles [25] and mobile-edge computing
networks [26] for powering computation offloading in energyconstrained devices.
Another important application of WPT is in sensing systems, where the MSs have a limited energy storage due to
their compact form-factor, but are expected to perform sophisticated sensing tasks cooperatively. Powering these sensors is
challenging, since frequent battery recharging/replacement is
impractical if not impossible. Recent breakthroughs in WPT
techniques offer a viable solution by powering MSs with
energy beamforming [27]–[31]. Specifically, a multi-antenna
wirelessly powered sensor networking (WPSN) testbed has
been developed in [28], where a power beacon transfers energy
to a sensor node via an adaptive energy-beamforming scheme.
For multiple WPSN sensors, efficient design of energy
beamforming is much more convoluted. In [29], accounting for
practical sensing and circuit power consumption at MSs, the
authors proposed a collaborative energy-beamforming scheme
for powering a cluster of sensors to transfer data by using
distributed information beamforming. Cooperative sensing was
further addressed in [30], where clusters of sensors forward
their information by utilizing relay and SWIPT techniques.
Finally, for low-complexity passive backscatter sensors, a
novel architecture for large-scale WPSN systems was recently
proposed in [31], where the sensors are powered by distributed
power beacons. However, the existing literature concentrates
on optimizing WPT efficiency to improve the sensing-andcommunication throughput, but disregards the design of user
incentivization and operator reward schemes. In this work, we
bridge this gap by integrating WPT with MCS to facilitate
human involvement into collective sensing activities, thus
simultaneously maximizing data utility and minimizing energy
consumption at the operator side.
C. Contributions and Organization
In this work, by leveraging the potential of WPT for
powering MSs and motivating MCS, we aim at addressing
the following two key issues.
•
•
The first one is how to incentivize and select MSs
for crowd-sensing operation. This includes designing a
new incentivization approach that can trade energy for
data depending on the data utility at different MSs. An
optimized sensor selection policy needs to be developed
at the operator to select a set of sensors to participate in
the MCS for the maximum reward.
The second one is how to jointly control wireless-power
allocation, data sensing, compression, and transmission
at the selected MSs. This requires the operator to allocate
wireless power for improved energy supply of each MS,
while the MSs should employ energy-efficient sensing,
compression, and transmission policies to minimize the
energy consumption.
To account for these two considerations, this paper addresses a multiuser WPCS system controlled by an operator
and comprising a multiple-antenna AP that transfers energy
to multiple single-antenna MSs. Each MS accumulates part
of such energy as a reward and utilizes the rest to conduct
the MCS tasks, including sensing, data compression, and
transmission of compressed data to the AP. We consider both
lossless and lossy (data) compression as well as design a joint
control policy for maximizing the data utility and minimizing
the energy consumption at the AP simultaneously. The main
contributions of this work are summarized as follows.
•
Problem Formulation and Iterative Solution: For lossless
compression, we formulate an optimization problem for
simultaneously maximizing data utility and minimizing
energy consumption at the operator. An iterative solution
that reaches the local optimum is then proposed for a joint
optimization of power allocation, sensing, compression,
and transmission.
3
Joint Optimization of Power Allocation, Sensing, and
Transmission: Given fixed compression ratios, the considered non-convex problem is reduced to a convex formulation. We first derive a semi-closed form expression for
the optimal sensor-transmission duration. The results are
used for deriving the optimal wireless-power allocation
and sensing-data sizes.
• Joint Optimization of Compression and Transmission:
Given fixed sensing-data sizes, the sensor-transmission
durations and compression ratios are optimized for minimizing energy consumption at the operator. The derived
optimal policy suggests that the MSs should perform
data compression only if the sensing-data size exceeds
a certain threshold.
• Optimal Control for Lossy Compression: The proposed
solution is further extended to lossy compression, while
the optimization problem is modified to account for the
respective data utility degradation. Given fixed compression ratios, the corresponding approach is similar to the
lossless case. For the optimization of compression to
preserve the data utility, we further optimize both the
sensing-data sizes and the compression ratios at the MSs.
The rest of this paper is organized as follows. The system model is introduced in Section II. Section III studies
efficient joint control in the WPCS system based on lossless
compression. In Section IV, this approach is extended for the
system with lossy compression. Simulation results are offered
in Section V, followed by the conclusions in Section VI.
•
II. S YSTEM M ODEL
Consider a multiuser WPCS system shown in Fig. 1 and
comprising multiple single-antenna MSs together with one
multi-antenna AP. The model of system operation is described
in below sub-sections.
A. WPCS Operation
We focus on a particular fixed time window for crowdsensing, which is divided into three phases: message passing,
WPT, and crowd sensing (see Fig. 1).
Consider the message-passing phase. Each sensor feeds
back to the AP its parameters including the effective channel power gain, sensing, and compression power. Given the
knowledge of all sensor parameters, the AP jointly optimizes
the power allocation for WPT and sensor operations (namely,
data sensing, compression, and transmission). The objective
is to simultaneously maximize the data utility and minimize
the energy consumption (see the performance metrics in the
sequel). Subsequently, the AP informs individual sensors on
their allocated power, optimal compression ratios, sensing-data
sizes, and time partitions for sensing, compression, and transmission. The power consumption at the AP for transmitting
these control parameters is negligible due to their small data
sizes.
Further, in the WPT phase, the AP employs an antenna
array to beam energy to the MSs as an incentive for sensing
(see Fig. 1). Upon harvesting energy, each MS stores a part
of it as a reward and applies the rest to operate the sensing
tasks, including sensing, data compression, and transmission
of compressed data to the AP.
Finally, in the crowd-sensing phase, the sensors simultaneously perform data sensing, compression, and transmission
based on the settings determined by the AP in the messagepassing phase. Parallel data collection is enabled via transmissions over orthogonal channels as allocated by the AP
to the sensors based on OFDMA. The time duration for
crowd sensing, denoted as T , is divided into three slots
(s)
(c)
with adjustable durations tn , tn , and tn , which are used
for sensing, compression, and transmission, respectively. This
introduces the following constraint:
(c)
(Time constraint) t(s)
n + tn + tn ≤ T.
(1)
B. Model of Wireless Power Transfer
In the WPT phase with a fixed duration of T0 (seconds),
the AP transfers energy simultaneously to N MSs by pointing
N radio beams to the corresponding MSs. Let gn denote
the effective channel power gain between the AP and MS
n using energy beamforming, while Pn is the transmission
power of the corresponding beam. The AP transmission power
is assumed to be fixed and represented by P0 , thus resulting
in the following constraint:
(Power constraint)
N
X
Pn ≤ P0 .
(2)
n=1
The energy transferred from the AP to MS n, denoted by
(h)
(h)
En (Pn ), is En (Pn ) = ηgn Pn T0 , where the constant 0 <
η < 1 represents the energy conversion efficiency.
Each of the sensors selected to participate in the crowd
sensing stores a part of received energy as its reward. Let the
(r) (s)
reward energy at sensor n, denoted by En (`n ), be propor(r) (s)
(r) (s)
tional to the size of sensing data, namely, En (`n ) = qn `n
(r)
with
qn being
a fixed scaling factor. The remaining energy,
(r)
(h)
En − En , is consumed by the sensor operations, including sensing, compression, and transmission, where energy
(s)
(c)
consumption is represented respectively by En , En and
(t)
En as specified in the sequel. Based on the above discussion,
each MS should satisfy the following energy constraint:
(Energy constraint) En(r) +En(s) +En(c) +En(t) ≤ ηgn Pn T0 . (3)
C. Mobile-Sensor Model
At each MS, crowd sensing comprises three sequential operations: data sensing, data compression, and data transmission
as shown in Fig. 2 and modeled as follows.
1) Data sensing model: Consider MS n. Let sn denote the
(s)
output data rate. Given the sensing time-duration tn , the size
(s)
(s)
(s)
of raw sensing data, denoted by `n (in bits), is `n = sn tn .
Based on the data-sensing energy consumption model in [32],
(s)
let qn denote the sensing energy consumption for generating
1-bit of data. The total energy consumption for sensing at MS
(s) (s)
(s) (s)
(s) (s)
n, denoted as En (`n ), is given as En (`n ) = qn `n .
4
Crowd-Sensing Phase
Data sensing
Data compression
Data transmission
Mobile Sensor
Harvested
energy
En(h)
Sensor
Compressor
Transmitter
En(s)
En(c)
En(t)
Battery
En(r)
Figure 2: MS architecture, operation, and energy consumption.
2) Data compression model: First, consider lossless compression of sensed data, for which the original data can be
perfectly reconstructed from the compressed data. Relevant
techniques include Huffman, run-length, and Lempel-Ziv encoding. It is assumed that all MSs deploy the same lossless
compression method with the maximum compression ratio
denoted as Rmax , namely, the maximum ratio between the
sizes of raw and compressed data. Let Rn ∈ [1, Rmax ] denote
the compression ratio chosen by MS n. Then, the size of
(s)
compressed data `n is given as `n = `n /Rn . Note that in
the current literature, there is still a lack of well-established
models for the computation complexity of data compression.
However, a tractable model can be inferred from the existing
measurement results for Zlib compression [33]. Specifically,
the required CPU cycles for compressing 1-bit of data can be
approximated as an exponential function of the compression
ratio Rn as:
(Compression complexity) C(Rn , ) = eRn − e ,
(4)
where is a positive constant depending on the compression
method. At Rn = 1, C(Rn , ) = 0, which corresponds to no
compression. Let fn denote the fixed CPU-cycle frequency at
(c)
(c)
MS n and tn is the compression time duration. Then, tn =
(s)
(`n C(Rn , ))/fn . Following the practical model in [34], each
(c)
CPU cycle consumes the energy of qn = αfn2 , where α is a
constant determined by the circuits. The energy consumption
(c) (s)
for data compression at MS n, denoted by En (`n , Rn ), is
(c) (s)
(c) (s)
given as En (`n , Rn ) = qn `n C(Rn , ) with C(Rn , ) in
(4). It follows that
(c) (s) Rn
(Compression energy) En(c) (`(s)
−e ).
n , Rn ) = qn `n (e
(5)
Further, consider lossy compression that achieves a higher
compression ratio than its lossless counterpart at the cost
of some information loss. Typical techniques involve data
truncation after a transform (e.g., discrete cosine transform) or
discarding the data immaterial for user perception. Our model
of compression energy consumption follows that in (5) with
the parameters {Rn , Rmax , } replaced by their counterparts
for lossy compression {Řn , Řmax , ˇ}. The information loss
can be measured with a quality factor, denoted by bn ∈ (0, 1]
and defined as the equivalent size of raw data for 1-bit
of compressed data in terms of utility. The quality factor,
however, has not been well-modeled in the existing literature.
To address it, one tractable model can bepinferred from the
simulation results in [35], given by bn = 1/ Řn . Specifically,
bn → 0 when Řn → ∞, and bn → 1 when Řn → 1, which
corresponds to no compression.
3) Data transmission model: Each selected MS transmits
(t)
its compressed data to the AP. Let Pn denote the transmission
power and tn is the transmission time duration. Assuming
channel reciprocity, the achievable transmission rate (in bits/s),
denoted by vn , can be given as vn = `n /tn = B log2 (1 +
gn Pn(t)
N0 ), where B is the bandwidth and N0 is the variance
of complex-white-Gaussian noise. Hence, the transmission
(t)
(t)
energy consumption denoted by En (`n ) follows: En (`n ) =
t
n
(t)
f (`n /tn ), where the function f (x) is defined as
Pn tn =
gn x
f (x) = N0 (2 B − 1).
4) Time and energy constraints (revisited): Finally, based
on the above models, the time constraint in (1) and energy
constraint in (3) can be rewritten as
(s)
(s)
`n C(Rn , )
`n
+
+ tn ≤ T,
sn
fn
h
i
tn
qn(r) +qn(s) +qn(c) C(Rn , ) `n(s) + f
gn
(6)
(s)
`n
tn Rn
!
≤ ηgn Pn T0 .
(7)
An important observation from the time constraint perspective is that given the sensing rate sn and the CPUcycle frequency fn , the partitioning of crowd-sensing time
of sensor n for sensing, compression, and transmission can
be determined by the sensing-data size `n , compression ratio
Rn , and transmission time tn to be optimized in the sequel.
D. Performance Metrics
Recall that the operator’s reward, which becomes our
performance measure, refers to the difference between the
sum weighted data utility and the weighted energy cost. Its
mathematical definition is provided as follows. Following a
commonly used model (see e.g., [36]), the utility of `n -bit data
delivered
by sensor
n is tackled by the logarithmic function
(s)
an log 1 + bn `n , where bn is the said information loss due
to compression and an is a weight factor depending on the type
of data. Note that the logarithmic function is chosen here to
model the diminishing return as the data size increases. Hence,
the sum data utility for the operator, denoted by U (`(s) ), is
U (`(s) ) =
N
X
an log(1 + bn `(s)
n ).
(8)
n=1
Then, the operator’s reward, denoted by R(`(s) , P ), can be
modeled as
R(`(s) , P ) =
N
X
n=1
an log(1 + bn `(s)
n )−c
N
X
Pn T0 ,
(9)
n=1
where c denotes the price of unit energy with respect to that
of unit data utility.
5
Access Point
Message passing
Joint optimization
of power allocation, sensing,
and transmission
Iterate
Joint optimization
of compression and transmission
Mobile
sensors
Effective channel power gains,
sensing rates,
sensing/compression power
Sensor transmission durations,
wireless power allocation,
sensing-data sizes,
compression ratios
Figure 3: Implementation of our iterative solution for optimal WPCS.
III. J OINT P OWER A LLOCATION , S ENSING ,
C OMPRESSION , AND T RANSMISSION
as demonstrated in Fig. 3. Specifically, the approach iterates
between solving two sub-problems, P1-A and P1-B as given
below, for optimizing two subsets of parameters, namely,
(s)
{Pn , `n , tn } and {Rn , tn }, respectively. The reason for such
separation of parametric optimization is in that it creates two
convex sub-problems as shown in the sequel. It is necessary
to include the transmission durations {tn } as variables in both
sub-problems, since they are affected by all other parameters.
First, with the compression ratios {Rn } fixed, problem
P1 reduces to the following problem P1-A, named joint
optimization of power allocation, sensing, and transmission:
N
X
max
In this section, we formulate and solve a non-convex
problem of jointly optimizing the power allocation, sensing,
lossless compression, and transmission by deriving the localoptimal policy as discussed in Section II. This yields important
guidelines for operating the proposed WPCS system. The
results are extended to the case of lossy compression in the
following section.
Pn ≥0,`(s)
n ≥0,
tn ≥0
s.t.
(P1-A)
an log(1 + `(s)
n )−c
n=1
N
X
Pn T0
n=1
N
X
P n ≤ P0 ,
n=1
βn `(s)
n + tn ≤
αn `(s)
n +
(r)
tn
f
gn
T,
∀ n,
(s)
`n
Rn tn
!
≤ ηgn Pn T0 ,
∀ n,
(c)
(s)
C(Rn ,)
1
.
sn +
fn
(s)
{Pn } ∪ {`n } ∪ {tn }, the sensing-
A. Problem Formulation
where αn = qn +qn +qn C(Rn , ) and βn =
The specific design problem here is to jointly optimize the
AP power allocation for WPT to sensors, {Pn }, the sizes
(s)
of sensing data, {`n }, the data compression ratios, {Rn },
and the partitioning of crowd-sensing time for sensing and
(s)
compression, determined by {tn } together with {`n } and
{Rn }. The objective is to maximize the operator’s reward in
(9) under the power constraint in (2), time constraint in (6), and
energy constraint in (7). For lossless compression, the quality
factor bn = 1. Mathematically, the optimization problem at
hand can be formulated as follows:
Next, given the parameters
data sizes `(s) are fixed and thus problem
is reduced to
PP1
N
minimizing the sum energy consumption: n=1 Pn T0 . Given
this criterion, the energy constraints in (7) are active at the
optimal point:
!
(s)
h
i
tn
`n
(r)
(s)
(c)
(s)
qn +qn +qn C(Rn , ) `n + f
= ηgn Pn T0 .
gn
tn Rn
(10)
After substituting this, problem P1-B, named joint optimization of compression and transmission, can be produced from
problem P1 as:
!
(s)
tn
`n
(r)
(s)
(c)
(s)
min
[qn +qn +qn C(Rn , )]`n + f
gn
tn Rn
Rn ∈[1,Rmax ],
N
X
max
(s)
Pn ≥0,`n
≥0,
Rn ∈[1,Rmax ],tn ≥0
s.t.
(P1)
an log(1 + `(s)
n )−c
n=1
N
X
n=1
(s)
`n
N
X
Pn T0
n=1
tn ≥0
P n ≤ P0 ,
+
(s)
`n C(Rn , )
(s)
(P1-B)
+ tn ≤ T,
sn
fn
(r)
(s)
[qn + qn + qn(c) C(Rn , )]`(s)
n
!
(s)
tn
`n
+ f
≤ ηgn Pn T0 ,
gn
tn Rn
∀ n,
s.t.
(s)
`n
`n C(Rn , )
+
+ tn ≤ T,
sn
fn
∀ n.
By iteratively solving these two problems, which are proved to
be convex in the sequel, the solution is guaranteed to converge
and reach a local maximum of the operator’s reward function
∀ n. in formulation P1.
B. Iterative Solution Approach
The convexity of problem P1 is characterized as follows.
Lemma 1. Problem P1 is a non-convex optimization problem.
This lemma
easily proved, since the optimization
n can be o
(s)
variables, `n , Rn , tn , are coupled at the constraints in
tn (s)
(s)
forms of `n C(Rn , ) and
f `n /(tn Rn ) . To establish
gn
the optimal policy for attaining a local maximum and characterize its structure, we propose an iterative solution approach
C. Joint Optimization of Power Allocation, Sensing, and
Transmission
To solve problem P1-A, we first prove that it can be reduced
to the problem of transmission-duration optimization. Then,
the solution of the reduced problem is used to characterize
the optimal wireless power allocation and sensing-data sizes.
1) Optimal Sensor Transmission: Problem P1-A is solved
as follows. First, observe that problem P1-A is always feasible,
(s)
since {Pn = 0, `n = 0, Rn = 1} for each n is a feasible solution. Next, some useful necessary conditions for the optimal
solution are provided as follows.
6
Lemma 2. As a result of solving problem P1-A, the optimal
power allocation {Pn∗ }, sensor transmission durations {t∗n },
(s)
and sensing-data sizes {(`n )∗ } satisfy the following:
!
(s)
∗ t ∗
(`n )∗
n
(s)
αn `n
+ f
= ηgn Pn∗ T0 ,
∀n,
gn
Rn t∗n
∗
βn `(s)
+ t∗n = T,
∀n.
n
Lemma 2 is proved in Appendix A. The equalities arise
from the fact that to maximize the operator’s reward, each
MS should fully utilize the transferred energy and the crowd(s)
sensing time. Based on these equalities, {Pn∗ } and {(`n )∗ }
can be written as functions of {t∗n }. Consequently, problem
P1-A can be equivalently transformed into problem P2 for
sensor-duration optimization having only {tn } as its variables:
max
tn ≥0
N
X
an log 1 +
n=1
T − tn
βn
N
X
αn (T − tn )
tn
T − tn
+ 2f
ηβn gn
ηgn
R n βn t n
n=1
N
X
T − tn
tn
αn (T − tn )
+ 2f
≤ P0 T0 ,
ηβn gn
ηgn
Rn βn tn
n=1
−c
(P2) s.t.
0 ≤ tn ≤ T,
∀ n.
The convexity of problem P2 is established in the following
lemma as proved in Appendix B.
Lemma 3. Problem P2 is a convex optimization problem.
As a result, problem P2 can be solved by the Lagrange
method. The corresponding partial Lagrange function is
L(tn , λ) =
N
X
n=1
αn (T −tn ) tn
T −tn
+ 2f
ηβn gn
ηgn
Rn βn tn
T −tn
− an log 1+
−λP0 T0 .
(11)
βn
(λ+c)
Define a function y(x) as y(x) = f (x) − (x + Rn1βn )f 0 (x)
and t∗n as the optimal solution for solving problem P2. Then,
applying Karush-Kuhn-Tucker (KKT) conditions leads to the
following necessary and sufficient conditions:
an
(λ∗ +cn ) 1
T − t∗n
αn
∂L
=
+
y
−
∂t∗n
β +(T −t∗n )
ηgn
gn
Rn βn t∗n
βn
n
∗
> 0, tn = 0,
(12a)
= 0, 0 < t∗n < T,
∗
< 0, tn = T,
!
N
X t∗ T −t∗ αn (T −t∗ )
n
n
n
∗
λ
f
+
−P0 T0 = 0.
ηgn2
Rn βn t∗n
ηβn gn
n=1
(12b)
Combing these conditions yields the optimal sensortransmission policy given as follows.
Proposition 1 (Optimal Sensor Transmission). Given fixed
compression ratios, for each MS selected for crowd sensing
Algorithm 1 A Bisection-Search Algorithm for Computing
the Optimal Sensor Transmission
•
•
Step 1 [Initialize]: Let λ` = 0, λh = λmax , and ξ > 0.
Obtain h tn,` and tn,h based
on (13),
icalculate E` =
PN
tn,`
T −tn,`
αn (T −tn,` )
+
f
and Eh =
2
n=1
ηβn gn
ηgn
Rn βn tn,`
PN h αn (T −tn,h ) tn,h T −tn,h i
+ ηg2 f Rn βn tn,h
respectively.
n=1
ηβn gn
n
Step 2 [Bisection search]: While E` 6= P0 T0 and Eh 6=
P0 T0 , update λ` and λh as follows.
(1) Define λm = (λ` + λh )/2, compute tn,m and Em =
PN h αn (T −tn,m ) tn,m T −tn,m i
+ ηg2 f Rn βn tn,m .
n=1
ηβn gn
n
(2) If Em < P0 T0 , let λh = λm , else λ` = λm .
Until λh − λ` < ξ. Return t∗n = tn,m .
with (0 < t∗n ≤ T ), the optimal transmission duration t∗n
satisfies:
T −t∗
n
(λ∗ + c)N0
T ln 2
BRn βn t∗
n − 1
1−
2
ηgn2
BRn βn t∗n
(λ∗ + c)αn
an
−
+
= 0, (13)
βn + (T − t∗n )
ηgn βn
where λ∗ satisfies the condition in (12b).
One can observe from Proposition 1 that the optimal sensortransmission duration t∗n has no closed form. However, the
value of t∗n can be efficiently computed by following the proposed bisection-search procedure summarized in Algorithm 1.
Based on Proposition 1, the dependence of the optimal sensortransmission durations on the utility weights and channels is
characterized in the following corollary.
Corollary 1 (Properties of Optimal Sensor Transmission). The
optimal sensor-transmission durations {t∗n } have the following
properties:
1) If the effective channel power gains {gn } are identical
but the utility weights satisfy a1 ≥ a2 · · · ≥ aN , then
t∗1 ≤ t∗2 · · · ≤ t∗N .
2) If the utility weights {an } are identical but the effective
channel power gains satisfy g1 ≥ g2 · · · ≥ gn , then t∗1 ≤
t∗2 · · · ≤ t∗N .
The proof is given in Appendix C. These cases indicate
that for a sensor with large utility weight and high effective
channel power gain, it is beneficial to have a short transmission
duration so that the sensing duration can be made longer to
increase the amount of sensing data, for larger sum data utility.
Last, the following result for the case of small optimal
transmission durations (t∗n T ) follows from Proposition 1
by substituting T − t∗n ≈ T .
Corollary 2 (Approximate Optimal Sensor Transmission).
Consider the optimal sensor-transmission duration t∗n in (13).
If t∗n T ,
T ln 2
t∗n ≈
,
(14)
BRn βn (θ̃ − 1)
ln 2
gn αn
an ηgn2
BRn βn −1 .
where θ̃ = W0
−
−1
e
(c+λ∗ )(βn +T )N0 N0 βn
7
Here, one can observe that the approximate result retains
the properties in Corollary 1. Moreover, this result shows that
the optimal sensor-transmission duration increases with the
growing sensing-and-compression power as represented by αn
and defined in problem P1-A.
2) Optimal Wireless-Power Allocation: Combining the optimal sensor-transmission policy in Proposition 1 with the
optimality conditions in Lemma 2 yields the optimal wirelesspower allocation policy formulated in the following proposition, which is proved in Appendix D.
Proposition 2 (Optimal Wireless-Power Allocation). The optimal wireless-power allocation policy {Pn∗ } that solves problem
P1-A is given as
∗
αn (T −t∗n )
tn
T −t∗n
1
+
f
, ϕn ≥ λ∗ ,
Rn βn t∗n
βn
Pn∗ = ηgn T0 gn
0,
ϕn < λ∗ ,
(15)
where t∗n and λ∗ are given in Proposition 1, while ϕn is named
the crowd-sensing priority function defined as
ϕn = κn − c,
where
4
κn =
(r)
qn
+
(s)
qn
+
an ηgn
(c)
qn C(Rn , )
(16)
+
N0 ln 2
gn BRn
.
(17)
Proposition 2 yields that the optimal wireless-power allocation policy has a threshold-based structure. In other words,
only the MSs with (crowd-sensing) priority functions exceeding the threshold λ∗ will be allocated wireless power or,
equivalently, selected for participating in the crowd sensing
operation. Specifically, both the operator’s configuration (including cost weight as well as provisioned power), and MSs’
individual parameters significantly affect the crowd-sensing
sensor selection as discussed in the following remarks.
Remark 1 (Effects of Operator’s Configuration on Sensor
Selection). First, it can be observed from (16) that if the cost
weight c is sufficiently large (c > max{κn }), the priority
function for each sensor is negative i.e., ϕn ≤ 0, ∀n. Since the
Lagrange multiplier λ∗ ≥ 0, it follows that ϕn < λ∗ , ∀n, thus
implying that no sensor is selected for crowd sensing. On the
other hand, if the energy cost is negligible (i.e., c = 0), each
sensor has a positive priority function. In this case, provided
that the operator supplies sufficient power for crowd sensing,
which leads to λ∗ = 0 (see (12a)), all of the sensors will
be selected to participate in the crowd sensing to increase
the operator’s reward. For the general case of finite c and
λ∗ > 0, it is desirable for the operator to choose only a subset
of sensors with the priorities above λ∗ for crowd sensing.
Remark 2 (Effects of Sensor Parameters on Sensor Selection). Individual parameters of sensors also affect the sensorselection policy by the term of κn as defined in (17). To be
specific, a sensor with smaller sensing-and-compression power
and larger effective channel power gain has higher crowdsensing priority and thus is more likely to be selected. Such
sensors with larger effective channel power gains can not only
harvest more energy, but also consume less transmission power
for their data transmission. Last, to compete for crowd sensing
with other sensors, a sensor can reduce its amount of reward
(r)
power, qn , so as to increase its own priority function.
Finally, the sum allocated power is affected by the cost
weight as shown below.
PN
Corollary 3. The sum allocated power n=1 Pn∗ T0 decreases
PN
with the cost weight c. In particular, n=1 Pn∗ T0 = P0 T0 , if
c = 0.
This corollary is aligned with intuition and suggests that
the optimal wireless power allocation balances the tradeoff
between obtaining data utility and energy with the price
measured by c.
3) Optimal Sensing-Data Sizes: Using a similar approach
for deriving the optimal wireless power allocation in Proposition 2, the optimal sensing-data sizes can be derived in the
following proposition by using Proposition 1 and Lemma 2.
Proposition 3 (Optimal Sensing-Data Sizes). The optimal
(s)
sensing-data sizes {(`n )∗ } that solve problem P1-A are given
as:
∗
T − tn , ϕ ≥ λ∗ ,
n
(s) ∗
βn
(18)
(`n ) =
0,
ϕn < λ∗ ,
where t∗n and λ∗ are given in Proposition 1, while ϕn is defined
in (16).
The effects of parameters on the sensing-data size are
characterized as follows.
Corollary 4 (Properties of Optimal Sensing-Data Sizes).
(s)
The optimal sensing-data sizes {(`n )∗ } have the following
properties:
1) If the effective channel power gains {gn } are identical
but the utility weights satisfy a1 ≥ a2 · · · ≥ aN , then
(s)
(s)
(s)
(`1 )∗ ≥ (`2 )∗ · · · ≥ (`n )∗ .
2) If the utility weights {an } are identical but the effective
channel power gains satisfy g1 ≥ g2 · · · ≥ gn , then
(s)
(s)
(s)
(`1 )∗ ≥ (`2 )∗ · · · ≥ (`n )∗ .
This corollary reflects that it is desirable to increase the
sensing-data sizes for sensors with high effective channel
power gains, since they consume less transmission energy.
D. Joint Optimization of Compression and Transmission
Consider the optimization formulation in problem P1-B (see
Section III-B). First, the problem convexity is established in
the following lemma, which can be easily proved by using the
property of perspective function [37].
Lemma 4. Problem P1-B is a convex optimization problem.
Hence, problem P1-B can be solved by the Lagrange
method. The Lagrange function is
!
(s)
t
`
n
n
b = Q(Rn )`(s) + f
b n , tn , λ)
L(R
n
gn
tn Rn
!
(s)
(s)
`n C(Rn , ) `n
b
+ λ tn +
+
− T , (19)
fn
sn
8
(r)
(s)
(c)
where Q(Rn ) = qn + qn + qn C(Rn , ). Let g(x) =
0
f (x) − xf (x) and t∗n , Rn∗ be the optimal solution for tackling
problem P1-B. Directly applying the KKT conditions results
in the following necessary and sufficient conditions:
!
!
(s)
(s)
b∗
b
∗
λ
`
`
∂L
n
n
Rn
= qn(c) +
`(s)
−
f0 ∗ ∗
n e
∂Rn∗
fn
gn Rn∗2
tn Rn
∗
> 0, Rn = 1,
(20a)
= 0, 1 < Rn∗ < Rmax ,
< 0, Rn∗ = Rmax ,
∗
> 0, tn = 0,
b
1
∂L
`sn
∗
b
=
(20b)
g ∗ ∗ +λ
= 0, 0 < t∗n < T,
∂t∗n
gn
tn Rn
∗
< 0, tn = T,
!
(s)
(s)
∗
`n C(Rn , ) `n
∗
∗
b
λ tn +
+
− T = 0.
(20c)
fn
sn
Therefore, the optimal sensor decisions on compress-or-not
are derived as follows, which is proved in Appendix E.
Lemma 5 (Compress or Not?). Given fixed sensing-data sizes,
sensor n should perform data compression (i.e., Rn > 1) for
minimizing its energy consumption, if and only if its sensingdata size satisfies:
T
`(s)
,
n >
θb
1
+
where θb = (c) ln 2
, and
s
n
qn gn fn
B W0 −
ϑn =
fn ln 2
Be
N0
+1e−ϑn +ϑn +1
+ 1.
Lemma 5 suggests that the optimal decision on compressor-not has a threshold-based structure. To be specific, data
compression is preferred only if the sensing-data size exceeds
a threshold that depends on the effective channel power gain
and compression power among others. This is due to the fact
that reducing the data sizes by compression can substantially
lower the exponentially-increasing energy required for transmitting larger amounts of data.
Based on the conditions in (20a)-(20c), the key results of
this subsection are derived in the following proposition.
Proposition 4 (Optimal Compression and Transmission). The
solution of problem P1-B is as follows.
1) The optimal sensor-transmission durations {t∗n } are
∗
(s)
(s)
`n eRn − e
`n
∗
tn = T −
−
,
∀ n,
(21)
fn
sn
which corresponds to full utilization of the crowd-sensing
duration.
2) The optimal compression ratios {Rn∗ } are
n
n
o o
bn , Rmax , 1 ,
Rn∗ = max min R
∀ n, (22)
bn satisfies: z(R
bn ) = 0 with the function z(Rn )
where R
defined by
1
1
g
eRn
z(Rn ) = qn(c) −
gn fn
d(Rn )Rn
Algorithm 2 Optimal Lossless Compression Ratio Algorithm
•
•
Step 1 [Initialize]: Let R` = 1, Rh = Rmax , and ξ > 0.
Obtain z(R` ) and z(Rh ).
Step 2 [Bisection search]: While z(R` ) 6= 0 and z(Rh ) 6=
0, update R` and Rh as follows.
(1) Define Rn∗ = (R` + Rh )/2, compute z(Rn∗ ).
(2) If z(Rn∗ ) > 0, let Rh = Rn∗ , else R` = Rn∗ .
Until Rh − R` < ξ.
Algorithm 3 Efficient Algorithm for Lossless Compression
Solving Problem P1
•
•
Step 1: Initialize the compression ratio Rn∗ .
Step 2: Repeat
(1) Given fixed compression ratio, find the optimal power
(s)
allocation Pn∗ and the sensing-data size (`n )∗ by using
Algorithm 1.
(2) Given fixed sensing-data size, update the compression
ratio Rn∗ by using Algorithm 2.
Until convergence or a maximum number of iterations
has been reached.
−
4
and d(Rn ) =
T
(s)
`n
−
1
f0
gn Rn2
1
d(Rn )Rn
,
(23)
1
1
− C(Rn , ) > 0.
sn
fn
bn has no closed
The proof is given in Appendix F. Though R
form, it can be computed by the bisection-search method in
Algorithm 2. This is due to the monotonicity of function z(Rn )
as shown in the lemma below and proved in Appendix G.
Lemma 6. z(Rn ) is a monotone-increasing function of Rn .
Consequently, the effects of parameters on the optimal
compression ratio are detailed below.
Corollary 5 (Properties of Optimal Compression Ratios).
The optimal compression ratios {Rn∗ } have the following
properties:
(c)
1) If the compression powers {qn } are identical but the
effective channel power gains satisfy g1 ≥ g2 · · · ≥ gn ,
∗
then R1∗ ≤ R2∗ · · · ≤ RN
.
2) If the effective channel power gains {gn } are identical
(c)
(c)
(c)
but the compression powers satisfy q1 ≥ q2 · · · ≥ qn ,
∗
∗
∗
then R1 ≤ R2 · · · ≤ RN .
Proposition 5 is proved in Appendix H. Accordingly, it
is more energy-efficient for a sensor to perform more aggressive compression in the case of a poor channel or small
compression-power consumption.
E. Complexity Analysis
Combining the results of the preceding subsections, the
efficient control policy for achieving the maximum operator’s
reward can be computed by an iterative procedure as summarized in Algorithm 3 with the complexity analyzed as follows.
In the outer iteration, the algorithm alternatively optimizes the
9
control policy given by fixed compression ratios and sensingdata sizes. First, given fixed compression ratios, the optimal
control policy can be computed by the one-dimensional search
for λ∗ described in Algorithm 1, which invokes an inner
iteration. The complexity of the one-dimensional search can
be represented by O(log(1/ξ)) given the solution accuracy of
ξ > 0.
Moreover, the complexity of the sensor-transmission,
power-allocation, and sensing-data size is O(N ) in each inner
iteration. Hence, the total complexity of the control policy
given fixed compression ratios is O(N log(1/ξ)). Next, given
fixed sensing-data sizes, the optimal compression ratio for each
sensor can be computed by another one-dimensional search in
Algorithm 2. Therefore, each iteration also has the computation complexity of O(N log(1/ξ)). Last, the complexity of the
outer iteration can be characterized by O(log(1/ξ)). In summary, the total complexity of Algorithm 3 is O(N log2 (1/ξ)),
which is feasible in practical systems.
IV. E XTENSION TO L OSSY C OMPRESSION
The preceding section considered lossless compression. In
this section, those results are extended to the case of lossy
compression. To this aim, problem P1 is modified to include
the quality factor bn = √1 for lossy compression, which
Řn
leads to its lossy-compression counterpart as follows:
!
N
N
X
X
1 ˇ(s)
an log 1+ p `n −c
max
P̌n T0
(s)
P̌n ≥0,`ˇn
≥0,
Řn
n=1
n=1
Řn ∈[1,Řmax ],ťn ≥0
s.t.
(P3)
N
X
P̌n ≤ P0 ,
n=1
(s)
`ˇn
Then, the corresponding optimization problem can be formulated as follows.
!
i ť
h
ˇ(s)
`
n
n
ˇ
Ř
ˇ
(r)
(s)
(c)
(s)
+ f
min
`ˇn qn +qn +qn e n −e
gn
Řn ťn
Řn ∈[1,Řmax ],
ťn ≥0,`ˇ(s)
n ≥0
s.t.
(P4)
1
p `ˇ(s)
n = un ,
Řn
(s)
(s)
`ˇn ˇŘn
`ˇn
− eˇ ≤ T,
e
+
ťn +
sn
fn
∀ n,
∀ n.
This problem is solved in the following sub-section.
A. Joint Optimization of Sensing-Data Sizes, Compression,
and Transmission
Problem P4 can be shown to be non-convex but may also
be transformed into a convex formulation as p
follows. Here, we
define a set of new
variables
{r
}
as
r
=
Řn ∈ [1, rmax ],
n
n
p
where rmax = Řmax . Hence, `ˇsn = un rn and Řn = rn2 .
Substituting them into problem P4 yields:
h
2
i ť u
n
n
min
un rn qn(r) +qn(s) +qn(c) eˇrn −eˇ + f
gn
rn ∈[1,rmax ],
ťn rn
ťn ≥0
!
2
1
eˇrn − eˇ
(P5) s.t. ťn + un rn
≤ T,
∀ n.
+
sn
fn
Using a similar approach as that for deriving Lemma 4, it can
be proved that problem P5 is a convex optimization problem.
Hence, applying the Lagrange method as for solving problem
P1-B, the key results of this subsection are derived as stated
in the following corollary.
(s)
`ˇn (eˇŘn −eˇ)
Proposition 5 (Optimal Lossy Compression and Transmis+
+ ťn ≤ T,
∀ n,
sion). The solution of P5 is as follows.
sn
fn
h
i
(r)
(s)
(c) ˇŘn
ˇ
1) The optimal sensor-transmission durations {ť∗n } are
`ˇ(s)
q
+
q
+
q
(e
−
e
)
n
n
n
n
∗2
!
(s)
un rn∗ eˇrn − eˇ
`ˇn
ťn
un rn∗
≤ ηgn Pˇn T0 , ∀ n.
+
−
,
∀ n, (25)
ť∗n = T −
gn Řn ťn
fn
sn
Problem P3 has a similar form as P1 and it is also nonconvex. Leveraging the iterative solution approach developed
in the preceding section, problem P3 can be decomposed into
two sub-problems for optimizing two corresponding subsets
(s)
(s)
of parameters, {P̌n , `ˇn , ťn } and {Řn , `ˇn , ťn }. The first subproblem has the same form as problem P1-A in the losslesscompression counterpart; thus, it can be solved in a similar
way (see Section III-C). The second sub-problem (see problem
P4 below) differs from problem P1-B due to the addition of
(s)
sensing-data sizes {`ˇn } to the optimization variables. The
reason is that lossy-compression causes information loss and
thereby reduces data utility, which has to be compensated for
by increasing the sensing-data sizes. To address this issue and
also for better tractability, we apply the constraint that the data
utility contributed by different sensors is fixed in problem P4
as {un }, which results from solving the first sub-problem.
Consequently, problem P4 follows from the problem P3 as:
1 ˇ(s)
`n . (24)
(Data utility constraint) un = bn `ˇ(s)
n = p
Řn
which corresponds to full utilization of the crowd-sensing
duration.
2) The optimal square roots of compression ratios {rn∗ } are
rn∗ = max {min {b
rn , rmax } , 1} ,
∀ n,
(26)
where rbn satisfies: ž(b
rn ) = 0 with the function ž(rn )
defined by
1
1 0
ž(rn ) = rn Q0 (rn )+Q(rn )−
f(
)
2
ˇ
gn r
d(rn )rn
n
1
1
− g
[V (rn )+rn V 0 (rn )] , (27)
ˇ
gn
d(rn )rn
2
(r)
(s)
(c)
where Q(rn ) = qn + qn + qn eˇrn −eˇ , V (rn ) =
2
2
1
eˇrn−eˇ
ˇ n ) = T −rn 1 + eˇrn−eˇ .
+
and
d(r
sn
fn
un
fn
sn
This proposition can be proved by a similar method as that
for deriving Proposition 4. Even though rbn has no closed
form, it can be computed by a bisection method due to the
10
0.07
monotonicity of function ž(rn ) as shown in the lemma below,
which is proved in Appendix I.
Following a similar approach as that for deriving Corollary 5, it can be proved that the optimal compression ratio
increases with the effective channel power gain and decreases
with the compression power.
Allocated power for MS 1 (W)
Lemma 7. ž(rn ) is a monotone-increasing function of rn .
0.06
0.05
0.04
0.03
0.02
0.01
Remark 3 (Lossy vs. Lossless Compression). Given the same
sensing-data size, as compared to lossless compression, the
lossy operation consumes less energy for compression and results in smaller size of compressed data, which further reduces
transmission-energy consumption. The disadvantage, however,
is that it sacrifices a part of data utility due to information loss.
Hence, there exists a tradeoff between the energy consumption
and the data utility for lossy compression. The decision of
choosing lossless or lossy compression is intractable due to
the lack of closed-form expression for the maximum reward
(see Proposition 4 and Corollary 5). However, it can be observed that lossy compression is preferred under the following
conditions: poor channel, high compression power, or low
sensing power. In these cases, lossy compression can achieve
significant energy savings.
V. S IMULATIONS AND D ISCUSSION
The simulation parameters are set as follows unless specified
otherwise. Our WPCS system comprises 1 AP equipped with
Nt = 40 antennas and 10 MSs each with a single antenna.
The Nt antennas are divided into 10 groups, each having
Nt /10 = 4 antennas and pointing to one specific MS by
transmit beamforming. The time duration of the WPT phase
is T0 = 1 s. For each MS n, the channel vector hn ∈ C4×1
is assumed to follow Rician fading and thus is modeled as
r
r
Ωn K
Ωn b
h̄n +
hn ,
hn =
1+K
1+K
where the Rician factor K = 10, the average fading power
gain Ωn = 5 × 10−4 × dn −2 , the distance between the AP and
MS dn is distributed uniformly in the range [1, 5] m, the lineof-sight (LOS) component h̄n has all elements equal to one,
b n is a 4 × 1 i.i.d. CN (0, 1) vector representing smalland h
scale fading. The effective channel power gain gn = |hn |2
as follows from transmit/receive beamforming. The energy
conversion efficiency is set as η = 0.5.
Further, the effective bandwidth B = 10 KHz and the variance of complex-white-Gaussian-channel noise N0 = 10−9
W. For the system reward, the utility weights for all the MSs
are the same: an = 0.04 and the cost weight c = 0.6. For
each MS, the sensing rate follows a uniform distribution with
sn ∈ [1, 10] × 104 bits/s and the energy consumption per bit is
(s)
distributed uniformly by qn ∈ [1, 10] × 10−12 J/bit [38]. For
the data compression, the required number of CPU cycles per
bit, CPU-cycle frequency, and energy consumption per cycle
follow the uniform distribution with Cn ∈ [0, 3000] cycles/bit,
(c)
fn ∈ [0.1, 1] GHz, and qn ∈ [1, 10] × 10−14 J/cycle, respectively. In addition, the energy reward per bit is distributed
(r)
uniformly as: qn ∈ [1, 10] × 10−12 J/bit. The maximum
0
0
0.5
1
1.5
2
2.5
3
3.5
Effective channel power gain of MS 1
4
4.5
5
10 -5
Figure 4: Impact of effective power channel gain on allocated wireless
power.
compression ratios for lossless and lossy compression are set
as Rmax = 3 and Řmax = 25 with = 4 and ˇ = 0.1,
respectively.
A. Wireless Power Allocation
Numerical results produced by using Proposition 2 are
presented in Fig. 4 to offer insights into the optimal policy
for wireless-power allocation. The effective channel power
gain of sensor 1 varies, while those of sensors 2, · · · , N
are fixed as 10−5 . One can observe that if the channel is
poor, sensor 1 does not receive wireless power due to the
small priority function. Otherwise, if the effective channel
power gain exceeds a threshold (of about 10−6 ), the allocated
wireless power first increases and then decreases. This curious
observation indicates that when the effective channel power
gain is not sufficiently large, the sensor should receive enough
power for increasing its data utility. However, equipped with
a good channel, the sensor can contribute to large data utility
even with smaller received power, which allows for saving
more power for other sensors in order to increase the sum
data utility.
B. Iteration Convergence Rate
The convergence of our proposed algorithm is assessed in
Fig. 5, where the allowed error is set as ξ = 10−5 . Observe
from Fig. 5(a) that the operator’s reward increases monotonically as the iteration evolves and tends to converge after only
a few iterations (about 3 times for lossless compression and 7
for lossy compression). Next, let R(k) denote the operator’s
reward after k-th iteration and R∗ is the maximum operator’s
reward of our proposed algorithm. Define Rgap as the reward
gap between R(k) and R∗ given as Rgap = |R(k) − R∗ |.
Fig. 5(b) displays the curves of the reward gap verses the
number of iterations. We can observe that the gap of our
proposed algorithm almost linearly decreases as the iteration
evolves, thus our proposed solution can be approximated as a
first-order algorithm with linear local convergence rate [39].
C. WPCS Performance
Three baseline policies are considered for our performance
comparison. The first one is the baseline policy with fixed
11
3.85
3.85
3.8
3.8
Operator's reward
Operator's reward
Lossless compression
Lossy compression
3.75
3.75
3.7
3.65
Our proposed policy
Baseline policy with EPA
Baseline policy with FCR
Baseline policy without compression
3.7
3.6
3.65
0
1
2
3
4
5
6
7
8
9
3.55
0.1
10
0.15
Number of iterations
0.3
0.35
0.4
0.45
0.5
(a) Lossless compression
100
3.85
Lossless compression
Lossy compression
10-1
3.8
10-2
Operator's reward
Reward gap
0.25
Maximum amount of transferred energy (J)
(a) Operator’s reward vs. iteration
10-3
10-4
3.75
3.7
3.65
Our proposed policy
Baseline policy with EPA
Baseline policy with FCR
Baseline policy without compression
3.6
10-5
10-6
0.2
0
1
2
3
4
5
6
7
8
9
10
Number of iterations
3.55
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
0.5
Maximum amount of transferred energy (J)
(b) Reward gap vs. iteration
(b) Lossy compression
Figure 5: Convergence of our proposed algorithm.
Figure 6: Operator’s reward vs. maximum amount of transferred
energy.
compression ratio (FCR) that jointly optimizes the power allocation, sensor transmission, and sensing-data size by following
the approach in Section III-C. Specifically, the compression
ratios for lossless and lossy compression are set as Rn = 1.5
and Řn = 4, respectively, which can achieve the maximum operator’s reward for the baseline policy with FCR. The second
one is the baseline policy with equal power allocation (EPA),
which jointly optimizes the sensor transmission, sensing-data
size, and compression ratio given the power allocation policy.
The third one is the baseline policy without compression,
which can be regarded as a special case of the FCR with
the compression ratio set as zero.
The curves for the operator’s reward versus the maximum
amount of transferred energy are displayed in Fig. 6 with
the time duration for crowd sensing set as T = 1 s. For
lossless compression, it can be observed from Fig. 6(a) that
the operator’s reward for our proposed policy is monotoneincreasing with the growth of transferred energy in the small
transferred-energy regime. However, when the amount of
energy exceeds a threshold (of about 0.25 J), the performance
cannot improve further by increasing the transferred energy.
The reason is that the bottleneck in this case is no longer
energy but other settings, such as the crowd-sensing duration.
Note that our proposed policy has significant performance
gain over the policy with FCR and without compression, even
when the maximum amount of transferred energy is sufficient.
This shows the necessity of optimizing the compression ratios
to improve system performance when the transferred energy
is sufficient. In addition, among the three baseline policies,
the one with EPA has the highest operator’s reward and approaches close-to-optimal performance in the large transferred
energy regime. This is due to the fact that optimizing compression ratios can substantially reduce transmission energy
consumption, thus allowing for using more energy to sense
data for better data utility. For lossy compression, one can
observe from Fig. 6(b) that our proposed policy yields higher
operator’s reward with respect to the lossless counterpart for
the case of small and large transferred energy, respectively.
Other observations are similar to those in Fig. 6(a).
Fig. 7 demonstrates the curves of the operator’s reward versus the crowd-sensing duration for both compression methods.
One can observe from Fig. 7(a) that for lossless compression,
the operator’s reward is monotone-increasing with the extension of crowd-sensing duration, as it can reduce transmissionenergy consumption so as to save more energy for data
sensing. However, the growth rate of the operator’s reward
slows down with the crowd-sensing duration. It indicates that
extending a short crowd-sensing duration can substantially
increase the operator’s reward. Moreover, in long crowdsensing duration regime, lossy compression yields higher
operator’s reward than in the lossless case. Other observations
are similar to those in Fig. 6.
Finally, the impact of effective channel power gain on the
compression method selection is evaluated in Fig. 8. Several
observations are made as follows. First, the operator’s reward
is observed to be a monotone-increasing function of the effective channel power gain. Next, lossy and lossless compression
12
3.7
3.6
Operator's reward
3.5
3.4
3.3
3.2
Our proposed policy
Baseline policy with EPA
Baseline policy with FCR
Baseline policy without compression
3.1
3
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Time slot for crowd sensing (s)
showed that it should select the MSs for participating in the
crowd sensing operation based on its configuration and the
individual parameters of MSs. In addition, the optimal power
allocation policy has a threshold-based structure depending
on the derived crowd sensing priority function. For MSs,
each selected sensor should control its compression ratio,
sensor transmission, and sensing-data size for minimizing
the resultant energy consumption. This work can be further
extended to other interesting scenarios, such as cooperation
among multiple operators and online control of multiuser
WPCS systems.
(a) Lossless compression
A PPENDIX
3.8
A. Proof of Lemma 2
3.7
Operator's reward
3.6
This lemma
by contradiction. First, asn can
∗ be proved
o
(s)
∗
∗
sume that
`n
,t ,P
is the optimal solution satisfying
n(s) ∗n
∗
∗
(` )
(s)
t
< ηgn Pn∗ T0 . Then, there exists
αn `n
+ gnn f Rnn t∗
3.5
3.4
3.3
n
3.2
3.1
3
0.1
(1)
another feasible power allocation policy, denoted
Pn ,
by
∗
(`(s)
(s) ∗
(1)
t∗
n )
∗
n
which satisfies Pn < Pn and αn (`n ) + gn f Rn t∗ =
Our proposed policy
Baseline policy with EPA
Baseline policy with FCR
Baseline policy without compression
0.2
0.3
0.4
0.5
0.6
0.7
0.8
n
0.9
1
Time slot for crowd sensing (s)
(b) Lossy compression
(1)
ηgn Pn T0 . It leads to the following
N
X
Figure 7: Operator’s reward vs. time slot for crowd sensing.
∗
an log(1 + (`(s)
n ) )−c
n=1
Operator's reward
Pn(1) T0
n=1
4
>
3.8
N
X
N
X
an log(1 + (`n(s) )∗ ) − c
N
X
Pn∗ T0 ,
n=1
n=1
3.6
which contradicts the
Similarly, if the optimal
∗
optimality.
(s)
solution leads to βn `n
+t∗n < T , then there exists another
3.4
3.2
3
Lossless compression
Lossy compression
2.8
2.6
1
2
3
4
5
6
7
8
Effective channel power gain
9
T −t∗
(s)
(s)
feasible policy with (`n )(1) that satisfies (`n )(1) = βn n >
∗
(s)
. This new policy yields higher operator’s reward as
`n
10
10-4
Figure 8: Operator’s reward vs. effective channel power gain.
is preferred for relatively small and large effective channel
power gains, respectively. The reason is that for poor channels,
transmission-energy can dominate other components (e.g.,
sensing), and it can be substantially reduced by lossy compression due to efficient data-compression performance. However,
when the channel is good enough, lossless compression can
contribute to higher data utility with small compression energy,
and thereby may be preferred over the lossy operation.
N
X
an log(1 + (`n(s) )(1) ) − c
n=1
This work proposed a multiuser WPCS system model
and investigated a joint control of power allocation, sensing,
compression, and transmission for maximizing the operator’s
reward. It offered several insights into the control policies of
the operator and the MSs. To be specific, for the operator, we
Pn∗ T0
n=1
>
N
X
an log(1 + `n(s)
∗
n=1
)−c
N
X
Pn∗ T0 ,
n=1
thus contradicting the assumption. Combining the above together leads to the desired result.
B. Proof of Lemma 3
First, it is easy to see that
VI. C ONCLUDING R EMARKS
N
X
concave and
PN
n=1
an log 1 +
T −tn
βn
is
αn (T −tn )
is linear for tn . Then, the second
ηgn
βn
T −tn
(T ln 2)2
tn
T −tn
BRn βn tn > 0
f
is η(gNn0BR
2
2 3 2
ηgn
Rn βn tn
n β n ) tn
derivative of
for 0 < tn < T ; thus, it is a convex function. It follows
that the objective function, being a summation over a set of
convex functions, preserves the convexity. Combining it with
the convex constraints yields the desired result.
13
C. Proof of Corollary 1
E. Proof of Lemma 5
From (13), the utility weight can be expressed as:
T ln 2
ln 2
αn gn
T ln 2
− BR
BRn βn t∗
n βn +
n
an =
−
1
e
+
1
BRn βn t∗n
N0 βn
(λ∗ + c)N0 [βn + (T − t∗n )]
.
×
ηgn2
Given a1 < a2 , it can be directly observed from the above
equation that t1 > t2 . As for the effective channel power gain,
it should satisfy the following equation:
an ηgn
αn
gn
−
(λ∗ + c)N0 [βn + (T − t∗n )] N0 βn
T ln 2
T ln 2
− ln 2
−
1
e BRn βn t∗n BRn βn + 1. (28)
=
∗
BRn βn tn
Note that both the right hand-side of (28) and N0 [βn +(T −t∗n )]
are decreasing with tn . It means that if g1 < g2 , it should be
satisfied that t1 ≥ t2 , thus completing the proof.
D. Proof of Proposition 2
∗
(s)
First, combing αn `n
+
∗
(s)
and βn `n
+ t∗n = T gives
Pn∗ =
t∗n
f
gn
1
ηgn T0
t∗
n
gn f
T − t∗n
Rn βn t∗n
∗
(`(s)
n )
Rn t∗
n
+
= ηgn Pn∗ T0
αn (T − t∗n )
.
βn
Next, the first and the second derivative of Pn are calculated
as follows.
T −tn
αn
N0
T ln 2
∂Pn
= 2
1−
2 BRn βn tn −1 −
,
∂tn ηgn T0
BRn βn tn
ηβn gn T0
2
T −tn
∂ 2 Pn
N0 (T ln 2)
=
2 BRn βn tn .
∂t2n
ηT0 (gn BRn βn )2 t3n
2
It can be observed that ∂∂tP2n > 0 for 0 ≤ tn ≤ T . Hence,
n
∂Pn
∂tn is monotone-increasing for 0 ≤ tn ≤ T and
∂Pn
∂tn
=
max
∂Pn
∂tn
=−
tn =T
Nn ln 2
αn
−
< 0.
ηgn2 BRn βn T0 ηβn gn T0
It means that Pn is monotone decreasing with tn and
(Pn )min = Pn |tn =T = 0. Combining (12a) and 0 ≤ tn ≤ T
yields
an ηgn
βn +(T −t∗
n)
λ∗ >
αn
βn
=
−
N0
gn
1−
T ln 2
BRn βn t∗
n
2
T −t∗
n
BRn βn t∗
n
an ηgn
(r)
qn
+
(s)
qn
(c)
+ qn eRn +
N0 ln 2
gn BRn
−c
−1
− c = ϕn .
If λ∗ < ϕn , the optimal sensing time will violate the time
constraint t∗n > T ; thus, the sensor will not be selected and no
power will be allocated to it i.e., Pn = 0, which completes
the proof.
Based on (20a), for Rn∗ = 1, it can be established
(s)
`n
b∗
(c)
2 t∗
λ
nB −
that Ng0nln
2
< 0. Combining it
q
+
n
B
fn e
(s)
`
b∗ =
with" the optimal conditions,
t∗n = T − snn and λ
#
(s)
`(s)
n
`n ln 2
N0
−
1
2 t∗n B + 1 , leads to the desired result.
∗
gn
t B
n
F. Proof of Proposition 4
b∗ = 0
Based on (20c), atleast one of the two equations λ
∗
or
t∗n
+
`(s)
eRn −e
n
fn
`(s)
+ snn −
∗
`(s)
eRn −e
n
T = 0 should be satisfied.
`(s)
Assume that t∗n +
+ snn − T 6= 0, then it has
fn
∗
∗
b = 0, which leads to R → ∞ according to (20b). This,
λ
n
however, contradicts the finite compression
ratio. Hence, it
∗
`(s)
eRn −e
n
`(s)
should satisfy: t∗n +
+ snn = T . The optimal
fn
solution should satisfy (20a) as
!
!
(s)
b∗
∗
λ
`n
1
(c)
0
Rn
qn +
f
= 0.
e
−
fn
gn (Rn∗ )2
t∗n Rn∗
b∗ =
Combining the above with the optimal conditions, λ
∗
(s)
(s)
Rn
(s)
`
e
−e
n
`
`
− snn according
− g1n g t∗nR∗ and t∗n = T −
fn
n n
to (20b) and (20c), produces the desired result.
G. Proof of Lemma 6
The first derivative of function z(Rn ) is
∂z(Rn ) N0 ln 2 2
ln 2[Rn d(Rn )]0
Rn d0 (Rn )
=
+
1−
∂Rn
BRn2 gn Rn
B[Rn d(Rn )]2
d(Rn )
ln 2
1
1
(c)
BR
d(R
)
n
n
×e
+ qn −
g
2 eRn.
gn fn
d(Rn )Rn
0
First, it has [Rn d(Rn )] = d(Rn ) − Rn d0 (Rn ). Hence,
[Rn d(Rn )]0 > 0 when d(Rn ) > Rn d0 (Rn ) and [Rn d(Rn )]0 <
0 when d(Rn ) < Rn d0 (Rn ). Combining these together leads
to:
Rn d0 (Rn )
ln 2[Rn d(Rn )]0
1
−
> 0, for Rn ∈ [1, Rmax ].
B[Rn d(Rn )]2
d(Rn )
Further, the first derivative of function g(x) is
g 0 (x) = −xf 00 (x) = −
N0 ln 22 x
2 B < 0, for x > 0.
B2
It means that g(x) is monotone-decreasing with x and thus
g(x) ≤ g(x)|x=0 = 0 for x > 0. Since d(Rn1)Rn > 0, it has
h
i
(c)
qn − gn1fn g d(Rn1)Rn
> 0 for x > 0. Combining the
above considerations leads to
thus completing the proof.
∂z(Rn )
∂Rn
> 0 for Rn ∈ [1, Rmax ],
14
H. Proof of Corollary 5
(c)
qn gn .
Define a function ϕ(Rn ) as ϕ(Rn ) =
According to
(23), it has
1
1
1
1
ϕ(Rn ) =
g
+ 2 Rn f 0
.
fn
d(Rn )Rn
Rn e
d(Rn )Rn
The derivation of ϕ(Rn ) is
ln 2[Rn d(Rn )]0
Rn eRn
2+
∂ϕ(Rn )
+
1−
=−
∂Rn
Rn eRn
B[Rn d(Rn )]2
fn d(Rn )
2
N0 ln 2 BR lnd(R
×
e n n) .
BRn2 eRn
2[Rn d(Rn )]0
Rn eRn
Since lnB[R
1
−
is always positive as
2
fn d(Rn )
n d(Rn )]
proved in Appendix G, the function ϕ(Rn ) is monotone(c)
(c)
decreasing for Rn ∈ [1, Rmax ]. Given q1 > q2 , it can be
derived that R1 < R2 . Using a similar approach, it can be
proved that R1 < R2 when g1 > g2 , thus completing the proof.
I. Proof of Lemma 7
The first derivative of ž(rn ) is
ˇ n )]0
ln 2
N0 ln 2 2
ln 2[rn d(r
rn dˇ0 (rn )
+
e Brn ď(rn ) +
1
−
2
2
ˇ
ˇ
Bgn rn rn
B[rn d(rn )]
d(rn )
2
1
(6 + 4rn2 )rn rn2
(6 + 4rn2 )rn qn(c) ern − g
e .
ˇ
gn fn
d(rn )rn
ˇ n ) is d(r
ˇ n ) − rn dˇ0 (rn ). Since
The derivative of function rn d(r
0
0
0
ˇ
ˇ
ˇ
ˇ
[rn d(rn )] > 0 when d(rn ) > rn d (rn ) and [r
n d(rn )]ˇ0 < 0
ˇ n )]0
d
(rn )
r
d(r
ln
2[r
n
n
0
ˇ n ) < rn dˇ (rn ), the item
1 − d(r
when d(r
ˇ n )]2
ˇ n)
B[rn d(r
is always positive. Moreover, g(x) < 0 when x > 0
1
according to Appendix G. Since d(r
ˇ n )rn > 0, the item
h
i
2
2
2
(c)
(6+4rn
)rn rn
1
(6 + 4rn2 )rn qn ern − g d(r
e
is alˇ
gn fn
)r
n
ways positive. Hence,
completing the proof.
∂ Ž(rn )
∂rn
n
> 0 for rn ∈ [1, rmax ], thus
R EFERENCES
[1] I. F. Akyildiz, W. Su, Y. Sankarasubramaniam, and E. Cayirci, “Wireless
sensor networks: A survey,” Comput. netw., vol. 38, no. 4, pp. 393–422,
2002.
[2] R. K. Ganti, F. Ye, and H. Lei, “Mobile crowdsensing: Current state
and future challenges,” IEEE Commun. Mag., vol. 49, no. 11, pp. 32–
39, 2011.
[3] O. Galinina, K. Mikhaylov, K. Huang, S. Andreev, and Y. Koucheryavy,
“Wirelessly powered urban crowd sensing over wearables: Trading energy for data,” [Online]. Available: https://arxiv.org/pdf/1611.05910.pdf.
[4] H. Ma, D. Zhao, and P. Yuan, “Opportunities in mobile crowd sensing,”
IEEE Commun. Mag., vol. 52, no. 8, pp. 29–35, 2014.
[5] V. Petrov, A. Samuylov, V. Begishev, D. Moltchanov, S. Andreev,
K. Samouylov, and Y. Koucheryavy, “Vehicle-based relay assistance
for opportunistic crowdsensing over narrowband IoT (NB-IoT),” IEEE
Internet Things J., vol. PP, no. 99, pp. 1–1, 2017.
[6] X. Zhang, Z. Yang, W. Sun, Y. Liu, S. Tang, K. Xing, and X. Mao,
“Incentives for mobile crowd sensing: A survey,” IEEE Commun.
Surveys Tuts., vol. 18, no. 1, pp. 54–67, 2016.
[7] Z. Feng, Y. Zhu, Q. Zhang, L. M. Ni, and A. V. Vasilakos, “TRAC:
Truthful auction for location-aware collaborative sensing in mobile
crowdsourcing,” in Proc. IEEE Infocom, 2014, pp. 1231–1239.
[8] L. Duan, T. Kubo, K. Sugiyama, J. Huang, T. Hasegawa, and J. Walrand,
“Incentive mechanisms for smartphone collaboration in data acquisition
and distributed computing,” in Proc. IEEE Infocom, 2012, pp. 1701–
1709.
[9] S. Yang, F. Wu, S. Tang, X. Gao, B. Yang, and G. Chen, “On designing
data quality-aware truth estimation and surplus sharing method for
mobile crowdsensing,” IEEE J. Sel. Areas Commun., vol. 35, no. 4,
pp. 832–847, 2017.
[10] M. A. Alsheikh, D. Niyato, D. Leong, P. Wang, and Z. Han, “Privacy
management and optimal pricing in people-centric sensing,” IEEE J. Sel.
Areas Commun., vol. 35, no. 4, pp. 906–920, 2017.
[11] X. Lu, P. Wang, D. Niyato, D. I. Kim, and Z. Han, “Wireless networks
with RF energy harvesting: A contemporary survey,” IEEE Commun.
Surveys Tuts., vol. 17, no. 2, pp. 757–789, 2015.
[12] Y. Zeng, B. Clerckx, and R. Zhang, “Communications and signals design
for wireless power transmission,” IEEE Trans. Commun., vol. 65, no. 5,
pp. 2264–2290, 2016.
[13] A. Ghazanfari, H. Tabassum, and E. Hossain, “Ambient RF energy
harvesting in ultra-dense small cell networks: Performance and tradeoffs,” IEEE Wireless Commun., vol. 23, no. 2, pp. 38–45, 2016.
[14] R. Zhang and C. K. Ho, “MIMO broadcasting for simultaneous wireless information and power transfer,” IEEE Trans. Wireless Commun.,
vol. 12, no. 5, pp. 1989–2001, 2013.
[15] J. Park and B. Clerckx, “Joint wireless information and energy transfer
in a two-user MIMO interference channel,” IEEE Trans. Wireless
Commun., vol. 12, no. 8, pp. 4210–4221, 2013.
[16] P. Popovski, A. M. Fouladgar, and O. Simeone, “Interactive joint transfer
of energy and information,” IEEE Trans. Commun., vol. 61, no. 5, pp.
2086–2097, 2013.
[17] Z. Ding, S. M. Perlaza, I. Esnaola, and H. V. Poor, “Power allocation
strategies in energy harvesting wireless cooperative networks,” IEEE
Trans. Wireless Commun., vol. 13, no. 2, pp. 846–860, 2014.
[18] Z. Ding, I. Krikidis, B. Sharif, and H. V. Poor, “Wireless information and
power transfer in cooperative networks with spatially random relays,”
IEEE Trans. Wireless Commun., vol. 13, no. 8, pp. 4440–4453, 2014.
[19] Z. Zheng, L. Song, D. Niyato, and Z. Han, “Resource allocation in
wireless powered relay networks: A bargaining game approach,” IEEE
Trans. Vehi. Techn., vol. 66, no. 7, pp. 6310–6323, 2017.
[20] D. W. K. Ng, E. S. Lo, and R. Schober, “Wireless information and
power transfer: Energy efficiency optimization in OFDMA systems,”
IEEE Trans. Wireless Commun., vol. 12, no. 12, pp. 6352–6370, 2013.
[21] S. Lee, R. Zhang, and K. Huang, “Opportunistic wireless energy
harvesting in cognitive radio networks,” IEEE Trans. Wireless Commun.,
vol. 12, no. 9, pp. 4788–4799, 2013.
[22] D. W. K. Ng, E. S. Lo, and R. Schober, “Multiobjective resource
allocation for secure communication in cognitive radio networks with
wireless information and power transfer,” IEEE Trans. Veh. Technol.,
vol. 65, no. 5, pp. 3166–3184, 2016.
[23] C.-F. Liu, M. Maso, S. Lakshminarayana, C.-H. Lee, and T. Q. Quek,
“Simultaneous wireless information and power transfer under different
CSI acquisition schemes,” IEEE Trans. Wireless Commun., vol. 14, no. 4,
pp. 1911–1926, 2015.
[24] M. Gregori and M. Payaró, “Energy-efficient transmission for wireless
energy harvesting nodes,” IEEE Trans. Wireless Commun., vol. 12, no. 3,
pp. 1244–1254, 2013.
[25] Y. Zeng, R. Zhang, and T. J. Lim, “Throughput maximization for
UAV-enabled mobile relaying systems,” IEEE Trans. Commun., vol. 64,
no. 12, pp. 4983–4996, 2016.
[26] C. You, K. Huang, and H. Chae, “Energy efficient mobile cloud
computing powered by wireless energy transfer,” IEEE J. Sel. Areas
Commun., vol. 34, no. 5, pp. 1757–1771, 2016.
[27] L. Xie, Y. Shi, Y. T. Hou, and A. Lou, “Wireless power transfer and
applications to sensor networks,” IEEE Wireless Commun., vol. 20, no. 4,
pp. 140–145, 2013.
[28] K. W. Choi, L. Ginting, P. A. Rosyady, A. A. Aziz, and D. I. Kim,
“Wireless-powered sensor networks: How to realize,” IEEE Trans.
Wireless Commun., vol. 16, no. 1, pp. 221–234, 2017.
[29] J. Xu, Z. Zhong, and B. Ai, “Wireless powered sensor networks: Collaborative energy beamforming considering sensing and circuit power
consumption,” IEEE Commun. Lett., vol. 5, no. 4, pp. 344–347, 2016.
[30] S. Guo, F. Wang, Y. Yang, and B. Xiao, “Energy-efficient cooperative
transmission for simultaneous wireless information and power transfer
in clustered wireless sensor networks,” IEEE Trans. Commun., vol. 63,
no. 11, pp. 4405–4417, 2015.
[31] K. Han and K. Huang, “Wirelessly powered backscatter communication
networks: modeling, coverage, and capacity,” IEEE Trans. Wireless
Commun., vol. 16, no. 4, pp. 2548–2561, 2017.
[32] M. Bhardwaj, T. Garnett, and A. P. Chandrakasan, “Upper bounds on
the lifetime of sensor networks,” in Proc. IEEE ICC, 2001, pp. 785–790.
15
[33] A. V. D. Ven, “Linux OS data compression options: Comparing behavior,” https://clearlinux.org/blogs/linux-os-data-compression-optionscomparing-behavior, 2017.
[34] A. P. Chandrakasan, S. Sheng, and R. W. Brodersen, “Low-power CMOS
digital design,” IEEE J. Solid-State Circuits, vol. 27, no. 4, pp. 473–484,
1992.
[35] C. J. V. den Branden Lambrecht and O. Verscheure, “Perceptual quality
measure using a spatiotemporal model of the human visual system,” in
Proc. SPIE, vol. 2668, 1996, pp. 450–461.
[36] D. Yang, G. Xue, X. Fang, and J. Tang, “Crowdsourcing to smartphones:
Incentive mechanism design for mobile phone sensing,” in Proc. MobiCom, 2012, pp. 173–184.
[37] S. Boyd and L. Vandenberghe, Convex Optimization.
Cambridge
University Press, 2004.
[38] A. Thiagarajan, L. Ravindranath, H. Balakrishnan, S. Madden, and
L. Girod, “Accurate, low-energy trajectory mapping for mobile devices,”
in Proc. NSDI, 2011, pp. 267–280.
[39] M. Schatzman, Numerical analysis: A mathematical introduction. Oxford University Press, 2002.
| 7cs.IT
|
On Asymptotic Inference in Stochastic Differential Equations
with Time-Varying Covariates
Trisha Maitra and Sourabh Bhattacharya∗
Abstract
arXiv:1605.03330v2 [math.ST] 13 Oct 2017
In this article, we introduce a system of stochastic differential equations (SDEs) consisting of
time-dependent covariates and consider both fixed and random effects set-ups. We also allow the
functional part associated with the drift function to depend upon unknown parameters. In this general set-up of SDE system we establish consistency and asymptotic normality of the M LE through
verification of the regularity conditions required by existing relevant theorems. Besides, we consider the Bayesian approach to learning about the population parameters, and prove consistency and
asymptotic normality of the corresponding posterior distribution. We supplement our theoretical investigation with simulated and real data analyses, obtaining encouraging results in each case.
Keywords: Asymptotic normality; Functional data; Gibbs sampling; Maximum likelihood estimator; Posterior consistency; Random effects.
1
Introduction
Systems of stochastic differential equations (SDEs) are appropriate for modeling situations where
“within” subject variability is caused by some random component varying continuously in time; hence,
SDE systems are also appropriate for modeling functional data (see, for example, Zhu et al. (2011),
Ramsay and Silverman (2005) for some connections between SDE and functional data analysis). When
suitable time-varying covariates are available, it is then appropriate to incorporate such information in
the SDE system. Some examples of statistical applications of SDE-based models with time-dependent
covariates are Oravecz et al. (2011), Overgaard et al. (2005), Leander et al. (2015).
However, systems of SDE based models consisting of time-varying covariates seem to be rare in the
statistical literature, in spite of their importance, and their asymptotic properties are hitherto unexplored.
Indeed, although asymptotic inference in single, fixed effects SDE models without covariates has been
considered in the literature as time tends to infinity (see, for example, Bishwal (2008)), asymptotic theory
in systems of SDE models is rare, and so far only random effects SDE systems without covariates have
been considered, as n, the number of subjects (equivalently, the number of SDEs in the system), tends
to infinity (Delattre et al. (2013), Maitra and Bhattacharya (2016c), Maitra and Bhattacharya (2015)).
Such models are of the following form:
dXi (t) = b(Xi (t), φi )dt + σ(Xi (t))dWi (t),
with Xi (0) = xi , i = 1, . . . , n,
(1.1)
where, for i = 1, . . . , n, Xi (0) = xi is the initial value of the stochastic process Xi (t), which is
assumed to be continuously observed on the time interval [0, Ti ]; Ti > 0 assumed to be known. The
function b(x, φ), which is the drift function, is a known, real-valued function on R × Rd (R is the real
line and d is the dimension), and the function σ : R 7→ R is the known diffusion coefficient. The
SDEs given by (1.1) are driven by independent standard Wiener processes {Wi (·); i = 1, . . . , n}, and
{φi ; i = 1, . . . , n}, which are to be interpreted as the random effect parameters associated with the n
individuals, which are assumed by Delattre et al. (2013) to be independent of the Brownian motions and
independently and identically distributed (iid) random variables with some common distribution.
For the sake of convenience Delattre et al. (2013) (see also Maitra and Bhattacharya (2016c) and
Maitra and Bhattacharya (2015)) assume b(x, φi ) = φi b(x). Thus, the random effect is a multiplicative
factor of the drift function; also, the function b(x) is assumed to be independent of parameters. In this
article, we generalize the multiplicative factor to include time-dependent covariates; we also allow b(x)
to depend upon unknown parameters.
∗
Trisha Maitra is a PhD student and Sourabh Bhattacharya is an Associate Professor in Interdisciplinary Statistical Research
Unit, Indian Statistical Institute, 203, B. T. Road, Kolkata 700108. Corresponding e-mail: [email protected].
1
Notably, such model extension has already been provided in Maitra and Bhattacharya (2016a) and
Maitra and Bhattacharya (2016b), but their goal was to develop asymptotic theory of Bayes factors for
comparing systems of SDEs, with or without time-varying covariates, emphasizing, when time-varying
covariates are present, simultaneous asymptotic selection of covariates and part of the drift function free
of covariates, using Bayes factors.
In this work, we deal with parametric asymptotic inference, both frequentist and Bayesian, in the
context of our extended system of SDEs. We consider, separately, fixed effects as well as random
effects. The fixed effects set-up ensues when coefficients associated with the covariates are the same
for all the subjects. On the other hand, in the random effects set-up, the subject-wise coefficients are
assumed to be a random sample from some distribution with unknown parameters.
It is also important to distinguish between the iid situation and the independent but non-identical
case (we refer to the latter as non-iid) that we consider. The iid set-up is concerned with the case where
the initial values xi and time limit Ti are the same for all i, and the coefficients associated with the
covariates are zero, that is, there are no covariates suitable for the SDE-based system. This set-up,
however, does not reduce to the iid set-up considered in Delattre et al. (2013), Maitra and Bhattacharya
(2016c) and Maitra and Bhattacharya (2015) because in the latter works b(x) was assumed to be free
of parameters, while in this work we allow this function to be dependent on unknown parameters. The
non-iid set-up assumes either or both of the following: presence of appropriate covariates and that xi
and Ti are not the same for all the subjects.
In the classical paradigm, we investigate consistency and asymptotic normality of the maximum
likelihood estimator (M LE) of the unknown parameters which we denote by θ, and in the Bayesian
framework we study consistency and asymptotic normality of the Bayesian posterior distribution of θ.
In other words, we consider prior distributions π(θ) of θ and study the properties of the corresponding
posterior
Q
π(θ) ni=1 fi (Xi |θ)
Qn
πn (θ|X1 , . . . , Xn ) = R
(1.2)
i=1 fi (Xi |ψ)dψ
ψ∈Θ π(ψ)
as the sample size n tends to infinity. Here fi (·|θ) is the density corresponding to the i-th individual and
Θ is the parameter space.
In what follows, after introducing our model and the associated likelihood in Section 2, we investigate asymptotic properties of M LE in the iid and non-iid contexts in Sections 3 and 4 respectively.
Then, in Sections 5 and 6 we investigate asymptotic properties of the posterior in the iid and non-iid
cases, respectively. In Section 7 we consider the random effects set-up and provide necessary discussion
to point towards validity of the corresponding asymptotic results. We demonstrate the applicability of
our developments to practical and finite-sample contexts using simulated and real data analyses in Sections 8 and 9, respectively. We summarize our contribution and provide further discussion in Section
10.
a.s.
P
L
Notationally, “ → ”, “→” and “→” denote convergence “almost surely”, “in probability” and “in
distribution”, respectively.
2
The SDE set-up
We consider the following system of SDE models for i = 1, 2, . . . , n:
dXi (t) = φi,ξ (t)bβ (Xi (t))dt + σ(Xi (t))dWi (t)
(2.1)
where Xi (0) = xi is the initial value of the stochastic process Xi (t), which is assumed to be continuously observed on the time interval [0, Ti ]; Ti > 0 for all i and assumed to be known. In the above, φi,ξ
is the parametric function consisting of the covariates and the unknown coefficients ξ associated with
the covariates, and bβ is a parametric function known up to the parameters β.
2
2.1
Incorporation of time-varying covariates
We assume that φi,ξ (t) has the following form:
φi,ξ (t) = φi,ξ (z i (t)) = ξ0 + ξ1 g1 (zi1 (t)) + ξ2 g2 (zi2 (t)) + · · · + ξp gp (zip (t)),
(2.2)
where ξ = (ξ0 , ξ1 , . . . , ξp ) is a set of real constants and z i (t) = (zi1 (t), zi2 (t), . . . , zip (t)) is the set of
available covariate information corresponding to the i-th individual, depending upon time t. We assume
that z i (t) is continuous in t, zil (t) ∈ Z l where Z l is compact and gl : Z l → R is continuous, for
l = 1, . . . , p. Let Z = Z 1 × · · · × Z p . We let
Z = {z(t) ∈ Z : t ∈ [0, ∞) such that z(t) is continuous in t} .
Hence, z i ∈ Z for all i. The function bβ is multiplicative part of the drift function free of the covariates.
Note that ξ consists of p + 1 parameters. Assuming that β ∈ Rq , where q ≥ 1, it follows that our
parameter set θ = (β, ξ) belongs to the (p + q + 1)-dimensional real space Rp+q+1 . The true parameter
set is denoted by θ 0 .
2.2
Likelihood
We first define the following quantities:
Z
Ui,θ =
0
Ti
φi,ξ (s)bβ (Xi (s))
dXi (s),
σ 2 (Xi (s))
Z
Ti
Vi,θ =
0
φ2i,ξ (s)b2β (Xi (s))
σ 2 (Xi (s))
ds
(2.3)
for i = 1, . . . , n.
Let C Ti denote the space of real continuous functions (x(t), t ∈ [0, Ti ]) defined on [0, Ti ], endowed
with the σ-field CTi associated with the topology of uniform convergence on [0, Ti ]. We consider the
distribution P xi ,Ti ,zi on (CTi , CTi ) of (Xi (t), t ∈ [0, Ti ]) given by (2.1) We choose the dominating
measure Pi as the distribution of (2.1) with null drift. So,
Vi,θ
dP xi ,Ti ,zi
.
(2.4)
= fi (Xi |θ) = exp Ui,θ −
dPi
2
3
Consistency and asymptotic normality of M LE in the iid set-up
In the iid set up we have xi = x and Ti = T for all i = 1, . . . , n. Moreover, the covariates are absent,
that is, ξi = 0 for i = 1, . . . , p. Hence, the resulting parameter set in this case is θ = (β, ξ0 ).
3.1
Strong consistency of M LE
Consistency of the M LE under the iid set-up can be verified by validating the regularity conditions of
the following theorem (Theorems 7.49 and 7.54 of Schervish (1995)); for our purpose we present the
version for compact parameter space.
Theorem 1 (Schervish (1995)) Let {Xn }∞
n=1
be conditionally iid given θ with density f1 (x|θ) with
respect to a measure ν on a space X 1 , B 1 . Fix θ 0 ∈ Θ, and define, for each M ⊆ Θ and x ∈ X 1 ,
Z(M, x) = inf log
ψ∈M
f1 (x|θ 0 )
.
f1 (x|ψ)
Assume that for each θ 6= θ 0 , there is an open set Nθ such that θ ∈ Nθ and that Eθ0 Z(Nθ , Xi ) > −∞.
Also assume that f1 (x|·) is continuous at θ for every θ, a.s. [Pθ0 ]. Then, if θ̂ n is the M LE of θ
corresponding to n observations, it holds that lim θ̂ n = θ 0 , a.s. [Pθ0 ].
n→∞
3
3.1.1
Assumptions
We assume the following conditions:
(H1) The parameter space Θ = B × Γ such that B and Γ are compact.
(H2) bβ (·) and σ(·) are C 1 (differentiable with continuous first derivative) on R and satisfy b2β (x) ≤
K1 (1 + x2 + kβk2 ) and σ 2 (x) ≤ K2 (1 + x2 ) for all x ∈ R, for some K1 , K2 > 0. Now, due to
(H1) the latter boils down to assuming b2β (x) ≤ K(1 + x2 ) and σ 2 (x) ≤ K(1 + x2 ) for all x ∈ R,
for some K > 0.
We further assume:
(H3) For every x, let bβ be continuous in β = (β1 , . . . , βq ) and moreover, for j = 1, . . . , q,
sup
β∈B
∂bβ (x)
∂βj
σ 2 (x)
≤ c (1 + |x|γ ) ,
for some c > 0 and γ ≥ 0.
(H4)
b2β (x)
σ 2 (x)
≤ Kβ 1 + x2 + kβk2 ,
(3.1)
where Kβ is continuous in β.
3.1.2
Verification of strong consistency of M LE in our SDE set-up
To verify the conditions of Theorem 1 in our case, note that assumptions (H1) – (H4) clearly imply
continuity of the density f1 (x|θ) in the same way as the proof of Proposition 2 of Delattre et al. (2013).
It follows that Uθ and Vθ are continuous in θ, the property that we use in our proceedings below.
Now consider,
f1 (X|θ 0 )
Z(Nθ , X) = inf log
θ 1 ∈Nθ
f1 (X|θ 1 )
Vθ 0
Vθ 1
− inf
Uθ1 −
= Uθ0 −
θ 1 ∈Nθ
2
2
Vθ
Vθ
≥ Uθ0 − 0 − inf
Uθ1 − 1
2
2
θ 1 ∈N̄θ
Vθ∗1 (X)
Vθ 0
= Uθ0 −
− Uθ∗1 (X) −
,
2
2
(3.2)
where Nθ is an appropriate open
subset ofthe relevant compact parameter space, and N̄θ is a closed
V
subset of Nθ . The infimum of Uθ1 − 2θ1 is attained at θ ∗1 = θ ∗1 (X) ∈ N̄θ due to continuity of Uθ
and Vθ in θ.
Let Eθ0 (Vθ1 ) = V̆θ1 and Eθ0 (Uθ1 ) = Ŭθ1 . From Theorem 5 of Maitra and Bhattacharya (2016c) it
4
follows that the above expectations are continuous in θ 1 . Using this we obtain
Vθ∗1 (X)=ϕ1
Vθ∗1 (X)
= Eθ∗1 (X)|θ0 EX|θ∗1 (X)=ϕ1 ,θ0 Uθ∗1 (X)=ϕ1 −
Eθ0 Uθ∗1 (X) −
2
2
!
V̆ϕ
= Eθ∗1 (X)|θ0 Ŭϕ1 − 1
2
!#
"
V̆ϕ1
≤ Eθ∗1 (X)|θ0 sup
Ŭϕ1 −
2
ϕ1 ∈N̄θ
!
V̆ϕ∗
= Eθ∗1 (X)|θ0 Ŭϕ∗1 − 1
2
!
V̆ϕ∗
= Ŭϕ∗1 − 1 ,
2
where
ϕ∗1
∈ N̄θ is where the supremum of Ŭϕ1 −
the last step (3.3) follows.
Noting that Eθ0 Uθ0 −
Vθ 0
2
and Ŭϕ∗1 −
V̆ϕ∗
1
2
V̆ϕ1
2
(3.3)
is achieved. Since ϕ∗1 is independent of X,
are finite due to Lemma 1 of Maitra and Bhata.s.
tacharya (2016a), it follows that Eθ0 Z(Nθ , X) > −∞. Hence, θ̂ n → θ 0 [Pθ0 ], as n → ∞. We
summarize the result in the form of the following theorem:
Theorem 2 Assume the iid setup and conditions (H1) – (H4). Then the M LE is strongly consistent in
a.s.
the sense that as n → ∞, θ̂ n → θ 0 [Pθ0 ].
3.2
Asymptotic normality of M LE
To verify asymptotic normality of M LE we invoke the following theorem provided in Schervish (1995)
(Theorem 7.63):
Theorem 3 (Schervish (1995)) Let Θ be a subset of Rp+q+1 , and let {Xn }∞
n=1 be conditionally iid
P
given θ each with density f1 (·|θ). Let θ̂ n be an M LE. Assume that θ̂ n → θ under Pθ for all θ. Assume
that f1 (x|θ) has continuous second partial derivatives with respect to θ and that differentiation can be
passed under the integral sign. Assume that there exists Hr (x, θ) such that, for each θ 0 ∈ int(Θ) and
each k, j,
sup
kθ−θ 0 k≤r
∂2
∂2
log f1 (x|θ 0 ) −
log f1 (x|θ) ≤ Hr (x, θ 0 ),
∂θk ∂θj
∂θk ∂θj
(3.4)
with
lim Eθ0 Hr (X, θ 0 ) = 0.
r→0
Assume that the Fisher information matrix I(θ) is finite and non-singular. Then, under Pθ0 ,
√
L
n θ̂ n − θ 0 → N 0, I −1 (θ 0 ) .
3.2.1
Assumptions
Along with the assumptions (H1) – (H4), we further assume the following:
(H5) The true value θ 0 ∈ int (Θ).
(H6) The Fisher’s information matrix I(θ) is finite and non-singular, for all θ ∈ Θ.
5
(3.5)
(3.6)
h
i
h
i
2
= ∂β∂k ∂βl bβ (x) for k, l = 1, . . . , q,
(H7) Letting b0β (x) = ∂β∂ k bβ (x) for k = 1, . . . , q; b00β (x)
kl
h
i k
∂3
=
and b000
(x)
b
(x)
for
k,
l,
m
=
1,
.
.
.
,
q,
there exist constants 0 < c < ∞, 0 <
β
β
∂βk ∂βl ∂βm
klm
γ1 , γ2 , γ3 , γ4 ≤ 1 such that for each combination of k, l, m = 1, . . . , q, for any β 1 , β 2 ∈ Rq , for
all x ∈ R,
bβ1 (x) − bβ2 (x) ≤ c k β 1 − β 2 kγ1 ;
h
i
h
i
≤ c k β 1 − β 2 kγ2 ;
b0β1 (x) − b0β2 (x)
k
k
h
i
h
i
≤ c k β 1 − β 2 kγ3 ;
b00β1 (x) − b00β2 (x)
kl
kl
h
i
h
i
000
≤ c k β 1 − β 2 kγ4 .
−
b
(x)
b000
(x)
β2
β1
klm
klm
3.2.2
Verification of the above regularity conditions for asymptotic normality in our SDE set-up
P
In Section 3.1.2 almost sure consistency of the M LE θ̂ n has been established. Hence, θ̂ n → θ under
Pθ for all θ. With assumptions (H1)–(H4), (H7), Theorem B.4 of Rao (2013) and the dominated convergence theorem, interchangability of differentiation and integration in case of stochastic integration and
usual integration respectively can be assured, from which it can be easily deduced that differentiation can
be passed under the integral sign, as required by Theorem 3. With the same arguments, it follows that in
2
our case ∂θ∂k ∂θj log f1 (x|θ) is differentiable in θ = (β, ξ0 ), and the derivative has finite expectation due
to compactness of the parameter space and (H7). Hence, (3.4) and (3.5) clearly hold.
In other words, asymptotic normality of the M LE, of the form (3.6), holds in our case. Formally,
Theorem 4 Assume the iid setup and conditions (H1) – (H7). Then, as n → ∞, the M LE is asymptotically normally distributed as (3.6).
4
Consistency and asymptotic normality of M LE in the non-iid set-up
We now consider the case where the processes Xi (·); i = 1, . . . , n, are independently, but not identically
distributed. In this case, ξ = (ξ0 , ξ1 , . . . , ξp ) where at least one of the coefficients ξ1 , . . . , ξp is non-zero,
guaranteeing the presence of at least one time-varying covariate. Hence, in this set-up θ = (β, ξ).
Moreover, following Maitra and Bhattacharya (2016c), Maitra and Bhattacharya (2015) we allow
the initial values xi and the time limits Ti to be different for i = 1, . . . , n, but assume that the sequences
{T1 , T2 , . . .} and {x1 , x2 , . . .} are sequences entirely contained in compact sets T and X, respectively.
Compactness ensures that there exist convergent subsequences with limits in T and X; for notational
convenience, we continue to denote the convergent subsequences as {T1 , T2 , . . .} and {x1 , x2 , . . .}.
Thus, let the limts be T ∞ ∈ T and x∞ ∈ X.
Henceforth, we denote the process associated with the initial value x and time point t as X(t, x) and
so for x ∈ X and T ∈ T, we let
Z T
φξ bβ (X(s, x))
Uθ (x, T, z) =
dX(s, x);
(4.1)
σ 2 (X(s, x))
0
Z T
φξ b2β (X(s, x))
Vθ (x, T, z) =
ds.
(4.2)
σ 2 (X(s, x))
0
Clearly, Uθ (xi , Ti , z i ) = Ui,θ and Vθ (xi , Ti , z i ) = Vi,θ , where Ui,θ and Vi,θ are given by (2.3). In this
non-iid set-up we assume, following Maitra and Bhattacharya (2016a), that
6
(H8) For l = 1, . . . , p, and for t ∈ [0, Ti ],
n
1X
gl (zil (t)) → cl (t);
n
(4.3)
1X
gl (zil (t))gm (zim (t)) → cl (t)cm (t),
n
(4.4)
i=1
and, for l, m = 1, . . . , p; t ∈ [0, Ti ],
n
i=1
as n → ∞, where cl (t) are real constants.
For x = xk , T = Tk and z = z k , we denote the Kullback-Leibler distance and the Fisher’s
information as Kk (θ 0 , θ) (Kk (θ, θ 0 )) and Ik (θ), respectively. Then the following results hold in the
same way as Lemma 11 of Maitra and Bhattacharya (2016a).
Lemma 5 Assume the non-iid set-up, (H1) – (H4) and (H8). Then for any θ ∈ Θ,
Pn
k=1 Kk (θ 0 , θ)
lim
= K(θ 0 , θ);
n→∞
n
Pn
k=1 Kk (θ, θ 0 )
= K(θ, θ 0 );
lim
n→∞
Pnn
k=1 Ik (θ)
lim
= I(θ),
n→∞
n
(4.5)
(4.6)
(4.7)
where the limits K(θ 0 , θ), K(θ, θ 0 ) and I(θ) are well-defined Kullback-Leibler divergences and Fisher’s
information, respectively.
Lemma 5 will be useful in our asymptotic investigation in the non-iid set-up. In this set-up, we first
investigate consistency and asymptotic normality of M LE using the results of Hoadley (1971).
4.1
Consistency and asymptotic normality of M LE in the non-iid set-up
Following Hoadley (1971) we define the following:
Ri (θ) = log
fi (Xi |θ)
fi (Xi |θ 0 )
=0
if fi (Xi |θ 0 ) > 0
otherwise.
(4.8)
Ri (θ, ρ) = sup {Ri (ψ) : kψ − θk ≤ ρ}
(4.9)
Vi (r) = sup {Ri (θ) : kθk > r} .
(4.10)
Following Hoadley (1971) we denote by ri (θ), ri (θ, ρ) and vi (r) toPbe expectations of Ri (θ), Ri (θ, ρ)
and Vi (r) under θ 0 ; for any sequence {ai ; i = 1, 2, . . .} we denote ni=1 ai /n by ān .
P
Hoadley (1971) proved that if the following regularity conditions are satisfied, then the MLE θ̂ n →
θ0 :
(1) Θ is a closed subset of Rp+q+1 .
(2) fi (Xi |θ) is an upper semicontinuous function of θ, uniformly in i, a.s. [Pθ0 ].
(3) There exist ρ∗ = ρ∗ (θ) > 0, r > 0 and 0 < K ∗ < ∞ for which
7
(i) Eθ0 [Ri (θ, ρ)]2 ≤ K ∗ ,
0 ≤ ρ ≤ ρ∗ ;
(ii) Eθ0 [Vi (r)]2 ≤ K ∗ .
(4)
(i) lim r̄n (θ) < 0,
n→∞
θ 6= θ 0 ;
(ii) lim v̄n (r) < 0.
n→∞
(5) Ri (θ, ρ) and Vi (r) are measurable functions of Xi .
Actually, conditions (3) and (4) can be weakened but these are more easily applicable (see Hoadley
(1971) for details).
4.1.1
Verification of the regularity conditions
Since Θ is compact in our case, the first regularity condition clearly holds.
For the second regularity condition, note that given Xi , fi (Xi |θ) is continuous by our assumptions
(H1) – (H4), as already noted in Section 3.1.2; in fact, uniformly continuous in θ in our case, since Θ is
compact. Hence, for any given > 0, there exists δi () > 0, independent of θ, such that kθ 1 − θ 2 k <
δi () implies |f (Xi |θ 1 ) − f (Xi |θ 2 )| < . Now consider a strictly positive function δx,T (), continuous
in x ∈ X and T ∈ T, such that δxi ,Ti () = δi (). Let δ() = inf δx,T (). Since X and T are compact,
x∈X,T ∈T
it follows that δ() > 0. Now it holds that kθ 1 − θ 2 k < δ() implies |f (Xi |θ 1 ) − f (Xi |θ 2 )| < , for
all i. Hence, the second regularity condition is satisfied.
Let us now focus attention on condition (3)(i).
Vi,θ
Vi,θ0
− Ui,θ0 +
2
2
Vi,θ
Vi,θ0
≤Ui,θ +
− Ui,θ0 +
.
2
2
Ri (θ) =Ui,θ −
(4.11)
Let us denote ψ ∈ Rp+q+1 : kψ − θk ≤ ρ by S(ρ, θ). Here 0 < ρ < ρ∗ (θ), and ρ∗ (θ) is so small
that S(ρ, θ) ⊂ Θ for all ρ ∈ (0, ρ∗ (θ)). It then follows from (4.11) that
Vi,θ0
Vi,θ
− Ui,θ0 +
.
(4.12)
sup Ri (ψ) ≤ sup
Ui,θ +
2
2
ψ∈S(ρ,θ)
θ∈S(ρ,θ)
The supremums in (4.12) are finite due to compactness of S(ρ, θ). Let the supremum be attained at some
θ ∗ where θ ∗ = θ ∗ (Xi ). Then, the expectation of the square of the upper bound can be calculated in the
same way as (3.3) noting that N̄θ in this case will be S(ρ, θ). Since under Pθ0 , finiteness of moments of
all orders of each term in the upper bound is ensured by Lemma 10 of Maitra and Bhattacharya (2016a),
it follows that
Eθ0 [Ri (θ, ρ)]2 ≤ Ki (θ),
(4.13)
where Ki (θ) = K(xi , Ti , z i , θ), with K(x, T, z, θ) being a continuous function of (x, T, z, θ), continuity being again a consequence of Lemma 10 of Maitra and Bhattacharya (2016a). Since because of
compactness of X, T and Θ,
Ki (θ) ≤
K(x, T, z, θ) < ∞,
sup
x∈X,T ∈T,z∈Z,θ∈Θ
regularity condition (3)(i) follows.
To verify condition (3)(ii), first note that we can choose r > 0 such that kθ 0 k < r and {θ ∈ Θ :
kθk > r} 6= ∅. It then follows that
sup
Ri (θ) ≤ sup Ri (θ) for every i ≥ 1. The right hand
{θ∈Θ:kθk>r}
θ∈Θ
side is bounded by the same expression as the right hand side of (4.12), with only S(ρ, θ) replaced with
Θ. The rest of the verification follows in the same way as verification of (3)(i).
8
To verify condition (4)(i) note that by (4.5)
Pn
i=1 Ki (θ 0 , θ)
lim r̄n = − lim
= −K(θ 0 , θ) < 0
n→∞
n→∞
n
for θ 6= θ 0 .
(4.14)
In other words, (4)(i) is satisfied.
Verification of (4)(ii) follows exactly in a similar way as verified in Maitra and Bhattacharya (2016c)
except that the concerned moment existence result follows from Lemma 10 of Maitra and Bhattacharya
(2016a). Regularity condition (5) is seen to hold by the same arguments as in Maitra and Bhattacharya
(2016c).
In other words, in the non-iid SDE framework, the following theorem holds:
Theorem 6 Assume the non-iid SDE setup and conditions (H1) – (H4) and (H8). Then it holds that
P
θ̂ n → θ 0 , as n → ∞.
4.2
Asymptotic normality of M LE in the non-iid set-up
0 (x, θ) =
Let ζi (x, θ) = log fi (x|θ); also, let ζi0 (x, θ) be the (p+q+1)×1 vector with k-th component ζi,k
∂
00
00
∂θk ζi (x, θ), and let ζi (x, θ) be the (p + q + 1) × (p + q + 1) matrix with (k, l)-th element ζi,kl (x, θ) =
∂2
∂θk ∂θl ζi (x, θ).
For proving asymptotic normality in the non-iid framework, Hoadley (1971) assumed the following
regularity conditions:
(1) Θ is an open subset of Rp+q+1 .
P
(2) θ̂ n → θ 0 .
(3) ζi0 (Xi , θ) and ζi00 (Xi , θ) exist a.s. [Pθ0 ].
(4) ζi00 (Xi , θ) is a continuous function of θ, uniformly in i, a.s. [Pθ0 ], and is a measurable function of
Xi .
(5) Eθ [ζi0 (Xi , θ)] = 0 for i = 1, 2, . . ..
(6) Ii (θ) = Eθ ζi0 (Xi , θ)ζi0 (Xi , θ)T = −Eθ [ζi00 (Xi , θ)], where for any vector y, y T denotes the
transpose of y.
(7) Īn (θ) → Ī(θ) as n → ∞ and Ī(θ) is positive definite.
0 (X , θ )
(8) Eθ0 ζi,k
i 0
3
≤ K2 , for some 0 < K2 < ∞.
(9) There exist > 0 and random variables Bi,kl (Xi ) such that
n
o
00 (X , ψ) : kψ − θ k ≤ ≤ B
(i) sup ζi,kl
i
0
i,kl (Xi ).
(ii) Eθ0 |Bi,kl (Xi )|1+δ ≤ K2 , for some δ > 0.
Condition (8) can be weakened but is relatively easy to handle. Under the above regularity conditions,
Hoadley (1971) prove that
√
L
n θ̂ n − θ 0 → N 0, Ī −1 (θ 0 ) .
(4.15)
9
4.2.1
Validation of asymptotic normality of M LE in the non-iid SDE set-up
Condition (1) holds also for compact Θ; see Maitra and Bhattacharya (2016c). Condition (2) is a simple
consequence of Theorem 6.
Conditions (3), (5) and (6) are clearly valid in our case because of interchangability of differentiation and integration, which follows due to (H1) – (H4), (H7) and Theorem B.4 of Rao (2013). Condition
(4) can be verified in exactly the same way as condition (2) of Section 4.1 is verified; measurability of ζi00 (Xi , θ) follows due to its continuity with respect to Xi . Condition (7) simply follows from
(4.7). Compactness, continuity, and finiteness of moments guaranteed by Lemma 10 of Maitra and
Bhattacharya (2016a) imply conditions (8), (9)(i) and 9(ii).
In other words, in our non-iid SDE case we have the following theorem on asymptotic normality.
Theorem 7 Assume the non-iid SDE setup and conditions (H1) – (H8). Then (4.15) holds, as n → ∞.
5
Consistency and asymptotic normality of the Bayesian posterior in the
iid set-up
5.1
Consistency of the Bayesian posterior distribution
As in Maitra and Bhattacharya (2015) here we exploit Theorem 7.80 presented in Schervish (1995),
stated below, to show posterior consistency.
Theorem 8 (Schervish (1995)) Let {Xn }∞
n=1
be conditionally iid given θ with density f1 (x|θ) with
respect to a measure ν on a space X 1 , B 1 . Fix θ 0 ∈ Θ, and define, for each M ⊆ Θ and x ∈ X 1 ,
Z(M, x) = inf log
ψ∈M
f1 (x|θ 0 )
.
f1 (x|ψ)
Assume that for each θ 6= θ 0 , there is an open set Nθ such that θ ∈ Nθ and that Eθ0 Z(Nθ , Xi ) > −∞.
Also assume that f1 (x|·) is continuous at θ for every θ, a.s. [Pθ0 ]. For > 0, define C = {θ :
K1 (θ 0 , θ) < }, where
f1 (X1 |θ 0 )
K1 (θ 0 , θ) = Eθ0 log
(5.1)
f1 (X1 |θ)
is the Kullback-Leibler divergence measure associated with observation X1 . Let π be a prior distribution
such that π(C ) > 0, for every > 0. Then, for every > 0 and open set N0 containing C , the
posterior satisfies
lim πn (N0 |X1 , . . . , Xn ) = 1, a.s. [Pθ0 ].
(5.2)
n→∞
5.1.1
Verification of posterior consistency
The condition Eθ0 Z(Nθ , Xi ) > −∞ of the above theorem is verified in the context of Theorem 1 in
Section 3.1.2. Continuity of the Kullback-Liebler divergence follows easily from Lemma 10 of Maitra
and Bhattacharya (2016a). The rest of the verification is the same as that of Maitra and Bhattacharya
(2015).
Hence, (5.2) holds in our case with any prior with positive, continuous density with respect to the
Lebesgue measure. We summarize this result in the form of a theorem, stated below.
Theorem 9 Assume the iid set-up and conditions (H1) – (H4). Let the prior distribution π of the
parameter θ satisfy dπ
dν = g almost everywhere on Θ, where g(θ) is any positive, continuous density on
Θ with respect to the Lebesgue measure ν. Then the posterior (1.2) is consistent in the sense that for
every > 0 and open set N0 containing C , the posterior satisfies
lim πn (N0 |X1 , . . . , Xn ) = 1,
n→∞
10
a.s. [Pθ0 ].
(5.3)
5.2
Asymptotic normality of the Bayesian posterior distribution
As in Maitra and Bhattacharya (2015), we make use of Theorem 7.102 in conjunction with Theorem
7.89 provided in Schervish (1995). These theorems make use of seven regularity conditions, of which
only the first four, stated below, will be required for the iid set-up.
5.2.1
Regularity conditions – iid case
(1) The parameter space is Θ ⊆ Rq+1 .
(2) θ 0 is a point interior to Θ.
(3) The prior distribution of θ has a density with respect to Lebesgue measure that is positive and
continuous at θ 0 .
(4) There exists a neighborhood N0 ⊆ Θ of θ 0 on which `n (θ) = log f (X1 , . . . , Xn |θ) is twice
continuously differentiable with respect to all co-ordinates of θ, a.s. [Pθ0 ].
Before proceeding to justify asymptotic normality of our posterior, we furnish the relevant theorem
below (Theorem 7.102 of Schervish (1995)).
Theorem 10 (Schervish (1995)) Let {Xn }∞
n=1 be conditionally iid given θ. Assume the above four
regularity conditions; also assume that there exists Hr (x, θ) such that, for each θ 0 ∈ int(Θ) and each
k, j,
sup
kθ−θ 0 k≤r
∂2
∂2
log f1 (x|θ 0 ) −
log f1 (x|θ) ≤ Hr (x, θ 0 ),
∂θk ∂θj
∂θk ∂θj
(5.4)
with
lim Eθ0 Hr (X, θ 0 ) = 0.
(5.5)
r→0
Further suppose that the conditions of Theorem 8 hold, and that the Fisher’s information matrix I(θ 0 )
is positive definite. Now denoting by θ̂ n the M LE associated with n observations, let
−`00n (θ̂ n ) if the inverse and θ̂ n exist
−1
Σn =
(5.6)
Iq+1
if not,
where for any t,
`00n (t)
=
∂2
`n (θ)
∂θi ∂θj
,
(5.7)
θ=t
−1
and Iq+1 is the identity matrix
of order
q + 1. Thus, Σn is the observed Fisher’s information matrix.
−1/2
Letting Ψn = Σn
it holds that
θ − θ̂ n , it follows that for each compact subset B of Rq+1 and each > 0,
lim Pθ0
n→∞
sup πn (Ψn |X1 , . . . , Xn ) − φ̃(Ψn ) > = 0,
(5.8)
Ψn ∈B
where φ̃(·) denotes the density of the standard normal distribution.
5.2.2
Verification of posterior normality
Observe that the four regularity conditions of Section 5.2.1 trivially hold. The remaining conditions of
Theorem 10 are verified in the context of Theorem 3 in Section 3.2.2. We summarize this result in the
form of the following theorem.
11
Theorem 11 Assume the iid set-up and conditions (H1) – (H7). Let the prior distribution π of the
parameter θ satisfy dπ
tothe
dν = g almost everywhere on Θ, where g(θ) is any density with respect
−1/2
Lebesgue measure ν which is positive and continuous at θ 0 . Then, letting Ψn = Σn
θ − θ̂ n , it
follows that for each compact subset B of Rq+1 and each > 0, it holds that
lim Pθ0 sup πn (Ψn |X1 , . . . , Xn ) − φ̃(Ψn ) > = 0.
n→∞
6
(5.9)
Ψn ∈B
Consistency and asymptotic normality of the Bayesian posterior in the
non-iid set-up
For consistency and asymptotic normality in the non-iid Bayesian framework we utilize the result presented in Choi and Schervish (2007) and Theorem 7.89 of Schervish (1995), respectively.
6.1
Posterior consistency in the non-iid set-up
We consider the following extra assumption for our purpose.
(H9) There exist strictly positive functions α1∗ (x, T, z, θ) and α2∗ (x, T, z, θ) continuous in (x, T, z, θ),
such that for any (x, T, z, θ),
Eθ [exp {α1∗ (x, T, z, θ)Uθ (x, T, z)}] < ∞,
and
Eθ [exp {α2∗ (x, T, z, θ)Vθ (x, T, z)}] < ∞,
Now, let
∗
α1,min
=
∗
α2,min
=
inf
α1∗ (x, T, z, θ),
(6.1)
inf
α2∗ (x, T, z, θ)
(6.2)
x∈X,T ∈T,z∈Z,θ∈Θ
x∈X,T ∈T,z∈Z,θ∈Θ
and
∗
∗
α = min α1,min
, α2,min
, c∗ ,
(6.3)
where 0 < c∗ < 1/16.
∗
∗
Compactness ensures that α1,min
, α2,min
> 0, so that 0 < α < 1/16. It also holds due to compactness that for θ ∈ Θ,
sup
Eθ [exp {αUθ (x, T, z)}] < ∞.
(6.4)
x∈X,T ∈T,z∈Z,θ∈Θ
and
sup
Eθ [exp {αVθ (x, T, z)}] < ∞.
(6.5)
x∈X,T ∈T,z∈Z,θ∈Θ
This choice of α ensuring (6.4) and (6.5) will be useful in verification of the conditions of Theorem
12, which we next state.
∞
Theorem 12 (Choi and Schervish (2007)) Let {Xi }∞
i=1 be independently distributed with densities {fi (·|θ)}i=1 ,
with respect to a common σ-finite measure, where θ ∈ Θ, a measurable space. The densities fi (·|θ) are
assumed to be jointly measurable. Let θ 0 ∈ Θ and let Pθ0 be the joint distribution of {Xi }∞
i=1 when θ 0
∞
is the true value of θ. Let {Θn }n=1 be a sequence of subsets of Θ. Let θ have prior π on Θ. Define the
12
following:
fi (Xi |θ 0 )
,
fi (Xi |θ)
Ki (θ 0 , θ) = Eθ0 (Λi (θ 0 , θ))
Λi (θ 0 , θ) = log
%i (θ 0 , θ) = V arθ0 (Λi (θ 0 , θ)) .
Make the following assumptions:
(1) Suppose that there exists a set B with π(B) > 0 such that
P∞ %i (θ0 ,θ)
(i)
< ∞, ∀ θ ∈ B,
i=1
i2
(ii) For all > 0, π (B ∩ {θ : Ki (θ 0 , θ) < , ∀ i}) > 0.
∞
(2) Suppose that there exist test functions {Φn }∞
n=1 , sets {Ωn }n=1 and constants C1 , C2 , c1 , c2 > 0
such that
P∞
(i)
n=1 Eθ 0 Φn < ∞,
(ii)
sup
θ∈Θcn ∩Ωn
Eθ (1 − Φn ) ≤ C1 e−c1 n ,
(iii) π (Ωcn ) ≤ C2 e−c2 n .
Then,
πn (θ ∈ Θcn |X1 , . . . , Xn ) → 0
6.1.1
a.s. [Pθ0 ].
(6.6)
Validation of posterior consistency
First note that, fi (Xi |θ) is given by (2.4). From the proof of Theorem 6, using finiteness of moments of
i |θ 0 )
all orders associated with Ui,θ and Vi,θ , it follows that log ffii(X
(Xi |θ) has an upper bound which has finite
first and second order moments under θ 0 , and is uniform for all θ ∈ B, where B is any compact subset
of Θ. Hence, for each i, %i (θ 0 , θ) is finite. Using compactness, Lemma 10 of Maitra and Bhattacharya
(2016a) and arguments similar to that of Section 3.1.1 of Maitra and Bhattacharya (2015), it easily
follows that %i (θ 0 , θ) < κ, for some 0 < κ < ∞, uniformly in i. Hence, choosing a prior that gives
positive probability to the set B, it follows that for all θ ∈ B,
∞
X
%i (θ 0 , θ)
i=1
i2
∞
X
1
<κ
< ∞.
i2
i=1
Hence, condition (1)(i) holds. Also note that (1)(ii) can be verified similarly as the verification of
Theorem 5 of Maitra and Bhattacharya (2015).
We now verify conditions (2)(i), (2)(ii) and (2)(iii). We let Ωn = Ω1n × Rp+1 , where Ω1n =
{β :k β k< Mn }, where Mn = O(en ). Note that
π (Ωcn ) = π (Ωc1n ) = π(k β k≥ Mn ) < Eπ (k β k) Mn−1 ,
(6.7)
so that (2)(iii) holds, assuming that the prior π is such that the expectation Eπ (k β k) is finite.
The verification of 2(i) can be checked in as in Maitra and Bhattacharya (2015) except the relevant
changes. So, here we only mention the corresponding changes, skipping detailed verification.
Kolmogorov’s strong law of large numbers for the non-iid case (see, for example, Serfling (1980)),
holds in our problem due to finiteness of the moments of Uθ (x, T, z) and Vθ (x, T, z) for every x, T ,
z and θ belonging to the respective compact spaces. Moreover, existence and boundedness of the third
order derivative of `n (θ) with respect to its components is ensured by assumption (H7) along with compactness assumptions. The results stated in Maitra and Bhattacharya (2016a) concerned with continuity
13
and finiteness of the moments of Uθ (x, T, z) and Vθ (x, T, z) for every x, T , z and θ belonging to
their respective compact spaces are needed here. The lower bound of log fi (Xi |θ 0 ) − log fi (Xi |θ̂ n ) is
denoted by C3 (Ui , Vi , θ̂ n ) where
C3 (Ui , Vi , θ̂ n ) = Ui,θ0 −
Vi,θ̂n
Vi,θ0
− Ui,θ̂n −
2
2
The rest of the verification is same as that of Maitra and Bhattacharya (2015) along with assumption
(H9).
To verify condition 2(ii) we define Θn = Θδ = {(β, ξ) : K(θ, θ 0 ) < δ}, where K(θ, θ 0 ), defined
as in (4.6), is the proper Kullback-Leibler divergence. This verification is again similar to that of Maitra
and Bhattacharya (2015). The result can be summarized in the form of the following theorem.
Theorem 13 Assume the non-iid SDE set-up. Also assume conditions (H1) – (H9). For any δ > 0,
let Θδ = {(β, ξ) : K(θ, θ 0 ) < δ}, where K(θ, θ 0 ), defined as in (4.6), is the proper Kullback-Leibler
divergence. Let the prior distribution π of the parameter θ satisfy dπ
dν = h almost everywhere on Θ,
where h(θ) is any positive, continuous density on Θ with respect to the Lebesgue measure ν. Then, as
n → ∞,
πn (θ ∈ Θcδ |X1 , . . . , Xn ) → 0 a.s. [Pθ0 ].
(6.8)
6.2
Asymptotic normality of the posterior distribution in the non-iid set-up
Below we present the three regularity conditions that are needed in the non-iid set-up in addition to the
four conditions already stated in Section 5.2.1, for asymptotic normality given by (5.8).
6.2.1
Extra regularity conditions in the non-iid set-up
(5) The largest eigenvalue of Σn goes to zero in probability.
(6) For δ > 0, define N0 (δ) to be the open ball of radius δ around θ 0 . Let ρn be the smallest
eigenvalue of Σn . If N0 (δ) ⊆ Θ, there exists K(δ) > 0 such that
!
lim Pθ0
n→∞
sup
ρn [`n (θ) − `n (θ 0 )] < −K(δ)
= 1.
(6.9)
θ∈Θ\N0 (δ)
(7) For each > 0, there exists δ() > 0 such that
lim Pθ0
1
2
!
1 + γ T Σn `00n (θ)Σn γ <
sup
n→∞
1
2
= 1.
(6.10)
θ∈N0 (δ()),kγk=1
Although intuitive explanations of all the seven conditions are provided in Schervish (1995), here we
briefly touch upon condition (7), which is seemingly somewhat unwieldy. First note that condition (6)
ensures consistency of the M LE θ̂ n , so that θ̂ n ∈ N0 (δ), as n → ∞. Thus, in (7), for sufficiently large
n and sufficiently small δ(), `00n (θ) ≈ `00n (θ̂ n ), for all θ ∈ N0 (δ()). Hence, from the definition of Σn−1
1
1
1
1
given by (5.6), it follows that Σn2 `00n (θ)Σn2 ≈ −Iq+1 so that, since kγk = 1, 1 + γ T Σn2 `00n (θ)Σn2 γ ≈
1 − kγk2 < , for all θ ∈ N0 (δ()) and for large enough n. Now it is easy to see that the role of
condition (7) is only to formalize the heuristic arguments.
6.2.2
Verification of the regularity conditions
Assumptions (H1) – (H9), along with Kolmogorov’s strong law of large numbers, are sufficient for
the regularity conditions to hold; the arguments remain similar as those in Section 3.2.2 of Maitra and
Bhattacharya (2015). We provide our result in the form of the following theorem.
14
Theorem 14 Assume the non-iid set-up and conditions (H1) – (H9). Let the prior distribution π of the
parameter θ satisfy dπ
dν = h almost everywhere on Θ, where h(θ) is any density with respect to
the
−1/2
Lebesgue measure ν which is positive and continuous at θ 0 . Then, letting Ψn = Σn
each compact subset B of
Rp+q+1
lim Pθ0
n→∞
7
θ − θ̂ n , for
and each > 0, the following holds:
sup πn (Ψn |X1 , . . . , Xn ) − φ̃(Ψn ) > = 0.
(6.11)
Ψn ∈B
Random effects SDE model
We now consider the following system of SDE models for i = 1, 2, . . . , n:
dXi (t) = φξi (t)bβ (Xi (t))dt + σ(Xi (t))dWi (t)
(7.1)
Note that this model is the same as described in Section 2 except that the parameters ξ i now depend
upon i. Indeed, now φξi (t) is given by
φξi (t) = φξi (z i (t)) = ξ0i + ξ1i g1 (zi1 (t)) + ξ2i g2 (zi2 (t)) + · · · + ξpi gp (zip (t)),
(7.2)
where ξ i = (ξ0i , ξ1i , . . . , ξpi )T is the random effect corresponding to the i-th individual for i = 1, . . . , n,
and z i (t) is the same as in Section 2.1. We let bβ (Xi (t), φξi ) = φξi (t)bβ (Xi (t)). Note that our
likelihood is the product over i = 1, . . . , n, of the following individual densities:
Vi,ξi ,β
,
fi,ξi ,β (Xi ) = exp Ui,ξi ,β −
2
where
Z
Ui,ξi ,β =
0
Ti
φξi (s)bβ (Xi (s))
σ 2 (Xi (s))
Z
dXi (s)
and
Vi,ξi ,β =
Ti
φ2ξi (s)b2β (Xi (s))
0
σ 2 (Xi (s))
ds.
β
β
T
Now, let mβ (z(t), x(t)) = (mβ
0 , m1 (z1 (t), x(t)), . . . , mp (zp (t), x(t))) be a function from Z ×
β
R → Rp+1 where mβ
0 ≡ 1 and mk (z(t), x(t)) = gk (zk (t))bβ (x(t)); k = 1, . . . , p. With this notation,
the likelihood can be re-written as the product over i = 1, . . . , n, of the following:
i T β i
fi,ξi ,β (Xi ) = exp((ξ i )T Aβ
i − (ξ ) B i ξ )
where
mβ (z(s), Xi (s))
dXi (s)
σ 2 (Xi (s))
(7.4)
T
mβ (z(s), Xi (s)) mβ (z(s), Xi (s))
ds
σ 2 (Xi (s))
(7.5)
Aβ
i =
and
Bβ
i
Z
=
0
Ti
Ti
(7.3)
Z
0
are (p + 1) × 1 random vectors and positive definite (p + 1) × (p + 1) random matrices respectively.
We assume that ξ i are iid Gaussian vectors, with expectation vector µ and covariance matrix Σ ∈
S p+1 (R) where S p+1 (R) is the set of real positive definite symmetric matrices of order p + 1. The
parameter set is denoted by θ = (µ, Σ, β) ∈ Θ ⊂ Rp+1 × S p+1 (R) × Rq .
To obtain the likelihood involving θ we refer to the multidimensional random effects set-up of Delattre et al. (2013). Following Lemma 2 of Delattre et al. (2013) it then follows in our case that, for each
β
β
−1
i ≥ 1 and for all θ, B β
i + Σ , Ip+1 + B i Σ, Ip+1 + ΣB i are invertible.
15
−1
−1 β
Setting Rβ
= (Ip+1 + B β
i
i Σ) B i we obtain
1
1
exp −
2
Bβ
i
µ−
fi (Xi |θ) = q
det(Ip+1 + B β
Σ)
i
T
−1
1
β
β
β
× exp
Ai
Ai
Bi
2
−1
Aβ
i
T
Rβ
i
−1
−1
β
β
µ − Bi
Ai
!
(7.6)
as our desired likelihood after integrating (7.3) with respect to the distrbution of ξ i .
With reference to Delattre et al. (2013) in our case
β
β
−1
γi (θ) = (Ip+1 + ΣB β
i ) (Ai − B i µ)
and
−1 β
Ii (Σ) = (Ip+1 + ΣB β
i ) Bi .
Hence, Proposition (10)(i) of Delattre et al. (2013) can be seen to be hold here in a similar way by
β
replacing Ui and Vi by Aβ
i and B i respectively.
Asymptotic investigation regarding consistency and asymptotic normality of M LE and Bayesian
posterior consistency and asymptotic posterior normality in both iid and non-iid set-ups can be established as in the one dimensional cases in Maitra and Bhattacharya (2016c) and Maitra and Bhattacharya
β
(2015) with proper multivariate modifications by replacing Ui and Vi with Aβ
i and B i respectively, and
exploiting assumptions (H1) – (H9).
8
Simulation studies
We now supplement our asymptotic theory with simulation studies where the data is generated from
a specific system of SDEs with one covariate, with given values of the parameters. Specifically, in
the classical case, we obtain the distribution of the M LEs using parametric bootstrap, along with the
95% confidence intervals of the parameters. We demonstrate in particular that the true values of the
parameters are well-captured by the respective 95% confidence intervals. In the Bayesian counterpart,
we obtain the posterior distributions of the parameters along with the respective 95% credible intervals,
and show that the true values fall well within the respective 95% credible intervals.
8.1
Distribution of M LE when n = 20
To demonstrate the finite sample analogue of asymptotic distribution of M LE as n → ∞, we consider
n = 20 individuals, where the i-th one is modeled by
dXi (t) = (θ1 + θ2 zi1 (t))(θ3 + θ4 Xi (t))dt + σdWi (t),
(8.1)
for i = 1, . . . , 20. We fix our diffusion coefficient as σ = 1. We consider the initial value X(0) = 0
and the time interval [0, T ] with T = 1. Further, we choose the true values as θ1 = 1, θ2 = −1, θ3 =
2, θ4 = −2.
We assume that the time dependent covariates zi1 (t) satisfy the following SDE
dzi1 (t) = ξi1 zi1 (t))dt + dWi (t),
iid
(8.2)
for i = 1, . . . , 20, where the coeffiicients ξi1 ∼ N (7, 1) for i = 1, . . . , 20. After simulating the
covariates using the system of SDEs (8.2), we generate the data using the system of SDEs (8.1). In
both the cases we discretize the time interval [0, 1] into 100 equispaced time points.
The distributions of the M LEs of the four parameters are obtained through the parametric bootstrap
method. In this method we simulated the data 1000 times by simulating as many paths of the Brownian
motion, where each data set consists of 20 individuals. Under each data set we perform the “blockrelaxation” method (see, for example, Lange (2010) and the references therein) to obtain the M LE.
16
In a nutshell, starting with some sensible initial value belonging to the parameter space, the blockrelaxation method iteratively maximizes the optimizing function (here, the log-likelihood), successively,
with respect to one parameter, fixing the others at their current values, until convergence is attained with
respect to the iterations. Details follow.
For the initial values of the θj for j = 1, . . . , 4, required to begin the block-relaxation method,
(0)
we simulate four N (0, 1) variates independently, and set them as the initial values θj ; j = 1, . . . , 4.
(i)
Denoting by θj the value of θj at the i-th iteration, for i ≥ 1, and letting L be the likelihood, the
block-relaxation method consists of the following steps:
Algorithm 1 Block-relaxation for M LE in SDE system with covariates
(i)
(1) At the i-th iteration, for j = 1, 2, 3, 4, obtain θj
(i)
(i)
by solving the equation
(i)
(i−1)
∂ log L
∂θj
= 0, con(i−1)
ditionally on θ1 = θ1 , θ2 = θ2 , . . . , θj−1 = θj−1 , θj+1 = θj+1 , . . . , θ4 = θ4
(i)
(i)
θ (i) = θ1 , . . . , θ4 .
. Let
(2) Letting k · k denote the Euclidean norm, if θ (i) − θ (i−1) ≤ 10−5 , set θ̂ = θ (i) , where θ̂ stands
for the maximum likelihood estimator of θ = (θ1 , . . . , θ4 ).
(3) If, on the other hand, θ (i) − θ (i−1) > 10−5 , increase i to i + 1 and continue steps (1) and (2).
Once we obtain the M LE by the above block-relaxation algorithm, we then plug-in θ = θ̂ in
(8.1) and generate 1000 data sets from the resulting system of SDEs, and apply the block-relaxation
algorithm to each such data set to obtain the M LE associated with the data sets. Thus, we obtain the
distribution of the M LE using the parametric bootstrap method.
The distributions of the components of θ̂ (denoted by θ̂i for i = 1, . . . , 4) are shown in Figure 8.1,
where the associated 95% confidence intervals are shown in bold lines. As exhibited by the figures, the
95% confidence intervals clearly contain the true values of the respective components of θ.
8.2
Posterior distribution of the parameters when n = 20
We now consider simulation study for the Bayesian counterpart, using the same data set simulated from
the system of SDEs given by (8.1), with same covariates simulated from
(8.2). We consider an empirical
Bayes prior based on the M LE such that for j = 1, . . . , 4, θj ∼ N θ̂j , σ̂j2 independently, where θ̂j
is the M LE of θj and σ̂j2 is such that the length of the 95% confidence interval associated with the
distribution of the M LE θ̂j , after adding one unit to both lower and upper ends of the interval, is the
same as the length of the 95% prior credible interval [θ̂j − 1.96σ̂j − 1, θ̂j + 1.96σ̂j + 1]. In other words,
we select σ̂j such that the length of the corresponding 95% prior credible interval is the same as that of
the enlarged 95% confidence interval associated with the distribution of the corresponding M LE.
To simulate from the posterior distribution of θ, we perform approximate Bayesian computation
(ABC) (Tavaŕe et al. (1997), Beaumont et al. (2002), Marjoram et al. (2003)), since the standard Markov
chain Monte Carlo (MCMC) based simulation techniques, such as Gibbs sampling and MetropolisHastings algorithms (see, for example, Robert and Casella (2004), Brooks et al. (2011)) failed to ensure
good mixing behaviour of the underlying Markov chains. Denoting the true data set by Xtrue , our
method of ABC is described by following steps. Figure 8.2 shows the posterior distribution of the
parameters θi , for i = 1, . . . , 4, where the 95% posterior credible intervals are shown in bold lines.
Observe that all the true values of θj ; j = 1, . . . , 4, fall comfortably within the respective 95% posterior
credible intervals.
17
Density of mle of (θ2)
600
density
400
60
0
0
20
200
40
density
80
800
100
1000
120
1200
Density of mle of (θ1)
0.990
0.995
1.000
θ^
1.005
1.010
−1.0010
−1.0005
1
−1.0000
θ^
−0.9995
−0.9990
2
Density of mle of (θ4)
density
0
0
5
1000
10
2000
15
density
20
3000
25
4000
30
35
5000
Density of mle of (θ3)
1.96
1.98
2.00
θ^
2.02
2.04
−2.0004
−2.0003
−2.0002
−2.0001
−2.0000
−1.9999
−1.9998
−1.9997
θ^4
3
Figure 8.1: Distributions of the M LEs.
Algorithm 2 ABC for SDE system with covariates
(1) For j = 1, . . . , 4, we simulate the parameters θj from their respective prior distributions.
(2) With the obtained values of the parameters we simulate the new data, which we denote by Xnew ,
using the system of SDEs (8.1).
(3) We calculate the average Euclidean distance between Xnew and Xtrue and denote it by dx .
(4) Until dx < 0.1, we repeat steps (1)–(3).
(5) Once dx < 0.1, we set the corresponding θ as a realization from the posterior of θ with approximation error 0.1.
(6) We obtain 10000 posterior realizations of θ by repeating steps (1)–(5).
18
Posterior of (θ2)
1.0
Posterior Density
0.0
0.0
0.5
0.5
Posterior Density
1.0
1.5
2.0
1.5
Posterior of (θ1)
0.5
1.0
1.5
2.0
−2.0
−1.5
(θ1)
−1.0
(θ2)
Posterior of (θ4)
0.6
Posterior Density
0.4
0.4
0.0
0.2
0.2
0.0
Posterior Density
0.6
0.8
1.0
0.8
1.2
Posterior of (θ3)
0.5
1.0
1.5
2.0
2.5
3.0
3.5
4.0
−3.5
(θ3)
−3.0
−2.5
−2.0
(θ4)
Figure 8.2: Posterior distributions of the parameters.
19
−1.5
−1.0
−0.5
9
Application to real data
We now consider application of our SDE system consisting of covariates to a real, stock market data
(467 observations from August 5, 2013, to June 30, 2015) for 15 companies. The data are available at
www.nseindia.com.
Each company-wise data is modeled by the availabe standard financial SDE models with the
“fitsde” package in R. The minimum value of BIC (Bayesian Information Criterion) is found corresponding to the CKLS (Chan, Karolyi, Longstaff and Sander; see Chan et al. (1992)) model. Denoting
the data by X(t), the CKLS model is described by
dX(t) = (θ1 + θ2 X(t))dt + θ3 X(t)θ4 dW (t).
In our application we treat the diffusion coefficient as a fixed quantity. So, we fix the values of θ3 and
θ4 as obtained by the “fitsde” function, We denote θ3 = A, θ4 = B.
We consider the “close price” of each company as our data X(t). IIP general index, bank interest
rate and US dollar exchange rate are considered as time dependent covariates which we incorporate in
the CKLS model.
The three covariates are denoted by c1 , c2 , c3 , respectively. Now, our considered system of SDE
models for national stock exchange data associated with the 15 companies is the following:
i
dXi (t) = (θ1 + θ2 c1 (t) + θ3 c2 (t) + θ4 c3 (t))(θ5 + θ6 Xi (t))dt + Ai Xi (t)B dWi (t),
(9.1)
for i = 1, . . . , 15.
9.1
Distribution of M LE
We first obtain the M LEs of the 6 parameters θj for j = 1, . . . , 6 by the block-relaxation algorithm
described by Algorithm 1, in Section 8.1. In this real data set up, the process starts with the initial
value θj = 1 for j = 1, . . . , 6 (our experiments with several other choices demonstrated practicability
of those choices as well) and in step (3) of Algorithm 1, the distance is taken as 0.1 instead of 10−5 .
Then taking the M LEs as the value of the parameters θj for j = 1, . . . , 6 we perform the parametric
bootstrap method where we generate the data 1000 times and with respect to each data set, obtain the
M LEs of the six parameters by the block-relaxation method as already mentioned. Figure 9.1 shows the
distribution of M LEs (denoted by θ̂j for j = 1, . . . , 6) where the respective 95% confidence intervals
are shown in bold lines. Among the covariates, c3 , that is, the US dollar exchange rate, seems to be less
significant compared to the others, since the distribution of the M LE of the associated coefficient, θ3 ,
has highest density around zero, with small variability, compared to the other coefficients. Also note
that the distribution of θ̂6 is highly concentrated around zero, signifying that the Xi (t) term in the drift
function of (9.1) is probably redundant.
9.2
Posterior Distribution of the parameters
In the Bayesian approach all the set up regarding the real data is exactly the same as in Section 9, that is,
each data is driven by the SDEs (9.1) where the covariates cj for j = 1, . . . , 3 are already mentioned
in that section. In this case, we consider the priors for the 6 parameters to be independent normal with
mean zero and variance 100. Since in real data situations the parameters are associated with greater
uncertainties compared to simulation studies, somewhat vague prior as we have chosen here, as opposed
to that in the simulation study case, makes sense.
The greater uncertainty in the parameters in this real data scenario makes room for more movement,
and hence, better mixing of MCMC samplers such as Gibbs sampling, in contrast with that in simulation
studies. As such, our application of Gibbs sampling, were the full conditionals are normal distributions
with appropriate means and variances, yielded excellent mixing. Although we chose the initial values
as θj = 0.1; j = 1, . . . , 6, other choices also turned out to be very much viable. We perform 100000
20
Density of mle of (θ3)
3
density
density
−0.9
−0.8
−0.7
−0.6
θ^1
−0.5
0
0
0
1
1
2
2
2
3
4
density
4
4
5
6
5
6
7
6
Density of mle of (θ2)
8
Density of mle of (θ1)
−0.4
0.0
0.2
0.4
0.6
−0.3
−0.2
−0.1
0.1
0.2
Density of mle of (θ5)
0.006
0.008
Density of mle of (θ6)
density
density
0.6
0.8
1.0
θ^4
1.2
1.4
0
0
0
1
50
1
2
100
2
density
3
150
3
4
200
4
5
250
5
Density of mle of (θ4)
0.0
θ^3
θ^2
−0.7
−0.6
−0.5
−0.4
−0.3
−0.2
−0.004
−0.002
θ^5
Figure 9.1: Distribution of M LEs for the real data.
21
0.000
0.002
θ^
6
0.004
0.3
Trace plot of (θ2)
Trace plot of (θ3)
θ3
−15
θ2
−30
−20
−10
−25
−15
−20
0
−10
θ1
−5
−10
10
−5
0
0
5
20
5
Trace plot of (θ1)
0
2000
4000
6000
8000
10000
0
2000
4000
Iteration
6000
8000
10000
0
2000
4000
Iteration
8000
10000
Trace plot of (θ5)
Trace plot of (θ6)
0.04
θ6
θ5
−10
−0.02
0
0.00
−1.0
0.02
10
θ4
−0.5
20
0.06
0.08
30
0.0
0.10
Trace plot of (θ4)
6000
Iteration
0
2000
4000
6000
8000
10000
0
Iteration
2000
4000
6000
8000
10000
Iteration
0
2000
4000
6000
8000
10000
Iteration
Figure 9.2: Trace Plot of the Parameters
Gibbs sampling iterations to obtain our required posterior distributions of the 6 parameters. Figure 9.2
shows the trace plots of the 6 parameters associated with 10000 thinned samples obtained by plotting
the output of every 10-th iteration. We emphasize that although we show the trace plots of only 10000
Gibbs sampling realizations to reduce the file size, our inference is based on all the 100000 realizations.
From the trace plots, convergence of the posterior distributions of the parameters is clearly observed.
Figure 9.2 displays the posterior densities of the 6 parameters, where the 95% credible intervals are
indicated by bold lines. The posterior distribution of θ3 is seen to include zero in the highest density
region; however, unlike the distribution of the M LE θ̂3 , the posterior of θ3 has a long left tail, so that
insignificance of c3 is not very evident. The posterior of θ6 is highly concentrated around zero, agreeing
with the M LE of θ6 that the term Xi (t) in the drift function is perhaps redundant. Note that the posterior
of θ5 also inclues zero in its high-density region, however, it has a long left tail, so that the significance
of θ5 , and hence, of the overall drift function, is not ruled out.
10
Summary and conclusion
In SDE based random effects model framework, Delattre et al. (2013) considered the linearity assumption in the drift function given by b(x, φi ) = φi b(x), assuming φi to be Gaussian random variables with
mean µ and variance ω 2 , and obtained a closed form expression of the likelihood of the above parameters. Assuming the iid set-up, they proved convergence in probability and asymptotic normality of the
maximum likelihood estimator of the parameters.
Maitra and Bhattacharya (2016a) and Maitra and Bhattacharya (2016b) extended their model by
incorporating time-varying covariates in φi and allowing b(x) to depend upon unknown parameters,
22
Posterior of (θ2)
Posterior of (θ3)
−20
−15
−10
−5
0
5
0.08
Posterior density
0.00
0.00
0.00
0.02
0.05
0.04
0.06
0.10
Posterior density
0.05
Posterior density
0.10
0.15
0.10
0.12
0.20
0.15
0.14
Posterior of (θ1)
−10
−5
0
5
θ1
10
15
−25
−20
−15
−10
Posterior of (θ5)
0
5
Posterior of (θ6)
200
Posterior density
Posterior density
5
0.04
0
0
10
20
θ4
30
0
100
0.02
0.00
Posterior density
10
300
0.06
400
15
Posterior of (θ4)
−5
θ3
θ2
−0.25
−0.20
−0.15
−0.10
−0.05
0.00
0.05
0.10
−0.010
−0.008
−0.006
−0.004
θ5
Figure 9.3: Posterior Distributions of the Parameters for real data
23
−0.002
θ6
0.000
0.002
0.004
but rather than inference regarding the parameters, they developed asymptotic model selection theory
based on Bayes factors for their purposes. In this paper, we developed asymptotic theories for parametric
inference for both classical and Bayesian paradigms under the fixed effects set-up, and provided relevant
discussion of asymptotic inference on the parameters in the random effects set-up.
As our previous investigations (Maitra and Bhattacharya (2016c), Maitra and Bhattacharya (2015),
for instance), in this work as well we distinguished the non-iid set-up from the iid case, the latter
corresponding to the system of SDEs with same initial values, time domain, but with no covariates.
However, as already noted, this still provides a generalization to the iid set-up of Delattre et al. (2013)
through generalization of b(x) to bβ (x); β being a set of unknown parameters. Under suitable assumptions we obtained strong consistency and asymptotic normality of the M LE under the iid set-up and
weak consistency and asymptotic normality under the non-iid situation. Besides, we extended our classical asymptotic theory to the Bayesian framework, for both iid and non-iid situations. Specifically, we
proved posterior consistency and asymptotic posterior normality, for both iid and non-iid set-ups.
In our knowledge, ours is the first-time effort regarding asymptotic inference, either classical or
Bayesian, in systems of SDEs under the presence of time-varying covariates. Our simulation studies
and real data applications, with respect to both classical and Bayesian paradigms, have revealed very
encouraging results, demonstrating the importance of our developments even in practical, finite-sample
situations.
Acknowledgment
We are sincerely grateful to the EIC, the AE and the referee whose constructive comments have led
to significant improvement of the quality and presentation of our manuscript. The first author also
gratefully acknowledges her CSIR Fellowship, Govt. of India.
References
Beaumont, M. A., Zhang, W., and Balding, D. J. (2002). Approximate Bayesian Computation in Population Genetics. Genetics, 162, 2025–2035.
Bishwal, J. P. N. (2008). Parameter Estimation in Stochastic Differential Equations. Lecture Notes in
Mathematics, 1923, Springer-Verlag.
Brooks, S., Gelman, A., Jones, G. L., and Meng, X.-L. (2011). Handbook of Markov Chain Monte
Carlo. Chapman and Hall, London.
Chan, K. C., Karolyi, G. A., Longstaff, F. A., and Sanders, A. B. (1992). An Empirical Comparison of
Alternative Models of the Short-Term Interest Rate. The Journal of Finance, 47, 1209–1227.
Choi, T. and Schervish, M. J. (2007). On Posterior Consistency in Nonparametric Regression Problems.
Journal of Multivariate Analysis, 98, 1969–1987.
Delattre, M., Genon-Catalot, V., and Samson, A. (2013). Maximum Likelihood Estimation for Stochastic Differential Equations with Random Effects. Scandinavian Journal of Statistics, 40, 322–343.
Hoadley, B. (1971). Asymptotic Properties of Maximum Likelihood Estimators for the Independent not
Identically Distributed Case. The Annals of Mathematical Statistics, 42, 1977–1991.
Lange, K. (2010). Numerical Analysis for Statisticians. Springer, New York.
Leander, J., Almquist, J., Ahlström, C., Gabrielsson, J., and Jirstrand, M. (2015). Mixed Effects Modeling Using Stochastic Differential Equations: Illustrated by Pharmacokinetic Data of Nicotinic Acid
in Obese Zucker Rats. The AAPS Journal, 17, 586–596.
24
Maitra, T. and Bhattacharya, S. (2015). On Bayesian Asymptotics in Stochastic Differential Equations with Random Effects. Statistics and Probability Letters, 103, 148–159. Also available at
“http://arxiv.org/abs/1407.3971”.
Maitra, T. and Bhattacharya, S. (2016a). Asymptotic Theory of Bayes Factor in Stochastic Differential
Equations: Part I. Available at “https://arxiv.org/abs/1503.09011”.
Maitra, T. and Bhattacharya, S. (2016b). Asymptotic Theory of Bayes Factor in Stochastic Differential
Equations: Part II. Available at “https://arxiv.org/abs/1504.00002”.
Maitra, T. and Bhattacharya, S. (2016c). On Asymptotics Related to Classical Inference in Stochastic
Differential Equations with Random Effects. Statistics and Probability Letters, 110, 278–288. Also
available at “http://arxiv.org/abs/1407.3968”.
Marjoram, P., Molitor, J., Plagnol, V., and Tavaŕe, S. (2003). Markov Chain Monte Carlo Without
Likelihoods. Proceedings of the National Academy of Sciences, 100, 15324–15328.
Oravecz, Z., Tuerlinckx, F., and Vandekerckhove, J. (2011). A hierarchical latent stochastic differential
equation model for affective dynamics. Psychological Methods, 16, 468–490.
Overgaard, R. V., Jonsson, N., Tornœ, C. W., and Madsen, H. (2005). Non-Linear Mixed-Effects Models with Stochastic Differential Equations: Implementation of an Estimation Algorithm. Journal of
Pharmacokinetics and Pharmacodynamics, 32, 85–107.
Ramsay, J. O. and Silverman, B. W. (2005). Functional Data Analysis. Springer, New York.
Rao, B. (2013). Semimartingales and their Statistical Inference. Chapman and Hall/CRC, Boca Ratan.
Robert, C. P. and Casella, G. (2004). Monte Carlo Statistical Methods. Springer-Verlag, New York.
Schervish, M. J. (1995). Theory of Statistics. Springer-Verlag, New York.
Serfling, R. J. (1980). Approximation Theorems of Mathematical Statistics. John Wiley & Sons, Inc.,
New York.
Tavaŕe, S., Balding, D. J., Griffiths, R. C., and Donelly, P. (1997). Inferring Coalescence Times from
DNA Sequence Data. Genetics, 145, 505–518.
Zhu, B., Song, P. X.-K., and Taylor, J. M. G. (2011). Stochastic Functional Data Analysis: A Diffusion
Model-Based Approach. Biometrics, 67, 1295–1304.
25
| 10math.ST
|
Social Aggregation
1
Simplicial models of social aggregation I
Mirco A. Mannucci, Lisa Sparks, Daniele C. Struppa
This article is dedicated to the memory of Jonathan M. Beck (d. 2006), who taught one of
us to search for the hidden, endless treasures of the simplicial ladder.
Abstract
This paper presents the foundational ideas for a new way of modeling social aggregation.
Traditional approaches have been using network theory, and the theory of random
networks. Under that paradigm, every social agent is represented by a node, and every
social interaction is represented by a segment connecting two nodes. Early work in family
interactions, as well as more recent work in the study of terrorist organizations, shows
that network modeling may be insufficient to describe the complexity of human social
structures. Specifically, network theory does not seem to have enough flexibility to
represent higher order aggregations, where several agents interact as a group, rather than
as a collection of pairs. The model we present here uses a well established mathematical
theory, the theory of simplicial complexes, to address this complex issue prevalent in
interpersonal and intergroup communication. The theory enables us to provide a richer
graphical representation of social interactions, and to determine quantitative mechanisms
to describe the robustness of a social structure. We also propose a methodology to create
random simplicial complexes, with the purpose of providing a new method to simulate
computationally the creation and disgregation of social structures. Finally, we propose
several measures which could be taken and observed in order to describe and study an
actual social aggregation occurring in interpersonal and intergroup contexts.
Social Aggregation
2
Toward a dynamic representation of interpersonal and intergroup communication:
Simplicial models of social aggregation
A topic of great interest in modern social sciences is the modeling of social
aggregation in a variety of forms. Social scientists are interested in the theory of group
formation, group evolution, and are looking for a variety of models which could help
them to both describe group static and its dynamic evolution. Scholars have been
particularly interested in examining group dynamics and decision-making in family, and
health risk communication contexts, particularly when it comes to the potential to offer
tools to act on groups in terms of either strengthening or weakening them (see e.g.,
Ayres, 1983; Fisher, 1953; James, 1953; Harwood & Sparks, 2003; Noller & Fitzpatrick,
1990; Nussbaum, 1989; Sillars & Wilmot, 1989; Sparks, 2003; Sparks, 2005; Sparks &
Harwood, in press; Sparks, Kreps, Botan, & Rowan, 2005; Sparks-Bethea, 2002; Wheeler
& Nezlek, 1977).
This paper proposes a new theoretical model that offers a dynamic representation
of interpersonal and intergroup communication by describing social aggregations,
providing a quantitative analysis of these social groups, and offering the possibility of
computer simulations of the evolution of such groups. Although various interpersonal and
intergroup contexts certainly possess unique characteristics (see e.g., Gallois & Giles,
1998), the simplicial model of social aggregation provides an important mathematically
based dynamic representation that is grounded in prior theoretical work in interpersonal
communication and intergroup relations.
An example which has regretfully become quite topical is the study of terrorist
networks, which only recently has been a topic of interest for communication scholars
(see e.g., Sparks, 2005; Sparks, Kreps, Botan, & Rowan, 2005). The assumption here is
that one can model a terrorist group (or a coalition of groups) via networks. In this model,
every terrorist is represented by a node in the network, and every time two terrorists are
interacting with each other, this is represented by an edge connecting the two nodes. This
model is interesting because mathematicians have developed a reliable and deep theory of
networks, and because one can add detail to the networks by allowing, for example,
Social Aggregation
3
multiple edges across nodes, as well as directed edges, colored edges, and finally
weighted edges. Thus, the quality of the simulation becomes increasingly good. The
purpose of modeling terrorist structures with networks is to provide our intelligence
agencies with visualization models, but also with tools to predict the evolution of the
network. On this point, the foremost theory is the so-called Random Graph Theory
(RGT) introduced by Erdos and Renyi (1960). Recently RGT has been partly superseded
by more realistic models, such as the theory of small-worlds (Watts 1999a, 1999b, 2004),
or the theory of scale-free networks (Barabasi, 2003; Barabasi et al., 2002): an excellent
and comprehensive survey is given by Dorogotsev and Mendes (2001). Of related interest
is the study of the topological properties of networks with the purpose of classifying
them, but also of finding ways to destroy or weaken such networks.
However, network theory as a way to discuss social aggregation is not limited to
military applications. As Sparks has pointed out in (Sparks-Bethea, 2002), graph and
network theory are used in the study of interpersonal communication as a tool to
describe, model, negotiate, re-negotiate, and control situations. Further, researchers of
relationships are frequently concerned with understanding marriage, romance, or
friendship; yet often forget about the matrix of associations in which individuals and any
one of their relationships are embedded. As Bateson (1984) suggested, to attempt to
understand an individual or any one of their relationships separate from their social
matrix is misleading. Although dyadic communication is essential to an understanding of
relationships (see e.g., Fisher, 1953; James, 1953; Wheeler & Nezlek, 1977), it is equally
essential that researchers of relationships gain a broader understanding of the dyad as it is
intertwined in other relationships. Thus, while interpersonal communication is grounded
in interactive processes, behaviors, symbolic exchanges, shared meaning, social
cognition, context, consciousness and intent, individual differences, and the like,
intergroup communication also takes an individual’s social categories into account.
Intergroup communication is most often defined as communication behaviors exhibited
by one or more individuals toward one or more other individuals that is based on the
individuals’ identification of themselves and others as belonging to different social
categories (see Harwood, Giles, & Ryan, 1995; Hajek & Giles, 2002). Thus, it is
important to understand the essential links between interpersonal communication and
Social Aggregation
4
intergroup communication. It is this broader understanding that we hope to be able to
offer with our model. To offer a specific example, the literature on family relationships
and family decision-making suggests that when the fundamental dyad (husband and wife)
is supplemented by the arrival, for example, of an aging parent the core relationship
undergoes a resilient modification of communicative behaviors to adjust to the new
situation (see Sparks-Bethea, 2002). The dyad becomes then a triangle, and the effects
can be quite unexpected. This particular example, however, is important because it shows
in which sense one may say that networks are inadequate as a tool to understand social
aggregation. Let us consider, again, the husband-wife dyad. Its natural network
representation is a graph with two nodes (H for husband, and W for wife), and one edge
connecting them, which represent the relationship and the communication between them.
When a third person enters this dyad, for example an elder parent, the network
representation allows only three different representations. In all cases we will have a
graph with three nodes (H, W, and P for parent), and with an edge HW between husband
and wife. However, we can represent the actual communicative structure by either having
an edge HP between H and P and an edge WP between W and P or not. So, three cases
arise: HP and WP are both present, or HP is present and WP is not, or finally WP is
present but HP is not.
The last two of these graphs are isomorphic, in the sense that they are simply connected
(i.e. they will become disconnected if node H or W are removed). The first graph,
however, has three nodes and three edges, and it cannot be disconnected by eliminating
one of the nodes (for example if one of the family members dies). For this reason, this is
a model of a stronger group. Nevertheless, social scientists have discovered (see e.g.,
Sparks-Bethea, 2002) that even when the model for the triad is the triangle, there are very
Social Aggregation
5
distinctive social situations which cannot be well modeled by the triangle shaped
network. In particular, we can discuss at least two different situations. In one, the three
members of the family talk to each other, on a one-to-one basis, but don’t really create a
unified family. W talks to H, and both of them talk to P, but there is little interaction of
the three as a family. There is no family dinner, so to speak. As it turns out, this situation
may have really negative effects on the strength of the connection between H and W (as it
could be indicated by a weakening of the weight of the edge HW). A very different
situation takes place when in addition to the three dyadic interactions, there is also an
interaction at the group level. This is the situation in which the three pairs have actually
merged into a family. As it is clear, networks cannot directly model such a case, though
at least some of the situations could be handled through the development of dynamical
networks with conditional probabilities
Once again, Sparks noted in (Sparks-Bethea, 2002) that the study of triadic
relationships offers an opportunity for researchers to move beyond the dyad as a standalone functioning unit. Wilmot and Sillars (1989) point out that "Just as individual
responses fail to reveal everything of interest about interpersonal relationships, dyadic
relationships fail to reveal everything of interests about a person's network of
relationships" (p.128). Studying the fluid formation of triadic relationships may add a
much-needed element to better understand the needs of the functioning marital dyad as it
is influenced by other relationships.
Communication scholars have generally neglected the study of “triadic”
relationships. Exceptions exist in the health communication domain (e.g., Beisecker,
1989), the family communication domain (e.g., Long & Mancini, 1990; McHale, 1995;
Wilmot, 1987; Sparks-Bethea, 2002), and the personal relationship domain (Klein &
Milardo, 1993). These studies, however, have failed to systematically examine triadic
relationships, particularly as they change and adapt across the life span. What may be
important, at least for individuals in these triadic relationships, is to understand the extent
to which individuals are influencing each other in these connected relationships.
If we go back to the model we were discussing concerning terrorist networks, we
see that also in that case the network description is insufficient. This is why, in everyday
language, we hear talk of ‘terror cells’, a notion which indicates a small group of
Social Aggregation
6
terrorists, who are in constant communication with each other and whose level of
cohesion (cognitive, ideological, organizational) is higher than the one offered by a
network model.
We need, therefore, a different model, which includes networks, but that also
allows for higher order groups, in which three or more people give rise to something
more than a set of dyadic conversations. Fortunately, such an object is well known in
mathematics, and in fact some would even go as far as claiming (see for instance Beck,
1979) that it is one of the fundamental building blocks of mathematics. We are referring
here to what is called “Abstract Simplicial Complex” and, which can be described (see
next section for a detailed and formally correct definition) as a higher dimensional
version of a network. When we look at isolated nodes, we have a 0-dimensional structure.
When we connect some (or all) of the nodes, we obtain a structure which is 1dimensional (though it may still have 0-dimensional components). In simplicial
terminology, nodes are known as 0-simplexes, while edges are known as 1-simplexes.
When we connect nodes with edges we have a simplicial complex. Moving to higher
dimension, a triangle is known as a 2-simplex. Note (very important point for what
follows), that a triangle (a 2-simplex) has three faces (we could call them sides) which are
segments, i.e. 1-simplexes. In turn, every 1-simplex has two faces (we could call them
border points) which are 0-simplexes. We can therefore think that our 2-simplexes must
be faces of something else we should call a 3-simplex. This is exactly the case, as 3simplexes are defined as tetrahedrons (i.e. pyramids with three faces, each of which is a
triangle).
Before we offer a few examples, and some formal definitions, we wish to point
out that the idea of using simplexes to describe and model social interactions is not totally
new, as it was developed by R. Atkin in his theory of Q-analysis (Atkin 1972, 1974,
1977) as well as in Legrand’s recent work (Legrand, 2002). However, as the reader will
see, there are many fundamental differences between our approach and Atkin’s. To begin
with, Atkin uses simplicial complexes to represent relations in a Cartesian product of two
sets (i.e. binary relations). In our point of view, simplicial complexes indicate actual
social entities, whereas relations (binary and n-ary alike) that comprise them are, literally,
their faces. Just as a human body is more than a mere aggregate of its organs, social
Social Aggregation
7
groups are more than their constituents. Even more crucial is the fact that Atkin’s model
is essentially static. We introduce, on the other hand, a probabilistic, dynamic approach to
the evolution of such simplexes, along the lines drawn by Erdos in his approach to
random graphs. Social groups (and the simplicial objects representing them), are to be
thought of as living, evolving, beings. Dynamics is, here, at the very core of the theory.
We will give the details of the construction we propose in the next section, but for
now we want to point out that a family triad as the one illustrated before can now be
modeled in four ways:
Nodes {H,W,P}; Edges {HW,HP}
Nodes {H,W,P}; Edges {HW,WP}
Nodes {H,W,P}; Edges {HW,HP,WP}
Social Aggregation
8
Nodes {H,W,P}; Edges {HW,HP,WP}; Triangles {HWP}.
The four representations are four different simplicial complexes. The first three
are one-dimensional with three 0-simplexes (the nodes) and either two or three 1simplexes (the edges). The last representation is the only two dimensional simplicial
complex, with three 0-simplexes, three 1-simplexes, and one 2-simplex (the triangle
which represents the family relationship). It is clear, even without developing a full
theory of simplicial models, that we now have at our disposal a more complex and
satisfactory way to describe the possible outcomes of the arrival of a parent in a dyadic
family.
In this paper, we want to offer the theory of simplexes as a new and robust model
to describe social aggregation. We intend this to be the first in a series of papers where
we approach several important issues, and we want to use this work to set the foundation
for our future research. To this purpose, section 2 will formalize and recall some notions
from the theory of simplicial complexes. Section 3 will discuss what could be called
random simplicial topology, namely a model of how we can generate random simplicial
complexes, in an attempt to offer a model for group formation and evolution. Finally,
Section 4 will begin a discussion on how these models can be implemented in a
computational fashion, which will make them useful for social scientists not only to
understand and visualize social aggregations, but also to offer theoretical tools to help
strengthen or weaken these aggregations.
Social Aggregation
9
An introduction to simplicial complexes
This section is devoted to a brief introduction to the theory of simplicial
complexes. This is a rather complex theory, which is very well established, and the reader
interested in more detail is referred to several standard references in algebraic and
combinatorial topology such as (Munkres, 1984). In this section, however, we will only
provide those notions which are necessary for the development of our models, and we
will recall some of the fundamental results, of course without proving them. The next few
paragraphs may appear a bit technical, and in particular the reader may be baffled as to
why we are using higher dimensional spaces (instead of limiting ourselves to the well
known three-dimensional space). The reason for that will be apparent once we start
introducing some concrete examples. We are asking, therefore, patience while reading the
next few pages.
The notion of simplex is a generalization to higher dimension of the intuitive
notions of triangles and pyramids. But let us proceed in order. A simplex of dimension
zero is a point P. A simplex of dimension one is a segment, including its end points. If P
and Q are the end points, we denote by [P,Q] this one-simplex. A simplex of dimension
two, or two-simplex, is the inside of an equilateral triangle [P,Q,R] whose vertices are the
points P, Q, and R. As for the 1-simplex, we ask that the triangle itself is part of the 2simplex. A 3-simplex (or a simplex of dimension three) [P,Q,R,S] is the inside of a
tetrahedron (the fancy mathematical name for the pyramid with four faces, each of which
is an equilateral triangle), again including the faces. The next steps are a bit harder to
visualize as they use higher dimensions. We will introduce the simplex of dimension
four, and then we will give the general definition of a simplex in dimension n. To
construct the simplex in dimension four, we start with a regular tetrahedron [P,Q,R,S]
(the 3-simplex) and then we take a point T in the fourth dimension, such that the distance
(in the four-dimensional space) between T and each of the points P,Q,R, and S is the
same. We then fill up the object [P,Q,R,S,T] in the four-space by taking all the points in
the four space which are “inside” the boundary created by the points P,Q,R,S, and T (in
the same way in which a triangle is the set of points “inside” three given points P,Q, and
Social Aggregation
10
R. The resulting object is a 4-simplex or pentatope. Mathematicians have an expression
to indicate this filling-up process, namely they say that the pentatope is the convex hull of
P,Q,R,S, and T. Let us describe what convex hull is, since this notion will be used in the
sequel. Let us start with two points P and Q. The convex hull of these two points is the
segment between them. If we now take three points P, Q, and R, the convex hull of them
is obtained by taking the convex hull of each pair, and then the convex hull of all the
pairs of points which have been obtained through this procedure, and so on, until there is
nothing else left to do. So, if three points are given, the first step is to take the convex hull
of P and Q, i.e. the segment [P,Q], then the convex hull of Q and R, i.e. the segment
[Q,R], and then the convex hull of P and R, i.e. the segment [P,R]. We have obtained the
boundary of the triangle with vertices P, Q, and R. We now continue the process by
constructing the convex hull of each pair of points taken from the boundary. So, for
example, we take a point on [P,Q] and a point on [Q,R] and we build their convex hull.
This will give us a segment which falls “inside” the boundary. Once we are finished with
this process, we obtain the entire triangle [P,Q,R]. At that point, even if we try to build
more convex hulls, we don’t find any new points, and the process is actually over. The
triangle [P,Q,R] is the convex hull of the three points P,Q, and R. Similarly, the 3simplex [P,Q,R,S], i.e. the tetrahedron, is the convex hull of the points P, Q, R, and S.
The road to generalization is now obvious. In dimension n, the n-simplex is the convex
hull of n+1 points P0, P1, P2, …, Pn . For the sake of completeness, we need to mention
that the points need to be chosen in such a way that the directions P1-P0, …, Pn-P0 are all
different.
We should mention that this way of describing simplexes can be at once
simplified and generalized if one has enough language from the theory of categories, and
the definition becomes a bit quicker in that case. The mathematically oriented reader is
referred to the classic book (May, 1993) for more details. We wish to point out that using
the categorical machinery it is possible to define simplicial objects in environments that
have more structure than just being a set. This is important, because it allows us to model
groups wherein individuals and their relations (dyadic and polyadic) carry additional
pieces of information (for instance, weights, labels, different types, etc.). This
Social Aggregation
11
“categorical” approach will be pursued by the authors in a more abstract and
mathematically oriented work (Anonymous, 2006).
The objects we are ultimately interested in, however, are not simplexes, but rather
simplicial complexes. To define what a simplicial complex is, we need to introduce the
notion of “face” of a simplex. This is the natural generalization of the notion of “face” of
a polyhedron. So, the faces of a 1-simplex are its two end points. The faces of a 2simplex (a triangle) are its three sides, the faces of a 3-simplex (a pyramid) are its four
faces (each of which is a triangle), the faces of a 4-simplex (a pentatope) [P,Q,R,S,T] are
the five 3-simplexes which one obtain from the pentatope by ignoring one vertex at a
time (i.e. they are [P,Q,R,S]. [P,Q,R,T], [P,Q,S,T], [P,R,S,T], and [Q,R,S,T]), and so on
and so forth. Note that the face of an n-simplex is always an (n-1)-simplex.
Given a simplex, we can consider its faces, and then the faces of its faces, and so
on. This leads us to the notion of m-face of a simplex. Quite simply, given an n-simplex
S=[P0,P1,…,Pn], we say that the points Pj are the 0-faces of S, we say that the 1-simplexes
[Pi,Pj] are the 1-faces, the 2-simplexes [Pi,Pj,Pk] are the 2-faces, etc.
We can now define a simplicial complex. By this word we mean the union of
several simplexes (possibly of different dimensions) such that if the intersection of two
simplexes is not empty, then the intersection itself is an m-face for both simplexes. Here
below we have a picture of a simplicial complex, and a pictures of an object which, while
the union of simplexes, is NOT a simplicial complex.
Social Aggregation
12
Before we introduce some of the computational tools which allow us to describe
properties of simplicial complexes, we would like to offer some examples which explain
why higher dimensions are needed to use simplicial complexes as models for social
aggregation. What follows are simply examples to show how simplicial complexes can
be used to visualize such aggregations. No computations of any type are attempted.
Example 1. A family is composed of dad, mom, and a child. This is a strong
family, in which the three components actually come together and work as a unit. The
simplex which describes this family is a 2-simplex. The nodes are D, M, C (dad, mom,
and child); the 1-faces describe the interaction between any two of them (D and M, or D
and C, etc.), and the entire simplex describe the interaction of the three components as a
family. The graphical representation is indicated below.
Example 2. Suppose now the family splits as the consequence of a divorce. The
new structure is a simplicial complex of dimension 1. It is made of three nodes (as
before), and three 1-simplexes (the relationship between mom and child, the relationship
Social Aggregation
13
between dad and child, and the relationship – albeit functioning differently than when
married and most often functioning negatively post-divorce – between mom and dad).
What is now missing is the 2-face, because the three components do not come together as
an intact family any more. The representation is below.
Example 3. Mom now marries again, and the child lives mostly with mom and her
new stepdad. The new family M, C, S (for stepdad) is a strong one, which operates as a
family. The representation now is a simplicial complex in which we have simplexes of
different dimensions. There are four nodes (D, M, C, S), and two triangles. One is the
triangle representing the original (now broken) family, and it involves nodes D, M, and
C. The dimension of this triangle is one, as only the edges exist. The second triangle
represents the new family, and it involves nodes M, C, and S. Because this is a strong
family, the representation is actually a 2-simplex (as we had originally in example 1).
Nodes S and D are not connected by an edge (they are the only ones in this picture not to
have a connection). So we have one simplex of dimension 2 (the family [M,C,S]), to
which we attach two simplexes of dimension 1 ([M,D] and [C,D]). These two 1simplexes are in turn attached on the 0-simplex D. See below for the picture describing
this situation.
Social Aggregation
14
Example 4. A variation on this example would occur if mom and dad reach a
point in which their relationship is so strained that they don’t even communicate any
more. This would destroy the edge connecting M and D, and the resulting picture would
be as follows:
Example 5. Building on example 3, let us now assume that the new family
[M,C,S] has a new baby B. The new baby is fully accepted by C, and the family now
must be represented as a 3-simplex [M,C,S,B]. On this simplex we have four faces, each
of which describes the interaction of the subfamily which one has when one of the
components is absent (when stepdad is away for work, [M,C,B] still work together as a
smaller family, and similarly, when C is away for visitation with her dad, [M,S,B] are a
fully functioning family, etc.). The new baby B and the original dad D are not connected,
and so the simplicial complex we need to represent this more complex family is now a
tetrahedron (representing [M,C,S,B]) and two 1-simplexes (the relations [M,D], and
[C,D]). See below for a graphic representation.
Social Aggregation
15
Example 6. Finally, a second little baby is born in this new family, let’s call her A.
The representation now requires us to use a pentatope [M,C,S,B,A] to describe the new
family, plus the old 1-simplexes. This last example, which we cannot graphically
represent, explains why we need higher dimensional simplexes when we want to describe
increasingly complex social structures.
The few examples we have given above indicate how this mathematical tool can be
used in a dynamic way to represent social aggregates, and their change through time. We
can think of a social structure as being a function F from the set of positive numbers with
values in the space of simplicial complexes1. F(0) represent the initial state of a social
structure, and as times t increases, F(t) represents the evolution of that social structure. As
new relationships are born, the complexity of F(t) will increase, but at the same time F(t)
can become simpler if some of the members of the social structure die (or leave the
structure in some other less dramatic way), or if relationships are broken (this reduces the
dimensionality of the simplicial complex, at least in some of its components). Thus,
simplicial complexes can easily be seen as helpful to the social scientist for at least the
following three applications:
1. Simplicial complexes can be used to model and visualize a social structure.
2. Computational tools from the theory of simplicial complexes can be used to
describe and modify a given social structure.
3. A dynamical theory of simplicial complexes can be used to simulate the growth
and structure of a given organization, with known initial conditions.
This current paper, as the first one in which the theory of simplicial complexes is
applied to social aggregation, focuses essentially on 1. Nevertheless, we want to give
some ideas of how 2 and 3 could be possible. Specifically, in this section we want to
discuss some of the computational tools which are available to describe simplicial
complexes. In the next section we will discuss how to set up a dynamical theory of
simplicial complexes. Both ideas will be then expanded in future papers. Let us start with
1
In a subsequent work (Mannucci, Sparks, Struppa, 2006b) we will consider more general frameworks for
social dynamics. For the mathematically savvy reader, we say that we shall employ sheaf models from a
small category ( generalized “time”) with values in the category of simplicial sets (or even simplicial
objects in some ambient universe).
Social Aggregation
16
some very simple examples, and then we will try to gain some theoretical ideas from
them.
Example 7. Consider simplex S made of two 0-simplexes [P] and [Q]. It
represents two people who do not interact with each other. Two strangers. Now consider
simplex T=[P,Q], which represents those two same people who now interact (maybe as
friends, maybe as enemies, maybe as lovers…). The question is: how do we distinguish
(from a mathematical point of view) these two simplexes? There are at least two ways of
doing this. One is to note that S has dimension zero (its components are all 0-simplexes),
while T is of dimension one. This teaches us that “dimension” is a numerical concept
which can be used to describe different simplexes or simplicial complexes. Another way
to distinguish S from T is to note that “S is made of two disconnected pieces”, while T is
just one piece (I can navigate along T, without ever lifting the pencil, which I clearly
cannot do on S). We say that T is connected, and that S is not. Or, we say that the 0-th
Betti number of S is two (S is made of two pieces) while the 0-th Betti number of T is
one (T is made of a single piece). The notion of connectedness, by itself, is not too useful,
as it would not distinguish a simplicial complex made of three unconnected nodes from
one made of two unconnected nodes. More relevant is the Betti number (the name comes
from a famous Italian mathematician of the end of the nineteenth century), which tells us
how many disjoint connected components make up the simplicial complex.
Example 8. Unfortunately, the 0-th Betti number is not enough to characterize a
simplicial complex. For example the complex S={[P],[Q]} and the complex
T={[P,Q],[R,S]} both have 0-th Betti numbers equal to two. Complex S is made of two
disconnected points. Complex T is made of two disconnected segments. Either way, they
are made of two disjoint components, and so they have the same 0-th Betti number. What
would diversify them, on the other hand, would be to note that S has no components of
dimension one. To do so, we introduce the 1-st Betti number, which counts exactly how
many “holes” of dimension one exist. In this way, we see that S’s 1-st Betti number is
zero, while T’s first Betti number is two. Thus, we have been able to differentiate these
two simplicial complexes, through the use of two Betti numbers. If we use the standard
notation b0 and b1 to denote the Betti numbers, we see that b0(S)=b0(T)=2, but b1(S)=0,
while b1(T)=2.
Social Aggregation
17
The definition of Betti numbers can be extended to higher dimensions, and this is
necessary to differentiate examples which otherwise would be undistinguishable.
Regretfully, the computation of these higher order Betti numbers (as well as their precise
definition) ends up being a bit more complicated than what appears from the previous
two examples. For example, if we consider the simplicial complex T of example 1 (the
triadic family before the divorce), we have b0(T)=1, b1(T)=0, and b2(T)=1. On the other
hand, if we consider the simplicial complex S of example 2 (a triadic family after a
divorce), we see that b0(S)=1, b1(S)=1, and b2(S)=0 (we could say that a one-dimensional
hole in the social tissue has been created).
But what is important from our point of view, is that a general theorem of
simplicial topology says that if two complexes have the same Betti numbers for every
dimension, then the two complexes are equivalent, in some suitable sense (they will have
the same degree of connectivity in each dimension). Equivalent complexes, from our
point of view, would represent social aggregations with the same degree of cohesiveness.
Unfortunately homological/homotopical techniques only detect differences up to socalled homotopy equivalence (i.e. two objects are considered the same if they can be
deformed into one another without breaking them), which is weaker than topological
isomorphism (here objects are considered indistinguishable if all of their topological
properties are indeed the same). Therefore it may be useful, in certain context, to consider
the finer notion of topological isomorphism to distinguish between social groups. For
instance, a couple and a single are homotopically equivalent (neither has any “holes” and
the single is a “retract” of the couple), but they are clearly very different social groups
(their topology is indeed quite different: the couple can break down, leaving two isolated
points).
The reader who is interested in more detail in the theory of Betti numbers (and the
related notions of homology and cohomology) is referred for example to standard
references in algebraic topology, such as (Munkres 1984).
Social Aggregation
18
Random simplicial complexes
In this section we want to set the stage for a theory designed to model the
evolution of simplicial complexes, and as such designed to model the evolution of social
aggregates. The ideas described in this section are novel and due to the authors. However,
we acknowledge the influence of the early fundamental work of Erdos and Renyi on
random graphs (Erdos & Renyi, 1960; Bollobas, 2001) The work of Erdos and Renyi was
extremely influential, not just within the mathematical community, but also because it
spawned a series of important works on evolving networks, beginning with (Watts 1999a,
1999b). It was the availability of these tools, we believe, that made it so appealing to try
to use network theory to describe and discuss terrorist groups. Unfortunately, as we have
pointed out, graph models seem insufficient, and we hope that the ideas contained in this
section will offer a more powerful and accurate model.
Let us begin by recalling the simple model of Erdos and Renyi. They consider a
graph with undirected edges (i.e. a simplicial complex of dimension at most one), and
they make the following basic assumptions:
(i) the total number of nodes in the graph is fixed, N;
(ii) the probability that two arbitrary nodes are connected equals a fixed number p.
It is obvious to see that, on the average, the graph will contain pN(N-1)/2 edges, and in
fact Erdos and Renyi go on to compute other important numerical characteristics of these
graphs. It is worth emphasizing that, despite the obvious simplicity of this structure,
random graphs actually turn out to be quite applicable, and a very interesting
mathematical object. It is also clear that more complicated objects can be obtained, for
example, if one allows the probability p not to be a constant on the graph, or if one allows
the number N to increase as time increases. We are now ready to introduce our first new
concept.
Definition 1. A random simplicial complex S(n,N,p) of size n, dimension at most
N, and constant probability p is constructed as follows. First we fix a number n of points,
which are the 0-simplexes of the complex. We call S0 the set of these numbers. Given
two vertices, the likelihood that they are connected by an edge is p. A realization of this
Social Aggregation
19
random process yields the 1-skeleton of the complex, which we call S1. Given three
vertices A, B, and C such that they are all connected to each other, the likelihood that
[A,B,C] is a simplex in the complex is p. The realization of this process now yields the 2skeleton of the complex, denoted by S2. We now continue this construction up to
simplexes of dimension N. We set S=SN.
Note that we use an abuse of language, since we call a random simplicial complex
what is actually a realization of a random process. We will be careful to distinguish the
two when and if the distinction becomes relevant. For now, let us give a few examples.
Example 1. A random simplicial complex S(n,1,p) is nothing but a random graph
according to Erdos and Renyi.
Example 2. Let us consider S(3,2,0.5), and let us study the possible realizations of
this random simplicial complex. First we note we start with three points A, B, C. The
likelihood that they are connected (two at a time) is 0.5, which means the following
realizations can occur at first stage (all of them with likelihood 0.125):
S0 = {A, B, C} (none of the possible edges exists);
S1 = {A, B, C, [A.B]}; S1 = {A, B, C, [A,C]}; S1 = {A, B, C, [B,C]} ;
S1 = {A, B, C, [A.B], [A,C]}; S1 = {A, B, C, [A.B], [B,C]};
S1 = {A, B, C, [A.C], [B,C]); and S1 = {A, B, C, [A.B],[A,C],[B,C]}.
In all cases, except the last one, the simplicial complex is complete, as no 2simplexes are possible, but the last case presents the possibility of creating a 2-simplex or
not, again with probability 0.5. Thus, the possible realizations of the complex are, in fact,
the first seven listed above (with probability 0.125) and then two more realizations (each
with probability 0.0625), namely S1 = {A, B, C, [A.B],[A,C],[B,C]}, and S2 = {A, B, C,
[A.B],[A,C],[B,C]}.
It is clear how quickly complex this procedure can become. The point here is that
we have a mechanism to create realizations of simplicial complexes, and that we can now
envision situations in which we can put in motion such a mechanism, to describe
evolving random simplicial complexes. This is the definition we are aiming for:
Social Aggregation
20
Definition 2. Let n(t) be a non-negative integer function defined on the set of
positive integers. Let p(t) be a functions, whose values are between 0 and 1, defined on
the set of positive integers. An evolving random simplicial complex is a discrete time
process S(t) where t assumes integer values, and for every value t, the expression S(t) is a
random simplicial complex obtained from S(t-1) by adding n(t) new points, and creating
new simplexes with probability p(t), according to the rules of definition 1.
Definitions 1 and 2, together with a series of variations which we will discuss
elsewhere, are the foundation of a theory of simulation for social aggregates. Clearly,
definition 1 provides the stage for the description of the possible outcomes when a group
of individuals have opportunities for the creation of relationships. However, it presents a
stable system, which does not account for arrival of new individuals, nor does it include
the possibility of relationships dissolving through time. Therefore, any realistic model
will have to account for a variety of different options. For example, instead of having a
probability p which does not depend on the dimension of the simplexes which are being
added, one may imagine that such a value may be an increasing function as the dimension
grows. A second option which must be considered has to do with the possibility to delete
existing connections, as time grows. So, instead of a single probability p, or p(t), we need
to envision two of them; one describes the likelihood that new connections or simplexes
are created, while the other describes the likelihood that an existing connection is
removed. Rules will have to regulate the implications of the removal of some connections
(for example if we remove [A,B] from [A,B,C], then we lose the 2-simplex, and not just
the 1-simplex [A,B]. This rule is a consequence of the fact that any modification should
still maintain the structure of simplicial complex.
It should be clear from the above that the simple-minded model and its variants
that we have sketched can be enriched as needed. For instance, just like recent years have
witnessed the shift from random graph to other forms of networks that accounts for hubs,
we envision random simplicial complexes where certain nodes or groups of nodes tend to
be hubs for higher-dimensional aggregates. More complicated simplicial models can be
generated by assuming some correlation between dynamics in two consecutive
dimensions (for instance, the probability of forming a group between individuals that are
already connected to one another should be much higher than the one afforded by a group
Social Aggregation
21
of individuals meeting at a party for the first time), as well as the possibility of using
conditional probabilities. While we intend to pursue a systematic study of these
possibilities in a subsequent work (Anonymous, 2006), we would like to close this
section with a description of one additional construct, which may prove useful in the
modeling we have proposed.
To begin with, we note that unless we have some information on the structure
which we want to model it will be impossible (or at least very hard) to create a reasonable
model. In the examples above, for instance, we have assumed that all nodes are
equivalent, and that all possible edges (and faces of higher dimensions) are equivalent, in
the sense that the probability distribution is uniform. But of course this is a rather coarse
approximation of reality. In real life, we know that some people are more likely than
others to create connections. In fact, the entire Small World theory discussed in (Watts,
1999b) is predicated on the existence of ‘hubs’ (i.e. of nodes that have a much larger
connectivity with other nodes). It is the existence of these hubs that allows the theory of
the six degrees of separation (Watts, 2004) to be formulated.
In our setting, these hubs must exist at higher dimensions as well. So, some nodes
are more likely to be connected to other nodes, and other nodes may have higher
likelihood to be parts of faces or even n-faces. This is clearly the characteristic of a leader
(if we are thinking in terms of social networks) or of an opinion leader. Thus, we can
define a notion of simplicial leadership rank. To do so, we need a small modification of
the definition of random simplicial complex. Recall, a random simplicial complex is a set
of vertices and a fixed, or constant, probability, which describes the likelihood that edges
connect given vertices, and that higher order simplexes exist. We now define the notion
of hierarchical random simplicial complex as follows.
Definition 3. A hierarchical random simplicial complex is a triple H=(n,N,p),
where n and N are positive integers as before, but p is an N-dimensional vector of
probabilities (real numbers between 0 and 1). The interpretation is that for every pair of
vertices Pi and Pj, the value p[(i,j)] describes the probability that Pi and Pj are connected
by an edge. Similarly, for every triple of vertices Pi, Pj, and Pk, the value p[(i,j,k)]
describes the probability that, if the three vertices are connected by three edges, the
complex actually contains also the triangle [Pi,Pj,Pk]. Thus, we see that p is defined on
Social Aggregation
22
any subset of length N<=n of the set of the n vertices. We also see that p is invariant
under permutations (so that p[(i,j)]=p[(j,i)], as well as p[(i,j,k)]=p[(i,k,j)], etc.).
In order to simplify the next definition (which is complicated only from a formal
point of view, but actually quite simple from an intuitive point of view), we will assume
that if V and W are two vertices in a complex, we will denote by T(V,W,k) the set of all
collections of k vertices in the same complex, which do not contain either V or W.
Definition 4. A vertex V in a random simplicial complex is said to be a simplicial
leader of dimension k (k any integer between 1 and N) and order t (t a number between 0
and 100), or simply a (k,t)-hub, if for any vertex W different from V, the number of
elements T in T(V,W,k) for which p[(V,T)] > p[(W,T)] is bigger than t% of the number
of elements in T(V,W,k). Obviously, this definition simply means that (if t is close to
one, at least), the vertex V has larger probabilities than most of the other vertices, as far
as k-simplexes are concerned.
As we said before, specific information is needed if we are to assign appropriate
leadership indices to different nodes. In (Sparks-Bethea, 2002) for example, the author
shows how to apply social sciences methodology to determine or at least approximate
possible indices to different components of family triads. Even though (Sparks-Bethea,
2002) predates this paper by several years, we think that many of its conclusions can be
neatly framed in this more mathematically rigorous model, and the ideas suggested there
can be useful to create reasonable models of family interactions.
It is also worth noticing that one could take issue with our specific definition of
simplicial leader. One may point out that such a definition reflects a sort of democracy in
the society described by our model. Indeed, we define a leader as somebody who is
involved in a majority of the social aggregations in the society. Different structures (for
example the structure of a terrorist organization such as Al Qaeda) may need to be
modeled differently. A leader such as Osama Bin Laden may not be involved in large
groups where ideas are discussed. Rather, he is probably at the top of a one-dimensional
pyramid which connects him to a large number of cells. Once he issues a declaration, he
is able to influence a large number of cells, who automatically react to his orders, even
though none of the members of these cells can be said to be part of the leadership group.
In other words, defining and operationalizing the term leader will likely result in
Social Aggregation
23
differential representations depending upon the social structure within which the leader is
embedded.
Some computational remarks
This last section contains an initial attempt to sketch a computational approach to
these ideas, as well as describe its possible application in concrete social analysis. The
notes below are the bare bones of our approach, and serve as an indication of future lines
of research.
The reader may recall from Section 3 that, assigned a social group (i.e. a
simplicial complex), its Betti numbers gauge its level of connectivity in the various
dimensions. A Betti number of dimension n measures how many n-dimensional holes are
contained in the group as a whole. What kind of information does this number convey to
the social analyst? Let us assume the complex has third Betti number equal 1. This mean
that somewhere in the complex a three dimensional subcomplex is missing (i.e. a three
dimensional object assembled from simplices of dimension at most three), though all of
its faces are already there. Depending on the context, the analyst may conclude that the
complex is weak in terms of triadic relations. He or she may also anticipate, depending of
course on the statistics governing the complex, that such group will “complete” itself by
“filling the hole” at some later stage. Strategic war analysts may be interested in the
current status of the complex in an attempt to disconnect it in some critical areas (i.e.
disrupt its cohesiveness). Unlike networks, general simplicial complexes can be
connected or disconnected in higher dimensions than 1, an extremely important fact for
both strategic warfare and social interactions. For instance, terrorist organizations may
want to destroy cohesiveness in such social structures, whereas in family relationships,
the goal may be to selectively increase cohesiveness.
Many more techniques of social engineering can be inferred by these calculations.
We should point out that Betti numbers are in a sense the easiest “homological
invariants” of simplicial complexes. More refined tools from homological algebra and
homotopy theory are available, representing more accurate gauges of the complex.
Social Aggregation
24
But, how can we turn these principles into actual computer calculations? The
good news is that homological algebra can be tackled computationally. Indeed, the
original motivation for combinatorial objects such as simplicial complexes was precisely
their intrinsic computational nature. A considerable amount of work has been done to
automatize homological computations (Kaczynski et al., 2004).
The Computational Homology Project (CHOMP, n.d.) has pioneered a software
library to efficiently perform homological calculations. Software tools such as CHOMP
will have to be enriched to include the capability of generating random simplicial
complexes and track down their evolution. We envision the creation of an integrated
software environment that will also enable the analyst to interactively manipulate a social
complex and visualize parts of it by selecting sub-complexes and set their properties.
The fundamental, ground-level, paradigm of programming, as in (Knuth, 1997) is
that programs are nothing but a set of data plus a set of instructions. High level
programming requires ways to structure data according to specific methods. Up to this
point, most data structures are either relational (such as in data bases) or trees, linked
lists, queues, or even simple arrays and vectors. Our idea is to use abstract simplicial
complexes as intrinsically hierarchical and dynamic data structures. Abstract simplicial
complexes can be thought of as “realizations” of the simplicial paradigm in different
settings. In fact, mathematicians would say that they are functors from the template
simplicial category ∆, to the category of the data which we want to structure. Data
assembled in this way allow the consideration of a hierarchical structure, which is natural
in the setting of social interactions (where higher order interactions find their foundations
in lower level interactions), as well as the modeling of their rich dynamics (creation,
destruction, and evolution of cells within larger social groups). Our forthcoming
(Mannucci, Mircea, Sparks, Struppa, 2006a) article will develop these ideas in detail.
Before we conclude, let us also offer a few suggestions for possible measures one
may want to look when modeling social aggregates through simplex theory. Let us
assume we are looking at an organization such as the membership of Facebook.com; a
way to understand the nature of its members is through a monitoring of the usage of the
software, and the number of messages exchanged. Similarly, when studying a terrorist
organization, we may be able to intercept emails and phone messages. Their numbers will
Social Aggregation
25
offer us an insight on the structure of the simplicial complex describing the organization.
We can use those data to analyze and make predictions on the leadership structure of the
organization. The velocity of diffusion is another important element that can be
monitored to gain information on a partially unknown simplicial complex. Suppose, for
example, that we are able to infiltrate two small cells in a large terrorist organization. We
have a glimpse of what the structure looks like in those cells, but we have no information
as to how the cells are connected. We could, however, send a message through one of our
informants, and see how long it takes for the message to reach the other informant, who is
working at a completely different physical location. If repeated several times, with
different entry points, and different exit points, one may use this information to infer
properties of the entire organization. Although such information may be difficult to
obtain, we can still input some data to see how long it takes the organization to get to a
“chatter” level or state and subsequently how long it would take for such “chatter” to
dissipate. Such analyses would reveal to what extent the individuals in the organizations
are connected to each other or not.
In conclusion, we have offered a mathematical model, which describes social
aggregations, provides a way to mathematically classify them, and finally is flexible
enough to allow for computational experiments to study the natural evolution of such
social aggregations. The model we have proposed can be modified to accommodate
different needs on the part of the social scientist, and particularly contributes to a more
precise understanding of the underlying functions of interpersonal and intergroup
communication.
Social Aggregation
26
References
Atkin, R. (1972). From cohomology in physics to q-connectivity in social science. International
Journal of Man-Machines Studies, 4, 341-362.
Atkin, R. (1974). Mathematical Structure in Human Affairs. London: Heinemann.
Atkin, R. (1977). Combinatorial Connectivities in Social Systems. Basel: Birkhauser Verlag.
Ayres, J. (1983) Strategies to maintain relationships: Their identification and perceived usage.
Communication Quarterly, 31, 62-67.
Barabasi, A.L. & Jeong, H. & Ravasz, E. & Neda, Z. & Schuberts, A. & Vicsek, T. (2002).
Evolution of the social network of scientific collaborations. Physica A, 311, 590-614.
Barabasi, A.L. (2003). Linked: How Everything Is Connected to Everything Else and What It
Means. Plume.
Bateson, M. C. (1984). With a daughter’s eye: A memoir of Margaret Mead and Gregory
Bateson. New York: Morrow.
Beck, J. (1979). Simplicial sets and the foundations of analysis. Applications of sheaves (Proc.
Res. Sympos. Appl. Sheaf Theory to Logic, Algebra and Anal., Univ. Durham, Durham,
1977), pp. 113--124, Lecture Notes in Math., 753, Springer, Berlin, 1979.
Beisecker, A. E.(1989). The influence of a companion on the doctor-elderly patient interaction.
Health Communication, 1, 55-70.
Bollobas, B. (2001). Random Graph Theory , Cambridge Studies in Advanced Mathematics ,
Cambridge University Press, 2001
CHOMP (n.d.). The Computational HOMology Project. Retrieved March 31, 2006, from
http://www.math.gatech.edu/~chomp/
Social Aggregation
27
Dorogotsev, S.N. & Mendes, J.F.F. (2001). Evolution of Networks. Retrieved March 31, 2006,
from http://xxx.lanl.gov/abs/cond-mat/0106144
Erdos, P. & Renyi, A. (1960). On the evolution of random graphs. Magyar Tud. Akad. Mat.
Kutató Int. Közl. 5, 17—61.
Fisher, P. H. (1953). An analysis of the primary group. Sociometry, 16, 272-276.
Gallois, C., & Giles, H. (1998). Accommodating mutual influence in intergroup encounters. In
M. T. Palmer & G. A. Barnett (Eds.), Mutual influence in interpersonal communication:
Theory and research in cognition, affect, and behavior (Vol. 20, Progress in
Communication Science, pp. 135-162). Stamford, UK: Ablex.
Gould, P. (1980). Q-analysis, or a language of structure: An introduction for social scientists,
geographers, and planners. International Journal of Man-Machine Studies, 13, 169-199.
Hajek, C. & Giles, H. (2002). The old man out: An intergroup analysis of intergenerational
communication in gay culture. Journal of Communication, 52, 698-714.
Harwood, J., Giles, H. & Ryan, E. B. (1995). Aging, communication, and intergroup theory:
Social identity and intergenerational communication. In J. F. Nussbaum & J. Coupland
(Eds.), Handbook of communication and ageing research (pp. 133-159). Hillsdale, NJ:
Lawrence Erlbaum Associates, Inc.
Harwood, J., & Sparks, L. (2003). Social identity and health: An intergroup communication
approach to cancer. Health Communication, 15, 145-170.
James, J. (1953). The distribution of free-forming small group size. American Sociological
Review, 18, 569-570.
Janson, S. & Luczak, T. & Rucinski, A. (2000). Random Graphs. New York, Wiley.
Social Aggregation
28
Kaczynski, T. & Mischaikow, K. & Mrozek, M. (2004). Computational Homology. New York:
Spinger Verlag.
Klein, R. & Milardo, R.M. (1993). Third party influence on the management of personal
relationships. In S. Duck (Ed.), Social context and relationships (pp. 55-77). Newbury
Park, CA: Sage.
Knuth, D.E. (1997), The Art of Computer Programming (3rd Ed.), Reading, MA: AddisonWesley.
Legrand, J. (2002). How far can Q-analysis go into social systems understanding? Retrieved
March 31, 2006, from www.afscet.asso.fr/resSystemica/Crete02/Legrand.pdf
Long, J.K. & Mancini, J.A. (1990).Aging couples and the family system. In T.H.Brubaker (Ed.),
Family relationships in later life (pp.29-47). Newburyt Park, CA: Sage.
Mannucci M., Mircea, V. Sparks, L., Struppa, D. (2006a) Simplicial Models of Social
Aggregation II. A computational perspective.
Mannucci, M., Sparks, L. , Struppa D. (2006b). Simplicial Models of Social Aggregation III.
A Categorical Framework for the Social Sciences, in preparation.
May, J.P. (1993). Simplicial Objects in Algebraic Topology. University Of Chicago Press.
McHale, J.P. (1995). Copoarenting and triadic interactions during infancy. The roles of marital
distress and child gender. Developmental Psychology, 31, 985-997.
Munkres, J.R. (1984). Elements of algebraic topology. Addison-Wesley Publishing Company,
Menlo Park, CA.
Newman, M.E.J. & Watts, D.J. & Strogatz S.H. (2002) Random graph models of social
networks. Proceedings of the National Academy of Sciences USA, 99, 2566-2572.
Social Aggregation
29
Noller P. & Fitzpatrick, M. A. (1990). Marital communication in the eighties. Journal of
Marriage and the Family, 52, 832-843.
Nussbaum, J. F. (1989). Life-span communication: Normative processes. Hillsdale, NJ:
Lawrence Erlbaum.
Sillars, A. L., & Wilmot, W. W. (1989). Marital communication across the life span. In J. F.
Nussbaum (Ed.), Life span communication: Normative processes (pp. 225-253 ).
Hillsdale, NJ: Lawrence Erlbaum.
Sparks, L. (Ed.). (2003). Cancer communication and aging [Special Issue]. Health
Communication, 15(2).
Sparks, L. (2005). Social identity and perceptions of terrorist groups: How others see them and
how they see themselves. In H. D. O'Hair, R. L. Heath, & G. R. Ledlow (Eds.),
Community preparedness and response to terrorism: Communication and the media (pp.
13-28). Westport, CT: Praeger Publishers.
Sparks, L. (forthcoming). Family decision-making. In W. Donsbach, J. Bryant, & R. Craig (Eds.)
International encyclopedia of communication. Oxford, England: Blackwell.
Sparks, L., & Harwood, J. (in press). Cancer, aging, and social identity: Development of an
integrated model of social identity theory and health communication. In L. Sparks, H. D.
O’Hair, & G. L. Kreps, (Eds.), Cancer communication and aging (pp. xx-xx). Cresskill,
NJ: Hampton Press.
Sparks, L., Kreps, G. L., Botan, C., & Rowan, K. (Eds.). (2005). Communication and terrorism
[Special Issue]. Communication Research Reports, 22(1).
Social Aggregation
30
Sparks-Bethea, L. (2002). The impact of an older adult parent on communicative satisfaction and
dyadic adjustment in the long-term marital relationship: Adult children and spouses’
retrospective accounts. Journal of Applied Communication Research, 30 (2), 107-125.
Watts, D.J. (1999a). Networks, dynamics and the small world phenomenon,
American Journal of Sociology, 105(2), 493-527.
Watts, D.J. (1999b). Small Worlds: The Dynamics of Networks Between Order and Randomness.
Princeton: Princeton University Press.
Watts, D.J. (2004). Six Degrees: The Science of a Connected Age. W. W. Norton & Company.
Wheeler L. & Nezlek, J. (1977). Sex differences in social participation. Journal of Personality
and Social Psychology, 35, 742-754.
Wilmot, W.W. (1987). Dyadic Communication (3rd ed.). New York: McGraw-Hill.
Wilmot, W.W. & Sillars, A.L. (1989). Developmental issues in personal relationships. In
J.F.Nussbaum (Ed.), Life-Span communication. Normative processes (pp. 120-132).
Hillsdale, NJ: Lawrence Erlbaum.
Mirco A. Mannucci
HoloMathics, LLC
[email protected]
Lisa Sparks & Daniele C. Struppa
George Mason University
[email protected]
[email protected]
Note: starting fall 2007, the last two authors will be at Chapman University.
Social Aggregation
Their new addresses will be:
[email protected]
[email protected]
31
| 5cs.CE
|
On the Maximal Halfspace Depth of
Permutation-invariant Distributions on the Simplex
arXiv:1705.04974v2 [math.ST] 21 Jun 2017
Davy Paindaveine1 , Germain Van Bever
Université libre de Bruxelles, ECARES and Département de Mathématique, CP 114/04,
50, Avenue F.D. Roosevelt, B-1050 Brussels, Belgium
Abstract
We compute the maximal halfspace depth for a class of permutation-invariant
distributions on the probability simplex. The derivations are based on stochastic ordering results that so far were only showed to be relevant for the
Behrens-Fisher problem.
Keywords: α-unimodality, Dirichlet distribution, Halfspace depth,
Majorization, Stochastic ordering
1. Introduction
Denoting as S k−1 := {x ∈ Rk : kxk2 = x0 x = 1} the unit sphere in Rk ,
the Tukey (1975) halfspace depth HD(θ, P ) = inf u∈S k−1 P [u0 (X − θ) ≥ 0]
measures the centrality of the k-vector θ with respect to a probability measure P = P X over Rk . Any probability measure P admits a deepest point,
that generalizes to the multivariate setup the univariate median; see, e.g.,
Proposition 7 in Rousseeuw and Ruts (1999). Parallel to the univariate
1
Corresponding author ([email protected])
Preprint submitted to some journal
June 22, 2017
median, this deepest point is not unique in general. Whenever a unique representative of the collection CP of P ’s deepest points is needed, the Tukey
median θP , that is defined as the barycentre of CP , is often considered. The
convexity of CP (see, e.g., the corollary of Proposition 1 in Rousseeuw and
Ruts, 1999) implies that θP has maximal depth. The depth of θP is larger
than or equal to 1/(k + 1); see Lemma 6.3 in Donoho and Gasko (1992).
In this paper, we determine the Tukey median and the corresponding
maximal depth for a class of permutation-invariant distributions on the probPk
ability simplex Dk := {x ∈ Rk : x1 , . . . , xk ≥ 0,
`=1 x` = 1}. The results
identify the most central location for some of the most successful models used
for compositional data. They have also recently proved useful in the context
of depth for shape matrices; see Paindaveine and Van Bever (2017).
The outline of the paper is as follows. In Section 2, we define the class of
distributions we will consider and state a stochastic ordering result on which
our derivations will be based. In Section 3, we state and prove the main
results of the paper. Finally, in Section 4, we illustrate the results through
numerical exercises and we shortly comment on open research questions.
2. Preliminaries
Let F be the collection of cumulative distribution functions F such that
(i) F (0) = 0 and (ii) F is concave on (0, +∞). In other words, F collects
the cumulative distribution functions of random variables that are (i) almost
surely positive and (ii) unimodal at 0. Any F in F admits a probability
2
density function f that is non-increasing on (0, +∞).
For an integer k ≥ 2 and F in F, consider the random k-vector
Xk = (Xk1 , . . . , Xkk )0 = (Vk1 , . . . , Vkk )0 /
Pk
`=1 Vk` ,
(1)
where Vk1 , . . . , Vkk are mutually independent and have cumulative distribution function F . The corresponding probability measure over Rk will be
denoted as Pk,F . Obviously, the random vector Xk takes its values in the
probability simplex Dk . This includes, for example, the Dirichlet distribution with parameter (α, . . . , α)0 ∈ Rk , obtained for the cumulative distribution function F = Fα of the Gamma(α, 21 ) distribution (that corresponds to
the probability density function x 7→ xα−1 exp(−x/2)/(2α Γ(α)) on (0, +∞),
where Γ is the Euler Gamma function). The unimodality constraint in (ii)
above imposes to restrict to α ≤ 1. Note that, irrespective of F , the mean
vector of Xk is µk = k −1 1k = k −1 (1, . . . , 1)0 .
To state the stochastic ordering result used in the sequel, we need to inP
P
troduce the following notation. For n-vectors a, b with n`=1 a` = n`=1 b` ,
we will say that a is majorized by b if and only if, after permuting the components of these vectors in such a way that a1 ≥ a2 ≥ . . . ≥ an and b1 ≥
P
P
b2 ≥ . . . ≥ bn (possible ties are unimportant below), r`=1 a` ≤ r`=1 b` , for
any r = 1, . . . , n−1; see, e.g., Marshall et al. (2011). For random variables Y1
and Y2 , we will say that Y1 is stochastically smaller than Y2 (Y1 ≤st Y2 ) if and
only if P [Y1 > t] ≤ P [Y2 > t] for any t ∈ R. To the best of our knowledge,
3
the following stochastic ordering result so far was only used in the framework
of the Behrens-Fisher problem; see Hájek (1962), Lawton (1968), and Eaton
and Olshen (1972).
Lemma 1 (Eaton and Olshen, 1972). Let W be a random variable with
a cumulative distribution function in F. Let Q1 , . . . , Qn be exchangeable
posPn
itiveP
random variables that are independent of W . Then, W/( `=1 a` Q` ) ≤st
W/( n`=1 b` Q` ) for any a, b ∈ Rn such that a is majorized by b.
In Eaton and Olshen (1972), the result is stated in a vectorial context
that requires the α-unimodality concept from Olshen and Savage (1970).
In the present scalar case, the minimal unimodality assumption in Eaton
√
and Olshen (1972) is that W is 2-unimodal about zero, which, in view of
Lemma 2 in Olshen and Savage (1970), is strictly equivalent to requiring
that W is unimodal about zero.
3. Main results
Our main goal is to determine the Tukey median of the probability measure Pk,F and the corresponding maximal depth. Permutation invariance
of Pk,F and affine invariance of halfspace depth allows to obtain
Theorem 1. The Tukey median of Pk,F is µk = k −1 1k .
Proof of Theorem 1. Let θ∗ = (θ∗1 , . . . , θ∗k )0 be a point maximizing HD(θ, Pk,F ) and let α∗ = HD(θ∗ , Pk,F ) be the corresponding maximal
depth. Of course, θ∗ ∈ Dk (if θ ∈
/ Dk , then HD(θ, Pk,F ) = 0). Denote by πi ,
4
i = 1, . . . , k!, the k! permutation matrices on k-vectors. By affine invariance of halfspace depth and permutation invariance of Pk,F , all πi θ∗ ’s have
maximal depth α∗ with respect to Pk,F . Now, for any ` = 1, . . . , k,
k!
k
k
1 X
1 X
1X
1
(πi θ∗ )` =
(k − 1)!θ∗` =
θ∗` = = (µk )` .
k! i=1
k! `=1
k `=1
k
Since this holds for any θ∗ maximizing HD(θ, Pk,F ), the result is proved.
Note that the unimodality of F about zero is not used in the proof of
Theorem 1, so that the result also holds at Fα with α ≥ 1. In contrast, the
proof of the following result, that derives the halfspace depth of the Tukey
median of Pk,F , requires unimodality.
Theorem 2. Let Xk = (Xk1 , . . . , Xkk )0 have distribution Pk,F with F ∈ F.
Then, HD(µk , Pk,F ) = P [Xk1 ≥ 1/k].
The proof requires the following preliminary result.
Lemma 2. For any positive integer k, let hk,F = P [Xk1 ≥ 1/k], where Xk =
(Xk1 , . . . , Xkk )0 has distribution Pk,F with F ∈ F. Then, the sequence (hk,F )
is monotone non-increasing.
Proof of Lemma 2. Since k −1 1k is majorized by the k-vector (k −
1)−1 (1, . . . , 1, 0)0 , Lemma 1 readily provides
k+1
k+1
k
1 X
1 X
Vk+1,` = P
Vk+1,1 ≥
Vk+1,`
= P Vk+1,1 ≥
k + 1 `=1
k+1
k + 1 `=2
Vk+1,1
Vk+1,1
= P 1 Pk+1
≥ 1 ≤ P 1 Pk
≥1
V
V
k+1,`
k+1,`
`=2
`=2
k
k−1
hk+1,F
5
k
k
1 X
1X
Vk` = P Vk1 ≥
Vk` = hk,F ,
k − 1 `=2
k `=1
= P Vk1 ≥
which establishes the result.
Note that this result shows that, for any F in F, the maximal depth, hk,F ,
in Theorem 2 is monotone non-increasing in k, hence converges as k goes
to infinity. Clearly, the law of large numbers and Slutzky’s theorem imply
P
that Vk1 /( k1 k`=1 Vk` ) → V11 /E[V11 ] in distribution, so that hk,F converges
to h∞,F = P [V11 ≥ E[V11 ]] as k goes to infinity. In particular, for F = Fα , the
limiting value is h∞,Fα = P [Zα > α], where Zα is Gamma(α, 1) distributed.
We can now prove the main result of this paper.
Proof of Theorem 2. We are looking for the infimum with respect
to u = (u1 , . . . , uk )0 in S k−1 , or equivalently in Rk \ {0}, of
X
k
1
u` Xk` ≥ ū
≥0 =P
u` Xk` −
p(u) := P
k
`=1
`=1
X
k
=P
X
k
u` Vk` ≥ ū
1
k
Vk` = P
`=1
`=1
where we wrote ū :=
k
X
Pk
`=1
X
k
(u` − ū)Vk` ≥ 0 ,
`=1
u` . Without loss of generality, we may assume
that u1 ≥ u2 ≥ . . . ≥ uk , which implies that u1 ≥ ū. Actually, if u1 = ū, then
all u` ’s must be equal to ū, which makes the probability p(u) equal to one.
Since this cannot be the infimum, we may assume that u1 > ū, which implies
that uk < ū. Therefore, denoting as m the largest integer for which um ≥ ū,
6
we have 1 ≤ m ≤ k − 1. Then, letting sm (u) =
Pm
`=1
(u` − ū), we may then
write
p(u) = P
X
m
k
X
(u` − ū)Vk` ≥
`=1
(ū − u` )Vk`
`=m+1
Pk
`=m+1 d` (u)Vk`
≤1 ,
= P Pm
`=1 c` (u)Vk`
where c` (u) = (u` − ū)/sm (u), ` = 1, . . . , m and d` (u) = (ū − u` )/sm (u),
P
` = m + 1, . . . , k are nonnegative and satisfy m
`=1 c` (u) = 1 and
k
X
P
Pk
(k − m)ū − k`=m+1 u`
`=m+1 (ū − u` )
Pm
=
= 1.
d` (u) = Pm
`=1 (u` − ū)
`=1 u` − mū
`=m+1
Since
Pm
`=1
d` (u)Vk` is unimodal at zero and since (c1 (u), c2 (u), . . . , cm (u))0
is majorized by (1, 0, . . . , 0)0 ∈ Rm , Lemma 2 yields
Pk
p(u) ≥ P
`=m+1
d` (u)Vk`
Vk1
≤ 1 = P Pk
Vkm
`=m+1
d` (u)Vk`
≥1 ,
where the lower bound is obtained for c1 (u) = 1 and c2 (u) = . . . = cm (u) =
0, that is, for u1 (> ū) arbitrary and u2 = . . . = um = ū. Now, since
(k − m)−1 1k−m is majorized by (dm+1 (u), . . . , dk (u))0 for any u, the same
result provides
p(u) ≥ P
1
k−m
Vkm
Pk
`=m+1
= P (k − m + 1)Vkm ≥
Vk`
k
X
≥ 1 = P (k − m)Vkm ≥
Vk`
k
X
`=m+1
Vk` = P Xk−m+1,1
`=m
7
1
≥
= hk−m+1,F ,
k−m+1
with the lower bound obtained for d` (u) = 1/(k − m), ` = m + 1, . . . , k, that
1
(u1 − ū), ` = m+1, . . . , k. Therefore, the global minimum
is, for u` = ū− k−m
is the minimum of hk−m+1,F , m = 1, . . . , k − 1, which, in view of Lemma 2,
is hk,F . This establishes the result.
Figure 1 plots the maximal depth hk,Fα as a function of k for several α,
where Fα still denotes the cumulative distribution function of the Gamma(α, 12 )
distribution. In accordance with Lemma 2, the maximal depth is decreasing
in k and is seen to converge to the limiting value h∞,Fα that was obtained
below that lemma. For any α, the maximal depth is equal to 1/2 if and only
if k = 2, which is in line with the fact that the (non-atomic) probability
measure Pk,Fα is (angularly) symmetric about µk if and only if k = 2; see
Rousseeuw and Struyf (2004). Interestingly, thus, the asymmetry of Pk,F
for k ≥ 3 would typically be missed by a test of symmetry that would reject
the null when the sample (Tukey) median is too far from the sample mean.
4. Numerical illustration
We conducted two numerical exercises to illustrate Theorems 1 and 2,
both in the trivariate case k = 3. In the first exercise, we generated N =
1, 000 random locations θ1 , . . . , θN from the uniform distribution over Dk
(that is the distribution Pk,F1 associated with the Gamma(1, 21 ) distribution
of the Vk` ’s; see, e.g., Proposition 2 in Bélisle, 2011). Our goal is to compare,
for various values of α, the depths HD(θi , Pk,Fα ), i = 1, . . . , N with the
depth HD(µk , Pk,Fα ). For each α, these N + 1 depth values were estimated
8
0.50
0.40
0.35
0.30
0.20
0.25
Maximal depth
0.45
α=1
α=0.5
α=0.25
10
20
30
40
50
k
Figure 1: Plots of the maximal depth hk,Fα as a function of k for several values of α,
where Fα denotes the cumulative distribution function of the Gamma(α, 21 ) distribution.
For each value of α considered, the limiting value as k goes to infinity is also showed.
9
by the depths HD(θi , Pn ), i = 1, . . . , N , and HD(µk , Pn ), computed with
respect to the empirical measure Pn of a random sample of size n = 100, 000
from Pk,Fα (these N +1 sample depth values were actually averaged over M =
100 mutually independent such samples). For each α, Figure 2 reports the
boxplots of the resulting estimates of HD(θi , Pk,Fα ), i = 1, . . . , N , and marks
the estimated value of HD(µk , Pk,Fα ). The results clearly support the claim
in Theorem 1 that µk = arg maxθ HD(θ, Pk,Fα ).
In the second numerical exercise, we generated, for various values of α
and n, a collection of M = 1, 000 mutually independent random samples of
size n from the distribution Pk,Fα . For each sample, we evaluated the halfspace depth of µk with respect to the corresponding empirical distribution Pn .
Figure 3 provides, for each α and n, the boxplot of the resulting M depth
values. Clearly, the results, through the consistency of sample depth, support
the theoretical depth values provided in Theorem 2. Actually, while the theorem was only proved above for α ≤ 1 (due to the unimodality condition in
Lemma 1), these empirical results suggest that the theorem might hold also
for α > 1. The extension of the proof of Theorem 2 to further distributions
is an interesting research question, that requires another approach (simulations indeed reveal that Lemma 1 does not hold if the cumulative distribution
function of W is Fα with α > 1). Of course, another challenge is to derive
a closed form expression for HD(θ, Pk,F ) for an arbitrary θ. After putting
some effort into this question, it seemed to us that such a computation calls
for a more general stochastic ordering result than the one in Lemma 1.
10
0.5
0.4
0.3
0.2
0.1
0.0
α=4
α=0.5
α=1
α=0.25
Figure 2: Boxplots, for k = 3 and for various values of α, of the sample depth values HD(θi , Pn ), i = 1, . . . , N , of N = 1, 000 locations randomly drawn from the uniform
distribution over the probability simplex Dk , where Pn denotes the empirical probability
measure associated with a random sample of size n = 100, 000 from the distributions Pk,Fα ,
with α = 4, 1, 0.5, 0.25. For each value of α considered, the sample depth HD(µk , Pn ) is
also provided (as explained in Section 4, these sample depths were actually averaged
over M = 100 mutually independent samples in each case).
11
α=0.5
n=1,000
n=10,000
n=100,000
n=1,000
n=10,000
n=100,000
0.50
0.45
0.35
0.35
0.40
0.45
0.50
α=0.25
0.40
0.45
0.40
0.35
0.35
0.40
0.45
0.50
α=1
0.50
α=4
n=1,000
n=10,000
n=100,000
n=1,000
n=10,000
n=100,000
Figure 3: Boxplots, for k = 3 and for various values of α and n, of M = 1, 000 mutually
independent values of the depth HD(µk , Pn ), where Pn denotes the empirical probability
measure associated with a random sample of size n from the same distributions Pk,Fα as
in Figure 2.
Acknowledgements
Davy Paindaveine’s research is supported by the IAP research network
grant nr. P7/06 of the Belgian government (Belgian Science Policy), the
Crédit de Recherche J.0113.16 of the FNRS (Fonds National pour la Recherche
Scientifique), Communauté Française de Belgique, and a grant from the National Bank of Belgium. Germain Van Bever’s research is supported by the
FC84444 grant of the FNRS.
References
Bélisle, C., 2011. On the polygon generated by n random points on a circle.
Statist. Probab. Lett. 81, 236–242.
Donoho, D. L., Gasko, M., 1992. Breakdown properties of location estimates
12
based on halfspace depth and projected outlyingness. Ann. Statist. 20,
1803–1827.
Eaton, M. L., Olshen, R. A., 1972. Random quotients and the Behrens-Fisher
problem. Ann. Math. Statist. 43, 1852–1860.
Hájek, J., 1962. Inequalities for the generalized Student’s distribution and
their applications. In: Select. Transl. Math. Statist. and Probability, Vol.
2. American Mathematical Society, Providence, R.I., pp. 63–74.
Lawton, W. H., 1968. Concentration of random quotients. Ann. Math.
Statist. 39, 466–480.
Marshall, A. W., Olkin, I., Arnold, B. C., 2011. Inequalities: theory of majorization and its applications, 2nd Edition. Springer Series in Statistics.
Springer, New York.
Olshen, R. A., Savage, L. J., 1970. A generalized unimodality. J. Appl.
Probab. 7, 21–34.
Paindaveine, D., Van Bever, G., 2017. Tyler shape depth. arXiv:1706.00666.
Rousseeuw, P. J., Ruts, I., 1999. The depth function of a population distribution. Metrika 49, 213–244.
Rousseeuw, P. J., Struyf, A., 2004. Characterizing angular symmetry and
regression symmetry. J. Statist. Plann. Inference 122, 161–173.
13
Tukey, J. W., 1975. Mathematics and the picturing of data. In: Proceedings
of the International Congress of Mathematicians (Vancouver, B. C., 1974),
Vol. 2. Canad. Math. Congress, Montreal, Que., pp. 523–531.
14
| 10math.ST
|
A Gradient Descent Technique Coupled with a Dynamic Simulation
to Determine the Near Optimum Orientation of Floor Plan Designs
Eugénio Rodrigues#1, Adélio Rodrigues Gaspar#2, Álvaro Gomes*3
#
ADAI–LAETA, Department of Mechanical Engineering, University of Coimbra
Rua Luís Reis Santos, Pólo II, 3030-788 Coimbra, Portugal
1
[email protected]
2
[email protected]
*
INESCC, Department of Electrical and Computer Engineering, University of Coimbra
Rua Luís Reis Santos, Pólo II, 3030-788 Coimbra, Portugal
3
[email protected]
Abstract
A prototype tool to assist architects during the early design stage of floor plans has
been developed, consisting of an Evolutionary Program for the Space Allocation
Problem (EPSAP), which generates sets of floor plan alternatives according to the
architect's preferences; and a Floor Plan Performance Optimization Program
(FPOP), which optimizes the selected solutions according to thermal performance
criteria. The design variables subject to optimization are window position and size,
overhangs, fins, wall positioning, and building orientation.
A procedure using a transformation operator with gradient descent, such as
behavior, coupled with a dynamic simulation engine was developed for the thermal
evaluation and optimization process. However, the need to evaluate all possible
alternatives regarding designing variables being used during the optimization
process leads to an intensive use of thermal simulation, which dramatically
increases the simulation time, rendering it unpractical. An alternative approach is a
smart optimization approach, which utilizes an oriented and adaptive search
technique to efficiently find the near optimum solution.
This paper presents the search methodology for the building orientation of floor
plan designs, and the corresponding efficiency and effectiveness indicators. The
calculations are based on 100 floor plan designs generated by EPSAP. All floor
plans have the same design program, location, and weather data, changing only
their geometry.
Dynamic simulation of buildings was effectively used together with the optimization
procedure in this approach to significantly improve the designs. The use of the
orientation variable has been included in the algorithm.
Keywords – building orientation; thermal performance; optimization; floor plans
1.
Introduction
An optimization procedure coupled with a building dynamic simulation
has traditionally encountered huge obstacles due to the computational time
that is required [1]. However, certain researchers have developed approaches
to overcome this difficulty, for example, Wang, Zmeureanu & Rivard [2]
presented a multi-objective genetic algorithm for building design
optimization coupled with dynamic simulation. To overcome the problem of
the necessary time to compute all the dynamic simulations, the methodology
was limited to one building design and only a few design variables, such as
building orientation, aspect ratio, window type, window-to-wall ratio, wall
type, and roof type. The methodology ignored possible arrangements for the
interior spaces and acknowledged the whole building as a single thermal
zone. In another approach, implemented by Peippo, Lund & Vartiainen [3], a
numerical multivariate optimization was employed. The procedure included
building orientation, building geometry (aspect ratio), thermal insulation,
collector slope, thermal mass, window type and area, collector type, PV
capability, energy storage volume, lightning type, lighting control, and
exhaust air heat recovery as design variables. The objective was to find the
economic optimum according to an energy consumption target. Haase &
Amato [4] determined the extent to which building location, climate and
orientation contributed to the thermal comfort of a building design. This
approach helped to determine the potential improvement of the design.
It may be evident that in order to overcome the problem of
computational resources and time needed to carry out a dynamic simulation,
the number of designs under optimization (population) should be limited or a
very abstract simulation model (fewer design variables) should be used for a
faster optimization procedure.
The building orientation plays a significant role in the overall thermal
performance of a building. Figure 1 illustrates three different building
performance curves for a floor plan design according to their angle of
orientation. The bottom axis represents the orientation angle, the top axis, the
cardinal direction, and the vertical axis, the amplitude of the curve. Every
curve is the result of exhaustively determining the building's thermal
discomfort performance for every orientation angle. Each point is the result
of the sum of the degree-hours of discomfort for all the spaces. As may be
observed, the curves differ greatly depending on several factors such as
window position and size, shape of the floor plan, self-projected shading,
space volume, thermal properties of the envelop elements, among others.
This procedure is very time consuming due to the necessity of carrying
out a dynamic simulation. Another drawback is that determining the
minimum value in the curve does not necessarily guarantee that this will be
the best orientation when the thermal optimization for the other design
variables has ended. The re-positioning of the windows and their sizing will
change the curve profile and consequently may shift the minimum point to
another position (angle). As each design variable - location, weather data,
constructive system, to name but a few, may modify the curve profile,
rendering it unique, it may be considered as the fingerprint of the building.
Building Orientation
East
South
West
North
1
North
0.5
0
0.1
0.2
0.3
0.4
amplitude
0.6
0.7
0.8
0.9
Floor Plan #1
Floor Plan #2
Floor Plan #3
0
15
30
45
60
75
90
105
120
135
150
165
180
195
210
225
240
255
270
285
300
315
330
345
360
angle
Figure 1: Three different floor plan designs (Floor Plan #1: One level three bedroom house,
Floor Plan #2: Two level three bedroom house, and Floor Plan #3: One level with two flats).
The researching team has developed a prototype tool to assist architects
in the early design stage of floor plans. It is made up of two modules, the
first module is an Evolutionary Program for the Space Allocation Problem
module (EPSAP) [5-7] which generates sets of different floor plan design
alternatives according to the architect's preferences and constraints. The
second module is a Floor Plan Performance Optimization Program module
(FPOP) [5], which improves each chosen solution according to thermal
performance criteria. The design variables subject to optimization in the
latter module are window position, window size, overhangs, fins, interior and
exterior wall positioning, and finally, the building orientation.
To overcome the problem associated to the time needed to carry out an
exhaustive enumeration using dynamic simulation, a type of gradient descent
technique was used as a transformation operator within the optimization
procedure to find a better building orientation within a margin of 2% of the
curve amplitude in relation to the global optimum. This signifies that on
average all floor plan design errors must be below this threshold.
Therefore, considering a problem consisting of 100 detailed floor plan
designs, in which one of the design variables to be optimized is the building
orientation, this paper presents the methodology that was used in the building
orientation transformation operator, the results of the efficiency and
effectiveness test, and the main conclusions.
2.
Methodology
There are many design variables, which may influence the identification
of the most suitable building orientation. Practitioner’s preferences,
constraints (minimum window size, window orientation, window-space area
ratio, etc.), the size of the walls and windows, the building system that was
used, are a few examples of these variables. The energy assessment of a
building requires the calculation of all possible combinations of all design
variables. This may become an impractical task due to the required
computational effort to carry out the dynamic simulation for all possible
combinations.
There are some rule of thumb principles and empirical knowledge in
relation to building orientation, which may be applied to potentially identify
the most adequate orientation (angle). However, the “best” orientation is
dependent on other design variables, which means that a detailed study
should be carried out for each set of (other) design variables. Accordingly, a
set of different design alternatives, with the same design program and user
preferences, may be subjected to an optimization procedure to predict their
performance in relation to the thermal performance criteria.
To avoid the previously mentioned problem - computational effort/time
- and the exhaustive enumeration of every building orientation angle, a type
of gradient descent technique which would change the variable until it
reached its near optimum value was employed as a transformation operator.
This transformation operator is invoked twice at different moments during
the optimization procedure for each floor plan design.
The FPOP optimization procedure consists of several sequential
transformation adaptive operators oriented to specific design variables where
the building orientation is one of those variables. The sequence of
transformation operators that were used included a window translation
operator, a building orientation operator, a window dimensioning operator, a
wall translation operator, a building orientation operator (second time), a
window’s overhang operator, and finally a window’s fins operator. The
building orientation operator is invoked at specific times during the
optimization procedure. After the re-positioning of the windows in relation
to the user’s predetermined orientation, and after the windows have been
dimensioned and the walls have been repositioned. The last operators place
the overhangs and fins according to the requirements and final orientation of
each space. This sequence of operators has proven to be the most effective
during the algorithm development.
The assessment of each design is completed using a cost function, which
penalizes the thermal discomfort measured in degree-hours for each space
(see Equation 1). The limits of thermal comfort conform to the European
Standard EN 15251 (2007) operative temperatures for buildings without
mechanical cooling systems [8]. When the outdoor mean temperature is
below 10ºC or above 30ºC, the algorithm will use those values to calculate
the adaptive operative temperature limits. The location and weather data for
all designs was for Coimbra, Portugal.
(1)
The hourly indoor air temperature results of each space, obtained by the
dynamic simulation, were then compared to the thermal comfort interval
temperatures. If the indoor temperature at a certain hour is outside the
interval, the cost function will penalize the design according to the deviation
from the comfort interval (T1 for the lower limit and T2 for the upper limit),
multiplied by a corresponding weighting value w1 for the lower limit and w2
for the upper limit (see Equation 2). In Equation 1, the Ns is the number of
spaces on the floor plan, the Nt is the number of hours in the year, Ti(t) is the
air temperature at the hour t for the space i, and fdf is the function that
calculates the difference between the hourly air temperature Ti(t) and the
thermal comfort limits according to Equation 2. The weight difference will
allow the user to specify their preferences for future cooling or heating
systems, if necessary.
(2)
In order to determine a suitable building orientation, the transformation
operator is a sort of gradient descent technique, which descends through the
building orientation curve (see Figure 1) from the initial orientation until it
reaches the lowest point possible (which corresponds to a better building
orientation for the set of design variables that have been considered). Each
step of the approach starts with the value of +30 degrees. This signifies that,
for instance, starting from the orientation angle 90 (East) it will determine if
the angle 120 has lower penalties. If this is correct, it continues with the
same direction. When it fails, the step becomes negative and decreases in
size, changing the direction of the search. Subsequently it continues until it
no further improvements can be found. When the termination condition of
this operator is reached the next operator is activated.
Due to the irregularities of the building orientation curve (see Figure 1)
and the benefit of the improvements, it was considered that the admissible
penalties deviations would be 2% of the penalties amplitude curve
(difference between maximum penalties and the minimum penalties points)
at this early design stage. This indicator was chosen instead of the angle
deviation indicator since the former could not be representative of the
algorithm performance. For example, the orientation curve of some buildings
could have two or more local minimums with similar values however with
greater angle differences. This occurs with floor plans with one or more axis
of geometric symmetry.
3.
Results and Discussion
A problem test was conducted to determine the effectiveness and
efficiency of the building orientation operator. A set of 100 floor plan
designs generated by the EPSAP module [5-7] were used. Every floor plan
had the same design program, one hall, corridor, kitchen, living room,
dinning room, two bathrooms, and three bedrooms. The topological
requirements were the same for all designs, and the initial window size was
10% of the corresponding space area in each design solution (see Figure 2
for a floor plan example).
Figure 2: The fittest individual of the FPOP module optimization.
The test was carried out with a dual-core computer at 3.2GHz with
1010MiB of RAM. The algorithm was built in JAVA programming language
and was implemented to take full advantage of parallel computing by using
two threads. The optimization procedure took 2 days, 13 hours and 33
minutes to conclude. Each simulation took approximately 28 seconds and
from a total of 7,833 simulations, the building orientation operator carried
out 993 simulations. This is nearly 2.7% of the simulations in comparison to
an exhaustive enumeration. This signifies that on average only 4 to 5 steps of
the operator are necessary to find the suitable orientation each time it is
invoked.
As one may observe on the right side of Figure 3, the building
orientation operator significantly impacts the improvement of the building’s
thermal performance. The relative contribution of this operator, compared to
the overall improvement of the floor plan design, is up to 14%, as illustrated
in left side of Figure 3.
Operators contribution by transformation
90
1400
1500
100
Operators relative contribution
800
700
penalties reduction
0
0
10
100
200
20
300
400
30
500
600
50
40
percentage
60
900
1000
70
1100
80
1200
1300
OT − Opening translation
BO − Building orientation
OD − Opening dimension
WT − Wall translate
OO − Opening overhang
OF − Opening fin
OT
BO
OD
WT
OO
OF
OT
BO
OD
WT
OO
OF
Figure 3: Operators’ relative contribution to the total improvement (left side) and operators’
contribution by transformation (right side).
The building orientation curves of the optimized individuals differ in
their profile. These curves are depicted in Figure 4 and the global optimum
(black dot) and determined optimization orientation values (red cross) are
marked. If the curves are adjusted to have their maximum value aligned to
the left side of the graph (see Figure 5) it is possible to observe the relative
position of the global minimums and the best optimization results. The
minimum value is usually located in the opposite area of the maximum value
within an interval of -90 to +90 degrees. This indicates that, after the
optimization procedure is concluded, the curve amplitude increases due to
the floor plan design being improved specifically for that location and
weather data. For instance, the re-positioning of the windows facing South,
West, and East will augment the penalties when the building is oriented with
the same façade to the North.
East
South
West
North
75000
80000
85000
Building Orientation
North
●
65000
penalties
70000
●
●
●
●
●
60000
●
●
●
●●
●
●
●●
●
●
55000
●
●
●
●
●
50000
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●● ● ●
●
●
●
●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
45000
●
●
Opt
0
15
30
45
60
75
90
105
120
135
150
165
180
195
210
225
240
255
270
285
300
Min
●
315
330
345
360
angle
Figure 4: Building orientation curves with corresponding minimum value and optimization
determined value.
Opposite
Max
75000
80000
85000
Building Relative Orientation
Max
●
●
65000
penalties
70000
●
●
●
●
60000
●
●
●
●
●
●
●
●
55000
●
●
●
●
●
●
●
●
●
●
●
●
●
●
50000
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
45000
●
●
Opt
0
15
30
45
60
75
90
105
120
135
150
165
180
165
150
135
120
105
90
75
60
●
45
30
Min
15
0
angle
Figure 5: Building orientation curves adjusted according to the maximum value.
Three indicators were used to analyze the effectiveness of the building
orientation operator. They are the difference in penalties in relation to the
global optimum (1), the angle of deviation in relation to the global optimum
(2), and finally, the percentage of error relatively to the curve amplitude (3).
The results for the indicators may be seen in Figure 6. The average
difference in penalties in relation to the global optimum (1) was
approximately 50, meaning that 50 degree-hours of air temperature divided
by all the spaces of the floor plan, and by the 8760 hours of the year, the
average temperature difference would be 0.0006ºC of improvement (top box
plot).
The second indicator (2) determines the average angle of deviation in
relation to the optimum angle (middle box plot of the Figure 6). The average
deviation in this case was 5 degrees and ranged from 0 to 15 degrees.
Some designs have a flat building orientation curve, meaning that the
deviation angle is not the best indicator of the algorithm success. For this
reason, the last indicator (3) assesses the percentage of error in relation to the
amplitude of the building orientation curve. In previous tests, the average
percentage of error was approximately 0.4%. This is less than half a percent
and it is below the previously stated threshold of admissible errors of 2%.
The range of this indicator varied from 0% up to 2.4%.
penalties
Difference in penalties from the global optimum
0
25
50
75
100
125
150
175
200
225
250
275
300
325
350
375
15
16
angle
Angle deviation from the global optimum
0
1
2
3
4
5
6
7
8
9
10
11
12
13
14
Percentage
Percentage of the curve amplitude
0.0%
0.2%
0.4%
0.6%
0.8%
1.0%
1.2%
1.4%
1.6%
1.8%
2.0%
2.2%
2.4%
Figure 6: Building orientation operator success indicators.
4.
Conclusion
In the approach presented in this paper, dynamic simulation of buildings
was effectively used together with an optimization procedure, allowing the
designs to be significantly improved. The use of an orientation variable was
also included in a building energy optimization procedure.
The capability of the FPOP module to improve a previously selected set
of floor plans according to thermal performance criteria equips the
practitioner with thermal information on their preferences and constraints
and the potential of each solution.
The building orientation is one of the design variables used in the
optimization of the FPOP module. A building orientation operator must take
two objectives into account, the correct solar orientation that minimizes
thermal discomfort penalties and it must allow other design variable
operators to fully reach their purpose. Each floor plan design has its unique
building orientation performance curve, meaning that any change in the
architectural elements such as walls, openings, spaces dimensions, and
materials alter the profile of the curve.
The proposed technique that was employed such as the building
orientation operator was tested for its effectiveness and efficiency using 100
floor plan designs with the same design program for the same location. The
test has shown that the deviation from the global optimum orientation falls
within the admissible error margin. Only 2.7% of all possible runs of the
dynamic simulation were necessary to find a near optimum building
orientation.
5.
References
[1] Magnier, L., Haghighat, F.. Multiobjective optimization of building design using TRNSYS
simulations, genetic algorithm, and Artificial Neural Network. Building and Environment. 45-3
(2010) 739–746.
[2] Wang, W., Zmeureanu, R., Rivard, H.. Applying multi-objective genetic algorithms in
green building design optimization. Building and Environment. 40 (2005) 1512–1525.
[3] Peippo, K., Lund, P.D., Vartiainen, E.. Multivariate optimization of design trade- offs for
solar low energy buildings. Energy and Buildings. 29-2 (1999) 189–205.
[4] Haase, M., Amato, A.. An investigation of the potential for natural ventilation and building
orientation to achieve thermal comfort in warm and humid climates. Solar Energy. 83-3 (2009)
389–399.
[5] Rodrigues, E., Gaspar, A., Gomes, A.. Evolutionary strategy enhanced with a local search
technique for the space allocation problem in architecture. Patent pending 20121000052347
(filing date July 2nd). Instituto Nacional de Propriedade Industrial (INPI). (2012).
[6] Rodrigues, E., Gaspar, A., Gomes, A.. An evolutionary strategy enhanced with a local
search technique for the space allocation problem in architecture, Part 1: methodology.
Computer Aided-Design. 45-5 (2013) 887-897. doi:10.1016/j.cad.2013.01.001
[7] Rodrigues, E., Gaspar, A., Gomes, A.. An evolutionary strategy enhanced with a local
search technique for the space allocation problem in architecture, Part 2: validation and
performance tests. Computer Aided-Design. 45-5 (2013) 898-910.
doi:10.1016/j.cad.2013.01.003
[8] EN 15251. Indoor environmental input parameters for design and assessment of energy
performance of buildings addressing indoor air quality, thermal environment, lighting and
acoustics. (2007).
| 5cs.CE
|
The geometry of characters of Hopf algebras
arXiv:1704.01099v4 [math.GR] 27 Sep 2017
Geir Bogfjellmo and Alexander Schmeding
Abstract Character groups of Hopf algebras can be used to conveniently describe
several species of “series expansions” such as ordinary Taylor series, B-series, arising in the study of ordinary differential equations, Fliess series, arising from control theory and rough paths. These ideas are a fundamental link connecting Hopf
algebras and their character groups to the topics of the Abelsymposium 2016 on
“Computation and Combinatorics in Dynamics, Stochastics and Control”. In this
note we will explain some of these connections, review constructions for Lie group
and topological structures for character groups and provide some new results for
character groups. Our main result is the construction and study of Lie group structures for Hopf algebras which are graded but not necessarily connected (in the sense
that the degree zero subalgebra is finite-dimensional).
Key words: infinite-dimensional Lie group, pro-Lie group, character group of a
Hopf algebra, regularity of Lie groups, continuous inverse algebra
MSC2010: 22E65 (primary); 16T05, 43A40, 58B25 (Secondary)
1
Foundations: Character groups and graded algebra . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2
Geometry of groups of characters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
3
Appendix: Infinite-dimensional calculus . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
4
Appendix: Characters and the exponential map . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
Geir Bogfjellmo
Chalmers tekniska högskola och Göteborgs universitet, Matematiska vetenskaper, 412 96
Göteborg, Sweden e-mail: [email protected]
Alexander Schmeding
NTNU Trondheim, Institutt for matematiske fag, 7491 Trondheim, Norway e-mail:
[email protected]
1
2
Geir Bogfjellmo and Alexander Schmeding
Character groups of Hopf algebras appear in a variety of mathematical contexts.
For example, they arise in non-commutative geometry, renormalisation of quantum
field theory [14], numerical analysis [10] and the theory of regularity structures for
stochastic partial differential equations [25]. A Hopf algebra is a structure that is
simultaneously a (unital, associative) algebra, and a (counital, coassociative) coalgebra that is also equipped with an antiautomorphism known as the antipode, satisfying a certain property. In the contexts of these applications, the Hopf algebras
often encode combinatorial structures and serve as a bookkeeping device.
Several species of “series expansions” can then be described as algebra morphisms from a Hopf algebra H to a commutative algebra B. Examples include ordinary Taylor series, B-series, arising in the study of ordinary differential equations,
Fliess series, arising from control theory and rough paths, arising in the theory of
stochastic ordinary equations and partial differential equations. An important fact
about such algebraic objects is that, if B is commutative, the set of algebra morphisms Alg(H , B), also called characters, forms a group with product given by
convolution
a ∗ b = mB ◦ (a ⊗ b) ◦ ∆H .
These ideas are the fundamental link connecting Hopf algebras and their character
groups to the topics of the Abelsymposium 2016 on “Computation and Combinatorics in Dynamics, Stochastics and Control”. In this note we will explain some of
these connections, review constructions for Lie group and topological structures for
character groups and provide some new results for character groups.
Topological and manifold structures on these groups are important to applications in the various fields outlined above. In many places in the literature the character group is viewed as “an infinite dimensional Lie group” and one is interested in
solving differential equations on these infinite-dimensional spaces (we refer to [6]
for a more detailed discussion and further references). This is due to the fact that the
character group admits an associated Lie algebra, the Lie algebra of infinitesimal
characters1
g(H , B) := {φ ∈ HomK (H , B) | φ (xy) = φ (x)εH (y) + εH (x)φ (y), ∀x, y ∈ H },
whose Lie bracket is given by the commutator bracket with respect to convolution.
As was shown in [5], character groups of a large class of Hopf algebras are infinitedimensional Lie groups. Note however, that in ibid. it was also shown that not every
character group can be endowed with an infinite-dimensional Lie group structure. In
this note we extend these results to a larger class Hopf algebras. To this end, recall
that a topological algebra is a continuous inverse algebra (or CIA for short) if the set
of invertible elements is open and inversion is continuous on this set (e.g. a Banach
algebra). Then we prove the following theorem.
1 Note that this Lie algebra is precisely the one appearing in the famous Milnor-Moore theorem in
Hopf algebra theory [41].
The geometry of characters of Hopf algebras
3
Theorem A Let H = ⊕n∈N0 Hn be a graded Hopf algebra such that dim H0 < ∞
and B be a commutative CIA. Then G (H , B) is an infinite-dimensional Lie group
whose Lie algebra is g(H , B).
As already mentioned, in applications one is interested in solving differential
equations on character groups (see e.g. [42] and compare [6]). These differential equations turn out to be a special class of equations appearing in infinitedimensional Lie theory in the guise of regularity for Lie groups. To understand this
and our results, we recall this concept now for the readers convenience.
Regularity (in the sense of Milnor) Let G be a Lie group modelled on a locally
convex space, with identity element e, and r ∈ N0 ∪ {∞}. We use the tangent map of
the left translation λg : G → G, x 7→ gx by g ∈ G to define g.v := Te λg (v) ∈ Tg G for
v ∈ Te (G) =: L(G). Following [20], G is called Cr -semiregular if for each Cr -curve
γ : [0, 1] → L(G) the initial value problem
(
η ′ (t) = η (t).γ (t)
η (0) = e
has a (necessarily unique) Cr+1 -solution Evol(γ ) := η : [0, 1] → G. If furthermore
the map
evol: Cr ([0, 1], L(G)) → G, γ 7→ Evol(γ )(1)
is smooth, G is called Cr -regular.2 If G is Cr -regular and r ≤ s, then G is also Cs regular. A C∞ -regular Lie group G is called regular (in the sense of Milnor) – a
property first defined in [40]. Every finite-dimensional Lie group is C0 -regular (cf.
[43]). In the context of this paper our results on regularity for character groups of
Hopf algebras subsume the following theorem.
Theorem B Let H = ⊕n∈N0 Hn be a graded Hopf algebra such that dim H0 < ∞
and B be a sequentially complete commutative CIA. Then G (H , B) is C0 -regular.
Recently, also an even stronger notion regularity called measurable regularity
has been considered [19]. For a Lie group this stronger condition induces many Lie
theoretic properties (e.g. validity of the Trotter product formula). In this setting, L1 regularity means that one can solve the above differential equations for absolutely
continuous functions (whose derivatives are L1 -functions). A detailed discussion of
these properties can be found in [19]. However, we will sketch in Remark 19 a proof
for the following proposition.
Proposition C Let H = ⊕n∈N0 Hn be a graded Hopf algebra with dim H0 < ∞
which is of countable dimension, e.g. H is of finite type. Then for any commutative
Banach algebra B, the group G (H , B) is L1 -regular.
One example of a Hopf algebra whose group of characters represent a series expansion is the Connes–Kreimer Hopf algebra or Hopf algebra of rooted trees HCK .
Brouder [10] established a very concrete link between HCK and B-series. Bseries, due to Butcher [12], constitute an algebraic structure for the study of inte2
Here we consider Cr ([0, 1], L(G)) as a locally convex vector space with the pointwise operations
and the topology of uniform convergence of the function and its derivatives on compact sets.
4
Geir Bogfjellmo and Alexander Schmeding
grators for ordinary differential equations. In this context, the group of characters
G (HCK , R) is known as the Butcher group. The original idea was to isolate the
numerical integrator from the concrete differential equation, and even from the surrounding space (assuming only that it is affine), thus enabling a study of the integrator an sich.
Another example connecting character groups to series expansions arises in the
theory of regularity structures for stochastic partial differential equations (SPDEs)
[25, 11]. In this theory one studies singular SPDEs, such as the continuous parabolic
Anderson model (PAM, cf. the monograph [33]) formally given by
∂
− ∆ u(t, x) = u(t, x)ζ (x) (t, x) ∈]0, ∞[×R2 , ζ spatial white noise.
∂t
We remark that due to the distributional nature of the noise, the product and thus
the equation is ill-posed in the classical sense (see [25, p.5]). To make sense of the
equation, one wants to describe a potential solution by “local Taylor expansion” with
respect to reference objects built from the noise terms. The analysis of this “Taylor
expansion” is quite involved, since products of noise terms are not defined. However,
it is possible to obtains Hopf algebras which describe the combinatorics involved.
Their R-valued character group G is then part of a so called regularity structure
(A , T , G ) ([11, Definition 5.1]) used in the renormalisation of the singular SPDE.
See Example 25 for a discussion of the Lie group structure for these groups.
1 Foundations: Character groups and graded algebra
In this section we recall basic concepts and explain the notation used throughout
the article. Whenever in the following we talk about algebras (or coalgebras or bialgebras) we will assume that the algebra (coalgebra, bialgebra) is unital (counital or
unital and counital in the bialgebra case). Further K ∈ {R, C} will always denote
either the field of real or complex numbers (though many of the following concepts
make sense over general base fields).
Definition 1. A Hopf algebra (over K) H is a K−bialgebra (H , m, 1H , ∆ , ε )
equipped with an antiautomorphism S, called the antipode, such that the diagram
The geometry of characters of Hopf algebras
H ⊗H
5
S⊗id
H ⊗H
∆
m
ε
H
K
u
H
m
∆
H ⊗H
id⊗S
H ⊗H
commutes.
In the diagram u : K → H , k 7→ k1H is the unit map of H , i.e. the map which
sends scalars to multiples of the unit 1H ∈ H . We refer to [35, 13, 37, 46] for basic
information on Hopf algebras.
Let B be a commutative algebra. The set of linear maps HomK (H , B) forms a
new algebra with the convolution product
φ ⋆ ψ = mB ◦ (φ ⊗ ψ )∆H ,
and unit uB ◦ εH (where uB is the unit map of B).
Recall that the invertible elements or units of an algebra A form a group, which
we denote A× .
Definition 2. A linear map φ : H → B is called
1. a (B-valued) character if φ (ab) = φ (a)φ (b) for all a, b ∈ H . The set of all
characters is denoted G (H , B).
2. a (B-valued) infinitesimal character if φ (ab) = εH (b)φ (a) + εH (a)φ (b) for all
a, b ∈ H . The set of all infinitesimal characters is denoted g(H , B).
Lemma 3 ([37, Proposition 21 and 22]).
1. G (H , B) is a subgroup of the group of units HomK (H , B)× . On G (H , B), the
inverse is given by
φ ⋆−1 = φ ◦ S
2. g(H , B) is a Lie subalgebra of the commutator Lie algebra HomK (H , B), [·, ·],
where the bracket is given by
[φ , ψ ] = φ ⋆ ψ − ψ ⋆ φ
An algebraic property of characters and infinitesimal characters is that the algebraic exponential
∞
1
exp⋆ (φ ) = ∑ φ ⋆n
n!
n=0
6
Geir Bogfjellmo and Alexander Schmeding
is a map from g(H , B) to G (H , B). [37, Proposition 22].
In order to study the topological aspects of characters and infinitesimal characters
of Hopf algebras, we need to assume at this step that B is a topological algebra, i.e.,
that B is topological vector space and that the product in B is a continuous bilinear
function
µB : B × B → B
We can then endow the space HomK (H , B) with the topology of pointwise convergence. The sets g(H , B) and G (H , B) are then closed subsets of HomK (H , B),
and carry the induced topologies.
Proposition 4. Let H be a Hopf algebra, and B a commutative, topological algebra. Endow HomK (H , B) with the topology of pointwise convergence. Then
• (HomK (H , B), ⋆) is a topological algebra,
• G (H , B) is a topological group,
• g(H , B) is a topological Lie algebra.
Proof. It is sufficient to prove that ⋆ is continuous. Since HomK (H , B) is endowed with the topology of pointwise convergence, it suffices to test convergence when evaluating at an element h ∈ H . Using Sweedler notation, we get
φ ⋆ ψ (h) = ∑(h) φ (h(1) ) ⋆ ψ (h(2) ) where the multiplication is carried out in B. As
point evaluations are continuous on HomK (H , B), and multiplication is continuous
in B, ⋆ is continuous.
⊓
⊔
The definition of ⋆ does not depend on the algebra structure of H , only the
coalgebra structure. We therefore get as a corollary:
Corollary 5. Let C be a coalgebra, and B a commutative, topological algebra. Then
(HomK (C, B), ⋆), equipped with the topology of pointwise convergence, is a topological algebra.
In Section 2 we will be able to state more about the topology and geometry
of groups of characters, under further assumptions on H and B. In particular, we
are interested in cases where G (H , B) is an (infinite dimensional) Lie group, or a
projective limit of finite dimensional Lie groups, i.e. a pro-Lie group. Both of these
classes of topological groups can to some extent claim to be the generalization of
finite dimensional Lie groups, and have been studied extensively for this reason (see
e.g. [29, 43, 27]).
For many arguments later on gradings will be important. We recall the following
basic definitions
L
Definition 6. Recall that a graded Hopf algebra H = n∈N0 H0 is a Hopf algebra together with a grading as algebra and coalgebra (i.e. Hn · Hm ⊆ Hn+m and
L
∆ (Hn ) ⊆ k+l=n Hk ⊗ Hl ). In particular, H0 becomes a Hopf subalgebra of H .
Note that for a graded Hopf algebra H , identifying a mapping f : H → B with
its components on the grading induces a natural isomorphisms of topological vector
spaces (with respect to the topologies of pointwise convergence)
The geometry of characters of Hopf algebras
HomK (H , B) = HomK (
M
7
Hn , B) ∼
=
n∈N0
∏ HomK (Hn , B).
n∈N0
Hence A = HomK (H , B) becomes a densely graded topological vector space (see
[5, Appendix B]). We denote by An := HomK (Hn , B) the parts of the dense grading.
Note that A0 becomes a locally convex subalgebra of A by definition of the grading.
2 Geometry of groups of characters
In this section, we review results on geometry, topology and Lie theory for character
groups of Hopf algebras G (H , B). Further, in Section 2.1 we prove a new result
which establishes a Lie group structure for character groups of non-connected Hopf
algebras.
In general, the existence of a Lie group structure on the character group of a Hopf
algebra H depends on structural properties of the underlying Hopf algebra (e.g. we
need graded and connected Hopf algebras), the table below provides an overview of
the topological and differentiable structures related to these additional assumptions.
Hopf algebra H
commutative algebra B
Structure on G (H , B)
arbitrary
weakly complete
pro-Lie group (cf. Remark 7)
graded and dim H0 < ∞ continuous inverse algebra ∞-dim. Lie group (Section 2.1)
graded and connected
locally convex algebra ∞-dim. Lie group (Section 2.2)
Fig. 1 Overview of topological and Lie group structures on character groups of Hopf algebras.
In general, the character group need not admit a Lie group structure as was shown
in [5, Example 4.11]. There we exhibited a character group of the group algebra of
an abelian group of infinite rank which can not be a Lie group.
Remark 7. If the target algebra B is a weakly complete algebra, e.g. a finite dimensional algebra, the character group G (H , B) is always a projective limit of finitedimensional Lie groups. In [5, Theorem 5.6] we have proved that for an arbitrary
Hopf algebra H and B a weakly complete algebra, G (H , B) is a special kind of
topological group, a so called pro-Lie group (see the survey [29]). A pro-Lie group is
closely connected to its pro-Lie algebra which turns out to be isomorphic to g(H , B)
for the pro-Lie group G (H , B). Although pro-Lie groups do not in general admit a
differentiable structure, a surprising amount of Lie theoretic properties carries over
to pro-Lie groups (we refer to the monograph [27] for a detailed account).
Often the character group of a Hopf algebra will have more structure than a
topological group. As we will see in the next example character groups often carry
Lie group structures.
8
Geir Bogfjellmo and Alexander Schmeding
Example 8. Let G be a compact Lie group. Then we consider the set R(G) of representative functions, i.e. continuous functions f : G → R such that the set of right
translates Rx f : G → R, Rx f (y) = f (yx) generates a finite-dimensional subspace of
C(G, R), cf. [28, Chapter 3] or [13, Section 3] for more information and examples.
Using the group structure of G and the algebra structure induced by C(G, R)
(pointwise multiplication), R(G) becomes a Hopf algebra (see [23, pp. 42-43]).
Following Remark 7, we know that G (R(G), R,) becomes a topological group.
It follows from Tannaka-Kreĭn duality that as compact groups G (R(G), R) ∼
=
G, whence G (R(G), R,) inherits a posteriori a Lie group structure [23, Theorem
1.30 and 1.31].3 Observe that the Lie group structure induced on G (R(G), R,) via
Tannaka-Kreĭn duality coincides with the ones discussed in Sections 2.1 and 2.2
(when these results are applicable to the Hopf algebra of representative functions).
Example 9 (The Butcher group). Let T denote the set of rooted trees, and HCK =
hhT ii the free commutative algebra generated by T . The Grossman–Larson coproduct is defined on trees as
∆ (τ ) = τ ⊗ 1 + ∑(τ \ σ ) ⊗ σ
σ
where the sum goes over all connected subsets σ of τ containing the root. Together
with the algebra structure, The Grossman–Larson coproduct defines a graded, connected bialgebra structure on HCK , and therefore also a Hopf algebra structure.
The characters G (HCK , R) are the algebra morphisms HomAlg (HCK , R). Clearly,
we can identify
G (HCK , R) ≃ RT
In numerical analysis, the character group G (HCK , R) is known as the Butcher
group [12, 24]. This group is closely connected to a class of numerical integrators for
ordinary differential equations. Namely, we let ẏ = f (y) be an autonomous ordinary
differential equation on an affine space E. Many numerical integrators 4 can be
expanded in terms of the elementary differentials of the vector field f , i.e. as a series
′
yn+1 = yn + a( )h f (yn ) + a( )h2 f f (yn ) + · · ·
(1)
The elementary differentials are in a natural one-to-one correlation with T , and the
series (1) thus defines an element in G (HCK , R). The crucial observation is that,
(after a suitable scaling,) the convolution product in G (HCK , R) corresponds to the
composition of numerical integrators.
In the following, it will be established that G (HCK , R) is a R-analytic, C0 -regular
Fréchet Lie group as well as a pro-Lie group. See [8, 5] for further details on the Lie
group structure of G (HCK , R).
3 This is only a glimpse at Tannaka-Kreĭn duality, which admits a generalisation to compact topological groups (using complex representative functions, see [26, Chapter 7, §30] and cf. [23, p. 46]
for additional information in the Lie group case). Also we recommend [28, Chapter 6] for more
information on compact Lie groups.
4 To be exact: The class of integrators depending only on the affine structure (cf. [38]).
The geometry of characters of Hopf algebras
9
However, in some sense, the Butcher group G (HCK , R) is too big to properly analyze numerical integrators. For every numerical integrator, the coefficients
a: T → R satisfy a growth bound |a(τ )| ≤ CK |τ | . Elements of G (HCK , R) satisfying such growth bounds form a subgroup, and even a Lie subgroup. However, the
modelling space now becomes a Silva space5 . This group is studied in the article
[7].
In the next section we begin with establishing general results on the infinitedimensional Lie group structure of Hopf algebra character groups. These manifolds
will in general be modelled on spaces which are more general then Banach spaces.
Thus the usual differential calculus has to be replaced by the so called Bastiani
calculus (see [2], i.e. differentiability means existence and continuity of directional
derivatives). For the readers convenience, Appendix 3 contains a brief recollection
of this calculus.
2.1 Character groups for H graded with finite dimensional H0
and B a continuous inverse algebra
In this section we consider graded but not necessarily connected Hopf algebras. In
general, character groups of non-connected Hopf algebras do not admit a Lie group
structure. Recall from example from [5, Example 4.11 (b)] that the character group
of the group algebra of an infinite-group does in general not admit a Lie group
structure. However, if the Hopf algebra is not too far removed from being connected
(i.e. the 0-degree subspace H0 is finite-dimensional) and the target algebra is at
least a continuous inverse algebra, we can prove that the character group G (H , B)
is an infinite-dimensional Lie group. This result is new and generalises [5] where
only character groups of graded and connected Hopf algebras were treated (albeit
the target algebra in the graded connected case may be a locally convex algebra).
2.1. Let (A, ·) be a (real or complex) locally convex algebra (i.e. the locally convex
topology of A makes the algebra product jointly continuous). We call (A, ·) continuous inverse algebra (or CIA for short) if its unit group A× is open and inversion
A× → A, a 7→ a−1 is continuous.
The class of locally convex algebras which are CIAs are of particular interest to
infinite-dimensional Lie theory as their unit groups are in a natural way (analytic)
Lie groups (see [16, 21]).
Before we establish the structural result, we have to construct an auxiliary Lie
group structure in which we can realise the character group as subgroup. To construct the auxiliary Lie group, we can dispense with the Hopf algebra structure and
5 Silva spaces arise as special inductive limits of Banach spaces, see [15] for more information.
They are also often called (DFS)-space in the literature, as they can be characterised as the duals
of Fréchet-Schwartz spaces.
10
Geir Bogfjellmo and Alexander Schmeding
consider only (graded) coalgebras at the moment. The authors are indebted to K.–H.
Neeb for providing a key argument in the proof of the following Lemma.
Lemma 10. Let (C, ∆ ) be a finite-dimensional coalgebra and B be a CIA. Then
(HomK (C, B), ⋆) is a CIA.
Proof. Consider the algebra (A := HomK (C, K), ⋆A ), where ⋆A is the convolution product. Then the algebraic tensor product T := B ⊗K A with the product
(b ⊗ ϕ ) · (c ⊗ ψ ) := bc ⊗ ϕ ⋆A ψ is a locally convex algebra. Since C is a finitedimensional coalgebra, A is a finite-dimensional algebra. Due to an argument which
was communicated to the authors by K.–H. Neeb, the tensor product of a finitedimensional algebra and a CIA is again a CIA.6 Thus it suffices to prove that the
linear mapping defined on the elementary tensors via
κ : T → HomK (C, B), b ⊗ ϕ 7→ (x 7→ ϕ (x)b)
is an isomorphism of unital algebras. Since A is finite-dimensional, it is easy to
see that κ is an isomorphism of locally convex spaces. Thus it suffices to prove
that κ is a algebra morphism. To this end let ε be the counit of C. We recall that
1A = ε , 1HomK (C,B) = (x 7→ ε (x) · 1B ) and 1T = 1B ⊗ 11 = 1B ⊗ ε are the units
in A, HomK (C, B) and T , respectively. Now κ (1T ) = (x 7→ ε (x)1B ) = 1HomK (C,B) ,
whence κ preserves the unit.
As the elementary tensors span T , it suffices to establish multiplicativity of κ on
elementary tensors b1 ⊗ ϕ , b2 ⊗ ψ ∈ T . For c ∈ C we use Sweedler notation to write
∆ (c) = ∑(c) c1 ⊗ c2 . Then
κ ((b1 ⊗ ϕ ) · (b2 ⊗ ψ ))(c) = κ (b1 b2 ⊗ ϕ ⋆A ψ )(c) = ϕ ⋆A ψ (x)b1 b2
= ∑ ϕ (c1 )ψ (c2 )b1 b2 = ∑(ϕ (c1 )b1 )(ψ (c2 )b2 )
(c)
(c)
= ∑ κ (b1 ⊗ ϕ )(c1 )κ (b2 ⊗ ψ )(c2 )
(c)
= κ (b1 ⊗ ϕ ) ⋆ κ (b2 ⊗ ψ )(c)
shows that the mappings agree on each c ∈ C, whence κ is multiplicative. Summing
up, HomK (C, B) is a CIA as it is isomorphic to the CIA B ⊗ A.
⊓
⊔
L
Proposition 11 (A× is a regular Lie group). Let C = n∈N0 Cn be a graded coalgebra with dim C0 < ∞ and B be a CIA. Then A = (HomK (H , B), ⋆) is a CIA whose
6
We are not aware of a literature reference of this fact apart from the forthcoming book by
Glöckner and Neeb [22]. To roughly sketch the argument from [22]: Using the regular representation of A one embeds B ⊗K A in the matrix algebra Mn (B) (where n = dim A). Now as A is a
commutant in EndK (A), the same holds for B ⊗ A in Mn (B). The commutant of a set in a CIA is
again a CIA, whence the assertion follows as matrix algebras over a CIA are again CIAs (cf. [45,
Corollary 1.2]).
The geometry of characters of Hopf algebras
11
unit group A× is Baker–Campbell–Hausdorff–Lie group (BCH–Lie group)7 with Lie
algebra (A, [·, ·]), where [·, ·] denotes the commutator bracket with respect to ⋆.
If in addition B is Mackey complete, then A is Mackey complete and the Lie group
A× is C1 -regular. If B is even sequentially complete, so is A and the Lie group A× is
C0 -regular. In both cases the associated evolution map is even K-analytic and the
Lie group exponential map is given by the exponential series.
Proof. Recall from [5, Lemma 1.6 (c) and Lemma B.7] that the locally convex
algebra A is a Mackey complete CIA since A0 is such a CIA by Lemma 10 (both
CIAs are even sequentially complete if B is so). Now the assertions concerning the
Lie group structure of the unit group are due to Glöckner (see [16]).
To see that the Lie group A× is Ck -regular (k = 1 for Mackey complete and
k = 0 of sequentially complete CIA B), we note that the regularity of the unit group
follows from the so called called (GN)-property (cf. [5] and see Definition 12 below) and (Mackey) completeness of A. Having already established completeness,
we recall from [5, Lemma 1.10] that A has the (GN)-property if A0 has the (GN)property. Below in Lemma 13 we establish that A0 has the (GN)-property if B has
the (GN)-property. Now B is a commutative Mackey complete CIA, whence B has
the (GN)-property by [21, Remark 1.2 and the proof of Corollary 1.3]. Summing up,
A has the (GN)-property and thus [21, Proposition 4.4] asserts that A× is Ck -regular
with analytic evolution map.
⊓
⊔
Before we can establish the (GN)-property for A0 as in the proof of Proposition
11 we need to briefly recall this condition.
Definition 12 ((GN)-property). A locally convex algebra A is said to satisfy the
(GN)-property, if for every continuous seminorm p on A, there exists a continuous
seminorm q and a number M ≥ 0 such that for all n ∈ N, we have the estimate:
(n)
µA
(n)
(n)
p,q
:= sup{p(µA (x1 , . . . , xn )) | q(xi ) ≤ 1, 1 ≤ i ≤ n} ≤ M n .
(2)
(n)
Here, µA is the n-linear map µA : A × · · · × A → A, (a1 , . . . , an ) 7→ a1 · · · an .
| {z }
n times
Lemma 13. Let C be a finite-dimensional coalgebra and B be a CIA with the (GN)property. Then the following holds:
1. (A := HomK (C, B), ⋆) has the (GN)-property.
2. If (B, k·k) is even a Banach algebra, then A is a Banach algebra.
Proof. Choose a basis ei , 1 ≤ i ≤ d for H0 . Identifying linear maps with the coefficients on the basis, we obtain A = HomK (C, B) ∼
= Bd (isomorphism of topological
vector spaces). For every continuous seminorm p on B, we obtain a corresponding
7
BCH-Lie groups derive their name from the fact that there is a neighborhood in their Lie algebra on which the Baker–Campbell–Hausdorff series converges and yields an analytic map. See
Definition 35.
12
Geir Bogfjellmo and Alexander Schmeding
seminorm p∞ : A → R, φ 7→ max1≤i≤d p(φ (ei )) and these seminorms form a generating set of seminorms for the locally convex algebra A. Let us now write the
coproduct of the basis elements as
∆ (ei ) = ∑ νijk e j ⊗ ek ,
jk
for νi ∈ K.
1≤i≤d
j,k
To establish the properties of A we need an estimate of the structural constants,
whence we a constant K := d 2 maxi, j,k {|νijk |, 1}
1. It suffices to establish the (GN)-property for a set of seminorms generating the
topology of A. Hence it suffices to consider seminorms q∞ induced by a continuous seminorm q on B. Since B has the (GN)-property there are a continuous
seminorm p on B and a constant M ≥ 0 which satisfy (2) with respect to the chosen q. We will now construct a constant such that (2) holds for the seminorms
q∞ and p∞ taking the rôle of q and p.
Observe that q∞ (ψ ) ≤ 1 implies that q(ψ (ei )) ≤ 1 for each 1 ≤ i ≤ d. Thus
by choice of the constants a trivial computation shows that the constant KM
satisfies (2) for q∞ , p∞ and n = 1. We will show that this constant satisfies the
inequality also for all n > 1 and begin for the readers convenience with the case
n = 2. Thus for ψ1 , ψ2 ∈ A with q∞ (ψl ) ≤ 1, l = 1, 2 and 1 ≤ k ≤ d we have
!
p(ψ1 ⋆ ψ2 (ei )) = p
∑ νijk ψ1 (e j )ψ2 (ek )
j,k
≤ ∑ |νijk | p(ψ1 (ei )ψ2 (e j ))
{z
}
|
j,k
≤M2
2
2
≤ |{z}
K M ≤ (KM)
≥1
As the estimate neither depends on i nor on the choice of ψ1 , ψ2 (we only used
(2)
q∞ (ψl ) ≤ 1), KM satisfies µA
≤ (KM)2 . Now for general n ≥ 2 we
p∞ ,q∞
choose ψl ∈ A with q∞ (ψl ) ≤ 1 and 1 ≤ l ≤ n. As convolution is associative,
ψ1 ⋆ · · · ⋆ ψn is obtained from applying ψ1 ⊗ · · · ⊗ ψn to the iterated coproduct
∆ n := idC ⊗ ∆ n−1 , ∆ 1 := ∆ and subjecting the result to the n-fold multiplication
map B ⊗ B ⊗ · · ·⊗ B → B of the algebra B. Hence one obtains the formula
p(ψ1 ⋆ ψ2 ⋆ · · · ⋆ ψn (ei ))
≤
∑ ∑ ··· ∑
jn−1 ,kn−1
j1 ,k1 j2 ,k2
|
{z
}
#of terms=d 2 ·d 2 ···d 2 =(d 2 )n−1
≤K n−1 M n ≤ (KM)n .
kn−2
|ν ij1 ,k1 ||ν kj21,k2 | · · · |ν jn−1
,kn−2 | p
|
{z
jk
≤(maxi, j,k {|νi |,1})n−1
}
ψ1 (ek1 )
∏
2≤r≤n
!
ψr (ekr )
Again the estimate does neither depend on ei nor on the choice of ψ1 , . . . , ψn ,
whence we see that one can take KM in general as a constant which satisfies
The geometry of characters of Hopf algebras
13
(2) for q∞ and p∞ . We conclude that A has the (GN)-property if B has the (GN)property.
2. Let now (B, k·k) be a Banach algebra, then A ∼
= Bd is a Banach space with the
norm kφ k∞ := max1≤i≤d kφ (ei )k. To prove that A admits a submultiplicative
norm, define the norm kα kK := K kα k∞ (for K the constant chosen above). By
construction k·kK is equivalent to k·k∞ and we will prove that k·kK is submultiplicative. Consider α , β ∈ A and compute the norm of α ⋆ β on a basis element
jk
kα ⋆ β (ei )k = kmB ◦ (α ⊗ β ) ◦ ∆ (ei)k ≤ ∑ |νi | α (e j )β (ek )
j,k
≤∑
jk
|νi |
α (e j ) kβ (ek )k ≤ K kα k∞ kβ k∞ .
j,k
In passing from the first to the second row we have used that the norm on B is
submultiplicative. Summing up, we obtain kα ⋆ β kK ≤ kα kK kβ kK whence A
is a Banach algebra.
⊓
⊔
In case the Hopf algeba H is only of countable dimension, e.g. a Hopf algebra
of finite type, and B is a Banach algebra, the unit group A× satisfies an even stronger
regularity condition.
L
Lemma 14. Let C = n∈N0 be a graded coalgebra with dim C0 < ∞ and B be a
Banach algebra. Assume in addition that C is of countable dimension, then the Lie
group A× from Proposition 11 is L1 -regular.
Proof. If C is of countable dimension, then A = HomK (C, B) is a Fréchet space (as
it is isomorphic as a locally convex space to a countable product of Banach spaces).
Now [19, Proposition 7.5] asserts that the unit group A× of a continuous inverse
algebra which is a Fréchet space will be L1 -regular if A is locally m-convex. However, by the fundamental theorem for coalgebras [39, Theorem 4.12], C = lim→ Σn
is the direct limit of finite dimensional subcoalgebras. Dualising, Lemma 13 shows
that A = lim← HomK (Σn , B) is the projective limit of Banach algebras. As Banach
algebras are locally m-convex and the projective limit of a system of locally mconvex algebras is again locally m-convex (cf. e.g. [36, Chapter III, Lemma 1.1]),
the statement of the Lemma follows.
⊓
⊔
To establish the following theorem we will show now that the Lie group A×
induces a Lie group structure on the character group of the Hopf algebra. Note that
the character group is a closed subgroup of A× , but, contrary to the situation for
finite dimensional Lie groups, closed subgroups do not automatically inherit a Lie
group structure (see [43, Remark IV.3.17] for a counter example).
L
Theorem 15. Let H = n∈N0 Hn be a graded Hopf algebra with dim H0 < ∞.
Then for any commutative CIA B, the group G (H , B) of B-valued characters of H
is a (K−analytic) Lie group. Furthermore, we observe the following.
(i) The Lie algebra L(G (H , B)) of G (H , B) is the Lie algebra g(H , B) of infinitesimal characters with the commutator bracket [ φ , ψ ] = φ ⋆ ψ − ψ ⋆ φ .
14
Geir Bogfjellmo and Alexander Schmeding
(ii)G (H , B) is a BCH–Lie group which is locally exponential, i.e. the Lie group
x⋆n
exponential map exp : g(H , B) → G (H , B), x 7→ ∑∞
n=0 n! is a local K-analytic
diffeomorphism.
Proof. Recall from Proposition 11 and Proposition 4 that G (H , B) is a closed subgroup of the locally exponential Lie group (A× , ⋆). We will now establish the Lie
group structure using a well-known criterion for locally exponential Lie groups: Let
expA be the Lie group exponential of A× and consider the subset
Le (G (H , B)) := {x ∈ L(A× ) = A | expA (Rx) ⊆ G (H , B)}.
We establish in Lemma 37 that g(H , B) is mapped by expA to G (H , B), whence
g(H , B) ⊆ Le (G (H , B)). To see that Le (G (H , B)) only contains infinitesimal
characters, recall from Lemma 37 that there is an open 0-neighborhood Ω ⊆ A such
that expA maps g(H , B)∩ Ω bijectively to expA (Ω )∩G (H , B). If x ∈ Le (G (H , B)
then we can pick t > 0 so small that tx ∈ Ω . By definition of Le (G (H , B))
we see that then expA (tx) ∈ G (H , B) ∩ expA (Ω ). Therefore, we must have tx ∈
Ω ∩ g(H , B), whence x ∈ g(H , B). This entails that Le (G (H , B)) = g(H , B) and
then [43, Theorem IV.3.3] implies that G (H , B) is a locally exponential closed Lie
subgroup of (A× , ⋆) whose Lie algebra is the Lie algebra of infinitesimal characters g(H , B). Moreover, since (A× , ⋆) is a BCH–Lie group, so is G (H, B) (cf. [43,
Definition IV.1.9]).
⊓
⊔
Note that the Lie group G (H , B) constructed in Theorem 15 will be modelled
on a Fréchet space if H is of countable dimension (e.g. if H is of finite type) and B
is in addition a Fréchet algebra. If H is even finite-dimensional and B is a Banach
algebra, then G (H , B) will even be modelled on a Banach space.
Example 16. 1. The characters of a Hopf algebra of finite-type, i.e. the components Hn in the grading H = ⊕n∈N0 Hn are finite-dimensional, are infinitedimensional Lie groups by Theorem 15. Most natural examples of Hopf algebras appearing in combinatorial contexts are of finite-type.
2. Every finite-dimensional Hopf algebra H can be endowed with the trivial grading H0 := H . Thus Theorem 15 implies that character groups (with values in
a commutative CIA) of finite-dimensional Hopf algebras (cf. [3] for a survey)
are infinite-dimensional Lie groups.
3. Graded and connected Hopf algebras (see next section) appear in the ConnesKreimer theory of perturbative renormalisation of quantum field theories. However, recently in [32] it has been argued that instead of the graded and connected
Hopf algebra of Feynman graphs considered traditionally (see e.g. the exposition in [14]) a non connected extension of this Hopf algebra should be used.
The generalisation of the Hopf algebra then turns out to be a Hopf algebra with
dim H0 < ∞, whence its character groups with values in a Banach algebra turn
out to be infinite-dimensional Lie groups.
These classes of examples could in general not be treated by the methods developed in [5].
The geometry of characters of Hopf algebras
15
L
Remark 17. Recall that by definition of a graded Hopf algebra H = n∈N0 Hn ,
the Hopf algebra structure turns H0 into a sub-Hopf algebra. Therefore, we always
obtain two Lie groups G (H , B) and G (H0 , B) if dim H0 < ∞. It is easy to see
that the restriction map q : G (H , B) → G (H0 , B), φ 7→ φ |H0 is a morphism of Lie
groups with respect to the Lie group structures from Theorem 15. Its kernel is the
normal subgroup
kerq = {φ ∈ G (H , B) | φ |H0 = 1|H0 },
i.e. the group of characters which coincide with the unit on degree 0.
We now turn to the question whether the Lie group constructed in Theorem 15 is
a regular Lie group.
Theorem 18. Let H be a graded Hopf algebra with dim H0 < ∞ and B be a Banach algebra. Then the Lie group G (H , B) is C0 -regular and the associated evolution map is even a K-analytic map.
Proof. In Theorem 15 the Lie group structure on the character group was identified
as a closed subgroup of the Lie group A× . By definition of the character group (cf.
Definition (2)), G (H , B) can be described as
G (H , B) = {φ ∈ A× | φ ◦ mH = mB ◦ (φ ⊗ φ )}
As discussed in Remark 36 Equation (3), the map (mH )∗ : A× → A×
⊗ , φ 7→ φ ◦ mH
is a Lie group morphism. Now we consider the map θ : A → A⊗ , φ 7→ mB ◦ (φ ⊗ φ ).
Observe that θ = β ◦ diag, where β is the continuous bilinear map (4) and diag : A →
A × A, φ 7→ (φ , φ ). Since β is smooth (as a continuous bilinear map), and the diagonal map is clearly smooth, θ is smooth. Further, (5) shows that θ (φ ⋆ ψ ) =
θ (φ ) ⋆At θ (ψ ). Thus θ restricts to a Lie group morphism θA× : A× → At × .
Summing up, we see that G (H , B) = {φ ∈ A× | (mH )∗ (φ ) = θA× (φ )}. Since
by Proposition 11 the Lie group A× is C0 -regular, the closed subgroup G (H , B) is
also C0 -regular by [20, Theorem G].
⊓
⊔
Remark 19. In Theorem 18 we have established that the character group G (H , B)
inherits the regularity properties of the ambient group of units. If the Hopf algebra
H is in addition of countable dimension, e.g. a Hopf algebra of finite type, then
Lemma 14 asserts that the ambient group of units is even L1 -regular. Unfortunately,
Theorem [20, Theorem G] only deals with regularity of type Ck for k ∈ N0 ∪ {∞}.
However, since G (H , B) is a closed Lie subgroup of HomK (H , B)× , it is easy
to see that the proof of [20, Theorem G] carries over without any changes to the
L1 -case.8 Hence, we can adapt the proof of Theorem 18 to obtain the following
statement:
Corollary Let H be a graded Hopf algebra of countable dimension with
dim H0 < ∞ and B be a Banach algebra. Then the Lie group G (H , B) is L1 -regular.
8 One only has to observe that a function into the Lie subgroup is absolutely continuous if and only
if it is absolutely continuous as a function into the larger group. On the author’s request, a suitable
version of [20, Theorem G] for L1 -regularity will be made available in a future version of [19].
16
Geir Bogfjellmo and Alexander Schmeding
Note that the results on L1 -regularity of character groups considerably strengthen
the results which have been available for regularity of character groups (see [5]).
2.2 Character groups for a graded and connected Hopf algebra H
and B a locally convex algebra
In many interesting cases, the Hopf algebra is even connected graded (i.e. H0 is onedimensional). In this case, we can weaken the assumption on B to be only locally
convex.
Theorem 20 ([5, Theorem 2.7]). Let H be a graded and connected Hopf algebra
and B be a locally convex algebra. Then the manifold structure induced by the global
⋆n
parametrisation exp : g(H , B) → G (H , B), x 7→ ∑n∈N0 xn! turns G (H , B) into a
K-analytic Lie group.
The Lie algebra associated to this Lie group is g(H , B). Moreover, the Lie group
exponential map is given by the real analytic diffeomorphism exp.
Remark 21. Note that Theorem 20 yields more information for the graded and connected case then the new results discussed in Section 2.1: In the graded and connected case, the Lie group exponential is a global diffeomorphism, whereas the theorem for the non-connected case only establishes that exp induces a diffeomorphism
around the unit.
Note that the connectedness of the Hopf algebra is the important requirement
here. Indeed, we can drop the assumption of an integer grading and generalise to the
grading by a suitable monoid.
Definition 22. Let M be a submonoid of (R, +) with 0 ∈ M. We call M an index
monoid if every initial segment Im := {n ∈ M | n ≤ m} is finite. Note that this forces
the monoid to be at most countable.
As in the N0 graded case, we say a Hopf algebra H is graded by an index monoid
L
M, if H = m∈M Hm and the algebra, coalgebra and antipode respect the grading
in the usual sense.
Example 23. The monoid (N0 , +) is an index monoid.
A source for (more interesting) index monoids is Hairer’s theory of regularity
structures for locally subcritical semilinear stochastic partial differential equations
[25, Section 8.1] (cf. in particular [25, Lemma 8.10] and see Example 25 below).
Note that the crucial property of an index monoid is the finiteness of initial segments. This property allows one to define the functional calculus used in the proof
of [5, Theorem B] to this slightly more general setting. Changing only trivial details
in the proof of loc.cit., one immediately deduces that the statement of Theorem 20
generalises from an N0 -grading to the more general situation
The geometry of characters of Hopf algebras
17
L
Corollary 24. Let H = m∈M Hm be a graded and connected Hopf algebra
graded by an index monoid M, B a sequentially complete locally convex algebra.
⋆n
Then the manifold structure induced by exp : g(H , B) → G (H , B), x 7→ ∑n∈N0 xn!
turns G (H , B) into a K-analytic Lie group. This Lie group is C0 -regular, its Lie
algebra is g(H , B) and the Lie group exponential map is the real analytic diffeomorphism exp.
2.2.1 Application to character groups in the renormalisation of SPDEs
Hopf algebras graded by index monoids appear in Hairer’s theory of regularity structures for locally subcritical semilinear SPDEs [25, Section 8.1]. Character groups of
these Hopf algebras (and their quotients) then appear as structure group in Hairer’s
theory (cf. [25, Theorem 8.24] and [11]) of regularity structures. Recall that a regularity structure (A, T, G) in the sense of Hairer (cf. [25, Definition 2.1]) is given
by
• an index set A ⊆ R L
with 0 ∈ A, which is bounded from below and locally finite,
• a model space T = α ∈A Tα which is a graded vector space with T0 ∼
= R (denote
by 1 its unit element) and Tα a Banach space for every α ∈ A,
• a structure group G of linear operators acting on T such that, for every Γ ∈ G,
α ∈ A and a ∈ Tα one has
Γa−a ∈
M
Tβ and Γ 1 = 1.
β <α
We sketch now briefly the theory developed in [11], where for a class of examples singular SPDEs the structure group was recovered as the character group of a
connected Hopf algebra graded by an index monoid.
Example 25. The construction outlined in [11] builds first a general Hopf algebra of
decorated trees in the category of bigraded spaces. Note that this bigrading does not
induce a suitable N0 -grading (or grading by index monoid) for our purposes. This
Hopf algebra encodes the combinatorics of Taylor expansions in the SPDE setting
and it needs to be tailored to the specific SPDE. This is achieved by choosing another
ingredient, a so called subcritical and complete normal rule, i.e. a device depending
on the SDE in question which selects a certain sub-Hopf algebra (see [11, Section
5] for details). Basically, the rule collects all admissible terms (= abstract decorated
trees) which appear in the local Taylor expansion of the singular SPDE.9
Using the rule, we can select an algebra of decorated trees T+ex admissible with
respect to the rule. Here the “+” denotes that we only select trees which are positive
with respect to a certain grading | · |+ (cf. [11, Remark 5.3 and Definition 5.29]).
Then [11, Proposition 5.34] shows that (T+ex , | · |+ ) is a graded and connected Hopf
algebra of decorated trees. Note however, that the grading | · |+ is in general not
integer valued, i.e. T+ex is graded by a submonoid M of [0, ∞[.
9
see [11, Section 5.5] for some explicit examples of this procedure, e.g. for the KPZ equation.
18
Geir Bogfjellmo and Alexander Schmeding
Since we are working with a normal rule which is complete and subcritical, the
submonoid M satisfies |{γ ∈ M | γ < c}| < ∞ for each c ∈ R, i.e. M is an index
monoid. The reason for this is that by construction T+ex is generated by tree products
of trees which strongly conform to the rule [11, Lemma 5.25 and Definition 5.29].
As the rule is complete and subcritical, there are only finitely many trees τ with
|τ |+ < c (for c ∈ R) which strongly conform to the rule. Now the tree product is the
Hopf algebra product (i.e. the product respects the grading), whence the poperty for
M follows.
Now by [11, Proposition 5.39] the graded space T ex = (hB0 i, | · |+ ) (suitably
generated by a certain subset of the strongly conforming trees) together with the
index set Aex = {|τ |+ | τ ∈ B0 } and the character group G+ex := G (T+ex , R) form a
regularity structure (Aex , T ex , G+ex ). In conclusion, Corollary 24 yields an infinitedimensional Lie group structure for the structure group G+ex of certain subcritical
singular SPDEs.
Remark 26. In the full renormalisation procedure outlined in [11] two groups are
involved in the renormalisation. Apart from the structure group outlined above, also
a so called renormalisation group G−ex is used (cf. [11, Section 5]). This group G−ex
is (in the locally subcritical case) a finite-dimensional group arising as the character
group of another Hopf algebra. However, it turns out that the actions of G+ex and
G−ex interact in an interesting way (induced by a cointeraction of underlying Hopf
algebras, think semidirect product). We hope to return to these actions and use the
(infinite-dimensional) Lie group structure to study them in future work.
Acknowledgements This research was partially supported by the European Unions Horizon 2020
research and innovation programme under the Marie Skłodowska-Curie grant agreement No.
691070 and by the Knut and Alice Wallenberg Foundation grant agreement KAW 2014.0354.
We are indebted to K.–H. Neeb and R. Dahmen for discussions which led to Lemma 10. Further,
we would like to thank L. Zambotti and Y. Bruned for explaining their results on character groups
in the renormalisation of SPDEs. Finally, we thank K.H. Hofmann for encouraging and useful
comments and apologize to him for leaving out [28] at first.
3 Appendix: Infinite-dimensional calculus
In this section basic facts on the differential calculus in infinite-dimensional spaces
are recalled. The general setting for our calculus are locally convex spaces (see
[44, 30]).
Definition 27. Let E be a topological vector space over K ∈ {R, C}. E is called
locally convex space if there is a family {pi | i ∈ I} of continuous seminorms for
some index set I, such that
i. the topology is the initial topology with respect to {pr pi : E → E pi | i ∈ I}, i.e. the
E-valued map f is continuous if and only if pri ◦ f is continuous for each i ∈ I,
where E pi := E/p−1
i (0) is the normed space associated to the pi and pri : E →
E pi is the canonical projection,
The geometry of characters of Hopf algebras
19
ii. if x ∈ E with pi (x) = 0 for all i ∈ I, then x = 0 (i.e. E is Hausdorff).
Many familiar results from finite-dimensional calculus carry over to infinite dimensions if we assume that all spaces are locally convex.
As we are working beyond the realm of Banach spaces, the usual notion of
Fréchet-differentiability cannot be used.10 Moreover, there are several inequivalent
notions of differentiability on locally convex spaces (see again [31]).
We base our investigation on the so called Bastiani calculus, [2]. The notion of
differentiability we adopt is natural and quite simple, as the derivative is defined via
directional derivatives.
Definition 28. Let K ∈ {R, C}, r ∈ N ∪ {∞} and E, F locally convex K-vector
spaces and U ⊆ E open. Moreover we let f : U → F be a map. If it exists, we
define for (x, h) ∈ U × E the directional derivative
d f (x, h) := Dh f (x) :=
lim t −1 ( f (x + th) − f (x)).
K× ∋t→0
r
We say that f is CK
if the iterated directional derivatives
d (k) f (x, y1 , . . . , yk ) := (Dyk Dyk−1 · · · Dy1 f )(x)
exist for all k ∈ N0 such that k ≤ r, x ∈ U and y1 , . . . , yk ∈ E and define continuous
maps d (k) f : U × E k → F. If it is clear which K is meant, we simply write Cr for
r . If f is C∞ , we say that f is holomorphic and if f is C∞ we say that f is smooth.
CK
C
R
For more information on our setting of differential calculus we refer the reader
to [17, 31]. Another popular choice for infinite-dimensional calculus is the so called
“convenient setting” of global analysis outlined in [34]. On Fréchet spaces (i.e. complete metrisable locally convex spaces) our notion of differentiability coincides with
differentiability in the sense of convenient calculus. Note that differentiable maps in
our setting are continuous by default (which is in general not true in the convenient
setting). We encounter analytic mappings between infinite-dimensional spaces, as a
preparation for this, note first:
Remark 29. A map f : U → F is of class CC∞ if and only if it is complex analytic
i.e., if f is continuous and locally given by a series of continuous homogeneous
polynomials (cf. [4, Proposition 7.4 and 7.7]). We then also say that f is of class
CCω .
To introduce real analyticity, we have to generalise a suitable characterisation
from the finite-dimensional case: A map R → R is real analytic if it extends to
a complex analytic map C ⊇ U → C on an open R-neighbourhood U in C. We
proceed analogously for locally convex spaces by replacing C with a suitable complexification.
10
The problem here is that the bounded linear operators do not admit a good topological structure
if the spaces are not normable. In particular, the chain rule will not hold for Fréchet-differentiability
in general for these spaces (cf. [31]).
20
Geir Bogfjellmo and Alexander Schmeding
Definition 30 (Complexification of a locally convex space). Let E be a real locally
convex topological vector space. Endow EC := E × E with the following operation
(x + iy).(u, v) := (xu − yv, xv + yu)
for x, y ∈ R, u, v ∈ E.
The complex vector space EC with the product topology is called the complexification of E. We identify E with the closed real subspace E × {0} of EC .
Definition 31. Let E, F be real locally convex spaces and f : U → F defined on an
open subset U. Following [40] and [17], we call f real analytic (or CRω ) if f extends
to a CC∞ -map f˜ : Ũ → FC on an open neighbourhood Ũ of U in the complexification
EC .11
Note that many of the usual results of differential calculus carry over to our setting. In particular, maps on connected domains whose derivative vanishes are constant as a version of the fundamental theorem of calculus holds. Moreover, the chain
rule holds in the following form:
Lemma 32 (Chain Rule [17, Propositions 1.12, 1.15, 2.7 and 2.9]). Fix k ∈ N0 ∪
k -maps f : E ⊇ U → F and g : H ⊇ V → E
{∞, ω } and K ∈ {R, C} together with CK
defined on open subsets of locally convex spaces. Assume that g(V ) ⊆ U. Then f ◦ g
k
is of class CK
and the first derivative of f ◦ g is given by
d( f ◦ g)(x; v) = d f (g(x); dg(x, v))
for all x ∈ U, v ∈ H.
The differential calculus developed so far extends easily to maps which are defined on non-open sets. This situation occurs frequently in the context of differential
equations on closed intervals (see [1] for an overview).
Having the chain rule at our disposal we can define manifolds and related constructions which are modelled on locally convex spaces.
Definition 33. Fix a Hausdorff topological space M and a locally convex space E
over K ∈ {R, C}. An (E-)manifold chart (Uκ , κ ) on M is an open set Uκ ⊆ M together with a homeomorphism κ : Uκ → Vκ ⊆ E onto an open subset of E. Two
such charts are called Cr -compatible for r ∈ N0 ∪ {∞, ω } if the change of charts
map ν −1 ◦ κ : κ (Uκ ∩ Uν ) → ν (Uκ ∩ Uν ) is a Cr -diffeomorphism. A Cr -atlas of M
is a set of pairwise Cr -compatible manifold charts, whose domains cover M. Two
such Cr -atlases are equivalent if their union is again a Cr -atlas.
A locally convex Cr -manifold M modelled on E is a Hausdorff space M with an
equivalence class of Cr -atlases of (E-)manifold charts.
Direct products of locally convex manifolds, tangent spaces and tangent bundles
as well as Cr -maps of manifolds may be defined as in the finite dimensional setting
(see [43, I.3]). The advantage of this construction is that we can now give a very
simple answer to the question, what an infinite-dimensional Lie group is:
11
If E and F are Fréchet spaces, real analytic maps in the sense just defined coincide with maps
which are continuous and can be locally expanded into a power series. See [18, Proposition 4.1].
The geometry of characters of Hopf algebras
21
∞Definition 34. A (locally convex) Lie group is a group G equipped with a CK
manifold structure modelled on a locally convex space, such that the group operations are smooth. If the manifold structure and the group operations are in addition
(K-) analytic, then G is called a (K-) analytic Lie group.
We recommend [43] for a survey on the theory of locally convex Lie groups.
However, the Lie groups constructed in this article have strong structural properties
as they belong to the class of Baker–Campbell–Hausdorff–Lie groups.
Definition 35 (Baker–Campbell–Hausdorff (BCH-)Lie groups and Lie algebras).
1. A Lie algebra g is called Baker–Campbell–Hausdorff–Lie algebra (BCH–Lie
algebra) if there exists an open 0-neighbourhoodU ⊆ g such that for x, y ∈ U the
BCH-series ∑∞
n=1 Hn (x, y) converges and defines an analytic map U × U → g.
The Hn are defined as H1 (x, y) = x + y, H2 (x, y) = 12 [x, y] and for n ≥ 3 by linear
combinations of iterated brackets, see [43, Definition IV.1.5.] or [9, Chapter 2,
§6].
2. A locally convex Lie group G is called BCH–Lie group if it satisfies one of the
following equivalent conditions (cf. [43, Theorem IV.1.8])
(i) G is a K-analytic Lie group whose Lie group exponential function is Kanalytic and a local diffeomorphism in 0.
(ii) The exponential map of G is a local diffeomorphism in 0 and L(G) is a
BCH–Lie algebra.
BCH–Lie groups share many of the structural properties of Banach Lie groups
while not necessarily being Banach Lie groups themselves.
4 Appendix: Characters and the exponential map
Fix for the rest of this section a K-Hopf algebra H = (H , mH , uH , ∆H , εH , SH )
and a commutative continuous inverse algebra B. Furthermore, we assume that the
L
Hopf algebra H is graded, i.e. H = n∈N0 H and dim H0 < ∞. The aim of this
section is to prove that the Lie group exponential map expA× of A := HomK (H , B)
restricts to a bijection from the infinitesimal characters to the characters.
Remark 36 (Cocomposition with Hopf multiplication). Let H ⊗ H be the tensor
Hopf algebra (cf. [35, p. 8]). We regard the tensor product H ⊗ H as a graded and
L
connected Hopf algebra with the tensor grading, i.e. H ⊗ H = n∈N0 (H ⊗ H )n
L
where for all n ∈ N0 the nth degree is defined as (H ⊗ H )n = i+ j=n Hi ⊗ H j .
Since dim H0 < ∞ we see that dim (H ⊗ H )0 = dim H0 ⊗ H0 < ∞ Thus with
respect to the topology of pointwise convergence and the convolution product, the
algebras
A := HomK (H , B)
A⊗ := HomK (H ⊗ H , B)
22
Geir Bogfjellmo and Alexander Schmeding
become continuous inverse algebras (see 6 and Lemma 10). This structure turns
· ◦ mH : HomK (H , B → HomK (H ⊗ H , B),
φ 7→ φ ◦ mH .
into a continuous algebra homomorphism. Hence its restriction
(mH )∗ : A× → A×
⊗,
φ 7→ φ ◦ mH
(3)
is a Lie group morphism with L((mH )∗ ) := Te (mH )∗ = · ◦ mH .
⋆n
x
Lemma 37. The Lie group exponential expA : L(A× ) = A → A× , x 7→ ∑∞
n=0 n!
maps g(H , B) to G (H , B). Further, there is a 0-neighborhood Ω ⊆ A such that
expA maps g(H , B) ∩ Ω bijectively onto G (H , B) ∩ expA (Ω ). 12
Proof. We denote by ⋆A⊗ the convolution product of the CIA A⊗ and let expA⊗
be the Lie group exponential of A×
⊗ . Let mB : B ⊗ B → B, b1 ⊗ b2 7→ b1 · b2 be the
multiplication in B. Define the continuous bilinear map (cf. [5, proof of Lemma
B.10] for detailed arguments)
β : A × A → A⊗ ,
(φ , ψ ) 7→ φ ⋄ ψ := mB ◦ (φ ⊗ ψ ).
(4)
We may use β to rewrite the convolution in A as ⋆A = β ◦ ∆ and obtain
(φ1 ⋄ ψ1 ) ⋆A⊗ (φ2 ⋄ ψ2 ) = (φ1 ⋆A φ2 ) ⋄ (ψ1 ⋆A ψ2 ).
(5)
Recall, that 1A := uB ◦ εH is the neutral element of the algebra A. From equation
(5), it follows at once, that the continuous linear maps
β (·, 1A ) : A → A⊗ ,
φ 7→ φ ⋄ 1A
and
β (1A , ·) : A → A⊗ ,
φ 7→ 1A ⋄ φ
are algebra homomorphisms which restrict to Lie group morphisms
β ρ : A× → A×
⊗,
φ 7→ β (φ , 1A ) and β λ : A× → A×
⊗,
φ 7→ β (1A , φ )
(6)
with L(β ρ ) = β (·, 1A ) and L(β λ ) = β (1A , ·). Let φ ∈ A be given and recall from
(5) that (φ ⋄ 1A ) ⋆A⊗ (1A ⋄ φ ) = φ ⋄ φ = (1A ⋄ φ ) ⋆A⊗ (φ ⋄ 1A ). As a consequence we
obtain
(7)
expA⊗ (φ ⋄ 1A + 1A ⋄ φ ) = expA⊗ (φ ⋄ 1A) ⋆A⊗ expA⊗ (1A ⋄ φ ).
since every Lie group exponential function transforms addition into multiplication
for commuting elements.
Note that it suffices to check multiplicativity of expA (φ ) as expA (φ )(1H ) = 1B
is automatically satisfied. For an infinitesimal character φ ∈ A we have by definition
φ ◦ mH = φ ⋄ 1A + 1A ⋄ φ . Using now the naturality of the Lie group exponentials
12
Note that apart from the locallity and several key arguments, the proof follows the general idea
of the similar statement [5, Lemma B.10]. For the readers convenience we repeat the arguments to
exhibit how properties of the Lie group exponential replace the functional calculus used in [5].
The geometry of characters of Hopf algebras
23
(i.e. for a Lie group morphism f : G → H we have expH ◦L( f ) = f ◦ expG ), we
derive the following:
Def
φ ∈ g(H , B) ⇐⇒ φ ◦ mH = φ ⋄ 1A + 1A ⋄ φ
(!)
=⇒ expA⊗ (φ ◦ mH ) = expA⊗ (φ ⋄ 1A + 1A ⋄ φ )
(7)
⇐⇒ expA⊗ (φ ◦ mH ) = expA⊗ (φ ⋄ 1A) ⋆A⊗ expA⊗ (1A ⋄ φ )
(6)
⇐⇒ expA⊗ (φ ◦ mH ) = expA (φ ) ⋄ 1A ⋆A⊗ 1A ⋄ expA (φ )
(5)
⇐⇒ expA⊗ (φ ◦ mH ) = expA (φ ) ⋆A 1A ⋄ 1A ⋆A expA (φ )
⇐⇒ expA⊗ (φ ◦ mH ) = expA (φ ) ⋄ expA (φ )
(3)
⇐⇒ expA (φ ) ◦ mH = expA (φ ) ⋄ expA (φ )
Def
⇐⇒ expA (φ ) ∈ G (H , B).
This shows that infinitesimal characters are mapped by the Lie group exponential to
elements in the character group.
Now we observe that in general the implication from the first to the second line
will not be an equivalence (as the Lie group exponential is not a global diffeomorphism unlike the connected Hopf algebra case discussed in [5, Lemma B.10]).
We exploit now that A× and A×
⊗ are locally exponential Lie groups, whence locally around 0 in A and A⊗ the Lie group exponentials induce diffeomorphisms.
Hence there are open neighborhoods of 0 and the units of A× and A×
⊗ , such that
expA× : A ⊇ V → W ⊆ A× and expA× : A⊗ ⊇ V⊗ → W⊗ ⊆ A×
are diffeomorphisms.
⊗
⊗
Since A⊗ is a locally convex space, there is an open 0-neighborhood Ω⊗ ⊆ A⊗ such
that Ω⊗ + Ω⊗ ⊆ V⊗ . By continuity of β and · ⊗ mH , we obtain an open open 0neighborhood
Ω := V ∩ (· ⊗ mH )−1 (V⊗ ) ∩ β (·, 1A )−1 (Ω⊗ ) ∩ β (1A , ·)−1 (Ω⊗ ) ⊆ A.
Now by construction elements in g(H , B) ∩ Ω are mapped by · ◦ mH into V× and
by β (·, 1A ) + β (1A, ·) into Ω⊗ + Ω⊗ ⊆ V⊗ . Since expA⊗ induces a diffeomorphism
on V⊗ , the implication (!) becomes an equivalence for elements in g(H , B) ∩ Ω .
We have thus established that expA maps g(H , B) ∩ Ω bijectively to expA (Ω ) ∩
G (H , B).
⊓
⊔
References
1. Alzaareer, H., Schmeding, A.: Differentiable mappings on products with different degrees
of differentiability in the two factors. Expo. Math. 33(2), 184–222 (2015). DOI 10.1016/j.
exmath.2014.07.002
2. Bastiani, A.: Applications différentiables et variétés différentiables de dimension infinie. J.
Analyse Math. 13, 1–114 (1964)
24
Geir Bogfjellmo and Alexander Schmeding
3. Beattie, M.: A survey of Hopf algebras of low dimension. Acta Appl. Math. 108(1), 19–31
(2009). DOI 10.1007/s10440-008-9367-3
4. Bertram, W., Glöckner, H., Neeb, K.H.: Differential calculus over general base fields and
rings. Expo. Math. 22(3), 213–282 (2004). DOI 10.1016/S0723-0869(04)80006-9
5. Bogfjellmo, G., Dahmen, R., Schmeding, A.: Character groups of Hopf algebras as infinitedimensional Lie groups. Ann. Inst. Fourier (Grenoble) 66(5), 2101–2155 (2016)
6. Bogfjellmo, G., Dahmen, R., Schmeding, A.: Overview of (pro-)Lie group structures on Hopf
algebra character groups. In: M. Barbero, K. Ebrahimi-Fard, D.M. de Diego (eds.) Discrete
Mechanics, Geometric Integration and Lie-Butcher Series, Springer Proceedings in Mathematics & Statistics (2017)
7. Bogfjellmo, G., Schmeding, A.: The tame Butcher group. J. Lie Theory 26, 1107–1144 (2016)
8. Bogfjellmo, G., Schmeding, A.: The Lie group structure of the Butcher group. Found. Comput.
Math. 17(1), 127–159 (2017). DOI 10.1007/s10208-015-9285-5
9. Bourbaki, N.: Lie groups and Lie algebras. Chapters 1–3. Elements of Mathematics (Berlin).
Springer-Verlag, Berlin (1998). Translated from the French, Reprint of the 1989 English
translation
10. Brouder, C.: Trees, renormalization and differential equations. BIT Num. Anal. 44, 425–438
(2004)
11. Bruned, Y., Hairer, M., Zambotti, L.: Algebraic renormalisation of regularity structures (2016).
URL http://arxiv.org/abs/1610.08468v1
12. Butcher, J.C.: An algebraic theory of integration methods. Math. Comp. 26, 79–106 (1972)
13. Cartier, P.: A primer of Hopf algebras. In: Frontiers in number theory, physics, and geometry.
II, pp. 537–615. Springer, Berlin (2007)
14. Connes, A., Marcolli, M.: Noncommutative geometry, quantum fields and motives, American
Mathematical Society Colloquium Publications, vol. 55. American Mathematical Society,
Providence, RI; Hindustan Book Agency, New Delhi (2008)
15. Floret, K.: Lokalkonvexe Sequenzen mit kompakten Abbildungen. J. Reine Angew. Math.
247, 155–195 (1971)
16. Glöckner, H.: Algebras whose groups of units are Lie groups. Studia Math. 153(2), 147–177
(2002). DOI 10.4064/sm153-2-4. URL http://dx.doi.org/10.4064/sm153-2-4
17. Glöckner, H.: Infinite-dimensional Lie groups without completeness restrictions. In: Geometry and analysis on finite- and infinite-dimensional Lie groups (Bedlewo, 2000), Banach
‘
Center Publ., vol. 55, pp. 43–59. Polish Acad. Sci. Inst. Math., Warsaw
(2002). DOI
10.4064/bc55-0-3
18. Glöckner, H.: Instructive examples of smooth, complex differentiable and complex analytic
mappings into locally convex spaces. J. Math. Kyoto Univ. 47(3), 631–642 (2007). DOI
10.1215/kjm/1250281028. URL http://dx.doi.org/10.1215/kjm/1250281028
19. Glöckner, H.: Measurable regularity properties of infinite-dimensional Lie groups (2015).
URL http://arxiv.org/abs/1601.02568v1
20. Glöckner, H.: Regularity properties of infinite-dimensional Lie groups, and semiregularity
(2015). URL http://arxiv.org/abs/1208.0715v3
21. Glöckner, H., Neeb, K.H.: When unit groups of continuous inverse algebras are regular Lie groups. Studia Math. 211(2), 95–109 (2012). DOI 10.4064/sm211-2-1. URL
http://dx.doi.org/10.4064/sm211-2-1
22. Glöckner, H., Neeb, K.H.: Infinite-dimensional Lie Groups. General Theory and Main Examples (2018). Unpublished
23. Gracia-Bondı́ a, J.M., Várilly, J.C., Figueroa, H.: Elements of noncommutative geometry.
Birkhäuser Advanced Texts: Basler Lehrbücher. [Birkhäuser Advanced Texts: Basel Textbooks]. Birkhäuser Boston, Inc., Boston, MA (2001). DOI 10.1007/978-1-4612-0005-5
24. Hairer, E., Lubich, C., Wanner, G.: Geometric Numerical Integration, Springer Series in Computational Mathematics, vol. 31. Springer Verlag (2 2006)
25. Hairer,
M.:
A
theory
of
regularity
structures.
Invent.
Math.
198(2), 269–504 (2014).
DOI 10.1007/s00222-014-0505-4.
URL
http://dx.doi.org/10.1007/s00222-014-0505-4
The geometry of characters of Hopf algebras
25
26. Hewitt, E., Ross, K.A.: Abstract harmonic analysis. Vol. II: Structure and analysis for compact
groups. Analysis on locally compact Abelian groups. Die Grundlehren der mathematischen
Wissenschaften, Band 152. Springer-Verlag, New York-Berlin (1970)
27. Hofmann, K.H., Morris, S.A.: The Lie theory of connected pro-Lie groups, EMS Tracts in
Mathematics, vol. 2. EMS, Zürich (2007). DOI 10.4171/032
28. Hofmann, K.H., Morris, S.A.: The structure of compact groups, De Gruyter Studies in Mathematics, vol. 25. De Gruyter, Berlin (2013). DOI 10.1515/9783110296792. A primer for the
student—a handbook for the expert, Third edition, revised and augmented
29. Hofmann, K.H., Morris, S.A.: Pro-Lie groups: A survey with open problems. Axioms 4,
294–312 (2015). DOI 10.3390/axioms4030294
30. Jarchow, H.: Locally convex spaces. B. G. Teubner, Stuttgart (1981). Mathematische
Leitfäden. [Mathematical Textbooks]
31. Keller, H.: Differential Calculus in Locally Convex Spaces. Lecture Notes in Mathematics
417. Springer Verlag, Berlin (1974)
32. Kock, J.: Perturbative renormalisation for not-quite-connected bialgebras. Lett. Math. Phys.
105(10), 1413–1425 (2015). DOI 10.1007/s11005-015-0785-7
33. König, W.: The parabolic Anderson model. Pathways in Mathematics. Birkhäuser/Springer,
[Cham] (2016). DOI 10.1007/978-3-319-33596-4. Random walk in random potential
34. Kriegl, A., Michor, P.W.: The convenient setting of global analysis, Mathematical Surveys and
Monographs, vol. 53. AMS (1997)
35. Majid, S.: Foundations of quantum group theory. Cambridge University Press, Cambridge
(1995). DOI 10.1017/CBO9780511613104
36. Mallios, A.: Topological algebras. Selected topics, North-Holland Mathematics Studies, vol.
124. North-Holland Publishing Co., Amsterdam (1986). Notas de Matemática [Mathematical
Notes], 109
37. Manchon, D.: Hopf algebras in renormalisation. In: Handbook of algebra. Vol. 5, Handb.
Algebr., vol. 5, pp. 365–427. Elsevier/North-Holland, Amsterdam (2008). DOI 10.1016/
S1570-7954(07)05007-3
38. McLachlan, R.I., Modin, K., Munthe-Kaas, H., Verdier, O.: B-series methods are exactly
the affine equivariant methods. Numer. Math. 133(3), 599–622 (2016). DOI 10.1007/
s00211-015-0753-2. URL http://dx.doi.org/10.1007/s00211-015-0753-2
39. Michaelis, W.: Coassociative coalgebras. In: Handbook of algebra, Vol. 3, pp. 587–
788. North-Holland, Amsterdam (2003). DOI 10.1016/S1570-7954(03)80072-4. URL
http://dx.doi.org/10.1016/S1570-7954(03)80072-4
40. Milnor, J.: Remarks on infinite-dimensional Lie groups. In: Relativity, groups and topology,
II (Les Houches, 1983), pp. 1007–1057. North-Holland, Amsterdam (1984)
41. Milnor, J.W., Moore, J.C.: On the structure of Hopf algebras. Ann. of Math. (2) 81, 211–264
(1965). DOI 10.2307/1970615. URL http://dx.doi.org/10.2307/1970615
42. Murua, A., Sanz-Serna, J.M.: Computing normal forms and formal invariants of dynamical
systems by means of word series. Nonlinear Anal. 138, 326–345 (2016). DOI 10.1016/j.na.
2015.10.013. URL http://dx.doi.org/10.1016/j.na.2015.10.013
43. Neeb, K.H.: Towards a Lie theory of locally convex groups. Jpn. J. Math. 1(2), 291–468
(2006). DOI 10.1007/s11537-006-0606-y
44. Schaefer, H.H.: Topological vector spaces. Springer-Verlag, New York-Berlin (1971). Third
printing corrected, Graduate Texts in Mathematics, Vol. 3
45. Swan, R.G.: Topological examples of projective modules. Trans. Amer. Math. Soc. 230, 201–
234 (1977). DOI 10.2307/1997717. URL http://dx.doi.org/10.2307/1997717
46. Sweedler, M.E.: Hopf algebras. Mathematics Lecture Note Series. W. A. Benjamin, Inc., New
York (1969)
| 4math.GR
|
Covariance Functions for Multivariate Gaussian Fields
Evolving Temporally Over Planet Earth
arXiv:1701.06010v2 [math.ST] 21 Nov 2017
Alfredo Alegrı́a1 , Emilio Porcu1,2 , Reinhard Furrer3 and Jorge Mateu4
1
School of Mathematics & Statistics, Newcastle University,
Newcastle Upon Tyne, United Kingdom.
2
Departamento de Matemática, Universidad Técnica Federico Santa Maria,
Valparaı́so, Chile.
3
Department of Mathematics and Department of Computer Science,
University of Zurich, Zurich, Switzerland.
4
Department of Mathematics, University Jaume I,
Castellón, Spain.
November 23, 2017
Abstract
The construction of valid and flexible cross-covariance functions is a fundamental task for modeling multivariate space-time data arising from climatological and oceanographical phenomena.
Indeed, a suitable specification of the covariance structure allows to capture both the space-time
dependencies between the observations and the development of accurate predictions. For data
observed over large portions of planet Earth it is necessary to take into account the curvature
of the planet. Hence the need for random field models defined over spheres across time. In particular, the associated covariance function should depend on the geodesic distance, which is the
most natural metric over the spherical surface. In this work, we propose a flexible parametric
family of matrix-valued covariance functions, with both marginal and cross structure being of
the Gneiting type. We additionally introduce a different multivariate Gneiting model based on
the adaptation of the latent dimension approach to the spherical context. Finally, we assess the
performance of our models through the study of a bivariate space-time data set of surface air
temperatures and precipitations.
Keywords: Geodesic; Gneiting classes; Latent dimensions; Precipitations; Space-time; Sphere;
Temperature.
1
Introduction
Monitoring several georeferenced variables is a common practice in a wide range of disciplines
such as climatology and oceanography. The phenomena under study are often observed over large
portions of the Earth and across several time instants. Since there is only a finite sample from the
involved variables, geostatistical models are a useful tool to capture both the spatial and temporal
interactions between the observed data, as well as the uncertainty associated to the limited available
information (Cressie, 1993; Wackernagel, 2003; Gneiting et al., 2007). The geostatistical approach
1
consists in modeling the observations as a partial realization of a space-time multivariate random
field (RF), denoted as {Z(x, t) = (Z1 (x, t), . . . , Zm (x, t))> : (x, t) ∈ D × T }, where > is the
transpose operator and m is a positive integer representing the number of components of the field.
If m = 1, we say that Z(x, t) is a univariate (or scalar-valued) RF, whereas for m > 1, Z(x, t)
is called an m-variate (or vector-valued) RF. Here, D and T denote the spatial and temporal
domains, respectively. Throughout, we assume that Z(x, t) is a zero mean Gaussian field, so that
a suitable specification of its covariance structure is crucial to develop both accurate inferences and
predictions over unobserved sites (Cressie, 1993).
Parametric families of matrix-valued covariance functions are typically given in terms of Euclidean
distances. The literature for this case is extensive and we refer the reader to the review by Genton
and Kleiber (2015) with the references therein. The main motivation to consider the Euclidean
metric is the existence of several methods for projecting the geographical coordinates, longitude
and latitude, onto the plane. However, when a phenomenon is observed over large portions of
planet Earth, the approach based on projections generates distortions in the distances associated
to distant locations on the globe. The reader is referred to Banerjee (2005), where the impact of
the different types of projections with respect to spatial inference is discussed.
Indeed, the geometry of the Earth must be considered. Thus, it is more realistic to work under
the framework of RFs defined spatially on a sphere (see Marinucci and Peccati, 2011). Let d be
a positive integer. The d-dimensional unit sphere is denoted as Sd := {x ∈ Rd+1 : kxk = 1},
where k · k represents the Euclidean norm. The most accurate metric in the spherical scenario is
the geodesic (or great circle) distance, which roughly speaking corresponds to the arc joining any
two points located on the sphere, measured along a path on the spherical surface. Formally, the
geodesic distance is defined as the mapping θ : Sd ×Sd → [0, π] given by θ := θ(x, y) = arccos(x> y).
The construction of valid and flexible parametric covariance functions in terms of the geodesic
distance is a challenging problem and requires the application of the theory of positive definite
functions on spheres (Schoenberg, 1942; Yaglom, 1987; Hannan, 2009; Berg and Porcu, 2017). In
the univariate and merely spatial case, Huang et al. (2011) study the validity of some specific
covariance functions. The essay by Gneiting (2013) contains a wealth of results related to the
validity of a wide range of covariance families. Other related works are the study of star-shaped
random particles (Hansen et al., 2011) and convolution roots (Ziegel, 2014). However, the spatial
and spatio-temporal covariances in the multivariate case are still unexplored, with the work of
Porcu et al. (2016) being a notable exception.
The Gneiting class (Gneiting, 2002) is one of the most popular space-time covariance families and
some adaptations in terms of geodesic distance have been given by Porcu et al. (2016). In this
paper, we extend to the multivariate scenario the modified Gneiting class introduced by Porcu
et al. (2016). Furthermore, we adapt the latent dimension approach to the spherical context
(Apanasovich and Genton, 2010) and we then generate additional Gneiting type matrix-valued
covariances. The proposed models are non-separable with respect to the components of the field
nor with respect to the space-time interactions. To obtain these results, we have demonstrated
several technical results that can be useful to develop new research in this area. Our findings are
illustrated through a real data application. In particular, we analyze a bivariate space-time data
set of surface air temperatures and precipitations. These variables have been obtained from the
Community Climate System Model (CCSM) provided by NCAR (National Center for Atmospheric
Research).
The remainder of the article is organized as follows. Section 2 introduces some concepts and
2
construction principles of cross-covariances on spheres across time. In Section 3 we propose some
multivariate Gneiting type covariance families. Section 4 contains a real data example of surface
air temperatures and precipitations. Section 5 concludes the paper with a discussion.
2
Fundaments and Principles of Space-time Matrix-valued Covariances
This section is devoted to introduce the basic material related to multivariate fields on spheres
across time. Let Z(x, t), for x ∈ Sd and t ∈ R, be an m-variate space-time field and let C : [0, π] ×
R → Rm×m be a continuous matrix-valued mapping, whose elements are defined as Cij (θ, u) =
cov{Zi (x, t + u), Zj (y, t)}, where θ = θ(x, y). Following Porcu et al. (2016), we say that C is
geodesically isotropic in space and stationary in time. Throughout, the diagonal elements Cii are
called marginal covariances, whereas the off-diagonal members Cij are called cross-covariances. Any
parametric representation of C must respect the positive definite condition, i.e. it must satisfy
n X
n
X
a>
` C(θ(x` , xr ), t` − tr )ar ≥ 0,
(2.1)
`=1 r=1
for all positive integer n, {(x1 , t1 ), . . . , (xn , tn )} ⊂ Sd × R and {a1 , . . . , an } ⊂ Rm . Appendix A
contains a complete background material on matrix-valued positive definite functions on Euclidean
and spherical domains.
We call the mapping C space-time m-separable if there exists two mappings CS : [0, π] → Rm×m
and CT : R → Rm×m , being merely spatial and temporal matrix-valued covariances, respectively,
such that
C(θ, u) = CS (θ) ◦ CT (u),
(θ, u) ∈ [0, π] × R,
where ◦ denotes the Hadamard product. We call the space-time m-separability property complete
if there exists a symmetric, positive definite matrix A ∈ Rm×m , a univariate spatial covariance
CS : [0, π] → R , and a univariate temporal covariance CT : R → R, such that
C(θ, u) = ACS (θ)CT (u),
(θ, u) ∈ [0, π] × R.
We finally call m-separable the mapping C for which there exists a univariate space-time covariance
C : [0, π] × R → R, and a matrix A, as previously defined, such that
(θ, u) ∈ [0, π] × R,
C(θ, u) = AC(θ, u),
and of course the special case C(θ, u) = CS (θ)CT (u) offers complete space-time m-separability as
previously discussed.
Separability is a very useful property for both modeling and estimation purposes, because the
related covariance matrices admit nice factorizations, with consequent alleviation of the computational burdens. At the same time, it is a very unrealistic assumption and the literature has been
focussed on how to develop non-separable models. How to escape from separability is a major deal,
and we list some strategies that can be adapted from others proposed in Euclidean spaces. To
the knowledge of the authors, none of these strategies have ever been implemented on spheres or
spheres across time.
3
Linear Models of Coregionalization Let q be a positive integer. Given a collection of matrices
Ak , k = 1, . . . , q, and a collection of univariate space-time covariances Ck : [0, π] × R → R,
the linear model of coregionalization (LMC) has expression
C(θ, u) =
q
X
Ak Ck (θ, u),
(θ, u) ∈ [0, π] × R,
k=1
where a simplification of the type Ck (θ, u) = Ck,S (θ)CT (u) might be imposed. This model
has several drawbacks that have been discussed in Gneiting et al. (2010) as well as in Daley
et al. (2015).
Lagrangian Frameworks Let Z be an m-variate Gaussian field on the sphere with covariance
CS : [0, π] → Rm×m . Let R be a random orthogonal matrix with a given probability law. Let
Y (x, t) := Z Rt x ,
(x, t) ∈ Sd × R.
Then, Y is a RF with transport effect over the sphere. Clearly, Y is not Gaussian and
evaluation of the corresponding covariance might be cumbersome, as shown in Alegrı́a and
Porcu (2017).
Multivariate Parametric Adaptation Let p be a positive integer. Let C(·, ·; λ), for λ ∈ Rp ,
be a univariate space-time covariance. Let λij ∈ Rp , for i, j = 1, . . . , m. For |ρij | ≤ 1 and
ρii = 1, and σii > 0, find the parametric restrictions such that C : [0, π] × R → Rm×m defined
through
Cij (θ, u) = σii σjj ρij C(θ, u; λij ),
(θ, u) ∈ [0, π] × R,
is a valid covariance. Sometimes the restriction on the parameters can be very severe, in
particular when m is bigger than 2. In Euclidean spaces this strategy has been adopted by
Gneiting et al. (2010) and Apanasovich et al. (2012) for the Matérn model, and by Daley
et al. (2015) for models with compact support.
Scale Mixtures Let (Ω, A, µ) be a measure space. Let C : [0, π] × R × Ω → Rm×m such that
1. C(·, ·, ξ) is a valid covariance for all ξ in Ω;
2. C(θ, u, ·) ∈ L1 (Ω, A, µ) for all (θ, u) ∈ [0, π] × R.
Then,
Z
C(θ, u, ξ)µ(dξ)
Ω
is still a valid covariance (Porcu and Zastavnyi, 2011). Of course, simple strategies can be very
effective. For instance, one might assume that C(·, ·, ξ) = C(·, ·)A(ξ), with C a univariate
covariance and A(ξ) ∈ Rm×m being a positive definite matrix for any ξ ∈ Ω and such that
the hypothesis of integrability above is satisfied.
3
Matrix-valued Covariance Functions of Gneiting Type
This section provides some general results for the construction of multivariate covariances for fields
defined on Sd ×R. The most important feature of the proposed models is that they are non-separable
mappings, allowing more flexibility in the study of space-time phenomena. In particular, we focus
4
Table 3.1: Some examples of completely monotone functions. Here, Kν denotes the modified Bessel
function of second kind of degree ν.
Function
γ)
g(t) = exp(−ct
√
√
g(t) = (2ν−1 Γ(ν))−1 (c t)ν Kν (c t)
g(t) = √
(1 + ctγ )−ν √
g(t) = 2ν (exp(c t) + exp(−c t))−ν
Parameters restriction
c > 0, 0 < γ ≤ 1
c > 0, ν > 0
c > 0, 0 < γ ≤ 1, ν > 0
c > 0, ν > 0
Table 3.2: Some examples of Bernstein functions.
Function
f (t) = (atα + 1)β
f (t) = ln(atα + b)/ ln(b)
f (t) = (atα + b)/(b(atα + 1))
Parameters restriction
a > 0, 0 < α ≤ 1, 0 ≤ β ≤ 1
a > 0, b > 1 0 < α ≤ 1
a > 0, 0 < b ≤ 1, 0 < α ≤ 1
on covariances with a Gneiting’s structure (Gneiting, 2002). See also Bourotte et al. (2016) for
multivariate Gneiting classes on Euclidean spaces.
Recently, Porcu et al. (2016) have proposed some versions of the Gneiting model for RFs with
spatial locations on the unit sphere. Let us provide a brief review. The mapping C : [0, π] × R → R
defined as
1
|u|2
C(θ, u) =
,
θ ∈ [0, π], u ∈ R,
(3.1)
g
1/2
f (θ)|[0,π]
f (θ)|
[0,π]
is called Gneiting model. Arguments in Porcu et al. (2016) show that C is a valid univariate
covariance, for all d ∈ N, for g : [0, ∞) → [0, ∞) being a completely monotone function, i.e.
g is infinitely differentiable on (0, ∞) and (−1)n g (n) (t) ≥ 0, for all n ∈ N and t > 0, and f :
[0, ∞) → (0, ∞) is a positive function with completely monotone derivative. Such functions f are
called Bernstein functions (Porcu and Schilling, 2011). Here, f (θ)|[0,π] denotes the restriction of the
mapping f to the interval [0, π]. Tables 3.1 and 3.2 contain some examples of completely monotone
and Bernstein functions, respectively. Additional properties about these functions are studied in
Porcu and Schilling (2011).
We also pay attention to the modified Gneiting class (Porcu et al., 2016) defined through
C(θ, u) =
1
g (θf (|u|)) ,
f (|u|)n+2
(θ, u) ∈ [0, π] × R,
(3.2)
where n ≤ 3 is a positive integer, g : [0, ∞) → [0, ∞) is a completely monotone function and
f : [0, ∞) → (0, ∞) is a strictly increasing and concave function on the positive real line. The
mapping (3.2) is a valid covariance for any d ≤ 2n + 1.
3.1
Multivariate Modified Gneiting Class on Spheres Acros Time
We are now able to illustrate the main result within this section. Specifically, we propose a multivariate space-time class generating Gneiting functions with different scale parameters.
Theorem 3.1. Let m ≥ 2 and n ≤ 3 be positive integers. Let g : [0, ∞) → [0, ∞) be a completely
monotone function. Consider f : [0, ∞) → (0, ∞) being strictly increasing and concave. Let σii > 0,
5
|ρij | ≤ 1 and cij > 0, for i, j = 1, . . . , m, be constants yielding the additional condition
X
|ρij |(cii /cij )n+1 ≤ 1,
(3.3)
i6=j
for each i = 1, . . . , m. Then, the mapping C, whose members Cij : [0, π] × R → R are defined
through
σii σjj ρij
θf (|u|)
g
,
(θ, u) ∈ [0, π] × R,
(3.4)
Cij (θ, u) =
f (|u|)n+2
cij
is a matrix-valued covariance for any d ≤ 2n + 1.
The condition (3.3) comes from the arguments in Daley et al. (2015). The proof of Theorem 3.1 is
deferred to Appendix B, coupled with the preliminary notation and results introduced in Appendix
A.
For example, taking n = 1, g(t) = exp(−3t) and f (t) = 1+1.7t/cT , for cT > 0, we can use Equation
(3.4), and the restrictions of Theorem 3.1, to generate a model of the form
1.7|u|
3θ
1
+
cT
σii σjj ρij
Cij (θ, u) =
exp
−
,
(θ, u) ∈ [0, π] × R.
(3.5)
3
cij
1.7|u|
1 + cT
The special parameterization used in the covariance (3.5) ensures that Cij (θ, 0)/(σii σjj ) < 0.05 and
Cij (0, |u|)/(σii σjj ) < 0.05, for θ > cij and |u| > cT , respectively.
3.2
A Multivariate Gneiting Family Based on Latent Dimension Approaches
We now propose families of matrix-valued covariances obtained from univariate models. The
method used is known as latent dimension approach and has been studied in the Euclidean case
by Porcu et al. (2006), Apanasovich and Genton (2010) and Porcu and Zastavnyi (2011). Next, we
illustrate the spirit of this approach.
Consider a univariate Gaussian RF defined on the product space Sd × R × Rk , for some positive
integers d and k, namely {Y (x, t, ξ) : (x, t, ξ) ∈ Sd × R × Rk }. Suppose that there exists a mapping
K : [0, π] × R × Rk → R such that cov{Y (x, t, ξ1 ), Y (y, s, ξ2 )} = K(θ(x, y), t − s, ξ1 − ξ2 ), for all
x, y ∈ Sd , t, s ∈ R and ξ1 , ξ2 ∈ Rk . The idea is to define the components of an m-variate RF on
Sd × R as
Zi (x, t) = Y (x, t, ξi ),
x ∈ Sd , t ∈ R, ξi ∈ Rk ,
for i = 1, . . . , m. Thus, the resulting covariance, C(·) = [Cij (·)]m
i,j=1 , associated to Z(x, t) is given
by
Cij (θ, u) = K(θ, u, ξi − ξj ),
θ ∈ [0, π], u ∈ R.
Here, the vectors ξi are handled as additional parameters of the model. The following theorem
allows to construct different versions of the Gneiting model, using the latent dimension technique.
Theorem 3.2. Consider the positive integers d, k and l. Let g be a completely monotone function
and fi , i = 1, 2, Bernstein functions. Then,
2
1
kuk
i ,
K(θ, u, v) =
g h
(θ, u, v) ∈ [0, π] × Rl × Rk ,
n h
io
k/2
kvk2
2
kvk
f1 f2 (θ)|
{f2 (θ)|[0,π] }l/2 f1
f2 (θ)|[0,π]
[0,π]
(3.6)
6
is a univariate covariance for fields defined on Sd × Rl × Rk .
The proof of Theorem 3.2 requires technical lemmas and is deferred to Appendix C, coupled with
the preliminary results introduced in Appendix A.
In order to avoid an excessive number of parameters, we follow the parsimonious strategy proposed
by Apanasovich and Genton (2010). Consider k = 1 and the scalars {ξ1 , . . . , ξm }. We can consider
the parameterization ζij = |ξi − ξj |2 , with ζij = ζji > 0 and ζii = 0, for all i and j.
Following Apanasovich and Genton (2010), a LMC based on the latent dimension approach can
be used to achieve different marginal structures. Indeed, suppose that the components of a bivariate field are given by Z1 (x, t) = a11 Y (x, t, ξ1 ) and Z2 (x, t) = a21 Y (x, t, ξ2 ) + a22 W (x, t), where
Y (x, t, ξ) is a RF with covariance
3|u|/cT,1
1
K(θ, u, ζij ) =
(3.7)
1/2 exp −
1/2 ,
1/2
400θ
ζij + 1
ζij + 1
cS,1 + 1
400θ
400θ
+1
+1
cS,1
cS,1
generated from Equation (3.6) and the first entries in Tables 3.1 and 3.2. The field W (x, t) is
independent of Y (x, t, ξ), with covariance function of Gneiting type as in Equation (3.1),
3|u|/cT,2
1
(3.8)
R(θ, u) =
1/2 exp −
1/2 .
400θ
400θ
+
1
+
1
cS,2
cS,2
Thus, the covariance function of Z(x, t), with K and R defined by (3.7) and (3.8), respectively, is
given by
C11 (θ, u) = a211 K(θ, u, 0)
C22 (θ, u) = a221 K(θ, u, 0) + a222 R(θ, u)
C12 (θ, u) = a11 a21 K(θ, u, ζ12 ).
(3.9)
Note that R(θ, 0) < 0.05 and R(0, |u|) < 0.05, for θ > cS,2 and |u| > cT,2 , respectively. Moreover,
since C12 (0, 0) = a11 a21 /(ζ12 + 1)1/2 , the parameter ζ12 is related to the cross-scale as well as the
collocated correlation coefficient between the fields.
4
Data Example: Temperatures and Precipitations
We illustrate the use of the proposed models on a climatological data set. We consider a bivariate
space-time data set of surface air temperatures (Variable 1) and precipitations (Variable 2) obtained
from a climate model provided by NCAR (National Center for Atmospheric Research), Boulder,
CO, USA. Specifically, the data come from the Community Climate System Model (CCSM4.0) (see
Gent et al., 2011). Temperatures and precipitations are physically related variables and their joint
modeling is crucial in climatological disciplines.
7
Figure 4.1: Residuals of the surface air temperatures (left) and precipitations (right) for December
2006, in the area with longitudes between 50 and 150 degrees, and latitudes between −50 and 0
degrees.
In order to attenuate the skewness of precipitations, we work with its cubic root. The units are
degrees Kelvin for temperatures and centimeters for the transformed precipitations. The spatial
grid is formed by longitudes and latitudes with 2.5 × 2.5 degrees of spatial resolution (10368 grid
points). We model planet Earth as a sphere of radius 6378 kilometers. We focus on December,
between the years 2006 and 2015 (10 time instants), and the region with longitudes between 50 and
150 degrees, and latitudes between −50 and 0 degrees. The maximum distance between the points
inside this region is approximately half of the maximum distance between any pair of points on
planet Earth. Indeed, we are considering a large portion of the globe. For each variable and time
instant, we use splines to remove the cyclic patterns along longitudes and latitudes. The residuals
are approximately Gaussian distributed with zero mean (see Figure 4.1). We attenuate the computational burden by taking a sub grid with 10 degrees of resolution in the longitud direction. We
thus work with only 200 spatial sites and the resulting data set consists of 4000 observations. The
empirical variances are 2.033 and 4.809×10−5 , for Variables 1 and 2, respectively. In addition, these
variables are moderately positively correlated. The empirical correlation between the components
is 0.279.
We are interested in showing that in real applications a non-separable model can produce significant
improvements with respect to a separable one. For our purposes, we consider the following models:
(A) An m-separable covariance based on the univariate Gneiting class (3.1):
σii σjj ρij
−3|u|/cT
Cij (θ, u) =
1/2 exp
1/2 ,
400θ
400θ
cS + 1
cS + 1
2 , σ 2 , ρ , c , c )> .
where the vector of parameters is given by (σ11
22 12 S T
(B) An m-separable version of the modified Gneiting class (3.5) with the parameterization cS :=
c11 = c22 = c12 . The vector of parameters is the same as in Model A.
(C) A non-separable version of the modified Gneiting class (3.5), with the parsimonious parame2 , σ 2 , ρ , c , c , c )> .
terization c12 = max{c11 , c22 }. The vector of parameters is given by (σ11
22 12 11 22 T
8
Figure 4.2: Empirical spatial semi-variograms versus theoretical Models C and D. At the top, we
consider the temporal lag as |u| = 0, whereas at the bottom |u| = 1.
Table 4.1: CL estimates for Models A-C. For the separable Models A and B, there is a single spatial
scale parameter: cS := c11 = c22 = c12 . For Model C, the cross scale parameter c12 is the maximum
between the marginal spatial scales c11 and c22 .
Model A
Model B
Model C
2
σ
b11
1.85
1.84
1.85
2
σ
b22
4.71 × 10−5
4.72 × 10−5
4.69 × 10−5
ρb12
0.28
0.28
0.28
b
c11
102
2602
2901
b
c22
102
2602
2245
b
c12
102
2602
2901
b
cT
11.88
22.58
22.92
(D) The non-separable LMC (3.9) based on the latent dimension approach. Here, the vector of
parameters is given by (a11 , a21 , a22 , cS,1 , cS,2 , cT,1 , cT,2 , ζ12 )> .
We estimate the models parameters using the pairwise composite likelihood (CL) method developed
for multivariate RFs by Bevilacqua et al. (2016). This method is a likelihood approximation
and offers a trade-off between statistical and computational efficiency. We only consider pairs of
observations with spatial distances less than 1275.6 kilometers (equivalent to 0.2 radians on the
unit sphere) and temporal distance less than 4 years.
The CL estimates for Models A-C are reported in Table 4.1, whereas Table 4.2 reports the results
for Model D. In addition, Table 4.3 contains the Log-CL values attained at the optimum for each
model. As expected, the non-separable covariances C and D exhibit the highest values of the LogCL. In Figure 4.2, we compare the empirical spatial semi-variograms, for the temporal lags |u| = 0
and |u| = 1, versus the theoretical models C and D. Both models seem to capture the behavior of
the empirical variograms at the limit values (sills). However, as noted by Gneiting et al. (2010),
disagreements between empirical and theoretical fits are typically observed in practice, and it can
be associated to biases in the empirical estimators.
9
Table 4.2: CL estimates for Model D.
Model D
b
a11
1.38
b
a21
3.98 × 10−3
b
a22
5.58 × 10−3
b
cS,1
47133
b
cS,2
30805
b
cT,1
41.36
b
cT,2
4.03
ζb12
1.66
Table 4.3: Comparison between Models A-D in terms of Log-CL and cross-validation scores. Here,
MSEi and CRPSi denote respectively the MSE and CRPS associated to Variable i, for i = 1, 2.
Model
Model
Model
Model
A
B
C
D
Log-CL ×10−5
5.79
5.80
5.81
5.81
MSE1
0.27
0.13
0.11
0.11
MSE2 × 105
2.07
1.62
1.61
1.84
CRPS1
1.21
0.67
0.62
0.63
CRPS2
1.52
1.02
0.93
0.62
Finally, we compare the predictive performance of the covariances through a cross-validation study
based on the cokriging predictor. We use a drop-one prediction strategy and quantify the discrepancy between the real and predicted values for each variable through the Mean Squared Error
(MSE) and the Continuous Ranked Probability Score (CRPS) (Zhang and Wang, 2010). Smaller
values of these indicators imply better predictions. Table 4.3 displays the results, where MSEi
denotes the MSE associated to the variable i, for i = 1, 2. The interpretation of CRPSi is similar. Apparently, the performance of Model A is quite poor in comparison to the other models.
Moreover, the non-separable models, C and D, have better results than the m-separable model B.
Note that although Model B has a smaller MSE2 value than Model D, D is globally superior. If we
choose C instead of B, we have a significant improvement in the prediction of Variable 1, since the
corresponding MSE1 ratio is approximately 0.84. Similarly, the CRPS1 ratio is 0.93. Also, direct
inspection of MSE2 and CRPS2 show that Model C provides improvements in the prediction of
Variable 2 in comparison to Model B. Indeed, for this specific data set, Model C shows the best
results in terms of MSE, outperforming even Model D. A nice property of the multivariate modified
Gneiting class C is that it has physically interpretable parameters. Naturally, the proposed models
can be combined to provide more flexible models.
5
Discussion
In the paper, we have discussed several construction principles for multivariate space-time covariances, which allow to escape from separability. In particular, we have proposed Gneiting type
families of cross-covariances and their properties have been illustrated through a real data example
of surface air temperatures and precipitations. The proposed models have shown a good performance using separable models as a benchmark. We believe that the methodology used to prove
our theoretical results can be adapted to find additional flexible classes of matrix-valued covariances and to develop new applications on univariate or multivariate global phenomena evolving
temporally. Specifically, our proposals can be used as a building block for the construction of more
sophisticated models, such as non-stationary RFs.
10
Acknowledgments
Alfredo Alegrı́a’s work was initiated when he was a PhD student at Universidad Técnica Federico
Santa Marı́a. Alfredo Alegrı́a was supported by Beca CONICYT-PCHA/Doctorado Nacional/201621160371. Emilio Porcu is supported by Proyecto Fondecyt Regular number 1170290. Reinhard
Furrer acknowledges support of the Swiss National Science Foundation SNSF-144973. Jorge Mateu
is partially supported by grant MTM2016-78917-R. We additionally acknowledge the World Climate
Research Programme’s Working Group on Coupled modeling, which is responsible for Coupled
Model Intercomparison Project (CMIP).
Appendices
A
Background on Positive Definite Functions
We start by defining positive definite functions, that arise in statistics as the covariances of Gaussian
RFs as well as the characteristic functions of probability distributions.
Definition A.1. Let E be a non-empty set and m ∈ N. We say that the matrix-valued function
F : E × E → Rm×m is positive definite if for all integer n ≥ 1, {e1 , . . . , en } ⊂ E and {a1 , . . . , an } ⊂
Rm , the following inequality holds:
n X
n
X
a>
` F(e` , er )ar ≥ 0.
(A.1)
`=1 r=1
We denote as P m (E) the class of such mappings F satisfying Equation (A.1).
Next, we focus on the cases where E is either Rd , Sd or Sd ×Rk , for d, k ∈ N. For a clear presentation
of the results, Table A.1 summarizes the notation introduced along this Appendix.
A.1
Matrix-valued Positive Definite Functions on Euclidean Spaces: The Classes Φm
d,S
and Φm
d,I
This section provides a brief review about matrix-valued positive definite functions on the Euclidean
space E = Rd . Specifically, we expose some characterizations for the stationary and Euclidean
isotropic members of the class P m (Rd ).
We say that F ∈ P m (Rd ) is stationary if there exists a mapping ϕ̃ : Rd → Rm×m such that
F(x, y) = ϕ̃(x − y) = [ϕ̃ij (x − y)]m
i,j=1 ,
x, y ∈ Rd .
(A.2)
We call Φm
d,S the class of continuous mappings ϕ̃ such that F in (A.2) is positive definite. Cramér’s
Theorem (Cramer, 1940) establishes that ϕ̃ ∈ Φm
d,S if and only if it can be represented through
Z
ϕ̃(h) =
exp{ıh> ω}dΛ̃d (ω),
h ∈ Rd ,
(A.3)
Rd
√
where ı = −1 ∈ C and Λ̃d : Rd → Cm×m is a matrix-valued mapping, with increments being
Hermitian and positive definite matrices, and whose elements, Λ̃d,ij (·), for i, j = 1, . . . , m, are functions of bounded variation (see Wackernagel, 2003). In particular, the diagonal terms, Λ̃d,ii (ω), are
11
Table A.1: Summary of the notation used along the Appendices. Throughout, in the univariate
case (m = 1) we omit the super index: P(E), Φd,S , Φd,I , Ψd,I and Υd,k are used instead of P 1 (E),
Φ1d,S , Φ1d,I , Ψ1d,I and Υ1d,k , respectively.
Notation
P m (E)
Φm
d,S
Φm
d,I
Ψm
d,I
Υm
d,k
Description
Positive definite matrix-valued (m × m) functions on E × E.
Continuous, bounded and stationary elements of P m (Rd ).
Continuous, bounded and Euclidean isotropic elements of P m (Rd ).
Continuous, bounded and geodesically isotropic elements of P m (Sd ).
Elements in P m (Sd × Rk ) being, continuous, bounded, geodesically
isotropic in the spherical variable and stationary in the Euclidean variable.
real, non-decreasing and bounded, whereas the off-diagonal elements are generally complex-valued.
Cramer’s Theorem is the multivariate version of the celebrated Bochner’s Theorem (Bochner, 1955).
If the elements of Λ̃d (·) are absolutely continuous, then Equation (A.3) simplifies to
Z
ϕ̃(h) =
exp{ıh> ω}λ̃d (ω)dω,
h ∈ Rd ,
Rd
d
with λ̃d (ω) = [λ̃d,ij (ω)]m
i,j=1 being Hermitian and positive definite, for any ω ∈ R . The mapping
λ̃d (ω) is known as the matrix-valued spectral density and classical Fourier inversion yields
Z
1
exp{−ıh> ω}ϕ̃(h)dh,
ω ∈ Rd .
λ̃d (ω) =
(2π)d Rd
Finally, the following inequality between the elements of ϕ̃ is true
|ϕ̃ij (h)|2 ≤ ϕ̃ii (0)ϕ̃jj (0),
for all h ∈ Rd .
However, the maximum value of the mapping ϕ̃ij (h), with i 6= j, is not necessarily reached at
h = 0. In general, ϕ̃ij is not itself a scalar-valued positive definite function when i 6= j.
Consider an element F in P m (Rd ) and suppose that there exists a continuous and bounded mapping
ϕ : R+ → Rm×m such that
F(x, y) = ϕ(kx − yk),
x, y ∈ Rd .
Then, F is called stationary and Euclidean isotropic (or radial). We denote as Φm
d,I the class of
m
bounded, continuous, stationary and Euclidean isotropic mappings ϕ(·) = [ϕij (·)]i,j=1 .
When m = 1, characterization of the class Φd,I was provided through the celebrated paper by
Schoenberg (1938). Alonso-Malaver et al. (2015) characterize the class Φm
d,I through the continuous
members ϕ having representation
Z
ϕ(r) =
Ωd (rω)dΛd (ω),
r ≥ 0,
[0,∞)
where Λd : [0, ∞) → Rm×m is a matrix-valued mapping, with increments being positive definite
matrices, and elements Λd,ij (·) of bounded variation, for each i, j = 1, . . . , m. Here, the function
Ωd (·) is defined as
Ωd (z) = Γ(d/2)(z/2)−(d−2)/2 J(d−2)/2 (z),
z ≥ 0,
12
with Γ being the Gamma function and Jν the Bessel function of the first kind of degree ν (see
Abramowitz and Stegun, 1970). If the elements of Λd (·) are absolutely continuous, then we have
an associated spectral density λd : [0, ∞) → Rm×m as in the stationary case, which is called,
following Daley and Porcu (2014), a d-Schoenberg matrix.
The classes Φm
d,I are non-increasing in d, and the following inclusion relations are strict
∞
\
Φm
∞,I :=
m
m
Φm
d,I ⊂ · · · ⊂ Φ2,I ⊂ Φ1,I .
d=1
The elements ϕ in the class Φm
∞,I can be represented as
Z
exp(−r2 ω 2 )dΛ(ω),
ϕ(r) =
r ≥ 0,
[0,∞)
where Λ is a matrix-valued mapping with similar properties as Λd .
A.2
Matrix-valued Positive Definite Functions on Sd : The Class Ψm
d,I
In this section, we pay attention to matrix-valued positive definite functions on the unit sphere.
m d
Consider F = [Fij ]m
i,j=1 ∈ P (S ). We say that F is geodesically isotropic if there exists a bounded
and continuous mapping ψ : [0, π] → Rm×m such that
F(x, y) = ψ(θ(x, y)),
x, y ∈ Sd .
The continuous mappings ψ are the elements of the class Ψm
d,I and the following inclusion relations
are true:
∞
\
m
m
(A.4)
Ψm
=
Ψm
∞,I
d,I ⊂ · · · ⊂ Ψ2,I ⊂ Ψ1,I ,
d=1
where Ψm
isotropic
∞,I is the class of geodesically
P
2
∞
N
sphere S = {(xn )n∈N ∈ R : n∈N xn = 1}.
positive definite functions being valid on the Hilbert
The elements of the class Ψm
d,I have an explicit connection with Gegenbauer (or ultraspherical)
polynomials (Abramowitz and Stegun, 1970). Here, Gnλ denotes the λ-Gegenbauer polynomial of
degree n, which is defined implicitly through the expression
∞
X
1
=
rn Gnλ (cos θ),
(1 + r2 − 2r cos θ)λ
θ ∈ [0, π],
r ∈ (−1, 1).
n=0
1/2
In particular, Tn := Gn0 and Pn := Gn
of degree n.
are respectively the Chebyshev and Legendre polynomials
The following result (Hannan, 2009; Yaglom, 1987) offers a complete characterization of the classes
m
Ψm
d,I and Ψ∞,I , and corresponds to the multivariate version of Schoenberg’s Theorem (Schoenberg,
1942). Equalities and summability conditions for matrices must be understood in a componentwise
sense.
Theorem A.1. Let d and m be positive integers.
13
(1) The mapping ψ is a member of the class Ψm
d,I if and only if it admits the representation
ψ(θ) =
∞
X
(d−1)/2
Bn,d
n=0
Gn
(cos θ)
(d−1)/2
Gn
(1)
,
θ ∈ [0, π],
(A.5)
where {Bn,d }∞
n=0 is a sequence of symmetric, positive definite and summable matrices.
(2) The mapping ψ is a member of the class Ψm
∞,I if and only if it can be represented as
ψ(θ) =
∞
X
Bn (cos θ)n ,
θ ∈ [0, π],
n=0
where {Bn }∞
n=0 is a sequence of symmetric, positive definite and summable matrices.
Using orthogonality properties of Gegenbauer polynomials (Abramowitz and Stegun, 1970) and
through classical Fourier inversion we can prove that
Z
1 π
B0,1 =
ψ(θ)dθ,
π 0
Z
2 π
Bn,1 =
cos(nθ)ψ(θ)dθ,
for n ≥ 1,
(A.6)
π 0
whereas for d ≥ 2, we have
Bn,d
(2n + d − 1) [Γ((d − 1)/2)]2
=
Γ(d − 1)
23−d π
Z
π
Gn(d−1)/2 (cos θ)(sin θ)d−1 ψ(θ)dθ,
n ≥ 0,
(A.7)
0
where integration is taken componentwise. The matrices {Bn,d }∞
n=0 are called Schoenberg’s matrices. For the case m = 1, such result is reported by Gneiting (2013).
A.3
Matrix-valued Positive Definite Functions on Sd × Rk : The Class Υm
d,k
Let d, k and m be positive integers. We now focus on the class of matrix-valued positive definite
functions on Sd × Rk , being bounded, continuous, geodesically isotropic in the spherical component
and stationary in the Euclidean one. The case k = 1 is particularly important, since P m (Sd × R)
can be interpreted as the class of admissible space-time covariances for multivariate Gaussian RFs,
with spatial locations on the unit sphere.
Consider F ∈ P m (Sd × Rk ) and suppose that there exists a bounded and continuous mapping
C : [0, π] × Rk → Rm×m such that
F((x, t), (y, s)) = C(θ(x, y), t − s),
x, y ∈ Sd , t, s ∈ Rk .
(A.8)
Such mappings C are the elements of the class Υm
d,k . These classes are non-increasing in d and we
have the inclusions
∞
\
m
m
m
Υ∞,k :=
Υm
d,k ⊂ · · · ⊂ Υ2,k ⊂ Υ1,k .
d=1
Ma (2016) proposes the generalization of Theorem A.1 to the space-time case. Theorem A.2 below
m
offers a complete characterization of the class Υm
d,k and Υ∞,k , for any m ≥ 1. Again, equalities and
summability conditions must be understood in a componentwise sense.
14
Theorem A.2. Let d, k and m be positive integers and C : [0, π] × Rk → Rm×m a continuous
matrix-valued mapping, with Cii (0, 0) < ∞, for all i = 1, . . . , m.
(1) The mapping C belongs to the class Υm
d,k if and only if
∞
X
C(θ, u) =
(d−1)/2
ϕ̃n,d (u)
Gn
(cos θ)
(d−1)/2
Gn
(1)
n=0
(θ, u) ∈ [0, π] × Rk ,
,
(A.9)
m
with {ϕ̃n,d (·)}∞
n=0 being a sequence of members of the class Φk,S , with the additional requirement that the sequence of matrices {ϕ̃n,d (0)}∞
n=0 is summable.
(2) The mapping C belongs to the class Υm
∞,k if and only if
C(θ, u) =
∞
X
ϕ̃n (u)(cos θ)n ,
(θ, u) ∈ [0, π] × Rk ,
(A.10)
n=0
m
with {ϕ̃n (·)}∞
n=0 being a sequence of members of the class Φk,S , with the additional require∞
ment that the sequence of matrices {ϕ̃n (0)}n=0 is summable.
Again, using orthogonality arguments, we have
Z
1 π
ϕ̃0,1 (u) =
C(θ, u)dθ,
π 0
Z
2 π
ϕ̃n,1 (u) =
cos(nθ)C(θ, u)dθ,
π 0
for n ≥ 1,
(A.11)
whereas for d ≥ 2,
ϕ̃n,d (u) =
B
(2n + d − 1) [Γ((d − 1)/2)]2
Γ(d − 1)
23−d π
Z
π
Gn(d−1)/2 (cos θ)(sin θ)d−1 C(θ, u)dθ,
n ≥ 0. (A.12)
0
Proof of Theorem 3.1
In order to illustrate the results following subsequently, a technical Lemma will be useful. We do
not provide a proof because it is obtained following the same arguments as in Porcu and Zastavnyi
(2011).
Lemma B.1. Let m, d and k be strictly positive integers. Let (X, B, µ) be a measure space, for
X ⊂ R and B being the Borel sigma algebra. Let ψ : [0, π] × X → R and ϕ : [0, ∞) × X → R be
continuous mappings satisfying
1. ψ(·, ξ) ∈ Ψm
d,I a.e. ξ ∈ X;
2. ψ(θ, ·) ∈ L1 (X, B, µ) for any θ ∈ [0, π];
3. ϕ(·, ξ) ∈ Φm
k,I a.e. ξ ∈ X;
4. ϕ(u, ·) ∈ L1 (X, B, µ) for any u ∈ [0, ∞).
Let C : [0, π] × [0, ∞) → Rm×m be the mapping defined through
Z
C(θ, u) =
ψ(θ, ξ)ϕ(u, ξ)µ(dξ),
(θ, u) ∈ [0, π] × [0, ∞).
X
Then, C is continuous and bounded. Further, C belongs to the class Υm
d,k .
15
(B.1)
Of course, Lemma B.1 is a particular case of the scale mixtures introduced in Section 2.
Proof of Theorem 3.1 Let (X, B, µ) as in Lemma B.1 and consider X = R+ with µ the Lebesgue
measure. We offer a proof of the constructive type. Let us define the function ψ(θ, ξ) with members
ψij (·, ·) defined through
θ n+1
ψij (θ, ξ) = σii σjj ρij 1 −
,
ξcij +
(θ, ξ) ∈ [0, π] × X,
i, j = 1, . . . , m,
where, as asserted, the constants σii , ρij and cij are determined according to condition (3.3). Let
us now define the mapping (u, ξ) 7→ ϕ(u, ξ) := ξ n+1 (1 − ξf (u))`+ , with (u, ξ) ∈ [0, ∞) × X. It can
be verified that both ψ and ϕ satisfy requirements 1–4 in Lemma B.1. In particular, Condition 1
yields thanks to Lemmas 3 and 4 in Gneiting (2013), as well as Theorem 1 in Daley et al. (2015).
Also, arguments in Porcu et al. (2016) show that Condition 3 holds for any ` ≥ 1. We can now
apply Lemma B.1, so that we have that
Z
Ci,j,n,` (θ, u) :=
ψ(θ, ξ)ϕ(u, ξ)dξ,
[0, π] × [0, ∞)
X
is a member of the class Υm
2n+1,1 for any ` ≥ 1. Pointwise application of an elegant scale mixture
argument as in Proposition 1 of Porcu et al. (2016) shows that
Ci,j,n,` (θ, u) = B(n + 2, ` + 1)
ρij σii σjj
f (u)n+2
1−
θf (u)
cij
n+`+1
,
(θ, u) ∈ [0, π] × [0, ∞),
(B.2)
+
where B denotes the Beta function (Abramowitz and Stegun, 1970). We now omit the factor
B(n + 2, ` + 1) since it does not affect positive definiteness. Now, standard convergence arguments
show that
ρij σii σjj
θf (u)
exp −
, (θ, u) ∈ [0, π] × [0, ∞),
lim Ci,j,n,` (θ/`, u) =
`→∞
f (u)n+2
cij
with the convergence being uniform in any compact set. The proof is then completed in view of
Bernstein’s theorem (Feller, 1966).
C
Proof of Theorem 3.2
Before we state the proof of Theorem 3.2, we need to introduce three auxiliary lemmas.
Lemma C.1. Let C : [0, π] × Rk → R be a continuous, bounded and integrable function, for some
positive integer k. Then C ∈ Υd,k , for d ≥ 1, if and only if the mapping ψω : [0, π] → R defined as
Z
1
exp{−ıω > v}C(θ, v)dv,
θ ∈ [0, π],
(C.1)
ψω (θ) =
(2π)k Rk
belongs to the class Ψd,I , for all ω ∈ Rk .
We do not report the proof of Lemma C.1 since the arguments are the same as in the proof of
Lemma C.2 below. Note that this lemma is a spherical version of the result given by Cressie and
Huang (1999).
16
Lemma C.2. Let C : [0, π] × Rl × Rk → R be a continuous, bounded and integrable function,
for some positive integers l and k. Then, C ∈ Υd,k+l , with d ≥ 1, if and only if the mapping
Cω : [0, π] × Rl → R defined as
Z
1
Cω (θ, u) =
exp{−ıω > v}C(θ, u, v)dv,
(θ, u) ∈ [0, π] × Rl ,
(C.2)
(2π)k Rk
belongs to the class Υd,l , for all ω ∈ Rk .
Proof of Lemma C.2 Suppose that C ∈ Υd,k+l , then the characterization of Berg and Porcu
(2017) implies that
∞
(d−1)/2
X
Gn
(cos θ)
C(θ, u, v) =
,
(C.3)
ϕ̃n,d (u, v)
(d−1)/2
Gn
(1)
n=0
P∞
where {ϕ̃n,d (·, ·)}∞
n=0 is a sequence of functions in Φk+l,S , with
n=0 ϕ̃n,d (0, 0) < ∞. Therefore,
!
Z
∞
(d−1)/2
X
1
G
(cos
θ)
n
dv
exp{−ıω > v}
Cω (θ, u) =
ϕ̃n,d (u, v)
(d−1)/2
(2π)k Rk
Gn
(1)
n=0
(d−1)/2
Z
∞
X
1
Gn
(cos θ)
>
=
,
exp{−ıω v}ϕ̃n,d (u, v)dv
k
(d−1)/2
(2π) Rk
Gn
(1)
n=0
where the last step is justified by dominated convergence. We need to prove that for each fixed
ω ∈ Rk , the sequence of functions
Z
1
u 7→ λ̃n,d (u; ω) :=
exp{−ıω > v}ϕ̃n,d (u, v)dv,
n ≥ 0,
(2π)k Rk
belongs to the class Φl,S , a.e. ω ∈ Rk . In fact, we have that
Z
Z Z
1
1
>
exp{−ıτ u}λ̃n,d (u; ω)du =
exp{−ıτ > u − ıω > v}ϕ̃n,d (u, v)dvdu. (C.4)
(2π)l Rl
(2π)k+l Rl Rk
Since ϕ̃n,d (·, ·) belongs to Φk+l,S , Bochner’s Theorem implies that the right side in Equation (C.4)
is non-negative everywhere.
P∞ This implies that λ̃n,d (·; ω) kbelongs to the class Φl,S . Also, direct
inspection shows that n=0 λ̃n,d (0; ω) < ∞, for all ω ∈ R . The necessary part is completed.
On the other hand, suppose that for each ω ∈ Rk the function Cω (θ, u) belongs to the class Υd,l ,
k
then there exists a sequence of mappings {λ̃n,d (·; ω)}∞
n=0 in Φl,S for each ω ∈ R , such that
Cω (θ, u) =
∞
X
(d−1)/2
λ̃n,d (u; ω)
n=0
Gn
(cos θ)
(d−1)/2
Gn
(1)
.
Thus,
Z
C(θ, u, v) =
exp{ıω > v}
∞
X
(d−1)/2
(cos θ)
!
dω
(d−1)/2
Gn
(1)
(d−1)/2
∞ Z
X
Gn
(cos θ)
>
=
exp{ıω v}λ̃n,d (u; ω)dω
.
(d−1)/2
Rk
Gn
(1)
n=0
Rk
λ̃n,d (u; ω)
Gn
n=0
We conclude the proof by invoking again Bochner’s Theorem and the result of Berg and Porcu
(2017).
17
Lemma C.3. Let d and k be two positive integers. Consider g and f be completely monotone and
Bernstein functions, respectively. Then,
kvk2
1
g
K(θ, v) =
,
(θ, v) ∈ [0, π] × Rk ,
(C.5)
k/2
f
(θ)|
[0,π]
f (θ)|[0,π]
belongs to the class Υd,k , for any positive integer d.
Proof of Lemma C.3 By Lemma C.1, we must show that ψω , defined through Equation (C.1),
belongs to the class Ψd,I , for all ω ∈ Rk . In fact, we can assume that C is integrable, since the
general case is obtained with the same arguments given by Gneiting (2002). Bernstein’s Theorem
establishes that g can be represented as the Laplace transform of a bounded measure G, then
Z
Z
1
rkvk2
>
exp{−ıω v}
exp −
dG(r)dv
ψω (θ) =
k/2
f (θ)|[0,π]
Rk
[0,∞)
f (θ)|[0,π]
Z
kωk2
k/2
exp −
= π
f (θ)|[0,π] dG̃(r),
4r
[0,∞)
where the last equality follows from Fubini’s Theorem and dG(r) = rk/2 dG̃(r). In addition, the
composition between a negative exponential and a Bernstein function is completely monotone on
the real line (Feller, 1966). Then, for any ω and r, the mapping θ 7→ exp{−kωk2 f (θ)|[0,π] /(4r)}
is the restriction of a completely monotone function to the interval [0, π]. Theorem 7 in Gneiting
(2013) implies that such mapping, and thus ψω , belongs to the class Ψd,I , for any d ∈ N and
ω ∈ Rk .
Proof of Theorem 3.2 By Lemma C.2, we must show that (C.2) belongs to the class Υd,l , for all
d ∈ N. In fact, assuming again that C is integrable and invoking Bernstein’s Theorem we have
Z
Z
2
1
rkvk
i
h
Cω (θ, u) =
exp{−ıω > v}
exp
dG(r)dv
−
n h
io
k/2
kuk2
f
kuk2
Rk
[0,∞)
1 f2 (θ)|
{f2 (θ)|[0,π] }l/2 f1 f2 (θ)|
[0,π]
[0,π]
Z
2
2
kuk
1
kωk
f1
= π k/2
exp −
dG̃(r),
l/2
4r
f2 (θ)|[0,π]
{f2 (θ)|[0,π] }
[0,∞)
where the last equality follows from Fubini’s Theorem and dG(r) = rk/2 dG̃(r). In addition, for any
ω, the mapping
Z
kωk2
f1 (·) dG̃(r)
gω (·) :=
exp −
4r
[0,∞)
is completely monotone (Feller, 1966). Therefore,
Cω (θ, u) = π
k/2
1
gω
{f2 (θ)|[0,π] }l/2
kuk2
f2 (θ)|[0,π]
,
and by Lemma C.3, we have that Cω ∈ Υd,l , for all d ∈ N.
References
Abramowitz, M. and Stegun, I. A., editors (1970). Handbook of Mathematical Functions. Dover, New York.
18
Alegrı́a, A. and Porcu, E. (2017). The dimple problem related to space–time modeling under the Lagrangian framework. Journal of Multivariate Analysis, 162:110 – 121.
Alonso-Malaver, C., Porcu, E., and Giraldo, R. (2015). Multivariate and multiradial Schoenberg measures with their
dimension walks. Journal of Multivariate Analysis, 133:251 – 265.
Apanasovich, T. V. and Genton, M. G. (2010). Cross-covariance functions for multivariate random fields based on
latent dimensions. Biometrika, 97:15 –30.
Apanasovich, T. V., Genton, M. G., and Sun, Y. (2012). A valid Matérn class of cross-covariance functions for
multivariate random fields with any number of components. Journal of the American Statistical Association,
107(497):180–193.
Banerjee, S. (2005). On geodetic distance computations in spatial modeling. Biometrics, 61(2):617–625.
Berg, C. and Porcu, E. (2017). From Schoenberg coefficients to Schoenberg functions. Constructive Approximation,
45(2):217–241.
Bevilacqua, M., Alegria, A., Velandia, D., and Porcu, E. (2016). Composite likelihood inference for multivariate
Gaussian random fields. Journal of Agricultural, Biological, and Environmental Statistics, 21(3):448–469.
Bochner, S. (1955). Harmonic Analysis and the Theory of Probability. California Monographs in mathematical
sciences. University of California Press.
Bourotte, M., Allard, D., and Porcu, E. (2016). A flexible class of non-separable cross-covariance functions for
multivariate space–time data. Spatial Statistics, 18:125–146.
Cramer, H. (1940). On the theory of stationary random processes. Annals of Mathematics, 41(1):215–230.
Cressie, N. (1993). Statistics for Spatial Data. Wiley, New York, revised edition.
Cressie, N. and Huang, H. (1999). Classes of nonseparable, spatiotemporal stationary covariance functions. Journal
of the American Statistical Association, 94:1330–1340.
Daley, D. and Porcu, E. (2014). Dimension walks and Schoenberg spectral measures. Proceedings of the American
Mathematical Society, 142(5):1813–1824.
Daley, D., Porcu, E., and Bevilacqua, M. (2015). Classes of compactly supported covariance functions for multivariate
random fields. Stochastic Environmental Research and Risk Assessment, 29(4):1249–1263.
Feller, W. (1966). An Introduction to Probability Theory and its Applications. Number v. 2 in Wiley Mathematical
Statistics Series. Wiley.
Gent, P. R., Danabasoglu, G., Donner, L. J., Holland, M. M., Hunke, E. C., Jayne, S. R., Lawrence, D. M., Neale,
R. B., Rasch, P. J., Vertenstein, M., Worley, P. H., Yang, Z.-L., and Zhang, M. (2011). The community climate
system model version 4. Journal of Climate, 24(19):4973–4991.
Genton, M. G. and Kleiber, W. (2015). Cross-covariance functions for multivariate geostatistics. Statistical Science,
30(2):147–163.
Gneiting, T. (2002). Stationary covariance functions for space-time data. Journal of the American Statistical Association, 97(458):590–600.
Gneiting, T. (2013). Strictly and non-strictly positive definite functions on spheres. Bernoulli, 19(4):1327–1349.
Gneiting, T., Genton, M. G., and Guttorp, P. (2007). Geostatistical space-time models, stationarity, separability
and full symmetry. In Finkenstadt, B., Held, L., and Isham, V., editors, Statistical Methods for Spatio-Temporal
Systems, pages 151–175. Chapman & Hall/CRC, Boca Raton: FL.
Gneiting, T., Kleiber, W., and Schlather, M. (2010). Matérn Cross-Covariance functions for multivariate random
fields. Journal of the American Statistical Association, 105:1167–1177.
Hannan, E. (2009). Multiple Time Series. Wiley Series in Probability and Statistics. Wiley.
Hansen, L. V., Thorarinsdottir, T. L., Gneiting, T., and Richards, D. (2011). Lévy particles: Modelling and simulating
star-shaped random sets. CSGB Research Reports, Department of Mathematics, Aarhus University.
19
Huang, C., Zhang, H., and Robeson, S. (2011). On the validity of commonly used covariance and variogram functions
on the sphere. Mathematical Geosciences, 43:721–733.
Ma, C. (2016). Time varying isotropic vector random fields on spheres. Journal of Theoretical Probability.
Marinucci, D. and Peccati, G. (2011). Random Fields on the Sphere: Representation, Limit Theorems and Cosmological Applications. Cambridge University Press.
Porcu, E., Bevilacqua, M., and Genton, M. G. (2016). Spatio-temporal covariance and cross-covariance functions of
the great circle distance on a sphere. Journal of the American Statistical Association, 111(514):888–898.
Porcu, E., Gregori, P., and Mateu, J. (2006). Nonseparable stationary anisotropic space–time covariance functions.
Stochastic Environmental Research and Risk Assessment, 21(2):113–122.
Porcu, E. and Schilling, R. L. (2011). From Schoenberg to Pick-Nevanlinna: Towards a complete picture of the
variogram class. Bernoulli, 17(1):441–455.
Porcu, E. and Zastavnyi, V. (2011). Characterization theorems for some classes of covariance functions associated to
vector valued random fields. Journal of Multivariate Analysis, 102(9):1293 – 1301.
Schoenberg, I. J. (1938). Metric spaces and completely monotone functions. Annals of Mathematics, 39(4):811–841.
Schoenberg, I. J. (1942). Positive definite functions on spheres. Duke Math. J., 9(1):96–108.
Wackernagel, H. (2003). Multivariate Geostatistics: An Introduction with Applications. Springer, New York, 3rd
edition.
Yaglom, A. M. (1987). Correlation Theory of Stationary and Related Random Functions. Volume I: Basic Results.
Springer, New York.
Zhang, H. and Wang, Y. (2010). Kriging and cross-validation for massive spatial data. Environmetrics, 21(3-4):290–
304.
Ziegel, J. (2014). Convolution roots and differentiability of isotropic positive definite functions on spheres. Proceedings
of the American Mathematical Society, 142(6):2063–2077.
20
| 10math.ST
|
Adaptive local surface refinement based on LR NURBS and its
application to contact
Christopher Zimmermann, Roger A. Sauer1
Aachen Institute for Advanced Study in Computational Engineering Science (AICES),
RWTH Aachen University, Templergraben 55, 52062 Aachen, Germany
Published2 in Computational Mechanics, DOI: 10.1007/s00466-017-1455-7
Submitted on 18. April 2017, Revised on 13. July 2017, Accepted on 21. July 2017
arXiv:1701.08742v3 [cs.CE] 18 Sep 2017
Abstract
A novel adaptive local surface refinement technique based on Locally Refined Non-Uniform Rational
B-Splines (LR NURBS) is presented. LR NURBS can model complex geometries exactly and are
the rational extension of LR B-splines. The local representation of the parameter space overcomes
the drawback of non-existent local refinement in standard NURBS-based isogeometric analysis. For a
convenient embedding into general finite element code, the Bézier extraction operator for LR NURBS
is formulated. An automatic remeshing technique is presented that allows adaptive local refinement
and coarsening of LR NURBS.
In this work, LR NURBS are applied to contact computations of 3D solids and membranes. For solids,
LR NURBS-enriched finite elements are used to discretize the contact surfaces with LR NURBS finite
elements, while the rest of the body is discretized by linear Lagrange finite elements. For membranes,
the entire surface is discretized by LR NURBS. Various numerical examples are shown, and they
demonstrate the benefit of using LR NURBS: Compared to uniform refinement, LR NURBS can
achieve high accuracy at lower computational cost.
Keywords: Adaptive local refinement, computational contact mechanics, isogeometric analysis, LR
B-splines, nonlinear finite element analysis
1
Introduction
A wide range of engineering applications, especially those that are governed by local surface effects,
necessitate an accurate surface description. An example are contact problems. Due to their strong
nonlinear behavior, those are usually only solvable by numerical methods. For this, isogeometric
finite element discretizations are advantageous, since they offer at least C 1 -continuity across element
boundaries. Standard NURBS based discretizations lack local refinement, but their ability to discretize complex geometries exactly makes them powerful.
Isogeometric Analysis (IGA) was developed by Hughes et al. (2005) and the work of Cottrell et al.
(2009) summarizes the concept of IGA. Borden et al. (2011) introduced the Bézier extraction operator, which allows a suitable embedding of isogeometric analysis into existing finite element codes. The
work of Scott et al. (2011) introduces the extension of the Bézier extraction operator to T-splines.
This offers a local representation of the Bézier extraction operator. B-splines and NURBS are the
most widespread element types in IGA. Those element types are only globally refinable. Hierarchical
B-splines were introduced by Forsey and Bartels (1988) and allow local refinement. An analysissuitable approach that bases on hierarchichal NURBS discretization can be found in, e.g. Schillinger
et al. (2012). The work of Sederberg et al. (2003) introduces T-splines as an approach to discretize
1
2
corresponding author, email: [email protected]
This pdf is the personal version of an article whose final publication is available at http://link.springer.com
1
surfaces more efficiently than hierarchical B-splines. Scott et al. (2012) developed a local refinement
approach based on analysis-suitable T-splines.
Locally Refined B-splines (LR B-splines) are developed as a new approach to allow local refinement in
IGA. LR B-splines were introduced by Dokken et al. (2013) and further advanced by Bressan (2013)
and Johannessen et al. (2014). Johannessen et al. (2015b) studied the similarities and differences of
LR B-spline, classical hierarchical and truncated hierarchical discretizations. The classical hierarchical
approach leads to a much denser sparsity pattern of matrices than LR B-spline or truncated hierarchical discretizations. Whether an LR B-spline or a truncated hierarchical discretization is beneficial
is depending on the ratio of the locally refined domain w.r.t. the entire mesh. In the IGA community, T-splines are more widespread than LR splines and commercial software for the automatic mesh
generation is available, e.g. Rhinoceros (McNeel, 2012) with a T-splines plug-in (Autodesk, 2015). In
contrast to T-splines, the refinement by LR B-splines is directly performed in the parameter domain,
which reflects the piecewise polynomial structure. This is more convenient than the multiple vertex
grids of T-splines. With an LR B-spline description new possibilites are achieved on how the computational mesh can be generated. This makes them interesting in many fields of engineering. In the
works of Nørtoft and Dokken (2014) and Johannessen et al. (2015a) the LR B-splines are successfully
applied to the computation of Navier-Stokes flows. To the best of the authors’ knowledge rational
basis functions in the context of LR splines have not been used so far. In this work, LR B-splines are
extended to LR NURBS to combine the advantages of local refinement and the ability of modeling
complex geometries exactly.
The numerical examples presented here consider 3D solids and membranes using a classical finite
element formulation for solids (Wriggers, 2008) and the membrane finite element formulation of Sauer
et al. (2014). This work is based on computational contact mechanics (Laursen, 2002; Wriggers,
2006) using isogeometric finite elements (De Lorenzis et al., 2011, 2012; Lu, 2011; Temizer et al.,
2011) and an unbiased contact algorithm (Sauer and De Lorenzis, 2013, 2015). In the literature
several approaches can be found that account for local refinement and adaptive meshing. Lee and
Oden (1994) present an adaptive method for h-p refined meshes in the context of frictional contact
and non-isogeometric finite elements. The work of Hager et al. (2012) presents a non-conforming
domain decomposition method. The domain decomposition consists of a global coarse mesh and an
overlapping fine mesh for the contact computation. In the context of IGA and frictionless contact a
recent work based on analysis-suitable T-splines is presented in Dimitri et al. (2014). The work by
Dimitri and Zavarise (2017) is based on analysis-suitable T-splines and investigates frictional contact
and mixed mode debonding problems. In the work of Temizer and Hesch (2016) hierarchical NURBS
are used to allow local refinement. In this work, LR NURBS are applied to frictionless and frictional
sliding contact.
LR B-splines have been successfully used in 2D, since they are linearly independent in 2D. In 3D, the
linear independence for arbitrary, locally refined meshes has not been proven yet and requires further
study. To overcome this issue, this work uses 2D LR NURBS and combines them with an enrichment
technique (Corbett and Sauer, 2014, 2015) along the third direction. This combines the high accuracy, which is achieved by isogeometric analysis and the efficiency of standard finite elements. This
formulation is adapted for LR NURBS-enrichment to gain the possibility of local refinement. In this
work we present a novel technique for adaptive local surface refinement and coarsening that
• considers the extension of LR B-spline finite elements to LR NURBS finite elements
• uses a local formulation of the Bézier extraction operator
• is automatically controlled by a proposed refinement criterion
• is extended to LR NURBS-enriched finite elements for 3D solids
• is applied to frictionless and frictional sliding contact.
2
In Sec. 2.1 a brief summary of the nonlinear finite element formulation for membranes and 3D solids
is given. A short introduction to computational contact mechanics follows in Sec. 2.2. The fundamentals of LR B-splines are presented in Sec. 2.3. The geometric modeling using LR NURBS and the
formulation of the Bézier extraction operator is presented in Sec. 3. The technique of adaptive local
refinement and coarsening using LR NURBS discretizations is presented in Sec. 4. Numerical results
are shown in Sec. 5. The performance of LR NURBS within the proposed local refinement technique
is compared with uniform meshes using NURBS discretizations. A conclusion is given in Sec. 6.
2
Preliminaries
This section presents an overview of the fundamental mathematical formulations that are used within
this work. A general finite element formulation for membranes and 3D solids in the context of contact
computations is discussed in Sec. 2.1 and Sec. 2.2. The fundamentals of LR B-splines are given in
Sec. 2.3.
2.1
Finite element formulation
The finite element formulation for nonlinear membranes and 3D solids is similar. Their governing
equations (e.g. equilibrium) are different but obtaining their discretized weak form follows the same
approach. The quasi-static weak form of a system of two deformable bodies in contact is described
by
II h
i
X
(2.1)
δΠkint − δΠkc − δΠkext = 0 ,
k=I
I, II)3 .
considering two objects (k =
This describes the equilibrium between the internal virtual work
δΠkint , the virtual contact work δΠkc and the external virtual work. In our examples δΠkext = 0
is considered. The computational formulation for nonlinear membranes in R3 in the framework of
curvilinear coordinates is taken from Sauer et al. (2014). A general constitutive setting and finite
element formulation for solids in R3 can be found in Wriggers (2008). The internal virtual work for
membranes Sk in R3 and solids Bk in R3 is
Z
for Sk ,
δϕk,α · σkαβ akβ dak , ∀ δϕk ∈ Vk ,
S
k
k
δΠint = Z
(2.2)
grad(δϕk ) : σ k dvk , ∀ δϕk ∈ Vk , for Bk .
Bk
Here, σkαβ are the contra-variant components of the Cauchy stress tensor σ with α = 1, 2 and β = 1, 2.
The parametric derivative of the virtual displacement field is denoted by δϕk,α and akα are the co-variant
tangent vectors. The virtual contact work for the contact surfaces Γkc is expressed as
Z
k
δΠc =
δϕk · tkc dak , ∀ δϕk ∈ Vk ,
(2.3)
Γkc
where tkc denotes the contact traction on the surface Γkc ⊂ Sk . The computation of the contact
integrals is performed using the two-half pass algorithm of Sauer and De Lorenzis (2013). This
algorithm evaluates Eq. (2.3) on each body equivalently. System (2.1) is solved for the unknown
deformation field ϕk ∈ Vk , with Vk as a suitable space of kinematically admissible variations. The
k
membrane and the contact surface are discretized
S k into a kset ofk membrane
S k elements Γe and a set of
k
h
h
contact elements Γc e such that Sk ≈ Sk = e Γe and Γc ≈ Γc = e Γc e , respectively. Here, the
3
the index k is used as super- and subscript but the position has no special meaning
3
element domains Γke and Γkc e are considered equal and have the same finite element discretization.
k
Similarly, the solids are discretized S
into a set of volume elements
S k Ωe and a set of surface contact
k
h
k
k
k
h
elements Γc e such that Bk ≈ Bk = e Ωe and Γc ≈ Γc = e Γc e , respectively. The superscript h
denotes the approximated discrete setting. With a set of ne basis functions N = [N1 I, ..., Nne I] with
identity matrix I in R3 and discrete points xe , a point x ∈ Bk is interpolated within each element as
x ≈ xh = N xe .
(2.4)
The same holds for the reference configuration X ≈ X h , the displacement field u ≈ uh and the virtual
displacement field δϕ ≈ v h . In here we drop the index k for convenience. In standard finite elements
the discrete points x are known as nodes, while in IGA they are called control points. For standard
finite element basis functions Lagrange polynomials are used, while for IGA the basis functions will
be discussed in Sec. 2.3. As a consequence of the above definitions, the dicretized weak form can be
written as
vT [fint − fc ] = 0, ∀ v ∈ V h ,
(2.5)
with the internal forces
fint
Z
αβ
NT
N,β dak ,
,α σ
Sk
= Z
N,x σ dvk ,
for Sk ,
(2.6)
for Bk ,
Bk
and the contact forces
Z
fc =
NT tc dak .
(2.7)
Γkc
A brief overview of computational contact mechanics is given in the next section.
2.2
Computational contact formulation
This section gives a brief summary of computational contact mechanics. For more detailed information, the monograph by Wriggers (2006) is recommended. Here, the penalty method is used in order
to enforce contact. The gap vector g between two points xI and xII on the surfaces SI and SII is given
by
g = xI − xII .
(2.8)
The closest projection point xp of xI is obtained by orthogonal projection of xI onto SII . The contact
traction is expressed as tc = tn + tt , i.e. it is decomposed into its normal and tangential components.
According to the penalty formulation, the normal traction
(
−εn gn np , gn < 0 ,
tn (xk ) =
(2.9)
0,
gn ≥ 0 ,
is applied at each contact point. The traction is proportional to the normal gap gn = (xI − xp ) · np ,
the surface outward normal np at xp and the constant penalty parameter εn . k is the index for
the body under consideration. The tangential contact traction tt during frictional sliding is given by
Coulomb’s law
ġ
tt = −µ p t .
(2.10)
||ġ t ||
Here, we find the friction coefficient µ, the contact pressure p = ||tn || and the relative tangential
sliding velocity ġ t . The tangential contact slip
g t = ∆g e + g s ,
4
(2.11)
consists of the reversible ∆g e and irreversible g s part. The slip criterion
fs = ||tt || − µ p ,
(2.12)
is used to distinguish between sticking and slipping. Sliding occurs for fs = 0 and sticking for fs < 0.
The examples presented below deal with frictionless and frictional contact. For frictionless contact,
the tangential traction tt vanishes and the contact traction is given by tc = tn . The implementation
and formulation of frictional contact is taken from Sauer and De Lorenzis (2015).
2.3
LR B-spline fundamentals
The core idea of LR B-splines is to disband the tensor-product mesh structure of classical B-splines in
order to obtain a local representation of the parameter domain. As introduced in the B-spline theory,
a knot vector Ξ of size n + p + 1 generates n linearly independent basis functions of order4 p. The
local representation of the parameter space and the geometric discretization are obtained by splitting
the global knot vector Ξ = [ξ1 , ..., ξn+p+1 ] into local knot vectors Ξi = [ξi , ..., ξi+p+1 ] each constructing
a single basis function. With the local knot vector representation, the domain of the basis function
is minimal, i.e. the basis has minimal support. The basis functions are defined by the Cox-de Boor
recursion formula, which depends on the entries of the local knot vector Ξi and desired polynomial
order p, see Cox (1971) and De Boor (1972). For p = 0
(
1 , if ξi ≤ ξ < ξi+1 ,
Ni0 (ξ) =
(2.13)
0 , otherwise ,
and for p > 0
Nip (ξ) =
ξi+p+1 − ξ
ξ − ξi
p−1
Nip−1 (ξi , ..., ξp+i ) +
Ni+1
(ξi+1 , ..., ξp+i+1 ) .
ξi+p − ξi
ξi+p+1 − ξi+1
(2.14)
The LR B-spline basis Bipjq (ξ, η) of order p and q in 2D is defined as a separable function B : R2 → R
Bipjq (ξ, η) = Nip (ξ) Mjq (η) .
(2.15)
Nip (ξ) and Mjq (η) are the shape functions in each parametric direction. The properties of LR Bsplines follow directly from standard isogeometric analysis. The basis is non-negative and forms a
partition of unity, and the resulting geometry lies within the convex hull of the control points. The
continuity across the element boundaries is C p−mv , i.e. it is defined by the polynomial order p and
the multiplicity of the knot vector entry mv . The LR B-spline surface is then constructed by a set of
control points xi j and the local basis functions Bipjq (ξ, η), as
x(ξ, η) =
n X
m
X
Bipjq (ξ, η) xi j ,
(2.16)
i=1 j=1
with n, m as the number of single basis functions in each parametric dimension. To ensure that LR Bsplines keep the partition of unity property during local refinement the basis functions are multiplied
by a scaling factor γi j ∈ (0, 1]
x(ξ, η) =
n X
m
X
Bipjq (ξ, η) xi j γi j .
(2.17)
i=1 j=1
With the above formulation, the global B-spline representation is split into a local representation
consisting of a set of locally refinable B-splines. In the following we use the terminology LR B-spline
for both locally refinable and locally refined B-splines. It is important to note that the global B-splines
do not exist anymore. Global B-splines do not allow local refinement. The ability is only given by
the LR B-splines. Locally refined meshes and the local refinement procedure are discussed in the
following sections.
4
also known as degree
5
2.3.1
Locally refined meshes
The local representation of the parametric domain leads to new possibilities on how the finite element
mesh can be constructed. The ability of local refinement is the major advantage in comparison to
classical tensor meshes. The LR mesh is the result of a series of meshline insertions into an initial
tensor mesh. Fig. 2.1a shows an example of an LR mesh. The meshlines never stop in the center of an
element (knot span). As a knot vector of an LR B-spline basis function has size of p + 2 knot vector
entries, the meshlines cross at least p + 2 knots. A meshline insertion can be either a new meshline,
an elongation of an existing one, a joining of two meshlines or the increase of the multiplicity of a
meshline. Increasing the multiplicity decreases the continuity of the LR B-spline. In the bi-variate
case, horizontal and vertical meshlines can be inserted. Any type of meshline insertion implies that
an LR B-spline loses the property of minimal support. An LR B-spline has minimal support if the
support domain of the basis functions is not fully crossed by any other meshline. The basis function
domain of the LR B-spline in Figure 2.1b, marked in gray, has minimal support because its support
domain is not fully crossed by any meshline insertion completely. The LR B-spline basis marked in the
Fig. 2.1c does not have minimal support. Its basis function domain is crossed by the vertical meshline,
which is spanned by η = [0.2, 0.8] at ξ = 0.5. If an LR B-spline loses the property of minimal support
refinement is performed as discussed in the next section.
0
0.2 0.4 0.6 0.8
(a) LR mesh
1
0
0.2 0.4 0.6 0.8
(b) Minimal support
1
0
0.2 0.4 0.6 0.8
1
(c) No minimal support
Figure 2.1: LR meshes contructed by meshline insertion (marked by thick lines) into a tensor mesh of
a bivariate parameter domain. (a) Example of an LR mesh. (b) Support domain of the LR B-spline
basis, marked in gray, which has minimal support. (c) Support domain of the LR B-spline basis,
marked in gray, that does not have minimal support. The meshline extension, which is spanned by
η = [0.2, 0.8] at ξ = 0.5 fully crosses the support domain.
2.3.2
Procedure of local refinement
Local refinement is perfomed if an LR B-spline loses the property of minimal support due to meshline
extensions. In general, the procedure of local refinement is realized by single knot insertion. From the
classical theory it is known that this enriches the basis while the geometry remains unchanged. The
insertion of a single knot ξˆ into the knot span ξ i = [ξi−1 , ξi ] of a local knot vector Ξ of size p + 2 leads
to a knot vector of size p + 3.h Consequently, two locally
i refined B-splines are generated by splitting
ˆ
the enlarged knot vector Ξ = ξ1 , ..., ξi−1 , ξ, ξi , ..., ξp+2 into the local knot vectors of size p + 2
h
i
ˆ ξi , ..., ξp+1 ,
Ξ1 = ξ1 , ..., ξi−1 , ξ,
h
i
(2.18)
ˆ ξi , ..., ξp+2 .
Ξ2 = ξ2 , ..., ξi−1 , ξ,
6
The relation for an LR B-spline basis in one parametric direction on the LR mesh is then given by
γNΞp (ξ) = γ1 NΞp1 (ξ) + γ2 NΞp2 (ξ) .
(2.19)
γ1 = α1 γ ,
γ2 = α2 γ .
(2.20)
with
The associated α1 and α2 are determined by
1 ,
ξˆ − ξ1
α1 =
ξp+1 − ξ1 ,
ξp+2 − ξˆ ,
α2 = ξp+2 − ξ2
1 ,
ξp+1 ≤ ξˆ < ξp+2 ,
ξ1 < ξˆ < ξp+1 ,
(2.21)
ξ2 < ξˆ < ξp+2 ,
ξ1 < ξˆ ≤ ξ2 .
Consider the knot ξˆ = 0.375 is inserted into the knot vector Ξ = [0, 0.25, 0.5, 0.75]. The knot vector
entries construct the quadratic basis function NΞ2 . As illustrated in Fig. 2.2, the original basis function
is split into two new functions each describing the LR B-spline basis functions γ α1 NΞ21 and γ α2 NΞ22 .
The refinement process for bivariate functions is performed in one parametric direction at a time, i.e.
1
0.8
0.6
0.4
0.2
0
0
0.25
0.375
0.5
0.75
1
Figure 2.2: The original basis function of the knot vector Ξ is split by knot insertion at ξ = 0.375
into two local basis functions characterized by the knot vectors Ξ1 and Ξ2 .
first all horizontal meshlines and then all vertical meshlines are inserted (or vice versa). Consider two
local knot vectors Ξ and H, one in each parametric direction. For a weighted, bivariate LR B-spline
pq
basis γ BΞ
(ξ, η) with Ξ = [Ξ, H] the following relation is given
pq
q
γ BΞ
(ξ, η) = γ NΞp (ξ) MH
(η) ,
q
= γ (α1 NΞp1 (ξ) + α2 NΞp2 (ξ)) MH
(η) ,
=
pq
γ1 BΞ
(ξ, η)
1
+
pq
γ2 BΞ
(ξ, η)
2
(2.22)
,
for the case that Ξ is split into Ξ1 and Ξ2 . The refinement process is given by two steps. At first
all LR B-spline basis functions whose support domain is crossed by meshline extensions will be split.
The second step is to check every newly created basis function if it has minimal support. In the case
that a newly created LR B-spline does not have minimal support an additional splitting is performed.
pq
pq
By splitting γ BΞ
(ξ, η) into γj BΞ
(ξ, η), with j = 1, 2, the following cases can occur:
j
7
1. The LR B-spline does not exist and a new LR B-spline is created. In this case the new control
point xj is a copy of the control point x of the parent LR B-spline xj = x. The weight γj is set
by Eq. (2.20).
pq
2. If BΞ
(ξ, η) already exists, the control point and the weight are set to
j
xj =
xj γ j + x γ α j
,
γj + αj γ
(2.23)
and
γj = γj + γαj .
(2.24)
pq
After splitting the former LR B-spline, γ BΞ
(ξ, η) is deleted in both cases. The algorithm proceeds
with the second step and checks if the support domain of the new local basis functions is fully crossed
by any existing meshline. If one does not have minimal support the first step is performed again. Note
that at every step of the refinement process the partition of unity is maintained and the geometric
mapping is left unchanged. For the use of LR B-splines in IGA it is required to ensure that the
resulting spline space is linearly independent. In the bivariate case this can be guaranteed by only
using primitive meshline extensions5 , see Mourrain (2014) and Johannessen et al. (2014). In this
work, all meshline extensions are formulated as primitives.
3
Geometric modeling using LR NURBS
So far, non-rational LR B-splines have been discussed. In this section, their extension to Locally
Refined Non-Uniform Rational B-Splines (LR NURBS) is introduced. With a NURBS representation
of objects one gains the ability to describe many geometries, which cannot be represented by polynomials, exactly. Especially conic sections, ellipsoids, spheres, cylinders, etc. can be constructed by a
projective transformation of piecewise quadratic curves exactly, see Farin (1992). This is one of the
defining features of NURBS.
3.1
LR NURBS
The fundamentals of standard NURBS in isogeometric analysis can be found in, e.g. Hughes et al.
(2005). Combining NURBS with the LR B-spline theory from Sec. 2.3, the extension of LR B-splines
to LR NURBS follows. An LR NURBS object in Rd is constructed by the projective transformation
of an LR B-spline entity in Rd+1 . The first d entries of a projective control point xw
i represent the
spatial coordinates and the d + 1 entry the weight, e.g. xw
=
[x
,
y
,
z
,
w
]
for
d
=
3. The control
i i i
i
i
points of the LR NURBS object result from the projective transformation
(xi )k =
(xw
i )k
,
wi
k = 1, ..., d ,
where
wi = (xw
i )d+1 ,
(3.1)
with (xi )k as the k th component of the vector xi and wi as the ith weight. To generalize this relation
the weighting function for the bivariate case is introduced by
W (ξ, η) =
n X
m
X
Bipjq (ξ, η) wi j .
(3.2)
i=0 j=0
5
A primitive meshline extension is (a) a meshline spanning p + 1 elements, (b) elongating a meshline by one element
or (c) raising the multiplicity of a meshline (length of p + 1 elements).
8
Bipjq (ξ, η) are the standard LR B-spline basis functions and the transformation is applied by dividing
every point of the curve by W (ξ, η). Each element of surface S is a polynomial, which is divided by
another polynomial of the same order. The resulting LR NURBS basis is then defined as
Ripjq (ξ, η) =
Bipjq (ξ, η) wi j
W (ξ, η)
Bipjq (ξ, η) wi j
.
= Pn Pm
B p q (ξ, η) wî ĵ
ĵ=0 î ĵ
î=0
(3.3)
With a set of control points xi j and the scaling factors γi j the LR NURBS surface is then defined as
x(ξ, η) =
n X
m
X
Ripjq (ξ, η) xi j γi j .
(3.4)
i=0 j=0
The properties of the LR NURBS basis, such as its continuity and support, follow directly from the
knot vectors as before. The basis is still non-negative and it still forms a partition of unity. This
leads to the strong convex hull property of the LR NURBS. Note that the weights are separated from
any explicit geometric representation. The weights are each associated with a specific control point
and a manipulation of them leads to a change of the resulting geometry. In the case that all weights
are equal, the surface is again a polynomial and consequently Ripjq (ξ, η) = Bipjq (ξ, η). Therefore, LR
B-splines are a special case of LR NURBS. In order to locally refine an LR NURBS object it is
necessary to take special care of the weights. Before the refinement procedure starts, the projective
control points xw
i j of the LR NURBS are computed by Eq. (3.1). The weights wi j are included in
w
vector xi j and treated as the fourth part of the control points. This results in
xw
i j = [xi j wi j , yi j wi j , zi j wi j , wi j ] .
(3.5)
After the refinement process, the control points are transformed back by dividing the spatial coordinates by their associated weight. Note that while the refinement is performed, wi j is treated in the
same way as the control points. The linear independence of LR NURBS spaces follow directly from
the LR B-splines.
3.2
Bézier extraction operator for LR NURBS
This section discusses the Bézier extraction of LR NURBS. The Bézier extraction is an advantageous
technique for the decomposition of splines into their Bézier elements. It allows for the embedding
of isogeometric analysis into a standard finite element framework. The Bézier extraction of NURBS
is introduced by Borden et al. (2011) and advanced to T-splines by Scott et al. (2011). The Bézier
extraction of LR NURBS shares similarities with the formulation of the Bézier extraction of T-splines.
Both element types have a local parametric representation. The Bézier decomposition of each local
basis function, that is nonzero within an element, is performed separately. The basis function in one
parametric direction of an LR NURBS element Γe within the knot span ξ can be expressed in terms
of a set of Bernstein polynomials B(ξ) and a linear operator cea . Each local basis function Nae (ξ) of
an element Γe can then be expressed by
Nae (ξ) = cea B(ξ) .
(3.6)
Here, a = 1, 2, ..., ne where ne is the number of nonzero basis functions of Γe . In matrix-vector form
this becomes
Ne (ξ) = Ce B(ξ) .
(3.7)
Ce is the element-wise Bézier extraction operator. The linear operator cea is of dimension [1 × (p + 1)]
and represents a single row of Ce . The coefficients of cea are calculated by an algorithm presented in
Scott et al. (2011). The algorithm inserts knots into the knot vector until all entries have multiplicity
p. The knot insertion performs the spline decomposition into its Bézier elements. With cea and the
Bernstein polynomials the relation in Eq. 3.6 is obtained. Due to the unstructured LR meshes it is
required to account for two aspects:
9
1. To determine cea it is required that the knot vectors are open, i.e. the first and last knot vector
entry have multiplicity of p + 1. The knot vectors of LR NURBS are in general not open. For
this, the local knot vectors are extended, which is simply done by adding knot vector entries.
The extension of the knot vectors does not affect the basis of LR NURBS itself because it is
only used for the Bézier decomposition.
2. The unstructured meshes of LR NURBS or T-splines necessitate a special consideration. An
example is shown in Fig. 3.1. Consider the element Γe with ξ̃ = [0, 0.1] marked in yellow in
Fig. 3.1 on the left. The support domain of one bivariate local basis function is marked in gray.
In ξ-direction, the basis function Na (ξ) is spanned by the local knot vector Ξ = [0, 0.2, 0.4, 0.6]
as shown in Fig. 3.1 on the right. The basis function is non-zero within the domain of Γe .
From Ξ follows that cea is defined within the knot span ξ = [0, 0.2] and exceeds ξ̃ = [0, 0.1] (the
domain of Γe ). Those non-matching domains must be considered and treated. For T-splines
an additional knot is inserted into the knot vector so that cea can be determined for the correct
parametric domain ξ̃ = [0, 0.1]. In contrast to T-splines, no additional knot insertion is required
for the formulation for LR NURBS.
0
0.2 0.4 0.6 0.8
1
Figure 3.1: Bivariate LR mesh (left) and the local basis function of a quadratic LR NURBS in one
parametric direction (right). The support width of Nae (ξ) is represented by cea B(ξ) in the knot span
ξ = [0, 0.2] exceeding the actual domain of Γe .
For LR NURBS the linear operator cea is mapped by using an additional linear operator Tea to obtain
a new c̃ea such that
Nae (ξ̃) = c̃ea B(ξ) .
(3.8)
With c̃ea = cea Tea
Nae (ξ̃) = cea Tea B(ξ) ,
(3.9)
follows. The matrix Tea follows from the relation
c̃ea B(ξ̃) = cea B(ξ) .
(3.10)
After some algebraic manipulations this equation is expressed as
c̃ea
=
cea
B(ξ̃)
−1
T
B(ξ)
such that
Tea
:=
B(ξ̃)
−1
10
,
(3.11)
T
B(ξ)
.
(3.12)
If the knot spans coincide, i.e. ξ̃ = ξ, then
B(ξ̃) = B(ξ) ,
(3.13)
Tea = I ,
(3.14)
and c̃ea = cea follows. With this formulation, the basis functions having support in each element Γe
are expressed by the Bernstein polynomials B(ξ) and the linear operators c̃ea .
4
Adaptive local refinement and coarsening
This section presents a technique for adaptive local refinement and coarsening for LR NURBS discretizations. The technique is formulated in the context of frictional contact. This requires the
mapping of contact variables from one mesh to another.
4.1
Adaptive local refinement
An adaptive local refinement technique necessitates an indicator for refinement. Commonly used
indicators use a posteriori error estimation that provide a reliable error distribution, see e.g. Ainsworth
and Oden (1997), Johannessen et al. (2014) and Kumar et al. (2015). In this work, a refinement
indicator based on the contact state is used. If new contact is detected in an element, refinement is
considered. If contact is lost in an element, coarsening is considered. After the contact domain is
determined, the local refinement is performed by meshline extensions in the parameter domain. The
refinement is performed until a prescribed refinement depth or smallest element size, is obtained. For
sliding contact it is desirable to automatically control the technique of adaptive local refinement and
coarsening. A changing contact domain can lead to a high number of newly detected contacts and
lost contacts. To reduce the number of refinement and coarsening events, three parameters are used.
In Fig. 4.1 four cases are sketched to illustrate the automatic control by using the parameters ddref ,
ddsafe and ddcrs . The superscript d denotes the current refinement depth that is performed.
• The parameter ddref ≥ 0 enlarges the refinement domain by a number of unrefined elements (see
Fig. 4.1a). In this example, ddref = 3 de1−d in both directions. Here,
de1−d =
d0e
,
21−d
(4.1)
is the minimum element length of the previous refinement depth and d0e is the original element
length. The elements within the enlarged contact domain are flagged for refinement. The LR
mesh is the result of splitting all flagged elements equally.
• The parameter ddsafe ≥ 0 sets a safety domain at the boundary between coarse and refined
elements. If contact is detected within this domain, local refinement is performed. An example
with ddsafe = d1−d
in both directions is shown in Fig. 4.1b. This parameter ensures that at every
e
time-step the contact domain is surrounded by refined elements. ddsafe detects a refinement event
and does not specify a domain for refinement.
• The parameter ddcrs ≥ 0 controls the coarsening. An example with ddcrs = 5 de1−d in both directions
is shown in Fig. 4.1c. Coarsening is performed if refined elements are detected beyond this
distance from the contact domain. ddcrs detects a coarsening event and does not specify a domain
for coarsening.
11
Figure 4.1: Automatic control of the refinement and coarsening procedure based on the parameters
ddref , ddsafe and ddcrs . (a.) Mesh refinement: During a refinement step, mesh refinement is considered
within the enlarged contact domain. (b.) Refinement detection: The need for mesh refinement is
detected once the contact domain overlaps with the safety domain. Mesh refinement is then applied by
redetermining the domain as shown in (a.). (c.) Coarsening detection: The need for mesh coarsening
is detected when refined elements are detected beyond the distance ddcrs from the contact domain.
Mesh coarsening is then applied as described in Sec. 4.2. (d.) Combined setup of the parameters ddref ,
ddsafe and ddcrs .
Fig. 4.1d shows the three parameters combined. The parameters ensure that the contact domain
and the neighboring elements are at every time-step represented in the desired discretization. The
parameters ddref , ddsafe and ddcrs can be set arbitrarily and a good choice is depending on the specific
problem setup. A good compromise of the size of the locally refined domain and the number of
refinement/coarsening events is desirable. The coarsening is discussed in the next section.
12
4.2
Adaptive coarsening
Coarsening LR meshes in the context of contact computations is challenging. The reasons are the
unstructured parametric representation of LR NURBS and that frictional contact is history dependent.
The contact variables need to be preserved during the process of coarsening and refinement. A
technique for NURBS and T-spline coarsening is presented in Thomas et al. (2015). This technique
has not been advanced to LR NURBS yet. To overcome this here, the coarsening is combined with
local refinement. Therefore, an intermediate step is introduced, in which the entire mesh is coarsened.
Based on the coarse mesh, the local refinement is then performed to obtain the desired LR mesh. The
coarsening and refinement process requires that the contact domain has the same discretization and
identical contact variables before and after the process. Due to the intermediate step, the contact
domain is temporary coarsened. To preserve the contact variables a mapping is required. The contact
variables are stored before the mesh is coarsened, and then mapped to the new mesh after the
refinement step. The technique consists of six main steps to perform a consistent adaptive coarsening
within the context of contact computations. The technique is sketched in Fig. 4.2 and the steps are:
1. Coarsening is detected when refined elements are detected beyond the distance ddcrs (Fig. 4.2a).
Store the current configuration and the contact variables for each element in contact.
2. Recover the undeformed, reference mesh (Fig. 4.2b).
3. Interpolate the control points of the mesh in Fig. 4.2a to obtain the deformed, coarse mesh
(Fig. 4.2c). The interpolation leads to a geometric approximation error, which is addressed in
step 6.
4. Perform local refinement within the enlarged contact domain ddref (Fig. 4.2d) to obtain the
desired LR mesh (Fig. 4.2e).
5. Preserve the contact variables by mapping them to the current configuration.
6. Reduce the geometric approximation error by recovering control points from the configuration
in Fig. 4.2a: By comparing the meshes in Fig. 4.2a and Fig. 4.2e it turns out that their discretization only differ slightly. The control points of the domains, that remain unchanged in
their discretization, are recovered.
As Fig. 4.2 shows, the discretization of the contact domain, at a given time-step, is unaffected by the
coarsening and refinement procedure. With the adaptive local refinement and coarsening technique,
a refinement/coarsening event is not detected at every computational step. A refinement/coarsening
event is performed only when it is detected according to the criteria of Fig. 4.1. The results of this
technique can be seen in the following examples, e.g. Fig. 5.8.
13
Figure 4.2: Sketch of the proposed coarsening and refinement procedure. (a.) LR mesh where coarsening is detected (shown are the deformed configuration and the corresponding parametric representation). (b.) Undeformed, reference mesh in its coarse representation. (c.) Deformed configuration
in its coarse representation. (d.) Local refinement detected. (e.) Deformed, locally refined mesh and
parametric representation of the mesh.
5
Numerical examples
The performance of the adaptive local refinement and coarsening technique using LR NURBS elements
is illustrated in this section by several numerical examples. The first example is used for validation.
In the remaining examples, the performance of LR NURBS for frictionless and frictional contact is
investigated. In the last example, LR NURBS-enriched contact elements are considered for 3D friction
of two deformable solids. To investigate the benefit of LR NURBS discretization, comparisons to
reference models with uniform NURBS discretization are made.
5.1
Inflation of a hemisphere
The first example considers the inflation of a spherical membrane balloon. The inflation can be described by an analytical formula, which is used for validation. The system is modeled as a hemisphere
and the boundary conditions are chosen such that the symmetry of the model is maintained during
inflation, see the left side of Fig. 5.1. The discretized geometry consists of five patches and the NURBS
weights differ for each. This chosen geometry is not perfectly spherical but still an approximation.
The material is described by the incompressible Neo-Hookean membrane material model (Sauer et al.,
2014)
µ
aαβ
αβ
αβ
σ =
A − 2
.
(5.1)
J
J
14
Here, µ is the shear modulus, J is the surface area change and Aαβ and aαβ are the contra-variant
components of the metric tensor in the reference and current configuration. The problem setup is
similar to the one of Sauer et al. (2014) in which the balloon is modeled as 1/8th of a sphere with
NURBS and Lagrange discretizations. Initially, the balloon has the volume V0 = 4πR3 /3 with the
radius R. The volume is increased step-wise until it reaches V = 50 V0 . The analytical pressure-volume
relation is given by
1 7 !
pint R
V0 3
V0 3
−
=2
,
(5.2)
µ
V
V
with the internal pressure pint . We now investigate the behavior of the proposed local refinement pro-
Figure 5.1: Inflation of a hemispherical membrane: Model setup in the initial configuration (left).
Blue lines denote the allowed direction of deformation. Black supports denote the fixed directions.
Current configuration at V = 50 V0 , two times locally refined (right).
cedure using quadratic LR NURBS elements. Five different uniform meshes with quadratic NURBS
discretizations are taken each as the starting point. The number of Gaussian quadrature points is
nqp = 3 × 3 for all elements. As an homogeneous hemisphere has no domain of major interest, the
center is chosen to apply local refinement of depth 2, see the right side of Fig. 5.1. The performance of
the five meshes is compared to quadratic NURBS and quadratic Lagrange discretizations examining
the pressure error, see Fig. 5.2. The isogeometric elements behave as observed in Sauer et al. (2014)
and show better convergence behavior than standard Lagrange elements. The LR NURBS meshes
behave equal well than the uniform NURBS meshes. Better behavior can not be expected in this
example, since the hemisphere has a homogeneous behavior and local refinement does not improve
the numerical results significantly. But with this first example we can conclude that the applied local
refinement procedure works successful and the computations with LR NURBS are valid. In the following examples, the performance of adaptive local refinement with respect to the accuracy and the
computational cost is investigated.
15
10 -3
10 -6
10 -9
10 -12
10 3
10 4
10 5
10 6
Figure 5.2: Inflation of a hemispherical membrane: Pressure error for meshes discretized by quadratic
Lagrange, quadratic NURBS and quadratic LR NURBS elements in comparison to the analytical
solution. There is no benefit in local refinement for a problem with a uniform solution and hence LR
NURBS perform equally well than uniform NURBS.
5.2
Rigid sphere in contact with a square membrane sheet
The second example considers a square membrane sheet with dimension 2 L0 × 2 L0 that is initially
pre-stretched by λ = 1.1. The pre-stretching is applied to avoid the membrane instability. A rigid
sphere, located at X = [0, 0, L0 ] with radius R = L0 , is pushed gradually downwards until the bottom
of the sphere reaches z = −L0 /2 as shown in Fig. 5.3. The initial mesh consists of 4 × 4 LR NURBS
elements with 5 × 5 Gaussian quadrature points each. The material behavior is the same as in the
previous example. The boundaries at X = 2 L0 and Y = 2 L0 are fixed in all directions. At X = 0 and
Y = 0 the deformation is constrained to be zero in x- and y-direction, respectively. Contact between
the rigid sphere and the membrane is treated by the penalty method with the penalty parameter
l0x l0y p−1
el
0
εn = εn ·
,
(5.3)
lelx lely
depending on the individual element size. Here, lelx and lely denote the current element lengths, and
l0x and l0y the initial element lengths. p is the polynomial order6 and ε0n = 10 E0 /L0 is the constant
penalty parameter. A mesh and order depending penalty parameter was studied in Roohbakhshan
and Sauer (2016) and Sauer and De Lorenzis (2015) giving a good balance between accuracy and
convergence behavior. The contact domain is locally refined by the procedure described in Sec. 4. A
series of resulting meshes for quadratic LR NURBS elements is illustrated in Fig. 5.4 from refinement
depth 1 to 5. A strong local aggregation of elements is obtained in the contact domain. In this
example, the parameter ddref is taken as zero. The performance of LR NURBS meshes and uniform
meshes using NURBS discretizations is investigated by examining the relative error
erel
n =
| fnref − fn |
,
| fnref |
(5.4)
of the normal contact force fn w.r.t. a reference solution. For the reference solution, a uniform mesh
6
with cubic NURBS discretizations is taken. This mesh has more than 5 · 104 dofs and εel
n > 10 E0 /L0 ,
which is highly accurate. The relative contact force error of LR NURBS and homogeneous meshes is
shown in Fig. 5.5. The proposed local refinement technique using LR NURBS elements show very good
convergence behavior, both for quadratic and cubic elements. It turns out that for cubic elements the
6
the order in both parametric directions is set equally, i.e. q = p
16
b.
a.
c.
Figure 5.3: Rigid sphere in contact with a square membrane sheet: (a) Initial membrane discretization,
boundary conditions and rigid sphere. (b), (c) Deformed membrane surface from two different points
of view. The coloring shows the downward displacement of the membrane.
Figure 5.4: Rigid sphere in contact with a square membrane sheet: Deformation series with a locally refined contact domain. A refinement depth from 1 to 5 is applied. The coloring shows the
displacement of the membrane.
obtained relative error decreases to ≈ 1.63 ·10−5 , while for quadratic elements ≈ 2.1 ·10−4 is achieved.
Since the penalty parameter is increased by Eq. (5.3) and since it is depending on the polynomial
order, higher accuracy for cubic elements can be expected. The dofs are reduced by a factor of more
than 16 for LR NURBS meshes with quadratic elements at the 5th refinement depth. For LR NURBS
meshes with cubic elements at the 4th refinement depth the dofs are reduced by a factor of more than
37. This is a huge decrease of the computational cost. In Tab. 5.1 the mesh properties and numerical
results for cubic elements are listed. This example shows that with the local refinement procedure, a
high accuracy is achieved while decreasing the computational cost.
LR mesh
depth 0
depth 1
depth 2
depth 3
depth 4
dofs
96
280
416
701
1340
] el.
16
64
112
211
430
rel. error
1.95 · 10−1
1.72 · 10−2
1.45 · 10−3
1.09 · 10−4
1.63 · 10−5
uniform mesh
depth 0
depth 1
depth 2
depth 3
depth 4
dofs
96
280
936
3400
12936
] el.
16
64
256
1024
4096
rel. error
1.95 · 10−1
1.72 · 10−2
1.45 · 10−3
1.09 · 10−4
1.62 · 10−5
Table 5.1: Rigid sphere in contact with a square membrane sheet: Comparison of the mesh parameters
and the computational error between the LR mesh and the uniformly refined mesh. Cubic discretization is considered. Corresponding refinement depths have almost equal error but hugely different
number of dofs.
17
10 0
10 0
10 -1
10 -1
10 -2
10 -2
10 -3
10 -3
10 -4
10 -4
10 -5
10 -5
10
2
10
3
10
4
10 2
10 3
10 4
Figure 5.5: Rigid sphere in contact with a square membrane sheet: Relative error of the normal
contact force for quadratic discretizations (left) and cubic discretizations (right). The error is defined
w.r.t. a uniformly discretized mesh with cubic NURBS elements that is taken as the reference.
In a second part of this example, the performance of LR NURBS elements for different sizes of the
contact domain is investigated. Computations for different radii of the rigid sphere are performed.
The initial radius is set to R = 1/10 L0 and is increased step-wise to R = L0 . The sphere is positioned
at X = Y = 0 and lowered gradually to z = −R/2. The performance of LR NURBS meshes using
cubic elements is investigated. Local refinement of depth 5 is applied. The relative error of the
normal contact force is defined w.r.t. uniformly discretized meshes using cubic NURBS elements. The
smallest element size of each mesh is identical, leading to the same value for εel
n . The percentage of
the used dofs with respect to the radius R is shown in Fig. 5.6. The relative contact force error for all
radii R is in the range of 5 · 10−8 and 3 · 10−7 . The LR NURBS meshes capture the reference solutions
nicely. A high accuracy is achieved while only using few dofs. Consequently, for R = 1/10L0 less
than 2% of the time is used in comparison to the reference model. The time for the local refinement
procedure is negligible. This reduction of the computational effort that still achieves high accuracy
demonstrates the benefit of the local refinement technique using LR NURBS.
7
6
5
4
3
2
1
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 1.1
Figure 5.6: Rigid sphere in contact with a square membrane sheet: Percentage of the used dofs with
respect to the radius R. The y-axis shows the ratio of the used dofs with respect to the reference
model.
18
5.3
Frictionless sliding contact
In the third example the performance of the adaptive local refinement and coarsening technique from
Sec. 4 is investigated. For this, frictionless sliding contact is considered. The problem setup consists
of a rectangular membrane sheet having dimension 8 λ L0 × 2 λ L0 with isotropic pre-stretch λ = 1.25.
The boundaries are clamped. A rigid sphere with radius L0 is initially located at X 0 = λ · [L0 , L0 , L0 ]
and first moved downwards to x0 = λ · [L0 , L0 , L0 /2]. This is followed by horizontal motion until the
sphere reaches x0 = λ · [7 L0 , L0 , L0 /2], see left side of Fig. 5.7. The volume enclosed by the membrane
is constrained to be constant so that it behaves like a cushion (Sauer et al., 2014). The material and
the computational parameters are the same as in the previous example. During frictionless sliding
the contact domain changes, and the adaptive local refinement and coarsening procedure from Sec. 4
is applied to obtain highly resolved meshes in the contact domain.
Figure 5.7: Frictionless sliding contact: Initial problem setup (left). The membrane is initially discretized by 4 × 16 elements. The LR NURBS mesh, after the sphere is moved downward, is shown
on the right figure. Local refinement of depth 4 is applied leading to a highly resolved mesh in the
contact domain.
The right side of Fig. 5.7 shows the deformed configuration after the downward displacement of the
rigid sphere has been applied. The parameters for the automatic control from Sec. 4 are ddref = 3 d1−d
e ,
d = 4 d0 . In Fig. 5.8 two meshes resulting from the adaptive local refinement and
and
d
ddsafe = 2 d1−d
e
crs
e
coarsening technique are illustrated. It can be observed that during sliding only the contact domain
has a high aggregation of LR NURBS elements.
Figure 5.8: Frictionless sliding contact: Adaptive local refinement and coarsening of the membrane
during frictionless sliding in top and side view. Highly resolved meshes within the local contact domain
are obtained while the periphery is still represented with the coarse, initial mesh.
In the following, the performance of adaptive local refinement and coarsening is investigated. For
this, LR NURBS meshes are compared to uniformly discretized meshes using standard NURBS. The
normal and tangential contact forces for the LR NURBS, the uniformly refined and the initial, coarse
mesh are illustrated in Fig. 5.9. The normal and tangential contact forces for the LR NURBS and
the uniformly refined mesh match each other nicely. The contact forces of the initial, coarse mesh
are oscillating. The oscillations arise from the coarse discretization. The normal contact force of the
initial, coarse mesh is much smaller than the normal contact forces of the refined meshes. The reason
19
2.2
0.04
2
0.02
1.8
1.6
0
1.4
-0.02
1.2
1
-0.04
0
1.5
3
4.5
6
7.5
0
1.5
3
4.5
6
7.5
Figure 5.9: Frictionless sliding contact: Normal (left) and tangential (right) contact forces for LR
NURBS, uniformly refined and initial, coarse meshes. The LR NURBS mesh uses less than 30% of
the dofs in comparison with the uniformly refined mesh.
is the mesh dependent penalty parameter, see Eq.(5.3). The net tangential contact force ft is small
in comparison to the net normal contact force fn , as the sliding is considered to be frictionless. In
10 -4
10 -4
10 -7
10 -7
10 -10
10 -10
10 -13
10 -13
0
1.5
3
4.5
6
7.5
0
1.5
3
4.5
6
7.5
Figure 5.10: Frictionless sliding contact: Relative contact force error of LR NURBS. Quadratic LR
NURBS elements are compared to quadratic, uniformly refined NURBS elements (left). Cubic LR
NURBS elements are compared to cubic, uniformly refined NURBS elements (right). The LR NURBS
mesh uses less than 30% of the dofs in comparison with the uniform mesh.
Fig. 5.10 the relative contact force error of ft and fn is illustrated. It can be seen that the error is
below 10−4 for quadratic LR NURBS and below 10−6 for cubic LR NURBS. The LR NURBS meshes
capture the reference solution nicely with less than 30% dofs in both examples. The high relative error
of quadratic LR NURBS elements in comparison to cubic elements is caused by the coarse periphery.
The evolution of the relative error in tangential direction shows abrupt changes. These result from
the geometric approximation error of the coarsening steps. The setting of the parameters for the
automatic control leads to an almost constant evolution of the relative normal contact force error.
This example shows that the adaptive local refinement and coarsening technique using LR NURBS
elements leads to highly accurate results while decreasing the computational cost.
20
5.4
Frictional contact of two deformable membranes
In the next example, the performance of LR NURBS is investigated for frictional contact considering
two deformable membranes. This example addresses two new aspects. First, the preservation of
sliding variables during coarsening and refinement. Second, the local refinement and coarsening of
two objects. We consider two deformable rectangular solid membranes with dimension 2 Lu0 × 0.5 Lu0
and 0.5 Ll0 × 2 Ll0 . They are fixed at their boundaries in all parametric directions. An isotropic prestretch of λ = 1.5 in longitudinal direction is applied to avoid membrane instabilities (like wrinkling)
during frictional sliding. The left side of Fig. 5.11 shows the problem setup. Both surfaces are inflated
until they almost touch. This is followed by sliding the upper membrane along the lower one. During
sliding, the inflated volume of each membrane is constrained to be constant. The adaptive locally
refined surfaces at initial contact detection are illustrated on the right side of Fig. 5.11. The friction
coefficient is set to µf = 0.25 and the enclosed volume of both surfaces is constrained to be constant
after inflation. The material behavior is the same as in the previous examples. The number of
Gaussian quadrature points is 5 × 5 for each element. Contact is treated by the penalty method. The
penalty parameter is increased by Eq. (5.3) with ε0 = 80 E0 /L0 and it is taken equal in tangential and
normal direction. The two-half-pass algorithm (Sauer and De Lorenzis, 2015) is used to evaluate the
contact forces. Due to frictional sliding the contact domain changes and the presented adaptive local
refinement and coarsening procedure is applied on both surfaces. In Fig. 5.12 two meshes resulting
from the adaptive local refinement and coarsening technique are illustrated.
Figure 5.11: Frictional contact of two deformable membranes: Initial model setup (left). The blue
arrows indicate the pressure associated with the inflation. The black arrows indicate the sliding
direction. Initially, each membrane surface is discretized by 8 × 32 elements. LR NURBS mesh at the
time-step of first contact detection (right). Local refinement of depth 2 is applied.
Figure 5.12: Frictional contact of two deformable membranes: Adaptive local refinement and coarsening of the membranes during frictional sliding. Additional zoom into the locally refined contact
domain. Highly resolved meshes within the local contact domain are obtained while the periphery is
still represented with the coarse, initial mesh.
The performance of quadratic and cubic LR NURBS is investigated. The LR NURBS meshes are
compared to their uniform counterparts using quadratic and cubic NURBS discretizations separately.
21
The normal contact force fn and the tangential contact force ft acting on the lower surface are shown
in Fig. 5.13. Their behavior is as expected, and both discretizations match each other nicely.
0.4
0.1
0.08
0.3
0.06
0.2
0.04
0.1
0.02
0
0
0
0.5
1
1.5
2
2.5
3
0
1
2
3
Figure 5.13: Frictional contact of two deformable membranes: Normal (left) and tangential (right)
contact forces for both uniform and LR NURBS mesh. The LR NURBS mesh uses less than 30% of
the dofs in comparison with the uniform mesh.
Next, the relative contact force errors
erel
t =
| ftref − ft |
,
| fnmax |
and erel
n =
| fnref − fn |
,
| fnmax |
(5.5)
are investigated. They are illustrated in Fig. 5.14. The evolution of the relative error in normal
direction shows a similar behavior as in the previous example. The cubic LR NURBS elements
perform better than quadratic elements due to the coarse periphery. The evolution of the relative
error in tangential direction is more smooth than in the previous example, see Sec. 5.3. The reason is
that the geometric approximation error shows no remarkable influence in this example.
10-2
10 -2
10-4
10 -4
10-6
10 -6
10-8
0
1
2
3
10 -8
0
1
2
3
Figure 5.14: Frictional contact of two deformable membranes: Relative contact force error of LR
NURBS. Quadratic LR NURBS elements are compared to quadratic, uniformly refined NURBS elements (left). Cubic LR NURBS elements are compared to cubic, uniformly refined NURBS elements (right). The LR NURBS mesh uses less than 30% of the dofs in comparison with the uniform
mesh.
22
A maximum relative error of ≈ 3 · 10−4 for quadratic LR NURBS discretizations is achieved. The
relative error for cubic LR NURBS elements is much smaller and a maximum of ≈ 3 · 10−6 is achieved.
d
1−d and dd = 4 d0 .
The parameters for the automatic control from Sec. 4 are ddref = 3 d1−d
e , dsafe = 2 de
crs
e
With this setting less than 30% of the dofs of the reference meshes are used for both quadratic and
cubic LR meshes. This example shows that the technique of adaptive local refinement and coarsening
can be successfully applied to frictional contact of membranes. The computational cost is reduced
while still achieving high accuracy. In the next example the adaptive local refinement and coarsening
technique is applied to LR NURBS-enriched volume elements considering frictional contact.
5.5
Frictional ironing contact between two deformable solids
Local refinement for general 3D isogeometrical models is challenging and the linear independency
of LR meshes in 3D has not been developed yet. To overcome this issue, the NURBS-enrichment
technique by Corbett and Sauer (2014) is used to extend LR NURBS surface discretizations to 3D.
5.5.1
LR NURBS-enriched finite elements
The core idea of this technique is to use standard finite Lagrange elements in the bulk domain and
enrich the surface by isogeometric finite elements. The local refinement of NURBS-enriched elements
is performed in two main steps
1. Local refinement of the isogeometric surface mesh using LR NURBS. The adpative local refinement technique from Sec. 4 is applied.
2. Extend the local refinement through all the Lagrange elements in the thickness direction
This work considers the local refinement of NURBS-enriched and Lagrange elements in two dimensions
only. The refinement of the thickness direction is not considered here, but it is possible to use standard
refinement procedures. The local refinement of the Lagrange elements typically involve the generation
of hanging nodes. For the treatment of these hanging nodes the approach of Demkowicz et al. (1989)
is used.
5.5.2
Numerical results
The last example considers a deformable hollow hemisphere in frictional contact with a deformable
block. Both solids have a quadratic LR NURBS-enriched contact surface. The hemisphere is discretized with three elements in its thickness direction and 7 × 7 elements on the contact surface. The
block is discretized with eight elements in its thickness direction and 4 × 40 elements on the contact
surface. Each contact element has 5 × 5 Gaussian quadrature points. The bulk domain is discretized
by linear Lagrange finite elements. Initially, the block has dimension L0 × L0 × 10 L0 and the hollow
hemisphere has the inner radius R0in = L0 /6 and the outer radius R0out = L0 /2, see Fig. 5.15. The
base of the block is fixed in all directions. Periodic boundary conditions are used on the left and
right boundary of the block. First, a downward displacement of 2/3R0out is applied to the top of the
hollow hemisphere, which is followed by a displacement of 6 L0 in tangential direction. The downward
displacement is considered to be frictionless while the tangential displacement is considered to be frictional. The solids follow the isotropic, non-linear Neo-Hookean material model (e.g. see Zienkiewicz
et al. (2005))
Λ
µ
σ=
ln (det(F )) I +
F FT − I ,
(5.6)
det(F )
det(F )
23
with the deformation gradient F and the identity tensor I. The Lamé constants Λ and µ can be
expressed in terms of the Poisson’s ratio ν and Young’s modulus E by
µ=
E
,
2 (1 + ν)
and
Λ=
2µν
.
1 − 2ν
(5.7)
Considered here are ν1 = ν2 = 0.3 for both solids and E1 = E0 for the block and E2 = 10 E0 for the
hollow hemisphere. Contact is computed by the two-half-pass algorithm using the penalty method
el
with ε0 = 80 E0 /L0 and increasing εel
n = εt by Eq. (5.3). The friction coefficient is set to µf = 0.25.
Figure 5.15: Frictional ironing: Initial problem setup of the hemisphere and the block.
Next, the performance of adaptive local refinement and coarsening using quadratic LR NURBSenriched finite elements is investigated. For this, LR NURBS meshes are compared to uniformly
discretized meshes using quadratic NURBS-enriched finite elements. Local refinement of depth 2
is applied. Fig. 5.16 shows the locally refined and deformed block after the hollow hemisphere has
been moved downward. The coloring shows the normalized stress invariant I1 = tr σ/E0 . It can be
observed that the area of major interest is locally refined while the periphery remains coarse. As the
surface of the hemisphere is almost entirely in contact with the block the local refinement leads to a
uniformly refined hemisphere, see Fig. 5.17.
Figure 5.16: Frictional ironing: Deformed, LR NURBS mesh of the block. Refinement of depth 2 is
applied. The coloring shows the normalized stress invariant I1 = tr σ/E0 .
The normal and tangential contact forces for LR NURBS and uniform meshes are shown in Fig. 5.18.
These are evaluated at the block’s contact surface. The contact forces for the two meshes only differ
slightly. By first investigating the normal contact force an oscillatory and periodic behavior can be
observed. The oscillations can be reduced by further mesh refinement. The normal contact force of
the LR NURBS mesh shows abrupt increases at several points. This is caused by an adaptive local
refinement step. The same can be observed for the tangential contact force, see right side of Fig. 5.18.
24
Figure 5.17: Frictional ironing: Refined model of the hollow hemisphere. Top view (left) and bottom
view (right). The coloring shows I1 = tr σ/E0 .
The absolute and relative contact force errors in normal and tangential direction are illustrated in
Fig. 5.19. The relative contact force errors are approximately in the range of 4 · 10−4 to 2 · 10−3 . The
errors show a continuous increase followed by an abrupt decrease. The errors reflect the influence of
the refinement and coarsening technique also seen in Fig. 5.18. Comparing the relative contact force
error to the results of the previous example, one order of magnitude is lost (in case of quadratic LR
NURBS elements). The third dimension of the solids affect the numerical results strongly. For further
improvements one could take local error measures into account. It is expected that by using cubic
LR NURBS-enrichment the error would decrease but this would not reduce the influence of the solid
domain. The parameters for the automatic control from Sec. 4 are ddref = 5 de1−d , ddsafe = 4 d1−d
and
e
d
0
dcrs = 6 de . With this setting the LR NURBS mesh has still ≈ 40% of the dofs of the reference mesh.
Reducing the size of the locally refined domain would lead to an increase of the error of the contact
forces. Still, the adaptive local refinement and coarsening of LR NURBS-enriched elements show the
benefits of this procedure.
0.181
0.05
0.1808
0.05
0.04
0.04995
0.0499
0.1806
0.04985
0.03
0.0498
0.04975
0.1804
0.02
0.0497
0.57
0.1802
0.18
0.58
0.59
0.6
0.01
0
2
4
0
6
0
2
4
6
Figure 5.18: Frictional ironing: Normal (left) and tangential (right) contact forces of the LR NURBS
and uniform meshes. The LR NURBS mesh uses ≈ 40% of the dofs in comparison with the uniform
mesh.
25
10 -2
10 -2
10 -4
10 -4
10 -6
10 -6
10 -8
10 -8
0
2
4
0
6
2
4
6
Figure 5.19: Frictional ironing: Absolute (left) and relative (right) error of the tangential and normal
contact forces of the LR NURBS mesh w.r.t. the reference solution. The LR NURBS mesh uses ≈ 40%
of the dofs in comparison with the uniform mesh.
6
Conclusion
This work presents a novel concept of adaptive local surface refinement using LR NURBS elements in
the framework of computational contact mechanics. The Bézier extraction of LR NURBS elements
is obtained by an additional mapping of the Bézier extraction operator. With this a convenient
embedding of LR NURBS elements into general finite element codes is achieved.
The adaptive local refinement and coarsening technique is automatically controlled by the presented
refinement indicators. The numerical results show that the geometric approximation error arising
from the coarsening is negligible. The technique is applied to pure isogeometric surface elements
formulation and to isogeometrically-enriched volume elements. The numerical examples show the
benefit of using LR NURBS. Due to the local refinement, the computational cost decreases, while still
achieving high accuracy. A good convergence behavior for LR NURBS elements is also achieved.
The numerical examples in Sec. 5.1 and Sec. 5.5 consider curved surfaces that truly use NURBS
instead of B-splines. So they show that the extension from LR B-splines to LR NURBS is implemented
successfully. The automatic control of the adaptive local refinement and coarsening technique works
robustly for frictionless and frictional contact. As seen in the examples, quadratic LR NURBS achieve
less accuracy than cubic LR NURBS. The reason for this is that the latter capture the solution in the
coarse (i.e. unrefined) mesh more accurately. This motivates the incorporation of local error measures
that indicate elements for refinement. Especially, the computations performed with quadratic LR
NURBS-enriched elements show the demand for such an error measure. The refinement depth is
prescribed in the presented examples. By using error measures the refinement depth can be specified
automatically to achieve a desired accuracy. This work considers the local refinement of LR NURBSenriched and Lagrange elements in two dimensions. More complex 3D objects call for an enhanced
adaptive local refinement technique, which includes local refinement in the third dimension.
In this work, only bivariate LR NURBS are considered. Their extension to trivariate LR NURBS
would be interesting, since they allow local refinement of 3D LR NURBS elements. This is a topic of
its own because of the complex structure of trivariate LR NURBS meshes. Ensuring that the resulting
LR mesh is linearly independent has not been proven yet and is challenging for arbitrary meshline
extensions.
26
Acknowledgements
The authors are grateful to the German Research Foundation (DFG) for supporting this research
through project GSC 111.
References
Ainsworth, M. and Oden, J. (1997). A posteriori error estimation in finite element analysis. Comput.
Meth. Appl. Mech. Engrg., 142(1-2):1 – 88.
Autodesk (2015). T-splines plug-in. http://www.autodesk.com.
Borden, M. J., Scott, M. A., Evans, J. A., and Hughes, T. J. R. (2011). Isogeometric finite element
data structures based on Bezier extraction of NURBS. Int. J. Numer. Meth. Engrg., 87:15–47.
Bressan, A. (2013). Some properties of LR-splines. Comp. Aid. Geom. Des., 30(8):778–794.
Corbett, C. J. and Sauer, R. A. (2014). NURBS-enriched contact finite elements. Comput. Meth.
Appl. Mech. Engrg., 275:55 – 75.
Corbett, C. J. and Sauer, R. A. (2015). Three-dimensional isogeometrically enriched finite elements
for frictional contact and mixed-mode debonding. Comput. Meth. Appl. Mech. Engrg., 284:781–806.
Cottrell, J. A., Hughes, T. J. R., and Bazilevs, Y. (2009). Isogeometric Analysis: Toward Integration
of CAD and FEA. Wiley.
Cox, G. (1971). The Numerical Evaluation of B-Splines. DNAC. National Physical Laboratory,
Division of Numerical Analysis and Computing.
De Boor, C. (1972). On calculating with B-splines. Journal of Approximation Theory, 6(1):50 – 62.
De Lorenzis, L., Temizer, I., Wriggers, P., and Zavarise, G. (2011). A large deformation frictional contact formulation using NURBS-based isogeometric analysis. Int. J. Numer. Meth. Engrg., 87:1278–
1300.
De Lorenzis, L., Wriggers, P., and Zavarise, G. (2012). A mortar formulation for 3D large deformation
contact using NURBS-based isogeometric analysis and the augmented Lagrangian method. Comp.
Mech., 49:1–20.
Demkowicz, L., Oden, J., Rachowicz, W., and Hardy, O. (1989). Toward a universal h-p adaptive
finite element strategy, part 1. constrained approximation and data structure. Comput. Meth. Appl.
Mech. Engrg., 77(1-2):79 – 112.
Dimitri, R., Lorenzis, L. D., Wriggers, P., and Zavarise, G. (2014). NURBS- and T-spline-based
isogeometric cohesive zone modeling of interface debonding. Comput. Mech., 54(2):369–388.
Dimitri, R. and Zavarise, G. (2017). Isogeometric treatment of frictional contact and mixed mode
debonding problems. Computational Mechanics, (doi:10.1007/s00466-017-1410-7).
Dokken, T., Lyche, T., and Pettersen, K. F. (2013). Polynomial splines over locally refined boxpartitions. Comp. Aid. Geom. Des., 30(3):331–356.
Farin, G. (1992). From conics to NURBS: A tutorial and survey. IEEE Comput. Graph. Appl.,
12(5):78–86.
27
Forsey, D. R. and Bartels, R. H. (1988). Hierarchical B-spline refinement. SIGGRAPH Comput.
Graph., 22(4):205–212.
Hager, C., Hauret, P., Le Tallec, P., and Wohlmuth, B. I. (2012). Solving dynamic contact problems
with local refinement in space and time. Comput. Meth. Appl. Mech. Engrg., 201:25–41.
Hughes, T. J. R., Cottrell, J. A., and Bazilevs, Y. (2005). Isogeometric analysis: CAD, finite elements,
NURBS, exact geometry and mesh refinement. Comput. Meth. Appl. Mech. Engrg., 194:4135–4195.
Johannessen, K. A., Kumar, M., and Kvamsdal, T. (2015a). Divergence-conforming discretization for
stokes problem on locally refined meshes using LR B-splines. Comput. Meth. Appl. Mech. Engrg.,
293:38 – 70.
Johannessen, K. A., Kvamsdal, T., and Dokken, T. (2014). Isogeometric analysis using LR B-splines.
Comput. Meth. Appl. Mech. Engrg., 269:471–514.
Johannessen, K. A., Remonato, F., and Kvamsdal, T. (2015b). On the similarities and differences
between Classical Hierarchical, Truncated Hierarchical and LR B-splines. Comput. Meth. Appl.
Mech. Engrg., 291:64 – 101.
Kumar, M., Kvamsdal, T., and Johannessen, K. A. (2015). Simple a posteriori error estimators in
adaptive isogeometric analysis. Computers and Mathematics with Applications, 70(7):1555 – 1582.
High-Order Finite Element and Isogeometric Methods.
Laursen, T. A. (2002). Computational Contact and Impact Mechanics: Fundamentals of modeling
interfacial phenomena in nonlinear finite element analysis. Springer.
Lee, C. and Oden, J. (1994). A posteriori error estimation of h-p finite element approximations of
frictional contact problems. Comput. Meth. Appl. Mech. Engrg., 113(1):11 – 45.
Lu, J. (2011). Isogeometric contact analysis: Geometric basis and formulation for frictionless contact.
Comput. Meth. Appl. Mech. Engrg., 200(5-8):726–741.
McNeel (2012). Rhinoceros 5. http://www.mcneel.com.
Mourrain, B. (2014). On the dimension of spline spaces on planar T-meshes. Mathematics Of Computation, 83(286):847–871.
Nørtoft, P. and Dokken, T. (2014). Isogeometric Analysis of Navier-Stokes Flow Using Locally Refinable B-Splines. Springer International Publishing, Cham.
Roohbakhshan, F. and Sauer, R. A. (2016). Isogeometric nonlinear shell elements for thin laminated composites based on analytical thickness integration. J. Micromech. Molec. Phys., 01(03&
04):1640010 1–24.
Sauer, R. A. and De Lorenzis, L. (2013). A computational contact formulation based on surface
potentials. Comput. Meth. Appl. Mech. Engrg., 253:369 – 395.
Sauer, R. A. and De Lorenzis, L. (2015). An unbiased computational contact formulation for 3D
friction. Int. J. Numer. Meth. Engrg., 101(4):251–280.
Sauer, R. A., Duong, T. X., and Corbett, C. J. (2014). A computational formulation for constrained
solid and liquid membranes considering isogeometric finite elements. Comput. Meth. Appl. Mech.
Engrg., 271:48–68.
Schillinger, D., Dede, L., Scott, M. A., Evans, J. A., Borden, M. J., Rank, E., and Hughes, T.
J. R. (2012). An isogeometric design-through-analysis methodology based on adaptive hierarchical
refinement of NURBS, immersed boundary methods, and T-spline CAD surfaces. Comput. Meth.
Appl. Mech. Engrg., (doi:10.1016/j.cma.2012.03.017).
28
Scott, M., Li, X., Sederberg, T., and Hughes, T. (2012). Local refinement of analysis-suitable Tsplines. Comput. Meth. Appl. Mech. Engrg., 213:206 – 222.
Scott, M. A., Borden, M. J., Verhoosel, C. V., Sederberg, T. W., and Hughes, T. J. R. (2011).
Isogeometric finite element data structures based on Bézier extraction of T-splines. Int. J. Numer.
Meth. Engrg., 88(2):126–156.
Sederberg, T. W., Zheng, J., Bakenov, A., and Nasri, A. (2003). T-splines and T-NURCCs. ACM
Trans. Graph., 22(3):477–484.
Temizer, I. and Hesch, C. (2016). Hierarchical NURBS in frictionless contact. Comput. Meth. Appl.
Mech. Engrg., 299:161 – 186.
Temizer, I., Wriggers, P., and Hughes, T. J. R. (2011). Contact treatment in isogeometric analysis
with NURBS. Comput. Meth. Appl. Mech. Engrg., 200:1100–1112.
Thomas, D., Scott, M., Evans, J., Tew, K., and Evans, E. (2015). Bézier projection: A unified
approach for local projection and quadrature-free refinement and coarsening of NURBS and Tsplines with particular application to isogeometric design and analysis. Comput. Meth. Appl. Mech.
Engrg., 284(0):55 – 105.
Wriggers, P. (2006). Computational Contact Mechanics. Springer, 2nd edition.
Wriggers, P. (2008). Nonlinear Finite Element Methods. Springer.
Zienkiewicz, O. C., Taylor, R. L., and Zhu, J. Z. (2005). The Finite Element Method: Its basis and
fundamentals. Butterworth-Heinemann, 6th edition.
29
| 5cs.CE
|
Cyclic Codes over the Matrix Ring M2(Fp)
and Their Isometric Images over Fp2 + uFp2 ∗
arXiv:1409.7228v2 [cs.IT] 29 Jun 2017
Dixie Falcunit, Jr.
IMSP, UP Los Baños
College 4031, Laguna, Philippines
e-mail: [email protected]
Virgilio Sison†
IMSP, UP Los Baños
College 4031, Laguna, Philippines
e-mail: [email protected]
February 18, 2018
Abstract
Let Fp be the prime field with p elements. We derive the homogeneous
weight on the Frobenius matrix ring M2 (Fp ) in terms of the generating character. We also give a generalization of the Lee weight on the finite chain ring
Fp2 +uFp2 where u2 = 0. A non-commutative ring, denoted by Fp2 +vp Fp2 , vp
an involution in M2 (Fp ), that is isomorphic to M2 (Fp ) and is a left Fp2 - vector
space, is constructed through a unital embedding τ from Fp2 to M2 (Fp ). The
elements of Fp2 come from M2 (Fp ) such that τ (Fp2 ) = Fp2 . The irreducible
polynomial f (x) = x2 + x + (p − 1) ∈ Fp [x] required in τ restricts our study of
cyclic codes over M2 (Fp ) endowed with the Bachoc weight to the case p ≡ 2 or
3 mod 5. The images of these codes via a left Fp -module isometry are additive
cyclic codes over Fp2 + uFp2 endowed with the Lee weight. New examples of
such codes are given.
Keywords: Frobenius matrix ring, finite chain ring, homogeneous weight, cyclic
codes.
∗
This paper was presented at the International Zurich Seminar on Communications, Zurich,
Switzerland, February 2014.
†
This author gratefully acknowledges financial support from Commission on Higher Education,
UPLB Academic Development Fund, and UP Research Dissemination Grant.
1
1
Introduction
The theory of codes over finite rings has gained much attention since the significant
result in [8] showed that several well-known families of good nonlinear binary codes
can be identified as Gray images of linear codes over the quaternary ring Z4 of
integers modulo 4. Several recent papers dealt with codes over finite Frobenius
rings. These rings are considered the most appropriate coding alphabet since the two
classical theorems, namely the extension theorem and the MacWilliams identities,
generalize neatly in the case of finite Frobenius rings.
Let p be a prime and r ≥ 1 an integer. We denote by Fpr the Galois field of
order pr and characteristic p. In this study we restrict ourselves to a small class of
finite Frobenius rings, the matrix rings over a finite field, in particular the ring of
2 × 2 matrices over Fp , denoted by M2 (Fp ). The multiplicative group GL(2, p) of
invertible matrices in M2 (Fp ) will be of much use in the ensuing discussion as well.
Until now very few publications on codes over non-commutative rings have been
seen. It was only in 2012 that the theory of cyclic codes over M2 (F2 ) was developed
[1]. The idea for the construction of cyclic codes over M2 (F2 ) came from [2] in which
was defined an isometric map φ from F24 onto M2 (F2 ) where
φ((a + bω, c + dω)) =
a+d
b+c
b+c+d a+b+d
using the usual Hamming weight wHam on F4 extended component-wise, and the
Bachoc weight wB on M2 (F2 ) such that wHam (α) = wB (φ(α)) for all α in F24 . Here ω
is a root of the monic irreducible polynomial x2 + x + 1 ∈ F2 [x] such that F4 is seen
as an extension of F2 by ω. The Bachoc weight on M2 (Fp ) as given in [2] is defined
as follows,
0 if A = 0
wB (A) = 1 if A ∈ GL(2, p)
p otherwise.
The study of codes over Z4 and M2 (F2 ) reveals the importance of weight functions
that are different from the Hamming weight. Here we derive the homogeneous
weight on M2 (Fp ) using the formula introduced by T. Honold for arbitrary finite
Frobenius rings [9]. Likewise we extend the definition of the Lee weight on F2 + uF2 ,
u2 = 0 given in [4] to the finite chain ring Fp2 + uFp2 , u2 = 0. The connection
between the minimal left ideals and the idempotent elements of M2 (Fp ) is used to
generalize the homogeneous weight on M2 (Fp ). We also employ the well known
representation of the field by matrices by giving a unital embedding τ from Fp2 to
M2 (Fp ) to construct a non-commutative ring that is isomorphic to M2 (Fp ) and is a
left Fp2 -vector space. This ring is denoted by Fp2 + vp Fp2 where vp is an involution
in M2 (Fp ) and the elements of Fp2 come from M2 (Fp ) such that τ (Fp2 ) ∼
= Fp2 . The
unital embedding τ comes from a characterization of Fp in terms of an irreducible
2
polynomial f (x) = x2 + x + (p − 1) ∈ Fp [x]. The property if this polynomial restricts
our study to the case where p ≡ 2 or 3 mod 5. As a consequence certain structural
properties of cyclic codes over M2 (Fp ) that are similar to those of cyclic codes over
M2 (F2 ) are derived. The structure theorems used the transformation of the noncommutative ring Fp2 + vp Fp2 to Fp2 + up Fp2 by introducing a matrix ip ∈ M2 (Fp )
such that up = ip + vp , where u2p is the zero matrix. Also we define a left Fp -module
isometry from M2 (Fp ) to Fp2 + uFp2 using their respective Bachoc weight and Lee
weight.
2
Homogeneous Weight on M2(Fp)
Let R be a finite ring and R the set of real numbers. A weight function w : R −→ R
is called left homogeneous provided w(0) = 0 and the following hold:
(H1) If Rx = Ry for x, y ∈ R, then w(x) = w(y).
(H2) There exists Γ > 0 such that for every nonzero x ∈ R there holds
X
w(y) = Γ|Rx|.
y∈Rx
The definition for a right homogeneous weight follows analogously, and we say that
w is homogeneous if it is both left homogeneous and right homogeneous. The number
Γ is called the average value of w. The weight w is said to be normalized if Γ = 1.
It is well known that the normalized homogeneous weight on Fq , q = pr , is given by
0
if x = 0
q
wnhom (x) =
if x 6= 0.
q−1
This idea comes from the generalization of the homogeneous weight on a finite chain
ring [7]. But our goal is to give a generalization of the homogeneous weight on
M2 (Fp ) which is not a finite chain ring but is a finite (non-commutative) Frobenius
ring. We shall use the generating character instead of the Möbius inversion formula
for homogeneous weight that was employed in [5].
For a finite Frobenius ring R, Honold [9] observed that every homogeneous weight
on R with generating character χ must have the form
#
"
1 X
χ(ux)
w : R −→ R, x 7→ Γ 1 − ×
|R |
×
u∈R
where R× is the group of units of R. Note that every finite Frobenius ring has a
generating character [12]. The generating character of Mn (Fq ) is
(
)
2πi · tr(T r(A))
χ(A) = exp
p
3
r−1
where tr is the trace map from Fq down to Fp , that is, tr(α) = α+αp +· · ·+αp for
α ∈ Fq , and T r is the classical trace of the matrix A ∈ Mn (Fq ). The homogeneous
weight on Mn (Fq ) is given by
"
#
X
1
w : Mn (Fq ) −→ R, A 7→ Γ 1 −
χ(uA)
|GL(n, q)|
u∈GL(n,q)
where GL(n, q) is the group of nonsingular matrices in Mn (Fq ). It is known that
Q
|GL(n, q)| = q n(n−1)/2 ni=1 (q i − 1) [3].
The main concern in this section is to derive the homogeneous weight on M2 (Fp ).
First we discuss the structure of M2 (Fp ).
Remark 2.1 The matrix ring Mn (Fq ) has no proper ideals but it has proper left
ideals [10]. In particular M2 (Fp ) has p + 1 minimal left ideals [2]. This is essential
in this section so we take it as a theorem.
Theorem 2.2 M2 (Fp ) has p + 1 minimal left ideals and each minimal left ideal
contains p2 elements.
0 0
1 r
a0 a1
are
and
. Note that
Proof: Let A ∈ M2 (Fp ) where A =
0 1
0 0
a2 a3
nonzero nonunit idempotents of M2 (Fp ) where r ∈ Fp . Thus the proper left ideals
0 a1
a0 ra0
. Hence there are p + 1 minimal left ideals
and
are of the form
0 a3
a2 ra2
in M2 (Fp ) since the intersection of any two minimal left ideals of M2 (Fp ) is the zero
matrix and every minimal left ideal of M2 (Fp ) has p2 elements. ✷
In order to generalize the homogeneous weight on M2 (Fp ) we need to get the
P
value of the sum
u∈GL(2,p) χ(uA) where A ∈ M2 (Fp ). The case when A is the
zero matrix is obvious. Theorem 2.3 below deals with the invertible matrices while
Theorem 2.4 involves the zero divisors.
Theorem 2.3
P
u∈GL(2,p) χ(u) =
P
u∈GL(2,p)
χ(uA) = p where A ∈ GL(2, p).
Proof: Let D be the set of all the zero divisors in M2 (Fp ). We have
0 [9]. So,
X
X
χ(B) − χ(0)
χ(u) = −
P
A∈M2 (Fp )
χ(A) =
B∈D
u∈GL(2,p)
and since M2 (Fp ) has p + 1 minimal left ideals, χIL (A) = χ(A) for all A ∈ IL and
χIL (0) = 1 in [9], where χIL is a character of the minimal left ideal IL of M2 (Fp ).
Hence,
X
X
χ(u) = −(p + 1)
χIL (B) − 1 = −(p + 1)(−1) − 1 = p.
u∈GL(2,p)
B∈IL \{0}
✷
4
Theorem 2.4
Proof:
−
P
uk ∈GL(2,p)
X
χ(uk B) = p − p2 for all B ∈ IL \{0}.
"
χ(uk B) =
uk ∈GL(2,p)
X
χ(Bj )
Bj ∈IL \{0}
X
χ(uk B)
uk ∈GL(2,p)
X
=
#"
X
#
χ(uk B)χ(Bj )
Bj ∈IL \{0} uk ∈GL(2,p)
X
=
X
χ(uk B + Bj )
Bj ∈IL \{0} uk ∈GL(2,p)
X
=
X
χ(uk B + Bj )
uk ∈GL(2,p) Bj ∈IL \{0}
For each ur ∈ GL(2, p), there exists Bs ∈ IL \{0} such that ur B + Bs = 0 (Note: Bs
is not unique for every ur ). Thus, given that uk B + Bj 6= 0 we have
−
X
X
χ(uk B) =
uk ∈GL(2,p)
χ(ur B + Bs ) +
ur ∈GL(2,p)
X
=
X
X
χ(uk B + Bj )
uk ∈GL(2,p) Bj ∈IL \{0}
χ(0) +
ur ∈GL(2,p)
X
X
χ(uk B + Bj )
uk ∈GL(2,p) Bj ∈IL \{0}
= |GL(2, p)| +
X
X
χ(uk B + Bj ).
uk ∈GL(2,p) Bj ∈IL \{0}
For every Bt ∈ IL \{0, Bs } we have ur B + Bt ∈ IL \{0, ur B} and for fixed Bt and ur
we can always find l such that ul 6= ur and ur B + Bt = ur B. Thus, we can collect
all the elements of IL \{0}. And since |IL \{0}| divides |GL(2, p)|,
−
X
χ(uk B) = |GL(2, p)| +
uk ∈GL(2,p)
2
|GL(2, p)||IL \{0}| − |GL(2, p)|
|IL \{0}|
2
2
X
Bj ∈IL \{0}
2
= (p − p)(p − 1) + (p − p)(p − 2)(−1)
= p2 − p.
Thus,
P
uk ∈GL(2,p)
χ(uk B) = p − p2 .✷
Theorem 2.5 The homogeneous weight on M2 (Fp ) is given by
0
if A = 0
!
1
Γ 1 −
if A ∈ GL(2, p)
2
(p − 1)(p − 1)
whom (A) =
!
2
p
otherwise.
Γ 2
p −1
Proof: The proof is straightforward from the two preceding theorems. ✷
5
χ(Bj )
Table 1: Bachoc Weight and Normalized Homogeneous Weight on M2 (F2 )
M (F )
2 p
0 0
0 0
1 0
0 1
0 1
1 1
1 1
1 0
1 0
1 1
1 1
0 1
0 1
1 0
0 0
1 0
3
Lee Weight on
wB
wnhom
0
0
1
2/3
1
2/3
1
2/3
1
2/3
1
2/3
1
2/3
2
4/3
M (F )
2 p
0 1
0 1
1 1
0 0
0 0
0 1
1 0
0 0
1 1
1 1
0 0
1 1
1 0
1 0
0 1
0 0
wB
wnhom
2
4/3
2
4/3
2
4/3
2
4/3
2
4/3
2
4/3
2
4/3
2
4/3
Fp2 + uFp2 , u2 = 0
In [4] the Lee weight wL of x = (x1 , . . . , xn ) ∈ (F2 +uF2 )n is defined as n1 (x)+2n2 (x),
where n2 (x) and n1 (x) are, respectively, the number of u symbols and the number
of 1 or 1 + u symbols present in x. So when n = 1, wL (0) = 0, wL (1) = wL (1 + u) = 1
and wL (u) = 2.
Consider the finite chain ring F3 + uF3 , u2 = 0 then we can define wL (x) =
n1 (x) + 3n2 (x), for all x ∈ F3 + uF3 , where n2 (x) and n1 (x) are, respectively, the
number of u symbols and the number of 1 or 1 + u symbols present in x, as can be
seen in Table 2.
Now consider the subset B2 of F4 + uF4 , u2 = 0 where
B2 = {(αa1 + αb1 ω) + u(βa1 + βb1 ω)|α = 1, a1 , b1 , β ∈ F2 }.
Similarly we can define the Lee weight on F4 + uF4 , u2 = 0 to be wL (x) = n1 (x) +
2n2 (x) where again n2 (x) and n1 (x) are, respectively, the number of u symbols and
the number of 1 or 1 + u symbols present in x.
0 if x = 0
wL (x) =
1 if A ∈ B2 \{0}
2 otherwise
6
Table 2: Lee weight on F3 + uF3 , u2 = 0
F3 + uF3
wL
0
0
1
1
2
1
1+u
1
2 + 2u = 2(1 + u) 1
u
3
2+u
3
2u
3
1 + 2u
3
This can also be seen in Table 3. In general we can define the Lee weight on
Fp2 + uFp2 , u2 = 0 as wL (x) = n1 (x) + pn2 (x).
wL (x) =
0 if x = 0
1 if A ∈ Bp \{0}
p otherwise
where Bp = {(αa1 + αb1 ω) + u(βa1 + βb1 ω)|α ∈ F×
p , a1 , b1 , β ∈ Fp }.
4
Fp2 -Linear Map
In this section we give the conditions on the finite field Fp for the polynomial f (x) =
x2 +x+(p−1) to be irreducible over Fp . Using the well known representation of fields
by matrices, Theorem 4.2 shows the corresponding cyclic algebra that is isomorphic
to M2 (Fp ) and is a left Fp2 -vector space.
Lemma 4.1 Let p ≡ 2 or 3 (mod 5) then the polynomial f (x) = x2 + x + (p − 1) is
irreducible over Fp .
Proof: The case when p = 2 is trivial. Note that the discriminant of the polynomial
f (x) is equal to 5 ∈ Fp . Then f (x) is reducible over Fp if there exists y ∈ Fp such
that y 2 ≡ 5 (mod p). By the Law of Quadratic Reciprocity of elementary number
theory, when p is odd, y 2 ≡ 5 (mod p) is solvable if and only if p ≡ 1 or -1 (mod 5).
✷
Pn
i
Theorem 4.2 Let f (x) =
i=0 ai x ∈ Fq [x] be a monic irreducible polynomial.
Then the mapping π : Fq [x] → Mn (Fq ), g(x) 7→ g(X) induces a unital embedding of
7
Table 3: Lee Weight on F4 + uF4 , u2 = 0
F4 + uF4
wL
0
0
1
1
ω
1
1+ω
1
1+u
1
ω + uω = ω(1 + u)
1
(1 + ω) + u(1 + ω) = (1 + ω)(1 + u) 1
u
2
ω+u
2
(1 + ω) + u
2
uω
2
1 + uω
2
(1 + ω) + uω
2
u(1 + ω)
2
1 + u(1 + ω)
2
ω + u(1 + ω)
2
Fq [x]/(f ) into Mn (Fq ) where
0 0
1 0
X = 0 1
. .
.. . .
0 0
· · · 0 −a0
· · · 0 −a1
· · · 0 −a2
.
..
. . ..
. .
.
· · · 1 −an−1
Remark 4.3 The matrix X is known as the companion matrix.
Corollary 4.4 Let Fp2 = Fp [ω] where ω 2 + ω + (p − 1) = 0 then τ : Fp2 −→ M2 (Fp )
defined by
a
b
a + bω →
7
b a + (p − 1)b
is an embedding.
Proof: The proof follows immediately from Lemma 4.1 and Theorem 4.2. ✷
Theorem 4.5 If ω is a root of f (x) = x2 + x + (p − 1) then ω p ≡ (p − 1)ω + (p −
1)(mod(ω 2 + ω + (p − 1)).
8
Proof: First we show that (p − 1)ω + (p − 1) is also a root of f (x), that is,
f [(p − 1)ω + (p − 1)] = [(p − 1)ω + (p − 1)]2 + [(p − 1)ω + (p − 1)] + (p − 1)
= [(p − 1)2 ω 2 + 2ω + 1] + [(p − 1)ω + (p − 1)] + (p − 1)
= ω 2 + 2ω + 1 − ω − 2
= ω 2 + ω + (p − 1)
= 0.
Now, let h(x) = xp . By the Division Algorithm, there exist g(x) and r1 x + r2
such that h(x) = g(x)f (x) + r1 x + r2 where r1 x + r2 is the remainder when h(x) is
divided by f (x). Since ω and (p − 1)ω + (p − 1) are roots of f (x) then we have
ω p = r1 ω + r2
and
[(p − 1)ω + (p − 1)]p = r1 [(p − 1)ω + (p − 1)] + r2
or equivalently,
(p − 1)ω p + (p − 1) = r1 (p − 1)ω + r1 (p − 1) + r2 .
Since the characteristic of Fp is p, then
[(p − 1)ω + (p − 1)]p = [(p − 1)ω]p + (p − 1)p = [(p − 1)p ω p ] + (p − 1)p .
By Fermat’s Little Theorem,
[(p − 1)p ω p ] + (p − 1)p = (p − 1)ω p + (p − 1).
Adding equations ω p = r1 ω + r2 and (p − 1)ω p + (p − 1) = r1 (p − 1)ω + r1 (p − 1) + r2
modulo p, the resulting equation is (p−1) = r1 (p−1)+2r2 or simply r1 +(p−2)r2 = 1.
Note that gcd(1, p−2) = 1. And we have 1 = (p−1)−(p−2) = (p−1)+(p−2)(p−
1). So, r1 = p−1 and r2 = p−1. Thus, ω p ≡ (p−1)ω+(p−1)( mod (ω 2 +ω+(p−1)).
✷
p−1 p−1
p
Theorem 4.6 τ (ω) =
.
p−1
0
Proof: Since τ is a homomorphism we have τ (ω p ) = τ p (ω). And from Corollary
4.4 and Theorem 4.5 we have
τ p (ω) = τ [(p − 1)ω + (p − 1)]
= τ [(p − 1)ω] + τ (p − 1)
✷
= τ (p − 1)τ (ω) + τ (p − 1)
p−1
0
0
1
p−1
0
=
+
0
p−1
1 p−1
0
p−1
0
p−1
p−1
0
=
+
p−1
1
0
p−1
p−1 p−1
=
.
p−1
0
9
Theorem 4.7 Let Fp be the set of all scalar matrices in M2 (Fp ), p ≡ 2 or 3 mod 5,
1
0
. Then vp τ (ω) = τ p (ω)vp , Fp [τ (ω)] = Fp2
τ (Fp2 ) = Fp2 and vp =
p−1 p−1
and M2 (Fp ) = Fp2 + vp Fp2 .
It is easy to show that vp τ (ω) = τ p (ω)vp and Fp [τ (ω)] = Fp2 since
0 0
2
.
τ (ω) + τ (ω) + τ (p − 1) =
0 0
Proof:
M2 (Fp ) = Fp2 + vp Fp2
)
(
a+c
b+d
|a, b, c, d ∈ Fp .
=
b − c − d a − b − c)
✷
5
Cyclic Codes over M2(Fp)
Structure theorems for cyclic codes over A2 = M2 (F2 ) were established in [1] by
introducing two matrices τ (ω) and v in A2 satisfying the relation vτ (ω) = τ 2 (ω)v.
0 1
and
A possible choice would be those given by Bachoc [2] which are v =
1 0
0 1
τ (ω) =
such that A2 = F2 [τ (ω)] + vF2 [τ (ω)] where F2 [τ (ω)] = F4 ∼
= F4
1 1
0
0
and A2 = F4 + uF4 . Alamadhi
and v 6= v2 . Setting u = τ (1) + v gives u2 =
0 0
et.al. [1] used the ring F4 + uF4 to develop structure theorems for cyclic codes over
M2 (F2 ) by simply extending from cyclic codes over F2 + uF2 , u2 = 0 [4].
It seems that a construction of cyclic codes over Fp + uFp , u2 = 0, will result in
the construction of cyclic codes over Ap = M2 (Fp ). Fortunately, Qian, Zhang and
Zhu [11] solved an open-ended question given in [4], that is, to extend the cyclic
codes over F2 + uF2 , u2 = 0 to Fp + uFp + · · · + uk−1Fp , uk = 0. Thus, the case when
k = 2 gives the cyclic codes over Fp + uFp , u2 = 0.
Fp2 + uFp2
|
uFp2
|
(0)
Figure 1: Lattice of ideals of Fp2 + uFp2 , u2 = 0
10
0 0
p−1 0
2
and
and up = vp + ip . Then up =
Let p ≡ 2 or 3 (mod 5), ip =
0 0
0
1
Ap = Fp2 + up Fp2
(
)
a
b
=
|a, b, c, d ∈ Fp .
b−c a−b−d
Fp2 + up Fp2
|
(0)
Figure 2: Lattice of ideals of Fp2 + up Fp2 ,u2p is the zero matrix
Let Ap [X] be the ring of polynomials over Ap . We have a natural homomorphic
mapping from Ap to its field Fp2 . For any a ∈ Ap , let â denote the polynomial
reduction modulo up . Now define a polynomial reduction mapping µ : Ap [X] −→
Fp2 [X] such that
r
r
X
X
f (X) =
ai X j 7→
âi X j .
i=0
i=0
A monic polynomial f over Ap [X] is said to be a basic irreducible polynomial if
its projection µ(f ) is irreducible over Fp2 [X]. An Ap -linear code C of length n is an
Ap -submodule of Anp . As left modules we have the expansion Rp,n = Ap [x]/(xn −1) =
Q
⊕tj=1 Ap,j , where the Ap,j = Ap [x]/(fj ) are quotient Ap -modules and xn −1 = tj=1 fj
where fj ’s are irreducible polynomials over Fp2 .
We shall prove the lemma and the theorem below using the same techniques in
[1] and [11] given the condition that p is not divisible by n.
Lemma 5.1 If f is an irreducible polynomial over Fp2 the only left A-modules of
Rp (f ) = Ap [X]/(f ) are (τ (0)), (up ) and (τ (1)). In particular this quotient ring is
a non-commutative chain ring.
Proof: Let I 6= (τ (0)) be an ideal of Rp (f ). Pick g in Ap [X] such that g + (f ) ∈ I,
but g ∈
/ (f ). Because f is irreducible the gcd of µg and f can only take two values,
τ (1) and f . In the first case g is invertible mod f and I = (τ (1)) = Rp (f ). If this
does not happen, I ⊆ up + (f ). To show the reverse inclusion, let g = up r with
up r + (f ) ⊆ I and up r + (f ) 6= τ (0). We can assume by the latter condition that
µr ∈
/ (f ). Hence by the irreducibility of f we have that gcd(µr, f ) = τ (1). This
entails the existence of a, b, c ∈ Ap [X] such that ra + f b = τ (1) + up c. Multiplying
both sides by up we get up ra = up + up f b. The left hand side is in I, a right sided
ideal. Thus the reverse inclusion follows. ✷
Theorem 5.2 Suppose C is a cyclic code of length n over Ap = Fp2 + up Fp2 where
p is not divisible by n. Then there are unique monic polynomials F0 , F1 , F2 such
11
1 0
, F̂1 = F0 F2 , F̂2 = F0 F1 , and
that C = hF̂1 , up F̂2 i, where F0 F1 F2 = X −
0 1
|C| = p2s where s = 2 deg F1 + deg F2 .
n
1 0
1 0
n
= f1 f2 . . . fr be the unique factorization of X −
Proof: Let X −
0 1
0 1
into a product of monic basic irreducible pairwise coprime polynomials. Note that
C is a direct sum of right Ap -modules of the form (ujp fˆi ), 0 ≤ j ≤ 1, 0 ≤ i ≤ r
Q
where fˆi = nj=1,j6=i fj . After reordering, we can assume that C is a direct sum of
any of the following
n
(fˆt1 +1 ), (fˆt1 +2 ), . . . , (fˆt1 +t2 ), (up fˆt1 +t2 +1 ), . . . , (up fˆr ).
That is,
C = hf1 f2 f3 . . . ft1 ft1 +t2 +1 . . . fr , up f1 f2 f3 . . . ft1 +t2 ft1 +t2 +t3 i.
Let F̂1 = f1 f2 f3 . . . ft1 ft1 +t2 +1 . . . fr and F̂2 = f1 f2 f3 . . . ft1 +t2 ft1 +t2 +t3 . where t1 , t2 ≥
0 and t1 + t2 + 1 ≤ r.
Then
Fi =
(
1
ti+1 = 0
ft0 +t1 +···+ti +1 . . . ft0 +t1 +···+ti+1
ti+1 6= 0,
where t0 = 0 , 0 ≤ i ≤ 2.
n
Then by our construction, it is clear that C = hF̂1 , up F̂2 i and X −
F0 F1 F2 = f1 f2 . . . fr .
1 0
=
0 1
To prove uniqueness, we assume that G0 , G
1 , G2 are pairwise coprime monic
1 0
polynomials in Ap [X] such that G0 G1 G2 = X n −
and C = hĜ1 , up Ĝ2 i. Thus,
0 1
C = (Ĝ1 ) + (up Ĝ2 ). Now there exist nonnegative integers m0 = 0, m1 , . . . , md+1
with m0 + m1 + · · · + md+1 = r, and a permutation of {f1 , f2 , . . . , fr } such that
Gi = fm0 +···+mi +1 . . . fm0 +···+mi+1 for i = 0, 1, 2. Hence,
C = (fˆm1 +1 ) ⊕ · · · ⊕ (fˆm1 +m2 ) ⊕ (up fˆm1 +m2 +1 ) · · · ⊕ (up fˆr ).
It follows that mi = ti for i = 0, 1, 2. Furthermore, (fm0 +···+md +1 , . . . , fm0 +···+md+1 ) is
a permutation of {ft0 +···+td +1 , . . . , ft0 +···+td+1 }. Therefore, Fi = Gi for i = 0, 1, 2. To
calculate the order of C, note that
C = hF̂1 , up F̂2 i = (F̂1 ) ⊕ (up F̂2 ).
Hence, |C| = (p2 )2(n−degF̂1 ) (p2 )n−degF̂2 = p2s . ✷
12
6
Left
Fp-module isometry
Recall that M2 (Fp ) = Fp2 + up Fp2 and Bp = {(αa1 + αb1 ω) + u(βa1 + βb1 ω)|α ∈
F×
p , a1 , b1 , β ∈ Fp } is a subset of Fp2 + uFp2 Consider the mapping Φp defined as
Φp : M2 (Fp ) −→ Fp2 + uFp2
where
Φp
"
a
b
b − c a − b − d)
#
= (a + bω) + u(c + dω).
It is easy to show that Φp is a left Fp -module isomorphism. Now let Dp be the set
of matrices in M2 (Fp ) with entries a = αa1 , b = αb1 , c = βa1 and d = βb1 where
−1
α ∈ F×
p and a1 , b1 , β ∈ Fp then Φp (Bp \{0}) = GL(2, p). Therefore, Φp is a left
Fp - module isometry such that wB (A) = wL (Φp (A)) for all A ∈ M2 (Fp ). Thus, if
C is a cyclic code over M2 (Fp ) with minimum Bachoc distance dB (C), the image
Φp (C) is an additive cyclic code over Fp2 + uFp2 , u2 = 0 with minimum Lee distance
dL (Φp (C)) = dB (C).
7
Examples
For the following examples, MAGMA routines were created to construct cyclic codes
over M2 (Fp ) and their isometric images.
3
1 0
0 1
Example 7.1 Let p = 2 and n = 3. Then x −
= F0 F1 F2 where F0 =
1 0
0 1
1 1
, F1 =
and F2 =
. Then C1 = hF̂1 , u2 F̂2 i is cyclic code of
0 1
1 1
1 0
length 3 with |C1 | = 26 = 64, minimum normalized homogeneous distance dnhom = 2,
minimum Bachoc distance dB = 3 and minimum Hamming distance dHam = 2. The
image Φ2 (C1 ) is an additive cyclic code over F4 + uF4 of length 3, order 64, and
minimum Lee distance dL = 3.
1 0
= f1 f2 f3 f4 where f1 =
Example 7.2 Let p = 3 and n = 4. Then x −
0 1
1 0
1 0
2 1
1 2
x−
, f2 = x +
, f3 = x +
and f4 = x +
. If we
0 1
0 1
1 1
2 2
let F0 = f2 f4 , F1 = f3 and F2 = f1 then C2 = hF̂1 , u3 Fˆ2 i is a cyclic code of
length 4 of order |C2 | = 93 = 729 with minimum normalized homogeneous distance
dnhom = 27/8, minimum Bachoc distance dB = 4 and minimum Hamming distance
dHam = 3. The image Φ3 (C2 ) is an additive cyclic code over F9 + uF9 with length 4,
cardinality 729 and minimum Lee distance dL = 4.
4
13
References
[1] A. Alamadhi, H. Sboui, P. Solé and O. Yemen, “Cyclic Codes over M2 (F2 ),”
available on arXiv:1201.6533v1, 2012.
[2] C. Bachoc, “Application of coding theory to the construction of modular
lattices,” J. Combinatorial Theory, vol. 78, pp. 92-119, 1997.
[3] C. Biggs and I. White, Permutation groups and combinatorial structure,
London: Cambridge University Press, 1979.
[4] A. Bonnecaze and P. Udaya, “Cyclic codes and self-dual codes over F2 +uF2 ,”
IEEE Trans. Inf. Theory, vol. 45, no. 4, pp. 1250-1255,1999.
[5] M. Greferath and S.E. Schmidt, “Linear Codes and Rings of Matrices,” Proceedings of AAECC 13 Hawaii Springer LNCS 1719, pp. 160-169, 1999.
[6] M. Greferath and S.E. Schmidt, ” Finite-Ring Combinatorics and
MacWilliams’ Equivalence Theorem,” Journal of Combinatorial Theory, Series A, vol. 92, pp. 17-28, 2000.
[7] M. Greferath and S.E. Schmidt, “Gray isometries for finite chain rings and
a nonlinear ternary (36, 312, 15) code,” IEEE Trans. Inf. Theory, vol. 45, no.
7, pp. 2522–2524, November 2001.
[8] A. R. Hammons, Jr., P. V. Kumar, A. R. Calderbank, N. J. A. Sloane, and P.
Solé, “The Z4 -linearity of Kerdock, Preparata, Goethals and related codes,”
IEEE Trans. Inform. Theory, vol. 40, no. 2, pp. 301 - 319, January 1994.
[9] T. Honold, “ A characterization of finite Frobenius rings,” Arch. Math.
(Basel), vol. 76, pp. 406-415, 2001.
[10] T. Hungerford, Algebra (Graduate Texts in Mathematics 73). New York:
Springer-Verlag, 1974.
[11] J. Qian, L. Zhang and S. Zhu, “Cyclic codes over Fp + uFp + · · · + uk−1Fp ,”
IEICE Trans. Fundamentals, vol. E88-A, no. 3, pp. 795-797, March 2005.
[12] J. Wood, “Duality for modules over finite rings and applications to coding
theory,” Amer. J. Math, vol. 121, pp. 555-575 ,1999.
14
| 7cs.IT
|
Can Machines Truly Think?
Murat Okandan
Sandia National Laboratories, PO Box 5800 MS1080, Albuquerque, NM 87185 [email protected]
Can machines truly think? This question and its answer have many implications that depend, in large
part, on any number of assumptions underlying how the issue has been addressed or considered
previously. A crucial question, and one that is almost taken for granted, is the starting point for this
discussion: Can “thought” be achieved or emulated by algorithmic procedures?
After seventy years of intense development, improvement, and utilization as foreseen by many early
pioneers of the field such as von Neumann and Turing, let us start by considering how we use our
computers and our brains.
What do we use computers for?
-
to predict the future
How do we do that with our computers? We design algorithms, write programs based on our knowledge
of the physical phenomena we would like to investigate and predict with better accuracy, run the
programs, and interpret the results. The results have a meaning only for the human observer—
computers have no knowledge or understanding of the data. This specific limitation has existed from the
earliest calculating machines to our current supercomputers. Computers perform symbolic manipulation
on a data set based on a set of rules provided by the programmer. The Turing machine embodies that
functionality.
What are brains for?
-
to ensure survival by controlling the body and movement to capture energy and to avoid
becoming someone else’s energy source
The higher the success rate of the neural system’s primary function (i.e. survival), the higher the chances
of further development and evolution. If the neural control system embodies the learning and
prediction functions, then the need for brute force alone is lessened as the entity can position itself for
maximum benefit, i.e. survival.1 As humans and other animals learn through experience how inanimate
and animate objects behave, we develop an internal model of ourselves and the physical environment
around us that enables us to predict continuously what is going to happen next without being
consciously aware of it.2 This relationship between survival and prediction appears to be true to varying
degrees in all animals that possess a complex enough nervous system.
1
Bozek, K., et al., “Exceptional Evolutionary Divergence of Human Muscle and Brain Metabolomes Parallels Human Cognitive
and Physical Uniqueness,” PLoS Biol 12(5): e1001871. doi:10.1371/journal.pbio.1001871
2
Kahneman, D.,Thinking, Fast and Slow, NY: Farrar, Straus and Giroux, 2013.
1
How does the brain do it?
We don’t know exactly how the brain performs its functions, but we have been gathering increasingly
detailed information on the basic building blocks of the neural systems, all the way from ion channels to
neural circuits and their implications for observed behavior. We can say, however, with some degree of
certainty, that the brain is not functioning as an equation-solver based on static symbolic representation
and an overriding program. In fact, the concepts of “hardware” and “software” that we have almost
internalized due to our electronic, digital experience have no meaning in a neural system. One might be
able to describe the functionality of the brain as a biologically, genetically encoded and instantiated
control system that is driven to acquire and use energy to reduce entropy, with the overarching goal of
propagation.
Can a digital computer do what the brain does?
A digital computer can emulate some of the functionality of a neural system, but it might not be possible
to describe algorithmically the complete system needed to achieve the most valuable, higher levels of
functionality, such as internal model generation and maintenance, learning, sense making,
understanding, and intuition. A simulation approach can capture the exquisite detail that is embedded
in the neural systems; however, this comes at a huge cost in terms of energy and time required to run
those simulations. Simulating neural systems is very useful and will provide further insights into how
neural systems might be functioning, but probably it will not lead to a deployable system in a
conventional computing configuration. Earlier work on symbolic computation has provided critical
insights towards improving our understanding of how the brain and the mind might function, but this
has not lead to a system that is conveniently accessible for rapid and large scale experimentation.
A digital computer is purpose-built for doing symbolic manipulation, running programs, and simulating
phenomena described by mathematical equations. The critical question then, is this: Can the higher
levels of functionality of the neural systems be completely described by algorithms and a mathematical
framework which then can be implemented on a digital computer? Unfortunately, that answer is still
not available.
Can a different type of system do what the brain does?
It might be necessary to move to a completely new architecture to achieve the low power, high speed,
and volumetric and energy efficiency that is so tantalizingly promised by the example of biological
neural systems. Biology had to use the available substrate, i.e., messy and very needy biochemical
processes in an environment with limited energy resources. The building blocks are incredibly noisy and
inaccurate; things happen in a stochastic yet driven manner. In contrast, the substrate of our modern
computers is unimaginably precise and accurate. Modern microelectronic devices achieve levels of
precision, reliability, and complexity that were unfathomable during the early days of computers, yet we
have not been able to achieve either the level of functionality we desire or that which we have
predicted in earlier attempts at machine intelligence. This problem might be resolved by even higher
levels of performance wrought from the scaling of microelectronic devices, but without the realization
of an elegant and energy efficient architecture. Or we could entertain the idea of non-symbolic
2
information representation, processing, storage, and recall as might exist in neural systems, this time
instantiated on an extremely precise and flexible solid-state micro-opto-electronic substrate.
On Spiking, Spatio-temporal, Sparse, Hierarchical Representation of Information–Energy Efficiency
A critical debate has been centered on the spiking waveforms (“action potentials”) observed in neural
circuits. Most of the earlier neural models have relied on a rate-coding scheme, which assumes that the
individual spikes and timing of spikes are not important and the information is carried in the rates of
firing of neurons. Recently, however, the potential benefits of using the actual spatio-temporal nature of
the spiking behavior in concert with the population dynamics of the massively interconnected,
reconfigurable neural circuit have attracted more attention. Such sparse, spatio-temporal, hierarchical
representation will likely produce the ultimate benefit of higher energy efficiency.3 As discussed above,
if such a control system is able to provide higher survival chances to the entity implementing that
approach, that system will have the evolutionary advantage to progress and explore further options.
In thinking about how the higher energy efficiency is achieved by a spiking neural system, a helpful
analogy might be found in the packet-switched and circuit-switched communication networks. Packetswitched networks rely on an addressing scheme and associated network machinery that enable any
element in the network to reach any other element; this is also the underlying model for the Internet
Protocol. However, this flexibility is costly when implemented for a neural circuit—current estimates
and measurements point to several to tens of picojoules per spike processed (Heidelberg, SpiNNaker).
Such a network’s generality and maximum reconfigurability is very attractive for exploring how neural
networks might be functioning, but it might not provide the necessary energy efficiency and the
desirable form factors for next generation systems. On the other hand, a circuit-switched network relies
on dedicated lines among predetermined network elements to deliver the information, which, in this
case, is a single spike, essentially the arrival of a single transition at the receiving node. Such a network
would not have the level of reconfigurability available in a packet-switched network, but could have
much higher energy efficiency and a smaller physical footprint. An early estimate of the potential
improvement in energy efficiency is on the order of 1000x (tens of femtojoules per spike). A detailed
evaluation of this potential energy efficiency advantage is in progress and will be presented later. The
unconventional, non-symbolic nature of information representation, processing, storage, and recall in
the system described here, contrasts greatly with conventional computing systems.
How can we learn to build such a system?
Systems have been built specifically for understanding how neural systems might work. Two examples of
large-scale implementations of these neuro-inspired/neuromorphic platforms are the SpiNNaker system
(University of Manchester) and the BrainScaleS/NM-PM system (University of Heidelberg). IBM’s Escape
System is in the design phase; its new capabilities will complement the existing platforms. A promising
development path would be to use these platforms in combination with information from
3
Hawkins, J. and Blakeslee, S., On Intelligence, NY: St. Martin's Griffin, 2005.
3
neuroscience/computational neuroscience to inform theories and architectures, test (prove/disprove)
those theories and architectures, and improve these systems, theories and models and develop the next
generation systems. These first generation platforms are not intended to be the deployable systems. We
hope the next generation systems will allow us to trade off some of the neural (biological) system detail
and fidelity for lower power and reduced footprint, while achieving the desired higher levels of
performance. Current systems provide the reduced timescales and effort required to implement and
test the architecture and theory iterations. Built with those requirements in mind, they have the
necessary tools in place to run the many experiments and produce datasets that support statistically
significant analysis options.
Taking a wider view, there are essentially three ways of building such systems:
1) Software
Software is being developed by those with significant levels of interest and support in machine
learning/deep learning fields. A necessary condition here is the existence and a method for
efficient implementation of an algorithm.
2) Modified, but still digital (symbolic manipulation) systems
Many researchers are attempting to improve the current bottlenecks observed in machine
learning while keeping the inherent architecture of the system and tools intact in terms of
symbolic computation.
3) New systems that inherently implement the sparse, spatio-temporal encoding of information
as their mode of operation
This is potentially a paradigm shift in computing which is in its very early stages. The sparse,
spatio-temporal encoding could be implemented in purely digital or mixed digital-analog
systems but they all have the distinct differentiation from conventional computing approaches.
What will we do with such systems?
1) Use them to understand neural systems
A key benefit of understanding the neural systems better would be to help alleviate the
problems encountered due to ailments and deficiencies of biological/neural systems. It might
also be possible to further improve the brain-machine interfaces (neural prosthesis, etc.) with
the information gained from the neuro-inspired systems.
2) Use them as data analysis, prediction, and control systems where we are not able to use our
current computers
Many of the largest challenges in our current computing, analysis, and control systems would be
significantly changed if we could implement what is being done by humans as an integral part of
those systems.
3) Don’t ask. (Just kidding.)
4
Path Forward
An early scoping portion of our evaluation study is now determining the available platforms,
performance metrics, and evaluation methods for assessing the potential impact of neuroinspired/neuromorphic computing approaches. Ideally, the evaluation study will use the existing
platforms, data sets, and evaluation approaches to create the first “test station” whereby many of the
algorithms, architectures, and theories could be tested rapidly for remote sensing applications. A
parallel evaluation and development path will use fitness/survival simulations implemented on the
neuro-inspired/neuromorphic platforms which will be carried out in virtual environments in order to
assess more general, higher-level functions.
The next stage of the study envisions further improvements of the individual platforms and design,
implementation, and testing of new systems with the information gained from the earlier iterations.
How do we measure the value or determine the presence of higher-level functionality, as discussed
above? In spirit of the imitation game, we can offer an observation task: if a human observes the entity
in question and concludes that there is “intelligence” embedded in that entity, we might declare that we
have achieved our goal.
To summarize, we still don’t know if algorithmic approaches can achieve (1) the higher levels of
functionality we observe in neural systems, and (2) do so with high enough energy efficiency to enable
deployable systems. We might be able to achieve that desired functionality with a new approach and
substrate that implements the sparse, spatio-temporal, hierarchical information representation,
processing, storage, and recall, which could be one of the critical developments that lets us understand
how neural (biological) systems function. We are in the early stages of an evaluation study that will use
existing neuro-inspired/neuromorphic platforms to determine performance parameters and metrics,
and will guide the development of next generation data analysis, prediction, and control systems.
One last note of hope and caution: we have had as our playthings what the early pioneers of computing
could only dream of and we have used them to create and maintain a standard of living that many
would have thought impossible. Let us hope that remains to be the case for future generations.
Sandia National Laboratories is a multi-program laboratory managed and operated by Sandia
Corporation, a wholly owned subsidiary of Lockheed Martin Corporation, for the U.S. Department of
Energy’s National Nuclear Security Administration under contract DE-AC04-94AL85000.
5
| 9cs.NE
|
DeepLung: 3D Deep Convolutional Nets for
Automated Pulmonary Nodule Detection and Classification
Wentao Zhu, Chaochun Liu, Wei Fan, Xiaohui Xie
arXiv:1709.05538v1 [cs.CV] 16 Sep 2017
University of California, Irvine and Baidu Research
{wentaoz1,xhx}@ics.uci.edu and {liuchaochun,fanwei03}@baidu.com
Abstract
In this work, we present a fully automated lung CT cancer
diagnosis system, DeepLung. DeepLung contains two parts,
nodule detection and classification. Considering the 3D nature of lung CT data, two 3D networks are designed for the
nodule detection and classification respectively. Specifically,
a 3D Faster R-CNN is designed for nodule detection with a
U-net-like encoder-decoder structure to effectively learn nodule features. For nodule classification, gradient boosting machine (GBM) with 3D dual path network (DPN) features is
proposed. The nodule classification subnetwork is validated
on a public dataset from LIDC-IDRI, on which it achieves
better performance than state-of-the-art approaches, and surpasses the average performance of four experienced doctors.
For the DeepLung system, candidate nodules are detected
first by the nodule detection subnetwork, and nodule diagnosis is conducted by the classification subnetwork. Extensive
experimental results demonstrate the DeepLung is comparable to the experienced doctors both for the nodule-level and
patient-level diagnosis on the LIDC-IDRI dataset.
Introduction
Lung cancer is the most common cause of cancer-related
death in men. Low-dose lung CT screening provides an effective way for early diagnosis, which can sharply reduce the
lung cancer mortality rate. Advanced computer-aided diagnosis systems (CADs) are expected to have high sensitivities
while at the same time maintaining low false positive rates.
Recent advances in deep learning enable us to rethink the
ways of clinician lung cancer diagnosis.
Current lung CT analysis research mainly includes nodule detection (Dou et al. 2017; Ding et al. 2017), and nodule
classification (Shen et al. 2015; 2016; Hussein et al. 2017;
Yan et al. 2016). There is few work on building a complete
lung CT cancer diagnosis system for fully automated lung
CT cancer diagnosis, integrating both nodule detection and
nodule classification. It is worth exploring a whole lung CT
cancer diagnosis system and understanding how far the performance of current technology differs from that of experienced doctors. To our best knowledge, this is the first work
for a fully automated and complete lung CT cancer diagnosis system.
Copyright c 2018, Association for the Advancement of Artificial
Intelligence (www.aaai.org). All rights reserved.
The emergence of large-scale dataset, LUNA16 (Setio et
al. 2017), accelerates the nodule detection related research.
Typically, nodule detection consists of two stages, region
proposal generation and false positive reduction. Traditional
approaches generally require manually designed features
such as morphological features, voxel clustering and pixel
thresholding (Murphy et al. 2009; Jacobs et al. 2014). Recently, deep convolutional networks, such as Faster R-CNN
(Ren et al. 2015) and fully convolutional networks (Long,
Shelhamer, and Darrell 2015), are employed to generate candidate bounding boxes (Ding et al. 2017; Dou et al. 2017).
In the second stage, more advanced methods or complex features, such as more carefully designed texture features, are
designed to remove false positive nodules. Due to the 3D nature of CT data and the effectiveness of Faster R-CNN for
object detection in 2D natural images (Huang et al. 2016b),
we design a 3D Faster R-CNN for nodule detection with Unet-like encoder-decoder structure to effectively learn latent
features (Ronneberger, Fischer, and Brox 2015).
Before the era of deep learning, feature engineering followed by classifiers is a general pipeline for nodule classification (Han et al. 2013). After the public large-scale dataset,
LIDC-IDRI (Armato et al. 2011), becomes available, deep
learning based methods have become dominant for nodule classification research (Shen et al. 2016). Multi-scale
deep convolutional network with shared weights on different scales has been proposed for the nodule classification
(Shen et al. 2015). The weight sharing scheme reduces the
number of parameters and forces the multi-scale deep convolutional network to learn scale-invariant features. Inspired
by the recent success of dual path network (DPN) on ImageNet (Chen et al. 2017; Deng et al. 2009), we propose a
novel framework for CT nodule classification. First, we design a 3D deep dual path network to extract features. Then
we use gradient boosting machines (GBM) with deep 3D
dual path features, nodule size and cropped raw nodule CT
pixels for the nodule classification (Friedman 2001).
At last, we build a fully automated lung CT cancer diagnosis system, DeepLung, by combining the nodule detection
network and nodule classification network together, as illustrated in Fig. 1. For a CT image, we first use the detection
subnetwork to detect candidate nodules. Next, we employ
the classification subnetwork to classify the detected nodules
into either malignant or benign. Finally, the patient-level di-
Figure 1: The framework of DeepLung. DeepLung first employs 3D Faster R-CNN to generate candidate nodules. Then it uses
3D deep dual path network (DPN) to extract deep features from the detected and cropped nodules. Lastly, gradient boosting
machine (GBM) with deep features, detected nodule size, and raw nodule pixels are employed for classification. Patient-level
diagnosis can be achieved by fusing the classification results of detected nodules in the CT.
agnosis result can be achieved for the whole CT by fusing
the diagnosis result of each nodule.
Our main contributions are as follows: 1) To our best
knowledge, DeepLung is the first work for a complete
automated lung CT cancer diagnosis system. 2) Two 3D
deep convolutional networks are designed for nodule detection and classification. Specifically, inspired by the effective of Faster R-CNN for object detection and deep dual
path network’s success on ImageNet (Huang et al. 2016b;
Chen et al. 2017), we propose 3D Faster R-CNN for nodule
detection, and 3D deep dual path network for nodule classification. 3) Our classification framework achieves superior
performance compared with state-of-the-art approaches, and
the performance surpasses the average performance of experienced doctors on the largest public dataset, LIDC-IDRI
dataset. 4) The DeepLung system is comparable to the average performance of experienced doctors both on nodulelevel and patient-level diagnosis.
Related Work
Traditional nodule detection requires manually designed
features or descriptors (Lopez Torres et al. 2015). Recently,
several works have been proposed to use deep convolutional networks for nodule detection to automatically learn
features, which is proven to be much more effective than
hand-crafted features. Setio et al. proposes multi-view convolutional network for false positive nodule reduction (Setio et al. 2016). Due to the 3D nature of CT scans, some
work propose 3D convolutional networks to handle the challenge. The 3D fully convolutional network (FCN) is proposed to generate region candidates, and deep convolution
network with weighted sampling is used in the false positive candidates reduction stage (Dou et al. 2017). Ding
et al. uses the Faster R-CNN to generate candidate nodules, followed by 3D convolutional networks to remove
false positive nodules (Ding et al. 2017). Due to the effective performance of Faster R-CNN (Huang et al. 2016b;
Ren et al. 2015), we design a novel network, 3D Faster RCNN, for the nodule detection. Further, U-net-like encoderdecoder scheme is employed for 3D Faster R-CNN to effectively learn the features (Ronneberger, Fischer, and Brox
2015).
Nodule classification has traditionally been based on segmentation (El-Baz et al. 2011; Zhu and Xie 2016) and manual feature design (Aerts et al. 2014). Several works designed 3D contour feature, shape feature and texture feature
for CT nodule diagnosis (Way et al. 2006; El-Baz et al. 2011;
Han et al. 2013). Recently, deep networks have been shown
to be effective for medical images. Artificial neural network
was implemented for CT nodule diagnosis (Suzuki et al.
2005). More computationally effective network, multi-scale
convolutional neural network with shared weights for different scales to learn scale-invariant features, is proposed
for nodule classification (Shen et al. 2015). Deep transfer learning and multi-instance learning is used for patientlevel lung CT diagnosis (Shen et al. 2016). A comparison
on 2D and 3D convolutional networks is conducted and
shown that 3D convolutional network is superior than 2D
convolutional network for 3D CT data (Yan et al. 2016).
Further, a multi-task learning and transfer learning framework is proposed for nodule diagnosis (Hussein et al. 2017;
Zhu et al. 2017). Different from their approaches, we propose a novel classification framework for CT nodule diagnosis. Inspired by the recent success of deep dual path network (DPN) on ImageNet (Chen et al. 2017), we design a
novel totally 3D deep dual path network to extract features
from raw CT nodules. Then, we employ gradient boosting
machine (GBM) with the deep DPN features, nodule size,
and raw nodule CT pixels for the nodule diagnosis. Patientlevel diagnosis can be achieved by fusing the nodule-level
diagnosis.
DeepLung Framework
The fully automated lung CT cancer diagnosis system,
DeepLung, consists of two parts, nodule detection and nodule classification. We design a 3D Faster R-CNN for nodule
detection, and propose gradient boosting machine with deep
3D dual path network features, raw nodule CT pixels and
nodule size for nodule classification.
Figure 2: The 3D Faster R-CNN framework contains residual blocks and U-net-like encoder-decoder
structure. The model employs 3 anchors and multitask learning loss, including coordinates (x, y, z) and
diameter d regression, and candidate box classification. The numbers in boxes are feature map sizes in
the format (#slices*#rows*#cols*#maps). The numbers above the connections are in the format (#filters
#slices*#rows*#cols).
3D Faster R-CNN for Nodule Detection
The 3D Faster R-CNN with U-net-like encoder-decoder
structure is illustrated in Fig. 2. Due to the GPU memory
limitation, the input of 3D Faster R-CNN is cropped from
3D reconstructed CT images with pixel size 96×96×96. The
encoder network is derived from ResNet-18 (He et al. 2016).
Before the first max-pooling, two convolutional layers are
used to generate features. After that, four residual blocks are
employed in the encoder subnetwork. We integrate the Unet-like encoder-decoder design concept in the detection to
learn the deep networks efficiently (Ronneberger, Fischer,
and Brox 2015). In fact, for the region proposal generation,
the 3D Faster R-CNN conducts pixel-wise multi-scale learning and the U-net is validated as an effective way for pixelwise labeling. This integration makes candidate nodule generation more effective. In the decoder network, the feature
maps are processed by deconvolution layers and residual
blocks, and are subsequently concatenated with the corresponding layers in the encoder network (Zeiler et al. 2010).
Then a convolutional layer with dropout (dropout probability 0.5) is used for the second last layer. In the last layer, we
design 3 anchors, 5, 10, 20, for scale references which are
designed based on the distribution of nodule sizes. For each
anchor, there are 5 parts in the loss function, classification
loss Lcls for whether the current box is a nodule or not, regression loss Lreg for nodule coordinates x, y, z and nodule
size d.
If an anchor overlaps a ground truth bounding box with
the intersection over union (IoU) higher than 0.5, we consider it as a positive anchor (p? = 1). On the other hand, if
an anchor has IoU with all ground truth boxes less than 0.02,
we consider it as a negative anchor (p? = 0). The multi-task
loss function for the anchor i is defined as
L(pi , ti ) = λLcls (pi , p?i ) + p?i Lreg (ti , ti ? ),
(1)
where pi is the predicted probability for current anchor i
being a nodule, ti is the regression predicted relative coordinates for nodule position, which is defined as
ti = (
x − xa y − ya z − za
d
,
,
, log( )),
da
da
da
da
(2)
where (x, y, z, d) are the predicted nodule coordinates and
diameter in the original space, (xa , ya , za , da ) are the coordinates and scale for the anchor i. For ground truth nodule
position, it is defined as
x? − xa y ? − ya z ? − za
d?
t?i = (
,
,
, log( )),
(3)
da
da
da
da
where (x? , y ? , z ? , d? ) are nodule ground truth coordinates
and diameter. The λ is set as 0.5. For Lcls , we used binary
cross entropy loss function. For Lreg , we used smooth l1
regression loss function (Girshick 2015).
Gradient Boosting Machine with 3D Dual Path Net
Feature for Nodule Classification
Inspired by the success of dual path network on the ImageNet (Chen et al. 2017; Deng et al. 2009), we design a 3D
deep dual path network framework for lung CT nodule classification in Fig. 4.
Dual path connection benefits both from the advantage
of residual learning and that of dense connection (He et al.
2016; Huang et al. 2016a). The shortcut connection in residual learning is an effective way to eliminate gradient vanishing phenomenon in very deep networks. From a learned
feature weights sharing perspective, residual learning enables feature reuse, while dense connection has an advantage of keeping exploiting new features (Chen et al. 2017).
And densely connected network has fewer parameters than
residual learning because there is no need to relearn redundant feature maps. The assumption of dual path connection
is that there might exist some redundancy in the exploited
features. And dual path connection uses part of feature maps
for dense connection and part of them for residual learning.
In implementation, the dual path connection splits its feature
maps into two parts. One part, F(x)[d :], is used for residual
learning, the other part, F(x)[: d], is used for dense connection as shown in Fig. 3. Here d is the hyper-parameter for
deciding how many new features to be exploited. The dual
path connection can be formulated as
y = G([F(x)[: d], F(x)[d :] + x]),
(4)
where y is the feature map for dual path connection, G is
used as ReLU activation function, F is convolutional layer
functions, and x is the input of dual path connection block.
Dual path connection integrates the advantages of the two
advanced frameworks, residual learning for feature reuse
and dense connection for keeping exploiting new features,
into a unified structure, which obtains success on the ImageNet dataset.
For CT data, advanced method should be effective to extract 3D volume feature (Yan et al. 2016). We design a 3D
deep dual path network for the 3D CT lung nodule classification in Fig. 4. We firstly crop CT data centered at predicted
nodule locations with size 32 × 32 × 32. After that, a convolutional layer is used to extract features. Then 4 dual path
blocks are employed to learn higher level features. Lastly,
the 3D average pooling and binary logistic regression layer
are used for benign or malignant diagnosis.
The 3D deep dual path network can be used as a classifier for nodule diagnosis directly. And it can also be employed to learn effective features. We construct feature by
Figure 3: Illustration of dual path connection (Chen et al.
2017), which benefits both from the advantage of residual learning (He et al. 2016) and that of dense connection
(Huang et al. 2016a) from network structure design intrinsically.
Figure 4: The 3D deep dual path network framework in the
nodule classification subnetwork, which contains 4 dual path
connection blocks. After the training, the deep 3D dual path
network feature is extracted for gradient boosting machine
to do nodule diagnosis. The numbers in the figure are of the
same formats as Fig. 2.
concatenating the learned 3D deep dual path network features (the second last layer, 2,560 dimension), nodule size,
and raw 3D cropped nodule pixels. Given complete and effective features, gradient boosting machine (GBM) is a superior method to build an advanced classifier from these
features (Friedman 2001). We validate the feature combining nodule size with raw 3D cropped nodule pixels, employ
GBM as a classifier, and obtain 86.12% test accuracy averagely. Lastly, we employ GBM with the constructed feature
and achieve the best diagnosis performance.
DeepLung System: Fully Automated Lung CT
Cancer Diagnosis
The DeepLung system includes the nodule detection using
the 3D Faster R-CNN, and nodule classification using gradient boosting machine (GBM) with constructed feature (deep
dual path features, nodule size and raw nodule CT pixels) in
Fig. 1.
Due to the GPU memory limitation, we first split the
whole CT into several 96 × 96 × 96 patches, process them
through the detector, and combine the detected results together. We only keep the detected boxes of detection probabilities larger than 0.12 (threshold as -2 before sigmoid
function). After that, non-maximum suppression (NMS) is
adopted based on detection probability with the intersection
over union (IoU) threshold as 0.1. Here we expect to not
miss too many ground truth nodules.
After we get the detected nodules, we crop the nodule
with the center as the detected center and size as 32×32×32.
The detected nodule size is kept for the classification model
as a part of features. The 3D deep dual path network is employed to extract features. We use the GBM and construct
features to conduct diagnosis for the detected nodules. For
CT pixel feature, we use the cropped size as 16 × 16 × 16
and center as the detected nodule center in the experiments.
For patient-level diagnosis, if one of the detected nodules is
positive (cancer), the patient is a cancer patient, and if all
the detected nodules are negative, the patient is a negative
(non-cancer) patient.
Experiments
We conduct extensive experiments to validate the DeepLung
system. We perform 10-fold cross validation using the detector on LUNA16 dataset. For nodule classification, we use the
LIDC-IDRI annotation, and employ the LUNA16’s patientlevel dataset split. Finally, we also validate the whole system
based on the detected nodules both on patient-level diagnosis and nodule-level diagnosis.
In the training, for each model, we use 150 epochs in total
with stochastic gradient descent optimization and momentum as 0.9. The used batch size is relied on the GPU memory. We use weight decay as 1 × 10−4 . The initial learning
rate is 0.01, 0.001 at the half of training, and 0.0001 after the
epoch 120.
Datasets
LUNA16 dataset is a subset of the largest public dataset for
pulmonary nodules, the LIDC-IDRI dataset (Armato et al.
2011). LUNA16 dataset only has the detection annotations,
while LIDC-IDRI contains almost all the related information for low-dose lung CTs including several doctors’ annotations on nodule sizes, locations, diagnosis results, nodule texture, nodule margin and other informations. LUNA16
dataset removes CTs with slice thickness greater than 3mm,
slice spacing inconsistent or missing slices from LIDC-IDRI
dataset, and explicitly gives the patient-level 10-fold cross
validation split of the dataset. LUNA16 dataset contains 888
low-dose lung CTs, and LIDC-IDRI contains 1,018 lowdose lung CTs. Note that LUNA16 dataset removes the annotated nodules of size smaller than 3mm.
For nodule classification, we extract nodule annotations
from LIDC-IDRI dataset, find the mapping of different doctors’ nodule annotations with the LUNA16’s nodule annotations, and get the ground truth of nodule diagnosis by taking
different doctors’ diagnosis averagely (Do not count the 0
score for diagnosis, which means N/A.). If the final average
score is equal to 3 (uncertain about malignant or benign),
we remove the nodule. For the nodules with score greater
than 3, we label them as positive. Otherwise, we label them
as negative. For doctors’ annotations, we only keep those of
four doctors who label most of the nodules for comparison.
We only keep the CTs within LUNA16 dataset, and use the
same cross validation split as LUNA16 for classification.
Preprocessing
We firstly clip the raw data into [−1200, 600]. Secondly, we
transform the range linearly into [0, 1]. Thirdly, we use the
Table 1: Nodule classification comparisons on LIDC-IDRI
dataset.
Nodule Size
All feat.
3D DPN
m-CNN
Models
+Pixel+GBM +GBM
Acc. (%)
88.74
86.12
90.44
86.84
Table 2: Nodule-level diagnosis accuracy (%) between nodule classification subnetwork in DeepLung and experienced
doctors on doctor’s individually confident nodules.
Dr 1
Dr 2
Dr 3
Dr 4 Average
Doctors
93.44 93.69 91.82 86.03
91.25
DeepLung 93.55 93.30 93.19 90.89
92.74
Figure 5: Sensitivity (Recall) rate with respect to false positives per scan. The FROC (average recall rate at the false
positives as 0.125, 0.25, 0.5, 1, 2, 4, 8) is 83.4%. The 3D
Faster R-CNN has a recall rate 94.6% for all the nodules.
The dash lines are lower bound and upper bound FROC
for 95% confidence interval using bootstrapping with 1,000
bootstraps (Setio et al. 2017). The solid line is the interpolated FROC based on prediction.
LUNA16’s given segmentation ground truth and remove the
useless background. It is worth visualizing the processed
data during processing because some segmentations are imperfect, and need to be removed. After that, the data can be
processed by the deep networks.
DeepLung for Nodule Detection
We train and evaluate the detector on LUNA16 dataset
following 10-fold cross validation with given patient-level
split. In the training, we use the flipping, randomly scale
from 0.75 to 1.25 for the cropped patches to augment the
data. The evaluation metric, FROC, is the average recall rate
at the average number of false positives as 0.125, 0.25, 0.5,
1, 2, 4, 8 per scan, which is the official evaluation metric for
LUNA16 dataset.
The FROC performance on LUNA16 is visualized in Fig.
5. The solid line is interpolated FROC based on true prediction, and the dash lines are upper bound and lower bound for
the bootstrapped FROC performance. The 3D Faster R-CNN
achieves 83.4% FROC without any false positive nodule reduction stage, which is comparable with 83.9% for the best
method, which uses two-stage training (Dou et al. 2017).
DeepLung for Nodule Classification
We validate the nodule classification performance of the
DeepLung system on the LIDC-IDRI dataset with the
LUNA16’s split principle, 10-fold patient-level cross validation. There are 1,004 nodules left and 450 nodules are
positive. In the training, we firstly pad the nodules of size
32 × 32 × 32 into 36 × 36 × 36, randomly crop 32 × 32 × 32
from the padded data, horizontal flip, vertical flip, z-axis flip
the data for augmentation, randomly set 4 × 4 × 4 patch as
0, and normalize the data with the mean and std. obtained
from training data. The total number of epochs is 1,050. The
learning rate is 0.01 at first, then became 0.001 after epoch
525, and turned into 0.0001 after epoch 840. Due to time
and resource limitation for training, we use the fold 1, 2, 3,
4, 5 for test, and the final performance is the average performance on the five test folds. The nodule classification performance is concluded in Table 1.
From the table 1, our 3D deep dual path network (DPN)
achieves better performance than that of multi-scale CNN
(Shen et al. 2015) because of the power of 3D structure and
deep dual path network. Because of the power of gradient
boosting machine (GBM), GBM with nodule size and raw
nodule pixels with crop size as 16 × 16 × 16 achieves comparable performance as multi-scale CNN (Shen et al. 2015).
Finally, we construct feature with 3D deep dual path network features, nodule size and raw nodule pixels, and obtain
90.44% accuracy.
Compared with Experienced Doctors on Their Individually Confident Nodules We compare our predictions with
those of four experienced doctors on their individually confident nodules (with individual score not as 3). Note that about
1/3 nodules are labeled as 3 for each doctor. Comparison results are concluded in Table 2.
From Table 2, these doctors’ confident nodules are easy
to be diagnosed nodules from the performance comparison
between our model’s performances in Table 1 and Table 2.
To our surprise, the average performance of our model is
1.5% better than that of experienced doctors. In fact, our
model’s performance is better than 3 out of 4 doctors (doctor
1, 3, 4) on the nodule diagnosis task. The result validates
deep network surpasses human-level performance for image
classification (He et al. 2016), and the DeepLung is better
suited for nodule diagnosis than experienced doctors.
DeepLung for Fully Automated Lung CT Cancer
Diagnosis
We also validate the DeepLung for fully automated lung CT
cancer diagnosis on the LIDC-IDRI dataset with the same
protocol as LUNA16’s patient-level split. Firstly, we employ
the 3D Faster R-CNN to detect suspicious nodules. Then we
retrain the model from nodule classification model on the
detected nodules dataset. If the center of detected nodule is
Table 3: Comparison between DeepLung’s nodule classification on all detected nodules and doctors on all nodules.
Method TP Set FP Set
Doctors
Acc. (%) 81.42
97.02 74.05-82.67
Table 4: Patient-level diagnosis accuracy(%) between
DeepLung and experienced doctors on doctor’s individually
confident CTs.
Dr 1
Dr 2
Dr 3
Dr 4 Average
Doctors
83.03 85.65 82.75 77.80
82.31
DeepLung 81.82 80.69 78.86 84.28
81.41
within the ground truth positive nodule, it is a positive nodule. Otherwise, it is a negative nodule. Through this mapping from the detected nodule and ground truth nodule, we
can evaluate the performance and compare it with the performance of experienced doctors. We adopt the test fold 1,
2, 3, 4, 5 to validate the performance the same as that for
nodule classification.
Different from pure nodule classification, the fully automated lung CT nodule diagnosis relies on nodule detection.
We evaluate the performance of DeepLung on the detection
true positive (TP) set and detection false positive (FP) set individually in Table 3. If the detected nodule of center within
one of ground truth nodule regions, it is in the TP set. If the
detected nodule of center out of any ground truth nodule regions, it is in FP set. From Table 3, the DeepLung system using detected nodule region obtains 81.42% accuracy for all
the detected TP nodules. Note that the experienced doctors
obtain 78.36% accuracy for all the nodule diagnosis on average. The DeepLung system with fully automated lung CT
nodule diagnosis still achieves above average performance
of experienced doctors. On the FP set, our nodule classification subnetwork in the DeepLung can reduce 97.02% FP
detected nodules, which guarantees that our fully automated
system is effective for the lung CT cancer diagnosis.
Compared with Experienced Doctors on Their Individually Confident CTs We employ the DeepLung for patientlevel diagnosis further. If the current CT has one nodule that
is classified as positive, the diagnosis of the CT is positive.
If all the nodules are classified as negative for the CT, the
diagnosis of the CT is negative. We evaluate the DeepLung
on the doctors’ individually confident CTs for benchmark
comparison in Table 4.
From Table 4, DeepLung achieves 81.41% patient-level
diagnosis accuracy. The performance is 99% of the average performance of four experienced doctors, and the performance of DeepLung is better than that of doctor 4. Thus
DeepLung can be used to help improve some doctors’ performance, like that of doctor 4, which is the goal for computer aided diagnosis system. We also employ Kappa coefficient, which is a common approach to evaluate the agreement between two raters, to test the agreement between
DeepLung and the ground truth patient-level labels (Landis
and Koch 1977). For comparison, we valid those of 4 individual doctors on their individual confident CTs. The Kappa
coefficient of DeepLung is 63.02%, while the average Kappa
coefficient of doctors is 64.46%. It shows the predictions
of DeepLung are of good agreement with ground truths for
patient-level diagnosis, and are comparable with those of experienced doctors.
Discussion
In this section, we are trying to explain the DeepLung by
visualizing the nodule detection and classification results.
Nodule Detection
We randomly pick nodules from test fold 1 and visualize
them in the first row and third row in Fig. 6. Detected nodules are visualized in the second, fourth and fifth row. The
numbers below detected nodules are with the format (ground
truth slice number-detected slice number-detection probability). The fifth row with red detection probabilies are the false
positive detections. Here we only visualize the slice of center z. The yellow circles are the ground truth nodules or detected nodules respectively. The center of the circle is the
nodule center and the diameter of the circle is relative to the
nodule size.
From the first four rows in Fig. 6, we observe the 3D
Faster R-CNN works well for the nodules from test fold 1.
The detection probability is also very high for these nodules.
From the detected false positive nodules in the last row, we
can find these false positives are suspicious to nodules visually if we observe them carefully. And the sizes of these
false positive nodules are very small. Note that the nodules
less than 3mm are removed in the ground truth annotations.
Nodule Classification
We also visualize the nodule classification results from test
fold 1 in Fig. 7. The nodules in the first four rows are those
the DeepLung predicted right, but some doctors predicted
wrong. The first number is the DeepLung predicted malignant probability. The second number is which doctor predicted wrong. If the probability is large than 0.5, it is a malignant nodule. Otherwise, it is a benign nodule. The rest
nodules are those the DeepLung predicted wrong with red
numbers indicating the DeepLung predicted malignant probability. For an experienced doctor, if a nodule is big and has
irregular shape, it has a high probability to be a malignant
nodule.
From the first 4 rows in Fig. 7, we can observe that doctors mis-diagnose some nodules. The reason is that, humans
are not good at processing 3D CT data, which is of low signal to noise ratio. In fact, even for high quality 2D natural
image, the performance of deep network surpasses that of
humans (He et al. 2016). From the doctors’ individual experience point of view, some doctors intend to be optimistic to
some nodules, and some doctors intend to be pessimistic to
them. And they can just observe one slice each time. Some
irregular boundaries are vague. The machine learning based
methods can learn these complicated rules and high dimensional features from these doctors’ annotations, and avoid
such bias.
Figure 6: Visualization of central slices for nodule ground
truths and detection results. We randomly choose nodules
(the first row and third row) from test fold 1. Detection results are shown in the second row and fourth row. The numbers below detected nodules are with the format (ground
truth slice number-detected slice number-detection probability). The last row shows the false positive detected nodules
and detection probabilities. Note that the ground truth nodules are of diameter greater than 3mm. The DeepLung performs well for detection.
The rest nodules with red numbers are the nodules the
DeepLung predicted wrong for test fold 1 in Fig. 7. If the
predicted malignant probability is large than 0.5, the ground
truth label is benign. Otherwise, the ground truth label is
malignant. From the central slices, we can hardly decide the
DeepLung’s predictions are wrong. Maybe the DeepLung
cannot find some weak irregular boundaries as for the bottom right 4 nodules, which is the possible reason why the
DeepLung predicts them as benign. From the above analysis, the DeepLung can be considered as a tool to assist the
diagnosis for doctors. Combing the DeepLung and doctor’s
own diagnosis is an effective way to improve diagnosis accuracy.
Conclusion
In this work, we propose a fully automated lung CT cancer diagnosis system, DeepLung. DeepLung consists two
parts, nodule detection and classification. For nodule detec-
Figure 7: Visualization of central slices for nodule classification results on test fold 1. We choose nodules that is
predicted right, but annotated wrong by some doctors in the
first 4 rows. The numbers below the nodules are (model predicted malignant probability-which doctor is wrong). The
last 2 rows show the nodules our model predicted wrong.
The red number is the wrongly predicted malignant probabilities (> 0.5, predict malignant; < 0.5, predict benign).
The DeepLung avoids doctors’ individual bias and achieves
better performance than the average performance of doctors.
tion, we design a 3D Faster R-CNN with U-net-like encoderdecoder structure to detect suspicious nodules. Then we input the detected nodules into the nodule classification network. The 3D deep dual path network is designed to extract
features. Further, gradient boosting machine with different
features combined together to achieve the state-of-the-art
performance, which surpasses that of experienced doctors.
Extensive experimental results on public available largescale datasets, LUNA16 and LIDC-IDRI datasets, demonstrate the superior performance of the DeepLung.
In the future, we will develop more advanced nodule detection method to further boost the performance of
DeepLung. Further, a unified framework for automated lung
CT cancer diagnosis is expected for nodule detection and
classification. Integrating multi-modality data, such as electronic medical report history, into the DeepLung is another
direction to enhance the diagnosis performance.
References
Aerts, H. J.; Velazquez, E. R.; Leijenaar, R. T.; Parmar, C.; Grossmann, P.; Cavalho, S.; Bussink, J.; Monshouwer, R.; Haibe-Kains,
B.; Rietveld, D.; et al. 2014. Decoding tumour phenotype by noninvasive imaging using a quantitative radiomics approach. Nature
communications 5.
Armato, S. G.; McLennan, G.; Bidaut, L.; McNitt-Gray, M. F.;
Meyer, C. R.; Reeves, A. P.; Zhao, B.; Aberle, D. R.; Henschke,
C. I.; Hoffman, E. A.; et al. 2011. The lung image database consortium (lidc) and image database resource initiative (idri): a completed reference database of lung nodules on ct scans. Medical
physics 38(2):915–931.
Chen, Y.; Li, J.; Xiao, H.; Jin, X.; Yan, S.; and Feng, J. 2017. Dual
path networks. arXiv preprint arXiv:1707.01629.
Deng, J.; Dong, W.; Socher, R.; Li, L.-J.; Li, K.; and Fei-Fei, L.
2009. Imagenet: A large-scale hierarchical image database. In
CVPR, 248–255. IEEE.
Ding, J.; Li, A.; Hu, Z.; and Wang, L. 2017. Accurate pulmonary
nodule detection in computed tomography images using deep convolutional neural networks. arXiv preprint arXiv:1706.04303.
Dou, Q.; Chen, H.; Jin, Y.; Lin, H.; Qin, J.; and Heng, P.-A.
2017. Automated pulmonary nodule detection via 3d convnets with
online sample filtering and hybrid-loss residual learning. arXiv
preprint arXiv:1708.03867.
El-Baz, A.; Nitzken, M.; Khalifa, F.; Elnakib, A.; Gimelfarb, G.;
Falk, R.; and El-Ghar, M. A. 2011. 3d shape analysis for early diagnosis of malignant lung nodules. In Biennial International Conference on Information Processing in Medical Imaging, 772–783.
Springer.
Friedman, J. H. 2001. Greedy function approximation: a gradient
boosting machine. Annals of statistics 1189–1232.
Girshick, R. 2015. Fast r-cnn. In ICCV, 1440–1448.
Han, F.; Zhang, G.; Wang, H.; Song, B.; Lu, H.; Zhao, D.; Zhao,
H.; and Liang, Z. 2013. A texture feature analysis for diagnosis
of pulmonary nodules using lidc-idri database. In Medical Imaging Physics and Engineering (ICMIPE), 2013 IEEE International
Conference on, 14–18. IEEE.
He, K.; Zhang, X.; Ren, S.; and Sun, J. 2016. Deep residual learning for image recognition. In CVPR, 770–778.
Huang, G.; Liu, Z.; Weinberger, K. Q.; and van der Maaten, L.
2016a. Densely connected convolutional networks. arXiv preprint
arXiv:1608.06993.
Huang, J.; Rathod, V.; Sun, C.; Zhu, M.; Korattikara, A.; Fathi,
A.; Fischer, I.; Wojna, Z.; Song, Y.; Guadarrama, S.; et al. 2016b.
Speed/accuracy trade-offs for modern convolutional object detectors. arXiv preprint arXiv:1611.10012.
Hussein, S.; Cao, K.; Song, Q.; and Bagci, U. 2017. Risk stratification of lung nodules using 3d cnn-based multi-task learning. In
IPMI, 249–260. Springer.
Jacobs, C.; van Rikxoort, E. M.; Twellmann, T.; Scholten, E. T.;
de Jong, P. A.; Kuhnigk, J.-M.; Oudkerk, M.; de Koning, H. J.;
Prokop, M.; Schaefer-Prokop, C.; et al. 2014. Automatic detection
of subsolid pulmonary nodules in thoracic computed tomography
images. Medical image analysis 18(2):374–384.
Landis, J. R., and Koch, G. G. 1977. The measurement of observer
agreement for categorical data. biometrics 159–174.
Long, J.; Shelhamer, E.; and Darrell, T. 2015. Fully convolutional
networks for semantic segmentation. In CVPR, 3431–3440.
Lopez Torres, E.; Fiorina, E.; Pennazio, F.; Peroni, C.; Saletta, M.;
Camarlinghi, N.; Fantacci, M.; and Cerello, P. 2015. Large scale
validation of the m5l lung cad on heterogeneous ct datasets. Medical physics 42(4):1477–1489.
Murphy, K.; van Ginneken, B.; Schilham, A. M.; De Hoop, B.;
Gietema, H.; and Prokop, M. 2009. A large-scale evaluation of
automatic pulmonary nodule detection in chest ct using local image features and k-nearest-neighbour classification. Medical image
analysis 13(5):757–770.
Ren, S.; He, K.; Girshick, R.; and Sun, J. 2015. Faster r-cnn:
Towards real-time object detection with region proposal networks.
In NIPS, 91–99.
Ronneberger, O.; Fischer, P.; and Brox, T. 2015. U-net: Convolutional networks for biomedical image segmentation. In MICCAI,
234–241. Springer.
Setio, A. A. A.; Ciompi, F.; Litjens, G.; Gerke, P.; Jacobs, C.; van
Riel, S. J.; Wille, M. M. W.; Naqibullah, M.; Sánchez, C. I.; and
van Ginneken, B. 2016. Pulmonary nodule detection in ct images:
false positive reduction using multi-view convolutional networks.
IEEE transactions on medical imaging 35(5):1160–1169.
Setio, A. A. A.; Traverso, A.; de Bel, T.; Berens, M. S.; van den Bogaard, C.; Cerello, P.; Chen, H.; Dou, Q.; Fantacci, M. E.; Geurts,
B.; et al. 2017. Validation, comparison, and combination of algorithms for automatic detection of pulmonary nodules in computed
tomography images: the luna16 challenge. Medical Image Analysis
42:1–13.
Shen, W.; Zhou, M.; Yang, F.; Yang, C.; and Tian, J. 2015. Multiscale convolutional neural networks for lung nodule classification.
In IPMI, 588–599. Springer.
Shen, W.; Zhou, M.; Yang, F.; Dong, D.; Yang, C.; Zang, Y.; and
Tian, J. 2016. Learning from experts: Developing transferable
deep features for patient-level lung cancer prediction. In MICCAI,
124–131. Springer.
Suzuki, K.; Li, F.; Sone, S.; and Doi, K. 2005. Computer-aided
diagnostic scheme for distinction between benign and malignant
nodules in thoracic low-dose ct by use of massive training artificial neural network. IEEE Transactions on Medical Imaging
24(9):1138–1150.
Way, T. W.; Hadjiiski, L. M.; Sahiner, B.; Chan, H.-P.; Cascade,
P. N.; Kazerooni, E. A.; Bogot, N.; and Zhou, C. 2006. Computeraided diagnosis of pulmonary nodules on ct scans: Segmentation and classification using 3d active contours. Medical Physics
33(7):2323–2337.
Yan, X.; Pang, J.; Qi, H.; Zhu, Y.; Bai, C.; Geng, X.; Liu, M.; Terzopoulos, D.; and Ding, X. 2016. Classification of lung nodule
malignancy risk on computed tomography images using convolutional neural network: A comparison between 2d and 3d strategies.
In ACCV, 91–101. Springer.
Zeiler, M. D.; Krishnan, D.; Taylor, G. W.; and Fergus, R. 2010.
Deconvolutional networks. In CVPR, 2528–2535. IEEE.
Zhu, W., and Xie, X. 2016. Adversarial deep structural networks for mammographic mass segmentation. arXiv preprint
arXiv:1612.05970.
Zhu, W.; Lou, Q.; Vang, Y. S.; and Xie, X. 2017. Deep multiinstance networks with sparse label assignment for whole mammogram classification. In MICCAI, 603–611. Springer.
| 1cs.CV
|
arXiv:1109.4593v2 [math.AC] 1 Aug 2012
HILBERT DEPTH OF GRADED MODULES OVER
POLYNOMIAL RINGS IN TWO VARIABLES
JULIO JOSÉ MOYANO-FERNÁNDEZ AND JAN ULICZKA
Abstract. In this article we mainly consider the positively Z–
graded polynomial ring R = F[X, Y ] over an arbitrary field F
and Hilbert series of finitely generated graded R–modules. The
central result is an arithmetic criterion for such a series to be the
Hilbert series of some R–module of positive depth. In the generic
case, that is deg(X) and deg(Y ) being coprime, this criterion can
be formulated in terms of the numerical semigroup generated by
those degrees.
1. Introduction and Review
We want to investigate how some of the results of [1] for Hilbert series
of finitely generated graded modules over the standard Z–graded polynomial ring can be generalised to the case where the ring of polynomials
is endowed with an arbitrary positive Z–grading.
Let R = F[X1 , . . . , Xn ] be the positively Z–graded polynomial ring
over some field F, i. e. each Xi has degree di ≥ 1 for every i = 1, . . . , n.
Moreover, let M be a finitely generated graded R–module. Every homogeneous component of M is a finite–dimensional F–vector space,
and since R is positively graded and M is finitely generated, Mj = 0
for j ≪ 0. Hence the Hilbert function of M
H(M, −) : Z → Z, j 7→ dimF (Mj ) ,
1991 Mathematics Subject Classification. Primary 13D40; Secondary 16XW50,
20M99.
Key words and phrases. Commutative graded ring, Hilbert series, numerical
semigroup, finitely generated module, Hilbert depth.
The first author was partially supported by the Spanish Government Ministerio
de Educación y Ciencia (MEC) grant MTM2007-64704 in cooperation with the
European Union in the framework of the founds “FEDER”, and by the Deutsche
Forschungsgemeinschaft (DFG).
1
HILBERT DEPTH OF GRADED MODULES OVER POLYNOMIAL RINGS
2
is a well–defined integer Laurent function (see [2], Definitions 5.1.1 and
5.1.12). The formal Laurent series associated to H(M, −)
X
X
HM (t) =
H(M, j)tj =
(dimF Mj ) tj ∈ Z[[t]][t−1 ]
j∈Z
j∈Z
is called the Hilbert series of M. Obviously it has no negative coefficients; such a series will be called nonnegative for short.
By the Theorem of Hilbert–Serre (see [3, Thm. 4.1.1]), HM may be
written as a fraction of the form
Q (t)
Qn M
,
di
i=1 (1 − t )
with some QM ∈ Z[t, t−1 ]. As a consequence of this theorem and a well–
known result in the theory of generating functions, see Proposition 4.4.1
of [4], there exists a quasi–polynomial P of period d := lcm(d1 , . . . , dn )
such that dimF Mj = P (j) for j ≫ 0.
The ring R is ∗ local, that is, it has a unique maximal graded ideal,
namely m := (X1 , . . . , Xn ). The depth of M is defined as the maximal
length of an M–regular sequence in m, i. e. the grade of m on M, and
denoted by depth(M) rather than grade(m, M). This deviation from
the standard terminology, where “depth” is used exclusively in the context of true local rings, may be justified by the fact that grade(m, M)
agrees with depth (Mm ), see [5, Prop. 1.5.15].
It is easy to see that (contrary to the Krull dimension) the depth of
a module M is not encoded in its Hilbert series. Therefore it makes
sense to introduce
There is a f. g. gr. R–module N
;
Hdep(M) := max r ∈ N
with HN = HM and depth(N) = r.
this number is called the Hilbert depth of M.
If the ring R is standard graded, then Hdep(M) turns out to coincide
with the arithmetical invariant
p(M) := max {r ∈ N | (1 − t)r HM (t) is nonnegative} ,
called the positivity of M, see Theorem 3.2 of [1]. The inequality
Hdep(M) ≤ p(M) follows from general results on Hilbert series and
regular sequences. The converse can be deduced from the main result
of [1], Theorem 2.1, which states the existence of a representation
dim(M )
HM (t) =
X
j=0
Qj (t)
with nonnegative Qj ∈ Z[t, t−1 ].
(1 − t)j
HILBERT DEPTH OF GRADED MODULES OVER POLYNOMIAL RINGS
3
We begin our investigation by establishing a similar decomposition theorem for Hilbert series of modules over any positively Z–graded polynomial ring. This result has some consequences for the Hilbert depth,
but it does not lead to an analogue of the equation Hdep(M) = p(M)
– the occurence of different factors in the denominator of HM complicating matters. In section 3 we restrict our attention to polynomial
rings in two variables. For this special case we deduce an arithmetic
characterisation of positive Hilbert depth. This criterion, surprinsingly
related to the theory of numerical semigroups, is the main result of our
paper.
2. Preliminary results
2.1. A decomposition theorem. Let F be a field. We consider the
polynomial ring R = F[X1 , . . . , Xn ], endowed with a general positive
Z–graded structure, i. e. deg(Xi ) = di ≥ 1 for every i = 1, . . . , n, and
let M be a finitely generated graded R–module. The Hilbert series of
M admits a decomposition analogous to that in the standard graded
case (cf. [1, Thm. 2.1]). This can be proven using certain filtrations,
similar to the argument in the proof of [6, Prop. 2.13]:
Theorem 2.1. Let M be a finitely generated graded module over the
positively graded ring of polynomials R. The Hilbert series of M can
be written in the form
X
QI (t)
Q
HM (t) =
(1)
di
i∈I (1 − t )
I⊆{1,...,n}
with nonnegative QI ∈ Z[t, t−1 ].
Proof. We use induction on n. For n = 0, i. e. R = F, the module M
is just a finite–dimensional graded vector space, and hence HM itself
is a Laurent polynomial. Assuming the claim to be true for modules
over the polynomial ring in at most n − 1 indeterminates, we consider
a finitely generated module M over R = F[X1 , . . . , Xn ] and define a
descending sequence of submodules Ui of M by Un+1 := M and
Ui := m ∈ Ui+1 | Xij m = 0 for some j > 0
(2)
for i = n, . . . , 1. Then for each i we have a short exact sequence
0 −→ Ui −→ Ui+1 −→ Ui+1 /Ui −→ 0
| {z }
=:Ni
and therefore
HUi+1 = HUi + HNi .
HILBERT DEPTH OF GRADED MODULES OVER POLYNOMIAL RINGS
4
Combining these equations yields
HM = HU 1 +
n
X
H Ni .
(3)
i=1
Among all these Hilbert series, the first one is harmless, because it is
easy to see that U1 coincides with the local cohomology Hm0 (M), and
so it has finite length, see [5, Prop. 3.5.4]. Therefore it is enough to
show that each series HNi admits a decomposition of the form (1). By
construction, Xi is not a zerodivisor on Ni , thus we have further exact
sequences
·X
i
0 −→ Ni (−di ) −→
Ni −→ Ni /Xi Ni −→ 0,
| {z }
=:Vi
and it follows that
HVi (t) = 1 − tdi HNi (t).
Since Xi acts trivially on Vi , we may regard Vi as a module over the
ci , . . . , Xn ] in n − 1 indeterminates. By
ring of polynomials F[X1 , . . . , X
the induction hypothesis we can write the corresponding Hilbert series
in the form
X
QI (t)
Q
.
HVi (t) =
dj
j∈I (1 − t )
I⊆{1,...,ı̂,...,n}
di
Division by 1 − t yields a presentation of the required form for the
Hilbert series of Ni .
A formal Laurent series admitting a decomposition of the form (1) will
be called (d1 , . . . , dn )–decomposable. Note that such a decomposition is
by no means unique. We define an important invariant of such a series:
H admits a decomp. of form (1)
ν(H) := max r ∈ N
(4)
with QI = 0 for all I with |I| < r.
It is easily seen that any (d1 , . . . , dn )–decomposable series H is in fact
the Hilbert series of some finitely
R–module: Choose a dePqgenerated
I
k
composition of H with QI (t) = k=pI hI,k t , and write JI for the ideal
generated by the Xi with i ∈
/ I, then the R–module
!
qI
hI,k
M
M
(5)
(R/JI ) (−k)
N :=
I⊆{1,...,n}
k=pI
has Hilbert series H. Hence we have shown the following.
HILBERT DEPTH OF GRADED MODULES OVER POLYNOMIAL RINGS
5
Corollary 2.2. A formal Laurent series H ∈ Z[[t]][t−1 ] is the Hilbert
series of a finitely generated graded module over the ring F[X1 , . . . , Xn ]
with deg(Xi ) = di if and only if it is (d1 , . . . , dn )–decomposable.
Remark 2.3. Note that this result is not the complete analogue of Corollary 2.3 in [1], since it remains open whether any nonnegative series of
is (d1 , . . . , dn )–decomposable.
the form Qn Q(t)
di
i=1 (1−t )
2.2. Hilbert Depth and Positivity. Let R = F[X1 , . . . , Xn ] be positively Z–graded with deg(Xi ) = di ≥ 1, and let d := lcm(d1 , . . . , dn ).
Moreover, let M be a finitely generated graded R-module with Hilbert
series HM . As in [1] we define the Hilbert depth of M by
Hdep(M) := max r ∈ N
There exists a f. g. gr. R–module N
with HN = HM and depth(N) = r.
.
On the other hand, the notion of positivity has to be adjusted to our
new situation, since in general there is no element of degree 1 and a
fortiori no M–regular sequence consisting of such elements, hence one
cannot expect a relationship between p(M) and Hdep(M). Instead of
p(M) we consider
r
pd (M) := max r ∈ N | 1 − td HM (t) is nonnegative .
This is an upper bound for Hdep(M) for the same reason as p(M) is in
the standard graded case: Since extension of the ground field does not
affect either the depth of a module or its Hilbert series, we may assume
that F is infinite, and in this case a maximal M–regular sequence can
be
S composed of elements of degree d. (This can be seen by considering
p∈Ass(M ) p ⊆ m in degree d: Since a vector space over an infinite field
cannot be the union of proper subspaces, equality means that m agrees
with some p ∈ Ass(M) in degree d, and this implies depth(M) = 0).
It remains to check whether we still have equality. An inspection of the
proof in the standard graded case shows that it is advisable to consider
a third invariant, namely
ν(M) := ν(HM ),
which is well–defined by Theorem 2.1. Note that in [6] this number is
called the Hilbert depth.
For any decomposition
of the form (1) the R–module N given in (5)
has depth(N) = min |I| QI =
6 0 . Therefore we have an inequality
ν(M) ≤ Hdep(M) ≤ pd (M).
HILBERT DEPTH OF GRADED MODULES OVER POLYNOMIAL RINGS
6
In the standard graded case, [1, Thm. 2.1] also yields p(M) ≤ ν(M) and
hence the equality of all three numbers, as already mentioned above.
This reasoning cannot be carried over to the general situation, since
the denominators in (1) are different from 1 − td .
2.3. The case Hdep(M) = 1. The method used in the proof of Theorem 2.1 also yields Hdep(M) = ν(M) in the special case Hdep(M) = 1.
Proposition 2.4. Let M be a finitely generated graded R-module with
Hdep(M) ≥ 1. Then we have ν(M) ≥ 1.
Proof. We may assume depth(M) = Hdep(M) ≥ 1. Define submodules
Ui of M and quotients Ni and Vi as in the proof of Theorem 2.1. By
assumption we have U1 ∼
= Hm0 (M) = 0, see [5, Prop. 3.5.4], so we get
the equation
n
X
HM =
H Ni .
di
The relation HVi (t) = 1 − t
ν(M) ≥ mini {ν(Ni )} ≥ 1.
i=1
HNi (t) implies ν(Ni ) ≥ 1, and hence
Thus, Hdep(M) and ν(M) coincide at least in three special cases:
Corollary 2.5. Let M be a finitely generated graded module over the
positively graded polynomial ring F[X1 , . . . , Xn ]. If Hdep(M) ≤ 1 or
Hdep(M) = n, then ν(M) = Hdep(M).
Proof. For Hdep(M) ≤ 1, this follows from ν(M) ≤ Hdep(M) and
Proposition 2.4, and if Hdep(M) = n, then HM is the Hilbert series of
a free module over R.
3. The case of the polynomial ring in two variables
From now on, we only consider modules over the ring F[X, Y ] with
α := deg(X) and β := deg(Y ) being coprime; we may assume α < β.
We will deduce an arithmetic characterisation of positive Hilbert depth,
which leads to an analogue of the equation ν(M) = Hdep(M) = p(M)
at least for the special case of Hdep(M) = 1.
The necessary condition
pd (M) = pαβ (M) > 0
alone is not sufficient, as the following example shows.
(6)
HILBERT DEPTH OF GRADED MODULES OVER POLYNOMIAL RINGS
7
Example 3.1. Let R = F[X,Y ] with α = 3 and β = 5 and consider
the module M := R ⊕ R/m (−1) ⊕ R/m (−2). We have ν(M) =
Hdep(M) = 0 and p15 (M) = 1.
Proof. It is easily checked that H(M, n + 15) ≥ H(M, n) holds for any
n ∈ N. This implies p15 (M) ≥ 1, and by H(M, 1) = H(M, 16) we
obtain p15 (M) = 1.
Now we show that any R–module with the same Hilbert series as M
has to have depth 0. Assume on the contrary that there is a module N
with HN = HM and depth(N) > 0. As mentioned earlier (Subsection
2.2), the field F may be assumed to be infinite, and so R contains an
N–regular element z of degree 15. It turns out that such an element
cannot exist due to the fact that, by elementary linear algebra, any
F–linear map N1 ⊕ N2 → N7 ⊕ N11 has a nontrivial kernel: First the
element X 5 cannot be N–regular, as one sees by considering the map
N1 ⊕ N2 → N7 ⊕ N11 , (n1 , n2 ) 7→ X 5 n2 , X 10 n1 .
We may therefore assume z = λX 5 + µY 3 with λ, µ ∈ F, µ 6= 0, and
consider the map f : N1 ⊕ N2 → N7 ⊕ N11 with
f (n1 , n2 ) = λX 2 n1 + µY n2 , µY 2 n1 − µX 3 n2 .
There is also a non–zero element (a, b) ∈ ker f , i. e.
λX 2 a + µY b = 0
µY 2 a − µX 3 b = 0.
We multiply the first equation by X 3 , the second by Y and add both
of them. This yields
λX 5 + µY 3 a = 0,
hence a = 0, since λX 5 + µY 3 was assumed to be N–regular. But
this implies that b 6= 0 is annihilated by powers of X and Y , hence
Hm0 (N) 6= 0, contradicting depth(N) > 0.
Using similar arguments as above, one can deduce additional necessary
conditions for positive Hilbert depth of aP
module M over F[X, Y ] with
deg(X) = 3, deg(Y ) = 5. Let HM (t) = n hn tn , then the coefficients
have to satisfy
hn + hn+1
hn + hn+2
hn + hn+4
hn + hn+7
≤
≤
≤
≤
hn+6 + hn+10 ,
hn+12 + hn+5 ,
hn+9 + hn+10 ,
hn+12 + hn+10
(7)
(8)
(9)
(10)
HILBERT DEPTH OF GRADED MODULES OVER POLYNOMIAL RINGS
8
for all n ∈ Z. Our next example shows that these additional inequalities are still not sufficient to ensure positive Hilbert depth: Let
M := R/m (−1) ⊕ R/(Y ) ⊕ R/(Y ) (−7) ⊕ R/(Y ) (−8),
then its Hilbert series
3
HM (t) = 1 + t + t +
∞
X
tn
n=6
satisfies conditions (6)–(10), but ν(M) = 0. Assume on the contrary
pi ti
that it is possible to decompose HM into summands of the form 1−t
3
q tj
j
and 1−t
5 with pi , qj ∈ N. Since h5 = 0 such a decomposition requires
1
the summand 1−t
3 , but then the remainder
H̃(t) =
∞
X
n=1
h̃n tn := HM (t) −
1
= t + t7 + t8 + t10 + · · ·
1 − t3
cannot be decomposed further, because h̃1 = 1 and h̃4 = h̃6 = 0; note
that H̃ does not satisfy (10) for n = −6. Hence ν(M) = 0 and, by
Proposition 2.4, also Hdep(M) = 0.
A computation using Normaliz (see [7]) reveals two other necessary
conditions for positive Hilbert depth, namely
hn + hn+1 + hn+2 ≤ hn+4 + hn+5 + hn+6 ,
hn + hn+2 + hn+4 ≤ hn+5 + hn+7 + hn+9
(11)
(12)
for all n ∈ Z. The example above does not satisfy the first one for
n = −1.
One might observe that the constants in conditions (6) – (12) which
are not multiples of 3 or 5 are the numbers 1, 2, 4 and 7. These are
the only positive integers not contained in h3, 5i := 3N0 + 5N0 , the
so-called numerical semigroup generated by 3 and 5. This is not a
mere coincidence: The necessary and sufficient conditions for positive
Hilbert depth to be developed in the sequel will turn out to be closely
related to the theory of numerical semigroups, so it seems advisable to
recall some basic facts of this theory here.
3.1. Numerical semigroups generated by two elements. Let S
be a sub–semigroup of N0 such that the greatest common divisor of
all its elements is equal to 1. Then the subset N0 \ S has only finitely
many elements, which are called the gaps of S. Such a semigroup is
said to be numerical. The smallest element c = c(S) ∈ S such that
HILBERT DEPTH OF GRADED MODULES OVER POLYNOMIAL RINGS
9
n ∈ S for all n ∈ N with n ≥ c is called the conductor of S. The
number of gaps is called the genus of S and is denoted by g(S).
We are interested in numerical semigroups generated by two elements.
Let α, β ∈ N with gcd(α, β) = 1; we write hα, βi := N0 α + N0 β and
denote the set of gaps of this semigroup by L.
Lemma 3.2 (Cf. [8], Prop. 2.13). The semigroup hα, βi generated by
two positive integers α, β with gcd(α, β) = 1 is numerical. Its conductor
and genus are given by
c
c = c (hα, βi) = (α − 1)(β − 1) and g (hα, βi) = .
2
Lemma 3.3 ([9], Lemma 1). Let e ∈ Z. Then e ∈
/ hα, βi if and only
if there exist k, ℓ ∈ N such that e = αβ − kα − ℓβ.
Corollary 3.4. Let k, ℓ ∈ N such that 1 ≤ k < β and 1 ≤ ℓ < α, then
|αβ − kα − ℓβ| ∈ L.
Proof. This follows immediately from the preceding lemma, since we
have either |αβ − kα − ℓβ| = αβ − kα − ℓβ or
|αβ − kα − ℓβ| = kα + ℓβ − αβ = αβ − (β − k)α − (α − ℓ)β.
Corollary 3.5. Any integer n > 0 has a unique presentation
n = pαβ − aα − bβ
with integers p > 0, 0 ≤ a < β and 0 ≤ b < α.
Proof. Since the gaps of hα, βi are covered by Lemma 3.3, we have to
show the existence of the presentation for integers n = kα + ℓβ with
k, ℓ ≥ 0. Let k = qβ + r, ℓ = q ′ α + r ′ with 0 ≤ r < β and 0 ≤ r ′ < α,
then
n = kα + ℓβ = (q + q ′ + 2)αβ − (β − r)α − (α − r ′ )β.
The uniqueness follows easily since gcd(α, β) = 1. Let
pαβ − aα − bβ = n = p′ αβ − a′ α − b′ β,
then
(p − p′ )β − a + a′ α = (b − b′ )β,
so α has to divide |b − b′ | < α and hence b = b′ . But this implies
|p − p′ |β = |a − a′ | < β and therefore a = a′ and p = p′ as well.
HILBERT DEPTH OF GRADED MODULES OVER POLYNOMIAL RINGS 10
The presentation mentioned above will be of particular importance for
the gaps of hα, βi. In the sequel we will frequently use the notation
e = αβ − a(e) · α − b(e) · β.
(13)
Let n be a nonzero element of S. The set
Ap(S, n) := {h ∈ S | h − n ∈
/ S}
is called the Apéry set of n in S.
Lemma 3.6 ([8], Lemma 2.4). Let S be a numerical semigroup and let
n ∈ S \ {0}. Then
Ap(S, n) = {0 = w(0), w(1), . . . , w(n − 1)},
where w(i) := min{x ∈ S | x ≡ i mod n} for 0 ≤ i ≤ n − 1.
Lemma 3.7. Let n ∈ S \ {0}, i ∈ {0, . . . , n − 1}. Let x be an integer
congruent to w(i) modulo n. Then x ∈ S if and only if w(i) ≤ x.
Proof. By assumption we have x ≡ w(i) ≡ i mod n. Hence x ∈ S
implies x ≤ w(i) by the definition of w(i). Conversely, let x ∈ Z
be such that x ≥ w(i). This implies x − w(i) = λn ≥ 0, and so
x = λn + w(i) ∈ S.
Lemma 3.8 ([10], Theorem 1). Let S = hα, βi, then
Ap(S, α) = {sβ | s = 0, . . . , α − 1} ,
Ap(S, β) = {rα | r = 0, . . . , β − 1} .
As an immediate consequence of the preceding lemmas we get
Corollary 3.9. Let 0 < r < β (resp. 0 < s < α), and let x be an
integer such that 0 < x < rα and x ≡ rα mod β (resp. such that
0 < x < sβ and x ≡ sβ mod α). Then x is a gap of hα, βi.
3.2. Arithmetical conditions for Hdep(M) > 0. Next we introduce
the announced necessary and sufficient condition for positive Hilbert
depth. We begin by reconsidering the special case α = 3, β = 5 and
HILBERT DEPTH OF GRADED MODULES OVER POLYNOMIAL RINGS 11
the inequalities mentioned above:
hn ≤
hn + hn+1 ≤
hn + hn+2 ≤
hn + hn+4 ≤
hn + hn+7 ≤
hn + hn+1 + hn+2 ≤
hn + hn+2 + hn+4 ≤
hn+15 ,
hn+6 + hn+10 ,
hn+12 + hn+5 ,
hn+9 + hn+10 ,
hn+12 + hn+10 ,
hn+5 + hn+6 + hn+7 ,
hn+5 + hn+7 + hn+9 .
We note some observations on the structure of these inequalities.
(1) For each index i on the left there are indices j, j ′ on the right
such that i ≡ j mod 3, i ≡ j ′ mod 5 and i < j, j ′ .
(2) In each inequality the constants appearing on the left hand side
are gaps of the semigroup h3, 5i, and the difference of any two
of these gaps is also a gap.
(3) The constants appearing on the right hand side are either gaps
of h3, 5i or multiples of 3 or 5.
This motivates the following definition.
Definition 3.10. Let α, β > 0 be coprime integers and let L denote
the set of gaps of hα, βi. An (α, β)–fundamental couple [I, J] consists
m
of two integer sequences I = (ik )m
k=0 and J = (jk )k=0 , such that
(0) i0 = 0.
(1) i1 , . . . , im , j1 , . . . , jm−1 ∈ L and j0 , jm ≤ αβ.
ik ≡ jk
(2) jk ≡ ik+1
jm ≡ i0
mod α
mod β
mod β
and
and
and
ik < jk
jk > ik+1
jm ≥ i0 .
for k = 0, . . . , m;
for k = 0, . . . , m − 1;
(3) |ik − iℓ | ∈ L for 1 ≤ k < ℓ ≤ m.
The number |I| = m + 1 will sometimes be called the length of the
couple. The set of (α, β)–fundamental couples will be denoted by Fα,β .
Fundamental couples will be discussed in more detail in the next subsection; here we just note the simplest nontrivial example:
Remark
3.11. Let e be a gap
of hα, βi with e = αβ − aα − bβ, then
(0, e), ((β − a)α, (α − b)β) form an (α, β)–fundamental couple.
HILBERT DEPTH OF GRADED MODULES OVER POLYNOMIAL RINGS 12
Remark 3.12. The number of (α, β)–fundamental couples grows surprisingly with increasing α and β. We give some examples:
|Fα,β | g(S)
S = hα, βi
h4, 5i
14
6
h4, 7i
30
9
728 25
h6, 11i
h11, 13i 104 006 60
The main result of this paper is the following theorem:
Theorem 3.13. Let R = F[X, Y ] be the polynomial ring in two variables such that deg(X) = α, deg(Y ) = β with gcd(α, β) = 1. Let M
be a finitely generated graded R–module.
P Then M has positive Hilbert
depth if and only if its Hilbert series n hn tn satisfies the condition
P
P
(⋆)
i∈I hi+n ≤
j∈J hj+n for all n ∈ Z, [I, J] ∈ Fα,β .
It is easily seen that condition (⋆) is indeed necessary for positive
Hilbert depth. First we note some elementary remarks concerning (⋆).
Lemma 3.14. a) Let H, H ′ ∈ Z[[t]][t−1 ] be nonnegative Laurent series
satisfying condition (⋆), then the same holds for H +H ′ and ti H, where
i ∈ Z.
1
1
b) The series 1−t
α and 1−tβ satisfy condition (⋆).
P
∞
n
c) Let H(t) =
n=b hn t be a nonnegative Laurent series satisfying
1
1
condition (⋆), then the same holds for 1−t
α H(t) and 1−tβ H(t).
Proof. Assertion a) is obvious, and b) is also clear in view of the definition of an (α, β)–fundamental couple. To prove c) it is, by symmetry,
enough to consider
! ∞
∞ X
∞
X
X
X
1
n
nβ
hn−kβ tn
hn t =
· H(t) =
·
t
1 − tβ
n=0
k≥0
n=b
n=b
with a finite inner sum. If H fulfills condition (⋆), then for each (α, β)–
fundamental couple [I, J] the inequalities
X
X
hn−kβ+i ≤
hn−kβ+j
i∈I
j∈J
hold for all n ∈ Z and all k ∈ N0 . Summing up yields
XX
XX
hn−kβ+i ≤
hn−kβ+j ,
k≥0 i∈I
k≥0 j∈J
HILBERT DEPTH OF GRADED MODULES OVER POLYNOMIAL RINGS 13
and hence
XX
hn−kβ+i ≤
i∈I k≥0
as desired.
XX
hn−kβ+j ,
j∈J k≥0
Let M be a finitely generated graded R–module with Hdep(M) > 0.
By Proposition 2.4, its Hilbert series admits a decomposition of the
form
Q2 (t)
QX (t) QY (t)
HM (t) =
+
+
(1 − tα )(1 − tβ ) 1 − tα 1 − tβ
with nonnegative Q2 , QX , QY ∈ Z[t, t−1 ]. These three summands in
the decomposition above fulfill (⋆) by the previous lemma, and so does
their sum HM . Hence we have proven:
Proposition 3.15. Let M be a finitely generated graded R–module
with Hdep(M) > 0. Then HM satisfies condition (⋆).
The proof of the converse is more elaborate and requires several steps.
First
show that we may restrict our attention to Laurent series
P we
n
h
t
whose
coefficients eventually become (periodically) constant.
n
n
Since the Hilbert series of an R–module M with dim(M) = 1 is of this
form, in the sequel such a series will be called a series of dimension 1
for short.
Let M be a finitely generated graded R–module with Hilbert series
satisfying condition (⋆). By Theorem 2.1, we have a decomposition of
the form
QX (t) QY (t)
Q2 (t)
HM (t) = Q0 (t) +
+
+
1 − tα 1 − tβ (1 − tα )(1 − tβ )
with nonnegative Q0 , QX , QY , Q2 ∈ Z[t, t−1 ]. The reduction to the case
of a series of dimension 1 is based on the following observation: For
any r > 0 the series
Q2 (t)
(14)
(1 − tα )(1 − tβ )
Prβ−1 jα
X
QX (t) QY (t)
j=0 t
+
+
Q
(t)
·
=:
h′n tn
= Q0 (t) +
2
α
β
β
1−t
1−t
1−t
n
H ′(t) := HM (t) − trαβ
is of dimension 1; therefore we have to show that for some r the series
H ′ also satisfies condition (⋆). To this end, choose r ∈ N large enough
such that rαβ > deg(Q0 ) + αβ. Then h′n = hn for n < deg(Q0 ) + αβ,
and so all inequalities of (⋆) which are influenced by Q0 are valid by
HILBERT DEPTH OF GRADED MODULES OVER POLYNOMIAL RINGS 14
assumption. The remaining inequalities of (⋆) are valid since for indices
n > deg(Q0 ) the coefficients of H ′ agree with those of the series
Prβ−1 jα
QX (t) QY (t)
j=0 t
+
+ Q2 (t) ·
,
α
β
1−t
1−t
1 − tβ
and the latter satisfies condition (⋆) by Lemma 3.14. Hence it remains
to prove the following assertion:
Proposition 3.16. Let H(t) = (1−tαQ(t)
be a nonnegative Laurent
)(1−tβ )
series of dimension 1. If H satisfies condition (⋆), then the series is
(α, β)–decomposable with ν(H) = 1.
An idea how to prove this
is not far to seek: Without loss
Pproposition
n
of generality let H(t) = n≥0 hn t with h0 > 0. If we can show that
at least one of the series
1
1
Hα (t) := H(t) −
and
H
(t)
:=
H(t)
−
β
1 − tα
1 − tβ
is nonnegative and satisfies condition (⋆), then a simple inductive argument would complete the proof.
Indeed condition (⋆) ensures that one of these series must be nonnegative:
P
Q(t)
n
Proposition 3.17. Let ∞
n=0 hn t = (1−tα )(1−tβ ) be a nonnegative Laurent series satisfying condition (⋆). If h0 > 0, then at least one of the
numbers cα := min{hrα | r > 0} and cβ := min{hrβ | r > 0} is positive.
Proof. Let cα = hkα and cβ = hℓβ . Since hn ≤ hn+αβ for all n ∈ Z and
h0 > 0 we may assume 0 < k < β and 0 < ℓ < α. We consider
e := αβ − (β − k)α − (α − ℓ)β.
If this number is a gap of hα, βi, then (0, e), (kα, ℓβ) is a fundamental
couple. Hence we have
0 < h0 ≤ h0 + he ≤ hkα + hℓβ ,
and therefore at least one term on the right must be positive. Otherwise
e < 0 and, by Corollary 3.4, −e
= αβ − kα − ℓβ is a gap. In this case
(0, −e), ((β − k)α, (α − ℓ)β) is a fundamental couple, and hence the
inequality
hn + hn−e ≤ hn+(β−k)α + hn+(α−ℓ)β
holds for all n ∈ Z. For n = e this yields
0 < h0 ≤ he + h0 ≤ hℓβ + hkα ,
so again the right hand side must be positive.
HILBERT DEPTH OF GRADED MODULES OVER POLYNOMIAL RINGS 15
The question whether condition (⋆) is still valid for Hα or Hβ is more
1
delicate. Subtraction of, say, 1−t
α diminishes all coefficients hrα with
r ≥ 0 by 1; therefore all inequalities of (⋆) containing such a coefficient either on both sides or not at all are preserved. But there are
(finitely many) inequalities where an index rα ≥ 0 on the right has
a counterpart r ′ α < 0 on the left. We introduce a name for such an
inequality.
P
n
Definition 3.18. Let H(t) = ∞
n=0 hn t be a formal Laurent series,
and let m ∈ Z such that there exist i ∈ I, j ∈ J with m + i < 0,
m + j ≥ 0 and m + i ≡ m + j ≡ 0 mod α. Then the inequality
X
X
hm+i ≤
hm+j
i∈I
j∈J
is called α–critical. A β–critical inequality is defined analogously.
1
Subtraction of 1−t
α diminishes only the right hand side of an α–critical
inequality, hence condition (⋆) remains valid for Hα if and only if for
H all α–critical inequalities are strict, or likewise with β. Therefore
we have to investigate whether for every Laurent series satisfying (⋆)
the critical inequalities of at least one type hold strictly. This requires
some technical machinery to be developed in the next subsection.
3.3. Fundamental and balanced couples. The structure of the fundamental couples is widely determined by the following facts:
Lemma 3.19. Let i1 , i2 be gaps of hα, βi, and let ak = a(ik ), bk = b(ik )
for k = 1, 2 denote the coefficients in the presentation of ik according
to Corollary 3.5. Then
a) The difference |i1 − i2 | is a gap if and only if (a2 − a1 )(b2 − b1 ) < 0.
b) There exists a gap j ≥ i1 , i2 with i1 ≡ j mod α and i2 ≡ j mod β
if and only if a1 ≥ a2 and b1 ≤ b2 , and this gap j is uniquely determined
to be αβ − a2 α − b1 β.
Proof. a) We may assume |i1 − i2 | = i1 − i2 = (a2 − a1 )α + (b2 − b1 )β.
If this number is a gap, then a := a2 − a1 and b := b2 − b1 must bear
different signs. For the converse note that |a| < β and |b| < α, hence
by Lemma 3.3
i1 − i2 = αβ − (β − a)α + bβ = αβ + aα − (α − b)β
is a gap if a > 0 and b < 0 or vice versa.
b) Let j := αβ − a2 α − b1 β, then every solution of the congruence
system in question is of the form j + rαβ with some r ∈ Z (Chinese
HILBERT DEPTH OF GRADED MODULES OVER POLYNOMIAL RINGS 16
remainder theorem). By Lemma 3.3, j is the only solution which is
possibly a gap. On the other hand j ≥ i1 , i2 if and only if a1 ≥ a2 and
b1 ≤ b2 , and in this case j is indeed a gap.
The condition in part b) of the lemma will occur frequently in the
sequel. Therefore we introduce a relation as follows:
Definition 3.20. For gaps i1 , i2 of the semigroup hα, βi we define
i1 i2 : ⇐⇒ a(i1 ) ≥ a(i2 ) ∧ b(i1 ) ≤ b(i2 )
and
i1 ≺ i2 : ⇐⇒ a(i1 ) > a(i2 ) ∧ b(i1 ) < b(i2 ).
Note that, deviating from the usual convention, i1 ≺ i2 is a stronger
assertion than just i1 i2 ∧ i1 6= i2 .
Obviously defines a partial ordering on the set of gaps. Together with
the second part of Lemma 3.19 this yields the announced structural
result for fundamental couples.
m
Corollary 3.21. a) Let [(ik )m
k=0 , (jk )k=0 ] be a fundamental couple, then
ik ≺ ik+1 for k = 1, . . . , m − 1.
b) An (α, β)–fundamental couple has length at most α.
c) Let i1 ≺ i2 ≺ · · · ≺ im be gaps of hα, βi, then there exists a unique
sequence J = (jk )m
k=0 such that (i0 := 0, i1 , . . . , im ), (j0 , . . . , jm ) is a
fundamental couple.
d) Let L′ = {ℓ1 , . . . , ℓm } be a subset of L := N\ hα, βi with |ℓn −ℓp | ∈ L
for all n 6= p. Then there exists a unique fundamental couple [I, J] =
m
′
[(ik )m
k=0 , (jk )k=0 ] such that L = {ik | 1 ≤ k ≤ m}.
Proof. Part a) is clear by Lemma 3.19, and so is c), since there are
unique gaps jk , 0 < k < m such that ik ≡ jk mod α, ik+1 ≡ jk mod β
and jk > ik , ik+1 , and the fundamental couple can only be completed
by setting j0 := (β − a(i1 ))α and jm := (α − b(im ))β. Part b) follows,
because the integers 0 < b(ik ) < α, k = 1, . . . , m − 1 must be distinct.
For d) one notes that all elements of L′ are ≺–comparable. Hence they
can be ordered by this relation; it remains to apply the previous parts
a) and c).
P
P
The fact that the series n≥0 tnα and n≥0 tnβ satisfy condition (⋆)
only depends on the second requirement in the definition of a fundamental couple. This suggests this property to be the most important
for our purpose. We introduce a notion for couples of integer sequences
with just this property:
HILBERT DEPTH OF GRADED MODULES OVER POLYNOMIAL RINGS 17
Definition 3.22. Let α, β > 0 be coprime integers.
a) An (α, β)–balanced couple [I, J] consists of two integer sequences
m
I = (ik )m
k=0 and J = (jk )k=0 , such that
ik ≡ jk mod α
jk ≡ ik+1 mod β
jm ≡ i0 mod β
and
and
and
ik ≤ jk
jk ≥ ik+1
jm ≥ i0 .
for k = 0, . . . , m;
for k = 0, . . . , m − 1;
The number m + 1 will again be called the length of the couple.
b) An (α, β)–balanced couple is called reduced, if it satisfies the additional condition
min {jk−1 − ik , jk − ik }
min {jm − i0 , j0 − i0 }
min {jk − ik , jk − ik+1 }
min {jm − im , jm − i0 }
<
<
<
<
αβ
αβ
αβ
αβ
for k = 1, . . . , m
for k = 0, . . . , m − 1
and the inequalities in a) hold strictly.
Remark 3.23. a) Note that the length of a balanced couple is not
bounded above, because there is no restriction against repetition of
residue classes; even the same integer may appear several times.
b) Any fundamental couple except [(0), (αβ)] is also a reduced balanced
couple.
c) Reducedness of the balanced couple [I, J] implies that each jk is the
smallest solution of the congruence system x ≡ ik mod α ∧ x ≡ ik+1
mod β fulfilling the additional requirement x ≥ ik , ik+1 . In particular, for any reduced balanced couple [I, J], the sequence J is uniquely
determined by I.
m
Lemma 3.24. Let [(ik )m
k=0 , (jk )k=0 ] be a reduced (α, β)–balanced couple
with nonnegative elements, i0 = 0 and m ≥ 1.
a) At least one of the elements i1 , im is a gap of hα, βi.
b) If ik , ik+1 are gaps with ik ik+1 , then ik ≺ ik+1 .
c) If i1 , . . . , im are gaps such that i1 i2 · · · im , then the couple
is fundamental.
d) Any reduced (α, β)–balanced couple [(0, i1 ), (j0 , j1 )] is fundamental.
Proof. a) The couple [I, J] is reduced, therefore we have
min {jm − i0 , j0 − i0 } = min {jm = sβ, j0 = rα} < αβ.
HILBERT DEPTH OF GRADED MODULES OVER POLYNOMIAL RINGS 18
Since i1 ≡ j0 mod β and im ≡ jm mod α the claim follows immediately from Corollary 3.9.
b) If a(ik ) = a(ik+1 ), then ik ≡ ik+1 mod β and hence jk = ik , a
contradiction; the equality b(ik ) = b(ik+1 ) is treated analogously.
c) By the previous part we even have i1 ≺ i2 ≺ · · · ≺ im . Hence there
exists a fundamental couple with this I by Corollary 3.21, and it is the
only reduced balanced couple with this I.
d) This follows immediately from a) and c).
m
Lemma 3.25. Let [(ik )m
k=0 , (jk )k=0 ] be an (α, β)–balanced couple with
nonnegative elements, and let ik = αβ − ak α − bk β be a gap of hα, βi
for some 0 < k < m. Then (α − bk )β < jk unless ik+1 is a gap with
ik ik+1 , and vice versa (β − ak )α < jk−1 unless ik−1 is a gap with
ik−1 ik .
Proof. Let ik+1 = pαβ − ak+1 α − bk+1 β according to Corollary 3.5, then
jk = rαβ − ak+1 α − bk β
for some r ∈ N. If ik+1 is a gap then, by assumption, ik 6 ik+1 , and
so, as already mentioned in the proof of part b) of Lemma 3.19,
jk ≥ 2αβ − ak+1 α − bk β > αβ − bk β;
otherwise p > 1, and since
(p − 1)αβ − ak+1 α − bk β = pαβ − ak+1 α − (α + bk )β < ik+1 ,
we have r ≥ p and therefore
jk ≥ pαβ − ak+1 α − bk β ≥ 2αβ − ak+1 α − bk β > αβ − bk β.
The second assertion can be proven analogously.
The next result provides the key for showing Proposition 3.16. Its
intricate proof is the technically most challenging step on the way to
our main result.
P
Theorem 3.26. Let H(t) = n hn tn be a nonnegative formal Laurent
series satisfying condition (⋆). Then the inequality
X
X
hi ≤
hj
i∈I
j∈J
holds for any (α, β)–balanced couple [I, J].
Proof. We may assume that [I, J] is reduced: A perhaps necessary
replacement of an i ∈ I with i + αβ or a j ∈ J with j − αβ is harmless
since hn ≤ hn+αβ for all n ∈ Z, while any elements ik = jk or jk =
HILBERT DEPTH OF GRADED MODULES OVER POLYNOMIAL RINGS 19
ik+1 can be removed from I and J without affecting the inequality
in question. Therefore we may in particular assume ik 6= ik+1 for
m
k = 0, . . . , m − 1. Since [(ik − x)m
k=0 , (jk − x)k=0 ] is a reduced balanced
couple as well for any x ∈ Z, we may also assume min I = 0. Finally,
we may shift the numbering of the elements in I and J such that i0 = 0.
Throughout this proof ak and bk denote the coefficients of α resp. β in
the presentation of ik according to Corollary 3.5.
The proof uses induction on m, the case m = 0 being trivial, while
m = 1 is covered by Lemma 3.24.
Let therefore m ≥ 2 and assume that the result is already proven
for balanced couples of length ≤ m. The general idea is to insert an
auxiliary element x into I and J, which allows to split the amended
couple into smaller balanced couples [I ′ , J ′ ] and [I ′′ , J ′′ ] with I ′ ∪ I ′′ =
I ∪ {x} and J ′ ∪ J ′′ = J ∪ {x}. The inequalities
X
i∈I ′
hi ≤
X
j∈J ′
hj and
X
i∈I ′′
hi ≤
X
hj
j∈J ′′
then hold by the induction hypothesis, so we get our desired inequality
by adding them and cancelling hx .
Since [I, J] is reduced, at least one of the elements j0 = rα, jm = sβ
has to be less than αβ. We distinguish three cases:
I) j0 < αβ and jm ≥ αβ: In this case i1 is a gap by Corollary 3.9,
while im is not. Let M be the largest index k such that i1 , . . . , ik are
gaps with i1 ≺ i2 · · · ≺ ik . Then iM +1 is not a gap with iM iM +1 .
Hence Lemma 3.25 implies that x := (α − bM )β < jM , and of course
x < αβ ≤ jm as well. Since x ≡ iM ≡ jM mod α and x ≡ jm mod β
we have two balanced couples
(i0 , . . . , iM ), (j0 , . . . , jM −1 , x)
(15)
and
(x, iM +1 , . . . , im ), (jM , . . . , jm ) .
Of these, the first one is already fundamental (by Lemma 3.24), while
the second has length m − M + 1 ≤ m, so the induction hypothesis can
be applied.
II) jm < αβ and j0 ≥ αβ: This case is mirror–imaged to the first.
Now im is a gap and i1 is not, so there is a smallest index N such
that iN , . . . , im are gaps with iN ≺ · · · ≺ im . Then iN −1 is not a gap
with iN −1 iN , so x := (β − aN )α < jN −1 by Lemma 3.25, and also
x < αβ ≤ j0 . Since x ≡ iN ≡ jN −1 mod β and x ≡ j0 mod α we have
HILBERT DEPTH OF GRADED MODULES OVER POLYNOMIAL RINGS 20
two balanced couples
and
(x, i1 , . . . , iN −1 ), (j0 , . . . , jN −1 )
(iN , . . . , im , i0 ), (jN , . . . , jm , x) ,
(16)
the first being of length N ≤ m, and the second being a cyclic permutation of a fundamental couple.
III) j0 , jm < αβ: In this case both i1 and im are gaps. We choose
M and N as in case I) resp. case II). If M = m, then the couple is
fundamental and we are done. We may therefore assume M < m and
thus N > M. Two subcases can be treated analogously to the cases
above:
If bM ≥ bm , then x := (α − bM )β ≤ (α − bm )β = jm , so we may adopt
the reasoning of case I) and split [I, J] into the couples given in (15).
If aN ≥ a1 then x := (β − aN )α ≤ (β − a1 )α = j0 , so we may adopt
the reasoning of case II) and split [I, J] into the couples given in (16).
We may therefore assume that bM < bm and aN < a1 and hence am < a1
and bm > b1 .
This case is treated recursively: Starting with p0 := M, q0 := N,
u1 := 1 and v1 := m we construct two nonincreasing integer sequences
(pr ), (vr ) and two nondecreasing integer sequences (qr ), (ur ) such that
bpr−1 < bvr
pr−1 ≥ ur
aur > avr
and
and
and
aqr−1 < aur ,
qr−1 ≤ vr ,
bur < bvr .
(17)
(18)
(19)
If pr−1 , qr−1 , ur and vr are already constructed for some r > 0, then
we continue by defining
pr := max{k ≤ pr−1 | ak > avr } ≥ ur ,
qr := min{k ≥ qr−1 | bk > bur } ≤ vr .
(20)
(21)
Note that if pr < M, we have
apr +1 ≤ avr ,
(22)
bqr −1 ≤ bur
(23)
and similarly
if qr > N.
By construction we have ipr ≺ ivr and iur ≺ iqr . According to Lemma
3.19 there exists a connecting gap for each of these pairs, and we investigate whether one of these gaps is suitable as the auxiliary element
x. If both of them fail to fit, then we continue our recursive procedure:
HILBERT DEPTH OF GRADED MODULES OVER POLYNOMIAL RINGS 21
(A) The pair ipr ≺ ivr : Insertion of x := αβ − avr α − bpr β allows to
split the couple [I, J] in the following way:
jvr −1
jm
ivr −1
ivr
im
0
jpr
x
j0
i1
ipr
x
ipr +1
The first part
(i0 , . . . , ipr , ivr , . . . , im ), (j0 , . . . , jpr −1 , x, jvr , . . . , jm )
is a fundamental couple. If its counterpart
(x, ipr +1 , . . . , ivr −1 ), (jpr , . . . , jvr −1 )
is a balanced couple too, then the induction hypothesis applies
to it since it is of length
1 + vr − 1 − (pr + 1) + 1 = vr − pr ≤ m − 1.
The required congruences are satisfied, hence it remains to
check whether x ≤ jpr , jvr −1 . The first inequality
x ≤ αβ − apr +1 α − bpr β = jpr
is clear for pr < M because of (22), and if pr = M it holds since
in this case one has jM ≥ 2αβ − aM +1 α − bM β, compare the
proof of Lemma 3.25. Similarly, the second inequality is clear if
vr = N, since in this case we have jvr −1 ≥ 2αβ − avr α − bvr −1 β.
Otherwise, i. e., if vr > N, then jvr −1 is a gap. We have
jvr −1 = αβ − avr α − bvr −1 β,
hence x ≤ jvr −1 if and only if bpr ≥ bvr −1 . Note that this
inequality holds in particular if pr = ur and qr = vr , since
either qr = N or (23) is valid.
(B) The pair iur ≺ iqr : This case is mirror–imaged to (A). Insertion
of x := αβ − aqr α − bur β yields a splitting of [I, J] into the
fundamental couple
(0, iur , iqr , . . . , im ), (j0 , . . . , jur −1 , x, jqr , . . . , jm )
HILBERT DEPTH OF GRADED MODULES OVER POLYNOMIAL RINGS 22
and its counterpart
(x, iur +1 , . . . , iqr −1 ), (jur , . . . , jqr −1 ) .
Again we are done if the latter is a balanced couple. Here we
have to check the inequalities x ≤ jqr−1 , jur . The first of these
x ≤ αβ − aqr α − bqr −1 β
follows from (23) if qr > N, and in the case qr = N one has
jN −1 ≥ 2αβ − aN α − bN −1 β; similarly, the second inequality is
clear if vr = M, since then one has jur ≥ 2αβ − aur +1 α − bur β.
Otherwise, i. e., if ur < M, we have
jur = αβ − aur +1 α − bur β,
thus x ≤ jur if and only if aqr ≥ aur +1 .
(C) By the previous discussion it remains to deal with the following
situation:
(A)
(B)
bpr < bvr −1 , vr > N and aqr < aur +1 , ur < M.
We continue by defining the next elements of the sequences (ur )
and (vr ) by
ur
for ur = pr
ur+1 :=
ur + 1 otherwise.
vr
for vr = qr
vr+1 :=
vr − 1 otherwise.
Note that we cannot have ur+1 = ur and vr+1 = vr simultaneously, since the case ur = pr and vr = qr is covered by (A).
It is easy to see that the inequalities (17) – (19) also hold
for r + 1: For the first and the second this is trivial by our
assumption resp. by definition of (ur ), (vr ). Since ur+1 ≤ ur + 1
and vr+1 ≥ vr − 1 we have aur+1 ≥ aur +1 and bvr+1 ≥ bvr −1 ;
together with (18) this implies (19). Hence we may continue
with the construction of pr+1 and qr+1 .
By construction, it is clear that this recursive procedure will
eventually terminate, namely with one of the cases ur = M,
ur = pr ∧ vr = qr or vr = N, which are covered by the discussion
above.
HILBERT DEPTH OF GRADED MODULES OVER POLYNOMIAL RINGS 23
3.4. Proof of the main result. After the previous rather technical
subsection we return to the proof of the main result, which now finally
can be completed with the aid of Theorem 3.26:
P
Q(t)
n
Let H(t) := ∞
n=0 hn t = (1−tα )(1−tβ ) be a nonnegative Laurent series
satisfying condition (⋆) with h0 > 0. We want to show that at least
one of the series
Hα (t) = H(t) −
1
,
1 − tα
Hβ (t) = H(t) −
1
1 − tβ
is nonnegative and satisfies (⋆) as well. Since by Proposition 3.17 at
least one of the numbers
cα := min{hrα | r > 0}
cβ := min{hrβ | r > 0}
is positive, there are two cases: If only one of these series, say Hβ , is
nonnegative, then we have to show that the β–critical inequalities hold
strictly. If both series are nonnegative, then we have to show that all
critical inequalities of one type hold strictly. We begin with the first
case.
P
n
Proposition 3.27. Let H(t) := ∞
n=0 hn t be a nonnegative Laurent
series satisfying condition (⋆) and h0 > 0. If cα = 0 (resp. cβ = 0),
then the β–critical (resp. the α–critical) inequalities hold strictly.
Proof. Assume cα = 0, thus hrα = 0 for some 0 < r < β. Let
X
i∈I
hn+i ≤
X
hn+j
j∈J
be a β–critical inequality, so n + ip = −s′ β for some 0 ≤ p ≤ m, s′ > 0.
We define a balanced couple
m
[I ′ , J ′ ] := [(ik + n)m
k=0 , (jk + n)k=0 ] .
Choose some integer ℓ < rα, jp′ such that ℓ ≡ rα mod β and ℓ ≡ jp′
mod α. We construct another balanced couple [I ′′ , J ′′ ] by replacing
i′p = −s′ β with the sequence 0 → rα → ℓ, as the following picture
illustrates:
HILBERT DEPTH OF GRADED MODULES OVER POLYNOMIAL RINGS 24
i′p−1
jp′
rα
sβ
−s′ β
0
i′p+1
ℓ
P
P
By Theorem 3.26 we have i∈I ′′ hi ≤ j∈J ′′ hj . This implies
X
X
X
X
X
X
hn+i < h0 +hℓ +
hi =
hi ≤
hj = hrα +
hj =
hn+j ,
i∈I
i∈I ′
i∈I ′′
j∈J ′′
j∈J ′
j∈J
so the original β–critical inequality holds strictly. The case cβ = 0 is
treated analogously.
The basic idea for solving the second case is quite similar:
P
n
Proposition 3.28. Let H(t) := ∞
n=0 hn t be a nonnegative Laurent
series satisfying condition (⋆) and h0 > 0. If cα , cβ 6= 0, then the
α–critical or the β–critical inequalities hold strictly.
Proof. Assume on the contrary that there is a non–strict α–critical
inequality, i. e.
X
X
hnα +j
hnα +i =
i∈Iα
j∈Jα
as well as a non–strict β–critical inequality
X
X
hnβ +̃
hnβ +ı̃ =
ı̃∈Iβ
̃∈Jβ
mβ
mβ
mα
α
where [Iα , Jα ] = [(ik )m
k=0 , (jk )k=0 ] and [Iβ , Jβ ] = (ı̃k )k=0 , (̃k )k=0 . By
definition of a critical inequality there exist 0 ≤ p ≤ mα , 0 ≤ q ≤ mβ ,
and r ′ , s′ > 0 such that nα + ip = −r ′ α and nβ + ı̃q = −s′ β. We define
balanced couples
ˆ J]
ˆ := [(ik + nα )mα , (jk + nα )mα ]
[I,
k=0
k=0
ˇ Jˇ := (ı̃k + nβ )mβ , (̃k + nβ )mβ .
I,
k=0
k=0
ˆ J]
ˆ
We construct
another balanced couple [I, J] by glueing together [I,
ˇ
ˇ
and I, J , as illustrated in the following picture:
HILBERT DEPTH OF GRADED MODULES OVER POLYNOMIAL RINGS 25
ı̌q−2
ı̂p+2
̌q−2
̂p+1
ı̌q−1
ı̂p+1
rα
sβ
0
−s′ β
−r ′ α
̂p−1
̌q
−rα − sβ
ı̌q+1
ı̂p−1
̌q+1
̂p−2
ı̌q+2
ı̂p−2
P
P
By Theorem 3.26, we have i∈I hi ≤ j∈J hj . Since h−r′ α = h−s′ β = 0
we have
X
X
X
X
X
X
X
X
hi +
hi < h0 +
hi +
hi =
hi ≤
hj =
hj +
hj ,
i∈Iˆ
i∈Iˇ
i∈Iˆ
i∈Iˇ
i∈I
hnβ +ı̃ =
X
but this contradicts
X
X
hi +
hi
i∈Iˇ
i∈Iˆ
=
X
hnα +i +
ı̃∈Iβ
i∈Iα
=
X
j∈Jˆ
X
hj +
X
j∈Jα
j∈Jˆ
j∈J
hnα +j +
X
j∈Jˇ
hnβ +̃
̃∈Jβ
hj .
j∈Jˇ
After these final preparatory steps we are ready to prove the essential
assertion, Proposition 3.16.
Proof of Prop. 3.16: Since H is of dimension 1, there existsPan integer
N such that hn = hn+αβ holds for all n ≥ N. Then the sum n+αβ−1
hk
k=n
has the same value for every n ≥ N; we denote this value by σ(H).
We prove the assertion by induction on s := σ(H), starting with the
vacuous case s = 0. For s > 0 we may assume h0 > 0 and hk = 0
for k < 0. Let cα and cβ be defined as above. We distinguish two
cases: If cα vanishes, then, by Propositions 3.17 and 3.27, Hβ (t) is a
HILBERT DEPTH OF GRADED MODULES OVER POLYNOMIAL RINGS 26
nonnegative series satisfying condition (⋆). Since σ(Hβ ) < σ(H) we
are done; the same argument works with α and β interchanged. If
cα , cβ > 0, then both series Hα , Hβ are nonnegative, and at least one
of them also satisfies condition (⋆) by Proposition 3.28, so we may
apply the induction hypothesis to it.
As mentioned above, this result implies the converse of Proposition
3.15 for any R–module; therefore our main result, Theorem 3.13, is
completely proven.
The closing example of this section confirms the importance of Proposition 3.28.
Example 3.29. Let α = 3 and β = 4. For
1 + t + t6 + t7 + t8
= 1 + t + 0t2 + t3 + t4 + 0t5 + t6 + · · · ,
1 − t3
we have c4 = 1, but not all the 4–critical inequalities hold strictly.
tk
Hence there exists no decomposition of H into summands 1−t
3 and
tk
1
containing 1−t4 .
1−t4
H(t) :=
Proof. Obviously we have h4r ≥ 1 for all r ≥ 0, but the 4–critical
inequality
h−4 + h1 ≤ h4 + h5
does not hold strictly. Therefore
1
H4 (t) := H(t) −
= 0 + t + 0t2 + 0t3 + 0t4 + 0t5 + t6 + · · ·
1 − t4
is nonnegative, but does not satisfy condition (⋆), and so ν(H4 ) = 0;
t
t
the latter is easily seen, since neither H4 (t) − 1−t
3 nor H4 (t) − 1−t4 is
nonnegative.
3.5. Remarks. i) We point out that Theorem 3.13 is also valid in the
degenerate case α = 1. Since the semigroup h1, βi = N0 has no gaps at
all, condition (⋆) collapses to the single inequality hn ≤ hn+β ∀n ∈ Z.
This criterion could be deduced alternatively by applying [1, Thm. 2.1]
to the nonnegative series (1 − tβ )HM (t).
ii) The case of deg(X) and deg(Y ) having a common divisor > 1 can
be reduced to the case of coprime degrees by standard methods. Hence
Theorem 3.13 provides a criterion for positive Hilbert depth also in the
general case:
HILBERT DEPTH OF GRADED MODULES OVER POLYNOMIAL RINGS 27
Let deg(X) = α′ = αδ and deg(Y ) = β ′ = βδ with δ >L
1 and
gcd(α, β) = 1. Any finitely generated graded R–module M = n Mn
decomposes into a direct sum of Veronese submodules
M=
M
M (k) , where M (k) :=
∞
M
Mnδ+k , k = 0, . . . , δ − 1.
n=0
k
We change the grading of R and M (k) by setting Rnδ resp. Mnδ+k as
the nth component in the new grading. Then M (k) is still a graded
R–module. Rewriting the conditions for positive Hilbert depth given
by Theorem 3.13 in terms of the original grading yields
P
P
(⋆k )
i∈I hid+nδ+k ≤
j∈J hjd+nδ+k for all n ∈ Z, [I, J] ∈ Fα,β .
Since
Hdep(M) = min Hdep M (k) ,
k
we have Hdep(M) > 0 if and only if HM satisfies conditions (⋆k ) for
k = 0, . . . , δ − 1.
iii) Theorem 3.13 also holds for modules over a larger polynomial ring
F[X1 , . . . , Xr , Y1, . . . , Ys ] where each variable is assigned one of two coprime degrees: The proof given above can be extended to a proof by
induction on the dimension of the module, since a reductive step similar
to (14) also works for higher dimensions.
iv) Let M be a finitely generated graded R–module of positive depth.
As explained above, Theorem 2.1 implies that HM satisfies condition
(⋆), but from this argument it is not immediately clear why the existence of an M–regular element forces these inequalities. The only
obvious exception is the minimal inequality hn ≤ hn+αβ . There is
also an alternative explanation for one special inequality with maximal
number of terms: The condition
hn + hn+1 + · · · + hn+α−1 ≤ hn+β + hn+β+1 + · · · + hn+β+α−1
(24)
can be deduced as follows.
Let S = F[U, V ] be the standard graded polynomial ring, then we may
identify R with the subalgebra F[U α , V β ] of S, and in this sense S is a
finite free R–module with basis {U i V j | 0 ≤ i < α, 0 ≤ j < β}. Hence
M̃ := M ⊗R S is a finite graded S–module of the same depth as M
with
!
β−1
α−1
X
X
X
HM̃ (t) =
tj · HM (t) =:
h̃n tn .
ti ·
i=0
j=0
n
HILBERT DEPTH OF GRADED MODULES OVER POLYNOMIAL RINGS 28
Since depthS (M̃) > 0 we have p(M̃) > 0, i. e. h̃n ≤ h̃n+1 for all n ∈ Z,
and rewriting this inequality in terms of hn yields exactly (24).
References
[1] J. Uliczka, Remarks on Hilbert Series of Graded Modules over Polynomial
Rings, Manuscr. math. 132 (2010) 159–168.
[2] M. Kreuzer, L. Robbiano, Computational Commutative Algebra 2, Springer,
2005.
[3] L. Smith, Polynomial Invariants of Finite Groups, second edition, A. K. Peters
Ltd., Wellesley, MA, 1997.
[4] R. P. Stanley, Enumerative Combinatorics vol. I, Wadsworth & Brooks/Cole,
1986.
[5] W. Bruns, J. Herzog, Cohen-Macaulay rings, revised edition, Cambridge University Press, Cambridge, 1998.
[6] W. Bruns, Chr. Krattenthaler, and J. Uliczka, Stanley decompositions and
Hilbert depth in the Koszul complex, J. of Comm. Alg. 2 (2010) 327–357.
[7] W. Bruns, B. Ichim, Normaliz: algorithms for rational cones and affine
monoids, J. of Algebra 324 (2010) 1098–1113.
[8] J. C. Rosales, P. A. Garcı́a Sánchez, Numerical semigroups, Dev. in Math. vol.
20, Springer, Berlin-Heidelberg-New York, 2010.
[9] J. C. Rosales, Fundamental gaps of numerical semigroups generated by two
elements, Linear Algebra and its Appl. 405 (2005) 200–208.
[10] S. Ilhan, On Apéry Sets of Symmetric Numerical Semigroups, Int. Math. Forum 1 (2006) 481–484.
Institut für Mathematik, Universität Osnabrück. Albrechtstraße
28a, D-49076 Osnabrück, Germany
E-mail address: [email protected]
Institut für Mathematik, Universität Osnabrück. Albrechtstraße
28a, D-49076 Osnabrück, Germany
E-mail address: [email protected]
| 0math.AC
|
From analytical mechanical problems
to rewriting theory through M. Janet
arXiv:1801.00053v1 [math.HO] 29 Dec 2017
Kenji Iohara
Philippe Malbos
Abstract – This note surveys the historical background of the Gröbner basis theory for D-modules
and linear rewriting theory. The objective is to present a deep interaction of these two fields largely
developed in algebra throughout the twentieth century. We recall the work of M. Janet on the
algebraic analysis on linear partial differential systems that leads to the notion of involutive division.
We present some generalizations of the division introduced by M. Janet and their relations with
Gröbner basis theory.
M.S.C. 2010 – 01-08, 13P10, 12H05, 35A25, 58A15, 68Q42.
1
2
3
4
5
Exterior differential systems
1.1 Pfaff’s problem . . . . . . . . . . . . . . . . . . . . . . . . . .
1.2 The Cartan-Kähler theory . . . . . . . . . . . . . . . . . . . . .
Monomial partial differential equations systems
2.1 Ring of partial differential operators and multiplicative variables
2.2 Completion procedure . . . . . . . . . . . . . . . . . . . . . .
2.3 Inverse of derivation . . . . . . . . . . . . . . . . . . . . . . .
Monomial involutive bases
3.1 Involutive division . . . . . . . . . . . . . . . . . . . . . . . .
3.2 Involutive completion procedure . . . . . . . . . . . . . . . . .
3.3 Others involutive approaches . . . . . . . . . . . . . . . . . . .
Polynomial partial differential equations systems
4.1 Parametric and principal derivatives . . . . . . . . . . . . . . .
4.2 First order PDE systems . . . . . . . . . . . . . . . . . . . . . .
4.3 Higher-order finite linear PDE systems . . . . . . . . . . . . . .
4.4 Completely integrable higher-order linear PDE systems . . . . .
4.5 Canonical forms of linear PDE systems . . . . . . . . . . . . .
4.6 Reduction of a PDE system to a canonical form . . . . . . . . .
4.7 Algebra, geometry and PDE . . . . . . . . . . . . . . . . . . .
4.8 Involutive systems . . . . . . . . . . . . . . . . . . . . . . . . .
4.9 Conclusive remarks . . . . . . . . . . . . . . . . . . . . . . . .
Polynomial involutive bases
5.1 Involutive reduction on polynomials . . . . . . . . . . . . . . .
5.2 Involutive bases . . . . . . . . . . . . . . . . . . . . . . . . . .
5.3 Involutive bases and Gröbner bases . . . . . . . . . . . . . . . .
. . .
. . .
. . .
. . .
. . .
. . .
. . .
. . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
. . .
. . .
. . .
10
10
12
14
15
18
21
24
24
25
28
29
29
31
32
34
36
38
47
49
51
51
52
56
60
Introduction
Several lectures of the Kobe-Lyon summer school1 recalled a deep interaction between Gröbner bases for
D-modules and linear rewriting theory. The objective of this note is to survey the historical background
of these two fields largely developed in algebra throughout the twentieth century and to present their
deep relations. Completion methods are the main streams for these computational theories. In Gröbner
bases theory, they were motivated by algorithmic problems in elimination theory such as computations
in quotient polynomial rings modulo an ideal, manipulating algebraic equations and computing Hilbert
series. In rewriting theory, they were motivated by computation of normal forms and linear basis for
algebras and computational problems in homological algebra.
In this note we present the precursory ideas of the french mathematician M. Janet on the algebraic
formulation of completion methods for polynomial systems. Indeed, the problem of completion already
appear in the seminal work of M. Janet in 1920 in his thesis [Jan20], that proposed a very original
approach by formal methods in the study of linear partial differential equations systems, PDE systems
for short. Its constructions were formulated in terms of polynomial systems, but without the notion of
ideal and of Noetherian induction. These two notions were introduced by E. Noether in 1921 [Noe21]
for commutative rings.
The work of M. Janet was forgotten for about a half-century. It was rediscovered by F. Schwarz
in 1992 in [Sch92]. Our exposition in this note does not follow the historical order. The first section
deals with the problems that motivate the questions on PDE undertaken by M. Janet. In Section 2, we
present completion for monomial PDE systems as introduced by Janet in his monograph [Jan29]. This
completion used an original division procedure on monomials. In Section 3, we present axiomatisation
of this Janet’s division, called involutive division, and due to V. P. Gerdt. The last two sections concern
the case of polynomial PDE systems, with the Janet’s completion method used to reduce a linear PDE
system to a canonical form and the axiomatisation of the reductions involved in terms of rewriting theory.
From analytical mechanical problems to involutive division
From Lagrange to Janet. The analysis on linear PDE systems was mainly motivated in 18th century
by resolution of analytical mechanical problems. The seminal work of J.-L. Lagrange gave the first
systematic study of PDE systems launched by such problems. The case of PDE of one unknown function
of several variables has been treated by J. F. Pfaff. The Pfaff problem will be recalled in 1.1. This
theory was developed in two different directions: toward the general theory of differential invariants and
the existence of solutions under given initial conditions. The differential invariants approachs will be
discussed in 1.1 and 1.1.4. The question of the existence of solution satisfying some initial conditions
was formulated in the Cauchy-Kowalevsky theorem recalled in 1.1.3.
Exterior differential systems. Following the work of H. Grassmann in 1844 exhibiting the rules of the
exterior algebra computation, É. Cartan introduced exterior differential calculus in 1899. This algebraic
calculus allowed him to describe a PDE system by an exterior differential system that is independent of the
1
2
Summer School On Quivers : Computational Aspects and Geometric Applications, July 21-31, 2015, Kobe, Japan.
choice of coordinates. This leaded to the so called Cartan-Kähler theory, that we will review in Section
1.2. We will present a geometrical property of involutivity on exterior differential systems in 1.2.6, that
motivates the formal methods introduced by M. Janet for analysis on linear PDE systems.
Generalizations of Cauchy-Kowalevsky’s theorem. Another origin of the work of M. Janet is the
Cauchy-Kowalevsky’s theorem that gives the initial conditions of solvability of a family of PDE systems
that we describe in 1.1.3. É. Delassus, C. Riquier and M. Janet attempted to generalize this result to a
wider class of linear PDE systems which in turn leaded them to introduce the computation of a notion of
normal form for such systems.
The Janet monograph. Section 2 presents the historical motivations that leaded M. Janet to introduce
an algebraic algorithm in order to compute normal form of linear PDE systems. In particular, we
recall the problem of computation of inverse of derivation introduced in the monograph of M. Janet,
! Leçons sur les systèmes d’équations aux dérivées partielles " on the analysis on linear PDE systems,
published in 1929, [Jan29]. In this monograph M. Janet introduced formal methods based on polynomial
computations for analysis on linear PDE systems. He developed an algorithmic approach for analyzing
ideals in the polynomial ring K[ BxB1 , . . . , BxBn ] of differential operators with constant coefficients. Having
the ring isomorphism between this ring and the ring K[x1 , . . . , xn ] of polynomials with n variables
in mind, M. Janet gave its algorithmic construction in this latter ring. He began by introducing some
remarkable properties of monomial ideals. In particular, he recovered the Dickson’s Lemma, [Dic13], on
the finiteness generation of monomial ideal. This property is essential for Noetherian properties on the
set of monomials. Note that, M. Janet wasn’t familiar with the axiomatisation of the algebraic structure
of ideal and the property of Noetherianity already introduced by E. Noether in [Noe21] and [Noe23].
Note also that the Dickson lemma was published in 1913 in a paper on numbers theory in an American
journal. Due to the first world war, it would take long times before these works were accessible to french
mathematical community. The Janet’s algebraic constructions given in his monograph will be recalled in
Section 2 for monomial systems and in Section 4 for polynomial systems.
Janet’s multiplicative variables. The computation on monomial and polynomial ideals performed by
M. Janet are founded on the notion of multiplicative variable that he introduced in his thesis, [Jan20].
Given an ideal generated by a set of monomials, he distinguished the family of monomials contained
in the ideal and those contained in the complement of the ideal. The notion of multiplicative and nonmultiplicative variables appear in order to stratify these two families of monomials. We will recall this
notion of multiplicativity on variables in 2.1.10. This leads to a refinement of the classical division on
monomials, nowadays called Janet’s division.
Involutive division and Janet’s completion procedure. The notion of multiplicative variable is local
in the sense that it is defined with respect to a subset U of the set of all monomials. A monomial u in U
is said to be a Janet divisor of a monomial w with respect to U, if w = uv and all variables occurring
in v are multiplicative with respect to U. In this way, we distinguish the set coneJ (U) of monomials
having a Janet divisor in U, called multiplicative or involutive cone of U, to the set cone(U) of multiple of
monomials in U for the classical division. The Janet division being a refinement of the classical division,
the set coneJ (U) is a subset of cone(U). M. Janet called a set of monomials U complete when this
inclusion is an equality.
3
For a monomial PDE system (Σ) of the form
Bα1 +...+αn
ϕ = fα (x1 , x2 , . . . , xn ),
Bxα1 1 . . . Bxαnn
where (α1 , . . . , αn ) belongs to a subset I of Nn , M. Janet associated the set of monomials
lm(Σ) = {xα1 1 . . . xαnn | (α1 , . . . , αn ) P I}. The compatibility conditions of the system (Σ) corresponds to the factorizations of the monomials ux in coneJ (lm(Σ)), where u is in lm(Σ) and x is a
non-multiplicative variable of u with respect to lm(Σ), in the sense given in 2.3.1. By definition, for any
monomial u in lm(Σ) and x non-multiplicative variable of u with respect to lm(Σ), the monomial ux
admits such a factorization if and only if lm(Σ) is complete, see Proposition 2.2.5.
The main procedure presented in Janet’s monograph [Jan29] completes in finite number of operations
r that contains U. This procedure consists
a finite set of monomials U into a complete set of monomials U
in analyzing all the local default of completeness, by adding all the monomials ux where u belongs to U
and x is a non-multiplicative variable for u with respect to U. This procedure will be recalled in 2.2.9.
A generalization of this procedure to any involutive division was given by V. P. Gerdt in [Ger97], and
recalled in 3.2.12.
Extending this procedure to a set of polynomials, M. Janet applied it to linear PDE systems, giving
a procedure that transforms a linear PDE system into a complete PDE system having the same set of
solutions. This construction is presented in Section 4.6. In Section 5, we present such a procedure for
an arbitrary involutive division given by V. P. Gerdt and Y. A. Blinkov in [GB98a] and its relation to the
Buchberger completion procedure in commutative polynomial rings, [Buc65].
The space of boundary conditions. In order to stratify the complement of the involutive cone coneJ (U)
M. Janet introduced the notion of complementary monomial, see 2.1.9, as the monomials that generate
this complement in a such a way that the involutive cone of U and the involutive cone of the set UK of
complementary monomials form a partition of the set of all monomials, see Proposition 2.2.2.
For each complementary monomial v in lm(Σ)K each analytic function in the multiplicative variables of v with respect to lm(Σ)K provides a boundary condition of the PDE system (Σ) as stated by
Theorem 2.3.3.
Polynomial partial differential equations systems. In Section 4, we present the analysis on polynomial
PDE systems as M. Janet described in his monograph, [Jan29]. To deal with polynomials he defined
some total orders on the set of derivatives, corresponding to total orders on the set of monomials. We
recall them in Section 4.1. The definitions on monomial orders given by M. Janet clarified the same
notion previously introduced by C. Riquier in [Riq93]. In particular, he made more explicit the notion of
parametric and principal derivatives in order to distinguish the leading derivative in a polynomial PDE. In
this way, he extended its algorithms on monomial PDE systems to the case of polynomial PDE systems.
In particular, using these notions, he defined the property for a polynomial PDE system to be complete.
Namely, a polynomial PDE system is complete if the associated set of monomials corresponding to
leading derivatives of the system is complete. Moreover, he extended also the notion of complementary
monomials to define the notion of boundary condition for a polynomial PDE system as in the monomial
case.
Boundary conditions. In this way, the notion of completeness is a suitable framework to discuss the
existence and the unicity of the boundary conditions for a linear PDE system. M. Janet proved that if a
4
linear polynomial PDE system of the form
Di ϕ =
ÿ
ai,j Di,j ϕ,
i P I,
j
of one unknown function ϕ and all the functions ai,j are supposed to be analytic in a neighborhood of a
point P in Cn and is complete with respect to some a total order, then it admits at most one analytic solution
satisfying the boundary condition formulated in terms of complementary monomials, see Theorems 4.3.4
and 4.3.6.
Integrability conditions. A linear polynomial PDE system of the above form is said to be completely
integrable if it admits an analytic solution for any given boundary condition. M. Janet gave an algebraic
characterization of complete integrability by introducing integrability conditions formulated in terms
of factorization of leading derivative of the PDE by non-multiplicative variables. These integrability
conditions are given explicitly in 4.4.4 as generalization to the polynomial situation of the integrability
conditions formulated above for monomial PDE systems in Subsection 2.3. M. Janet proved that a linear
polynomial PDE system is completely integrable if and only if any integrability condition is trivial, as
stated in Theorem 4.4.7.
Janet’s procedure of reduction of linear PDE systems to a canonical form. In order to extend
algorithmically Cauchy-Kowalevsky’s theorem on the existence and uniqueness of solutions of initial
condition problems as presented in 1.1.3, M. Janet considered normal forms of linear PDE systems with
respect to a suitable total order on derivatives, satisfying some analytic conditions on coefficients and a
complete integrability condition on the system, as defined in 4.5.2. Such normal forms of PDE systems
are called canonical by M. Janet.
Janet’s procedure, recalled in 7, decides if a linear PDE system can be transformed into a completely
integrable system. If the system cannot be reduced to a canonical form, the procedure returns the
obstructions of the system to be transformed into a completely integrable system. This procedure
depends on a total order on derivatives of unknown functions of the PDE system. For this purpose,
M. Janet introduced a general method to define a total order on derivatives using a parametrization of
a weight order on variables and unknown functions, as recalled in 4.1.5. The Janet procedure uses a
specific weight order called canonical and defined in 4.6.2.
The first step of Janet’s procedure consists in applying autoreduction procedure, defined in 4.6.4, in
order to reduce any PDE of the system with respect to the total order on derivatives. Namely two PDE of
the system cannot have the same leading derivative, and any PDE of the system is reduced with respect
to the leading derivatives of the others PDE, as defined in 5.
The second step consists in applying the completion procedure, defined in 6. That is, the set of leading
derivatives of the system defines a complete set of monomials in the sense given in 4.3.2.
Having transformed the PDE system to an autoreduced and complete system, one can discuss about its
integrability conditions. M. Janet shown that this set of integrability conditions is a finite set of relations
that does not contain principal derivative, as explained in 4.4.4. Hence, these integrability conditions are
J-normal forms and uniquely defined. By Theorem 4.4.7, if all of these normal forms are trivial, then the
system is completely integrable. Otherwise, if there is a non-trivial condition in the set of integrability
conditions that contains only unknown functions and variables, then this relation imposes a relation on
the boundary conditions of the system, else if there is no such relation, the procedure is applied again on
5
the PDE system completed by all the integrability conditions. Note that this procedure depends on the
Janet division and on a total order on the set of derivatives.
By this algorithmic method, M. Janet has generalized in certain cases Cauchy-Kowalevsky’s theorem
at the time where the algebraic structures have not been introduced to compute with polynomial ideals.
This is pioneering work in the field of formal approaches to analysis on PDE systems. Algorithmic
methods to deals with polynomial ideals were developed throughout the twentieth century and extended
to wide range of algebraic structures. In the next subsection, we present some milestones on these formal
mathematics.
Constructive methods and rewriting in algebra through the twentieth century
The constructions developed by M. Janet in his formal theory of linear partial differential equation systems
are based on the structure of ideal, that he called module of forms. This notion corresponds to those
introduced previously by D. Hilbert in [Hil90] with the terminology of algebraic form. Notice that N.
M. Gunther dealt with such a structure in [Gün13b]. The axiomatization of the notion of ideal on an
arbitrary ring were given by E. Noether in [Noe21]. As we will explain in this note, M. Janet introduced
algorithmic methods to compute a family of generators of an ideal having the involutive property and
called involutive bases. This property is used to obtain a normal form of linear partial differential equation
systems.
Janet’s computation of involutive bases is based on a refinement of classical polynomial division
called involutive division. He defined a division that was suitable for reduction of linear partial differential
equation systems. Thereafter, other involutive divisions were introduced in particular by J. M. Thomas
[Tho37] and by J.-F. Pommaret [Pom78].
The main purpose is to complete a generating family of an ideal into an involutive bases with respect
to a given involutive division. This completion process is quite similar to those introduced with the
classical division in Gröbner bases theory. In fact, involutive bases appears to be particular cases of
Gröbner bases. The principle of completion had been developed independently in rewriting theory,
that proposes a combinatorial approach of equivalence relation motivated by several computational and
decision problems in algebra, computer science and logic.
Some milestones on algebraic rewriting and constructive algebra. The main results in the work of
M. Janet rely on constructive methods in linear algebra using the principle of computing normal forms by
rewriting and the principle of completion of a generating set of an ideal. These two principles have been
developed during all of the twentieth century in many algebraic contexts with different formulations and
at several occasions. We review below some important milestones in this long and wealth history from
Kronecker to the more recent developments.
1882. L. Kronecker introduced the notion of resultant of polynomials in [Kro82] and gave the first result
in elimination theory using this notion.
1886. K. Weierstrass proved a fundamental result called preparation theorem on the factorization of
analytic functions by polynomials. As an application he showed a division theorem for rings of
convergent series, [Wei86].
6
1890. D. Hilbert proved that any ideals of a ring of commutative polynomials on a finite set of variables
over a field and ring of integers are finitely generated, [Hil90]. This is the first formulation of the
Hilbert basis theorem stating that a polynomial ring over a Noetherian ring is Noetherian.
1913. In a paper on number theory, L.E. Dickson proved a monomial version of the Hilbert basis theorem
by a combinatorial method, [Dic13, Lemma A].
1913. In a serie of forgotten papers, N. Günther develop algorithmic approaches for polynomials rings,
[Gün13c, Gün13b, Gün13a]. A review of the Günther theory can be found in [Gun41].
1914. M. Dehn described the word problem for finitely presented groups, [Deh10]. Using systems of
transformations rules, A. Thue studied the problem for finitely presented semigroups, [Thu14]. It
was only much later in 1947, that the problem for finitely presented monoids was shown to be
undecidable, independently by E. L. Post [Pos47] and A. Markov [Mar47a, Mar47b].
1916. F. S. Macaulay was one of the pioneers in commutative algebra. In his book The algebraic theory
of modular systems, [Mac16], following the fundamental Hilbert basis theorem, he initiated an
algorithmic approach to treat generators of polynomial ideals. In particular, he introduced the
notion of H-basis corresponding to a monomial version of Gröbner bases.
1920. M. Janet defended his doctoral thesis, [Jan20], that presents a formal study of systems of partial
differential equations following works of Ch. Riquier and É. Delassus. In particular, he analyzed
completly integrable systems and Hilbert functions on polynomial ideals.
1921. In her seminal paper, Idealtheorie in Ringbereichen, [Noe21], E. Noether gave the foundation of
general commutative ring theory, and gave one of the first general definitions of a commutative
ring. She also formulated the theorem of finite chains [Noe21, Satz I, Satz von der endlichen Kette].
1923. E. Noether stated in [Noe23, Noe24] concepts of elimination theory in the language of ideals that
she had introduced in [Noe21].
1926. G. Hermann, a student of E. Noether [Her26], initiated purely algorithmic approaches on ideals,
such as ideal membership problem and primary decomposition ideals. This work appears as a
fundamental contribution for emergence of computer algebra.
1927. F. S. Macaulay showed in [Mac27] that the Hilbert function of a polynomial ideal I is equal to the
Hilbert function of the monomial ideal generated by the set of leading monomials of polynomials
in I with respect a monomial order. As a consequence the coefficients of the Hilbert function of a
polynomial ideal are polynomial for sufficiently big degree.
1937. Based on early works by Ch. Riquier and M. Janet, in [Tho37] J. M. Thomas reformulated in
the algebraic language of B. L. van der Waerden, Moderne Algebra, [van30, van31], the theory of
normal forms of systems of partial differential equations.
1937. In [Grö37], W. Gröbner formulated the isomorphism between the ring of polynomials with coefficients in an arbitrary field and the ring of differential operators with constant coefficients, see
Proposition 2.1.2. The identification of these two rings was used before in the algebraic study of
systems of partial differential equations but without being explicit.
7
1942. In a seminal paper on rewriting theory, M. Newman presented rewriting as a combinatorial
approach to study equivalence relations, [New42]. He proved a fundamental rewriting result stating
that under termination hypothesis, the confluence properties is equivalent to local confluence.
1949. In its monograph Moderne algebraische Geometrie. Die idealtheoretischen Grundlagen, [Grö49],
W. Gröbner surveyed algebraic computation on ideal theory with applications to algebraic geometry.
1962. A. Shirshov introduced in [Shi62] an algorithmic method to compute normal forms in a free Lie
algebra with respect to a family of elements of the Lie algebra satisfying a confluence property, the
method is based on a completion procedure and he proved a version of Newman’s lemma for Lie
alegbras, called composition lemma. He deduced a constructive proof of the Poincaré-Birkhoff-Witt
theorem.
1964. H. Hironaka introduced in [Hir64] a division algorithm and introduced the notion of standard
basis, that is analogous to the notion of Gröbner basis, for power series rings in order to solve
problems of resolution of singularities in algebraic geometry.
1965. Under the supervision of W. Gröbner, B. Buchberger developed in his PhD thesis the algorithmic
theory of Gröbner bases for commutative polynomial algebras, [Buc65, Buc70, Buc06]. Buchberger
gave a characterization of Gröbner bases in terms of S-polynomials and an algorithm to compute
such bases, with a complete implementation in the assembler language of the computer ZUSE Z
23 V.
1967. D. Knuth and P. Bendix defined in [KB70] a completion procedure that complete with respect to
a termination a set of equations in an algebraic theory into a confluent term rewriting system. The
procedure is similar to the Buchberger’s completion procedure. We refer the reader to [Buc87] for
an historical account on critical-pair/completion procedures.
1972. H. Grauert introduced in [Gra72] a generalization of Weierstrass’s preparation division theorem
in the language of Banach algebras.
1973. M. Nivat formulated a critical pair lemma for string rewriting systems and proved that for a
terminating rewriting system, the local confluence is decidable, [Niv73].
1976, 1978. L. Bokut in [Bok76] and G. Bergman in [Ber78] extended Gröbner bases and Buchberger
algorithm to associative algebras. They obtained the confluence Newman’s Lemma for rewriting
systems in free associative algebras compatible with a monomial order, called respectively Diamond
Lemma for ring theory and composition Lemma.
1978. J.-F. Pommaret introduced in [Pom78] a global involutive division simpler than those introduced
by M. Janet.
1980. F.-O. Schreyer in his PhD thesis [Sch80] gave a method that computes syzygies in commutative
multivariate polynomial rings using the division algorithm, see [Eis95, Theorem 15.10].
1980. G. Huet gave in [Hue80] a proof of Newman’s lemma using a Noetherian well-founded induction
method.
8
1985. Gröbner basis theory was extended to Weyl algebras by A. Galligo in [Gal85].
1997. V. P. Gerdt introduced in [Ger97] an axiomatization of the involutive monomial division.
1999, 2002. J.-C. Faugère developed efficient algorithms for computing Gröbner bases, algorithm F4,
[Fau99] then and algorithm F5, [Fau02].
Conventions and notations
The set of non-negative integers is denoted by N. In this note, K[x1 , . . . , xn ] denotes the polynomial
ring on the variables x1 , . . . , xn over a field K of characteristic zero. For a subset G of polynomials
of K[x1 , . . . , xn ], we will denote by Id(G) the ideal of K[x1 , . . . , xn ] generated by G. A polynomial is
either zero or it can be written as a sum of a finite number of non-zero terms, each term being the product
of a scalar in K by a monomial.
Monomials. We will denote by M(x1 , . . . , xn ) the set of monomials in the ring K[x1 , . . . , xn ]. For a
subset I of {x1 , . . . , xn } we will denote by M(I) the set of monomials in M(x1 , . . . , xn ) whose variables
lie in I. A monomial u in M(x1 , . . . , xn ) is written as u = xα1 1 . . . xαnn , were the αi are non-negative
integers. The integer αi is called the degree of the variable xi in u, it will be also denoted by degi (u).
For α = (α1 , . . . , αn ) in Nn , we denote xα = xα1 1 . . . xαnn and |α| = α1 + . . . + αn .
For a set U of monomials of M(x1 , . . . , xn ) and 1 ď i ď n, we denote by degi (U) the largest possible
degree in variable xi of the monomials in U, that is
degi (U) = max degi (u) | u P U .
We call the cone of the set U the set of all multiple of monomials in U defined by
ď
cone(U) =
uM(x1 , . . . , xn ) = { uv | u P U, v P M(x1 , . . . , xn ) }.
uPU
Homogeneous polynomials. An homogenous polynomial of K[x1 , . . . , xn ] is a polynomial all of whose
non-zero terms have the same degree. An homogenous polynomial is of degree p all of whose non-zero
terms have degree p. We will denote by K[x1 , . . . , xn ]p the space of homogenous polynomials of degree
p. The dimension of this space is given by the following formula:
(p + 1)(p + 2) . . . (p + n − 1)
.
Γnp := dim K[x1 , . . . , xn ]p =
1 ¨ 2 ¨ . . . ¨ (n − 1)
Monomial order. Recall that a monomial order on M(x1 , . . . , xn ) is a relation ď on M(x1 , . . . , xn )
satisfying the following three conditions
i) ď is a total order on M(x1 , . . . , xn ),
ii) ď is compatible with multiplication, that is, if u ď u 1 , then uw ď u 1 w for any monomials u, u 1 , w
in M(x1 , . . . , xn ),
iii) ď is a well-order on M(x1 , . . . , xn ), that is, every nonempty subset of M(x1 , . . . , xn ) has a smallest
element with respect to ď.
9
1. Exterior differential systems
The leading term, leading monomial and leading coefficient of a polynomial f of K[x1 , . . . , xn ], with
respect to a monomial order ď, will be denoted respectively by ltď (f), lmď (f) and lcď (f). For a set F of
polynomials in K[x1 , . . . , xn ], we will denote by lmď (F) the set of leading monomials of the polynomials
in F. For simplicity, we will use notations lt(f), lm(f), lc(f) and lm(F) if there is no possible confusion.
1. Exterior differential systems
Motivated by problems in analytical mechanics, L. Euler (1707 - 1783) and J.-L. Lagrange (1736 - 1813)
initiated the so-called variational calculus, cf. [Lag88], which led to the problem of solving partial
differential equations, PDE for short. In this section, we briefly explain the evolutions of these theory to
serve as a guide to the M. Janet contributions. We present the historical background of exterior differential
systems and of the questions on PDE. For a deeper discussion of the theory of differential equations and
the Pfaff problem, we refer the reader to [For90, vW00] or [Car99].
1.1. Pfaff’s problem
1.1.1. Partial differential equations for one unknown function. In 1772, J.-L. Lagrange [Lag72]
considered a PDE of the following form
F(x, y, ϕ, p, q) = 0 with
p=
Bϕ
Bx
and
q=
Bϕ
,
By
(1)
i.e. a PDE of one unknown function ϕ of two variables x and y. Lagrange’s method to solve this PDE
can be summarized as follows.
i) Express the PDE (1) in the form
q = F1 (x, y, ϕ, p)
ii) ‘Temporally, forget the fact p =
Bϕ
Bx ’
with p =
Bϕ
Bx
and
q=
Bϕ
.
By
(2)
and consider the following 1-form
Ω = dϕ − pdx − qdy = dϕ − pdx − F1 (x, y, ϕ, p)dy,
by regarding p as some (not yet fixed) function of x, y and ϕ.
iii) If there exist functions M and Φ of x, y and ϕ satisfying MΩ = dΦ, then Φ(x, y, ϕ) = C for some
constant C. Solving this new equation, we obtain a solution ϕ = ψ(x, y, C) to the equation (2).
1.1.2. Pfaffian systems. In 1814-15, J. F. Pfaff (1765 - 1825) [Pfa02] has treated a PDE for one unknown
function of n variables, which was then succeeded to C. G. Jacobi (1804 - 1851) (cf. [Jac27]). Recall
that any PDE of any order is equivalent to a system of PDE of first order. Thus we may only think of
system of PDE of first order with m unknown function
Fk x1 , . . . , xn , ϕ1 , . . . , ϕm ,
10
Bϕa
(1 ď a ď m, 1 ď i ď n) = 0,
Bxi
for
1 ď k ď r.
1.1. Pfaff’s problem
Introducing the new variables pai , the system is defined on the space with coordinates (xi , ϕa , pai ) and is
given by
Fk (xi , ϕa , pai ) = 0,
n
ÿ
dϕa − pai dxi = 0,
i=1
dx ∧ . . . ∧ dx ‰ 0.
n
1
Noticed that the last condition means that the variables x1 , . . . , xn are independent. Such a system is
called a Pfaffian system. One is interested in the questions, whether this system admits a solution or not,
and if there exists a solution whether it is unique under some conditions. These questions are Pfaff’s
problems.
1.1.3. Cauchy-Kowalevsky’s theorem. A naive approach to Pfaff’s problems, having applications to
mechanics in mind, is the question of the initial conditions. In series of articles published in 1842, A.
Cauchy (1789 - 1857) studied the system of PDE of first order in the following form:
m ÿ
n
ÿ
Bϕa
Bϕb
= fa (t, x1 , ¨ ¨ ¨ , xn ) +
, for 1 ď a ď m,
fia,b (t, x1 , . . . , xn )
Bt
Bxi
b=1 i=1
where fa , fia,b and ϕ1 , . . . , ϕm are functions of n+1 variables t, x1 , . . . , xn . S. Kowalevsky (1850 - 1891)
[von75] in 1875 considered the system of PDE in the following form: for some ra P Zą0 (1 ď a ď m),
m
ÿ
Bra ϕa
=
Btra
b=1
rÿ
a −1
fj,α
a,b (t, x1 , . . . , xn )
j=0
j+|α|ďra
Bj+|α| ϕb
,
Btj Bxα
1
m are functions of n + 1 variables t, x , . . . , x , and where for
where, fj,α
n
1
a,b and ϕ , . . . , ϕ
ř
n
α = (α1 , ¨ ¨ ¨ , αn ) in (Zě0 ) , we set |α| = ni=1 αi and Bxα = Bxα1 1 . . . Bxαnn . They showed that
under the hypothesis on the analyticity of the coefficients, such a system admits a unique analytic local
solution satisfying a given initial condition, that is now called the Cauchy-Kowalevsky theorem.
1.1.4. Completely integrable systems. A first geometric approach to this problem was taken over by
G. Frobenius (1849 - 1917) [Fro77] and independently by G. Darboux (1842 - 1917) [Dar82]. Let X be
a differentiable manifold of dimension n. We consider the Pfaffian system:
ωi = 0
1 ď i ď r,
where ωi ’s are 1-forms defined on a neighbourhood V of a point x in X. Suppose that the family
{(ωi )y }1ďiďr Ă Ty˚ X
is linearly independent for y in V. For 0 ď p ď n, let us denote by ΩpX (V) the space of differentiable
p-forms on V. A p-dimensional distribution D on X is a subbundle of TX whose fibre is of dimension p.
A distribution D is involutive if, for any vector field ξ and η taking values in D, the Lie bracket
[ξ, , η] := ξη − ηξ
takes values in D as well. Such a Pfaffian system is called completely À
integrable.
p
n
G. Frobenius and G. Darboux showed that the ideal I of
p=0 ΩX (V), generated by the
1-forms ω1 , . . . , ωr is a differential ideal, i.e. dI Ă I, if and only if the distribution D on V defined as
the annihilator of ω1 , . . . , ωr is involutive.
11
1. Exterior differential systems
1.2. The Cartan-Kähler theory
Here, we give a brief exposition of the so-called Cartan-Kähler theory from view point of its history. In
particular, we will present the notion of systems in involution. For the expositions by the founders of the
theory, we refer the reader to [Car45] and [Käh34], for a modern introduction by leading experts, we refer
to [BCG+ 91].
1.2.1. Differential forms. H. Grassmann (1809 - 1877), [Gra44], introduced in 1844 the first equational
formulation of the structure of exterior algebra with the anti-commutativity rules,
x ∧ y = −y ∧ x.
Using this notion, É. Cartan (1869 - 1951), [Car99] defined in 1899 the exterior differential and differential
p-form. He showed that these notions are invariant with respect to any coordinate transformation. Thanks
to this differential structures, several results obtained in 19th century were reformulated in a clear manner.
1.2.2. Exterior differential Ť
systems. An exterior differential system Σ is a finite set of homogeneous
differential forms, i.e. Σ Ă p ΩpX . É. Cartan, [Car01], in 1901 studied exterior differential systems
generated by 1-forms, i.e. Pfaffian systems. Later, E. Kähler (1906 - 2000) [Käh34] generalized the
Cartan theory to any differential ideal I generated by an exterior differential system. By this reason, the
general theory on exterior differential systems is nowadays called the Cartan-Kähler theory.
In the rest of this subsection, we briefly describe the existence theorem for such a system. Since the
argument developed here is local and we need the Cauchy-Kowalevsky theorem, we assume that every
function is analytic in x1 , . . . , xn unless otherwise stated.
1.2.3. Integral elements. Let Σ be an exterior differential system on a real analytic manifold X of
dimension n such that the ideal generated by Σ is an differential ideal. For 0 ď p ď n, set Σp = Σ X ΩpX .
We fix x in X. For p ą 0, the pair (Ep , x), for a p-dimensional vector subspace Ep Ă Tx X is called
an integral p-element if ω|Ep = 0 for any ω in Σpx := Σp X ΩpX,x , where ΩpX,x denotes the space of
differentila p-forms defined on a neighbourhood of x P X. We denote the set of integral elements of
dimension p by IΣpx .
An integral manifold Y is a submanifold of X whose tangent space Ty Y at any point y in Y is an
integral element. Since the exterior differential system defined by Σ is completely integrable, there
exists independent r-functions ϕ1 (x), ¨ ¨ ¨ , ϕr (x), called integral of motion or first integral, defined on a
neighbourhood V of a point x P U such that their restrictions on V X Y are constants.
The polar space H(Ep ) of an integral element Ep of Σ at origin x is the vector subspace of Tx X
generated by those ξ P Tx X such that Ep + Rξ is an integral element of Σ.
1.2.4. Regular integral elements. Let E0 be the real analytic subvariety of X defined as the zeros of Σ0
and let U the subset of smooth points. A point in E0 is called integral point. A tangent vector ξ in Tx X
is called linear integral element if ω(ξ) = 0 for any ω P Σ1x with x P U. We define inductively the
properties called "regular" and "ordinary" as follows:
(i) The 0th order character is the integer s0 = maxxPU {dim RΣ1x }. A point x P E0 is said to be regular
if dim RΣ1x = s0 , and a linear integral element ξ P Tx X is called ordinary if x is regular.
12
1.2. The Cartan-Kähler theory
(ii) Set E1 = Rξ, where ξ is an ordinary linear integral element. The 1st order character is the
integer s1 satisfying s0 + s1 = maxxPU {dim H(E1 )}. The ordinary integral 1-element (E1 , x) is
said to be regular if dim H(E1 ) = s0 + s1 . Any integral 2-element (E2 , x) is called ordinary if it
contains at least one regular linear integral element.
(iii) Assume that all these are defined up to (p − 1)th step and that s0 + s1 + ¨ ¨ ¨ + sp−1 ă n − p + 1.
The pth order character is the integer sp satisfying
p
ÿ
si = max {dim H(Ep )}.
xPU
i=0
An integral p-element (Ep , x) is said to be regular if
p
ÿ
si = dim H(Ep ).
i=0
The integral p-element (Ep , x) is said to be ordinary if it contains at least one regular integral
element (Ep−1 , x).
ř
Let h be the smallest positive integer such that hi=0 si = n − h. In such a case, there does not exist an
integral (h + 1)-element. The integer h is called the genus of the system Σ. In such a case, for 0 ă p ď h,
one has
p−1
ÿ
si ď n − p.
i=0
1.2.5. Theorem. Let 0 ă p ď h be an integer.
řp−1
(i) The case i=0
si = n − p : let (Ep , x) be an ordinary integral p-element and let Yp−1 be an
integral manifold of dimension p − 1 such that (Tx Yp−1 , x) is a regular integral (p − 1)-element
contained in (Ep , x). Then, there exists a unique integral manifold Yp of dimension p containing
Yp−1 such that Tx Yp = Ep .
řp−1
(ii) The case i=0
si ă n − p : let (Ep , x) be an integral p-element and let Yp−1 be an integral
manifold of dimension p − 1 such that (Tx Yp−1 , x) is a regular integral (p − 1)-element contained
řp−1
in (Ep , x). Then, for each choice of n − p − i=0
si differentiable functions on x1 , ¨ ¨ ¨ , xp , there
exists a unique integral manifolds Yp of dimension p containing Yp−1 such that Tx Yp = Ep .
This theorem states that a given chain of ordinary integral elements
(E0 , x) Ă (E1 , x) Ă ¨ ¨ ¨ Ă (Eh , x),
dim Ep = p
(0 ď p ď h),
one can inductively find an integral manifold Yp of dimension p such that Y0 = {x}, Yp−1 Ă Yp
and Tx Yp = Ep . Notice that to obtain Yp from Yp−1 , one applies the Cauchy-Kowalevsky theorem to the
system of PDE defined by Σp and the choice of arbitrary differentiable functions in the above statement
provide initial data consisting of
13
2. Monomial partial differential equations systems
1.2.6. Systems in involution. In many applications, the exterior differential systems one considers admit
p-independent variables x1 , . . . , xp . In such a case, we are only interested in the p-dimensional integral
manifolds among which it imposes no additional relation between x1 , . . . , xp . In general, an exterior
differential system Σ for n − p unknown functions and p independent variables x1 , . . . , xp is said to be
in involution if it satisfies the two following conditions
(i) its genus is more than or equal to p,
(ii) the defining equations of the generic ordinary integral p-element introduce no linear relation
among dx1 , . . . , dxp .
1.2.7. Reduced characters. Consider a family F of integral elements of dimensions 1, 2, ¨ ¨ ¨ , p − 1
than can be included in an integral p-element at a generic integral point x P X. Take a local chart of with
origin x. The reduced polar system Hred (Ei ) of an integral element x is the polar system of the restriction
of the exterior differential system Σ to the submanifold
{x1 = x2 = ¨ ¨ ¨ = xp = 0}.
1
, called the reduced characters, are defined in such a way that s01 +s11 +¨ ¨ ¨+si1
The integers s01 , s11 , ¨ ¨ ¨ , sp−1
is the dimension of the reduced polar system Hred (Ei ) at a generic integral element. For convenience,
1
one sets sp1 = n − p − (s01 + s11 + ¨ ¨ ¨ + sp−1
).
Let Σ be an exterior differential system of n − p unknown functions of p independent variables such that
the ideal generated by Σ is an differential ideal. É. Cartan showed that it is a system in involution iff the
most general integeral p-element in F depends upon s11 + 2s21 + ¨ ¨ ¨ + psp1 independent parameters.
1.2.8. Recent developments. In 1957, M. Kuranishi (1924- ), [Kur57], considered the problem of the
prolongation of a given exterior differential system and treated the cases what É. Cartan called total.
Here, M. Kuranishi as well as É. Cartan studied locally in analytic category. After an algebraic approach
to the integrability due to V. Guillemin and S. Sternberg, [GS64], in 1964, I. Singer and S. Sternberg,
[SS65], in 1965 studied some classes of infinite dimensional which is even applicable to C∞ -category.
In 1970’s, with the aid of Jet bundles and the Spencer cohomology, J. F. Pommaret (cf. [Pom78])
reworked on the formal integrable involutive differential systems which generalized works of M. Janet,
in the language of sheaf theory. For other geometric aspects not using sheaf theory, see the books by
P. Griffiths (1938-), [Gri83], and R. Bryant et al., [BCG+ 91].
2. Monomial partial differential equations systems
In this section, we present the method introduced by M. Janet called inverse calculation of the derivation
in his monograph [Jan29]. In [Jan29, Chapter I] M. Janet considered monomial PDE, that is PDE of the
form
Bα1 +α2 +...+αn ϕ
= fα1 α2 ...αn (x1 , x2 , . . . , xn ),
(3)
Bxα1 1 Bxα2 2 . . . Bxαnn
where ϕ is an unknown function and the fα1 α2 ...αn are several variables analytic functions. Its objective
is to compute an analytic function u solution of the system.
14
2.1. Ring of partial differential operators and multiplicative variables
2.1. Ring of partial differential operators and multiplicative variables
2.1.1. Historical context. In the beginning of 1890’s, following collaboration with C. Méray (18351911), C. Riquier (1853-1929) initiated his research on finding normal forms of systems of (infinitely
many) PDE for finitely many unknown functions with finitely many independent variables (see [Riq10]
and [Riq28] for more details).
In 1894, A. Tresse [Tre94] showed that such systems can be always reduced to systems of finitely
many PDE. This is the first result on Noeterianity of a module over a ring of differential operators. Based
on this result, É. Delassus (1868 - 19..) formalized and simplified Riquier’s theory. In these works, one
already finds an algorithmic approach analysing ideals of the ring K[ BxB1 , . . . , BxBn ].
It was M. Janet (1888 - 1983), already in his thesis [Jan20] published in 1920, who had realized that
the latter ring is isomorphic to the ring of polynomials with n variables K[x1 , ¨ ¨ ¨ , xn ] at the time where
several abstract notions on rings introduced by E. Noether in Germany had not been known by M. Janet
in France. It was only in 1937 that W. Gröbner (1899-1980) proved this isomorphism.
2.1.2. Proposition ([Grö37, Sect. 2.]). There exists a ring isomorphism
Φ : K[x1 , . . . , xn ] −→ K[
B
B
,...,
],
Bx1
Bxn
from the ring of polynomials with n variables x1 , . . . , xn with coefficients in an arbitrary field K to the
ring of differential operators with constant coefficients.
2.1.3. Derivations and monomials. M. Janet considers monomials in the variables x1 , . . . , xn and use
implicitly the isomorphism Φ of Proposition 2.1.2. To a monomial xα = xα1 1 xα2 2 . . . xαnn he associates
the differential operator
B|α|
Dα := Φ(xα ) = α1 α2
.
Bx1 Bx2 . . . Bxαnn
In [Jan29, Chapter I], M. Janet considered finite monomial PDE systems. The equations are of the
form (3) and the system having a finitely many equations, the set of monomials associated to the PDE
system is finite. The first result of the monograph is a finiteness result on monomials stating that a
sequence of monomials in which none is a multiple of an earlier one is necessarily finite. He proved this
result by induction on the number of variables. We can formulate this result as follows.
2.1.4. Lemma ([Jan29, §7]). Let U be a subset of M(x1 , . . . , xn ). If, for any monomials u and u 1 in U,
the monomial u does not divide u 1 , then the set U is finite.
This result corresponds to Dickson’s Lemma, [Dic13], which asserts that any monomial ideal
of K[x1 , . . . , xn ] is finitely generated.
2.1.5. Stability of the multiplication. M. Janet paid a special attention to families of monomials with
the following property. A subset of monomial U of M(x1 , . . . , xn ) is called multiplicatively stable if for
any monomial u in M(x1 , . . . , xn ) such that there exists u 1 in U that divides u, then u is in U. In other
words, the set U is closed under multiplication by monomials in M(x1 , . . . , xn ).
As a consequence of Lemma 2.1.4, if U is a multiplicatively stable subset of M(x1 , . . . , xn ), then it
contains only finitely many elements which are not multiples of any other elements in U. Hence, there
exists a finite subset Uf of U such that for any u in U, there exists uf in Uf such that uf divides u.
15
2. Monomial partial differential equations systems
2.1.6. Ascending chain condition. M. Janet observed an other consequence of Lemma 2.1.4: the
ascending chain condition on multiplicatively stable monomial sets that he formulated as follows. Any
ascending sequence of multiplicatively stable subsets of M(x1 , . . . , xn )
U1 Ă U2 Ă . . . Ă Uk Ă . . .
is finite. This corresponds to the Noetherian property on the set of monomials in finitely-many variables.
2.1.7. Inductive construction. Let us fix a total order on variables xn ą xn−1 ą . . . ą x1 . Let U be a
finite subset of M(x1 , . . . , xn ). Let us define, for every 0 ď αn ď degn (U),
[αn ] = {u P U | degn (u) = αn }.
The family ([0], . . . , [degn (U)]) forms a partition of U. We define for every 0 ď αn ď degn (U)
[αn ] = {u P M(x1 , . . . , xn−1 ) | uxαnn P U }.
We set for every 0 ď i ď degn (U)
ď
Ui1 =
{u P M(x1 , . . . , xn−1 ) | there exists u 1 P [αn ] such that u 1 |u }.
0ďαn ďi
We set
and M(U) =
Ť
{ uxkn | u P Uk1 } if k ă degn (U),
Uk =
{ uxkn | u P Ua1 } if k ě degn (U).
Uk . By this inductive construction, M. Janet obtains the monomial ideal generated
kě0
by U. Indeed, M(U) consists in the following set of monomial
{ u P M(x1 , . . . , xn ) | there exists u 1 in U such that u 1 |u }.
2.1.8. Example. Consider the subset U = { x3 x22 , x33 x21 } of monomials in M(x1 , x2 , x3 ). We have
[0] = H,
[1] = {x3 x22 },
[2] = H,
[3] = {x33 x21 }.
[1] = {x22 },
[2] = H,
[3] = {x21 }.
Hence,
[0] = H,
The set M(U) is defined using of the following subsets:
U01 = H,
U11 = {xα1 1 xα2 2 | α2 ě 2},
U21 = U11 ,
U31 = {xα1 1 xα2 2 | α1 ě 2 ou α2 ě 2}.
2.1.9. Complementary monomials. Let U be a finite subset of M(x1 , . . . , xn ). The set of complementary monomials of U is the set of monomial denoted by UK defined by
ď
UK =
UK(i) ,
1ďiďn
where
UK(n) = {xβn | 0 ď β ď degn (U) and [β] = H},
and for every 1 ď i ă n
αi+1
UK(i) = xβi xi+1
. . . xαnn [αi+1 , . . . , αn ] ‰ H, 0 ď β ă degi ([αi+1 , . . . , αn ]), [β, αi+1 , . . . , αn ] = H .
Note that, for i ‰ j we have UK(i) X UK(j) = H.
16
2.1. Ring of partial differential operators and multiplicative variables
2.1.10. Janet’s multiplicative variables, [Jan20, §7]. Let us fix a total order xn ą xn−1 ą . . . ą x1 on
variables. Let U be a finite subset of M(x1 , . . . , xn ). For all 1 ď i ď n, we define the following subset
of U:
[αi , . . . , αn ] = {u P U | degj (u) = αj for all i ď j ď n}.
That is [αi , . . . , αn ] contains monomials of U of the form vxαi i . . . xαnn , with v in M(x1 , . . . , xi−1 ). The
sets [αi , . . . , αn ], for αi , . . . , αn in N, form a partition of U. Moreover, for all 1 ď i ď n − 1, we
have [αi , αi+1 , . . . , αn ] Ď [αi+1 , . . . , αn ] and the sets [αi , . . . , αn ], where αi P N, form a partition
of [αi+1 , . . . , αn ].
Given a monomial u in U, the variable xn is said to be multiplicative for u in the sense of Janet if
degn (u) = degn (U).
For i ď n − 1, the variable xi is said to be multiplicative for u in the sense of Janet if
u P [αi+1 , . . . , αn ]
and
degi (u) = degi ([αi+1 , . . . , αn ]).
We will denote by MultU
J (u) the set of multiplicative variables of u in the sense of Janet with respect to
the set U, also called J-multiplicative variables.
Note that, by definition, for any u and u 1 in [αi+1 , . . . , αn ], we have
U
1
{xi+1 , . . . , xn } X MultU
J (u) = {xi+1 , . . . , xn } X MultJ (u ).
As a consequence, we will denote by MultU
J ([αi+1 , . . . , αn ]) this set of multiplicative variables.
2.1.11. Example. Consider the subset U = {x2 x3 , x22 , x1 } of M(x1 , x2 , x3 ) with the order x3 ą x2 ą x1 .
We have deg3 (U) = 1, hence the variable x3 is J-multiplicative for x3 x2 and not J-multiplicative for x22
and x1 .
For α P N, we have [α] = {u P U | deg3 (u) = α}, hence
[0] = {x22 , x1 },
[1] = {x2 x3 }.
We have deg2 (x22 ) = deg2 ([0]), deg2 (x1 ) ‰ deg2 ([0]) and deg2 (x2 x3 ) = deg2 ([1]), hence the variable x2
is J-multiplicative for x22 and x2 x3 and not J-multiplicative for x1 . We have
[0, 0] = {x1 },
[0, 2] = {x22 },
[1, 1] = {x2 x3 }
and deg1 (x22 ) = deg1 ([0, 2]), deg1 (x1 ) = deg1 ([0, 0]) and deg1 (x3 x2 ) = deg1 ([1, 1]), hence the variable x1 is J-multiplicative for x1 , x22 and x3 x2 .
2.1.12. Janet divisor. Let U be a subset of M(x1 , . . . , xn ). A monomial u in U is called Janet divisor
of a monomial w in M(x1 , . . . , xn ) with respect to U, if there is a decomposition w = uv, where any
variable occurring in v is J-multiplicative with respect to U.
2.1.13. Proposition. Let U be a subset of M(x1 , . . . , xn ) and w be a monomial in M(x1 , . . . , xn ).
Thenw admits in U at most one Janet divisor with respect to U.
17
2. Monomial partial differential equations systems
Proof. If u is a Janet divisor of w with respect to U, there is v in M(MultU
J (u)) such that w = uv. We
have degn (v) = degn (w) − degn (u). If degn (w) ě degn (U), then the variable xn is J-multiplicative
and degn (v) = degn (w) − degn (U). If degn (w) ă degn (U), then xn cannot be J-multiplicative
and degn (v) = 0.
As a consequence, for any Janet divisors u and u 1 in U of w, we have degn (u) = degn (u 1 )
and u, u 1 P [α] for some α P N.
Suppose now that u and u 1 are two distinct Janet divisor of w in U. There exists 1 ă k ď n such
that u, u 1 P [αk , . . . , αn ] and degk−1 (u) ‰ degk−1 (u 1 ). Suppose that degk−1 (u) ą degk−1 (u 1 ), then
the variable xk−1 cannot be J-multiplicative for u 1 with respect to U. It follows that u 1 cannot be a Janet
divisor of w. This leads to a contradiction, hence u = u 1 .
2.1.14. Multiplicative variables of complementary monomials. For any monomial u in UK(n) , we
have
K(n)
MultU
(u) = {x1 , . . . , xn−1 }.
J
For 1 ď i ď n − 1, for any monomial u in UK(i) , there exists αi+1 , . . . , αn such that u P [αi+1 , . . . , αn ].
Then
K(i)
MultJU (u) = {x1 , . . . , xi−1 } Y MultU
J ([αi+1 , . . . , αn ]).
Finally, for u in UK , there exists an unique 1 ď iu ď n such that u P UK(iu ) . Then we set
K(iu )
K
U
MultU
J (u) = MultJ
(u).
2.1.15. Example, [Jan29, p. 17]. Consider the subset U = { x33 x22 x21 , x33 x31 , x3 x2 x31 , x3 x2 } of monomials
in M(x1 , x2 , x3 ) with the order x3 ą x2 ą x1 . The following table gives the multiplicative variables for
each monomial:
x33 x22 x21
x33 x31
x3 x2 x31
x3 x2
x3
x3
x2
x2
x2
x1
x1
x1
The set of complementary monomials are
UK(3) = {1, x23 },
UK(2) = {x33 x2 , x3 },
UK(1) = {x33 x22 x1 , x33 x22 , x33 x21 , x33 x1 , x33 , x3 x2 x21 , x3 x2 x1 }.
The following table gives the multiplicative variables for each monomial:
1, x23
x33 x2
x3
2
3
x3 x2 x1 , x33 x22
x33 x21 , x3 x1 , x33
x3 x2 x21 , x3 x2 x1
x2
x3
x3
x3
x1
x1
x1
x2
x2
2.2. Completion procedure
In this subsection, we present the notion of complete system introduced by M. Janet in [Jan29]. In
particular, we recall the completion procedure that he gave in order to complete a finite set of monomials.
18
2.2. Completion procedure
2.2.1. Complete systems. Let U be a set of monomials of M(x1 , . . . , xn ). For a monomial u in U, M.
Janet defined the involutive cone of u with respect to U as the following set of monomials:
coneJ (u, U) = { uv | v P M(MultU
J (u)) }.
The involutive cone of the set U is defined by
coneJ (U) =
ď
coneJ (u, U).
uPU
M. Janet called complete a set of monomials U when cone(U) = coneJ (U). An involutive cone is called
class in Janet’s monograph [Jan29]. The terminology "involutive" first appear in [Ger97] by V. P. Gerdt
and became standard now. We refer the reader to [Man96] for a discussion on relation between this notion
with the notion of involutivity in the work of É. Cartan, see 1.2.6, and Kroneker, see 1.2.6.
2.2.2. Proposition ([Jan29, p. 18]). For any finite set U of monomials of M(x1 , . . . , xn ), we have the
following partition
M(x1 , . . . , xn ) = coneJ (U) > coneJ (UK ).
2.2.3. A proof of completeness by induction. Let U be a finite set of monomials in M(x1 , . . . , xn ).
We consider the partition [0], . . . , [degn (U)] of monomials in U by their degrees in xn . Let order
α1 ă α2 ă . . . ă αk be the positive integers such that [αi ] is non-empty. Recall that [αi ] is the set of
monomials u in M(x1 , . . . , xn−1 ) such that uxαni is in U. With these notations, the following result gives
an inductive method to prove that a finite set of monomials is complete.
2.2.4. Proposition ([Jan29, p. 19]). The finite set U is complete if and only if the two following conditions
are satisfied:
i) the sets [α1 ], . . . , [αk ] are complete,
ii) for any 1 ď i ă k, the set [αi ] is contains in coneJ ([αi + 1]).
As an immediate consequence of this proposition, M. Janet obtained the following characterisation.
2.2.5. Proposition ([Jan29, p. 20]). A finite set U of monomials of M(x1 , . . . , xn ) is complete if and
only if, for any u in U and any x non-multiplicative variable of u with respect to U, ux is in coneJ (U).
2.2.6. Example, [Jan29, p. 21]. Consider the subset U = { x5 x4 , x5 x3 , x5 x2 , x24 , x4 x3 , x23 } of M(x1 , . . . , x5 ).
The multiplicative variables are given by the following table
x5 x4
x5 x3
x5 x2
x24
x3 x4
x23
x5
x5
x5
x4
x3
x3
x4
x3
x3
x3
x2
x2
x2
x2
x2
x2
x1
x1
x1
x1
x1
x1
In order to prove that this set of monomials is complete, we apply Proposition 2.2.5. The completeness
follows from the identities:
19
2. Monomial partial differential equations systems
x5 x3 .x4 = x5 x4 .x3 ,
x5 x2 .x4 = x5 x4 .x2 , x5 x2 .x3 = x5 x3 .x2 ,
x24 .x5 = x5 x4 .x4 ,
x4 x3 .x5 = x5 x4 .x3 , x4 x3 .x4 = x24 .x3 ,
x23 .x5 = x5 x3 .x3 , x23 .x4 = x4 x3 .x3 .
2.2.7. Examples. For every 1 ď p ď n, the set of monomials of degree p is complete. Any finite set of
monomials of degree 1 is complete.
2.2.8. Theorem (Janet’s Completion Lemma, [Jan29, p. 21]). For any finite set U of monomials
of M(x1 , . . . , xn ) there exists a finite set J(U) satisfying the following three conditions:
i) J(U) is complete,
ii) U Ď J(U),
iii) cone(U) = cone(J(U)).
2.2.9. Completion procedure. From Proposition 2.2.5, M. Janet deduced the completion procedure
Complete(U), Procedure 1, that computes a completion of finite set of monomials U, [Jan29, p. 21].
M. Janet does not give a proof of the termination of this procedure. We will present a proof of the
correction and termination of this procedure in Section 3.2.
Procedure 1: Complete(U)
Input: U a finite set of monomials in M(x1 , . . . , xn )
Output: A finite set J(U) satisfying the condition of Theorem 2.2.8.
begin
r←U
U
r and x P NMultUr (u) such that ux is not in coneJ (U)
r do
while exists u P U
J
Choose such u and x,
r←U
r Y {ux}.
U
end
end
2.2.10. Example, [Jan29, p. 28]. Consider the set U = { x3 x22 , x33 x21 } of monomials of M(x1 , x2 , x3 )
with the order x3 ą x2 ą x1 . The following table gives the multiplicative variables for each monomial:
x33 x21
x3 x22
x3
x2
x2
x1
x1
r ← U Y {x2 x2 }
We complete the set U as follows. The monomial x3 x22 .x3 is not in coneJ (U), we set U
3 2
r
and we compute multiplicative variables with respect to U:
x33 x21
x23 x22
x3 x22
20
x3
x2
x2
x2
x1
x1
x1
2.3. Inverse of derivation
r but x2 x2 .x3 is not in coneJ (U),
r then we set U
r←U
r Y {x3 x2 }. The
The monomial x3 x22 .x3 is in coneJ (U)
3 2
3 2
multiplicative variable of this new set of monomials is
x23 x22
x33 x21
x23 x22
x3 x22
x3
x3
x2
x1
x1
x1
x1
x2
x2
r the other products are in coneJ (U),
r and we prove that the
The monomial x3 x21 .x2 is not in coneJ (U),
system
r = { x3 x2 , x3 x2 , x3 x2 , x3 x2 x2 , x2 x2 }
U
2
3 1
3 2
3
1
3 2
is complete.
2.3. Inverse of derivation
We study resolution of monomial PDE systems of the form
(Σ)
Dα ϕ = fα (x1 , x2 , . . . , xn )
α P Nn ,
(4)
where ϕ is an unknown function and the fα are analytic functions of several variables. As recall in 2.1.1,
an infinite set of partial differential equations can be always reduced to a finite set of such equations. By
this reason, we can assume that the system (Σ) is finite without loss of generality. Using Proposition 2.1.2,
M. Janet associated to each differential operator Dα a monomial xα in M(x1 , . . . , xn ). In this way, to a
PDE system (Σ) on variables x1 , . . . , xn he associated a finite set lm(Σ) of monomials. By Theorem 2.2.8,
any such a set lm(Σ) of monomials can be completed into a finite complete set J(lm(Σ)) having the same
cone as lm(Σ).
2.3.1. Computation of inverse of derivation. Let us now assume that the set of monomials lm(Σ) is
finite and complete. The cone of lm(Σ) being equal to the involutive cone of lm(Σ), for any monomial u
lm(Σ)
in lm(Σ) and non-multiplicative variable xi in NMultJ (u), there exists a decomposition
uxi = vw,
lm(Σ)
where v is in lm(Σ) and w belongs to M(MultJ (v)). For any such a decomposition, it corresponds a
compatibility condition of the PDE system (Σ), that is, for u = xα , v = xβ and w = xγ with α, β and γ
in Nn ,
Bfα
= Dγ fβ .
Bxi
Let us denote by (CΣ ) the set of all such compatibility conditions. M. Janet showed that with the
completeness hypothesis this set of compatibility conditions is sufficient for the PDE system (Σ) to be
integrable.
21
2. Monomial partial differential equations systems
2.3.2. The space of boundary conditions. Let us consider the set lm(Σ)K of complementary monomials
of the finite complete set lm(Σ). Suppose that the PDE system (Σ) satisfies the set (CΣ ) of compatibility
conditions. M. Janet associated to each monomial v = xβ in lm(Σ)K with β P Nn an analytic function
ϕβ (xi1 , . . . , xikv ),
lm(Σ)K
(v). By Proposition 2.2.2, the set of such analytic functions provides
where {xi1 , . . . , xikv } = MultJ
a compatible boundary condition. Under these assumptions, M. Janet proved the following result.
2.3.3. Theorem ([Jan29, p. 25]). Let (Σ) be a finite PDE system such that lm(Σ) is complete. If (Σ)
satisfies the compatibility conditions (CΣ ), then it always admits a unique solution with boundary
conditions given for any v = xβ in lm(Σ)K with β P Nn by
Dβ ϕ
lm(Σ)K
where {xi1 , . . . , xikv } = MultJ
lm(Σ)K
xj =0 @xj PNMultJ
(v)
= ϕβ (xi1 , . . . , xikv ),
(v).
These boundary conditions are called initial conditions by M. Janet. The method to obtain this
boundary conditions is illustrated by the two following examples.
2.3.4. Example, [Jan29, p. 26]. Consider the following PDE system (Σ) of unknown function ϕ of
variables x1 , . . . , x5 :
B2 ϕ
= f1 (x1 , . . . , x5 ),
Bx5 Bx4
B2 ϕ
= f4 (x1 , . . . , x5 ),
Bx24
B2 ϕ
= f2 (x1 , . . . , x5 ),
Bx5 Bx3
B2 ϕ
= f5 (x1 , . . . , x5 ),
Bx4 Bx3
B2 ϕ
= f3 (x1 , . . . , x5 ),
Bx5 Bx2
B2 ϕ
= f6 (x1 , . . . , x5 ).
Bx23
The set (CΣ ) of compatibility relations of the PDE system (Σ) is a consequence of the identities used in
Example 2.2.6 to prove the completeness of the system:
x5 x3 .x4 = x5 x4 .x3 ,
x5 x2 .x4 = x5 x4 .x2 , x5 x2 .x3 = x5 x3 .x2 ,
x24 .x5 = x5 x4 .x4 ,
x4 x3 .x5 = x5 x4 .x3 , x4 x3 .x4 = x24 .x3 ,
x23 .x5 = x5 x3 .x3 , x23 .x4 = x4 x3 .x3 ,
Bf2
Bf1
Bx2 = Bx3 ,
Bf3
Bf1 Bf3
Bx4 = Bx2 , Bx3 =
Bf4
Bf1
Bx5 = Bx4 ,
Bf1
Bf5
Bf5
Bx5 = Bx3 , Bx4 =
Bf6
Bf2
Bf6
Bx5 = Bx3 , Bx4 =
Bf2
Bx2 ,
Bf4
Bx3 ,
Bf5
Bx3 .
The boundary conditions are obtained using the multiplicative variables of the set lm(Σ)K of complementary monomials of lm(Σ). We have
lm(Σ)K(5) = lm(Σ)K(4) = lm(Σ)K(1) = H,
lm(Σ)K(3) = {1, x3 , x4 },
lm(Σ)K(2) = {x5 }.
The multiplicative variables of these monomials are given by the following table
22
2.3. Inverse of derivation
1, x3 , x4
x5
x1 , x2 ,
x1 , x5 .
By Theorem 2.3.3, the PDE system (Σ) admits always a unique solution with any given boundary
conditions of the following type
Bϕ
Bx4
x3 =x4 =x5 =0
Bϕ
Bx3
x3 =x4 =x5 =0
= ϕ0,0,0,1,0 (x1 , x2 )
= ϕ0,0,1,0,0 (x1 , x2 )
ϕ|x3 =x4 =x5 =0 = ϕ0,0,0,0,0 (x1 , x2 )
Bϕ
Bx5
= ϕ0,0,0,0,1 (x1 , x5 ).
x2 =x3 =x4 =0
2.3.5. Example. In a last example, M. Janet considered a PDE system where the partial derivatives of
the left hand side do not form a complete system. The PDE system (Σ) of unknown function ϕ of
variables x1 , x2 , x3
B5 ϕ
B3 ϕ
=
f
(x
,
x
,
x
),
= f2 (x1 , x2 , x3 ).
1
1
2
3
Bx22 Bx3
Bx21 Bx33
We consider the set of monomials lm(Σ) = {x3 x22 , x33 x21 }. In Example 2.2.10, we complete lm(Σ) into the
following complete set of monomials
J(lm(Σ)) = { x3 x22 , x33 x21 , x33 x22 , x33 x2 x21 , x23 x22 }.
The complementary set of monomials are
J(lm(Σ))K(3) = {1},
J(lm(Σ))K(2) = {x23 x2 , x23 , x3 x2 , x3 },
J(lm(Σ))K(1) = {x33 x2 x1 , x33 x2 , x33 x1 , x33 }.
The multiplicative variables of these monomials are given by the following table
J(lm(Σ))K(3)
J(lm(Σ))K(2)
J(lm(Σ))K(1)
x1 , x2 ,
x1 .
x3 .
By Theorem 2.3.3, the PDE system (Σ) admits always a unique solution with any given boundary
conditions of the following type
ϕ|x3 =0 = ϕ0,0,0 (x1 , x2 ),
B2 ϕ
Bx23
B4 ϕ
Bx33 Bx1
= ϕ0,0,2 (x1 ),
x2 =x3 =0
= ϕ1,0,3 (x3 ),
x1 =x2 =0
Bϕ
Bx3
= ϕ0,0,1 (x1 ),
x2 =x3 =0
B3 ϕ
Bx23 Bx2
B4 ϕ
Bx33 Bx2
B2 ϕ
Bx3 Bx2
= ϕ0,1,2 (x1 ),
x2 =x3 =0
= ϕ0,1,3 (x3 ),
x1 =x2 =0
= ϕ0,1,1 (x1 )
x2 =x3 =0
B3 ϕ
Bx33
= ϕ0,0,3 (x3 ),
x1 =x2 =0
B5 ϕ
Bx33 Bx2 Bx1
= ϕ1,1,3 (x3 ).
x1 =x2 =0
23
3. Monomial involutive bases
3. Monomial involutive bases
In this section, we recall a general approach of involutive monomial divisions introduced by V. P. Gerdt
in [Ger97], see also [GB98a, GB98b]. In particular, we give the axiomatic properties of an involutive
division. The partition of variables into multiplicative and non-multiplicative can be deduced from this
axiomatic. In this way, we explain how the notion of multiplicative variable in the sense of Janet can be
deduced from a particular involutive division.
3.1. Involutive division
3.1.1. Involutive division. An involutive division I on the set of monomials M(x1 , . . . , xn ) is defined
by a relation |U
I in U ˆ M(x1 , . . . , xn ), for every finite subset U of M(x1 , . . . , xn ), satisfying, for all
monomials u, u 1 in U and v, w in M(x1 , . . . , xn ), the following six conditions
i) u|U
I w implies u|w,
ii) u|U
I u, for all u in U,
U
U
iii) u|U
I uv and u|I uw if and only if u|I uvw,
1U
U 1
1U
iv) if u|U
I w and u |I w, then u|I u or u |I u,
U
1
1U
v) if u|U
I u and u |I w, then u|I w,
1
U
vi) if U 1 Ď U and u P U 1 , then u|U
I w implies u|I w.
When no confusion is possible, the relation |U
I will be also denoted by |I .
3.1.2. Multiplicative monomial. If u|U
I w, by i) there exists a monomial v such that w = uv. We say
that u is an I-involutive divisor of w, w is an I-involutive multiple of u and v is I-multiplicative for u
with respect to U. When the monomial uv is not an involutive multiple of u with respect to U, we say
that v is I-non-multiplicative for u with respect to U.
We define in a same way the notion of multiplicative (resp. non-multiplicative) variable. We denote
U
by MultU
I (u) (resp. NMultI (u)) the set of multiplicative (resp. non-multiplicative) variables for the
division I of a monomial u with respect to U. We have
U
MultU
I (u) = { x P {x1 , . . . , xn } u|I ux }
and thus a partition of the set of variables { x1 , . . . , xn } into sets of multiplicative and non-multiplicative
variables. An involutive division I is thus entirely defined by a partition
U
{x1 , . . . , xn } = MultU
I (u) \ NMultI (u),
for any finite subset U of M(x1 , . . . , xn ) and any u in U, satisfying conditions iv), v) and vi) of
Definition 3.1.1. The involutive division I is then defined by setting u |U
I w if w = uv and the
U
monomialv belongs to M(MultI (u)). Conditions i), ii) and iii) of Definition 3.1.1 are consequence of
this definition.
24
3.2. Involutive completion procedure
3.1.3. Example. Consider U = {x1 , x2 } in M(x1 , x2 ) and suppose that I is an involutive division such
U
that MultU
I (x1 ) = {x1 } and MultI (x2 ) = {x2 }. Then we have
x1 ∤I x1 x2 ,
and
x2 ∤I x1 x2 .
3.1.4. Autoreduction. A subset U of M(x1 , . . . , xn ) is said to be autoreduced with respect to an
involutive division I, or I-autoreduced, if it does not contain a monomial I-divisible by another monomial
of U.
In particular, by definition of the involutive division, for any monomials u, u 1 in U and monomial w
in M(x1 , . . . , xn ), we have u|I w and u 1 |I w implies u|I u 1 or u 1 |I u. As a consequence, if a set of
monomials U is I-autoreduced, then any monomial in M(x1 , . . . , xn ) admits at most one I-involutive
divisor in U.
3.1.5. The Janet division. We call Janet division the division on M(x1 , . . . , xn ) defined by the multiplicative variables in the sense of Janet defined in 2.1.10. Explicitely, for a subset U of M(x1 , . . . , xn )
and monomials u in U and w in M(x1 , . . . , xn ), we define u|U
J w if u is a Janet divisor of w as defined
U
in 2.1.12, that is w = uv, where v P M(MultJ (u)) and MultU
J (u) is the set of Janet’s multiplicative
variables defined in 2.1.10.
By Proposition 2.1.13, for a fixed subset of monomial U, any monomial of M(x1 , . . . , xn ) has a unique
Janet divisor in U with respect to U. As a consequence, the conditions iv) and v) of Definition3.1.1 trivially
hold for the Janet division. Now suppose that U 1 Ď U and u is a monomial in U 1 . If u|U
J w there is a
1
U
U1
U
decomposition w = uv with v P M(MultJ (u)). As MultJ (u) Ď MultJ (u), this implies that u|U
J w.
Hence, the conditions vi) of Definition 3.1.1 holds for the Janet division. We have thus proved
3.1.6. Proposition ([GB98a, Proposition 3.6]). The Janet division is involutive.
3.2. Involutive completion procedure
3.2.1. Involutive set. Let I be an involutive division on M(x1 , . . . , xn ) and let U be a set of monomials.
The involutive cone of a monomial u in U with respect to the involutive division I is defined by
coneI (u, U) = { uv v P M(x1 , . . . , xn ) and u|U
I uv }.
The involutive cone of U with respect to the involutive division I is the following subset of monomials:
ď
coneI (U) =
coneI (u, U).
uPU
Note that the inclusion coneI (U) Ď cone(U) holds for any set U. Note also that when the set U is
I-autoreduced, by involutivity this union is disjoint.
A subset U of M(x1 , . . . , xn ) is I-involutive if the following equality holds
cone(U) = coneI (U).
In other words, a set U is I-involutive if any multiple of an element u in U is also I-involutive multiple
of an element v of U. Note that the monomial v can be different from the monomial u, as we have seen
in Example 2.2.6.
25
3. Monomial involutive bases
3.2.2. Involutive completion. A completion of a subset U of monomials of M(x1 , . . . , xn ) with respect
r satisfying the following
to an involutive division I, or I-completion for short, is a set of monomials U
three conditions
r is involutive,
i) U
r
ii) U Ď U,
r = cone(U).
iii) cone(U)
3.2.3. Noetherianity. An involutive division I is said to be noetherian if all finite subset U of M(x1 , . . . , xn )
r
admits a finite I-completion U.
3.2.4. Proposition ([GB98a, Proposition 4.5]). The Janet division is noetherian.
3.2.5. Prolongation. Let U be a subset of M(x1 , . . . , xn ). We call prolongation of an element u of U
a multiplication of u by a variable x. Given an involutive division I, a prolongation ux is multiplicative
(resp. non-multiplicative) if x is a multiplicative (resp. non-multiplicative) variable.
3.2.6. Local involutivity. A subset U of M(x1 , . . . , xn ) is locally involutive with respect to an involutive
division I if any non-multiplicative prolongation of an element of U admit an involutive divisor in U.
That is
@u P U @xi P NMultU
I (u)
Dv P U
such that
v|I uxi .
3.2.7. Example. By definition, if U is I-involutive, then it is locally I-involutive. The converse is false
in general. Indeed, consider the involutive division I on M = M(x1 , x2 , x3 ) defined by
MultM
I (x1 ) = {x1 , x3 },
MultM
I (x2 ) = {x1 , x2 },
MultM
I (x3 ) = {x2 , x3 },
M
with MultM
I (1) = {x1 , x2 , x3 } and MultI (u) is empty for deg(u) ě 2. Then the set {x1 , x2 , x3 } is locally
I-involutive but not I-involutive.
3.2.8. Continuity. An involutive division I is continuous if for all finite subset U of M(x1 , . . . , xn ) and
any finite sequence (u1 , . . . , uk ) of elements in U such that, there exists xij in NMultU
I (uj ) such that
uk |I uk−1 xik−1 , . . . , u3 |I u2 xi2 , u2 |I u1 xi1 ,
then ui ‰ uj , for any i ‰ j.
For instance, the involutive division in Example 3.2.7 is not continuous. Indeed, there exists the
following cycle of divisions:
x2 |I x1 x2 ,
26
x1 |I x3 x1 ,
x3 |I x2 x3 ,
x2 |I x1 x2 .
3.2. Involutive completion procedure
3.2.9. From local to global involutivity. Any I-involutive subset U of M(x1 , . . . , xn ) is locally Iinvolutive. When the division I is continuous the converse is also true. Indeed, suppose that U is locally
I-involutive. Let us show that U is I-involutive when the division I is continuous.
Given a monomial u in U and a monomial w in M(x1 , . . . , xn ), let us show that the monomial uw
admits an I-involutive divisor in U. If u|I uw the claim is proved. Otherwise, there exists a nonmultiplicative variable xk1 in NMultU
I (u) such that xk1 |w. By local involutivity, the monomial uxk1
admits an I-involutive divisor v1 in U. If v1 |I uw the claim is proved. Otherwise, there exists a
uw
non-multiplicative variable xk2 in NMultU
I (v1 ) such that xk2 divides v1 . By local involutivity, the
monomial v1 xk2 admits an I-involutive divisor v2 in U.
In this way, we construct a sequence (u, v1 , v2 , . . .) of monomials in U such that
v1 |I uxk1 ,
v2 |I v1 xk2 ,
v3 |I v2 xk3 ,
...
By continuity hypothesis, all monomials v1 , v2 , . . . are distinct. Moreover, all these monomials are divisor
of uw, that admits a finite set of distinct divisors. As a consequence, previous sequence is finite. It
follows, that its last term vk is an I-involutive monomial of uw. We have thus proved the following result.
3.2.10. Theorem ([GB98a, Theorem 4.10]). Let I be a continuous involutive division. A subset of
monomials of M(x1 , . . . , xn ) is locally I-involutive if and only if it is I-involutive.
3.2.11. Proposition ([GB98a, Corollary 4.11]). The Janet division is continuous.
3.2.12. Involutive completion procedure. Procedure 2 generalizes Janet’s completion procedure given
in 2.2.9 to any involutive division. Let us fix a monomial order ď on M(x1 , . . . , xn ). Given a set
of monomials U, the procedure completes the set U by all possible non-involutives prolongations of
monomials in U.
Procedure 2: Involutive completion procedure.
Input: U a finite set of monomials of M(x1 , . . . , xn )
begin
r←U
U
r
r and x P NMultU (u) such that ux does not have I-involutive divisor in U
r do
while exist u P U
I
Choose such a u and x corresponding to the smallest monomial ux with respect to the
monomial order ď
r←U
r Y {ux}
U
end
end
r the minimal involutive completion of the set U.
Output: U
By introducing the notion of constructive involutive division, V. P. Gerdt and Y. A. Blinkov gave
in [GB98a] some conditions on the involutive division I in order to show the correction and the termination
of this procedure. A continuous involutive division I is constructive if for any subset of monomials U
of M(x1 , . . . , xn ) and for any non-multiplicative prolongation ux of a monomial u in U satisfying the
following two conditions
27
3. Monomial involutive bases
i) ux does not have an I-involutive divisor in U,
ii) any non-multiplicative prolongation vy ‰ ux of a monomial v in U that divides ux has an I-involutive
divisor in U,
the monomial ux cannot be I-involutively divided by a monomial w in coneI (U) with respect to U Y {w}.
If I is a constructive division, then the completion procedure completes the set U into an involutive
set. We refer the reader to [GB98a, Theorem 4.14] for a proof of correctness and termination of the
completion procedure under these hypothesis.
3.2.13. Example. An application of this procedure on the set of monomials U = { x3 x22 , x33 x21 } given by
M. Janet in [Jan29] is developed in 2.2.10.
3.3. Others involutive approaches
For analysis on differential systems several other notion of multiplicative variables were introduced by
J. M. Thomas 1937 and J.-F. Pommarret in 1978. Others examples of involutive divisions can be found
in [GB98b].
3.3.1. Thomas division. In [Tho37], Thomas introduced an involutive division that differs from those
of M. Janet also used in the analysis on differential systems. The multiplicative variables in the sense of
Thomas’s division for a monomial u with of a finite subset U of M(x1 , . . . , xn ) are defined as follows:
xi P MultU
T (u)
if
degi (u) = degi (U).
In particular, we have u|U
T w if w = uv and for all variable xi in v, we have degi (u) = degi (U). The
Thomas division is a noetherian and continuous involutive division. We refer the reader to [GB98a] for
detailed proofs of this results. Note also that the Janet division is a refinement of Thomas division in the
sense that for any finite set of monomials U and any monomial u in U, the following inclusions hold
U
MultU
T (u) Ď MultJ (u)
and
U
NMultU
J (u) Ď NMultT (u).
3.3.2. Pommaret division. In [Pom78], Pommaret introduced an involutive division that is defined
globally, that is the multiplicative variables for the Pommaret division does not depend of a given subset
of monomials. In this way, Pommaret’s division can be defined on an infinite set of monomials.
We fix an order on the variables x1 ą x2 ą . . . ą xn . Given a monomial u = xα1 1 . . . xαk k ,
with αk ą 0, the Pommaret multiplicative variables for u are defined by
M(x1 ,...,xn )
xj P MultP
(u),
if j ě k,
and
M(x1 ,...,xn )
xj P NMultP
(u),
M(x ,...,x )
if j ă k.
We set MultP 1 n (1) = {x1 , . . . , xn }. The Pommaret division is a continuous involutive division that
is not noetherian, [GB98a]. The Janet division is also a refinement of the Pommaret division, that is, for
an autoreduced finite set of monomials U, the following inclusions hold for any monomial u in U,
U
MultU
P (u) Ď MultJ (u)
28
and
U
NMultU
J (u) Ď NMultP (u).
4. Polynomial partial differential equations systems
4. Polynomial partial differential equations systems
In this section, we extend the results presented in Section 2 on monomial systems to linear (polynomial) systems. All PDE systems are considered in analytic categories, namely all unknown functions,
coefficients and boundary conditions are supposed to be analytic. In a first part, we recall the notion of
principal derivative with respect to an order on derivatives introduced by M. Janet. This notion is used to
give an algebraic characterization of complete integrability conditions of a PDE system. Then we present
a procedure that decides whether a given finite linear PDE system can be transformed into a completely
integrable linear PDE system. Finally, we recall the algebraic formulation of involutivity introduced by
M. Janet in [Jan29].
4.1. Parametric and principal derivatives
4.1.1. Motivations. In [Jan29, Chapter 2], M. Janet first considered the following PDE of one unknown
function on Cn :
n
ÿ
ÿ
ÿ
Bϕ
B2 ϕ
B2 ϕ
B2 ϕ
=
+
+
+ c(x)ϕ + f(x),
br (x)
a
(x)
a
(x)
i,j
i
2
Bxn
Bxi Bxj 1ďiăn
Bxi Bxn r=1
Bxr
1ďi, jăn
(5)
where the functions ai,j (x), ai (x), br (x), c(x) and f(x) are analytic functions in a neighborhood of a
point P = (x01 , . . . , x0n ) in Cn . Given two analytic functions ϕ1 and ϕ2 in a neighborhood UQ of a
point Q = (x01 , . . . , x0n−1 ) in Cn−1 , M. Janet studied the existence of solutions of equation (5) under the
following initial condition:
Bϕ
= ϕ2 ,
(6)
ϕ|xn =x0n = ϕ1 ,
Bxn xn =x0n
holds in a neighborhood of the point Q. In 4.4.2, we will formulate such condition for higher-order linear
PDE systems with several unknown functions, called boundary condition.
4.1.2. Principal and parametric derivatives. In order to analyse the existence and the uniqueness of
a solution of equation (5) under the initial condition (6), M. Janet introduced the notions of parametric
and principal derivative defined as follows. The partial derivatives Dα ϕ, with α = (α1 , . . . , αn ), of an
analytic function ϕ are determined by
i) ϕ1 and its derivatives for αn = 0,
ii) ϕ2 and its derivatives for αn = 1,
in the neighborhood UQ . These derivatives for αn = 0 and αn = 1 are called parametric, those
2
derivatives for αn ě 2, i.e. the derivative of BBxϕ
2 , are called principal. Note that the values of the
n
principal derivative at the point P are entirely given by ϕ1 and ϕ2 and by their derivatives thanks to
equation (5). Note that the notion of parametric derivative corresponds to a parametrization of initial
conditions of the system.
29
4. Polynomial partial differential equations systems
4.1.3. Janet’s orders on derivatives. Let α = (α1 , . . . , αn ) and β = (β1 , . . . , βn ) be in Nn . Let ϕ be
an analytic function. The derivative Dα ϕ is said to be posterior (resp. anterior) to Dβ ϕ if
|α| ą |β|
(resp. |α| ă |β|)
or
|α| = |β| and αn ą βn
(resp. αn ă βn ).
Obviously, any derivative of ϕ admits only finitely many anterior derivatives of ϕ. Using this notion
of posteriority, M. Janet showed the existence and unicity problem of equation (5) under the initial
condition (6).
In his monograph, M. Janet gave several generalizations of the previous posteriority notion. The first
one corresponds to the degree lexicographic order, [Jan29, §22], formulated as follows:
i) for |α| ‰ |β|, the derivative Dα ϕ is called posterior (resp. anterior) to Dβ ϕ, if |α| ą |β| (resp.
|α| ă |β|),
ii) for |α| = |β|, the derivative Dα ϕ is called posterior (resp. anterior) to Dβ ϕ if the first non-zero
difference
αn − βn , αn−1 − βn−1 , . . . , α1 − β1 ,
is positive (resp. negative).
4.1.4. Generalization. Let us consider the following generalization of equation (5):
Dϕ =
ÿ
ai Di ϕ + f,
(7)
iPI
where D and the Di are differential operators such that Di ϕ is anterior to Dϕ for all i in I. The
derivative Dϕ and all its derivatives are called principal derivatives of the equation (7). All the other
derivative of u are called parametric derivatives of the equation (7).
4.1.5. Weight order. Further generalization of these order relations were given by M. Janet by introducing the notion of cote, that corresponds to a parametrization of a weight order defined as follows. Let us
fix a positive integer s. We define a matrix of weight
C1,1 . . . Cn,1
..
C = ...
.
C1,s . . . Cn,s
that associates to each variable xi non negative integers Ci,1 , . . . , Ci,s , called the s-weights of xi .
This notion was called cote by M. Janet in [Jan29, §22] following the terminology introduced by
Riquier, [Riq10]. For each derivative Dα ϕ, with α = (α1 , . . . , αn ) of an analytic function ϕ, we
associate a s-weight Γ (C) = (Γ1 , . . . , Γs ) where the Γk are defined by
Γk =
n
ÿ
αi Ci,k .
i=1
Given two monomial partial differential operators Dα and Dβ as in 4.1.3, we say that Dα ϕ is posterior
(resp. anterior) to Dβ ϕ with respect to a weigh matrix C if
30
4.2. First order PDE systems
i) |α| ‰ |β| and |α| ą |β| (resp. |α| ă |β|),
ii) otherwise |α| = |β| and the first non-zero difference
Γ1 − Γ11 ,
Γ2 − Γ21 ,
...
, Γs − Γs1 ,
is positive (resp. negative).
In this way, we define an order on the set of monomial partial derivatives, called weight order. Note that,
we recover the Janet order defined in 4.1.3 by setting Ci,k = δi+k,n+1 .
4.2. First order PDE systems
We consider first resolution of first order PDE systems.
4.2.1. Complete integrability. In [Jan29, §36], M. Janet considered the following first order PDE system
Bϕ
= fλ (y1 , ¨ ¨ ¨ , yh , z1 , ¨ ¨ ¨ , zk , ϕ, q1 , ¨ ¨ ¨ , qk )
(1 ď λ ď h)
(8)
Byλ
where ϕ is an unknown function of independent variables y1 , . . . , yh , z1 , . . . , zk , with h + k = n
Bϕ
and qi = Bz
. Moreover, we suppose that the functions fλ are analytic in a neighborhood of a point P.
i
M. Janet wrote down explicitly the integrability condition of the PDE systems (Σ) defined in (8) namely
by the following equality
Bϕ
Bϕ
B
B
=
,
Byλ Byµ
Byµ Byλ
for any 1 ď λ, µ ď h. Following (8), we deduce that
k
B
Bϕ Bfµ ÿ Bfµ B2 ϕ
Bϕ
Bfµ
+
+
,
=
Byλ Byµ
Byλ Byλ Bϕ i=1 Bqi Byλ Bzi
ÿ
k
k
Bfµ
Bfµ ÿ Bfµ Bfλ
Bfλ
Bfλ Bfµ B2 ϕ
=
+ fλ
+
+ qi
.
+
Byλ
Bϕ i=1 Bqi Bzi
Bϕ
Bqi Bqj Bzi Bzj
i,j=1
(Σ)
Hence, the integrability condition is expressed as
Bϕ
Bϕ
B
B
−
Byλ Byµ
Byµ Byλ
k
k
Bfµ ÿ Bfµ Bfλ
Bfµ
Bfµ
Bfλ
Bfλ ÿ Bfλ Bfµ
Bfλ
+ fλ
+
+ qi
− fµ
−
+ qi
=
−
Byλ
Bϕ i=1 Bqi Bzi
Bϕ
Byµ
Bϕ i=1 Bqi Bzi
Bϕ
(9)
=0,
for any 1 ď λ ‰ µ ď h. When the PDE system (Σ) defined in (8) satisfies relation (9), the system (Σ) is
said to be completely integrable.
4.2.2. Theorem. Suppose that the PDE system (Σ) in (8) is completely integrable. Let P be a point in Cn
and ϕ(z1 , . . . , zk ) be an analytic function in the neighborhood of the point π(P), where π : Cn → Ck
denotes the canonical projection (y1 , . . . , yh , z1 , . . . zk ) Þ→ (z1 , . . . , zk ). Then, the system (Σ) admits
only one analytic solution satisfying u = π˚ ϕ in a neighborhood of the point P.
31
4. Polynomial partial differential equations systems
4.3. Higher-order finite linear PDE systems
In [Jan29, §39], M. Janet discussed the existence of solutions of a finite linear PDE system of one
unknown function ϕ in which each equation is of the following form:
ÿ
(Σ)
Di ϕ =
ai,j Di,j ϕ, i P I.
(10)
j
All the functions ai,j are supposed analytic in a neighborhood of a point P in Cn .
4.3.1. Principal and parametric derivatives. Consider Janet’s order ďJ on derivatives as defined
in 4.1.3. We suppose that each equation of the system (Σ) defined by (10) satisfies the following two
conditions:
i) Di,j ϕ is anterior to Di ϕ, for any i in I,
ii) all the Di ’s for i in I are distinct.
We extend the notion of principal derivative introduced in 4.1.4 for one PDE equation to a system
of the form (10) as follows. The derivative Di ϕ, for i in I, and all its derivatives are called principal
derivatives of the PDE system (Σ) given in (10) with respect to Janet’s order. Any other derivative of ϕ
is called parametric derivative.
4.3.2. Completeness with respect to Janet’s order. Let us fix an order xn ą xn−1 ą . . . ą x1 on
variables. By the isomorphism of Proposition 2.1.2, that identifies monomial partial differential operators
with monomials in M(x1 , . . . , xn ), we associate to the set of operators Di ’s, i in I, defined in 4.3.1, a
set lmďJ (Σ) of monomials. By definition, the set lmďJ (Σ) contains the monomials associated to leading
derivatives of the PDE system (Σ) with respect to Janet’s order.
The PDE system (Σ) is said to be complete with respect to Janet’s order ďJ if the set of monomials lmďJ (Σ) is complete in the sense of 2.2.1. In 6 we will give a completion procedure that transforms a
finite linear PDE system into an equivalent complete linear PDE system.
By definition the set of principal derivatives corresponds, by isomorphism of Proposition 2.1.2, to
the multiplicative cone of the monomial set lmďJ (Σ). Hence, when (Σ) is complete, the set of principal
derivatives corresponds to the involutive cone of lmďJ (Σ). By Proposition 2.2.2, there is a partition
M(x1 , . . . , xn ) = coneJ (lmďJ (Σ)) > coneJ (lmďJ (Σ)K ).
It follows that set of parametric derivatives of a complete system (Σ) corresponds to the involutive cone
of the set of monomials lmďJ (Σ)K .
4.3.3. Boundary conditions. Consider the set lmďJ (Σ)K of complementary monomials of lmďJ (Σ), as
defined in 2.1.9. To a monomial xβ in lmďJ (Σ)K , with β = (β1 , . . . , βn ) in Nn and
lmďJ (Σ)K
MultJ
(xβ ) = {xi1 , . . . , xikβ },
we associate an arbitrary analytic function
ϕβ (xi1 , . . . , xikβ ).
32
4.3. Higher-order finite linear PDE systems
Using these functions, M. Janet defined a boundary condition:
(Cβ )
Dβ ϕ
lmď (Σ)K
J
xj =0 @xj PNMultJ
(xβ )
= ϕβ (xi1 , . . . , xikβ ).
Then he formulated a boundary condition of the equation (10) with respect to Janet’s order as the following
set
{ Cβ | xβ P lmďJ (Σ)K }.
(11)
4.3.4. Theorem ([Jan29, §39]). If the PDE system (Σ) in (10) is complete with respect to Janet’s
order ďJ , then it admits at most one analytic solution satisfying the boundary condition (11).
4.3.5. PDE systems with several unknown functions. The construction of boundary conditions given
in 4.3.3 for one unknown function can be extended to linear PDE systems on Cn with several unknown functions using a weight order. Let us consider a linear PDE system of m unknown analytic
functions ϕ1 , . . . , ϕm of the following form
ÿ
β s
(Σ)
Dα ϕr =
α P Ir ,
(12)
ar,s
α,β D ϕ ,
(β,s)PNn ˆ{1,2,...,m}
for 1 ď r ď m, where Ir is a finite subset of Nn and the ar,s
α,β are analytic functions.
For such a system, we define a weight order as follows. Let us fix a positive integer s. To any variable
xi we associate s + 1 weights Ci,0 , Ci,1 , . . . , Ci,s by setting Ci,0 = 1 and the Ci,1 , . . . , Ci,s as defined in
(j)
(j) (j)
4.1.5. For each unknown function ϕj , we associate s + 1 weights T0 , T1 . . . , Ts . With these data, we
(j)
(j) (j)
define the s + 1 weights Γ0 , Γ1 , . . . , Γs of the partial derivative Dα ϕj with α = (α1 , . . . , αn ) in Nn
by setting
n
ÿ
(j)
(j)
Γk =
αi Ci,k + Tk
(0 ď k ď s).
i=1
We define the notions of anteriority and posteriority on derivatives with respect to this weight order,
denoted by ďwo , as it is done in 4.3.1 for systems of one unknown function. In particular, we define the
notions of principal and parametric derivatives in a similar way to systems of one unknown function.
Now suppose that the system (12) is written in the form
ÿ
β s
(Σ)
Dα ϕr =
α P Ir .
(13)
ar,s
α,β D ϕ ,
(β,s)PNn ˆ{1,2,...,m}
Dβ ϕs ăwo Dα ϕr
We can formulate the notion of completeness with respect to the weight order ďwo as in 4.3.2. Let
consider lmďwo (Σ, ϕr ) be the set of monomials associated to leading derivatives Dα of all PDE in (Σ)
such that α belongs to Ir . The PDE system (Σ) is complete with respect to ďwo , if for any 1 ď r ď m,
the set of monomials lmďwo (Σ, ϕr ) is complete in the sense of 2.2.1. Finally, we can formulate as in (11)
a boundary condition for the linear PDE system (13) with respect to such a weight order.
4.3.6. Theorem ([Jan29, §40]). If the PDE system (Σ) in (13) is complete with respect to a weight
order ďwo , then it admits at most one analytic solution satisfying the boundary condition (11).
M. Janet said that this result could be proved in a way similar to the proof of Theorem 4.3.4.
33
4. Polynomial partial differential equations systems
4.4. Completely integrable higher-order linear PDE systems
In this subsection we will introduce integrability conditions for higher-order linear PDE systems of
several unknown functions. The main result, Theorem 4.4.7, algebraically characterizes the complete
integrability property for complete PDE systems. It states that, under the completeness property, the
complete integrability condition is equivalent to have all integrability conditions trivial. In this subsection,
we will assume that the linear PDE systems are complete. Note that, we will see in 6 a procedure that
transforms a linear PDE system of the form (13) into a complete linear PDE system with respect to a
weight order.
4.4.1. Formal solutions. Let consider a linear PDE system (Σ) of the form (13) of unknown functions ϕ1 , . . . , ϕm and independent variables x1 , . . . , xn . We suppose that (Σ) is complete, hence the set
of monomials lmďwo (Σ, ϕr ) = {xα | α P Ir } is complete for all 1 ď r ď m. For the remaining part of
this subsection, we will denote lmďwo (Σ, ϕr ) by Ur . Let denote by (coneJ,ďwo (Σ)) the following PDE
system, for 1 ď r ď m,
ÿ
β s
Φ(u)(Dα ϕr ) =
Φ(u) ar,s
D
ϕ
,
α,β
(β,s)PNn ˆ{1,2,...,m}
Dβ ϕs ăwo Dα ϕr
for α P Ir and u P M(Mult(xα , Ur )).
We use the PDE system (coneJ,ďwo (Σ)) to compute the values of the principal derivative at a
point P0 = (x01 , . . . , x0n ) of Cn . We call formal solutions of the PDE system (Σ) at the point P0 the
elements ϕ1 , . . . , ϕm in C[[x1 − x01 , . . . , xn − x0n ]] which are solutions of (Σ). If the system (Σ) admits
an analytic solution then these formal solutions are convergent series and give analytic solutions of (Σ)
on a neighbourhood of the point P0 .
4.4.2. Boundary conditions. The question is to determine under which condition the system (Σ) admits
a solution for any given boundary condition. These boundary conditions are parametrized by the set UK
r
of complementary monomials of the set of monomials Ur as in 4.3.3. Explicitly, for 1 ď r ď m, to a
UK
β
n
r
monomial xβ in UK
r , with β in N and MultJ (x ) = {xi1 , . . . , xikr }, we associate an arbitrary analytic
function
ϕβ,r (xi1 , . . . , xikr ).
Formulating boundary condition as the following data:
(Cβ,r )
Dβ ϕr
UK
xj =x0j @xj PNMultJ r (xβr )
= ϕβ,r (xi1 , . . . , xikr ).
We set the boundary condition of the system (Σ) in (12) to be the following set
ď
{ Cβ,r | xβr P UK
r }.
(14)
1ďrďm
Note that M. Janet call degree of generality of the solution of the PDE system (Σ) the dimension of the
boundary conditions of the system, that is
UK
Max MultJ r (u) .
uPVUr
34
4.4. Completely integrable higher-order linear PDE systems
4.4.3. J-normal form. Suppose that the PDE system (Σ) is complete. Given a linear equation E amongst
the unknown functions ϕ1 , . . . , ϕm and variables x1 , . . . , xn . A J-normal form of E with respect to the
system (Σ) is an equation obtained from E by the reduction process that replaces principal derivatives by
parametric derivatives with a similar procedure to RightReduce given in 5.
α
r
4.4.4. Integrability conditions. Given 1 ď r ď m and α P Ir , let xi be in NMultU
J (x ) a nonmultiplicative variable. Let us differentiate the equation
ÿ
β s
Dα ϕr =
ar,s
α,β D ϕ
(β,s)PNn ˆ{1,2,...,m}
Dβ ϕs ăwo Dα ϕr
by the partial derivative Φ(xi ) =
α
r
Φ(xi )(D ϕ ) =
B
Bxi .
We obtain the following PDE
ÿ
(β,s)PNn ˆ{1,2,...,m}
Dβ ϕs ăwo Dα ϕr
Bar,s
α,β
Bxi
β
s
D ϕ +
β s
ar,s
α,β Φ(xi )(D ϕ )
.
(15)
Using system (coneJ,ďwo (Σ)), we can rewrite the PDE (15) into an PDE formulated in terms of parametric
derivatives and independent variables. The set of monomials Ur being complete, there exists α 1 in Nn
1
α
α 1 as a
α
α1
α1
r
with xα in Ur and u in M(MultU
J (x )) such that xi x = ux . Then Φ(xi )D = Φ(u)D
consequence, we obtain the following equation
r,s
ÿ
ÿ
Baα,β β s
r,s
β s
β1 s
D ϕ + aα,β Φ(xi )(D ϕ ) =
Φ(u)(ar,s
α 1 ,β 1 D ϕ ).
Bxi
n
1
n
(β,s)PN ˆ{1,2,...,m}
Dβ ϕs ăwo Dα ϕr
(β ,s)PN ˆ{1,2,...,m}
1
1
Dβ ϕs ăwo Dα ϕr
(16)
Using equations of system (coneJ,ďwo (Σ)), we replace all principal derivatives in the equation (16) by
parametric derivatives and independent variables. The order ďwo being well-founded this process is
terminating. Moreover, when the PDE system (Σ) is complete this reduction process is confluent in the
sense that any transformations of an equation (16) ends on a unique J-normal forms. This set of J-normal
forms is denoted by IntCondJ,ďwo (Σ).
4.4.5. Remarks. The system (Σ) being complete any equation (16) is reduced to a unique normal form.
Such a normal form allows us to judge whether a given integrability condition is trivial or not.
Recall that the parametric derivatives correspond to the boundary conditions. Hence, a non-trivial
relation in IntCondJ,ďcwo (Σ) provides a non-trivial relation among the boundary conditions. In this way,
we can decide whether the system (Σ) is completely integrable or not.
4.4.6. Completely integrable systems. A complete linear PDE system (Σ) of the form (13) is said to
be completely integrable if it admits an analytic solution for any given boundary condition (14). For the
geometrical interpretation of these condition, we refer the reader to 1.1.4.
4.4.7. Theorem ([Jan29, §42]). Let (Σ) be a complete finite linear PDE system of the form (13). Then
the system (Σ) is completely integrable if and only if any relation in IntCondJ,ďwo (Σ) is a trivial identity.
A proof of this result is given in [Jan29, §43]. Note that the later condition is equivalent to say that
any relation (16) is an algebraic consequence of a PDE equation of the system (coneJ,ďwo (Σ)).
35
4. Polynomial partial differential equations systems
4.5. Canonical forms of linear PDE systems
In this subsection, we recall from [Jan29] the notion of canonical linear PDE system. A canonical system
is a normal form with respect to a weight order on derivatives, and satisfying some analytic conditions,
allowing to extend the Cauchy-Kowalevsky’s theorem given in 1.1.3. Note that this terminology refers
to a notion of normal form, but it does not correspond to the well known notion for a rewriting system
meaning both terminating and confluence. In this notes, we present canonical systems with respect to
weight order as it done in Janet’s monograph [Jan29], but we notice that this notion can be defined with
any total order on derivative.
4.5.1. Autoreduced PDE systems. Let (Σ) be a finite linear PDE system. Suppose that a weight
order ďwo is fixed on the set of unknown functions ϕ1 , . . . , ϕm of (Σ) and their derivatives, as defined
in 4.3.5. We suppose also that each equation of the system (Σ) can be expressed in the following form
ÿ
(α,r)
(Σ(α,r) )
Dα ϕr =
a(β,s) Dβ ϕs ,
(β,s)PNn ˆ{1,2,...,m}
Dβ ϕs ăwo Dα ϕr
so that
(Σ) =
ď
Σ(α,r) ,
(17)
(α,r)PI
the union being indexed by a multiset I. The support of the equation (Σ(α,r) ) is defined by
(α,r)
Supp(Σ(α,r) ) = { (β, s) | a(β,s) ‰ 0 }.
For 1 ď r ď m, consider the set of monomials lmďwo (Σ, ϕr ) corresponding to leading derivatives,
that is monomials xα such (α, r) belongs to I. The system (Σ) is said to be
i) J-left-reduced with respect to ďwo if for any (α, r) in I there is no (α 1 , r) in I and non-trivial
1
lm
(Σ,ϕr ) α 1
monomial xγ in M(MultJ ďwo
(x )) such that xα = xγ xα ,
ii) J-rigth-reduced with respect to ďwo if, for any (α, r) in I and any (β, s) in Supp(Σ(α,r) ), there is
1
lm
(Σ,ϕr ) α 1
no (α 1 , s) in I and non-trivial monomial xγ in M(MultJ ďwo
(x )) such that xβ = xγ xα ,
iii) J-autoreduced with respect to ďwo if it is both J-left-reduced and J-right-reduced with respect to
ďwo .
4.5.2. Canonical PDE systems. A PDE system (Σ) is said to be J-canonical with respect a weight
order ďwo if it satisfies the following five conditions
i) it consists of finitely many equations and each equation can be expressed in the following form
ÿ
(α,r)
Dα ϕr =
a(β,s) Dβ ϕs ,
(β,s)PNn ˆ{1,2,...,m}
Dβ ϕs ăwo Dα ϕr
ii) the system (Σ) is J-autoreduced with respect to ďwo ,
36
4.5. Canonical forms of linear PDE systems
iii) the system (Σ) is complete,
iv) the system (Σ) is completely integrable,
(α,r)
v) the coefficients a(β,s) of the equations in i) and the boundary conditions of (Σ) are analytic.
Under these assumptions, the system (Σ) admits a unique analytic solution satisfying appropriate boundary
conditions parametrized by complementary monomials as in 4.3.3.
4.5.3. Example. In [Jan29, §44], M. Janet studied the following linear PDE system of one unknown
function ϕ
p54 = p11 ,
p53 = p41 ,
p = p ,
52
31
(Σ)
p
=
p
44
52 ,
p43 = p21 ,
p = p ,
33
42
B2 ϕ
. In Example 2.2.6, we have shown that the left hand sides of the equations of
Bxi Bxj
this system form a complete set of monomials. Let us define the following weights for the variables:
where pi,j denotes
x1
1
0
x2
0
0
x3
1
0
x4
1
1
x5
2
1
We deduce the following weights for the second derivatives:
p22
p21
p32
p42
0
0
1
0
1
1
p11
p31
p33
2
0
p52
p41
p43
2
1
p44
p51
p53
p54
p55
2
2
3
1
3
2
4
2
As seen in Example 2.3.4, given any four analytic functions
ϕ0 (x1 , x2 ),
ϕ3 (x1 , x2 ),
ϕ4 (x1 , x2 ),
ϕ5 (x1 , x5 ),
there exists a unique solution of the PDE system (Σ). Note that the boundary condition is given by
ϕ|x3 =x0 ,x4 =x0 ,x5 =x0 = ϕ0,0,0,0,0 (x1 , x2 ),
3
4
5
Bϕ
Bx3
x3 =x03 ,x4 =x04 ,x5 =x05
Bϕ
Bx4
x3 =x03 ,x4 =x04 ,x5 =x05
Bϕ
Bx5
x2 =x02 ,x3 =x03 ,x4 =x04
= ϕ0,0,1,0,0 (x1 , x2 ),
= ϕ0,0,0,1,0 (x1 , x2 ),
= ϕ0,0,0,0,1 (x1 , x5 ).
We set
37
4. Polynomial partial differential equations systems
A = p54 − p11
B = p53 − p41
C = p52 − p31
D = p44 − p52
E = p43 − p21
F = p33 − p42
x5
x5
x5
x4
x3
x3
x4
x3
x3
x3
x2
x2
x2
x2
x2
x2
x1
x1
x1
x1
x1
x1
where the variable on the right correspond to the multiplicative variables of the first term. In order to
decide if the system (Σ) is completely integrable it suffices to check if the following terms
B4 , C4 , C3 , D5 , E5 , E4 , F5 , F4
are linear combinations of derivative of the terms A, B, C, D, E, F with respect to their multiplicative
variables. Here Yi denotes the derivative BxB i Y of a term Y. Finally, we observe that
B4 = A3 − D1 − C1 ,
C4 = A2 − E1 ,
C3 = B2 − F1 ,
D5 = A4 − B1 − C5 ,
E5 = A3 − C1 ,
E4 = D3 + B2 ,
F5 = B3 − A2 + E1 ,
F4 = E3 − D2 − C2 .
As a consequence the system (Σ) is completely integrable, hence it is J-canonical.
4.6. Reduction of a PDE system to a canonical form
In his monograph [Jan29], M. Janet did not mention about the correctness of the procedures that he
introduced in order to reduce a finite linear PDE system to a canonical form. In this section, we explain
how to transform a finite linear PDE system with several unknown functions by derivation, elimination
and autoreduction, into an equivalent linear PDE system that is either in canonical form or in incompatible
system. For linear PDE systems with constant coefficients, the correctness of the procedure can be verified
easily.
4.6.1. Equivalence of PDE system. Janet’s procedure transforms by reduction and completion a finite
linear PDE system into a new PDE system. The PDE system obtained in this way is equivalent to the
original system. In his work, M. Janet dit not explain this notion of equivalence that can be described as
follows. Consider two finite linear PDE systems with m unknown functions and n independent variables
l
(Σ )
m
ÿ
pli,j ϕj = 0,
i P Il ,
j=1
for l = 1, 2, where pli,j are linear differential operators. We say that the PDE systems (Σ1 ) and (Σ2 ) are
equivalent if the set of solutions of the two systems are the same. This notion can be also formulated by
saying that the D-modules generated by the families of differentials operators (p1i,1 , . . . , p1i,m ) for i P I1
and (p2i,1 , . . . , p2i,m ) for i P I2 are equals.
38
4.6. Reduction of a PDE system to a canonical form
4.6.2. A canonical weight order. Consider a finite linear PDE system (Σ) of m unknown functions ϕ1 , . . . , ϕm of independent variables x1 , . . . , xn . To these variables and functions we associate the
following weights
x1
1
0
0
0
..
.
x2
1
0
0
0
..
.
...
...
...
...
...
xn−1
1
0
0
1
..
.
xn
1
0
1
0
..
.
ϕ1
0
1
0
0
..
.
ϕ2
0
2
0
0
..
.
...
...
...
...
...
ϕm
0
m
0
0
..
.
0
1
1
0
...
...
0
0
0
0
0
0
0
0
...
...
0
0
The weight order on monomial partial derivatives defined in 4.1.5 induced by this weight system is total.
This order is called canonical weight order following M. Janet and denoted by ďcwo .
4.6.3. Combination of equations. Consider the PDE system (Σ) with the canonical weight order ďcwo
defined in 4.6.2. We suppose that the system (Σ) is given in the same form as (17) and that each equation
of the system is written in the following form
ÿ
(α,r)
(β,s)
Dα ϕr =
(Ei )
i P I(α,r) .
a(α,r),i Dβ ϕs ,
(β,s)PNn ˆ{1,2,...,m}
Dβ ϕs ăcwo Dα ϕr
(α,r)
will be denoted by ldegďcwo (Eα,r
For such an equation, the leading pair (α, r) of the equation Ei
i ).
n
We will denote by Ldegďcwo (Σ) the subset of N ˆ {1, . . . , m} consisting of leading pairs of equations
of the system (Σ):
Ldegďcwo (Σ) = { ldegďcwo (E) | E is an equation of Σ }.
The canonical weight order ďcwo induces a total order on Nn ˆ {1, . . . , m} denoted by ălp . We will
(α,r)
denote by K(α, r, i) the set of pairs (β, s) of running indices in the sum of the equation Ei . Given i
and j in I(α,r) , we set
We define
(β,s)
(β,s)
(αi,j , ri,j ) = Max (β, s) P K(α, r, i) Y K(α, r, j) | a(α,r),i ‰ a(α,r),j .
(αi,j ,ri,j )
b(α,r)
(α ,r )
i,j i,j
a(α,r),i
(αi,j ,ri,j )
= −a(α,r),i
a(αi,j ,ri,j ) − a(αi,j ,ri,j )
(α,r),i
(α,r)
and we denote by Ei,j
(α,r),i
if (αi,j , ri,j ) P K(α, r, i) \ K(α, r, j),
if (αi,j , ri,j ) P K(α, r, j) \ K(α, r, i),
(18)
if (αi,j , ri,j ) P K(α, r, i) X K(α, r, j),
the equation
Dαi,j ϕri,j =
ÿ
(β,s)PK(α,r,j)
(β,s)ălp (αi,j ,ri,j )
(β,s)
c(αi,j ,ri,j ),j Dβ ϕs −
ÿ
(β,s)
c(αi,j ,ri,j ),i Dβ ϕs ,
(19)
(β,s)PK(α,r,i)
(β,s)ălp (αi,j ,ri,j )
39
4. Polynomial partial differential equations systems
where, for any k = i, j,
(β,s)
(β,s)
(α
,ri,j )
i,j
c(αi,j ,ri,j ),k = a(α,r),k {b(α,r)
.
(α,r)
The equation (19) corresponds to a combination of the two equations Ei
(α,r) (α,r)
Combineďcwo (Ei , Ej ).
denoted by
by combination.
(α,r)
and Ej
and it will be
Procedure 3 adds to a set of PDE equations (Σ) an equation E
Procedure 3: Addďcwo (Σ, E)
Input:
- A canonical weight order ďcwo for ϕ1 , . . . , ϕm and x1 , . . . , xn .
- (Σ) a finite linear PDE system with unknown functions ϕ1 , . . . , ϕm of independent variables
x1 , . . . , xn given in the same form as (17) such that the leading derivatives are different.
- E be a linear PDE in the same form as (17).
begin
Γ ←Σ
(β, s) ← ldegďcwo (E)
if (β, s) R Ldegďcwo (Γ ) then
Γ ← Γ Y {E}
end
else
let E(β,s) be the equation of the system (Σ) whose leading pair is (β, s).
C ← Combineďcwo (E(β,s) , E)
Addďcwo (Γ, C)
end
end
Output: Γ a PDE system equivalent to the system obtained from (Σ) by adding equation E.
Note that at each step of the procedure RightReduceJ,ďcwo the running system Γ remains J-left
reduced. As consequence by combining this procedure with the procedure LeftReduceJ,ďcwo we obtain
the following autoreduce procedure that transform a PDE system into a autoreduced PDE system.
4.6.4. Procedure AutoreduceJ,ďcwo (Σ). Let us fix a canonical weight order ďcwo for ϕ1 , . . . , ϕm
and x1 , . . . , xn . Let (Σ) be a finite linear PDE system given in the same form as (17) with unknown
functions ϕ1 , . . . , ϕm of independent variables x1 , . . . , xn . We suppose that the leading derivatives of (Σ)
are all different. The procedure AutoreduceJ,ďcwo transforms the PDE system (Σ) into an J-autoreduced
PDE system that is equivalent to (Σ) by applying successively the procedures LeftReduceJ,ďcwo and
RightReduceJ,ďcwo . An algebraic version of this procedure is given in 9. Let us remark that the
autoreduction procedure given in Janet’s monographs corresponds to the LeftReduceJ,ďcwo , it does not
deal with right reduction of equations.
Note that, the procedure AutoreduceJ,ďcwo fails if and only if the procedure Combineďcwo fails.
(α,r)
(α,r)
and some
and Ej
This occurs when the procedure Combineďcwo is applied on equations Ei
(α ,r
)
i,j i,j
coefficients b(α,r)
, as defined in (18), vanish on some point of Cn . In particular, the procedure AutoreduceJ,ďcwo does not fail when all the coefficients are constant. This constraint on the
40
4.6. Reduction of a PDE system to a canonical form
Procedure 4: LeftReduceJ,ďcwo (Σ)
Input:
- A canonical weight order ďcwo for ϕ1 , . . . , ϕm and x1 , . . . , xn .
- (Σ) a finite linear PDE system with unknown functions ϕ1 , . . . , ϕm of independent variables
x1 , . . . , xn given in the same form as (17) such that the leading derivatives are different.
begin
Γ ←Σ
I ← Ldegďcwo (Γ )
Ur ← {xα | (α, r) P I}
1
r
(xα )) such that
while exist (α, r), (α 1 , r) in I and non-trivial monomial xγ in M(MultU
J
1
xα = xγ xα do
Γ ← Γ \ {E(α,r) }
1
1
Let Dγ E(α ,r) be the equation obtained from the equation E(α ,r) by applying the operator
Dγ to the two sides.
1
C ← Combineďcwo (E(α,r) , Dγ E(α ,r) )
Addďcwo (Γ, C)
end
end
Output: Γ a J-left-reduced PDE system with respect to ďcwo that is equivalent to (Σ).
coefficients of the system concerns only the left reduction and were not discussed in Janet’s monograph.
As a consequence, we have the following result.
4.6.5. Theorem. If (Σ) is a finite linear PDE system with constant coefficients, the procedure
AutoreduceJ,ďcwo terminates and produces a finite autoreduced PDE system that is equivalent to (Σ).
4.6.6. Completion procedure of a PDE system. Consider a finite linear PDE system (Σ) with the
canonical weight order ďcwo given in 4.6.2. If the system (Σ) is J-autoreduced, then the following
procedure CompleteJ,ďcwo (Σ) transforms the system (Σ) into a finite complete J-autoreduced linear
PDE system. This procedure of completion appears in Janet’s monograph [Jan29] but not given in an
explicit way.
4.6.7. Completion and integrability conditions. In the completion procedure 6, the set Pr contains all
the obstructions of the system to be complete. The procedure CompleteJ,ďcwo add to the system the
necessary equations in order to eliminate all these obstructions. The equations added to the system have
the following form
(δ,r)
(δ,r)
Dβ ϕr = Rhs(E(β,r) ) − a(β,r) Dδ ϕr + a(β,r) Dγ (Rhs(E(α,r) ))
with δ ‰ β and lead to the definition of new integrability condition of the form (16) by using the
construction given in 4.4.4.
41
4. Polynomial partial differential equations systems
Procedure 5: RightReduceJ,ďcwo (Σ)
Input:
- A canonical weight order ďcwo for ϕ1 , . . . , ϕm and x1 , . . . , xn .
- (Σ) a finite linear PDE system with unknown functions ϕ1 , . . . , ϕm of independent variables
x1 , . . . , xn given in the same form as (17) and that is J-left reduced with respect to ďcwo .
begin
Γ ←Σ
Γ1 ← Γ
I ← Ldegďcwo (Γ )
{{ The canonical weight order ďcwo induces a total
{{ order on the set I of leading pairs denoted by ďlp
(δ, t) ← max(I) with respect to ďlp
while Γ 1 ‰ H do
Γ 1 ← Γ 1 \ {E(δ,t) }
I ← I \ {(δ, t)}
S ← Supp(E(δ,t) )
Ur ← {xα | (α, r) P I}
α
r
while exist (β, r) in S, (α, r) in I and non-trivial monomial xγ in M(MultU
J (x )) such
that xβ = xγ xα do
Γ ← Γ \ {E(δ,t) }
(β,r)
(β,r)
C ← E(δ,t) − a(δ,t) Dβ ϕr + a(δ,t) Dγ (Rhs(E(α,r) ))
Addďcwo (Γ, C)
end
end
end
Output: Γ a J-right-reduced PDE system with respect to ďcwo that is equivalent to (Σ).
42
4.6. Reduction of a PDE system to a canonical form
Procedure 6: CompleteJ,ďcwo (Σ)
Input:
- A canonical weight order ďcwo for ϕ1 , . . . , ϕm and x1 , . . . , xn .
- (Σ) a finite J-autoreduced linear PDE system with unknown functions ϕ1 , . . . , ϕm of independent
variables x1 , . . . , xn given in the same form as (17) and whose leading derivatives are different.
begin
Γ ←Σ
Ξ←H
for r = 1, . . . , m do
while Ξ = H do
I ← Ldegďcwo (Γ )
Ur ←
{xα | (α, r) P I}
Ur δ
δ
Pr ← BE
Bx | E P Γ, x P NMultJ (x ) with (δ, r) = ldeg(E) and xx R coneJ (Ur )
C←0
while Pr ‰ H and C = 0 do
choose E(β,r) in Pr , whose leading pair (β, r) is minimal with respect to ďcwo .
Pr ← Pr \ {E(β,r) }
C ← E(β,r)
SC ← Supp(C)
α
δ
γ α
r
while exist (δ, r) in SC , (α, r) in I and xγ in M(MultU
J (x )) such that x = x x
do
(δ,r)
(δ,r)
C ← C − a(β,r) Dδ ϕr + a(β,r) Dγ (Rhs(E(α,r) ))
SC ← Supp(C)
end
end
if C ‰ 0 then
Γ ← AutoreduceJ,ďcwo (Γ Y {C})
end
else
Ξ←Γ
end
end
end
end
Output: (Ξ) a linear J-autoreduced PDE system equivalent to (Σ) and that is complete with
respect to ďcwo .
43
4. Polynomial partial differential equations systems
4.6.8. Janet’s procedure. Given a finite linear PDE system (Σ) with the canonical weight order ďcwo
defined in 4.6.2, Janet’s procedure JanetJ,ďcwo either transforms the system (Σ) into a PDE system (Γ )
that is J-canonical with respect to ďcwo or computes an obstruction to transform the system (Σ) to such a
form. In the first case, the solutions of the J-canonical system (Γ ) are solutions of the initial system (Σ).
In the second case, the obstruction corresponds to a non trivial relation on the boundary conditions. We
refer the reader to [Sch92] or [Rob14] for a deeper discussion on this procedure and its implementations.
Applying successively the procedures AutoreduceJ and CompleteJ , the first step of the procedure
consists in reducing the PDE system (Σ) into a PDE system (Γ ) that is J-autoreduced and complete with
respect to ďcwo .
Then it computes the set IntCondJ,ďcwo (Γ ) of integrability conditions of the system (Γ ). Recall
from 4.4.4 that this set is a finite set of relations that does not contain principal derivative. Hence,
these integrability conditions are J-normal forms with respect to Γ . The system (Γ ) being complete,
these normal forms are unique and by Theorem 4.4.7, if all of these normal forms are trivial, then the
system (Γ ) is completely integrable. Otherwise, the procedure takes a non-trivial condition R in the
set IntCondJ,ďcwo (Γ ) and distinguishes two cases. If the relation R is among functions ϕ1 , . . . , ϕm and
variables x1 , . . . , xn , then this relation imposes a relation on the boundary conditions of the system (Γ ).
In the other case, the set IntCondJ,ďcwo (Γ ) contains at least one PDE having a derivative of one of the
functions ϕ1 , . . . , ϕm and the procedure JanetJ,ďcwo is applied again to the PDE system (Σ) completed
by all the PDE equations in IntCondJ,ďcwo (Γ ).
4.6.9. Remarks. If the procedure stops at the first loop, that is when C consists only trivial identities,
then the system (Σ) is reducible to the J-canonical form (Γ ) equivalent to (Σ).
When the set C contains an integrability condition having at least one derivative of the unknown
functions, the procedure is applied again to the system (Σ) Y {C}. Notice that, it could be also possible to
recall the procedure on (Γ ) Y {C}, but as done in Janet’s monograph, we choose to restart the procedure
on (Σ) Y {C} in order to have a PDE system where each equation has a clear meaning, either it comes
from the initial problem or the integrability condition.
Finally, note that the procedure JanetJ,ďcwo fails on a PDE system (Σ) if and only if the procedure
AutoreduceJ,ďcwo fails on (Σ) and the potential non trivial relations among the unknown functions and
variables added during the process, as explained in 4.6.4.
4.6.10. Example. In [Jan29, §47], M. Janet studied the following PDE system:
p33 = x2 p11 ,
(Σ)
p22 = 0,
Bk ϕ
of an unknown function ϕ of independent variBxi1 . . . Bxik
ables x1 , x2 , x3 . The set of monomials of the left hand side of the system (Σ) is U = {x23 , x22 }. The
set U is not complete. Indeed, for instance the monomial x3 x22 is not in the involutive cone coneJ (U).
r := U Y {x3 x2 }. The PDE
If we complete the set U by the monomial x3 x22 we obtain a complete set U
2
system (Σ) is then equivalent to the following PDE system
p33 = x2 p11 ,
(Γ )
p322 = 0,
p22 = 0.
where pi1 ...ik denotes the derivative
44
4.6. Reduction of a PDE system to a canonical form
Procedure 7: JanetJ,ďcwo (Σ)
Input:
- A canonical weight order ďcwo for ϕ1 , . . . , ϕm and x1 , . . . , xn .
- (Σ) a finite linear PDE system with unknown functions ϕ1 , . . . , ϕm of independent variables
x1 , . . . , xn given in the same form as (17) and whose leading derivatives are different.
begin
Γ ← AutoreduceJ,ďcwo (Σ)
Γ ← CompleteJ,ďcwo (Γ )
C ← IntCondJ,ďcwo (Γ )
if C consists only of trivial identities then
return The PDE system (Σ) is transformable to a J-canonical system (Γ ).
end
if C contains a non trivial relation R among functions ϕ1 , . . . , ϕm and variables x1 , . . . , xn
then
return The PDE system (Σ) is not reducible to a J-canonical system and the relation R
imposes a non trivial relation on the boundary conditions of the system (Γ ).
end
else
// C contains a non trivial relation among functions ϕ1 , . . . , ϕm , variables x1 , . . . , xn ,
// and at least one derivative of one of the functions ϕ1 , . . . , ϕm .
Σ ← Σ Y {C}
JanetJ,ďcwo (Σ).
end
end
Output: Complete integrability of the system (Σ) and its obstructions to be reduced to a
J-canonical form with respect to ďcwo .
45
4. Polynomial partial differential equations systems
r is given by
Note that p322 = Bx3 p22 = 0. The table of multiplicative variables with respect to the set U
x23
x3 x22
x22
x3
x2
x2
x2
x1
x1
x1
We deduce that there exists only one non trivial compatibility condition, formulated as follows
p3322 =Bx3 p322 = Bx22 p33 ,
(x3 .x3 x22 = (x2 )2 .x23 )
=Bx22 (x2 p11 ) = 2p211 + x2 p2211 = 2p211 = 0,
(p2211 = Bx21 p22 = 0).
Hence, p211 = 0 is a non trivial relation of the system (Γ ). As a consequence, the PDE system (Σ) is not
completely integrable. Then, we consider the new PDE system given by
(Σ 1 )
p33 = x2 p11 ,
p22 = 0,
p211 = 0.
The associated set of monomials U 1 = {x23 , x22 , x2 x21 } is not complete. It can be completed into the
Ă1 := U 1 Y {x3 x2 , x3 x2 x2 }. The PDE system (Σ 1 ) is then equivalent to the following PDE
complete set U
2
1
system
p33 =
x2 p11 ,
p322 = 0,
1
(Γ )
p3211 = 0,
p22 =
0,
p221 = 0.
Note that p322 = Bx3 p22 and p3211 = Bx3 p211 . The multiplicative variables with respect to the set of
monomials U 1 is given by the following table
x23
x3 x22
x3 x2 x21
x22
x2 x21
x3
x2
x2
x2
x1
x1
x1
x1
x1
We deduce that the only non-trivial compatibility relation is
p33211 =Bx3 (p3211 ) = 0
=Bx21 Bx2 (p33 ) = Bx21 Bx2 (x2 p11 )
=Bx21 (p11 + x2 p211 ) = p1111
46
(p211 = 0).
4.7. Algebra, geometry and PDE
We deduce that p1111 = 0 is a non-trivial relation of the system (Γ 1 ). Hence, the system (Σ 1 ) is not
completely integrable. Then, we consider the new PDE system given by
(Σ 2 )
p33 =
p =
22
p
211 =
p
1111 =
x2 p11 ,
0,
0,
0.
The associated set of monomials U 2 = {x23 , x22 , x2 x21 , x41 } is not complete. It can be completed into the
Ă2 := U 2 Y {x3 x2 , x3 x2 x2 , x3 x4 }. The PDE system (Σ 2 ) is equivalent to the following
set of monomials U
2
1
1
system
p33 =
x2 p11 ,
p322 =
0,
p
31111 = 0,
(Γ 2 )
p22 =
0,
p211 =
0,
p
0.
1111 =
Note that p322 = Bx2 p22 and p31111 = Bx3 p1111 . All the compatibility conditions are trivial identities,
by Theorem 4.4.7 we deduce that the PDE (Σ 2 ), obtained from the initial PDE system (Σ) by adding
compatibility conditons, is completely integrable.
4.6.11. Remark. Let us mention, that using a similar procedure presented in this section, M. Janet
in [Jan29, §48] gave a constructive proof of a result obtained previously by A. Tresse [Tre94], that a
infinite linear PDE system can be reduced to a finite linear PDE system.
4.7. Algebra, geometry and PDE
The notion of ideal first appeared in the work of R. Dedekind. This notion appeared also in a seminal
paper [Hil90] of D. Hilbert, were he developed the theory of ideals in polynomial rings. In particular,
he proved noetherianity results as the noetherianity of the ring of polynomials over a field, now called
Hilbert’s basis theorem. In its works on PDE systems, [Jan22a, Jan22b, Jan24], M. Janet used the notion
of ideal generated by homogeneous polynomials under the terminology of module of forms, that he
defined as follows. He called form a homogeneous polynomial with several variables and he defined a
module of forms as an algebraic system satisfying the two following conditions:
i) if a form f belongs to the system, then the form hf belongs to the system for every form h,
ii) if f and g are two forms in the system of the same order, then the form f + g belongs to the system.
Finally, in [Jan29, §51], M. Janet recall Hilbert’s basis theorem.
47
4. Polynomial partial differential equations systems
4.7.1. Characteristic functions of homogeneous ideals. In [Jan29, §51], M. Janet recalled the Hilbert
description of the problem of finding the number of independent conditions so that a homogenous
polynomial of order p belongs to a given homogeneous ideal. This independent conditions correspond to
the independent linear forms that vanish all homogenous polynomials of degree p in the ideal. M. Janet
recalled from [Hil90] that this number of independent conditions is expressed as a polynomial in p for
sufficiently big p.
Let I be a homogenous ideal of K[x1 , . . . , xn ] generated by polynomials f1 , . . . , fk . Given a monomial
order on M(x1 , . . . , xn ), we can suppose that all the leading coefficients
are equal to 1. For any p ě 0,
À
consider the homogenous component of degree p so that I = p Ip with
Ip := I X K[x1 , . . . xn ]p .
Let us recall that
dim Ip ď dim K[x1 , . . . , xn ]p = Γnp .
The number of independent conditions so that a homogenous polynomial of order p belongs to the ideal I
is given by the difference
χ(p) := Γnp − dim Ip .
This is the number of monomials of degree p that cannot be divided by the monomials lm(f1 ), . . . , lm(fk ).
The function χ(p) corresponds to a coefficient of the Hilbert series of the ideal I and is called characteristic
function of the ideal I by M. Janet, or postulation in [Jan29, §52]. In the Section 4.8, we will show that
the function χ(p) is polynomial for sufficiently big p. Finally, note that the set of monomials that cannot
be divided by the monomials lm(f1 ), . . . , lm(fk ) forms a finite number of classes of complementary
monomials.
4.7.2. Geometrical remark. M. Janet gave the following geometrical observation about the characteristic function. Suppose that p is sufficiently big so that the function χ(p) is polynomial. Let λ − 1 be the
degree of the leading term of the polynomial χ(p). Consider the projective variety V(I) defined by
V(I) = {a P Pn−1 | f(a) = 0 for all f in I }.
The integer µ = lc(χ(p))(λ − 1)! corresponds to the degree of the variety V(I), [Hil90]. If χ(p) = 0
then the variety V(I) is empty, in the others cases V(I) is a sub-variety of Pn−1 of dimension λ − 1.
4.7.3. Example, [Jan29, §53]. Consider the monomial ideal I of K[x1 , x2 , x3 ] generated by x21 , x1 x2 , x22 .
The characteristic function χ(p) of the ideal I is constant equal to 3. The unique point that annihilates the
ideal I is (0, 0, 1) with multiplicity 3. This result is compatible with the fact that the zeros of the ideal J
generated by the following polynomials
(x1 − ax3 )(x1 − bx3 ),
(x1 − ax3 )(x2 − cx3 ),
(x2 − cx3 )(x2 − dx3 ).
consists of the three points
(a, c, 1),
48
(a, d, 1),
(b, c, 1).
4.8. Involutive systems
4.7.4. The ideal-PDE dictionary. Let I be a homogeneous ideal of K[x1 , . . . , xn ] generated by a
set F = {f1 , . . . , fk } of polynomials. For a fixed monomial order on M(x1 , . . . , xn ), we set U = lm(F).
Consider the ring isomorphism Φ from K[x1 , ¨ ¨ ¨ , xn ] to K[ BxB1 , ¨ ¨ ¨ , BxBn ] given in 2.1.2. To any
polynomial f in I we associate a PDE Φ(f)ϕ = 0. In this way, the ideal I defines a PDE system (Σ(I)).
Let λ and µ be the integers associated to the characteristic function χ(p) as defined in 4.7.2. The maximal
number of arguments of the arbitrary analytic functions used to define the boundary conditions
{ Cβ | xβ P UK }
of the PDE system (Σ(I)), as defined in (11), corresponds to λ, explicitly
K
λ = max |MultU
J (v)|,
vPVU
where UK denotes the set of complementary monomials of U. Moreover, the number of arbitrary analytic
functions with λ arguments in the boundary conditions { Cβ | xβ P UK } is equal to µ, that is
K
µ = { v P UK | |MultU
J (v)| = λ } .
Conversely, let (Σ) be a PDE system with one unknown function ϕ of independent variables x1 , . . . , xn .
Consider the set, denoted by ldo(Σ), made of differential operators associated to the principal derivatives
of PDE in (Σ), with respect to Janet’s order on derivatives defined in 4.1.3. By isomorphism Φ,
|α|
to any monomial differential operator α1B αn in ldo(Σ), we associate a monomial xα1 1 . . . xαnn
Bx1 ¨¨¨Bxn
in M(x1 , . . . , xn ).
Let us denote by I(Σ) the ideal of K[x1 , . . . , xn ] generated by Φ−1 (ldo(Σ)). Note that, by construction
the ideal I(Σ) is monomial and for any monomial u in I(Σ) the derivative Φ(u)ϕ is a principal derivative
of the PDE system (Σ) as defined in Section 4.3.1. In [Jan29, §54], M. Janet called characteristic form
any element of the ideal I(Σ).
In this way, M. Janet concluded that the degree of generality of the solutions of a linear PDE system
with one unknown function is described by the leading term of the charateristic function of the ideal of
characteristic forms defined in 4.7.1.
4.7.5. The particular case of first order systems. Consider a completely integrable first order linear
PDE system (Σ). The number λ, defined in 4.7.4, that is equal to the maximal number of arguments of
the arbitrary functions used to define the boundary conditions of the system (Σ), is also equal in this case
to the cardinal of the set UK of complementary monomials of the set of monomials U = Φ−1 (ldo(Σ)).
4.8. Involutive systems
In this subsection, we recall the algebraic formulation of involutive systems as introduced by M. Janet.
This formulation first appeared in its work in [Jan22a] and [Jan22b]. But notice that this notion comes
from the work of É. Cartan in [Car04].
49
4. Polynomial partial differential equations systems
4.8.1. Characters and derived systems. Let I be a proper ideal of K[x1 , . . . , xn ] generated by homogeneous polynomials. M. Janet introduced the characters of the homogeneous component Ip as the non
negative integers σ1 , σ2 , . . . , σn defined inductively by the following formula
dim Ip +
h
ÿ
K[x1 , ¨ ¨ ¨ , xn ]p−1 xi
i=1
!!
= dim(Ip ) + σ1 + . . . + σh ,
1 ď h ď n.
Note that the sum σ1 + σ2 + . . . + σn corresponds to the codimension of Ip in K[x1 , ¨ ¨ ¨ , xn ]p .
Given a positive integer λ, we set
Jp+λ = K[x1 , . . . , xn ]λ Ip .
(λ)
(λ)
(λ)
We define the non negative integers σ1 , σ2 , . . . , σn by the relations
dim Jp+λ +
h
ÿ
K[x1 , ¨ ¨ ¨ , xn ]p+λ−1 xi
i=1
!!
(λ)
(λ)
= dim(Jp+λ ) + σ1 + . . . + σh
1 ď h ď n.
For λ = 1, M. Janet called Jp+1 the derived system of Ip . Let us mention some properties on these
numbers proved by M. Janet.
(1)
(2)
4.8.2. Lemma. We set σh1 = σh and σh2 = σh for 1 ď h ď n.
i) σ11 + σ21 + . . . + σn1 ď σ1 + 2σ2 + . . . + nσn .
ii) If σ11 + σ21 + . . . + σn1 = σ1 + 2σ2 + . . . + nσn , the two following relations hold:
a) σ12 + σ22 + . . . + σn2 = σ11 + 2σ21 + . . . + nσn1 .
b) σh1 = σh + σh+1 + . . . + σn .
We refer the reader to [Jan29] for a proof of the relations of Lemma 4.8.2.
4.8.3. Involutive systems. The homogenous component Ip is said to be in involution when the following
equality holds:
σ11 + σ21 + . . . + σn1 = σ1 + 2σ2 + . . . + nσn .
Following properties ii)-a) of Lemma 4.8.2, if the component Ip is in involution, then the component Ip+k is in involution for all k ě 0.
4.8.4. Proposition ([Jan29, §56 & §57]). The characters of a homogeneous component Ip satisfy the
two following properties
i) σ1 ě σ2 ě . . . ě σn .
ii) if Ip ‰ {0}, then σn = 0.
50
4.9. Conclusive remarks
4.8.5. Polynomiality of characteristic function. Suppose that the homogeneous component Ip is in
involution. We show that the characteristic function χ(P) defined in 4.7.1 is polynomial for P ě p. Using
Lemma 4.8.2, we show by induction that for any 1 ď h ă n and any positive integer λ, we have the
following relation:
n−h−1
ÿ λ + k − 1
(λ)
σh =
σh+k .
k
k=0
The codimension of Ip+λ in K[x1 , ¨ ¨ ¨ , xn ]p+λ is given by
n−1
ÿ
h=1
(λ)
σh
!
n−1
ÿ i−1
ÿ λ + k − 1
λ+k−1
=
σh+k =
σi
k
k
h=1 k=0
i=1 k=0
!
n−1
n−1
ÿ i−1
ÿ P − p + k − 1
ÿ P − p + i − 1
=
σi =
σi .
k
i−1
i=1 k=0
i=1
n−1
ÿ n−h−1
ÿ
This proves the polynomiality of the characteristic function of the ideal I for sufficiently big p.
4.9. Conclusive remarks
Recall that the so-called Cartan-Kähler theory is about the Pfaffian systems on a differentiable (or analytic)
manifold and its aim is to judge whether a given system is prolongeable to a completely integrable system
or an incompatible system. Their method relies on a geometrical argument, which is to construct integral
submanifolds of the system inductively. Here, a step of the induction is to find an integral submanifold
of dimension i + 1 containing the integral submanifold of dimension i, and their theory does not allow
one to see whether such step can be achieved or not.
Janet’s method is, even if it works only locally, completely algebraic and algorithmic so that it partially
completes the parts where one cannot treat with Cartan-Kähler theory.
By these works, there are two seemingly different notions of involutivity; the one by G. Frobenius,
G. Darboux and É. Cartan and the other by M. Janet. The fact is that at each step of the induction in the
Cartan-Kähler theory, one has to study a system of PDE. Its system is called in involution (cf. compare
those in Sections 1.2.6 with 4.8) if it can be written in a canonical system, as defined in 4.5.2, if necessary
after a change of coordinates. Following the algebraic definition of involutivity by M. Janet, several
involutive methods were developed for polynomial and differential systems, [Tho37, Pom78]. In these
approaches, a differential system is involutive when its non-multiplicative derivatives are consequences
of multiplicative derivatives. In [Ger97, GB98a], V. P. Gerdt gave an algebraic charaterization of the
involutivity for polynomial systems. The Gerdt’s approach is developed in the next section.
5. Polynomial involutive bases
In this section, we present the algebraic definition of involutivity for polynomial systems given by
V. P. Gerdt in [Ger97, GB98a]. In particular, we relate the notion of involutive basis for a polynomial
ideal to the notion of Gröbner basis.
51
5. Polynomial involutive bases
5.1. Involutive reduction on polynomials
5.1.1. Involutive basis. Recall that a monomial ideal I of K[x1 , . . . , xn ] is an ideal generated by monomials. An involutive basis of the ideal I with respect to an involutive division division I is an involutive
set of monomials U that generates I. By Dickson Lemma, [Dic13], any monomial ideal I admits a finite
set of generators. When the involutive division I is noetherian as defined in 3.2.3, this generating set
admits a finite I-completion that forms an involutive basis of the ideal I. As a consequence, we deduce
the following result.
5.1.2. Proposition. Let I be a noetherian involutive division on M(x1 , . . . , xn ). Any monomial ideal
of K[x1 , . . . , xn ] admits an I-involutive basis.
The objective of this section is to show how to extend this result to polynomial ideals with respect
to a monomial order. In the remainder of this subsection we assume that a monomial order ď is fixed
on M(x1 , . . . , xn ).
5.1.3. Multiplicative variables for a polynomial. Let I be an involutive division on M(x1 , . . . , xn ).
Let F be a set of polynomials of K[x1 , . . . , xn ] and f be a polynomial in F. We define the set of
I-multiplicative (resp. I-non-multiplicative) variables of the polynomial f with respect to F and the
monomial order ď by setting
lmď (F)
MultFI,ď (f) = MultI
(lmď (f)),
lmď (F)
( resp. NMultFI,ď (f) = NMultI
(lmď (f)) ).
Note that the I-multiplicative variables depend on the monomial order ď used to determine leading
monomials of polynomials of F.
5.1.4. Polynomial reduction. The polynomial division can be describe as a rewriting operation as
follows. Given polynomials f and g in K[x1 , . . . , xn ], we say that f is reducible modulo g with respect
to ď, if there is a term λu in f whose monomial u is divisible by lmď (g) for the usual monomial division.
gď
/ h, where
In that case, we denote such a reduction by f
h=f−
λu
g.
ltď (g)
For a set G of polynomials of K[x1 , . . . , xn ], we define a rewriting system corresponding to the division
Gď
/ defined by
modulo G by considering the relation reduction
Gď
/ =
ď gď
/.
gPG
We will denote by
52
Gď
/ ˚ the reflexive and transitive closure of the relation
Gď
/.
5.1. Involutive reduction on polynomials
5.1.5. Involutive reduction. In a same way, we define a notion of reduction with respect to an involutive
division I on M(x1 , . . . , xn ). Let g be a polynomial in K[x1 , . . . , xn ]. A polynomial f in K[x1 , . . . , xn ]
is said to be I-reducible modulo g with respect to the monomial order ď, if there is a term λu of f,
with λ P K − {0} and u P M(x1 , . . . , xn ), such that
u = lmď (g)v and
Such a I-reduction is denoted by f
gď
lmď (G)
v P M(MultI
(g)).
/ I h, where
h=f−
λu
λ
gv = f −
g.
lcď (g)
ltď (g)
5.1.6. Involutive normal forms. Let G be a set of polynomials of K[x1 , . . . , xn ]. A polynomial f is said
to be I-reducible modulo G with respect to the monomial order ď, if there exists a polynomial g in G
Gď
/ I this reduction relation defined by
such that f is I-reducible modulo g. We will denote by
Gď
/I =
ď gď
/ I.
gPG
The polynomial f is said to be in I-irreducible modulo G if it is not I-reducible modulo G. A I-normal
form of a polynomial f is a I-irreducible polynomial h such that there is a sequence of reductions from f
to h:
Gď
G
G
Gď
/ I h,
/ I f1 ď / I f2 ď / I . . .
f
The procedure InvReductionI,ď (f, G) computes a normal form of f modulo G with respect to the
division I. The proofs of its correctness and termination can be achieved as in the case of the division
procedure for the classical polynomial division, see for instance [BW93, Proposition 5.22].
Procedure 8: InvReductionI,ď (f, G)
Input: a polynomial f in K[x1 , . . . , xn ] and a finite subset G of K[x1 , . . . , xn ].
begin
h←f
lmď (G)
while exist g in G and a term t of h such that lmď (g)|I
t
lcď (t)
do
choose such a g
h ← h − ltďt(g) g
end
end
Output: h a I-normal form of the polynomial f with respect to the monomial order ď
53
5. Polynomial involutive bases
5.1.7. Remarks. Note that the involutive normal form of a polynomial f is not unique in general, it
depends on the order in which the reductions are applied. Suppose that, for each polynomial f we
have a I-normal form with respect to the monomial order ď, that is denoted by nf G
I,ď (f). Denote
G
by nf ď (f) a normal form of a polynomial f obtained by the classical division procedure. In general,
G
the equality nf G
ď (f) = nf I,ď (f) does not hold. Indeed, suppose that G = {x1 , x2 } and consider the
G
Thomas division T defined in 3.3.1. We have nf G
ď (x1 x2 ) = 0, while nf T,ď (x1 x2 ) = x1 x2 because the
monomial x1 x2 is a T-irreducible modulo G.
5.1.8. Autoreduction. Recall from 3.1.4 that a set of monomials U is I-autoreduced with respect to an
involutive division I if it does not contain a monomial I-divisible by another monomial of U. In that
case, any monomial in M(x1 , . . . , xn ) admits at most one I-involutive divisor in U.
A set G of polynomials of K[x1 , . . . , xn ] is said to be I-autoreduced with respect to the monomial
order ď, if it satisfies the two following conditions:
i) (left I-autoreducibility) the set of leading monomials lmď (G) is I-autoreduced,
ii) (right I-autoreducibility) for any g in G, there is no term λu ‰ ltď (g) of g, with λ ‰ 0
and u P coneI (lmď (G)).
Note that the condition i), (resp. ii)) corresponds to the left-reducibility (resp. right-reducibility)
property given in 4.5.2. Any finite set G of polynomials of K[x1 , . . . , xn ] can be transformed into a finite
I-autoreduced set that generates the same ideal by Procedure 9. The proofs of correctness and termination
are immediate consequences of the property of involutive division.
Procedure 9: AutoreduceI,ď (G)
Input: G a finite subset of K[x1 , . . . , xn ].
begin
H←G
H1 ← H
while exists h P H and g P H \ {h} such that h is I-reducible modulo g with respect to ď do
choose such a h
H 1 ← H \ {h}
1
h 1 ← nf H
I,ď (h)
if h 1 = 0 then
H ← H1
end
else
H ← H 1 Y {h 1 }
end
end
end
Output: H an I-autoreduced set generating the same ideal as G does.
54
5.1. Involutive reduction on polynomials
5.1.9. Proposition ([GB98a, Theorem 5.4]). Let G be an I-autoreduced set of polynomials
of K[x1 , . . . , xn ] and f be a polynomial in K[x1 , . . . , xn ]. Then nf G
I,ď (f) = 0 if and only if the polynomialf
can be written in the form
ÿ
f=
βi,j gi vi,j ,
i,j
lmď (G)
where gi P G, βi,j P K and vi,j P M(MultI
(lmď (gi ))), with lmď (vi,j ) ‰ lmď (vi,k ) if j ‰ k.
Proof. Suppose that nf G
I,ď (f) = 0, then there exists a sequence of involutive reductions modulo G:
f = f0
g1
/ I f1
g2
/ I f2
g3
/I ...
gk−1
/
I
fk = 0,
lm (G)
λ
ď
terminating on 0. For any 1 ď i ď k, we have fi = fi−1 − lcďi,j
(lmď (gi ))).
(gi ) gi vi,j , with vi,j P M(MultI
This show the equality.
Conversely, suppose that f can be written in the given form. Then the leading monomial lmď (f)
admits an involutive I-divisor in lmď (G). Indeed, the leading monomial of the decomposition of f has
the following form:
!
ÿ
lmď
gi vi,j = lmď (gi0 )vi0 ,j0 .
i,j
The monomial lmď (gi0 ) is an involutive divisor of lmď (f) and by autoreduction hypothesis, such a divisor
is unique. Hence the monomial lmď (gi0 )vi0 ,j0 does not divide other monomial of the form lmď (gi )vi,j .
g i0 ď
/ I 0 on the decomposition. In this way, we define a sequence of
We apply the reduction gi0 vi0 ,j0
reductions ending on 0. This proves that nf G
I,ď (f) = 0.
5.1.10. Unicity and additivity of normal forms. From decomposition 5.1.9, we deduce two important
properties on involutive normal forms. Let G be a I-autoreduced set of polynomials of K[x1 , . . . , xn ]
G
and f be a polynomial. Suppose that h1 = nf G
I,ď (f) and h2 = nf I,ď (f) are two involutive normal forms
of f. From the involutive reduction procedure that computes this two normal forms, we deduces two
decompositions
ÿ
ÿ
1
1
.
gi vi,j
h1 = f − βi,j gi vi,j ,
h2 = f − βi,j
i,j
i,j
As a consequence, h1 − h2 admits a decomposition as in Proposition 5.1.9, hence nf G
I,ď (h1 − h2 ) = 0.
The polynomial h1 − h2 being in normal form, we deduce that h1 = h2 . This shows the unicity of the
involutive normal form modulo an autoreduced set of polynomials.
In a same manner we prove the following additivity formula for any polynomial f and f 1 :
1
G
G
1
nf G
I,ď (f + f ) = nf I,ď (f) + nf I,ď (f ).
Finally, for any reduction procedure that computes a normal form nf G
ď (f) and any involutive reduction
G
procedure that computes an involutive normal form nf I,ď (f), as a consequence of unicity of the involutive
normal form and Proposition 5.1.11, we have
G
nf G
ď (f) = nf I,ď (f).
55
5. Polynomial involutive bases
5.1.11. Proposition. Let I be an involutive division on K[x1 , . . . , xn ] and G be a subset of K[x1 , . . . , xn ].
A polynomial of K[x1 , . . . , xn ] is reducible with respect to G if and only if it is I-reducible modulo G.
Proof. Let f be a polynomial of K[x1 , . . . , xn ]. By definition of involutive reduction, if f is I-reducible
Gď
/ . Conversely, suppose that f is reducible by a
modulo G, then it is reducible for the relation
polynomial g in G. That is there exists a term λu in f, where λ is a nonzero scalar and u is a monomial
of M(x1 , . . . , xn ) such that u = lmď (g)v, where v P M(x1 , . . . , xn ). The set G being involutive, we
have nf G
I,ď (gv) = 0. Following Proposition 5.1.9, the polynomial gv can written in the form:
ÿ
gv =
βi,j gi vi,j ,
i,j
lmď (G)
where gi P G, βi,j P K and vi,j P M(MultI
u admits an involutive divisor in G.
(lmď (gi ))). In particular, this shows that the monomial
5.2. Involutive bases
We fix a monomial order ď on M(x1 , . . . , xn ).
5.2.1. Involutive bases. Let I be an ideal of K[x1 , . . . , xn ]. A set G of polynomials of K[x1 , . . . , xn ] is
an I-involutive basis of the ideal I with respect the monomial order ď, if G is I-autoreduced and satisfies
the following property:
@g P G, @u P M(x1 , . . . , xn ),
nf G
I,ď (gu) = 0.
In other words, for any polynomial g in G and monomial u in M(x1 , . . . , xn ), there is a sequence of
involutive reductions:
gk−1 ď
g2 ď
g3 ď
g1 ď
/ I f2
/I ...
/ I 0,
/ I f1
gu
with gi in G. In particular, we recover the notion of involutive sets of monomials given in 3.2.1. Indeed,
if G is an I-involutive basis, then lmď (G) is an I-involutive set of monomials of M(x1 , . . . , xn ).
5.2.2. Example. We set U = {x1 , x2 }. We consider the deglex order induced by x2 ą x1 and the Thomas
division T. The monomial x1 x2 is T-irreducible modulo U. Hence, it does not admits zero as T-normal
form and the set U cannot be an T-involutive basis of the ideal generated by U. In turn the set {x1 , x2 , x1 x2 }
is a T-involutive basis of the ideal generated by U.
We now consider the Janet division J. We have deg2 (U) = 1, [0] = {x1 } and [1] = {x2 }. The
J-multiplicative variables are given by the following table:
u
x1
x2
MultU
J (u)
x1
x1
x2
It follows that the monomial x1 x2 is not J-reducible by x1 modulo U. However, it is J-reducible by x2 .
Hence the set U form a J-involutive basis.
As an immediate consequence of involutive bases, the involutive reduction procedure provides a
decision method of the ideal membership problem, as stated by the following result.
56
5.2. Involutive bases
5.2.3. Proposition ([GB98a, Corollary 6.4]). Let I be an ideal of K[x1 , . . . , xn ], and G be an
I-involutive basis of I with respect to a monomial order ď. For any polynomial f of K[x1 , . . . , xn ],
we have f P I if and only if nf G
I,ď (f) = 0.
Proof. If nf G
I,ď (f) = 0, then the polynomial f can be written in the form 5.1.9. This shows that f belongs
to the ideal I. Conversely, suppose that f belongs to I, then it can be decomposed in the form
ÿ
f=
h i gi ,
i
ř
where hi = j λi,j ui,j P K[x1 , . . . , xn ]. The set G being I-involutive, we have nf G
I,ď (ui,j gi ) = 0, for any
G
monomials ui,j and gi in G. By linearity of the operator nf I,ď (−), we deduce that nf G
I,ď (f) = 0.
5.2.4. Local involutivity. V. P. Gerdt and Y. A. Blinkov introduced in [GB98a] the notion of local
involutivity for a set of polynomials. A set G of polynomials of K[x1 , . . . , xn ] is said to be locally
involutive if the following condition holds
lmď (G)
@g P G, @x P NMultI
(lmď (g)),
nf G
I,ď (gx) = 0.
For a continuous involutive division I, they prove that a I-autoreduced set of polynomials is involutive if
and only if it is locally involutive, [GB98a, Theorem 6.5]. This criterion of local involutivity is essential
for computing the completion of a set of polynomials into an involutive basis. Note that this result is
analogous to the critical pair lemma in rewriting theory stating that a rewriting system is locally confluent
if and only if all its critical pairs are confluent, see e.g. [GM16, Sect. 3.1]. Together with the Newman
Lemma stating that for terminating rewriting, local confluence and confluence are equivalent properties,
this gives a constructive method to prove confluence in a terminating rewriting system by analyzing the
confluence of critical pairs.
5.2.5. Completion procedure. For a given monomial order ď on M(x1 , . . . , xn ) and a continuous and
constructive involutive division I, as defined in [GB98a, Definition 4.12], the Procedure 10 computes an
I-involutive basis of an ideal from a set of generators of the ideal. We refer the reader to [GB98a, Sect. 8]
or [Eva06, Sect. 4.4] for correctness of this procedure and conditions for its termination. This procedure
is in the same vein as the completion procedure for rewriting systems by Knuth-Bendix, [KB70], and
completion procedure for commutative polynomials by Buchberger, [Buc65].
5.2.6. Example: computation of an involutive basis. Let I be the ideal of Q[x1 , x2 ] generated by the
set F = {f1 , f2 }, where the polynomial f1 and f2 are defined by
f1 = x22 − 2x1 x2 + 1,
f2 = x1 x2 − 3x21 − 1.
We compute an involutive basis of the ideal I with respect to the Janet division J and the deglex order
induced by x2 ą x1 . We have lm(f1 ) = x22 et lm(f2 ) = x1 x2 , hence the following J-reductions
x22
f1
/ J 2x1 x2 − 1,
x1 x2
f2
/ J 3x2 + 1.
1
57
5. Polynomial involutive bases
Procedure 10: InvolutiveCompletionBasis I,ď (F)
Input: F a finite set of polynomials in K[x1 , . . . , xn ].
begin
F 1 ← AutoreduceI,ď (F)
G←H
while G = H do
1
P ← {fx | f P F 1 , x P NMultFI,ď (f)}
p1 ← 0
while P ‰ H and p 1 = 0 do
choose p in P such that lmď (p) is minimal with respect to ď.
P ← P \ {p}
p 1 ← InvReductionI,ď (p, F 1 )
end
if p 1 ‰ 0 then
F 1 ← AutoreduceI,ď (F 1 Y {p 1 })
end
else
G ← F1
end
end
end
Output: G an I-involutive basis of the ideal generated by F with respect to the monomial order ď.
58
5.2. Involutive bases
The polynomial f1 is J-reducible by f2 , we have
f1
f2
/ J x2 − 2(3x2 + 1) + 1 = x2 − 6x2 − 1.
2
1
2
1
Thus, we set f3 = x22 − 6x21 − 1 and we consider the reduction
x22
f3
/ J 6x2 + 1.
1
The set F 1 = {f2 , f3 } is J-autoreduced and generates the ideal I.
Let us compute the multiplicative variables of the polynomials f2 and f3 . We have deg2 (F 1 ) =
deg2 ({x22 , x1 x2 }) = 2, [1] = {x1 x2 } and [2] = {x22 }. Hence the J-multiplicative variables are given by the
following table:
f
f2
f3
lm(f)
x1 x2
x22
1
MultFJ (f)
x1
x1
x2
The polynomial f2 x2 = x1 x22 − 3x21 x2 − x2 is the only non-multiplicative prolongation to consider.
This prolongation can be reduced as follows
f2 x2
f3
f2
/ J 6x3 + x1 − 3x2 x2 − x2
1
1
/ J − 3x3 − 2x1 − x2 .
1
We set f4 = −3x31 − 2x1 − x2 , whose associated reduction is
x31
f4
/ J − 2 x1 − 1 x2 ,
3
3
and we set F 1 = {f2 , f3 , f4 }. We have deg2 (F 1 ) = 2, [0] = {x31 }, [1] = {x1 x2 } and [2] = {x22 }. Hence the
J-multiplicative variables are given by the following table:
f
f2
f3
f4
lm(f)
x1 x2
x22
x31
1
MultFJ (f)
x1
x1
x1
x2
There are two non-multiplicative prolongations to consider:
f2 x2 = x1 x22 − 3x21 x2 − x2 ,
f4 x2 = −3x31 x2 − 2x1 x2 − x22 .
We have lm(f2 x2 ) = x1 x22 ă lm(f4 x2 ) = x31 x2 . Hence the prolongation f2 x2 must first be examined. We
have the following reductions:
f2 x2
f3
/ J 6x3 + x1 − 3x2 x2 − x2
1
1
f2
/ J − 3x3 − 2x1 − x2
1
f4
/ J 0.
59
5. Polynomial involutive bases
Hence, there is no polynomial to add. The other non-multiplicative prolongation is f4 x2 , that can be
reduced to an J-irreducible polynomial as follows:
f4 x2
f2
/ J − 3x3 x2 − 6x2 − x2 − 2
2
1
1
f2
f3
/ J − 3x3 x2 − 12x2 − 3
1
1
/ J −9x4 −15x2 −3
1
1
f4
/ J 3x1 x2 −9x2 −3
1
f2
/ J 0.
All the non-multiplicative prolongations are J-reducible to 0, it follows that the set F 1 is a Janet basis of
the ideal I.
5.3. Involutive bases and Gröbner bases
In this subsection, we recall the notion of Gröbner basis and we show that any involutive basis is a Gröbner
basis. We fix a monomial order ď on M(x1 , . . . , xn ).
5.3.1. Gröbner bases. A subset G of K[x1 , . . . , xn ] is a Gröbner basis with respect to the monomial
order ď if it is finite and satisfies one of the following equivalent conditions
Gď
i)
Gď
ii)
Gď
iii)
Gď
iv)
/ is Church-Rosser,
/ is confluent,
/ is locally confluent,
/ has unique normal forms,
Gď
v) f
/ ˚ 0, for all polynomial f in Id(G),
vi) every polynomial f in Id(G) \ {0} is reducible modulo G,
vii) for any term t in ltď (Id(G)), there is g in G such that ltď (g) divides t,
viii) Sď (g1 , g2 )
G
/ ˚ 0 for all g1 , g2 in G, where
Sď (g1 , g2 ) =
µ
µ
g1 −
g2 ,
ltď (g1 )
ltď (g2 )
with µ = ppcm(lmď (g1 ), lmď (g2 )), is the S-polynomial of g1 and g2 with respect to the monomial
order ď,
60
REFERENCES
ix) any critical pair
✉ µ ■■■ µ g2
■■lt(g2 )
✉✉
✉
✉
■■
■■
✉✉
✉
■$
z✉
✉
µ
g
lt(g1 ) 1
µ−
µ
lt(g1 ) g1
with µ = ppcm(lm(g1 ), lm(g2 )), of the relation
µ−
µ
lt(g2 ) g2
G /
is confluent.
We refer the reader to [BW93, Theorem 5.35] for proofs of these equivalences, see also [GHM17,
Section 3]. The equivalence of conditions i)-iv) are classical results for terminating rewriting systems.
Note that condition viii) corresponds to the Buchberger criterion, [Buc65], and condition ix) is a formulation of this criterion in rewriting terms. We refer to [BN98, Chapter 8] for the rewriting interpretation
of the Buchberger algorithm.
A Gröbner basis of an ideal I of K[x1 , . . . , xn ] with respect to a monomial order ď is a Gröbner basis
with respect to ď that generates the ideal I. This can be also be formulated saying that G is a generating
set for I such that Id(lt(G)) = Id(lt(I)).
5.3.2. Involutive bases and Gröbner bases. Let I be an ideal of K[x1 , . . . , xn ]. Suppose that G is an
involutive basis of the ideal I with respect to an involutive division I and the monomial order ď. In
particular, the set G generates the ideal I. For every g1 and g2 in G, we consider the S-polynomial
Sď (g1 , g2 ) with respect to ď. By definition, the polynomial Sď (g1 , g2 ) belongs to the ideal I. By
involutivity of the set G and following 5.1.10 and Proposition 5.2.3, we have
nf G (Sď (g1 , g2 )) = nf G
I (Sď (g1 , g2 )) = 0.
In this way, G is a Gröbner basis of the ideal I by the Buchberger criterion viii). We have thus proved the
following result due to V.P. Gerdt and Y.A. Blinkov.
5.3.3. Theorem ([GB98a, Corollary 7.2]). Let ď be a monomial order on M(x1 , . . . , xn ) and I be an
involutive division on K[x1 , . . . , xn ]. Any I-involutive basis of an ideal I of K[x1 , . . . , xn ] is a Gröbner
basis of I.
References
[BCG+ 91] R. L. Bryant, S. S. Chern, R. B. Gardner, H. L. Goldschmidt, and P. A. Griffiths. Exterior differential
systems, volume 18 of Mathematical Sciences Research Institute Publications. Springer-Verlag, New
York, 1991.
[Ber78]
G. M. Bergman. The diamond lemma for ring theory. Adv. in Math., 29(2):178–218, 1978.
[BN98]
F. Baader and T. Nipkow. Term rewriting and all that. Cambridge University Press, 1998.
[Bok76]
L. A. Bokut. Imbeddings into simple associative algebras. Algebra i Logika, 15(2):117–142, 245,
1976.
61
REFERENCES
[Buc65]
B. Buchberger. Ein Algorithmus zum Auffinden der Basiselemente des Restklassenringes nach einem
nulldimensionalen Polynomideal (An Algorithm for Finding the Basis Elements in the Residue Class
Ring Modulo a Zero Dimensional Polynomial Ideal). PhD thesis, University of Innsbruck, Austria,
1965.
[Buc70]
B. Buchberger. Ein algorithmisches Kriterium für die Lösbarkeit eines algebraischen Gleichungssystems. Aequationes Math., 4:374–383, 1970.
[Buc87]
B. Buchberger. History and basic features of the critical-pair/completion procedure. J. Symbolic
Comput., 3(1-2):3–38, 1987. Rewriting techniques and applications (Dijon, 1985).
[Buc06]
B. Buchberger. An algorithm for finding the basis elements of the residue class ring of a zero
dimensional polynomial ideal. J. Symbolic Comput., 41(3-4):475–511, 2006. Translated from the
1965 German original by Michael P. Abramson.
[BW93]
T. Becker and V. Weispfenning. Gröbner bases, volume 141 of Graduate Texts in Mathematics.
Springer-Verlag, New York, 1993. A computational approach to commutative algebra, In cooperation
with Heinz Kredel.
[Car99]
É. Cartan. Sur certaines expressions différentielles et le problème de Pfaff. Ann. Sci. Éc. Norm. Supér.
(3), 16:239–332, 1899.
[Car01]
É. Cartan. Sur l’intégration des systèmes d’équations aux différentielles totales. Ann. Sci. Éc. Norm.
Supér. (3), 18:241–311, 1901.
[Car04]
É. Cartan. Sur la structure des groupes infinis de transformations. Ann. Sci. Éc. Norm. Supér. (3),
21:153–206, 1904.
[Car45]
É. Cartan. Les systèmes différentiels extérieurs et leurs applications géométriques. Actualités Sci. Ind.,
no. 994. Hermann et Cie., Paris, 1945.
[Dar82]
G. Darboux. Sur le problème de Pfaff. Bull. Sci. Math. et Astro., 6:14–36, 49–68, 1882.
[Deh10]
M. Dehn. Über die Topologie des dreidimensionalen Raumes. Math. Ann., 69(1):137–168, 1910.
[Dic13]
L. E. Dickson. Finiteness of the Odd Perfect and Primitive Abundant Numbers with n Distinct Prime
Factors. Amer. J. Math., 35(4):413–422, 1913.
[Eis95]
D. Eisenbud. Commutative algebra, volume 150 of Graduate Texts in Mathematics. Springer-Verlag,
New York, 1995. With a view toward algebraic geometry.
[Eva06]
G. A. Evans. Noncommutative Involutive Bases. PhD thesis, University of Wales, Bangor, February
2006. PhD Thesis, 2006.
[Fau99]
J.-C. Faugère. A new efficient algorithm for computing Gröbner bases (F4 ). J. Pure Appl. Algebra,
139(1-3):61–88, 1999. Effective methods in algebraic geometry (Saint-Malo, 1998).
[Fau02]
J.-C. Faugère. A new efficient algorithm for computing Gröbner bases without reduction to zero (F5 ).
In Proceedings of the 2002 International Symposium on Symbolic and Algebraic Computation, pages
75–83. ACM, New York, 2002.
[For90]
A. R. Forsyth. Theory of differential equations. Part I. Exact equations and Pfaff’s problem. Cambridge.
University Press. XIII + 340 S. 8˝ (1890)., 1890.
[Fro77]
G. Frobenius. Ueber das Pfaffsche Problem. J. Reine Angew. Math., 82:230–315, 1877.
[Gal85]
A. Galligo. Some algorithmic questions on ideals of differential operators. In EUROCAL ’85, Vol. 2
(Linz, 1985), volume 204 of Lecture Notes in Comput. Sci., pages 413–421. Springer, Berlin, 1985.
62
REFERENCES
[GB98a]
V. P. Gerdt and Y. A. Blinkov. Involutive bases of polynomial ideals. Math. Comput. Simulation, 45(56):519–541, 1998. Simplification of systems of algebraic and differential equations with applications.
[GB98b]
V. P. Gerdt and Y. A. Blinkov. Minimal involutive bases. Math. Comput. Simulation, 45(5-6):543–560,
1998. Simplification of systems of algebraic and differential equations with applications.
[Ger97]
V. P. Gerdt. Gröbner bases and involutive methods for algebraic and differential equations. Math.
Comput. Modelling, 25(8-9):75–90, 1997. Algorithms and software for symbolic analysis of nonlinear
systems.
[GHM17] Y. Guiraud, E. Hoffbeck, and P. Malbos. Convergent presentations and polygraphic resolutions of
associative algebras. arXiv:1406.0815, December 2017.
[GM16]
Y. Guiraud and P. Malbos. Polygraphs of finite derivation type. Mathematical Structures in Computer
Science, pages 1–47, 009 2016.
[Gra44]
H. E. Grassmann. Die lineale Ausdehnungslehre. Leipzig: Verlag von Otto Wigand. 324 p. , 1844.
[Gra72]
H. Grauert. Über die Deformation isolierter Singularitäten analytischer Mengen. Invent. Math.,
15:171–198, 1972.
[Gri83]
P. A. Griffiths. Exterior differential systems and the calculus of variations, volume 25 of Progress in
Mathematics. Birkhäuser, Boston, Mass., 1983.
[Grö37]
W. Gröbner. Über das macaulaysche inverse system und dessen bedeutung für die theorie der linearen
differentialgleichungen mit konstanten koeffizienten. Abh. Math. Sem. Univ. Hamburg, 12(1):127–132,
1937.
[Grö49]
W. Gröbner. Moderne algebraische Geometrie. Die idealtheoretischen Grundlagen. Springer-Verlag,
Wien und Innsbruck, 1949.
[GS64]
V. W. Guillemin and S. Sternberg. An algebraic model of transitive differential geometry. Bull. Amer.
Math. Soc., 70:16–47, 1964.
[Gün13a]
N. M. Günter. Über die Elimination. St. Petersburg, Sborn. Inst. put. Soobšč. 83, 1-10 (1913), auch
separat (1914) (1913,1914)., 1913.
[Gün13b]
N. Günther. Über die kanonische Form der Systeme kanonischer homogener Gleichungen. Samml.
des Inst. der Verkehrswege 82, 22 S. l ß (1913)., 1913.
[Gün13c]
N. Günther. Über einige Zusammenhänge zwischen den homogenen Gleichungen. Samml. des Inst.
der Verkehrwege 82, 20 S. (1913)., 1913.
[Gun41]
N. M. Gunther. Sur les modules des formes algébriques. Tr. Tbilis. Mat. Inst. 9, 97-206 (1941)., 1941.
[Her26]
G. Hermann. Die Frage der endlich vielen Schritte in der Theorie der Polynomideale. Math. Ann.,
95(1):736–788, 1926.
[Hil90]
D. Hilbert. Ueber die Theorie der algebraischen Formen. Math. Ann., 36(4):473–534, 1890.
[Hir64]
H. Hironaka. Resolution of singularities of an algebraic variety over a field of characteristic zero. I, II.
Ann. of Math. (2) 79 (1964), 109–203; ibid. (2), 79:205–326, 1964.
[Hue80]
G. Huet. Confluent reductions: abstract properties and applications to term rewriting systems. J.
Assoc. Comput. Mach., 27(4):797–821, 1980.
[Jac27]
C.G.J. Jacobi. Ueber die Integration der partiellen Differentialgleichungen erster Ordnung. J. Reine
Angew. Math., 2:317–329, 1827.
63
REFERENCES
[Jan20]
M. Janet. Sur les systèmes d’équations aux dérivées partielles. Journal de mathématiques pures et
appliquées, 8(3):65–151, 1920.
[Jan22a]
M. Janet. Les caractères des modules de formes et les systèmes d’équations aux dérivées partielles. C.
R. Acad. Sci., Paris, 174:432–434, 1922.
[Jan22b]
M. Janet. Sur les formes canoniques invariantes des systèmes algébriques et différentiels. C. R. Acad.
Sci., Paris, 174:991–994, 1922.
[Jan24]
M. Janet. Les modules de formes algébriques et la théorie générale des systèmes différentielles. Ann.
Sci. Éc. Norm. Supér. (3), 41:27–65, 1924.
[Jan29]
M. Janet. Leçons sur les systèmes d’équations aux dérivées partielles. VIII + 124 p. Paris, GauthierVillars (Cahiers scientifiques publiés sous la direction de G. Julia, fasc. IV.) (1929)., 1929.
[Käh34]
E. Kähler. Einführung in die Theorie der Systeme von Differentialgleichungen. (Hamburg. Math.
Einzelschr. 16) Leipzig, Berlin: B. G. Teubner IV, 80 S. (1934)., 1934.
[KB70]
D. Knuth and P. Bendix. Simple word problems in universal algebras. In Computational Problems in
Abstract Algebra (Proc. Conf., Oxford, 1967), pages 263–297. Pergamon, Oxford, 1970.
[Kro82]
L. Kronecker. Grundzüge einer arithmetischen Theorie der algebraischen Grössen. (Festschrift zu
Herrn Ernst Eduard Kummers fünfzigjährigem Doctor-Jubiläum, 10 September 1881). J. Reine Angew.
Math., 92:1–122, 1882.
[Kur57]
M. Kuranishi. On E. Cartan’s prolongation theorem of exterior differential systems. Amer. J. Math.,
79:1–47, 1957.
[Lag72]
J.-L. Lagrange. Sur l’intégration des équations à différences partielles du premier ordre. Mém. Acad.
Sci. et Belles-Lettres de Berlin, pages 353–372, 1772.
[Lag88]
J.-L. Lagrange. Méchanique Analitique. Desaint, 1788.
[Mac16]
F. S. Macaulay. The algebraic theory of modular systems. Cambridge: University press, XIV u. 112
S. 8˝ (1916)., 1916.
[Mac27]
F. S. Macaulay. Some properties of enumeration in the theory of modular systems. Proc. Lond. Math.
Soc. (2), 26:531–555, 1927.
[Man96]
E. L. Mansfield. A simple criterion for involutivity. J. London Math. Soc. (2), 54(2):323–345, 1996.
[Mar47a]
A. Markov. On the impossibility of certain algorithms in the theory of associative systems. Doklady
Akad. Nauk SSSR (N.S.), 55:583–586, 1947.
[Mar47b]
A. Markov. On the impossibility of certain algorithms in the theory of associative systems. II. Doklady
Akad. Nauk SSSR (N.S.), 58:353–356, 1947.
[New42]
M. Newman. On theories with a combinatorial definition of “equivalence”. Ann. of Math. (2),
43(2):223–243, 1942.
[Niv73]
M. Nivat. Congruences parfaites et quasi-parfaites. In Séminaire P. Dubreil, 25e année (1971/72),
Algèbre, Fasc. 1, Exp. No. 7, page 9. Secrétariat Mathématique, Paris, 1973.
[Noe21]
E. Noether. Idealtheorie in ringbereichen. Mathematische Annalen, 83:24–66, 1921.
[Noe23]
E. Noether. Eliminationstheorie und allgemeine Idealtheorie. Math. Ann., 90(3-4):229–261, 1923.
[Noe24]
E. Noether. Eliminationstheorie und Idealtheorie. Jahresber. Dtsch. Math.-Ver., 33:116–120, 1924.
64
REFERENCES
[Pfa02]
J. F. Pfaff. Allgemeine Methode, partielle Differentialgleichungen zu integrieren (1815). Aus dem
Lateinischen übersetzt und herausgegeben von G. Kowalewski.. 84 S. 8vo (Ostwalds Klassiker No. 129)
(1902)., 1902.
[Pom78]
J.-F. Pommaret. Systems of partial differential equations and Lie pseudogroups, volume 14 of Mathematics and its Applications. Gordon & Breach Science Publishers, New York, 1978. With a preface
by André Lichnerowicz.
[Pos47]
E. L. Post. Recursive unsolvability of a problem of Thue. J. Symbolic Logic, 12:1–11, 1947.
[Riq93]
Ch. Riquier. De l’existence des intégrales dans un système différentiel quelconque. Ann. Sci. Éc. Norm.
Supér. (3), 10:65–86, 1893.
[Riq10]
Ch. Riquier. Les systèmes d’équations aux dérivées partielles. XXVII - 590 p. Paris, Gauthier-Villars.,
1910.
[Riq28]
Ch. Riquier. La méthode des fonctions majorantes et les systèmes d’équations aux dérivées partielles.
Paris: Gauthier-Villars (Mémorial des sciences mathématiques, fasc. 32). 63 p. , 1928.
[Rob14]
D. Robertz. Formal algorithmic elimination for PDEs, volume 2121 of Lecture Notes in Mathematics.
Springer, Cham, 2014.
[Sch80]
F.-O. Schreyer. Die Berechnung von Syzygien mit dem verallgemeinerten Weierstrass’schen Divisionssatz. PhD thesis, Universität Hamburg, Germany, 1980.
[Sch92]
F. Schwarz. An algorithm for determining the size of symmetry groups. Computing, 49(2):95–115,
1992.
[Shi62]
A. I. Shirshov. Some algorithmic problems for Lie algebras. Sib. Mat. Zh., 3:292–296, 1962.
[SS65]
I. M. Singer and S. Sternberg. The infinite groups of Lie and Cartan. I. The transitive groups. J.
Analyse Math., 15:1–114, 1965.
[Tho37]
J. M. Thomas. Differential systems. IX + 118 p. New York, American Mathematical Society (American
Mathematical Society Colloquium Publications Vol. XXI) (1937)., 1937.
[Thu14]
A. Thue. Probleme über Veränderungen von Zeichenreihen nach gegebenen Regeln. Kristiania Vidensk.
Selsk, Skr., 10:493–524, 1914.
[Tre94]
A. Tresse. Sur les invariants différentie1s des groupes continus de transformations. Acta Math.,
18:1–88, 1894.
[van30]
B. L. van der Waerden. Moderne Algebra. Bd. I. Unter Benutzung von Vorlesungen von E. Artin und E.
Noether. Die Grundlehren der mathematischen Wissenschaften in Einzeldarstellungen mit besonderer
Berücksichtigung der Anwendungsgebiete. Bd. 23. Berlin: J. Springer, viii, 243 S. (1930)., 1930.
[van31]
B.L. van der Waerden. Moderne Algebra. II. Teil. Unter Benutzung von Vorlesungen von E. Artin
und E. Noether. Die Grundlehren der mathematischen Wissenschaften in Einzeldarstellungen mit
besonderer Berücksichtigung der Anwendungsgebiete. Bd. 34. Berlin: Julius Springer. vii, 216 S.
(1931)., 1931.
[von75]
S. von Kowalevsky. Zur Theorie der partiellen Differentialgleichungen. J. Reine Angew. Math.,
80:1–32, 1875.
[vW00]
E. von Weber. Vorlesungen über das Pfaff’sche Problem und die Theorie der partiellen Differentialgleichungen erster Ordnung. Leipzig: B. G. Teubner. XI + 622 S. gr. 8˝ (1900)., 1900.
[Wei86]
K. Weierstrass. Abhandlungen aus der Functionenlehre. Berlin. J. Springer. IV. u. 262 S. gr. 8˝ (1886).,
1886.
65
REFERENCES
Kenji Iohara
[email protected]
Philippe Malbos
[email protected]
— January 3, 2018 - 1:27 —
66
Université de Lyon,
Institut Camille Jordan, CNRS UMR 5208
Université Claude Bernard Lyon 1
43, boulevard du 11 novembre 1918,
69622 Villeurbanne cedex, France
Université de Lyon,
Institut Camille Jordan, CNRS UMR 5208
Université Claude Bernard Lyon 1
43, boulevard du 11 novembre 1918,
69622 Villeurbanne cedex, France
| 0math.AC
|
The Semantics of Graph Programs
Detlef Plump
Sandra Steinert
Department of Computer Science
The University of York, UK
Department of Computer Science
The University of York, UK
GP (for Graph Programs) is a rule-based, nondeterministic programming language for solving graph
problems at a high level of abstraction, freeing programmers from handling low-level data structures.
The core of GP consists of four constructs: single-step application of a set of conditional graphtransformation rules, sequential composition, branching and iteration. We present a formal semantics
for GP in the style of structural operational semantics. A special feature of our semantics is the use
of finitely failing programs to define GP’s powerful branching and iteration commands.
1
Introduction
This paper defines the semantics of GP, an experimental nondeterministic programming language for
high-level problem solving in the domain of graphs. The language is based on conditional rule schemata
for graph transformation (introduced in [16]) and thereby frees programmers from handling low-level
data structures for graphs. The prototype implementation of GP compiles graph programs into bytecode
for the York abstract machine, and comes with a graphical editor for programs and graphs [11].
GP has a simple syntax as its core contains only four commands: single-step application of a set of
rule schemata, sequential composition, branching and as-long-as-possible iteration. Despite its simplicity, GP is computationally complete in that every computable function on graphs can be programmed
[8]. A major goal of the GP project is the development of a practical graph-transformation language that
comes with a concise formal semantics, to facilitate program verification and other formal reasoning on
programs. Also, a formal semantics provides implementors with a rigorous definition of the language
that does not depend on a compiler or machine.
To define the meaning of GP programs, we adopt Plotkin’s method of structural operational semantics
[14]. This approach is well established for imperative programming languages [13] but is novel in the
field of graph transformation. In brief, the method consists in devising inference rules which inductively
define the effect of commands on program states. Whereas a classic state consists of the values of all
program variables at a certain point in time, the analogue for graph transformation is the graph on which
the rules of a program operate.
As GP is nondeterministic, our semantics assigns to a program P and an input graph G all graphs that
can result from executing P on G. A special feature of the semantics is the use of failing computations
to define powerful branching and iteration constructs. (Failure occurs when a set of rule schemata to
be executed is not applicable to the current graph.) While the conditions of branching commands in
traditional programming languages are boolean expressions, GP uses arbitrary programs as conditions.
The evaluation of a condition C succeeds if there exists an execution of C on the current graph that
produces a graph. On the other hand, the evaluation of C is unsuccessful if all executions of C on the
current graph result in failure. In this case C finitely fails on the current graph.
In logic programming, finite failure (of SLD resolution) is used to define negation [4]. In the case
of GP, it allows to “hide” destructive executions of the condition C of a statement if C then P else Q.
This is because after evaluating C, the resulting graph is discarded and either P or Q is executed on the
I. Mackie and A. Martins Moreira (Eds.): Tenth International
Workshop on Rule-Based Programming (RULE 2009)
EPTCS 21, 2010, pp. 27–38, doi:10.4204/EPTCS.21.3
c D. Plump & S. Steinert
Semantics of Graph Programs
28
graph with which the branching statement was entered. Finite failure also allows to elegantly lift the
application of as-long-as-possible iteration from sets of rule schemata (as in [16]) to arbitrary programs:
the body of a loop can no longer be applied if it finitely fails on the current graph.
Control constructs which allow programmers to write “strategies” for applying rewrite rules have
long been present in term-rewriting languages such as Elan [2] and Stratego [3]. These languages allow
recursive definitions of strategies whereas GP is based on a small set of built-in, non-recursive constructs.
(See [19] for an extension of GP with recursive procedures.)
Another difference between GP and languages such as Elan and Stratego is that strategies in the
latter languages rely on the structure of the objects that they manipulate, that is, on the tree structure of
terms. In both languages, term-rewrite rules are applied at the root of a term so that traversal operations
are needed to apply rules and strategies deep inside terms. In contrast, the semantics of GP’s control
constructs does not depend on the structure of graphs and is completely orthogonal to the semantics
of rule schemata. This provides a clear separation of concerns between rules and the control of rules,
making it easy to adapt GP’s semantics to different formats of rules or graphs.1
The contributions of this paper can be summarised as follows:
• A graph-transformation language with simple syntax and semantics, facilitating understanding by
programmers and formal reasoning on programs. Our experience so far is that very often short
and easy to understand programs can be written to solve problems on graphs (see [15] for various
small case studies).
• The first formal operational semantics for a graph-transformation language (to the best of our
knowledge). Well-known languages such as AGG [6], Fujaba [12] and GrGen [7] have no formal
semantics. The only graph-transformation language with a complete formal semantics that we
are aware of is PROGRES [18]. Its semantics, given by Schürr in his dissertation [17], translates
programs into control-flow diagrams and consists of more than 300 rules (including the definition
of the static semantics) .
• A powerful branching construct based on the concept of finite failure, allowing to conveniently
express complex destructive tests on input graphs. In addition, finite failure enables an elegant
definition of as-long-as-possible iteration. These definitions do not depend on the structure of
graphs and can be used for string- or term-based rewriting languages, too.
The rest of this paper is structured as follows. The next section reviews the graph-transformation
formalism underlying GP, the so-called double-pushout approach with relabelling. Section 3 introduces
conditional rule schemata as the building blocks of GP programs. In Section 4, we discuss an example
program for graph colouring and define the abstract syntax of graph programs. Section 5 presents our
formal semantics of GP in the style of structural operational semantics. In Section 6, we conclude and
mention some topics for future work.
2
Graph Transformation
We briefly review the model of graph transformation underlying GP, the double-pushout approach with
relabelling [9]. Our presentation is tailored to GP in that we consider graphs over a fixed label alphabet,
and rules in which only the interface may contain unlabelled nodes.
GP programs operate on graphs labelled with sequences of integers and strings. (The reason for using
sequences will become clear in Section 4.) To formalise this, let Z be the set of integers and Char be a
1 In
the extreme, one could even replace the underlying formalism of graph-transformation with some other rule-based
framework, such as string or term rewriting.
D. Plump & S. Steinert
29
finite set of characters—we may think of Char as the characters that can be typed on a keyboard. We fix
the label alphabet L = (Z ∪ Char∗ )+ consisting of all nonempty sequences made up from integers and
character strings.
A partially labelled graph over L (or graph for short) is a system G = (VG , EG , sG ,tG , lG , mG ), where
VG and EG are finite sets of nodes (or vertices) and edges, sG ,tG : EG → VG are the source and target
functions for edges, lG : VG → L is the partial node labelling function and mG : EG → L is the (total)
edge labelling function. Given a node v, we write lG (v) =⊥ to express that lG (v) is undefined. Graph G
is totally labelled if lG is a total function.
The set of all totally labelled graphs over L is denoted by G . GP programs operate on the graphs
in G , unlabelled nodes occur only in the interfaces of rules (see below) and are necessary in the doublepushout approach to relabel nodes. There is no need to relabel edges as they can always be deleted and
reinserted with changed labels.
A graph morphism g : G → H between graphs G and H consists of two functions gV : VG → VH
and gE : EG → EH that preserve sources, targets and labels (that is, sH ◦ gE = gV ◦ sG , tH ◦ gE = gV ◦ tG ,
mH ◦ gE = mG , and lH (g(v)) = lG (v) for all v such that lG (v) 6=⊥). Morphism g is an inclusion if g(x) = x
for all nodes and edges x. It is injective if gV and gE are injective.
A rule r = (L ← K → R) consists of two inclusions K → L and K → R where L and R are totally
labelled graphs. Graph K is the interface of r. Intuitively, an application of r to a graph will remove the
items in L − K, preserve K, add the items in R − K, and relabel the unlabelled nodes in K. Given a graph
G in G , an injective graph morphism g : L → G is a match for r if it satisfies the dangling condition: no
node in g(L) − g(K) is incident to an edge in G − g(L). In this case G directly derives the graph H in G
that is constructed from G as follows:2
1. Remove all nodes and edges in g(L) − g(K).
2. Add disjointly all nodes and edges from R − K, keeping their labels. For e ∈ ER − EK , sH (e) is
sR (e) if sR (e) ∈ VR −VK , otherwise gV (sR (e)). Targets are defined analogously.
3. For each node v in K with lK (v) = ⊥, lH (gV (v)) becomes lR (v).
We write G ⇒r,g H (or just G ⇒r H) if G directly derives H as above.
Figure 1 shows an example of a direct derivation. The rule in the upper row is applied to the left
graph of the lower row, resulting in the right graph of the lower row. For simplicity, we do not depict
edge labels and assume that they are all the same. The node identifiers 1 and 2 in the rule specify the
inclusions of the interface. The middle graph of the lower row is an intermediate result (omitted in
the above construction). This diagram represents a double-pushout in the category of partially labelled
graphs over L .
To define conditional rules, we equip rules with predicates that restrict sets of matches. A conditional
rule q = (r, P) consists of a rule r and a predicate P on graph morphisms. Given totally labelled graphs
G, H and a match g : L → G for q, we write G ⇒q,g H (or just G ⇒q H) if P(g) holds and G ⇒r,g H. For
a set of conditional rules R, we write G ⇒R H if there is some q in R such that G ⇒q H.
3
Conditional Rule Schemata
A GP program is essentially a list of declarations of conditional rule schemata together with a command
sequence for controlling the application of the schemata. Rule schemata generalise rules in that labels
can contain expressions over parameters of type integer or string. In this section, we give an abstract
2 See
[9] for an equivalent definition by graph pushouts.
Semantics of Graph Programs
30
1
1
1
1
←
→
2
1
↓
2
2
3
1
2
↓
↓
1
1
←
1
→
1
2
3
1
1
Figure 1: A direct derivation
syntax for the textual components of conditional rule schemata and interpret them as sets of conditional
rules.
Figure 2 shows an example for the declaration of a conditional rule schema. It consists of the identifier bridge followed by the declaration of formal parameters, the left and right graphs of the schema
which are labelled with expressions over the parameters, the node identifiers 1, 2, 3 determining the
interface of the schema, and the keyword where followed by the condition.
bridge(a, b, x, y, z : int)
x
1
a
y
2
b
a+b
z
3
⇒
x
a
1
y
2
b
z
3
where a >= 0 and b >= 0 and not edge(1, 3)
Figure 2: A conditional rule schema
In the GP programming system [11], rule schemata are constructed with a graphical editor. Figure
3 gives a grammar in Extended Backus-Naur Form for node and edge labels in the left and right graph
of a rule schema (categories LeftLabel and RightLabel).3 Labels can be sequences of expressions separated by underscores, as will be demonstrated by Example 1 in Section 4. We require that labels in the
left graph must be simple expressions because their values at execution time are determined by graph
matching. All variable identifiers in the right graph must also occur in the left graph. Every expression
in category Exp has type int or string, where arithmetical operators expect arguments of type int and
the type of variable identifiers is determined by their declarations.
The condition of a rule schema is a boolean expression built from expressions of category Exp and
the special predicate edge, see Figure 4. Again, all variable identifiers occurring in the condition must
3 The
grammars in Figure 3 and Figure 4 are ambiguous, we use parentheses to disambiguate expressions where necessary.
D. Plump & S. Steinert
31
LeftLabel
RightLabel
SimpleExp
Exp
ArithOp
Num
String
::=
::=
::=
::=
::=
::=
::=
SimpleExp [’ ’ LeftLabel]
Exp [’ ’ RightLabel]
[’-’] Num | String | VarId
SimpleExp | Exp ArithOp Exp
’+’ | ’-’ | ’∗’ | ’/’
Digit {Digit}
’ ” ’ {Char} ’ ” ’
Figure 3: Syntax of node and edge labels
BoolExp
::=
Node
RelOp
BoolOp
::=
::=
::=
edge ’(’ Node ’,’ Node ’)’ | Exp RelOp Exp
| not BoolExp | BoolExp BoolOp BoolExp
Digit {Digit}
’=’ | ’\=’ | ’>’ | ’<’ | ’>=’ | ’<=’
and | or
Figure 4: Syntax of conditions
also occur in the left graph of the schema. The predicate edge demands the (non-)existence of an
edge between two nodes in the graph to which the rule schema is applied. For example, the expression
not edge(1, 3) in the condition of Figure 2 forbids an edge from node 1 to node 3 when the left graph is
matched.
We interpret a conditional rule schema as the (possibly infinite) set of conditional rules that is obtained by instantiating variables with any values and evaluating expressions. To define this, consider a
declaration D of a conditional rule-schema. Let L and R be the left and right graphs of D, and c the
condition. We write Var(D) for the set of variable identifiers occurring in D. Given x in Var(D), type(x)
denotes the type associated with x. An assignment is a mapping α : Var(D) → (Z ∪ Char∗ ) such that for
each x in Var(D), type(x) = int implies α (x) ∈ Z, and type(x) = string implies α (x) ∈ Char∗ .
Given a label l of category RightLabel occuring in D and an assignment α , the value l α ∈ L is
inductively defined. If l is a numeral or a sequence of characters, then l α is the integer or character string
represented by l (which is independent of α ). If l is a variable identifier, then l α = α (l). Otherwise, l α
is obtained from the values of l’s components. If l has the form e1 ⊕ e2 with ⊕ in ArithOp and e1 , e2 in
Exp, then l α = eα1 ⊕Z eα2 where ⊕Z is the integer operation represented by ⊕.4 If l has the form e m with
e in Exp and m in RightLabel, then l α = eα mα (the concatenation of eα and mα ). Note that our definition
of l α covers all labels in D since LeftLabel is a subcategory of RightLabel.
The value of the condition c in D not only depends on an assignment but also on a graph morphism.
For, if c contains the predicate edge, we need to consider the structure of the graph to which we want to
apply the rule schema. Consider an assignment α and let Lα be obtained from L by replacing each label
l with l α . Let g : Lα → G be a graph morphism with G ∈ G . Then for each Boolean subexpression b of
c, the value bα ,g in B = {tt, ff} is inductively defined. If b has the form e1 ⊲⊳ e2 with ⊲⊳ in RelOp and
e1 , e2 in Exp, then bα ,g = tt if and only if eα1 ⊲⊳Z eα2 where ⊲⊳Z is the relation on integers represented by
4 For
simplicity, we consider division by zero as an implementation-level issue.
Semantics of Graph Programs
32
⊲⊳. If b has the form not b1 with b1 in BoolExp, then bα ,g = tt if and only if b1α ,g = ff. If b has the
form b1 ⊕ b2 with ⊕ in BoolOp and b1 , b2 in BoolExp, then bα ,g = b1α ,g ⊕B b2α ,g where ⊕B is the Boolean
operation on B represented by ⊕. A special case is given if b has the form edge(v, w) where v, w are
identifiers of interface nodes in D. We then have
tt if there is an edge from g(v) to g(w),
α ,g
b =
ff otherwise.
Let now r be the rule-schema identifier associated with declaration D. For every assignment α , let
rα = (Lα ← K → Rα , Pα ) be the conditional rule given as follows:
• Lα and Rα are obtained from L and R by replacing each label l with l α .
• K is the discrete subgraph of L and R determined by the node identifiers for the interface, where
all nodes are unlabelled.
• Pα is defined by: Pα (g) if and only if g is a graph morphism Lα → G such that G ∈ G and
cα ,g = tt.
The interpretation of r is the rule set I(r) = {rα | α is an assignment}. For notational convenience, we
sometimes denote the relation ⇒I(r) by ⇒r . Note that I(r) is a (possibly infinite) set of conditional rules
in the sense of Section 2, grounding rule schemata in the theory of the double-pushout approach with
relabelling [9].
For example, the upper rows of Figure 5 show the rule schema bridge of Figure 2 (without condition) and its instance bridgeα , where α (x) = 0, α (y) = α (z) = 1, α (a) = 3 and α (b) = 2. The
condition c of bridge evaluates to the predicate Pα which is true for a match g of the left-hand graph
if and only if there is no edge from g(1) to g(3). (The subexpressions a >= 0 and b >= 0 evaluate to
tt and hence can be ignored.) The lower rows of Figure 5 show an application of bridgeα by a graph
morphism satisfying Pα .
a+b
Schema:
a
x
1
y
b
z
2
⇒
3
a
x
1
y
b
z
2
↓α
3
↓α
5
Instance:
3
0
1
1
2
1
2
⇒
3
3
0
1
1
2
1
2
↓
3
↓
5
3
0
1
1
2
2
1
0
⇒
3
0
1
1
2
Figure 5: Application of a rule schema using instantiation
2
1
0
D. Plump & S. Steinert
4
33
Graph Programs
We start by discussing an example program for graph colouring.
Example 1 (Computing a 2-colouring). A colouring for a graph is an assignment of colours (integers)
to nodes such that the source and target of each edge have different colours. A graph is 2-colourable
(or bipartite) if it possesses a colouring with at most two colours. The program 2-colouring in Figure 6 generates a 2-colouring for nonempty, connected input graphs without loops if such a colouring
exists—otherwise the input graph is returned. The program consists of five rule-schema declarations, the
macro colour representing the rule-schema set {colour1, colour2}, and the main command sequence
following the key word main.
main = choose; colour!; if illegal then undo!
colour = {colour1, colour2}
choose(x : int)
x
⇒
illegal(a, i, x, y : int)
x0
xi
1
1
1
colour1(a, i, x, y : int)
xi
a
1
y
⇒
2
a
yi
2
⇒
xi
1
a
yi
2
undo(i, x : int)
xi
a
1
y 1−i
2
xi
1
⇒
x
1
colour2(a, i, x, y : int)
xi
1
a
y
2
⇒
xi
1
a
y 1−i
2
Figure 6: The program 2-colouring
Given an integer-labelled input graph, the program first uses the rule schema choose to pick any
node and replace its label x with x 0. The underscore operator allows to add a tag to a label, used
here to add colours to labels. In general, a tagged label consists of a sequence of expressions joined by
underscores. After the first node has been coloured, the command colour! applies the rule schemata
colour1 and colour2 nondeterministically as long as possible to colour all remaining nodes. In each
iteration of the loop, an uncoloured node adjacent to an already coloured node v gets the colour in {0, 1}
that is complementary to v’s colour. If the input graph is connected, the graph resulting from colour!
is correctly coloured if and only if the rule schema illegal is not applicable. The latter is checked
by the if-statement. If illegal is applicable, then the input must contain an undirected cycle of odd
length and hence is not 2-colourable (see for example [10]). In this case the loop undo! removes all tags
to return the input graph unmodified. Note that the number of rule-schema applications performed by
2-colouring is linear in the number of input nodes.
To make 2-colouring applicable to graphs that are possibly empty or disconnected, we can insert
Semantics of Graph Programs
34
a nested loop:
main = (choose; colour!)!; if illegal then undo!.
Now if the input graph is empty, choose fails which causes the outer loop to terminate and return the
current (empty) graph. On the other hand, if the input consists of several connected components, the
body of the outer loop is repeatedly called to colour each component.
Figure 7 shows the abstract syntax of GP programs.5 A program consists of a number of declarations
of conditional rule schemata and macros, and exactly one declaration of a main command sequence. The
rule-schema identifiers (category RuleId) occurring in a call of category RuleSetCall refer to declarations
of conditional rule schemata in category RuleDecl (see Section 3). Semantically, each rule-schema
identifier r stands for the set I(r) of conditional rules induced by that identifier. A call of the form
S
{r1 , . . . , rn } stands for the union ni=1 I(ri ).
Prog
Decl
MacroDecl
MainDecl
ComSeq
Com
::=
::=
::=
::=
::=
::=
RuleSetCall
MacroCall
::=
::=
Decl {Decl}
RuleDecl | MacroDecl | MainDecl
MacroId ’=’ ComSeq
main ’=’ ComSeq
Com {’;’ Com}
RuleSetCall | MacroCall
| if ComSeq then ComSeq [else ComSeq]
| ComSeq ’!’
| skip | fail
RuleId | ’{’ [RuleId {’,’ RuleId}] ’}’
MacroId
Figure 7: Abstract syntax of GP
Macros are a simple means to structure programs and thereby to make them more readable. Every
program can be transformed into an equivalent macro-free program by replacing macro calls with their
associated command sequences (recursive macros are not allowed). In the next section we use the terms
“program” and “command sequence” synonymously, assuming that all macro calls have been replaced.
The commands skip and fail can be expressed through the other commands (see next section),
hence the core of GP includes only the call of a set of conditional rule schemata (RuleSetCall), sequential
composition (’;’), the if-then-else statement and as-long-as-possible iteration (’!’).
5
Semantics of Graph Programs
We present a formal semantics of GP in the style of Plotkin’s structural operational semantics [14]. As
usual for this approach, inference rules inductively define a small-step transition relation → on configurations. In our setting, a configuration is either a command sequence together with a graph, just a graph
or the special element fail:
→ ⊆ (ComSeq × G ) × ((ComSeq × G ) ∪ G ∪ {fail}).
5 Where
necessary we use parentheses to disambiguate programs.
D. Plump & S. Steinert
35
Configurations in ComSeq × G represent unfinished computations, given by a rest program and a state in
the form of a graph, while graphs in G are proper results of computations. In addition, the element fail
represents a failure state. A configuration γ is terminal if there is no configuration δ such that γ → δ .
Each inference rule in Figure 8 consists of a premise and a conclusion separated by a horizontal
bar. Both parts contain meta-variables for command sequences and graphs, where R stands for a call in
category RuleSetCall, C, P, P′ , Q stand for command sequences in category ComSeq and G, H stand for
S
graphs in G . Given a rule-set call R, let I(R) = {I(r) | r is a rule-schema identifier in R} (see Section
3 for the definition of I(r)). The domain of ⇒I(R) , denoted by Dom(⇒I(R) ), is the set of all graphs G in
G such that G ⇒I(R) H for some graph H. Meta-variables are considered to be universally quantified.
For example, the rule [Call1 ] should be read as: “For all R in RuleSetCall and all G, H in G , G ⇒I(R) H
implies hR, Gi → H.”
Figure 8 shows the inference rules for the core constructs of GP. We write →+ and →∗ for the
transitive and reflexive-transitive closures of →. A command sequence C finitely fails on a graph G ∈
G if (1) there does not exist an infinite sequence hC, Gi → hC1 , G1 i → . . . and (2) for each terminal
configuration γ such that hC, Gi →∗ γ , γ = fail. In other words, C finitely fails on G if all computations
starting from (C, G) eventually end in the configuration fail.
[Call1 ]
G ⇒I(R) H
hR, Gi → H
[Call2 ]
G 6∈ Dom(⇒I(R) )
hR, Gi → fail
[Seq1 ]
hP, Gi → hP′ , Hi
hP; Q, Gi → hP′ ; Q, Hi
[Seq2 ]
hP, Gi → H
hP; Q, Gi → hQ, Hi
[Seq3 ]
hP, Gi → fail
hP; Q, Gi → fail
[If1 ]
hC, Gi →+ H
hif C then P else Q, Gi → hP, Gi
[Alap1 ]
hP, Gi →+ H
hP!, Gi → hP!, Hi
[If2 ]
C finitely fails on G
hif C then P else Q, Gi → hQ, Gi
[Alap2 ]
P finitely fails on G
hP!, Gi → G
Figure 8: Inference rules for core commands
The concept of finite failure stems from logic programming where it is used to define negation as
failure [4]. In the case of GP, we use it to define powerful branching and iteration constructs. In particular,
our definition of the if-then-else command allows to “hide” destructive tests.
Example 2 (Recognizing series-parallel graphs). A graph is series-parallel if it reduces to a graph consisting of two nodes and an edge between them by the following two operations [1, 5]: (1) Replace a
pair of parallel edges by an edge from their source to their target. (2) Given a node v with exactly one
incoming edge e1 and exactly one outgoing edge e2 such that the source of e1 and the target of e2 are
distinct, replace e1 , e2 and v by an edge from the source of e1 to the target of e2 .
Suppose that we want to check whether a connected, integer-labelled graph G is series-parallel and,
depending on the result, execute either a program P or a program Q on G. We can do this with the
program
main = if {par, seq}!; base then P else Q
whose rule schemata par, seq and base are shown in Figure 9. The subprogram {par, seq}! applies
Semantics of Graph Programs
36
as long as possible the operations (1) and (2) to the input graph G, then the rule schema base checks if
the resulting graph consists of two nodes connected by an edge. Graph G is series-parallel if and only
if base is applicable to the reduced graph. (Note that {par, seq}! preserves connectedness and that, by
the dangling condition, base is applicable only if the images of its left-hand nodes have degree one.) It
is important to note that by the inference rules [If1 ] and [If2 ], the main program executes P or Q on the
input graph G whereas the graph resulting from the test is discarded.
par(a, b, x, y : int)
a
y
⇒
x
b
1
2
0
x
1
y
2
seq(a, b, x, y, z : int)
x
a
b
y
⇒
z
1
2
x
1
0
z
2
base(a, x, y : int)
x
a
y
⇒
0/
Figure 9: Rule schemata for recognizing series-parallel graphs
The meaning of the remaining GP commands is defined in terms of the meaning of the core commands, see Figure 10. We refer to these commands as derived commands.
[Skip]
hskip, Gi → hr, Gi
where r is an identifier for the rule schema 0/ ⇒ 0/
[[Fail]
hfail, Gi → h{}, Gi
[If3 ]
hif C then P, Gi → hif C then P else skip, Gi
Figure 10: Inference rules for derived commands
We can now summarise the meaning of GP programs by a semantic function J K which assigns to
each program P the function JPK mapping an input graph G to the set of all possible results of running P
on G. The result set may contain, besides proper results in the form of graphs, the special value ⊥ which
indicates a nonterminating or stuck computation. The semantic function J K : ComSeq → (G → 2G ∪{⊥} )
is defined by6
+
JPKG = {H ∈ G | hP, Gi → H} ∪ {⊥ | P can diverge or get stuck from G}
where P can diverge from G if there is an infinite sequence hP, Gi → hP1 , G1 i → hP2 , G2 i → . . . , and P
can get stuck from G if there is a terminal configuration hQ, Hi such that hP, Gi →∗ hQ, Hi.
6 We
write JPKG for the application of JPK to a graph G.
D. Plump & S. Steinert
37
Note that JPKG = 0/ if and only if P finitely fails on G. In Example 2, for instance, we have
J{par, seq}!; baseKG = 0/ for every connected graph G containing a cycle. This is because the graph
resulting from {par, seq}! is still connected and cyclic, so the rule schema base is not applicable.
A program can get stuck only in two situations: either it contains a subprogram if C then P else Q
where C both can diverge from some graph and cannot produce a proper result from that graph, or it contains a subprogram B! where the loop’s body B possesses the said property of C. The evaluation of these
subprograms will get stuck because the inference rules for branching and iteration are not applicable.
6
Conclusion
GP is an experimental rule-based language for high-level problem solving in the domain of graphs,
freeing programmers from handling low-level data structures. The hallmark of GP is syntactic and
semantic simplicity. Conditional rule schemata for graph transformation allow to express application
conditions and computations on labels, in addition to structural changes. The semantics of rule schemata
is orthogonal to the semantics of control constructs, making it easy to change the format of rules or
graphs.
The operational semantics of programs describes the effect of GP’s control constructs in a natural
way and captures the nondeterminism of the language. In particular, powerful branching and iteration
commands have been defined using the concept of finite failure. Destructive tests on the current graph
can be hidden in the condition of the branching command, and nested loops can be coded since arbitrary
subprograms can be iterated as long as possible.
Future extensions of GP may include recursive procedures for writing complex algorithms (see [19]),
and a type concept for restricting the shape of graphs. Our goal is to support formal reasoning on graph
programs by developing static analyses for properties such as termination and confluence (uniqueness of
results), and a calculus and tool support for program verification.
References
[1] Jørgen Bang-Jensen and Gregory Gutin. Digraphs: Theory, Algorithms and Applications. Springer-Verlag,
2000.
[2] Peter Borovanský, Claude Kirchner, Hélène Kirchner, and Pierre-Etienne Moreau. ELAN from a rewriting
logic point of view. Theoretical Computer Science, 285(2):155–185, 2002.
[3] Martin Bravenboer, Arthur van Dam, Karina Olmos, and Eelco Visser. Program transformation with scoped
dynamic rewrite rules. Fundamenta Informaticae, 69(1–2):123–178, 2006.
[4] Keith L. Clark. Negation as failure. In Herve Gallaire and Jack Minker, editors, Logic and Data Bases, pages
293–322. Plenum Press, 1978.
[5] R. J. Duffin. Topology of series-parallel networks. Journal of Mathematical Analysis and Applications,
10:303–318, 1965.
[6] Claudia Ermel, Michael Rudolf, and Gabi Taentzer. The AGG approach: Language and environment. In
H. Ehrig, G. Engels, H.-J. Kreowski, and G. Rozenberg, editors, Handbook of Graph Grammars and Computing by Graph Transformation, volume 2, chapter 14, pages 551–603. World Scientific, 1999.
[7] Rubino Geiß, Gernot Veit Batz, Daniel Grund, Sebastian Hack, and Adam M. Szalkowski. GrGen: A fast
SPO-based graph rewriting tool. In Proc. International Conference on Graph Transformation (ICGT 2006),
volume 4178 of Lecture Notes in Computer Science, pages 383–397. Springer-Verlag, 2006.
38
Semantics of Graph Programs
[8] Annegret Habel and Detlef Plump. Computational completeness of programming languages based on graph
transformation. In Proc. Foundations of Software Science and Computation Structures (FOSSACS 2001),
volume 2030 of Lecture Notes in Computer Science, pages 230–245. Springer-Verlag, 2001.
[9] Annegret Habel and Detlef Plump. Relabelling in graph transformation. In Proc. International Conference
on Graph Transformation (ICGT 2002), volume 2505 of Lecture Notes in Computer Science, pages 135–147.
Springer-Verlag, 2002.
[10] Jon Kleinberg and Éva Tardos. Algorithm Design. Addison Wesley, 2006.
[11] Greg Manning and Detlef Plump. The GP programming system. In Proc. Graph Transformation and Visual
Modelling Techniques (GT-VMT 2008), volume 10 of Electronic Communications of the EASST, 2008.
[12] Ulrich Nickel, Jörg Niere, and Albert Zündorf. The FUJABA environment. In Proc. International Conference
on Software Engineering (ICSE 2000), pages 742–745. ACM Press, 2000.
[13] Hanne Riis Nielson and Flemming Nielson. Semantics with Applications: An Appetizer. Springer-Verlag,
2007.
[14] Gordon D. Plotkin. A structural approach to operational semantics. Journal of Logic and Algebraic Programming, 60–61:17–139, 2004.
[15] Detlef Plump. The graph programming language GP. In Proc. Algebraic Informatics (CAI 2009), volume
5725 of Lecture Notes in Computer Science, pages 99–122. Springer-Verlag, 2009.
[16] Detlef Plump and Sandra Steinert. Towards graph programs for graph algorithms. In Proc. International
Conference on Graph Transformation (ICGT 2004), volume 3256 of Lecture Notes in Computer Science,
pages 128–143. Springer-Verlag, 2004.
[17] Andy Schürr. Operationales Spezifizieren mit programmierten Graphersetzungssystemen. Deutscher Universitäts-Verlag, 1991. In German.
[18] Andy Schürr, Andreas Winter, and Albert Zündorf. The PROGRES approach: Language and environment.
In H. Ehrig, G. Engels, H.-J. Kreowski, and G. Rozenberg, editors, Handbook of Graph Grammars and
Computing by Graph Transformation, volume 2, chapter 13, pages 487–550. World Scientific, 1999.
[19] Sandra Steinert. The Graph Programming Language GP. PhD thesis, The University of York, 2007.
| 6cs.PL
|
Published as a conference paper at ICLR 2018
M EMORY AUGMENTED C ONTROL N ETWORKS
Arbaaz Khan, Clark Zhang, Nikolay Atanasov, Konstantinos Karydis,
Vijay Kumar, Daniel D. Lee
GRASP Laboratory, University of Pennsylvania
arXiv:1709.05706v6 [cs.AI] 14 Feb 2018
A BSTRACT
Planning problems in partially observable environments cannot be solved directly
with convolutional networks and require some form of memory. But, even memory
networks with sophisticated addressing schemes are unable to learn intelligent
reasoning satisfactorily due to the complexity of simultaneously learning to access
memory and plan. To mitigate these challenges we propose the Memory Augmented Control Network (MACN). The network splits planning into a hierarchical
process. At a lower level, it learns to plan in a locally observed space. At a higher
level, it uses a collection of policies computed on locally observed spaces to learn
an optimal plan in the global environment it is operating in. The performance of
the network is evaluated on path planning tasks in environments in the presence of
simple and complex obstacles and in addition, is tested for its ability to generalize
to new environments not seen in the training set.
1
I NTRODUCTION
A planning task in a partially observable environment involves two steps: inferring the environment
structure from local observation and acting based on the current environment estimate. In the past,
such perception-action loops have been learned using supervised learning with deep networks as
well as deep reinforcement learning (Daftry et al., 2016), (Chebotar et al., 2016), (Lee et al., 2017).
Popular approaches in this spirit are often end-to-end (i.e. mapping sensor readings directly to motion
commands) and manage to solve problems in which the underlying dynamics of the environment or
the agent are too complex to model. Approaches to learn end-to-end perception-action loops have
been extended to complex reinforcement learning tasks such as learning how to play Atari games
(Mnih et al., 2013a), as well as to imitation learning tasks like controlling a robot arm (Levine et al.,
2015).
Purely convolutional architectures (CNNs) perform poorly when applied to planning problems due
to the reactive nature of the policies learned by them (Zhang et al., 2016b), (Giusti et al., 2016).
The complexity of this problem is compounded when the environment is only partially observable as
is the case with most real world tasks. In planning problems, when using a function approximator
such as a convolutional neural network, the optimal actions are dependent on an internal state. If one
wishes to use a state-less network (such as a CNN) to obtain the optimal action, the input for the
network should be the whole history of observations and actions. Since this does not scale well, we
need a network that has an internal state such as a recurrent neural network or a memory network.
(Zhang et al., 2016a) showed that when learning how to plan in partially observable environments,
it becomes necessary to use memory to retain information about states visited in the past. Using
recurrent networks to store past information and learn optimal control has been explored before
in (Levine, 2013). While (Siegelmann & Sontag, 1995) have shown that recurrent networks are
Turing complete and are hence capable of generating any arbitrary sequence in theory, this does not
always translate into practice. Recent advances in memory augmented networks have shown that it
is beneficial to use external memory with read and write operators that can be learned by a neural
network over recurrent neural networks (Graves et al., 2014), (Graves et al., 2016). Specifically,
we are interested in the Differentiable Neural Computer (DNC) (Graves et al., 2016) which uses
an external memory and a network controller to learn how to read, write and access locations in
the external memory. The DNC is structured such that computation and memory operations are
separated from each other. Such a memory network can in principle be plugged into the convolutional
1
Published as a conference paper at ICLR 2018
architectures described above, and be trained end to end since the read and write operations are
differentiable. However, as we show in our work, directly using such a memory scheme with CNNs
performs poorly for partially observable planning problems and also does not generalize well to new
environments.
To address the aforementioned challenges we propose the Memory Augmented Control Network
(MACN), a novel architecture specifically designed to learn how to plan in partially observable
environments under sparse rewards.1 Environments with sparse rewards are harder to navigate since
there is no immediate feedback. The intuition behind this architecture is that planning problem can
be split into two levels of hierarchy. At a lower level, a planning module computes optimal policies
using a feature rich representation of the locally observed environment. This local policy along with
a sparse feature representation of the partially observed environment is part of the optimal solution
in the global environment. Thus, the key to our approach is using a planning module to output
a local policy which is used to augment the neural memory to produce an optimal policy for the
global environment. Our work builds on the idea of introducing options for planning and knowledge
representation while learning control policies in MDPs (Sutton et al., 1999). The ability of the
proposed model is evaluated by its ability to learn policies (continuous and discrete) when trained in
environments with the presence of simple and complex obstacles. Further, the model is evaluated on
its ability to generalize to environments and situations not seen in the training set.
The key contributions of this paper are:
1. A new network architecture that uses a differentiable memory scheme to maintain an estimate
of the environment geometry and a hierarchical planning scheme to learn how to plan paths
to the goal.
2. Experimentation to analyze the ability of the architecture to learn how to plan and generalize
in environments with high dimensional state and action spaces.
2
M ETHODOLOGY
Section 2.1 outlines notation and formally states the problem considered in this paper. Section 2.2 and
2.3 briefly cover the theory behind value iteration networks and memory augmented networks. Finally,
in section 2.4 the intuition and the computation graph is explained for the practical implementation
of the model.
2.1
P RELIMINARIES
Consider an agent with state st ∈ S at discrete time t. Let the states S be a discrete set [s1 , s2 , . . . , sn ].
For a given action at ∈ A, the agent evolves according to known deterministic dynamics: st+1 =
f (st , at ). The agent operates in an unknown environment and must remain safe by avoiding collisions.
Let m ∈ {−1, 0}n be a hidden labeling of the states into free (0) and occupied (−1). The agent has
access to a sensor that reveals the labeling of nearby states through an observations zt = H(st )m ∈
{−1, 0}n , where H(s) ∈ Rn×n captures the local field of view of the agent at state s. The local
observation consists of ones for observable states and zeros for unobservable states. The observation
zt contains zeros for unobservable states. Note that m and zt are n × 1 vectors and can be indexed
by the state st . The agent’s task is to reach a goal region S goal ⊂ S, which is assumed obstacle-free,
i.e., m[s] = 0 for all s ∈ S goal . The information available to the agent at time t to compute its action
at is ht := s0:t , z0:t , a0:t−1 , S goal ∈ H, where H is the set of possible sequences of observations,
states, and actions. Our problem can then be stated as follows :
Problem 1. Given an initial state s0 ∈ S with m[s0 ] = 0 (obstacle-free) and a goal region S goal ,
find a function µ : S → A such that applying the actions at := µ(st ) results in a sequence of states
s0 , s1 , . . . , sT satisfying sT ∈ S goal and m[st ] = 0 for all t = 0, . . . , T .
Instead of trying to estimate the hidden labeling m using a mapping approach, our goal is to learn a
policy µ that maps the sequence of sensor observations z0 , z1 , . . . zT directly to actions for the agent.
The partial observability requires an explicit consideration of memory in order to learn µ successfully.
A partially observable problem can be represented via a Markov Decision Process (MDP) over
1
In this work an agent receives a reward only when it has reached the goal prescribed by the planning task.
2
Published as a conference paper at ICLR 2018
the history space H. More precisely, we consider a finite-horizon discounted MDP defined by
M(H, A, T , r, γ), where γ ∈ (0, 1] is a discount factor, T : H × A → H is a deterministic transition
function, and r : H → R is the reward function, defined as follows:
T (ht , at ) = (ht , st+1 = f (st , at ), zt+1 = H(st+1 )m, at )
r(ht , at ) = zt [st ]
The reward function definition stipulates that the reward of a state s can be measured only after its
occupancy state has been observed.
P
Given observations z0:t , we can obtain an estimate m̂ = max{ τ zτ , −1} of the map of the environment and use it to formulate a locally valid, fully-observable problem as the MDP Mt (S, A, f, r, γ)
with transition function given by the agent dynamics f and reward r(st ) := m̂[st ] given by the map
estimate m̂.
2.2
VALUE I TERATION N ETWORKS
The typical algorithm to solve an MDP is Value Iteration (VI) (Sutton & Barto, 1998). The value of a
state (i.e. the expected reward over the time horizon if an optimal policy is followed) is computed
iteratively by calculating an action value function Q(s, a) for each state. The value for state s can
then be calculated by V (s) := maxa Q(s, a). By iterating multiple times over all states and all
actions possible in each state, we can get a policy π = arg maxa Q(s, a). Given a transition function
Tr (s0 |s, a), the update rule for value iteration is given by (1)
X
Vk+1 (s) = max[r(s, a) + γ
Tr (s0 |s, a)Vk (s)] .
(1)
a
s0
A key aspect of our network is the inclusion of this network component that can approximate this
Value Iteration algorithm. To this end we use the VI module in Value Iteration Networks (VIN) (Tamar
et al., 2016). Their insight is that value iteration can be approximated by a convolutional network
with max pooling. The standard form for windowed convolution is
V (x) =
x+w
X
V (k)u(k) .
(2)
k=x−w
P
(Tamar et al., 2016) show that the summation in (2) is analogous to s0 T (s0 |s, a)Vk (s) in (1).
When (2) is stacked with reward, max pooled and repeated K times, the convolutional architecture
can be used to represent an approximation of the value iteration algorithm over K iterations.
2.3
E XTERNAL MEMORY N ETWORKS
Recent works on deep learning employ neural networks with external memory (Graves et al.,
2014), (Graves et al., 2016), (Kurach et al., 2015), (Parisotto & Salakhutdinov, 2017). Contrary to
earlier works that explored the idea of the network learning how to read and access externally fixed
memories, these recent works focus on learning to read and write to external memories, and thus
forgo the task of designing what to store in the external memory. We are specifically interested in the
DNC (Graves et al., 2016) architecture. This is similar to the work introduced by (Oh et al., 2016)
and (Chen et al., 2017). The external memory uses differentiable attention mechanisms to determine
the degree to which each location in the external memory M is involved in a read or write operation.
The DNC makes use of a controller (a recurrent neural network such as LSTM) to learn to read and
write to the memory matrix. A brief overview of the read and write operations follows.2
2.3.1
R EAD AND W RITE O PERATION
The read operation is defined as a weighted average over the contents of the locations in the memory.
This produces a set of vectors defined as the read vectors. R read weightings {wtread,1 , . . . , wtread,R }
are used to compute weighted averages of the contents of the locations in the memory. At time t the
read vectors {re1t , . . . , reR
t } are defined as :
reit = Mt> wtread,i
2
We refer the interested reader to the original paper (Graves et al., 2016) for a complete description.
3
(3)
Published as a conference paper at ICLR 2018
where wtread,i are the read weightings, ret is the read vector, and Mt is the state of the memory at
time t. These read vectors are appended to the controller input at the next time step which provides it
access to the memory. The write operation consists of a write weight wtW , an erase vector et and
a write vector vt . The write vector and the erase vector are emitted by the controller. These three
components modify the memory at time t as :
W >
Mt = Mt−1 (1 − wtW e>
t ) + wt vt .
(4)
Memory addressing is defined separately for writing and reading. A combination of content-based
addressing and dynamic memory allocation determines memory write locations, while a combination
of content-based addressing and temporal memory linkage is used to determine read locations.
2.4
M EMORY AUGMENTED C ONTROL M ODEL
(a) 2D Grid World
(b) Sensor observation.
(c) Policy for sensor observation.
Figure 1: 2D Environment a) Let the agent (blue square) operate in a 2D environment. The goal
region is represented by the red square and the orange square represents the agents observation b)
Agents observation. The gray area is not observable. c) It is possible to plan on this locally observed
space since it is a MDP.
Consider the 2D grid world in Fig 1a. The agent is spawned randomly in this world and is represented
by the blue square. The goal of the agent is to learn how to navigate to the red goal region. Let
this environment in Fig 1a be represented by a MDP M. The key intuition behind designing this
architecture is that planning in M can be decomposed into two levels. At a lower level, planning
is done in a local space within the boundaries of our locally observed environment space. Let this
locally observed space be z 0 . Fig 1b represents this locally observed space. As stated before in
Section 2.1, this observation can be formulated as a fully observable problem Mt (S, A, f, r, γ). It
is possible to plan in Mt and calculate the optimal policy for this local space, πl∗ independent of
previous observations (Fig 1c). It is then possible to use any planning algorithm to calculate the
optimal value function Vl∗ from the optimal policy πl∗ in z 0 . Let Π = [πl1 , πl2 , πl3 , πl4 , . . . , πln ] be the
list of optimal policies calculated from such consecutive observation spaces [z0 , z1 , . . . zT ]. Given
these two lists, it is possible to train a convolutional neural network with supervised learning.The
network could then be used to compute a policy πlnew when a new observation z new is recorded.
This policy learned by the convolutional network is purely reactive as it is computed for the z new
observation independent of the previous observations. Such an approach fails when there are local
minima in the environment. In a 2D/3D world, these local minima could be long narrow tunnels
culminating in dead ends (see Fig 2). In the scenario where the environment is populated with
tunnels, (Fig 2) the environment is only partially observable and the agent has no prior knowledge
about the structure of this tunnel forcing it to explore the tunnel all the way to the end. Further,
when entering and exiting such a structure, the agent’s observations are the same, i.e z1 = z2 , but
the optimal actions under the policies πl1 and πl2 (computed by the convolutional network) at these
time steps are not the same, i.e aπ1 6= aπ2 . To backtrack successfully from these tunnels/nodes,
information about previously visited states is required, necessitating memory.
To solve this problem, we propose using a differentiable memory to estimate the map of the environment m̂. The controller in the memory network learns to selectively read and write information
4
Published as a conference paper at ICLR 2018
Figure 2: Environment with local minima. The agents observation when entering the tunnel to
explore it and when backtracking after seeing the dead end are the same. Using a reactive policy for
such environments leads to the agent getting stuck near the dead end .
to the memory bank. When such a differentiable memory scheme is trained it is seen that it keeps
track of important events/landmarks (in the case of tunnel, this is the observation that the dead end
has been reached) in its memory state and discards redundant information. In theory one can use a
CNN to extract features from the observation z 0 and pass these features to the differentiable memory.
Instead, we propose the use of a VI module (Tamar et al., 2016) that approximates the value iteration
algorithm within the framework of a neural network to learn value maps from the local information.
We hypothesize that using these value maps in the differential memory scheme provides us with
better planning as compared to when only using features extracted from a CNN. This architecture is
shown in Figure 3.
The VI module is setup to learn how to plan on the local observations z. The local value maps (which
can be used to calculate local policies) are concatenated with a low level feature representation of the
environment and sent to a controller network. The controller network interfaces with the memory
through an access module (another network layer) and emits read heads, write heads and access
heads. In addition, the controller network also performs its own computation for planning. The output
from the controller network and the access module are concatenated and sent through a linear layer
to produce an action. This entire architecture is then trained end to end. Thus, to summarize, the
planning problem is solved by decomposing it into a two level problem. At a lower level a feature
rich representation of the environment (obtained from the current observation) is used to generate
local policies. At the next level, a representation of the histories that is learned and stored in the
memory, and a sparse feature representation of the currently observed environment is used to generate
a policy optimal in the global environment.
Computation Graph: To explain the computation graph, consider the case of a 2D grid world with
randomly placed obstacles, a start region and a goal region as shown in Fig 1a. The actions for
this grid world are considered to be discrete. The 2D grid world is presented in the form of an
image I of size m × n to the network. Let the goal region be [mgoal , ngoal ] and the start position be
[mstart , nstart ]. At any given instant, only a small part of I is observed by the network and the rest
of the image I is blacked out. This corresponds to the agent only observing what is visible within the
range of its sensor. In addition to this the image is stacked with a reward map Rm as explained in
(Tamar et al., 2016). The reward map consists of an array of size m × n where all elements of the
array except the one corresponding to index [mgoal , ngoal ] are zero. Array element corresponding
to [mgoal , ngoal ] is set to a high value(in our experiments it is set to 1) denoting reward. The input
image of dimension [m × n × 2] is first convolved with a kernel of size (3 × 3), 150 channels and
stride of 1 everywhere. This is then convolved again with a kernel of size (1,1), 4 channels and stride
of 1. Let this be the reward layer R. R is convolved with another filter of size (3,3) with 4 channels.
This is the initial estimate of the action value function or Q(s, a). The initial value of the state V (s)
is also calculated by taking max over Q(s, a). The operations up to this point are summarized by the
"Conv" block in Figure 3. Once these initial values have been computed, the model executes a for
loop k times (the value of k ranges based on the task). Inside the for loop at every iteration, the R and
V are first concatenated. This is then convolved with another filter of size (3,3) and 4 channels to get
the updated action value of the state, Q(s, a). We find the value of the state V(s) by taking the max
5
Published as a conference paper at ICLR 2018
Figure 3: MACN Architecture. The architecture proposed uses convolutional layers to extract
features from the environment. The value maps are generated with these features. The controller
network uses the value maps and low level features to emit read and write heads in addition to doing
its own planning computation.
of the action value function. The values of the kernel sizes are constant across all three experiments.
The updated value maps are then fed into a DNC controller. The DNC controller is a LSTM (hidden
units vary according to task) that has access to an external memory. The external memory has 32
slots with word size 8 and we use 4 read heads and 1 write head. This varies from task to task since
some of the more complex environments need more memory. The output from the DNC controller
and the memory is concatenated through a linear layer to get prediction for the action that the agent
should execute. The optimizer used is the RMSProp and we use a learning rate of 0.0001 for our
experiments.
This formulation is easy enough to be extended to environments where the state space is larger than
two dimensions and the action space is larger. We demonstrate this in our experiments.
3
E XPERIMENTS
To investigate the performance of MACN, we design our experiments to answer three key questions:
• Can it learn how to plan in partially observable environments with sparse rewards?
• How well does it generalize to new unknown environments?
• Can it be extended to other domains?
We first demonstrate that MACN can learn how to plan in a 2D grid world environment. Without
loss of generality, we set the probability of all actions equal. The action space is discrete, A :=
{down, right, up, lef t}. This can be easily extended to continuous domains since our networks
output is a probability over actions. We show this in experiment 3.4. We then demonstrate that our
network can learn how to plan even when the states are not constrained to a two dimensional space
and the action space is larger than four actions.
3.1
NAVIGATION IN PRESENCE OF SIMPLE OBSTACLES
We first evaluate the ability of our network to successfully navigate a 2D grid world populated with
obstacles at random positions. We make the task harder by having random start and goal positions.
The full map shown in Fig. 4 is the top down view of the entire environment. The input to the
network is the sensor map, where the area that lies outside the agents sensing abilities is grayed out as
explained before. VIN: With just the VI module and no memory in place, we test the performance of
the value iteration network on this 2D partially observable environment. CNN + Memory: We setup
6
Published as a conference paper at ICLR 2018
Figure 4: Performance on grid world environment . In the map the blue square represents the start
position while the red square represents the goal. The red dotted line is the ground truth and the blue
dash line represents the agents path. The maps shown here are from the test set and have not been
seen before by the agent during training.
Model
VIN
CNN + Memory
MACN (LSTM)
MACN
Performance
Success(%)
Test Error
Success(%)
Test Error
Success (%)
Test Error
Success(%)
Test Error
16 × 16
0
0.63
0.12
0.43
88.12
0.077
96.3
0.02
32 × 32
0
0.78
0
0.618
73.4
0.12
85.91
0.08
64 × 64
0
0.81
0
0.73
64
0.21
78.44
0.13
Table 1: Performance on 2D grid world with simple obstacles: All models are tested on maps
generated via the same random process, and were not present in the training set. Episodes over 40
(for a 16 × 16 wide map), 60 (for 32 × 32) and 80 (for 64 × 64) time steps were terminated and
counted as a failure. Episodes where the agent collided with an obstacle were also counted as failures.
a CNN architecture where the sensor image with the reward map is forward propagated through four
convolutional layers to extract features. We test if these features alone are enough for the memory
to navigate the 2D grid world. A natural question to ask at this point is can we achieve planning in
partially observable environments with just a planning module and a simple recurrent neural network
such as a LSTM. To answer this we also test MACN with a LSTM in place of the memory scheme.
We present our results in Table 1. These results are obtained from testing on a held out test-set
consisting of maps with random start, goal and obstacle positions.
Our results show that MACN can learn how to navigate partially observable 2D unknown environments. Note that the VIN does not work by itself since it has no memory to help it remember past
actions. We would also like to point out that while the CNN + Memory architecture is similar to (Oh
et al., 2016), its performance in our experiments is very poor due to the sparse rewards structure.
MACN significantly outperforms all other architectures. Furthermore, MACN’s drop in testing
accuracy as the grid world scales up is not as large compared to the other architectures. While these
results seem promising, in the next section we extend the experiment to determine whether MACN
actually learns how to plan or it is overfitting.
3.2
NAVIGATION IN PRESENCE OF LOCAL MINIMA
The previous experiment shows that MACN can learn to plan in 2D partially observable environments.
While the claim that the network can plan on environments it has not seen before stands, this is
weak evidence in support of the generalizability of the network. In our previous experiment the test
environments have the same dimensions as in the training set, the number of combinations of random
obstacles especially in the smaller environments is not very high and during testing some of the
wrong actions can still carry the agent to the goal. Thus, our network could be overfitting and may
not generalize to new environments. In the following experiment we test our proposed network’s
capacity to generalize.
7
Published as a conference paper at ICLR 2018
Figure 5: Grid world environment with local minima. Left: In the full map the blue square
represents the current position while the red square represents the goal. Center-left: The partial map
represents a stitched together version of all states explored by the agent. Since the agent does not
know if the tunnel culminates in a dead end, it must explore it all the way to the end. Center-right:
The sensor input is the information available to the agent. Right: The full map that we test our agent
on after being trained on smaller maps. The dimensions of the map as well as the tunnel are larger.
The environment is setup with tunnels. The agent starts off at random positions inside the tunnel.
While the orientation of the tunnel is fixed, its position is not. To comment on the the ability of our
network to generalize to new environments with the same task, we look to answer the following
question: When trained to reach the goal on tunnels of a fixed length, can the network generalize to
longer tunnels in bigger maps not seen in the training set?
The network is set up the same way as before. The task here highlights the significance of using
memory in a planning network. The agent’s observations when exploring the tunnel and exiting the
tunnel are the same but the actions mapped to these observations are different. The memory in our
network remembers past information and previously executed policies in those states, to output the
right action. We report our results in Table 2. To show that traditional deep reinforcement learning
performs poorly on this task, we implement the DQN architecture as introduced in (Mnih et al.,
2013b). We observe that even after one million iterations, the DQN does not converge to the optimal
policy on the training set. This can be attributed to the sparse reward structure of the environment.
We report similar findings when tested with A3C as introduced in (Mnih et al., 2016). We also
observe that the CNN + memory scheme learns to turn around at a fixed length and does not explore
the longer tunnels in the test set all the way to the end.
Model
DQN
A3C
CNN + Memory
VIN
MACN
Success (%)
0
0
12
0
100
Maximum generalization length
0
0
20
0
330
Table 2: Performance on grid world with local minima: All models are trained on tunnels of
length 20 units. The success percentages represent the number of times the robot reaches the goal
position in the test set after exploring the tunnel all the way to the end. Maximum generalization
length is the length of the longest tunnel that the robot is able to successfully navigate after being
trained on tunnels of length 20 units.
These results offer insight into the ability of MACN to generalize to new environments. Our network
is found capable of planning in environments it has not seen in the training set at all. On visualizing
the memory (see supplemental material), we observe that there is a big shift in the memory states
only when the agent sees the end of the wall and when the agent exits the tunnel. A t-sne (Maaten &
Hinton, 2008) visualization over the action spaces (see Fig. 6) clearly shows that the output of our
network is separable. We can conclude from this that the network has learned the spatial structure of
the tunnel, and it is now able to generalize to tunnels of longer length in larger maps.
8
Published as a conference paper at ICLR 2018
Thus, we can claim that our proposed model is generalizable to new environments that are structurally
similar to the environments seen in the training set but have not been trained on. In addition to
this in all our experiments are state and action spaces have been constrained to a small number of
dimensions. In our next experiment we show that MACN can learn how to plan even when the state
space and action space are scaled up.
(a) Raw Image Data
(b) Output from VI module
(c) Output from MACN
Figure 6: T-sne visualization on 2D grid worlds with tunnels. a) T-sne visualization of the raw
images fed into the network. Most of the input images for going into the tunnel and exiting the
tunnel are the same but have different action labels. b) T-sne visualization from the outputs of the pre
planning module. While it has done some separation, it is still not completely separable. c) Final
output from the MACN. The actions are now clearly separable.
3.3
G ENERAL G RAPH S EARCH
In our earlier experiments, the state space was constrained in two
dimensions, and only four actions were available. It is nearly impossible to constrain every real world task to a two dimensional space
with only four actions. However, it is easier to formulate a lot of
partially observable planning problems as a graph.We define our
environment as an undirected graph G = (V, E) where the connections between the nodes are generated randomly (see Fig. 7). In Fig
7 the blue node is the start state and the red node is the goal state.
Each node represents a possible state the agent could be in. The
agent can only observe all edges connected to the node it currently
is in thus making it partially observable. The action space for this
state is then any of the possible nodes that the agent can visit next.
As before, the agent only gets a reward when it reaches the goal. We
also add in random start and goal positions. In addition, we add a Figure 7: 9 Node Graph
transition probability of 0.8. (For training details and generation of Search. Blue is start and Red
graph see Appendix.) We present our results in Table 3. On graphs is goal.
with small number of nodes, the reinforcement learning with DQN
and A3C sometimes converge to the optimal goal due to the small
state size and random actions leading to the goal node in some of the cases. However, as before
the MACN outperforms all other models. On map sizes larger than 36 nodes, performance of our
network starts to degrade. Further, we observe that even though the agent outputs a wrong action
at some times, it still manages to get to the goal in a reasonably small number of attempts. From
these results, we can conclude that MACN can learn to plan in more general problems where the state
space is not limited to two dimensions and the action space is not limited to four actions.
3.4
C ONTINUOUS CONTROL DOMAIN
Learning how to navigate in unknown environments, where only some part of the environment
is observable is a problem highly relevant in robotics. Traditional robotics solve this problem by
creating and storing a representation of the entire environment. However, this can quickly get memory
intensive. In this experiment we extend MACN to a SE2 robot. The SE2 notation implies that the
9
Published as a conference paper at ICLR 2018
Model
VIN
A3C
DQN
CNN + Memory
MACN (LSTM)
MACN
9 Nodes
0.57, 23.39
NA, 10
NA, 12
0.25, 81.5
0.14, 98
0.1, 100
Test Error, Success(%)
16 Nodes
25 Nodes
0.61, 14
0.68, 0
NA, 7
NA, 0
NA, 5.2
NA, 0
0.32, 63
0.56, 19
0.19, 96.27 0.26, 84.33
0.18, 100
0.22, 95.5
36 Nodes
0.71, 0
NA, 0
NA,0
0.68, 9.7
0.29, 78
0.28, 89.4
Table 3: Performance on General Graph Search. Test error is not applicable for the reinforcement
learning models A3C and DQN
(b) Laser Scan
(a) Robot Environment
(c) Top Down View of Environment
Figure 8: Navigation in a 3D environment on a continuous control robot. a) The robot is spawned
in a 3d simulated environment. b) Only a small portion of the entire map is visible at any given point
to the robot c) The green line denotes ground truth and red line indicates the output of MACN.
robot is capable of translating in the x-y plane and has orientation. The robot has a differential drive
controller that outputs continuous control values. The robot is spawned in the environment shown in
Fig (8a). As before, the robot only sees a small part of the environment at any given time. In this case
the robot has a laser scanner that is used to perceive the environment.
It is easy to convert this environment to a 2D framework that the MACN needs. We fix the size
of the environment to a m × n grid. This translates to a m × n matrix that is fed into the MACN.
The parts of the map that lie within the range of the laser scanner are converted to obstacle free and
obstacle occupied regions and added to the matrix. Lastly, an additional reward map denoting a high
value for the goal location and zero elsewhere as explained before is appended to the matrix and fed
into the MACN. The network output is used to generate way points that are sent to the underlying
controller. The training set is generated by randomizing the spawn and goal locations and using a
suitable heuristic. The performance is tested on a held out test set of start and goal locations. More
experimental details are outlined in the appendix.
10
Published as a conference paper at ICLR 2018
We observe in Table 4 that the proposed architecture
is able to find its way to the goal a large number
of times and its trajectory is close to the ground
truth. This task is more complex than the grid world
navigation due to the addition of orientation. The
lack of explicit planning in the CNN + Memory
architecture hampers its ability to get to the goal
Table 4: Performance on robot world
in this task. In addition to this, as observed before
deep reinforcement learning is unable to converge
to the goal. We also report some additional results in Fig 9. In Fig 9a we show that MACN converges
faster to the goal than other baselines.
Model
DQN,A3C
VIN
CNN + Memory
MACN
Success (%)
0
57.60
59.74
71.3
In addition to rate of convergence, one of the biggest advantages of MACN over other architectures,
for a fixed memory size is its ability to scale up when the size of the environment increases. We show
that MACN is able to beat other baselines when scaling up the environment. In this scenario, scaling
up refers to placing the goal further away from the start position. While the success percentage
gradually drops to a low value, it is observed that when the memory is increased accordingly, the
success percentage increases again. Lastly, in Fig 10 we observe that in the robot world, the
performance of MACN scales up to goal positions further away by adjusting the size of the external
memory in the differentiable block accordingly.
(a) Dist. to goal vs number of steps
(b) Success % vs goal distance
Figure 9: Performance on simulated environment. a) We report a plot of the number of steps left
to the goal as the agent executes the learned policy in the environment (Lower is better). In this plot,
the agent always starts at a position 40 steps away from the goal. b) The biggest advantage of MACN
over other architectures is its ability to scale. We observe that as the distance to the goal increases,
MACN still beats other baselines at computing a trajectory to the goal.(Higher success % is better)
Figure 10: Effect of Memory in Robot World MACN scales well to larger environments in the
robot world when memory is increased suitably.
11
Published as a conference paper at ICLR 2018
3.5
C OMPARISON WITH ANALYTICAL MOTION PLANNING BASELINES
In this section, we analyze the performance of the proposed network against traditional motion
planning baselines. As stated before, for the grid world environments and the tunnel task, we obtain
expert trajectories by running A∗ on the environment. In the case of the continuous control domain,
we use the the Human Friendly Navigation (HFN) paradigm Guzzi et al. (2013) which uses a variant
of A∗ along with a constraint for safe distances from obstacles to plan paths from start location to
goal location. For the grid worlds (both with simple obstacles and local minima), we compute the
ratio of path length predicted by network architecture to the path length computed by A∗ . Our results
are presented in Table 5.
The VIN alone is unable to reach the goal in a fixed number of steps. This behavior is consistent
across all grid worlds. In the case of the tunnels, the VIN gets stuck inside the local minima and is
unable to navigate to the goal. Thus, the ratio of path length produced by VIN to the path length
produced by A∗ is infinite. In the case of the CNN+Memory, the network is able to navigate to the
goal only when the grid world is small enough. In the case of the tunnels, the CNN+Memory learns
to turn around at a fixed distance instead of exploring the tunnel all the way to the end. For example,
when trained on tunnels of length 20 units and tested on tunnels of length 32 and 64 units, the
CNN+Memory turns around after it has traversed 20 units in the tunnel. For this task, to demonstrate
the ineffectiveness of the CNN+Memory model, we placed the goal just inside the tunnel at the dead
end. Thus, the ratio of path length produced by CNN+Memory to A∗ is ∞ since the agent never
explored the tunnel all the way to the end. For the case of the MACN, we observe performance
close to A∗ for the small worlds. The performance gets worse when the size of the grid world is
increased. However, the dropoff for MACN with the DNC is lesser than that of the MACN with
LSTM. For the tunnel world environment, both network architectures are successfully able to emulate
the performance of A∗ and explore the tunnel all the way to the end.
It is important to note here that A∗ is a model based approach and requires complete knowledge
of the cost and other parameters such as the dynamics of the agent (transition probabilities). In
addition, planners like A∗ require the user to explicitly construct a map as input, while MACN learns
to construct a map to plan on which leads to more compact representations that only includes vital
parts of the map (like the end of the tunnel in the grid world case). Our proposed method is a model
free approach that learns
1. A dynamics model of the agent,
2. A compact map representation,
3. How to plan on the learned model and map.
This model free paradigm also allows us to move to different environments with a previously trained
policy and be able to perform well by fine-tuning it to learn new features.
Model
VIN
CNN + Memory
MACN (LSTM)
MACN
G(16 × 16)
∞
1.43
1.2
1.07
A∗ Ratio
G(32 × 32) G(64 × 64)
∞
∞
2.86
∞
1.4
1.62
1.11
1.47
T(L =32)
∞
∞
1.0
1.0
T(L = 64)
∞
∞
1.0
1.0
Table 5: Comparison to A∗ . G corresponds to grid world with simple obstacles with the size of the
world specified inside the parenthesis. L corresponds to grid worlds with local minima/tunnels with
the length of the tunnel specified inside the parenthesis. All ratios are computed during testing. For
the worlds with tunnels, the network is trained on tunnels of length 20 units.
4
R ELATED W ORK
Using value iteration networks augmented with memory has been explored before in (Gupta et al.,
2017). In their work a planning module together with a map representation of a robot’s free space is
12
Published as a conference paper at ICLR 2018
used for navigation in a partially observable environment using image scans. The image scans are
projected into a 2D grid world by approximating all possible robot poses. This projection is also
learned by the model. This is in contrast to our work here in which we design a general memory
based network that can be used for any partially observed planning problem. An additional difference
between our work and that of (Gupta et al., 2017) is that we do not attempt to build a 2D map of the
environment as this hampers the ability of the network to be applied to environments that cannot be
projected into such a 2D environment. We instead focusing on learning a belief over the environment
and storing this belief in the differentiable memory. Another similar work is that of (Oh et al., 2016)
where a network is designed to play Minecraft. The game environment is projected into a 2D grid
world and the agent is trained by RL to navigate to the goal. That network architecture uses a CNN to
extract high level features followed by a differentiable memory scheme. This is in contrast to our
paper where we approach this planning by splitting the problem into local and global planning. Using
differential network schemes with CNNs for feature extraction has also been explored in (Chen et al.,
2017). Lastly, a recently released paper Neural SLAM (Zhang et al., 2017) uses the soft attention
based addressing in DNC to mimic subroutines of simultaneous localization and mapping. This
approach helps in exploring the environment robustly when compared to other traditional methods.
A possible extension of our work presented here, is to use this modified memory scheme with the
differentiable planner to learn optimal paths in addition to efficient exploration. We leave this for
future work.
5
C ONCLUSION
Planning in environments that are partially observable and have sparse rewards with deep learning has
not received a lot of attention. Also, the ability of policies learned with deep RL to generalize to new
environments is often not investigated. In this work we take a step toward designing architectures
that compute optimal policies even when the rewards are sparse, and thoroughly investigate the
generalization power of the learned policy. In addition we show our network is able to scale well to
large dimensional spaces.
The grid world experiments offer conclusive evidence about the ability of our network to learn how
to plan in such environments. We address the concern of oversimplifying our environment to a 2D
grid world by experimenting with planning in a graph with no constraint on the state space or the
action space. We also show our model is capable of learning how to plan under continuous control.
In the future, we intend to extend our policies trained in simulation to a real world platform such
as a robot learning to plan in partially observable environments. Additionally, in our work we use
simple perfect sensors and do not take into account sensor effects such as occlusion, noise which
could aversely affect performance of the agent. This need for perfect labeling is currently a limitation
of our work and as such cannot be applied directly to a scenario where a sensor cannot provide direct
information about nearby states such as a RGB camera. We intend to explore this problem space in
the future, where one might have to learn sensor models in addition to learning how to plan.
ACKNOWLEDGEMENT
We gratefully acknowledge the support of ARL grants W911NF-08-2-0004 and W911NF-10-2-0016,
ARO grant W911NF-13-1-0350, ONR grants N00014-07-1-0829, N00014-14-1-0510, DARPA grant
HR001151626/HR0011516850 and DARPA HR0011-15-C-0100, USDA grant 2015-67021-23857
NSF grants IIS-1138847, IIS-1426840, CNS-1446592, CNS-1521617, and IIS-1328805. Clark Zhang
is supported by the National Science Foundation Graduate Research Fellowship Program under Grant
No. DGE-1321851. The authors would like to thank Anand Rajaraman, Steven Chen, Jnaneswar Das,
Ekaterina Tolstoya and others at the GRASP lab for interesting discussions related to this paper.
R EFERENCES
Yoshua Bengio, Jérôme Louradour, Ronan Collobert, and Jason Weston. Curriculum learning. In
Proceedings of the 26th annual international conference on machine learning, pp. 41–48. ACM,
2009.
13
Published as a conference paper at ICLR 2018
Yevgen Chebotar, Karol Hausman, Zhe Su, Gaurav S Sukhatme, and Stefan Schaal. Self-supervised
regrasping using spatio-temporal tactile features and reinforcement learning. In Intelligent Robots
and Systems (IROS), 2016 IEEE/RSJ International Conference on, pp. 1960–1966. IEEE, 2016.
Steven W Chen, Nikolay Atanasov, Arbaaz Khan, Konstantinos Karydis, Daniel D Lee, and Vijay
Kumar. Neural network memory architectures for autonomous robot navigation. arXiv preprint
arXiv:1705.08049, 2017.
Shreyansh Daftry, J Andrew Bagnell, and Martial Hebert. Learning transferable policies for monocular
reactive mav control. In International Symposium on Experimental Robotics, pp. 3–11. Springer,
2016.
Alessandro Giusti, Jérôme Guzzi, Dan C Cireşan, Fang-Lin He, Juan P Rodríguez, Flavio Fontana,
Matthias Faessler, Christian Forster, Jürgen Schmidhuber, Gianni Di Caro, et al. A machine
learning approach to visual perception of forest trails for mobile robots. IEEE Robotics and
Automation Letters, 1(2):661–667, 2016.
Alex Graves, Greg Wayne, and Ivo Danihelka.
arXiv:1410.5401, 2014.
Neural turing machines.
arXiv preprint
Alex Graves, Greg Wayne, Malcolm Reynolds, Tim Harley, Ivo Danihelka, Agnieszka GrabskaBarwińska, Sergio Gómez Colmenarejo, Edward Grefenstette, Tiago Ramalho, John Agapiou, et al.
Hybrid computing using a neural network with dynamic external memory. Nature, 538(7626):
471–476, 2016.
Saurabh Gupta, James Davidson, Sergey Levine, Rahul Sukthankar, and Jitendra Malik. Cognitive
mapping and planning for visual navigation. arXiv preprint arXiv:1702.03920, 2017.
Jérôme Guzzi, Alessandro Giusti, Luca M Gambardella, Guy Theraulaz, and Gianni A Di Caro.
Human-friendly robot navigation in dynamic environments. In Robotics and Automation (ICRA),
2013 IEEE International Conference on, pp. 423–430. IEEE, 2013.
Karol Kurach, Marcin Andrychowicz, and Ilya Sutskever. Neural random-access machines. arXiv
preprint arXiv:1511.06392, 2015.
Alex X Lee, Sergey Levine, and Pieter Abbeel. Learning visual servoing with deep features and fitted
q-iteration. arXiv preprint arXiv:1703.11000, 2017.
Sergey Levine. Exploring deep and recurrent architectures for optimal control. arXiv preprint
arXiv:1311.1761, 2013.
Sergey Levine, Chelsea Finn, Trevor Darrell, and Pieter Abbeel. End-to-end training of deep
visuomotor policies. arXiv preprint arXiv:1504.00702, 2015.
Laurens van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of Machine
Learning Research, 9(Nov):2579–2605, 2008.
Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Alex Graves, Ioannis Antonoglou, Daan
Wierstra, and Martin Riedmiller. Playing atari with deep reinforcement learning. arXiv preprint
arXiv:1312.5602, 2013a.
Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Alex Graves, Ioannis Antonoglou, Daan
Wierstra, and Martin Riedmiller. Playing atari with deep reinforcement learning. arXiv preprint
arXiv:1312.5602, 2013b.
Volodymyr Mnih, Adria Puigdomenech Badia, Mehdi Mirza, Alex Graves, Timothy Lillicrap, Tim
Harley, David Silver, and Koray Kavukcuoglu. Asynchronous methods for deep reinforcement
learning. In International Conference on Machine Learning, pp. 1928–1937, 2016.
Junhyuk Oh, Valliappa Chockalingam, Satinder Singh, and Honglak Lee. Control of memory, active
perception, and action in minecraft. arXiv preprint arXiv:1605.09128, 2016.
Emilio Parisotto and Ruslan Salakhutdinov. Neural map: Structured memory for deep reinforcement
learning. arXiv preprint arXiv:1702.08360, 2017.
14
Published as a conference paper at ICLR 2018
Stuart J. Russell and Peter Norvig. Artificial Intelligence: A Modern Approach. Pearson Education, 2
edition, 2003. ISBN 0137903952.
Hava T Siegelmann and Eduardo D Sontag. On the computational power of neural nets. Journal of
computer and system sciences, 50(1):132–150, 1995.
Richard S Sutton and Andrew G Barto. Reinforcement learning: An introduction, volume 1. MIT
press Cambridge, 1998.
Richard S Sutton, Doina Precup, and Satinder Singh. Between mdps and semi-mdps: A framework
for temporal abstraction in reinforcement learning. Artificial intelligence, 112(1-2):181–211, 1999.
Aviv Tamar, Yi Wu, Garrett Thomas, Sergey Levine, and Pieter Abbeel. Value iteration networks. In
Advances in Neural Information Processing Systems, pp. 2154–2162, 2016.
Sebastian Thrun and John J Leonard. Simultaneous localization and mapping. In Springer handbook
of robotics, pp. 871–889. Springer, 2008.
Jingwei Zhang, Lei Tai, Joschka Boedecker, Wolfram Burgard, and Ming Liu. Neural slam. arXiv
preprint arXiv:1706.09520, 2017.
Marvin Zhang, Zoe McCarthy, Chelsea Finn, Sergey Levine, and Pieter Abbeel. Learning deep neural
network policies with continuous memory states. In Robotics and Automation (ICRA), 2016 IEEE
International Conference on, pp. 520–527. IEEE, 2016a.
Tianhao Zhang, Gregory Kahn, Sergey Levine, and Pieter Abbeel. Learning deep control policies for
autonomous aerial vehicles with mpc-guided policy search. In Robotics and Automation (ICRA),
2016 IEEE International Conference on, pp. 528–535. IEEE, 2016b.
15
Published as a conference paper at ICLR 2018
APPENDIX
A
G RID W ORLD E XPERIMENTS
For the grid world we define our sensor to be a 7 × 7 patch with the agent at the center of this patch.
Our input image I to the VI module is [m × n × 2] image where m,n are the height and width of
the image. I[:, :, 0] is the sensor input. Since we set our rewards to be sparse, I[:, :, 1] is the reward
map and is zero everywhere except at the goal position (mgoal , ngoal ). I is first convolved to obtain a
reward image R of dimension [n × m × u] where u is the number of hidden units (vary between
100-150). This reward image is sent to the VI module. The value maps from the VI module after K
iterations are fed into the memory network controller. The output from the network controller (here a
LSTM with 256 hidden units) and the access module is concatenated and sent through a linear layer
followed by a soft max to get a probability distribution over A.
During training and testing we roll out our sequence state by state based on the ground truth or the
networks output respectively. Further, the set of transitions from start to goal to are considered to
be an episode. During training at the end of each episode the internal state of the controller and the
memory is cleared. The size of the external memory is 32 × 8 the grid world task. An additional
hyperparameter is the number of read heads and write heads. This parameter controls the frequency
of reads vs frequency of writes. For the the grid world task, we fix the number of read heads to 4
and the number of write heads to 1. For the grid world with simple obstacles, we observe that the
MACN performs better when trained with curriculum (Bengio et al., 2009). This is expected since
both the original VIN paper and the DNC paper show that better results are achieved when trained
with curriculum. For establishing baselines, the VIN and the CNN+Memory models are also trained
with curriculum learning. In the grid world environment it is easy to define tasks that are harder
than other tasks to aid with curriculum training. For a grid world with size (m, n) we increase the
difficulty of the task by increasing the number of obstacles and the maximum size of the obstacles.
Thus, for a 32 × 32 grid, we start with a maximum of 2 obstacles and the maximum size being 2 × 2.
Both parameters are then increased gradually. The optimal action in the grid world experiments is
generated by A star (Russell & Norvig, 2003). We use the Manhattan distance between our current
position and the goal as a heuristic. Our error curves on the test set for the MACN with the LSTM
and the addressable memory scheme are shown in Figure 11.
Figure 11: Testing Error on grid world (32 x 32)
16
Published as a conference paper at ICLR 2018
B
E NVIRONMENTS WITH T UNNELS
We setup the network the same way as we did for the grid world experiments with blob shaped
obstacles. Due to the relatively simple structure of the environment, we observe that we do not really
need to train our networks with curriculum. Additionally, the read and write heads are both set to 1
for this experiment.
B.1
VI MODULE + PARTIAL M APS
We observe that for the tunnel shaped obstacle when just the VI module is fed the partial map (stitched
together version of states explored) as opposed to the sensor input, it performs extremely well and is
able to generalize to new maps with longer tunnels without needing any memory. This is interesting
because it proves our intuition about the planning task needing memory. Ideally we would like the
network to learn this partial map on its own instead of providing it with a hand engineered version
of it. The partial map represents an account of all states visited in the past. We argue that not all
information from the past is necessary and the non redundant information that is required for planning
in the global environment can be learned by the network. This can be seen in the memory ablation.
B.2
DQN
As stated in the main paper, we observe that the DQN performs very poorly since the rewards are very
sparse. The network is setup exactly as described in (Mnih et al., 2013a). We observe that even after
1 million iterations, the agent never reaches the goal and instead converges to a bad policy. This can
be seen in Figure 12. It is clear that under the random initial policy the agent is unable to reach the
goal and converged to a bad policy. Similar results are observed for A3C. Further, it is observed that
even when the partial map instead of the sensor input is fed in to DQN, the agent does not converge.
Figure 12: Learning optimal policy with Deep Q networks. In this case the input to the DQN
is the sensor input. State after 1 million iterations. The agent gets stuck along the wall (left wall
between 40 and 50)
B.3
V ISUALIZING THE M EMORY S TATE
For the tunnel task, we use an external memory with 32 rows (N = 32) and a word size of 8 (W = 8).
We observe that when testing the network, the memory registers a change in its states only when
important events are observed. In Figure 11, the left hand image represents the activations from the
memory when the agent is going into the tunnel. We observe that the activations from the memory
remain constant until the agent observes the end of the tunnel. The memory states change when
the agent observes the end of the tunnel, when it exits the tunnel and when it turns towards its goal
(Figure 13). Another key observation for this task is that the MACN is prone to over fitting for this
task. This is expected because ideally, only three states need to be stored in the memory; entered the
tunnel, observe end of tunnel and exit tunnel. To avoid overfitting we add L2 regularization to our
memory access weights.
17
Published as a conference paper at ICLR 2018
Figure 13: Shift in memory states just before and after the end of the tunnel is observed. Once the
agent has turned around, the memory state stays constant till it reaches the end of the tunnel.
C
G RAPH E XPERIMENTS
For the graph experiment, we generate a random connected undirected graph with N nodes. We will
call this graph G = (V, E), with nodes V = {V1 , V2 , . . . , VN } and edges, E. The agent, at any point
in the simulation, is located at a specific node Vi and travels between nodes via the edges. The agent
can take actions from a set U = {u1 , u2 , . . . , uN } where choosing action ui will attempt to move to
node Vi . We have transition probabilities
0, if k 6= j or (Vi , Vj ) 6∈ E
p(Vj |Vi , uk ) =
(5)
1, if k = j and (Vi , Vj ) ∈ E
At each node, the agent has access to the unique node number (all nodes are labeled with a unique
ID), as well as the (A)i , the ith row of the adjacency matrix A. It also has access to the unique node
number of the goal (but no additional information about the goal).
(a) A random connected graph
(b) Adjacency matrix for graph
Figure 14: a) A random connected undirected graph with a the goal given by the star-shaped node,
and the current state given by the blue node b) Shows the corresponding adjacency matrix where
white indicates a connection and black indicates no connection. The goal is given by the row shaded
in green, and the current state is given by the row shaded in blue.
To train the network to navigate this graph, we used supervised training with an expert demonstrating
an intended behavior (breadth first search). Training samples were generated by running breadth first
search (and connecting nodes that are explored by traveling previously explored nodes of the graph).
Thus, for each state of the node and goal, we obtain a desired action. To fit this into the framework of
our network and 2D convolutions, we reshaped the row vector of the matrix into a matrix that could
18
Published as a conference paper at ICLR 2018
use the same convolution operation. The reward prior is also a row vector with a 1 at the index of the
goal node and zero everywhere else. This row vector is reshaped and stacked with the observation.
We train the graph by giving example paths between pairs of nodes. We then test on pairs of nodes
not shown during training. The training network is setup as before in the grid world navigation task.
Due to the increased action space and state space, this task is significantly more complex than the
grid world navigation task. We train MACN and the baselines with curriculum training. In the graph
task it is easy to define a measure of increasing complexity by changing the number of hops between
the start state and the goal state. Additionally, for the graph task the number of read heads and write
heads are set to 1 and 4 respectively.
D
C ONTINUOUS C ONTROL
Navigating an unknown environment is a highly
relevant problem in robotics. The traditional
methodology localizes the position of the robot
in the unknown world and tries to estimate a
map. This approach is called Simultaneous Localization and Mapping (SLAM) and has been
explored in depth in robotics (Thrun & Leonard,
2008). For the continuous control experiment,
we use a differential drive robot (Figure 15). The
robot is equipped with a head mounted LIDAR
and also has a ASUS Xtion Pro that can provide
the depth as well as the image from the front
facing camera. In this work, we only use the
information from the LIDAR and leave the idea
Figure 15: Simulated ground robot
of using data from the camera for future work.
The ground truth maps are generated by using
Human Friendly Navigation (HFN) (Guzzi et al., 2013) which generates reactive collision avoidance
paths to the goal. Given a start and goal position, the HFN algorithm generates way points that are
sent to the controller. For our experiment, we generate a tuple of (x, y, θ) associated with every
observation. To train the network, a m × n matrix (environment matrix) corresponding to the m × n
environment is initialized. A corresponding reward array (reward matrix) also of size m × n with a 1
at the goal position and zero elsewhere is concatenated with the environment matrix. The observations
corresponding to the laser scan are converted to a j × k matrix (observation matrix) where j < m
and k < n. The values at the indices in the environment array corresponding to the local observation
are updated with the values from the observation matrix. At every iteration, the environment matrix
is reset to zero to ensure that the MACN only has access to the partially observable environment.
For the continuous world we define our observation to be a 10 × 10 matrix with the agent at the
bottom of this patch. We change our formulation in the previous cases where our agent was at the
center since the LIDAR only has a 270 degree field of view and the environment behind the robot is
not observed. Our input image I to the VI module is [m × n × 2] image where m = 200,n = 200
are the height and width of the environment. I[:, :, 0] is the sensor input. I is first convolved to obtain
a reward image R of dimension [n × m × u] where u is the number of hidden units (200 in this
case). The K (parameter corresponding to number of iterations of value iteration) here is 40. The
network controller is a LSTM with 512 hidden units and the external memory has 1024 rows and a
word size of 512. We use 16 write heads and 4 read heads in the access module. The output from
the access module is concatenated with the output from the LSTM controller and sent through a
linear layer followed by a soft max to get probability distributions for (x, y, θ). We sample from
these distributions to get the next waypoint. These way points are then sent to the controller. The
waypoints are clipped to ensure that the robot takes incremental steps.
For this task, we find that the performance increases when trained by curriculum training. MACN in
addition to the baselines is first trained on maps where the goal is close and later trained on maps
where the goal is further away. An additional point here, is that due to the complexity of the task, we
train and test on the same map. Maps in the train set and test set differ by having random start and
goal regions.
19
| 2cs.AI
|
Statistical estimation of the Oscillating Brownian Motion
Antoine Lejay∗
Paolo Pigato†
arXiv:1701.02129v1 [math.PR] 9 Jan 2017
January 10, 2017
Abstract
We study the asymptotic behavior of estimators of a two-valued, discontinuous
diffusion coefficient in a Stochastic Differential Equation, called an Oscillating Brownian
Motion. Using the relation of the latter process with the Skew Brownian Motion, we
propose two natural consistent estimators, which are variants of the integrated volatility
estimator and take the occupation times into account. We show the stable convergence of
the renormalized errors’ estimations toward some Gaussian mixture, possibly corrected
by a term that depends on the local time. These limits stem from the lack of ergodicity
as well as the behavior of the local time at zero of the process. We test both estimators
on simulated processes, finding a complete agreement with the theoretical predictions.
Keywords: Oscillating Brownian Motion, Gaussian mixture, local time, occupation time, Arcsine
distribution, Skew Brownian Motion
1
Introduction
Diffusion processes with discontinuous coefficients attract more and more attention for
simulation and modelling purposes (see references in [23]). Many domains are actually
concerned, such as geophysics [31], population ecology [4, 5], finance [6, 28], . . . among others.
From a theoretical point of view, diffusions with discontinuous coefficients are an instance of
Stochastic Differential Equations (SDE) with local time — also called skew diffusion — for
which many results are contained in the work of J.-F. Le Gall [20].
Estimation come together with simulation and modelling, as models need to be calibrated.
This article deals with the parametric estimation of the coefficients of a one-dimensional SDE
of type
Z
Z
t
Yt = x +
t
σ(Ys ) dWs +
0
b(Ys ) ds
(1.1)
0
where W is a Brownian motion and x 7→ σ(x) takes two values {σ+ , σ− } according to the
sign of x, when the process Y is observed at discrete times iT /n, i = 0, . . . , n up to a time T .
∗
Université de Lorraine, IECL, UMR 7502, Vandœuvre-lès-Nancy, F-54500, France;
CNRS, IECL, UMR 7502, Vandœuvre-lès-Nancy, F-54500, France;
Inria, Villers-lès-Nancy, F-54600, France
email: [email protected]
†
Université de Lorraine, IECL, UMR 7502, Vandœuvre-lès-Nancy, F-54500, France;
CNRS, IECL, UMR 7502, Vandœuvre-lès-Nancy, F-54500, France;
Inria, Villers-lès-Nancy, F-54600, France
email: [email protected]
1
In a first part, we consider that (1.1) contains no drift (b = 0). The solution Y to (1.1) is
called an Oscillating Brownian Motion (OBM). This process was studied by J. Keilson and
J.A. Wellner in [18] who give some of its main properties and close form expressions for its
density and occupation time.
We provide two very simple estimators which generalize the integrated volatility (or
averaged squared) estimator (in finance, σ is called the volatility when Y represents the
logarithm of the price of an asset). For two processes Z, Z 0 , we set
[Z, Z 0 ]nT :=
n
X
i=1
n
0
0
(Zi,n − Zi−1,n )(Zi,n
− Zi−1,n
) and Q̄nT (Z, +) :=
1X
1Zi,n ≥0
n
i=1
Q̄nT (Z, +)
with Zi,n = ZiT /n . The quantity
is an approximation of the occupation time of Z on
n )2 of σ 2 is then [Y + , Y + ]/Q̄n (Y, +), where
the positive axis up to time T . Our estimator (σ̂+
+
T
+
2 . Although an analytic
Y is the positive part of Y . A similar estimator is defined for σ−
form is known for the density, this estimator is simpler to implement than the Maximum
Likelihood Estimator. Besides, it also applies when b =
6 0, while explicit expressions for the
density become cumbersome, at best [21, 29].
n is a consistent estimator of σ . Yet it is asymptotically biased. We
We show first that σ̂+
+
√
n
2 ) converges stably to a mixture of Gaussian distributions (in
also prove that n((σ̂+ )2 − σ+
which, unsurprisingly, the occupation time of the positive side appears) plus an explicit term
giving the bias.
When estimating σ+ , the actual size of the “useful” sample is proportional to the occupation
time of R+ . Therefore, a dependence on the occupation time is to be expected in any reasonable
estimator. The law of the occupation time for the OBM follows an arcsine type distribution,
which generalizes the one of the Brownian motion. Since these laws carry much mass close
to the extremes, this amounts to say that many trajectories of this process spend most of
the time on the same side of 0. Therefore, with a high probability, either σ+ or σ− is only
based on few observations. This affects our central limit theorem as well as the quality of the
estimation of (σ+ , σ− ), meaning that the limit law will not be a Gaussian one, as one would
expect from the approximation of quadratic variation, but a Gaussian mixture displaying
heavy tails.
Another tool of importance in this framework, strictly connected with the occupation
time, is the local time. Given a stochastic process Y , its local time at a point x, denoted by
{Lxt (Y )}t≥0 , represents the time spent by Y at x, properly re-scaled. It has a fundamental
role in the study of SDEs with discontinuous coefficients. Intuitively, the local time appears
when dealing with discontinuous coefficients because it helps to quantify what happens locally
at the discontinuity. A Lamperti’s type transform applied with the help of the Itô-Tanaka
formula shows that Y is intimately related to the Skew Brownian Motion (SBM, see [10, 24])
X, the solution to the SDE
Xt = x + Wt + θL0t (X), −1 < θ < 1,
through a simple deterministic transform X = Φ(Y ) [8, 26].
In the present paper, the local time plays an important role for two reasons. First,
because we use the transform X = Φ(Y ) to apply some convergence results which extend
to thePSBM some results of J. Jacod on the asymptotic behavior of quantities of type
n−1/2 ni=1 f (Xi−1,n , Xi,n ). Second, because the local time itself appears in the limit of the
above quantities. Actually, the asymptotic bias is related to the local time.
2
2 , defined as (mn )2 = [Y ± , Y ]n /Q̄n (Y, ±)
We then provide a second simple estimator of σ±
±
T
T
√
2 ) converges stably to a Gaussian
n
which is also consistent. We show that n((m± )2 − σ±
√
n )2 −σ 2 ) is removed through the introduction
mixture. The asymptotic bias observed in n((σ̂±
±
of the quadratic term [Y + , Y − ] which is related to the local time. The variance of the former
limit is not larger than the one of the latter.
In Corollary 3.8, we also generalize these convergence results in presence of a bounded
drift term. We prove that the estimators mentioned above converge to analogous limit random
variables, depending on the occupation time of the SDE under study. Unlike for the OBM,
the limit law is not explicit for SDEs with general drift, since the law of the occupation time
is not know, if not in special cases (see e.g. [17, 37, 38]).
The novelty of the paper lies in the treatment of a discontinuous diffusion coefficient. This
implies a drastic difference with the case of regular coefficients, as the situation cannot be
reduced to a Brownian one (the measure of the SBM being singular with respect to the one
n , which
of the Brownian motion [20]). This explains the presence of an asymptotic bias for σ̂±
n
is removed by a correction (leading to m± ) which involves only a fraction of order n−1/2 of
the observations.
Besides, the framework is not the one of ergodic processes, like for many estimators, but
of null recurrent ones. On the last point, our study does not fit the situations considered e.g.,
in [1, 11, 12].
With respect to many estimators constructed for diffusion processes, the lack of ergodicity
of the process explains the presence of a mixed normal distribution in the limit. For diffusions,
asymptotic convergence involving a mixture of normal distributions (with different type of
limits) is already observed in the works of F. Florens-Zmirou [9] for non-parametric estimation,
and of J. Jacod [14, 15], from which we borrow and adapt the general treatment. The core of
our proof requires the adaptation to the SBM of some results on the convergence toward the
local time given in [14].
Content of the paper
In Section 2 we define the Oscillating Brownian Motion (OBM) and recall some useful
properties. In Section 3 we define our estimators and state precisely the convergence theorems.
These results are then proved is Section 4. In Section 5 we consider the Oscillating Random
Walk, a discrete process that can be used to construct the OBM, and study an estimator
on this discrete process. Section 6 is devoted to the implementation of the estimators of the
OBM, and contains numerical experiments showing the good behavior of the estimators in
practice.
Notations
For notational convenience, we work on the time interval [0, 1]. Our results can be extended
to a general time interval via a space-time re-scaling (see Remark 3.7). Throughout the paper,
p
law
we use the following notation for convergence of random variables: −
→ in probability; −−→
sl
law
in law; −
→ stable in law; = denotes equality in law. The Lebesgue measure is written
Leb. The positive and negative parts of x ∈ R are denoted by x+ = x ∨ 0, x− = (−x) ∨ 0.
For any continuous semimartingale M , we write hM i for its quadratic variation process. For
y ∈ R we define the (symmetric) local time of M at y as the process (Lyt (M ))t∈[0,1] , with
3
(See [33, Corollary VI.1.9, p. 227])
Lyt (M ) = lim
ε↓0
1
2ε
Z
0
t
1{y−ε≤Ms ≤y+ε} dhM is almost surely.
When we do not specify y, we mean the local time at 0: Lt (M ) = L0t (M ).
For fixed n ∈ N, we consider the discretization of [0, 1] given by 0, 1/n, . . . , 1. For any
process (Mt )t∈[0,1] , we write Mk,n = Mk/n . For any processes M, M̄ we also set the “discrete
bracket”
n
X
[M, M̄ ]n1 =
(Mk,n − Mk−1,n )(M̄k,n − M̄k−1,n ).
k=1
We also write [M ]n1 = [M, M ]n1 .
2
Oscillating Brownian Motion
For two parameters σ+ , σ− > 0, we define the diffusion coefficient σ as follows:
σ(y) = σ+ 1(y ≥ 0) + σ− 1(y < 0), ∀y ∈ R.
(2.1)
Let W be a Brownian motion with its (completed) natural filtration (Gt )t≥0 on a probability
space (Ω, F, P). From now on, we denote by Y the unique strong solution to
t
Z
Yt = Y0 +
σ(Ys ) dWs .
(2.2)
0
Strong existence and uniqueness of Y is granted by the results of [20]. Following the
terminology of [18], we call Y an Oscillating Brownian Motion (OBM)
We recall some known properties of Y , proved in [18].
For (t, x, y) ∈ (0, 1] × R2 , let p(t, x, y) be the density of Yt in y, with initial condition
Y0 = x. In [18] explicit formulas for the transition density are given. In particular, when
sgn x 6= sgn y,
1 −( σx+ − σy− )2 2t1
2σ+
√
for x > 0, y < 0
σ− (σ+ + σ− ) 2πt e
p(t, x, y) =
2σ−
1 −( σy+ − σx− )2 2t1
√
e
for x < 0, y > 0.
σ+ (σ+ + σ− ) 2πt
Integrating the previous equations we obtain
√
2σ+
Φ −Y0 /(σ+ t) ,
σ+ + σ−
√
2σ−
P[Yt > 0 | Y0 < 0] =
Φ Y0 /(σ− t) ,
σ− + σ+
P(Yt < 0|Y0 > 0) =
(2.3)
where Φ is the cumulative density function of the standard Gaussian. The law of the
occupation time of R+ is also computed in [18]. Let the occupation time of R+ be defined as
Q+
t = Leb{s ≤ t : Ys ≥ 0}.
4
(2.4)
law
+
+
The distribution of Q+
t , with Y0 = 0 a.s., is explicit. The scaling Q1 = Qt /t holds, and
P(Q+
1 ∈ du) =
1
σ+ /σ−
1
p
du for 0 < u < 1.
π u(1 − u) 1 − (1 − (σ+ /σ− )2 )u
(2.5)
This generalizes the arcsine law for the occupation time of the Brownian Motion. The
+
occupation time Q− on R− is easily computed from Q+ since obviously, Q−
t + Qt = t for
any t ≥ 0.
We introduce the process (Xt )t≥0 , setting Xt = Yt /σ(Yt ) for t ≥ 0. It follows from the
Itô-Tanaka formula that X is a Skew Brownian Motion (SBM, see [24]), meaning that X
satisfies the following SDE:
Xt = X0 + Bt + θLt (X),
(2.6)
where B is a Brownian Motion, Lt (X) is the symmetric local time of X at 0, X0 = Y0 /σ(Y0 )
and the coefficient θ is given by
σ− − σ+
.
(2.7)
θ=
σ− + σ+
We write from now on BM for Brownian Motion, SBM for Skew Brownian Motion, OBM for
Oscillating Brownian Motion. The local times of X and Y are related by
Lt (X) =
σ+ + σ−
Lt (Y )
2σ+ σ−
(2.8)
(see [24] for a special case from which we easily recover this formula).
3
3.1
Main results
The stable convergence
Before stating our results, we need to recall the notion of stable convergence, which was
introduced by A. Rényi [32]. We refer to [16] or [15] for a detailed exposition. Let Zn a
sequence of E-valued random variables defined on the same probability space (Ω, F, P). Let Z
be an E-valued random variable defined on an extension, (Ω̃, F̃, P̃). We then say that Zn
sl
converges stably to Z (and write Zn −−−→ Z) if:
n→∞
E[Y f (Zn )] −−−→ Ẽ[Y f (Z)]
n→∞
for all bounded continuous functions f on E and all bounded random variables Y on (Ω, F)
(or, equivalently, for all Y as above and all functions f which are bounded and Lipschitz).
This notion of convergence is stronger than convergence in law, but weaker than convergence
in probability. We use in this paper the following crucial result: for random variables Yn , Zn
(n ≥ 1), Y and Z,
sl
p
sl
n→∞
n→∞
n→∞
if Zn −−−→ Z and Yn −−−→ Y then (Yn , Zn ) −−−→ (Y, Z).
5
3.2
Estimators for the parameters the Oscillating Brownian motion
Let us assume that we observe the process Y solution to (2.2) at the discrete times 0, 1/n, . . . , 1.
We want to estimate σ+ , σ− from these observations.
A natural estimator for the occupation time Q+
1 defined in (2.4) is given by the Riemann
sums (see Section 4.4):
n
X
1(Yk ≥ 0)
Q̄n1 (Y, +) =
.
(3.1)
n
k=1
We define now
n
σ̂+
as
s
n
σ̂+
=
[Y + ]n1
,
Q̄n1 (Y, +)
(3.2)
which we show to be a consistent estimator for σ+ . Similarly, we set
s
n
X
[Y − ]n1
1(Y
<
0)
k
n
,
Q̄n1 (Y, −) =
= 1 − Q̄n1 (Y, +) and σ̂−
=
n
Q̄n1 (Y, −)
k=1
Finally, we define our estimator of the vector (σ+ , σ− )0 as
n
σ̂+
n
σ̂ =
n .
σ̂−
Theorem 3.1. Let Y be solution of (2.2) with Y0 = 0 a.s., and σ̂ n defined in (3.2). Then
(i) The estimator is consistent:
σ+
.
σ̂ −−−→
n→∞
σ−
n
p
(ii) There exists an extension (Ω̃, F̃ , P̃) of (Ω, F, P) carrying a Brownian motion B̄ independent
from W such that
√
2 R
2σ+
1
n 2
1 ! √
2
0 1(Ys > 0) dB̄s
+
√
2 2
σ− σ+
(σ̂+ ) − σ+
sl
√Q+
Q
1
1
√
n
−−−→
L1 (Y ).
−
1
n )2 − σ 2
2 R
n→∞
(σ̂−
2σ−
3 π σ+ + σ−
1
−
1−Q+
1
1(Y
<
0)
d
B̄
s
s
0
1 − Q+
1
(3.3)
The stable limit in the above result depends on the path of Y through its local time L1 (Y )
and its occupation time Q±
1 . By identifying the distribution of the limit, we can rewrite
(3.3) as a convergence in distribution involving elementary random variables as
√ 2
2σ+
n 2
2
√
N1
√
(σ̂+ ) − σ+
law
−
n
−−−→ √2σΛ2
n
2
2
n→∞
(σ̂− ) − σ−
√ −N
1−Λ
q
√
1−Λ
2
8 2 (σ− σ+ )
ξ
q Λ
√
q
Λ
σ
+
σ
3
π
2
2
+
−
(1 − Λ)σ− + Λσ+
2
1−Λ
√
√
2σ 2
8
1 √ ξ 1−Λ
√ + N1 − √
3 π r+1 (1−Λ)+Λr2
Λ
, (3.4)
√ 2
=
√
2σ−
1 √ ξ Λ
√
N2 − √8 1/r+1
2
1−Λ
6
3 π
Λ+(1−Λ)/r
where r = σ+ /σ− , ξ, N1 , N2 , Λ are mutually independent, ξ ∼ exp(1), N1 , N2 ∼ N (0, 1)
and Λ follows the modified arcsine law (2.5) with density
pΛ (τ ) =
r
1
.
1/2
− τ ) 1 − (1 − r2 )τ
πτ 1/2 (1
Remark 3.2. The Brownian case is σ =: σ+ = σ− , r = 1. The limit law is
√
√ 2 √1Λ N1 − 3√4 π ξ 1 − Λ
2σ
√
4
√1
√
N2 − 3 π ξ Λ
1−Λ
where Λ follows the classical arcsine law (see [27, 33]).
Remark 3.3. In (4.21) we prove
√
+
n[Y , Y
sl
− n
]1 −−−→
n→∞
√
2 2 σ+ σ−
√
L1 (Y ).
3 π σ+ + σ−
Actually, each term of type (Yt+
− Yt+
)(Yt−
− Yt−
) vanishes unless sgn(Yti ) 6= sgn(Yti−1 ).
i
i−1
i
i−1
√
Thus, n[Y + , Y − ]n1 provides us with as an estimator of the local time since it somehow
counts the number of crossings of zero (cf. [14, 22]).
Remark 3.4. We assume Y0 = 0 a.s. because we need Y0 to visit both R+ and R− . This
happens a.s. in any right neighborhood of 0 if the diffusion starts from 0. If the initial
condition is not 0, we shall wait for the first time at which the process reaches 0, say τ0 , and
consider the (random) interval [τ0 , T ].
We define now a different estimator for σ± by
s
s
+ , Y ]n
[Y
[Y − , Y ]n1
n
1
mn+ :=
,
m
:=
−
Q̄n1 (Y, +)
Q̄n1 (Y, −)
n
m+
.
and m :=
mn−
n
(3.5)
Theorem 3.5. Let Y be solution of (2.2) with Y0 = 0 a.s., and mn defined in (3.5). The
following convergence holds for n → ∞:
√ 2
2σ+ R 1
n 2
2
√
(m+ ) − σ+
sl
+ 0 1(Ys > 0) dB̄s
n
−−−→ √Q2σ1 2 R
,
n
2
2
1
(m− ) − σ− n→∞
−
+ 0 1(Ys < 0) dB̄s
1−Q
1
where B̄ is a BM independent of Y on an extension (Ω̃, F̃, P̃) of (Ω, F, P). We can rewrite
such convergence as follows:
√ 2
2σ+
n 2
2
√
N1
√
(m+ ) − σ+
law
,
(3.6)
n
−−−→ √2σΛ2
2
n→∞
(mn− )2 − σ−
√ −N
1−Λ
2
where N1 , N2 , Λ are mutually independent, N1 , N2 ∼ N (0, 1) and Λ follows the modified
arcsine law with density given by (2.5), with r = σ + /σ− .
7
Remark 3.6. Comparing Theorems 3.1 and 3.5, we see that an asymptotic bias is present
√
in σ̂ n , but not in mn . This bias has the same order (∼ 1/ n) as the “natural fluctuations” of
n underestimates σ .
the estimator. Because the local time is positive, it is more likely that σ̂±
±
For a more quantitative comparison between the convergence of the two estimators, see
Remark 4.18. In Section 6, we compare the two estimators in practice.
Remark 3.7. Theorem 3.5 gives the asymptotic behavior for an estimator of (σ+ , σ− ) in
the presence of high frequency data, yet with fixed time horizon T = 1. The OBM enjoys a
√
scaling property: if Y is an OBM issued from 0, then ( cYt/c )t∈R+ is an OBM issued form
0, for any c > 0 constant (see [33, Exercise IX.1.17, p. 374]). Using this fact, we can easily
generalize Theorem 3.5 to the case of data on a time interval [0, T ] for some fixed T > 0. We
set
s
s
+ , Y ]n
[Y
[Y − , Y ]nT
n,T
T
mn,T
:=
,
m
:=
.
(3.7)
+
−
Q̄nT (Y, +)
Q̄nT (Y, −)
The estimator is consistent and we have the following convergence:
√ 2
!
2T σ+ R T
n,T
1(Y
>
0)
d
B̄
2
2
s
s
√
0
(m+ ) − σ+
sl
Q+
n
−−−→ √2TTσ2 R
,
n,T 2
2
T
−
(m− ) − σ− n→∞
1(Ys < 0) dB̄s
+
T −QT
0
where B̄ is a BM independent of W on an extension of the underlying probability space. The
limiting random variable follows the law given in (3.6), which actually does not depend on T .
A slightly different approach is to imagine that our data are not in high frequency,
but that we observe the process at regular time intervals, for a long time. In this case
it is more reasonable to consider an OBM (Yt )t∈R+ , construct an estimator depending on
(Yi )i=0,1,...,T −1,T , T ∈ N, and then take the limit in long time. We set
s
s
+ , Y ]T
[Y
[Y − , Y ]TT
T
T
µT+ :=
,
µ
:=
.
−
Q̄TT (Y, +)
Q̄TT (Y, −)
Using again Theorem 3.5 and the diffusive scaling of the OBM, we have the following
convergence:
√ 2
2σ+
T 2
2
√
√
N1
(µ+ ) − σ+
law
.
T
−−−−→ √2σΛ2
T
2
2
(µ− ) − σ− T →∞
√ −N
1−Λ
2
The limit distribution is again the law given in (3.6). Theorem (3.1) can also be generalized
to high frequency data on an interval [0, T ] and to equally spaced data in long time, using
the diffusive scaling and (4.19)-(4.20). For example, analogously to (3.7), we define
s
s
[Y + , Y + ]nT
[Y − , Y − ]nT
n,T
n,T
σ̂+ :=
,
σ̂
:=
.
−
Q̄nT (Y, +)
Q̄nT (Y, −)
Again, the limit law does not change and is the one given in (3.4).
8
3.3
A generalization to OBM with drift
We consider now a wider class of processes, adding a drift term to equation (2.2). Formally,
let now ξ be the strong solution to
dξt = b(ξt ) dt + σ(ξt ) dWt ,
(3.8)
with ξ0 = 0, σ defined in (2.1) and b measurable and bounded. Again, strong existence and
uniqueness of the solution to (3.8) is ensured by the results of [20].
Let mn (ξ) be defined as in (3.5):
s
s
n
+ , ξ]n
[ξ
[ξ − , ξ]n1
m+ (ξ)
n
n
n
1
, m− (ξ) =
and m (ξ) :=
.
m+ (ξ) =
n
n
mn− (ξ)
Q̄1 (ξ, +)
Q̄1 (ξ, −)
Let us also denote Q+
1 (ξ) =
R1
0
1(ξs > 0) ds.
Corollary 3.8. The following convergence holds for n → ∞:
√ 2
2σ+ R 1
n
2
√
(m+ (ξ))2 − σ+
0 1(ξs > 0) dB̄s
sl
√+
1 (ξ)
−−−→ Q
n
,
2
R1
n
2
2
2σ
(m− (ξ)) − σ− n→∞
−
1(ξ
<
0)
d
B̄
s
s
1−Q+ (ξ) 0
1
where B̄ is a BM independent of W on an extension of the underlying probability space. We
can rewrite such convergence as follows:
√ 2
2σ+
n
2
2
√
N
√
1
(m+ (ξ)) − σ+
law
,
−−−→ √2σΘ2
n
2
n→∞
(mn− (ξ))2 − σ−
√ −N
1−Θ
2
law
where N1 , N2 , Θ are mutually independent, N1 , N2 ∼ N (0, 1) and Θ = Q+
1 (ξ).
Remark 3.9. Unlike for the OBM, the limit law is not explicit in Corollary 3.8, since the
law of the occupation time of the positive axes is not know in general (See e.g., [17, 19, 37, 38]).
On the other hand, some information on the law of Θ can be obtained, at least in some
special cases, via Laplace transform.
We also stress that this dependence on the occupation time is due the actual sample size
of the data giving us useful information. Indeed, when estimating σ+ , the number of intervals
that we can use is proportional to the occupation time of R+ . Analogously for the negative
part.
Remark 3.10. Actually, Corollary 3.8 holds under weaker assumptions. An example of
model fitting into this framework is the SET-Vasicek model [6], a generalization of the Vasicek
interest rate model to a non-constant volatility, given exactly by (2.1):
dξt = −α(ξt − β) dt + σ(ξt ) dWt .
Remark 3.11. The scaling property described in Remark 3.7 no longer holds in this situation,
so that the estimator can only be used in the “high frequency” setting.
9
4
Proofs of the convergence theorem
This section is devoted to the proof Theorems 3.1 and 3.5. We first deal with some general
approximation results which are well known for diffusions with regular coefficients (see [14–16]),
but not for the framework considered here with discontinuous coefficients (when θ 6= 0, the
law of the SBM is singular with respect to the one of the BM [20]).
Following [8, 26], we use the connection between the OBM and the SBM through a
Lamperti-type transform. Hence, we apply the results of [22] to the convergence of estimators
of quadratic variation, covariation and occupation time for these processes. Finally, we use
all these results to prove the main Theorems 3.1 and 3.5.
4.1
Approximation results
Let us write
Z
∞
Leb(φ) =
φ(α) dα
−∞
for the Lebesgue integral of a function. In [22], the following approximation result, borrowed
from [14], is proved for the SBM solution of (2.6).
R
Lemma 4.1. Let f be a bounded function such that |x|k |f (x)| dx < ∞ for k = 0, 1, 2 and X
be a SBM of parameter θ ∈ [−1, 1] (i.e., the solution to (2.6)). Then for any a > 0,
"
#
n−1
√
1 X
P √
f (Xi,n n) − λ̄θ (f )L1 (X) > a → 0,
n
i=1
where {Lt (X)}t≥0 is the symmetric local time at 0 of the SBM and
λ̄θ (f ) = (1 + θ) Leb(f + ) + (1 − θ) Leb(f − ).
(4.1)
Remark 4.2. In particular, when θ = 0, X is BM and the coefficient in front of the local
time is simply Leb(f ). We recover there a special case of a theorem by J. Jacod (see [14],
Theorem 4.1).
We prove now an approximation result for the OBM.
Lemma 4.3. Let Y be the OBM in (2.2). Let f be a bounded function such that
∞ for k = 0, 1, 2. Then for any a > 0,
#
"
n−1
√
1 X
P √
f (Yi,n n) − λσ (f )L1 (Y ) > a → 0,
n
i=1
where Lt (Y ) is the local time of Y and
Leb(f + ) Leb(f − )
λσ (f ) = 2
+
.
2
2
σ+
σ−
10
R
|x|k |f (x)| dx <
Proof. Recall Yt /σ(Yt ) = Xt , (2.8) and (2.7). Let f˜(x) = f (σ(x)x). We have
λ̄θ (f˜) = (1 + θ) Leb(f˜+ ) + (1 − θ) Leb(f˜− )
2σ−
2σ+
2σ+ σ−
=
Leb(f + ) +
Leb(f − ) =
λσ (f ),
σ+ (σ− + σ+ )
σ− (σ− + σ+ )
(σ− + σ+ )
so λσ (f )Lt (Y ) = λ̄θ (f˜)Lt (X). Therefore, from Lemma 4.1, for any a > 0,
"
P
#
n−1
√
1 X
√
f (Yi,n n) − λσ (f )Lt (Y ) > a
n
i=1
"
#
n−1
√
1 X˜
=P √
f (Xi,n n) − λ̄θ (f˜)L1 (X) > a −−−→ 0.
n→∞
n
i=1
This concludes the proof.
We state now a very special case of Theorem 3.2 in [13], that we apply several times in
this work. The version in [13] holds for semimartingales, not only martingales, the processes
involved can be multi-dimensional, and the limit process is not necessarily 0. Anyways, we
do not need this general framework here. Stating the theorem only for one-dimensional
martingales converging to 0 allows us to keep a simpler notation, which we introduce now: for
each càdlàg process J we write ∆ni J = Ji/n − J(i−1)/n . Consider a filtered probability space
(Ω, F, F , P) carrying a Brownian motion B. The filtration F = (Ft )t∈[0,1] is the natural
(completed) one for the Brownian motion. We define the filtration F n as the “discretization”
defined by Ftn = F[nt]/n . We consider a F n -martingale in R, i.e., a process of the form
Z1n
=
n
X
χni ,
i=1
where each χni is Fi/n measurable, square-integrable, and E[χni | F i−1 ] = 0.
n
Theorem 4.4 (Simplified form of Theorem 3.2 in [13]). Suppose that
n
h
i p
X
E |χni |2 F i−1 −−−→ 0,
and
i=1
n
X
i=1
n
n→∞
i p
h
E χni ∆ni B F i−1 −−−→ 0.
n
n→∞
(4.2)
(4.3)
Then Z1n converges to 0 in probability as n → ∞.
Remark 4.5. In [13] some kind of uniform integrability is assumed in the limit, whereas here
we do not ask explicitly for such a condition. The reason is that the uniform integrability
assumption is implied by the facththat the limit in
i (4.2) is 0.
Pn
n
n
It is also required that i=1 E χi ∆i N F i−1 converges to 0 for any bounded martingale
n
N orthogonal to B on (Ω, F, F, P). As we have considered the Brownian motion with its
natural (completed) filtration, this set is reduced to the constant ones.
11
4.2
Scaled quadratic variation of Brownian local time
Let (βt )(t∈[0,1]) be a BM and L(β) its local time at 0. Let us recall the diffusive scaling
√
law √
property (βct , Lt (β))t>0 = ( cβ, cL(β)) for any c > 0 (see e.g. [33, Exercise 2.11, p. 244]).
Let H = (Ht )t∈[0,1] be the natural (completed) filtration of β.
For i = 1, . . . , n, we write Hi,n = Hi/n .
Lemma 4.6. Let L(β) be the Brownian local time at 0. The following convergence holds:
√
n[L(β)]n1
√
n
√ X
4 2
p
2
= n
(Li,n (β) − Li−1,n (β)) −−−→ √ L1 (β).
n→∞ 3 π
i=1
We split the proof of this result in the next tree lemmas. We start with the explicit
computations on the moments of the Brownian local time.
Lemma 4.7. For p ≥ 1, we set φp (α) := E[L1 (β)p | β0 = α]. We have
Leb(φp ) =
2
E|N |p+1 ,
p+1
(4.4)
where N denotes a standard Gaussian random variable. Besides, the following tail estimates
hold for p = 2, 4:
√
2
2
e−α /2
16 2 e−α /4
√
√
φ2 (α) ≤
and φ4 (α) ≤
.
(4.5)
α
π
α 2π
Remark 4.8. These functions φp will be useful when applying Lemma 4.1 and Remark 4.2,
taking, for fixed p, f = φp . Inequalities (4.5) imply that the integrability condition for f is
satisfied and the theorem can be applied.
Proof. Formula (6) in [35] gives the following expression for the moments of the Brownian
local time
Z ∞
φp (α) = 2p
xp−1 P(N ≥ |α| + x) dx.
0
To apply Remark 4.2 we need to compute the following integral
Z
∞
Leb(φp ) =
∞
Z
φp (α) dα = 2
−∞
∞
Z
φp (α) dα = 2
0
Z
2p
0
∞
x
p−1
0
Z
∞
0
2
e−t /2
√
1[t≥α+x] dt dx dα.
2π
Changing the order of integration by Fubini-Tonelli’s theorem,
Z
Leb(φp ) = 4p
0
∞
2
e−t /2
√
dt
2π
Z
∞
x
p−1
Z
dx
0
∞
1[t≥α+x] dα =
0
2
E|N |p+1 ,
p+1
so (4.4) is proved. We now use the following bound for Gaussian tails:
We apply it twice and find the upper bound for p = 2:
Z
φ2 (α) = 4
∞
Z
x
0
0
∞
2
2
R∞
e−t /2
e−α /2
√
1[t≥α+x] dt dx ≤ √ .
2π
α 2π
12
x
2
2
/2
e−t
√
2π
dt ≤
/2
e−x
√
.
x 2π
For p = 4 we apply the same inequality:
Z
∞
x
φ4 (α) = 8
3
Z
∞
0
0
2
e−t /2
√
1[t≥α+x] dt dx
2π
Z ∞
Z ∞
2
2
e−(α+x) /2
e−(α+x) /2
√ dx ≤ 8
x3
x2 √
≤8
dx.
(α + x) 2π
2π
0
0
Now, since xe−x ≤ e−1 for all x ≥ 0,
φ4 (α) ≤ 16
Z
0
∞
2 /4
e−(α+x)
√
2π
2
e−α /4
dx ≤ 32 √ .
α 2π
Hence the result.
We consider now the quadratic sum in Lemma 4.6, and write
n
n
√ X
√ X
n
(Li,n (β) − Li−1,n (β))2 = n
E[(Li,n (β) − Li−1,n (β))2 |Hi−1,n ]
i=1
i=1
n
√ X
+ n
(Li,n (β) − Li−1,n (β))2 − E[(Li,n (β) − Li−1,n (β))2 |Hi−1,n ] .
i=1
In the next two lemmas we prove the convergence of the two summands. Lemma 4.6 follows
directly.
Lemma 4.9. Let L(β) be the Brownian local time at 0. The following convergence holds:
√
n
√ X
4 2
p
2
n
E[(Li,n (β) − Li−1,n (β)) |Hi−1,n ] −−−→ √ L1 (β).
n→∞ 3 π
i=1
Proof. The diffusive scaling property of (β, L(β)) implies that for any p ≥ 1,
E[(Li,n (β) − Li−1,n (β))p | Hi−1,n ] =
1
np/2
√
E L1 (β)p β0 = nβi−1,n .
(4.6)
Setting p = 2 and since Leb(φ2 ) = 23 E|N |3 from (4.4), Remark 4.2 below Lemma 4.1 implies
that
√
n
n
X
√ X
√
1
4 2
p
2
√ φ2 (βi−1,n n) −−−→ √ L1 (β).
n
E[(Li,n (β) − Li−1,n (β)) |Hi−1,n ] =
n→∞ 3 π
n
i=1
i=1
Hence the result.
We consider now the martingale part.
Lemma 4.10. With
Hi,n := (Li,n (β) − Li−1,n (β))2 − E[(Li,n (β) − Li−1,n (β))2 |Hi−1,n ],
it holds that
√ Pn
p
n i=1 Hi,n −−−→ 0.
n→∞
13
Proof. The statement is proved using Theorem 4.4 by setting χni :=
√
nHi,n .
• We prove fist (4.2). From (4.6) with p = 4,
2
√
1
E Hi,n
Hi−1,n ≤ E (Li,n (β) − Li−1,n (β))4 Hi−1,n = 2 φ4 (βi−1,n n).
n
P
√
With Remark 4.2 below Lemma 4.1, n−1/2 ni=1 φ4 (βi−1,n n) converges in probability to
Leb(φ4 )L1 (β) because of (4.5). Thus,
!
n
n
X
2
√
1
1 X
p
√
n
E Hi,n Hi−1 ≤ √
φ4 (βi−1,n n) −−−→ 0.
n→∞
n
n
i=1
i=1
• We take B = β in (4.3). We have
n
X
i=1
E[Hi,n (βi,n − βi−1,n ) | Hi−1,n ] =
n
X
E (Li,n (β) − Li−1,n (β))2 (βi,n − βi−1,n ) Hi−1,n
i=1
− E (Li,n (β) − Li−1,n (β))2 Hi−1,n E[(βi,n − βi−1,n ) | Hi−1,n ].
Since E[(βi,n − βi−1,n ) | Hi−1,n ] = 0, we only estimate the first summand:
E (Li,n (β) − Li−1,n (β))2 (βi,n − βi−1,n ) Hi−1,n
1/2
1/2
≤ E (Li,n (β) − Li−1,n (β))4 Hi−1,n
E (βi,n − βi−1,n )2 Hi−1,n
.
We estimate the two factors:
1/2
1
E (βi,n − βi−1,n )2 Hi−1,n
≤√
n
and from (4.6) with p = 4,
1/2
√
1
E (Li,n (β) − Li−1,n (β))4 Hi−1,n
≤ φ4 (βi−1,n n)1/2 .
n
Therefore, from Remark 4.2, that can be applied because of (4.5),
√
n
n
X
i=1
1
E[Hn,i (βi,n − βi−1,n ) | Hi−1,n ] ≤ √
n
n
√
1 X
√
φ4 (βi−1,n n)1/2
n
i=1
!
p
−−−→ 0.
n→∞
The proof is then complete.
4.3
Scaled quadratic covariation of skew Brownian motion and its local
time
We now give some results on the scaled quadratic covariation between the SBM and its local
time. For the Brownian motion W with the filtration G = (Gt )t≥0 of Section 2, we consider X
the strong solution to Xt = x + Wt + θLt (X) for θ ∈ [−1, 1] and L(X) its local time (apart
from the results in [20], strong existence for the SBM has been proved first in [10]).
14
Lemma 4.11. For X and L(X) as above, the following convergence holds:
√
n
√ X
p
= n
(Xi,n − Xi−1,n )(Li,n (X) − Li−1,n (X)) −−−→ 0,
(4.7)
n
√ X
p
= n
(|Xi,n | − |Xi−1,n |)(Li,n (X) − Li−1,n (X)) −−−→ 0,
(4.8)
n[X, L(X)]n1
n→∞
i=1
√
n[|X|, L(X)]n1
n→∞
i=1
√
n[X
+
, L(X)]n1
n
√ X
p
+
+
= n
(Xi,n
− Xi−1,n
)(Li,n (X) − Li−1,n (X)) −−−→ 0.
(4.9)
n→∞
i=1
We set
Zi,n := (Xi,n − Xi−1,n )(Li,n (X) − Li−1,n (X))
and write
(4.10)
n
n
n
√ X
√ X
√ X
n
Zi,n = n
E[Zi,n |Gi−1,n ] + n
Zi,n − E[Zi,n |Gi−1,n ] .
i=1
i=1
i=1
We prove (4.7) in the next two lemmas. Once (4.7) is proved, (4.8) follows since |X| is a
SBM with parameter θ = 1, while (4.9) follows from a combination of (4.7) and (4.8) since
X + = |X|+X
.
2
Lemma 4.12. With Zi,n defined in (4.10), the following convergence holds:
n
√ X
p
n
E[Zi,n | Gi−1,n ] −−−→ 0.
n→∞
i=1
(4.11)
Proof. We express first E(Xt − x)2 as a function of x using the law of the SBM. The density
transition function of the SBM is [24, 36]
pθ (t, x, y) := p(t, x − y) + sgn(y)θp(t, |x| + |y|)
where p(t, x) = (2πt)−1/2 e−x
2 /(2t)
, the Gaussian density. Therefore
Z ∞
2
√
e−(|x|+|y|) /2
2
2
2
√
E(Xt − x) = EBt + θtψ(x/ t) with ψ(x) :=
(x − y) sgn(y)
dy.
2π
−∞
We compute ψ for x > 0:
Z 0
Z ∞
−(x−y)2 /2
−(x+y)2 /2
2e
2e
√
√
ψ(x) = −
dy +
(x − y)
dy
(x − y)
2π
2π
−∞
0
Z ∞
Z ∞
−z 2 /2
−(x+y)2 /2
2e
2e
√
=−
z √
dz +
(x − y)
dy
2π
2π
x
0
and
Z ∞
Z ∞
Z ∞ −(x+y)2 /2
−(x+y)2 /2
−(x+y)2 /2
e
2e
2e
√
√
(x − y)
dy =
(x + y)
dy − 4x
y √
dy
2π
2π
2π
0
0
0
Z ∞
Z ∞
Z ∞ −(x+y)2 /2
2
−z 2 /2
e−(x+y) /2
e
2e
2
√
=
z √
dz − 4x
(y + x) √
dy + 4x
dy
2π
2π
2π
x
0
0
Z ∞
Z ∞ −z 2 /2
Z ∞ −z 2 /2
−z 2 /2
e
e
2e
2
√
=
z √
dz − 4x
z √
dz + 4x
dz.
2π
2π
2π
x
x
x
15
So for x > 0
ψ(x) = −4x
∞
Z
x
and
2
e−z /2
z √
dz + 4x2
2π
∞
Z
ψ(x) dx = 2
0
Z
∞
x
2
e−z /2
√
dz = 4x(x(1 − Φ(x)) − p(1, x))
2π
√
E[|N |3 ]
2 2
− E[|N |] = − √ .
3
3 π
With the change of variable y → −y, we see that ψ is an odd function. Thus,
Recall now (2.6). Writing (Xt − x) − θLt (X) = Bt ,
R∞
−∞ ψ(x) dx
= 0.
(Xt − x)2 + θ2 Lt (X)2 − 2θ(Xt − x)Lt (X) = Bt2 .
law
Recall that (|X|, L(X)) = (|β|, L(β)), where β is a BM. Moreover, φ2 defined in Lemma 4.7
is symmetric. So
√
√
ELt (X)2 = ELt (β)2 = tφ2 (β0 / t) = tφ2 (X0 / t).
Therefore
√
√
θ
1
tθ
t
E(Xt − x)Lt (X) = ELt (X)2 + (E(Xt − x)2 − EBt2 ) = φ2 (x/ t) + ψ(x/ t)
2
2θ
2
2
and
E[(Xi,n − Xi−1,n )(Li,n (X) − Li−1,n (X)) | Gi−1,n ] =
√
√
1
(θφ2 (Xi−1,n n) + ψ(Xi−1,n n)).
2n
Since φ2 is symmetric and applying (4.1),
λ̄θ (φ2 ) = Leb(φ2 ) = 2
E[|N |3 ]
.
3
Since ψ is anti-symmetric and (4.1)
−
+
+
λ̄θ (ψ) = (1 + θ) Leb(ψ ) + (1 − θ) Leb(ψ ) = 2θ Leb(ψ ) = 4θ
so
E[|N |3 ]
− E[|N |]
3
θφ2 ψ
λ̄θ
+
= θ E[|N |3 ] − 2E[|N |] = 0.
2
2
R
It is straightforward to check that |x|k ( 2θ φ2 (x) + 12 ψ(x)) dx < ∞ for k = 0, 1, 2. With
Lemma 4.1, this proves (4.11).
Lemma 4.13. With Zi,n defined by (4.10), the following convergence holds:
n
√ X
p
n
(Zi,n − E[Zi,n | Gi−1,n ]) −−−→ 0.
n→∞
i=1
16
Proof. We mean to apply Theorem 4.4. We first prove (4.2):
E (Zi,n − E[Zi,n | Gi−1,n ])2 Gi−1,n
≤ E (Xi,n − Xi−1,n )2 (Li,n (X) − Li−1,n (X))2 Gi−1,n
1/2
1/2
≤ E (Xi,n − Xi−1,n )4 Gi−1,n
E (Li,n (X) − Li−1,n (X))4 Gi−1,n
1/2
and we upper bound the two factors. We know E (Xi,n − Xi−1,n )4 |Gi−1,n
≤
C
n.
Recall
law
again that (|X|, L(X)) = (|β|, L(β)), where β is a BM, and that φ4 is symmetric. From
(4.6),
1/2
√
1
E (Li,n (X) − Li−1,n (X))4 Gi−1,n
≤ φ4 (Xi−1,n n)1/2 .
n
Because of (4.5), we apply Lemma 4.1 so that
n
√
C X 1
p
√ φ4 (Xi−1 n)1/2 −−−→ 0.
nE (Zi,n − E[Zi,n | Gi−1,n ])2 Gi−1,n ≤ √
n→∞
n
n
i=1
Since (2.2) has a strong solution, we take B = W , the BM driving (2.2), in (4.3). Since
E[(Wi,n − Wi−1,n ) | Gi−1,n ] = 0,
n
√ X
n
E[(Zi,n − E[Zi,n | Gi−1,n ])(Wi,n − Wi−1,n ) | Gi−1,n ]
i=1
=
n
√ X
n
E[Zi,n (Wi,n − Wi−1,n ) | Gi−1,n ].
i=1
Since Zi,n defined in (4.10) is the product of the increments of X with the ones of the local
time,
1/4
|E[Zi,n (Wi,n − Wi−1,n ) | Gi−1,n ]| ≤ E (Xi,n − Xi−1,n )4 Gi−1,n
h
i
1/2
× E (Wi,n − Wi−1,n )4 Gi−1,n ]1/4 E (Li,n (X) − Li−1,n (X))2 Gi−1,n
.
Now,
1/4
1/4
1
1
E (Wi,n − Wi−1,n )4 Gi−1,n
≤ √ and E (Xi,n − Xi−1,n )4 Gi−1,n
≤√ .
n
n
law
From (4.6), (|X|, L(X)) = (|β|, L(β)), with β BM, and φ2 symmetric
1/2
√
1
E (Li,n (X) − Li−1,n (X))2 Gi−1,n
≤ √ φ2 (Xi−1,n n)1/2 .
n
Therefore
√
1
E[Zi,n (Wi,n − Wi−1,n ) | Gi−1,n ] ≤ √ φ2 (Xi−1,n n)1/2 .
n n
17
From Lemma 4.1,
n
√ X
n
E[(Zi,n − E[Zi,n | Gi−1,n ])(Wi,n − Wi−1,n ) | Gi−1,n ]
i=1
1
≤√
n
n
√
1 X
√
φ2 (Xi−1,n n)1/2
n
!
i=1
p
−−−→ 0.
n→∞
Hence the result.
4.4
Approximation of occupation time
In this section we extend the result in [30], which is proved for diffusions with smooth
coefficients, to the OBM. We consider approximating the occupation time of [0, ∞) up to
time 1:
Z
1
Q+
1 = Leb(s ∈ [0, 1] : Ys ≥ 0) =
0
1{Ys ≥0} ds.
As previously, we suppose that we know the values Yi,n of Y on a grid of time lag 1/n.
Theorem 4.14. Let Y be given in (2.2) and Q̄n1 (Y, +) be given by (3.1). The following
convergence holds:
Z t
√
p
n
n Q̄1 (Y, +) −
1{Ys ≥0} −−−→ 0.
0
n→∞
!
Z
For i = 1, . . . , n, we consider
Ji,n =
1
1
−
n {Yi−1,n ≥0}
i
n
Z
i−1
n
= sgn(Yi−1,n )
1{Ys ≥0} ds
i
n
i−1
n
1{Yi−1,n Ys <0} ds,
(4.12)
Ui,n = Ji,n − E[Ji,n | Gi−1,n ].
The Ui,n are martingale increments. We write
√
Z
n
n Q̄1 (Y, +) −
0
1
1{Ys ≥0}
=
n
n
√ X
√ X
n
E[Ji,n | Gi−1,n ] + n
Ui,n .
1=1
1=1
In the following lemmas we prove the convergence of the two summands.
Remark 4.15. In [30] it is proved that the estimator times n3/4 is tight, so the speed of
convergence proved there, holding only for smooth coefficients, is faster than the speed proved
here. P
We are actually able to prove that n3/4 is the speed of convergence for the martingale
part n1=1 Ui,n also for
Pthe OBM (and other diffusions with discontinuous coefficients), but
not for the drift part n1=1 E[Ji,n | Gi−1,n ]. We would need the central limit theorem giving
the speed of convergence corresponding to the law of large numbers proved in Lemma (4.3),
but this looks quite hard to get in the case of discontinuous diffusion coefficients. Anyways,
for our goal of estimating the parameters of the OBM, the fact that our estimator multiplied
√
with the “diffusive scaling” n converges to 0 is enough. Actually, this result depends on
a compensation between two terms in Lemma 4.16 which holds for this particular diffusion
but for which we do not have results holding for a wider class of SDEs with discontinuous
coefficients.
18
Lemma 4.16. With Ji,n defined by (4.12), the following convergence holds:
n
√ X
p
n
E[Ji,n | Gi−1,n ] −−−→ 0.
n→∞
1=1
Proof. From (4.12),
#
"Z i
n
√
nE
sgn(Yi−1,n )1{Yi−1,n Ys <0} ds Gi−1,n
i−1
n
√
= n
Z
i
n
i−1
n
h
i
sgn(Yi−1,n )E 1{Yi−1,n Ys <0} Gi−1,n ds
=
√
i
n
Z
n
i−1
n
sgn(Yi−1,n )P[Yi−1,n Ys < 0 | Gi−1,n ] ds.
Using the Markov property and (2.3) we can compute this quantity. When Y0 > 0
√
Z
n
1
n
1
n
√
2σ−
Φ(Y0 /(σ− s)) ds
0 σ− + σ+
√
Z 1
√ dt
√
√
n2σ−
1
=
Φ(Y0 n/(σ− t))
= √ f (Y0 n),
σ− + σ+ 0
n
n
Z
sgn(Y0 )P (Y0 Ys < 0) ds =
0
where
f (x) :=
2σ+
σ− + σ+
1
Z
√
Φ(−x/(σ+ t)) dt for x > 0.
0
Now, for Y0 < 0, we find
√
n
Z
0
where
and
√
1
sgn(Y0 )P (Y0 Ys < 0) ds = √ f (Y0 n),
n
−2σ−
f (x) :=
σ− + σ+
We can compute
Z ∞
Z
f (x) dx =
0
1
n
0
∞Z 1
0
Z
1
√
Φ(x/(σ− t)) dt for x < 0.
0
2
√
2σ+
2σ+
Φ(−x/(σ+ s)) ds dx =
σ− + σ+
σ− + σ+
2
−2σ−
f (x) dx =
σ− + σ+
−∞
Z
0
Z
0
−∞
Z
1
Z
∞Z 1
0
√
Φ(−x/ s) ds dx
0
√
Φ(x/ s) ds dx.
0
Therefore computing the coefficient in (4.3) we find
Z
Z
2
2
λσ (f ) = 2
f+ + 2
f − = 0.
σ+
σ−
Moreover,
|f (x)| ≤ C
for some constant C, so
statement.
R
Z
0
1
√
2
Φ(Cx/ t)) dt ≤ Ce−Cx
|x|k f (x) dx < ∞ for all k ≥ 0. Applying Lemma 4.3, we prove the
19
Lemma 4.17. With Ui,n defined by (4.12), the following convergence holds:
n
√ X
p
n
Ui,n −−−→ 0.
n→∞
i=1
Proof. We consider (4.2). We have
2
2
2
E Ui,n
Gi−1,n = E Ji,n
Gi−1,n − E[Ji,n | Gi−1,n ]2 ≤ E Ji,n
Gi−1,n .
From (4.12),
#
"Z i
1
2
n
1{(Yi−1,n Ys )<0} ds Gi−1,n
E Ji,n Gi−1,n ≤ E
i−1
n
n
Z i
√
1 n
1
=
P[Yi−1,n Ys < 0 | Gi−1,n ] ds = 2 |f (Yi−1,n n)|.
n i−1
n
n
Therefore, in (4.3),
2
λσ (|f |) = 2
σ+
Z
2
f − 2
σ−
+
Z
f
−
4
=
σ− + σ+
Z
∞
Z
−∞
1
√
Φ(x/ s) ds dx.
0
Lemma 4.3 implies
n
n
n
X
2
X
√
1
p
Gi−1,n =
E Ji,n
f (Yi−1,n n) −−−→ 0.
n→∞
n
i=1
i=1
Now we check (4.3). Let W be the BM driving (2.2).
2
1/2
1/2
Gi−1,n
|E[Ui,n (Wi,n − Wi−1,n ) | Gi−1,n ]| ≤ E Ui,n
E (Wi,n − Wi−1,n )2 Gi−1,n
q
√
1
1
≤
|f (Y0 n)| √
n
n
and
√
n|
[nt]
X
i=1
[nt]
1 X 1
√
E[Ui,n (Wi,n − Wi−1,n )|Gi−1,n ]| ≤ √
n
n
i=1
q
√
p
|f (Y0 n)| −−−→ 0
n→∞
from another application of Lemma 4.3. The condition of integrability is satisfied because
p
2
|f (x)| ≤ Ce−Cx ,
p
R
for some constant C, so |x|k |f (x)| dx < ∞ for all k ≥ 0.
4.5
Proof of the main results
In this section we use all the results proved so far to prove Theorems 3.1 and 3.5.
20
Proof of Theorem 3.1. We set
Z t
Z t
1(Ys < 0) dWs .
1(Ys > 0) dWs and ηt = σ−
ξt = σ+
0
0
Itô-Tanaka’s formula (see [33]) gives the following equation for the positive and negative part
of Y :
1
1
Yt+ = ξt + Lt (Y ), and Yt− = −ηt + Lt (Y ).
(4.13)
2
2
Moreover, ξ is a martingale with quadratic variation
Z t
Z t
2
1(Ys > 0) ds.
σ(Ys )2 1(Ys > 0) ds = σ+
hξit =
0
0
It is well known that the quadratic variation of a martingale can be approximated with the
sum of squared increments over shrinking partitions. Thus,
Z 1
p
n
2
2 +
[ξ]1 −−−→ hξi1 =
σ+
1(Ys > 0) ds = σ+
Q1 (Y ).
(4.14)
n→∞
0
From (4.13),
[L(Y )]n1
+ [Y + , L(Y )]n1 .
4
The local time L(Y ) is of finite variation, Y + is continuous. Thus [L(Y )]n1 as well as
[Y + , L(Y )]n1 converge to 0 almost surely. From (4.14),
Z 1
p
+ n
2
2 +
[Y ]1 −−−→
σ+
1(Ys > 0) ds = σ+
Q1 (Y ).
(4.15)
[Y + ]n1 = [ξ]n1 −
n→∞
0
Recall the definition of Q̄n1 (Y, +) in (3.1). Then
Z 1
a.s.
n
Q̄1 (Y, +) −−−→
1(Ys ≥ 0) ds = Q+
1 (Y ).
n→∞
0
p
p
n −
n −
n , σ̂ n )
From (4.15) and (4.5), σ̂+
→ σ+ , and similarly σ̂−
→ σ− . Therefore, the vector (σ̂+
−
n
n
converges in probability to (σ+ , σ− ). The estimator (σ̂+ , σ̂− ) is then consistent.
We consider now the rate of convergence. From (4.13) applied to Y − , we have as in (4.5)
that
+ n n +
[Y ]1
[ξ]1
[Y , L(Y )]n1
1 [L(Y )]n1
=
+
−
.
[Y − ]n1
[η]n1
[Y − , L(Y )]n1
1
4
We consider separately the tree summands. From the central limit theorem for martingales
(see for example [15], (5.4.3) or Theorem 5.4.2), since 1(Ys > 0)1(Ys < 0) = 0,
√ Z 1 σ 2 1(Ys > 0)
√
[ξ]n1
[ξ, η]n1
hξi1
0
0
sl
+
n
−
−−−→ 2
2 1(Y < 0) dB̄s ,
n→∞
[η, ξ]n1
[η]n1
0
hηi1
0
σ−
s
0
where B̄ is a Brownian motion independent of the filtration of W . Therefore it is also
independent of L(Y ). Consider now the second summand. The OBM Y is linked to a SBM
X solution to (2.6) through Yt = Xt σ(Xt ). With (2.8) and (4.9) in Lemma 4.11,
√
n[Y + , L(Y )]n1 =
√
n
2σ
2σ+
p
−
[X + , L(X)]n1 −−−→ 0.
n→∞
σ+ + σ−
21
Clearly this also holds for [Y − , L(Y )]n1 , and we obtain the convergence in probability of
√
n([Y + , L(Y )]n1 , [Y − , L(Y )]n1 ) to (0, 0).
We use Lemma 4.6 for dealing with the third summand:
√
√
2σ+ σ− 2
n
n[L(Y )]1 = n
[L(X)]n1
σ+ + σ−
√
√
4 2
2σ+ σ− 2
2σ+ σ−
4 2
p
−−−→ √
L1 (X) = √
L1 (Y ). (4.16)
n→∞ 3 π
σ+ + σ−
3 π σ+ + σ−
We obtain, using (3.1),
+ n
√
[Y ]1
hξi1
n
−
[Y − ]n1
hηi1 )
!
√ R1 2
2 0 σ+ 1(Ys > 0) dB̄s
sl
−−−→ √ R 1 2
−
n→∞
2 0 σ− 1(Ys < 0) dB̄s
1
1
! √
2σ+ σ−
2
√
L1 (Y ). (4.17)
3 π σ+ + σ−
We write now
2 Q̄n (Y, +)
[Y + ]n1 − σ+
1
n (Y, +)
Q̄
1
=
2 Q̄n (Y, −)
[Y − ]n1 − σ−
1
Q̄n1 (Y, −)
+n
2 (Q+ − Q̄n (Y, +))
σ+
[Y ]1 − hξi1
1
1
Q̄n (Y, +)
Q̄n1 (Y, +)
.
1
= −n
+ 2
+
n
[Y ]1 − hηi1
σ− (1 − Q1 − Q̄1 (Y, −))
Q̄n1 (Y, −)
Q̄n1 (Y, −)
2
n )2 − σ
(σ̂+
+
n )2 − σ 2
(σ̂−
−
−
+
Recall that Q̄n1 (Y, +) and Q̄n1 (Y, −) converge almost surely to Q+
1 (Y ) and Q1 (Y ) = 1 − Q1 .
+
Besides, 0 < Q1 < 1 a.s., because Y0 = 0. Therefore, from Theorem 4.14,
2 (Q+ − Q̄n (Y, +))
σ+
1
1
n
p
√
0
Q̄1 (Y, +)
.
n 2
−−−→
+
n
σ− (1 − Q1 − Q̄1 (Y, −)) n→∞ 0
Q̄n1 (Y, −)
Using again (3.1) and (4.17),
√
2 R
2σ+
1
1
√
n 2
2
+
0 1(Ys > 0) dB̄s
√
2
2σ
σ
σ̂ ) − σ+
sl
√Q+
Q
+
−
1
1
√
n
−−−→
L1 (Y ).
−
2
2 R
n→∞
σ̂ n )2 − σ−
2σ−
1 3 π σ+ + σ−
1
0 1(Ys < 0) dB̄s
1 − Q+
1
1 − Q+
1
(4.18)
The statement is now proved, but we would like to get a more explicit expression for the law
+
of the limit random variable. Recall Q+
t (Y ) = Qt (X). From Corollary 1.2 in [2], standard
computations give that the joint density of (Lt (X), Q+
t ) is, for b > 0, τ ∈ [0, t]:
(1 − θ2 )b
(1 + θ)2 b2 (1 − θ)2 b2
pLt (X),Q+ (b, τ ) =
exp
−
−
.
t
8τ
8(t − τ )
4πτ 3/2 (t − τ )3/2
22
We set now
Lt (X)
Zt =
4
s
(1 + θ)2 (1 − θ)2
+
.
Q+
t − Q+
t
t
Changing variable in the integration, the joint density of (Zt , Q+
t ) is
2
x
1 − θ2
1
pZt ,Q+ (x, τ ) = x exp −
.
t
2
πτ 1/2 (t − τ )1/2 (1 + θ)2 (t − τ ) + (1 − θ)2 τ
We also find the joint density of (Zt , t − Q+
t ) as
2
x
1
1 − θ2
.
pZt ,t−Q+ (x, τ ) = x exp −
t
2
πτ 1/2 (t − τ )1/2 (1 + θ)2 τ + (1 − θ)2 (t − τ )
(4.19)
(4.20)
As we can factorize pZt ,Q+ (x, τ ) = pZt (x)pQ+ (τ ), Zt and Q+
t are independent and their laws
t
t
2
are explicit. In particular from (2.5), for t = 1, pZ1 (x) = x exp(− x2 ),
1 − θ2
1
×
πτ 1/2 (1 − τ )1/2 (1 + θ)2 (t − τ ) + (1 − θ)2 τ
1
σ+ /σ−
=
×
,
1/2
1/2
1 − (1 − (σ+ /σ− )2 )τ
πτ (1 − τ )
pQ+ (τ ) =
1
and
1
1 − θ2
×
πτ 1/2 (1 − τ )1/2 (1 + θ)2 τ + (1 − θ)2 (1 − τ )
σ− /σ+
1
×
=
.
1/2
1/2
1 − (1 − (σ− /σ+ )2 )τ
πτ (1 − τ )
p1−Q+ (τ ) =
1
Let now Λ be a random variable with the same law of Q+
1 , and let ξ be an independent
exponential random variable of parameter 1. From (4.19), (4.20)
q
q
1 !
1−Λ
1−Λ
+
2(σ
+
σ
)ξ
4ξ
law
+
−
Q1
q Λ = q
q Λ .
L1 (X) = p
1
2 + Λ(1 − θ)2
Λ
Λ
2
2
+
(1
−
Λ)(1
+
θ)
1−Q1
(1 − Λ)σ− + Λσ+
1−Λ
1−Λ
Moreover,
√
2
2σ+
√Q+
2σ1 2
−
1−Q+
1
√ 2
2σ+
1(Y
>
0)
d
B̄
√
N1
s
s law
0
,
= √2σΛ2
R1
−
√
N
0 1(Ys < 0) dB̄s
1−Λ 2
R1
where N1 , N2 are standard Gaussian random variables independent of ξ and Λ. Therefore
the limit law has the expression given in the statement.
Proof of Theorem 3.5. Using (4.13), we obtain
n
1
1
+
n
[Y , Y ]1 = ξ + L(Y ), Y
= [ξ, Y ]n1 + [L(Y ), Y ]n1 .
2
2
1
23
From the Central Limit Theorem for martingales [15, Theorem 5.4.2] and ξt + ηt = Yt ,
√
√
[ξ, Y ]n1
[ξ, ξ]n1 − [ξ, η]n1
n
= n
[η, Y ]n1
[η, ξ]n1 − [η, η]n1
√ Z 1 σ 2 1(Ys > 0)
√ Z 1 σ 2 1(Ys > 0)
sl
+
+
2
−−−→ 2
2 1(Y < 0) dB̄s =
2 1(Y < 0) dB̃s ,
n→∞
−σ−
σ−
s
s
0
0
where B̃ is another BM independent of the filtration of W . Both W and B̃ are defined on an
extension (Ω̃, F̃, P̃) of (Ω, F, P) with P0 = P ⊗ P0 where P0 carries the BM B̃. Moreover,
[L(Y ), Y ]n1 = [L(Y ), Y + ]n1 − [L(Y ), Y − ]n1 =
so
√
2σ
2
2σ+
2σ+ σ−
−
[X + , L(X)]n1 −
[X − , L(X)]n1
σ+ + σ−
σ+ + σ−
p
n[L(Y ), Y ] −
→ 0 because of Lemma 4.11. Finally,
+
√ Z 1 σ 2 1(Ys > 0)
√
[Y , Y ]n1
+
n
= 2
2 1(Y < 0) dB̄s .
[Y − , Y ]n1
σ−
s
0
This is the analogous of (4.17) in the proof of Theorem 3.1. From now on the proof follows
as in Theorem 3.1, but without the local time part.
Remark 4.18. We look for the origin of the asymptotic bias present in σ̂ n , but not in mn .
Consider first the difference between the approximation of quadratic variation used in the
two different estimators:
[Y + , Y ]n1 = [Y + , Y + ]n1 − [Y + , Y − ]n1 = [Y + ]n1 − [Y + , Y − ]n1 .
From (4.13),
1
1
1
[Y + , Y − ]n1 = −[ξ, η]n1 + [ξ, L(Y )]n1 − [L(Y ), η]n1 + [L(Y )]n1
2
2
4
1 +
1
1
n
= −[ξ, η]1 + [Y , L(Y )]n1 + [L(Y ), Y − ]n1 − [L(Y )]n1 .
2
2
4
From the central limit theorem for martingales [15, Theorem 5.4.2],
√
Since
√
sl
n[ξ, η]n1 −−−→
n→∞
√ Z
2
1
σ+ 1(Ys > 0)σ− 1(Ys < 0) dB̄s = 0.
0
n[Y ± , L(Y )]n1 converges in probability to 0, using (4.16) we obtain
√
√
2 2 σ+ σ−
sl
+
− n
n[Y , Y ]1 −−−→ √
L1 (Y ).
n→∞ 3 π σ+ + σ−
2 is related to the bracket [Y + , Y − ]n .
We then see that the asymptotic bias in σ̂±
1
24
(4.21)
4.6
Proof of Corollary 3.8: adding a drift term via Girsanov Theorem
Let us start with a remark on the stability of the stable convergence under a Girsanov
transform.
Lemma 4.19. For two probability spaces (Ω, F, P) and (Ω0 , F 0 , P0 ), let us define an extension
(Ω̃, F̃, P̃) by of (Ω, F, P) of the form
Ω̃ = Ω × Ω0 , F̃ = F ⊗ F 0 and P̃ = P ⊗ P0 .
Assume that (Ω, F, P) and (Ω0 , F 0 , P0 ) carry respectively Brownian motions W and W 0 with
natural (completed) filtrations F = (Ft )t>0 and F 0 = (Ft0 )t>0 . Assume also that W and W 0
are independent.
On (Ω, F, P), let G be an exponential F -martingale which is uniformly integrable. Let Q
= Gt .
be the measure such that dQ
dP
Ft
Suppose now that a sequence Zn on (Ω, F, P) of FT -measurable random variables converges
stably to a random variable Z = AWB0 on the extension (Ω̃, F̃, P̃) of (Ω, F, P) where A and B
are FT -random variables on (Ω, F, P).
Then Zn converges stably to Z = AWB0 on (Ω̃, F̃, Q ⊗ P0 ) where W 0 is a Brownian motion
independent from A and B (the laws of A and B are of course changed).
Proof. Let us write Q̃ = Q ⊗ P0 . The Girsanov weight GT is FT -measurable and integrable
with respect to P. Hence, it is easily shown that for any bounded, FT -measurable random
variable Y and any bounded, continuous function f , E[GY f (Zn )] converges to EP̃ [GY f (Z)].
Under Q̃, hW 0 , W i = 0 as W and W 0 are independent and the bracket does not change
under a Girsanov transform. This implies that W 0 is still a Brownian motion under Q̃. Hence
the result.
Proof of Corollary 3.8. Let ξ be solution to dξt = σ(ξt ) dWt with an underlying Brownian
motion W on (Ω, F, P). We denote by (Gt )t≥0 the filtration of W . Thus, ξ is an OBM.
The Girsanov theorem is still valid for discontinuous coefficients [20]. Let us set
!
Z t
Z
b(ξs )
1 t b(ξs ) 2
Gt = exp
dWs −
ds .
2 0 σ(ξs )
0 σ(ξs )
dQ
dP G
t
Since b is bounded, we define a new measure Q by
= Gt . Under Q, the process ξ is
Rt
solution to dξt = σ(ξt ) dW̃t + b(ξt ) dt for a Brownian motion W̃t = Wt − 0 b(ξs )σ(ξs )−1 ds,
t ≥ 0.
Theorems 3.1 and 3.5 hold for ξ under P. Therefore, Lemma 4.19 applies here. Thus,
√ 2
2σ+ R 1
n
2
+
√
(m+ (ξ))2 − σ+
0 1(ξs > 0) dB̄s
Q̃-sl
√1 (ξ)
n
−−−→ Q
,
2
n
2
2
2σ− R 1
(m− (ξ)) − σ− n→∞
1(ξ
<
0)
d
B̄
s
s
+
0
1−Q (ξ)
1
where B̄ is a BM independent of W and W̃ also under Q̃.
25
5
Oscillating Random Walk
In [18] the OBM is constructed also as a limit of discrete processes, called Oscillating Random
Walks (ORW), analogously to how the BM is constructed as a limit of Random Walks. The
aim of this section is to examplify the phenomena of dependence on the occupation of the
limit law, in a simpler framework and with a non-technical proof.
We define the ORW as the following process. Fix 0 < p, q ≤ 1. For k ∈ N, we introduce
the following random variables:
q
Uk iid, P(Uk = 1) = P(Uk = −1) = , P(Uk = 0) = 1 − q,
2
p
Vk iid, P(Vk = 1) = P(Vk = −1) = , P(Vk = 0) = 1 − p,
2
1
Zk iid, P(Zk = 1) = P(Zk = −1) = .
2
Now we set Y0∗ = 0 and
∗
Yk+1
∗
Yk + Uk+1
= Yk∗ + Vk+1
∗
Yk + Zk+1
if Yk∗ > 0,
if Yk∗ < 0,
if Yk∗ = 0.
We consider the re-normalized process
∗
Ytn = n−1/2 Y[nt]
.
For all K > 0, we have the following convergence:
p
sup |Ytn − Yt | −−−→ 0.
n→∞
0≤t≤K
The convergence in probability holds if the processes Y n are constructed as in [34], and Y is an
2 = q, σ 2 = p. This means that in this setting we have 0 < σ , σ ≤ 1,
OBM of parameters σ+
− +
−
but we do not loose in generality because we can always re-scale time and space.
In this appendix, we recover from the observations of Y n for some large n an estimator
for the parameters of the OBM.
n > 0}, αn = #{k ∈ N, k ≤ n : Y n > 0, Y n
We set β n = #{k ∈ N, k ≤ n : Yk/n
k/n
(k+1)/n 6=
n
2
Yk/n }, and introduce the following estimator of q = σ+ :
q̂ n =
αn
.
βn
(5.1)
Theorem 5.1. Let q̂ n be the estimator defined above. The following convergence holds:
r
√
q(1 − q)
law
n
n (q̂ − q) −−−→ N
,
n→∞
Λ
where Λ follows the law in (2.5), N is a standard Gaussian and they are independent.
n
n
n
n
n
Proof. When Yk/n
> 0, Y(k+1)/n
6= Yk/n
with probability q, and Y(k+1)/n
= Yk/n
with
probability 1 − q. We can compute the log-likelihood and maximize it as in the statistics of
26
Binomial variables, finding that the maximum likelihood estimator for q is q̂ n in (5.1). In [18]
it is proved that
#{k ≤ n : Yk ≥ 0} law
−−−→ Λ,
n→∞
n
where Λ follows the law in (2.5). This easily implies
β n law
−−−→ Λ.
n n→∞
(5.2)
Conditioning to β n , we have that αn follows is a binomial distribution with parameters q, β n .
We write the event
s
)
(s
n
n
√
β
β
.
n (q̂ n − q) ≤ x =
(q̂ n − q) ≤ x
q(1 − q)
nq(1 − q)
From Berry-Essen inequality [3, 7], we have
s
√
P n(q̂ n − q) ≤ x β n − Φ x
βn
nq(1 − q)
!
p
≤ Cq / β n ,
for some constant Cq . Now, from (5.2) and Portmanteau Lemma,
s
s
"
!#
!#
"
βn
Λ
E Φ x
−−−→ E Φ x
.
n→∞
nq(1 − q)
q(1 − q)
√
Moreover, E[Cq / β n ] → 0. Recalling
√
√
P n(q̂ n − q) ≤ x = E P n (q̂ n − q) ≤ x β n ,
we obtain the following convergence
s
"
√
P n (q̂ n − q) ≤ x −−−→ E Φ x
n→∞
Λ
q(1 − q)
!#
,
which implies the statement.
6
Empirical evidence
In this section we implement estimators σ̂ n , mn and use them on simulated data. For doing
so, we reduce the OBM (2.2) to a SBM (2.6), and we simulate it through the simulation
method given in [23]. This method gives the successive positions {Xk∆t }k≥0 of the SBM,
hence the OBM, on a time grid of size ∆t.
n,T
Recall Remark 3.7, in particular estimators mn,T
± , σ̂± , for which we have central limit
theorems with the same limit laws of (3.4), (3.6). We use the parameters: T = 5, ∆t = 0.01
(thus n = 500), σ− = 0.5, σ+ = 2 (so that θ = −0.48 in (2.6), pushing the process to the
negative side). In Figure 1, we plot the density of
√
√
n,T 2
2
2
n
2
M±n := n((mn,T
n((σ̂±
) − σ±
)
± ) − σ± ) and S± :=
27
for N realizations of these estimators (meaning the simulation of N = 10 000 paths of the
OBM). Their empirical densities are compared with the ones of
√
√
2σ 2
8
1 √ ξ 1−Λ
√ 2
√ + N1 − √
3 π r+1 (1−Λ)+Λr2
2σ N
Λ
S+
,
√ 2
M± := √ ± and
:=
√
2σ
S−
1 √ ξ Λ
Λ
√ −
N2 − √8 1/r+1
2
3 π
1−Λ
Λ+(1−Λ)/r
given in (3.6) and (3.4), with N ∼ N (0, 1) and ξ ∼ exp(1). The densities of M and S (which
do not depend on T ) are obtained by simulation. The occupation time Λ is simulated by
inverting its distribution function [19, 25]:
2V
σ−
law
2 Uπ
Λ = 2
, U uniform on [0, 1).
with
V
=
sin
2 (1 − V )
2
σ− V + σ+
Negative side
Biased
Biased vs Unbiased
1
1
1
0.8
0.8
0.8
0.6
0.6
0.6
0.4
0.4
0.4
0.2
0.2
0.2
0
−2
−1
(a)
−2
6
Positive side
Unbiased
·10
0
n
S−
1
0
vs S−
6
−2
·10
(b)
−2
−1
0
n
M−
1
0
vs M−
6
4
4
2
2
2
(d)
0
n
S+
20
vs S+
·10
−1
0
n
S−
vs
1
n
M−
0
0
−20
(c)
−2
4
0
−2
−20
0
(e)
n
M+
20
vs M+
−20
(f)
0
n
S+
vs
20
n
M+
Figure 1: Densities of:
n
(a), (d) normalized error of the estimators M±
(solid line) and theoretical limits M± (dashed lines);
n
(b), (e) normalized error of the estimators S± (solid line) and theoretical limits S± (dashed lines);
n
n
(c), (f) normalized error of the estimators M±
and normalized error of the estimators S±
.
We see that the limit distribution on the positive side has a larger variance than the
one in the negative side. This is due to the sample size, proportional to the occupation
time, which is on average larger on the side where the diffusion coefficient is smaller. We
also obtain a good agreement of the normalized empirical error with the prediction of the
28
central limit theorem. On the pictures on the right, we observe the difference between biased
and non-biased estimator; the grey area is the effect given by the shift to the left of the
distribution, caused by the local time term. This shift is more visible when the diffusion
coefficient is smaller.
√ [Y +,Y − ]nT
)
We have also checked that n
has a distribution close to the one of LT√(Y
=
T
T
2
√
2σ+ σ−
√2
L1 (X), which is straightforward since the density of L1 (X) is known (for this,
3 π σ+ +σ−
we use P. Lévy’s identity which relates the local time to the supremum of the Brownian
motion whose density is explicitly known). The agreement is good.
Finally, the same simulation work can be done using the random walks defined in Section 5
using the simple approach. Again, the numerical results are in good agreements with the
theory, although some instabilities appear due to the fact that the occupation time may take
small values with high probability.
References
[1] R. Altmeyer and J. Chorowski. Estimation error for occupation time functionals of
stationary Markov processes. ArXiv e-prints, October 2016.
[2] T. Appuhamillage, V. Bokil, E. Thomann, E. Waymire, and B. Wood. Occupation and
local times for skew Brownian motion with applications to dispersion across an interface.
Ann. Appl. Probab., 21(1):183–214, 2011.
[3] Andrew C. Berry. The accuracy of the Gaussian approximation to the sum of independent
variates. Trans. Amer. Math. Soc., 49:122–136, 1941.
[4] R S Cantrell and C Cosner. Diffusion Models for Population Dynamics Incorporating
Individual Behavior at Boundaries: Applications to Refuge Design. Theoretical Population
Biology, 55(2):189–207, 1999.
[5] R.S. Cantrell and C. Cosner. Skew Brownian motion: a model for diffusion with
interfaces? In Proceedings of the International Conference on Mathematical Models in
the Medical and Health Sciences, pages 73–78. Vanderbilt University Press, 1998.
[6] Marc Decamps, Marc Goovaerts, and Wim Schoutens. Self exciting threshold interest
rates models. Int. J. Theor. Appl. Finance, 9(7):1093–1122, 2006.
[7] Carl-Gustav Esseen. On the Liapounoff limit of error in the theory of probability. Ark.
Mat. Astr. Fys., 28A(9):19, 1942.
[8] Pierre Étoré. On random walk simulation of one-dimensional diffusion processes with
discontinuous coefficients. Electron. J. Probab., 11:no. 9, 249–275 (electronic), 2006.
[9] Daniéle Florens-Zmirou. On estimating the diffusion coefficient from discrete observations.
J. Appl. Probab., 30(4):790–804, 1993.
[10] J. M. Harrison and L. A. Shepp. On skew Brownian motion. Ann. Probab., 9(2):309–313,
1981.
29
[11] R. Höpfner and E. Löcherbach. Limit theorems for null recurrent Markov processes.
Mem. Amer. Math. Soc., 161(768), 2003.
[12] Reinhard Höpfner. Asymptotic statistics. De Gruyter Graduate. De Gruyter, Berlin,
2014. With a view to stochastic processes.
[13] Jean Jacod. On continuous conditional Gaussian martingales and stable convergence in
law. In Séminaire de Probabilités, XXXI, volume 1655 of Lecture Notes in Math., pages
232–246. Springer, Berlin, 1997.
[14] Jean Jacod. Rates of convergence to the local time of a diffusion. Ann. Inst. H. Poincaré
Probab. Statist., 34(4):505–544, 1998.
[15] Jean Jacod and Philip Protter. Discretization of processes, volume 67 of Stochastic
Modelling and Applied Probability. Springer, Heidelberg, 2012.
[16] Jean Jacod and Albert N. Shiryaev. Limit theorems for stochastic processes, volume 288
of Grundlehren der Mathematischen Wissenschaften. Springer-Verlag, Berlin, second
edition, 2003.
[17] Yuji Kasahara and Yuko Yano. On a generalized arc-sine law for one-dimensional diffusion
processes. Osaka J. Math., 42(1):1–10, 2005.
[18] Julian Keilson and Jon A. Wellner. Oscillating Brownian motion. J. Appl. Probability,
15(2):300–310, 1978.
[19] John Lamperti. An occupation time theorem for a class of stochastic processes. Trans.
Amer. Math. Soc., 88:380–387, 1958.
[20] J.-F. Le Gall. One-dimensional stochastic differential equations involving the local times
of the unknown process. Stochastic Analysis. Lecture Notes Math., 1095:51–82, 1985.
[21] A. Lejay, Lenôtre, and G. Pichot. One-dimensional skew diffusions: explicit expressions
of densities and resolvent kernel, 2015. Preprint.
[22] A. Lejay, E. Mordecki, and S. Torres. Convergence of estimators for the skew Brownian
motion with application to maximum likelihood estimation, 2017. In preparation.
[23] A. Lejay and G. Pichot. Simulating diffusion processes in discontinuous media: a
numerical scheme with constant time steps. Journal of Computational Physics, 231:7299–
7314, 2012.
[24] Antoine Lejay. On the constructions of the skew Brownian motion. Probab. Surv.,
3:413–466, 2006.
[25] Antoine Lejay. Simulation of a stochastic process in a discontinuous layered medium.
Electron. Commun. Probab., 16:764–774, 2011.
[26] Antoine Lejay and Miguel Martinez. A scheme for simulating one-dimensional diffusion
processes with discontinuous coefficients. Ann. Appl. Probab., 16(1):107–139, 2006.
[27] Paul Lévy. Sur certains processus stochastiques homogènes. Compositio Math., 7:283–339,
1939.
30
[28] Alex Lipton and Artur Sepp. Filling the gaps. Risk Magazine, pages 66–71, 2011-10.
[29] S. Mazzonetto. On the Exact Simulation of (Skew) Brownian Diffusion with Discontinuous
Drift. Phd thesis, Postdam University & Université Lille 1, 2016.
[30] Hoang-Long Ngo and Shigeyoshi Ogawa. On the discrete approximation of occupation
time of diffusion processes. Electron. J. Stat., 5:1374–1393, 2011.
[31] J. M. Ramirez, E. A. Thomann, and E. C. Waymire. Advection–dispersion across
interfaces. Statist. Sci., 28(4):487–509, 2013.
[32] Alfréd Rényi. On stable sequences of events. Sankhyā Ser. A, 25:293 302, 1963.
[33] Daniel Revuz and Marc Yor. Continuous martingales and Brownian motion, volume 293
of Grundlehren der Mathematischen Wissenschaften. Springer-Verlag, Berlin, 3 edition,
1999.
[34] Charles Stone. Limit theorems for random walks, birth and death processes, and diffusion
processes. Illinois J. Math., 7:638–660, 1963.
[35] Lajos Takács. On the local time of the Brownian motion. Ann. Appl. Probab., 5(3):741–
756, 1995.
[36] J. B. Walsh. A diffusion with discontinuous local time. In Temps locaux, volume 52-53,
pages 37–45. Société Mathématique de France, 1978.
[37] Shinzo Watanabe. Generalized arc-sine laws for one-dimensional diffusion processes and
random walks. In Stochastic analysis (Ithaca, NY, 1993), volume 57 of Proc. Sympos.
Pure Math., pages 157–172. Amer. Math. Soc., Providence, RI, 1995.
[38] Shinzo Watanabe, Kouji Yano, and Yuko Yano. A density formula for the law of time
spent on the positive side of one-dimensional diffusion processes. J. Math. Kyoto Univ.,
45(4):781–806, 2005.
31
| 10math.ST
|
Combining Monte-Carlo and Hyper-heuristic methods for the
Multi-mode Resource-constrained Multi-project Scheduling
Problem
arXiv:1511.04387v2 [cs.DS] 8 Sep 2016
Shahriar Astaa , Daniel Karapetyana,b,∗, Ahmed Kheiria , Ender Özcana ,
Andrew J. Parkesa
a University of Nottingham, School of Computer Science
Jubilee Campus, Wollaton Road, Nottingham, NG8 1BB, UK
b University of Essex, Institute for Analytics and Data Science
Wivenhoe Park, Colchester, CO4 3SQ, UK
Abstract
Multi-mode resource and precedence-constrained project scheduling is a well-known
challenging real-world optimisation problem. An important variant of the problem requires scheduling of activities for multiple projects considering availability of local and
global resources while respecting a range of constraints. A critical aspect of the benchmarks addressed in this paper is that the primary objective is to minimise the sum of the
project completion times, with the usual makespan minimisation as a secondary objective.
We observe that this leads to an expected different overall structure of good solutions
and discuss the effects this has on the algorithm design. This paper presents a carefully
designed hybrid of Monte-Carlo tree search, novel neighbourhood moves, memetic algorithms, and hyper-heuristic methods. The implementation is also engineered to increase
the speed with which iterations are performed, and to exploit the computing power of
multicore machines. Empirical evaluation shows that the resulting information-sharing
multi-component algorithm significantly outperforms other solvers on a set of “hidden”
instances, i.e. instances not available at the algorithm design phase.
Keywords: metaheuristics; hybrid heuristics; hyper-heuristics; Monte Carlo tree search;
permutation based local search; multi-project scheduling
1. Introduction
Project scheduling has been of long-standing interest to academics as well as practitioners. Solving such a problem requires scheduling of interrelated activities (jobs),
potentially each using or sharing scarce resources, subject to a set of constraints, and
with one or several of a variety of objective functions. There are various project scheduling problems and many relevant surveys in the literature, e.g. see [4, 22, 21, 48, 39, 20, 65].
The best-known problem class is the Resource Constrained Project Scheduling Problem
∗ Corresponding
author
Preprint submitted to Elsevier
September 9, 2016
(RCPSP) in which activities have fixed usages of the resources, there are fixed precedence constraints between them, and often the objective is simple minimisation of the
makespan (completion time of last activity). These problems have been proven to be
NP-hard [2], and a well-known benchmark suite, PSPLIB, is provided in [29].
A generalisation of the RCPSP is to also consider ‘Multi-mode RCPSP’ (MRCPSP)
in which activities can be undertaken in one of a set of modes, with each mode potentially
using different sets of resources. Furthermore, there are many options besides makespan
for the objective function(s); a typical one is that a weighted sum of completion times
is minimised. As common in optimisation problems, exact methods perform best on
smaller instances and on larger instances heuristics and metaheuristics become necessary.
Recent works on the MRCPSP range from exact approaches, such as, MILP [33], and
branch-and-bound [59], to metaheuristics, such as, differential evolution [11], estimation
of distribution algorithms [61], evolutionary algorithms [14, 55, 17], swarm intelligence
methods [31], and others [8, 60].
This paper presents our winning approach submitted to MISTA 2013 challenge1 on a
further extension called ‘multi-mode resource-constrained multi-project scheduling’ (MRCMPSP) and the results on the associated benchmark/competition instances. The full
description of this problem domain can be found on the competition website and in
[62]; however, for completeness we also summarise it in Section 2. The broad aim is to
schedule a set of different and partially interacting projects, with each project consisting of a set of activities. There are no precedence constraints between the activities of
different projects however they can compete for resources. Also, the objective function
is extended to be a mix of a kind of weighted completion time and makespan. The
MRCMPSP is hence interesting in that it has a mix of structures and requirements that
are a step towards modelling the complexity of real-world scheduling problems. The
high real-world relevance of the multi-project version of scheduling is well-known, e.g. a
survey [35] found that “84% of the companies which responded to the survey indicated
that they worked with multiple projects”. However, the majority of scheduling work is
on the single project version, though there is some existing work on the multi-project
case, e.g. see [35, 18, 32, 36].
Our approach searches the space of sequences of activities, from which schedules are
constructed and then the quality of each schedule is evaluated using the objective function. The search process on the set of sequences operates in two phases in a “construct
and improve” fashion. In the first phase, a heuristic constructor creates initial sequences
of activities. A novel proposal in this paper is to investigate the overall global structure
of the solutions and use this to motivate constructing the initial sequences using a MonteCarlo Tree Search (MCTS) method, e.g. see [3]. This construction phase is followed by
an improvement phase which makes use of a large and diverse set of heuristic neighbourhood moves. The search process during the improvement phase is carefully controlled
by a combination of methods arising from a standard metaheuristic, namely memetic
algorithm, and also an extension of existing hyper-heuristic components [27, 46].
There is an interesting potential for dual views of the overall problem. It is defined as a
multi-project problem, but it can be also viewed as a single project (multi-mode) RCPSP,
in which the precedence graph has a particular structure, consisting of disjoint clusters.
1 http://gent.cs.kuleuven.be/mista2013challenge/
2
There is a sense in which we work with both views together. Some neighbourhood moves
treat the problem in a single-project fashion and work on the constituent activities; other
neighbourhood operators explicitly consider the multi-project nature of the problem, and
focus on moves of projects. Both views, and kinds of operators, are used and work
together to improve the overall project-level structure as well as the detailed activity
level structure. A discussion and a computational study on both approaches can be
found in [35].
The primary contributions of this paper are:
• Observation and investigation of how the primary objective function being essentially a “sum of project completion times” leads to good solutions having inherently
different structure to those with makespan as the primary objective. In particular,
minimisation of project completion times subject to limited global resource results
in partial ordering of projects; this does somewhat reduce the effective size of the
search space, but also may lead to good solutions being more widely separated. Understanding of this significantly affected our algorithm design, including an MCTS
construction method aiming to create solutions having such structure.
• Novel neighbourhood moves, including those that are designed specifically for
smoother navigation through the search space of the multi-project extension of
MRCPSP – reflecting our observation of the effect that the main objective function has on the solution structures.
• An adaptive hybrid hyper-heuristic system to effectively control the usage of the
rich set of neighbourhood moves.
• Evidence of the effectiveness based on successful results on a range of benchmark
problems. This includes winning a competition, in which some problems were
hidden at the algorithm design/tuning phase. We also tested our algorithm on
single-project instances from PSPLIB. Although our algorithm was not designed
to work on single-project instances, it demonstrated good performance in these
tests, and was competitive with the state-of-the-art methods tailored to the singleproject case. Furthermore, it improved 3 best solutions on these PSPLIB instances
during these experiments.
These contributions are directed towards a system that is both robust and flexible;
with the potential to be effective at handling a wide variety of problem requirements
and instances. Arguably, one of the lessons of this paper is that greater complexity and
richness of such scheduling problems needs to be matched with a greater complexity and
richness of the associated algorithms; especially when not all instances are known in
advance, and so algorithms should not over-specialise to a particular data set.
Regarding the structure of the paper; in Section 2 we describe the problem to be
solved. In Section 3 we discuss how we have carefully chosen the appropriate data structures and implemented algorithms operating with those data structures efficiently in
order to construct the schedule from a given sequence as fast as possible. (To build
an effective system, one has to pay attention to all of its components.) However, most
of the contribution of this paper arises from choosing and combining the effective algorithmic components including the search control algorithm and then (partially) tuning
the relevant parameters within the overall approach. These consist of the MCTS-based
3
constructor given in Section 4, the neighbourhoods given in Section 5, and the improvement phase given in Section 6. The computational experiments and competition results
are presented and analysed in Section 7; including some reports of performance on a
multi-mode, though single project, benchmark set from PSPLIB. Section 8 concludes
the paper.
2. Problem Description
The problem consists of a set P of projects, where each project p ∈ P is composed of
a set of activities, denoted as Ap , a partition from all activities A. Each project p ∈ P
has a release time ep , which is the earliest start time for the activities Ap .
The activities are interrelated by two different types of constraints: the precedence
constraints, which force each activity j ∈ A to be scheduled to start no earlier than
all the immediate predecessor activities in set Pred (j) are completed; and the resource
constraints, in which the processing of the activities is subject to the availability of
resources with limited capacities. There are three different types of the resources: local
renewable, local non-renewable and global renewable. Renewable resources (denoted
using the superscript ρ) have a fixed capacity per time unit. Non-renewable resources
(denoted using the superscript ν) have a fixed capacity for the whole project duration.
Global renewable resources are shared between all the projects while local resources are
specified independently for each project.
Renewable and non-renewable resources are denoted using the superscript ρ and ν,
respectively. Rpρ is the set of local renewable resources associated with a project p ∈ P ,
ρ
and Rpk
is the capacity of k ∈ Rpρ , i.e. the amount of the resource k available at each
time unit. Rpν is the set of local non-renewable resources associated with a project p ∈ P ,
ν
and Rpk
is the capacity of k ∈ Rpν , i.e. the amount of the resource k available for the
whole duration of the project. G ρ is the set of the global renewable resources, and Gρk is
the capacity of the resource k ∈ G ρ .
Each activity j ∈ Ap , p ∈ P , has a set of execution modes Mj . Each mode m ∈ Mj
determines the duration of the activity djm and the activity resource consumptions. For
ρ
a local renewable resource k ∈ Rpρ , the resource consumption is rjkm
; for a local nonν
ν
renewable resource k ∈ Rp , the resource consumption is rjkm ; for a global renewable
ρ
resources k ∈ G ρ , the resource consumption is gjkm
.
Schedule D = (T, M ) is a pair of time and mode vectors, each of size n. For an
activity j ∈ A, values Tj and Mj indicate the start time and the execution mode of j,
respectively. Schedule D = (T, M ) is feasible if:
• For each p ∈ P and each j ∈ Ap , the project release time is respected: Tj ≥ ep ;
• For each project p ∈ P and each local non-renewable resource k ∈ Rpν , the total
ν
resource consumption does not exceed its capacity Rpk
.
• For each project p ∈ P , each time unit t and each local renewable resource k ∈ Rpρ ,
ρ
the total resource consumption at t does not exceed the resource capacity Rpk
.
• For each time unit t and each global renewable resource k ∈ Gpρ , the total resource
consumption at t does not exceed the resource capacity Gρk .
4
• For each j ∈ A, the precedence constraints hold: Tj ≥ maxj ′ ∈Prec(j) Tj ′ + dj ′ Mj′ .
The objective of the problem is to find a feasible schedule D = (T, M ) such that it
minimises the so-called total project delay (TPD), defined by using the time for total
project completion (TPC)
X
TPC =
Cp
(1)
p∈P
and
TPD ≡ fd (D) = TPC − L =
X
p∈P
where Cp is the completion time of project p
Cp − L ,
Cp = max Tj + djMj .
(2)
(3)
j∈Ap
The constant L is a lower bound calculated as
X
L=
(CPDp + ep ) ,
(4)
p∈P
with CPDp being a given pre-calculated value. Since L is a constant, then it does not
affect the optimisation (it was presumably introduced in the competition just to make
the output numbers smaller and easier to interpret). Specifically, since L is the lower
bound (though not necessarily a tight bound), fd (D) ≥ 0 for any feasible solution D.
Note that this primary objective
is an instance of the standard “weighted
completion
P
P
time”, usually denoted by “ j wj Cj ”, but specialised to the case, “ p wp Cp ”, in which
only the completion times of the projects are used2 (in the case of TPD all the weights
are assigned to be one).
The tie-breaking secondary objective is to minimise the total makespan, (TMS), which
is the finishing time of the last activity (or equivalently of the last project):
(5)
TMS ≡ fm (D) = max Tj + djMj = max Cp .
p∈P
j∈A
In our implementation, we combine the objective functions fd (D) and fm (D) into
one function f (D) that gives the necessary ranking to the solutions:
f (D) = fd (D) + γfm (D) ,
(6)
where 0 < γ ≪ 1 is a constant selected so that γfm (D) < 1 for any solution D produced
by the algorithm. In fact, we sometimes use γ = 0 to disable the second objective. For
details, see Section 6.1.
Under
it could perhaps be described as “MP+ S |
Pthe conventional “α|β|γ” labelling,
+
prec | ( p wp Cp , Cmax )” using ‘MP S’ to denote ‘Multi-mode Multi-Project Scheduling’.
2 If there is no unique activity marking the end of a project, then a dummy empty end activity can
always be added, without changing the problem.
5
3. Schedule Generator
Designing an algorithm for solving the multi-mode resource-constrained multi-project
scheduling problem requires an appropriate solution representation. There are two ‘natural’ solution representations in the scientific literature:
Schedule-based: A direct representation using the assignment times, and also modes,
of activities, i.e. vectors T and M .
Sequence-based: This is based on selecting a total order on all the activities. Given
such a sequence, a time schedule is constructed by taking the activities one at a
time in the order of the sequence and placing each one at the earliest time slot such
that feasibility of the solution would be preserved. This approach is called serial
schedule generation scheme (e.g., see [4]).
The schedule-based representation is perhaps the most natural one for a mathematical
programming approach, but we believe that it could make the search process difficult for
a metaheuristic method, in particular, generating a feasible solution at each step could
become more challenging. As is common for heuristic approaches [62], we preferred the
sequence-based representation, since it provides the ease of producing schedules that are
both feasible and for which no activity can be moved to an earlier time without moving
some other activities (the schedule is then said to be ‘active’).
The sequence-based representation is a pair S = (π, M ), where π is a permutation of
all the activities A, and M is a modes vector, same as in the direct representation. The
permutation π has to obey all the precedence relations, i.e., π(j) > π(j ′ ) for each j ∈ A
and j ′ ∈ Pred (j). The modes vector is feasible if Mj ∈ Mj for each j ∈ A and the local
non-renewable resource constraints are satisfied for each project p ∈ P .
In order to evaluate a solution S, it has to be converted into the direct representation D. By definition, the sequence-based representation S = (π, M ) corresponds to
a schedule produced by consecutive allocation of activities π(1), π(2), . . . , π(n) to the
earliest available slot. The corresponding procedure, which we call schedule generator, is
formalised in Algorithms 1, 2 and 3. Note that the procedure guarantees feasibility of
the resulting schedule as it schedules every activity in such a way that feasibility of the
whole schedule is preserved.
Algorithm 1: Serial schedule generation scheme
1
2
3
4
5
Let S = (π, M ) be the sequence-based solution;
for i ← 1, 2, . . . , n do
Let j ← π(i);
Schedule j in mode Mj to the earliest available slot such that feasibility of the
schedule is preserved;
end
The worst case time complexity of this implementation is O(n(ζ + T ρd)), where
ζ = maxj∈A |Prec(j)| is the maximum length of the precedence relation list, T is the
makespan, ρ = |G ρ | + maxp∈P |Rpρ | is the maximum number of local and global renewable resources and d = maxj∈A djMj is the maximum activity duration. The first term of
6
Algorithm 2: Scheduling an activity to the earliest available slot.
1
2
3
4
5
6
Let j be the activity to be scheduled;
Let m be the mode associated with j;
Let p be an index such that j ∈ Ap ;
Calculate the earliest start time of j as t0 ← max ep ,
max
j ′ ∈Prec(j)
(Tj ′ + dj ′ m ) ;
t ← TestSlot(j, t0 );
Allocate activity j at t in mode m and update the remaining capacities;
Algorithm 3: A naive implementation of the TestSlot (j, t) function. The function
returns the earliest time slot at or after t feasible for scheduling activity j.
1
2
3
4
5
6
7
8
9
10
11
12
13
Let m be the mode associated with j;
Let p be an index such that j ∈ Ap ;
for t′ ← t, t + 1, . . . , t + djm − 1 do
for k ∈ Rpρ do
Let a be the remaining capacity of k at t′ ;
ρ
if rjkm
> a then return TestSlot (j, t + 1)
end
for k ∈ G ρ do
Let a be the remaining capacity of k at t′ ;
ρ
if rjkm
> a then return TestSlot (j, t + 1)
end
end
return t;
the sum corresponds to handling precedence relations, and the second term corresponds
to scanning slots and testing resource availability. Note that ζ < n, and the maximum
number ρ of resources is a constant in our benchmark instances. Also, T is typically
linear in n, and, hence, the time complexity is quadratic. The schedule generator is the
performance bottleneck of our solver (note that most of the local search moves described
in this paper are no worse than linear time complexity). In our experiments, schedule
generation was usually taking over 98% of the CPU time. By introducing several improvements (based on information sharing and caching of partial solutions) described
below we reduced the running times of the schedule generator by a factor of around ten
compared to our initial routine implementation. That significantly increased the number
of iterations the higher level algorithm was able to run within a given time.
3.1. Issues in Efficient Implementation
In this section, we briefly discuss algorithmic and implementational issues that do
not directly affect the number of sequence evaluations, but that are designed to increase
the rate that evaluations are performed. This is of importance for application of the
methods to real-world problems. However, it can also be of potential importance to
choices between different heuristics or other algorithmic components.
7
The methods for evaluation and comparison of algorithms are not necessarily clear.
In particular, two aspects that arise with respect to this work, and scheduling in general,
are “hidden instances”, and “termination criteria”. Hidden instances (meaning ones that
are not available until after the implementation is finished) act against the danger that
occurs with open instances of the techniques becoming tailored to the specific instances.
For reliability and verification of results this generally means the implementations must
be finalised before the release of the instances. In practice, this seems to be rarely applied
outside of the context of a competition. In such cases, one might regard competitions
are a way to enforce the scientific good practice of fully finalising the algorithm and
implementation before the testing.
The other aspect is evaluating algorithms’ performance purely in terms of their (wallclock) runtimes, or instead using some attempt at implementing independent “counting”
measure of steps taken. The advantages of the former “runtime” method is that it relates
to what real world users would usually care about, and also might be the only real option
when no sensible pure counting methods are available. The advantage of the counting is
that it hides the implementation efficiency and hence allows to compare “pure algorithm
designs”. In some research situations the classes of algorithms are sufficiently similar for
a counting based comparison being viable, and then may well be standard e.g. in genetic
algorithms the number of fitness evaluations is commonly used.
Hence, the counting-based approach encourages/supports rapid explorations of ideas
for new algorithms, however exclusive usage would discourage developing new practical
methods of improving algorithm performance. For example, counting fitness evaluations
can miss the advantages of the incremental evaluation techniques routinely used in metaheuristics, and that are vital for their effectiveness. Runtime-based approach on contrary
encourages the researchers to exploit methods that are practical in real circumstances,
taking into account incremental evaluation, parallelism and other considerations crucial
for real-world systems. We note here that the associated added complexity of algorithm
engineering could potentially be partly addressed by hyper-heuristics, as they could provide feedback to the programmer regarding the extra value of low-level heuristics if their
implementation were improved. This is particularly relevant in the context of a solver
employing multiple neighbourhoods, like the one presented in this paper.
We believe that there is no simple answer to which of these two algorithm evaluation
approaches is best in general and so both of them have their place. However in the context
of the very well studied project scheduling we do believe that engineering questions need
to be accounted for. As an example of the importance of such ‘engineering issues’ we
refer to another well-studied area of solving propositional satisfiability (SAT) problems
and that has been active for many decades3 . An important part of development of
SAT solvers was the development of ‘watched literals’ [42]. The technique only directly
affected the standard and routine ‘unit propagation’ procedure in SAT solvers (which
is the CPU-intensive portion, analogous to the schedule generator), but it did so in a
fashion that meant new heuristics were then practical, leading to new algorithm designs.
In real-world usages of scheduling, an important aspect is the software engineering
aspect of the time and cost of implementing and maintaining the software. An initial
implementation of a neighbourhood is often relatively easy; however, the practical problems can arise when effective use requires that it is implemented in a fashion that uses
3 E.g.
competitions have been held for many years, see http://www.satcompetition.org/
8
incremental or delta evaluation (so that the objective function does not require a full
re-evaluation). To support the incremental evaluation it is often necessary to implement
appropriate data structures that are more sophisticated, and so harder to implement and
maintain. With serial generation the majority of the CPU time is spent in the generation
of the schedule from the sequence. Hence, we naturally found that significant improvements were achieved by modifying the generator algorithm. Good ‘engineering’ of the
serial generation also has the important advantage that it helps all of the neighbourhoods.
(If all the neighbourhoods were to rely on entirely separate implementations, then there
would be much more pressure, for practical software engineering issues, to reduce to a
smaller set.)
Observe (see Algorithm 2) that the schedule generation algorithm spends most of the
time finding the first available slot for an activity. To speed up this phase, we use a
modification of the Knuth-Morris-Pratt substring search algorithm. By testing resource
availability in the reversed order, we can use early exploration of insufficient resources
to skip several values of t, see Algorithm 4.
Algorithm 4: An improved implementation of the TestSlot (j, t) function. The
function returns the earliest time slot at or after t feasible for scheduling activity j.
1
2
3
4
5
6
7
8
9
10
11
for t′ ← t + djm − 1, t + djm − 2, . . . , t do
for k ∈ Rpρ do
Let a be the remaining capacity of k at t′ ;
ρ
if rjkm
> a then return TestSlot (j, t′ + 1)
end
for k ∈ G ρ do
Let a be the remaining capacity of k at t′ ;
ρ
if rjkm
> a then return TestSlot (j, t′ + 1)
end
end
return t
Another speed-up heuristic is exploiting the nature of the neighbourhood moves; we
noted that any two solutions tested consequently are likely to share a prefix. Let S 1 =
(π 1 , M 1 ) be some solution, S 2 = (π 2 , M 2 ) be its neighbour, and D1 = (T 1 , M 1 ) and D2 =
(T 2 , M 2 ) be their direct representations. According to our assumption, π 1 (i) = π 2 (i) = j
and Mj1 = Mj2 for i = 1, 2, . . . , x, where x is the prefix length (which is likely to be
significant). Then, by construction, Tj1 = Tj2 for each j = π 2 (1), π 2 (2), . . . , π 2 (x). Hence,
knowing D1 , we do not need to calculate the values Tj2 for j = π 2 (1), π 2 (2), . . . , π 2 (x).
For details, see Algorithm 5.
This gives a form of incremental evaluation that has the advantage that it applies to
all the neighbourhoods used. Potentially, this engineering optimisation could have impact
on the design of the improvement algorithm in that the selection of neighbourhood moves
could benefit from exploiting changes at the end of the schedule being faster to evaluate
than those at the beginning. This is something that can potentially be captured by a
hyper-heuristic, as one of the long-term intentions of hyper-heuristics is that they should
monitor the CPU times taken by different moves, and combine this with monitoring of
9
their effects, in order to give better adaptive control of the improvement phase.
Algorithm 5: Serial schedule generation scheme with prefix detection.
1
2
3
4
5
6
7
8
9
10
11
12
Let S 2 = (π 2 , M 2 ) be the new solution;
Let S 1 = (π 1 , M 1 ) be the previous solution and D1 = (T 1 , M 1 ) be the
corresponding direct representation;
Let prefix ← true;
for i ← 1, 2, . . . , n do
Let j ← π 2 (i);
if prefix = true and π 1 (i) = j and Mj1 = Mj2 then
Allocate activity j to Tj1 in mode Mj2 and update the remaining capacities;
else
Schedule j in mode Mj2 to the earliest available slot;
prefix ← false;
end
end
In addition to the algorithmic improvements, we used several standard programming
techniques to optimise the implementation performance, with the primary ones being:
• Incremental maintenance of auxiliary data: Note that the schedule generator
has to maintain the amount of remaining
resource
for each renewable resource
P
and each time unit, i.e. T · |G ρ | + p∈P |Rpρ | values. Since T is not known
in advance, the corresponding auxiliary data structure has to be large enough to
fit a schedule of size Tmax , where Tmax is the upper bound of T . Considering
the initialisation of this auxiliary data structure, the real time
P complexity of the
algorithm is O(n(ζ + T ρd) + Tmax ρ′ ), where ρ′ = |G ρ | + p∈P |Rpρ | = O(ρ|P |).
Usually Tmax ≫ T and ρ′ ≫ 1 and, hence, the last term of the generator complexity
has a major impact on the real performance of the procedure. By reusing the same
auxiliary data structures and only re-initialising the portion that was altered in the
previous run of the generator, we speed up the initialisation phase by a factor of
about Tmax /T .
• Local memory access: Most of the solution evaluations happen in our metaheuristic within local search procedures. Each instance of the local search algorithm is assigned a CPU core, and a dedicated copy of the schedule generator is
maintained for it. This is particularly important for the prefix reuse heuristic and
CPU cache efficiency.
4. Solution Structures
This section primarily explains why having “Total Project Delay” (TPD) as the
main objective can be expected to lead to the general structure of solutions being very
different to the structure obtained with the more standard makespan (TMS) objective.
Our observations of the TPD-driven structure had an important influence on our design of
the overall algorithm. The structure motivates many of the “project level” moves that are
10
considered in Section 5. In this section, we also propose a heuristic method to construct
the initial solutions which feed into the improvement phase later. As standard in such
“construct and improve” optimisation, the concern was that the desirable structures might
be difficult to achieve by the improvement unless the initial constructor heuristically tried
to get close, by doing an appropriate “heuristically-guided global sampling” of the space
of approximate structures. Accordingly, with the intent to increase the robustness of
the solver, to handle such structures in unseen instances, a specialised MCTS-based
constructor is proposed and developed.
4.1. TPD-driven Solution Structures
(a) Only makespan is minimised (the TPD objective is disabled), that is minimising the latest
completion times of projects. TPC = 925, TMS = 119.
(b) Both objectives are enabled, that is, the primary objective is the TPD – minimising the sum
of completion times of projects. TPC = 654, TMS = 128.
Figure 1: An example of overall project-level structure of good solutions (using instance
B-1). Each horizontal line shows the duration allocated to each project in the schedule.
In this subsection, we report on the investigation of the natural question of what
constitutes a ‘good’ approximate structure. We firstly look at the structure of a high
quality schedule for a benchmark instance, and then we elucidate the observed structures
using some small examples. The critical message is that optimising TPD leads to different
structures of the solutions than when optimising TMS; we believe that algorithm design
needs to take account of this difference.4
In particular, we have observed that in many cases, dominance by the TPD objective
frequently leads to approximate ordering of the projects. A typical example of this is
4 Similar
effects are also observed in [63].
11
given in Figure 1 using a good solution to instance B-1. Figure 1a shows the structure
obtained when only the standard makespan is minimised, but the structure is very different in Figure 1b with the required TPD-dominated objective. That is, the pattern of
completion times of each project, the Cp of (3), depends on whether they are driven by
TPD, effectively minimising the average of the Cp , or instead driven by the makespan,
reducing the maximum of the Cp . The example shows how, with TPD dominating, there
are time periods in the schedule when the general focus is on relatively few projects and
during the schedule this focus changes between projects.
Hence, the evidence from Figure 1, and other multi-project cases we have looked at,
suggests that the approximate ordering is common. Overall, these structures naturally
arises from the combination of the TPD (2) objective with the limited global resources.
Suppose that some P is the final finishing project, and so its last activity determines the
TMS and P’ is an earlier finishing project. It may well be that P’ can move some of its
activities earlier by delaying activities of P, though without changing the last activity of
P. Such moves will improve the TPD without worsening the TMS.
That is, in contrast to the makespan objective, the TPD objective, together with
limited shared resources, has a natural side-effect of encouraging unfairness between the
finish times of projects, and will drive some projects to finish as early as possible.
To illustrate the way in which an approximate ordering might arise, and TPD-driven
and TMS-driven solutions can be quite different, we give a set of small example instances.
We denote these example instances by “2[x–y–x]”, consisting of two projects. Each
project consists of a precedence-constrained chain of just 3 activities of durations x, y
and x, respectively. There are two (renewable) shared resources with activities 1 and 3
using the resource 2, and activity 2 using resource 1. We focus on values of x and y that
lead to resource 2 being the bottleneck, and so the main driving force of the makespan –
corresponding to the shared global resource in the MISTA benchmarks. Figure 2a shows
four schedules of this parametrised instance, along with their corresponding values of C1
and C2 . Solutions S4 and S3 arise from simply swapping the order of the projects in
solutions S1 and S2 respectively.
The solutions S1 and S4 do not interleave the two projects, and so leave gaps in the
bottleneck resource 2. In contrast, solutions S2 and S3 interleave the projects and so lead
to full usage of the bottleneck resource, and are hence automatically the best solutions
for the TMS objective. However, S2 and S3 are not always the best solutions for the
TPD objective. Changing the solution leads to a tradeoff between C1 and C2 , and this is
illustrated, in Figure 2b for various choices of x and y. In this figure, the values of x and
y are selected so that we always have 2x + y = 100; this makes the effect of changing the
relative balance of x and y clearer, and also means that the values can be interpreted as
percentages of the project time.
Before proceeding further, we observe that we can also interpret the project completion times, Cp , from the point of view of ‘multi-objective optimisation’, by regarding
each of them as separate objectives. In particular, in standard fashion, one can say that
a set of values for Cp dominates another set provided that none of the Cp values are
worse (larger) and at least one is better (smaller). One can hence discuss the cases in
Figure 2b as being the Cp -Pareto Front, of non-dominated sets of Cp . We emphasise this
view is for obtaining insight into the space of solutions, and is not used directly within
our algorithm; we did not perform multi-objective optimisation over the Cp , though believe it would be worthy of future investigation. Also, note this view is different from
12
(a) Case “2[x–y–x]”. Showing four possible solutions, S1 , . . . , S4 , and associated project completion times C1 and C2 .
(b) Tradeoff between completion times of
projects. We select a range of different values for x and y, but with the invariant that
2x + y = 100; so that the effect of varying
the balance between x and y is made clear.
Figure 2: Simple 2-project examples to illustrate concave and convex non-dominated
sets. The notation “2[x–y–x]” means two projects each consisting of a chain of 3 tasks
of durations [x, y, x] with shared resources, and the tasks 1 and 3 in each chain using
resource 2, and task 2 using resource 1.
taking the pair (TPD, TMS) as a bi-objective problem, and having a (TPD, TMS)-Pareto
Front (this would correspond again to a different class of algorithms than we consider
here, but is again worthy of future investigation). Incidentally, there are two distinct
‘multi-objective views’ on the space of solutions:
• Cp -Pareto Front, of dimension p, formed from the p completion times
• (TPD, TMS)-Pareto Front, of dimension 2, and formed from the aggregates (sum
and max) of the Cp
However, it is important to realise that although the Cp and (TPD,
TMS) views are
P
different, they are tightly linked. Both the TPD, arising from p Cp , and the TMS,
from maxp Cp , are monotone (non-decreasing) with respect to the Cp and so preserve
Pareto dominance. That is, if solution Cp dominates solution Cp′ , then the resulting
pair (TPD, TMS) dominates (TPD’, TMS’) – though note that the converse does not
13
apply. Hence, the (TPD, TMS)-Front can be extracted from the Cp -Front. However, this
‘projection’ loses information; the Cp -Front gives more insight into the space of solutions.
Consequently, the “Cp -Pareto Front”, or set of non-dominated Cp values, can give
some insight into the effect of TPD compared to TMS on the structure of the solutions.
Hence, we can regard Figure 2b as showing the effect of the x–y balance on the Cp -Pareto
Front. In particular, we see that for two of the cases, 2[44–12–44] and 2[48–4–48] the
Cp -Pareto Front is concave in terms of the set of feasible Cp values (on the top right).
When the Front is concave then the two solutions minimising the TPD are S1 and S4 and
are at opposite ends of the Cp -Front. In contrast, the solutions S2 and S3 minimising
the TMS are in the middle.
P
Generally, with a concave Cp -Pareto Front the solutions (locally) minimising p Cp
are naturally at the ends of the Front, and furthermore solutions at the ends of the Front
will be more likely to have some Cp values small and others large. This matches with the
way that the observed solutions minimising TPD do indeed sequence the projects. That
is, when the projects are roughly equal size, and the Cp Pareto Front is concave, then it
becomes reasonable that the TPD objective prefers the end points, and so the candidate
good solutions are more likely to be widely separated in the search space.
Also, notice that as the fraction of activities using the bottleneck resource 2 (that is
the value of 2x) increases, then the Cp -Pareto Front becomes more concave, and the TPDdriven solutions become different from the TMS-driven ones. This is consistent with the
our general experience of the properties of MISTA instances; the shared global resource
tends to be a bottleneck, and consequently they show the approximate ordering as seen
in Figure 1. Future work could well use such features related to the tightness of bottleneck resources to predict these effects, and so be used to select appropriate algorithm
components. We also remark that the interesting structure of the tradeoff between the
completion times does suggest that future work might well use multi-objective methods,
e.g. see [10], and of course could study the effect of different weights for the completion
times.
The simple example above is atypical in that the two projects are identical. However,
if the projects are similar in size then it is reasonable to expect that the Cp -Pareto Front
may be similar in overall structure but slightly distorted. When the front is concave but
still roughly symmetric with respect to swapping the Cp , then the TPD-driven solutions
are likely to still be towards the ends of the Front and so widely separated. However,
because the front is not totally symmetric the ends are likely to have slightly different
values of TPD. Hence, in such circumstances, it is reasonable to expect that there may
be many widely separated local optima in TPD, and also that the optima correspond to
different (approximate) orderings of the projects.
4.2. MCTS
We have seen above that the TPD objective tends to favour a different solution
structure than standard makespan minimisation, and in particular can drive a partial
orderings of the projects, and there may be many good solutions that are widely separated in the solution space. Consequently, our construction method attempts to do a
broad exploration of the solution space and to create initial sequences with a broadly
similar partial order. The reason for including this was simply that we expected the
partial ordering to lead to widely separated local minima in the search space, with large
barriers between them – as swapping the order of two projects might mean going through
14
interleaved intermediate states in which the TPD is worse. We included a method to
sample the space of approximate project orderings, with the intent to avoid starting in
a poor ordering and being trapped.
We expect that only a partial ordering is needed because we can assume that the
subsequent improvement phase can make small or medium size adjustments to the overall
project ordering structure. However, the improvement phase could have more difficulty,
and take more iterations, if the general structure of the project ordering were not close
to the structure expected in best solutions. Consequently, and for simplicity, we decided
that a reasonable approximation would be to use a 3-way partition of the projects taken
to correspond to ‘start’, ‘middle’ and ‘end’ parts of the overall project time. We required
the numbers of projects in each part (of the partition) to be equal – or with a difference
of at most one when the total number of projects is not a multiple of 3.
The problem then is how to quickly select good partition of the projects, and the
method we selected is a version of Monte-Carlo Tree Search (MCTS) methods [3]. The
general idea of MCTS is to search a tree of possibilities, but the evaluation of leaves is
not done using a predefined heuristic, but instead by sampling the space of associated
solutions. The sampling is performed using multiple invocations of a “rollout” which
is designed to be fast and unbiased. It needs to be fast so that multiple samples can
be taken; also rather than trying to produce “best solutions” it is usually designed to
be unbiased – the (now standard) idea being that it should provide reliable branching
decisions in the tree, but is not directly trying to find good solutions.
In our case, the tree search corresponds to decisions about which projects should be
placed in which part of the partition. The rollout is a fast way to sample the feasible
activity sequences consistent with the candidate choice for the partition of the projects.
Specifically, the tree search works in two levels; firstly to select the projects to be placed
in the end part and then to select the partition between the start and middle parts.
The first stage considers5 100 random choices for the partition of the projects, and
then selects between these using 120 samples or the rollout6 . The rollout consists of two
main stages:
1. Randomly select a total ordering of the activities consistent with the precedences
and with the candidate partitioning. Specifically, within each partition we effectively consider a dispatch policy that randomly selects between activities that are
available to be scheduled because their preceding activities (if any) are already
scheduled.
2. Randomly select modes for the activities. If the result is not feasible then this can
only be because of the mode selection causing a shortfall in some non-renewable
resources. Hence, it is repaired using a local search on the space of mode selections.
We use moves that randomly flip one mode at a time, and an objective function
that measures the degree of infeasibility by the shortfall in resources. Since the
non-renewable resources are not shared between projects, this search turned out to
be fast and reliable. As a measure of precaution, if the procedure fails to obtain
5 The parameters for the MCTS were the result of some mild tuning, and used in the competition
submission, but of course are adjustable.
6 The number 120 was selected so that the rollouts could be evenly distributed between 2, 4, 6, 8 or
12 cores of the machine
15
a feasible selection of modes after a certain number of local search iterations, we
restart it with a random selection of modes.
The first stage ends by making a selection of the best partitioning, using the quality of
the 25th percentile of the final solution qualities (the best quartile) of the results of the
rollouts. The ‘end’ part is then fixed to that of the best partition. The decision to fix
the ‘end’ part also arose out of the observation that in good solutions the end projects
are least interleaved. The MCTS proceeds to the second stage, and follows the same
rollout procedure but this time to select the contents of the middle (and hence start)
parts. This entire process usually completes within only a few seconds, and was used
as the construction stage before the much longer improvement phase. We emphasise
that the TPD-structure observed earlier had an important influence of the design of the
neighbourhoods used in the key improvement phase – in particular, the motivation that
some moves should affect the project level structure.
5. Neighbourhood Operators
In this section, we describe our neighbourhood operators, used by several components
during the improvement phase. The operators (also referred to as low-level heuristics or
simply moves, depending on the algorithm which makes use of them) are categorised into
three groups. This categorisation is mainly based on the common nature of the strategy
the operators employ while manipulating the solution. Some of the moves are similar to
those used by other submissions; see [62], and [16] and [57] for few examples. To make
the paper self-contained, below we provide brief descriptions of all the moves used in our
algorithm.
We guarantee that all of our operators preserve feasibility of the solution. Also, all of
the moves are randomised so that they could be repeatedly used in “simulated annealing”like improvement or applied as mutation operators. All the random selections are made
at uniform unless specified otherwise.
5.1. Activity-level Operators
Operators in this category involve basic operations, widely used in the literature, such
as changing the mode of a single activity or swapping the positions of two activities. On
top of that, we implemented a limited first improvement local search procedure for each
of the basic moves.
To describe the moves in this category, we will need additional notations. Let pos(j)
be the position of activity j ∈ A within a given solution. If activity j is shifted to a
different position, we say that the feasible range of its new position is [ℓ(j), u(j)], where
ℓ(j) and u(j) can be computed as follows:
(
maxj∈Pred(j1 ) pos(j) + 1 if Pred(j1 ) 6= ∅,
ℓ(j1 ) =
(7)
1
otherwise
and
ℓ(j1 ) =
(
minj∈Succ(j1 ) pos(j) − 1 if Succ(j1 ) 6= ∅,
n
otherwise,
where Succ(j1 ) is the set of successors of j1 .
16
(8)
• Swap activities: swap two activities in the sequence. Select an activity j1 ∈
A randomly. Select another activity j2 6= j1 ∈ A randomly such that ℓ(j1 ) ≤
pos(j2 ) ≤ u(j1 ). If ℓ(j2 ) ≤ pos(j1 ) ≤ u(j2 ) then swap j1 and j2 . Otherwise leave
the solution intact.
• Shift: shift an activity to a new location in the sequence. Shift a randomly selected
activity j ∈ A to a new position randomly selected from pos(j) ∈ [ℓ(j), u(j)].
• Change mode: change the mode of a single activity. Select an activity j randomly
at uniform, and if |Mj | = 1 then leave the solution intact. Otherwise select a new
mode m 6= Mj ∈ Mj and, if this does not lead to a violation of non-renewable
resource constraints, update Mj to m.
• FILS swap activities: apply the first improvement local search (FILS) procedure
based on the swap move. The operator has one parameter: the width W > 1 of
the window to be scanned. Select an activity j1 ∈ A randomly. Define the window
[ℓ′ , u′ ] for activity j2 as follows. If u(j) − ℓ(j) < W , let ℓ′ = ℓ(j) and u′ = u(j).
Otherwise select ℓ′ ∈ [ℓ(j), u(j) − W + 1] and set u′ = ℓ′ + W − 1. Then, for
every j2 ∈ A \ {j1 } such that ℓ′ ≤ pos(j2 ) ≤ u, attempt to swap j1 and j2 (see
SwapActivities). If the attempt is successful and it reduces the objective value
of the solution then accept it and stop the search. Otherwise roll back the move
and proceed to the next j2 if any.
• FILS shift: apply the first improvement local search (FILS) procedure based on
the shift move. The operator is implemented very similarly to FILS swapActivities, i.e. it attempts to shift a randomly selected activity to an new position within
a window of a given size.
• FILS change mode: apply the first improvement local search (FILS) procedure
based on the change mode move (the operator has no parameters). Select an
activity j ∈ A randomly. For m ∈ Mj \ {Mj }, produce a new solution by setting
Mj = m. If the resulting solution is feasible and provides an improvement over the
original solution, accept the move and stop the local search. Otherwise proceed to
the next m ∈ Mj \ {Mj } if any.
5.2. Ruin & Recreate Operators
The Ruin & Recreate (R&R) operators are widely used in metaheuristics as mutations
or strong local search moves but, to the best of our knowledge, they are relatively new to
serial generation in scheduling. As the name suggests, such operators have two phases:
the “ruin” phase removes some elements of the solution and the “create” phase reshuffles
those elements (for example, randomly) and then inserts them back. The number of
elements to remove and re-insert is a parameter of a R&R operator, which controls its
strength (i.e. the average distance between the original and the resulting solutions).
We implemented three different types of R&R operators, and several strategies to
select activities involved in the move. One can arbitrary combine any type of the R&R
operator with any activity selection strategy.
The move types are Reshuffle Positions, Reshuffle modes and Reshuffle
positions and modes. Reshuffle positions move removes all the selected activities
17
A′ ⊂ A from the solution (leaving |A′ | gaps in the sequence) and then re-inserts them in
a random order while respecting the precedence relations. To determine a feasible order,
we compute which activities V i can be placed in each gap i = 1, 2, . . . , |A′ | (in terms of
precedence relations between the activities in A′ and activities in A\A′ ), and also produce
a precedence relations sub-graph induced by A′ . Then we apply a backtracking algorithm.
In each iteration, it fills one gap, starting from the earliest gaps in the sequence. For gap
i it randomly selects an activity j ∈ A′ such that there are no incoming arcs to j in the
precedence sub-graph. If such an activity exists, the algorithm removes j from A′ and
from the precedence sub-graph. Otherwise it rolls back to attempt another activity on
the previous level of search. The depth of roll backs is unlimited, i.e., in the worst case,
the algorithm will performs the depth first search of the whole search tree (observe that
there is always at least one feasible arrangement of the activities).
Reshuffle modes move changes the modes of selected activities while keeping their
positions intact. Each of the selected activities j ∈ A′ is assigned a randomly chosen
mode, either equal or not to the previous mode Mj . Observe that the new mode selection
may cause infeasibility in terms of non-renewable resources. Thus, we use the multi-start
metaheuristic, simply repeating the above step until a feasible mode selection is found.
Reshuffle positions and modes combines the above two moves, which is trivial
to implement as modes feasibility is entirely independent of the sequence feasibility.
We implemented several activity selection strategies some of which exploit our knowledge of the problem structure.
• Uniform: as the name suggests, A′ ⊂ A is selected randomly at uniform. The
number of elements in A′ is a parameter of the move.
• Project: the activities A′ are selected within a single project, i.e. A′ ⊂ Ap . The
project p ∈ P is selected randomly.
• Local: the selection of activities is biased to those scheduled near a certain time
slot in the direct representation D. We randomly sample the activities accepting
an activity j with probability
probability (j) =
1
|Tj −τ |
width
+1
,
(9)
where 0 ≤ τ ≤ fm (D) (see (5)) is the randomly selected time slot (the centre of
the distribution) and width is a parameter defining the spread of the distribution.
The sampling stops when A′ reaches the prescribed cardinality.
• Global resource driven: the selection of activities is biased to the ones scheduled to time slots that under-utilise the global resources. The rationale is that
global resources usually present a bottleneck in minimising the project completion
times, and inefficiencies in the global resources consumption should be addressed
when polishing the solution. In this selection strategy, as in Local selection,
we use random sampling, but the acceptance probability for an activity j ∈ A is
defined by
P
ρ remaining k (Tj )
,
(10)
probability (j) = k∈G P
ρ
k∈G ρ Gk
18
where remaining k (t) is the remaining (unutilised) capacity of global resource k at
the time slot t.
• Ending biased: the selection of activities is biased to the last activities within
the projects. The rationale is that, in an unpolished solution, the completion of a
project might be improved by careful packing the last activities such that all the
last activities end at roughly the same time, effectively maximising the utilisation
of the local resources and redistributing the global resources between projects. As
in Local selection and Global resource driven selection, we use random
sampling of activities. The probability of accepting an activity j ∈ A is
probability (j) =
project -pos(j)
,
|Ap |
(11)
where p is the project of activity j and project -pos(j) is the position of activity j
among the activities of project p.
5.3. Project-level Operators
It was shown in Section 4 that the TPD objective function tends to create a partial
ordering of projects in high-quality solutions. The danger, however, is that the ordering to
which the solution converges might be sub-optimal. Consider two well-polished solutions
S1 and S2 having different project orderings. The distance between S1 and S2 is likely
to be significant with respect to the activity-level and R&R operators as many activities
need to be moved to convert S1 into S2 . Moreover, the transitional solutions in such a
conversion will have significantly poorer quality compared to that of S1 and S2 . This
indicates that, once the algorithm converges to a certain project ordering, it needs a lot
of effort to leave the corresponding local minimum and change the project ordering.
The project-level operators are designed to overcome such barriers, allowing faster
exploration of the rough landscape of the MRCMPSP. They perform on the project
level and thus one move of any of operators from this category usually results into a
change in the project ordering. That being said, the project-level moves are likely to
corrupt the solution causing many activity-level inefficiencies that need to be treated
with activity-level and R&R moves. Thus, they ought to be used rarely within local
search, but they can serve well as mutation operators.
In some of the project-level moves we use the formal concept of project ordering P (S).
To extract P (S) from a solution S, we compute the “centre of mass” centre(p) for each
project p ∈ P as
1 X
centre(p) =
pos(j)
|Ap |
j∈Ap
and define the ordered set P (S) of projects according to their centres of mass.
• Swap two projects: swap two randomly selected projects in the sequence. Select
two projects p1 6= p2 ∈ P randomly. Remove all the activities belonging to projects
p1 and p2 from the sequence. Fill the gaps with all the p2 activities and then
all the p1 activities preserving the original order within each of the projects. If,
in the original solution, p1 was located earlier than p2 then the move swaps the
projects (for instance, (1, 1, 1, 2, 1, 2, 2, 2) turns into (2, 2, 2, 2, 1, 1, 1, 1)). Otherwise
19
it simply separates them more clearly (for instance, (2, 2, 2, 1, 2, 1, 1, 1) turns into
(2, 2, 2, 2, 1, 1, 1, 1)).
• Swap neighbour projects: swap two projects adjacent in the project ordering.
Extract the project ordering P (S), randomly select 1 ≤ i < q and set p1 = P (S)i
and p2 = P (S)i+1 . Then apply the Swap two projects move.
• Compress project: place all the activities of a project adjacently to a new location x in the sequence while preserving their original ordering. Randomly select a
project p ∈ P and remove all the activities j ∈ Pp from the sequence, squeezing the
gaps (the resulting sequence will contain n− |Pp | activities). Insert all the activities
j ∈ Pp consecutively at position ⌈x(n − |Pp |)⌉, where the parameter 0 ≤ x ≤ 1 is
the new relative location of Pp .
• Shift project: shift all the activities of a project by some offset. Randomly select
a project p ∈ P and calculate pos min = minj∈Pp pos j and pos max = maxj∈Pp pos j .
Randomly select an offset −pos min < δn − pos max and shift every activity j ∈ Pp
by δ positions toward the end of the sequence.
• Flush projects: flush the activities of one or several projects to the beginning or
ending of the sequence. Compute the project ordering P (S) and select x consecutive
projects P ′ from P (S), where 1 ≤ x < |P | is a parameter. Flush all the activities
in projects P to either the beginning or the ending of the sequence (defined by an
additional parameter of the move).
6. Improvement Phase
Most of the time our algorithm spends on improving the initial solutions. We use
a multi-threaded implementation of a simple memetic algorithm with a powerful local
search procedure based on a hyper-heuristic which controls moves discussed in the previous section.
6.1. Memetic Algorithm
A genetic algorithm is a population based metaheuristic combining principles of natural evolution and genetics for problem solving [54]. A pool of candidate solutions (individuals) for a given problem is evolved to obtain a high quality solution at the end. A
fitness function is used to measure the quality of each solution. Mate/parent selection,
recombination, mutation and replacement are the main operators of an evolutionary algorithm. However, the usefulness of recombination is still under debate in the research
community [13, 38]. A recent study showed that recombination can be useful at a certain
stage during the search process, if the mutations do not change the quality of resultant
individuals leading to a population containing different individuals with the same fitness
[56]. The choice of the recombination operator can influence the best setting for the rate
of mutation depending on the problem dealt with. Although the study is rigorous, it
is still limited considering that some benchmark functions, such as OneMax are used
for the proofs. A memetic algorithm (MA) hybridises a genetic algorithm with local
search which is commonly applied after mutation on the new individuals [40, 41]. Many
improvements for MAs have been suggested, for example the population sizing [24] and
20
interleaved mode of operation [47]. MAs have been successfully applied to many different
problems ranging from generalised travelling salesman [19] to nurse rostering [44].
The improvement phase of our algorithm is controlled by a simple multi-threaded MA
which manages the solution pool and effectively utilises all the cores of the CPU. Our
MA is based on quantitative adaptation at a local level according to the classification in
[43].
Within the MA, we use a powerful local search procedure that takes a few seconds on
a single core to converge to a good local minimum. (Running the local search for a longer
time still improves the solution but the pace of improvements slows down.) As the local
search procedure has to be applied to every solution of the population in each generation,
and the local search is by far the most intensive time consumer in our algorithm, the
total running time of the algorithm can be estimated as
num-gen · pop-size · ls-time
,
cores
where num-gen is the number of generations, pop-size is the size of the population,
ls-time is the time taken by the local search procedure on one core per solution and
cores is the number of available CPU cores. Our algorithm is designed to converge
within a few minutes. This implies that, to have a sufficient number of generations, the
size of the population ought to be of the same order as the number of cores. On the other
hand, as the memetic algorithm requires a barrier synchronisation at each generation,
and our local search procedure cannot utilise any more than one CPU core, the size of
the population has to be pop-size = cores · i to fully utilise all the cores, where i is a
positive integer. In our implementation, we fixed the running time of the local search
procedure to five seconds, and the size of the population to the number of CPU cores
available to the algorithm.
Because of the small population size and limited number of generations, we decided
to use a simple version of the MA, see Algorithm 6. We denote the population as
S = {S1 , S2 , . . . , Scores }, and the subroutines used in the algorithm are as follows:
• Construct() returns a new random solution with the initial partial project sequence
obtained during the construction phase (see Section 4).
• Accept(Si ) returns true if the solution Si is considered ‘promising’ and false otherwise. The function returns false in two cases: (1) f (Si ) > 1.05f (Si′ ) for some
i′ ∈ {1, 2, . . . , |S|} or (2) the solution was created at least three generations ago and
Si is among the worst three solutions. Ranking of solutions is performed according
to fd (Si ) + idle, where idle is the number of consecutive generations that did not
improve the solution Si .
• Select (S) returns a solution from the population chosen with the tournament selection based on two randomly picked individuals.
• Mutate(X) returns a new solution produced from solution X by applying a mutation operator. The mutation operator to be applied is selected randomly and
uniformly among the available options:
– Apply the Reshuffle positions and modes move with the Local selection
strategy (see Section 5.2). The cardinality of the selection is |A′ | = 3. Repeat
21
–
–
–
–
the procedure 20 times, each time randomly selecting the centre of distribution
1 ≤ τ ≤ fm (D), see (9). The width parameter is taken as width = 0.1fm (D).
Apply the Swap neighbour projects operator once (see Section 5.3).
Apply the Flush projects operator with the number of selected projects
being one and flushing to the end of the sequence (see Section 5.3).
Apply the Flush projects operator with the number of selected projects
being two and flushing to the beginning of the sequence. It was noted that
the project ordering is usually less explicit at the beginning of the sequence
in good solutions, and for that reason we did not use flushing single projects
to the beginning in our mutations.
Same as the last mutation except that the number of selected projects is three.
It was noted in Section 4 that the components fd (D) and fm (D) of the objective
function f (D) (see (6)) are competing. Indeed, minimisation of the total makespan
favours solutions with projects running in parallel as such solutions are more likely to
achieve higher utilisation of the global resources. At the same time, minimisation of
the TPD favours solutions with the activities grouped by projects. Hence, the second
objective creates a pressure for the local search that pushes the solutions away from the
local minima with regards to the first (main) objective. To avoid this effect, we initially
disable the second objective (γ ← 0, see Section 2) and re-enable it only after 70% of the
given time is elapsed.
The parameters of the memetic algorithm (such as the ones used in the Accept (Si )
function, or the number of solutions in the tournament in Select(S)) have been chosen
using a parameter tuning procedure. It should be noted, however, that the algorithm is
not very sensitive to any of those parameters, which makes it efficient on a wide range
of instances, and this conclusion was supported by our empirical tests, see Section 7.
6.2. A Dominance based Hyper-heuristic Using an Adaptive Threshold Move Acceptance
There is a growing interest towards self-configuring, self-tuning, adaptive and automated search methodologies. Hyper-heuristics are such high level approaches which
explore the space of heuristics (i.e. move operators) rather than the solutions in problem
solving [5]. There are two common types of hyper-heuristics in the scientific literature
[7]: selection methodologies that choose/mix heuristics from a set of preset low-level
heuristics (which can both improve or worsen the solution) and attempt to control those
heuristics during the search process; and generation methodologies that aim to build new
heuristics from a set of preset components. The main constituents of an iterative selection hyper-heuristic are heuristic selection and move acceptance methods. At each step,
an input solution is modified using a selected heuristic from a set of low-level heuristics.
Then the move acceptance method is used to decide whether to accept or reject the new
solution. More on different types of hyper-heuristics, their components and application
domains can be found in [5, 6, 53, 45].
In this study, we combine two selection hyper-heuristics under a single point based
iterated local search framework by employing them adaptively in a staged manner. The
approach extends the heuristic selection and move acceptance methods introduced in
[46] and [28], respectively. The pseudocode of our adaptive iterated multi-stage hyperheuristic is given in Algorithm 7. lines 5–21 and 22–25 provide the high level design of
the first and second stage hyper-heuristics, respectively.
22
Algorithm 6: Improvement Phase.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
γ ← 0;
for i ← 1, 2, . . . , |S| do
Si ← Construct ();
end
while elapsed -time ≤ given-time do
if elapsed -time ≥ 0.7given-time then
γ ← 0.000001 (enable secondary objective function);
end
for i ← 1, 2, . . . , |S| (multi-threaded) do
Si ← LocalSearch(Si );
end
for i ← 1, 2, . . . , |S| do
if Accept(Si ) = false then
X ← Select (S);
Si ← Mutate(X);
end
end
end
6.2.1. First stage hyper-heuristic
The first stage hyper-heuristic maintains an active pool of low-level heuristics LLH ⊆
LLH all and a score score h associated with each heuristic h ∈ LLH . In each iteration,
it randomly selects a low-level heuristic from the active pool with probability of picking
h ∈ LLH being proportional to score h (line 5). Then the selected heuristic is applied
to the current solution (line 6). Initially, each heuristic has a score of 1, hence the
selection probability of a heuristic is equally likely. The first stage hyper-heuristic always
maintains the best solution found so far, denoted as Sbest (lines 9–11) and keeps track
of the time since last improvement.
The move acceptance component of this hyper-heuristic (lines 7–17) is an adaptive
threshold acceptance method controlled by a parameter ǫ accepting all improving moves
(lines 7–12) and sometimes non-improving moves (lines 14–16). If the quality of a new
solution S ′ is better than (1 + ǫ)f (Sbest ) (line 14), even if this is a non-improving move,
S ′ gets accepted becoming the current solution S. Whenever Sbest can no longer be
improved for elapsed -time 2 (in our implementation elapsed -time 2 = 1 sec), the parameter
ǫ gets updated as follows:
ǫ(Sbest ) =
⌈log(f (Sbest ))⌉ + rand
f (Sbest )
(12)
where 1 ≤ rand ≤ ⌈log(x)⌉ is selected randomly at uniform. Note that 0 is a lower bound
for f (S) (see Section 2) and, hence, the algorithm will terminate if f (Sbest ) = 0.
6.2.2. Second stage hyper-heuristic
The second stage hyper-heuristic dynamically starts operating (lines 22–25 of Algorithm 7) whenever there is no improvement in f (Sbest ) for elapsed -time 3 (in our imple23
Algorithm 7: LocalSearch(Si )
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
Let LLH all = {LLH 1 , LLH 2 , . . . , LLH M } represent set of all low level heuristics
with each heuristic being associated with a score, initially set to 1;
Let Sbest represent the best schedule;
S ← Si ; Sbest ← Si ; LLH ← LLH all ; ǫ ← ǫ(Sbest );
repeat
h ← SelectLowLevelHeuristic(LLH);
S ′ ← h(S);
if f (S ′ ) < f (S) then
S ← S ′;
if f (S ′ ) < f (Sbest ) then
Sbest ← S ′ ;
end
end
else
if f (S ′ ) < (1 + ǫ)f (Sbest ) then
S ← S ′;
end
end
if NoImprovement (elapsed -time 2 ) then
S ← Sbest ;
ǫ ← ǫ(Sbest );
end
if NoImprovement (elapsed -time 3 ) then
ǫ ← ǫ(Sbest );
(S, LLH ) ← SecondStage(Sbest , LLH all , elapsed -time 1 );
end
until timeLimitExceeded (elapsed -time 1 );
return Sbest ;
24
mentation elapsed -time 3 = 3 sec) in line 22. The hyper-heuristic in this stage updates
the active pool LLH of heuristics. LLH ⊆ LLHall is formed based on the idea of a
dominance-based heuristic selection as introduced in [46] reflecting the trade-off between
the objective value achieved by each low-level heuristic and number of steps involved.
The method considers that a low-level heuristic producing a solution with a small improvement in a small number of steps has a similar performance to a low level heuristic
generating a large improvement in large number of steps. This hyper-heuristic not only
attempts to reduce the set of low-level heuristics but also assigns a score for each lowlevel heuristic in the reduced set, dynamically. Those scores are used as a basis for the
heuristic selection probability of each low level heuristic to be used in the first stage
hyper-heuristic.
In the second stage hyper-heuristic, firstly, ǫ is updated in the same manner as in
the first stage hyper-heuristic and never gets changed during this phase. Then a greedy
strategy is employed using all heuristics in LLH all for a certain number of steps. Each
move in this stage is accepted using the same adaptive threshold acceptance method as
described in Section 6.2.1. LLH all is partitioned into three subsets LLH small , LLH medium,
LLH large considering the number of activities processed (e.g., number of swaps) by a given
heuristic.
Small: Swap activities, Shift and Change mode.
Medium:
• Reshuffle modes and Reshuffle positions and modes with the Uniform
selection strategy;
• Reshuffle modes and Reshuffle positions and modes with the Local selection strategy;
• Reshuffle positions and modes with the Global resource driven selection
strategy;
• Reshuffle positions and modes with the End biased selection strategy;
• Reshuffle positions and modes with the Project selection strategy;
• FILS swap activities, FILS shift and FILS change mode;
Large:
• Flush projects applied to one project; the direction is picked randomly;
• Swap two projects;
• Compress project;
• Shift project.
At each step, each low-level heuristic is applied to the same input solution for a fixed
number of iterations (5n/q for small, n/q for medium and 1 for large heuristics).7 If
7 Similar to the memetic algorithm parameters, this partition of LLH
all and the associated numbers
of iterations were selected by intuition and parameter tuning.
25
a low-level heuristic produces a new solution identical to the input, that invocation is
ignored. Otherwise, the objective of the new solution together with the low-level heuristic
which produced that solution gets recorded. Once all heuristics are applied to the same
input and get processed at a given step, the best resultant solution propagates as input
to the next greedy step.
A solution is considered to be a ‘non-dominated’ solution at a given step, if the
quality of the solution is better than the best solution obtained in the previous steps.
The active pool LLH of low-level heuristics is formed using the heuristics achieving the
non-dominated solutions obtained at each step during the greedy phase. The score of
a low-level heuristic is set to the number of non-dominated solutions that it produced.
Note that a non-dominated solution could be generated by multiple low-level heuristics,
in which case, their scores get increased.
Figure 3 illustrates a run of the second stage hyper-heuristic with LLH all = {LLH 1 ,
LLH 2 , LLH 3 , LLH 4 } for four steps. The plot shows that the set of non-dominated
solutions contains 3 points marked in black. The best solution achieved by LLH 3 in
the last step is ignored, since that solution is dominated by the first three best solutions
obtained in the first three steps. The first, second and third points in the set of nondominated solutions are associated with {LLH 1 , LLH 2 }, {LLH 1 } and {LLH 1 , LLH 3 },
respectively. Hence, LLH becomes {LLH 1 , LLH 2 , LLH 3 }. The scores of LLH 1 , LLH 2
and LLH 3 are assigned to 3, 1 and 1, respectively, yielding a heuristic selection probability
of 60%, 20% and 20% for each heuristic. More details on the components of the proposed
hyper-heuristic can be found in [26, 25].
objective value
LLH3,
LLH4
LLH2
LLH1
LLH2
LLH4
S'best
LLH1,
LLH2
LLH3
LLH1
LLH4
LLH2
LLH1,
LLH3
input: S'best
LLH4
LLH3
steps
Figure 3: An illustration of how the second stage greedy hyper-heuristic operates.
7. Experiments
In this section we report the results of our computational experiments as well as
discuss the MISTA 2013 Challenge and its outcomes. For additional analysis of the
performance of our and other submitted algorithms we refer to [62].
7.1. Test Instances
In our experiments, we used three sets of MRCPSP instances provided by the organisers of the MISTA 2013 Challenge. The first set (set A) was provided at the beginning of
the competition. The second set (set B) was provided after the qualification phase. The
26
third set (set X) was hidden from the participants of the competition until the announcement of the results; it was used to evaluate the performance of the submitted algorithms.
While designing and tuning our algorithm, we had no information on what properties
would the hidden instances have, which drove our decisions towards self-adaptive methods and rich range of moves.
All the three sets of instances were produced by combining PSPLIB benchmark instances [29] and can be downloaded from allserv.kahosl.be/mista2013challenge. In
Table 1 we provide some properties of all the A, B and X instances.
7.2. Experimental Results
During the competition, the submitted algorithms were tested by running each of
the algorithms on each of the B and X instances 10 times. The results for X (hidden)
instances were used to rank the teams. The detailed discussion of the ranking scheme
and the analysis of the competition results are reported in [62].
To further analyse the performance of our algorithm, we ran it for 5 minutes on each
of the B and X instances 2500 times. We used a machine with similar configuration8 to
that employed by the competition organisers.
The aggregated results of our experiment are reported in Table 2. The average and
best solutions are obtained from our 2500 runs, while the competition (“Comp” column)
result is the best solution found during the final phase of the competition by our or some
other algorithm.
It is worth mentioning here that during the final phase of the competition where the
submitted algorithms were tested on B and X instances, our approach outperformed, on
average, all the other algorithms in 18 out of 20 cases, and produced the best solution
for 17 instances. [62] note the consistency of the performance of our algorithm, and also
provide ranks with several different ranking methodologies, showing that our approach
would win the competition even if the rules of the competition were different.
We conducted an additional experiment to investigate the effect of the runtime on
the ranking results, and so an indication of whether or not the code improvements as
discussed in Section 3.1, might have been the predominant reason, as opposed to the
algorithm itself, of achieving the first rank during the competition. In this experiment,
we ran our algorithm (in the style which was used during the competition) with reduced
time limits. After each run, we determined the rank of our algorithm among all other
competitors (the data required for this ranking was received from competition organisers).
Figure 4 shows the result of the experiment for various time limits ranging from 1 to 25
seconds. For the first two or three seconds the algorithm runs MCTS thus not generating
high quality solutions (recall that the goal of MCTS is to find a high-quality partial
ordering of projects). However, as soon as the memetic algorithm starts, the solution
quality rapidly improves. In fact, our algorithm manages to achieve the first rank after
just 12 seconds. Given that the time limit during the competition was 300 seconds, this
means that the algorithm would have still ranked highest even if it had run 25 times
slower. In contrast the improvements of Section 3.1 did not give that factor of a speed
up. Assuming that the other submissions were reasonably implemented then we conclude
that efficient implementation was not the only factor, but that the algorithm design also
had a major role in the success of the algorithm.
8 Intel
i7 3.2 GHz CPU, 16 GB of RAM, Microsoft Windows 7 x64 operating system
27
q
n
avg
d
avg
|M |
avg
|Pred |
avg
|R ρ |
avg
|R ν |
|G ρ |
avg
Rρ
avg
Cν
avg
Gρ
avg
CPD
H
max
height
avg
SP
A-1
A-2
A-3
A-4
A-5
A-6
A-7
A-8
A-9
A-10
2
2
2
5
5
5
10
10
10
10
20
40
60
50
100
150
100
200
300
300
5.53
4.63
5.51
4.37
5.43
5.13
6.03
5.67
5.61
5.53
3
3
3
3
3
3
3
3
3
3
1.20
1.70
1.73
1.20
1.70
1.73
1.20
1.70
1.73
1.73
1
1
1
1
1
1
0
0
1
1
2
2
2
2
2
2
2
2
2
2
1
1
1
1
1
1
2
2
1
1
18.5
23.5
38.5
15.2
24.0
23.8
0.0
0.0
27.5
25.9
51.3
117.3
154.8
44.9
92.4
175.4
48.4
110.8
168.0
158.2
16.0
23.0
49.0
12.0
13.0
13.0
11.5
22.5
27.0
15.0
14.5
22.5
33.5
14.2
23.0
25.6
16.8
24.6
29.6
30.7
167
303
498
409
844
1166
787
1569
2353
2472
4
8
9
5
8
9
6
9
11
10
0.33
0.34
0.26
0.40
0.32
0.26
0.37
0.31
0.27
0.26
B-1
B-2
B-3
B-4
B-5
B-6
B-7
B-8
B-9
B-10
10
10
10
15
15
15
20
20
20
20
100
200
300
150
300
450
200
400
600
420
5.33
5.67
5.52
5.03
6.02
4.62
4.87
5.48
5.31
5.28
3
3
3
3
3
3
3
3
3
3
1.20
1.70
1.73
1.20
1.70
1.73
1.20
1.70
1.73
1.66
1
0
1
1
1
1
1
0
1
0
2
2
2
2
2
2
2
2
2
2
1
2
1
1
1
1
1
2
1
2
17.1
0.0
28.5
17.5
20.7
25.0
14.7
0.0
26.6
0.0
44.8
94.0
144.4
52.3
99.6
141.8
49.6
104.7
154.8
115.9
11.0
21.0
28.0
10.0
17.0
34.0
10.0
10.0
10.0
18.0
12.9
23.9
29.5
15.8
22.5
31.1
15.4
23.7
30.1
24.5
821
1628
2391
1216
2363
3582
1596
3163
4825
3340
5
7
10
6
8
11
5
8
12
12
0.34
0.28
0.26
0.36
0.31
0.26
0.37
0.30
0.26
0.31
X-1
X-2
X-3
X-4
X-5
X-6
X-7
X-8
X-9
X-10
10
10
10
15
15
15
20
20
20
20
100
200
300
150
300
450
200
400
600
410
5.53
5.53
4.98
5.70
5.52
5.49
5.03
5.53
5.54
5.30
3
3
3
3
3
3
3
3
3
3
1.20
1.70
1.73
1.20
1.70
1.73
1.20
1.70
1.73
1.65
0
1
1
0
1
1
1
1
1
1
2
2
2
2
2
2
2
2
2
2
2
1
1
2
1
1
1
1
1
1
0.0
24.0
27.9
0.0
19.9
24.6
13.9
22.2
23.9
20.0
47.6
105.6
167.0
54.5
100.1
163.7
53.9
104.5
146.3
101.2
12.5
14.0
33.0
13.5
12.0
20.0
10.0
15.0
11.0
10.0
14.9
23.0
29.9
14.9
23.6
29.9
15.0
24.5
28.9
24.1
783
1588
2404
1204
2360
3597
1542
3217
4699
3221
5
8
11
5
8
10
5
8
10
9
0.36
0.32
0.27
0.36
0.32
0.26
0.33
0.32
0.26
0.30
Instance
Table 1: Characteristics of the instances: q is the number of projects, n is the number of
activities, avg d is the average duration of activities in all possible modes, avg |M | is the
average number of modes for each activity, avg |Pred | is the average number of the predecessor activities for each activity, avg |R ρ | is the average number of the local renewable
resources per project, avg |R ν | is the average number of the local non-renewable resources
per project, |G ρ | is the number of global renewable resources, avg Rρ is the average renewable resource capacities, avg C ν is the average non-renewable resource capacities, avg
Gρ is the average global renewable resource capacities, avg CPD is the average CPD per
project, H is the upper bound on the time horizon. The last two columns give an indication of the properties of the network (precedence graph); max height is the length of the
longest path, e.g. see https://users.soe.ucsc.edu/~manfred/pubs/J2.pdf and avg
SP is a measure of the ‘series vs. parallel’ nature [58].
28
TPD
TMS
Instance
Avg
Comp
Best
Avg
Comp
Best
A-1
A-2
A-3
A-4
A-5
A-6
A-7
A-8
A-9
A-10
1
2
0
65
155
141
605
292
208
880
1
2
0
65
153
147
596
302
223
969
1
2
0
65
150
133
590
272
197
836
23
41
50
42
105
808
201
153
128
313
23
41
50
42
105
96
196
155
119
314
23
41
50
42
103
99
190
148
122
303
B-1
B-2
B-3
B-4
B-5
B-6
B-7
B-8
B-9
B-10
352
452
554
1299
832
950
802
3323
4247
3290
349
434
545
1274
820
912
792
3176
4192
3249
345
431
526
1252
807
905
782
3048
4062
3140
128
167
210
283
255
232
232
545
754
455
127
160
210
289
254
227
228
533
746
456
124
158
200
275
245
225
225
523
738
436
X-1
X-2
X-3
X-4
X-5
X-6
X-7
X-8
X-9
X-10
405
356
329
960
1785
730
866
1256
3272
1613
392
349
324
955
1768
719
861
1233
3268
1600
386
345
310
907
1727
690
831
1201
3155
1573
143
164
193
209
373
238
233
288
648
383
142
163
192
213
374
232
237
283
643
381
137
158
187
201
362
226
220
279
632
373
Table 2: Summary of experimental results. (Note that the results for the A instances
are for the algorithm of the first stage of the competition, and are included only for
completeness.)
To illustrate the performance of our approach in time, two boxplots are provided
in Figures 5a and 5b for instances B-1 and X-10, respectively. These instances have
been chosen to demonstrate the performance of our algorithm on relatively small and
large instances. However, our experiments show that the algorithm behaves in a similar
manner with respect to the other instances. The central dividing line of each box in
each of the figures presents the median objective value obtained by our algorithm. The
29
Figure 4: The rank achieved by our algorithm with various time limits in the assumption
that the other algorithms are given 300 seconds. Given just 12 seconds, our algorithm
would already achieve the rank 1.
edges of each box refer to 25th and 75th percentiles while the whiskers (demonstrated by
a + marker) are the extreme objective values which are not considered as outliers. Also,
the curve which passes through the plot demonstrates the average performance of the
algorithm. Our approach continuously makes improvement indicating that the algorithm
does not prematurely converge.
It would obviously be interesting to link together the performance of the algorithm(s)
with the features of the instances; for example, see [50, 37] and others. There are many
potential features that could be used to characterise the instances, and these have often
been used as part of the process of generating instances, for example [30, 12]. For example,
one can characterise the precedence graph in various ways [52, 23, 15] and associated
measures such as ‘complexity’ and ‘Series-Parallel’ indices could also be included. This
would be far outside the scope of this paper, and warrants a future study. However, we
can make some initial observations by analysing the instance properties in Table 1 and
the results in Table 2, and in more detail in Table 33 of the final competition summary
[64]. In particular, we focus on the property of the number of local renewable resources
in the column “avg |R ρ |”, and that there are none for instances B-2, B-8, B-10, X-1 and
X-4. These seem to correlate with the performance in that the algorithm here only fails
to rank first on instances B-2 and X-1, both of which have no local renewable resources.
We remark that having no local renewables is a rather special case, and so if these
were removed the relative performance would have further improved. However, it seems
reasonably likely that the number of local renewable resources would be good instance
feature for selecting algorithms, or for future studies comparing different algorithms in
more details.
7.3. Results on Single Project (PSPLIB) Instances
The main aim of this work is to develop methods for the MISTA benchmark problems
having multiple projects. However, it is also of interest to study the methods on other
scheduling benchmarks; in particular, the PSPLIB library9 of instances [29] has been
widely used and studied in the academic scheduling community. Unfortunately, despite it
being long known that the majority of real world problems are multi-project [35], it seems
9 http://www.om-db.wi.tum.de/psplib/
30
380
average
average
1800
375
1750
objective value
objective value
370
365
360
1700
1650
355
350
1600
345
100
time (sec)
200
300
100
(a) B-1
time (sec)
200
300
(b) X-10
Figure 5: Performance dynamics of the proposed approach. The horizontal axis corresponds to the time spent by the algorithm on optimisation and the vertical axis shows
the solution quality. The results are averaged over 2500 runs and represented in the box
plot form.
that much of the work in project scheduling has been on single project instances, and
PSPLIB does not currently contain any multi-project instances. Despite this limitation
of PSPLIB, it has functioned well as a standard set of benchmarks, and so it is likely to
interest how well the methods of this paper perform on these instances. The most relevant
set of instances in PSPLIB is provided by the multi-mode single project instances10 . The
study of those instances is also limited considering the number of tasks which goes up
to 30 tasks, whereas the MISTA competition included instances with up to 600 tasks.
(Other PSPLIB instances are either much simpler as they are just single mode, or else
introduce extra constraints of objectives for the timing, and so are outside the scope of
this current paper.) Specifically, we studied the J30 instances from PSPLIB; they all
have precisely 3 modes for each of the 30 tasks, with 2 renewable and 2 non-renewable
resources, and we use the 552 of them that are satisfiable (a feasible solution is known).
For the special case of single project, minimising the TPD objective becomes the same
as minimising the makespan, TMS. Hence, the PSPLIB multi-mode instances can be
regarded as a special single-project case of the MISTA formulation and so our methods
could be directly applied to these instances to minimise the makespan. However, doing
so would be rather inefficient as many portions of our methods are designed to handle
multiple-projects and the associated TPD objective. Hence, we partially specialised our
code in order to handle these simpler PSPLIB instances. In particular, we inactivated
the Monte-Carlo Tree Search as it is only there in order to handle the structures arising
from the TPD objective, replacing it with a simple random generation heuristic. We also
inactivated the neighbourhoods that are specific to working on a specific project. Since
making comparisons to previous work also needed a very limited time period and just
a single thread, we also disabled the memetic component as it was designed to exploit
10 http://www.om-db.wi.tum.de/psplib/getdata.cgi?mode=mm
31
multi-cores and longer runs; to partially compensate for the resulting expected loss in
diversity in the search, we also added a simple restart of the search every 10k generated
schedules.
These instances have been studied in various works. [49] used a Genetic algorithm
and also Forward-Backward Iterations (FBI) [34] along with a simple method to repair
infeasible mode choices. [9] again used FBI along with a more expensive SAT-based
method to ensure that only feasible mode decisions were made (recently, [51] reported
that this was the best method on these instances). In these works, the standard fashion
is to impose a limit on the number of schedules generated. The reported performance
measure is then the gap, averaged over all instances, from either a known lower bound
or else from the current “Best Known Solutions (BKS)”. Often, the number of schedules
permitted is rather small, such a 5–10k, which would take our code less than a second,
and so would not really match the intent of our methods. Hence, in Table 3 we compare
with previous results at 50k and 500k from Table 4 of [9]. We also note that with
500k schedules, and under 2s, our heuristic found the best-known solutions in more
than 85% of instances, and in 99% it was within 1 time unit of the BKS. We also
performed much longer runs to exploit the full power of all the algorithm components:
up to 500M generated schedules within the concurrent memetic, though still only taking
up to about 400 seconds, due to the concurrency, and the more efficient implementations
(see Section 3). We were able to find 3 new best solutions; these have been added to
PSPLIB, and so giving the only improvements to these multi-mode J30 instances since
2011.
Schedules generated:
This work
Coelho and Vanhoucke (2011)
Van Peteghem and Vanhoucke (2010)
50k
500k
Gap, %
Time, sec
Gap, %
Time, sec
13.68
12.77
13.31
0.12
25.1
2.46
12.84
12.41
13.09
1.14
210.1
18.03
Table 3: Summary of experimental results on PSPLIB multi-mode J30 instances for 50k
and 500k generated schedules. Giving the average gap from the lower bound (the lower
the better) and the average running time (the lower the better). Note that the runtimes
are only indicative as the older results will be on machines that might well be 2x slower.
Hence, we conclude that the methods here perform competitively on these multi-mode
instances. We believe this gives evidence of the power and flexibility of the overall approach. This is particularly noteworthy, as our methods were originally designed for the
more general multi-project case with the extra delay based objective, which has quite
different properties, as already discussed in Section 4; the difference between minimising
TPD and TMS can lead to a significant difference in the nature of the solutions, and
reasonable algorithms. For example, both of the compared approaches [49, 9] used the
successful forward-backward iterations, however our methods for did not include it, because it is not naturally applicable to the primary objective being TPD. Including it in
our suite of improvement operators, might well improve performance for the special case
of improving makespan.
32
Finally, we remark that recent work on these multi-mode single project instances by
[37] has considered the problem of selecting algorithms. This is perhaps the closest in
spirit to our goals, in that the aim is to take multiple options for algorithms and then to
use intelligent or machine learning techniques in order to make the selection. In contrast
to a more traditional approach in which the exploration of the combinations of many
components is done manually.
8. Conclusions
This paper described the components of a multi-component hybrid approach to the
MRCMPSP. Broadly, our algorithm is built on the serial generation scheme, in which sequences of activities are used to construct final schedules for their quality evaluation. We
use a two-phase construct-and-improve method to search the space of activity sequences,
with a variety of novel contributions, including algorithm mechanisms specifically designed to handle the multi-project structure and associated objectives functions.
Firstly, the primary objective is a sum of completion times of the individual projects,
and we give evidence and arguments that this tends to lead to an imbalance in the
completion times; some projects finish much earlier than others. Hence, the nature of
the problem itself is such that an approximate partial ordering of projects occurs in good
solutions, and this is quite different from the structure expected with makespan as the
primary objective. We expect such approximate partial ordering may well be common
in real-world multi-project problems as well. Accordingly, in the construction phase, a
novel Monte-Carlo Tree Search method is given that creates and selects initial solutions
with an approximately-similar structure.
Secondly, in the improvement phase, multiple novel neighbourhoods are included
that are designed to work at the ‘project-level’ and so enable any needed changing of
the approximate partial ordering of the projects. These are complemented with a wide
range of neighbourhood moves also designed to work at lower levels within projects. The
improvement phase itself is organised and controlled using a novel hybrid of a memetic
algorithm and a hyper-heuristic, and in a fashion that furthermore makes effective use
of a multi-core machine when it is available.
Perhaps a distinguishing characteristic of the methods here is that there are far more
neighbourhoods than in other submissions to the competition, or indeed the literature
in general. From Table 6 in [62], other submissions have around 1–4 neighbourhoods
but we have 13–17 (though many neighbourhoods have different variants, and are used
in different stages, and so there is not a single meaningful count). At first this might
seem to be a step towards extra complexity, and that might make the algorithm more
difficult to use in real-world practice. However, in many problem domains there is a
drive towards a larger number of neighbourhoods, for example, in realistic educational
timetabling where 10–15 neighbourhoods is not uncommon (for example [28, and others]).
Having many neighbourhoods is a natural reaction to the increasing complexity of the
structure of the problem. They are designed to work on each level of the roughly hierarchical structure of many projects with each project containing many activities. We only
eliminated neighbourhoods in a stage when they clearly never helped; instead preferring
to keep neighbourhoods, with the intention that the diversity would help in robustness
and effectiveness (though admittedly at the cost of ‘neatness’). Also, the objective itself
is a type of weighted completion time and so is sensitive to the internal structure of
33
solutions. Hence, it seems reasonable that many neighbourhoods are needed; and this
seems to be be borne out by the much better performance. Furthermore, in the case
of scheduling using the serial generation there is an advantage and opportunity in that
the neighbourhoods are mostly ‘lightweight’ to implement. They work by modifying the
activity sequence, and when the majority of the runtime is spent on the evaluation of a
sequence, then designing and efficiently implementing many different neighbourhoods is
much less onerous then many other problem domains.
Another significant difference to other submissions to the challenge, and as described
in [62], it seems that many others in the competition used forward-backward iteration
(FBI) methods. We did not use it due to the extra complexity, and also uncertainty of
how to best use it for the TPD (weighted completion time) rather than the standard
makespan. Also, the results in [51] said it did not “distinguish between a good or bad
performing metaheuristic”. However, it would be interesting future work to also include
FBI moves; possibly with appropriate modification for the TPD objective.
The effectiveness of the resulting hybrid algorithm was demonstrated when it convincingly won the MISTA 2013 challenge, outperforming, on average, the other algorithms
on 18 of the 20 instances [62]. The many-component and many-neighbourhood structure
was a deliberate decision on our part, and was aimed at giving a more robust solver. In
particular, part of the design goal was that the methods would be robust and so work
well on the hidden instances, and it did indeed do particular well on those; ranking first
on 9 out of the 10 hidden instances. We also observed that the performance was weakest
(though still high ranking) on those instances with no local renewable resources. Such
instances are rather special, and arguably less realistic, however, it does suggest that
there is good potential for future work to investigate the links between features of the
instances and the the algorithm performance. Nevertheless, Section 7.3 gives additional
evidence of the robustness of our algorithm as even a much-reduced version performs
well on the simple single-project but multi-mode instances from PSPLIB.
The broader contribution of this work is to give evidence for the general approach
and methodology of investigating algorithms that consist of many components (such as
many neighbourhoods, but also other aspects) and that aims to exploit the potential
benefits of simple, but carefully controlled, interactions between them. It is possible
that the algorithm could be simplified (e.g. by reducing the number of neighbourhood
moves); however, we suspect that doing so is not worth the risk of it only working well
on the seen set of instances, and would lose its robustness and hence performance on
the unseen instances. Hence, we believe that the work gives evidence that individual
components should not be prematurely discarded. Instead, ultimately, the decision as to
which combination works best over a particular suite should be automated; that is, as a
form of algorithm assembly, for example in the style of [1]. Of course, such assembly can
also use a form of algorithm selection for the appropriate scheduling algorithm, based on
the features of instances (e.g. [50, 37]).
Such ‘automated assembly’ work would be outside the scope of this particular paper;
however, by providing a broad range of options and components that have the proven
potential to give a successful and robust solver, this work hence gives a good foundation
for such studies in the MRCMSP, and other variants of project scheduling, and arguably
intelligently-coordinated metaheuristic methods in general.
Acknowledgements This work was supported in part by EPSRC Grants EP/F033214/1
34
and EP/H000968/1.
References
[1] Leonardo C. T. Bezerra, Manuel López-Ibáñez, and Thomas Stützle. Automatic design of evolutionary algorithms for multi-objective combinatorial optimization. In Thomas Bartz-Beielstein,
Jürgen Branke, Bogdan Filipič, and Jim Smith, editors, PPSN 2014, volume 8672 of Lecture Notes
in Computer Science, pages 508–517. Springer, 2014.
[2] J. Blazewicz, J.K. Lenstra, and A.H.G.Rinnooy Kan. Scheduling subject to resource constraints:
classification and complexity. Discrete Applied Mathematics, 5(1):11 – 24, 1983.
[3] C.B. Browne, E. Powley, D. Whitehouse, S.M. Lucas, P.I. Cowling, P. Rohlfshagen, S. Tavener,
D. Perez, S. Samothrakis, and S. Colton. A survey of Monte Carlo tree search methods. Computational Intelligence and AI in Games, IEEE Transactions on, 4(1):1–43, 2012.
[4] Peter Brucker, Andreas Drexl, Rolf Mohring, Klaus Neumann, and Erwin Pesch. Resourceconstrained project scheduling: Notation, classification, models, and methods. European Journal
of Operational Research, 112(1):3–41, 1999.
[5] E. K. Burke, M. Gendreau, M. Hyde, G. Kendall, G. Ochoa, E. Özcan, and R. Qu. Hyper-heuristics:
A survey of the state of the art. Journal of the Operational Research Society, 2013.
[6] E. K. Burke, E. Hart, G. Kendall, J. Newall, P. Ross, and S. Schulenburg. Hyper-heuristics:
An emerging direction in modern search technology. In F. Glover and G. Kochenberger, editors,
Handbook of Metaheuristics, pages 457–474. Kluwer, 2003.
[7] Edmund K. Burke, Matthew Hyde, Graham Kendall, Gabriela Ochoa, Ender Özcan, and John R.
Woodward. A classification of hyper-heuristics approaches. In Michel Gendreau and Jean-Yves
Potvin, editors, Handbook of Metaheuristics, volume 57 of International Series in Operations Research & Management Science, chapter 15, pages 449–468. Springer, 2nd edition, 2010.
[8] Wang Chen, Yan-jun Shi, Hong-fei Teng, Xiao-ping Lan, and Li-chen Hu. An efficient hybrid
algorithm for resource-constrained project scheduling. Information Sciences, 180(6):1031 – 1039,
2010. Special Issue on Modelling Uncertainty.
[9] J. Coelho and M. Vanhoucke. Multi-mode resource-constrained project scheduling using rcpsp and
sat solvers. European Journal of Operational Research, 213(1):73–82, 2011.
[10] Carlos A. Coello Coello. A comprehensive survey of evolutionary-based multiobjective optimization
techniques. Knowledge and Information Systems, 1(3):269–308, 1999.
[11] N. Damak, B. Jarboui, P. Siarry, and T. Loukil. Differential evolution for solving multi-mode
resource-constrained project scheduling problems. Comput. Oper. Res., 36(9):2653–2659, 2009.
[12] Erik Demeulemeester, Mario Vanhoucke, and Willy Herroelen. RanGen: A random network generator for activity-on-the-node networks. Journal of Scheduling, 6:17–38, 2003.
[13] Benjamin Doerr, Edda Happ, and Christian Klein. Crossover can provably be useful in evolutionary computation. In Proceedings of the 10th Annual Conference on Genetic and Evolutionary
Computation, GECCO ’08, pages 539–546, 2008.
[14] Sonda Elloumi and Philippe Fortemps. A hybrid rank-based evolutionary algorithm applied to
multi-mode resource-constrained project scheduling problem. European Journal of Operational
Research, 205(1):31 – 41, 2010.
[15] Frank Emmert-Streib and Matthias Dehmer. Exploring statistical and population aspects of network
complexity. PLoS One, 7(5), 2012.
[16] Martin Josef Geiger. Iterated variable neighborhood search for the resource constrained multi-mode
multi-project scheduling problem. CoRR, abs/1310.0602, 2013.
[17] Helton Cristiano Gomes, Francisco de Assis das Neves, and Marcone Jamilson Freitas Souza. Multiobjective metaheuristic algorithms for the resource-constrained project scheduling problem with
precedence relations. Computers & Operations Research, 44:92–104, 2014.
[18] J.F. Gonçalves, J.J.M. Mendes, and M.G.C. Resende. A genetic algorithm for the resource constrained multi-project scheduling problem. European Journal of Operational Research, 189(3):1171–
1190, 2008.
[19] Gregory Gutin and Daniel Karapetyan. A memetic algorithm for the generalized traveling salesman
problem. Natural Computing, 9(1):47–60, 2009.
[20] Sönke Hartmann and Dirk Briskorn. A survey of variants and extensions of the resource-constrained
project scheduling problem. European Journal of Operational Research, 207(1):1 – 14, 2010.
[21] Willy Herroelen and Roel Leus. Project scheduling under uncertainty: Survey and research potentials. European Journal of Operational Research, 165(2):289 – 306, 2005.
35
[22] Willy Herroelen, Bert De Reyck, and Erik Demeulemeester. Resource-constrained project scheduling: A survey of recent developments. Computers & Operations Research, 25(4):279 – 302, 1998.
[23] Jerzy Kamburowski, David J. Michael, and Matthias F.M. Stallmann. Minimizing the complexity
of an activity network. Networks, 36(1):47–52, 2000.
[24] Daniel Karapetyan and Gregory Gutin. A new approach to population sizing for memetic algorithms:
a case study for the multidimensional assignment problem. Evolutionary computation, 19(3):345–71,
2011.
[25] Ahmed Kheiri. Multi-stage Hyper-heuristics for Optimisation Problems. PhD thesis, University of
Nottingham, School of Computer Science, 2014.
[26] Ahmed Kheiri and Ender Özcan. An iterated multi-stage selection hyper-heuristic. European
Journal of Operational Research, 250(1):77–90, 2016.
[27] Ahmed Kheiri, Ender Özcan, and Andrew J. Parkes. HySST: Hyper-heuristic search strategies and
timetabling. In Proceedings of the Ninth International Conference on the Practice and Theory of
Automated Timetabling (PATAT 2012), 2012.
[28] Ahmed Kheiri, Ender Özcan, and Andrew J. Parkes. A stochastic local search algorithm with
adaptive acceptance for high-school timetabling. Annals of Operations Research, 239(1):135–151,
2016.
[29] Rainer Kolisch and Arno Sprecher. PSPLIB - a project scheduling problem library : OR software ORSEP operations research software exchange program. European Journal of Operational Research,
96(1):205–216, 1997.
[30] Rainer Kolisch, Arno Sprecher, and Andreas Drexl. Characterization and generation of a general
class of resource-constrained project scheduling problems. Management Science, 41(10):1693–1703,
1995.
[31] Georgios Koulinas, Lazaros Kotsikas, and Konstantinos Anagnostopoulos. A particle swarm optimization based hyper-heuristic algorithm for the classic resource constrained project scheduling
problem. Information Sciences, 277:680 – 693, 2014.
[32] Doreen Krüger and Armin Scholl. A heuristic solution framework for the resource constrained
(multi-)project scheduling problem with sequence-dependent transfer times. European Journal of
Operational Research, 197(2):492–508, 2009.
[33] Thomas S. Kyriakidis, Georgios M. Kopanos, and Michael C. Georgiadis. MILP formulations for
single- and multi-mode resource-constrained project scheduling problems. Computers & Chemical
Engineering, 36(0):369 – 385, 2012.
[34] K.Y. Li and R.J. Willis. An iterative scheduling technique for resource-constrained project scheduling. European Journal of Operational Research, 56(3):370 – 379, 1992.
[35] Antonio Lova, Concepci??n Maroto, and Pilar Tormos. A multicriteria heuristic method to improve resource allocation in multiproject scheduling. European Journal of Operational Research,
127(2):408–424, 2000.
[36] J.J.M. Mendes, J.F. Gon?alves, and M.G.C. Resende. A random key based genetic algorithm for the
resource constrained project scheduling problem. Computers & Operations Research, 36(1):92–109,
2009.
[37] Tommy Messelis and Patrick De Causmaecker. An automatic algorithm selection approach for
the multi-mode resource-constrained project scheduling problem. European Journal of Operational
Research, 233(3):511–528, 2014.
[38] Melanie Mitchell, John H. Holland, and Stephanie Forrest. When will a genetic algorithm outperform hill climbing. In Jack D. Cowan, Gerald Tesauro, and Joshua Alspector, editors, NIPS, pages
51–58. Morgan Kaufmann, 1993.
[39] Rolf H. Möhring, Andreas S. Schulz, Frederik Stork, and Marc Uetz. Solving project scheduling
problems by minimum cut computations. Management Science, 49(3):330–350, 2003.
[40] Pablo Moscato. On evolution, search, optimization, genetic algorithms and martial arts: Towards
memetic algorithms. Caltech concurrent computation program, C3P Report, 826:1989, 1989.
[41] Pablo Moscato and Michael G Norman. A memetic approach for the traveling salesman problem implementation of a computational ecology for combinatorial optimization on message-passing
systems. Parallel Computing and Transputer Applications, 1:177–186, 1992.
[42] M.W. Moskewicz, C.F. Madigan, Ying Zhao, Lintao Zhang, and S. Malik. Chaff: engineering an
efficient sat solver. In Design Automation Conference, 2001. Proceedings, pages 530–535, 2001.
[43] Yew-Soon Ong, Meng-Hiot Lim, Ning Zhu, and Kok-Wai Wong. Classification of adaptive memetic
algorithms: A comparative study. Trans. Sys. Man Cyber. Part B, 36(1):141–152, 2006.
[44] Ender Özcan. Memes, self-generation and nurse rostering. In Edmund K. Burke and Hana Rudova,
editors, Practice and Theory of Automated Timetabling VI, volume 3867 of Lecture Notes in Com-
36
puter Science, pages 85–104. Springer Berlin Heidelberg, 2007.
[45] Ender Özcan, Burak Bilgin, and Emin Erkan Korkmaz. A comprehensive analysis of hyperheuristics. Intelligent Data Analysis, 12(1):3–23, 2008.
[46] Ender Özcan and Ahmed Kheiri. A hyper-heuristic based on random gradient, greedy and dominance. In Erol Gelenbe, Ricardo Lent, and Georgia Sakellari, editors, Computer and Information
Sciences II, pages 557–563. Springer London, 2012.
[47] Ender Özcan, Andrew J. Parkes, and Alpay Alkan. The interleaved constructive memetic algorithm
and its application to timetabling. Comput. Oper. Res., 39(10):2310–2322, October 2012.
[48] Linet Özdamar and Gündüz Ulusoy. A survey on the resource-constrained project scheduling problem. IIE Transactions, 27(5):574–586, 1995.
[49] Vincent Van Peteghem and Mario Vanhoucke. A genetic algorithm for the preemptive and nonpreemptive multi-mode resource-constrained project scheduling problem. European Journal of Operational Research, 201(2):409 – 418, 2010.
[50] Vincent Van Peteghem and Mario Vanhoucke. Using resource scarceness characteristics to solve the
multi-mode resource-constrained project scheduling problem. Journal of Heuristics, 17(6):705–728,
2011.
[51] Vincent Van Peteghem and Mario Vanhoucke. An experimental investigation of metaheuristics for
the multi-mode resource-constrained project scheduling problem on new dataset instances. European
Journal of Operational Research, 235(1):62–72, 2014.
[52] Bert De Reyck and Willy Herroelen. On the use of the complexity index as a measure of complexity
in activity networks. European Journal of Operational Research, 91(2):347–366, 1996.
[53] P. Ross. Hyper-heuristics. In E. K. Burke and G. Kendall, editors, Search Methodologies: Introductory Tutorials in Optimization and Decision Support Techniques, chapter 17, pages 529–556.
Springer, 2005.
[54] Kumara Sastry, David E. Goldberg, and Graham Kendall. Genetic algorithms. In Edmund K.
Burke and Graham Kendall, editors, Search Methodologies, pages 93–118. Springer US, 2014.
[55] M H Sebt, Y Alipouri, and Y Alipouri. Solving resource-constrained project scheduling problem
with evolutionary programming. Journal of the Operational Research Society, 64:1327–1335, 2013.
[56] Dirk Sudholt. Crossover speeds up building-block assembly. In Proceedings of the 14th Annual
Conference on Genetic and Evolutionary Computation, GECCO ’12, pages 689–702, New York,
NY, USA, 2012. ACM.
[57] Túlio Toffolo, Haroldo Santos, Marco Carvalho, Janniele Soares, Tony Wauters, and Greet Vanden Berghe. An integer programming approach to a generalized project scheduling problem. In
Matheuristics. Hamburg, 11-13 June 2014, 2014.
[58] Mario Vanhoucke. An overview of recent research results and future research avenues using simulation studies in project management. ISRN Computational Mathematics, 2013, 2013.
[59] Mario Vanhoucke, Erik Demeulemeester, and Willy Herroelen. An exact procedure for the resourceconstrained weighted earliness?tardiness project scheduling problem. Annals of Operations Research, 102:179–196, 2001.
[60] Ling Wang and Chen Fang. An effective shuffled frog-leaping algorithm for multi-mode resourceconstrained project scheduling problem. Information Sciences, 181(20):4804 – 4822, 2011. Special
Issue on Interpretable Fuzzy Systems.
[61] Ling Wang and Chen Fang. An effective estimation of distribution algorithm for the multi-mode
resource-constrained project scheduling problem. Computers & Operations Research, 39(2):449 –
460, 2012.
[62] Tony Wauters, Joris Kinable, Pieter Smet, Wim Vancroonenburg, Greet Vanden Berghe, and Jannes
Verstichel. The multi-mode resource-constrained multi-project scheduling problem. Journal of
Scheduling, 19(3):271–283, 2016.
[63] Tony Wauters, Katja Verbeeck, Patrick De Causmaecker, and Greet Vanden Berghe. A learningbased optimization approach to multi-project scheduling. Journal of Scheduling, 18(1):61–74, 2015.
[64] Tony Wauters, Jannes Verstichel, Katja Verbeeck, and Greet Vanden Berghe. A learning metaheuristic for the multi mode resource constrained project scheduling problem. In Learning and
Intelligent OptimizatioN (LION 2009): Trento, 2009.
[65] Jan Weglarz, Joanna Józefowska, Marek Mika, and Grzegorz Waligóra. Project scheduling with
finite or infinite number of activity processing modes - a survey. European Journal of Operational
Research, 208(3):177–205, 2011.
37
| 8cs.DS
|
arXiv:1802.10277v1 [math.AC] 28 Feb 2018
DEGENERATIONS OVER (A∞ )-SINGULARITIES AND CONSTRUCTION OF
DEGENERATIONS OVER COMMUTATIVE RINGS
NAOYA HIRAMATSU, RYO TAKAHASHI, AND YUJI YOSHINO
Abstract. We give a necessary condition of degeneration via matrix representations, and consider
degenerations of indecomposable Cohen-Macaulay modules over hypersurface singularities of type (A∞ ).
We also provide a method to construct degenerations of finitely generated modules over commutative
rings.
1. Introduction
The notion of degenerations of modules has been a central subject in the representation theory of
algebras. Let k be a field, and let R be a k-algebra. Let M, N be R-modules with the same finite dimension
as k-vector spaces. We say that M degenerates to N if in the module variety the point corresponding
to M belongs to the Zariski closure of the orbit of the point corresponding to N . This definition is only
available for modules of finite length. The third author [9] thus introduces a scheme-theoretic definition of
degenerations for arbitrary finitely generated modules, and extends the characterization of degenerations
due to Riedtmann [5] and Zwara [12]. Interest of this subject is to describe relations of degenerations of
modules [2, 5, 8, 9, 11, 12]. In [3], the first and third authors give a complete description of degenerations
over a ring of even-dimensional simple hypersurface singularity of type (An ).
The first purpose of this paper is to give a necessary condition for the degenerations of maximal CohenMacaulay modules by considering it via matrix representations over the base regular local ring. As an
application, we give the description of degenerations of indecomposable Cohen-Macaulay modules over
hypersurface singularities of type (A∞ ).
Theorem 1.1. Let k be an algebraically closed field.
(1) Let R = k[[x, y]]/(x2 ) be a hypersurface of dimension one. Then (x, y a )R degenerates to (x, y b )R if
and only if a ≤ b and a ≡ b mod 2.
(2) Let R = k[[x, y, z]]/(xy) be a hypersurface of dimension two. Then for all a < b, (x, z a )R and (y, z a )R
do not degenerate to (x, z b )R and (y, z b )R, respectively.
The proof of this theorem requires the fact that the Bass-Quillen conjecture holds for a formal power
series ring over a field ([4, Chapter V Theorem 5.1]), which is the only place where the assumption that R
is complete and equicharacteristic is necessary. Our proof also requires the base field to be algebraically
closed. Note that a complete equicharacteristic local hypersurface R has countable Cohen-Macaulay
representation type if and only if R has either of the two forms in Theorem 1.1.
The second purpose of this paper is to provide a method to construct degenerations. We obtain the
following theorem.
Theorem 1.2. Let R be a commutative noetherian algebra over a field, and let L be a finitely generated
R-module. Let α be an endomorphism of L with Im α = Ker α, and let x ∈ R be an L-regular element.
Then the following statements hold.
(1) Any submodule M of L containing α(L) + xL degenerates to N := α(M ) + x2 L.
(2) If L, M are Cohen-Macaulay, then so is N .
Date: March 1, 2018.
Key words and phrases. degeneration of modules, (maximal) Cohen-Macaulay module, hypersurface singularity of type
(A∞ ), countable Cohen-Macaulay representation type, Knörrer’s periodicity.
2010 Mathematics Subject Classification. 13C14, 14D06, 16G60.
NH was supported by JSPS Grant-in-Aid for Young Scientists 15K17527. RT was supported by JSPS Grant-in-Aid for
Scientific Research 16K05098. YY was supported by JSPS Grant-in-Aid for Scientific Research 26287008.
1
2
NAOYA HIRAMATSU, RYO TAKAHASHI, AND YUJI YOSHINO
This theorem is actually a corollary of our main result in this direction, which holds over arbitrary
commutative rings and is proved by elementary calculations of matrices; we do not use matrix representations. We also give various applications, one of which gives an alternative proof of the ‘if’ part of
Theorem 1.1(1).
The organization of this paper is as follows. In Section 2, we review the notion of degenerations
of Cohen-Macaulay modules by means of matrix representations, and give a necessary condition for
degenerations (Corollary 2.8). Using this, in Section 3, we describe degenerations over hypersurface
singularities of type (A∞ ) with dimension at most two (Theorems 3.1 and 3.2). In Section 4, we give
several ways to construct degenerations (Theorem 4.2), and provide applications. In Section 5 we attempt
to extend Theorem 1.1 to higher dimension (Example 5.5).
2. Matrix representation
We begin with recalling the definition of degenerations of finitely generated modules. For details, we
refer the reader to [8, 9, 10].
Definition 2.1. Let R be a noetherian algebra over a field k. Let M, N be finitely generated left Rmodules. We say that M degenerates to N (or N is a degeneration of M ) if there exists a discrete
valuation k-algebra (V, tV, k) and a finitely generated left R ⊗k V -module Q such that Q is flat as a
V -module, Q/tQ ∼
= M ⊗k Vt as an R ⊗k Vt -module. When this is the case,
= N as an R-module, and Qt ∼
we say that M degenerates to N along V .
Remark 2.2. Let R be a noetherian algebra over a field k and let M , N and L be finitely generated left
R-modules.
(1) M degenerates to N if and only if there exists a short exact sequence of finitely generated left
R-modules
h
g
0 −−−−→ Z −−−−→ M ⊕ Z −−−−→ N −−−−→ 0,
where the endomorphism g of Z is nilpotent. See [12, Theorem 1] and [9, Theorem 2.2]. Note that if M
and N is Cohen-Macaulay, then so is Z (see [9, Remark 4.3]).
(2) Suppose that there is an exact sequence 0 → L → M → N → 0. Then M degenerates to L ⊕ N . See
[9, Remark 2.5] for the detail.
(3) Assume that R is commutative and suppose that M degenerates to N . Then the ith Fitting ideal of
M contains that of N for all i ≥ 0. Namely, denoting the ith Fitting ideal of an R-module M by FiR (M ),
we have FiR (M ) ⊇ FiR (N ) for all i ≧ 0. See [10, Theorem 2.5].
Remark 2.3. We can always take the localization k[t](t) of a polynomial ring as a discrete valuation ring
V as in Definition 2.1. Moreover, let T = k[t]\(t) and T ′ = k[t]\{0}. Then we also have R⊗k V = T −1 R[t]
and R ⊗k Vt = T ′−1 R[t]. See [9, Corollary 2.4].
Now let us recall the definition of a matrix representation.
Definition 2.4. Let R be a commutative noetherian local ring. Let S be a Noether normalization of R,
that is, S is a subring of R which is a regular local ring such that R is a finitely generated S-module.
Then R is Cohen-Macaulay if and only if R is S-free, and a finitely generated R-module is (maximal)
Cohen-Macaulay1 if and only if M is S-free.
Let R, M be Cohen-Macaulay. Then M ∼
= S n for some n ≥ 0, and we have a k-algebra homomorphism
∼ S n×n .
R → EndS (M ) =
This is called a matrix representation of M over S.
Example 2.5. (1) Let S be a regular local ring and R = S[[x]]/(f ) where f = xn + a1 xn−1 + · · · + an
with ai ∈ S. Then R is a Cohen-Macaulay ring with an S-basis {1, x, x2 , ..., xn−1 } and each matrix
representation of Cohen-Macaulay R-module M comes from the action of x on M .
(2) Let R = k[[x, y]]/(x2 ) with a field k of characteristic not two. It is known that the isomorphism classes
of indecomposable Cohen-Macaulay R-modules are the following:
R,
R/(x),
(x, y n ) n ≥ 1.
1Throughout this paper, “Cohen-Macaulay” means “maximal Cohen-Macaulay”.
DEGENERATIONS OF MODULES OVER COMMUTATIVE RINGS
3
Giving the matrix representations over k[[y]] of these modules is equivalent to giving the following squarezero matrices; see also [7, (6.5)].
0 yn
n ≥ 1.
( 00 10 ) , (0),
0 0
Proposition 2.6. Let (R, m, k) be a commutative noetherian complete equicharacteristic local ring. Then
one can take a formal power series ring S over k as a Noether normalization of R. Assume that k is
algebraically closed. Let M, N be Cohen-Macaulay R-modules such that M degenerates to N along V . Let
Q be a finitely generated R⊗k V -module which gives the degeneration. Then Q is free as an S⊗k V -module.
Proof. Since V = T −1 k[t] where T = k[t]\(t) by Remark 2.3, we can take a finitely generated R[t]submodule Q′ of Q such that T −1 Q′ = Q. Then Q′ is flat over k[t], Q′0 ∼
= M for each
= N and Q′c ∼
×
′
′
′
c ∈ k . Here Qc is defined to be Q /(t − c)Q for an element c ∈ k; see [9, Theorem 3.2]. We show that
Q′ is free as an S[t]-module. Set T ′ = k[t]\{0} and p be a prime ideal of S[t].
Suppose T ′ ∩ p 6= ∅. Since k is an algebraically closed field, there is c ∈ k such that t − c ∈ p. Then,
Q′p /(t − c)Q′p ∼
= (Q′ /(t − c)Q′ ) p .
By the choice of Q′ , the right-hand side is isomorphic to N p or M p , so that S p -free. Thus Q′p /(t − c)Q′p
is free as an S p = (S[t]/(t − c)) p -module. Thus, Q′p is a Cohen-Macaulay module over the regular local
ring S[t]p . Hence, Q′p is free as an S[t]p -module.
Suppose T ′ ∩ p = ∅. Then
Q′p ∼
= (M ⊗k Vt )p′ .
= (T ′−1 Q)p′ ∼
= (T ′−1 (T −1 Q′ )p′ ∼
= (T ′−1 Q′ )p′ ∼
Hence Q′p is a free S[t]p -module.
Therefore Q′ is a projective S[t]-module. By the fact on the Bass-Quillen conjecture [4, Chapter V
Theorem 5.1], each projective S[t]-module is S[t]-free since S is a formal power series ring. Hence Q′ is
free as an S[t]-module, so that Q = T −1 Q′ is free as a T −1 S[t] = S ⊗k V -module.
The above proposition enables us to consider the matrix representation of Q over S ⊗k V because it is
free. For matrix representations µ, ν : R → S n×n , we write µ ∼
= ν if there exists an invertible matrix α
such that α−1 µ(r)α = ν(r) for each r ∈ R. Such a matrix α is called an interwining matrix of µ and ν.
Corollary 2.7. Let R, k, S, V be as in Proposition 2.6. Let M and N be Cohen-Macaulay R-modules
and µ and ν the matrix representations of M and N over S respectively. Then M degenerates to N
along V if and only if there exists a matrix representation ξ over S ⊗k V such that ξ ⊗V V /tV ∼
= ν and
µ
⊗
V
hold.
ξ ⊗V Vt ∼
=
k t
Under the situation of the proposition above, R is a finitely generated S-module. Set u1 , ..., ut be
generators of R as an S-module. Then giving a matrix representation µ over S is equivalent to giving a
tuple of matrices (µ1 , ..., µt ), where µi = µ(ui ) for i = 1, ..., t.
Corollary 2.8. Let R, k, S, V be as in Proposition 2.6 and let M be a Cohen-Macaulay R-module. Let
Q be a free S ⊗k V -module and suppose that Qt is isomorphic to M ⊗k Vt . We denote by (µ1 , . . . , µt )
(resp. (ξ1 , . . . , ξt )) the matrix representation of M (resp. Q) over S (resp. S ⊗k V ). Then we have the
following equalities in S ⊗k V for i = 1, . . . , t:
(1) tr(ξi ) = tr(µi ) and det(ξi ) = det(µi ),
′
(2) For all j ≥ 0, there exist l, l′ such that Ij (ξi ) = tl Ij (µi ) and tl Ij (ξi ) = Ij (µi ).
Here, Ij (−) stands for the ideal generated by the j-minors.
Proof. By the assumption, there exists an invertible matrix α with entries in S ⊗k Vt such that α−1 ξi α =
µi . Thus tr(ξi ) = tr(µi ) and det(ξi ) = det(µi ). Moreover we have Ij (ξi ) = Ij (µi ) in S ⊗k Vt . Hence we
have the equalities in the lemma.
Taking Corollary 2.7 into account, Corollary 2.8 gives a necessary condition of the degeneration. In
the next section, we apply the condition to (Ad∞ )-singularities of dimension at most two.
4
NAOYA HIRAMATSU, RYO TAKAHASHI, AND YUJI YOSHINO
3. Degenerations of Cohen-Macaulay modules over (A∞ )-singularities
Recall that a Cohen-Macaulay local ring R is said to have countable Cohen-Macaulay representation
type if there exist infinitely but only countably many isomorphism classes of indecomposable CohenMacaulay R-modules. Let R be a complete equicharacteristic local hypersurface with residue field k of
characteristic not two. Then R has countable Cohen-Macaulay representation type if and only if R is
isomorphic to the ring k[[x0 , x1 , x2 , · · · , xd ]]/(f ), where f is either of the following:
(
x21 + · · · + x2d
(A∞ ),
f=
x20 x1 + x22 + · · · + x2d (D∞ ).
Moreover, when this is the case, the indecomposable Cohen-Macaulay R-modules are classified completely;
we refer the reader to [1, §1] for more information.
In this section, we describe the degenerations of indecomposable Cohen-Macaulay R-modules in the
case where R has type (A∞ ) and dim R = 1, 2. Throughout the rest of this section, we assume that k is
an algebraically closed field of characteristic not two. Let us start by the 1-dimensional case.
Theorem 3.1. Let R = k[[x, y]]/(x2 ). Then (x, y a ) degenerates to (x, y b ) if and only if a ≤ b and a ≡ b
mod 2.
Proof. First we notice that a ≤ b if (x, y a ) degenerates to (x, y b ) by Remark
2.2(3). As mentioned in
n
Example 2.5(2), the matrix representations of (x, y n ) for n ≥ 0 are 00 y0 . Here (x, y 0 ) = R. Suppose
that a ≡ b mod 2. We consider R ⊗k V -module Q whose matrix representation ξ over S ⊗k V is
a+b
yb
ty 2
.
ξ=
a+b
2 a
−t y
−ty
2
Then one can show that Q gives the degeneration.
To show the converse, we prove the following claim.
2m+1
, that is, R never degenerates to (x, y 2m+1 ).
Claim. ( 00 10 ) never degenerates to 00 y 0
Proof of Claim. After applying elementary row and column operations, we may assume that the matrix
representation of Q over S ⊗k V which gives the degeneration is of the form:
2m+1
y 2m+1 +tβ
.
+ t αγ βδ = tα
ξ = 00 y 0
tγ
tδ
Then ξ ⊗ Vt ∼
= µ ⊗ Vt . By Corollary 2.8, we have tδ = −tα since tr(ξ) = tr(µ) = 0. Moreover,
det ξ = −t2 α2 − tγ(y 2m+1 + tβ) = 0,
(3.1)
I1 (ξ) = (tα, tγ, y 2m+1 + tβ) ⊇ (tl ) for some l.
Note that the above equalities are obtained in S ⊗k V . From the equation (3.1), we have tα2 = γ(y 2m+1 +
tβ). Since t does not divide y 2m+1 + tβ, t divides γ, so that γ = tγ ′ for some γ ′ ∈ S ⊗k V . Hence we also
have the equality in S ⊗k V :
α2 = γ ′ (y 2m+1 + tβ).
(3.2)
Since S is factorial, so is S[t]. Thus S ⊗k V = T −1 S[t] is also factorial. Take the unique factorization
into prime elements of y 2m+1 + tβ:
y 2m+1 + tβ = P1e1 P2e2 · · · Pnen .
Then, there exists i such that ei is an odd number. Since the equation (3.2) holds, Pi divides α, so that
Pi also divides γ ′ . Therefore, (Pi ) ⊇ Ii (ξ) ⊇ (tl ), so that Pi = t. This makes contradiction since t cannot
divide y 2m+1 + tβ.
♦
a
b
Now we assume that 00 y0 degenerates to 00 y0 . Then we also have the following matrix representation of Q over S ⊗k V :
b
β
y b +tβ .
= tα
ξ = 00 y0 + t αγ −α
tγ −tα
Then we also have t2 α2 = tγ(y b + tβ), (tα, tγ, y b + tβ) ⊇ tl (y a ) for some l, and
(3.3)
′
tl (tα, tγ, y b + tβ) ⊆ (y a ) for some l′ .
DEGENERATIONS OF MODULES OVER COMMUTATIVE RINGS
5
From (3.3), we see that y a divides tα, tγ and y b + tβ. This yields that
2
t2 α′ = tγ ′ (y b−a + tβ ′ ) and (tα′ , tγ ′ , y b−a + tβ ′ ) ⊇ (tl ) for some l,
where α = y a α′ , γ = y a γ ′ and y b + tβ = y a (y b−a + tβ ′ ). If a 6≡ b mod 2, b − a is an odd number. As in
the proof of the claim above, this never happen. Therefore a ≡ b mod 2.
Next we consider the 2-dimensional case, that is, let R = k[[x, y, z]]/(xy). The nonisomorphic indecomposable Cohen-Macaulay R-modules are R and the following ideals; see [1, Proposition 2.2].
(x),
(y),
(x, z n ),
(y, z n ) n ≥ 1.
Theorem 3.2. Let R = k[[x, y, z]]/(xy). Then (x, z a ) (resp. (y, z a )) never degenerates to (x, z b ) and
(y, z b ) for all a < b.
Proof. Replacing Y with x − y, we may consider k[[x, Y, z]]/(x2 − Y x) as R. We consider the ideals (x, z a )
and (x − Y, z a ) instead of the ones in the assertion.
We rewrite
Y by y. The matrix representations of
a
a
(x, z a ) and (x − y, z a ) over S = k[[y, z]] are y0 z0 and 00 zy respectively. See Remark 3.3 below.
a
b
Suppose that µ = y0 z0 degenerates to y0 z0 along V , and let Q be a finitely generated R ⊗k V module which gives the degeneration. By elementary row and column operations, we may assume that
the matrix representation ξ of Q over S ⊗k V which gives the degeneration is of the form:
b
z b +tβ
.
ξ = y0 z0 + t αγ βδ = y+tα
tγ
tδ
Since ξ ⊗ Vt ∼
= µ ⊗ Vt , by Corollary 2.8, we have
(3.4)
tr(ξ) = y,
b
(3.5)
(3.6)
det(ξ) = (y + tα)tδ − tγ(z + tβ) = 0,
I1 (ξ) = (y + tα, tγ, z b + tβ, tδ) ⊇ tl (z a , y) for some l.
By (3.4) and (3.5), we have
(3.7)
−α(y + tα) = γ(z b + tβ).
Set P = z b + tβ. Then, since P ≡ z b modulo t, P divides α. Consider the equation (3.7) in S ⊗k V /y(S ⊗k
V ):
−tα · α = γ · z b + tβ.
Since P divides α, P divides α. Thus P also divides γ. This yields that I1 (ξ) = (y +tα, tγ, z b +tβ, −tα) ⊆
(P, y). By (3.6), tl (z a , y) ⊆ (P, y). Let p be a minimal prime ideal of (P, y). Notice that p has height 2
and contains tl z and y. Suppose that p contains t. Then p also contains z since p contains P = z b + tβ,
so that p = (z, y, t). This makes contradiction. Hence p = (z, y). Consider the above inclusion of ideals
in S ⊗k V /y(S ⊗k V ) again, and then we see that (P, y) = (P ) ∋ tl z a . Since (P ) are also contained
in p = (z), P = P ′ z for some P ′ ∈ S ⊗k V /y(S ⊗k V ). Thus (P ) = (P ′ z) ∋ tl z a , which implies that
(P ′ ) ∋ tl z a−1 as an ideal in S ⊗k V /y(S ⊗k V ). Note that P ′ = z b−1 + tP ′′ in S ⊗k V and the minimal
prime ideal of (P ′ ) in S ⊗k V /y(S ⊗k V ) is also (z). Hence we can repeat the procedure and we obtain
the inclusion of ideals in S ⊗k V /y(S ⊗k V ):
(z b−a + tP ′′′ ) ⊇ (tl ).
However this never happen if b − a > 0. Therefore we obtain the conclusion. The same proof is valid for
showing the remaining claim that (x, z a ) (resp. (y, z a ) ) does not degenerate to (y, z b ) (resp. (x, z b ) or
(y, z b )).
Remark 3.3. Let R = k[[x, y, z]]/(x2 −xy) and let M = (x, z n ), N = (x−y, z n ) be ideals. Set S = k[[y, z]].
Note that the matrix representations of M and N over S are obtained from the action of x on M and N
respectively. Since M has a basis x and z n as an S-module, that is, M ∼
= xS ⊕ z n S, the multiplication
map ax by x on M induces the correspondence:
ax
ayx+bz n x
n ) 7→
.
ax : xS ⊕ z n S → xS ⊕ z n S; ( bz
0
6
NAOYA HIRAMATSU, RYO TAKAHASHI, AND YUJI YOSHINO
Hence the matrix representation of M over S is
S-module, the multiplication map bx on N induces
y zn
0 0
. Similarly, since N ∼
= (x − y)S ⊕ z n S as an
bx : (x − y)S ⊕ z n S → (x − y)S ⊕ z n S;
n
so that the matrix representation of N is 00 zy .
a(x−y)
bz n
7→
bz n (x−y)
byz n
,
Remark 3.4. Araya, Iima and the second author [1] show that the Cohen-Macaulay modules which
appear in Theorem 3.1 (resp. Theorem 3.2) are obtained from the extension by R/(x) and itself (resp.
R/(x) and R/(y)). In the case, we have the degeneration by Remark 2.2(2). Summing up this fact,
we obtain complete description of the degenerations of indecomposable Cohen-Macaulay modules over
k[[x, y]]/(x2 ) and k[[x, y, z]]/(xy).
4. Construction of degenerations of modules
In this section we provide a construction of a nontrivial degeneration over an arbitrary commutative
noetherian algebra over a field (without matrix representations). We begin with establishing a lemma.
β
α
β
α
Lemma 4.1. Let R be a commutative ring, and let · · · −
→L−
→L−
→L−
→ · · · be an exact sequence of
R-modules. Let x ∈ R, and set Z = α(L) + xL and W = β(L) + xL.
(1) There is an exact sequence
( α0 βx )
· · ·❇
❇❇
❇❇
Z
β −x
0 α
// L ⊕ L
;;
■■
■■
✈✈
✈
■$$
✈
✈
W
( α0 βx )
// L ⊕ L
::
❍❍
✉✉
❍❍
✉
✉
❍##
✉
Z
β −x
0 α
// L ⊕ L
;;
■■
■■
✈✈
✈
■$$
✈
✈
W
// · · · .
==
④④
④
④
(2) Suppose that x is L-regular. Then there are exact sequences:
β −x
0 α
(α x)
L ⊕ L −−−−−→ L ⊕ L −−−−→ L → L/Z → 0,
( α0 βx )
( β −x )
L ⊕ L −−−−→ L ⊕ L −−−−−→ L → L/W → 0.
Proof. (1) Clearly, ( α0 βx ) β0 −x
= 0. Let s, t ∈ L with ( α0 βx ) ( st ) = ( 00 ). Then α(s) + xt = 0 = β(t),
α
and t ∈ ker β = Im α. We have t = α(u) for some u ∈ L, and 0 = α(s) +xt = α(s+ xu), which implies
v
s + xu ∈ ker α = Im β. Writing s + xu = β(v) with v ∈ L, we get ( st ) = β(v)−xu
= β0 −x
( u ). Thus
α(u)
α
α x
α x
β −x
β −x
Im 0 α = ker ( 0 β ). A symmetric argument shows Im ( 0 β ) = ker 0 α .
(2) Note that ( α x ) is a submatrix of ( α0 βx ); we have ( α x ) β0 −x
= 0. Let s, t ∈ L with ( α x ) ( st ) =
α
0
( 0 ). Then α(s) + xt = 0, and 0 = β(α(s) + xt) = xβ(t).
Since x is L-regular, β(t) = 0. The
β −x
=
ker
( α x ). A symmetric argument shows
.
Thus
Im
above argument shows ( st ) ∈ Im β0 −x
0 α
α
α x
Im ( 0 β ) = ker ( β −x ).
Recall that a finitely generated module M over a commutative noetherian ring R is Cohen-Macaulay
if depthRp Mp ≥ dim Rp for all prime ideals p of R. The main result of this section is the following.
α
β
α
β
Theorem 4.2. Let R be a commutative ring, and let · · · −
→L−
→L−
→L−
→ · · · be an exact sequence
of R-modules. Let x be an L-regular element such that β(L) ⊆ α(L) + xL =: Z. Let M be an R-module
with Z ⊆ M ⊆ L, and set N = β(M ) + xZ.
(1) The sequence
θ
η
( β −x )
0 → Z −−−→ M ⊕ Z −−−−−→ N → 0
is exact, where θ is the inclusion map and η(α(s) + xt) = β(t) for s, t ∈ L.
(2) Suppose that R contains a field, Z, M are finitely generated and β is nilpotent. Then M degenerates
to N .
(3) Assume R is noetherian. If L, M are Cohen-Macaulay, then so are Z, N .
DEGENERATIONS OF MODULES OVER COMMUTATIVE RINGS
7
Proof. (1) As the image of β is contained in Z, so is that of η. Take any element z ∈ Z. Then z = α(s)+xt
for some s, t ∈ L, and we have β(z) = β(α(s) + xt) = xβ(t). Since x is L-regular and β(t) ∈ β(L) ⊆ Z,
the assignment z 7→ β(t) gives a map η : Z → Z, and
β(z) = xη(z) for all z ∈ Z.
As θ is injective, so is the map
, whose image is contained in the kernel of the map ( β −x ) because
α(s)+xt
θ
( β −x ) η (α(s) + xt) = ( β −x )
= β(xt) − xβ(t) = 0. The definition of N shows that ( β −x )
β(t)
is surjective.
It remains to show that the kernel of ( β −x ) is contained in the image of ηθ . Let ( m
z ) ∈ M ⊕Z
m
β −x ) ( m
such
that
(
)
=
0.
Lemma
4.1
implies
that
there
exist
s,
t
∈
L
such
that
(
)
=
( α0 βx ) ( st ) =
z
z
α(s)+xt
= ηθ (α(s) + xt). Thus we are done.
β(t)
(2) The module N is finitely generated, and β n = 0 for some n > 0. It follows by (4.1) that 0 =
n
β (z) = (xη)n (z) = xn η n (z) for all z ∈ Z. As xn is L-regular, we get η n (z) = 0, that is, η is nilpotent.
We see from (1) and Remark 2.2(1) that M degenerates to N .
(3) Set W = β(L) + xL. Combining (1) and Lemma 4.1, we get a commutative diagram
(4.1)
θ
η
( β −x )
0 −−−−→ Z −−−−→ L ⊕ L −−−−−→ L −−−−→ L/W −−−−→ 0
x
i 0
0 j
θ
η
( β −x )
0 −−−−→ Z −−−−→ M ⊕ Z −−−−−→ L −−−−→ L/N −−−−→ 0
with exact rows, where i, j are inclusion maps. The snake lemma gives an exact sequence
(4.2)
0 → L/M ⊕ L/Z → L/N → L/W → 0.
From now on, assuming that R is noetherian, we show that if L, M are Cohen-Macaulay, then so are
Z, N . For this, we may assume that R is a local ring of Krull dimension d. The first exact sequence
in Lemma 4.1 shows that Z, W are Cohen-Macaulay. The depth lemma implies that the R-modules
L/M, L/Z, L/W have depth at least d − 1. It follows from (4.2) that L/N also has depth at least d − 1,
which implies that N is Cohen-Macaulay.
As a direct corollary of the theorem, we obtain the following result.
Corollary 4.3. Let R be a commutative noetherian ring containing a field, and let L be a finitely
generated R-module. Let α be an endomorphism of L with Im α = ker α, and let x ∈ R be an L-regular
element. Then the following statements hold.
(1) An R-module M with α(L) + xL ⊆ M ⊆ L degenerates to N := α(M ) + x2 L.
(2) If L, M are Cohen-Macaulay, then so is N .
Proof. Note α2 = 0. Letting α = β in Theorem 4.2 shows that M degenerates to α(M ) + xα(L) + x2 L,
and if the former is Cohen-Macaulay, then so is the latter. As xL ⊆ M , we have xα(L) = α(xL) ⊆ α(M ).
Hence α(M ) + xα(L) + x2 L = N .
From now on, we give applications of Corollary 4.3.
α
β
α
Corollary 4.4. Let R be a commutative noetherian ring containing a field, and let · · · −
→L−
→L−
→
β
L−
→ · · · be an exact sequence of R-modules. Let x be an L-regular element, and set Z = α(L) + xL and
W = β(L) + xL. Let M, N be R-submodules of L containing Z, W respectively. Then M ⊕ N degenerates
to K := (α(N ) + x2 L) ⊕ (β(M ) + x2 L). If L, M and N are Cohen-Macaulay, then so is K.
Proof. The endomorphism γ = β0 α0 : L⊕2 → L⊕2 is such that Im γ = ker γ. Putting X = {( m
n) | m ∈
M, n ∈ N }, we have γ(L⊕2 ) + xL⊕2 ⊆ X ⊆ L⊕2 . Note that X ∼
= K. Now
= M ⊕ N and γ(X) + x2 L⊕2 ∼
the assertion follows from Corollary 4.3.
Remark 4.5. Let R be a hypersurface local ring. Let X be a Cohen-Macaulay R-module. Then X has
β
α
α
a free resolution of the form · · · −
→F −
→F −
→ F → X → 0, and one can apply Corollary 4.4 to L = F .
8
NAOYA HIRAMATSU, RYO TAKAHASHI, AND YUJI YOSHINO
Corollary 4.6. Let R be a commutative noetherian ring containing a field and L a finitely generated
R-module. Let α be an endomorphism of L with Im α = ker α, and let x ∈ R be an L-regular element.
Then α(L) + xj L degenerates to α(L) + x2i−j L for all i ≥ j ≥ 0. If L is Cohen-Macaulay, then so are
α(L) + xj L and α(L) + x2i−j L.
Proof. We have α(L) + xi L ⊆ α(L) + xj L ⊆ L. Corollary 4.3 implies that α(L) + xj L degenerates to
α(α(L) + xj L) + x2i L = xj α(L) + x2i L ∼
= α(L) + x2i−j L, where the isomorphism follows from the fact
that xj is L-regular. Fix h ≥ 0. Lemma 4.1 implies that C := α(L) + xh L is isomorphic to the cokernel
α −xh
α −xh
α xh
h
0 α
0 α
0 α
2
2
α
−x
of 0 α , and that the sequence · · · −−−−−−→ L −−−−−→ L −−−−−−→ · · · is exact. Using this exact
sequence, we easily observe that for each prime ideal p of R, an Lp -sequence is a Cp -sequence. Thus, if
L is Cohen-Macaulay, then so is C.
We present a couple of examples. For such an element x as in the second statement it is said that
[x, x] is an exact pair of zerodivisors.
Example 4.7. Let R be a commutative noetherian ring containing a field.
(1) Let L be a finitely generated R-module having a submodule C such that L/C ∼
= C. Let x ∈ R be an
L-regular element. Then, for all i ≥ j ≥ 0, C + xj L degenerates to C + x2i−j L.
(2) Let x ∈ R be such that (0 : x) = (x), and let y ∈ R be a non-zerodivisor. Then for all integers
i ≥ j ≥ 0 (x, y j ) degenerates to (x, y 2i−j ), where (x, y j ) and (x, y 2i−j ) are ideals of R.
Proof. The composition α : L ։ L/C ∼
= C ֒→ L satisfies Im α = ker α. The first assertion is shown by
Corollary 4.6. The second assertion follows from the first one.
Remark 4.8. Example 4.7(2) immediately recovers the degeneration obtained in [3, Remark 2.5] and
the ‘if’ part of Theorem 3.1.
5. Remarks on degenerations over hypersurface rings
In this section, we consider extending Theorem 1.1 to higher dimension. Let k be an algebraically closed
field of characteristic not two and let R = S/(f ) be a hypersurface, where S = k[[x0 , x1 , · · · , xn ]] is a formal
power series ring with maximal ideal mS = (x0 , x1 , · · · , xn ) and f ∈ mS . We define R♯ = S[[u]]/(f + u2 )
and R♯♯ = S[[u, v]]/(f + u2 + v 2 ). In what follows, for a certain hypersurface R, we investigate the
connection among the degenerations of Cohen-Macaulay modules over R, R♯ and R♯♯ .
Remark 5.1. Let R = S[[y]]/(y 2 + f ) with f ∈ mS .
(1) There is a one-to-one correspondence between
• the isomorphism classes of Cohen-Macaulay R-modules, and
• the equivalence classes of square matrices ϕ with entries in S such that ϕ2 = −f .
This follows by considering matrix representations over S.
(2) Let M be a Cohen-Macaulay R-module, and let ΩM be the first syzygy of M with respect to the
minimal R-free resolution of M . As remarked in [7, Lemma 12.2], the matrix representation of ΩM over
S is −µ, where µ is that of M .
Let R be as in the above remark. Let µ be a square matrix with entries in S such that µ2 =
−f
µ u
♯
and M a Cohen-Macaulay
R-module
which
corresponds
to
µ.
Consider
the
matrices
µ
=
and
−u −µ
√
√
µ ζ
♯ 2
µ♯♯ = −η
,
where
ζ
=
u
+
−1v
and
η
=
u
−
−1v.
By
Remark
5.1(1)
we
have
(µ
)
=
−(f + u2 )
−µ
(resp. (µ♯♯ )2 = −(f + u2 + v 2 )), and µ♯ (resp. µ♯♯ ) corresponds to a Cohen-Macaulay module over R♯
(resp. R♯♯ ), which we denote by M ♯ (resp. M ♯♯ ). Note that (M ♯ )♯ is not always isomorphic to M ♯♯ .
To show our main result, we state and prove a lemma.
Lemma 5.2. Let R be a commutative noetherian ring and M, N finitely generated R-modules. If M
degenerates to N , then M/xM also degenerates to N/xN for an N -regular element x.
Proof. Suppose that M degenerates to N along V . As mentioned in Remark 2.2(1), we have an exact
sequence 0 → Z → M ⊕ Z → N → 0, where the endomorphism g of Z is nilpotent. Since x is N -regular,
we also have 0 → Z/xZ → M/xM ⊕ Z/xZ → N/xN → 0. The endomorphism g ⊗ R/xR is also nilpotent,
so that M/xM degenerates to N/xN along V .
DEGENERATIONS OF MODULES OVER COMMUTATIVE RINGS
9
The following is the main result of this section.
Proposition 5.3. Let R = S[[y]]/(y 2 +f ) be a hypersurface, and let M, N be Cohen-Macaulay R-modules.
Consider the following six conditions.
(1) M degenerates to N .
(2) M ♯♯ degenerates to N ♯♯ .
(3) M ♯ degenerates to N ♯ .
♯♯
♯♯
(1’) M degenerates to ΩN . (2’) M degenerates to (ΩN ) . (4) M ⊕ ΩM degenerates to N ⊕ ΩN .
Then the implications (1) ⇒ (2) ⇒ (3) ⇒ (4) and (1’) ⇒ (2’) ⇒ (3) ⇒ (4) hold true.
Proof. (1) ⇒ (2): Assume that M degenerates to N along V . Let µ and ν be the matrix representations
of M and N over S. We have the matrix representation
ξ over S ⊗k V such that ξ ⊗V Vt ∼
= µ ⊗k Vt and
√
√
ξ
ζ
♯♯
∼ ν. Now consider the matrix ξ =
ξ ⊗V V /tV =
−1v
and
η
= u − −1v.
,
where
ζ
=
u
+
−η −ξ
∼ µ♯♯ and ξ ♯♯ ⊗V V /tV ∼
Then one can show that ξ ♯♯ ⊗V Vt =
= ν ♯♯ . In fact, since α−1 (ξ ⊗V Vt )α = µ and
−1
β (ξ ⊗V V /tV )β = ν respectively for some suitable invertible matrices α and β, we have
−1
−1
α (ξ⊗ V )α α−1 (ζ⊗ V )α
µ ζ
♯♯
α
0
(ξ ♯♯ ⊗V Vt ) ( α0 α0 ) = α−1 (−η⊗V Vt )α α−1 (−ξ⊗V Vt )α = −η
−µ = µ .
0 α−1
V
Similarly,
−1
β
0
0
β
−1
(ξ ♯♯ ⊗V V /tV )
β 0
0 β
=
t
V
t
β −1 (ξ⊗V V /tV )β β −1 (ζ⊗V V /tV )β
β −1 (−η⊗V V /tV )β β −1 (−ξ⊗V V /tV )β
=
ν ζ
−η −ν
= ν ♯♯ .
According to Corollary 2.7, we conclude that M ♯♯ degenerates to N ♯♯ .
(1′ ) ⇒ (2′ ): The similar argument to the proof of (1) ⇒ (2) shows this.
(2) ⇒ (3): By Lemma 5.2, we see that M ♯♯ /vM ♯♯ degenerates to N ♯♯ /vN ♯♯ . Clearly, M ♯♯ /vM ♯♯ ∼
= M♯
♯
♯♯
♯♯ ∼
and N /vN = N . Thus we obtain the assertion.
♯♯
5.2.
(2′ ) ⇒ (3): We also have the degeneration from M ♯♯ /vM ♯♯ ∼
= M ♯ to (ΩN )♯♯ /v(ΩN
) by Lemma
−ν u
♯♯
♯
♯♯
♯♯ ∼
of
(ΩN
)
/v(ΩN
)♯♯ ,
N
.
Actually,
for
the
matrix
representation
We note that (ΩN
)
/v(ΩN
)
=
−u ν
−ν u 0 1
ν u
0 −1
we have 1 0
−1 0 = ( −u −ν ) . Thus, the implication holds.
−u ν
(3) ⇒ (4): It follows from Lemma 5.2 since M ♯ /uM ♯ ∼
= N ⊕ ΩN .
= M ⊕ ΩM and N ♯ /uN ♯ ∼
Remark 5.4. In Proposition 5.3, (4) does not necessarily imply (1) or (1’). Let R = k[[x, y, z]]/(x3 +
y 2 + z 2 ) and p, q, r be ideals (x, y), (x, y 2 ), (x, y 3 ) of R respectively. Then we have the exact sequence
0 → q → p ⊕ r → q → 0, so that p ⊕ r degenerates to q ⊕ q. Notice that r ∼
= Ωq. However q
= Ωp and q ∼
never degenerates to p or q. See [3, Theorem 3.1] for the details.
Recall that Knörrer’s periodicity theorem [9, Theorem 12.10] gives an equivalence Φ : CM(R) →
CM(R♯♯ ) of triangulated categories. We call this functor a Knörrer’s periodicity functor. Using Proposition 5.3, we generalize the ‘if’ part of Theorem 3.1.
Example 5.5. Consider the hypersurface singularity of type (A∞ ) with odd dimension:
k[[x0 , x1 , . . . , xn ]]/(x21 + x22 + · · · + x2n ).
Let M (h) be the image of (x0 , z h ) by the composition of Knörrer’s periodicity functors
CM(k[[x0 , x1 ]]/(x21 )) → CM(k[[x0 , x1 , x2 , x3 ]]/(x21 + x22 + x23 )) → · · · → CM(k[[x0 , x1 , . . . , xn ]]/(x21 + x22 +
· · · + x2n )). Iterated application of Proposition 5.3 shows that M (j) degenerates to M (2i − j).
Proposition 5.6. Let R = S/(f ) with f ∈ mS and let α : S n → S n be an endomorphism such that
(α, α) is a matrix factorization of f over S. Let z√∈ R be a non-zerodivisor,
and consider the map
√
( α zh ) : Rn ⊕ Rn → Rn for each h ≥ 0. Set ζ = u + −1v and η = u − −1v. Then the following hold.
ζ zh 0
(a) The images Im ( α zh ) and Im αη −α
are Cohen-Macaulay modules over R and R♯♯ , respech
0 z
tively.
ζ zh 0
(b) The functor Φ sends Im ( α zh ) to Im αη −α
.
0 zh
(c) The R-module Im ( α zh ) is nonfree and indecomposable if and only if so is the R♯♯ -module
ζ zh 0
Im αη −α
.
0 zh
ζ zj 0
(d) For all i ≥ j ≥ 0 one can show that Im ( α zj ) degenerates to Im ( α z2i−j ) and Im αη −α
j
0 z
ζ z 2i−j
0
♯♯
degenerates to Im αη −α
as
Cohen-Macaulay
modules
over
R,
R
respectively.
0 z 2i−j
10
NAOYA HIRAMATSU, RYO TAKAHASHI, AND YUJI YOSHINO
Proof. We regard α and α′ =
α ζ
η −α
2n
as the endomorphisms of Rn and R♯♯ , respectively. Note that
Im α = ker α and Im α′ = ker α′ . Assertions (a) and (d) follow from Corollary 4.6. By [7, (12.8.1)], the
!
!
α ζ −z h 0
α −z h ζ
0
h
h
0 α
0
ζ
η
−α
0
−z
α
−z
cokernel of 0 α
is sent by Φ to that of η 0 −α −zh , which is transformed into 0 0 α ζ
0
η
0
−α
0
0
η
−α
by elementary row and column operations. By Lemma 4.1, the cokernel of the latter matrix is the image
ζ zh 0
of αη −α
. Thus (b) follows. As Φ is an equivalence of additive categories, (c) follows from (b).
0 zh
Remark 5.7. (1) Since, for each h > 0, the ideal (x0 , z h ) of the ring k[[x0 , z]]/(x20 ) is nonfree and indecomposable and coincides with Im ( x0 zh ), Example 5.5 can also be obtained from the above proposition.
(2) Let R be an algebra over a field. Let 0 → L → M → N → 0 be an exact sequence of finitely
generated R-modules. Then M degenerates to L ⊕ N by Remark 2.2(2). Such a degeneration is called
a degeneration by an extension in [3, Definition 2.4]. Thus, it is important to investigate the existence
of degenerations that cannot be obtained by extensions. The degeneration in Example 5.5 cannot be
obtained by (iterated) extensions, because M (2i − j) is an indecomposable R-module.
References
[1] T. Araya, K. Iima and R. Takahashi, On the structure of Cohen-Macaulay modules over hypersurfaces of countable
Cohen-Macaulay representation type. J. Algebra 361 (2012), 213–224.
[2] K. Bongartz, On degenerations and extensions of finite-dimensional modules. Adv. Math. 121 (1996), 245–287.
[3] N. Hiramatsu and Y. Yoshino, Examples of degenerations of Cohen-Macaulay modules, Proc. Amer. Math. Soc. 141
(2013), no. 7, 2275–2288.
[4] T.Y. Lam, Serre’s problem on projective modules. Springer Monographs in Mathematics. Springer-Verlag, Berlin, 2006.
xxii+401 pp.
[5] C. Riedtmann, Degenerations for representations of quivers with relations. Ann. Scient. École Norm. Sup. 4e sèrie
19 (1986), 275–301.
[6] F-O. Schreyer, Finite and countable CM-representation type. Singularities, representation of algebras, and vector
bundles, Lecture Notes in Math., 1273, Springer, Berlin, 1987, 9–34,
[7] Y. Yoshino, Cohen-Macaulay Modules over Cohen-Macaulay Rings, London Mathematical Society Lecture Note Series
146. Cambridge University Press, Cambridge, 1990. viii+177 pp.
[8] Y. Yoshino, On degenerations of Cohen-Macaulay modules. J. Algebra 248 (2002), 272–290.
[9] Y. Yoshino, On degenerations of modules. J. Algebra 278 (2004), 217–226.
[10] Y. Yoshino, Stable degenerations of Cohen-Macaulay modules, J. Algebra 332 (2011), 500–521.
[11] G. Zwara, Degenerations for modules over representation-finite algebras. Proc. Amer. Math. Soc. 127 (1999), 1313–
1322.
[12] G. Zwara, Degenerations of finite-dimensional modules are given by extensions. Compositio Math. 121 (2000), 205–
218.
Department of general education, National Institute of Technology, Kure College, 2-2-11, Agaminami,
Kure Hiroshima, 737-8506 Japan
E-mail address: [email protected]
Graduate School of Mathematics, Nagoya University, Furocho, Chikusaku, Nagoya, Aichi 464-8602, Japan
E-mail address: [email protected]
URL: http://www.math.nagoya-u.ac.jp/~takahashi/
Department of Mathematics, Graduate School of Natural Science and Technology, Okayama University,
Okayama 700-8530, Japan
E-mail address: [email protected]
| 0math.AC
|
DIFFERENCE BASES IN DIHEDRAL GROUPS
arXiv:1704.02472v2 [math.CO] 13 Apr 2017
TARAS BANAKH AND VOLODYMYR GAVRYLKIV
Abstract. A subset B of a group G is called a difference basis of G if each element g ∈ G can be written as
the difference g = ab−1 of some elements a, b ∈ B. The smallest cardinality |B| ofpa difference basis B ⊂ G is
called the difference size of G and is denoted by ∆[G]. The fraction ð[G] := ∆[G]/ |G| is called the difference
characteristic √
of G. We prove that for every n ∈ N the dihedral group D2n of order 2n has the difference
characteristic 2 ≤ ð[D2n ] ≤ √48 ≈ 1.983. Moreover, if n ≥ 2 · 1015 , then ð[D2n ] < √4 ≈ 1.633. Also we
586
6
calculate the difference sizes and characteristics of all dihedral groups of cardinality ≤ 80.
A subset B of a group G is called a difference basis for a subset A ⊂ G if each element a ∈ A can be written
as a = xy −1 for some x, y ∈ B. The smallest cardinality of a difference basis for A is called the difference size of
A and is denoted by ∆[A]. For example, the set {0, 1, 4, 6} is a difference basis for the interval A = [−6, 6] ∩ Z
witnessing that ∆[A] ≤ 4.
The p
definition of a difference basis B for a set A in a group G implies that |A| ≤ |B|2 and gives a lower
bound |A| ≤ ∆[A]. The fraction
∆[A]
ð[A] := p
≥1
|A|
is called the difference characteristic of A.
For a real number x we put
⌈x⌉ = min{n ∈ Z : n ≥ x} and ⌊x⌋ = max{n ∈ Z : n ≤ x}.
The following proposition is proved in [1, 1.1].
Proposition 1. Let G be a finite group. Then
√
1+ 4|G|−3
≤ ∆[G] ≤ |G|+1
,
(1)
2
2
(2) ∆[G] ≤ ∆[H] · ∆[G/H] and ð[G] ≤ ð[H] · ð[G/H] for any normal subgroup H ⊂ G;
(3) ∆[G] ≤ |H| + |G/H| − 1 for any subgroup H ⊂ G.
In [8] Kozma and Lev proved (using the classification of finite simple groups) that each finite group G has
difference characteristic ð[G] ≤ √43 ≈ 2.3094.
In this paper we shall evaluate the difference characteristics of dihedral groups and prove that each diherdal
group D2n has ð[D2n ] ≤ √48
≈ 1.983. Moreover, if n ≥ 2 · 1015 , then ð[D2n ] < √46 ≈ 1.633. We recall that
586
the dihedral group D2n is the isometry group of a regular n-gon. The dihedral group D2n contains a normal
cyclic subgroup of index 2. A standard model of a cyclic group of order n is the multiplicative group
Cn = {z ∈ C : z n = 1}
of n-th roots of 1. The group Cn is isomorphic to the additive group of the ring Zn = Z/nZ.
Theorem 2. For any numbers n, m ∈ N the dihedral group D2nm has the difference size
√
2 nm ≤ ∆[D2nm ] ≤ ∆[D2n ] · ∆[Cm ]
√
and the difference characteristic 2 ≤ ð[D2nm ] ≤ ð[D2n ] · ð[Cm ].
Proof. It is well-known that the dihedral group D2nm contains a normal cyclic subgroup of order nm, which
can be identified with the cyclic group Cnm . The subgroup Cm ⊂ Cnm is normal in D2mn and the quotient
group D2mn /Cm is isomorphic to D2n . Applying Proposition 1(2), we obtain the upper bounds ∆[D2n ] ≤
∆[D2nm /Cm ] · ∆[Cm ] = ∆[D2n ] · ∆[Cm ] and ð[D2nm ] ≤ ð[D2n ] · ð[Cm ].
√
Next, we prove the lower bound 2 nm ≤ ∆[D2nm ]. Fix any element s ∈ D2nm \ Cnm and observe that
s = s−1 and sxs−1 = x−1 for all x ∈ Cnm . Fix a difference basis D ⊂ D2nm of cardinality |D| = ∆[D2nm ] and
1991 Mathematics Subject Classification. 05B10, 05E15, 20D60.
Key words and phrases. dihedral group, difference-basis, difference characteristic.
1
2
TARAS BANAKH AND VOLODYMYR GAVRYLKIV
write D as the union D = A ∪ sB for some sets A, B ⊂ Cnm ⊂ D2nm . We claim that AB −1 = Cnm . Indeed,
for any x ∈ Cnm we get xs ∈ sCnm ∩ (A ∪ sB)(A ∪ sB)−1 = AB −1 s−1 ∪ sBA−1 and hence
x ∈ AB −1 s−1 s−1 ∪ sBA−1 s−1 = AB −1 ∪ B −1 A = AB −1 .
So, Cnm = AB −1 and hence nm ≤√|A| · |B|. Then ∆[D2nm ] = |A| + |B| ≥ min{l + k : l, k ∈ N, lk ≥ nm} ≥
√
√
]
√ 2nm ≥ 2
√ nm =
2 nm and ð[D2nm ] = ∆[D
2.
2nm
2nm
Corollary 3. For any number n ∈ N the dihedral group D2n has the difference size
√
2 n ≤ ∆[D2n ] ≤ 2 · ∆[Cn ]
√
√
and the difference characteristic 2 ≤ ð[D2n ] ≤ 2 · ð[Cn ].
The difference sizes of finite cyclic groups were evaluated in [2] with the help of the difference sizes of the
order-intervals [1, n] ∩ Z in the additive group Z of integer numbers. For a natural number n ∈ N by ∆[n] we
√
shall denote the difference size of the order-interval [1, n] ∩ Z and by ð[n] := ∆[n]
its difference characteristic.
n
∞
The asymptotics of the sequence (ð[n])n=1 was studied by Rédei and Rényi [9], Leech [7] and Golay [6] who
eventually proved that
r
q
q
128
sin(ϕ)
4
< ð[6] = 83 .
< 2 + max 2 ϕ+π
≤ lim ð[n] = inf ð[n] ≤ ð[6166] = √
2 + 3π
n→∞
0<ϕ<2π
n∈N
6166
In [2] the difference sizes of the order-intervals [1, n] ∩ Z were applied to give upper bounds for the difference
sizes of finite cyclic groups.
Proposition 4. For every n ∈ N the cyclic group Cn has difference size ∆[Cn ] ≤ ∆ ⌈ n−1
2 ⌉ , which implies
that
64
2
1
<√ .
lim sup ð[Cn ] ≤ √ inf ð[n] ≤ √
2 n∈N
3083
3
n→∞
The following upper bound for the difference sizes of cyclic groups were proved in [2].
Theorem 5. For any n ∈ N the cyclic group Cn has the difference characteristic:
(1) ð[Cn ] ≤ ð[C4 ] = 32 ;
√
(2) ð[Cn ] ≤ ð[C2 ] =√ð[C8 ] = 2 if n 6= 4;
(3) ð[Cn ] ≤ √1273 < 2 if n ≥ 9;
(4) ð[Cn ] ≤ √24
< √1273 if n ≥ 9 and n 6= 292;
293
(5) ð[Cn ] < √23 if n ≥ 2 · 1015 .
For some special numbers n we have more precise upper bounds for ∆[Cn ]. A number q is called a prime
power if q = pk for some prime number p and some k ∈ N.
The following theorem was derived in [2] from the classical results of Singer [11], Bose, Chowla [3], [4] and
Rusza [10].
Theorem 6. Let p be a prime number and q be a prime power. Then
(1) ∆[Cq2 +q+1 ] = q + 1;
√
(2) ∆[Cq2 −1 ] ≤ q − 1 + ∆[Cq−1 ] ≤ q − 1 + 23 q − 1;
√
√
(3) ∆[Cp2 −p ] ≤ p − 3 + ∆[Cp ] + ∆[Cp−1 ] ≤ p − 3 + 23 ( p + p − 1).
The following Table 1 of difference sizes and characteristics of cyclic groups Cn for ≤ 100 is taken from [2].
Using Theorem 6(1), we shall prove that for infinitely many numbers n the lower and upper bounds given
in Theorem 2 uniquely determine the difference size ∆[D2n ] of D2n .
Theorem 7. If n = 1 + q + q 2 for some prime power q, then
m
√ lp
∆[D2n ] = 2 · ∆[Cn ] = 2 n =
2|D2n | = 2 + 2q.
Proof. By Theorem 6(1), ∆[Cn ] = 1 + q. Since
p
√
2 q 2 + q + 1 = 2 n ≤ ∆[D2n ] ≤ ∆[D2 ] · ∆[Cn ] = 2 · ∆[Cn ] = 2 + 2q,
p
p
it suffices to check that (2 + 2q) − 2 q 2 + q + 1 < 1, which is equivalent to q 2 + q + 1 > q +
q 2 + q + 1 > q 2 + q + 14 .
1
2
and to
DIFFERENCE BASES IN DIHEDRAL GROUPS
3
Table 1. Difference sizes and characteristics of cyclic groups Cn for n ≤ 100
n ∆[Cn ]
1
1
2
2
3
2
4
3
5
3
6
3
7
3
8
4
9
4
10
4
11
4
12
4
13
4
14
5
15
5
16
5
17
5
18
5
19
5
20
6
21
5
22
6
23
6
24
6
25
6
ð[Cn ]
1
1.4142...
1.1547...
1.5
1.3416...
1.2247...
1.1338...
1.4142...
1.3333...
1.2649...
1.2060...
1.1547...
1.1094...
1.3363...
1.2909...
1.25
1.2126...
1.1785...
1.1470...
1.3416...
1.0910...
1.2792...
1.2510...
1.2247...
1.2
n ∆[Cn ]
26
6
27
6
28
6
29
7
30
7
31
6
32
7
33
7
34
7
35
7
36
7
37
7
38
8
39
7
40
8
41
8
42
8
43
8
44
8
45
8
46
8
47
8
48
8
49
8
50
8
ð[Cn ]
1.1766...
1.1547...
1.1338...
1.2998...
1.2780...
1.0776...
1.2374...
1.2185...
1.2004...
1.1832...
1.1666...
1.1507...
1.2977...
1.1208...
1.2649...
1.2493...
1.2344...
1.2199...
1.2060...
1.1925...
1.1795...
1.1669...
1.1547...
1.1428...
1.1313...
n ∆[Cn ]
51
8
52
9
53
9
54
9
55
9
56
9
57
8
58
9
59
9
60
9
61
9
62
9
63
9
64
9
65
9
66
10
67
10
68
10
69
10
70
10
71
10
72
10
73
9
74
10
75
10
ð[Cn ]
n ∆[Cn ]
ð[Cn ]
1.1202... 76
10
1.1470...
1.2480... 77
10
1.1396...
1.2362... 78
10
1.1322...
1.2247... 79
10
1.1250...
1.2135... 80
11
1.2298...
1.2026... 81
11
1.2222...
1.0596... 82
11
1.2147...
1.1817... 83
11
1.2074...
1.1717... 84
11
1.2001...
1.1618... 85
11
1.1931...
1.1523... 86
11
1.1861...
1.1430... 87
11
1.1793...
1.1338... 88
11
1.1726...
1.125
89
11
1.1659...
1.1163... 90
11
1.1595...
1.2309... 91
10
1.0482...
1.2216... 92
11
1.1468...
1.2126... 93
12
1.2443...
1.2038... 94
12
1.2377...
1.1952... 95
12
1.2311...
1.1867... 96
12
1.2247...
1.1785... 97
12
1.2184...
1.0533... 98
12
1.2121...
1.1624... 99
12
1.2060...
1.1547... 100
12
1.2
A bit weaker result holds also for the dihedral groups D8(q2 +q+1) .
Proposition 8. If n = 1 + q + q 2 for some prime power q, then
4q + 3 ≤ ∆[D8n ] ≤ 4q + 4.
Proof. By Theorem 6(1), ∆[Cn ] = 1 + q. Since ∆[D8 ] = 4 (see Table 2), by Theorem 2,
p
√
4 q 2 + q + 1 = 2 4n ≤ ∆[D8n ] ≤ ∆[D8 ] · ∆[Cn ] = 4(1 + q).
p
To p
see that 4q + 3 ≤ ∆[D8n ] ≤ 4q + 4, it suffices to check that (4 + 4q) − 4 q 2 + q + 1 < 2, which is equivalent
to q 2 + q + 1 > q + 12 and to q 2 + q + 1 > q 2 + q + 41 .
In Table 2 we present the results of computer calculation
of the difference sizes and characteristics of
√
dihedral groups of order ≤ 80. In this table lb[D2n ] := ⌈ 4n ⌉ is the lower bound given in Theorem 2. With
the boldface font we denote the numbers 2n ∈ {14, 26, 42, 62}, equal to 2(q 2 + q + 1) for a prime power q. For
these numbers we know that ∆[D2n ] = lb[D2n ] = 2q + 2. For q = 2 and n = q 2 + q + 1 = 7 the table shows
that ∆[D56 ] = ∆[D8n ] = 11 = 4q + 3, which means that the lower bound 4q + 3 in Proposition 8 is attained.
Theorem 9. For any number n ∈ N the dihedral group D2n has the difference characteristic
√
48
2 ≤ ð[D2n ] ≤ √
≈ 1.983.
586
Moreover, if n ≥ 2 · 1015 , then ð[D2n ] < √46 ≈ 1.633.
√
√
by Theorem 5(4),
Proof. By Corollary 3, 2 ≤ ð[D2n ] ≤ 2 · ð[Cn ]. If n ≥ 9 and n 6= 292, then ð[Cn ] ≤ √24
293
√
√
24
48
√
√
and hence ð[D2n ] ≤ 2 · ð[Cn ] ≤ 2 · 293 = 586 . If n = 292, then known values ð[C73 ] = √973 (given in
√
Table 1), ð[D8 ] = √48 = 2 (given in Table 2) and Theorem 2 yield the upper bound
ð[D2·292 ] = ð[D8·73 ] ≤ ð[D8 ] · ð[C73 ] =
√
48
9
.
2· √ < √
73
586
4
TARAS BANAKH AND VOLODYMYR GAVRYLKIV
Table 2. Difference sizes and characteristics of dihedral groups D2n for 2n ≤ 80.
2n lb[D2n ] ∆[D2n ]
2
2
2
4
3
3
6
4
4
8
4
4
10
5
5
12
5
5
14
6
6
16
6
6
18
6
7
20
7
7
22
7
8
24
7
7
26
8
8
28
8
8
30
8
8
32
8
9
34
9
9
36
9
9
38
9
10
40
9
9
2∆[Cn ]
2
4
4
6
6
6
6
8
8
8
8
8
8
10
10
10
10
10
10
12
ð[D2n ]
1.4142...
1.5
1.6329...
1.4142...
1.5811...
1.4433...
1.6035...
1.5
1.6499...
1.5652...
1.7056...
1.4288...
1.5689...
1.5118...
1.4605...
1.5909...
1.5434...
1.5
1.6222...
1.4230...
2n lb[D2n ] ∆[D2n ]
42
10
10
44
10
10
46
10
11
48
10
10
50
10
11
52
11
11
54
11
12
56
11
11
58
11
12
60
11
12
62
12
12
64
12
12
66
12
13
68
12
13
70
12
12
72
12
13
74
13
14
76
13
14
78
13
14
80
13
14
Analyzing the data from Table 2, one can check that ð[D2n ] ≤
If n ≥ 2 · 1015 , then ð[Cn ] < √23 by Theorem 5(5), and hence
ð[D2n ] ≤
√48
586
2∆[Cn ]
10
12
12
12
12
12
12
12
14
14
12
14
14
14
14
14
14
16
14
16
ð[D2n ]
1.5430...
1.5075...
1.6218...
1.4433...
1.5556...
1.5254...
1.6329...
1.4699...
1.5756...
1.5491...
1.5240...
1.5
1.6001...
1.5764...
1.4342...
1.5320...
1.6274...
1.6059...
1.5851...
1.5652...
≈ 1.983 for all n ≤ 8.
√
4
2 · ð[Cn ] < √ .
6
Question 10. Is supn∈N ð[D2n ] = ð[D22 ] =
√8
22
≈ 1.7056?
To answer Question 10 affirmatively, it suffices to check that ð[D2n ] ≤ √822 for all n < 1 212 464.
√
Proposition 11. The inequality ð[D2n ] ≤ 2 · ð[Cn ] ≤ √822 holds for all n ≥ 1 212 464.
Proof. It suffices to prove that ð[Cn ] ≤ √411 for all n ≥ 1 212 464. To derive a contradiction, assume that
ð[Cn ] > √411 for some n ≥ 1 212 464. Let (qk )∞
k=1 be an increasing enumeration of prime powers. Let k ∈ N
2
2
be the unique number such that 12qk + 14qk + 15 < n ≤ 12qk+1
+ 14qk+1 + 15. By Corollary 4.9 of [2],
4
∆[Cn ] ≤ 4(qk+1 + 1). The inequality ð[Cn ] > √11 implies
q
4
4 √
n≥ √
12qk2 + 14qk + 16.
4(qk+1 + 1) ≥ ∆[Cn ] > √
11
11
By Theorem 1.9 of [5], if qk ≥ 3275, then qk+1 ≤ qk + 2 lnq2k(qk ) . On the other hand, using WolframAlpha
√
computational knowledge engine it can be shown that the inequality 1 + x + 2 lnx2 (x) ≤ √111 12x2 + 14x + 16
holds for all x ≥ 43. This implies that qk < 3275.
Analysing the table1 of (maximal gaps between) primes, it can be shows that 11(qk+1 +1)2 ≤ 12qk2 + 14qk + 16
2
2
if qk ≥ 331. So, q√
k ≤ 317, qk+1 ≤ 331 and 11 · (qk+1 + 1) = 11 · 332 = 1 212 464 ≤ n, which contradicts
4
4(qk+1 + 1) > √11 n.
1See https://primes.utm.edu/notes/GapsTable.html and https://primes.utm.edu/lists/small/1000.txt
DIFFERENCE BASES IN DIHEDRAL GROUPS
5
References
[1] T. Banakh, V. Gavrylkiv, O. Nykyforchyn, Algebra in superextension of groups, I: zeros and commutativity, Algebra Discrete
Math. 3 (2008), 1–29.
[2] T. Banakh, V. Gavrylkiv, Difference bases in cyclic groups, preprint (https://arxiv.org/abs/1702.02631).
[3] R.C. Bose, An affine analogue of Singer’s theorem, Journal of the Indian Mathematical Society 6 (1942), 1–15.
[4] R.C. Bose, S. Chowla, Theorems in the additive theory of numbers, Comment. Math. Helvetici 37 (1962-63) 141–147.
[5] P. Dusart, Autour de la fonction qui compte le nombre de nombres premiers, Ph.D. Thesis, Univ. de Limoges, 1998;
(http://www.unilim.fr/laco/theses/1998/T1998\01.pdf).
[6] M. Golay, Notes on the representation of 1, 2, . . . , n by differences, J. London Math. Soc. (2) 4 (1972) 729–734.
[7] J. Leech, On the representation of 1, 2, . . . , n by differences, J. London Math. Soc. 31 (1956), 160–169.
[8] G. Kozma, A. Lev, Bases and decomposition numbers of finite groups, Arch. Math. (Basel) 58:5 (1992), 417–424.
[9] L. Rédei, A. Rényi, On the representation of the numbers 1, 2, . . . , N by means of differences, Mat. Sbornik N.S. 24(66) (1949),
385–389.
[10] I.Z. Ruzsa, Solving a linear equation in a set of integers I, Acta Arithmetica LXV. 3 (1993) 259–282.
[11] J. Singer, A theorem in finite projective geometry and some applications to number theory, Trans. Amer. Math. Soc. 43:3
(1938), 377–385.
(T. Banakh) Ivan Franko National University of Lviv (Ukraine), and
Institute of Mathematics, Jan Kochanowski University in Kielce (Poland)
E-mail address: [email protected]
(V. Gavrylkiv) Vasyl Stefanyk Precarpathian National University, Ivano-Frankivsk, Ukraine
E-mail address: [email protected]
| 4math.GR
|
arXiv:1706.04387v1 [math.GR] 14 Jun 2017
TOPOLOGICAL FINITENESS PROPERTIES OF MONOIDS
PART 1: FOUNDATIONS
ROBERT D. GRAY AND BENJAMIN STEINBERG
Abstract. We initiate the study of higher dimensional topological finiteness properties of
monoids. This is done by developing the theory of monoids acting on CW complexes. For this
we establish the foundations of M -equivariant homotopy theory where M is a discrete monoid.
For projective M -CW complexes we prove several fundamental results such as the homotopy
extension and lifting property, which we use to prove the M -equivariant Whitehead theorems.
We define a left equivariant classifying space as a contractible projective M -CW complex. We
prove that such a space is unique up to M -homotopy equivalence and give a canonical model for
such a space via the nerve of the right Cayley graph category of the monoid. The topological
finiteness conditions left-Fn and left geometric dimension are then defined for monoids in terms
of existence of a left equivariant classifying space satisfying appropriate finiteness properties.
We also introduce the bilateral notion of M -equivariant classifying space, proving uniqueness
and giving a canonical model via the nerve of the two-sided Cayley graph category, and we define
the associated finiteness properties bi-Fn and geometric dimension. We explore the connections
between all of the these topological finiteness properties and several well-studied homological
finiteness properties of monoids which are important in the theory of string rewriting systems,
including FPn , cohomological dimension, and Hochschild cohomological dimension. We also
develop the corresponding theory of M -equivariant collapsing schemes (that is, M -equivariant
discrete Morse theory), and among other things apply it to give topological proofs of results of
Anick, Squier and Kobayashi that monoids which admit presentations by complete rewriting
systems are left- right- and bi-FP∞ .
1. Introduction
The study of the higher dimensional finiteness properties of groups was initiated fifty years
ago by C. T. C. Wall [Wal65] and Serre [Ser71]. An Eilenberg–MacLane complex K(G, 1) for a
discrete group G, also called a classifying space, is an aspherical CW complex with fundamental
group G. Such a space can always be constructed for any group G (e.g. via the bar construction) and it is unique up to homotopy equivalence. While useful for theoretical purposes, this
canonical K(G, 1)-complex is very big and is often not useful for practical purposes, specifically
if one wants to compute the homology of the group. It is therefore natural to seek a ‘small’
K(G, 1) for a given group by imposing various finiteness conditions on the space. Two of the
most natural and well-studied such conditions are the topological finiteness property Fn and
the geometric dimension gd(G) of the group.
Property Fn was introduced by C. T. C. Wall in [Wal65]. A group G is said to be of type
Fn if it has an Eilenberg–MacLane complex K(G, 1) with finite n-skeleton. It is easily verified
that a group is finitely generated if and only if it is of type F1 and is finitely presented if
Date: June 15, 2017.
2010 Mathematics Subject Classification. 20M50, 20M05, 20J05, 57M07, 20F10, 20F65.
Key words and phrases. CW complex, monoid, equivariant homotopy theory, homological finitenss property
FPn , cohomological dimension, Hoschild cohomology, rewriting system, collapsing scheme, discrete Morse theory.
This work was supported by the EPSRC grant EP/N033353/1 ‘Special inverse monoids: subgroups, structure,
geometry, rewriting systems and the word problem’. The second author was supiported in part by United StatesIsrael Binational Science Foundation #2012080 and NSA MSP #H98230-16-1-0047.
1
2
ROBERT D. GRAY AND BENJAMIN STEINBERG
and only if it is of type F2 . Thus property Fn generalises the two fundamental finiteness
properties of being finitely generated, or finitely presented, to higher dimensions. The geometric
dimension of G, denoted gd(G), is the smallest non-negative integer n such that there exists
an n-dimensional K(G, 1) complex. If no such n exists, then we set gd(G) = ∞. For more
general background on higher dimensional finiteness properties of groups we refer the reader to
the books [Bro94, Chapter 8], [Geo08, Chapters 6-9], or the survey article [Bro10].
Each of these topological finiteness properties has a natural counterpart in homological algebra given in terms of the existence of projective resolutions of ZG-modules. The analogue
of Fn in this context is the homological finiteness property FPn , while geometric dimension
corresponds to the cohomological dimension of the group. Recall that a group G is said to be
of type FPn (for a positive integer n) if there is a projective resolution P of Z over ZG such
that Pi is finitely generated for i ≤ n. We say that G is of type FP∞ if there is a projective
resolution P of Z over ZG with Pi finitely generated for all i. The property FPn was introduced
for groups by Bieri in [Bie76] and since then has received a great deal of attention in the literature; see [BB97,BH01,Bra99,BW07,LS06]. For groups, Fn and FPn are equivalent for n = 0, 1,
while important results of Bestvina and Brady [BB97] show that FP2 is definitely weaker than
F2 . For higher n there are no further differences, in that a group G is of type Fn (2 ≤ n ≤ ∞)
if and only if it is finitely presented and of type FPn . One natural line of investigation has
been the study of the closure properties of FPn . Examples include results about the behaviour
of FPn under taking: finite index subgroups or extensions, direct (and semidirect) products,
wreath products, HNN extensions, amalgamated free products, and quasi-isometry invariance;
see for example [Alo94, BBG98, Bie76]. In [BHMS02] it is shown that if G is a subgroup of a
direct product of n surface groups, then if G is of type FPn then G has a subgroup of finite
index which is the direct product of at most n finitely generated surface groups. Thompson’s
groups, and several interesting generalisations of these groups, have all be shown to be of type
FP∞ ; see for example [Bro87, FMWZ13, Ste92, GS06].
The cohomological dimension of a group G, denoted cd(G), is the smallest non-negative
integer n such that there exists a projective resolution P = (Pi )i≥0 of Z over ZG of length ≤ n,
i.e., satisfying Pi = 0 for i > n. (Or, if no such n exists, then we set cd(G) = ∞.) The geometric
dimension of a group provides an upper bound for the cohomological dimension. It is easily
seen that gd(G) = cd(G) = 0 if and only if G is trivial. It follows from important results of
Stallings [Sta68] and Swan [Swa69] that gd(G) = cd(G) = 1 if and only if G is non-trivial free
group. Eilenberg and Ganea [EG57] proved that for n ≥ 3 the cohomological and the geometric
dimension of a group are the same. The famous Eilenberg–Ganea problem asks whether this
also holds in dimension two.
Working in the more general context of monoids, and projective resolutions of left ZMmodules, gives the notion of left-FPn , and left cohomological dimension, of a monoid M . There
is an obvious dual notion of monoids of type right-FPn , working with right ZM -modules.
Working instead with bimodules resolutions of the (ZM , ZM )-bimodule ZM one obtains the
notion bi-FPn introduced and studied in [KO01]. Property bi-FPn is of interest from the point of
view of Hochschild cohomology, which is the standard notion of cohomology for rings; [Hoc45],
[Wei94, Chapter 9], or [Mit72]. For monoids all these notions of FPn are known to be different,
while for groups they are all equivalent; see [Coh92, Pri06]. Similarly there is a dual notion of
the right cohomological dimension of a monoid which again is in general not equal to the left
cohomological dimension; see [GP98]. The two-sided notion is the Hochschild cohomological
dimension [Mit72].
TOPOLOGICAL FINITENESS PROPERTIES
3
In monoid and semigroup theory the property FPn arises naturally in the study of string
rewriting systems (i.e. semigroup presentations). The history of rewriting systems in monoids and groups is long and distinguished, and has roots in fundamental work of Dehn and
Thue. A central topic in this area is the study of complete rewriting systems and in methods for computing normal forms. A finite complete rewriting system is a finite presentation for a monoid of a particular form (both confluent and Noetherian) which in particular
gives a solution of the word problem for the monoid; see [BO93]. It is therefore of considerable interest to develop an understanding of which monoids are presentable by such rewriting systems. Many important classes of groups are known to be presentable by finite complete rewriting systems, including surface groups, Coxeter groups, and many closed threemanifold groups. Rewriting systems continue to receive a lot of attention in the literature;
see [CS09a,CS09b,Cho06,DDM09,GS08,HS99,PO05]. The connection between complete rewriting systems and homological finiteness properties is given by a result of Anick [Ani86] (see
also [Bro92]) which shows that a monoid that admits such a presentation must be of type
left- and right-FP∞ ; the special case of FP3 was also handled by Squier [Squ87]. More generally Kobayashi [Kob05] proved that any such monoid is of type bi-FP∞ . A range of other
interesting homotopical and homological finiteness properties have been studied in relation to
monoids defined by compete rewriitng systems including finite homological type, finite derivation type, and higher dimensional analogues FDTn ; see [SOK94, PO04, PO05, GM12]. More
background on the importance the property FPn (and other related finiteness conditions) in
semigroup theory, and the connections with the theory of string rewriting systems may be found
in the survey articles [Coh97, OK97]. Results on cohomology, and cohomological dimension, of
monoids include [AR67, GP96, Nic69, Nic72, CS80, Nun95, GP98] and [Nov98]. The cohomological dimension of left regular bands was recently considered in [MSS15b] and [MSS15a] where
connections with the Leray number of simplicial complexes [KM05] and the homology of cell
complexes was obtained.
It is often easier to establish the topological finiteness properties Fn for a group than the
homological finiteness properties FPn , especially if there is a suitable geometry or topological
space available on which the group acts cocompactly. The desired homological finiteness properties can then be derived by the above-mentioned result for groups, that Fn (for n ≥ 2) is
equivalent to being finitely presented and of type FPn . In contrast, no corresponding theory of
Fn for monoids currently exists. Similarly, there is currently no analogue of geometric dimension of monoids in the literature. The study of homological finiteness properties of monoids
should greatly profit from the development of a corresponding theory of topological finiteness
properties of monoids. The central aim of the present article is to lay the foundations of such
a theory.
For such a theory to be useful in the study of homological finiteness properties of monoids
there are certain properties that any definition of left-Fn , and left geometric dimension, should
certainly satisfy. Specifically left-Fn should imply left-FPn , and the left geometric dimension
should provide an upper bound for the left cohomological dimension. The fundamental question
that needs to be addressed when developing this theory is to determine the correct analogue
of the K(G, 1)-complex in the theory for monoids? There is a natural notion of classifying
space |BM | of a monoid M . This is obtained by viewing M as a one-point category, letting
BM denote the nerve of this category, and setting |BM | as the geometric realisation of the
nerve; see Section 5 for full details of this construction. For a group G this space |BG| is a
K(G, 1)-complex, it is the canonical complex for G mentioned earlier in the introduction. Since
K(G, 1)s are unique up to homotopy equivalence the finiteness conditions Fn and cohomological
dimension can all be defined in terms of existence of CW complexes homotopy equivalent to
4
ROBERT D. GRAY AND BENJAMIN STEINBERG
|BG| satisfying the appropriate finiteness property. Indeed in group theory it is a common
approach in the study of these topological finiteness properties to begin with the space |BG|
and then seek transformations on the space which preserve the homotopy equivalence class, but
make the space smaller. This is, for example, the basic idea behind the theory of collapsing
schemes (which will be discussed in more detail in Section 8). It could be regarded as natural
therefore to try define and study topological finiteness properties of a monoid M in terms of
the space |BM |. We note that there is an extensive literature on the study of classifying spaces
|BM | of monoids and related topics; see for instance [Fie84,Seg78,Nun95,McC69,McD79,MS76,
LN01, Pup58, Pup59, KT76, Hur89].
It turns out, however, that using |BM | to define topological finiteness properties of monoids
is not the right approach in the sense that it would lead to a definition of Fn for monoids which
does not imply left- or right-FPn , and there are similar issues for the corresponding definition
of geometric dimension. This essentially comes does to the fact that the space |BM | does not
contain enough information about the monoid M to recover the corresponding homological
finiteness properties.
In more detail, by applying results of MacDuff [McD79] it is possible to show that there are
examples of monoids which are not of type left-FP1 even though |BM | is contractible. For
example, if M is an infinite left zero semigroup (a semigroup with multiplication xy = x for
all elements x and y) with an identity adjoined then by [McD79, Lemma 5] the space |BM |
is contractible while it is straightforward to show that M does not even satisfy the property
left-FP1 (this also follows from Theorem 6.13 below). This shows that one should not define
property Fn for monoids using the space |BM |. Similar comments apply to attempts to define
geometric dimension–if one tries to define geometric dimension using |BM | then if M is any
monoid with a left zero element but no right zero element, the left cd(M ) would not equal zero
(by Proposition 6.28) while the geometric dimension would be zero.
This issue in fact arose in work of Brown [Bro92] when he introduced the theory of collapsing
schemes. In that paper Brown shows that if a monoid M admits a presentation by a finite
complete rewriting system, then |BM | has the homotopy type of a CW complex with only
finitely many cells in each dimension. When M is a group this automatically implies that the
group is of type FP∞ . Brown goes on to comment
“We would like, more generally, to construct a ‘small’ resolution of this type
for any monoid M with a good set of normal forms, not just for groups. I do
not know any way to formally deduce such a resolution from the existence of the
homotopy equivalence for |BM | above”.
As the comments above show, just knowing about the homotopy equivalence class of |BM |
will never suffice in order to deduce that the monoid is left- (or right-) FP∞ . It turns out
that the correct framework for studying topological finiteness properties of monoids is to pass
−−→
to the universal bundle |EM | over |BM |, which has a concrete description as the geometric
realisation of the right Cayley graph category of the monoid (this will be explained in detail
−−→
in Section 5). The space |EM | is contractable and the monoid has a natural action by left
multiplication on this space. This action is free and sends n-cells to n-cells. It turns out that
this space is the correct canonical model of what we shall call a left equivariant classifying space
for the monoid. In fact we are able to work in the more general context of projective M -sets,
and shall define a left equivariant classifying space as a contractable projective M -CW complex
(see Section 2 for the definition of projective M -CW complex). The corresponding finiteness
conditions left-Fn and left geometric dimension are then defined in the obvious natural way
in terms of the existence of a left equivariant classifying space satisfying appropriate finiteness
properties. Consequently, in order to define and study the topological finiteness properties of
TOPOLOGICAL FINITENESS PROPERTIES
5
monoids we are interested in, it will first be necessary for us to develop some of the foundations
of M -equivariant homotopy theory. Equivariant homotopy theory, and cohomology theory, for
groups is an important and well-established area; see for example [May96] for an introduction.
In this way, we are interested in studying homological finiteness properties of monoids by
investigating their actions on CW complexes.
The paper is structured as follows. The notions of free and projective M -CW complexes
are introduced in Section 2. For projective M -CW complexes we then go on to prove an
M -equivariant version of HELP (homotopy extension and lifting property), from which we
deduce the M -equivariant Whitehead theorems. We also apply HELP to prove that every
M -equivariant continuous mapping of projective M -CW complexes is M -homotopy equivalent
to a cellular one (the cellular approximation theorem). In Section 3 some useful base change
theorems are established. Section 4 is concerned with monoids acting on simplicial sets. In
particular we show how (rigid) M -CW complexes arise from (rigid) M -simplicial sets by taking
the geometric realisation. In Section 5 we recall the definition of the nerve of a category. By
considering three natural categories associated with a monoid M , namely the one-point category, the right Cayley graph category, and the two-sided Cayley graph category, we associate
−−→
←−→
three CW complexes with the monoid, denoted |BM |, |EM | and |EM |. We go through these
−−→
←−→
constructions in detail in that section. In particular the spaces |EM | and |EM | are the canonical models of M -equivariant (respectively two-sided M -equivariant) classifying spaces of the
monoid. In Section 6 we will define left-equivariant, and dually right-equivariant, classifying
spaces for a monoid M . We prove that such spaces always exist, and that they are unique
up to M -homotopy equivalence. Using the notion of left-equivariant classifying space we define property left-Fn for monoids. We prove several fundamental results about this property,
including results showing its relationship with property left-FPn , and its connection with the
properties of being finitely generated and finitely presented. We prove some results about the
closure properties of property left-Fn , in particular results relating the property holding in M
to it holding in certain maximal subgroups of M . We also introduce the notion of left geometric
dimension of a monoid in this section. We show that it is possible to find an M -finite equivariant classifying space for M which is projective when no M -finite free equivariant classifying
space exists, justifying our choice to work with projective M -CW complexes. The geometric dimension is proved to provide an upper bound for the cohomological dimension, and we
characterize monoids with left geometric dimension equal to zero. In Section 7 we introduce
the bilateral notion of a classifying space in order to introduce the stronger property of bi-Fn .
We prove results for bi-Fn analogous to those previously established for left- and right-Fn . In
particular we show that bi-Fn implies bi-FPn which is of interest from the point of view of
Hochschild cohomology. We also define the geometric dimension as the minimum dimension of
a bi-equivariant classifying space and show how it relates to the Hochschild dimension. In Sections 8 and 9 we develop the theory of M -equivariant collapsing schemes [Bro92]. Equivalently,
this may be viewed as the development of M -equivariant discrete Morse theory in the sense
of Forman [For02]. We show in Section 10 that this theory can be applied to monoids which
admit a, so-called, guarded collapsing scheme. Then, in Section 11, we identify some classes of
monoids which admit guarded collapsing schemes, and in particular recover a topological proof
of the Anick’s theorem, and more generally Kobayashi’s theorem, that monoids defined by finite
complete rewriting systems are of type bi-FP∞ (by proving that they are all of type bi-F∞ ).
Brown proved Anick’s theorem by developing a theory of collapsing schemes (or discrete Morse
theory) for chain complexes, which has been rediscovered by other authors [Koz05] later on.
Our approach obviates the need to develop a chain complex analogue in this setting.
6
ROBERT D. GRAY AND BENJAMIN STEINBERG
Applications of the topological approach set out in this paper will be given in a future
article by the current authors [GS]. Among other things in that paper we shall show how
our topological approach can be used to prove results about the closure properties of left-Fn
and bi-Fn for (i) amalgamated free products of monoids (simplifying and vastly improving on
some results of Cremanns and Otto [CO98]) (ii) HNN-like extensions in the sense of Otto and
Pride [PO04] (in particular establishing results which generalise [PO04, Theorem 1] to higher
dimensions), and (iii) HNN extensions of the sort considered by Howie [How63]. For example,
we prove that if A, B are monoids containing a common submonoid C such that A and B are
free right C-set, then if A and B are of type left-Fn and C is of type left-Fn−1 , then A ∗C B is of
type left-Fn . An analogous result is proved for the homological finiteness property FPn , under
the weaker hypothesis that ZA and ZB are flat as ZC-modules. Monoid amalgamated products
are much more complicated than group ones. For instance, the amalgamated free product of
finite monoids can have an undecidable word problem, and the factors do not have to embed
or intersect in the base monoid.
Additionally, we shall give a topological proof that a free inverse monoid on one or more
generators is neither of type left-FP2 nor right-FP2 generalising a classical result of Schein
[Sch75] that such monoids are not finitely presented.
Finally, in [GS] we shall apply our methods to prove results about the homological finiteness
properties of special monoids, that is, monoids defined by finite presentations of the form
hA | w1 = 1, . . . , wk = 1i. We prove that if M is a special monoid with group of units G then
if G is of type FPn with 1 ≤ n ≤ ∞, then M is of type left- and right-FPn ; and moreover that
cd G ≤ cd M ≤ max{2, cd G}. As a corollary we obtain that all special one-relation monoids are
of type left- and right-FP∞ , answering a special case of a question of Kobayashi [Kob00], and
we recover Kobayashi’s result that if the relator is not a proper power then the cohomological
dimension is at most 2. Specifically we show that if M is a special one-relation monoid then
M is of type left- and right-FP∞ , and if M = hA | w = 1i with w not a proper power, then
cd M ≤ 2; otherwise cd M = ∞.
2. Projective M -CW complexes and M -homotopy theory
2.1. CW complexes and topological finiteness properties of groups. For background
on CW complexes, homotopy theory, and algebraic topology for group theory, we refer the
reader to [Geo08] and [May99]. Throughout B n will denote the closed unit ball in Rn , S n−1
the (n − 1)-sphere which is the boundary ∂B n of the n-ball, and I = [0, 1] the unit interval. We
use en to denote an open n-cell, homeomorphic to the open n ball B̊ n = B n − ∂B n , ∂e denotes
the boundary of e and ē = cl(e) the closure of e, respectively. We identify I r = I r × 0 ⊂ I r+1 .
A CW complex is a space X which is a union of subspaces Xn such that, inductively, X0 is
a discrete set of points, and Xn is obtained from Xn−1 by attaching balls B n along attaching
maps j : S n−1 → Xn−1 . The resulting maps B n → Xn are called the characteristic maps. So
Xn is the quotient space obtained from Xn−1 ∪ (Jn × Bn ) by identifying (j, x) with j(x) for
x ∈ S n , where Jn is the discrete set of such attaching maps. Thus Xn is obtained as a pushout
of spaces:
Jn × S n−1
Xn−1
Jn × B n
Xn .
The topology of X should be that of the inductive limit X = lim Xn
−→
TOPOLOGICAL FINITENESS PROPERTIES
7
S
A CW complex X is then equal as a set to the disjoint union of (open) cells X = α eα
where the eα are the images of B̊ n under the characteristic maps. Indeed, an alternative way of
defining CW complex, which shall be useful for us to use later on, is as follows. A CW complex
is a Hausdorff space K along with a family {eα } of open cells of various dimensions such that,
letting
[
K j = {eα : dim eα ≤ j},
the following conditions are satisfied
[
eα and eα ∩ eβ = ∅ for α 6= β.
(CW1) K =
α
(CW2) For each cell eα there is a map ϕα : B n → K (called the characteristic map) where B n
is a topological ball of dimension n = dim eα , such that
(a) ϕα |B̊ n is a homeomorphism onto eα ;
(b) ϕα (∂B n ) ⊂ K n−1 .
(CW3) Each eα0 is contained in a union of finitely many eα .
(CW4) A set A ⊂ K is closed in K if and only if A ∩ eα is closed in eα for all eα .
Note that each characteristic map ϕ : B n → K gives rise to a characteristic map ϕ′ : I n → K
be setting ϕ′ = ϕh for some homeomorphism h : I n → B n . So we can restrict our attention to
characteristic maps with domain I n when convenient. If ϕ : B n → K is a characteristic map
map for e. A subcomplex is a subset L ⊂ K with a
for a cell e then ϕ|∂B n is called an attaching
[
subfamily {eβ } of cells such that L =
eβ and every eβ is contained in L. If L is a subcomplex
of K we write L < K and call (K, L) a CW pair. If e is a cell of K which does not lie in (and
hence does not meet) L we write e ∈ K − L. An isomorphism between CW complexes is a
homeomorphism that maps cells to cells.
Let M be a monoid. We shall define notions of free and projective M -CW complexes and
then use these to study topological finiteness properties of M . The notion of a free M -CW
complex is a special case of a free C-CW complex for a category C considered by Davis and
Lück in [DL98] and so the cellular approximation theorem, HELP Theorem and Whitehead
Theorem in this case can be deduced from their results. The HELP Theorem and Whitehead
Theorem for for projective M -CW complexes can be extracted with some work from the more
general results of Farjoun [DFZ86] on diagrams of spaces but to keep things elementary and
self-contained we present them here .
2.2. The category of M -sets. A left M -set consists of a set X and a mapping M × X → X
written (m, x) 7→ mx called a left action, such that 1x = x and m(nx) = (mn)x for all m, n ∈ M
and x ∈ X. Right M -sets are defined dually, they are the same thing as left M op -sets. A biM -set is an M × M op -set. There is a category of M -sets and M -equivariant mappings, where
f : X → Y is M -equivariant if f (mx) = mf (x) for all x ∈ X, m ∈ M .
A (left) M -set X is said to be free on a set A if there is a mapping ι : A → X such that for
any mapping f : A → Y with Y an M -set, there is a unique M -equivariant map F : X → Y
such that
A ι X
F
f
Y
8
ROBERT D. GRAY AND BENJAMIN STEINBERG
commutes. The mapping ι is necessarily injective. If X is an M -set and A ⊆ X, then A is a
free basis for X if and only if each element of X can be uniquely expressed as ma with m ∈ M
and a ∈ A.
The free left M -set on A exists and can be realised as the set M × A with action m(m′ , a) =
(mm′ , a) and ι is the map a 7→ (1, a). Note that if G is a group, then a left G-set X is free if
and only if G acts freely on X, that is, each element of X has trivial stabilizer. In this case,
any set of orbit representatives is a basis.
An M -set P is projective if any M -equivariant surjective mapping f : X → P has an M equivariant section s : P → X with f ◦ s = 1P . Free M -sets are projective and an M -set is
projective if and only if it is a retract of a free one.
`
Each projective M -set P is isomorphic to an M -set of the form a∈A M ea (disjoint union,
which is the coproduct in the category of M -sets) with ea ∈ E(M ). Here E(M ) denotes the
set of idempotents of the monoid M . In particular, projective G-sets are the same thing as free
G-sets for a group G. (See [Kna72] for more details.)
2.3. Equivariant CW complexes. A left M -space is a topological space X with a continuous
left action M × X → X where M has the discrete topology. A right M -space is the same thing
as an M op -space and a bi-M -space is an M × M op -space. Each M -set can be viewed as a
discrete M -space. Note that colimits in the category of M -spaces are formed by taking colimits
in the category of spaces and observing that the result has a natural M -action.
Let us define a (projective) M -cell of dimension n to be an M -space of the form M e × B n
where e ∈ E(M ) and B n has the trivial action; if e = 1, we call it a free M -cell. We will define
a projective M -CW complex in an inductive fashion by imitating the usual definition of a CW
complex but by attaching M -cells M e × B n via M -equivariant maps from M e × S n−1 to the
(n − 1)-skeleton.
Formally, a projective (left) relative M -CW complex is a pair (X, A) of M -spaces such that
X = lim Xn with in : Xn → Xn+1 inclusions, X−1 = A, X0 = P0 ∪ A with P0 a projective M -set
−→
and where Xn is obtained as a pushout of M -spaces
Pn × S n−1
Xn−1
(2.1)
Pn × B n
Xn
with Pn a projective M -set and B n having a trivial M -action for n ≥ 1. As usual, Xn is called
the n-skeleton of X and if Xn = X and Pn 6= ∅, then X is said to have dimension n. Notice
that since Pn is isomorphic to a coproduct of M -sets of the form M e with e ∈ E(M ), we are
indeed attaching M -cells at each step. If A = ∅, we call X a projective M -CW complex. Note
that a projective M -CW complex is a CW complex and the M -action is cellular (in fact, takes
n-cells to n-cells). We can define projective right M -CW complexes and projective bi-M -CW
complexes by replacing M with M op and M × M op , respectively. We say that X is a free
M -CW complex if each Pn is a free M -set. If G is a group, a CW complex with a G-action is a
free G-CW complex if and only if G acts freely, cellularly, taking cells to cells, and the setwise
stabilizer of each cell is trivial [Geo08, Appendix of Section 4.1].
More generally we define an M -CW complex in the same way as above except that the
Pi are allowed to be arbitrary M -sets. Most of the theory developed below is only valid in
the projective setting, but there will be a few occasions (e.g. when we discuss M -simplicial
sets) where it will be useful for us to be able to refer to M -CW complexes in general. For
future reference we should note here that, just as for the theory of G-CW complexes, there
is alternative way of defining M -CW complex in terms of monoids acting on CW complexes.
TOPOLOGICAL FINITENESS PROPERTIES
9
This follows the same lines as that of groups, see for example [Geo08, Section
3.2 and page 110]
S
or [May96]. Let Y be a left M -space where M is a monoid and Y = α eα is a CW complex
with characteristic maps ϕα : B n → Y . We say that Y is a rigid left M -CW complex if it is:
• Cellular and dimension preserving: For every eα and m ∈ M there exists an eβ such that
meα = eβ and dim(eβ ) = dim(eα ); and
• Rigid on cells: If meα = eβ then mϕα (k′ ) = ϕβ (k′ ) for all k′ ∈ B n − ∂B n .
If the action of M on the set of n-cells is free (respectively projective) then we call Y a free
(respectively projective) rigid left M -CW complex. The inductive process described above for
building (projective, free) left M -CW complexes is easily seen to give rise to rigid (projective,
free) left M -CW complexes, in the above sense. Conversely every rigid (projective, free) left
M -CW complex arises in this way. In other words, the two definitions are equivalent. For an
explanation of this in the case of G-CW complexes see, for example, [Geo08, page 110]. The
proof for monoids is analogous and is omitted. Similar comments apply for rigid right M -CW
complexes and rigid bi-M -CW complexes.
We say that a projective M -CW complex X is of M -finite type if Pn is a finitely generated
projective M -set for each n and we say that X is M -finite if it is finite dimensional and of
M -finite type (i.e., X is constructed from finitely many M -cells).
Notice that if m ∈ M , then mX is a subcomplex of X for all m ∈ M with n-skeleton mXn .
Indeed, mX0 = mP0 is a discrete set of points and mXn is obtained from mXn−1 via the
pushout diagam
mPn × S n−1
mXn−1
mPn × B n
mXn .
A projective M -CW subcomplex of X is an M -invariant subcomplex A ⊆ X which
` ′′is a union
′
of M -cells of X. In other words, each Pn (as above) can be written Pn = Pn Pn with the
images of the Pn′ × B n giving the cells of A. Notice that if A is a projective M -CW subcomplex
of X, then (X, A) can be viewed as a projective relative M -CW complex in a natural way. Also
note that a cell of X belongs to A if and only if each of its translates do.
A projective {1}-CW complex is the same thing as a CW complex and {1}-finite type ({1}finite) is the same thing as finite type (finite).
If e ∈ E(M ) is an idempotent and m ∈ M e, then left multiplication by m induces an
isomorphism Hn ({e} × B n , {e} × S n−1 ) → Hn ({m} × B n , {m} × S n−1 ) (since it induces a
homeomorphism {e} × B n /{e} × S n−1 → {m} × B n /{m} × S n−1 ) and so if we choose an
orientation for the n-cell {e} × B n , then we can give {m} × B n the orientation induced by this
isomorphism. If m ∈ M and m′ ∈ M e, then the isomorphism
Hn ({e} × B n , {e} × S n−1 ) → Hn ({mm′ } × B n , {mm′ } × S n−1 )
induced by mm′ is the composition of the isomorphism
Hn ({e} × B n , {e} × S n−1 ) → Hn ({m′ } × B n , {m′ } × S n−1 )
induced by m′ and the isomorphism
Hn ({m′ } × B n , {m′ } × S n−1 ) → Hn ({mm′ } × B n , {mm′ } × S n−1 )
induced by m and so the action of m preserves orientation. We conclude that the degree n
component of the cellular chain complex for X is isomorphic to ZPn as a ZM -module and
10
ROBERT D. GRAY AND BENJAMIN STEINBERG
L
`
∼
∼
hence is projective (since Z
a∈A ZM ea and ZM = ZM e ⊕ ZM (1 − e) for any
a∈A M ea =
idempotent e ∈ E(M )).
If X is a projective M -CW complex then so is Y = M × I `
where I is given the trivial action.
∼
If we retain the above notation, then Y0 = X`
×
∂I
X
X0 . The n-cells for n ≥ 1 are
=
0
0
n and P
n−1 × I. Notice that
obtained from attaching Pn × B n × ∂I ∼
(P
P
)
×
B
= n
n
n−1 × B
X × ∂I is a projective M -CW subcomplex of X × I.
If X, Y are M -spaces, then an M -homotopy between M -equivariant continuous maps f, g : X →
Y is an M -equivariant mapping H : X × I → Y with H(x, 0) = f (x) and H(x, 1) = g(x) for
x ∈ X where I is viewed as having the trivial M -action. We write f ≃M g in this case. We say
that X, Y are M -homotopy equivalent, written X ≃M Y , if there are M -equivariant continuous
mappings (called M -homotopy equivalences) f : X → Y and g : Y → X with gf ≃M 1X and
f g ≃M 1Y . We write [X, Y ]M for the set of M -homotopy classes of M -equivariant continuous
mappings X → Y .
Lemma 2.1. Let X, Y be projective M -CW complexes and A a projective M -CW subcomplex
`
of X. Let f : A → Y be a continuous M -equivariant cellular map. Then the pushout X A Y
is a projective M -CW complex.
`
Proof. It is a standard result that X A Y is a CW complex whose n-cells are the`n-cells of Y
together with the n-cells of X not belonging to A. In more detail, let q : X → X A Y be the
canonical mapping and view Y as a subspace of the adjunction space. Then the attaching map
of a cell coming from Y is the original attaching map, whereas the attaching map of a cell of
X not belonging to A is the composition of q with its original attaching mapping. `
It follows
from the definition of a projective M -CW subcomplex and the construction that X A Y is a
projective M -CW complex. Here it is important that a translate by M of a cell from X \ A is
a cell of X \ A.
A free M -CW subcomplex of a free M -CW complex X is an M -invariant subcomplex A ⊆ X
which is a union of M -cells of X.
The proof of Lemma 2.1 yields the following.
Lemma 2.2. Let X, Y be free M -CW complexes and A a free M -CW subcomplex
` of X. Let
f : A → Y be a continuous M -equivariant cellular map. Then the pushout X A Y is a free
M -CW complex.
A continuous mapping f : X → Y of spaces is an n-equivalence if
f∗ : πq (X, x) → πq (Y, f (x))
is a bijection for 0 ≤ q < n and a surjection for q = n where π0 (Z, z) = π0 (Z) (viewed as a
pointed set with base point the component of z). It is a weak equivalence if it is an n-equivalence
for all n, i.e., f∗ is a bijection for all q ≥ 0. We will consider a weak equivalence as an ∞equivalence. We shall see later that an M -equivariant weak equivalence of projective M -CW
complexes is an M -homotopy equivalence.
Let Top(X, Y ) denote the set of continuous maps X → Y for spaces X, Y and TopM (X, Y )
denote the set of continuous M -equivariant maps X → Y between M -spaces X, Y .
Proposition 2.3. Let X be a space with a trivial M -action, e ∈ E(M ) and Y an M -space.
Then there is a bijection between TopM (M e × X, Y ) and Top(X, eY ). The bijection sends
g: Me × X → Y
f : M e × X → Y to f : X → eY given by f (x) = f (e, x) and g : X → eY to b
given by gb(m, x) = mg(x).
TOPOLOGICAL FINITENESS PROPERTIES
11
Proof. If x ∈ X, then f (x) = f (e, x) = f (e(e, x)) = ef (e, x) ∈ eY . Clearly, f is continuous. As
gb is the composition of 1M e × g with the action map, it follows that gb is continuous. We show
that the two constructions are mutually inverse. First we check that
fb(m, x) = mf (x) = mf (e, x) = f (m(e, x)) = f (me, x) = f (m, x)
for m ∈ M e and x ∈ X. Next we compute that
gb(x) = gb(e, x) = eg(x) = g(x)
since g(x) ∈ eY . This completes the proof.
Proposition 2.3 is the key tool to transform statements about projective M -CW complexes
into statement about CW complexes. We shall also need the following lemma relating equivariant n-equivalences and n-equivalences.
Lemma 2.4. Let Y, Z be M -spaces and let k : Y → Z be an M -equivariant n-equivalence with
0 ≤ n ≤ ∞. Let e ∈ E(M ) and k′ = k|eY : eY → eZ. Then k′ is an n-equivalence.
Proof. First note that k(ey) = ek(y) and so k|eY does indeed have image contained in eZ. Let
y ∈ eY and q ≥ 0. Let α : eY → Y and β : eZ → Z be the inclusions. Then note that the
action of e gives retractions Y → eY and Z → eZ. Hence we have a commutative diagram
k∗
πq (Y, y)
α∗
πq (Z, k(y))
e∗
β∗
πq (eY, y)
k∗′
e∗
πq (eZ, k(y))
with e∗ α∗ and e∗ β∗ identities. Therefore, if k∗ is surjective, then k∗′ is surjective and if k∗ is
injective, then k∗′ is injective. The lemma follows.
2.4. Whitehead’s theorem. With Lemma 2.4 in hand, we can prove an M -equivariant version
of HELP (homotopy extension and lifting property) [May99, Page 75], which underlies most of
the usual homotopy theoretic results about CW complexes. If X is a space, then ij : X → X ×I,
for j = 0, 1, is defined by ij (x) = (x, j).
Theorem 2.5 (HELP). Let (X, A) be a projective relative M -CW complex of dimension at
most n ∈ N ∪ {∞} and k : Y → Z an M -equivariant n-equivalence of M -spaces. Then given M equivariant continuous mappings f : X → Z, g : A → Y and h : A × I → Z such that kg = hi1
and f i = hi0 (where i : A → X is the inclusion), there exist M -equivariant continuous mappings
ge and e
h making the diagram
i0
A
A×I
i1
A
h
g
Y
e
h
f
X
k
Z
i
i0
i
g
e
X ×I
i1
X
commute.
Proof. Proceeding by induction on the skeleta and adjoining an M -cell at a time, it suffices to
handle the case that
(X, A) = (M e × B q , M e × S q−1 )
12
ROBERT D. GRAY AND BENJAMIN STEINBERG
with 0 ≤ q ≤ n. By Proposition 2.3 it suffices to find continuous mappings ge and e
h making the
diagram
i0
S q−1
S q−1 × I
i1
S q−1
h
g
eY
e
h
f
Bq
k
eZ
i
i0
i
g
e
Bq × I
i1
Bq
commute where we have retained the notation of Proposition 2.3. The mapping k : eY → eZ is
an n-equivalence by Lemma 2.4 and so we can apply the usual HELP theorem [May99, Page 75]
for CW complexes to deduce the existence of ge and e
h. This completes the proof.
As a consequence we may deduce the M -equivariant Whitehead theorems.
Theorem 2.6 (Whitehead). If X is a projective M -CW complex and k : Y → Z is an M equivariant n-equivalence of M -spaces, then the induced mapping k∗ : [X, Y ]M → [X, Z]M is a
bijection if dim X < n or n = ∞ and a surjection if dim X = n < ∞.
Proof. For surjectivity we apply Theorem 2.5 to the pair (X, ∅). If f : X → Z, then e
g: X → Y
satisfies kg ≃M f . For injectivity, we apply Theorem 2.5 to the pair (X × I, X × ∂I) and note
that X ×I has dimension one larger than X.
` Suppose that p, q : X → Y are such that kp ≃M kq
via a homotopy f : X × I → Z. Put g = p q : X × ∂I → Y and define h : X × ∂I × I → Z by
h(x, s, t) = f (x, s). Then e
g : X × I → Y is a homotopy between p and q.
Corollary 2.7 (Whitehead). If k : Y → Z an M -equivariant weak equivalence (n-equivalence)
between projective M -CW complexes (of dimension less than n), then k is an M -homotopy
equivalence.
Proof. Under either hypothesis, k∗ : [Z, Y ]M → [Z, Z]M is a bijection by Theorem 2.6 and so
kg ≃M 1Z for some M -equivariant g : Z → Y . Then kgk ≃M k and hence, since k∗ : [Y, Y ] →
[Y, Z] is a bijection by Theorem 2.6, we have that gk ≃M 1Y . This completes the proof.
2.5. Cellular approximation. Our next goal is to show that every M -equivariant continuous
mapping of projective M -CW complexes is M -homotopy equivalent to a cellular one. We
shall need the well-known fact that if Y is a CW complex, then the inclusion Yn ֒→ Y is an
n-equivalence for all n ≥ 0 [May99, Page 76].
Theorem 2.8 (Cellular approximation). Let f : X → Y be a continuous M -equivariant mapping with X a projective M -CW complex and Y a CW complex with a continuous action of M
by cellular mappings. Then f is M -homotopic to a continuous M -equivariant cellular mapping.
Any two cellular approximations are homotopy equivalent via a cellular M -homotopy.
Proof. We prove only the first statement. The second is proved using a relative version of the
first whose statement and proof we omit. Note that Yn is M -equivariant for all n ≥ 0 because
M acts by cellular mappings. We construct by induction M -equivariant continuous mappings
fn : Xn → Yn such that f |Xn ≃M fn |Xn via an M -homotopy hn and fn , hn extend f`n−1 , hn−1 ,
respectively (where we take X−1 = ∅). We have, without loss of generality, X0 = a∈A M ea .
Since ea Y is a subcomplex of Y with 0-skeleton ea Y0 and f (ea ) ∈ ea Y , we can find a path pa
in ea Y from f (ea ) to an element ya ∈ ea Y0 . Define f0 (mea ) = mya and h0 (mea , t) = mpa (t),
cf. Proposition 2.3.
TOPOLOGICAL FINITENESS PROPERTIES
13
Assume now that fn , hn have been defined. Since the inclusion Yn+1 → Y is an M -equivariant
(n + 1)-equivalence, Theorem 2.5 gives a commutative diagram
i0
Xn
Xn × I
i1
Xn
hn
fn
Y
i
Yn+1
hn+1
f
Xn+1
i0
i
fn+1
Xn+1 × I
i1
Xn+1
thereby establishing the inductive step. We obtain our desired cellular mapping and M homotopy by taking the colimit of the fn and hn .
3. Base change
If A is a right M -set and B is a left M -set, then A ⊗M B is the quotient of A × B by the least
equivalence relation ∼ such that (am, b) ∼ (a, mb) for all a ∈ A, b ∈ B and m ∈ M . We write
a ⊗ b for the class of (a, b) and note that the mapping (a, b) 7→ a ⊗ b is universal for mappings
f : A × B → X with X a set and f (am, b) = f (a, mb). If M happens to be a group, then M
acts on A × B via m(a, b) = (am−1 , mb) and A ⊗M B is just the set of orbits of this action.
The tensor product A ⊗M () preserves all colimits because it is a left adjoint to the functor
X 7→ X A .
If B is a left M -set there is a natural preorder relation ≤ on B where x ≤ y if and only if
M x ⊆ M y. Let ≈ denote the symmetric-transitive closure of ≤. That is, x ≈ y if there is a
sequence z1 , z2 , . . . , zn of elements of B such that for each 0 ≤ i ≤ n − 1 either zi ≤ zi+1 or
zi ≥ zi+1 . This is clearly an equivalence relation and we call the ≈-classes of B the weak orbits
of the M -set. This corresponds to the notion of the weakly connected components of a directed
graph. If B is a right M -set then we use B/M to denote the set of weak orbits of the M -set.
Dually, if B is a left M -set we use M \B to denote the set of weak orbits. Note that if 1 denotes
the trivial right M -set and B is a left M -set, then we have M \B = 1 ⊗M B.
Let M, N be monoids. An M -N -biset is an M × N op -set. If A is an M -N -biset and B is a
left N -set, then the equivalence relation defining A ⊗N B is left M -invariant and so A ⊗N B is
a left M -set with action m(a ⊗ b) = ma ⊗ b.
Proposition 3.1. Let A be an M -N -biset that is (finitely generated) projective as an M -set
and let B be a (finitely generated) projective N -set. Then A ⊗N B is a (finitely generated)
projective M -set.
Proof. As B is a (finite) coproduct of N -sets N e with e ∈ E(N ), it suffices to handle the case
B = N e. Then A ⊗N N e ∼
= Ae via a ⊗ n 7→ an with inverse a 7→ a ⊗ e for a ∈ Ae. Now
define r : A → Ae by r(a) = ae. Then r is an M -equivariant retraction. So Ae is a retract of a
(finitely generated) projective and hence is a (finitely generated) projective.
If X is a left M -space and A is a right M -set, then A ⊗M X is a topological space with the
quotient topology. Again the functor A ⊗M () preserves all colimits. In fact, A ⊗M X is the
coequalizer in the diagram
a
a
X⇒
X → A ⊗M X
A×M
A
where the top map sends x in the (a, m)-component to mx in the a-component and the bottom
map sends x in the (a, m)-component to x in the am-component.
14
ROBERT D. GRAY AND BENJAMIN STEINBERG
Corollary 3.2. If A is an M -N -biset that is projective as an M -set and X is a projective
N -CW complex, then A ⊗N X is a projective M -CW complex. If A is in addition finitely
generated as an M -set and X is of N -finite type, then A ⊗N X is of M -finite type. Moreover,
dim A ⊗N X = dim X.
Proof. Since A ⊗N () preserves colimits, A ⊗N X = lim A ⊗N Xn . Moreover, putting X−1 = ∅,
−→
we have that if Xn is obtained as per the pushout square (2.1), then A ⊗N Xn is obtained from
the pushout square
(A ⊗N Pn ) × S n−1
A ⊗N Xn−1
(A ⊗N Pn ) × B n
A ⊗N Xn
by preservation of colimits and the observation that if C is a trivial left N -set and B is a left
N -set, then A ⊗N (B × C) ∼
= (A ⊗N B) × C via a ⊗ (b, c) 7→ (a ⊗ b, c). The result now follows
from Proposition 3.1
By considering the special case where M is trivial and A is a singleton, and observing that
a projective M -set P is finitely generated if and only if M \P is finite, we obtain the following
corollary.
Corollary 3.3. Let X be a projective M -CW complex. Then M \X is a CW complex. Moreover,
X is M -finite (of M -finite type) if and only if M \X is finite (of finite type).
The following observation will be used many times.
Proposition 3.4. Let X be a locally path connected N -space and A an M -N -biset. Then π0 (X)
is an N -set and π0 (A ⊗N X) ∼
= A ⊗N π0 (X).
Proof. Note that the functor X 7→ π0 (X) is left adjoint to the inclusion of the category of N -sets
into the category of locally path connected M -spaces and hence it preserves all colimits. The
result now follows from the description of tensor products as coequalizers of coproducts.
The advantage of working with M -homotopies is that they behave well under base change.
Proposition 3.5. Let A be an M -N -biset and let X, X ′ be N -homotopy equivalent N -spaces.
Then A ⊗N X ′ is M -homotopy equivalent to A ⊗N X ′ . In particular, if Y, Z are M -spaces and
Y ≃M Z, then M \Y ≃ M \Z.
Proof. It suffices to prove that if Y, Z are N -spaces and f, g : Y → Z are N -homotopic N equivariant maps, then
A ⊗N f, A ⊗N g : A ⊗N Y → A ⊗N Z
are M -homotopic. This follows immediately from the identification of A ⊗N (Y × I) with
(A ⊗N Y ) × I. For if H : Y × I → Z is an N -homotopy between f and g, then A ⊗N H provides
the M -homotopy between A ⊗N f and A ⊗N g.
The following base change lemma, and its dual, is convenient for dealing with bisets.
Lemma 3.6. Let A be an M × M op -set and consider the right M × M op -set M with the
right action m(mL , mR ) = mmL . Then A/M is a left M -set and there is an M -equivariant
isomorphism A/M → M ⊗M ×M op A.
TOPOLOGICAL FINITENESS PROPERTIES
15
Proof. Clearly, A/M = A ⊗M 1 is a left M -set. Write [a] for the class of a in A/M . Define
f : A/M → M ⊗M ×M op A by f ([a]) = 1 ⊗ a. This is well defined and M -equivariant because if
a ∈ A and m ∈ M , then 1 ⊗ am = 1 ⊗ (1, m)a = 1(1, m) ⊗ a = 1 ⊗ a and 1 ⊗ ma = 1 ⊗ (m, 1)a =
1(m, 1) ⊗ a = m ⊗ a. Define G : M × A → A/M by G(m, a) = [ma]. If m, mL , mR ∈ M , then
G(m(mL , mR ), a) = G(mmL , a) = [mmL a] and G(m, (mL , mR )a) = [mmL amR ] = [mmL a].
Therefore, G induces a well defined mapping g : M ⊗M ×M op A → A/M . Then we check that
gf ([a]) = g(1⊗ a) = [a] and f g(m ⊗ a) = f ([ma]) = 1⊗ ma = 1⊗ (m, 1)a = 1(m, 1)⊗ a = m ⊗ a.
Thus f and g are inverse isomorphisms.
The following basic result will be used later.
Proposition 3.7. Let G be a group. Then G × G is a (G × Gop )-G-biset that is free as right
G-set on cardinality of G generators under the right action (g, g ′ )h = (gh, h−1 g ′ ).
Proof. It is easy to check that right action of G is indeed an action commuting with the left
action of G × Gop . Moreover, the right action of G is free and two elements (g1 , g2 ) and (g1′ , g2′ )
are in the same right G-orbit if and only if g1 g2 = g1′ g2′ . This completes the proof.
Corollary 3.8. Let M be a monoid and X a projective M × M op -CW complex.
(1) X/M is a projective M -CW complex and M \X is a projective M op -CW complex.
(2) If X is of M × M op -finite type, then X/M is of M -finite type and dually for M \X.
(3) dim X/M = dim X = dim M \X.
(4) If X, Y are M × M op -homotopic projective M × M op -CW complexes, then X/M and
Y /M (respectively, M \X and M \Y ) are M -homotopic projective M -CW complexes
(respectively, M op -homotopic projective M op -CW complexes).
Proof. The first three items follow from Corollary 3.2 and Lemma 3.6 (and their duals). The
final statement follows from Lemma 3.6 and Proposition 3.5.
We shall frequently use without comment that if A is an M -N -biset and B is an N -set, then
Z[A ⊗N B] ∼
= ZA ⊗ZN ZB as left ZM -modules. Indeed, there are natural isomorphisms of
abelian groups
HomZM (ZA ⊗ZN ZB, V ) ∼
= HomZN (ZB, HomZM (ZA, V ))
∼
= HomN (B, HomM (A, V ))
∼
= HomM (A ⊗N B, V )
∼
= HomZM (Z[A ⊗N B], V )
for a ZM -module V and so we can apply Yoneda’s Lemma.
4. Simplicial sets
An important source of examples of rigid M -CW complexes will come from simplicial sets
which admit suitable monoid actions. In this section we introduce the notion of a rigid M simplicial set, and we show how these give rise to rigid M -CW complexes via the geometric
realisation functor. For further background on simplicial sets we refer the reader to [Wei94,
Chapter 8] or [May67].
Let ∆ denote the simplicial category. It has objects all the finite linearly ordered sets [n] =
{0, 1, . . . , n − 1} (n ≥ 0) and morphisms given by (non-strictly) order-preserving maps. A
simplicial set X is then a functor X : ∆op → Set from ∆op to the category of sets. For each n,
the image of [n] under X is denoted Xn and is called the set of n-simplicies of the simplicial
16
ROBERT D. GRAY AND BENJAMIN STEINBERG
set. Any simplicial set X may be defined combinatorially as a collection of sets Xn (n ≥ 0) and
functions di : Xn → Xn−1 and si : Xn → Xn+1 (0 ≤ i ≤ n) satisfying
di dj = dj−1 di
(i < j)
si sj = sj+1 si (i ≤ j)
i = j, j + 1
1
di sj = sj−1 di i < j
sj di−1 i > j + 1.
Here the di are called the face maps and the si are called the degeneracy maps. We say that
an n-simplex x ∈ Xn is degenerate if it is the image of some degeneracy map.
A simplicial morphism f : X → Y between simplicial sets is a natural transformation between
the corresponding functors, i.e., a sequence of functions fn : Xn → Yn for each n ≥ 0 such that
fn−1 di = di fn and fn+1 sj = sj fn . There is a functor | · | : SSet → CG, called the geometric
realization functor, from the category SSet of simplicial
sets and the category CG of compactlyS
generated Hausdorff topological spaces. Let K = i≥0 Ki be a simplicial set with degeneracy
and face maps di , si . The geometric realisation |K| of K is the CW complex constructed from
K in the following way. Let
n
o
X
∆n = (t0 , . . . , tn ) : 0 ≤ ti ≤ 1,
ti = 1 ⊆ Rn+1
denote the standard topological n-simplex. Define
δi :
∆n−1
→
∆n
(t0 , . . . , tn−1 ) 7→ (t0 , . . . , ti−1 , 0, ti , . . . , tn−1 ),
σi :
∆n+1
→
∆n
(t0 , . . . , tn+1 ) 7→ (t0 , . . . , ti + ti+1 , . . . , tn−1 ).
and
Then
|K| =
G
n≥0
Kn × ∆n /∼
where ∼ is the equivalence relation generated by
(x, δi (u)) ∼ (di (x), u),
We give
G
0≤n≤q
(x, σi (v)) ∼ (si (x), v).
Kn × ∆n /∼
the quotient topology for all q and take the inductive limit of the resulting topologies. The
geometric realisation |K| is a CW complex whose cells are in natural bijective correspondence
with the non-degenerate simplicies of K. To see this, write
G
K=
Kn × ∆n .
n≥0
Then a point (k, x) ∈ K is called non-degenerate if k is a non-degenerate simplex and x is an
interior point. The following is [Mil57, Lemma 3].
Lemma 4.1. Each point (k, x) ∈ K is ∼-equivalent to a unique non-degenerate point.
TOPOLOGICAL FINITENESS PROPERTIES
17
In each case, the point in question is determined by the maps δi , di , σi and si (see [Mil57] for
details). This lemma is the key to proving that that |K| is a CW complex: we take as n-cells
of |K| the images of the non-degenerate n-simplices of K, and the above lemma shows that the
interiors of these cells partition |K|. The remaining properties of a CW complex are then easily
verified. The following lemma shows that geometric realisation defines a functor from SSet to
CG.
The next result is [Mil57, Lemma 4].
S
S
Lemma 4.2. If K = Ki and L = Li are simplicial sets and f : K → L is a simplicial
morphism then f given by
f n : Kn × ∆n → Ln × ∆n ,
(x, u) 7→ (f (x), u)
is continuous, and induces a well-defined continuous map
|f | : |K| → |L|,
(x, u)/∼ 7→ (f (x), u)/∼
of the corresponding geometric realizations, which is cellular.
A left M -simplicial set is a simplicial set equipped with a left action of M by simplicial
morphisms. In order to construct rigid M -CW complexes we shall need the following special
kind of M -simplicial set.
S
Definition 4.3 (Rigid M -simplicial set). Let K = i≥0 Ki be a simplicial set with degeneracy
and face maps di , si , and let M be a monoid. We call K a rigid left M -simplicial set if K
comes equipped with an action of M × K → K such that
• M is acting by simplicial morphisms, i.e., M maps n-simplicies to n-simplicies, and
commutes with di and si ;
• M preserves non-degeneracy, i.e., for every non-degenerate n-simplex x and every m ∈
M the n-simplex mx is also non-degenerate.
A rigid right M -simplicial set is defined dually, and a rigid bi-M -simplicial set is simultaneously both a left and a right M -simplicial set, with commuting actions. A bi-M -simplicial set
is the same thing as a left M × M op -simplicial set. Note that it follows from the condition that
M acts by simplicial morphisms that, under the action of M , degenerate n-simplicies are sent
to to degenerate n-simplicies. The geometric realisation construction defines a functor from the
category of left M -simplicial sets (with M -equivariant simplicial morphisms) to the category of
left M -spaces. In particular, this functor associates with each rigid left M -simplicial set a rigid
M -CW complex. Corresponding statements hold for both rigid right and bi-M -simplicial sets.
S
Lemma 4.4. For any rigid left M -simplicial set K = i≥0 Ki the geometric realisation |K| is
a rigid left M -CW complex with respect to the induced action given by
m · [(x, u)/∼] = (m · x, u)/∼.
Proof. It follows from Lemma 4.2 that the action is continuous. By the definition of rigid left
M -simplicial set the M -action maps non-degenerate simplices to non-degenerate simplices, and
the cells of |K| are in natural bijective correspondence with the non-degenerate simplicies of K.
It follows that the action of M on |K| sends n-cells to n-cells. The action is rigid by definition.
Thus |K| is a rigid M -CW complex.
There are obvious right- and bi-M -simplicial set analogues of Lemma 4.4 obtained by replacing M by M op and M × M op , respectively.
18
ROBERT D. GRAY AND BENJAMIN STEINBERG
5. Standard constructions of projective M -CW complexes
In this section we shall give a fundamental method that, for any monoid M , allows us to
construct in a canonical way free left-, right- and bi-M -CW complexes. These constructions
will be important when we go on to discuss M -equivariant classifying spaces later on in the
article. Each of the constructions in this section is a special case of the general notion of the
nerve of a category.
To any (small) category C we can associate a simplicial set N (C) called the nerve of the
category. For each k ≥ 0 we let N (C)k denote the set of all sequences (f1 , . . . , fk ) composable
arrows
f1
f2
f
k
A0 −→ A1 −→ · · · −→
Ak
(5.1)
where we allow objects to repeat in these sequences. The objects of C are the 0-simplices. The
face map di : N (C)k → N (C)k−1 omits Ai , so it carries the above sequence to
f1
f2
fi+1 ◦fi
fi−1
fi+2
f
k
A0 −→ A1 −→ · · · −−−→ Ai−1 −−−−→ Ai+1 −−−→ · · · −→
Ak
while the degeneracy map si : N (C)k → N (C)k+1 carries it to
f1
f2
fi
idA
fi+1
fi+2
f
k
i
A0 −→ A1 −→ · · · −
→ Ai −−→
Ai −
−−
→ Ai+1 −
−−
→ · · · −→
Ak
The classifying space of a (small) category C is the geometric realisation |N (C)| of the nerve
N (C) of C.
The nerve is a functor from Cat (the category of small categories) to SSet (the category
of simplicial sets, with simplicial morphisms) given by applying the functor to the diagram
(5.1). From this it follows that a functor between small categories C and D induces a map of
simplicial sets N (C) → N (D), which in turn induces a continous map between the classifying
spaces |N (C)| → |N (D)|. Also, a natural transformation between two functors between C
and D induces a homotopy between the induced maps on the classifying spaces. In particular,
equivalent categories have homotopy equivalent classifying spaces. Any functor which is left or
right adjoint induces a homotopy equivalence of nerves. Consequently, |N (C)| is contractible if
C admits an initial or final object. (For a proof of this see [Sri96, Corollary 3.7].)
It is obvious from the nerve construction that the nerve of a category which is not connected
is the disjoint union of the nerves of the connected components of the category. Thus, if every
component of C admits an initial or final object, then each of the components of |N (C)| will
be contractible.
It is well know that the geometric realisations of the nerve of a category C and its reversal
C op are homeomorphic.
5.1. The classifying space |BM | of a monoid M . In the general context above, given a
monoid M we can construct a category with a single object, one arrow for every m ∈ M , and
composition given by multiplication. The classifying space |BM | of the monoid M is then the
geometric realisation of the nerve of the category corresponding to M op (the reversal is for
the technical reason of avoiding reversals in the face maps). In more detail, the nerve of this
category is the simplicial set BM with n-simplices: σ = (m1 , m2 , ..., mn ), n-tuples of elements
of M . The face maps are given by
i=0
(m2 , . . . , mn )
di σ = (m1 , . . . , mi−1 , mi mi+1 , mi+2 , . . . , mn ) 0 < i < n
(m1 , . . . , mn−1 )
i = n,
TOPOLOGICAL FINITENESS PROPERTIES
19
and the degeneracy maps are given by
si σ = (m1 , . . . , mi , 1, mi+1 , . . . , mn ) (0 ≤ i ≤ n).
The geometric realisation |BM | is called the classifying space of M . Then |BM | is a CW
complex with one n-cell for every non-degenerate n-simplex of BM , i.e., for every n-tuple all
of whose entries are different from 1. As mentioned in the introduction, classifying spaces of
monoids have received some attention in the literature.
5.2. Right Cayley graph category. Let Γr (M ) denote the right Cayley graph category for
M , which has
• Objects: M ;
• Arrows: (x, m) : x → xm; and
• Composition of arrows: (xm, n) ◦ (x, m) = (x, mn).
The identity at x is (x, 1). This composition underlies our use of M op in defining BM .
−−→
−−→
Let EM be the nerve of the category Γr (M ). The n-simplies of EM may be written using
the notation m(m1 , m2 , ..., mn ) = mτ where τ = (m1 , m2 , ..., mn ) is an n-simplex of BM . Here
m(m1 , m2 , ..., mn ) denotes the n-tuple of composable arrows in the category Γr (M ) where we
start at m and the follow the path labelled by m1 , m2 , ..., mn .
−−→
The face maps in EM are given by
i=0
mm1 (m2 , ..., mn )
di (m(m1 , m2 , ..., mn )) = m(m1 , m2 , ..., mi mi+1 , ..., mn ) 0 < i < n
m(m1 , m2 , ..., mn−1 )
i=n
and the degeneracy maps are given by
si σ = m(m1 , . . . , mi , 1, mi+1 , . . . , mn ) (0 ≤ i ≤ n).
where σ = m(m1 , ..., mn ).
−−→
−−→
−−→
Let |EM | denote the geometric realisation of EM . So |EM | is a CW complex with one
−−→
n-cell for every non-degenerate n-simplex of EM , that is, for every m(m1 , m2 , . . . , mn ) with
−−→
mj 6= 1 for 1 ≤ j ≤ n. As a consequence, by an n-cell of EM we shall mean a non-degenerate
n-simplex.
Consider the right Cayley graph category Γr (M ). For each m ∈ M there is precisely one
morphism (1, m) from 1 to m. Since the category has an initial object we conclude that the
−−→
geometric realisation of its nerve |EM | is contractible.
Applying the nerve functor to the projection functor from the category Γr (M ) to the onepoint category M op , which identifies all the vertices of Γr (M ) to a point, gives a simplicial
−−→
morphism π : EM → BM between the corresponding nerves, which maps
m(m1 , m2 , ..., mn ) 7→ (m1 , m2 , ..., mn ).
−−→
Observe that, for each n, the projection π maps the set of n-cells of EM onto the set of n-cells
−−→
BM . If we then apply the geometric realisation functor we obtain a projection π : |EM | →
|BM | (we abuse notation slightly by using the same notation π to denote this map).
The monoid M acts by left multiplication on the category Γr (M ). By functoriality of the
−−→
nerve, it follows that M acts on the left of EM n by simplicial morphisms via
s · m(m1 , m2 , ..., mn ) = sm(m1 , m2 , ..., mn ).
−−→
Under this action EM n is a free left M -set with basis BMn . Also, if we restrict to the n-cells
(i.e., non-degenerate simplices), then we obtain a free left M -set with basis the set of n-cells of
20
ROBERT D. GRAY AND BENJAMIN STEINBERG
BM . It is an easy consequence of the definitions that this is an action by simplicial morphisms
and that it preserves non-degeneracy in the sense that s · mσ is an n-cell if and only if mσ is an
−−→
−−→
n-cell for all s ∈ M and mσ ∈ EM . Therefore EM is a rigid left M -simplicial set. Combining
−−→
these observations with Lemma 4.4 we conclude that |EM | is a free left M -CW complex which
is contractible.
←−−
Dually, we use EM to denote the nerve of the left Cayley graph category Γl (M ). The
−−→
←−−
simplicial set EM satisfies all the obvious dual statements to those above for EM . In particular
←−−
←−−
M acts freely via right multiplication action on EM by simplicial morphisms, and |EM | is a
free right M -CW complex which is contractible.
←−−→
5.3. Two-sided Cayley graph category. Let Γ(M ) denote the two-sided Cayley graph category for M , which has
• Objects: M × M ;
• Arrows: M × M × M where (mL , m, mR ) : (mL , mmR ) → (mL m, mR ); and
• Composition of arrows: (nL , n, nR ) ◦ (mL , m, mR ) = (mL , mn, nR ) where (mL m, mR ) =
(nL , nnR ). Equivalently this is the same as the composition (mL m, n, nR )◦(mL , m, nnR ) =
(mL , mn, nR ) and corresponds to the path
(mL , mnnR ) → (mL m, nnR ) → (mL mn, nR ).
This is in fact the kernel category of the identity map, in the sense of Rhodes and Tilson [RT89].
←−−→
There is a natural M × M op action of the category Γ(M ).
←−−→
←−→
←−→
Let EM be the nerve of the category Γ(M ). The simplicial set EM parallels the two-sided
←−→
geometric bar construction of J. P. May; see [May72, May75]. The n-simplies of EM may
be written using the notation m(m1 , m2 , ..., mn )s = mτ s where τ = (m1 , m2 , ..., mn ) is an
n-simplex of BM .
←−−→
Here m(m1 , m2 , ..., mn )s denotes the n-tuple of composable morphisms in the category Γ(M )
where we start at (m, m1 m2 . . . mn s) and follow the path
(m, m1 m2 m3 . . . mn s) → (mm1 , m2 m3 . . . mn s) → . . .
. . . (mm1 m2 , m3 . . . mn s) → . . . (mm1 m2 . . . mn , s)
labelled by the edges
(m, m1 , m2 m3 . . . mn s), (mm1 , m2 , m3 . . . mn s), . . . ,
(mm1 m2 . . . mn−2 , mn−1 , mn s), (mm1 m2 . . . mn−1 , mn , s)
←−→
and finish at (mm1 m2 . . . mn , s). The face maps in the nerve EM are given by
i=0
mm1 (m2 , ..., mn )s
di (m(m1 , m2 , ..., mn )s) = m(m1 , m2 , ..., mi mi+1 , ..., mn )s 0 < i < n
m(m1 , m2 , ..., mn−1 )mn s
i=n
and the degeneracy maps are given by
si σ = m(m1 , . . . , mi , 1, mi+1 , . . . , mn )s (0 ≤ i ≤ n),
where σ = m(m1 , ..., mn )s.
←−→
←−→
←−→
Let |EM | denote the geometric realisation of EM . So |EM | is a CW complex with one n-cell
←−→
for every non-degenerate n-simplex of EM . Observe that (mL , mR ) and (m′L , m′R ) are in the
←−−→
same component of the two-sided Cayley graph category Γ(M ) if and only if mL mR = m′L m′R .
Moreover, for each m ∈ M the vertex (1, m) is initial in its component. It follows from
TOPOLOGICAL FINITENESS PROPERTIES
21
←−→
←−→
these observations that π0 (|EM |) ∼
= M as an M × M op -set, and each component of |EM | is
←−−→
contractible. There is a natural projection from Γ(M ) to the one-point category M op mapping
(mL , m, mR ) to its middle component m. Applying the nerve functor to this projection gives a
←−→
simplicial morphism π : EM → BM given by
m(m1 , m2 , ..., mn )s 7→ (m1 , m2 , ..., mn ).
←−→
As in the one-sided case, this projection sends n-cells to n-cells and induces a map π : |EM | →
|BM | between the corresponding geometric realisations.
←−→
The monoid M has a natural two-sided action on EM n via
x · [m(m1 , m2 , ..., mn )s] · y = xm(m1 , m2 , ..., mn )sy.
←−→
Under this action EM is a free rigid bi-M -simplicial set. Combining these observations with
←−→
←−→
Lemma 4.4 we conclude that |EM | is a free bi-M -CW complex such that π0 (|EM |) ∼
= M as an
←−→
op
M × M -set and each component of |EM | is contractible
6. One-sided classifying spaces and finiteness properties
We will define left and right equivariant classifying spaces for a monoid M . Two-sided
equivariant classifying spaces will be defined in the next section. As we shall see, the examples
discussed in Section 5 will serve as the standard models of such spaces.
We say that a projective M -CW complex X is a (left) equivariant classifying space for M if it
is contractible. A right equivariant classifying space for M will be a left equivariant classifying
space for M op . Notice that the augmented cellular chain complex of an equivariant classifying
space for M provides a projective resolution of the trivial (left) ZM -module Z.
Example 6.1. The bicyclic monoid is the monoid B with presentation ha, b | ab = 1i. It is not
hard to see that each element of B is uniquely represented by a word of the form bi aj where
i, j ≥ 0. Figure 1 shows an equivariant classifying space for B. The 1-skeleton is the Cayley
graph of B and there is a 2-cell glued in for each path labelled ab. This example is a special
case of far more general results about equivariant classifying spaces of special monoids which
will appear in a future paper of the current authors [GS].
..
.
..
.
a
b
a
b
a
b
a
b
b
b
···
Figure 1. An equivariant classifying space for the bicyclic monoid
Our first goal is to show that any two equivariant classifying spaces for M are M -homotopy
equivalent.
Lemma 6.2. Let X be an equivariant classifying space for M and let Y be a contractible
M -space. Then there exists a continuous M -equvariant mapping f : X → Y .
22
ROBERT D. GRAY AND BENJAMIN STEINBERG
Proof. The proof constructs by induction M -equivariant
` continuous mappings fn : Xn → Y with
fn extending fn−1 . To define f0 , observe Q
that X0 = a∈A M ea (without loss of generality) and
so, by Proposition 2.3, TopM (X0 , Y ) ∼
= a∈A ea Y 6= ∅ and so we can define f0 . Assume that
fn : Xn → Y has been defined. Let Z be the one-point space with the trivial M -action and let
k : Y → Z be the unique M -equivariant map. Then k is a weak equivalence. So by Theorem 2.5
we can construct a commutative diagram
Xn
i0
Xn × I
i1
i
Z
k
Y
Xn
fn
i
fn+1
Xn+1
i0
Xn+1 × I
i1
Xn+1
with fn+1 M -equivariant. Now take f to be the colimit of the fn .
Theorem 6.3. Let X, Y be equivariant classifying spaces for M . Then X and Y are M homotopy equivalent by a cellular M -homotopy equivalence.
Proof. By Corollary 2.7 and Theorem 2.8 it suffices to construct an M -equivariant continuous
mapping f : X → Y . But Lemma 6.2 does just that.
Next we now give an elementary proof that contractible free M -CW complexes exist and
hence there are equivariant classifying spaces for M . A more canonical construction, using
simplicial sets, was given in the previous section.
Lemma 6.4. Let M be a monoid.
(1) If X0 is a non-empty projective (free) M -set, then there is a connected projective (free)
M -graph X with vertex set X0 .
(2) If X is a connected projective (free) M -CW complex such that πq (X) = 0 for 0 ≤ q < n,
then there exists a projective M -CW complex Y containing X as a projective M -CW
subcomplex and such that Yn = Xn and πq (Y ) = 0 for 0 ≤ q ≤ n.
(3) If X is a connected projective (free) M -CW complex such that πq (X) is trivial for
0 ≤ q < n, then there exists a contractible projective (free) M -CW complex Y containing
X as a projective M -CW subcomplex and such that Yn = Xn .
Proof. For the first item, fix x0 ∈ X0 . The edge set of X will be bijection with M × X0 with the
edge corresponding to (m, x) connecting mx0 to mx. Then X is a projective (free) M -graph
and each vertex x is connected to x0 via the edge (1, x).
For the second item, we show that we can add free M -cells of dimension n + 1 to X to obtain
a new projective M -CW complex Y with πn (Y ) = 0. If πn (X) = 0, then take Y = X. So
assume that πn (X) is non-trivial. Fix a base point x0 ∈ X0 and let fa : S n → X, a ∈ A, be
mappings whose based homotopy classes form a set of generators for πn (X, x0 ). As Xn → X is
an n-equivalence, we may assume without loss of generality that fa : S n → Xn . Suppose that
X is constructed from pushouts as per (2.1). Note that M × A, where A has the trivial action,
is a free M -set. Let us define Y by putting Yk = Xk for 0 ≤ k ≤ n, defining Yn+1 to be the
pushout
`
(Pn+1 × S n ) (M × A × S n )
Xn
(Pn+1 × B n+1 )
`
(M × A × B n+1 )
Yn+1 ,
TOPOLOGICAL FINITENESS PROPERTIES
23
where the top map is the union of the attaching map for X with the mapping (m, a, x) 7→ mfa (x)
(cf. Proposition 2.3), and putting Yk = Xk ∪ Yn+1 for k > n + 1. Then Y is a projective M -CW
complex containing X as a projective M -CW subcomplex and with Yn = Xn . Moreover, since
Xn = Yn → Y is an n-equivalence, it follows that the based homotopy classes of the fa : S n → Y
generate πn (Y, x0 ). By construction the fa can be extended to B n+1 → Y and so their classes
are trivial in πn (Y, x0 ). Thus πn (Y ) = 0. Also, because Xn = Yn → Y is an n-equivalence, we
have that πq (Y ) = 0 for 0 ≤ q < n.
The final item follows from Whitehead’s theorem, iteration of the second item and that
Yn → Y is an n-equivalence for all n ≥ 0.
Corollary 6.5. Let M be a monoid. Then there exists a contractible free M -CW complex.
Proof. Put X0 = M . Then by Lemma 6.4 we can find a connected free M -graph X with vertex
set X0 . Now applying the third item of Lemma 6.4 we can find a contractible free M -CW
complex with 1-skeleton X.
Example 6.6. It follows from the definitions and results in Section 5 that the geometric realisa−−→
tion |EM | of the nerve of the right Cayley graph category of M is a left equivariant classifying
space for M .
Corollary 6.7. If X, Y are equivariant classifying spaces for M , then M \X and M \Y are
homotopy equivalent. In particular, M \X ≃ |BM |. Therefore, if G is a group and X is an
equivariant classifying space for G, then G\X is an Eilenberg-Mac Lane space for G. Conversely, the universal cover of any Eilenberg-Mac Lane space for G is an equivariant classifying
space for G.
Proof. The first statement follows from Theorem 6.3 and Proposition 3.5. The second statement
−−→
follows from the first as |EM | is an equivariant classifying space for M . The group statements
then follow from the previous statements and classical covering space theory.
If M and N are monoids, then E(M ×N ) = E(M )×E(N ) and (M ×N )(e, f ) = M e×N f . It
follows that if P is a (finitely generated) projective M -set and Q a (finitely generated) projective
N -set, then P × Q is a (finitely generated projective) M × N -set.
Proposition 6.8. Let M, N be monoids and let X, Y be equivariant classifying spaces for M, N ,
respectively. Then X × Y is an M × N -equivariant classifying space, which is of M × N -finite
type whenever X is of M -finite type and Y is of N -finite type.
Proof. Assume that X is obtained via attaching projective M -cells Pn × B n and that Y is
obtained by
N -cells Qn × B n . Then the n-cells of X × Y are obtained by
`nattaching projective
adjoining i=0 Pi × Qn−i × B n and hence X × Y is an M × N -projective CW complex which
is of M × N -finite type whenever X is of M -finite type and Y is of N -finite type. As X × Y is
contractible, we deduce that it is an M × N -equivariant classifying space.
6.1. Monoids of type left-Fn . A key definition for this paper is the following. A monoid M
is of type left-Fn if there is an equivariant classifying space X for M such that Xn is M -finite,
i.e., such that M \X has finite n-skeleton. We say that M is of type left-F∞ if M has an
equivariant classifying space X that is of M -finite type, i.e., M \X is of finite type. The monoid
M is defined to have type right-Fn if M op is of type left-Fn for 0 ≤ n ≤ ∞. The following
proposition contains some basic facts.
Proposition 6.9. Let M be a monoid.
(1) A group is of type left-Fn if and only if it is of type Fn in the usual sense for any
0 ≤ n ≤ ∞.
24
ROBERT D. GRAY AND BENJAMIN STEINBERG
(2) For 0 ≤ n ≤ ∞, if M is of type left-Fn , then it is of type left-FPn .
(3) If M is of type left-F∞ , then it is of type left-Fn for all n ≥ 0.
Proof. The first item follows from Corollary 6.7 and Corollary 3.3. The second is immediate
using that the augmented cellular chain complex of an equivariant classifying space X gives a
projective ZM -resolution of the trivial ZM -module since if X is built up from pushouts as per
(2.1), then the nth -chain module is isomorphic to ZPn . The final item is trivial.
−−→
Note that, trivially, if M is a finite monoid then |EM | has finitely many cells in each dimension
and thus M is of type left-F∞ .
Sometimes it will be convenient to use the following reformulation of the property left-Fn .
Proposition 6.10. Let M be a monoid. The following are equivalent for 0 ≤ n < ∞.
(1) M is of type left-Fn
(2) There is a connected M -finite projective M -CW complex X of dimension at most n with
πq (X) = 0 for 0 ≤ q < n.
Proof. If Y is an equivariant classifying space for M such that Yn is M -finite, then since Yn → Y
is an n-equivalence, we deduce that X = Yn is as required for the second item. Conversely, if
X is as in the second item, we can construct by Lemma 6.4 an equivariant classifying space Y
for M with Yn = X. Thus M is of type left-Fn .
Recall that the fundamental group of |BM | is isomorphic to the universal group (or maximal
group image, or group completion) U (M ) of M i.e., the group with generators M and relations
the multiplication table of M (cf. [GZ67]).
Corollary 6.11. Let M be a monoid. If M is of type left-F1 , then U (M ) is finitely generated.
If M is of type left-F2 , then U (M ) is finitely presented.
Proof. By Corollary 6.7, |BM | in the first case is homotopy equivalent to a CW complex with
finite 1-skeleton and in the second case to a CW complex with finite 2-skeleton by Corollary 3.3.
Thus U (M ) ∼
= π1 (BM ) has the desired properties in both cases.
Recall that an inverse monoid is a monoid M with the property that for every m ∈ M there
is a unique element m′ ∈ M such that mm′ m = m and m′ mm′ = m′ . For more on inverse
monoids, and other basic concepts from semigroup theory we refer the reader to [How95].
Corollary 6.12. Let M be a monoid such that |BM | is an Eilenberg-Mac Lane space (e.g., if
M is cancellative with a left or right Ore condition or if M is an inverse monoid) and suppose
that 0 ≤ n ≤ ∞. If M is of type left-Fn , then U (M ) is of type Fn .
Proof. If X is an equivariant classifying space for M , then M \X is homotopy equivalent to
|BM | by Corollary 6.7 and hence is an Eilenberg-Mac Lane space for U (M ). The result now
follows from Corollary 3.3.
Since, as already mentioned above, D. McDuff [McD79] has shown that every path-connected
space has the weak homotopy type of the classifying space of some monoid, not every monoid
|BM | is an Eilenberg-Mac Lane space. So not every monoid satisfies the hypotheses of Corollary 6.12. The fact that if M is cancellative with a left or right Ore condition then |BM | is
an Eilenberg-Mac Lane space is well known. If M is an inverse monoid then |BM | may also
be shown to be an Eilenberg-Mac Lane space. Both of these results can easily be proved by
appealing to Quillen’s theorem A, see [Wei13, Chapter 4], and should be considered folklore.
TOPOLOGICAL FINITENESS PROPERTIES
25
The converse of Corollary 6.12 does not hold. For example, the free inverse monoid on one
generator is not of type left-F2 while its maximal group image Z is F∞ (this proof of the fact
that the free inverse monoid on one generator is not left-F2 will appear in [GS]).
For groups, being of type F1 is equivalent to finite generation. For monoids, the condition of
being left-F1 is considerably weaker. Recall that if M is a monoid and A ⊆ M , then the (right)
Cayley digraph Γ(M, A) of M with respect to A is the graph with vertex set M and with edges
in bijection with M × A where the directed edge (arc) corresponding to (m, a) starts at m and
ends at ma. Notice that Γ(M, A) is a free M -CW graph and is M -finite if and only if A is
finite.
Theorem 6.13. Let M be a monoid. The following are equivalent.
(1) M is of type left-F1 .
(2) M is of type left-FP1 .
(3) There is a finite subset A ⊆ M such that Γ(M, A) is connected as an undirected graph.
In particular, any finitely generated monoid is of type left-F1 .
Proof. Item (1) implies (2) by Proposition 6.9, whereas (2) implies (3) by a result due to
Kobayashi [Kob07]. For completeness, let us sketch the proof that (2) implies (3). Let ε : ZM →
Z be the augmentation map; the ideal I = ker ε is called the augmentation ideal. If M is of
type left-FP1 , then I must be finitely generated because the augmentation map gives a partial
free resolution. But I is generated by all elements of the form m − 1 with m ∈ M . Hence there
is a finite subset A ⊆ M such that the elements a − 1 with a ∈ A generate I. Consider the
Cayley digraph Γ(M, A). Then M acts cellularly on Γ(M, A) and hence acts on π0 (Γ(M, A)).
There is a surjective ZM -module homomorphism η : ZM → Zπ0 (Γ(M, A)) mapping m ∈ M to
the connected component of the vertex m of Γ(M, A). Moreover, the augmentation ε factors
through η. Thus to show that Γ(M, A) is connected, it suffices to show that I = ker η. By
construction ker η ⊆ I. But if a ∈ A, then a and 1 are in the same connected component of
Γ(M, A) and thus a − 1 ∈ ker η. Since the a − 1 with a ∈ A generate I, we deduce that I ⊆ ker η
and hence Γ(M, A) is connected.
Finally, (3) implies (1) by Proposition 6.10 as Γ(M, A) is an M -finite connected free M -CW
complex of dimension at most 1.
We next show that a finitely presented monoid is of type left-F2 . In fact, we shall see later
that finitely presented monoids are of type bi-F2 , which implies left-F2 , but the proof of this
case is instructive.
Theorem 6.14. Let M be a finitely presented monoid. Then M is of type left-F2
Proof. Suppose that M is generated by a finite set A with defining relations u1 = v1 , . . . , un =
vn . Let us construct a 2-dimensional, M -finite, free M -CW complex X with 1-skeleton the
Cayley graph Γ(M, A) by attaching a free M -cell M × B 2 for each relation. Let pi , qi be the
paths from 1 to mi labelled by ui and vi , respectively, where mi is the image of ui (and vi ) in M .
Then we glue in a disk di with boundary path pi qi−1 and glue in M ×B 2 using Proposition 2.3 (so
{m} × B 2 is sent to mdi ). Then X is an M -finite connected free M -CW complex of dimension
at most 2. See Figure 1 for this construction for the bicyclic monoid. By Proposition 6.10, it
suffices to prove that X is simply connected.
A digraph is said to be rooted if there is a vertex v so that there is a directed path from v to
any other vertex. For instance, Γ(M, A) is rooted at 1. It is well known that a rooted digraph
admits a spanning tree, called a directed spanning tree, such that the geodesic from the root to
any vertex is directed. Let T be a directed spanning tree for Γ(M, A) rooted at 1 (this is the
same thing as a prefix-closed set of normal forms for M so, for instance, shortlex normal forms
26
ROBERT D. GRAY AND BENJAMIN STEINBERG
a
would do). Let e = m −→ ma be a directed edge not belonging to T . Then the corresponding
generator of π1 (X, 1) is of the form peq −1 where p and q are directed paths from 1 to m and
ma, respectively. Let u be the label of p and v be the label of q. Then ua = v in M . Thus it
suffices to prove that if x, y ∈ A∗ are words which are equal in M to an element m′ , then the
loop ℓ labelled xy −1 at 1, corresponding to the pair of parallel paths 1 to m′ labelled by x and
y, is null homotopic. By induction on the length of a derivation from x to y, we may assume
that x = wui w′ and y = wvi w′ for some i = 1, . . . , n. Let m0 be the image of w in M . Then
m0 di is a 2-cell with boundary path the loop at m0 labeled by ui vi−1 . It follows that ℓ is null
homotopic. This completes the proof.
The converse of Theorem 6.14 is not true, e.g., the monoid (R, ·) is of type left-F2 (by
Corollary 6.23) but is not even finitely generated. It is natural to ask whether there is a nice
characterisation, analogous to Theorem 6.13(3), for left-F2 in terms of the right Cayley graph
together with the left action of M . We would guess that M is of type left-F2 if and only if it
has a finite subset A ⊆ M such that Γ(M, A) is connected, and finitely many free M -2-cells
can be adjoined to make a simply connected 2-complex.
It is well known that, for finitely presented groups, the properties Fn and FPn are equivalent
for 3 ≤ n ≤ ∞. We now provide the analogue in our context. Here we replace finitely presented
by left-F2 .
Theorem 6.15. Let M be a monoid of type left-F2 . Then M is of type left-Fn if and only if
M is of type left-FPn for 0 ≤ n ≤ ∞.
Proof. We prove that if there is a connected M -finite projective M -CW complex X of dimension
at most n with πq (X) = 0 for 0 ≤ q < n with n ≥ 2 and M is of type left-FPn+1 , then there
is a connected M -finite projective M -CW complex Y of dimension at most n + 1 with Yn = X
and πq (Y ) = 0 for all 0 ≤ q < n + 1. This will imply the theorem by Proposition 6.10,
Proposition 6.9 and induction.
Since X is simply connected, Hq (X) = 0 for 1 ≤ q < n and Hn (X) ∼
= πn (X) by the Hurewicz
theorem. Therefore, the augmented cellular chain complex of X gives a partial projective
resolution of Z of length at most n, which is finitely generated in each degree. Therefore,
since M is of type left-FPn+1 , it follows that Hn (X) = ker dn : Cn (X) → Cn−1 (X) is finitely
generated as a left ZM -module. Choose representatives of fa : S n → X, with a ∈ A, of a
finite set of elements of πn (X) that map to a finite ZM -module generating set of Hn (X) under
the Hurewicz isomorphism. Then form Y by adjoining M × A × B n+1 to X via the attaching
map M × A × S n → Xn given by (m, a, x) 7→ mfa (x). Then Y is an M -finite projective
M -CW complex of dimension n + 1 with Yn = X. Since the inclusion of X = Yn into Y is
an n-equivalence, we deduce that πq (Y ) = 0 for 1 ≤ q < n and that the inclusion X → Y is
surjective on πn . But since the Hurewicz map in degree n is natural and is an isomorphism for
both X and Y , we deduce that the inclusion X → Y induces a surjection Hn (X) → Hn (Y ).
But, by construction, the images of the ZM -module generators of Hn (X) are trivial in Hn (Y )
(since they represent trivial elements of πn (Y )). We deduce that Hn (Y ) = 0 and hence, by the
Hurewicz theorem, πn (Y ) = 0. This completes the induction.
Notice that Theorem 6.15 implies that M is of type left-F∞ if and only if M is of type left-Fn
for all n ≥ 0.
Proposition 6.16. If M is of type left-Fn with n ≥ 1, then M has a free contractible M -CW
complex X such that Xn is M -finite.
Proof. This is clearly true for n = 1 by Theorem 6.13. Note that Lemma 6.4 and the construction in the proof of Theorem 6.15 show that if Y is a simply connected M -finite free M -CW
TOPOLOGICAL FINITENESS PROPERTIES
27
complex of dimension at most 2 and M is of type left-FPn , then one can build a contractible
free M -CW complex X with X2 = Y such that Xn is M -finite. Thus it remains to prove that
if there is a simply connected M -finite projective M -CW complex Y of dimension 2, then there
is a simply connected
` M -finite free M -CW complex X of dimension 2.
Note that Y0 = a∈A M ea with A a finite set. Define X0 = M × A. Identifying M ea with
M ea × {a}, we may view Y0 as an M -subset of X0 . Using this identification, we can define X1
to consists of Y1 (the edges of Y ) along with some new edges. We glue in an edge from (m, a)
to (mea , a) for each m ∈ M and a ∈ A; that is we glue in a free M -cell M × A × B 1 where
the attaching map takes (m, a, 0) to (m, a) and (m, a, 1) to (mea , a). Notice that all vertices of
X0 \ Y0 are connected to a vertex of Y0 in X1 and so X1 is connected as Y1 was connected.
To define X2 , first we keep all the two-cells from Y2 . Notice that if T is a spanning tree for Y ,
then a spanning tree T ′ for X can be obtained by adding to T all the edges (m, a) −→ (mea , a)
with m ∈
/ M ea (all vertices of X0 \ Y0 have degree one). Thus the only edges in X1 \ Y1 that do
not belong to T ′ are the loop edges (m, a) −→ (mea , a) for m ∈ M ea that we have added. So if
we attach M × A × B 2 to X1 by the attaching map M × A × S 1 → X1 mapping {m} × {a} × S 1
to the loop edge (mea , a) −→ (mea , a) from X1 \ Y1 , then we obtain a simply connected free
M -CW complex X which is M -finite. This completes the proof.
In light of Proposition 6.16, one might wonder why we bother allowing projective M -CW
complexes rather than just free ones. The reason is because projective M -CW complexes are
often easier to construct and, as we are about to see, sometimes it is possible to find an M finite equivariant classifying space for M which is projective when no M -finite free equivariant
classifying space exists. This will be relevant when considering geometric dimension.
Let M be a non-trivial monoid with a right zero element z. Then M z = {z} is a one-element
set with the trivial action. Since z is idempotent, it follows that the trivial M -set is projective
but not free. Therefore, the one-point space with the trivial M -action is an M -finite equivariant
classifying space for M , which is not free. We will show that if M has a finite number of right
zeroes (e.g., if M has a zero element), then there is no finite free resolution of the trivial module
which is finitely generated in each degree. In this case, every free equivariant classifying space
for M of M -finite type will be infinite dimensional.
A finitely generated projective module P over a ring R is said to be stably free if there are
finite rank free modules F, F ′ such that P ⊕ F ′ ∼
= F . The following lemma is well known, but
we include a proof for completeness.
Lemma 6.17. Let P be a finitely generated projective (left) module over a ring R. Then P has
a finite free resolution, finitely generated in each degree, if and only if P is stably free.
Proof. Suppose that P is stably free, say P ⊕ F ′ ∼
= F with F, F ′ finite rank free R-modules.
Then the exact sequence
0 −→ F ′ −→ F −→ P −→ 0
provides a finite free resolution of P that is finitely generated in each degree.
Conversely, suppose that
0 −→ Fn −→ Fn−1 −→ · · · −→ F0 −→ P −→ 0
is a free resolution with Fi finitely generated for all 0 ≤ i ≤ n. We also have a projective
resolution
0 −→ Pn −→ Pn−1 −→ · · · −→ P0 −→ P −→ 0
with P0 = P and Pi = 0 for 1 ≤ i ≤ n because P is projective. By the generalized Schanuel’s
lemma, we have that
P0 ⊕ F1 ⊕ P2 ⊕ F3 ⊕ · · · ∼
= F0 ⊕ P1 ⊕ F2 ⊕ P3 ⊕ · · ·
28
ROBERT D. GRAY AND BENJAMIN STEINBERG
and hence
P ⊕ F1 ⊕ F3 ⊕ · · · ∼
= F0 ⊕ F2 ⊕ · · ·
and so P is stably free.
So we are interested in showing that the trivial module for ZM is not stably free if M is a
non-trivial monoid with finitely many right zeroes (and at least one).
Recall that a ring R is said to have the Invariant Basis Number property (IBN) if whenever
Rn ∼
= Rm as R-modules, one has m = n (where m, n are integers); in this definition it does not
matter if one uses left or right modules [Lam99].
Our first goal is to show that if M is a monoid with zero z, then the contracted monoid
ring ZM/Zz has IBN. This result is due to Pace Nielsen, whom we thank for allowing us to
reproduce it. It is equivalent to show that if M is a monoid and I is a proper ideal of M , then
ZM/ZI has IBN. The proof makes use of the Hattori-Stallings trace (see [Wei13, Chapter 2]).
Let ∼ be the least equivalence relation on a monoid M such that mn ∼ nm for all m, n ∈ M ;
this relation, often called conjugacy, has been studied by a number of authors.
Lemma 6.18. Let M be a monoid and e ∈ M an idempotent. Suppose that e is conjugate to
an element of an ideal I. Then e ∈ I.
Proof. Suppose that e is conjugate to m ∈ I. Then we can find elements x1 , . . . , xn , y1 , . . . , yn ∈
M with e = x1 y1 , yi xi = xi+1 yi+1 and yn xn = m. Then e = en+1 = (x1 y1 )n+1 = x1 (y1 x1 )n y1 =
x1 (x2 y2 )n y1 = x1 x2 (y2 x2 )n−1 y2 y1 = · · · = x1 · · · xn (yn xn )yn yn−1 · · · y1 ∈ I as yn xn = m ∈
I.
If R is a ring, then [R, R] denotes the additive subgroup generated by all commutators
ab − ba with a, b ∈ R. The abelian group R/[R, R] is called the Hattori-Stallings trace of R;
this is also the 0-Hochschild homology group of R. Cohn proved that if 1 + [R, R] has infinite
order in R/[R, R], then R has IBN (see [Lam99, Exercise 1.5]). The point is that if A is an
m × n matrix over R and B is an n × m-matrix over R with AB = Im and BA = In , then
m+[R, R] = T (AB) = T (BA) = n+[R, R] where, for a square matrix C over R, we define T (C)
to be the class of the sum of the diagonal entries of C in R/[R, R]. The following proposition
is an elaboration of Pace Nielsen’s argument (see [Nie]).
Proposition 6.19. Let M be a monoid and I a (possibly empty) ideal. Let R = ZM/ZI.
Then R/[R, R] is a free abelian group on the conjugacy classes of M that do not intersect I.
More precisely, if T is a transversal to the conjugacy classes of M not intersecting I, then the
elements of the form m + [R, R] with m ∈ T form a basis for R/[R, R].
Proof. We view R as having basis M \ I subject to the relations of the multiplication table
of M and that the elements of I are 0. Let A be the free abelian group on the conjugacy
classes of M that do not intersect I. Write [m] for the conjugacy class of m ∈ M . Define
an abelian group homomorphism f : A → R/[R, R] by f ([m]) = m + [R, R]. This is well
defined because xy + [R, R] = yx + [R, R] for x, y ∈ M with xy, yx ∈
/ I. To see that f
is surjective, note that if m ∈ M \ I with [m] ∩ I 6= ∅, then m + [R, R] = [R, R]. This
follows because if m = x1 y1 , yi xi = xi+1 yi+1 , for i = 1, . . . , n − 1, and yn xn ∈ I, then
[R, R] = yn xn + [R, R] = xn yn + [R, R] = · · · = y1 x1 + [R, R] = m + [R, R].
Let us define g : R → A on m ∈ M \ I by
(
[m], if [m] ∩ I = ∅
g(m) =
0,
else.
TOPOLOGICAL FINITENESS PROPERTIES
Then if a, b ∈ R with, say,
a=
X
cm m, b =
m∈M \I
then we have that
ab − ba =
X
29
dn n
n∈M \I
X
cm dn (mn − nm).
m,n∈M \I
Since mn ∼ nm, either both map to 0 under g or both map to [mn] = [nm]. Therefore,
ab − ba ∈ ker g and so g induces a homomorphism g′ : R/[R, R] → A. Clearly, if [m] ∩ I = ∅,
then gf ([m]) = g′ (m + [R, R]) = g(m) = [m]. It follows that f is injective and hence an
isomorphism. The result follows.
As a consequence we deduce the result of Nielsen.
Corollary 6.20. Let M be a monoid and I a proper ideal (possibly empty). Then ZM/ZI has
IBN. In particular, contracted monoid rings have IBN.
Proof. Put R = ZM/ZI. If I is a proper ideal, then 1 is not conjugate to any element of I by
Lemma 6.18. It follows from Proposition 6.19 that 1 + [R, R] has infinite order in R/[R, R] and
hence R has IBN.
Theorem 6.21. Let M be a non-trivial monoid with a finitely many right zeroes (and at least
one). Then the trivial left ZM -module Z is projective but not stably free and hence does not
have a finite free resolution that is finitely generated in each degree.
Proof. Let I be the set of right zero elements of M and fix z ∈ I. Observe that I is a proper
two-sided ideal. Note that z, 1 − z form a complete set of orthogonal idempotents of ZM and
so ZM ∼
= ZM z ⊕ ZM (1 − z) and hence ZM z ∼
= Z is projective. Suppose that Z is stably free,
′ free ZM -modules of rank r, r ′ , respectively.
F
with
F,
F
that is, Z ⊕ F ′ ∼
=
There is a exact functor from ZM -modules to zZM z-modules given by V 7→ zV . Note that
zZM z = Zz ∼
= Z as a ring. Also, zZM = ZI is a free abelian group (equals zZM z-module) of
rank |I| < ∞. Therefore,
′
′
Z|I|r ∼
= (ZI)r ∼
= zF ∼
= Z ⊕ zF ′ ∼
= Z ⊕ (ZI)r ∼
= Z1+|I|r
as Z-modules and hence r|I| = 1 + r ′ |I| as Z has IBN. But putting R = ZM/ZI and observing
that Z/ZI · Z = 0, we have that
′
Rr ∼
= R ⊗ZM F ∼
= (R ⊗ZM Z) ⊕ (R ⊗ZM F ′ ) ∼
= Rr
and hence r = r ′ as R has IBN by Corollary 6.20. This contradiction completes the proof.
There is, of course, a dual result for left zeroes. In particular, if M is non-trivial monoid with
a zero element, then Z is not stably free as either a left or right ZM -module. Thus, if M is a
non-trivial monoid with zero, it has no M -finite free left or right equivariant classifying space
but it has an M -finite projective one. This justifies considering projective M -CW complexes.
If L is a left ideal of M containing an identity e, then L = M e = eM e. Note that ϕ : M → M e
given by ϕ(m) = me is a surjective monoid homomorphism in this case since ϕ(1) = e and
ϕ(mn) = mne = mene = ϕ(m)ϕ(n) as ne ∈ M e = eM e. Also note that the left M -set structure
on M e given by inflation along ϕ corresponds to the left M -set structure on M e induced by left
multiplication because if m ∈ M and n ∈ M e, then n = en and so mn = men = ϕ(m)n. Notice
that if f ∈ E(M e), then f ∈ E(M ) and M ef = M f is projective as both an M e-set and an M set. Thus each (finitely generated) projective M e-set is a (finitely generated) projective M -set
via inflation along ϕ. Note that if A is a left M -set, then eM ⊗M A ∼
= eA via em ⊗ a 7→ ema.
30
ROBERT D. GRAY AND BENJAMIN STEINBERG
Proposition 6.22. Suppose that M and e ∈ E(M ) with M e = eM e and 0 ≤ n ≤ ∞. If M e
is of type left-Fn , then so is M . The converse holds if eM is a finitely generated projective left
eM e-set.
Proof. If X is a projective M e-CW complex constructed via pushouts as in (2.1) (but with
M replaced by M e), then each Pn is a projective M -set via inflation along ϕ and so X is a
projective M -CW complex. Moreover, if X is of M e-finite type (respectively, M e-finite), then
it is of M -finite type (respectively, M -finite). Thus if M e is of type left-Fn , then so is M .
Suppose that X is an equivariant classifying space for M and eM is a finitely generated
projective left eM e-set. Then eM ⊗M X ∼
= eX. Now eX is a projective eM e-CW complex and
if Xn is M -finite, then (eX)n = eXn is eM e-finite by Corollary 3.2. Moreover, since eX is a
retract of X as a CW complex and X is contractible, it follows that eX is contractible. Thus
eX is an equivariant classifying space for eM e. The result follows.
Our first corollary is that having a right zero guarantees the property left-F∞ (which can be
viewed as a defect of the one-sided theory).
Corollary 6.23. If M contains a right zero, then M is of type left-F∞ . Hence any monoid
with a zero is both of type left- and right-F∞ .
Proof. If e is a right zero, then M e = {e} = eM e and {e} is of type left-F∞ . Thus M is of type
left-F∞ by Proposition 6.22.
Recall that two elements m and n a monoid M are said to be L -related if and only if they
generate the same principal left ideal, i.e., if M m = M n. Clearly L is an equivalence relation
on M .
Corollary 6.24. Suppose that M is a monoid and e ∈ E(M ) with eM a two-sided minimal
ideal of M and 0 ≤ n ≤ ∞. Note that Ge = eM e is the maximal subgroup at e. If Ge is of
type-Fn , then M is of type left-Fn . If eM contains finitely many L -classes, then the converse
holds.
Proof. Note that M e = eM e = Ge and so the first statement is immediate from Proposition 6.22. For the converse, it follows from Green’s lemma [How95, Lemma 2.2.1] that eM is
free left Ge -set and that the orbits are the L -classes of eM . Thus the second statement follows
from Proposition 6.22.
Corollary 6.24 implies that if M is a monoid with a minimal ideal that is a group G, then
M is of type left-Fn if and only if G is of type Fn and dually for right-Fn .
The following is a slight extension of the fact that a finite index subgroup of a group of type
Fn is also of type Fn ; see [Bro94, Chapter VIII, Proposition 5.1].
Proposition 6.25. Let M be a monoid and N a submonoid such that M is a finitely generated
projective left N -set. If M is of type left-Fn , then N is of type left-Fn , as well.
Proof. Observe that each finitely generated free left M -set is a finitely generated projective
N -set. Hence each finitely generated projective M -set, being a retract of a finitely generated
free left M -set, is a finitely generated projective N -set. Thus any equivariant classifying space
for M is also an equivariant classifying space for N .
An immediate consequence of Proposition 6.8 is the following.
Proposition 6.26. Let M, N be monoids of type left-Fn . Then M × N is of type left-Fn .
TOPOLOGICAL FINITENESS PROPERTIES
31
6.2. Left geometric dimension. Let us define the left geometric dimension of M to be
the minimum dimension of a left equivariant classifying space for M . The right geometric
dimension is, of course, defined dually. Clearly, the geometric dimension is an upper bound on
the cohomological dimension cd M of M . Recall that the (left) cohomological dimension of M
is the projective dimension of the trivial module Z, that is, the shortest length of a projective
resolution of Z. As mentioned in the introduction, for groups of cohomological dimension
different than 2, it is known that geometric dimension coincides with cohomological dimension,
but the general case is open.
Theorem 6.27. Let M be a monoid. Then M has an equivariant classifying space of dimension
max{cd M, 3}.
Proof. If M has infinite cohomological dimension, then this is just the assertion that M has an
equivariant classiyfing space. So assume that cd M < ∞. Put n = max{cd M, 3}. Let Y be
an equivariant classifying space for M . As the inclusion Yn−1 → Y is an (n − 1)-equivalence,
we deduce that πq (Yn−1 ) is trivial for 0 ≤ q < n − 1. Also, as the augmented cellular chain
complex of Yn−1 provides a partial projective resolution of the trivial module of length n−1 and
cd M ≤ n, it follows that ker dn−1 = Hn−1 (Yn−1 ) is a projective ZM -module. By the Eilenberg
swindle, there is a free ZM -module F such that Hn−1 (Yn−1 ) ⊕ F ∼
= F . Suppose that F is free
on a set A. Fix a basepoint y0 ∈ Yn−1 . We glue a wedge of (n − 1)-spheres, in bijection with
A, into Yn−1 at y0 as well as freely gluing in its translates. That is we form a new projective
M -CW complex Z with Zn−2 = Yn−2 and where Z = Zn−1 consists of the (n − 1)-cells from
Yn−1 and M × A × B n−1 where the attaching map M × A × S n−2 is given by (m, a, x) 7→ my0 .
Notice that Cn−1 (Z) ∼
= Cn−1 (Yn−1 ) ⊕ F as a ZM -module and that the boundary map is zero
on the F -summand since the boundary of each of the new (n − 1)-cells that we have glued in is
a point and n ≥ 3. Therefore, Hn−1 (Z) = ker dn−1 = Hn−1 (Yn−1 ) ⊕ F ∼
= F . As the inclusion
Yn−2 = Zn−2 → Z is an (n−2)-equivalence, we deduce that πq (Z) is trivial for 0 ≤ q ≤ n−2. In
particular, Z is simply connected as n ≥ 3. By the Hurewicz theorem, πn−1 (Z, y0 ) ∼
= Hn−1 (Z).
Choose mappings fa : S n−1 → Z, for a ∈ A, whose images under the Hurewicz mapping from a
ZM -module basis for Hn−1 (Z) ∼
= F . Then form X by attaching M × A × B n to Z = Zn−1 via
the mapping
M × A × S n−1 → Z
sending (m, a, x) to mfa (x).
Note that X is an n-dimensional CW complex with Xn−1 = Zn−1 and hence the inclusion Z = Xn−1 → X is an (n − 1)-equivalence. Therefore, πq (X) = 0 = Hq (X) for
0 ≤ q ≤ n − 2. Also πn−1 (X, y0 ) ∼
= Hn−1 (X) via the Hurewicz isomorphism. Moreover,
as the inclusion Z = Xn−1 → X is an (n − 1)-equivalence, we deduce that the inclusion
induces a surjective homomorphism πn−1 (Z, y0 ) → πn−1 (X, y0 ) and hence a surjective homomorphism Hn−1 (Z) → Hn−1 (X). As the ZM -module generators of Hn−1 (Z) have trivial images
in Hn−1 (X) by construction and the Hurewicz map, we deduce that Hn−1 (X) = 0.
Recall that Cn (X) = Hn (Xn , Xn−1 ). By standard cellular homology Hn (Xn , Xn−1 ) is a free
ZM -module on the images of the generator of the relative homology of (B n , S n−1 ) under the
characteristic mappings
ha : ({1} × {a} × B n , {1} × {a} × S n−1 ) → (Xn , Xn−1 )
and the boundary map ∂n : Hn (Xn , Xn−1 ) → Hn−1 (Xn−1 ) sends the class corresponding to
a ∈ A to the image of the generator of S n−1 under the map on homology induced by the
attaching map fa : S n−1 → Xn−1 . Hence a free basis of Hn (Xn , Xn−1 ) is sent by ∂n bijectively
to a free basis for Hn−1 (Xn−1 ) and so ∂n is an isomorphism. The long exact sequence for
32
ROBERT D. GRAY AND BENJAMIN STEINBERG
reduced homology and the fact that an (n − 1)-dimensional CW complex has trivial homology
in degree n provides an exact sequence
∂
n
0 = Hn (Xn−1 ) −→ Hn (Xn ) −→ Hn (Xn , Xn−1 ) −→
Hn−1 (Xn−1 )
and so Hn (X) = Hn (Xn ) ∼
= ker ∂n = 0. As X is a simply connected n-dimensional CW
complex with Hq (X) = 0 for 0 ≤ q ≤ n, we deduce that X is contractible by the Hurewicz and
Whitehead theorems. Therefore, X is an n-dimensional equivariant classifying space for M ,
completing the proof.
We end this section by observing that monoids of left cohomological dimenison 0 are precisely
the monoids of left geometric dimension 0. The following result generalises [GP98, Lemma 1
and Theorem 1].
Proposition 6.28. Let M be a monoid. Then the following are equivalent.
(1) M has a right zero element.
(2) M has left cohomological dimension 0.
(3) M has left geometric dimension 0.
Proof. If M has a right zero z, then M z = {z} is a projective M -set and hence the one point
space is an equivariant classifying space for M . Thus M has left geometric dimension zero. If
M has left geometric dimension zero, then it has left cohomological dimension zero. If M has
left cohomological dimension zero, then Z is a projective ZM -module and so the augmentation
mapping ε : ZM → Z splits. Let P be the image of the splitting, so that ZM = P ⊕ Q. As P is
a retract of ZM and each endomorphism of ZM is induced by a right multiplication, we have
that Z ∼
= P = ZM e for some idempotent e ∈ ZM with ε(e) = 1. Then since me = e for all
m ∈ M and e has finite support X, we must have that M permutes X under left multiplication.
Let G be the quotient of M that identifies two elements if they act the same on X. Then G is
a finite group and Z must be a projective ZG-module. Therefore, G is trivial. But this means
that if x ∈ X, then mx = x for all m ∈ M and so M has a right zero.
We do not know whether left geometric dimension equals left cohomological dimension for
monoids of cohomological dimension one or two, although the former is true for groups by the
Stallings-Swan theorem.
7. Bi-equivariant classifying spaces
Let M be a monoid. We now introduce the bilateral notion of a classifying space in order
to introduce a stronger property, bi-Fn . It will turn out that bi-Fn implies both left-Fn and
right-Fn , but is strictly stronger. Moreover, bi-Fn implies bi-FPn which is of interest from the
point of view of Hochschild cohomology, which is the standard notion of cohomology for rings.
Many of the results are similar to the previous section, but the proofs are more complicated.
First recall that M is an M × M op -set via the action (mL , mR )m = mL mmR . We say that a
projective M × M op -CW complex X is a bi-equivariant classifying space for M if π0 (X) ∼
=M
as an M × M op -set and each component of X is contractible; equivalently, X has an M × M op equivariant homotopy equivalence to the discrete M × M op -set M .
We can augment the cellular chain complex of X via the canonical surjection ε : C0 (X) →
H0 (X) ∼
= Zπ0 (X) ∼
= ZM . The fact that each component of X is contractible guarantees that
this is a resolution, which will be a projective bimodule resolution of ZM and hence suitable for
computing Hochschild cohomology. We begin by establishing the uniqueness up to M × M op homotopy equivalence of bi-equivariant classifying spaces.
TOPOLOGICAL FINITENESS PROPERTIES
33
Lemma 7.1. Let X be a bi-equivariant classifying space for M and let Y be a locally path
connected M × M op -space with contractible connected components. Suppose that g : π0 (X) →
π0 (Y ) is an M ×M op -equivariant mapping. Then there exists a continuous M ×M op -equivariant
mapping f : X → Y such that the mapping f∗ : π0 (X) → π0 (Y ) induced by f is g.
Proof. Let r : X → π0 (X) and k : Y → π0 (Y ) be the projections to the set of connected
components. Then k and r are continuous M × M op -equivariant maps where π0 (X) and π0 (Y )
carry the discrete topology. Our goal will be to construct an M × M op -equivariant continuous
mapping f : X → Y such that the diagram
X
f
Y
r
π0 (X)
k
π0 (Y )
g
commutes. We construct, by induction, M ×M op -equivariant continuous mappings fn : Xn → Y
such that
Xn
fn
Y
r
π0 (X)
(7.1)
k
π0 (Y )
g
commutes and fn extends fn−1 . `
To define f0 , observe that X0 = a∈A M ea × e′a M . Choose ya ∈ Y with k(ya ) = g(r(ea , e′a )).
Then k(ea ye′a ) = ea k(ya )e′a = ea g(r(ea , e′a ))e′a = g(r(ea , e′a )) and so replacing ya by ea ye′a , we
may assume without loss of generality that ya ∈ ea Y e′a . Then by Proposition 2.3 there is an
M × M op -equivariant mapping X0 → Y given by (mea , e′a m′ ) 7→ mea ya e′a m′ for a ∈ A and
m ∈ M . By construction, the diagram (7.1) commutes.
Assume now that fn has been defined. The map k : Y → π0 (Y ) is M × M op -equivariant
and a weak equivalence (where π0 (Y ) has the discrete topology) because Y has contractible
connected components. So by Theorem 2.5 we can construct a commutative diagram
i0
Xn
Xn × I
i1
Xn
grπXn
π0 (Y )
i
k
fn
Y
Xn+1
i
fn+1
gr
i0
Xn+1 × I
i1
Xn+1
where fn+1 is M × M op -equivariant and πXn is the projection. Note that kfn+1 ≃ gr and hence,
since π0 (Y ) is a discrete space, we conclude that kfn+1 = gr. Now take f to be the colimit of
the fn . This completes the proof.
Theorem 7.2. Let X, Y be bi-equivariant classifying spaces for M . Then X and Y are M ×
M op -homotopy equivalent by a cellular M × M op -homotopy equivalence.
Proof. As π0 (X) ∼
=M ∼
= π0 (Y ) as M × M op -sets, there is an M × M op -equivariant isomorphism
g : π0 (X) → π0 (Y ). Then by Lemma 7.1, there is an M × M op -equivariant continuous mapping
f : X → Y inducing g on connected components. It follows that f is a weak equivalence as X
and Y both have contractible connected components. The result now follows from Corollary 2.7
and Theorem 2.8.
34
ROBERT D. GRAY AND BENJAMIN STEINBERG
Next we prove in an elementary fashion that bi-equivariant classifying spaces for M exist. A
more canonical construction, using simplicial sets, was described earlier.
Lemma 7.3. Let M be a monoid.
(1) If X is a projective (free) M ×M op -CW complex such that π0 (X) ∼
= M and πq (X, x) = 0
for all 1 ≤ q < n and x ∈ X, then there exists a projective M × M op -CW complex Y
containing X as a projective M × M op -CW subcomplex and such that Yn = Xn and
πq (Y, y) = 0 for all y ∈ Y and 1 ≤ q ≤ n.
(2) If X is a projective (free) M ×M op -CW complex such that π0 (X) ∼
= M and πq (X, x) = 0
for all 1 ≤ q < n and x ∈ X, then there exists a projective (free) M × M op -CW complex
Y with contractible connected components containing X as a projective M × M op -CW
subcomplex and such that Yn = Xn .
Proof. This is a minor adaptation of the proof of Lemma 6.4 that we leave to the reader.
Corollary 7.4. Let M be a monoid. Then there exists a free M × M op -CW complex X with
π0 (X) ∼
= M and each connected component of X contractible.
∼ M . We take
Proof. By Lemma 7.3 it suffices to construct a free M ×M op -graph X with π0 (X) =
X0 = M × M and we take an edge set in bijection with M × M × M . The edge (mL , mR , m) will
connect (mL , mmR ) to (mL m, mR ). Then X is a free M × M op -graph. Notice that if (m1 , m2 )
is connected by an edge to (m′1 , m′2 ), then m1 m2 = m′1 m′2 . On the other hand, (1, m2 , m1 ) is
an edge from (1, m1 m2 ) to (m1 , m2 ) and hence there is a bijection π0 (X) → M sending the
component of (m1 , m2 ) to m1 m2 and this mapping is an M × M op -equivariant bijection.
Example 7.5. It follows from the definitions and results in Section 5 that the geometric re←−→
alisation |EM | of the nerve of the two-sided Cayley graph category of M is a bi-equivariant
classifying space for M .
Corollary 7.6. If X is a bi-equivariant classifying space for M , then M \X/M ≃ |BM |.
←−→
Proof. We have M \|EM |/M ∼
= |BM |. The result now follows from Theorem 7.2 and Proposition 3.5.
Another important definition for this paper is the following. A monoid M is of type bi-Fn if
there is a bi-equivariant classifying space X for M such that Xn is M ×M op -finite, i.e., M \X/M
has finite n-skeleton. We say that M is of type bi-F∞ if M has a bi-equivariant classifying space
X that is of M × M op -finite type, i.e., M \X/M is of finite type. Clearly by making use of the
←−→
canonical two-sided classifying space |EM | we can immediately conclude that any finite monoid
is of type bi-F∞ .
Recall that a monoid M is said to be of type bi-FPn if there is a partial resolution of the
(ZM, ZM )-bimodule ZM
An → An−1 → · · · → A1 → A0 → ZM → 0
where A0 , A1 , . . . , An are finitely generated projective (ZM, ZM )-bimodules. Monoids of type
bi-FPn were studied by Kobayashi and Otto in [KO01]. We note that this differs from the
definition of bi-FPn considered in [AH03], which is called weak bi-FPn by Pride in [Pri06] where
it is shown to be equivalent to being simultaneously of type left- and right-FPn . In this paper by
bi-FPn we shall always mean bi-FPn in the above sense of Kobayashi and Otto. The property biFPn is of interest because of its connections with the study of Hochschild cohomology [Wei94,
Chapter 9]. Kobayashi investigated bi-FPn in [Kob05, Kob07, Kob10] proving, in particular,
that any monoid which admits a presentation by a finite complete rewriting system is of type
TOPOLOGICAL FINITENESS PROPERTIES
35
bi-FPn . This has applications for the computation of Hochschild cohomology. We shall recover
this theorem of Kobayashi below in Section 11 as an application of our results on equivariant
discrete Morse theory and collapsing schemes. See also [Pas08] for further related results on
bi-FPn .
The following result relates bi-Fn with bi-FPn .
Proposition 7.7. Let M be a monoid.
(1) For 0 ≤ n ≤ ∞, if M is of type bi-Fn , then it is of type bi-FPn .
(2) If M is of type bi-F∞ , then it is of type bi-Fn for all n ≥ 0.
(3) If M is of type bi-Fn for 0 ≤ n ≤ ∞, then M is of type left-Fn and type right-Fn .
(4) For 0 ≤ n ≤ ∞, a group is of type bi-Fn if and only if it is of type Fn .
Proof. The first item follows using that the cellular chain complex of a bi-equivariant classying
space X can be augmented, as discussed earlier, to give a bimodule resolution of ZM and that
if X is built up from pushouts as per (2.1) (with M × M op in place of M ), then the nth -chain
module is isomorphic to ZPn as a bimodule and hence is projective. The second item is trivial.
←−→
←−−
For the third item, one verifies that if EM is the two-sided bar construction, then EM ∼
=
←−→
←−−
EM /M where EM is the left bar construction. Suppose now that X is a bi-equivariant
←−→
classifying space for M such that Xn is M × M op -finite. Then X ≃M ×M op EM by Theo←−−
←−→
rem 7.2. Therefore, X/M ≃M |EM |/M = |EM | and X/M is a projective M -CW complex with
(X/M )n = Xn /M being M -finite by Corollary 3.8. Thus if M is of type bi-Fn for 0 ≤ n ≤ ∞,
then M is of type left-Fn and dually right-Fn .
If G is a group of type bi-Fn , then it is of type Fn by the previous item and Proposition 6.9.
Conversely, suppose that X is a free G-CW complex with G-finite n-skeleton. Then using the
right G-set structure on G × G from Proposition 3.7 we have that Y = (G × G) ⊗G X is a
projective G × Gop -CW complex by Proposition 3.2 such that Yn is G × Gop -finite. Moreover,
π0 (Y ) ∼
= (G×G)⊗G π0 (X) by Proposition 3.4. But π0 (X) is the trivial G-set and (G×G)⊗G 1 ∼
=
G as a G × Gop -set via (g, h) ⊗ 1 7→ gh. Finally, since G is a free
right
G-set
of
on
G-generators
`
by Proposition 3.7, it follows that as a topological space Y = G X and hence each component
of Y is contractible. This completes the proof.
The proof of Proposition 7.7 establishes the following proposition.
Proposition 7.8. If X is a bi-equivariant classifying space for M , then X/M is an equivariant
classifying space for M .
Sometimes it will be convenient to use the following reformulation of the property bi-Fn .
Proposition 7.9. Let M be a monoid. The following are equivalent for 0 ≤ n < ∞.
(1) M is of type bi-Fn
(2) There is an M × M op -finite projective M × M op -CW complex X of dimension at most
n with π0 (X) ∼
= M and πq (X, x) = 0 for 1 ≤ q < n and x ∈ X.
Proof. This is entirely analogous to the proof of Proposition 6.10.
←−−−−→
If M is a monoid and A ⊆ M , then the two-sided Cayley digraph Γ(M, A) is the digraph
with vertex set M × M and with edges in bijection with elements of M × M × A. The directed
edge (mL , mR , a) goes from (mL , amR ) to (mL a, mR ) and we draw it as
a
(mL , amR ) −→ (mL a, mR ).
←−−−−→
Note that Γ(M, A) is a free M × M op -graph and is M × M op -finite if and only if A is finite.
Also note that if (m1 , m2 ) is connected to (m′1 , m′2 ) by an edge, then m1 m2 = m′1 m′2 . Hence
36
ROBERT D. GRAY AND BENJAMIN STEINBERG
multiplication of the coordinates of a vertex induces a surjective M × M op -equivariant mapping
←−−−−→
π0 (Γ(M, A)) → M . If A is a generating set for M , then the mapping is an isomorphism because
if m1 , m2 ∈ M and u ∈ A∗ is a word representing m1 , then there is a directed path labelled by
u from (1, m) to (m1 , m2 ). Namely, if u = a1 · · · ak with ai ∈ A, then the path labelled by u
from (1, m) to (m1 , m2 ) is
a
a
a
a
(1, m) −−1→ (a1 , a2 · · · ak m2 ) −−2→ (a1 a2 , a3 · · · ak m2 ) −−3→ · · · −−k→ (m1 , m2 ).
(7.2)
A monoid M is said to be dominated by a subset A if whenever f, g : M → N are monoid
homomorphisms with f |A = g|A , one has f = g. In other words, the inclusion hAi ֒→ M is
an epimorphism (in the category theory sense). Of course, a generating set of M dominates
M . Note that if A is a subset of an inverse monoid M (e.g., a group), then A dominates M if
and only if A generates M as an inverse monoid. Hence M is finitely generated if and only if
M is finitely dominated. Kobayashi gives an example of an infinitely generated monoid that is
finitely dominated. See [Kob07] for details.
Theorem 7.10. The following are equivalent for a monoid M .
(1) M is of type bi-F1 .
(2) M is of type bi-FP1 .
←−−−−→
(3) There is a finite subset A ⊆ M such that the natural mapping π0 (Γ(M, A)) → M is an
isomorphism.
(4) There is a finite subset A ⊆ M that dominates M .
In particular, any finitely generated monoid is of type bi-F1 .
Proof. The equivalence of (2) and (4) was established by Kobayashi [Kob07] using Isbel’s zig-zag
lemma (actually, the equivalence of (3) and (4) is also direct from Isbel’s zig-zag lemma).
←−−−−→
Assume that (3) holds. Then Γ(M, A) is M × M op -finite and so M is of type bi-F1 by
Proposition 7.9. Proposition 7.7 shows that (1) implies (2).
Assume that M is of type bi-FP1 . Then we have a partial free resolution
µ
ZM ⊗ ZM −−→ ZM −→ 0
of finite type where µ is induced by the multiplication in ZM . Since M is of type bi-FP1 , ker µ
op -module by the
is finitely generated. It is well known that ker µP
is generated as a ZM ⊗ ZMP
elements m ⊗ 1 − 1 ⊗ m with m ∈ M . Indeed, if
ci mi ⊗ ni is in ker µ, then
ci mi ni = 0 and
so
X
X
X
ci mi ⊗ ni =
ci mi ⊗ ni −
ci (1 ⊗ mi ni )
X
=
ci (mi ⊗ ni − 1 ⊗ mi ni )
X
=
ci (mi ⊗ 1 − 1 ⊗ mi )ni .
Hence there is a finite subset A ⊆ M such that ker µ is generated by the elements a⊗1−1⊗a with
←−−−−→
a ∈ A. We claim that the natural surjective mapping π0 (Γ(M, A)) → M is an isomorphism.
Identifying Z[M ×M op ] with ZM ⊗ZM op as rings and Z[M ×M ] with ZM ⊗ZM as bimodules,
we have a bimodule homomorphism
←−−−−→
λ : Z[M × M ] → Zπ0 (Γ(M, A))
←−−−−→
sending (mL , mR ) to its connected component in Γ(M, A) and µ factors as λ followed by the
natural mapping
←−−−−→
Zπ0 (Γ(M, A)) → ZM.
TOPOLOGICAL FINITENESS PROPERTIES
37
Clearly, ker λ ⊆ ker µ and so to prove the result it suffices to show that ker µ ⊆ ker λ. Now ker µ
is generated by the elements (1, a) − (a, 1) with a ∈ A under our identifications. But (1, 1, a) is
an edge from (1, a) to (a, 1). Thus (1, a) − (a, 1) ∈ ker λ for all a ∈ A. This establishes that (2)
implies (3), thereby completing the proof.
If G is a group, then it follows from Theorem 7.10 that G ∪ {0} is of type bi-F1 if and only
if G is finitely generated. Indeed, G ∪ {0} is an inverse monoid and hence finitely dominated
if and only if finitely generated. But G ∪ {0} is finitely generated if and only if G is finitely
generated. On the other hand, G ∪ {0} is both of type left- and right-F∞ for any group G by
Corollary 6.23. Thus bi-Fn is a much stronger notion.
Remark 7.11. It can be shown that if M is a monoid and M 0 is the result of adjoining a 0 to M ,
then if M 0 is of type bi-Fn , then M is of type bi-Fn . The idea is that if X is a bi-equivariant
classifying space for M 0 , then the union Y of components of X corresponding to elements of M
is a bi-equivariant classifying space for M and Yn will be M × M op -finite if Xn is M 0 × (M 0 )op finite. More generally, if T is a submonoid of M such that M \ T is an ideal, then M being of
type bi-Fn implies T is also of type bi-Fn .
Next we show that finitely presented monoids are of type bi-F2 . The proof is similar to the
proof of Theorem 6.14, which is in fact a consequence.
Theorem 7.12. Let M be a finitely presented monoid. Then M is of type bi-F2 .
Proof. Suppose that M is generated by a finite set A with defining relations u1 = v1 , . . . , un =
vn . We construct an M × M op -finite 2-dimensional free M × M op -CW complex X with 1←−−−−→
skeleton the two-sided Cayley graph Γ(M, A) by attaching an M × M op -cell M × M × B 2 for
each relation. Suppose that ui and vi map to mi in M . Let pi , qi be the paths from (1, mi ) to
(mi , 1) labelled by ui and vi , respectively, cf. (7.2). Then we glue in a disk di with boundary
path pi qi−1 and glue in M × M × B 2 using Proposition 2.3 (so {(mL , mR )} × B 2 is sent to
mL di mR ). Then X is a free M × M op -CW complex of dimension at most 2 that is M × M op finite and π0 (X) ∼
= M . By Proposition 6.10, it suffices to prove that each connected component
of X is simply connected.
The connected component X(m) of X corresponding to m ∈ M is a digraph rooted at (1, m)
a
by (7.2). Let Tm be a directed spanning tree for X(m) rooted at (1, m). Let e = (n1 , an2 ) −→
(n1 a, n2 ) be a directed edge of X(m) not belonging to Tm . Then the corresponding generator of
π1 (X(m), (1, m)) is of the form peq −1 where p and q are directed paths from (1, m) to (n1 , an2 )
and (n1 a, n2 ), respectively. Let u be the label of p and v be the label of q. Then ua = v in
M . Thus it suffices to prove that if x, y ∈ A∗ are words which are equal in M to m′ labelling
respective paths from (1, m) to (m′ , m′′ ) with m′ m′′ = m, then the corresponding loop ℓ labelled
xy −1 at (1, m) is null homotopic.
By induction on the length of a derivation from x to y, we may assume that x = wui w′
and y = wvi w′ for some i = 1, . . . , n. Then the path labelled by w starting at (1, m) ends at
(w, mi w′ m′′ ) where we recall that mi is the image of ui , vi in M . Then wdi w′ m′′ is a 2-cell
bounded by parallel paths from (w, mi w′ m′′ ) to (wmi , w′ m′′ ) labeled by ui and vi , respectively.
It follows that the paths labelled by x and y from (1, m) to (m′ , m′′ ) are homotopic relative
to endpoints and hence ℓ is null homotopic. This completes the proof that X(m) is simply
connected.
Remark 7.13. We currently do not know the precise relationship between bi-F2 and finitely
presentability for monoids. Specifically we have the question: Is there a finitely generated bi-F2
monoid that is not finitely presented? Even for inverse monoids this question remains open.
38
ROBERT D. GRAY AND BENJAMIN STEINBERG
We next observe that finitely generated free monoids are bi-F∞ .
Proposition 7.14. Let A be a finite set. Then the free monoid A∗ is of type bi-F∞ .
←−−−−→
←−−−−→
Proof. Each connected component of Γ(M, A) is a tree and hence contractible. Thus Γ(M, A)
is an A∗ × (A∗ )op -finite bi-equivariant classifying space for A∗ .
Theorem 6.15 has an analogue for bi-Fn and bi-FPn with essentially the same proof, which
we omit.
Theorem 7.15. Let M be a monoid of type bi-F2 . Then M is of type bi-Fn if and only if M
is of type bi-FPn for 0 ≤ n ≤ ∞.
Observe that Theorem 7.15 implies that M is of type bi-F∞ if and only if M is of type bi-Fn
for all n ≥ 0. The analogue of Proposition 6.16 in our setting again admits a very similar proof
that we omit.
Proposition 7.16. If M is of type bi-Fn with n ≥ 1, then M has a free M × M op -CW complex
X that is a bi-equivariant classifying space for M where Xn is M × M op -finite.
Proposition 6.26 also has a two-sided analogue.
Proposition 7.17. If M, N are of type bi-Fn , then M × N is of type bi-Fn .
Let us turn to some inheritance properties for bi-Fn . If I is an ideal of M containing an
identity e, then e is a central idempotent and M eM = M e = eM = eM e. Indeed, em =
(em)e = e(me) = me as em, me ∈ I. If f, f ′ ∈ E(M ), then f e, f ′ e ∈ E(eM e) and e(M f ×
f ′ M )e = eM ef e × f ′ eM e as an eM e-eM e-biset and hence is finitely generated projective.
Thus if P is a (finitely generated) projective M × M op -set, then eP e is a (finitely generated)
projective eM e × eM eop -set.
Proposition 7.18. Let M be a monoid and 0 ≤ n ≤ ∞ and e ∈ E(M ) be a central idempotent.
If M is of type bi-Fn , then so is eM e.
Proof. Let X be a bi-equivariant classifying space of M such that Xn is M ×M op -finite. Suppose
that X is obtained via pushouts as per (2.1) (but with M ×M op in place of M ). Then each ePk e
is a projective eM e × eM eop -set and is finitely generated whenever Pk was finitely generated
by the observation preceding the proposition. Thus eXe is a projective eM e × eM eop -CW
complex and (eXe)n = eXn e is eM e × eM eop -finite. Also, since eXe ∼
= (eM × M e) ⊗M ×M op X,
∼
∼
we deduce that π0 (eXe) = eπ0 (X)e = eM e by Proposition 3.4. If X(m) is the component of X
corresponding to m ∈ eM e, then eX(m)e is the component of eXe corresponding to m in eXe
and is a retract of X(m). But X(m) is contractible and hence eX(m)e is contractible. This
shows that eXe is a bi-equivariant classifying space for eM e, completing the proof.
Two monoids M and N are Morita equivalent if the categories of left M -sets and left N -sets
are equivalent. It is known that this is the case if and only if there is an idempotent e ∈ E(M )
such that xy = 1 for some x, y ∈ M with ey = y and eM e ∼
= N [Kna72]. It follows easily
that if M and N are Morita equivalent, then so are M op and N op . Note that if e is as above,
then the functor A 7→ eA ∼
= eM ⊗M A from M -sets to N -sets (identifying N with eM e) is an
equivalence of categories with inverse B 7→ M e ⊗N B. This uses that M e ⊗eM e eM ∼
= M as
op
M × M -sets via the multiplication map (the inverse bijection takes m ∈ M to mxe ⊗ y) and
eM ⊗M M e ∼
= eM e as eM e × (eM e)op -sets (via the multiplication with inverse eme 7→ em ⊗ e).
It follows that if P is a (finitely generated) projective M -set, then eP is a (finitely generated)
projective N -set (as being projective is categorical and a projective is finitely generated if and
only if it is a coproduct of finitely many indecomposable projectives, which is also categorical).
In particular, eM is a finitely generated projective N -set.
TOPOLOGICAL FINITENESS PROPERTIES
39
Proposition 7.19. Let M and N be Morita equivalent monoids and 0 ≤ n ≤ ∞.
(1) M is of type left-Fn if and only if N is of type left-Fn .
(2) M is of type right-Fn if and only if N is of type right-Fn .
(3) M is of type bi-Fn if and only if N is of type bi-Fn .
Proof. By symmetry it suffice to prove the implications from left to right. We may assume
without loss of generality that N = eM e where 1 = xy with ey = y. Notice that 1 = xy = xey
and so replacing x by xe, we may assume that xe = x. To prove (1), suppose that X is an
equivariant classifying space for M such that Xn is M -finite. Then eM ⊗M X ∼
= eX is a
projective N -CW complex by Proposition 3.2 such that (eX)n = eXn is N -finite. But eX is a
retract of X and hence contractible. We deduce that N is of type left-Fn .
The proof of (2) is dual. To prove (3), observe that (e, e)(M × M op )(e, e) = eM e × (eM e)op
and that we have (x, y)(y, x) = (1, 1) and (e, e)(y, x) = (y, x) in M ×M op because xy = e, ey = y
and xe = x. Thus M ×M op is Morita equivalent to N ×N op and eM ×M e is a finitely generated
projective N × N op -set. Suppose that X is a bi-equivariant classifying space for M such that
Xn is M ×M op -finite. Then (eM ×M e)⊗M ×M op X ∼
= eXe is a projective N ×N op -CW complex
such that (eXe)n = eXn e is N × N op -finite by Corollary 3.2. Also, π0 (eXe) ∼
=N
= eπ(X)e ∼
by Proposition 3.4. Moreover, if m ∈ eM e and X(m) is the component of X corresponding to
m, then the component of eXe corresponding to m is eX(m)e, which is a retract of X(m) and
hence contractible. Thus eXe is a bi-equivariant classifying space of N . The result follows.
There are examples of Morita equivalent monoids that are not isomorphic; see [Kna72].
We define the geometric dimension of M to be the minimum dimension of a bi-equivariant
classifying space for M . The Hochschild cohomological dimension of M , which we write dim M ,
is the length of a shortest projective resolution of ZM as a Z[M × M op ]-module. Of course,
the Hochschild cohomological dimension bounds both the left and right cohomological dimension and the geometric dimension bounds the Hochschild cohomological dimension. Also the
geometric dimension bounds both the left and right geometric dimensions because if X is a biequivariant classifying space for M of dimension n, then X/M is a classifying space of dimension
n.
The following theorem has an essentially identical proof to Theorem 6.27.
Theorem 7.20. Let M be a monoid. Then M has a bi-equivariant classifying space of dimension max{dim M, 3}.
Free monoids have a forest for a bi-equivariant classifying space and hence have geometric
dimension 1. It is well known (see e.g. [Mit72]) that they have Hochschild cohomological
dimension 1.
It is known that a monoid has Hochschild cohomological dimension 0 if and only if it is a
finite regular aperiodic monoid with sandwich matrices invertible over Z (see [Che84]). For
instance, any finite aperiodic inverse monoid has Hochschild cohomogical dimension 0. A nontrivial monoid of Hochschild cohomological dimension 0 does not have geometric dimension 0
because M would have to be a projective M -biset. So M ∼
= M e × f M , with e, f ∈ E(M ),
via an equivariant map ϕ sending (e, f ) to 1 (as M being finite aperiodic implies that 1 is the
unique generator of M as a two-sided ideal). But then f = ϕ(e, f )f = ϕ(e, f ) = 1 and similarly
e = 1 and so M is trivial. Thus non-trivial monoids of Hochschild cohomological dimension 0
do not have geometric dimension 0.
40
ROBERT D. GRAY AND BENJAMIN STEINBERG
8. Brown’s theory of collapsing schemes
The theory of collapsing schemes was introduced by Brown in [Bro92]. Since then it has
become an important and often-used tool for proving that certain groups are of type F∞ . The
first place the idea appears is in a paper of Brown and Geoghegan [BG84] where they had a
cell complex with one vertex and infinitely many cells in each positive dimension, and they
showed how it could be collapsed to a quotient complex with only two cells in each positive
dimension. Brown went on to develop this idea further in [Bro92] formalising it in his theory
of collapsing schemes, and applying it to give a topological proof that groups which admit
presentations by finite complete rewriting systems are of type F∞ (see Section 11 below for
the definition of complete rewriting system). Brown’s theory of collapsing schemes was later
rediscovered under the name of discrete Morse theory [For95, For02], an important area in
algebraic combinatorialists. Chari [Cha00] formulated discrete Morse theory combinatorially
via Morse matchings, which turn out to be the same thing as collapsing schemes.
The basic idea of collapsing schemes for groups is a follows. Suppose we are given a finitely
presented group G and we would like to prove it is of type F∞ . Then we can first begin with the
big K(G, 1) complex |BG| with infinitely many n-cells for each n. Then in certain situations
it is possible to show how one can collapse away all but finitely many cells in each dimension
resulting in a K(G, 1) much smaller than the one we started with. The collapse is carried out
using a so-called collapsing scheme associated with the simplicial set BG. It turns out that any
group which is presentable by a finite complete rewriting system admits a collapsing scheme
that, using this process, can be used to prove the group is of type F∞ ; see [Bro92, page 147].
As mentioned in the introduction above, Brown in fact develops this theory for monoids in
general, and applies the theory of collapsing schemes to show that if M admits a presentation
by a finite complete rewriting system then its classifying space |BM | has the homotopy type of
a CW complex with only finitely many cells in each dimension. However, as discussed in detail
in the introduction to this article, this information about the space |BM | is not enough on its
own to imply that the monoid M is of type left-FP∞ .
Motivated by this, in this section we shall develop the theory of M -equivariant collapsing
schemes. We shall prove that if an M -simplicial set admits an M -equivariant collapsing scheme
of finite type then the monoid is of type left-F∞ . We then prove that if M admits a finite
−−→
complete rewriting system then EM admits an M -equivariant collapsing scheme of finite type,
thus giving a topological proof that such monoids are of type left-F∞ . To do this, we shall identify conditions under which a collapsing scheme for BM can be lifted to give an M -equivariant
−−→
collapsing scheme for EM . These conditions will hold in a number of different situations, including when M admits a presentation by a finite complete rewriting system and when M is a,
so-called, factorable monoid [HO14]. We also develop the two-sided theory. As a consequence
we also obtain a topological proof of the fact that such a monoid is of type bi-F∞ , recovering
a theorem of Kobayashi [Kob05].
S
8.1. Collapsing schemes. Let K = i≥0 Ki be a simplicial set and let X = |K| be its
geometric realisation. We identify the cells of X with the non-degenerate simplices of K. A
collapsing scheme for K consists of the following data:
• A partition of the cells of X into three classes, E, C, R, called the essential, collapsible
and redundant cells, respectively, where the collapsible cells all have dimension at least
one.
• Mappings c and i which associate with each redundant n-cell τ a collapsible (n + 1)-cell
c(τ ), and a number i(τ ), such that τ = di(τ ) (c(τ )).
TOPOLOGICAL FINITENESS PROPERTIES
41
Let σ = c(τ ). If τ ′ is a redundant n-cell such that τ ′ = dj σ for some j 6= i(τ ) then we call τ ′
an immediate predecessor of τ and write τ ′ ≺ τ . Furthermore, the conditions for a collapsing
scheme are satisfied, which means:
(C1) for all n, the mapping c defines a bijection between Rn (the redundant n-cells) and
Cn+1 (the collapsible (n + 1)-cells).
(C2) there is no infinite descending chain τ ≻ τ ′ ≻ τ ′′ ≻ · · · of redundant n-cells.
Condition (C2) clearly implies that there is a unique integer i such that τ = di (c(τ )) (otherwise
we would have τ ≻ τ , leading to an infinite descending chain). It also follows from (C2) that,
by Königs lemma, there cannot be arbitrarily long descending chains τ0 ≻ · · · ≻ τk . This is a
key fact in the proof of [Bro92, Proposition 1] since it gives rise to the notion of ‘height’:
Definition 8.1 (Height). The height of a redundant cell τ , written height(τ ), is the maximum
length of a descending chain τ = τ0 ≻ τ1 ≻ · · · ≻ τk .
We say that a collapsing scheme is of finite type if it has finitely many essential cells of each
dimension.
In the construction of the ‘small’ CW complex in the proof of [Bro92, Proposition 1] the
redundant n-cells are adjoined in order of their heights, guaranteeing in the proof that the
adjunction of τ and c(τ ) is an elementary expansion. This is the the key idea in the proof
which is that each pair (τ, c(τ )) of redundant and corresponding collapsible cells may be adjoined
without changing the homotopy type, and so in the end it is only the essential cells that matter.
More precisely, Brown proves that if K be a simplical set with a collapsing scheme, then its
geometric realisation X = |K| admits a canonical quotient CW complex Y , whose cells are
in 1–1 correspondence with the essential cells of X. This notion of height in Brown’s theory
relates to the values taken by the discrete Morse function in Forman’s theory (see [For02, page
10]). A discrete Morse function gives one a way to build the simplicial complex by attaching
the simplices in the order prescribed by the function, i.e., adding first the simplices which are
assigned the smallest values. Brown’s essential cells are called ‘critical’ in Forman’s theory.
9. M -equivariant collapsing schemes
In this section we develop the theory of M -equivariant collapsing schemes, or equivalently,
of M -equivariant discrete Morse theory. Results on G-equivariant discrete Morse theory, for G
a group, may
S be found in [Fre09].
Let K = i≥0 Ki be a simplicial set with degeneracy and face operators di , si , and equipped
with a collapsing scheme (E, R, C, c, i). Here E, R and C partition the cells (which are in
bijective correspondence with the non-degenerate simplices) of K.
Let M be a monoid acting on the simplicial set K with the following conditions satisfied:
(A1) The action of M maps n-simplicies to n-simplicies, and commutes with di and si , that
is, M is acting by simplicial morphisms.
(A2) For every n-simplex σ and m ∈ M , σ is a cell (i.e. is a non-degenerate simplex) if and
only if mσ is a cell, in which case σ ∈ E (respectively R, C) if and only if mσ ∈ E
(respectively R, C).
(A3) If (σ, τ ) ∈ Rn × Cn+1 is a matched redundant-collapsible pair (i.e. τ = c(σ)) then so is
the pair m(σ, τ ) = (mσ, mτ ) ∈ Rn × Cn+1 , i.e., c(mσ) = mc(σ) for σ ∈ Rn .
(A4) There is a subset B ⊆ E ∪ R ∪ C such that for all n the set of n-cells is a free left M -set
with basis Bn (where Bn is the set of n-cells in B). Let E B = E ∩ B, RB = R ∩ B and
C B = C ∩ B. Then En is a free left M -set with basis EnB , and similarly for Rn and Cn .
(A5) For every matched pair (σ, τ ) ∈ R × C, σ ∈ RB if and only if τ ∈ C B . In particular,
for every matched pair (σ, τ ) there is a unique pair (σ ′ , τ ′ ) ∈ RB × C B and m ∈ M
42
ROBERT D. GRAY AND BENJAMIN STEINBERG
such that (σ, τ ) = m(σ ′ , τ ′ ); namely, if σ = mσ ′ with σ ′ ∈ RB and τ ′ = c(σ ′ ), then
mτ ′ = mc(σ ′ ) = c(mσ ′ ) = c(σ) = τ .
(A6) For every redundant cell τ and every m ∈ M
height(τ ) = height(mτ ),
with height defined as in Definition 8.1 above.
These conditions imply that K is a rigid free left M -simplicial set and hence by Lemma 4.4 the
action of M on K induces an action of M on the geometric realisation |K| by continuous maps
making |K| into a free left M -CW complex. When the above axioms hold, we call (E, R, C, c, i)
an M -equivariant collapsing scheme for the rigid free left M -simplicial set K. Dually, given a
rigid free right M -simplicial set K with a collapsing scheme satisfying the above axioms for K
as an M op -simplicial set we call (E, R, C, c, i) an M -equivariant collapsing scheme for K. If K
is a bi-M -simplicial set we say (E, R, C, c, i) an M -equivariant collapsing scheme if the axioms
are satisfied for K as a left M × M op -simplicial set.
Our aim is to prove a result about the M -homotopy type of |K| when K has an M -equivariant
collapsing scheme. Before doing this we first make some observations about mapping cylinders
and the notion of elementary collapse.
9.1. Mapping cylinders and elementary collapse. If X is a subspace of a space Y then
D : Y → X is a strong deformation retraction if there is a map F : Y × I → Y such that, with
Ft : Y → Y defined by Ft (y) = F (y, t), we have (i) F0 = 1Y , (ii) Ft (x) = x for all (x, t) ∈ X × I,
and (iii) F1 (y) = D(y) for all y ∈ Y . If D : X → Y is a strong deformation retraction then D
is a homotopy equivalence, a homotopy inverse of which is the inclusion i : X ֒→ Y .
Definition 9.1 (Mapping cylinder). Let f : X → Y be a cellular map
`between CW complexes.
The mapping cylinder Mf is defined to be the adjunction complex Y f0 (X × I) where f0 : X ×
{0} is the map (x, 0) 7→ f (x). Let i1 : X → X × I, x 7→ (x, 1) and let i0 : X → X × I, x 7→ (x, 0).
Let p be the projection p : X × I → X, (x, i) 7→ x. Also set i = k ◦ i1 , with k as below. Thus
we have
X
p
f
Y
j
i1 i0
X ×I
k
Mf
The map from X × I to Y taking (x, t) to f (x), and the identity map on Y , together induce a
retraction r : Mf → Y .
The next proposition is [Geo08, Proposition 4.1.2].
Proposition 9.2. The map r is a homotopy inverse for j, so r is a homotopy equivalence.
Indeed there is a strong deformation retraction D : Mf × I → Mf of Mf onto Y such that
D1 = r.
The following result is the M -equivariant analogue of [Geo08, Proposition 4.1.2]. Recall that
if X is a projective (resp. free) M -CW complex then Y = M × I is a projective (resp. free)
M -CW complex, where I is given the trivial action.
TOPOLOGICAL FINITENESS PROPERTIES
43
Lemma 9.3. Let f : X → Y be a continuous M -equivariant cellular map between free (projective) M -CW complexes X and Y . Let Mf be the pushout of
X
f
Y
j
i0
X ×I
k
Mf
where i0 : X → X × I, x 7→ (x, 0). Then:
(i) Mf is a free (projective) M -CW complex; and
(ii) there is an M -equivariant strong deformation retraction r : Mf → Y . In particular Mf
and Y are M -homotopy equivalent.
Proof. It follows from Lemma 2.2 that Mf has the structure of a free (projective) M -CW
complex, proving part (i). For part (ii) first note that the map from X × I to Y taking (x, t)
to f (x), and the identity map on Y , together induce a retraction r : Mf → Y . It follows from
Proposition 9.2 that r is a homotopy equivalence with homotopy inverse j. By Corollary 2.7 to
show that r is an M -homotopy equivalence it suffices to verify that r is an M -equivariant map
between the sets Mf and Y . But M -equivariance of r follows from the definitions of r and Mf ,
the definition of the action of M on Mf which in turn is determined by the actions of M on
X × I and on Y , together with the assumption that f : X → Y is M -equivariant.
The fundamental idea of collapsing schemes, and discrete Morse theory, is that of a collapse.
The following definition may be found in [Coh73, page 14] and [FS05, Section 2]. We use the
same notation as in [Coh73]. In particular ≈ denotes homeomorphism of spaces.
Definition 9.4 (Elementary collapse). If (K, L) is a CW pair then K collapses to L by an
elementary collapse, denoted K ցe L, if and only if:
(1) K = L∪en−1 ∪en where en and en−1 are open cells of dimension n and n−1 respectively,
which are not in L, and
(2) there exists a ball pair (B n , B n−1 ) ≈ (I n , I n−1 ) and a map ϕ : B n → K such that
(a) ϕ is a characteristic map for en
(b) ϕ|B n−1 is a characteristic map for en−1
(c) ϕ(P n−1 ) ⊆ Ln−1 where P n−1 = cl(∂B n − B n−1 ).
Note that in this statement P n−1 is an (n − 1)-ball (i.e. is homeomorphic to I n−1 ). We
say that K collapses to L, writing K ց L, if L may be obtained from K by a sequence of
elementary collapses. We also say that K is an elementary expansion of L. An elementary
collapse gives a way of modifying a CW complex K, by removing the pair {en−1 , en }, without
changing the homotopy type of the space. We can write down a homotopy which describes such
an elementary collapse K ցe L as follows. Let (K, L) be a CW pair such that K ցe L. Set
ϕ0 = ϕ|pn−1 in the above definition. Then
ϕ0 : (P n−1 , ∂P n−1 ) → (Ln−1 , Ln−2 ),
(using the identification P n−1 ≈ I n−1 ) and
(K, L) ≈ (L
a
B n , L).
ϕ0
The following is [Coh73, (4.1)].
Lemma 9.5. If K ցe L then there is a cellular strong deformation retraction D : K → L.
44
ROBERT D. GRAY AND BENJAMIN STEINBERG
Indeed, let K = L ∪ en−1 ∪ en . There is a map ϕ0 : I n−1 ≈ P n−1 → Ln−1 such that
a
B n , L).
(K, L) ≈ (L
ϕ0
`
But L ϕ0 B n is the mapping cylinder Mϕ0 of ϕ0 : I n−1 → Ln−1 . Thus by Proposition 9.2 there
is a strong deformation retraction
a
In → L
D: K ≈ L
ϕ0
such that D(en ) = ϕ0 (I n−1 ) ⊂ Ln−1 . The map D is given by the map r in Definition 9.1.
We may now state and prove the main result of this section.
Theorem 9.6. Let K be a rigid free left M -simplicial set with M -equivariant collapsing scheme
(E, R, C, c, i) (that is, the conditions (A1)–(A6) are satisfied). Then, with the above notation,
there is a free left M -CW complex Y with Y ≃M |K| and such that the cells of Y are in bijective
correspondence with E, and under this bijective correspondence Yn is a free left M -set with basis
EnB for all n.
Proof. Let X be the geometric realisation |K| of the simplicial set K. By axiom (A1) we have
that K is a left M -simplical set, and it follows that X = |K| has the structure of a left M -CW
complex where the M -action is given by Lemma 4.4. In fact, by assumptions (A2)-(A6), X is
a free M -CW complex where, for each n, the set EnB ∪ RnB ∪ CnB is a basis for the n-cells.
Write X as an increasing sequence of subcomplexes
X0 ⊆ X0+ ⊆ X1 ⊆ X1+ ⊆ . . .
where, X0 consists of the essential vertices, Xn+ is obtained from Xn by adjoining the redundant
n-cells and collapsible (n + 1)-cells, and Xn+1 is obtained from Xn+ by adjoining the essential
(n + 1)-cells. We write Xn+ as a countable union
Xn = Xn0 ⊆ Xn1 ⊆ Xn2 ⊆ . . .
`
with Xn+ = i≥0 Xni where Xnj+1 is constructed from Xnj by adjoining (τ, c(τ )) for every redundant n-cell τ of height j. From assumptions (A1)-(A6), for every n and j, each of Xn+ , Xn and
Xnj is a free M -CW subcomplex of X.
As argued in the proof of [Bro92, Proposition 1], for every redundant n-cell τ of height j the
adjunction of (τ, c(τ )) is an elementary expansion. In this way Xnj+1 can be obtained from Xnj
by a countable sequence of simultaneous elementary expansions. The same idea, together with
Lemma 9.3, can be used to obtain an M -homotopy equivalence between Xnj and Xnj+1 . The
details are as follows.
Recall that Xnj+1 is obtained from Xnj by adjoining (τ, c(τ )) for every redundant n-cell τ of
height j. It follows from the axioms (A1)-(A6) that this set of pairs (τ, c(τ )) is a free M -set with
B
B
with height(τ ) = j,
: height(τ ) = j}. Let (τ, c(τ )) ∈ RnB × Cn+1
basis {(τ, c(τ )) ∈ RnB × Cn+1
and let m ∈ M . From the assumptions (A1)-(A6) it follows that
m · (τ, c(τ )) = (mτ, c(mτ )),
and
height(mτ ) = height(τ ) = j.
The pair
(Xnj+1 , Xnj+1 − {mτ, c(mτ )})
satisfies the conditions of an elementary collapse. Indeed (as argued in the proof of [Bro92,
Proposition 1]) every face of c(mτ ) other than mτ is either (i) a redundant cell of height
< j, (ii) is essential (so has height 0), or (iii) is collapsible or degenerate. It follows that the
adjunction of mτ and c(mτ ) is an elementary expansion. This is true for every pair (mτ, c(mτ ))
TOPOLOGICAL FINITENESS PROPERTIES
45
B
with height(τ ) = j and m ∈ M . Now Xnj+1 is obtained from Xnj
where (τ, c(τ )) ∈ RnB × Cn+1
B
: height(τ ) = j} and
by adjoining all such pairs (mτ, c(mτ )). Let A = {(τ, c(τ )) ∈ RnB × Cn+1
let M × A denote the free left M -set with basis {(1, a) : a ∈ A} and action m(n, a) = (mn, a).
Then (M × A) × I n+1 is a disjoint union of the free M -cells (M × {a}) × I n+1 with a ∈ A. The
characteristic maps for the collapsible n + 1 cells of height j combine to give an M -equivariant
map ϕ : (M × A) × I n+1 → Xnj+1 such that
(E1) ϕ restricted to the (m, a) × I n+1 , (m ∈ M, a ∈ A), gives characteristic maps for each of
the collapsible n + 1 cells of height j;
(E2) ϕ restricted to the (m, a) × I n , (m ∈ M, a ∈ A), gives characteristic maps for each
τ ∈ Rn such that c(τ ) is a collapsible n + 1 cell of height j;
(E3) ϕ((M × A) × P n ) ⊆ (Xnj+1 )≤n (where (Xnj+1 )≤n is the subcomplex of Xnj+1 of cells of
dimension ≤ n) where P n = cl(∂I n+1 − I n ).
Set ϕ0 = ϕ|(M ×A)×P n . Then
ϕ0 : (M × A) × P n → (Xnj+1 )≤n
is a continuous M -equivariant cellular map between free M -CW complexes (M × A) × P n and
(Xnj+1 )≤n . It follows that Xnj+1 is M -equivariantly isomorphic to
a
(9.1)
(Xnj+1 − {(τ, c(τ )) ∈ Rn × Cn+1 : height(τ ) = j}) ((M × A) × I n+1 ).
ϕ0
Pn
cl(∂I n+1
− I n)
In
But since
=
≈
we conclude that (9.1) is just the mapping cylinder of ϕ0 .
Thus we can apply Lemma 9.3 to obtain a strong deformation retraction rj : Xnj+1 → Xnj which
is also an M -homotopy equivalence. It follows that there is a retraction rn : Xn+ → Xn which
is an M -equivariant homotopy equivalence.
We build the space Y such that |K| ≃M Y inductively. First set Y0 = X0 . Now suppose
+
that we have an M -homotopy equivalence π n−1 : Xn−1
→ Yn−1 is given. Define Yn to be the
B
B
the M -CW complex Yn−1 ∪ (M × En ) where (M × En ) is a collection of free M -cells indexed
by EnB . These free M -cells are attached to Yn by composing with π n−1 the attaching maps
+
for the essential n-cells of X. This makes sense because Xn−1
contains the (n − 1)-skeleton
n−1
n−1 : X → Y in the obvious way. This
of X. Extend π
to an M -homotopy equivalence π[
n
n
+
is possible since Xn is obtained from Xn−1 by adjoining the essential n-cells. Composing rn
n−1 then gives an M -homotopy equivalence X + → Y . Passing to the union gives the
with π[
n
n
M -homotopy equivalence X ≃M Y stated in the theorem.
There is an obvious dual result for right simplicial sets which follows from the above result
by replacing M by M op . We also have the two-sided version.
Theorem 9.7. Let K be a rigid free bi-M -simplicial set with M -equivariant collapsing scheme
(E, R, C, c, i). Then, with the above notation, there is a free bi-M -CW complex Y with Y ≃M
|K| and such that the cells of Y are in bijective correspondence with E, and under this bijective
correspondence Yn is a free bi-M -set with basis EnB for all n.
Proof. This follows by applying Theorem 9.6 to the rigid free left M ×M op -simplicial set K.
10. Guarded collapsing schemes
In this section we introduce the idea of a left guarded collapsing scheme. We shall prove that
−−→
whenever BM admits a left guarded collapsing scheme then EM will admit an M -equivariant
collapsing scheme whose M -orbits of cells are in bijective correspondence with the essential
46
ROBERT D. GRAY AND BENJAMIN STEINBERG
cells of the given collapsing scheme for BM . Applying Theorem 9.6 it will then follow that
when BM admits a left guarded collapsing scheme of finite type then the monoid M is of
type left-F∞ . Analogous results will hold for right guarded and right-F∞ , and (two-sided)
guarded and bi-F∞ . In later sections we shall give some examples of monoids which admit
guarded collapsing schemes of finite type, including monoids with finite complete presentations
(rewriting systems), and factorable monoids in the sense of [HO14].
S
Definition 10.1 (Guarded collapsing schemes). Let K = i≥0 Ki be a simplicial set and let
X be its geometric realisation. We identify the cells of X with the non-degenerate simplices of
K, and suppose that K admits a collapsing scheme (E, C, R, c, i). We say that this collapsing
scheme is
• left guarded if for every redundant n-cell τ we have i(τ ) 6= 0;
• right guarded if for every redundant n-cell τ we have i(τ ) 6= n + 1;
• guarded if it is both left and right guarded.
In other words, the collapsing scheme is guarded provided the function i never takes either
of its two possible extreme allowable values. The aim of this section is to prove the following
result.
Theorem 10.2. Let M be a monoid and suppose that BM admits a collapsing scheme (E, C, R, c, i).
(a) If (E, C, R, c, i) is left guarded , then there is an M -equivariant collapsing scheme (E, R, C, κ, ι)
−−→
for the free left M -simplicial set EM such that, for each n, the set of essential n-cells En
is a free left M -set of rank |En |.
(b) If (E, C, R, c, i) is right guarded , then there is an M -equivariant collapsing scheme (E, R, C, κ, ι)
←−−
for the free right M -simplicial set EM such that, for each n, the set of essential n-cells En
is a free right M -set of rank |En |.
(c) If (E, C, R, c, i) is guarded , then there is an M -equivariant collapsing scheme (E, R, C, κ, ι)
←−→
for the free bi-M -simplicial set EM such that, for each n, the set of essential n-cells En is
a free M × M op -set of rank |En |.
Corollary 10.3. Let M be a monoid and let |BM | be its classifying space.
(a) If BM admits a left guarded collapsing scheme of finite type then M is of type left F∞ (and
therefore also is of type left-FP∞ ).
(b) If BM admits a right guarded collapsing scheme of finite type then M is of type right-F∞
(and therefore also is of type right-FP∞ ).
(c) If BM admits a guarded collapsing scheme of finite type then M is of type bi-F∞ (and
therefore also is of type bi-FP∞ ).
Proof. This follows directly from Theorem 9.6 and its dual, and Theorems 9.7 and 10.2.
We shall give some examples of monoids to which this corollary applies in the next section.
The rest of this section will be devoted to the proof of Theorem 10.2. Clearly part (b) of the
theorem is dual to part (a). The proofs of parts (a) and (c) are very similar, the only difference
being that the stronger guarded condiion is needed for (c), while only left guarded is needed
for (a). We will begin by giving full details of the proof of Theorem 10.2(a). Then afterwards
we will explain the few modifications in the proof necessary to obtain the two-sided proof for
(c), in particular highlighting the place where the guarded condition is needed.
10.1. Proof of Theorem 10.2(a). Let (E, C, R, c, i) be a left guarded collapsing scheme for
BM . We can now ‘lift’ this collapsing scheme to a collapsing scheme (E, R, C, κ, ι) for the
TOPOLOGICAL FINITENESS PROPERTIES
47
−−→
simplicial set EM in the following natural way. First observe that
m(m1 , m2 , ..., mn ) is an n-cell of EM
⇔ mi 6= 1 for all 1 ≤ i ≤ n
⇔ (m1 , m2 , ..., mn ) is an n-cell of BM .
−−→
Define an n-cell m(m1 , m2 , ..., mn ) of EM to be essential (respectively redundant, collapsible respectively) if and only if (m1 , m2 , ..., mn ) is essential (respectively redundant, collapsible
respectively) in the collapsing scheme (E, R, C, c, i). This defines the partition (E, R, C) of the
−−→
n-cells of EM for each n. We call these sets the essential, redundant and collapsible cells,
−−→
−−→
respectively, of EM . For the mappings κ and ι, given mτ ∈ EM where τ = (m1 , ..., mn ) is a
redundant n-cell of BM we define
ι(mτ ) = i(τ )
κ(mτ ) = m(c(τ )).
(10.1)
(10.2)
We claim that (E, R, C, κ, ι) is an M -equivariant collapsing scheme for the free left M -simplicial
−−→
set EM such that, for each n, the set of essential n-cells En is a free left M -set of rank |En |.
Once proved this will complete the proof of Theorem 10.2(a).
−−→
We begin by proving that (E, R, C, κ, ι) is a collapsing scheme for EM , and then we will
verify that all of the conditions (A1) to (A6) are satisfied.
10.1.1. Proving that (E, R, C, κ, ι) is a collapsing scheme.
Proposition 10.4. With the above definitions, (E, R, C, κ, ι) is a collapsing scheme for the
−−→
simplicial set EM .
−−→
We have already observed that (E, R, C) partitions the cells of EM . To complete the proof
of the proposition we need to verify that the conditions (C1) and (C2) in the definition of
collapsing scheme are both satisfied. For part of the proof we will find it useful to recast the
ideas in terms of certain bipartite digraphs. The idea of viewing collapsing schemes in terms
of matchings in bipartite graphs is a natural one and has been used in the literature; see for
example [Cha00]. Let us now introduce the terminology and basic observations about digraphs
that we shall need.
A directed graph D consists of: a set of edges ED, a set of vertices V D and functions α
and β from ED to V D. For e ∈ E we call α(e) and β(e) the initial and terminal vertices of
the directed edge e. A directed path of length n in D is a sequence of edges e1 e2 . . . en such
that β(ei ) = α(ei+1 ) for each directed edge. Note that edges in paths are allowed to repeat,
and vertices can also repeat in the sense that β(ei ) = β(ej ) is possible for distinct i and j
(in graph theory literature what we call a path here is often called a walk). By an infinite
directed path we mean a path (ei )i∈N . Note that an infinite directed path need not contain
infinitely many distinct edges. For example, if a digraph contains a directed circuit e1 e2 e3
then e1 e2 e3 e1 e2 e3 . . . would be an infinite directed path. A bipartite digraph D with bipartition
V D1 ∪ V D2 has vertex set V D = V D1 ∪ V D2 where V D1 and V D2 are disjoint, such that for
every e ∈ ED we have either α(e) ∈ V D1 and β(e) ∈ V D2 , or α(e) ∈ V D2 and β(e) ∈ V D1
(i.e., there are no directed edges between vertices in the same part of the bipartition). A
homomorphism ϕ : D → D ′ between digraphs is a map ϕ : (V D ∪ ED) → (V D ′ ∪ ED ′ ) which
maps vertices to vertices, edges to edges, and satisfies α(ϕ(e)) = ϕ(α(e)) and β(ϕ(e)) = ϕ(β(e)).
If p = e1 e2 . . . en is a path of length n in D and ϕ : D → D ′ is a digraph homomorphism then
we define ϕ(p) = ϕ(e1 )ϕ(e2 ) . . . ϕ(en ) which is a path of length n in D ′ . Note that in general a
48
ROBERT D. GRAY AND BENJAMIN STEINBERG
homomorphism is allowed to map two distinct vertices (resp. edges) of D to the same vertex
(resp. edge) of D ′ . Since digraph homomorphisms map paths to paths, we have the following
basic observation.
Lemma 10.5. Let ϕ : D → D ′ be a homomorphism of directed graphs. If D has an infinite
directed path than D ′ has an infinite directed path.
For each n ∈ N let Γ(n) (BM ) be the directed bipartite graph defined as follows. The vertex
set V Γ(n) (BM ) = Cn ∪ Cn+1 where Ci denotes the set of i-cells BM . The directed edges EV of
V are are of two types:
(DE1) A directed edge τ −→ dj (τ ) (with initial vertex τ and terminal vertex dj (τ )) for every
collapsible τ ∈ Cn+1 and j ∈ {0, . . . , n + 1} such that dj (τ ) is a redundant n-cell (i.e., is a
redundant non-degenerate n-simplex) and either c(dj (τ )) 6= τ or c(dj (τ )) = τ but j 6= i(dj (τ ));
(DE2) A directed edge σ −→ c(σ) (with initial vertex σ and terminal vertex c(σ)) for every
redundant σ ∈ Cn .
We sometimes refer to these two types of directed arcs as the “down-arcs” and the “uparcs” (respectively) in the bipartite graph. Note that condition (C2) in the collapsing scheme
definition is equivalent to saying that the digraph D(n, n + 1) does not contain any infinite
−−→
directed path, and in particular does not contain any directed cycles. Let Γ(n) (EM ) be the
corresponding directed bipartite graph defined in the same way using (E, R, C, κ, ι), with vertex
−−→
set the n and n + 1 cells of EM and directed edges determined by the maps κ and ι. To simplify
notation, let us fix n ∈ N and set Γ(BM ) = Γ(n) (BM ) and Γ(EM ) = Γ(n) (EM ).
−−→
−−→
Lemma 10.6. Let π : V Γ(EM ) ∪ EΓ(EM ) → V Γ(BM ) ∪ EΓ(BM ) be defined on vertices by:
m(m1 , . . . , mk ) 7→ (m1 , . . . , mk ) (k ∈ {n, n + 1})
and defined on edges (DE1) and (DE2) by π(x → y) = π(x) → π(y). Then π is a digraph
homomorphism.
−−→
Proof. We need to prove, for each directed edge x → y from EΓ(EM ), that π(x) → π(y) is a
directed edge in EΓ(BM ). There are two cases that need to be considered depending on arc
type. The two arc types depend on whether the arc is going downwards from level n + 1 to
level n (arc type 1) or upwards from level n to level n + 1 (arc type 2).
−−→
Case: Arc type 1. Let m(m1 , . . . , mn+1 ) be a collapsible n + 1 cell in EM and let j ∈
{0, . . . , n + 1} and suppose that
m(m1 , . . . , mn+1 ) −→ dj (m(m1 , . . . , mn+1 ))
−−→
−−→
is an arc in Γ(EM ). This means that dj (m(m1 , . . . , mn+1 )) is a redundant n-cell in EM and
if κ(dj (m(m1 , . . . , mn+1 ))) = m(m1 , . . . , mn+1 ), then j 6= ι(dj (m(m1 , . . . , mn+1 ))). Note that
j = 0 or j = n + 1 is possible here. We claim that
π(m(m1 , . . . , mn+1 )) −→ π(dj (m(m1 , . . . , mn+1 )))
−−→
in Γ(BM ). Indeed, we saw above in Section 5 that the projection π : EM → BM is a simplicial
morphism with
π(dj (m(m1 , . . . , mn+1 ))) = dj (m1 , . . . , mn+1 ) = dj (π(m(m1 , . . . , mn+1 ))).
TOPOLOGICAL FINITENESS PROPERTIES
49
Therefore
π(m(m1 , . . . , mn+1 ))
= (m1 , . . . , mn+1 ) → dj (m1 , . . . , mn+1 )
= π(dj (m(m1 , . . . , mn+1 ))),
in Γ(BM ), since if c(dj (m1 , . . . , mn+1 )) = (m1 , . . . , mn+1 ), then
κ(dj (m(m1 , . . . , mn+1 ))) = κ(mdj (m1 , . . . , mn+1 )) = mc(dj (m1 , . . . , mn+1 )) = m(m1 , . . . , mn+1 )
and hence by definition of ι we have
j 6= ι(dj (m(m1 , . . . , mn+1 ))) = i(dj (m1 , . . . , mn+1 )).
Case: Arc type 2. These are the arcs that go up from level n to level n + 1. A typical such
−−→
arc arises as follows. Let mσ ∈ Cn be a redundant cell in EM where σ is a redundant cell in
BM and m ∈ M . Then κ(mσ) ∈ Cn+1 is collapsible and
dι(mσ) (κ(mσ)) = mσ.
A typical type 2 arc has the form
mσ −→ κ(mσ) = m(c(σ)),
by definition of κ. Applying π to this arc gives
π(mσ) = σ −→ c(σ) = π(m(c(σ))),
which is a type 2 arc in Γ(BM ) completing the proof of the lemma.
Proof of Proposition 10.4. We must check that (E, R, C, κ, ι) satisfies the two collapsing scheme
conditions (C1) and (C2).
Verifying collapsing scheme condition (C1). We must prove that the map κ defines a
bijection from the redundant n-cells to the collapsible n + 1-cells.
The map is injective since κ(mτ ) = κ(m′ σ) implies mc(τ ) = m′ c(σ) so m = m′ and τ = σ
since c is injective by assumption. Also, given an arbitrary collapsible n + 1 cell mσ there exists
σ = c(τ ) where τ is a redundant n-cell an so mσ = κ(mτ ).
Moreover, for every redundant n-cell mτ , we have
dι(mτ ) (κ(mτ )) = di(τ ) (mc(τ )) (by definition)
= mdi(τ ) (c(τ )) (since i(τ ) 6= 0)
= mτ,
and this concludes the proof that collapsing scheme condition (C1) holds.
Note that it is in the second line of the above sequence of equations that we appeal to our
assumption that (E, C, R, c, i) is left guarded (which implies i(τ ) 6= 0). In fact, this will be the
only place in the proof of Theorem 10.2(a) that the left guarded assumption is used.
Verifying collapsing scheme condition (C2). To see that collapsing scheme condition
−−→
−−→
(C2) holds let Γ(BM ) and Γ(EM ) be the level (n, n + 1) bipartite digraphs of BM and EM ,
respectively, defined above. By Lemma 10.6 the mapping π defines a digraph homomorphism
−−→
−−→
from Γ(EM ) to Γ(BM ). If Γ(EM ) contained an infinite directed path then by Lemma 10.5
the image of this path would be an infinite directed path in Γ(BM ) which is impossible since
50
ROBERT D. GRAY AND BENJAMIN STEINBERG
−−→
(E, R, C, c, i) is a collapsing scheme. Therefore Γ(EM ) contains no infinite path and thus
(E, R, C, κ, ι) satisfies condition (C2).
This completes the proof of Proposition 10.4.
−−→
Remark 10.7. It follows from Proposition 10.4 that every down arc in Γ(EM ) (that is, every
arc of type (DE1)) is of the form τ → dj (τ ) where τ ∈ Cn+1 , dj (τ ) ∈ Rn , and κ(dj (τ )) 6= τ .
10.1.2. Proving (E, R, C, κ, ι) is a left M -equivariant collapsing scheme. To prove that (E, R, C, κ, ι)
−−→
is a left M -equivariant collapsing scheme for EM we need to verify that conditions (A1)–(A6)
−−→
hold. In Section 5 we proved that EM is a rigid free left M -simplicial set. In addition to
−−→
this, from the definitions an n-cell m(m1 , m2 , ..., mn ) of EM essential (redundant, collapsible
respectively) if and only if (m1 , m2 , ..., mn ) is essential (redundant, collapsible respectively) in
the collapsing scheme (E, R, C, c, i) of BM . These facts prove that (A1) and (A2) both hold:
−−→
(A1) The action of M on EM maps n-simplicies to n-simplicies, and commutes with di and
−−→
si , that is, M is acting by simplicial morphisms on EM .
(A2) For every n-simplex σ and m ∈ M , σ is a cell (i.e. is a non-degenerate simplex) if and
only if mσ is a cell, in which case σ ∈ E (respectively R, C) if and only if mσ ∈ E
(respectively R, C).
The next axiom we need to verify is:
(A3) If (σ, τ ) ∈ Rn × Cn+1 is a matched redundant-collapsible pair (i.e. τ = c(σ)) then so is
the pair m(σ, τ ) = (mσ, mτ ) ∈ Rn × Cn+1 .
−−→
We shall prove a little more than this. We consider the bipartite digraph Γ(EM ) between levels
n and n + 1 defined above. We want to prove that M acts on this digraph, that is, that arcs
are sent to arcs under the action. We note that the action of M on the vertices preserves the
bipartition. As above, there are two types of directed arcs that need to be considered, the
up-arcs and the down-arcs.
First consider the down-arcs. My Remark 10.7, a typical down-arc has the form
dj
mσ −→ dj (mσ)
where κ(dj (mσ)) 6= mσ. Let n ∈ M be arbitrary. Then we have
dj
nmσ −→ dj (nmσ) = ndj (mσ),
by definition of dj . This is a down-arc because if κ(dj (nmσ)) = nmσ, then from nmdj (σ) =
dj (nmσ) we deduce that nmc(dj (σ)) = nmσ and so c(dj (σ)) = σ, whence κ(dj (mσ)) =
κ(mdj (σ)) = mc(dj (σ)) = mσ, which is a contradiction.
Now consider up-arcs. A typical up-arc has the form mσ → κ(mσ). Let n ∈ M . Then
nmσ → κ(nmσ) = nm(c(σ)) = nκ(mσ),
which is an up-arc as required.
−−→
−−→
This covers all types of arcs in Γ(EM ) and we conclude that M acts on Γ(EM ) by digraph
endomorphisms. This fact together with (A2) then implies property (A3), since the up-arcs in
this bipartite graph are precisely the matched redundant-collapsible pairs. Next consider
TOPOLOGICAL FINITENESS PROPERTIES
51
(A4) There is a subset B ⊆ E ∪ R ∪ C such that for all n the set of n-cells is a free left M -set
with basis Bn (the n-cells in B). Let EB = E ∩ B, RB = R ∩ B and CB = C ∩ B. Then
En is a free left M -set with basis EB
n , and similarly for Rn and Cn .
−−→
We saw in Section 5 that the set of n-cells of EM is a free left M -set with basis the set of
n-cells
B = {(m1 , . . . , mn ) : mi 6= 1 for all i}
of BM . The last clause of (A4) then follows from (A2). Now we shall prove
(A5) For every matched pair (σ, τ ) ∈ R × C, σ ∈ RB if and only if τ ∈ CB . In particular, for
every matched pair (σ, τ ) there is a unique pair (σ ′ , τ ′ ) ∈ RB × CB and m ∈ M such
that (σ, τ ) = m(σ ′ , τ ′ ).
−−→
The matched pairs are the up-arcs in the graph Γ(EM ). A typical up-arc has the form
m(m1 , . . . , mn ) → κ(m(m1 , . . . , mn ) = mc(m1 , . . . , mn ).
So this pair is
m · ((m1 , . . . , mn ) → c(m1 , . . . , mn ))
where
(m1 , . . . , mn ) → c(m1 , . . . , mn )
is a uniquely determined matched basis pair. Also, if (σ, τ ) is a matched pair then σ =
(m1 , . . . , mn ) belongs to the basis if and only if κ(m1 , . . . , mn ) = c(m1 , . . . , mn ) belongs to
the basis, completing the proof of (A5). Finally we turn our attention to showing axiom
(A6) For every redundant cell τ and every m ∈ M
height(τ ) = height(mτ )
where height is taken with respect to the collapsing scheme (E, R, C, κ, ι).
The following lemma will be useful.
−−→
−−→
Lemma 10.8 (Path lifting property). Define a mapping π : V Γ(EM )∪EΓ(EM ) → V Γ(BM )∪
EΓ(BM ) defined on vertices by:
m(m1 , . . . , mk ) 7→ (m1 , . . . , mk ) (k ∈ {n, n + 1})
and defined on edges (DE1) and (DE2) by π(x → y) = π(x) → π(y). Let µ be a redundant
−−→
n-cell in EM . Then for each path p in Γ(BM ), with initial vertex π(µ) there is a path p̂ in
−−→
Γ(EM ), with initial vertex µ, such that π(p̂) = p.
Proof. We shall establish two claims, from which the lemma will be obvious by induction on
−−→
path length. First we claim that if y = mτ is a redundant n-cell of EM (with m ∈ M and τ and
n-cell of BM ) and σ ∈ V Γ(BM ) is a vertex such that there is a directed edge τ = π(y) → σ,
−−→
−−→
then there is a vertex z ∈ V Γ(EM ) such that y → z is a directed edge in EΓ(EM ), and
−−→
π(y → z) = π(y) → σ. Indeed, set z = κ(y). Then y → z is a directed edge in EΓ(EM ) by
definition and π(z) = π(κ(y)) = π(κ(mτ )) = π(m(c(τ ))) = c(τ ) = σ.
−−→
Next we claim that if x → y is an up-arc of EΓ(EM ) as in (DE2) and σ is a vertex in V Γ(BM )
−−→
such that π(y) → σ is a directed edge in EΓ(BM ), then there exists a vertex z ∈ V Γ(EM ) such
52
ROBERT D. GRAY AND BENJAMIN STEINBERG
−−→
that y → z is a directed edge in EΓ(EM ), and π(y → z) = π(y) → σ. Write x = mτ where
m ∈ M and τ is a redundant n-cell of BM . Then x → y is equal to mτ −→ κ(mτ ) = mc(τ ). In
Γ(BM ) we have the path π(x) → π(y) → σ which equals τ → c(τ ) → σ. Therefore c(τ ) → σ is
an arc in Γ(BM ) of type 1. Therefore, by Remark 10.7 applied to the graph Γ(BM ), it follows
that σ is a redundant n-cell with σ = dj (c(τ )) for some j ∈ {0, . . . , n + 1} and σ 6= τ (using that
c is a bijection). Set z = dj (y). We need to show that π(z) = σ and that y → z is a directed
−−→
edge in Γ(EM ).
For the first claim, since π is a simplicial morphism we have
π(z) = π(dj (y)) = dj (π(y)) = dj (c(τ )) = σ.
−−→
For the second claim, to verify that y → z is a directed edge in Γ(EM ) we just need to
show that z is a redundant cell and y 6= κ(z). From the definitions it follows that under the
mapping π any vertex in the preimage of a redundant cell is a redundant cell. Thus, since σ is
redundant and π(z) = σ it follows that z is redundant. Finally, if y = κ(z), then z = x because
κ is injective. Therefore, σ = π(z) = π(x) = τ , which is a contradiction.
We can now construct the path p̂ by induction on the length of p, where the inductive step
uses the first claim if the lift of the previous portion ends at a redundant cell and uses the
second claim if it ends at a collapsible cell.
Axiom (A6) is then easily seen to be a consequence of the following lemma.
Lemma 10.9. Let τ be a redundant cell in the collapsing scheme (E, R, C, κ, ι). Write τ = mσ
−→ (mσ) denote the height of mσ with respect to
where σ is a redundant cell in BM . Let height−
EM
the collapsing scheme (E, R, C, κ, ι), and let heightBM (σ) denote the height of σ with respect to
−→ (mσ) = heightBM (σ).
the collapsing scheme (E, R, C, c, i). Then height−
EM
Proof. Let
mσ = τ = τ0 ≻ τ1 ≻ · · · ≻ τk
be a descending chain of redundant n-cells from R. It follows that there is a directed path
p = τ0 → κ(τ0 ) → · · · → κ(τk−1 ) → τk
−−→
in Γ(EM ). Since π is a digraph homomorphism it follows that π(p) is a directed path in Γ(BM )
and hence
σ = π(τ0 ) ≻ π(τ1 ) ≻ · · · ≻ π(τk )
−→ (mσ) ≤ heightBM (σ).
is a descending chain of redundant n-cells in R. This proves that height−
EM
For the converse, let
σ = σ0 ≻ σ1 ≻ · · · ≻ σk
be a descending chain of redundant n-cells from R. Then there is a directed path
q = σ0 → c(σ0 ) → · · · → c(σk−1 ) → σk
−−→
in Γ(BM ). By Lemma 10.8 we can lift q to a path q̂ in Γ(EM ) with initial vertex mσ and such
that π(q̂) = q. But then the redundant cells in the path q̂ form a decending chain of length k
−→ (mσ) ≥ heightBM (σ).
starting at mσ, proving that height−
EM
This completes the proof of Theorem 10.2(a) and its dual Theorem 10.2(b).
TOPOLOGICAL FINITENESS PROPERTIES
53
10.2. Proof of Theorem 10.2(c). We shall explain how the above proof of Theorem 10.2(c)
is modified to prove the two-sided analogue Theorem 10.2(c).
←−→
Let (E, C, R, c, i) be a guarded collapsing scheme. Define an n-cell m(m1 , m2 , ..., mn )u of EM
to be essential (respectively redundant, collapsible respectively) if and only if (m1 , m2 , ..., mn ) is
essential (respectively redundant, collapsible respectively) in the collapsing scheme (E, R, C, c, i).
This defines (E, R, C).
←−→
For the mappings κ and ι, given mτ u ∈ EM where τ = (m1 , ..., mn ) is an n-cell of BM we
define
ι(mτ u) = i(τ )
κ(mτ u) = m(c(τ ))u.
(10.3)
(10.4)
With these definitions we claim that (E, R, C, κ, ι) is an M -equivariant collapsing scheme for
←−→
the free bi-M -simplicial set EM such that for each n the set of essential n-cells En is a free
bi-M -set of rank |En |.
10.2.1. Proving that (E, R, C, κ, ι) is a collapsing scheme.
Proposition 10.10. With the above definitions, (E, R, C, κ, ι) is a collapsing scheme for the
←−→
simplicial set EM .
The proof is analogous to the proof of Proposition 10.4. As in the proof of that proposition,
←−→
←−→
we have a digraph homomorphism π : V Γ(EM ) ∪ EΓ(EM ) → V Γ(BM ) → EΓ(BM ) given by
m(m1 , . . . , mk )n → (m1 , . . . , mk ),
π(x → y) = π(x) → π(y).
To verify collapsing scheme condition (C1) it suffices to observe that for every redundant n-cell
mτ we have
dι(mτ n) (κ(mτ n)) = di(τ ) (mc(τ )n) (by definition)
= mdi(τ ) (c(τ ))n (since i(τ ) 6= 0 and i(τ ) 6= n + 1)
= mτ n.
Here the second line appeals to the fact that the original collapsing scheme is guarded. Collapsing scheme condition (C2) holds again by applying Lemma 10.5 and the fact that π is a
digraph homomorphism.
10.2.2. Proving (E, R, C, κ, ι) is an M -equivariant collapsing scheme. We have already seen in
←−→
Section 5 that EM is a bi-M -simplicial set. We need to show that (E, R, C, κ, ι) is an M ←−→
equivariant collapsing scheme for EM . By definition, for this we need to verify that axioms
←−→
(A1)-(A6) are satisfied for EM as a left M × M op -simplicial set.
←−→
In Section 5 we saw that EM is a rigid free left M × M op -simplicial set. Together with the
←−→
definition of the action of M × M op on EM and the definition of E, R and C, axioms (A1) and
(A2) both then follow. Axioms (A3)-(A6) are then proved exactly as above just with M × M op
in place of M in the proof. This then completes the proof of Theorem 10.2(c).
11. Monoids admitting guarded collapsing schemes
In this section we give examples of classes of monoids to which the above theory of equivariant
collapsing schemes applies. In particular, this will allow us to use the theory developed above
54
ROBERT D. GRAY AND BENJAMIN STEINBERG
to give a topological proof of the fact that monoids which admit finite complete presentations
are of type bi-F∞ .
Let M be a monoid defined by a finite presentation hA | Ri with generators A and defining
∗ is the smallest congruence
∗ where ↔
relations R ⊆ A∗ × A∗ . Thus, M is isomorphic to A∗ / ↔
R
R
∗
on A containing R. We view hA | Ri as a string rewriting system, writing l → r for the pair
(l, r) ∈ R. We define a binary relation → on A∗ , called a single-step reduction, in the following
way:
u → v ⇔ u ≡ w1 lw2 and v ≡ w1 rw2
for some (l, r) ∈ R and w1 , w2 ∈ X ∗ . A word is called irreducible if no single-step reduction
∗ .
rule may be applied to it. The transitive and reflexive closure of →R is denoted by −→
R
This rewriting system is called noethereian if there are no infinite descending chains
w1 →R w2 →R w3 →R · · · →R wn →R · · · .
∗ v ′ there is a word w ∈ X ∗ such
∗ v and u −→
It is called confluent if whenever we have u −→
R
R
∗ w. If R is simultaneously noetherian and confluent we say that R is
∗ w and v ′ −→
that v −→
R
R
complete. The presentation hA | Ri is called complete if the rewriting system R is complete.
It is well-known (see for example [BO93]) that if hA | Ri is a finite complete presentation then
∗ -class of A∗ contains exactly one irreducible element. So the set of irreducible elements
each ↔
R
give a set of normal forms for the elements of the monoid M . In particular, if a monoid admits
a presentation by a finite compete rewriting system then the word problem for the monoid is
decidable.
In [Bro92, page 145] a method is given for constructing a collapsing scheme on BM for any
monoid M that is given by a finite complete rewriting system. It is easily observed from [Bro92,
page 145] that in the given collapsing scheme (E, R, C, c, i) the function i never takes either of its
two extreme allowable values, that is, the collapsing scheme for BM given in [Bro92] is guarded
in the sense of Definition 10.1. Also, as Brown observes (see [Bro92, page 147]), it follows easily
from his definition that there are only finitely many essential cells in each dimension. Thus we
have:
Proposition 11.1. Let M be a monoid. If M admits a presentation by a finite complete
rewriting system then BM admits a guarded collapsing scheme of finite type.
It follows that the theory of M -equivariant collapsing schemes developed in the previous
section applies to monoids with complete presentations, giving:
Corollary 11.2. Let M be a monoid that admits a presentation by a finite complete rewriting
system. Then M is of type left-F∞ , right-F∞ and bi-F∞ .
Proof. By Proposition 11.1 the simplicial set BM admits a guarded collapsing scheme of finite
type. Then the result follows from Corollary 10.3.
We obtain the following result of Kobayashi as a special case.
Corollary 11.3 ([Kob05]). Let M be a monoid that admits a presentation by a finite complete
rewriting system. Then M is of type bi-FP∞ .
Proof. Follows from Proposition 7.7 and Corollary 11.2.
More recently the class of, so-called, factorable monoids was introduced in work of Hess
and Ozornova [HO14]. Since it is quite technical we shall not give the definition of factorable
monoid here, we refer the reader to [HO14] to the definition, and we shall use the same notation
as there. In their work they show that a number of interesting classes of monoids admit
TOPOLOGICAL FINITENESS PROPERTIES
55
factorability structures. In some cases (e.g. Garside groups) the rewriting systems associated
with factorability structures are finite and complete, and so in these cases the monoids are
bi-F∞ . On the other hand, in [HO14, Appendix] they give an example of a factorable monoid
where the associated rewriting system is not noetherian and thus not complete (although it is
not discussed there whether the monoid admits a presentation by some other finite complete
presentation). So, there are some examples where factorability structures may be seen to exist,
even when the given presentation is not complete. In [HO14, Section 9] the authors construct
a matching on the reduced, inhomogeneous bar complex of a factorable monoid. As they say
in their paper, the matching they construct is very similar to the construction used by Brown
giving a collapsing scheme for monoids defined by finite complete rewriting systems [Bro92, page
145]. Details of the matching they construct for factorable monoids may be found on pages 27
and 28 of [HO14]. It is immediate from the definition of the mapping µ on page 28 that their
construction defines a guarded collapsing scheme for the simplicial set BM where M is a any
factorable monoid (M, E, η). If the generating set E for the monoid is finite, then the number of
essential cells in each dimension is finite, and so we have a guarded collapsing scheme of finite
type, giving:
Corollary 11.4. Let M be a monoid. If M admits a factorability structure (M, E, η) with finite
generating set E then BM admits a guarded collapsing scheme of finite type. In particular M
is of type left-F∞ , right-F∞ and bi-F∞ .
References
[AH03]
[Alo94]
[Ani86]
[AR67]
[BB97]
[BBG98]
[BG84]
[BH01]
[BHMS02]
[Bie76]
[BO93]
[Bra99]
[Bro87]
[Bro92]
J. M. Alonso & S. M. Hermiller. ‘Homological finite derivation type’. Internat. J. Algebra Comput.,
13, no. 3 (2003), pp. 341–359. doi: 10.1142/S0218196703001407.
J. M. Alonso. ‘Finiteness conditions on groups and quasi-isometries’. J. Pure Appl. Algebra, 95, no. 2
(1994), pp. 121–129. doi: 10.1016/0022-4049(94)90069-8.
D. J. Anick. ‘On the homology of associative algebras’. Trans. Amer. Math. Soc., 296, no. 2 (1986),
pp. 641–659. doi: 10.2307/2000383.
W. W. Adams & M. A. Rieffel. ‘Adjoint functors and derived functors with an application to the
cohomology of semigroups’. J. Algebra, 7 (1967), pp. 25–34. doi: 10.1016/0021-8693(67)90065-8.
M. Bestvina & N. Brady. ‘Morse theory and finiteness properties of groups’. Invent. Math., 129,
no. 3 (1997), pp. 445–470. doi: 10.1007/s002220050168.
G. Baumslag, M. R. Bridson, & K. W. Gruenberg. ‘On the absence of cohomological finiteness in
wreath products’. J. Austral. Math. Soc. Ser. A, 64, no. 2 (1998), pp. 222–230.
K. S. Brown & R. Geoghegan. ‘An infinite-dimensional torsion-free FP∞ group’. Invent. Math., 77,
no. 2 (1984), pp. 367–381. doi: 10.1007/BF01388451.
R. Bieri & J. Harlander. ‘On the FP3 -conjecture for metabelian groups’. J. London Math. Soc. (2),
64, no. 3 (2001), pp. 595–610.
M. R. Bridson, J. Howie, C. F. Miller, III, & H. Short. ‘The subgroups of direct products of surface
groups’. Geom. Dedicata, 92 (2002), pp. 95–103. Dedicated to John Stallings on the occasion of his
65th birthday. doi: 10.1023/A:1019611419598.
R. Bieri. Homological dimension of discrete groups. Mathematics Department, Queen Mary College,
London, 1976. Queen Mary College Mathematics Notes.
R. V. Book & F. Otto. String-rewriting systems. Texts and Monographs in Computer Science.
Springer-Verlag, New York, 1993.
N. Brady. ‘Branched coverings of cubical complexes and subgroups of hyperbolic groups’. J. London
Math. Soc. (2), 60, no. 2 (1999), pp. 461–480.
K. S. Brown. ‘Finiteness properties of groups’. In Proceedings of the Northwestern conference on cohomology of groups (Evanston, Ill., 1985), vol. 44, pp. 45–75, 1987. doi:
10.1016/0022-4049(87)90015-6.
K. S. Brown. ‘The geometry of rewriting systems: a proof of the Anick-Groves-Squier theorem’. In
Algorithms and classification in combinatorial group theory (Berkeley, CA, 1989), vol. 23 of Math.
Sci. Res. Inst. Publ., pp. 137–163. Springer, New York, 1992. doi: 10.1007/978-1-4613-9730-4-6.
56
[Bro94]
ROBERT D. GRAY AND BENJAMIN STEINBERG
K. S. Brown. Cohomology of groups, vol. 87 of Graduate Texts in Mathematics. Springer-Verlag,
New York, 1994. Corrected reprint of the 1982 original.
[Bro10]
K. S. Brown. ‘Lectures on the cohomology of groups’. In Cohomology of groups and algebraic Ktheory, vol. 12 of Adv. Lect. Math. (ALM), pp. 131–166. Int. Press, Somerville, MA, 2010.
[BW07]
K.-U. Bux & K. Wortman. ‘Finiteness properties of arithmetic groups over function fields’. Invent.
Math., 167, no. 2 (2007), pp. 355–378.
[Cha00]
M. K. Chari. ‘On discrete Morse functions and combinatorial decompositions’. Discrete Math., 217,
no. 1-3 (2000), pp. 101–113. Formal power series and algebraic combinatorics (Vienna, 1997). doi:
10.1016/S0012-365X(99)00258-7.
[Che84]
C. C.-a. Cheng. ‘Separable semigroup algebras’. J. Pure Appl. Algebra, 33, no. 2 (1984), pp. 151–158.
doi: 10.1016/0022-4049(84)90003-3.
[Cho06]
F. Chouraqui. ‘Rewriting systems in alternating knot groups’. Internat. J. Algebra Comput., 16,
no. 4 (2006), pp. 749–769.
[CO98]
R. Cremanns & F. Otto. ‘FP∞ is undecidable for finitely presented monoids with word problems
decidable in polynomial time’. Mathematische Schriften Kassel 11/98, Universitat Kassel, September
(1998) .
[Coh73]
M. M. Cohen. A course in simple-homotopy theory. Springer-Verlag, New York, 1973. Graduate
Texts in Mathematics, Vol. 10.
[Coh92]
D. E. Cohen. ‘A monoid which is right F P∞ but not left F P1 ’. Bull. London Math. Soc., 24, no. 4
(1992), pp. 340–342.
[Coh97]
D. E. Cohen. ‘String rewriting and homology of monoids’. Math. Structures Comput. Sci., 7, no. 3
(1997), pp. 207–240.
[CS80]
C. C.-a. Cheng & J. Shapiro. ‘Cohomological dimension of an abelian monoid’. Proc. Amer. Math.
Soc., 80, no. 4 (1980), pp. 547–551. doi: 10.2307/2043420.
[CS09a]
J. Cassaigne & P. V. Silva. ‘Infinite periodic points of endomorphisms over special confluent rewriting
systems’. Ann. Inst. Fourier, 59, no. 2 (2009), pp. 769–810.
[CS09b]
J. Cassaigne & P. V. Silva. ‘Infinite words and confluent rewriting systems: endomorphism extensions’. Internat. J. Algebra Comput., 19, no. 4 (2009), pp. 443–490.
[DDM09]
V. Diekert, A. Duncan, & A. Miasnikov. ‘Geodesic rewriting systems and pregroups’. To appear in Combinatorial and Geometric Group Theory, Dortmund and Carleton Conferences. Series:
Trends in Mathematics Bogopolski, O.; Bumagin, I.; Kharlampovich, O.; Ventura, E. (Eds.) 2009.,
arXiv:0906.2223v1, 2009.
[DFZ86]
E. Dror Farjoun & A. Zabrodsky. ‘Homotopy equivalence between diagrams of spaces’. J. Pure Appl.
Algebra, 41, no. 2-3 (1986), pp. 169–182. doi: 10.1016/0022-4049(86)90109-X.
[DL98]
J. F. Davis & W. Lück. ‘Spaces over a category and assembly maps in isomorphism conjectures in
K- and L-theory’. K-Theory, 15, no. 3 (1998), pp. 201–252. doi: 10.1023/A:1007784106877.
[EG57]
S. Eilenberg & T. Ganea. ‘On the Lusternik-Schnirelmann category of abstract groups’. Ann. of
Math. (2), 65 (1957), pp. 517–518. doi: 10.2307/1970062.
[Fie84]
Z. Fiedorowicz. ‘Classifying spaces of topological monoids and categories’. Amer. J. Math., 106,
no. 2 (1984), pp. 301–350. doi: 10.2307/2374307.
[FMWZ13] M. G. Fluch, M. Marschler, S. Witzel, & M. C. B. Zaremsky. ‘The Brin-Thompson groups sV are
of type F∞ ’. Pacific J. Math., 266, no. 2 (2013), pp. 283–295. doi: 10.2140/pjm.2013.266.283.
[For95]
R. Forman. ‘A discrete Morse theory for cell complexes’. In Geometry, topology, & physics, Conf.
Proc. Lecture Notes Geom. Topology, IV, pp. 112–125. Int. Press, Cambridge, MA, 1995.
[For02]
R. Forman. ‘A user’s guide to discrete Morse theory’. Séminaire Lotharingien de Combinatoire, B48c
(2002), pp. 1–35. url: http://www.emis.ams.org/journals/SLC/wpapers/s48forman.pdf.
[Fre09]
R. Freij. ‘Equivariant discrete Morse theory’. Discrete Math., 309, no. 12 (2009), pp. 3821–3829.
doi: 10.1016/j.disc.2008.10.029.
[FS05]
D. Farley & L. Sabalka. ‘Discrete Morse theory and graph braid groups’. Algebr. Geom. Topol., 5
(2005), pp. 1075–1109. doi: 10.2140/agt.2005.5.1075.
[Geo08]
R. Geoghegan. Topological methods in group theory, vol. 243 of Graduate Texts in Mathematics.
Springer, New York, 2008. doi: 10.1007/978-0-387-74614-2.
[GM12]
Y. Guiraud & P. Malbos. ‘Higher-dimensional normalisation strategies for acyclicity’. Adv. Math.,
231, no. 3-4 (2012), pp. 2294–2351. doi: 10.1016/j.aim.2012.05.010.
[GP96]
V. S. Guba & S. J. Pride. ‘Low-dimensional (co)homology of free Burnside monoids’. J. Pure Appl.
Algebra, 108, no. 1 (1996), pp. 61–79. doi: 10.1016/0022-4049(95)00038-0.
TOPOLOGICAL FINITENESS PROPERTIES
[GP98]
[GS]
[GS06]
[GS08]
[GZ67]
[HO14]
[Hoc45]
[How63]
[How95]
[HS99]
[Hur89]
[KM05]
[Kna72]
[KO01]
[Kob00]
[Kob05]
[Kob07]
[Kob10]
[Koz05]
[KT76]
[Lam99]
[LN01]
[LS06]
[May67]
[May72]
[May75]
[May96]
57
V. S. Guba & S. J. Pride. ‘On the left and right cohomological dimension of monoids’. Bull. London
Math. Soc., 30, no. 4 (1998), pp. 391–396. doi: 10.1112/S0024609398004676.
R. D. Gray & B. Steinberg. ‘Topological finiteness properties of monoids. Part 2: amalgamated free
products, HNN extensions, and special monoids’. in preparation.
V. S. Guba & M. V. Sapir. ‘Diagram groups and directed 2-complexes: homotopy and homology’.
J. Pure Appl. Algebra, 205, no. 1 (2006), pp. 1–47. doi: 10.1016/j.jpaa.2005.06.012.
O. Goodman & M. Shapiro. ‘On a generalization of Dehn’s algorithm’. Internat. J. Algebra Comput.,
18, no. 7 (2008), pp. 1137–1177.
P. Gabriel & M. Zisman. Calculus of fractions and homotopy theory. Ergebnisse der Mathematik
und ihrer Grenzgebiete, Band 35. Springer-Verlag New York, Inc., New York, 1967.
A. Hess & V. Ozornova. ‘Factorability, string rewriting and discrete morse theory’. arXiv:1412.3025,
2014.
G. Hochschild. ‘On the cohomology groups of an associative algebra’. Ann. of Math. (2), 46 (1945),
pp. 58–67. doi: 10.2307/1969145.
J. M. Howie. ‘Embedding theorems for semigroups’. Quart. J. Math. Oxford Ser. (2), 14 (1963), pp.
254–258. doi: 10.1093/qmath/14.1.254.
J. M. Howie. Fundamentals of semigroup theory. Academic Press [Harcourt Brace Jovanovich Publishers], London, 1995. L.M.S. Monographs, No. 7.
S. Hermiller & M. Shapiro. ‘Rewriting systems and geometric three-manifolds’. Geom. Dedicata, 76,
no. 2 (1999), pp. 211–228.
C. M. Hurwitz. ‘On the homotopy theory of monoids’. J. Austral. Math. Soc. Ser. A, 47, no. 2
(1989), pp. 171–185.
G. Kalai & R. Meshulam. ‘A topological colorful Helly theorem’. Adv. Math., 191, no. 2 (2005), pp.
305–311. doi: 10.1016/j.aim.2004.03.009.
U. Knauer. ‘Projectivity of acts and Morita equivalence of monoids’. Semigroup Forum, 3, no. 4
(1971/1972), pp. 359–370.
Y. Kobayashi & F. Otto. ‘On homotopical and homological finiteness conditions for finitely
presented monoids’. Internat. J. Algebra Comput., 11, no. 3 (2001), pp. 391–403. doi:
10.1142/S0218196701000577.
Y. Kobayashi. ‘Every one-relation monoid has finite derivation type’. In Proceedings of the Third
Symposium on Algebra, Languages and Computation (Osaka, 1999), pp. 16–20. Shimane Univ.,
Matsue, 2000.
Y. Kobayashi. ‘Gröbner bases of associative algebras and the Hochschild cohomology’. Trans. Amer.
Math. Soc., 357, no. 3 (2005), pp. 1095–1124 (electronic). doi: 10.1090/S0002-9947-04-03556-1.
Y. Kobayashi. ‘The homological finiteness property FP1 and finite generation of monoids’. Internat.
J. Algebra Comput., 17, no. 3 (2007), pp. 593–605. doi: 10.1142/S0218196707003743.
Y. Kobayashi. ‘The homological finiteness properties left-, right-, and bi-FPn of monoids’. Comm.
Algebra, 38, no. 11 (2010), pp. 3975–3986. doi: 10.1080/00927872.2010.507562.
D. N. Kozlov. ‘Discrete Morse theory for free chain complexes’. C. R. Math. Acad. Sci. Paris, 340,
no. 12 (2005), pp. 867–872. doi: 10.1016/j.crma.2005.04.036.
D. M. Kan & W. P. Thurston. ‘Every connected space has the homology of a K(π, 1)’. Topology,
15, no. 3 (1976), pp. 253–258.
T. Y. Lam. Lectures on modules and rings, vol. 189 of Graduate Texts in Mathematics. SpringerVerlag, New York, 1999. doi: 10.1007/978-1-4612-0525-8.
I. J. Leary & B. E. A. Nucinkis. ‘Every CW-complex is a classifying space for proper bundles’.
Topology, 40, no. 3 (2001), pp. 539–550. doi: 10.1016/S0040-9383(99)00073-7.
I. J. Leary & M. Saadetoğlu. ‘Some groups of finite homological type’. Geom. Dedicata, 119 (2006),
pp. 113–120.
J. P. May. Simplicial objects in algebraic topology. Van Nostrand Mathematical Studies, No. 11. D.
Van Nostrand Co., Inc., Princeton, N.J.-Toronto, Ont.-London, 1967.
J. P. May. The geometry of iterated loop spaces. Springer-Verlag, Berlin-New York, 1972. Lectures
Notes in Mathematics, Vol. 271.
J. P. May. ‘Classifying spaces and fibrations’. Mem. Amer. Math. Soc., 1, no. 1, 155 (1975), pp.
xiii+98.
J. P. May. Equivariant homotopy and cohomology theory, vol. 91 of CBMS Regional Conference Series
in Mathematics. Published for the Conference Board of the Mathematical Sciences, Washington, DC;
by the American Mathematical Society, Providence, RI, 1996. With contributions by M. Cole, G.
58
[May99]
[McC69]
[McD79]
[Mil57]
[Mit72]
[MS76]
[MSS15a]
[MSS15b]
[Nic69]
[Nic72]
[Nie]
[Nov98]
[Nun95]
[OK97]
[Pas08]
[PO04]
[PO05]
[Pri06]
[Pup58]
[Pup59]
[RT89]
[Sch75]
[Seg78]
[Ser71]
[SOK94]
ROBERT D. GRAY AND BENJAMIN STEINBERG
Comezaña, S. Costenoble, A. D. Elmendorf, J. P. C. Greenlees, L. G. Lewis, Jr., R. J. Piacenza, G.
Triantafillou, and S. Waner. doi: 10.1090/cbms/091.
J. P. May. A concise course in algebraic topology. Chicago Lectures in Mathematics. University of
Chicago Press, Chicago, IL, 1999.
M. C. McCord. ‘Classifying spaces and infinite symmetric products’. Trans. Amer. Math. Soc., 146
(1969), pp. 273–298.
D. McDuff. ‘On the classifying spaces of discrete monoids’. Topology, 18, no. 4 (1979), pp. 313–320.
doi: 10.1016/0040-9383(79)90022-3.
J. Milnor. ‘The geometric realization of a semi-simplicial complex’. Ann. of Math. (2), 65 (1957),
pp. 357–362.
B. Mitchell. ‘Rings with several objects’. Advances in Math., 8 (1972), pp. 1–161. doi:
10.1016/0001-8708(72)90002-3.
D. McDuff & G. Segal. ‘Homology fibrations and the “group-completion” theorem’. Invent. Math.,
31, no. 3 (1975/76), pp. 279–284.
S. Margolis, F. Saliola, & B. Steinberg. ‘Cell complexes, poset topology and the representation
theory of algebras arising in algebraic combinatorics and discrete geometry’. ArXiv e-prints, 2015.
arXiv: 1508.05446.
S. Margolis, F. Saliola, & B. Steinberg. ‘Combinatorial topology and the global dimension of algebras arising in combinatorics’. J. Eur. Math. Soc. (JEMS), 17, no. 12 (2015), pp. 3037–3080. doi:
10.4171/JEMS/579.
W. R. Nico. ‘On the cohomology of finite semigroups’. J. Algebra, 11 (1969), pp. 598–612. doi:
10.1016/0021-8693(69)90093-3.
W. R. Nico. ‘A counterexample in the cohomology of monoids’. Semigroup Forum, 4 (1972), pp.
93–94. doi: 10.1007/BF02570775.
P. Nielsen. ‘Can the trivial module be stably free for a monoid ring?’. MathOverflow.
URL:https://mathoverflow.net/q/248371
(version:
2016-08-26).
url:
https://mathoverflow.net/q/248371.
B. V. Novikov. ‘Semigroups of cohomological dimension one’. J. Algebra, 204, no. 2 (1998), pp.
386–393. doi: 10.1006/jabr.1997.7363.
M. Nunes. ‘Cohomological results in monoid and category theory via classifying spaces’. J. Pure
Appl. Algebra, 101, no. 2 (1995), pp. 213–244. doi: 10.1016/0022-4049(94)00056-O.
F. Otto & Y. Kobayashi. ‘Properties of monoids that are presented by finite convergent stringrewriting systems—a survey’. In Advances in algorithms, languages, and complexity, pp. 225–266.
Kluwer Acad. Publ., Dordrecht, 1997.
E. Pasku. ‘On some homotopical and homological properties of monoid presentations’. Semigroup
Forum, 76, no. 3 (2008), pp. 427–468. doi: 10.1007/s00233-007-9037-1.
S. J. Pride & F. Otto. ‘For rewriting systems the topological finiteness conditions FDT and
FHT are not equivalent’. J. London Math. Soc. (2), 69, no. 2 (2004), pp. 363–382. doi:
10.1112/S0024610703004903.
S. J. Pride & F. Otto. ‘On higher order homological finiteness of rewriting systems’. J. Pure Appl.
Algebra, 200, no. 1-2 (2005), pp. 149–161.
S. J. Pride. ‘Homological finiteness conditions for groups, monoids, and algebras’. Comm. Algebra,
34, no. 10 (2006), pp. 3525–3536. doi: 10.1080/00927870600796110.
D. Puppe. ‘Homotopie und Homologie in abelschen Gruppen- und Monoidkomplexen. I. II’. Math.
Z., 68 (1958), pp. 367–406, 407–421.
D. Puppe. ‘A theorem on semi-simplicial monoid complexes’. Ann. of Math. (2), 70 (1959), pp.
379–394.
J. Rhodes & B. Tilson. ‘The kernel of monoid morphisms’. J. Pure Appl. Algebra, 62, no. 3 (1989),
pp. 227–268. doi: 10.1016/0022-4049(89)90137-0.
B. M. Schein. ‘Free inverse semigroups are not finitely presentable’. Acta Math. Acad. Sci. Hungar.,
26 (1975), pp. 41–52. doi: 10.1007/BF01895947.
G. Segal. ‘Classifying spaces related to foliations’. Topology, 17, no. 4 (1978), pp. 367–382. doi:
10.1016/0040-9383(78)90004-6.
J.-P. Serre. ‘Cohomologie des groupes discrets’. (1971), pp. 77–169. Ann. of Math. Studies, No. 70.
C. C. Squier, F. Otto, & Y. Kobayashi. ‘A finiteness condition for rewriting systems’. Theoret.
Comput. Sci., 131, no. 2 (1994), pp. 271–294. doi: 10.1016/0304-3975(94)90175-9.
TOPOLOGICAL FINITENESS PROPERTIES
[Squ87]
[Sri96]
[Sta68]
[Ste92]
[Swa69]
[Wal65]
[Wei94]
[Wei13]
59
C. C. Squier. ‘Word problems and a homological finiteness condition for monoids’. J. Pure Appl.
Algebra, 49, no. 1-2 (1987), pp. 201–217. doi: 10.1016/0022-4049(87)90129-0.
V. Srinivas. Algebraic K-theory, vol. 90 of Progress in Mathematics. Birkhäuser Boston, Inc., Boston,
MA, second edition, 1996. doi: 10.1007/978-0-8176-4739-1.
J. R. Stallings. ‘On torsion-free groups with infinitely many ends’. Ann. of Math. (2), 88 (1968), pp.
312–334. doi: 10.2307/1970577.
M. Stein. ‘Groups of piecewise linear homeomorphisms’. Trans. Amer. Math. Soc., 332, no. 2 (1992),
pp. 477–514. doi: 10.2307/2154179.
R. G. Swan. ‘Groups of cohomological dimension one’. J. Algebra, 12 (1969), pp. 585–610. doi:
10.1016/0021-8693(69)90030-1.
C. T. C. Wall. ‘Finiteness conditions for CW-complexes’. Ann. of Math. (2), 81 (1965), pp. 56–69.
C. A. Weibel. An introduction to homological algebra, vol. 38 of Cambridge Studies in Advanced
Mathematics. Cambridge University Press, Cambridge, 1994. doi: 10.1017/CBO9781139644136.
C. A. Weibel. The K-book, vol. 145 of Graduate Studies in Mathematics. American Mathematical
Society, Providence, RI, 2013. An introduction to algebraic K-theory.
School of Mathematics,, University of East Anglia,, Norwich NR4 7TJ,, England and Department of Mathematics,, City College of New York,, Convent Avenue at 138th Street,, New York,
New York 10031,, USA
E-mail address: [email protected] and [email protected]
| 4math.GR
|
Distilling Abstract Machines (Long Version)
Beniamino Accattoli
Pablo Barenbaum
Damiano Mazza
arXiv:1406.2370v1 [cs.PL] 9 Jun 2014
January 16, 2018
Abstract
It is well-known that many environment-based abstract machines can
be seen as strategies in lambda calculi with explicit substitutions (ES).
Recently, graphical syntaxes and linear logic led to the linear substitution
calculus (LSC), a new approach to ES that is halfway between big-step
calculi and traditional calculi with ES. This paper studies the relationship between the LSC and environment-based abstract machines. While
traditional calculi with ES simulate abstract machines, the LSC rather
distills them: some transitions are simulated while others vanish, as they
map to a notion of structural congruence. The distillation process unveils
that abstract machines in fact implement weak linear head reduction, a
notion of evaluation having a central role in the theory of linear logic. We
show that such a pattern applies uniformly in call-by-name, call-by-value,
and call-by-need, catching many machines in the literature. We start by
distilling the KAM, the CEK, and the ZINC, and then provide simplified versions of the SECD, the lazy KAM, and Sestoft’s machine. Along
the way we also introduce some new machines with global environments.
Moreover, we show that distillation preserves the time complexity of the
executions, i.e. the LSC is a complexity-preserving abstraction of abstract
machines.
1
Introduction
In the theory of higher-order programming languages, abstract machines and
explicit substitutions are two tools used to model the execution of programs
on real machines while omitting many details of the actual implementation.
Abstract machines can usually be seen as evaluation strategies in calculi of
explicit substitutions (see at least [HM98, BD07a, Lan07, Cré07]), that can in
turn be interpreted as small-step cut-elimination strategies in sequent calculi
[ABS09].
Another tool providing a fine analysis of higher-order evaluation is linear
logic, especially via the new perspectives on cut-elimination provided by proof
nets, its graphical syntax. Explicit substitutions (ES) have been connected to
linear logic by Kesner and co-authors in a sequence of works [DCKP03, KL07,
KR09], culminating in the linear substitution calculus (LSC), a new formalism
with ES behaviorally isomorphic to proof nets (introduced in [AK10], developed
in [AP12, Acc12, ADL12, Acc13, ABKL14], and bearing similarities with calculi by De Bruijn [dB87], Nederpelt [Ned92], and Milner [Mil07]). Since linear
logic can model all evaluation schemes (call-by-name/value/need) [MOTW99],
1
the LSC can express them modularly, by minor variations on rewriting rules
and evaluation contexts. In this paper we revisit the relationship between
environment-based abstract machines and ES. Traditionally, calculi with ES
simulate machines. The LSC, instead, distills them.
A Bird’s Eye View. In a simulation, every machine transition is simulated
by some steps in the calculus with ES. In a distillation—that will be a formal
concept, not just an analogy—only some of the machine transitions are simulated, while the others are mapped to the structural equivalence of the calculus,
a specific trait of the LSC. Now, structural equivalence commutes with evaluation, i.e. it can be postponed. Thus, the transitions sent on the structural
congruence fade away, without compromising the result of evaluation. Additionally, we show that machine executions and their distilled representation in
the LSC have the same asymptotic length, i.e. the distillation process preserves
the complexity of evaluation. The main point is that the LSC is an arguably
simpler than abstract machines, and also—as we will show—it can uniformly
represent many different machines in the literature.
Traditional vs Contextual ES. Traditional calculi with ES (see [Kes09] for
a survey) implement β-reduction (lx.t)u →β t{x u} introducing an annotation
(the explicit substitution [x u]),
(lx.t)u →B
t[x u]
and percolating it through the term structure,
(tw)[x u] →@
(lx.t)[y u] →l
t[x u]w[x u]
lx.t[y u]
(1)
until they reach variable occurrences on which they finally substitute or get
garbage collected,
x[x u] →var u
y[x u] →≠ y
The LSC, instead, is based on a contextual view of evaluation and substitution, also known as at a distance. The idea is that one can get rid of the rules
percolating through the term structure—i.e. @ and l—by introducing contexts
C (i.e. terms with a hole ⟨⋅⟩) and generalizing the base cases, obtaining just two
rules, linear substitution (ls) and garbage collection (gc):
C⟨x⟩[x u] →ls
t[x u] →gc
C⟨u⟩[x u]
t
if x ∉ fv(t)
Dually, the rule creating substitutions (B) is generalized to act up to a context of substitutions [. . . . . .] ∶= [x1 w1 ] . . . [xk wk ] obtaining rule dB (B at a
distance):
(lx.t)[. . . . . .]u →dB t[x u][. . . . . .]
Logical Perspective on the LSC. From a sequent calculus point of view,
rules @ and l, corresponding to commutative cut-elimination cases, are removed
and integrated—via the use of contexts—directly in the definition of the principal cases B, var and ≠, obtaining the contextual rules dB, ls, and gc. This is
2
the analogous for terms of the removal of commutative cases provided by proof
nets. From a linear logic point of view, →dB can be identified with the multiplicative cut-elimination case →m , while →ls and →gc correspond to exponential
cut-elimination. Actually, garbage collection has a special status, as it can always be postponed. We will then identify exponential cut-elimination →e with
linear substitution →ls alone.
The LSC has a simple meta-theory, and is halfway between traditional calculi with ES—with whom it shares the small-step dynamics—and l-calculus—of
which it retains most of the simplicity.
Distilling Abstract Machines. Abstract machines implement the traditional approach to ES, by
1. Weak Evaluation: forbidding reduction under abstraction (no rule →l in
(1)),
2. Evaluation Strategy: looking for redexes according to some notion of weak
evaluation context E,
3. Context Representation: using environments e (aka lists of substitutions)
and stacks π (lists of terms) to keep track of the current evaluation context.
The LSC distills—i.e. factorizes—abstract machines. The idea is that one
can represent the strategy of an abstract machine by directly plugging the evaluation context in the contextual substitution/exponential rule, obtaining:
E
E⟨x⟩[x u] →e
E⟨u⟩[x u]
and factoring out the parts of the machine that just look for the next redex to
E
reduce. By defining ⊸ as the closure of →e and →m by evaluation contexts E,
one gets a clean representation of the machine strategy.
The mismatch between the two approaches is in rule →@ , that contextually—
by nature—cannot be captured. In order to get out of this cul-de-sac, the very
idea of simulation of an abstract machine must be refined to that of distillation.
The crucial observation is that the equivalence ≡ induced by →@ ∪ →gc
has the same special status of →gc , i.e. it can be postponed without affecting
reduction lengths. More abstractly, ≡ is a strong bisimulation with respect to
⊸, i.e. it verifies (note one step to one step, and viceversa)
t
≡
u
⇒ ∃q s.t.
t
≡
u
r
≡
q
⇒ ∃r s.t.
t
≡
u
r
≡
q
r
and
t
≡
u
q
Now, ≡ can be considered as a structural equivalence on the language. Indeed,
the strong bisimulation property states that the transformation expressed by
3
≡ is irrelevant with respect to ⊸, in particular ≡-equivalent terms have ⊸evaluations of the same length ending in ≡-equivalent terms (and this holds
even locally).
Abstract machines then are distilled : the logically relevant part of the substitution process is retained by ⊸ while both the search of the redex →@ and
garbage collection →≠ are isolated into the equivalence ≡. Essentially, ⊸ captures principal cases of cut-elimination while ≡ encapsulate the commutative
ones (plus garbage collection, corresponding to principal cut-elimination involving weakenings).
Case Studies. We will analyze along these lines many abstract machines.
Some are standard (KAM [Kri07], CEK [FF86], ZINC [Ler90]), some are new
(MAM, WAM), and of others we provide simpler versions (SECD [Lan64], Lazy
KAM [Cré07, DZ13], Sestoft’s [Ses97]). The previous explanation is a sketch
of the distillation of the KAM, but the approach applies mutatis mutandis to
all the other machines, encompassing most realizations of call-by-name, callby-value, and call-by need evaluation. The main contribution of the paper is
indeed a modular contextual theory of abstract machines. We start by distilling
some standard cases, and then rationally reconstruct and simplify non-trivial
machines as the SECD, the lazy KAM, and Sestoft’s abstract machine for callby-need (deemed SAM), by enlightening their mechanisms as different encoding
of evaluation contexts, modularly represented in the LSC.
Call-by-Need. Along the way, we show that the contextual (or at a distance)
approach of the LSC naturally leads to simple machines with just one global
environment, as the newly introduced MAM (M for Milner). Such a feature
is then showed to be a key ingredient of call-by-need machines, by using it to
introduce a new and simple call-by-need machine, the WAM (W for Wadsworth),
and then showing how to obtain (simplifications of) the Lazy KAM and the SAM
by simple tweaks.
Distillation Preserves Complexity. It is natural to wonder what is lost in
the distillation process. What is the asymptotic impact of distilling machine
executions into ⊸? Does it affect in any way the complexity of evaluation? We
will show that nothing is lost, as machine executions are only linearly longer
than ⊸. More precisely, they are bilinear, i.e. they are linear in 1) the length
of ⊸, and in 2) the size ∣t∣ of the starting term t. In other words, the search
of redexes and garbage collection can be safely ignored in quantitative (time)
analyses, i.e. the LSC and ⊸ provide a complexity-preserving abstraction of
abstract machines. While in call-by-name and call-by-value such an analysis
follows from an easy local property of machine executions, the call-by-need
case is subtler, as such a local property does not hold and bilinearity can be
established only via a global analysis.
Linear Logic and Weak Linear Head Reduction. Beyond the contextual
view, our work also unveils a deep connection between abstract machines and linear logic. The strategies modularly encoding the various machines (generically
noted ⊸ and parametric in a fixed notion of evaluation contexts) are in fact callby-name/value/need versions of weak linear head reduction (WLHR), a funda4
mental notion in the theory of linear logic [Acc13, MP94, DHR96, ER06, Cla11].
This insight —- due to Danos and Regnier for the KAM [DR04]—is not ours,
but we develop it in a simpler and tighter way, modularly lifting it to many
other abstract machines.
Call-by-Name. The call-by-name case (catching the KAM and the new
MAM) is in fact special, as our distillation theorem has three immediate corollaries, following from results about WLHR in the literature:
1. Invariance: it implies that the length of a KAM/MAM execution is an
invariant time cost model (i.e. polynomially related to, say, Turing machines, in both directions), given that in [ADL12] the same is shown for
WLHR.
2. Evaluation as Communication: we implicitly establish a link between the
KAM/MAM and the π-calculus, given that the evaluation of a term via
WLHR is isomorphic to evaluation via Milner’s encoding in the π-calculus
[Acc13].
3. Plotkin’s Approach: our study complements the recent [ABKL14], where
it is shown that WLHR is a standard strategy of the LSC. The two works
together provide the lifting to explicit substitutions of Plotkin’s approach
of relating a machine (the SECD machine in that case, the KAM/MAM
in ours) and a calculus (the call-by-value l-calculus and the LSC, respectively) via a standardization theorem and a standard strategy [Plo75].
Beyond Abstract Machines. This paper is just an episode—the one about
abstract machines—in the recent feuilleton about complexity analysis of functional languages via linear logic and rewriting theory, starring the linear substitution calculus. The story continues in [ADL14] and [ASC14]. In [ADL14], the
LSC is used to prove that the length of leftmost-outermost β-reduction is an
invariant cost-model for l-calculus (i.e. it is a measure polynomially related to
evaluation in classic computational models like Turing machines or random access machines), solving a long-standing open problem in the theory of l-calculus.
Instead, [ASC14] studies the asymptotic number of exponential steps (for ⊸)
in terms of the number of multiplicative steps, in the call-by-name/value/need
LSC. Via the results presented here, [ASC14] establishes a polynomial relationship between the exponential and the multiplicative transitions of abstract
machines, complementing our work and implying that distillation can be pushed
forward, forgetting exponential steps too.
Related Work. Beyond the already cited works, Danvy and coauthors have
studied abstract machines in a number of works. In some of them, they shows
how to extract a functional evaluator from an abstract machine via a sequence of
transformations (closure conversion, CPS, and defunctionalization) [ABDM03,
ADM04, Dan04]. Such a study is orthogonal in spirit to what we do here. The
only point of contact is the rational deconstruction of the SECD in [Dan04],
that is something that we also do, but in a different way. Another sequence
of works studies the relationship between abstract machines and calculi with
ES [BD07b, BD07a, DZ13], and it is clearly closer to our topic, except that 1)
5
[BD07b, BD07a] follow the traditional (rather than the contextual) approach to
ES; 2) none of these works deals with complexity analysis nor with linear logic.
On the other hand, [BD07a] provides a deeper analysis of Leroy’s ZINC machine,
as ours does not account for the avoidance of needless closure creations that is
a distinct feature of the ZINC. Last, what here we call commutative transitions
essentially corresponds to what Danvy and Nielsen call decompose phase in
[DN04].
The call-by-need calculus we use—that is a contextual re-formulation of
Maraist, Odersky, and Wadler’s calculus [MOW98]—is a novelty of this paper.
It is simpler than both Ariola and Felleisen’s [AF97] and Maraist, Odersky, and
Wadler’s calculi because it does not need any re-association axioms. Morally, it
is a version with let-bindings (avatars of ES) of Chang and Felleisen’s calculus
[CF12]. A similar calculus is used by Danvy and Zerny in [DZ13]. Another
call-by-need machine, with whom we do not deal with, appears in [GLS09].
Proofs. Some proofs have been omitted for lack of space. They can be found
in the longer version [ABM14].
2
Preliminaries on the Linear Substitution Calculus
Terms and Contexts. The language of the weak linear substitution calculus
(WLSC) is generated by the following grammar:
t, u, w, r, q, p
∶∶=
x ∣ v ∣ tu ∣ t[x u]
v
∶∶= lx.t
The constructor t[x u] is called an explicit substitution (of u for x in t). The
usual (implicit) substitution is instead denoted by t{x u}. Both lx.t and t[x u]
bind x in t, with the usual notion of α-equivalence. Values, noted v, do not
include variables: this is a standard choice in the study of abstract machines.
Contexts are terms with one occurrence of the hole ⟨⋅⟩, an additional constant. We will use many different contexts. The most general ones will be weak
contexts W (i.e. not under abstractions), which are defined by:
W, W ′
∶∶= ⟨⋅⟩ ∣ W u ∣ tW ∣ W [x u] ∣ t[x W ]
The plugging W ⟨t⟩ (resp. W ⟨W ′ ⟩) of a term t (resp. context W ′ ) in a context W
is defined as ⟨t⟩ ∶= t (resp. ⟨W ′ ⟩ ∶= W ′ ), (W t)⟨u⟩ ∶= W ⟨u⟩t (resp. (W t)⟨W ′ ⟩ ∶=
W ⟨W ′ ⟩t), and so on. The set of free variables of a term t (or context W )
is denoted by fv(t) (resp. fv(W )). Plugging in a context may capture free
variables (replacing holes on the left of substitutions). These notions will be
silently extended to all the contexts used in the paper.
Rewriting Rules. On the above terms, one may define several variants of the
LSC by considering two elementary rewriting rules, distance-β (dB) and linear
substitution (ls), each one coming in two variants, call-by-name and call-byvalue (the latter variants being abbreviated by dBv and lsv), and pairing them
in different ways and with respect to different evaluation contexts.
The rewriting rules rely in multiple ways on contexts. We start by defining
substitution contexts, generated by
L
∶∶= ⟨⋅⟩ ∣ L[x t].
6
↦m
Calculus Evaluation contexts
↦e
H
⊸m
⊸e
H
Name
H ∶∶= ⟨⋅⟩ ∣ Ht ∣ H[x t]
↦dB
ValueLR
V ∶∶= ⟨⋅⟩ ∣ V t ∣ L⟨v⟩V ∣ V [x t]
↦dBv ↦lsv V ⟨↦dB ⟩ V ⟨↦ls ⟩
ValueRL
S ∶∶= ⟨⋅⟩ ∣ SL⟨v⟩ ∣ tS ∣ S[x t]
↦dBv ↦lsv S⟨↦dB ⟩
Need
N ∶∶= ⟨⋅⟩ ∣ N t ∣ N [x t] ∣ N ′ ⟨x⟩[x N ] ↦dB
↦ls
H⟨↦dB ⟩ H⟨↦ls ⟩
V
S
N
V
S
S⟨↦ls ⟩
N
↦lsv N ⟨↦dB ⟩ N ⟨↦ls ⟩
Table 1: The four linear substitution calculi.
A term of the form L⟨v⟩ is an answer. Given a family of contexts C, the two
variants of the elementary rewriting rules, also called root rules, are defined as
follows:
L⟨lx.t⟩u ↦dB
L⟨lx.t⟩L′ ⟨v⟩ ↦dBv
C⟨x⟩[x u] ↦ls
C⟨x⟩[x L⟨v⟩] ↦lsv
L⟨t[x u]⟩
L⟨t[x L′ ⟨v⟩]⟩
C⟨u⟩[x u]
L⟨C⟨v⟩[x v]⟩
In the linear substitution rules, we assume that x ∈ fv(C⟨x⟩), i.e., the context C
does not capture the variable x, and we also silently work modulo α-equivalence
to avoid variable capture in the rewriting rules. Moreover, we use the notations
C
C
↦ls and ↦lsv to specify the family of contexts used by the rules, with C being
the meta-variable ranging over such contexts.
All of the above rules are at a distance (or contextual ) because their definition
involves contexts. Distance-β and linear substitution correspond, respectively,
to the so-called multiplicative and exponential rules for cut-elimination in proof
nets. The presence of contexts is how locality on proof nets is reflected on terms.
The rewriting rules decompose the usual small-step semantics for l-calculi,
by substituting one occurrence at the time, and only when such an occurrence is
in evaluation position. We emphasise this fact saying that we adopt a micro-step
semantics.
A linear substitution calculus is defined by a choice of root rules, i.e., one
of dB/dBv and one of ls/lsv, and a family of evaluation contexts. The chosen
distance-β (resp. linear substitution) root rule is generically denoted by ↦m (resp.
↦e ). If E ranges over a fixed notion of evaluation context, the context-closures
of the root rules are denoted by ⊸m ∶= E⟨↦m ⟩ and ⊸e ∶= E⟨↦e ⟩, where m (resp.
e) stands for multiplicative (exponential ). The rewriting relation defining the
calculus is then ⊸∶=⊸m ∪ ⊸e .
2.1
Calculi
We consider four calculi, noted Name, ValueLR , ValueRL , and Need, and defined
in Tab. 1. They correspond to four standard evaluation strategies for functional languages. We are actually slightly abusing the terminology, because—as
we will show—they are deterministic calculi and thus should be considered as
strategies. Our abuse is motivated by the fact that they are not strategies in
the same calculus. The essential property of all these four calculi is that they
are deterministic, because they implement a reduction strategy.
7
Proposition 2.1 (Determinism). The reduction relations of the four calculi of
Tab. 1 are deterministic: in each calculus, if E1 , E2 are evaluation contexts and
if r1 , r2 are redexes (i.e., terms matching the left hand side of the root rules
defining the calculus), E1 ⟨r1 ⟩ = E2 ⟨r2 ⟩ implies E1 = E2 and r1 = r2 , so that
there is at most one way to reduce a term.
Proof. See Sect. 13 in the appendix (page 33).
Call-by-Name (CBN). The evaluation contexts for Name are called weak
head contexts and—when paired with micro-step evaluation—implement a strategy known as weak linear head reduction. The original presentation of this strategy does not use explicit substitutions [MP94, DR04]. The presentation in use
here has already appeared in [Acc13, ABKL14] (see also [Acc12, ADL12]) as the
weak head strategy of the linear substitution calculus (which is obtained by considering all contexts as evaluation contexts), and it avoids many technicalities
of the original one. In particular, its relationship with the KAM is extremely
natural, as we will show.
Let us give some examples of evaluation. Let δ ∶= lx.(xx) and consider the
usual diverging term Ω ∶= δδ. In Name it evaluates—diverging—as follows:
δδ
⊸m
(xx)[x
(δx)[x
(yy)[y
(xy)[y
(δy)[y
(zz)[z
δ]
δ]
x][x
x][x
x][x
y][y
⊸e
⊸m
δ]
⊸e
δ]
⊸e
δ]
⊸m
x][x δ] ⊸e
...
Observe that according to our definitions both lx.Ω and xΩ are ⊸-normal
for Name, because evaluation does not go under abstractions, nor on the right
of a variable (but terms like xΩ will be forbidden, as we will limit ourselves to
closed terms). Now let show the use of the context L in rule ⊸m . Let I ∶= ly.y
and τ ∶= (lz.δ)I, and consider the following variation over Ω, where rule ⊸m is
applied with L ∶= ⟨⋅⟩[z I]:
ττ
⊸m
δ[z I]τ
⊸m
(xx)[x τ ][z I] ⊸e
...
Call-by-Value (CBV). For call-by-value calculi, left-to-right (ValueLR ) and
right-to-left (ValueRL ) refer to the evaluation order of applications, i.e. they
correspond to operator first and argument first, respectively. The two calculi we
consider here can be seen as strategies of a micro-step variant of the value substitution calculus, the (small-step) call-by-value calculus at a distance introduced
and studied in [AP12].
As an example, we consider again the evaluation of Ω. In ValueLR it goes as
follows:
δδ ⊸m (x1 x1 )[x1 δ]
⊸e
(δx1 )[x1 δ]
⊸e
(δδ)[x1 δ]
⊸m
(x2 x2 )[x2 δ][x1 δ] ⊸e
(δx2 )[x2 δ][x1 δ]
⊸e . . .
8
While in ValueRL it takes the following form:
δδ
⊸m
(x1 x1 )[x1 δ]
(x1 δ)[x1 δ]
(δδ)[x1 δ]
(x2 x2 )[x2 δ][x1 δ]
(x2 δ)[x2 δ][x1 δ]
⊸e
⊸e
⊸m
⊸e
⊸e . . .
Note that the CBV version of ⊸m and ⊸e employ substitution contexts L in a
new way. An example of their use is given by the term τ τ consider before for
CBN. For instance, in ValueLR we have:
ττ
⊸m
δ[z I]τ
⊸m
δ[z I](δ[z I])
⊸m
(xx)[x δ[z I]][z I] ⊸e
(δx)[x δ[z I]][z I] . . .
Call-by-Need (CBNeed). The call-by-need calculus Need is a novelty of
this paper, and can be seen either as a version at a distance of the calculi of
[MOW98, AF97] or as a version with explicit substitution of the one in [CF12].
It fully exploits the fact that the two variants of the root rules may be combined: the β-rule is call-by-name, which reflects the fact that, operationally,
the strategy is by name, but substitution is call-by-value, which forces arguments to be evaluated before being substituted, reflecting the by need content
of the strategy. Please note the definition of CBNeed evaluation contexts N
in Tab. 1. They extend the weak head contexts for call-by-name with a clause
(N ′ ⟨x⟩[x N ]) turning them into hereditarily weak head contexts. This new
clause is how sharing is implemented by the reduction strategy. The general
(non-deterministic) calculus is obtained by closing the root rules by all contexts, but its study is omitted. What we deal with here can be thought as its
standard strategy (stopping on a sort of weak head normal form).
Let us show, once again, the evaluation of Ω.
δδ
⊸m
(x1 x1 )[x1
(δx1 )[x1
(x2 x2 )[x2
(x2 x2 )[x2
(δx2 )[x2
(x3 x3 )[x3
(x3 x3 )[x3
(δx3 )[x3
δ]
δ]
x1 ][x1 δ]
δ][x1 δ]
δ][x1 δ]
x2 ][x2 δ][x1 δ]
δ][x2 δ][x1 δ]
δ][x2 δ][x1 δ]
⊸e
⊸m
⊸e
⊸e
⊸m
⊸e
⊸e
⊸m
...
Note the difference with CBN: hereditarily weak evaluation contexts allow
the micro-step substitution rule to replace variable occurrences in explicit substitutions.
As shown by the evaluation of Ω, the various calculi considered in this paper not only select different β-redexes, they are also characterised by different
substitution processes. Such processes are the object of a detailed analysis in
the companion paper [ASC14].
9
t[x u] ≡gc
t[x u][y w] ≡com
t[x u][y w] ≡[⋅]
t
if x ∉ fv(t)
t[y w][x u] if y ∉ fv(u) and x ∉ fv(w)
t[x u[y w]] if y ∈/ fv(t)
t[x u] ≡dup
(tw)[x u] ≡@
(tw)[x u] ≡@l
Figure 1: Axioms for structural equivalences. In ≡dup , t[y]x denotes a term
obtained from t by renaming some (possibly none) occurrences of x as y.
Structural equivalence. Another common feature of the four calculi is that
they come with a notion of structural equivalence, denoted by ≡. Consider
Fig. 1. For call-by-name and call-by-value calculi, ≡ is defined as the smallest
equivalence relation containing the closure by weak contexts of =α ∪ ≡gc ∪ ≡dup
∪ ≡@ ∪ ≡com ∪ ≡[⋅] where =α is α-equivalence. Call-by-need evaluates inside some
substitutions (those hereditarily substituting on the head) and thus axioms as
≡dup and ≡@ are too strong. Therefore, the structural equivalence for call-byneed, noted ≡Need , is the one generated by ≡@l ∪ ≡com ∪ ≡[⋅] .
Structural equivalence represents the fact that certain manipulations on explicit substitutions are computationally irrelevant, in the sense that they yield
behaviorally equivalent terms. Technically, it is a strong bisimulation:
Proposition 2.2 (≡ is a Strong Bisimulation). Let ⊸m , ⊸e and ≡ be the reduction relations and the structural equivalence relation of any of the calculi of
Tab. 1, and let x ∈ {m, e}. Then, t ≡ u and t ⊸x t′ implies that there exists u′
such that u ⊸x u′ and t′ ≡ u′ .
Proof. See Sect. 14 of the appendix (page 37).
The essential property of strong bisimulations is that they can be postponed.
In fact, it is immediate to prove the following, which holds for all four calculi:
Lemma 2.3 (≡ Postponement). If t (⊸m ∪ ⊸e ∪ ≡)∗ u then t (⊸m ∪ ⊸e )∗ ≡ u
and the number of ⊸m and ⊸e steps in the two reduction sequences is exactly
the same.
In the simulation theorems for machines with a global environment (see
Sect. 7.1 and Sect. 8) we will also use the following commutation property
between substitutions and evaluation contexts via the structural equivalence of
every evaluation scheme, proved by an easy induction on the actual definition
of evaluation contexts.
Lemma 2.4 (ES Commute with Evaluation Contexts via ≡). For every evaluation scheme let C denote an evaluation context s.t. x ∉ fv(C) and ≡ be its
structural equivalence. Then C⟨t⟩[x u] ≡ C⟨t[x u]⟩.
3
Preliminaries on Abstract Machines.
Codes. All the abstract machines we will consider execute pure λ-terms. In
our syntax, these are nothing but terms without explicit substitutions. Moreover, while for calculi we work implicitly modulo α, for machines we will not
consider terms up to α, as the handling of α-equivalence characterizes different
10
t[y]x [x u][y u]
t[x u]w[x u]
t[x u]w
i
approaches to abstract machines. To stress these facts, we use the metavariables
t, u, w, r for pure λ-terms (not up to α) and v for pure values.
States. A machine state s will have various components, of which the first
will always be the code, i.e. a pure l-term t. The others (environment, stack,
dump) are all considered as lists, whose constructors are the empty list ǫ and
the concatenation operator ∶∶. A state s of a machine is initial if its code t is
closed (i.e., fv(t) = ∅) and all other components are empty. An execution ρ is a
sequence of transitions of the machine s0 →∗ s from an initial state s0 . In that
case, we say that s is a reachable state, and if t is the code of s0 then t is the
initial code of s.
Invariants. For every machine our study will rely on a lemma about some
dynamic invariants, i.e. some properties of the reachable states that are stable
by executions. The lemma is always proved by a straightforward induction on
the length of the execution and the proof is omitted.
Environments and Closures. There will be two types of machines, those with
many local environments and those with just one global environment. Machines
with local environments are based on the mutually recursive definition of closure
(ranged over by c) and environment (e):
c
∶∶=
(t, e)
e
∶∶=
ǫ ∣ [x c] ∶∶ e
Global environments are defined by E ∶∶= ǫ ∣ [x t] ∶∶ E, and global environment
machines will have just one global closure (t, E).
Well-Named and Closed Closures. The explicit treatment of α-equivalence, is
based on particular representatives of α-classes defined via the notion of support.
The support ∆ of codes, environments, and closures is defined by:
• ∆(t) is the multiset of its bound names (e.g. ∆(lx.ly.lx.(zx))
= [x, x, y]).
• ∆(e) is the multiset of names captured by e (for
∆([x c1 ][y c2 ][x c3 ]) = [x, x, y]), and similarly for ∆(E).
example
• ∆(t, e) ∶= ∆(t) + ∆(e) and ∆(t, E) ∶= ∆(t) + ∆(E).
A code/environment/closure is well-named if its support is a set (i.e. a multiset with no repetitions). Moreover, a closure (t, e) (resp. (t, E)) is closed if
fv(t) ⊆ ∆(e) (resp. fv(t) ⊆ ∆(E)).
4
Distilleries
This section presents an abstract, high-level view of the relationship between
abstract machines and linear substitution calculi, via the notion of distillery.
Definition 4.1. A distillery D = (M, C, ≡, ⋅ ) is given by:
1. An abstract machine M, given by
(a) a deterministic labeled transition system → on states s;
(b) a distinguished class of states called initials (in bijection with closed
l-terms, and from which applying → one obtains the reachable states);
(c) a partition of the labels of the transition system → as:
11
• commutative transitions, noted →c ;
• principal transitions, in turn partitioned into
– multiplicative transitions, denoted by →m ;
– exponential transitions, denoted by →e ;
2. a linear substitution calculus C given by a pair (⊸m , ⊸e ) of rewriting
relations on terms with ES;
3. a structural equivalence ≡ on terms s.t. it is a strong bisimulation with
respect to ⊸m and ⊸e ;
4. a distillation ⋅ , i.e. a decoding function from states to terms, s.t. on
reachable states:
• Commutative: s →c s′ implies s ≡ s′ .
• Multiplicative: s →m s′ implies s ⊸m ≡ s′ ;
• Exponential: s →e s′ implies s ⊸e ≡ s′ ;
Given a distillery, the simulation theorem holds abstractly. Let ∣ρ∣ (resp. ∣d∣),
∣ρ∣m (resp. ∣d∣m ), ∣ρ∣e (resp. ∣d∣e ), and ∣ρ∣p denote the number of unspecified,
multiplicative, exponential, and principal steps in an execution (resp. derivation).
Theorem 4.2 (Simulation). Let D be a distillery. Then for every execution
ρ ∶ s →∗ s′ there is a derivation d ∶ s ⊸∗ ≡ s′ s.t. ∣ρ∣m = ∣d∣m , ∣ρ∣e = ∣d∣e , and
∣ρ∣p = ∣d∣.
Proof. By induction on ∣ρ∣ and by the properties of the decoding, it follows that
there is a derivation e ∶ s(⊸≡)∗ s′ s.t. the number ∣ρ∣p = ∣e∣. The witness d for
the statement is obtained by applying the postponement of strong bisimulations
(Lemma 2.3) to e.
Reflection. Given a distillery, one would also expect that reduction in the
calculus is reflected in the machine. This result in fact requires two additional
abstract properties.
Definition 4.3 (Reflective Distillery). A distillery is reflective when:
Termination: →c terminates (on reachable states); hence, by determinism,
every state s has a unique commutative normal form nfc (s);
Progress: if s is reachable, nfc (s) = s and s ⊸x t with x ∈ {m, e}, then there
exists s′ such that s →x s′ , i.e., s is not final.
Then, we may prove the following reflection of steps in full generality:
Proposition 4.4 (Reflection). Let D be a reflective distillery, s be a reachable
state, and x ∈ {m, e}. Then, s ⊸x u implies that there exists a state s′ s.t.
nfc (s) →x s′ and s′ ≡ u.
In other words, every rewriting step on the calculus can be also performed
on the machine, up to commutative transitions.
12
Proof. The proof is by induction on the number n of transitions leading from s
to nfc (s).
• Base case n = 0: by the progress property, we have s →x′ s′ for some
state s′ and x′ ∈ {m, e}. By Theorem 4.2, we have s ⊸x′ u′ ≡ s′ and we
may conclude because x′ = x and u′ = u by determinisim of the calculus
(Proposition 2.1).
• Inductive case n > 0: by hypothesis, we have s →c s1 . By Theorem 4.2, s ≡
s1 . The hypothesis and the strong bisimulation property (Proposition 2.2)
then give us s1 ⊸x u1 ≡ u. But the induction hypothesis holds for s1 ,
giving us a state s′ such that nfc (s1 ) →x s′ and s′ ≡ u1 ≡ u. We may now
conclude because nfc (s) = nfc (s1 ).
The reflection can then be extended to a reverse simulation.
Corollary 4.5 (Reverse Simulation). Let D be a reflective distillery and s an
initial state. Given a derivation d ∶ s ⊸∗ t there is an execution ρ ∶ s →∗ s′ s.t.
t ≡ s′ and ∣ρ∣m = ∣d∣m , ∣ρ∣e = ∣d∣e , and ∣ρ∣p = ∣d∣.
Proof. By induction on the length of d, using Proposition 4.4.
In the following sections we shall introduce abstract machines and distillations for which we will prove that they form reflective distilleries with respect
to the calculi of Sect. 2. For each machine we will prove 1) that the decoding is
in fact a distillation, and 2) the progress property. We will instead assume the
termination property, whose proof is delayed to the quantitative study of the
second part of the paper, where we will actually prove stronger results, giving
explicit bounds.
5
Call-by-Name: the KAM
The Krivine Abstract Machine (KAM) is the simplest machine studied in the
paper. A KAM state (s) is made out of a closure and of a stack (π):
π
∶∶=
ǫ ∣ c ∶∶ π
s ∶∶=
(c, π)
For readability, we will use the notation t ∣ e ∣ π for a state (c, π) where c = (t, e).
The transitions of the KAM then are:
tu
lx.t
x
e
e
e
π
c ∶∶ π
π
→c
→m
→e
t
t
t
e
[x c] ∶∶ e
e′
(u, e) ∶∶ π
π
π
where →e takes place only if e = e′′ ∶∶ [x (t, e′ )] ∶∶ e′′′ .
A key point of our study is that environments and stacks rather immediately
become contexts of the LSC, through the following decoding:
ǫ ∶= ⟨⋅⟩
(t, e) ∶= e⟨t⟩
t ∣ e ∣ π ∶= π⟨e⟨t⟩⟩
[x c] ∶∶ e ∶=
c ∶∶ π ∶=
e⟨⟨⋅⟩[x c]⟩
π⟨⟨⋅⟩c⟩
The decoding satisfies the following static properties, shown by easy inductions on the definition.
13
Lemma 5.1 (Contextual Decoding). Let e be an environment and π be a stack
of the KAM. Then e is a substitution context, and both π and π⟨e⟩ are evaluation
contexts.
Next, we need the dynamic invariants of the machine.
Lemma 5.2 (KAM Invariants). Let s = u ∣ e ∣ π be a KAM reachable state
whose initial code t is well-named. Then:
1. Closure: every closure in s is closed;
2. Subterm: any code in s is a literal subterm of t.
3. Name: any closure in s is well-named.
4. Environment Size: the length of any environment in s is bound by ∣t∣.
Abstract Considerations on Concrete Implementations. The name invariant
is the abstract property that allows to avoid α-equivalence in KAM executions.
In addition, forbidding repetitions in the support of an environment, it allows
to bound the length of any environment with the names in t, i.e. with ∣t∣. This
fact is important, as the static bound on the size of environments guarantees
that →e and →c —the transitions looking-up and copying environments—can
be implemented (independently of the chosen concrete representation of terms)
in at worst linear time in ∣t∣, so that an execution ρ can be implemented in
O(∣ρ∣ ⋅ ∣t∣). The same will hold for every machine with local environments.
The previous considerations are based on the name and environment size
invariants. The closure invariant is used in the progress part of the next theorem, and the subterm invariant is used in the quantitative analysis in Sect. 11
(Theorem 11.3), subsuming the termination condition of reflective distilleries.
Theorem 5.3 (KAM Distillation). (KAM, Name, ≡, ⋅ ) is a reflective distillery.
In particular, on a reachable state s we have:
1. Commutative: if s →c s′ then s ≡ s′ .
2. Multiplicative: if s →m s′ then s ⊸m s′ ;
3. Exponential: if s →e s′ then s ⊸e ≡ s′ ;
Proof. Properties of the decoding:
1. Commutative. We have tu ∣ e ∣ π →c t ∣ e ∣ (u, e) ∶∶ π, and:
tu ∣ e ∣ π
=
≡∗@
π⟨e⟨tu⟩⟩
π⟨e⟨t⟩e⟨u⟩⟩ = t ∣ e ∣ (u, e) ∶∶ π
2. Multiplicative. lx.t ∣ e ∣ c ∶∶ π →m t ∣ [x c] ∶∶ e ∣ π, and
lx.t ∣ e ∣ c ∶∶ π
=
⊸m
=
π⟨e⟨lx.t⟩c⟩
π⟨e⟨t[x c]⟩⟩
t ∣ [x c] ∶∶ e ∣ π
The rewriting step can be applied because by contextual decoding
(Lemma 5.1) it takes place in an evaluation context.
14
3. Exponential. x ∣ e′ ∶∶ [x (t, e)] ∶∶ e′′ ∣ π →e t ∣ e ∣ π, and
x ∣ e′ ∶∶ [x (t, e)] ∶∶ e′′ ∣ π
=
⊸e
≡∗gc
=
π⟨e′′ ⟨e′ ⟨x⟩[x e⟨t⟩]⟩⟩
π⟨e′′ ⟨e′ ⟨e⟨t⟩⟩[x e⟨t⟩]⟩⟩
π⟨e⟨t⟩⟩
t∣e∣π
Note that e′′ ⟨e′ ⟨e⟨t⟩⟩[x e⟨t⟩]⟩ ≡∗gc e⟨t⟩ holds because e⟨t⟩ is closed by point
1 of Lemma 5.2, and so all the substitutions around it can be garbage
collected.
Termination. Given by (forthcoming) Theorem 11.3 (future proofs of distillery
theorems will omit termination).
Progress. Let s = t ∣ e ∣ π be a commutative normal form s.t. s ⊸ u. If t is
• an application uw. Then a →c transition applies and s is not a commutative normal form, absurd.
• an abstraction lx.u: if π = ǫ then s = e⟨lx.u⟩, which is ⊸-normal, absurd.
Hence, a →m transition applies.
• a variable x: by point 1 of Lemma 5.2.1, we must have e = e′ ∶∶ [x c] ∶∶ e′′ ,
so a →e transition applies;
6
Call-by-Value: the CEK and the LAM
Here we deal with two adaptations to call-by-value of the KAM, namely Felleisen
and Friedman’s CEK machine [FF86] (without control operators), and a variant,
deemed Leroy abstract machine (LAM). They differ on how they behave with
respect to applications: the CEK implements left- to-right call-by-value, i.e. it
first evaluates the function part, the LAM gives instead precedence to arguments, realizing right-to-left call-by-value. The LAM owes its name to Leroy’s
ZINC machine [Ler90], that implements right-to-left call-by-value evaluation.
We introduce a new name because the ZINC is a quite more sophisticated machine than the LAM: it has a separate sets of instructions to which terms are
compiled, it handles arithmetic expressions, and it avoids needless closure creations in a way that it is not captured by the LAM.
The states of the CEK and the LAM have the same shape of those of the
KAM, i.e. they are given by a closure plus a stack. The difference is that
they use call-by-value stacks, whose elements are labelled either as functions or
arguments, so that the machine may know whether it is launching the evaluation
of an argument or it is at the end of such an evaluation. They are re-defined
and decoded by (c is a closure):
π
∶∶=
ǫ ∣ f (c) ∶∶ π ∣ a(c) ∶∶ π
ǫ
f (c) ∶∶ π
a(c) ∶∶ π
∶= ⟨⋅⟩
∶= π⟨c⟨⋅⟩⟩
∶= π⟨⟨⋅⟩c⟩
The states of both machines are decoded exactly as for the KAM, i.e. t ∣ e ∣ π ∶=
π⟨e⟨t⟩⟩.
15
6.1
Left-to Right Call-by-Value: the CEK machine.
The transitions of the CEK are:
tu
v
v
x
e
π
→c1
e a(u, e′ ) ∶∶ π →c2
e f (lx.t, e′ ) ∶∶ π →m
e
π
→e
t
e
a(u, e) ∶∶ π
u
e′
f (v, e) ∶∶ π
t [x (v, e)] ∶∶ e′
π
t
e′
π
where →e takes place only if e = e′′ ∶∶ [x (t, e′ )] ∶∶ e′′′ .
While one can still statically prove that environments decode to substitution
contexts, to prove that π and π⟨e⟩ are evaluation contexts we need the dynamic
invariants of the machine.
Lemma 6.1 (CEK Invariants). Let s = u ∣ e ∣ π be a CEK reachable state whose
initial code t is well-named. Then:
1. Closure: every closure in s is closed;
2. Subterm: any code in s is a literal subterm of t;
3. Value: any code in e is a value and, for every element of π of the form
f (u, e′ ), u is a value;
4. Contextual Decoding: π and π⟨e⟩ are left-to-right call-by-value evaluation
contexts.
5. Name: any closure in s is well-named.
6. Environment Size: the length of any environment in s is bound by ∣t∣.
We have everything we need:
Theorem 6.2 (CEK Distillation). (CEK, ValueLR , ≡, ⋅ ) is a reflective distillery.
In particular, on a reachable state s we have:
1. Commutative 1: if s →c1 s′ then s ≡ s′ ;
2. Commutative 2: if s →c2 s′ then s = s′ .
3. Multiplicative: if s →m s′ then s ⊸m s′ ;
4. Exponential: if s →e s′ then s ⊸e ≡ s′ ;
Proof. Properties of the decoding: in the following cases, evaluation will always
takes place under a context that by Lemma 6.1.4 will be a left-to-right call-byvalue evaluation context, and similarly structural equivalence will alway be used
in a weak context, as it should be.
1. Commutative 1. We have tu ∣ e ∣ π →c1 t ∣ e ∣ a(u, e) ∶∶ π, and:
tu ∣ e ∣ π
= π⟨e⟨tu⟩⟩
π⟨e⟨t⟩e⟨u⟩⟩
≡∗@
=
t ∣ e ∣ a(u, e) ∶∶ π
2. Commutative 2. We have v ∣ e ∣ a(u, e′ ) ∶∶ π →c2 u ∣ e′ ∣ f (v, e) ∶∶ π, and:
16
v ∣ e ∣ a(u, e′ ) ∶∶ π
= π⟨e⟨v⟩e′ ⟨u⟩⟩
u ∣ e′ ∣ f (v, e) ∶∶ π
=
3. Multiplicative. We have v ∣ e ∣ f (lx.t, e′ ) ∶∶ π →m u ∣ [x (v, e)] ∶∶ e′ ∣ π,
and:
v ∣ e ∣ f (lx.t, e′ ) ∶∶ π
= π⟨e′ ⟨lx.t⟩e⟨v⟩⟩
π⟨e′ ⟨t[x e⟨v⟩]⟩⟩
t ∣ [x (v, e)] ∶∶ e′ ∣ π
⊸m
=
4. Exponential. Let e = e′′ ∶∶ [x (t, e′ )] ∶∶ e′′′. We have x ∣ e ∣ π →e t ∣ e′ ∣ π,
and:
x∣e∣π
= π⟨e⟨x⟩⟩
π⟨e′′′ ⟨e′′ ⟨x⟩[x e′ ⟨t⟩]⟩⟩
π⟨e′′′ ⟨e′ ⟨e′′ ⟨t⟩[x t]⟩⟩⟩
π⟨e′ ⟨t⟩⟩
=
⊸e
≡∗gc
=
t ∣ e′ ∣ π
We can apply ⊸e since by Lemma 6.1.3, t is a value. We also use that
by Lemma 6.1.1, e′ ⟨t⟩ is a closed term to ensure that e′′ and e′′′ can be
garbage collected.
Progress. Let s = t ∣ e ∣ π be a commutative normal form s.t. s ⊸ u. If t is
• an application uw. Then a →c1 transition applies and s is not a commutative normal form, absurd.
• an abstraction v: by hypothesis, π cannot be of the form a(c) ∶∶ π ′ . Suppose it is equal to ǫ. We would then have s = e⟨v⟩, which is a call-by-value
normal form, because e is a substitution context. This would contradict
our hypothesis, so π must be of the form f (u, e′ ) ∶∶ π ′ . By point 3 of
Lemma 6.1, u is an abstraction, hence a →m transition applies.
• a variable x: by point 1 of Lemma 6.1, e must be of the form e′ ∶∶ [x c] ∶∶
e′′ , so a →e transition applies;
6.2
Right-to-Left Call-by-Value: the Leroy Abstract Machine
The transitions of the LAM are:
tu
v
lx.t
x
e
e
e
e
π
f (t, e′ ) ∶∶ π
a(c) ∶∶ π
π
→c1
→c2
→m
→e
u
t
t
t
e
e′
[x c] ∶∶ e
e′
f (t, e) ∶∶ π
a(v, e) ∶∶ π
π
π
where →e takes place only if e = e′′ ∶∶ [x (t, e′ )] ∶∶ e′′′ .
We omit all the proofs (that can be found in the appendix, page 57) because
they are minimal variations on those for the CEK.
Lemma 6.3 (LAM Invariants). Let s = u ∣ e ∣ π be a LAM reachable state whose
initial code t is well-named. Then:
17
tu
v
v
x
e
e
e
e ∶∶ [x (v, e′ )] ∶∶ e′′
π
(t, e′ ) ∶∶ π
ǫ
π
D
D
((lx.t, e′ ), π) ∶∶ D
D
→c1
→c2
→m
→e
t
t
t
v
Figure 2: The Split CEK, aka the revisited SECD.
1. Closure: every closure in s is closed;
2. Subterm: any code in s is a literal subterm of t;
3. Value: any code in e is a value and, for every element of π of the form
a(u, e′ ), u is a value;
4. Contexts Decoding: π and π⟨e⟩ are right-to-left call-by-value evaluation
contexts.
5. Name: any closure in s is well-named.
6. Environment Size: the length of any environment in s is bound by ∣t∣.
Theorem 6.4 (LAM Distillation). (LAM, ValueRL , ≡, ⋅ ) is a reflective distillery. In particular, on a reachable state s we have:
1. Commutative 1: if s →c1 s′ then s ≡ s′ ;
2. Commutative 2: if s →c2 s′ then s = s′ .
3. Multiplicative: if s →m s′ then s ⊸m s′ ;
4. Exponential: if s →e s′ then s ⊸e ≡ s′ ;
7
Towards Call-by-Need:
Split CEK
the MAM and the
In this section we study two further machines:
1. The Milner Abstract Machine (MAM), that is a variation over the KAM
with only one global environment and without the concept of closure.
Essentially, it unveils the content of distance rules at the machine level.
2. The Split CEK (SCEK), obtained disentangling the two uses of the stack
(for arguments and for functions) in the CEK. The split CEK can be seen
as a simplification of Landin’s SECD machine [Lan64].
The ideas at work in these two case studies will be combined in the next section,
obtaining a new simple call-by-need machine.
18
e
e′
[x (v, e)] ∶∶ e′
e′
(u, e) ∶∶ π
ǫ
π
π
7.1
Milner Abstract Machine
The linear substitution calculus suggests the design of a simpler version of the
KAM, the Milner Abstract Machine (MAM), that avoids the concept of closure.
At the language level, the idea is that, by repeatedly applying the axioms ≡dup
and ≡@ of the structural equivalence, explicit substitutions can be folded and
brought outside. At the machine level, the local environments in the closures
are replaced by just one global environment that closes the code and the stack,
as well as the global environment itself.
Of course, naively turning to a global environment breaks the well-named
invariant of the machine. This point is addressed using an α-renaming in the
variable transition, i.e. when substitution takes place. Here we employ the
global environments E of Sect. 3 and we redefine stacks as π ∶∶= ǫ ∣ t ∶∶ π. A state
of the MAM is given by a code t, a stack π and a global environment E. Note
that the code and the stack together now form a code.
The transitions of the MAM are:
tu
lx.t
x
π
u ∶∶ π
π
E
E
E
→c
→m
→e
t
t
α
t
u ∶∶ π
π
π
E
[x u] ∶∶ E
E
α
where →e takes place only if E = E ′′ ⟨E ′ [x t]⟩ and t is a well-named code
α
α-equivalent to t and s.t. any bound name in t is fresh with respect to those
in π and E 1 .
The decoding of a MAM state t ∣ π ∣ E is similar to the decoding of a KAM
state, but the stack and the environment context are applied in reverse order
(this is why stack and environment in MAM states are swapped with respect to
KAM states):
ǫ ∶=
t ∶∶ π ∶=
⟨⋅⟩
π⟨⟨⋅⟩t⟩
[x t] ∶∶ E
t∣π∣E
∶=
∶=
E⟨⟨⋅⟩[x t]⟩
E⟨π⟨t⟩⟩
To every MAM state t ∣ π ∣ E we associate the pair (tπ, E) and call it the global
closure of the state.
As for the KAM, the decoding of contexts can be done statically, i.e. it does
not need dynamic invariants.
Lemma 7.1 (Contextual Decoding). Let E be a global environment and π be a
stack of the MAM. Then E is a substitution context, and both π and π⟨E⟩ are
evaluation contexts.
For the dynamic invariants we need a different notion of closed closure.
Definition 7.2. Given a global environment E and a code t, we define by mutual
induction two predicates E is closed and (t, E) is closed as follows:
(t, E) is closed
fv(t) ⊆ ∆(E) ∧ E is closed
Ô⇒
Ô⇒
ǫ is closed
[x t] ∶∶ E is closed
(t, E) is closed
1
The well-named invariant can be restored also in another way. One can simply substitute
α
t (instead of t ) but modify →m as follows (with y fresh):
lx.t
u ∶∶ π
E
→m
t{x y}
19
π
[y u] ∶∶ E
The dynamic invariants are:
Lemma 7.3 (MAM invariants). Let s = u ∣ π ∣ E be a MAM state reached by
an execution ρ of initial well-named code t. Then:
1. Global Closure: the global closure (tπ, E) of s is closed;
2. Subterm: any code in s is a literal subterm of t;
3. Names: the global closure of s is well-named;
4. Environment Size: the length of the global environment in s is bound by
∣ρ∣m .
Abstract Considerations on Concrete Implementations. Note the new environment size invariant, whose bound is laxer than for local environment machines. Let ρ be a execution of initial code t. If one implements →e looking
for x in E sequentially, then each →e transition has cost ∣ρ∣m (more precisely,
linear in the number of preceding →m transitions) and the cost of implementing
ρ is easily seen to become quadratic in ∣ρ∣. An efficient implementation would
then employ a representation of codes such that variables are pointers, so that
looking for x in E takes constant time. The name invariant guarantees that
variables can indeed taken as pointers, as there is no name clash. Note that
the cost of a →e transition is not constant, as the renaming operation actually
makes →e linear in ∣t∣ (by the subterm invariant). So, assuming a pointer-based
representation, ρ can be implemented in time O(∣ρ∣ ⋅ ∣t∣), as for local machines,
and the same will hold for every global environment machine.
Theorem 7.4 (MAM Distillation). (MAM, Name, ≡, ⋅ ) is a reflective distillery.
In particular, on a reachable state s we have:
1. Commutative: if s →c s′ then s = s′ ;
2. Multiplicative: if s →m s′ then s ⊸m ≡ s′ ;
3. Exponential: if s →e s′ then s ⊸e =α s′ .
Proof. Properties of the decoding (progress is as for the KAM):
1. Commutative. In contrast to the KAM, →c gives a true identity:
tu ∣ π ∣ E
=
E⟨π⟨tu⟩⟩
= t ∣ u ∶∶ π ∣ E
2. Multiplicative. Since substitutions and evaluation contexts commute via
≡ (Lemma 2.4), →m maps to:
lx.t ∣ u ∶∶ π ∣ E
= E⟨π⟨(lx.t)u⟩⟩
E⟨π⟨t[x u]⟩⟩
E⟨π⟨t⟩[x u]⟩
t ∣ π ∣ [x u] ∶∶ E
⊸m
≡Lem.2.4
=
3. Exponential. The erasure of part of the environment of the KAM is replaced by an explicit use of α-equivalence:
20
tu
lx.t
x
v
π
u ∶∶ π
π
ǫ
D
D
D
(E1 , x, π) ∶∶ D
E
E
E1 ∶∶ [x t] ∶∶ E2
E2
→c1
→m
→c2
→e
t
t
t
vα
u ∶∶ π
π
ǫ
π
Figure 3: The Wadsworth Abstract Machine (WAM).
x ∣ π ∣ E ∶∶ [x u] ∶∶ E ′
= E ′ ⟨E⟨π⟨x⟩⟩[x u]⟩
E ′ ⟨E⟨π⟨u⟩⟩[x u]⟩
E ′ ⟨E⟨π⟨uα ⟩⟩[x u]⟩
uα ∣ π ∣ E ∶∶ [x u] ∶∶ E ′
⊸e
=α
=
Digression about ≡. Note that in the distillation theorem structural equivalence is used only to commute with stacks. The calculus and the machine in fact
form a distillery also with respect to the following simpler notion of structural
equivalence. Let ≡MAM be the smallest equivalence relation generated by the
closure by (call-by-name) evaluation contexts of the axiom ≡@l in Fig. 1 (page
10). The next lemma guarantees that ≡MAM is a strong bisimulation (the proof
is in the appendix, page 58), and so ≡MAM provides another MAM distillery.
Lemma 7.5. ≡MAM is a strong bisimulation with respect to ⊸.
7.2
The Split CEK, or Revisiting the SECD Machine
For the CEK machine we proved that the stack, that collects both arguments
and functions, decodes to an evaluation context (Lemma 6.1.4). The new CBV
machine in Fig. 2, deemed Split CEK, has two stacks: one for arguments and
one for functions. Both will decode to evaluation contexts. The argument
stack is identical to the stack of the KAM, and, accordingly, will decode to an
applicative context. Roughly, the function stack decodes to contexts of the form
H⟨v⟨⋅⟩⟩. More precisely, an entry of the function stack is a pair (c, π), where
c is a closure (v, e), and the three components v, e, and π together correspond
to the evaluation context π⟨e⟨v⟨⋅⟩⟩⟩. For the acquainted reader, this new stack
corresponds to the dump of Landin’s SECD machine [Lan64].
Let us explain the main idea. Whenever the code is an abstraction v and
the argument stack π is non-empty (i.e. π = c ∶∶ π ′ ), the machine saves the active
closure, given by current code v and environment e, and the tail of the stack π ′
by pushing a new entry ((v, e), π ′ ) on the dump, and then starts evaluating the
first closure c of the stack. The syntax for dumps then is
D ∶∶= ǫ ∣ (c, π) ∶∶ D
Every dump decodes to a context according to:
ǫ ∶=
⟨⋅⟩
((v, e), π) ∶∶ D
∶=
D⟨π⟨e⟨v⟨⋅⟩⟩⟩⟩
The decoding of terms, environments, closures, and stacks is as for the KAM.
The decoding of states is defined as t ∣ e ∣ π ∣ D ∶= D⟨π⟨e⟨t⟩⟩⟩. The proofs for the
Split CEK are in the appendix (page 58).
21
D
D
(E1 , x, π) ∶∶ D
D
E
[x u
E
E1 ∶∶ [x
Lemma 7.6 (Split CEK Invariants). Let s = u ∣ e ∣ π ∣ D be a Split CEK
reachable state whose initial code t is well-named. Then:
1. Closure: every closure in s is closed;
2. Subterm: any code in s is a literal subterm of t;
3. Value: the code of any closure in the dump or in any environment in s is
a value;
4. Contextual Decoding: D, D⟨π⟩, and D⟨π⟨e⟩⟩ are left-to-right call-by-value
evaluation context.
5. Name: any closure in s is well-named.
6. Environment Size: the length of any environment in s is bound by ∣t∣.
Theorem 7.7 (Split CEK Distillation). (Split CEK, ValueLR , ≡, ⋅ ) is a reflective distillery. In particular, on a reachable state s we have:
1. Commutative 1: if s →c1 s′ then s ≡ s′ ;
2. Commutative 2: if s →c2 s′ then s ≡ s′ ;
3. Multiplicative: if s →m s′ then s ⊸m s′ ;
4. Exponential: if s →e s′ then s ⊸e ≡ s′ .
8
Call-by-Need:
WAM
the WAM and the Merged
In this section we introduce a new abstract machine for call-by-need, deemed
Wadsworth Abstract Machine (WAM). The WAM arises very naturally as a
reformulation of the Need calculus of Sect. 2. The motivations behind the introduction of a new machine are:
1. Simplicity: the WAM is arguably simpler than all other CBNeed machines
in the literature, in particular its distillation is very natural;
2. Factorizing the Distillation of the Lazy KAM and of the SAM : the study
of the WAM will be followed by two sections showing how to tweak the
WAM in order to obtain (simplifications of) two CBNeed machines in the
literature, Cregut’s Lazy KAM and Sestoft’s machine (here called SAM ).
Expressing the Lazy KAM and the SAM as modifications of the WAM
helps understanding their design, their distillation (that would otherwise
look very technical), and their relationship;
3. Modularity of Our Contextual Theory of Abstract Machines: the WAM is
obtained by applying to the KAM the following two tweaks:
(a) Dump: the WAM uses the dump-like approach of the Split
CEK/SECD to evaluate inside explicit substitutions;
(b) Global Environments: the WAM uses the global environment approach of the MAM to implement memoization;
22
8.1
The WAM
The WAM is shown in Fig. 3. Note that when the code is a variable the transition is now commutative. The idea is that whenever the code is a variable x
and the environment has the form E1 ∶∶ [x t] ∶∶ E2 , the machine jumps to evaluate t saving the prefix of the environment E1 , the variable x on which it will
substitute the result of evaluating t, and the stack π. This is how hereditarily
weak head evaluation context are implemented by the WAM. In Sect. 10, we
will present a variant of the WAM that avoids the splitting of the environment
saving E1 in a dump entry.
The syntax for dumps is
D ∶∶= ǫ ∣ (E, x, π) ∶∶ D
Every dump stack decodes to a context according to:
ǫ ∶=
⟨⋅⟩
(E, x, π) ∶∶ D
∶=
E⟨D⟨π⟨x⟩⟩⟩[x ⟨⋅⟩]
The decoding of terms, environments, and stacks is defined as for the KAM.
The decoding of states is defined by t ∣ π ∣ D ∣ E ∶= E⟨D⟨π⟨t⟩⟩⟩. The decoding of
contexts is static:
Lemma 8.1 (Contextual Decoding). Let D, π, and E be a dump, a stack,
and a global environment of the WAM, respectively. Then D, D⟨π⟩, E⟨D⟩, and
E⟨D⟨π⟩⟩ are CBNeed evaluation contexts.
Closed closures are defined as for the MAM. Given a state s = t ∣ π ∣ D ∣ E0
with D = (E1 , x1 , π1 ) ∶∶ . . . ∶∶ (En , xn , πn ), its closures are (π⟨t⟩, E0 ) and, for
i ∈ {1, . . . , n},
(πi ⟨xi ⟩, Ei ∶∶ [xi πi−1 ⟨xi−1 ⟩] ∶∶ . . . ∶∶ [x1 π⟨t⟩] ∶∶ E0 ).
The dynamic invariants are:
Lemma 8.2 (WAM invariants). Let s = t ∣ π ∣ D ∣ E0 be a WAM reachable state
whose initial code t is well-named, and s.t. D = (E1 , x1 , π1 ) ∶∶ . . . ∶∶ (En , xn , πn ).
Then:
1. Global Closure: the closures of s are closed;
2. Subterm: any code in s is a literal subterm of t;
3. Names: the closures of s are well-named.
For the properties of the decoding function please note that, as defined in
Sect. 2, the structural congruence ≡Need for call-by-need is different from before.
Theorem 8.3 (WAM Distillation). (WAM, Need, ≡Need , ⋅ ) is a reflective distillery. In particular, on a reachable state s we have:
1. Commutative 1: if s →c1 s′ then s = s′ ;
2. Commutative 2: if s →c2 s′ then s = s′ ;
3. Multiplicative: if s →m s′ then s ⊸m ≡Need s′ ;
4. Exponential: if s →e s′ then s ⊸e =α s′ .
23
tu
lx.t
x
v
π
a(u) ∶∶ π
π
h(E1 , x) ∶∶ π
E
E
E1 ∶∶ [x t] ∶∶ E2
E2
→c1
→m
→c2
→e
t
t
t
vα
a(u) ∶∶ π
π
h(E1 , x) ∶∶ π
π
E
[x u] ∶∶ E
E2
E1 ∶∶ [x v] ∶∶ E2
Figure 4: The Merged WAM.
Proof.
1. Commutative 1.
t u ∣ π ∣ D ∣ E = E⟨D⟨π⟨t u⟩⟩⟩ = t ∣ u ∶∶ π ∣ D ∣ E
2. Commutative 2 :
x ∣ π ∣ D ∣ E1 ∶∶ [x t] ∶∶ E2
= E2 ⟨E1 ⟨D⟨π⟨x⟩⟩⟩[x t]⟩
t ∣ ǫ ∣ (E1 , x, π) ∶∶ D ∣ E2
=
3. Multiplicative.
lx.t ∣ u ∶∶ π ∣ D ∣ E
=
E⟨D⟨π⟨(lx.t) u⟩⟩⟩
E⟨D⟨π⟨t[x u]⟩⟩⟩
E⟨D⟨π⟨t⟩⟩[x u]⟩
t ∣ π ∣ D ∣ [x u] ∶∶ E
⊸m
≡Need Lem.
=
2.4
Note that to apply Lemma 2.4 we use the global closure invariant, as u,
being on the stack, is closed by E and so D does not capture its free
variables.
4. Exponential.
v ∣ ǫ ∣ (E1 , x, π) ∶∶ D ∣ E2
= E2 ⟨E1 ⟨D⟨π⟨x⟩⟩⟩[x v]⟩
E2 ⟨E1 ⟨D⟨π⟨v⟩⟩⟩[x v]⟩
E2 ⟨E1 ⟨D⟨π⟨v α ⟩⟩⟩[x v]⟩
v α ∣ π ∣ D ∣ E1 ∶∶ [x v] ∶∶ E2
⊸e
=α
=
Progress. Let s = t ∣ π ∣ D ∣ E be a commutative normal form s.t. s ⊸ u. If t is
1. an application uw. Then a →c1 transition applies and s is not a commutative normal form, absurd.
2. an abstraction v. The decoding s is of the form E⟨D⟨π⟨v⟩⟩⟩. The stack
π and the dump D cannot both be empty, since then s = E⟨v⟩ would be
normal. So either the stack is empty and a →e transition applies, or the
stack is not empty and a →m transition applies.
3. a variable x. By Lemma 8.2.1 it must be bound by E, so a →c2 transition
applies, and s is not a commutative normal form, absurd.
24
tu
lx.t
lx.t
x
v
π
u ∶∶ π
u ∶∶ π
π
ǫ
D
ǫ
(y, π ′ ) ∶∶ D
D
(x, π) ∶∶ D
E
E
E1 ∶∶ [y ◻] ∶∶ E2
E1 ∶∶ [x t] ∶∶ E2
E1 ∶∶ [x ◻] ∶∶ E2
→c1
→m1
→m2
→c2
→e
t
t
t
t
vα
u ∶∶ π
π
π
ǫ
π
Figure 5: The Pointing WAM.
9
The Merged WAM, or Revisiting the Lazy
KAM
Splitting the stack of the CEK machine in two we obtained a simpler form of
the SECD machine. In this section we apply to the WAM the reverse transformation. The result is a machine, deemed Merged WAM, having only one stack
and that can be seen as a simpler version of Cregut’s lazy KAM [Cré07] (but
we are rather inspired by Danvy and Zerny’s presentation in [DZ13]).
To distinguish the two kinds of objects on the stack we use a marker, as for
the CEK and the LAM. Formally, the syntax for stacks is:
π ∶∶= ǫ ∣ a(t) ∶∶ π ∣ h(E, x) ∶∶ π
where a(t) denotes a term to be used as an argument (as for the CEK) and
h(E, x, π) is morally an entry of the dump of the WAM, where however there
is no need to save the current stack. The transitions of the Merged WAM are
in Fig. 4.
The decoding is defined as follows
ǫ
[x t] ∶∶ E
h(E, x) ∶∶ π
a(t) ∶∶ π
t∣π∣E
∶=
∶=
∶=
∶=
∶=
⟨⋅⟩
E⟨⟨⋅⟩[x t]⟩
E⟨π⟨x⟩⟩[x ⟨⋅⟩]
π⟨⟨⋅⟩t⟩
E⟨π⟨t⟩⟩
Lemma 9.1 (Contextual Decoding). Let π and E be a stack and a global environment of the Merged WAM. Then π and E⟨π⟩ are CBNeed evaluation contexts.
The dynamic invariants of the Merged WAM are exactly the same of the
WAM, with respect to an analogous set of closures associated to a state (whose
exact definition is omitted). The proof of the following theorem—almost identical to that of the WAM—is in the appendix (page 60).
Theorem 9.2 (Merged WAM Distillation). (Merged WAM, Need, ≡Need , ⋅ ) is a
reflective distillery. In particular, on a reachable state s we have:
1. Commutative 1: if s →c1 s′ then s = s′ ;
2. Commutative 2: if s →c2 s′ then s = s′ ;
3. Multiplicative: if s →m s′ then s ⊸m ≡Need s′ ;
4. Exponential: if s →e s′ then s ⊸e =α s′ .
25
D
ǫ
(y, π ′ ) ∶∶ D
(x, π) ∶∶ D
D
E
[x u
E1 ∶∶ [y ◻] ∶∶
E1 ∶∶ [x
E1 ∶∶ [x
10
The Pointing WAM, or Revisiting the SAM
In the WAM, the global environment is divided between the environment of
the machine and the entries of the dump. On one hand, this choice makes
the decoding very natural. On the other hand, one would like to keep the
environment in just one place, letting the dump only collect variables and stacks.
This is what we do here, exploiting the fact that variable names can be taken
as pointers (see the discussion after the invariants in Sect. 7.1).
The new machine, called Pointing WAM, is in Fig. 5, and uses a new dummy
constant ◻ for the substitutions whose variable is in the dump. It can be seen as
a simpler version of Sestoft’s abstract machine [Ses97], here called SAM. Dumps
and environments are defined by:
D
E
∶∶=
∶∶=
ǫ ∣ (x, π) ∶∶ D
ǫ ∣ [x t] ∶∶ E ∣ [x ◻] ∶∶ E
A substitution of the form [x ◻] is dumped, and we also say that x is dumped.
Note that the variables of the entries in D appear in reverse order with
respect to the corresponding substitutions in E. We will show that fact is an
invariant, called duality.
Definition 10.1 (Duality ED). Duality ED between environments and
dumps is defined by
1. ǫǫ;
2. E ∶∶ [x t]D if ED;
3. E ∶∶ [x ◻](x, π) ∶∶ D if ED.
Note that in a dual pair the environment is always at least as long as the
dump. A dual pair ED decodes to a context as follows:
(E, ǫ) ∶=
(E ∶∶ [x ◻], (x, π) ∶∶ D) ∶=
(E ∶∶ [x t], (y, π) ∶∶ D) ∶=
E
(E, D)⟨π⟨x⟩⟩[x ⟨⋅⟩]
(E, (y, π) ∶∶ D)[x t]
The analysis of the Pointing WAM is based on a complex invariant that
includes duality plus a generalization of the global closure invariant. We need
an auxiliary definition:
Definition 10.2. Given an environment E, we define its slice E ↿ as the sequence of substitutions after the rightmost dumped substitution. Formally:
ǫ↿
(E ∶∶ [x t])↿
(E ∶∶ [x ◻])↿
∶=
∶=
∶=
ǫ
E ↿∶∶ [x t]
ǫ
Moreover, if an environment E is of the form E1 ∶∶ [x ◻] ∶∶ E2 , we define
E ↿x ∶= E1 ↿∶∶ [x ◻] ∶∶ E2 .
The notion of closed closure with global environment (Sect. 7.1) is extended
to dummy constants ◻ as expected.
Lemma 10.3 (Pointing WAM invariants). Let s = t ∣ E ∣ π ∣ D be a Pointing
WAM reachable state whose initial code t is well-named. Then:
26
1. Subterm: any code in s is a literal subterm of t;
2. Names: the global closure of s is well-named.
3. Dump-Environment Duality:
(a) (π⟨t⟩, E ↿) is closed;
(b) for every pair (x, π ′ ) in D, (π ′ ⟨x⟩, E ↿x ) is closed;
(c) ED holds.
4. Contextual Decoding: (E, D) is a call-by-need evaluation context.
Proof. In the appendix, page 60.
The decoding of a state is defined as t ∣ π ∣ D ∣ E ∶= (E, D)⟨π⟨t⟩⟩.
Theorem 10.4 (Pointing WAM Distillation). (Pointing WAM, Need, ≡Need , ⋅ )
is a reflective distillery. In particular, on a reachable state s we have:
1. Commutative 1 & 2: if s →c1 s′ or s →c2 s′ then s = s′ ;
2. Multiplicative 1 & 2: if s →m1 s′ or s →m2 s′ then s ⊸m ≡Need s′ ;
3. Exponential: if s →e s′ then s ⊸e =α s′ ;
Proof. Properties of the decoding:
1. Conmutative 1. We have
tu ∣ π ∣ D ∣ E
=
(E, D)⟨π⟨t u⟩⟩
= t ∣ u ∶∶ π ∣ D ∣ E
2. Conmutative 2. Note that E2 has no dumped substitutions, since E1 ∶∶
[x ◻] ∶∶ E2 (x, π) ∶∶ D. Then:
x ∣ π ∣ D ∣ E1 ∶∶ [x t] ∶∶ E2
E2 ⟨(E1 , D)⟨π⟨x⟩⟩[x t]⟩
t ∣ ǫ ∣ (x, π) ∶∶ D ∣ E1 ∶∶ [x ◻] ∶∶ E2
=
=
3. Multiplicative, empty dump.
lx.t ∣ u ∶∶ π ∣ ǫ ∣ E
= E⟨π⟨(lx.t) u⟩⟩
E⟨π⟨t[x u]⟩⟩
E⟨π⟨t⟩[x u]⟩
t ∣ π ∣ ǫ ∣ [x u] ∶∶ E
⊸m
≡∗@l
=
Lem. 2.4
4. Multiplicative, non-empty dump.
lx.t ∣ u ∶∶ π ∣ (y, π ′ ) ∶∶ D ∣ E1 ∶∶ [y ◻] ∶∶ E2
E2 ⟨(E1 , D)⟨π ′ ⟨y⟩⟩[y π⟨(lx.t) u⟩]⟩
E2 ⟨(E1 , D)⟨π ′ ⟨y⟩⟩[y π⟨t[x u]⟩]⟩
E2 ⟨(E1 , D)⟨π ′ ⟨y⟩⟩[y π⟨t⟩][x u]⟩
t ∣ π ∣ (y, π ′ ) ∶∶ D ∣ E1 ∶∶ [y ◻] ∶∶ [x u] ∶∶ E2
27
=
⊸m
≡Need
=
Lem. 2.4
5. Exponential.
v ∣ ǫ ∣ (x, π) ∶∶ D ∣ E1 ∶∶ [x ◻] ∶∶ E2
E2 ⟨(E1 , D)⟨π⟨x⟩⟩[x v]⟩
E2 ⟨(E1 , D)⟨π⟨v⟩⟩[x v]⟩
E2 ⟨(E1 , D)⟨π⟨v α ⟩⟩[x v]⟩
v α ∣ π ∣ D ∣ E1 ∶∶ [x v] ∶∶ E2
=
⊸e
=α
=
Progress. Let s = t ∣ π ∣ D ∣ E be a commutative normal form s.t. s ⊸ u. If t is
• an application uw. Then a →c1 transition applies and s is not a commutative normal form, absurd.
• a variable x. By the machine invariant, x must be bound by E ↿. So
E = E1 ∶∶ [x u] ∶∶ E2 , a →c2 transition applies, and s is not a commutative
normal form, absurd.
• an abstraction v. Two cases:
– The stack π is empty. The dump D cannot be empty, since if D = ǫ
we have that s = e⟨v⟩ is normal. So D = (x, π ′ ) ∶∶ D′ . By duality,
E = E1 ∶∶ [x ◻] ∶∶ E2 and a →e transition applies.
– The stack π is non-empty. If the dump D is empty, the first case of
→m applies. If D = (x, π ′ ) ∶∶ D′ , by duality E = E1 ∶∶ [x ◻] ∶∶ E2 and
the second case of →m applies.
11
Distillation Preserves Complexity
Here, for every abstract machine we bound the number of commutative steps
∣ρ∣c in an execution ρ in terms of
1. the number of principal steps ∣ρ∣p ,
2. the size ∣t∣ of the initial code t.
The analysis only concerns the machines, but via the distillation theorems it
expresses the length of the machine executions as a linear function of the length
of the distilled derivations in the calculi. For every distillery, we will prove that
the relationship is linear in both parameters, namely ∣ρ∣c = O(∣t∣ ⋅ ∣ρ∣p ) holds.
Definition 11.1. Let M be a distilled abstract machine and ρ ∶ s →∗ s′ be an
execution of initial code t. M is
1. Globally bilinear if ∣ρ∣c = O((∣t∣ + 1) ⋅ ∣ρ∣p ).
2. Locally linear if whenever s′ →kc s′′ then k = O(∣t∣).
The next lemma shows that local linearity is a sufficient condition for global
bilinearity.
Proposition 11.2 (Locally Linear ⇒ Globally Bilinear). Let M be a locally
linear distilled abstract machine, and ρ an execution of initial code t. Then M is
globally bilinear.
28
Proof. The execution ρ writes uniquely as →kc 1 →hp 1 . . . →kc m →hp m . By hypothesis
ki = O(∣t∣) for every i ∈ {1, . . . , m}. From m ≤ ∣ρ∣p follows that ∣ρ∣c = O(∣t∣ ⋅ ∣ρ∣p ).
We conclude with ∣ρ∣ = ∣ρ∣p + ∣ρ∣c = ∣ρ∣p + O(∣t∣ ⋅ ∣ρ∣p ) = O((∣t∣ + 1) ⋅ ∣ρ∣p ).
Call-by-name and call-by-value machines are easily seen to be locally linear,
and thus globally bilinear.
Theorem 11.3. KAM, MAM, CEK, LAM, and the Split CEK are locally linear,
and so also globally bilinear.
Proof.
1. KAM/MAM. Immediate: →c reduces the size of the code, that is
bounded by ∣t∣ by the subterm invariant.
2. CEK. Consider the following measure for states:
⎧
⎪
⎪∣u∣ + ∣w∣ if π = a(w, e′ ) ∶∶ π ′
#(u ∣ e ∣ π) ∶= ⎨
⎪
otherwise
∣u∣
⎪
⎩
By direct inspection of the rules, it can be seen that both →c1 and →c2
transitions decrease the value of # for CEK states, and so the relation
→c1 ∪ →c2 terminates (on reachable states). Moreover, both ∣u∣ and ∣w∣
are bounded by ∣t∣ by the subterm invariant (Lemma 6.1.2), and so k ≤
2 ⋅ ∣t∣ = O(∣t∣).
3. LAM and Split CEK. Minor variations over the CEK, see the appendix
(page 63).
Call-by-need machines are not locally linear, because a sequence of →c2 steps
(remember →c ∶=→c1 ∪ →c2 ) can be as long as the environment e, that is not
bound by ∣t∣ (as for the MAM). Luckily, being locally linear is not a necessary
condition for global bilinearity. We are in fact going to show that call-by-need
machines are globally bilinear. The key observation is that ∣ρ∣c2 is not only
locally but also globally bound by ∣ρ∣p , as the next lemma formalizes.
We treat the WAM. The reasoning for the Merged WAM and for the Pointing
WAM is analogous. Define ∣ǫ∣ ∶= 0 and ∣(E, x, π) ∶∶ D∣ ∶= 1 + ∣D∣. We have:
Lemma 11.4. Let s = t ∣ π ∣ D ∣ E be a WAM state, reached by the execution ρ.
Then
1. ∣ρ∣c2 = ∣ρ∣e + ∣D∣.
2. ∣E∣ + ∣D∣ ≤ ∣ρ∣m
3. ∣ρ∣c2 ≤ ∣ρ∣e + ∣ρ∣m = ∣ρ∣p
Proof.
1. Immediate, as →c2 is the only transition that pushes elements on
D and →e is the only transition that pops them.
2. The only rule that produces substitutions is →m . Note that 1) →c2 and
→e preserve the global number of substitutions in a state; 2) E and D
are made out of substitutions, if one considers every entry (E, x, π) of the
dump as a substitution on x (and so the statement follows); 3) the inequality is given by the fact that an entry of the dump stocks an environment
(counting for many substitutions).
29
3. Substitute Point 2 in Point 1.
Theorem 11.5. The WAM has globally linear commutations.
Proof. Let ρ be an execution of initial code t. Define →¬c1 ∶=→e ∪ →m ∪ →c2 and
note ∣ρ∣¬c1 the number of its steps in ρ. We estimate →c ∶=→c1 ∪ →c2 by studying
its components separately. For →c2 , Lemma 11.4.3 proves ∣ρ∣c2 ≤ ∣ρ∣p = O(∣ρ∣p ).
For →c1 , as for the KAM, the length of a maximal →c1 subsequence of ρ is
bounded by ∣t∣. The number of →c1 maximal subsequences of ρ is bounded
by ∣ρ∣¬c1 , that by Lemma 11.4.3 is linear in O(∣ρ∣p ). Then ∣ρ∣c1 = O(∣t∣ ⋅ ∣ρ∣p ).
Summing up,
∣ρ∣c2 + ∣ρ∣c1 = O(∣ρ∣p ) + O(∣t∣ ⋅ ∣ρ∣p ) = O((∣t∣ + 1) ⋅ ∣ρ∣p )
The analysis presented here is complemented by the study in [ASC14], where
the number of exponential steps ⊸e in a derivation d is shown to be polynomial
(actually quadratic in call-by-name and linear in call-by-value/need) in terms
of the number of multiplicative steps ⊸m in d. Given our distillation theorems,
the results in [ASC14] equivalently relate the exponential and multiplicative
transitions of the abstract machines studied in this paper. Note that these
derived results could very hardly be obtained directly on the machines, i.e.
without distillation.
12
Conclusions
The novelty of our study is the use of the linear substitution calculus (LSC) to
discriminate between abstract machine transitions: some of them—the principal
ones—are simulated, and thus shown to be logically relevant, while the others—
the commutative ones—are sent on the structural congruence and have to be
considered as bookkeeping operations. On one hand, the LSC is a sharp tool
to study abstract machines. On the other hand, it provides an alternative to
abstract machines which is simpler while being conservative at the level of
complexity analysis.
Acknowledgments
A special acknowledgement to Claudio Sacerdoti Coen, for many useful discussions, comments and corrections to the paper. In particular, we owe him
the intuition that a global analysis of call-by-need commutative rules may
provide a linear bound. This work was partially supported by the ANR
projects Logoi (10-BLAN-0213-02) and Coquas (ANR-12-JS02-006-01), by
the French-Argentinian Laboratory in Computer Science INFINIS, the FrenchArgentinian project ECOS-Sud A12E04, the Qatar National Research Fund
under grant NPRP 09-1107-1-168.
References
[ABDM03] Mads Sig Ager, Dariusz Biernacki, Olivier Danvy, and Jan Midtgaard. A functional correspondence between evaluators and abstract machines. In PPDP, pages 8–19, 2003.
30
[ABKL14]
Beniamino Accattoli, Eduardo Bonelli, Delia Kesner, and Carlos
Lombardi. A nonstandard standardization theorem. In POPL,
pages 659–670, 2014.
[ABM14]
Beniamino Accattoli, Pablo Barenbaum, and Damiano Mazza. Dinstilling Abstract Ma chines (Long Version). bla boa boa, 2014.
[ABS09]
Zena M. Ariola, Aaron Bohannon, and Amr Sabry. Sequent calculi
and abstract machines. ACM Trans. Program. Lang. Syst., 31(4),
2009.
[Acc12]
Beniamino Accattoli. An abstract factorization theorem for explicit
substitutions. In RTA, pages 6–21, 2012.
[Acc13]
Beniamino Accattoli. Evaluating functions as processes. In TERMGRAPH, pages 41–55, 2013.
[ADL12]
Beniamino Accattoli and Ugo Dal Lago. On the invariance of the
unitary cost model for head reduction. In RTA, pages 22–37, 2012.
[ADL14]
Beniamino Accattoli and Ugo Dal Lago. Beta Reduction is Invariant, Indeed. Accepted to LICS/CSL 2014, 2014.
[ADM04]
Mads Sig Ager, Olivier Danvy, and Jan Midtgaard. A functional
correspondence between call-by-need evaluators and lazy abstract
machines. Inf. Process. Lett., 90(5):223–232, 2004.
[AF97]
Zena M. Ariola and Matthias Felleisen. The call-by-need lambda
calculus. J. Funct. Program., 7(3):265–301, 1997.
[AK10]
Beniamino Accattoli and Delia Kesner. The structural λ-calculus.
In CSL, pages 381–395, 2010.
[AP12]
Beniamino Accattoli and Luca Paolini. Call-by-value solvability,
revisited. In FLOPS, pages 4–16, 2012.
[ASC14]
Beniamino Accattoli and Claudio Sacerdoti Coen. On the Value of
Variables. Accepted to WOLLIC 2014, 2014.
[BD07a]
Malgorzata Biernacka and Olivier Danvy. A concrete framework
for environment machines. ACM Trans. Comput. Log., 9(1), 2007.
[BD07b]
Malgorzata Biernacka and Olivier Danvy. A syntactic correspondence between context-sensitive calculi and abstract machines.
Theor. Comput. Sci., 375(1-3):76–108, 2007.
[CF12]
Stephen Chang and Matthias Felleisen. The call-by-need lambda
calculus, revisited. In ESOP, pages 128–147, 2012.
[Cla11]
Pierre Clairambault. Estimation of the length of interactions in
arena game semantics. In FOSSACS, pages 335–349, 2011.
[Cré07]
Pierre Crégut. Strongly reducing variants of the Krivine abstract
machine. Higher-Order and Symbolic Computation, 20(3):209–230,
2007.
31
[Dan04]
Olivier Danvy. A rational deconstruction of landin’s secd machine.
In IFL, pages 52–71, 2004.
[dB87]
Nicolaas G. de Bruijn. Generalizing Automath by Means of a
Lambda-Typed Lambda Calculus. In Mathematical Logic and Theoretical Computer Science, number 106 in Lecture Notes in Pure
and Applied Mathematics, pages 71–92. Marcel Dekker, 1987.
[DCKP03] Roberto Di Cosmo, Delia Kesner, and Emmanuel Polonovski.
Proof nets and explicit substitutions. Math. Str. in Comput. Sci.,
13(3):409–450, 2003.
[DHR96]
Vincent Danos, Hugo Herbelin, and Laurent Regnier. Game semantics & abstract machines. In LICS, pages 394–405, 1996.
[DN04]
Olivier Danvy and Lasse R. Nielsen. Refocusing in reduction semantics. Technical Report RS-04-26, BRICS, 2004.
[DR04]
Vincent Danos and Laurent Regnier. Head linear reduction. Technical report, 2004.
[DZ13]
Olivier Danvy and Ian Zerny. A synthetic operational account of
call-by-need evaluation. In PPDP, pages 97–108, 2013.
[ER06]
Thomas Ehrhard and Laurent Regnier. Böhm trees, Krivine’s machine and the Taylor expansion of lambda-terms. In CiE, pages
186–197, 2006.
[FF86]
Matthias Felleisen and Daniel P. Friedman. Control operators, the
SECD-machine, and the lambda-calculus. In 3rd Working Conference on the Formal Description of Programming Concepts, August
1986.
[GLS09]
Ronald Garcia, Andrew Lumsdaine, and Amr Sabry. Lazy evaluation and delimited control. In POPL, pages 153–164, 2009.
[HM98]
Thérèse Hardin and Luc Maranget. Functional runtime systems
within the lambda-sigma calculus. J. Funct. Program., 8(2):131–
176, 1998.
[Kes09]
Delia Kesner. A theory of explicit substitutions with safe and full
composition. Logical Methods in Computer Science, 5(3), 2009.
[KL07]
Delia Kesner and Stéphane Lengrand. Resource operators for
lambda-calculus. Inf. Comput., 205(4):419–473, 2007.
[KR09]
Delia Kesner and Fabien Renaud. The prismoid of resources. In
MFCS, pages 464–476, 2009.
[Kri07]
Jean-Louis Krivine. A call-by-name lambda-calculus machine.
Higher-Order and Symbolic Computation, 20(3):199–207, 2007.
[Lan64]
Peter John Landin. The Mechanical Evaluation of Expressions.
The Computer Journal, 6(4):308–320, January 1964.
32
[Lan07]
Frédéric Lang. Explaining the lazy Krivine machine using explicit
substitution and addresses. Higher-Order and Symbolic Computation, 20(3):257–270, 2007.
[Ler90]
Xavier Leroy. The ZINC experiment: an economical implementation of the ML language. Technical report 117, INRIA, 1990.
[Mil07]
Robin Milner. Local bigraphs and confluence: Two conjectures.
Electr. Notes Theor. Comput. Sci., 175(3):65–73, 2007.
[MOTW99] John Maraist, Martin Odersky, David N. Turner, and Philip
Wadler. Call-by-name, call-by-value, call-by-need and the linear
lambda calculus. Theor. Comput. Sci., 228(1-2):175–210, 1999.
[MOW98]
John Maraist, Martin Odersky, and Philip Wadler. The call-byneed lambda calculus. J. Funct. Program., 8(3):275–317, 1998.
[MP94]
Gianfranco Mascari and Marco Pedicini. Head linear reduction and
pure proof net extraction. Theor. Comput. Sci., 135(1):111–137,
1994.
[Ned92]
Robert. P. Nederpelt. The fine-structure of lambda calculus. Technical Report CSN 92/07, Eindhoven Univ. of Technology, 1992.
[Plo75]
Gordon D. Plotkin. Call-by-name, call-by-value and the lambdacalculus. Theor. Comput. Sci., 1(2):125–159, 1975.
[Ses97]
Peter Sestoft. Deriving a lazy abstract machine. J. Funct. Program,
7(3):231–264, 1997.
13
13.1
Technical Appendix: proofs of the determinism of the calculi (Proposition 2.1)
Call-by-Name
Let t = H1 ⟨r1 ⟩ = H2 ⟨r2 ⟩. By induction on the structure of t. Cases:
• Variable or an abstraction. Vacuously true, because there is no redex.
• Application. Let t = uw. Suppose that one of the two evaluation contexts, for instance H1 , is equal to ⟨⋅⟩. Then, we must have u = lx.u′ , but
in that case it is easy to see that the result holds, because H2 cannot have
its hole to the right of an application (in w) or under an abstraction (in
u′ ). We may then assume that none of H1 , H2 is equal to ⟨⋅⟩. In that case,
we must have H1 = H1′ w and H2 = H2′ w, and we conclude by induction
hypothesis.
• Substitution. Let t = u[x w]. This case is entirely analogous to the
previous one.
33
13.2
Left-to-Right Call-by-Value
We prove the following statement, of which the determinism of the reduction is
a consequence.
Lemma 13.1. Let t be a term. Then t has at most one subterm u that verifies
both (i) and (ii):
(i) Either u is a variable x, or u is an application L⟨v⟩L′ ⟨v ′ ⟩, for v, v ′ being
values.
(ii) u is under a left-to-right call-by-value evaluation context, i.e. t = V ⟨u⟩.
From the statement it follows that there is at most one ⊸-redex in t, i.e. ⊸
is deterministic.
Proof. by induction on the structure of t:
• t is a variable. There is only one subterm, under the empty evaluation
context.
• t is an abstraction. There are no subterms that verify both (i) and (ii),
since the only possible evaluation context is the empty one.
• t is an application w r. There are three possible situations:
– The left subterm w is not of the form L⟨v⟩. Then u cannot be at the
root, i.e. u ≠ t. Since w⟨⋅⟩ is not an evaluation context, u must be
internal to ⟨⋅⟩r, which is an evaluation context. We conclude by i.h..
– The left subterm w is of the form L⟨v⟩ with v a value, but the right
subterm r is not. Then u cannot be a subterm of w, and also u ≠ t.
Hence, if there is a subterm u as in the statement, it must be internal
to the evaluation context w⟨⋅⟩. We conclude by i.h..
– Both subterms have that form, i.e. w = L⟨v⟩ and r = L′ ⟨v ′ ⟩ with v and
v ′ values. The only subterm that verifies both (i) and (ii) is u = t.
• t is a substitution w[x r]. Any occurrence of u must be internal to w
(because w[x ⟨⋅⟩] is not an evaluation context). We conclude by i.h. that
there is at most one such occurrence.
13.3
Right-to-Left Call-by-Value
Exactly as in the case for left-to-right call-by-value, we prove the following
property, from which determinism of the reduction follows.
Lemma 13.2. Let t be a term. Then t has at most one subterm u that verifies
both (i) and (ii):
(i) u is either a variable x or an application L⟨v⟩L′ ⟨v ′ ⟩, where v and v ′ are
values.
(ii) u is under a right-to-left call-by-value evaluation context, i.e. t = S⟨u⟩.
34
As a corollary, any term t has at most one ⊸-redex.
Proof. By induction on the structure of t:
• Variable or abstraction. Immediate.
• Application. If t = w r, there are three cases:
– The right subterm r is not of the form L′ ⟨v ′ ⟩. Then u cannot be at
the root. Since ⟨⋅⟩ r is not an evaluation context, u must be internal
to r and we conclude by i.h..
– The right subterm r is of the form L′ ⟨v ′ ⟩ but the left subterm w is not.
Again u cannot be at the root. Moreover, r has no applications or
variables under an evaluation context. Therefore u must be internal
to w and we conclude by i.h..
– Both subterms have that form, i.e. w = L⟨v⟩ and r = L′ ⟨v ′ ⟩. We
first note that w and r have no applications or variables under an
evaluation context. The only possibility that remains is that u is at
the root, i.e. u = t.
• Substitution. If t = w[x r] is a substitution, u must be internal to w
(because w[x ⟨⋅⟩] is not an evaluation context), and we conclude by i.h..
13.4
Call-by-Need
We first need an auxiliary result:
Lemma 13.3. Let t ∶= N ⟨x⟩ for an evaluation context N such that x ∈ fv(t).
Then:
1. for every substitution context L and abstraction v, t ≠ L⟨v⟩;
2. for every evaluation context N ′ and variable y, t = N ′ ⟨y⟩ implies N ′ = N
and y = x;
3. t is a call-by-need normal form.
Proof. In all points we use a structural induction on N . For point 1:
• N = ⟨⋅⟩: obvious.
• N = N1 u: obvious.
• N = N1 [y w]: suppose that L = L′ [y w] (for otherwise the result is
obvious); then we apply the induction hypothesis to N1 to obtain N1 ⟨x⟩ ≠
L′ ⟨v⟩.
• N = N1 ⟨y⟩[y N2 ]: suppose that L = L′ [y N2 ⟨x⟩] (for otherwise the
result is obvious); then we apply the induction hypothesis to N1 to obtain
N1 ⟨y⟩ ≠ L′ ⟨v⟩.
For point 2:
• N = ⟨⋅⟩: obvious.
35
• N = N1 u: we must necessarily have N ′ = N1′ u and we conclude by induction hypothesis.
• N = N1 [z u]: in principle, there are two cases. First, we may have
N ′ = N1′ [z u], which allows us to conclude immediately by induction hypothesis, as above. The second possibility would be N ′ = N1′ ⟨z⟩[z N2′ ],
with N2′ ⟨y⟩ = u, but this is actually impossible. In fact, it would imply N1 ⟨x⟩ = N1′ ⟨z⟩, which by induction hypothesis would give us z = x,
contradicting the hypothesis x ∈ fv(t).
• N = N1 ⟨z⟩[z N2 ]: by symmetry with the above case, the only possibility is N ′ = N1 ⟨z⟩[z N2′ ], which allows us to conclude immediately by
induction hypothesis.
For point 3, let r be a redex (i.e., a term matching the left hand side of
↦dB or ↦lsv ) and let N ′ be an evaluation context. We will show by structural
induction on N that t ≠ N ′ ⟨r⟩. We will do this by considering, in each inductive
case, all the possible shapes of N ′ .
• N = ⟨⋅⟩: obvious.
• N = N1 u: the result is obvious unless N ′ = ⟨⋅⟩ or N ′ = N1′ u. In the latter
case, we conclude by induction hypothesis (on N1 ). In the former case,
since r is a redex, we are forced to have r = L⟨v⟩u′ for some abstraction v,
substitution context L and term u′ . Now, even supposing u′ = u, we are
still allowed to conclude because N1 ⟨x⟩ ≠ L⟨v⟩ by point 1.
• N = N1 [y u]: the result is obvious unless:
– N ′ = ⟨⋅⟩: this time, the fact that r is a redex forces r = N1′ ⟨y⟩[y u].
Even if we admit that u = L⟨v⟩, we may still conclude because x ≠ y
(by the hypothesis x ∈ fv(t)), hence N1 ⟨x⟩ ≠ N1′ ⟨y⟩ by point 2.
– N ′ = N1′ [y u]: immediate by induction hypothesis on N1 .
– N ′ = N1′ ⟨y⟩[y N2′ ]: even if N2′ ⟨r⟩ = u, we may still conclude because,
again, x ≠ y implies N1 ⟨x⟩ ≠ N1′ ⟨y⟩ by point 2.
• N = N1 ⟨y⟩[y N2 ]: again, the result is obvious unless:
– N ′ = ⟨⋅⟩: the fact that r is a redex implies r = N1′ ⟨y⟩[y L⟨v⟩]. Even
assuming N1′ = N1 , we may still conclude because N2 ⟨x⟩ ≠ L⟨v⟩ by
point 1.
– N ′ = N1′ [y N2 ⟨x⟩]: since y ∈ fv(N1 ⟨y⟩), we conclude because the
induction hypothesis gives us N1 ⟨y⟩ ≠ N1′ ⟨r⟩.
– N ′ = N1 ⟨y⟩[y N2′ ]: we conclude at once by applying the induction
hypothesis to N2 .
Now, the proof of Proposition 2.1 is by structural induction on t ∶= N1 ⟨r1 ⟩ =
N2 ⟨r2 ⟩. Cases:
• Variable or abstraction. Impossible, since variables and abstractions
are both call-by-need normal.
36
• Application, i.e. t = uw. This case is treated exactly as in the corresponding case of the proof of Proposition 2.1.
• Substitution, i.e. t = u[x w]. Cases:
– Both contexts have their holes in u or w. It follows from the i.h..
– One of the contexts—say N1 —is empty, i.e. u = N3 ⟨x⟩, w = L⟨v⟩,
and r1 = N3 ⟨x⟩[x L⟨v⟩]. This case is impossible. Indeed, 1) the
hole of N2 cannot be in L⟨v⟩, because it is call-by-need normal, and
2) it cannot be inside N3 ⟨x⟩ because by Lemma 13.3.3 N ⟨x⟩ is callby-need normal.
– One of the contexts—say N1 —has its hole in w and the other one
has its hole in u, i.e. N1 = N3 ⟨x⟩[x N4 ] and N2 = N5 [x w]. This
case is impossible, because by Lemma 13.3.3 N3 ⟨x⟩ is call-by-need
normal.
14
14.1
Technical Appendix: proofs of strong bisimulation
Proof of Proposition 2.2 (≡ is a strong bisimulation)
for call-by-name
Before proving the main result, we need two auxiliary lemmas, proved by
straightforward inductions on H:
Lemma 14.1. Let t be a term, H be a call-by-name evaluation context not capturing any variable in fv(t), and x ∉ fv(H⟨y⟩). Then H⟨t[x u]⟩ ≡ H⟨t⟩[x u].
Lemma 14.2. The equivalence relation ≡ as defined for call-by-name preserves
the shape of H⟨x⟩. More precisely, if H⟨x⟩ ≡ t, with x not captured by H, then
t is of the form H ′ ⟨x⟩, with x not captured by H ′ .
Now we turn to the proof of Proposition 2.2 itself.
Let ⇚
⇛ be the symmetric closure of the union of the axioms defining ≡ for
call-by-name, that is of ≡gc ∪ ≡dup ∪ ≡@ ∪ ≡com ∪ ≡[⋅] . Note that ≡ is the
reflexive–transitive closure of ⇚
⇛. The proof is in two parts:
(I) Prove the property holds for ⇚
⇛, i.e. if t ⊸a u and t ⇚
⇛ w, there exists
r s.t. w ⊸a r and u ≡ r.
(II) Prove the property holds for ≡ (i.e. for many steps of ⇚
⇛) by resorting
to (I).
The proof of (II) is immediate by induction on the number of ⇚
⇛ steps. The
proof of (I) goes by induction on the rewriting step ⊸ (that, since ⊸ is closed
by evaluation contexts, becomes a proof by induction on the evaluation context
H). In principle, we should always consider the two directions of ⇚
⇛. Most of
the time, however, one direction is obtained by simply reading the diagram of
the other direction bottom-up, instead than top-down; these cases are simply
omitted, we distinguish the two directions only when it is relevant.
37
1. Base case 1: multiplicative root step t = L⟨lx.t′ ⟩u′ ↦dB L⟨t′ [x u′ ]⟩ =
u.
If the ⇚
⇛ step is internal to u′ or internal to one of the substitutions in L,
the pattern of the ⇚
⇛ redex does not overlap with the ↦dB step, and the
proof is immediate, the two steps commute. Otherwise, we consider every
possible case for ⇚
⇛:
(a) Garbage Collection ≡gc . The garbage collected substitution must
be one of the substitutions in L, i.e. L must be of the form
̂ ∶= L′ ⟨L′′ ⟩. Then:
L′ ⟨L′′ [y w′ ]⟩. Let L
L⟨lx.t′ ⟩u′
dB
L⟨t′ [x u′ ]⟩
≡gc
≡gc
′ ′
̂
L⟨lx.t
⟩u
dB
̂ ′ [x u′ ]⟩
L⟨t
(b) Duplication ≡dup . The duplicated substitution must be one of the
substitutions in L, i.e. L must be of the form L′ ⟨L′′ [y w′ ]⟩. Then:
dB
L′ ⟨L′′ ⟨lx.t′ ⟩[y w′ ]⟩u′
≡dup
t1
≡dup
dB
t2
t3
where
t1 ∶= L′ ⟨L′′ ⟨t′ [x u′ ]⟩[y w′ ]⟩,
t2 ∶= L′ ⟨ (L′′ ⟨lx.t′ ⟩)[z]y [y w′ ][z w′ ] ⟩u′ ,
t3 ∶= L′ ⟨ (L′′ ⟨t′ [x u′ ]⟩)[z]y [y w′ ][z w′ ] ⟩.
(c) Commutation with application ≡@ . Here ≡@ can only be applied in
one direction. The diagram is:
L⟨ly.t′ ⟩[x q ′ ]u′ [x q ′ ]
≡@
t1
dB
dB
t4
t2
=α
≡dup
t3
t5
≡∗com
t6
≡[⋅]
38
where
t1 ∶= (L⟨ly.t′ ⟩u′ )[x q ′ ],
t2 ∶= (L⟨t′ [y u′ ]⟩)[x q ′ ],
t3 ∶= (L⟨t′ [y u′ {x y}]⟩)[x q ′ ][y q ′ ],
t4 ∶= L⟨t′ [y u′ [x q ′ ]]⟩[x q ′ ],
t5 ∶= (L⟨t′ [y u′ {x y}[y q ′ ]]⟩)[x q ′ ],
t6 ∶= (L⟨t′ [y u′ {x y}][y q ′ ]⟩)[x q ′ ].
(d) Commutation of independent substitutions ≡com . The substitutions
that are commuted by the ≡com rule must be both in L, i.e. L
must be of the form L′ ⟨L′′ [y w′ ][z r′ ]⟩ with z ∈/ fv(w′ ). Let
̂ = L′ ⟨L′′ [z r′ ][y w′ ]⟩. Then:
L
L⟨lx.t′ ⟩u′
dB
≡com
L⟨t′ [x u′ ]⟩
≡com
′ ′
̂
L⟨lx.t
⟩u
dB
̂ ′ [x u′ ]⟩
L⟨t
(e) Composition of substitutions ≡[⋅] . The substitutions that appear in
the left-hand side of the ≡[⋅] rule must both be in L, i.e. L must
̂=
be of the form L′ ⟨L′′ [y w′ ][z r′ ]⟩ with z /∈ fv(L′′ ⟨lx.t′ ⟩). Let L
L′ ⟨L′′ [y w′ [z r′ ]]⟩. Exactly as in the previous case:
L⟨lx.t′ ⟩u′
dB
≡[⋅]
L⟨t′ [x u′ ]⟩
≡[⋅]
′ ′
̂
L⟨lx.t
⟩u
dB
̂ ′ [x u′ ]⟩
L⟨t
2. Base case 2:
exponential root step t = H ′ ⟨x⟩[x t′ ] ↦ls
′ ′
′
H ⟨t ⟩[x t ] = u. If the ⇚
⇛ step is internal to t′ , the proof is immediate, since there is no overlap with the pattern of the ↦ls redex. Similarly,
if the ⇚
⇛ step is internal to H⟨x⟩, the proof is straightforward by resorting
to Lemma 14.2.
Now we proceed by case analysis on the ⇚
⇛ step:
(a) Garbage collection ≡gc . Note that ≡gc cannot remove [x t′ ], because
by hypothesis x does occur in its scope. If the removed substitution
̂′ ∶= H ′′ ⟨H ′′′ ⟩. Then:
belongs to H ′ , i.e. H ′ = H ′′ ⟨H ′′′ [y u′ ]⟩. Let H
H ′ ⟨x⟩[x t′ ]
ls
≡gc
H ′ ⟨t′ ⟩[x t′ ]
≡gc
̂′ ⟨x⟩[x t′ ]
H
ls
̂′ ⟨t′ ⟩[x t′ ]
H
If ≡gc adds a substitution as topmost constructor the diagram is
analogous.
(b) Duplication ≡dup . Two sub-cases:
39
i. The equivalence ≡dup acts on a substitution internal to H ′ . This
case goes as for Garbage collection.
ii. The equivalence ≡dup acts on [x t′ ]. There are two further subcases:
• The substituted occurrence is renamed by ≡dup :
H ′ ⟨x⟩[x t′ ]
ls
H ′ ⟨t′ ⟩[x t′ ]
≡gc
≡gc
′
H[y]
⟨y⟩[x t′ ][y t′ ]
x
ls
t1
′
′
where t1 ∶= H[y]
⟨t′ ⟩[x t′ ][y t′ ] and H[y]
is the context
x
x
′
obtained from H by renaming some (possibly none) occurrences of x as y.
• The substituted occurrence is not renamed by ≡dup . Essentially as in the previous case:
H ′ ⟨x⟩[x t′ ]
ls
H ′ ⟨t′ ⟩[x t′ ]
≡dup
where t1 ∶=
≡dup
′
H[y]
⟨x⟩[x
x
t ][y t ]
′
H[y]
⟨t′ ⟩[x
x
t ][y t ].
′
ls
′
′
t1
′
(c) Commutation with application ≡@ . Two sub-cases:
i. The equivalence ≡@ acts on a substitution internal to H ′ . This
case goes as for Garbage collection.
ii. The equivalence ≡@ acts on [x t′ ]. It must be the case that H ′
is of the form H ′′ u′ . Then:
ls
(H ′′ ⟨x⟩u′ )[x t′ ]
≡@
t1
≡@
ls
t2
t3
where
t1 ∶= (H ′′ ⟨t′ ⟩u′ )[x t′ ],
t2 ∶= H ′′ ⟨x⟩[x t′ ]u′ [x t′ ],
t3 ∶= H ′′ ⟨t′ ⟩[x t′ ]u′ [x t′ ].
(d) Commutation of independent substitutions ≡com . Two sub-cases:
i. The equivalence ≡com acts on two substitutions internal to H ′ .
This case goes as for Garbage collection.
ii. The equivalence ≡com acts on [x t′ ]. It must be the case that
H ′ is of the form H ′′ . Then:
H ′′ ⟨x⟩[y u′ ][x t′ ]
ls
≡com
H ′′ ⟨t′ ⟩[y u′ ][x t′ ]
≡com
H ′′ ⟨x⟩[x t′ ][y u′ ]
40
ls
H ′′ ⟨t′ ⟩[x t′ ][y u′ ]
(e) Composition of substitutions ≡[⋅] . Two sub-cases:
i. The equivalence ≡[⋅] acts on two substitutions internal to H ′ .
This case goes as for Garbage collection.
ii. The equivalence ≡[⋅] acts on [x t′ ]. Note that the equivalence
≡[⋅] cannot be applied from left to right to [x t′ ], because H ′ ⟨x⟩
must be of the form H ′′ ⟨x⟩[y u′ ] with x ∉ fv(H ′′ ⟨x⟩), which is
clearly not possible. It can be applied from right to left. The
diagram is:
H ′ ⟨x⟩[x t′ [y u]]
≡@
t1
ls
ls
t4
t2
≡ by Lemma 14.1
t5
≡dup
t3
≡[⋅]
t6
≡com
=α
t7
where
t1 ∶= H ′ ⟨x⟩[x t′ ][y u],
t2 ∶= H ′ ⟨t′ ⟩[x t′ ][y u],
t3 ∶= H ′ ⟨t′ {y z}⟩[x t′ ][z u][y u],
t4 ∶= H ′ ⟨t′ [y u]⟩[x t′ [y u]],
t5 ∶= H ′ ⟨t′ ⟩[y u][x t′ [y u]],
t6 ∶= H ′ ⟨t′ ⟩[y u][x t′ ][y u],
t7 ∶= H ′ ⟨t′ {y z}⟩[z u][x t′ ][y u].
3. Inductive case 1: left of an application H = H ′ q. The situation is:
t = t′ q ⊸a u′ q = u
for terms t′ , u′ such that either t′ ⊸m u′ or t′ ⊸e u′ . Two sub-cases:
(a) The t ⇚
⇛ w step is internal to t′ . The proof simply uses the i.h.
applied to the (strictly smaller) evaluation context of the step t′ ⊸a
u′ .
(b) The t ⇚
⇛ w step involves the topmost application. The ⇚
⇛ step can
only be a commutation with the root application. Moreover, for t′ q
to match with the right-hand side of the ≡@ rule, t′ must have the
form w′ [x r′ ] and q the form q ′ [x r′ ], so that the ⇚
⇛ is:
w = (w′ q ′ )[x r′ ] ≡@ w′ [x r′ ]q ′ [x r′ ] = t
Three sub-cases:
i. The rewriting step is internal to w′ . Then the two steps trivially
commute. Let a ∈ {dB, ls}:
41
w′ [x r′ ]q ′ [x r′ ]
a
w′′ [x r′ ]q ′ [x r′ ]
a
(w q )[x r′ ]
≡@
≡@
(w q )[x r ]
′ ′
′
′′ ′
ii. dB-step not internal to w′ . Exactly as the multiplicative root
case 1c (read in the other direction).
(c) ls-step not internal to w′ . Not possible: the topmost constructor is
an application, consequently any ⊸e has to take place in w′ .
4. Inductive case 2: left of a substitution H = H ′ [x q]. The situation
is:
t = t′ [x q] ⊸ u′ [x q] = u
with t′ = H ′ ⟨t′′ ⟩. If the ⇚
⇛ step is internal to H ′ ⟨t′ ⟩, the proof we conclude
using the i.h.. Otherwise:
(a) Garbage Collection ≡gc . If the garbage collected substitution is [x q]
then:
t′ [x q]
u′ [x q]
≡gc
≡gc
t′
u′
If the substitution is introduced out of the blue, i.e. t′ [x q] ≡gc
t′ [x q][y q ′ ] or t′ [x q] ≡gc t′ [y q ′ ][x q] the diagram is analogous.
(b) Duplication ≡dup . If the duplicated substitution is [x q] then:
t′ [x q]
u′ [x q]
≡dup
t′[y]x [x
≡dup
q][y q]
u′[y]x [x
q]
If duplication is applied in the other direction, i.e. t′ = t′′ [y q] and
t′ [x q] = t′′ [y q][x q] ≡dup t′′ {y x}[x q] = t′ [x q]
the interesting case is when t′′ = H ′′ ⟨y⟩ and the step is exponential:
H ′′ ⟨y⟩[y q][x q]
ls
H ′′ ⟨q⟩[y q][x q]
≡dup
≡dup
H ′′ ⟨x⟩{y x}[x q]
ls
H ′′ ⟨q⟩{y x}[x q]
If t′ is H ′′ ⟨x⟩ it is an already treated base case and if t′ has another
form the rewriting step does not interact with the duplication, and
so they simply commute.
(c) Commutation with application ≡@ . Then t′ = t′′ u′′ . Three sub-cases:
i. The ⊸ step is internal to t′′ . Then:
42
(t′′ u′′ )[x q]
(t′′′ u′′ )[x q]
≡@
≡@
t [x q]u [x q]
t [x q]u′′ [x q]
′′
′′
′′′
ii. The ⊸ step is a multiplicative step. If t′′ = L⟨ly.t′′′⟩ then it goes
like the diagram of the multiplicative root case 1c (read in the
other direction).
iii. The ⊸ step is an exponential step. Then it must be [x q] that
substitutes on the head variable, but this case has already been
treated as a base case (case 2c).
(d) Commutation of independent substitutions ≡com . It must be t′ =
t′′ [y q ′ ] with x ∉ fv(q ′ ), so that t′′ [y q ′ ][x q] ≡com t′′ [x q][y q ′ ].
Three sub-cases:
i. Reduction takes place in t′′ . Then reduction and the equivalence
simply commute, as in case 4(c)i.
ii. Exponential steps involving [x q]. This is an already treated
base case (case 2(d)ii).
iii. Exponential step involving [y q ′ ]. This case is solved reading
bottom-up the diagram of case 2(d)ii.
(e) Composition of substitutions ≡[⋅] . It must be t′ = t′′ [y q ′ ] with x ∉
fv(t′′ ), so that t′′ [y q ′ ][x q] ≡[⋅] t′′ [x q[y q ′ ]]. Three sub-cases:
i. Reduction takes place in t′′ . Then reduction and the equivalence
simply commute, as in case 4(c)i.
ii. Exponential steps involving [x q]. This case is solved reading
bottom-up the diagram of case 2(e)ii.
iii. Exponential step involving [y q ′ ]. Impossible, because by hypothesis x ∉ fv(t′′ ).
14.2
Proof of Proposition 2.2 (≡ is a Strong Bisimulation)
for Left-to-Right Call-by-Value
We follow the structure of the proof in Sect. 14.1 for call-by-name. Structural
equivalence for call-by-value is defined exactly in the same way.
Before proving the main result, we need the following auxiliary lemmas,
proved by straightforward inductions on the contexts. Lemma 14.3.2 is the
adaptation of Lemma 14.2 already stated for call-by-name:
Lemma 14.3. The equivalence relation ≡ preserves the “shapes” of L⟨v⟩ and
V ⟨x⟩. Formally:
1. If L⟨v⟩ ≡ t, then t is of the form L′ ⟨v ′ ⟩.
2. If V ⟨x⟩ ≡ t, with x not bound by V , then t is of the form V ′ ⟨x⟩, with x
not bound by V ′ .
Lemma 14.4. L⟨t[x u]⟩ ≡ L⟨t[x L⟨u⟩]⟩
43
Proof. By induction on L. The base case is trivial. For L = L′ ⟨⋅⟩[y w], by i.h.
we have:
L′ ⟨t[x u]⟩[y w] ≡ L′ ⟨t[x L′ ⟨u⟩]⟩[y w]
Let (L′ ⟨u⟩)[z]y be the result of
Then:
L′ ⟨t[x
≡dup L′ ⟨t[x
≡∗com L′ ⟨t[x
≡[⋅]
L′ ⟨t[x
=α
L′ ⟨t[x
replacing all occurrences of y by z in L′ ⟨u⟩.
L′ ⟨u⟩]⟩[y w]
(L′ ⟨u⟩)[z]y ]⟩[y w][z w]
(L′ ⟨u⟩)[z]y ][z w]⟩[y w]
(L′ ⟨u⟩)[z]y [z w]]⟩[y w]
L′ ⟨u⟩[y w]]⟩[y w]
Now we prove the strong bisimulation property, by induction on ⊸.
1. Base case 1: multiplicative root step t = L⟨lx.t′ ⟩L′ ⟨v⟩ ↦dBv u =
L⟨t′ [x L′ ⟨v⟩]⟩. The nontrivial cases are when the ⇚
⇛ step overlaps the
pattern of the dBv-redex. Note that by Lemma 14.3.1, if the ⇚
⇛ is internal to L′ ⟨v⟩, the proof is direct, since the dBv-redex is preserved. More
precisely, if L′ ⟨v⟩ ⇚
⇛ L′′ ⟨v ′ ⟩, we have:
dBv
L⟨lx.t′ ⟩ L′ ⟨v⟩
L⟨t′ [x L′ ⟨v⟩]⟩
⇚
⇛
⇚
⇛
dBv
L⟨lx.t′ ⟩ L′′ ⟨v ′ ⟩
L⟨t′ [x L′′ ⟨v ′ ⟩]⟩
Consider the remaining possibilities for ⇚
⇛:
(a) Garbage collection ≡gc . The garbage collected substitution must
be in L, i.e. L must be of the form L1 ⟨L2 [y L′′ ⟨v ′ ⟩]⟩ with y ∈/
̂ ∶= L1 ⟨L2 ⟩. Then:
fv(L2 ⟨lx.t′ ⟩). Let L
dBv
L⟨lx.t′ ⟩ L′ ⟨v⟩
L⟨t′ [x L′ ⟨v⟩]⟩
≡gc
≡gc
′
̂
L⟨lx.t
⟩ L′ ⟨v⟩
̂ ′ [x L′ ⟨v⟩]⟩
L⟨t
dBv
(b) Duplication ≡dup . The duplicated substitution must be in L, i.e. L
̂ ∶= L1 ⟨⟨⋅⟩[y u′ ][z u′ ]⟩.
must be of the form L1 ⟨L2 [y u′ ]⟩. Let L
Then:
dBv
L⟨lx.t′ ⟩ L′ ⟨v⟩
L⟨t′ [x L′ ⟨v⟩]⟩
≡dup
≡dup
̂
L⟨(L
2 ⟨lx.t ⟩)[z]y ⟩ L ⟨v⟩
′
′
′
′
̂
where t1 ∶= L⟨(L
2 ⟨t [x L ⟨v⟩]⟩)[z]y ⟩.
44
dBv
t1
(c) Commutation with application ≡@ . The axiom can be applied only in
one direction and there must be the same explicit substitution [y q]
as topmost constructor of each of the two sides of the application.
The diagram is:
dBv
L⟨lx.t′ ⟩[x q] L′ ⟨v⟩[y q]
t1
≡@
≡
(L⟨lx.t ⟩ L ⟨v⟩)[y q]
′
′
dBv
t2
where
t1 ∶= L⟨t′ [x L′ ⟨v⟩[y q]]⟩[x q],
t2 ∶= L⟨t′ [x L′ ⟨v⟩]⟩[y q].
To prove the equivalence on the right, let L′ ⟨v⟩[z]x denote the result
of replacing all occurrences of x by a fresh variable z in L′ ⟨v⟩. The
equivalence holds because:
L⟨t′′ [y
L⟨t′′ [y
L⟨t′′ [y
L⟨t′′ [y
L⟨t′′ [y
≡dup
≡∗com
≡[⋅]
=α
L′ ⟨v⟩]⟩[x q]
L′ ⟨v⟩[z]x ]⟩[x q][z q]
L′ ⟨v⟩[z]x ][z q]⟩[x q]
L′ ⟨v⟩[z]x [z q]]⟩[x q]
L′ ⟨v⟩[x q]]⟩[x q]
(d) Commutation of independent substitutions ≡com . The commutation of substitutions must be in L, i.e. L must be of the form
̂ ∶= L1 ⟨L2 [z w′ ][y u′ ]⟩.
L1 ⟨L2 [y u′ ][z w′ ]⟩ with z /∈ fv(u′ ). Let L
Then:
dBv
L⟨lx.t′ ⟩ L′ ⟨v⟩
L⟨t′ [x L′ ⟨v⟩]⟩
≡com
≡com
′
̂
L⟨lx.t
⟩ L′ ⟨v⟩
dBv
̂ ′ [x L′ ⟨v⟩]⟩
L⟨t
(e) Composition of substitutions ≡[⋅] . The composition of substitutions
must be in L, i.e. L must be of the form L1 ⟨L2 [y u′ ][z w′ ]⟩ with
̂ ∶= L1 ⟨L2 [y u′ [z w′ ]]⟩. As in the previous
z ∈/ fv(L2 ⟨lx.t′ ⟩). Let L
case:
dBv
L⟨lx.t′ ⟩ L′ ⟨v⟩
L⟨t′ [x L′ ⟨v⟩]⟩
≡[⋅]
≡[⋅]
̂
L⟨lx.t
⟩ L ⟨v⟩
′
dBv
′
̂ [x L′ ⟨v⟩]⟩
L⟨t
′
2. Base case 2: exponential root step t = V ⟨x⟩[x L⟨v⟩] ↦lsv u =
L⟨V ⟨v⟩[x v]⟩. Consider first the case when the ⇚
⇛-redex is internal to
V ⟨x⟩. By Lemma 14.3.2 we know ⇚
⇛ preserves the shape of V ⟨x⟩, i.e.
V ⟨x⟩ ⇚
⇛ V̂ ⟨x⟩. Then:
lsv
V ⟨x⟩[x L⟨v⟩]
L⟨V ⟨v⟩[x v]⟩
≡
⇚
⇛
V̂ ⟨x⟩[x L⟨v⟩]
lsv
45
L⟨V̂ ⟨v⟩[x v]⟩
If the ⇚
⇛-redex is internal to one of the substitutions in L, the proof is
straightforward. Note that the ⇚
⇛-redex has always a substitution at the
root. The remaining possibilities are such that substitution is in L, or
that it is precisely [x L⟨v⟩]. Axiom by axiom:
(a) Garbage collection ≡gc . If the garbage collected substitution is in L,
̂ be L without such substitution. Then:
let L
lsv
V ⟨x⟩[x L⟨v⟩]
≡gc
L⟨V ⟨v⟩[x v]⟩
≡gc
̂
V ⟨x⟩[x L⟨v⟩]
lsv
̂ ⟨v⟩[x v]⟩
L⟨V
The garbage collected substitution cannot be [x L⟨v⟩], since this
would imply x ∈/ fv(V ⟨x⟩), which is a contradiction.
(b) Duplication ≡dup . If the duplicated substitution is in L, then L is of
̂ = L1 ⟨[y t′ ][z t′ ]⟩. Then:
the form L1 ⟨L2 [y t′ ]⟩. Let L
lsv
V ⟨x⟩[x L⟨v⟩]
≡dup
L⟨V ⟨v⟩[x v]⟩
≡dup
lsv
t1
t2
where
̂ 2[z] ⟨v[z] ⟩⟩],
t1 ∶= V ⟨x⟩[x L⟨L
y
y
̂
t2 ∶= L⟨L2[z]y ⟨V ⟨v[z]y ⟩[x v[z]y ]⟩⟩.
If the duplicated substitution is [x L⟨v⟩], there are two possibilities,
depending on whether the occurrence of x substituted by the ↦lsv
step is replaced by the fresh variable y, or left untouched. If it is not
replaced:
lsv
V ⟨x⟩[x L⟨v⟩]
L⟨V ⟨v⟩[x v]⟩
≡dup
≡dup
t2
≡ (Lemma 14.4)
lsv
t4
t3
where
t2 ∶= L⟨(V ⟨v⟩)[y]x [x v][y v]⟩,
t3 ∶= L⟨(V ⟨v⟩)[y]x [x v][y L⟨v⟩]⟩,
t4 ∶= (V ⟨x⟩)[y]x [x L⟨v⟩][y L⟨v⟩].
If the occurrence of x substituted by the ↦lsv step is replaced by the
fresh variable y, the situation is essentially analogous.
46
(c) Commutation with application ≡@ . The only possibility is that the
substitution [x L⟨v⟩] is commuted with the outermost application
in V ⟨x⟩. Two cases:
i. The substitution acts on the left of the application, i.e. V = V ′ t′ .
lsv
(V ′ ⟨x⟩ t′ )[x L⟨v⟩]
t1
≡∗@
t2
≡@
≡∗[⋅]
t4
lsv
t3
where
t1 ∶= L⟨(V ′ ⟨v⟩ t′ )[x v]⟩,
t2 ∶= L⟨V ′ ⟨v⟩[x v]⟩L⟨t′ [x v]⟩,
t3 ∶= V ′ ⟨x⟩[x L⟨v⟩]t′ [x L⟨v⟩],
t4 ∶= L⟨V ′ ⟨v⟩[x v]⟩t′ [x L⟨v⟩].
ii. The substitution acts on the right of the application, i.e. V =
L′ ⟨v ′ ⟩V ′ . Similar to the previous case:
lsv
(L′ ⟨v ′ ⟩ V ′ ⟨x⟩)[x L⟨v⟩]
t1
≡∗@
t2
≡@
≡∗[⋅]
t4
lsv
t3
where
t1 ∶= L⟨(L′ ⟨v ′ ⟩ V ′ ⟨v⟩)[x v]⟩,
t2 ∶= L⟨L′ ⟨v ′ ⟩[x v]⟩L⟨V ′ ⟨v⟩[x v]⟩,
t3 ∶= L′ ⟨v ′ ⟩[x L⟨v⟩]V ′ ⟨x⟩[x L⟨v⟩],
t4 ∶= L′ ⟨v ′ ⟩[x L⟨v⟩]L⟨V ′ ⟨v⟩[x v]⟩.
(d) Commutation of independent substitutions ≡com . If the commuted
̂ be the result of commuting
substitutions both belong to L, let L
them, and the situation is exactly as for Garbage collection.
The remaining possibility is that V = V ′ [y t′ ] and [x L⟨v⟩] commutes with [y t′ ] (which implies x /∈ fv(t′ )). Then:
V ′ ⟨x⟩[y t′ ][x L⟨v⟩]
lsv
≡com
V ⟨x⟩[x L⟨v⟩][y t ]
′
′
47
lsv
L⟨V ′ ⟨v⟩[y t′ ][x v]⟩
≡∗com
L⟨V ⟨v⟩[x v]⟩[y t′ ]
′
(e) Composition of substitutions ≡[⋅] . If the composed substitutions both
̂ be the result of composing them, and the situation
belong to L, let L
is exactly as for Garbage collection.
The remaining possibility is that [x L⟨v⟩] is the outermost substitution composed by ≡[⋅] . This is not possible if the rule is applied
from left to right, since it would imply that V ⟨x⟩ = V ′ ⟨x⟩[y t′ ] with
x ∈/ V ′ ⟨x⟩, which is a contradiction.
Finally, if the ≡[⋅] rule is applied from right to left, L is of the form
L′ [y t′ ] and:
V ⟨x⟩[x L′ ⟨v⟩[y t′ ]]
lsv
L′ ⟨V ⟨v⟩[x v]⟩[y t′ ]
=
≡[⋅]
V ⟨x⟩[x L′ ⟨v⟩][y t′ ]
lsv
L′ ⟨V ⟨x⟩[x v]⟩[y t′ ]
3. Inductive case 1: left of an application V = V ′ q. The situation is:
t = V ′ ⟨t′ ⟩ q ⊸ V ′ ⟨u′ ⟩ q = u
If the ⇚
⇛ step is internal to V ′ ⟨t′ ⟩, the result follows by i.h.. The proof is
also direct if ⇚
⇛ is internal to q. The nontrivial case is when the ⇚
⇛ step
overlaps V ′ ⟨t′ ⟩ and q. There are two possibilities. The first is trivial: ≡gc
is used to introduce a substitution out of the blue, but this case clearly
commutes with reduction.
The second is that the application is commuted with a substitution via
the ≡@ rule (applied from right to left). There are two cases:
(a) The substitution comes from t′ . That is, V ′ = ⟨⋅⟩ and t′ has a substitution at its root. Then t′ must be a ↦lsv -redex t′ = V ′′ ⟨x⟩[x L⟨v⟩].
Moreover q = q ′ [x L⟨v⟩]. We have:
V ′′ ⟨x⟩[x L⟨v⟩] q ′ [x L⟨v⟩]
lsv
≡@
t2
≡
lsv
where
t1 ∶= L⟨V ′′ ⟨v⟩[x v]⟩ q ′ [x L⟨v⟩],
t2 ∶= (V ′′ ⟨x⟩ q ′ )[x L⟨v⟩],
t3 ∶= L⟨(V ′′ ⟨v⟩ q ′ )[x v]⟩.
For the equivalence on the right note that:
≡∗[⋅]
≡∗@
≡@
L⟨V ′′ ⟨v⟩[x v]⟩ q ′ [x L⟨v⟩]
L⟨V ′′ ⟨v⟩[x v]⟩ L⟨q ′ [x v]⟩
L⟨V ′′ ⟨v⟩[x v] q ′ [x v]⟩
L⟨(V ′′ ⟨v⟩ q ′ )[x v]⟩
48
t1
t3
(b) The substitution comes from V ′ . That is: V ′ = V ′′ [x w′ ]. Moreover,
q = q ′ [x w′ ]. The proof is then straightforward:
V ′′ ⟨t′ ⟩[x w′ ] q ′ [x w′ ]
t1
≡@
≡@
t2
t3
where
t1 ∶= V ′′ ⟨u′ ⟩[x w′ ] q ′ [x w′ ],
t2 ∶= (V ′′ ⟨t′ ⟩ q ′ )[x w′ ],
t3 ∶= (V ′′ ⟨u′ ⟩ q ′ )[x w′ ].
4. Inductive case 2: right of an application V = L⟨v⟩V ′ . The situation
is:
t = L⟨v⟩ V ′ ⟨t′ ⟩ ⊸ L⟨v⟩ V ′ ⟨u′ ⟩ = u
Reasoning as in the previous case (left of an application), if the ⇚
⇛ step
is internal to V ′ ⟨t′ ⟩, the result follows by i.h., and if it is internal to L⟨v⟩,
it is straightforward to close the diagram by resorting to the fact that ≡
preserves the shape of L⟨v⟩ (Lemma 14.3).
The remaining possibility is that the ⇚
⇛ step overlaps both L⟨v⟩ and
V ′ ⟨t′ ⟩. As in the previous case, this can only be possible if ≡gc introduces
a substitution out of the blue, which is a trivial case, or because of a
Commutation with application rule (≡@ , from right to left). This again
leaves two possibilities:
(a) The substitution comes from t′ . That is, V ′ = ⟨⋅⟩ and t′ is a ↦lsv -redex
t′ = V ′′ ⟨y⟩[y L′ ⟨v ′ ⟩]. Moreover, L = L′′ [y L′ ⟨v ′ ⟩]. Then:
L′′ ⟨v⟩[y L′ ⟨v ′ ⟩] V ′′ ⟨y⟩[y L′ ⟨v ′ ⟩]
≡@
t2
lsv
t1
≡
lsv
t3
where
t1 ∶= L′′ ⟨v⟩[y L′ ⟨v ′ ⟩] L′ ⟨V ′′ ⟨v ′ ⟩[y v ′ ]⟩,
t2 ∶= (L′′ ⟨v⟩ V ′′ ⟨y⟩)[y L′ ⟨v ′ ⟩],
t3 ∶= L′ ⟨(L′′ ⟨v⟩ V ′′ ⟨v ′ ⟩)[y v ′ ]⟩.
Exactly as in the previous case, for the equivalence on the right consider:
L′′ ⟨v⟩[y L′ ⟨v ′ ⟩] L′ ⟨V ′′ ⟨v ′ ⟩[y v ′ ]⟩
∗
≡[⋅] L′ ⟨L′′ ⟨v⟩[y v ′ ]⟩ L′ ⟨V ′′ ⟨v ′ ⟩[y v ′ ]⟩
≡∗@ L′ ⟨L′′ ⟨v⟩[y v ′ ] V ′′ ⟨v ′ ⟩[y v ′ ]⟩
≡@ L′ ⟨(L′′ ⟨v⟩ V ′′ ⟨v ′ ⟩)[y v ′ ]⟩
49
(b) The substitution comes from V ′ . That is, V ′ = V ′′ [x w′ ]. Moreover,
L = L′ [x w′ ]. This case is then straightforward:
L′ ⟨v⟩[x w′ ] V ′′ ⟨t′ ⟩[x w′ ]
L′ ⟨v⟩[x w′ ] V ′′ ⟨u′ ⟩[x w′ ]
≡@
≡@
(L′ ⟨v⟩ V ′′ ⟨t′ ⟩)[x w′ ]
(L′ ⟨v⟩ V ′′ ⟨u′ ⟩)[x w′ ]
5. Inductive case 3: left of a substitution V = V ′ [x q]. The situation
is:
t = V ′ ⟨t′ ⟩[x q] ⊸ V ′ ⟨u′ ⟩[x q] = u
If the ⇚
⇛ step is internal to V ′ ⟨t′ ⟩, the result follows by i.h.. If it is
internal to q, the steps are orthogonal, which makes the diagram trivial.
If the equivalence ≡gc introduces a substitution out of the blue the steps
trivially commute.
The remaining possibility is that the substitution [x q] is involved in the
⇚
⇛ redex. By case analysis on the kind of the step ≡b :
(a) Garbage collection ≡gc . We know x /∈ fv(V ′ ⟨t′ ⟩) and therefore also
x ∈/ fv(V ′ ⟨u′ ⟩). We get:
V ′ ⟨t′ ⟩[x q]
V ′ ⟨u′ ⟩[x q]
≡gc
≡gc
V ⟨t ⟩
′
V ⟨u′ ⟩
′
′
(b) Duplication ≡dup . The important fact is that if V ′ ⟨t′ ⟩ ⊸ V ′ ⟨u′ ⟩ and
V ′ ⟨t′ ⟩[y]x denotes the result of renaming some (arbitrary) occurrences
of x by y in V ′ ⟨t′ ⟩, then V ′ ⟨t′ ⟩[y]x ⊸ V ′ ⟨u′ ⟩[y]x , where V ′ ⟨u′ ⟩[y]x
denotes the result of renaming some occurrences of x by y in V ′ ⟨u′ ⟩.
By this we conclude:
V ′ ⟨t′ ⟩[x q]
V ′ ⟨u′ ⟩[x q]
≡dup
≡dup
(V ⟨t ⟩)[y]x [x q][y q]
′
′
(V ⟨u′ ⟩)[y]x [x q][y q]
′
(c) Commutation with application ≡@ . V ′ ⟨t′ ⟩ must be an application.
This allows for three possibilities:
i. The application comes from t′ . That is, V ′ = ⟨⋅⟩ and t′ is a
↦dBv -redex t′ = L⟨ly.t′′⟩ L′ ⟨v⟩. The diagram is exactly as for the
multiplicative base case 1c (read bottom-up).
ii. The application comes from V ′ , left case. That is, V ′ = V ′′ w′ .
This case is direct:
(V ′′ ⟨t′ ⟩ w′ )[x q]
t1
≡@
≡@
t2
t3
50
where
t1 ∶= (V ′′ ⟨u′ ⟩ w′ )[x q],
t2 ∶= V ′′ ⟨t′ ⟩[x q] w′ [x q],
t3 ∶= V ′′ ⟨u′ ⟩[x q] w′ [x q].
iii. The application comes from V ′ , right case.
L⟨v⟩ V ′′ . Analogous to the previous case.
That is, V ′ =
(d) Commutation of independent substitutions ≡com . Since V ′ ⟨t′ ⟩ must
have a substitution at the root, there are two possibilities:
i. The substitution comes from t′ . That is, V ′ = ⟨⋅⟩ and t′ is a
↦lsv -redex t′ = V ′′ ⟨y⟩[y L⟨v⟩], with x ∈/ fv(L⟨v⟩). Then:
V ′′ ⟨y⟩[y L⟨v⟩][x q]
lsv
≡com
V ⟨y⟩[x q][y L⟨v⟩]
′′
lsv
L⟨V ′′ ⟨v⟩[y v]⟩[x q]
≡∗com
L⟨V ⟨v⟩[x q][y v]⟩
′′
ii. The substitution comes from V ′ . That is, V ′ = V ′′ [y w′ ] with
x ∈/ fv(w′ ). This case is direct:
V ′′ ⟨t′ ⟩[y w′ ][x q]
lsv
V ′′ ⟨u′ ⟩[y w′ ][x q]
lsv
V ⟨u ⟩[x q][y w′ ]
≡com
≡com
V ⟨t ⟩[x q][y w ]
′′
′
′
′′
′
(e) Composition of substitutions ≡[⋅] . As in the previous case, there are
two possibilities:
i. The substitution comes from t′ . That is, V ′ = ⟨⋅⟩ and t′ is a
↦lsv -redex t′ = V ′′ ⟨y⟩[y L⟨v⟩], with x ∈/ fv(V ′′ ⟨y⟩). Then:
V ′′ ⟨y⟩[y L⟨v⟩][x q]
lsv
L⟨V ′′ ⟨v⟩[y v]⟩[x q]
=
≡[⋅]
V ′′ ⟨y⟩[y L⟨v⟩[x q]]
lsv
L⟨V ′′ ⟨v⟩[y v]⟩[x q]
ii. The substitution comes from V ′ . That is, V ′ = V ′′ [y w′ ] with
x ∈/ fv(V ′′ ⟨t′ ⟩). The proof for this case is direct:
14.3
V ′′ ⟨t′ ⟩[y w′ ][x q]
V ′′ ⟨u′ ⟩[y w′ ][x q]
≡[⋅]
≡[⋅]
V ′′ ⟨t′ ⟩[y w′ [x q]]
V ′′ ⟨u′ ⟩[y w′ [x q]]
Proof of Proposition 2.2 (≡ is a Strong Bisimulation)
for Right-to-Left Call-by-Value
The proof is obtained as a minimal variation over the proof for left-to-right
call-by-value (previous subsection), and is therefore omitted.
51
14.4
Proof of Proposition 2.2 (≡ is a Strong Bisimulation)
for Call-by-Need
We need two preliminary lemmas, proved by straightforward inductions on N :
Lemma 14.5. Let t be a term, N be a call-by-need evaluation context not
capturing any variable in fv(t), and x ∉ fv(N ⟨y⟩). Then N ⟨t[x u]⟩ ≡Need
N ⟨t⟩[x u].
Lemma 14.6. The equivalence relation ≡Need preserves the shape of N ⟨x⟩. More
precisely, if N ⟨x⟩ ≡Need t, with x not captured by N , then t is of the form N ′ ⟨x⟩,
with x not captured by N ′ .
We follow the structure of the previous proofs of strong bisimulation, in
particular the proof is by induction on ⊸ and to ease the notation we write
≡ for ≡Need . Remember that for call-by-need the definition of the structural
equivalence is different, it is the one given only by axioms ≡@l , ≡com , and ≡[⋅] .
1. Base case 1: multiplicative root step t = L⟨lx.t′ ⟩q ↦dB u = L⟨t′ [x q]⟩.
Every application of ≡ inside q or inside one of the substitutions in L
trivially commutes with the step. The interesting cases are those where
structural equivalence has a critical pair with the step:
(a) Commutation with left of an application ≡@l . If L = L′ [y r] then
dB
L′ ⟨lx.t′ ⟩[y r]q
L′ ⟨t′ [x q]⟩[y r]
=
≡@l
dB
(L ⟨lx.t ⟩q)[y r]
′
′
L ⟨t [x q]⟩[y r]
′
′
(b) Commutation of independent substitutions ≡com . The substitutions
that are commuted by the ≡com rule must be both in L, i.e. L
must be of the form L′ ⟨L′′ [y w′ ][z r′ ]⟩ with z ∈/ fv(w′ ). Let
̂ = L′ ⟨L′′ [z r′ ][y w′ ]⟩. Then:
L
L⟨lx.t′ ⟩u′
dB
L⟨t′ [x u′ ]⟩
dB
̂ [x u′ ]⟩
L⟨t
≡com
≡com
̂
L⟨lx.t
⟩u
′
′
′
(c) Composition of substitutions ≡[⋅] . The substitutions that appear in
the left-hand side of the ≡[⋅] rule must both be in L, i.e. L must
̂=
be of the form L′ ⟨L′′ [y w′ ][z r′ ]⟩ with z /∈ fv(L′′ ⟨lx.t′ ⟩). Let L
′
′′
′
′
L ⟨L [y w [z r ]]⟩. Exactly as in the previous case:
L⟨lx.t′ ⟩u′
dB
≡[⋅]
L⟨t′ [x u′ ]⟩
≡[⋅]
′ ′
̂
L⟨lx.t
⟩u
52
dB
̂ ′ [x u′ ]⟩
L⟨t
2. Base case 2: exponential root step t = N ⟨x⟩[x L⟨v⟩] ↦lsv u =
L⟨N ⟨v⟩[x v]⟩. Consider first the case when the ⇚
⇛-redex is internal
to N ⟨x⟩. By Lemma 14.6 we know ⇚
⇛ preserves the shape of N ⟨x⟩, i.e.
̂ ⟨x⟩. Then:
N ⟨x⟩ ⇚
⇛N
lsv
N ⟨x⟩[x L⟨v⟩]
L⟨N ⟨v⟩[x v]⟩
≡
⇚
⇛
̂ ⟨x⟩[x L⟨v⟩]
N
̂ ⟨v⟩[x v]⟩
L⟨N
lsv
If the ⇚
⇛-redex is internal to one of the substitutions in L, the proof is
straightforward. Note that the ⇚
⇛-redex has always a substitution at the
root. The remaining possibilities are that such substitution is in L, or
that it is precisely [x L⟨v⟩]. Axiom by axiom:
(a) Commutation with the left of an application ≡@l . The only possibility
is that the substitution [x L⟨v⟩] is commuted with the outermost
application in N ⟨x⟩, i.e. N = N ′ t′ . The diagram is:
lsv
(N ′ ⟨x⟩ t′ )[x L⟨v⟩]
L⟨(N ′ ⟨v⟩ t′ )[x v]⟩
≡@l
≡∗@l
L⟨N ′ ⟨v⟩[x v]⟩ t′
lsv
N ′ ⟨x⟩[x L⟨v⟩]t′
(b) Commutation of independent substitutions ≡com . Two sub-cases:
̂ be the result
i. The commuted substitutions both belong to L. Let L
of commuting them, and the diagram is:
N ⟨x⟩[x L⟨v⟩]
lsv
≡com
L⟨N ⟨v⟩[x v]⟩
≡com
̂
N ⟨x⟩[x L⟨v⟩]
lsv
̂ ⟨v⟩[x v]⟩
L⟨N
ii. One of the commuted substitutions is [x L⟨v⟩]. Then N =
N ′ [y t′ ] and [x L⟨v⟩] commutes with [y t′ ] (which implies
x ∈/ fv(t′ )). Then:
N ′ ⟨x⟩[y t′ ][x L⟨v⟩]
lsv
≡com
N ′ ⟨x⟩[x L⟨v⟩][y t′ ]
lsv
L⟨N ′ ⟨v⟩[y t′ ][x v]⟩
≡∗com
L⟨N ′ ⟨v⟩[x v]⟩[y t′ ]
(c) Composition of substitutions ≡[⋅] . Two sub-cases:
i. The composed substitutions both belong to L. Analogous to case
2(b)i.
ii. One of the composed subtitutions is [x L⟨v⟩]. This is not possible if the rule is applied from left to right, since it would imply
that N ⟨x⟩ = N ′ ⟨x⟩[y t′ ] with x ∈/ N ′ ⟨x⟩, which is a contradiction.
53
Finally, if the ≡[⋅] rule is applied from right to left, L is of the
form L′ [y t′ ] and:
lsv
N ⟨x⟩[x L′ ⟨v⟩[y t′ ]]
L′ ⟨N ⟨v⟩[x v]⟩[y t′ ]
=
≡[⋅]
lsv
N ⟨x⟩[x L′ ⟨v⟩][y t′ ]
L′ ⟨N ⟨x⟩[x v]⟩[y t′ ]
3. Inductive case 1: left of an application N = N ′ q. The situation is:
t = N ′ ⟨t′ ⟩ q ⊸ N ′ ⟨u′ ⟩ q = u
If the ⇚
⇛ step is internal to N ′ ⟨t′ ⟩, the result follows by i.h.. The proof
is also direct if ⇚
⇛ is internal to q. The nontrivial cases are those where
⇚
⇛ overlaps N ′ ⟨t′ ⟩ and q. The only possible case is that a substitution
commutes with the topmost application via ≡@l (applied from right to
left). There are two cases:
(a) The substitution comes from t′ . That is, N ′ = ⟨⋅⟩ and t′ has a substitution at its root. Then t′ must be a ↦lsv -redex t′ = N ′′ ⟨x⟩[x L⟨v⟩].
We have:
lsv
N ′′ ⟨x⟩[x L⟨v⟩] q
≡@l
lsv
(N ′′ ⟨x⟩ q)[x L⟨v⟩]
L⟨N ′′ ⟨v⟩[x v]⟩ q
≡∗@l
L⟨(N ′′⟨v⟩ q)[x v]⟩
(b) The substitution comes from N ′ . That is: N ′ = N ′′ [x w′ ]. The
proof is then straightforward:
N ′′ ⟨t′ ⟩[x w′ ] q
N ′′ ⟨u′ ⟩[x w′ ] q
≡@l
≡@l
(N ′′ ⟨t′ ⟩ q)[x w′ ]
(N ′′ ⟨u′ ⟩ q)[x w′ ]
4. Inductive case 2: left of a substitution N = N ′ [x q]. The situation
is:
t = N ′ ⟨t′ ⟩[x q] ⊸ N ′ ⟨u′ ⟩[x q] = u
If the ⇚
⇛ step is internal to N ′ ⟨t′ ⟩, the result follows by i.h.. If it is
internal to q, the steps are orthogonal, which makes the diagram trivial.
The remaining possibility is that the substitution [x q] is involved in the
⇚
⇛ redex. By case analysis on the kind of the step ≡b :
(a) Commutation with the left of an application ≡@l . N ′ ⟨t′ ⟩ must be an
application. Two possibilities:
i. The application comes from t′ . That is, N ′ = ⟨⋅⟩ and t′ is a ↦dB redex t′ = L⟨ly.t′′ ⟩ r. This is exactly as the base case 1a (read
bottom-up).
ii. The application comes from N ′ , i.e. N ′ = N ′′ w′ . This is exactly
as the inductive case 3b (read bottom-up).
54
(b) Commutation of independent substitutions ≡com . Since N ′ ⟨t′ ⟩ must
have a substitution at the root, there are two possibilities:
i. The substitution comes from t′ . That is, N ′ = ⟨⋅⟩ and t′ is a
↦lsv -redex t′ = N ′′ ⟨y⟩[y L⟨v⟩], with x /∈ fv(L⟨v⟩). This case is
exactly as the base exponential case 2(b)ii (read bottom-up).
ii. The substitution comes from N ′ . That is, N ′ = N ′′ [y w′ ] with
x ∈/ fv(w′ ). The diagram is:
N ′′ ⟨t′ ⟩[y w′ ][x q]
lsv
N ′′ ⟨u′ ⟩[y w′ ][x q]
≡com
≡com
N ′′ ⟨t′ ⟩[x q][y w′ ]
lsv
N ′′ ⟨u′ ⟩[x q][y w′ ]
(c) Composition of substitutions ≡[⋅] . As in the previous case, there are
two possibilities:
i. The substitution comes from t′ . That is, N ′ = ⟨⋅⟩ and t′ is a
↦lsv -redex t′ = N ′′ ⟨y⟩[y L⟨v⟩], with x /∈ fv(N ′′ ⟨y⟩). This case
is exactly as the base exponential case 2(c)ii (read bottom-up).
ii. The substitution comes from N ′ . That is, N ′ = N ′′ [y w′ ] with
x ∈/ fv(N ′′ ⟨t′ ⟩). The diagram is:
N ′′ ⟨t′ ⟩[y w′ ][x q]
N ′′ ⟨u′ ⟩[y w′ ][x q]
≡[⋅]
≡[⋅]
N ⟨t ⟩[y w [x q]]
′′
′
N ⟨u ⟩[y w′ [x q]]
′
′′
′
5. Inductive case 3: inside a hereditary head substitution N =
N ′ ⟨x⟩[x N ′′ ]. The situation is:
t = N ′ ⟨x⟩[x N ′′ ⟨q⟩] ⊸ N ′ ⟨x⟩[x N ′′ ⟨q ′ ⟩] = u
If ⇚
⇛ is internal to N ′ ⟨x⟩ the two steps clearly commutes. If ⇚
⇛ is internal
to N ′′ ⟨q⟩ we conclude using the i.h.. The remaining cases are when ⇚
⇛
overlaps with the topmost constructor. Axiom by axiom:
(a) Commutation with the left of an application ≡@l . It must be that
N ′ ⟨x⟩ = N ′′′′ ⟨x⟩r with x ∉ fv(r). Then the two steps simply commute:
(N ′′′′⟨x⟩r)[x N ′′ ⟨q⟩]
(N ′′′′⟨x⟩r)[x N ′′ ⟨q ′ ⟩]
≡@l
N
′′′′
≡@l
⟨x⟩[x N ⟨q⟩]r
′′
N
′′′′
⟨x⟩[x N ′′ ⟨q ′ ⟩]r
(b) Commutation of independent substitutions ≡com . It must be that
N ′ ⟨x⟩ = N ′′′′ ⟨x⟩[y r] with x ∉ fv(r). Then the two steps simply
commute:
t1
N ′′′′⟨x⟩[y r][x N ′′ ⟨q⟩]
≡@l
≡@l
t2
t3
55
where
t1 ∶= N ′′′′⟨x⟩[y r][x N ′′ ⟨q ′ ⟩],
t2 ∶= N ′′′′⟨x⟩[x N ′′ ⟨q⟩][y r],
t3 ∶= N ′′′′⟨x⟩[x N ′′ ⟨q ′ ⟩][y r].
(c) Composition of substitutions ≡[⋅] . There are various sub-cases
i. [x N ′′ ⟨q⟩] enters in a substitution. It must be that N ′ ⟨x⟩ =
N1 ⟨y⟩[y N2 ⟨x⟩] with x ∉ fv(N1 ⟨y⟩). Then the diagram is:
N1 ⟨y⟩[y N2 ⟨x⟩][x N ′′ ⟨q⟩]
t1
≡[⋅]
≡[⋅]
t2
t3
t1 ∶= N1 ⟨y⟩[y N2 ⟨x⟩][x N ′′ ⟨q ′ ⟩],
t2 ∶= N1 ⟨y⟩[y N2 ⟨x⟩[x N ′′ ⟨q⟩]],
t3 ∶= N1 ⟨y⟩[y N2 ⟨x⟩[x N ′′ ⟨q ′ ⟩]].
ii. a substitution pops out of [x N ′′ ⟨q⟩]. Two sub-cases:
A. The substitution comes from N ′′ .
Then N ′′ ⟨q⟩
′′′′
N ⟨q⟩[y r]. The diagram is:
N ′ ⟨x⟩[x N ′′′′⟨q⟩[y r]]
t1
≡[⋅]
≡[⋅]
t2
t3
=
where
t1 ∶= N ′ ⟨x⟩[x N ′′′′⟨q ′ ⟩[y r]],
′
(2)
′′′′
t2 ∶= N ⟨x⟩[x N ⟨q⟩][y r],
t3 ∶= N ′ ⟨x⟩[x N ′′′′⟨q ′ ⟩][y r].
(3)
(4)
B. The substitution comes from q. Then N ′′ = ⟨⋅⟩ and q is a
↦lsv -redex t′ = N ′′′′ ⟨y⟩[y L⟨v⟩] and the diagram is:
N ′ ⟨x⟩[x N ′′′′⟨y⟩[y L⟨v⟩]]
lsv
≡[⋅]
t2
lsv
t1
≡∗[⋅]
t3
where
t1 ∶= N ′ ⟨x⟩[x L⟨N ′′′′⟨v⟩[y v]⟩],
t2 ∶= N ′ ⟨x⟩[x N ′′′′⟨y⟩][y L⟨v⟩],
(5)
(6)
t3 ∶= L⟨N ′ ⟨x⟩[x N ′′′′ ⟨v⟩][y v]⟩.
(7)
56
14.5
Proofs for the LAM
Invariants, Lemma 6.3. By induction on the length of the execution leading to
s, and straightforward inspection of the transition rules.
Distillation, Theorem 6.4.
e ∣ f (t, e) ∶∶ π, and:
tu ∣ e ∣ π
=
1. Commutative 1. We have tu ∣ e ∣ π →c1 u ∣
π⟨e⟨tu⟩⟩ ≡∗@
π⟨e⟨t⟩e⟨u⟩⟩
= u ∣ e ∣ f (t, e) ∶∶ π
As before, we use that π is a right-to-left call-by-value evaluation context,
which enables us to use the ≡@ rule.
2. Commutative 2. We have v ∣ e ∣ f (t, e′ ) ∶∶ π →c2 t ∣ e′ ∣ a(v, e) ∶∶ π, and:
v ∣ e ∣ f (t, e′ ) ∶∶ π
=
π⟨e′ ⟨t⟩e⟨v⟩⟩
= t ∣ e′ ∣ a(v, e) ∶∶ π
3. Multiplicative. We have lx.t ∣ e ∣ a(c) ∶∶ π →m t ∣ [x c] ∶∶ e ∣ π, and:
lx.t ∣ e ∣ a(c) ∶∶ π
= π⟨e⟨lx.t⟩c⟩ ⊸m
π⟨e⟨t[x c]⟩⟩
which is equal to t ∣ [x c] ∶∶ e ∣ π.
4. Exponential. Let e = e′′ ∶∶ [x (t, e′ )] ∶∶ e′′′. We have x ∣ e ∣ π →e t ∣ e′ ∣ π,
and::
x∣e∣π
=
π⟨e⟨x⟩⟩
⊸e
π⟨e′′′ ⟨e′ ⟨e′′ ⟨t⟩[x t]⟩⟩⟩
≡∗gc
π⟨e′ ⟨t⟩⟩ =
t ∣ e′ ∣ π
Note that by Lemma 6.3.3, t is an abstraction, and thus we are able to
apply ⊸e . Moreover, by Lemma 6.3.1, e binds variables to closures, and
e′ ⟨t⟩ is closed; this allows e′′ and e′′′ to be garbage collected. For doing so,
the ≡gc rule must be applied below a right-to-left call-by-value evaluation
context, which follows from Lemma 6.3.4.
Progress. Let s = t ∣ e ∣ π be a commutative normal form s.t. s ⊸ u. If t is
• an application uw. Then a →c1 transition applies and s is not a commutative normal form, absurd.
• an abstraction lx.u. Then s = π⟨e⟨lx.u⟩⟩ is not in normal form. There can
only be a ⊸m -redex, so π must be of the form π⟨⟨⋅⟩c⟩. This implies there
is a →m transition from s.
• a variable x. Then s = π⟨e⟨x⟩⟩ is not in normal form. There can only be a
⊸e -redex, and it must involve x, thus e = e′′′ ⟨e′′ [x e′ ⟨v⟩]⟩. This implies
there is a →e transition from s.
57
14.6
Proofs for the MAM
Proof. Let ⇚
⇛ be the symmetric and contextual closure of the ∼ rule by which
≡MAM is defined. Note ≡MAM is the reflexive–transitive closure of ⇚
⇛. It suffices
to show that the property holds for ⇚
⇛, i.e. that w ⇚
⇛⊸ u implies w ⊸≡MAM u.
The fact that ⇚
⇛∗ is a bisimulation then follows by induction on the number of
⇚
⇛ steps.
Let w ⇚
⇛ t ⊸ u. The proof of w ⊸≡MAM u goes by induction on the callby-need context N under which the ⊸-redex in t is contracted. Note that since
t1 ∼ t2 determines a bijection between the redexes of t1 and t2 , it suffices to check
the cases when ∼ is applied from left to right (i.e. t ∼ w). For the right-to-left
cases, all diagrams can be considered from bottom to top.
• Base case, i.e. empty context N = ⟨⋅⟩. Two cases, depending on the
⊸ step contracting a ↦dB or a ↦ls redex:
1. Case t = L⟨lx.t′ ⟩u′ ↦dB L⟨t′ [x u′ ]⟩. There are no ∼ redexes in t,
since any application in t must be either t itself or below lx, which
is not a call-by-name evaluation context.
2. Case t = N ⟨x⟩[x t′ ] ↦ls N ⟨t′ ⟩[x t′ ]. Any ∼ redex must be internal
to N , in the sense that N = N ′ ⟨(N ′′ u′ )[y w′ ]⟩ with y ∈/ fv(u′ ). Let
̂ = N ′ ⟨N ′′ [y w′ ]u′ ⟩. Then:
N
N ⟨x⟩[x t′ ]
ls
∼
N ⟨t′ ⟩[x t′ ]
∼
̂ ⟨x⟩[x t′ ]
N
ls
̂ ⟨t′ ⟩[x t′ ]
N
• Inductive case N = N ′ q. Since the application of ⇚
⇛ must be internal
to N ′ , the result follows directly by i.h..
• Inductive case N = N ′ [x q]. If the ⇚
⇛ step is internal to N ′ , the result
follows again by applying i.h.. The remaining possibility is that N ′ ⟨t⟩ is
an application. Here there are two cases:
1. N ′ = ⟨⋅⟩, i.e. ∼ interacts with a redex. The redex in question
must be a dB-redex, since it must have an application at the root.
The situation is the following, with x ∈/ fv(u′ ):
dB
(L⟨ly.t′⟩ u′ )[x q]
L⟨t′ [y u′ ]⟩[x q]
∼
L⟨ly.t ⟩[x q] u
′
=
dB
′
L⟨t [y u′ ]⟩[x q]
′
2. N ′ = N ′′ t′ , i.e. there is no interaction between ∼ and a redex.
This case is straightforward, since the contraction of the ⊸ redex
and the application of ∼ are orthogonal.
14.7
Proofs for the Split CEK
Split CEK Distillation, Theorem 7.7. Properties of the decoding:
58
1. Commutative 1. We have t u ∣ e ∣ π ∣ D →c1 t ∣ e ∣ (u, e) ∶∶ π ∣ D, and:
tu ∣ e ∣ π ∣ D
= D⟨π⟨e⟨t u⟩⟩⟩
D⟨π⟨e⟨t⟩ e⟨u⟩⟩⟩
t ∣ e ∣ (u, e) ∶∶ π ∣ D
≡∗@
=
2. Commutative 2. We have v ∣ e ∣ (t, e′ ) ∶∶ π ∣ D →c2 t ∣ e′ ∣ ǫ ∣ ((v, e), π) ∶∶ D,
and:
v ∣ e ∣ (t, e′ ) ∶∶ π ∣ D = D⟨π⟨e⟨v⟩ e′ ⟨t⟩⟩⟩
≡∗gc
′
≡∗@
D⟨π⟨e⟨v⟩ e⟨e ⟨t⟩⟩⟩⟩
′
D⟨π⟨e⟨v e ⟨t⟩⟩⟩⟩
=
t ∣ e′ ∣ ǫ ∣ ((v, e), π) ∶∶ D
3. Multiplicative. We have v ∣ e ∣ ǫ ∣ ((lx.t, e′ ), π) ∶∶ D →m t ∣ [x (v, e)] ∶∶ e′ ∣
π ∣ D, and:
v ∣ e ∣ ǫ ∣ ((lx.t, e′ ), π) ∶∶ D =
→m
D⟨π⟨e′ ⟨(lx.t) e⟨v⟩⟩⟩⟩
=
D⟨π⟨e′ ⟨t[x e⟨v⟩]⟩⟩⟩
t ∣ [x (v, e)] ∶∶ e′ ∣ π ∣ D
4. Exponential. We have x ∣ e1 ∶∶ [x (v, e)] ∶∶ e2 ∣ π ∣ D →e v ∣ e ∣ π ∣ D, and:
x ∣ e1 ∶∶ [x (v, e)] ∶∶ e2 ∣ π ∣ D
D⟨π⟨e2 ⟨e1 ⟨x⟩[x e⟨v⟩]⟩⟩⟩
D⟨π⟨e2 ⟨e⟨e1 ⟨v⟩[x v]⟩⟩⟩⟩
D⟨π⟨e⟨v⟩⟩⟩
v∣e∣π∣D
=
→e
≡∗gc
=
We use that e⟨v⟩ is closed by Lemma 7.6.1 to ensure that e1 , e2 , and [x v]
can be garbage collected.
Progress. Let s = t ∣ e ∣ π be a commutative normal form s.t. s ⊸ u. If t is
• an application uw. Then a →c1 transition applies and s is not a commutative normal form, absurd.
• an abstraction v. The decoding s = D⟨π⟨e⟨v⟩⟩⟩ must have a multiplicative
redex, because it must have a redex and v is not a variable. So v is
applied to something, i.e. there must be at least one application node in
D⟨π⟩. Moreover, the stack π must be empty, otherwise there would be an
administrative →c2 transition, contradicting the hypothesis. So D is not
empty. Let D = ((u, e′ ), π ′ ) ∶∶ D′ . By point 3 of Lemma 7.6, u must be a
value, and a →m transition applies.
• a variable x. By point 1 of Lemma 7.6, x must be bound by e, so e = e1 ∶∶
[x (u, e′ )] ∶∶ e2 and a →e transition applies.
59
14.8
Proofs for the Merged WAM
Distillation, Theorem 9.2.
a(u) ∶∶ π ∣ E, and:
1. Commutative 1. We have t u ∣ π ∣ E →c1 t ∣
t u ∣ π ∣ E = E⟨π⟨t u⟩⟩ = t ∣ a(u) ∶∶ π ∣ E
2. Commutative 2. We have x ∣ π ∣ E1 ∶∶ [x t] ∶∶ E2 →c2 t ∣ h(E1 , x) ∶∶ π ∣ E2 ,
and:
x ∣ π ∣ E1 ∶∶ [x t] ∶∶ E2 = E2 ⟨E1 ⟨π⟨x⟩⟩[x t]⟩ =
t ∣ h(E1 , x) ∶∶ π ∣ E2
3. Multiplicative. We have lx.t ∣ a(u) ∶∶ π ∣ E →m t ∣ π ∣ [x u] ∶∶ E, and:
lx.t ∣ a(u) ∶∶ π ∣ E
= E⟨π⟨(lx.t) u⟩⟩
E⟨π⟨t[x u]⟩⟩
E⟨π⟨t⟩[x u]⟩
t ∣ π ∣ [x u] ∶∶ E
⊸m
≡Need Lem. 2.4
=
4. Exponential. We have v ∣ h(E1 , x) ∶∶ π ∣ E2 →e v α ∣ π ∣ E1 ∶∶ [x v] ∶∶ E2 ,
and:
v ∣ h(E1 , x) ∶∶ π ∣ E2
= E2 ⟨E1 ⟨π⟨x⟩⟩[x v]⟩
E2 ⟨E1 ⟨π⟨v⟩⟩[x v]⟩
E2 ⟨E1 ⟨π⟨v α ⟩⟩[x v]⟩
v α ∣ π ∣ E1 ∶∶ [x v] ∶∶ E2
⊸e
=α
=
Progress. Let s = t ∣ π ∣ E be a commutative normal form s.t. s ⊸ u. If t is
1. an application uw. Then a →c1 transition applies and s is not a commutative normal form, absurd.
2. an abstraction v. The decoding s is of the form E⟨π⟨v⟩⟩. The stack π
cannot be empty, since then s = E⟨v⟩ would be normal. So either the a
→e or a →m transition applies.
3. a variable x. By the global closure invariant, x is bound by E. Then a →c2
transition applies and s is not a commutative normal form, absurd.
14.9
Proofs for the Pointing WAM
Pointing WAM Invariants, Lemma 10.3. By induction on the length of the execution. Points 1 and 2 are by direct inspection of the rules. Assuming ED,
point 4 is immediate by induction on the length of D.
Thus we are only left to check point 3. We use point 2, i.e. that substitutions in E bind pairwise distinct variables. Following we show that transitions
preserve the invariant:
1. Conmutative 1. We have:
t u ∣ π ∣ D ∣ E →c1 t ∣ u ∶∶ π ∣ D ∣ E
Trivial, since the dump and the environment are the same and (u ∶∶ π)⟨t⟩ =
π⟨t u⟩.
60
2. Conmutative 2. We have s →c2 s′ with:
s = x ∣ π ∣ D ∣ E1 ∶∶ [x t] ∶∶ E2
s′ = t ∣ ǫ ∣ (x, π) ∶∶ D ∣ E1 ∶∶ [x ◻] ∶∶ E2
Note that since by i.h. (π⟨x⟩, (E1 ∶∶ [x t] ∶∶ E2 ) ↿) is closed and x is
free in π⟨x⟩, there cannot be any dumped substitutions in E2 . Then
(E1 ∶∶ [x t] ∶∶ E2 )↿= E1 ↿∶∶ [x t] ∶∶ E2 and we know:
(π⟨x⟩, E1 ↿∶∶ [x t] ∶∶ E2 ) is closed
For 3a, note (E1 ∶∶ [x ◻] ∶∶ E2 ) ↿= E2 .
(t, E2 ) is closed, which is implied by (8).
(8)
Then we must show
For 3b, there are two cases:
• If the pair is (x, π), we must show
(π⟨x⟩, (E1 ∶∶ [x ◻] ∶∶ E2 )↿x ) is closed, i.e.
(π⟨x⟩, E1 ↿∶∶ [x ◻] ∶∶ E2 ) is closed
which is implied by (8).
• If the pair is (y, π ′ ) in D, with y ≠ x, note first that
(E1 ∶∶ [x t] ∶∶ E2 )↿y = E1 ↿y ∶∶ [x t] ∶∶ E2
And similarly for (E1 ∶∶ [x ◻] ∶∶ E2 ) ↿y . Moreover, by the invariant
on s we know
(π ′ ⟨y⟩, E1 ↿y ∶∶ [x t] ∶∶ E2 ) is closed
and this implies
(π ′ ⟨y⟩, E1 ↿y ∶∶ [x ◻] ∶∶ E2 ) is closed
as required.
For 3c, we have already observed that E2 has no dumped substitutions.
Then [x ◻] is the rightmost dumped substitution in the environment of
s′ , while (x, π) is the leftmost pair in the dump. We conclude by the fact
that the invariant already holds for s.
3. Multiplicative, empty dump. We have s →m s′ with:
s = lx.t ∣ u ∶∶ π ∣ ǫ ∣ E
s′ = t ∣ π ∣ ǫ ∣ [x u] ∶∶ E
First note that, since the environment and the dump are dual in s, there
are no dumped substitutions in E.
61
For point 3a, we know that:
(π⟨(lx.t) u⟩, E) is closed
(9)
and we have to check:
(π⟨t⟩, [x u] ∶∶ E) is closed
Let y ∈ fv(π⟨t⟩). Then either y = x, which is bound by [x u], or y ∈
fv(π⟨lx.t⟩), in which case y is bound by E. Moreover, since π is an
application context, by (9) we get (u, E) is closed.
Points 3b and 3c are trivial since the dump is empty and the environment
has no dumped substitutions.
4. Multiplicative, non-empty dump. We have s →m s′ with:
s = lx.t ∣ u ∶∶ π ∣ (y, π ′ ) ∶∶ D ∣ E1 ∶∶ [y ◻] ∶∶ E2
s′ = t ∣ π ∣ (y, π ′ ) ∶∶ D ∣ E1 ∶∶ [y ◻] ∶∶ [x u] ∶∶ E2
Note first that since the invariant holds for s, we know [y ◻] is the rightmost dumped substitution in the environment of both s and s′ . Therefore
(E1 ∶∶ [y ◻] ∶∶ E2 )↿= E2
For proving point 3a, we have:
(π⟨(lx.t) u⟩, E2 ) is closed
and we must show:
(π⟨t⟩, [x u] ∶∶ E2 ) is closed
The situation is exactly as in point 3a for the →m transition, empty dump
case.
For point 3b, let (z, π ′′ ) be any pair in (y, π ′ ) ∶∶ D. Let also
⎧
⎪
⎪E1 ↿
E1′ ∶= ⎨
⎪
E ↿
⎪
⎩ 1 z
if y = z
otherwise
and note that (E1 ∶∶ [y ◻] ∶∶ E) ↿y = E1′ ∶∶ [y ◻] ∶∶ E for any environment
E that contains no dumped substitutions. By the invariant on s, we have
that:
(π ′′ ⟨z⟩, E1′ ∶∶ [y ◻] ∶∶ E2 ) is closed
Moreover, from point 3a we know (u, E2 ) is closed. Both imply:
(π ′′ ⟨z⟩, E1′ ∶∶ [y ◻] ∶∶ [x u] ∶∶ E2 ) is closed
as required.
For point 3c, just note that the substitution [x u] added to the environment is not dumped, and so duality holds because it holds for s by
i.h..
62
5. Exponential. We have s →e s′ with:
s = v ∣ ǫ ∣ (x, π) ∶∶ D ∣ E1 ∶∶ [x ◻] ∶∶ E2
s′ = v α ∣ π ∣ D ∣ E1 ∶∶ [x v] ∶∶ E2
First note that since the environment and the dump are dual in s, we
know E2 has no dumped substitutions.
For proving point 3a, by resorting to point 3a on the state s, for which
the invariant already holds, we have that:
(v, E2 ) is closed
(10)
Moreover, by point 3b on s, specialized on the pair (x, π), we also know:
(π⟨x⟩, E1 ↿∶∶ [x ◻] ∶∶ E2 ) is closed
(11)
We must check that:
(π⟨v α ⟩, E1 ↿∶∶ [x v] ∶∶ E2 ) is closed
Any free variable in π⟨v α ⟩ is either free in π, in which case by (10) it
must be bound by E1 ↿∶∶ [x ◻] ∶∶ E2 , or free in v, in which case by (10) it
must be bound by E2 . In both cases it is bound by E1 ↿∶∶ [x v] ∶∶ E2 , as
required. To conclude the proof of point 3a, note that by combining (10)
and (11) we get E1 ↿∶∶ [x v] ∶∶ E2 is closed.
For proving point 3b, let (y, π ′ ) be a pair in D. Using that x ≠ y, by the
invariant on s we know:
(π ′ ⟨y⟩, E1 ↿y ∶∶ [x ◻] ∶∶ E2 ) is closed
and this implies:
(π ′ ⟨y⟩, E1 ↿y ∶∶ [x v] ∶∶ E2 ) is closed
as wanted.
Point 3c is immediate, given that the environment and the dump are
already dual in s.
14.10
Proofs for Distillation is Complexity Preserving
Theorem 11.3.
1. LAM. As for the CEK, using the corresponding subterm
invariant and the following measure:
⎧
⎪
⎪∣u∣ + ∣w∣ if π = f (w, e′ ) ∶∶ π ′
#(u ∣ e ∣ π) ∶= ⎨
⎪
otherwise
⎪
⎩∣u∣
2. Split CEK. As for the CEK, using the corresponding subterm invariant
and the following measure:
⎧
⎪
⎪∣u∣ + ∣w∣ if π = (w, e′ ) ∶∶ π ′
#(u ∣ e ∣ π ∣ D) ∶= ⎨
⎪
otherwise
∣u∣
⎪
⎩
63
| 6cs.PL
|
Model Checking Gene Regulatory Networks
?
Mirco Giacobbe∗ , Călin C. Guet∗ , Ashutosh Gupta∗† ,
Thomas A. Henzinger∗ , Tiago Paixão∗ , and Tatjana Petrov∗
arXiv:1410.7704v2 [cs.CE] 16 Jan 2015
IST Austria, Austria∗
TIFR, India†
Abstract. The behaviour of gene regulatory networks (GRNs) is typically analysed using simulation-based statistical testing-like methods. In
this paper, we demonstrate that we can replace this approach by a formal
verification-like method that gives higher assurance and scalability. We
focus on Wagner’s weighted GRN model with varying weights, which is
used in evolutionary biology. In the model, weight parameters represent
the gene interaction strength that may change due to genetic mutations.
For a property of interest, we synthesise the constraints over the parameter space that represent the set of GRNs satisfying the property. We
experimentally show that our parameter synthesis procedure computes
the mutational robustness of GRNs –an important problem of interest
in evolutionary biology– more efficiently than the classical simulation
method. We specify the property in linear temporal logics. We employ
symbolic bounded model checking and SMT solving to compute the space
of GRNs that satisfy the property, which amounts to synthesizing a set
of linear constraints on the weights.
1
Introduction
Gene regulatory networks (GRNs) are one of the most prevalent and fundamental
type of biological networks whose main actors are genes regulating other genes.
A topology of a GRN is represented by a graph of interactions among a finite set
of genes, where nodes represent genes, and edges denote the type of regulation
(activation or repression) between the genes, if any. In [21], Wagner introduced
a simple but useful model for GRNs that captures important features of GRNs.
In the model, a system state specifies the activity of each gene as a Boolean
value. The system is executed in discrete time steps, and all gene values are
synchronously and deterministically updated: a gene active at time n affects the
value of its neighbouring genes at time n + 1. This effect is modelled through two
kinds of parameters: threshold parameters assigned to each gene, which specify
?
This research was supported by the European Research Council (ERC) under grant
267989 (QUAREM), the Austrian Science Fund (FWF) under grants S11402-N23
(RiSE) and Z211-N23 (Wittgenstein Award), the European Union’s SAGE grant
agreement no. 618091, ERC Advanced Grant ERC-2009-AdG-250152, the People
Programme (Marie Curie Actions) of the European Union’s Seventh Framework
Programme (FP7/2007-2013) under REA grant agreement no. 291734, and the SNSF
Early Postdoc.Mobility Fellowship, the grant number P2EZP2 148797.
the strength necessary to sustain the gene’s activity, and weight parameters
assigned to pairs of genes, which denote the strength of their directed effect.
Some properties of GRNs can be expressed in linear temporal logic (LTL)(such
as reaching a steady-state), where atomic propositions are modelled by gene values. A single GRN may or may not satisfy a property of interest. Biologists
are often interested in the behavior of populations of GRNs, and in presence
of environmental perturbations. For example, the parameters of GRNs from a
population may change from one generation to another due to mutations, and
the distribution over the different GRNs in a population changes accordingly.
We refer to the set of GRNs obtained by varying parameters on a fixed topology as GRN space. For a given population of GRNs instantiated from a GRN
space, typical quantities of interest refer to the long-run average behavior. For
example, robustness refers to the averaged satisfiability of the property within a
population of GRNs, after an extended number of generations. In this context,
Wagner’s model of GRN has been used to show that mutational robustness can
gradually evolve in GRNs [10], that sexual reproduction can enhance robustness to recombination [1], or to predict the phenotypic effect of mutations [17].
The computational analysis used in these studies relies on explicitly executing
GRNs, with the purpose of checking if they satisfy the property. Then, in order
to compute the robustness of a population of GRNs, the satisfaction check must
be performed repeatedly for many different GRNs. In other words, robustness
is estimated by statistically sampling GRNs from the GRN space and executing
each of them until the property is (dis)proven. In this work, we pursue formal
analysis of Wagner’s GRNs which allows to avoid repeated executions of GRNs,
and to compute mutational robustness with higher assurance and scalability.
In this paper, we present a novel method for synthesizing the space of parameters which characterize GRNs that satisfy a given property. These constraints
eliminate the need of explicitly executing the GRN to check the satisfaction of the
property. Importantly, the synthesized parameter constraints allow to efficiently
answer questions that are very difficult or impossible to answer by simulation,
e.g. emptiness check or parameter sensitivity analysis. In this work, we chose
to demonstrate how the synthesized constraints can be used to compute the
robustness of a population of GRNs with respect to genetic mutations. Since
constraint evaluation is usually faster than executing a GRN, the constraints
pre-computation enables faster computation of robustness. This further allows
to compute the robustness with higher precision, within the same computational
time. Moreover, it sometimes becomes possible to replace the statistical sampling
with the exact computation of robustness.
In our method, for a given GRN space and LTL property, we used SMT
solving and bounded model checking to generate a set of constraints such that a
GRN satisfies the LTL property if and only if its weight parameters satisfy the
constraints. The key insight in this method is that the obtained constraints are
complex Boolean combinations of linear inequalities. Solving linear constraints
has been the focus of both industry and academia for some time. However, the
technology for solving linear constraints with Boolean structure, namely SMT
solving, has matured only in the last decade [3]. This technology has enabled us
to successfully apply an SMT solver to generate the desired constraints.
We have built a tool which computes the constraints for a given GRN space
and a property expressed in a fragment of LTL. In order to demonstrate the
effectiveness of our method, we computed the robustness of five GRNs listed
in [8], and for three GRNs known to exhibit oscillatory behavior. We first synthesized the constraints and then we used them to estimate robustness based
on statistical sampling of GRNs from the GRN space. Then, in order to compare the performance with the simulation-based methods, we implemented the
approximate computation of robustness, where the satisfiability of the property
is verified by executing the GRNs explicitly. The results show that in six out
of eight tested networks, the pre-computation of constraints provides greater
efficiency, performing up to three times faster than the simulation method.
Related Work Formal verification techniques are already used for aiding various aspects of biological research [11,13,16,15,22]. In particular, the robustness
of models of biochemical systems with respect to temporal properties has been
studied [18,6,19]. Our work is, to the best of our knowledge, the first application of formal verification to studying the evolution of mutational robustness
of gene regulatory networks and, as such, it opens up a novel application area
for the formal verification community. As previously discussed, with respect to
related studies in evolutionary biology, our method can offer a higher degree of
assurance, more accuracy, and better scalability than the traditional, simulationbased approaches. In addition, while the mutational robustness has been studied
only for invariant properties, our method allows to compute the mutational robustness for non-trivial temporal properties that are expressible in LTL, such as
bistability or oscillations between gene states.
1.1
Motivating example
In the following, we will illustrate the main idea of the paper on an example of a
GRN space T generated from the GRN network shown in Fig. 1(a). Two genes
A and B inhibit each other, and both genes have a self-activating loop. The
parameters (iA , iB ) represent constant inputs, which we assume to be regulated
by some other genes that are not presented in the figure. Each of the genes is assigned a threshold value (tA , tB ), and each edge is assigned a weight (wAA ,wAB ,
wBA , wBB ). The dynamics of a GRN-individual chosen from T depends on these
parameters. Genes are in either active or inactive state, which we represent with
Boolean variables. For a given initial state of all genes, and for fixed values of
weights and thresholds, the values of all genes evolve deterministically in discrete
time-steps that are synchronized over all genes. Let a (resp. b) be the Boolean
variable representing the activity of gene A (resp. B). We denote a GRN state by
a pair (a, b). Let τ be the function that governs the dynamics of G (see Def. 3):
τ (a, b) = (iA + awAA − bwBA > tA , iB + bwBB − awAB > tB )
The next state of a gene is the result of arithmetically adding the influence from
other active genes.
10+3
01
iA
wAA
A
tA
wAB
wBA
(a)
α0 , β¯1
iB
wBB α¯0 , β¯1
B
tB
α¯0 , β1
misa
10+2
00
time[s]
α0 , β1
10+1
10+0
10
11
α1 , β0
(b)
10-1
10+5
evaluation
execution
10+6
# of samples
10+7
(c)
Fig. 1: Motivating example. a) The topology of a GRN with mutual inhibition of
genes A and B. b) The labelled transition system where labels denote the linear
constraints which enable the transition. α0 , α1 , β0 , and β1 denote linear atomic
formulas iA − wBA ≤ tA , iB + wBB > tB , iB − wAB ≤ tB , and iA + wAA > tA
respectively. c) Run-times comparison between execution method (dashed lines)
and our evaluation method (solid lines).
The topology of mutually inhibiting pair of genes is known to be bistable:
whenever one gene is highly expressed and the other is barely expressed, the
system remains stable [14,19]. The bistability property can be written as the
following LTL formula (see Def. 4):
(A ∧ ¬B =⇒ (A ∧ ¬B)) ∧ (¬A ∧ B =⇒ (¬A ∧ B)).
Let us fix values for parameters tA = tB = 0.6, wAB = wBA = wBB = 0.3,
and iA = iB = 32 . Then, we can check that a GRN is bistable by executing
the GRN. Indeed, for the given parameter values, the states (0, 1) and (1, 0) are
fixed points of τ . In other words, the GRN with those parameter values have
two stable states: if they start in state (0, 1) (resp. (1, 0)), they remain there.
Now let us choose iA = 23 , iB = 13 . Again, by executing the GRN, we can
conclude that it does not satisfy the property: at state (0, 1), B does not have
sufficiently strong activation to surpass its threshold and the system jumps to
(0, 0). Intuitively, since the external activation of B is too small, the phenotype
has changed to a single stable state. In general, it is not hard to inspect that
the bistability property will be met by any choice of parameters satisfying the
following constraints:
{iA − wBA ≤ tA , iA + wAA > tA , iB − wAB ≤ tB , iB + wBB > tB }.
(1)
Let’s now suppose that we want to compute the robustness of T in presence
of variations on edges due to mutations. Then, for each possible value of parameters, one needs to verify if the respective GRN satisfies the property. Using the
constraints (1), one may verify GRNs without executing them.
Our method automatizes this idea to any given GRN topology and any property specified in LTL. We first encode T as a parametrized labelled transition
system, partly shown in Fig. 1(b). Our implementation does not explicitly construct this transition system, nor executes the GRNs (the implementation is
described in Section 5). Then, we apply symbolic model checking to compute
the constraints which represent the space of GRN’s from T that satisfy the
bi-stability property.
To illustrate the scalability of our method in comparison with the standard
methods, in Fig. 1(c), we compare the performance of computing the mutational robustness with and without precomputing the constraints (referred to
as evaluation and execution method respectively). We choose a mutation model
such that each parameter takes 13 possible values distributed according to the
binomial distribution (see Appendix for more details on the mutation model).
We estimate the robustness value by statistical sampling of the possible parameter values. For a small number of samples, our method is slower because we
spend extra time in computing the constraints. However, more samples may
be necessary for achieving the desired precision. As the number of samples increases, our method becomes faster, because each evaluation of the constraints
is two times faster than checking bistability by executing GRN-individuals. For
1.2 × 105 many simulations, execution and evaluation methods take same total
time, and the robustness value estimated from these many samples lies in the
interval (0.8871, 0.8907) with 95% confidence. Hence, for this GRN, if one needs
better precision for the robustness value, our method is preferred.
One may think that for this example, we may compute exact robustness because the number of parameter values is only 136 (four weights and two inputs).
For simplicity of illustration, we chose this example, and we later present examples with a much larger space of parameters, for which exact computation of
robustness is infeasible.
2
Preliminaries
In this section, we start by defining a GRN space, which will serve to specify
common features for GRNs from the same population. These common features
are types of gene interactions (topology), constant parameters (thresholds), and
ranges of parameter values that are subject to some environmental perturbation
(weights). Then, we formally introduce a model of an individual GRN from the
GRN space and temporal logic to express its properties.
2.1
Basic notation
R≥0 (resp. Q≥0 ) is the set of non-negative real (resp. rational) numbers. For
m < n, let m..n denote the set of integers from m to n. With abuse of notation,
we treat finite maps with ordered domain as vectors with size of the map.
Let V be an infinite set of variable names. For rationals k1 , . . . , kn , a rational
variable vector v = (v1 , . . . , vn ), and a rational t, let k1 v1 + · · · + kn vn + t denote
a linear term. Let k1 v1 + · · · + kn vn + t > 0 and k1 v1 + · · · + kn vn + t ≥ 0 be a
strict and non-strict inequality over v respectively. Let linear(v) be the set of all
the (non-)strict inequalities over v. Let polyhedra(v) be the set of all the finite
conjunctions of the elements of linear(v).
2.2
GRN space
The key characteristics of the behaviour of a GRN are typically summarised
by a directed graph where nodes represent genes and edges denote the type of
regulation between the genes. A regulation edge is either activation (one gene’s
activity increases the activity of the other gene) or repression (one gene’s activity
decreases the activity of the other gene) [20]. In Wagner’s model of a GRN, in
addition to the activation types between genes, each gene is assigned a threshold
and each edge (pair of genes) is assigned a weight. The threshold of a gene
models the amount of activation level necessary to sustain activity of the gene.
The weight on an edge quantifies the influence of the source gene on destination
gene of the edge.
We extend the Wagner’s model by allowing a range of values for weight
parameters. We call our model GRN space, denoting that all GRNs instantiated
from that space share the same topology, and their parameters fall into given
ranges. We assume that each gene always has some minimum level of expression
without any external influence. In the model, this constant input is incorporated
by a special gene which is always active, and activates all other genes from
the network. The weight on the edge between the special gene and some other
gene represents the minimum level of activation. The minimal activation is also
subject to perturbation.
Definition 1 (GRN space) A GRN space is a tuple
T = (G, gin ,
,
, t, wmax , W ),
where
– G = {g1 , . . . , gd } is a finite ordered set of genes,
– gin ∈ G is the special gene used to model the constant input for all genes,
–
⊆ G × G is the activation relation such that ∀g ∈ G \ {gin } (gin , g) ∈
and ∀g (g, gin ) ∈
/ ,
⊆ G × G is the repression relation such that ∩ = ∅ ∧ ∀g (g, gin ) ∈
/ ,
–
– t : G → Q is the threshold function such that ∀g ∈ G \ {gin } t(g) ≥ 0 and
t(gin ) < 0,
– wmax : ( ∪ ) → Q≥0 is the maximum value of an activation/repression,
– W = P(( ∪ ) → Q≥0 ) assigns a set of possible weight functions to each
activation/inhibition relation, so that w ∈ W ⇒ ∀(g, g 0 ) ∈ ∪ w(g, g 0 ) ≤
wmax (g, g 0 ).
In the following text, if not explicitly specified otherwise, we will be referring
to the GRN space T = (G, gin , , , t, wmax , W ).
2.3
GRN-individual
Definition 2 (GRN-individual) A GRN-individual G is a pair (T, w), where
w ∈ W is a weight function from the GRN space.
A state σ : G → B of a GRN-individual G = (T, w) denotes the activation
state of each gene in terms of a Boolean value. Let Σ(G) (resp. Σ(T )) denote the
set of all states of G (resp. T ), such that σ(gin ) = true. The GRN model executes
in discrete time steps by updating all the activation states synchronously and
deterministically according to the following rule: a gene is active at next time if
and only if the total influence on that gene, from genes active at current time,
surpasses its threshold.
Definition 3 (Semantics of a GRN-individual) A run of a GRN-individual
G = (T, w) is an infinite sequence of states σ0 , σ1 , . . . such that σn ∈ Σ(G) and
τ (σn ) = σn+1 for all n ≥ 0, where τ : Σ(G) → Σ(G) is a deterministic transition
function defined by
X
X
w(g)
−
w(g)
> t(g 0 ) .
(2)
τ (σ) := λg 0.
{g|σ(g)∧(g,g 0 )∈
}
{g|σ(g)∧(g,g 0 )∈
}
The language of G, denoted by JGK, is a set of all possible runs of G. Note that
a GRN-individual does not specify the initial state. Therefore, JGK may contain
more than one run.
2.4
Temporal properties
A GRN exists in a living organism to exhibit certain behaviors. Here we present
a linear temporal logic (LTL) to express the expected behaviors of GRNs.
Definition 4 (Syntax of Linear Temporal properties) The grammar of the
language of temporal linear logic formulae are given by the following rules
ϕ ::= g | (¬ϕ) | (ϕ ∨ ϕ) | (ϕUϕ),
where g ∈ G is a gene.
Linear temporal properties are evaluated over all (in)finite runs of states from
Σ(G). Let us consider a run r = σ1 , σ2 , σ3 , · · · ∈ Σ(G)∗ ∪ Σ(G)∞ . Let ri be the
suffix of r after i states and ri is the ith state of r. The satisfaction relation |=
between a run and an LTL formula is defined as follows:
r |= g if r1 (g),
i
r |= ¬ϕ if r 6|= ϕ,
j
r |= ϕ1 ∨ ϕ2 if r |= ϕ1 or r |= ϕ2 ,
r |= (ϕ1 Uϕ2 ) if ∃i.r |= ϕ2 and ∀j ≤ i.r |= ϕ1 .
Note that if |r| < i then ri has no meaning. In such a situation, the above
semantics returns undefined, i.e., r is too short to decide the LTL formula. We
say a language L |= ϕ if for each run r ∈ L, r |= ϕ, and a GRN G |= ϕ if
LJGK |= ϕ. Let ♦ϕ be shorthand of trueUϕ and ϕ be shorthand of ¬♦¬ϕ.
Note that we did not include next operator in the definition of LTL. This is
because a typical GRN does not expect something is to be done in strictly next
cycle.
3
Algorithm for parameter synthesis
In this section we present an algorithm for synthesizing the weights’ space corresponding to a given property in linear temporal logic. The method combines
LTL model checking [2] and satisfiability modulo theory (SMT) solving [4].
The method operates in two steps. First, we represent any GRN-individual
from the GRN space with a parametrized transition system. In this system,
a transition exists between every two states, and it is labelled by linear constraints, that are necessary and sufficient constraints to enable that transition
in a concrete GRN-individual (for example, see Fig. 1b)). We say that a run
of the parametrized transition system is feasible if the conjunction of all the
constraints labelled along the run is satisfiable. Second, we search for all the feasible runs that satisfy the desired LTL property and we record the constraints
collected along them. The disjunction of such run constraints fully characterizes
the regions of weights which ensure that LTL property holds in the respective
GRN-individual.
Definition 5 (Parametrized transition system) For a given GRN space T
and rational parameters map v : G → V , the parametrized transition system
(T, v) is a labelled transition system (Σ(T ), Φ), where the labeling of edges
Φ : Σ(T ) × Σ(T ) → polyhedra(v) is defined as follows:
Φ := λσσ 0 .
^
g 0 ∈G
X
v(g)
0
{g|σ(g)∧(g,g )∈
−
X
v(g) > t(g 0 ) ⇐⇒ σ 0 (g 0 ) .
} {g|σ(g)∧(g,g 0 )∈
}
Φ(σ, σ 0 ) says that a gene g 0 is active in σ 0 iff the weighted sum of activation and
suppression activity of the regulators of g 0 is above its threshold.
A run of (T, v) is a sequence of states σ0 , σ1 , . . . such that σn ∈ Σ(T ) for all
n ≥ 0, and Φ(σ0 , σ1 ) ∧ Φ(σ1 , σ2 ) ∧ . . . is said to be the run constraint of the run.
A run is feasible if its run constraint is satisfiable. We denote by J(T, v)K the set
of feasible traces for (T, v). For a weight function w, let Φ(σ, σ 0 )[w/v] denote the
formula obtained by substituting v by w and let (T, v)[w/v] = (Σ(T ), Φ0 ), where
Φ0 (σ, σ 0 ) = Φ(σ, σ 0 )[w/v] for each σ, σ 0 ∈ Σ(T ).
In the following text, we refer to the parametrized transition system (T, v)
and an LTL property ϕ. Moreover, we denote the run constraint of run r =
σ0 , σ1 , . . . ∈ J(T, v)K by cons(r).
Lemma 1. For a weight function w, the set of feasible runs of (T, v)[w/v] is
equal to J(T, w)K.
The proof of the above lemma follows from the definition of the semantics for
GRN-individual. Note that the run constraints are conjunctions of linear (non)strict inequalities. Therefore, we may apply efficient SMT solvers to analyze
(T, v).
function GenCons((T, v) = (Σ(T ), Φ), ϕ)
begin
1 goodCons := true
2 for each σ ∈ Σ(T ) do
3 goodCons := GenConsRec(σ, true, goodCons, Σ(T ), Φ, ϕ)
4 done
5 return goodCons
end
function GenConsRec(run.σ, runCons, goodCons, Σ(T ), Φ, ϕ)
begin
6 if |run.σ| < |Σ(T )| then
for each σ 0 ∈ Σ(T ) do
7
runCons0 := runCons ∧ Φ(σ, σ 0 )
8
if goodCons ∧ runCons0 is sat then
9
if run.σσ 0 |= ¬ϕ then
(∗ check may return undef ∗)
10
goodCons := goodCons ∧ ¬runCons0
11
else
12
goodCons := GenConsRec(run.σσ 0 , runCons0 , goodCons, Σ(T ), Φ, ϕ)
13
14
done
15 return goodCons
end
Fig. 2: Counterexample guided computation of the mutation space feasible wrt.
ϕ. Let “.” be an operator that appends two sequences. run.σσ 0 |= ¬ϕ can be
implemented by converting ¬ϕ into a Büchi automaton and searching for an
accepting run over run.σσ 0 . However, a finite path may be too short to decide
whether ϕ holds or not. In that case, the condition at line 10 fails. Since (T, v)
is finite, the finite runs are bound to form lassos within |Σ(T )| steps. If a finite
run forms a lasso, then the truth value of run.σσ 0 |= ¬ϕ will be determined.
3.1
Constraint generation via model checking
Now our goal is to synthesize the constraints over v which characterise exactly
the set of weight functions w, for which (T, w) satisfies ϕ. Each feasible run
violating ϕ reports a set of constraints which weight parameters should avoid.
Once all runs violating ϕ are accounted for, the desired region of weights is
completely characterized. More explicitly, the desired space of weights is obtained
by conjuncting negations of run constraints of all feasible runs that satisfy ¬ϕ.
In Fig. 2, we present our algorithm GenCons for the constraint generation.
GenCons unfolds (T, v) in depth-first-order manner to search for runs which
satisfy ¬ϕ. At line 3, GenCons calls recursive function GenConsRec to do the
unfolding for each state in Σ(T ). GenConsRec takes six input parameters. The
parameter run.σ and runCons are the states of the currently traced run and its
respective run constraint. The third parameter are the constraints, collected due
to the discovery of counterexamples, i.e., runs which violate ϕ. The forth, fifth
and sixth parameter are the description of the input transition system and the
LTL property ϕ. Since GRN-individuals have deterministic transitions, we only
need to look for the lassos upto length |Σ(T )| for LTL model checking. Therefore,
we execute the loop at line 7 only if the run.σ has length smaller than |Σ(T )|. The
loop iterates over each state in Σ(T ). The condition at line 9 checks if run.σσ 0
is feasible and, if it is not, the loop goes to another iteration. Otherwise, the
condition at line 10 checks if run.σσ 0 |= ¬ϕ. Note that run.σσ 0 |= ¬ϕ may also
return undefined because the run may be too short to decide the LTL property. If
the condition returns true, we add negation of the run constraint in goodCons.
Otherwise, we make a recursive call to extend the run at line 13. goodCons
tells us the set of values of v for which we have discovered no counterexample.
GenCons returns goodCons at the end.
Since run constraints are always a conjunction of linear inequalities, goodCons
is a conjunction of clauses over linear inequalities. Therefore, we can apply efficient SMT technology to evaluate the condition at line 9. The following theorem
states that the algorithm GenCons computes the parameter region which satisfies property ϕ.
Theorem 1. For every weight function w ∈ W , the desired set of weight functions for which a GRN-individual satisfies ϕ equals the weight functions which
satisfy the constraints returned by GenCons:
(T, w) |= ϕ iff w |= GenCons((T, v), ϕ).
Proof. The
T statement amounts to showing that the sets A = {w | (T, w) |= ϕ}
and B = r∈J(T,v)K∧r|=¬ϕ {w | w |= ¬cons(r)} are equivalent. Notice that
W \ A = {w | ∃r ∈ J(T, v)K such that w |= cons(r) ∧ r |= ¬ϕ}
[
=
{w | w |= cons(r)}
r∈J(T,v)K∧r|=¬ϕ
=W \
\
{w | w |= ¬cons(r)}.
r∈J(T,v)K∧r|=¬ϕ
We use the above presentation of the algorithm for easy readability. However, our implementation of the above algorithm differs significantly from the
presentation. We follow the encoding of [7] to encode the path exploration as
a bounded-model checking problem. Further details about implementation are
available in Section 5. The algorithm has exponential complexity in the size of
T . However, one may view the above procedure as the clause learning in SMT
solvers, where clauses are learnt when the LTL formula is violated [23]. Similar
to SMT solvers, in practice, this algorithm may not suffer from the worst-case
complexity.
Example 1. The GRN osc3 (shown in Fig. 3) was the model of a pioneering
work in synthetic biology [12], and it is known to provide oscillatory behaviour:
each gene should alternate its expression between ‘on’ and ‘off’ state:
^
ϕ3 =
(v ⇒ ♦¬v) ∧ (¬v ⇒ ♦v).
v∈{A,B,C}
The solutions are the constraints:
(T, w) |= ϕ3 iff (iA > tA ) ∧ (iB > tB ) ∧ (iC > tC )∧
(iB − wAB ≤ tB ) ∧ (iC − wBC ≤ tC ) ∧ (iA − wCA ≤ tA ).
4
Computing Robustness
In this section, we present an application of our parameter synthesis algorithm,
namely computing robustness of GRNs in presence of mutations. To this end,
we formalize GRN-population and its robustness. Then, we present a method to
compute the robustness using our synthesized parameters.
A GRN-population models a large number of GRN-individuals with varying
weights. All the GRN-individuals are defined over the same GRN space, hence
they differ only in their weight functions. The GRN-population is characterised
by the GRN space T and a probability distribution over the weight functions. In
the experimental section, we will use the range of weights W and the distribution
π based on the mutation model outlined in the Appendix.
Definition 6 (GRN-population) A GRN-population Z is a pair (T, π), where
π : W → [0, 1] is a probability distribution over all weight functions from GRN
space T .
We write ϕ(Z) ∈ [0, 1] to denote an expectation that a GRN instantiated
from a GRN-population Z = (T, π) satisfies ϕ. The value ϕ(Z) is in the interval
[0, 1] and we call it robustness.
Definition 7 (Robustness) Let Z = (T, π) be a GRN-population, and ϕ be
an LTL formula which expresses the desired LTL property. Then, robustness of
Z with respect to property ϕ is given by
X
ϕ(Z) :=
π(w)
{w|J(T,w)K|=ϕ}
The above definition extends that of [10], because it allows for expressing any
LTL property as a phenotype, and hence it can capture more complex properties
such as oscillatory behaviour.
In the following, we will present an algorithm for computing the robustness,
which employs algorithm GenCons.
4.1
Evaluating robustness
Let us suppose we get a GRN-population Z = (T, π) and LTL property ϕ as
input to compute robustness.
For small size of GRN space T , robustness can be computed by explicitly
enumerating all the GRN-individuals from T , and verifying each GRN-individual
against ϕ. The probabilities of all satisfying GRN-individuals are added up.
However, the exhaustive enumeration of the GRN space is often intractable
due to a large range of weight functions W in T . In those cases, the robustness
is estimated statistically: a number of GRN-individuals are sampled from T
according to the distribution π, and the fraction of satisfying GRN-individuals
is stored. The sampling experiment is repeated a number of times, and the
mean (respectively variance) of the stored values are reported as robustness
(respectively precision).
Depending on how a sampled GRN-individual is verified against the LTL
property, we have two methods:
– In the first method, which we will call execution method, each sampled GRNindividual is verified by executing the GRN-individual from all initial states
and checking if each run satisfies ϕ;
– In the second method, which we will call evaluation method, the constraints
are first precomputed with GenCons, and each sampled GRN-individual is
verified by evaluating the constraints.
Clearly, the time of computing the constraints initially renders the evaluation
method less efficient. This cost is amortized when verifying a GRN-individual by
constraint evaluation is faster than by execution. In the experimental section, we
compare the overall performance of the two approaches on a number of GRNs
from literature.
5
Experimental results
We implemented a tool which synthesizes the parameter constraints for a given
LTL property (explained in Section 3), and the methods for computing the
mutational robustness (explained in Section 4.1). We ran our tool on a set of
GRNs from literature.
5.1
Implementation
Our implementation does not explicitly construct the parametrised transition
system described in Section 3 (Dfn. 5 and Alg. 2). Instead, we encode the
bounded model-checking (Alg. 2) as a satisfiability problem, and we use an
SMT solver to efficiently find goodCons. More concretely, we initially build a
formula which encodes the parametrized transition system and it is satisfied if
and only if some run of (T, v) satisfies ¬ϕ. If the encoding formula is satisfiable,
the constraints cons(r) along the run are collected, and ¬cons(r) is added to
goodCons. Then, we expand the encoding formula by adding ¬cons(r), so as to
rule out finding the same run again. We continue the search until no satisfying
assignment of the encoding formula can be found. The algorithm always terminates because the validity of the property is always decided on finite runs (as
explained in Section 3).
The implementation involves 8k lines of C++ and we use Z3 SMT solver as
the solving engine. We use CUDD to reduce the size of the Boolean structure of
Property
Space size
mi:
Y
Z
(Y Z̄ =⇒ Y Z̄) ∧ (Z Ȳ =⇒ Z Ȳ )
4225
misa:
Y
Z
(Y Z̄ =⇒ Y Z̄) ∧ (Z Ȳ =⇒ Z Ȳ )
105625
Y
Z
(Y S Z̄ R̄ =⇒ Y S Z̄ R̄) ∧
(Ȳ S̄ZR =⇒ Ȳ S̄ZR)
≈ 109
♦X ∨ ♦X̄
≈ 1010
(Y QS Z̄ P̄ R̄ =⇒ Y QS Z̄ P̄ R̄) ∧
(Ȳ Q̄S̄ZP R =⇒ Ȳ Q̄S̄ZP R)
≈ 1018
X1 =⇒ ♦X̄1 ∧ X̄1 =⇒ ♦X1 ∧
X2 =⇒ ♦X̄2 ∧ X̄2 =⇒ ♦X2 ∧
...
XN =⇒ ♦X̄N ∧ X̄N =⇒ ♦XN
3:274625
5:≈ 109
7:≈ 1012
qi:
R
S
X
cc:
S
ncc:
Z
R
Q
S
Y
T
Z
P
R
X1
oscN:
X2
XN
Fig. 3: GRN benchmarks. mi, misa (mutual inhibition), qi (quad inhibition),
and ncc (cell cycle switch) satisfy different forms of bistability. For the networks
ci (cell cycle switch), the value of gene eventually stabilizes [9]. In osc3, also
known as the repressilator [12], the gene values alternate. osc5 and osc7 (not
shown) are generalizations of osc3, and also exhibit oscilating behavior.
the resulting formula. We ran the experiments on a GridEngine managed cluster
system. Our tool, as well as the examples and the simulation results are available
online.1 .
5.2
Performance evaluation
We chose eight GRN topologies as benchmarks for our tool. The benchmarks are
presented in Fig. 3. The first five of the GRN topologies are collected from [8].
On these benchmarks we check for the steady-state properties. On the final three
GRN topologies, we check for the oscillatory behavior. The results are presented
in Fig. 4.
We ran the robustness computation by the evaluation and execution methods
(the methods are described in Section 4.1). In order to obtain robustness and to
estimate the precision, we computed the mean of 100 experiments, each containing a number of samples ranging from 103 to 106 . The total computation time
in the execution methods linearly depends on the number of samples used. The
total computation time in the evaluation method depends linearly on the number of samples, but initially needs time to compute the constraints. Technically,
1
http://pub.ist.ac.at/~mgiacobbe/grnmc.tar.gz
10+1
1.00
0.82
10+0
0.57
10-1
10+5
mi misa qi
cc ncc osc3osc5osc7
mi
10
10
# of samples
qi
10+0
10+5
+8
10
+6
10
+4
10
+7
10
# of samples
osc3
10
+6
10
10
+7
10
# of samples
osc5
10
evaluation
execution
+6
10
+7
10
# of samples
+8
10
10+7
# of samples
osc7
10+8
10+3
10+2
10+0
10+5
10
10
10+1
10+0
evaluation
execution
+6
+4
time[s]
time[s]
10+1
10+2
10+0
10+5
+8
10+3
10+2
10+8
10+1
evaluation
execution
+4
10+3
10+7
# of samples
ncc
10+3
10+2
10+0
10+5
+8
10
10
time[s]
time[s]
evaluation
execution
evaluation
execution
+6
+4
10+1
10+0
10-1
10+5
10
+7
10+3
10+1
10-1
10+5
+6
10
10+2
10+2
10+1
evaluation
execution
+4
10+3
time[s]
10+3
10+2
time[s]
time[s]
~ratio execution/constraint evaluation
1.62
cc
10+4
10+3
2.01
+4
time[s]
misa
10+4
3.08
10+2
10+1
evaluation
execution
+6
10
+7
10
# of samples
+8
10
10+0
10+5
evaluation
execution
+6
10
10+7
# of samples
10+8
Fig. 4: The comparison in performance when mutational robustness is statistically estimated, and a property check is performed either by evaluation, or
by execution (see Section 4.1 for the description of the two methods). The bar
(top-left) shows the ratio of average times needed to verify the property of one
sampled GRN-individual. For example, for osc7, the performance of evaluation
method is more than three times faster. The other graphs show how the robustness computation time depends on the total number of sampled GRNs (in order
to obtain robustness and to estimate the precision, we computed the mean of 100
experiments, each containing a number of samples ranging from 103 to 106 ). The
graph is shown in log-log scale. The non-linear slope of the evaluation method
is most notable in examples mcc and osc7, and it is due to the longer time used
for compute the constraints.
the time needed to compute robustness by execution method is tex = kex p, and
the time needed to compute robustness by evaluation approach tev = kev p + tc ,
where p represents the total number of samples used, tc is the time to compute
the constraints, and kex (resp. kev ) is the time needed to verify the property by
evaluation (resp. execution). We used linear regression to estimate the paramein top-left position of Fig. 4. The
ters kex and kev , and we present the ratio kkex
ev
results indicate that on six out of eight tested networks, evaluation is more efficient than execution. For some networks, such as osc7, the time for computing
the constraints is large, and the gain in performance becomes visible only once
the number of samples is larger than 106 .
6
Conclusion and discussion
In this paper, we pursued formal analysis of Wagner’s GRN model, which allows
symbolic reasoning about the behavior of GRNs under parameter perturbations.
More precisely, for a given space of GRNs and a property specified in LTL, we
have synthesized the space of parameters for which the concrete, individual GRN
from a given space satisfies the property. The resulting space of parameters is
represented by complex linear inequalities. In our analysis, we have encoded a
bounded model-checking search into a satisfiability problem, and we used efficient SMT solvers to find the desired constraints. We demonstrated that these
constraints can be used to efficiently compute the mutational robustness of populations of GRNs. Our results have shown the cases in which the computation
can be three times faster than the standard (simulation) techniques employed in
computational biology.
While computing mutational robustness is one of the applications of our
synthesized constraints, the constraints allow to efficiently answer many other
questions that are very difficult or impossible to answer by executing the sampled
GRNs. In our future work, we aim to work on further applications of our method,
such as parameter sensitivity analysis for Wagner’s model. Moreover, we plan to
work on the method for exact computation of robustness by applying the point
counting algorithm [5].
The Wagner’s model of GRN is maybe the simplest dynamical model of a
GRN – there are many ways to add expressiveness to it: for example, by incorporating multi-state expression level of genes, non-determinism, asynchronous
updates, stochasticity. We are planning to study these variations and chart the
territory of applicability of our method.
References
1. R. B. R. Azevedo, R. Lohaus, S. Srinivasan, K. K. Dang, and C. L. Burch. Sexual
reproduction selects for robustness and negative epistasis in artificial gene networks. Nature, 440(7080):87–90, Mar. 2006.
2. C. Baier and J.-P. Katoen. Principles of model checking. MIT Press, 2008.
3. C. Barrett, M. Deters, L. de Moura, A. Oliveras, and A. Stump. 6 years of SMTCOMP. Journal of Automated Reasoning, 50(3):243–277, 2013.
4. C. W. Barrett, R. Sebastiani, S. A. Seshia, and C. Tinelli. Satisfiability modulo
theories. Handbook of satisfiability, 185:825–885, 2009.
5. A. Barvinok and J. E. Pommersheim. An algorithmic theory of lattice points in
polyhedra. New perspectives in algebraic combinatorics, 38:91–147, 1999.
6. G. Batt, C. Belta, and R. Weiss. Model checking genetic regulatory networks with
parameter uncertainty. In Hybrid systems: computation and control, pages 61–75.
Springer, 2007.
7. A. Biere, A. Cimatti, E. M. Clarke, O. Strichman, and Y. Zhu. Bounded model
checking. Advances in computers, 58:117–148, 2003.
8. L. Cardelli. Morphisms of reaction networks that couple structure to function.
9. L. Cardelli and A. Csikász-Nagy. The cell cycle switch computes approximate
majority. Scientific reports, 2, 2012.
10. S. Ciliberti, O. C. Martin, and A. Wagner. Robustness can evolve gradually in
complex regulatory gene networks with varying topology. PLoS Computational
Biology, 3(2), 2007.
11. V. Danos and C. Laneve. Formal molecular biology. Theoretical Computer Science,
325(1):69–110, 2004.
12. M. B. Elowitz and S. Leibler. A synthetic oscillatory network of transcriptional
regulators. Nature, 403(6767):335–338, 2000.
13. J. Fisher and T. A. Henzinger. Executable cell biology. Nature biotechnology,
25(11):1239–1249, 2007.
14. T. S. Gardner, C. R. Cantor, and J. J. Collins. Construction of a genetic toggle
switch in escherichia coli. Nature, 403(6767):339–342, 2000.
15. S. K. Jha, E. M. Clarke, C. J. Langmead, A. Legay, A. Platzer, and P. Zuliani. A
bayesian approach to model checking biological systems. In Computational Methods
in Systems Biology, pages 218–234. Springer, 2009.
16. M. Kwiatkowska, G. Norman, and D. Parker. Using probabilistic model checking in
systems biology. ACM SIGMETRICS Performance Evaluation Review, 35(4):14–
21, 2008.
17. T. MacCarthy, R. Seymour, and A. Pomiankowski. The evolutionary potential of
the drosophila sex determination gene network. Journal of Theoretical Biology,
225(4):461468, 2003.
18. R. Mateescu, P. T. Monteiro, E. Dumas, and H. De Jong. Ctrl: Extension of
ctl with regular expressions and fairness operators to verify genetic regulatory
networks. Theoretical Computer Science, 412(26):2854–2883, 2011.
19. A. Rizk, G. Batt, F. Fages, and S. Soliman. A general computational method for
robustness analysis with applications to synthetic gene networks. Bioinformatics,
25(12):i169–i178, 2009.
20. T. Schlitt and A. Brazma. Current approaches to gene regulatory network modelling. BMC bioinformatics, 8(Suppl 6):S9, 2007.
21. A. Wagner. Does evolutionary plasticity evolve? Evolution, 50(3):1008–1023, 1996.
22. B. Yordanov, C. M. Wintersteiger, Y. Hamadi, and H. Kugler. SMT-based analysis
of biological computation. In NASA Formal Methods, pages 78–92. Springer, 2013.
23. L. Zhang, C. F. Madigan, M. H. Moskewicz, and S. Malik. Efficient conflict driven
learning in a boolean satisfiability solver. In Proceedings of the 2001 IEEE/ACM
international conference on Computer-aided design, pages 279–285. IEEE Press,
2001.
1
p
p
3
A
a)
p
3
p
3
C
1
T
…
p
1
p
b)
no
mutati
on
G
p
3
mutati
on
p
3
Fig. 5: A model of changes of a single nucleotide from one generation to another
Mutation model
In this section, we present the mutation model that is grounded in evolutionary
biology research. Genetic mutations refer to events of changing base pairs in the
DNA sequence of the genes and they may disturb the regular functioning of the
host cell. Such mutations affect the weight functions and we assume their range
to be between a maximum weight and zero.
Modeling mutations in a nucleotide. The mutations of a single nucleotide
follow a discrete-time Markov chain (DTMC), illustrated in Fig. 5a). When the
DNA is passed from the mother to the daughter cell, a ‘correct’ nucleotide (in
figure it is the A) can mutate to each different value (T, C or G) with some
probability. In the model in Fig. 5a), the probability of each possible mutation is
p
3 , and hence, the probability of retaining the same nucleotide is 1−p. In Fig. 5b),
there is a process where all mutated states are lumped together - the probability
to get to the ‘correct’ nucleotide is equal to p3 , and to remain mutated is therefore
(1 − p3 ). We will refer to the general probabilities of the lumped process with a
matrix
p p
P id = 00 01 ,
p10 p11
with the intuition 0 being the non-mutated state and 1 the mutated state. Let
Xni ∈ {0, 1} be a process (DTMC) reporting whether the i-th nucleotide is
in its ‘correct’ configuration, or mutated configuration at n-th generation. The
number P (Xni = 1) can be interpreted also as that the fraction of mutated i-th
nucleotides observed in the population at generation n. Notably, for the values
shown in Fig. 5b), the stationary distribution of Xni is (0.75, 0.25), independently
of the probability of a single mutation p.
Modeling mutations in a single gene. Let us assume l to be the length of
the DNA binding region of a gene 2 g, and P id to be specify the probabilities of
single point mutations occurring in it. Moreover, we assume that the maximal
2
More precisely the length of the promoter of gene
weight for influence of gene g to g 0 , i.e. w(g, g 0 ), is achieved for a single sequence
of nucleotides, and that the weight will linearly decrease with the number of
mutated nucleotides (where ‘mutated’ refers to being different to the sequence
achieving the maximal weight). Therefore, if k out of l nucleotides in the promoter region of g 0 that is affected by g are mutated, the weight w(g, g 0 ) becomes
wmax (g, g 0 )(1 − kl ).
If the whole promoter sequence of one gene is modelled by a string a ∈
{A, T, C, G}l , its changes over generations are captured by an l-dimensional random process (Xn1 , . . . , Xnl )(a) ∈ {0, 1}l . Mutation events Xn1 , Xn2 ,. . . , Xnl are
assumed to happen independently within the genome, and independently of the
history of mutations. Then, a random process Mn := Xn1 + . . . + Xnl , such that
Mn = k ∈ {0, . . . , l}, denotes that the configuration at generation n differs from
the optimal configuration in exactly k points. The following lemma defines the
values of the transition matrix of the process {Mn }.
Lemma 2. The process {Mn } is a DTMC, with transition probabilities P (Mn+1 =
j | Mn = i) = T (i, j), where T : 0..l × 0..l → [0, 1] amounts to:
T (i, j) :=
i u i−u l − i j−u l−i−(j−u)
p p
p
p
.
u 11 10 j − u 01 00
min{i,j}
X
u=0
(3)
Moreover, {Mn } converges to a unique stationary distribution, which is a
i
binomial,
probability β = limn→∞ P (Xn = 1). In a special case
with success
1−p p
3
P̂ id =
p , the stationary distribution is β = 4 , independently of the
p
1
−
3
3
value of p.
Proof. (Lem. 2) It suffices to observe that u represents the number of mutated
nucleotides which remain mutated at time (n + 1), and (j − u) is the number of
unmated nucleotides which become mutated at time (n + 1). The existence and
convergence of the stationary trivially follows from that {Mn } is regular (irreducible and aperiodic). Each process {Xni } at a stationary behaves as a Bernoulli
process, with probability of being mutated equal to β = limn→∞ P (Xni = 1),
and a probability of remaining un-mutated 1 − β. The claim follows as the sum
of l independent Bernoulli processes is binomially distributed. The special case
follows because matrix P̂ id has a unique stationary distribution (0.25, 0.75) (denoting by a and b its coordinates, we obtain equations a(1 − p) + b p3 = a and
ap + b(1 − p3 ) = b, which can be satisfied only if a = 3b ).
Remark that in a special case when p00 = 0.25, the transition matrix of the
process {Mt } has the transition probabilities
P (Kn+1 = j | Kn = i) = pj11 pl−j
00
i
X
i
l−i
u j−u
u=0
l
=
0.75j 0.25l−j ,
j
which do not depend on i.
The special case of matrix T for l = 2 is illustrated in Fig. 6.
0 mutations 1 mutation
2 mutations
2 2
3
p00
2p00 p01
p201
0 mutations
4p00 p10 p00 p11 + p01 p10 p01 p11 5
1 mutation
2 mutations
p210
2p10 p11
p211
Fig. 6: The transition matrix for the evolutionary process of one gene with two
nucleotides. The transition probabilities are computed based on the assumption
that the single nucleotide mutations happen independently across the genome:
for example, the transition from zero to two mutation is equal to p201 , because it
occurs iff both nucleotides transition from a correct to a mutated state.
Modeling mutations in a GRN. We assume that mutations happen independently across the genome, and therefore, if the genome is in the configuration
k = (k1 , . . . , kd ), the probability that the next generation will be in the configuration k0 = (k10 , . . . , kd0 ) will be a product of transition probabilities from ki to
ki0 , for i = 1, . . . , d.
Each sequence k = (k1 , . . . , kd ) ∈ 0..l1 ×. . .×0..ld defines one weight function
k
wk , with wk (gi , gj ) = wmax (gi , gj )(1− ljj ). Hence, the domain of weight functions
is given by
W = {wk | k ∈ 0..l1 × . . . × 0..ld },
and the distribution is given by
π(wk ) =
d
Y
d
β ki (1 − β)d−ki .
k
i
i=1
| 5cs.CE
|
Deterministic Perturbations For Simultaneous Perturbation Methods
Using Circulant Matrices.
arXiv:1702.06250v2 [cs.SY] 15 Mar 2017
Chandramouli K ] , D. Sai Koti Reddy† , Shalabh Bhatnagar‡
Abstract— We consider the problem of finding optimal
parameters under simulation optimization setup. For a pdimensional parameter optimization, the classical KieferWolfowitz Finite Difference Stochastic Approximation (FDSA)
scheme uses p + 1 or 2p simulations of the system feedback
for one-sided and two-sided gradient estimates respectively.
The dependence on the dimension p makes FDSA impractical
for high dimensional problems. An alternative approach for
gradient estimation in high dimensional problems is the simultaneous perturbation technique. The main idea in this approach
is to estimate the gradient by using only two settings of the
p-dimensional parameter being optimized. The two settings
of the parameter are obtained by simultaneously perturbing
all the components of the parameter by adding a random
direction. A drawback of using random directions for the
gradient estimate is the very large or possibly infinite range
of these random directions (for e.g. ±1 symmetric Bernoulli
perturbations typically used in 1SPSA algorithm has a range
of cardinality 2p ). In this article we consider deterministic
perturbations with a range of cardinality p + 1 to improve
the convergence of these algorithms. A novel construction of
deterministic perturbations based on specially chosen circulant
matrix is proposed. Convergence analysis of the proposed
algorithms is presented along with numerical experiments.
I. INTRODUCTION
Simulation optimization problems frequently arise in engineering disciplines like transportation systems, machine
learning, service systems, manufacturing etc. Practical limitations, lack of model information and the large dimensionality
of these problems prohibit analytic solution of these problems and simulation is often employed to evaluate a system in
these disciplines. Hence simulation budget becomes critical
and one aims to converge to optimal parameters using as few
simulations as possible.
To state the problem formally, consider a system where
noise-corrupted feedback of the system is available i.e.,
given the system parameter θ the feedback that is available is h(θ, ξ) where ξ is the noise term inherent in the
system, pictorially shown in Figure 1. The objective in
these problems is to find a parameter vector that gives
the optimal expected performance of the system. Suppose
J(θ) = Eξ [h(θ, ξ)] then h(θ, ξ) = J(θ) + M (θ, ξ) where
M (θ, ξ) = h(θ, ξ) − Eξ [h(θ, ξ)] is a mean zero term. Hence,
one needs to solve the following optimization problem:
Find θ∗ = arg minp J(θ).
θ∈R
(1)
Analogous to deterministic optimization problems, where explicit analytic gradient of the objective function is available,
a solution approach could be to devise an algorithm that
mimics the familiar gradient descent algorithm. However in
our setting only noise corrupted samples of the objective are
available. So one essentially needs to estimate the gradient
of the objective function using simulation samples.
This adaptive system optimization framework has its roots
in the work by Robbins and Monro [1] and Kiefer and
Wolfowitz [2]. The work by Kiefer and Wolfowitz uses
stochastic approximation framework by Robbins and Monro
to optimize an objective using noisy samples. In their work
Kiefer and Wolfowitz [2] use one-sided or two-sided approximation of the gradient. This method has the drawback of
using p + 1 simulations for one-sided approximation and 2p
simulations for two-sided approximation of the gradient for a
p-dimensional problem. Later works [3],[4] have replaced the
gradient approximation using finite differences with random
perturbations.
Simultaneous perturbation is a useful technique to estimate
the gradient from the function samples, especially in high
dimensional problems - see [5], [6] for a comprehensive
treatment of this subject matter. The first-order Simultaneous
Perturbation Stochastic Approximation algorithm for simulation optimization, henceforth referred to as 1SPSA-2R, was
proposed in [4]. 1SPSA-2R uses two simulations per iteration
and random perturbations to perturb the parameter vector. A
one measurement variant of [4] is proposed in [7] which
we refer here as 1SPSA-1R. A closely related algorithm is
the first-order Random Directions Stochastic Approximation
henceforth referred to as 1RDSA-2R that appears in [3,
pp. 58-60]. The algorithm 1RDSA-2R differs from 1SPSA2R, both in the construction as well as convergence analysis
and performs poorly compared to 1SPSA-2R, see [8] for a
detailed comparative study.
To enhance the performance of 1SPSA-2R and 1SPSA-1R,
[9] proposed deterministic perturbations based on Hadamard
matrices. We refer here the variants of 1SPSA-2R and
1SPSA-1R that use Hadamard perturbations as 1SPSA-2H
and 1SPSA-1H. Analogous to their work we propose deterministic perturbations for 1RDSA-2R and its one simulation
variant (referred as 1RDSA-1R here) based on a specially
chosen circulant matrix.
II. M OTIVATION
Department of Computer Science and Automation, Indian Institute
of Science, Bangalore, E-Mail: ] [email protected].
† [email protected]. ‡ [email protected].
? Supported by Robert Bosch Center for Cyber-Physical Systems, IISc.
Here we motivate the necessary conditions that deterministic perturbations should satisfy. Analogous properties for
Hadamard matrix based perturbations are well motivated in
feedback
θ
System simulator
h(θ, ξ)
Fig. 1: Simulation Optimization Model
[5]. Both 1SPSA-2R and 1RDSA-2R iteratively update the
parameter vector and a typical update step is of the following
form:
b
θn+1 = θn − an ∇J(θ
n ),
(2)
where an is the step-size that satisfies standard stochastic
b
approximation conditions (see (A2) in section IV) and ∇J(θ)
is an estimate of the gradient of the objective function J.
Thus, (2) can be considered as the stochastic version of the
well-known gradient descent method for optimization.
Now consider the Taylor series expansion of J(θn +δn dn )
around θn , we have
J(θn + δn dn ) = J(θn ) + δn dTn ∇J(θn ) + o(δn ).
(3)
Algorithm 1 Basic structure of 1RDSA-2C and 1RDSA-1C
algorithms.
Input:
p
• θ0 ∈ R , initial parameter vector;
• δn , n ≥ 0, a sequence to approximate gradient;
• Matrix of perturbations
p
Q = p + 1[H −1/2 , −H −1/2 u],
with u = [1, 1, · · · , 1]T ;
• noisy measurement of cost objective J;
• an , n ≥ 0, step-size sequence satisfying assumption
(A2) of section IV;
for n = 1, 2, . . . nend do
Let dn be mod(n, p + 1)th column of Q.
Update the parameter as follows:
b
θn+1 = θn − an ∇J(θ
n ),
(11)
b
where ∇J(θ
n ) is chosen according to (13) or (14).
end for
Return θnend .
Similarly, an expansion of J(θn − δn dn ) around θn gives
J(θn − δn dn ) = J(θn ) − δn dTn ∇J(θn ) + o(δn ).
(4)
From (3) and (4), we have
J(θn + δn dn ) − J(θn − δn dn )
dn − ∇J(θn )
2δn
= (dn dTn − I)∇J(θn ) + o(δn ).
n=s+1
(5)
(6)
dn = 0.
n=s+1
(8)
(9)
III. S TRUCTURE OF THE ALGORITHM
− I)∇J(θn ) constitutes the bias in the
Note that
gradient estimate and from the assumption (A3) in section
IV
"
#
(7)
In the case of a one-simulation algorithm with parameter
θn + δn dn , a similar Taylor series expansion gives
J(θn + δn dn )
dn − ∇J(θn )
δn
J(θn )
dn + (dn dTn − I)∇J(θn ) + O(δn ).
=
δn
(P2)
s+P
P
In the following sections we construct deterministic perturbations which satisfy (P1) and (P2) and provide convergence
results for the resulting 1RDSA-2C and 1RDSA-1C algorithms.
The rest of the paper is organized as follows: In section
III, we describe the various ingredients in algorithm 1 and
the construction of perturbations dn . In section IV, we
present the convergence results for 1RDSA-2C and 1RDSA1C algorithms. In section V, we present the results from
numerical experiments and finally, in section VI, we provide
the concluding remarks.
(dn dTn
E (dn dTn − I)∇J(θn ) θn = 0.
(P1) Let Dn := dn dTn − I. For any s ∈ N there exists a
s+P
P
P ∈ N such that
Dn = 0 and,
One expects the following to hold in addition to (7) in the
case of random perturbations for one simulation algorithms:
"
#
J(θn )
E
dn θn = 0.
(10)
δn
Notice that (7) and (10) are achieved asymptotically for
a random perturbation direction dn , for example with ±1
symmetric Bernoulli components, dn has to cycle through
2p possible vectors in the range to achieve bias cancellation.
Clearly one is motivated to look for perturbations that satisfy
similar properties and have a smaller cycle length for faster
bias cancellation and thereby improve the performance of the
algorithm. Analogous to (7) and (10) we expect deterministic
sequence of perturbations to satisfy the following properties:
Algorithm 1 presents the basic structure of 1RDSA-2C
and 1RDSA-1C. We describe the individual components of
algorithm 1 below.
A. Deterministic Pertubations
Let δn , n ≥ 0 denote a sequence of diminishing positive
real numbers satisfying assumption A2 in section IV and
dn = (d1n , . . . , dpn )T denote (p + 1) mod n column of Q.
where Q is constructed as follows. Let H be the p × p
dimensional matrix defined as:
2 1 1···1
1 2 1···1
H=
(12)
.
.. .. .. ..
....
1 1 1···2
Observe that H = I + uuT , where u = [1, 1, · · · , 1]T , is a
−1/2
positive definite matrix. Hence
is well defined. Define
√H
p×(p+1) matrix Q as Q = p + 1[H −1/2 , −H −1/2 u]. The
columns of Q are used as perturbations dn .
Let yn+ , yn− be defined as below. yn+ = J(θn +δn dn )+Mn+
and yn− = J(θn − δn dn ) + Mn− , where the noise terms
Mn+ , Mn− satisfy E [ Mn+ − Mn− | Fn ] = 0 with Fn =
σ(θm , m ≤ n) denoting the underlying sigma-field. The
estimate of the gradient ∇J(θn ) is given by
+
(yn − yn− )dn
b
for 1RDSA-2C
(13)
∇J(θn ) =
2δn
b
∇J(θ
n) =
Lemma 1. Let I be a p × p identity matrix and u =
1
1
..
.
1
be a p × 1 vector of 1s, then
B. Gradient estimation
(yn+ )dn
δn
for 1RDSA-1C
(14)
Observe that while 1RDSA-2C algorithm uses two function
samples yn+ and yn− at θn + δn dn and θn − δn dn , 1RDSA-1C
algorithm uses only one function sample yn+ at θn + δn dn .
(I + uuT )−1/2 = I −
Proof. It is enough to show that
"
#2
uuT
uuT
T
= I.
(I + uu ) I −
+ p
p
p (1 + p)
Using kuk2 = uT u = p we have
"
#2
uuT
uuT
T
(I + uu ) I −
+ p
p
p (1 + p)
"
uuT
uuT
+ p
= (I + uuT ) I −
p
p (1 + p)
IV. CONVERGENCE ANALYSIS
We make the same assumptions as those used in the
analysis of [4], with a few minor alterations. The assumptions
are listed below. Also k.k denotes 2-norm.
(A1) The map J : Rp → R is Lipschitz continuous and is
differentiable with bounded second order derivatives.
Further, the map L : Rp → R defined as L(θ) =
−∇J(θ) is Lipschitz continuous.
(A2) The step-size sequences an , δn > 0, ∀n satisfy
X
X an 2
an , δn → 0,
an = ∞,
< ∞.
δn
n
n
a
Further, anj → 1 as n → ∞, for all j ∈ {n, n + 1, n +
2 · · · , n + M } for any given M > 0 and bn = aδnn is
b
such that bnj → 1 as n → ∞, for all j ∈ {n, n + 1, n +
2, · · · , n + M }.
(A3) maxn kdn k = C, maxn kDn k = C̄.
(A4) The iterates θn remain uniformly bounded almost
surely, i.e.,
sup kθn k < ∞, a.s.
n
(A5) The ODE θ̇(t) = −∇J(θ(t)) has a compact set G ⊂ Rp
as its set of asymptotically stable equilibria (i.e., the set
of local minima of J is compact).
(A6) The sequences (Mn+ , Fn ), (Mn− , Fn ), n ≥ 0 form martingale difference sequences. Further, (Mn+ , Mn− , n ≥
0) are square integrable random variables satisfying
±
E[kMn+1
k2 |Fn ] ≤ K(1 + kθn k2 ) a.s., ∀n ≥ 0,
for a given constant K ≥ 0.
The following lemma is useful in obtaining the negative
square root of H i.e., H −1/2 . Also note that it takes only
O(p) operations to compute H −1/2 using the lemma and
circulant structure of H −1/2 .
uuT
uuT
+ p
.
p
p (1 + p)
−
uuT uuT
uuT
uuT uuT
p
+
−
p
p2
p2 (1 + p)
uuT
uuT uuT
uuT uuT
+ p
− p
+ 2
p (1 + p) p2 (1 + p) p (1 + p)
"
uuT
uuT
T
= (I + uu ) I −
+ p
p
p (1 + p)
−
#
uuT
uuT
uuT
+
− p
p
p
p (1 + p)
uuT
uuT
uuT
+ p
− p
+
p (1 + p) p (1 + p) p(1 + p)
"
#
1
T
T 1
= (I + uu ) I − uu ( −
)
p p(1 + p)
"
#
uuT
T
= (I + uu ) I −
1+p
= I + uuT −
#
puuT
uuT
−
1+p 1+p
=I
proving the lemma.
as in (12) and Q
=
√ Let H−1/2defined
p + 1[H
, −H −1/2 u]. Let the pertubations dn be
the columns of Q.
Lemma 2. The perturbations dn chosen as columns of Q
satisfy properties (P1) and (P2).
Proof. Let P = p+1. Observe that as n goes through one cyp+1
p+1
P
P
cle from 1 to p+1 we have
dn dTn = QQT and
dn =
n=1
n=1
s+P
P
u
Q
. Now it is enough to show that
Dn = 0
1
n=s+1
s+P
P
and
dn = 0 i.e., Q
n=s+1
u
1
= 0 for the choice s = 0.
Consider
P
X
Dn =
n=0
=
P
X
convergence theorem, it follows that Nn , n ≥ 0 converges
n+k−1
P
almost surely. Hence k
aj Mj+1 k → 0 almost surely
j=n
P
X
(dn dTn
n=1
as n → ∞. Moreover
J(θj + δj dj ) − J(θj − δj dj ) dj
≤ J(θj + δj dj ) − J(θj − δj dj ) kdj k
≤ C |J(θj + δj dj )| + |J(θj − δj dj )| ,
− I)
(dn dTn ) − (p + 1)I
n=1
= QQT − (p + 1)I
= (p + 1)([H −1/2 , −H −1/2 u][H −1/2 , −H −1/2 u]T − I)
since kdj k ≤ C, ∀j ≥ 0. Note that
= (p + 1)([H −1 + H −1/2 uuT H −1/2 ] − I)
= (p + 1)(H −1/2 (I + uuT )H −1/2 − I)
|J(θj + δj dj )| − |J(0)| ≤ |J(θj + δj dj ) − J(0)|
= (p + 1)(H −1/2 (H)H −1/2 − I)
≤ B̂kθj + δj dj k,
= 0.
where B̂ is the Lipschitz constant of the function J(.).
Hence,
In addition,
p
u
u
−1/2
−1/2
, −H
u])
Q
= p + 1([H
1
1
p
−1/2
−1/2
= p + 1(H
u−H
u)
|J(θj + δj dj )| ≤ B̃(1 + kθj + δj dj k),
=0
for B̃ =max(|J(0)|, B̂). Similarly,
proving the lemma.
Lemma 3. Given any fixed integer P > 0, kθm+k −θm k → 0
w.p.1, as m → ∞, for all k ∈ {1, · · · , P }
|J(θj − δj dj )| ≤ B̃(1 + kθj − δj dj k).
From assumption (A1), it follows that
Proof. Fix a k ∈ {1, · · · , P }. Now
θn+k = θn −
n+k−1
X
!
J(θj + δj dj ) − J(θj − δj dj )
dj
2δj
aj
j=n
−
n+k−1
X
sup
J(θj + δj dj ) − J(θj − δj dj ) dj ≤ K̃ < ∞,
j
for some K̃ > 0. Thus,
aj Mj+1 .
j=n
kθn+k − θn k ≤ K̃
Thus,
n+k−1
X
j=n
kθn+k − θn k ≤
n+k−1
X
j=n
+
n+k−1
X
aj
J(θj + δj dj ) − J(θj − δj dj )
kdj k
2δj
n+k−1
X
aj
+k
aj Mj+1 k
δj
j=n
→ 0 a.s. with n → ∞
proving the lemma.
aj kMj+1 k
j=n
Lemma 4.
Now clearly,
Nn =
n−1
X
aj Mj+1 , n ≥ 1,
j=0
forms a martingale sequence with respect to the filtration
{Fn }. Further, from the assumption (A6) we have,
n
X
E[kNm+1 − Nm k2 |Fm ] =
m=0
≤
n
X
m=0
n
X
E[a2m kMm+1 k2 |Fm ]
a2m K(1 + kθm k2 )
m=0
From the assumption (A4), the quadratic variation process of
Nn , n ≥ 0 converges almost surely. Hence by the martingale
For any m ≥ 0,
m+P
X−1
n=m
m+P
X−1
n=m
an
Dn ∇J(θn ) → 0 and
am
bn
dn J(θn ) → 0, almost surely, as m → ∞.
bm
Proof. From lemma 3, it can be seen that kθm+s − θm k → 0
as m → ∞, for all s = 1, · · · , P. Also from assumption
(A1), we have k∇J(θm+s ) − ∇J(θm )k → 0 as m → ∞,
m+P
P−1
for all s = 1, · · · , P. Now from lemma 2,
Dn = 0
n=m
∀m ≥ 0. Hence Dm = −
m+P
P−1
Dn . Thus we have,
n=m+1
m+P
X−1
=
m+P
X−1
m+P
X−1
Dn
m+P
X−1
am
−
am
m+P
X−1
a
am
m+P
X−1
− 1 ∇J(θn ) + ∇J(θn ) − ∇J(θm )
n
am
bn
dn J(θn )
bm
n=m
m+P
X−1
m+P
X−1
m+P
X−1
bn
dn J(θm )
dn J(θn ) −
bm
n=m+1
n=m+1
m+P
X−1
=
dn
b
n=m+1
≤
m+P
X−1
kdn k
n
bm
m+P
X−1
n=m+1
=C
m+P
X−1
n=m+1
≤C
m+P
X−1
n=m+1
J(θn ) − J(θm )
b
n
bm
n=m+1
≤C
al ξ 1 (δl ) −
n+P
X−1
al Ml+1
l=n
Now the third term on the RHS can be written as
an
n+P
X−1
al
Dl ∇J(θl ) = an ξn ,
an
where ξn = o(1) from lemma 4. Thus, the algorithm is
asymptotically analogous to
Hence from chapter 2 of [10] we have that θn , n ≥ 0
converge to local minima of the function J.
Theorem 6. θn , n ≥ 0 obtained from 1RDSA-1C satisfy
θn → G almost surely.
Proof. Note that
bn
dn J(θn ) + dm J(θm )
bm
n=m+1
=
l=n
n+P
X−1
θn+1 = θn − an (∇J(θn ) + o(δ) + Mn+1 ).
The claim now follows from assumptions (A1) and (A2).
Now observe that kJ(θm+k ) − J(θm )k → 0 as m → ∞, for
all k ∈ {1, · · · , P } as a P
consequence of (A1) and lemma3.
m+P −1
Moreover from dm = − n=m+1 dn we have
=
al ∇J(θl )
al (dl dTl − I)∇J(θl )
l=n
a
n=m+1
m+P
X−1
n+P
X−1
l=n
− 1 ∇J(θn ) + ∇J(θn ) − ∇J(θm )
n
n=m+1
≤ C̄
−
n
∇J(θn ) − ∇J(θm )
n
∇J(θn ) − ∇J(θm )
n+P
X−1
l=n
a
n=m+1
= C̄
θn+P = θn −
a
n=m+1
≤ C̄
It follows that
m+P
X−1
an
Dn ∇J(θn ) −
Dn ∇J(θm )
am
n=m+1
n=m+1
=
l=n
an
Dn ∇J(θn ) + Dm ∇J(θm )
am
n=m+1
=
θn+P = θn −
n+P
X−1 J(θl + δl dl ) − J(θl − δl dl )
dl + Ml+1 .
al
2δl
an
Dn ∇J(θn )
am
n=m
m+P
X−1
Proof. Note that
θn+P = θn −
al
l=n
n+P
J(θ + δ d )
X−1
l
l l
dl −
al Ml+1 .
2δl
l=n
It follows that
θn+P = θn −
n+P
X−1
−
−
n+P
X−1
l=n
n+P
X−1
al
n+P
X−1
J(θl )
dl −
al (dl dTl − I)∇J(θl )
δl
al O(δl ) −
l=n
b
n
bm
b
n
bm
b
J(θn ) − J(θm )
n
al Ml+1
Now we observe that
n+P
X−1
− 1 J(θn ) + J(θn ) − J(θm )
l=n
al
n+P
X−1
J(θl )
dl =
bl J(θl )dl
δl
l=n
n+P
X−1
− 1 J(θn ) + J(θn ) − J(θm )
The claim now follows as a consequence of assumptions
(A1) and (A2).
l=n
n+P
X−1
l=n
= bn
bm
al ∇J(θl )
l=n
J(θn ) − J(θm )
n+P
X−1
l=n
where
ξn1
bl J(θl )
dl = bn ξn1 ,
bn δ l
= o(1) by lemma 4. Similarly
n+P
X−1
al (dl dTl − I)∇J(θl ) = an ξn2 ,
l=n
Theorem 5. θn , n ≥ 0 obtained from 1RDSA-2C satisfy
θn → G almost surely.
with ξn2 = o(1) by lemma 4. The rest follows as explained
in Theorem 5.
V. SIMULATION EXPERIMENTS
A. Implementation
We compare the performance of 1SPSA-2R, 1SPSA-2H
and 1RDSA-2C in the case of two simulation algorithms.
In the case of one simulation algorithms we compare the
performance of 1SPSA-1R, 1SPSA-1H and 1RDSA-1C. We
chose i.i.d Bernoulli ±1-valued perturbations for 1SPSA-2R
and 1SPSA-1R. 1
For the empirical evaluations, we use the following two
loss functions in p = 10 dimensions:
a) Quadratic loss:
TABLE I: NMSE values of two simulation methods for
quadratic objective (15) with and without noise for 2000
simulations: standard error from 100 replications shown after
± symbol
J(θ) = θT Aθ + bT θ.
Noise parameter σ = 0
N
N
X
X
J(θ) = θT AT Aθ + 0.1
(Aθ)3j + 0.01
(Aθ)4j .
(16)
j=1
The optimum θ∗ for the above J is θ∗ = 0, with J(θ∗ ) = 0.
In both functions, A is such that pA is an upper triangular
matrix with each nonzero entry equal to one, b is the N dimensional vector of ones and the noise structure is similar
to that used in [11]. For any θ, the noise is [θT , 1]z, where
z ≈ N (0, σ 2 I11×11 ). We perform experiments for noisy as
well as noise-less settings, with σ = 0.01 for the noisy
case. For all algorithms, we chose step sizes to have the
form δn = c/(n + 1)γ and an = 1/(n + B + 1)α . We set
α = 0.602 and γ = 0.101. These values for α and γ have
been used before (see [11]) and have demonstrated good
finite-sample performance empirically, while satisfying the
theoretical requirements needed for asymptotic convergence.
For all the algorithms, the initial point θ0 is the p-dimensional
vector of ones.
B. Results
We use Normalized Mean Square Error (NMSE) as the
performance metric for evaluating the algorithms. NMSE is
2
2
the ratio kθnend − θ∗ k / kθ0 − θ∗ k . Here nend denotes the
iteration number at the end of simulation budget.
Tables I–II present the normalized mean square values
observed for the three algorithms - 1SPSA-2R, 1SPSA-2H
and 1RDSA-2C for the quadratic and fourth-order loss functions respectively. Tables III–IV present the NMSE values
observed for the three algorithms - 1SPSA-1R, 1SPSA-1H
and 1SPSA-1C with quadratic and fourth order loss functions
respectively. The results in Table I are obtained after 2000
simulations and the results in Table II are obtained after
10000 simulations. The results in Tables –III–IV are obtained
after running the one simulation algorithms with a budget of
20000 simulations.
Figures 2 and 3 show plots of log10 (NMSE) as a function
of number of iterations with quadratic and fourth-order loss
1 The
implementation is available at
cs1070166/1RDSA-2Cand1RDSA-1C/
Method
https://github.com/
NMSE
10−3
± 2.473 × 10−3
1SPSA-2R
5.762 ×
1SPSA-2H
4.012 × 10−5 ± 1.654 × 10−5
1RDSA-2C
2.188 × 10−5 ± 9.908 × 10−6
(15)
For this particular choice of p the optimum θ∗ for the above
J is a 10×1 column vector with each entry equal to −0.9091,
and J(θ∗ ) = −4.55.
b) Fourth-order loss:
j=1
Noise parameter σ = 0.01
Method
NMSE
1SPSA-2R
5.755 × 10−3 ± 2.460 × 10−3
1SPSA-2H
1.601 × 10−5 ± 2.724 × 10−20
1RDSA-2C
2.474 × 10−8 ± 1.995 × 10−23
TABLE II: NMSE values of two simulation methods for
fourth order objective (16) with and without noise for 10000
simulations: standard error from 100 replications shown after
± symbol
Noise parameter σ = 0.01
Method
NMSE
1SPSA-2R
2.762 × 10−2 ± 1.415 × 10−2
1SPSA-2H
3.958 × 10−3 ± 4.227 × 10−4
1RDSA-2C
3.598 × 10−3 ± 4.158 × 10−4
Noise parameter σ = 0
Method
NMSE
10−2
± 1.413 × 10−2
1SPSA-2R
2.747 ×
1SPSA-2H
3.901 × 10−3 ± 4.359 × 10−18
1RDSA-2C
3.535 × 10−3 ± 1.743 × 10−18
TABLE III: NMSE values of one simulation methods for
quadratic objective (15) with and without noise for 20000
simulations: standard error from 100 replications shown after
± symbol
Noise parameter σ = 0.01
Method
NMSE
10−2
± 3.691 × 10−2
1SPSA-1R
8.582 ×
1SPSA-1H
2.774 × 10−2 ± 2.578 × 10−4
1RDSA-1C
8.225 × 10−3 ± 5.959 × 10−5
Noise parameter σ = 0
Method
NMSE
1SPSA-1R
8.584 × 10−2 ± 3.681 × 10−2
1SPSA-1H
2.770 × 10−2 ± 3.836 × 10−17
1RDSA-1C
8.225 × 10−3 ± 1.569 × 10−17
TABLE IV: NMSE values of one simulation methods for
fourth order objective (16) with and without noise for 20000
simulations: standard error from 100 replications shown after
± symbol
Fig. 3: log10 (NMSE)vs No of iterations for fourth order
objective objective (16) with noise (σ = 0.01) using 2
simulation methods.
Noise parameter σ = 0.01
Method
NMSE
10−1
± 1.836 × 10−1
1SPSA-1R
3.240 ×
1SPSA-1H
8.916 × 10−2 ± 1.896 × 10−2
1RDSA-1C
4.972 × 10−2 ± 9.812 × 10−3
Noise parameter σ = 0
Method
NMSE
1SPSA-1R
3.192 × 10−1 ± 1.991 × 10−1
1SPSA-1H
8.173 × 10−2 ± 1.255 × 10−16
1RDSA-1C
4.403 × 10−2 ± 9.066 × 10−17
Fig. 2: log10 (NMSE)vs No of iterations for quadratic objective (15) with noise (σ = 0.01) using 2 simulation methods.
objectives respectively, with σ = 0.01, when 2-simulation
methods are used.
Figures 4 and 5 plot the log10 (NMSE) as a function of
the iterations with quadratic and fourth-order loss objectives
respectively, with σ = 0.01, when 1-simulation methods are
used.
From the results in Tables I,II, III,IV and plots 2,3,4, 5,
we make the following observations:
Observation1 In the case of two simulation algorithms,
1RDSA-2C is slightly better than 1SPSA-2H, while both of
them outperform 1SPSA-2R.
Observation2 In the case of one simulation algorithms,
1RDSA-1C is better than both 1SPSA-1H and 1SPSA-1R.
VI. CONCLUSIONS
We presented a novel construction of deterministic
perturbations for 1RDSA-2R and 1RDSA-1R algorithms
Fig. 4: log10 (NMSE)vs No of iterations for quadratic order
objective objective (15) with noise (σ = 0.01) using 1
simulation methods.
and showed that the resulting algorithms 1RDSA-2C and
1RDSA-1C are provably convergent. The advantage with our
deterministic perturbation construction is that the same set
of perturbations can be used for both two simulation and one
simulation variants. These perturbations also have a smaller
cycle length compared to Hadmard matrix based perturbations. Numerical experiments demonstrated that 1RDSA2C (1RDSA-1C) outperforms 1SPSA-2R (1SPSA-1R) and
1SPSA-2H (1SPSA-1H). Also recent works [12], [13] use
asymmetric Bernoulli perturbations and feedback mechanisms to improve the performance of 2RDSA. As future
work, it would be interesting to derive deterministic perturbations in the context of 2RDSA and compare with the above
work. A challenging future direction would be to derive
Fig. 5: log10 (NMSE)vs No of iterations for fourth order
objective objective (16) with noise (σ = 0.01) using 1
simulation methods.
weak convergence results when deterministic perturbations
are used.
R EFERENCES
[1] H. Robbins and S. Monro, “A stochastic approximation method,”
Ann. Math. Statist., vol. 22, no. 3, pp. 400–407, 09 1951. [Online].
Available: http://dx.doi.org/10.1214/aoms/1177729586
[2] J. Kiefer and J. Wolfowitz, “Stochastic estimation of the maximum
of a regression function,” Ann. Math. Statist., vol. 23, no. 3, pp.
462–466, 09 1952. [Online]. Available: http://dx.doi.org/10.1214/
aoms/1177729392
[3] H. J. Kushner and D. S. Clark, Stochastic Approximation Methods for
Constrained and Unconstrained Systems. Springer Verlag, 1978.
[4] J. C. Spall, “Multivariate stochastic approximation using a simultaneous perturbation gradient approximation,” IEEE Trans. Auto. Cont.,
vol. 37, no. 3, pp. 332–341, 1992.
[5] S. Bhatnagar, H. L. Prasad, and L. A. Prashanth, Stochastic Recursive Algorithms for Optimization: Simultaneous Perturbation Methods
(Lecture Notes in Control and Information Sciences). Springer, 2013,
vol. 434.
[6] J. C. Spall, Introduction to Stochastic Search and Optimization:
Estimation, Simulation, and Control. Wiley, Hoboken, NJ, 2003.
[7] ——, “A one-measurement form of simultaneous perturbation stochastic approximation,” Automatica, vol. 33, pp. 109–112, 1997.
[8] D. C. Chin, “Comparative study of stochastic algorithms for system
optimization based on gradient approximations,” IEEE Transactions on
Systems, Man, and Cybernetics, Part B: Cybernetics, vol. 27, no. 2,
pp. 244–249, 1997.
[9] S. Bhatnagar, M. C. Fu, S. I. Marcus, I. Wang et al., “Two-timescale
simultaneous perturbation stochastic approximation using deterministic perturbation sequences,” ACM Transactions on Modeling and
Computer Simulation (TOMACS), vol. 13, no. 2, pp. 180–209, 2003.
[10] V. S. Borkar, Stochastic approximation: a dynamical systems viewpoint, 2008.
[11] J. C. Spall, “Adaptive stochastic approximation by the simultaneous
perturbation method,” IEEE Trans. Autom. Contr., vol. 45, pp. 1839–
1853, 2000.
[12] D. S. K. Reddy, L. Prashanth, and S. Bhatnagar, “Improved hessian
estimation for adaptive random directions stochastic approximation,”
in Decision and Control (CDC), 2016 IEEE 55th Conference on.
IEEE, 2016, pp. 3682–3687.
[13] L. Prashanth, S. Bhatnagar, M. C. Fu, and S. Marcus, “Adaptive system
optimization using random directions stochastic approximation,” IEEE
Transactions on Automatic Control, 2016.
| 3cs.SY
|
Adaptive Rejection of Periodic Disturbances Acting on Linear Systems
with Unknown Dynamics
arXiv:1603.05361v1 [cs.SY] 17 Mar 2016
Behrooz Shahsavari, Jinwen Pan and Roberto Horowitz
dynamics when the poles are on or very close to the stability
boundary [16]. Another limitation of IMP based repetitive
control is that the controller sampling frequency has to be
divisible by the fundamental frequency of the disturbance.
In general, adaptive feedforward control for this class of
problems does not have the above limitations. This type of
controllers commonly estimate a set of parameters that determine the control law. The estimated parameters converge
when the system is not stochastic or the adaptation gain
is vanishing in a stationary (or cyclostationary) stochastic
system. This implies that the estimated parameters can be
frozen after convergence and the control sequence becomes
a pure feedforward action that can be stored and then looked
up without a need to feeding the error to the controller [11].
This is an important advantage over the feedback schemes
because that type of controller should be constantly in the
loop to generate the control sequence. Another advantage is
that the Bode’s sensitivity integral theorem does not hold
true, which implies that perfect rejection can be achieved
without affecting the suppression level at other frequencies.
Nevertheless, analysis of the adaptive methods is, in general,
more complex and relies on a set of assumptions that may
not hold true in many situations.
Although rejection of sinusoidal disturbances is a classical
control problem, few algorithms exist for the case that the
system dynamics is unknown and possibly time–varying.
Gradient descent based algorithms that use online identification schemes to obtain a finite impulse response (FIR)
of the plant [17] have been proposed by both control and
signal processing communities. The number of estimated
parameters in these methods is usually large since low–
order FIR models cannot mimic complex dynamics. The
harmonic steady–state (HSS) control is another adaptive
method for rejection of sinusoidal disturbances which is easy
to understand and implement. However, it suffers from slow
convergence since it relies on averaging over batches of data
[18], [19].
This paper provides a novel direct adaptive feedforward
control that does not require a model of the plant dynamics
or batches of measurements. The proposed method does not
rely on any assumptions about the location of plant zeros and
can be applied to minimum and non–minimum phase plants.
The algorithm is a “direct” adaptive method, meaning that
the identification of system parameters and the control design
are not separate. It will be shown that the number of adapted
parameters in this scheme is significantly less than methods
that identify the plant frequency response when the number
of frequencies is large [19], [20], [21].
Abstract— This paper proposes a novel direct adaptive control method for rejecting unknown deterministic disturbances
and tracking unknown trajectories in systems with uncertain
dynamics when the disturbances or trajectories are the summation of multiple sinusoids with known frequencies, such as
periodic profiles or disturbances. The proposed algorithm does
not require a model of the plant dynamics and does not use
batches of measurements in the adaptation process. Moreover,
it is applicable to both minimum and non–minimum phase
plants. The algorithm is a “direct” adaptive method, in the sense
that the identification of system parameters and the control
design are performed simultaneously. In order to verify the
effectiveness of the proposed method, an add–on controller is
designed and implemented in the servo system of a hard disk
drive to track unknown nano–scale periodic trajectories.
I. I NTRODUCTION
Control methodologies for rejecting periodic and multi–
harmonic disturbances or tracking such trajectories have
attracted many researchers in the past two decades. There is
a multitude of applications, especially due to the dominating
role of rotary actuators and power generators, that crucially
rely on this type of control. A non–exhaustive list of these
applications include aircraft interior noise control [1], [2],
periodic load compensation in wind turbines [3], [4], wafer
stage platform control [5], [6], steel casting processes [7],
laser systems [8], milling machines [9], [10] and hard disk
drives [11], [12]. In this paper, we introduce a novel direct
adaptive control for rejecting deterministic disturbances and
tracking unknown trajectories in systems with unknown
dynamics when the disturbances or trajectories are the summation of multiple sinusoids with known frequencies. Note
that a periodic disturbance/trajectory with a known period
can be considered as a special case of the problems under
our study.
Control methods applied to this class of problems are typically categorized into two types, namely feedback methods
that are based on internal model principle (IMP) [13] and
feedforward algorithms that usually use an external model
[14] or a reference signal correlated with the disturbance
[15]. The classical form of internal model for periodic
disturbances introduces poles on the stability boundary which
can cause poor numerical properties and instability when
implemented on an embedded system with finite precision
arithmetic. Instability can also happen due to unmodeled
B. Shahsavari and R. Horowitz are with the Department of Mechanical Engineering, University of California, Berkeley, CA 94720, U.S.A.
{behrooz,horowitz}@berkeley.edu
J. Pan is with the Department of Automation, University of
Science and Technology of China, Hefei, 230027, P.R.China,
[email protected]
1
w
u
r
n
G
_
uF
uA
m
generality, we assume that it contaminates the plant output.
Let the adaptive controller sampling time be T . The class of
disturbances under our study can be written as
y
CF
e
r(k) =
n
X
αi sin (ωi kT + δi )
(1)
i=1
CA
where the amplitudes, αi , and phase shifts, δi , are unknown
but the frequencies, ωi , are known.
Our objective is to synthesize an adaptive controller that
only uses the scalar–valued error signal e(k) to generate a
feedforward control uA (k) such that it perfectly compensates
for the effect of r(k) on the error signal e(k). We call it a
feedforward controller because when the system dynamics
and disturbance profile are time–invariant, CA will not
depend on the error signal once the control law is learned.
Fig. 1: Closed loop system augmented by a plug–in adaptive
controller.
The remainder of this paper is organized as follows. We
first formalize the problem and explain the system under our
study in section II. Mathematical preliminaries and notations
are given in section III. The algorithm derivation is presented
in section IV and simulation results for rejecting periodic
disturbances in a hard disk drive nanopositioner servo system
are illustrated in section V. Conclusion remarks and future
work form section VI.
III. M ATHEMATICAL P RELIMINARIES AND N OTATIONS
Let R(z) be the single–input single–output system (in z–
domain) from the adaptive control injection points to the
error signal e
II. P ROBLEM S TATEMENT
R (z) : = G (z) S (z)
The adaptive controller proposed in this work is aimed
to be implemented in a plug–in fashion, meaning that it
is used to augment an existing robustly stable closed–
loop system in order to reject periodic disturbances (track
periodic trajectories) that are not well rejected (tracked)
by the existing controller. In this architecture, the original
controller can be designed without consideration of this
special control task. Moreover, the plug–in controller does
not alter the performance of the original control system. To
clarify this notion, we use a common Single-Input SingleOutput (SISO) plant–controller interconnection shown in
Fig. 1 as a running example. The blocks G and CF in the
figure respectively denote a linear time invariant (LTI) plant
and an LTI feedback compensator that form a stable closed–
loop sensitivity function
and let A(q −1 ) and B(q −1 ) be the polynomials that represent
R(·) in time–domain
S :=
R(q −1 ) =
B(q −1 )
.
A(q −1 )
(2)
We define an output disturbance, say w̄, on R(·) and a
polynomial C(q −1 ) that satisfies
1
[w̄(k)] = R [w(k)] + S [n(k) + m(k)]
A(q −1 )
We also define r̄(k) as a sequence on the output of R(.) that
has the same effect as r in the error signal e
r̄(k) = S [r(k)] ·
1
.
1 + GCF
When the same transfer function filters multiple input signals
i1 (k), i2 (k), · · · , im (k), we abuse the notation and use
1
T [i1 (k)]
i (k)
T [i2 (k)]
i2 (k)
=
T
.. .
..
.
.
One of the main contributions of the controller that will
be presented shortly is that it does not require the plant
and controller dynamics. Since our design does not depend
on whether the plant/nominal controller are continuous or
discrete time, we assume that both G and CF are discrete
time systems to make notations simpler. It is worth noting
that this interconnection is only a running example through
this paper and the proposed controller can be plugged to any
unknown and stable LTI system regardless of its internal
stabilization mechanism.
A general stochastic environment is considered for the system by appending input disturbance w, output disturbance n,
and contaminating measurement noise m to our framework.
Generally, the nominal feedback controller CF is designed
to compensate for these input and output noises. The special
deterministic disturbance that should be compensated by the
adaptive controller is denoted by r, and without loss of
T [im (k)]
im (k)
The disturbance r(k) in (1) can be factorized as the inner
product of a known “regressor” vector φ(k) and an unknown
vector of parameters θ
r(k) = θT φR (k)
where
θT : = [α1 cos(δ2 ), α1 sin(δ2 ), . . .
, αn cos(δn ), αn sin(δn )]
φTR (k)
: = [sin(ω1 kT ), cos(ω1 kT ), . . .
, sin(ωn kT ), cos(ωn kT )] .
2
(3)
Lemma 1
Consider r(k) as a general periodic signal and L(q −1 )
as a discrete-time linear system. The steady state response
r̃(k) := L(q −1 )[r(k)] is periodic. Moreover, when r(k) is
a linear combination of sinusoidal signals factorized similar
to (3), r̃(k) (in steady state) consists of sinusoidals with the
same frequencies
r̃(k) = L(q −1 ) θT φR (k) = θT L(q −1 ) [φR (k)] = θT φRL (k)
where
φTRL (k) := [mL1 sin(ω1 kT + δL1 ), mL1 cos(ω1 kT + δL1 ),
. . . , mLn sin(ωn kT + δLn ), mLn cos(ωn kT + δLn )] .
(4)
A(q −1 ) : = 1 + a1 q −1 + a2 q −2 + · · · + anA q −nA
B(q −1 ) : = b1 q −1 + b2 q −2 + · · · + bnB q −nB .
Without loss of generality, we assume that the relative degree
of the transfer function from the controllable input channel
to the output is 1, which implies that nA = nB . The analysis
for other non-negative relative degrees is very similar to the
sequel, but the notation would be more tedious due to differences in vector/matrix sizes. Let A∗ (q −1 ) := 1 − A(q −1 ),
then the error is given by
e(k) = A∗ (q −1 )e(k) + B(q −1 ) (u(k) + uA (k))
+ w̄(k) + A(q −1 )r̄(k)
Here, mLi and δLi are the magnitude and phase of
L(e−jωi T ) respectively. Define
mLi cos(δLi ) mLi sin(δLi )
DLi :=
,
(5)
−mLi sin(δLi ) mLi cos(δLi )
(9)
This equation can be represented purely in discrete time
domain in a vector form
T
T
e(k) = θA
φe (k) + θB
(φu (k) + φuA (k)) + r̃(k) + w̄(k)
(10)
φR (k) can be transformed to φRL (k) by a linear transformation
DL1
0
···
0
0
DL2 · · ·
0
(6)
φRL (k) = .
.
.. φR (k).
..
..
..
.
.
0
0
· · · DLn
{z
}
|
where
T
θA
: = −a1 , −a2 , · · · , −anA ,
T
θB
: = b1 , b2 , · · · , bnA ,
φTe (k) : = e(k − 1), e(k − 2), · · · , e(k − nA ) ,
φTu (k) : = u(k − 1), u(k − 2), · · · , u(k − nA ) ,
φTuA (k) : = uA (k − 1), uA (k − 2), · · · , uA (k − nA ) ,
(11)
DL
As a result
and r̃(k) := A(q −1 )r̄(k). Note that two regressors, denoted
by φu (k) and φuA (k), are considered for the excitation signal
u(k) and the adaptive control uA (k) separately although they
could be combined into a unique regressor. The rationale
behind this consideration will be explained later after (19).
Since disturbance r̄(k) is periodic and A(q −1 ) is a stable
filter – i.e. it operates as an FIR filter – the response r̃(k) is
also periodic by Lemma 1
T
T
r̃(k) = A(q −1 ) θR̄
φR (k) = θR̄
φRA (k)
T
r̃(k) = L(q −1 ) θT φR (k) = θT φRL (k) = θR̄
DL φ R .
Proof: Refer to [22] for a general formula for the
steady-state sinusoidal response of a linear time-invariant
system.
Using Lemma 1 the equivalent disturbance r̄ can be
factorized as
T
r̄(k) = θR̄
φR (k)
A and, B can be determined such that they give a finite vector
difference equation that describes the recorded data as well
as possible, i.e.
(7)
where θR̄ is unknown.
where
IV. A DAPTIVE C ONTROL S YNTHESIS
DA1
0
φRA (k) = .
..
A. Error Dynamics
The error sequence in time domain can be represented as
a function of the closed–loop dynamics, control signals and
disturbances
1
B(q −1 )
(u(k) + uA (k)) +
w̄(k) + r̄(k)
e(k) =
A(q −1 )
A(q −1 )
(8)
0
|
0
DA2
..
.
···
···
..
.
0
0
..
.
0
···
DAn
{z
DA
φR (k).
(12)
}
Accordingly, r̃(k) can be represented using the same regressor vector, φR (k)
where u(k) is an exogenous excitation signal and w̄(k) is an
unmeasurable wide–sense stationary sequence of independent random values with finite moments. We assume that
the nominal feedback controller is able to stabilize the open
loop plant, i.e. A(p) has all roots strictly outside the unit
circle. Although a real dynamic system cannot be exactly
described by finite order polynomials, in most of applications
T
T
T
φR (k).
r̃(k) = θR̄
φRA (k) = θR̄
DA φR (k) = θR
| {z }
T
θR
Substituting this expression in (10) yields
T
T
T
e(k) = θA
φe (k) + θB
(φu (k) + φuA (k)) + θR
φR (k) + w̄(k).
(13)
3
Equation (9) shows that an ideal control signal u∗A (k) should
satisfy
B(q −1 )u∗A (k) + A(q −1 )r̄(k) = 0.
T
T
time. As a result, the term θB
φuA (k) + θR
φR (k) in (13) can
be replaced by (18) which yields to
T
T
T
e(k) = θA
φe (k) + θB
φu (k) + θM
(k)φR (k) + wt (k) + w̄(k).
(19)
(14)
Again, since B(q −1 ) and A(q −1 ) are both LTI systems and
r̄(k) contains only sinusoidal signals, the ideal control signal
u∗A (k) has to have sinusoidal contents at frequencies equal
to ωi ’s. This motivates us to decompose the ideal control
signal into
Remark 1
The reason behind choosing two separate regressors for u(k)
and uA (k), as remarked earlier, is that the recent substitution
in the above equation is not feasible if the two regressors
were combined into a single regressor.
T
u∗A (k) = θD
φR (k).
B. Parameter Adaptation Algorithm
By this representation of the control signal, our goal will be
to estimate θD in an adaptive manner. We define the actual
control signal as
T
uA (k) = θ̂D
(k)φR (k)
The error dynamics shows that the information obtained
from measurements cannot be directly used to estimate θ̂D
as long as the closed loop system is unknown. We propose
an adaptive algorithm in this section that accomplishes the
estimation of the closed loop system and control synthesis
simultaneously.
Let θ̂A , θ̂B and θ̂M be the estimated parameters analogous
to (19). We denote the a–priori estimate of the error signal
at time k based on the estimates at k − 1 as
(15)
where θ̂D (k) is the vector of estimated parameters that
should ideally converge to θD . As a result, the residual in
(14) when θD is replaced by θ̂D is
B(q −1 )uA (k) + A(q −1 )r̄(k) =
h
i
T
T
B(q −1 ) θ̂D
(k)φR (k) + θR
φR (k).
(16)
T
T
T
ŷ(k) = θ̂A
(k − 1)φe (k) + θ̂B
(k − 1)φu (k) + θ̂M
(k − 1)φR (k).
(20)
and accordingly, the a–priori estimation error is defined as
Lemma 2
Let B(q −1 ) have a minimal realization B(q −1 ) =
−1
T
CB
(qI − AB ) BB . Then,
h
i
T
T
B(q −1 ) θ̂D
(k)φR (k) = θ̂D
(k) B(q −1 ) [φR (k)] + wt (k)
◦ (k) : = e(k) − ŷ(k)
(21)
Assume that the estimates at time k = 0 are initialized by
either zero or some “good” values when prior knowledge
about the system dynamics is available. We propose the
where
following adaptation algorithm for updating the estimated
h
ii
T
h
−1
−1
wt (k) := − M (q ) H(q ) φR (k + 1) θ̂(k + 1) − θ̂(k) parameters
θ̂A (k − 1)
θ̂A (k)
T
M (q −1 ) : = CB
(qI − A)−1
θ̂B (k) = θ̂B (k − 1)
H(q −1 ) : = (qI − A)−1 BB
θ̂M (k)
θ̂M (k − 1)
Proof: Refer to the discrete–time swapping lemma in
φe (k)
−1
γ1 (k)F (k)
0
[23].
φu (k) ◦ (k).
+
−1
0
γ
(k)f
(k)I
2
We define a new parameter vector
φR (k)
(22)
T
T
T
θM
(k) := θ̂D
(k)DB + θR
(17)
F (k) is a positive (semi) definite matrix with proper diwhere DB is a matrix similar to DA in (6), but its block mension and f (k) is a positive scalar. These gains, which
diagonal terms
are formed by the magnitude and phase of are usually known as learning factor or step size, can be
B e−jωi T rather than A(e−jωi T ). Since the vector θM (k) updated via either recursive least squares algorithm, least
corresponds to the imperfection in control synthesis, it is mean squares type methods or a combination of them. We
called the residual parameters vector throughout this section. use recursive least squares for the plant since the number
Substituting the result of Lemma 2 in (16) yields
of coefficients is usually “small”. On the other hand, for
−1
−1
large n, the recursive least squares algorithm requires major
B(q )uA (k) + A(q )r̄(k)
computations. Therefore, it is of interest to reduce the
T
T
= θ̂D
(k) B(q −1 )[φR (k)] + θR
φR (k) + wt (k)
computations, possibly at the price of slower convergence,
T
T
by replacing the recursive least squares update law by the
= θ̂D (k)DB φR (k) + θR φR (k) + wt (k)
T
= θM (k)φR (k) + wt (k)
(18) stochastic gradient method. It is well known that the step
size of adaptive algorithms in stochastic environments should
T
where the term θM
(k)φR (k) corresponds to the residual error converge to zero or very small values to avoid “excess error”
at the compensation frequencies. The term wt (k) represents caused by parameter variations due to noises. Therefore,
the transient excitation caused by the variation of θ̂D (k) over positive real valued decreasing scalar sequences γ1 (k) and
4
γ2 (k) are considered conjointly with the step sizes. More D̂B converge to the actual θM and DB – which will be
explicitly, the update rules for F and f are
proved later – the steady state residue is
T
T
lim θM (k) = DB
lim θ̂D (k) + θR
φe (k)
φe (k)
k→∞
k→∞
F (k) = F (k − 1) + γ1 (k) φu (k) φu (k) − F (k − 1)
−α
lim θM (k) + θR
=
φ (k)
φ (k)
1 − β k→∞
1−β
f (k) = f (k − 1) + γ2 (k) φTR (k)φR (k) − f (k − 1) .
=
θR .
1−β+α
This expression shows that there is a compromise between
the steady state attenuation level and robustness, and in order
to achieve both, the two gains should be chosen such that
Remark 2
φu (k) should be persistently exciting of order 2nA in order
to guarantee that F (k) is non–singular and (22) is not
susceptible to numerical problems. It is clear that f (k) is not
subjected to this issue since φTR (k)φR (k) is always strictly
positive.
0 < α β < 1.
Now that we have an update law for θ̂D (k), we have
a complete algorithm for synthesizing the control signal
(repeated from (15))
C. Control Synthesis
T
uA (k) = θ̂D
(k)φR (k).
Suppose that the parameter vector θM (k) and response
matrix DB are known at time step k. Then, a possible update
rule that satisfies (14) would be
Theorem 1
The control update rule outlined by (15), (22) and (24)
1−β
θR with probability 1, the only
make θM converge to 1−β+α
equilibrium point of the closed loop system is stable in the
sense of Lyapunov and it corresponds to θ̂A = θA , θ̂B = θB ,
1−β
and θ̂M = 1−β+α
θR if the following conditions are satisfied:
1) u(k) is persistently exciting of at least order 2nA .
∞
P
2)
γi (k) = ∞ and γi (k) → 0 as k → ∞ for i = 1, 2.
T
T
0 = θ̂D
(k + 1)DB + θR
T
T
T
= θ̂D
(k + 1)DB + θM
(k) − θ̂D
(k)DB ⇒ (23)
−1
T
T
T
θ̂D
(k + 1) = θ̂D
(k) − θM
(k)DB
.
Here, we have used the fact that DB is a block diagonal
combination of scaled rotation matrices, which implies that
it is full rank and invertible. This is an infeasible update rule
since neither θM (k), nor DB is known. We replace these
variables by their respective estimated values and use a small
step size α in order to avoid large transient and excess error
k=1
3) The estimated θ̂A (k) belongs to
nA
DA := θ̂A : 1 + â1 q + · · · + ânA q = 0 ⇒ |q| > 1 .
−1
T
T
T
θ̂D
(k + 1) = θ̂D
(k) − αθ̂M
(k)D̂B
(k).
infinitely often with probability one.
4) The estimated θ̂B (k) always belongs to
DB : = θ̂B : 0 < |b̂1 e−jωh + · · · + b̂nA e−jnA ωh |
Note that this update rule works as a first order system
that has a pole at 1. In order to robustify this difference
equation we alternatively propose using a Ridge solution for
(23). More formally, we are interested in minimizing the
instantaneous cost function
1
1 T
T −1 2
T
(k) + θR
DB k2 + λkθ̂D
(k)k22
Jc (k) := kθ̂D
2
2
α
|b1 e−jωh + · · · + bnA e−jnA ωh |,
1−β
and ∀h ∈ {1, . . . , n} .
<
−jωh
5) Real B̂B(e(e−jωh)) > 0 for all h ∈ {1, . . . , n} infinitely
k
often with probability one.
where λ is a (positive) weight for the penalization term. We
use a gradient descent algorithm to recursively update θ̂D .
Let β = 1 − αλ and the gradient of Jc (k) with respect to
T
θ̂D
(k) be denoted by
∂Jc (k)
−1
T
T
T
T
= θ̂D
(k) + θM
(k)DB
− θ̂D
(k) + λθ̂D
(k).
T (k)
∂ θ̂D
Proof: Only the sketch of the proof is outlined here due
to space limitation. The method of analysis of stochastic recursive algorithms developed by Ljung [24] can be deployed
to study the convergence and asymptotic behavior of the proposed adaptive algorithm with update and control rules given
in (22), (24) and (15). Three sets of regularity conditions
are proposed in [24] that target the analysis of deterministic
and stationary stochastic processes. The problem under our
study cannot be exactly outlined in these frameworks since
the input signal consists of stochastic and deterministic parts,
and as a matter of fact, it is a cyclostationary stochastic
process. However, “Assumptions C” in [24] can be adopted
and generalized to this case with minor modifications. It
Since the actual values of θM and DB are unknown, we use
the estimates and define the gradient descent update rule for
θ̂D as
−1
T
T
T
θ̂D
(k + 1) = β θ̂D
(k) − αθ̂M
(k)D̂B
(k).
(25)
(24)
This expression implies that a positive value of β less than
1 results in a bounded value of θ̂D in steady state as long
−1
T
as θ̂M
D̂B
stays bounded. Moreover, assuming that θ̂M and
5
TABLE I: Hyper Parameters of the adaptive control algorithm in the empirical study.
can be shown that these regularity conditions are satisfied
when the assumptions of theorem 1 are satisfied. Under
these regularity conditions, the results of theorem 2 in [24]
imply that the only convergence point of the system is the
stable equilibrium of the differential equation counterpart.
This equilibrium point corresponds to the actual values of
1−β
θR . Moreover, this theorem proves that the
plant and 1−β+α
estimated parameters converge with probability one to this
1−β
equilibrium point which results in θM → 1−β+α
θR with
probability one.
nA (11)
5
α (24)
4E -5
β (24)
1-(2E -7)
The signals w, r, n and m denote the (unknown) airflow disturbance, repeatable runout (RRO), non-repeatable runout
(NRRO) and measurement noise respectively. The measured
position error signal (PES) is denoted by e. The design
of CF is not discussed here, and it is assumed that this
compensator can robustly stabilize the closed loop system
and attenuate the broad band noises w, n and m (c.f. [26],
[27]). The plug–in controller that will be explained shortly
targets the RRO (r) which consists of sinusoids. An exact
dynamics of the actuators is not known for each individual
HDD. Furthermore, temperature variations and deterioration
over time can introduce more uncertainties [28]. We do
not use any information about the system dynamics or the
feedback controller. In other words, everything is unknown
to us except the position error signal (PES).
Bit patterned media (BPM) is an emerging magnetic
recording technology in which each data bit is recorded
on a single magnetic island in a gigantic array patterned
by lithography on the disk. This makes the servo system
a crucial component, and introduces significant new complexity. BPMR requires that the data tracks be followed
with significantly more accuracy than what is required in
conventional continuous media recording since the head has
to be accurately positioned over the single–domain islands
in order to read or write data. Unknown track shapes results
in written–in runout which becomes repeatable (RRO) from
the controller sight of view when the disk spins – i.e. r
in Fig. 1. In our setup, RRO has narrow–band contents at
the HDD spinning frequency (120Hz) and its 173 higher
harmonics. In other words, n in (1) is 174,
Remark 3
Assumption 2 can be satisfied by a broad range of gain
sequences γ(k). For instance, both regularity conditions hold
for γ(k) = kCα when 0 < C < ∞ and 0 < α ≤ 1.
Remark 4
Assumption 3 requires monitoring the roots of Â(q −1 )
polynomial. This is a common issue in adaptive control
and several methods have been proposed. For instance, the
estimates can be projected to the interior of DA whenever the
poles fall out of (or on) the unit circle. Assumption 4 requires
monitoring the magnitude of polynomials B(e−jωh ) at all
compensation frequencies. The left inequality guarantees that
D̂B is always invertible. The right inequality requires some
very rough knowledge about the plant magnitude because the
term α/(1 − β) is large according to (25). Both inequalities
can be satisfied by projecting the estimates into the interior
of DB whenever they do not belong to DB .
Assumption 5 is in general difficult to verify since B(q −1 )
polynomial is unknown. Based on theorem 1, the equilibrium
point of the closed loop system satisfies this assumption. It
can be shown that one of the factors that determines the
domain of attraction associated with this equilibrium point
is the excitation sequence u(k) intensity. This implies that
when no information about B(q −1 ) is available and the
estimated parameters are initialized with zeros, it may be
required to use a large excitation signal such that the domain
of attraction includes the initial values. However, a nominal
system is usually known in practice and this requirement
can be relaxed. Moreover, when the system is slowly timevarying, it is expected that this assumption is satisfied with a
significantly small excitation since the current estimates are
kept inside the domain of attraction of the slowly varying
equilibrium point.
ωi (rad/s) = i × 120(Hz) × 2π
i = 1, · · · , 174, (26)
and system sampling frequency is 41.760KHz (120Hz ×
348).
A. Computer Simulation Results
The magnitude response of the closed loop dynamics from
the VCM input to the PES decays notably after 2KHz,
which makes the VCM at frequencies above 7KHz ineffective. Accordingly, we only focus on tracking the first 58
harmonics (120Hz, 240Hz, ..., 6960Hz) in the design of CA
in Fig. 1. The remaining 115 harmonics should be allocated
to a higher bandwidth actuator which is beyond the scope of
this paper and is considered as one of our future work.
The design parameters of the adaptive control algorithm
are listed in Table I. The estimated coefficients for A(q −1 )
and B(q −1 ) that construct θ̂A and θ̂B are shown in Fig. 2.
The figure shows that the estimated parameters converge
to “some” values quickly. In order to evaluate the con−1
)
vergence point, we generated the transfer function B̂(q
Â(q −1 )
V. E MPIRICAL S TUDY
This section provides the experimental verification of
the proposed controller. The method is used to design a
plug-in controller for tracking nano-scale unknown periodic
trajectories with high frequency spectra in hard disk drives
(HDDs).
A so-called single-stage HDD uses a voice coil motor
(VCM) for movements of the read/write heads [25]. The
block diagram in Fig. 1 can be adopted for this mechatronic
system in track-following mode. The blocks G and CF refer
to the VCM and the nominal feedback controller respectively.
6
0.5
0.2
a3
a4
0.1
a5
b1
b2
0
0
b3
b4
-0.1
b5
-0.2
-5
Estimated coe/cients of B(q!1 )
Estimated coe/cients of A(q!1 )
a2
Amplitude Spectrum (nm)
a1
5
Adaptive Controller Off
Adaptive Controller On
0.4
0.3
0.2
0.1
0
20
40
60
80
100
120
140
160
Harmonic
1
2
3
4
5
6
7
8
9
Step
Fig. 5: Comparison of the position error spectrum before and
after plugging the adaptive controller. This figure shows the
amplitude of Fourier transformation only at harmonics – i.e.
other frequencies are removed.
10
#10 4
Fig. 2: Evolution of −θ̂A and θ̂B elements.
0.4
-40
Feedforward Control
Magnitude (dB)
0
-20
-60
-80
20
40
60
80
100
120
140
160
180
Phase (deg)
180
0.2
0
-0.2
90
0
-0.4
50
-90
150
200
250
300
Step
-180
20
40
60
80
100
120
140
160
Fig. 6: Feedforward control signal, uA (k), learned by the
adaptive controller.
180
Harmonic
Fig. 3: Frequency response comparison of the identified
model and the actual VCM loop. The shaded strip indicates
the compensation frequency range.
As mentioned earlier, the control signal uA learned by
the controller is periodic. Figure 6 depicts one period of this
signal, which can be saved as one period of a repetitive feedforward control sequence that is able to perfectly compensate
the first 58 harmonics.
#10 -3
6
4
3^M coe/cients
100
2
0
VI. C ONCLUSION
-2
-4
A novel direct adaptive control method for the rejection
of disturbances or tracking trajectories consisted of multiple
sinusoidals with selective frequencies was proposed. The
method is applicable to both minimum and non-minimum
phase linear systems with unknown dynamics. The adapted
parameters converge to the real values when a large enough
excitation signal is injected to the system. In the presence
of some rough knowledge about the system dynamics, the
excitation signal can be reduced considerably. The analysis
in this paper was performed for linear time-invariant systems.
However, similar results can be extended to systems with
slowly time–varying parameters.
We verified the effectiveness of the proposed control
algorithm in tracking unknown nano–scale periodic trajectories in hard disk drives by designing an add–on repetitive
controller that was able to track the first 58 harmonics of
the disk spinning frequency. Full spectrum compensation
was impossible in our running example due to the VCM
limited bandwidth. This issue can be addressed by deploying
a dual–stage mechanism that has a high–bandwidth actuator
in conjunction with the VCM. Extension of the proposed
-6
-8
1
2
3
4
5
Step
6
7
8
9
10
#10 4
Fig. 4: Estimated residue parameters, θ̂M .
that corresponds to these values. The frequency response of
this (5th order) transfer function is compared to the actual
transfer function of the VCM loop (a realistic 50th order
model) in Fig. 3. The shaded strip indicates the compensation
frequency interval where the adaptive controller was active.
The estimated residue parameters, θ̂M , are depicted in
Fig. 4. The plot shows that the residual disturbance converges
towards zero as the algorithm evolves. This can be verified in
frequency domain based on the spectrum of error too. The
amplitude spectrum of the error before and after plugging
the adaptive controller to the closed loop servo system are
depicted in Fig. 5. For clearness, the figure only shows the
amplitude of the error Fourier transformation at compensation frequencies.
7
method to multi–input single–output systems and experimental verification of the algorithm will form our future work.
[14] M. Tomizuka, K.-K. Chew, and W.-C. Yang, “Disturbance rejection
through an external model,” Journal of dynamic systems, measurement,
and control, vol. 112, no. 4, pp. 559–564, 1990.
[15] M. Bodson and S. C. Douglas, “Adaptive algorithms for the rejection
of sinusoidal disturbances with unknown frequency,” Automatica,
vol. 33, no. 12, pp. 2213–2221, 1997.
[16] M. Bodson, “Rejection of periodic disturbances of unknown and timevarying frequency,” International Journal of Adaptive Control and
Signal Processing, vol. 19, no. 2-3, pp. 67–88, 2005.
[17] M. Zhang, H. Lan, and W. Ser, “Cross-updated active noise control
system with online secondary path modeling,” Speech and Audio
Processing, IEEE Transactions on, vol. 9, no. 5, pp. 598–602, 2001.
[18] D. Patt, L. Liu, J. Chandrasekar, D. S. Bernstein, and P. P. Friedmann,
“Higher-harmonic-control algorithm for helicopter vibration reduction
revisited,” Journal of guidance, control, and dynamics, vol. 28, no. 5,
pp. 918–930, 2005.
[19] J. Chandrasekar, L. Liu, D. Patt, P. P. Friedmann, and D. S. Bernstein,
“Adaptive harmonic steady-state control for disturbance rejection,”
Control Systems Technology, IEEE Transactions on, vol. 14, no. 6,
pp. 993–1007, 2006.
[20] S. Pigg and M. Bodson, “Adaptive rejection of sinusoidal disturbances
of known frequency acting on unknown systems,” in American Control
Conference, 2006, pp. 5–pp, IEEE, 2006.
[21] S. Pigg and M. Bodson, “Adaptive algorithms for the rejection of
sinusoidal disturbances acting on unknown plants,” Control Systems
Technology, IEEE Transactions on, vol. 18, no. 4, pp. 822–836, 2010.
R EFERENCES
[1] U. Emborg, “Cabin noise control in the saab 2000 high-speed
turboprop aircraft,” in PROCEEDINGS OF THE INTERNATIONAL
SEMINAR ON MODAL ANALYSIS, vol. 1, pp. 13–26, KATHOLIEKE
UNIVERSITEIT LEUVEN, 1998.
[2] J. Wilby, “Aircraft interior noise,” Journal of Sound and Vibration,
vol. 190, no. 3, pp. 545–564, 1996.
[3] K. A. Stol and M. J. Balas, “Periodic disturbance accommodating
control for blade load mitigation in wind turbines,” Journal of solar
energy engineering, vol. 125, no. 4, pp. 379–385, 2003.
[4] I. Houtzager, J.-W. van Wingerden, and M. Verhaegen, “Rejection of
periodic wind disturbances on a smart rotor test section using lifted
repetitive control,” Control Systems Technology, IEEE Transactions on,
vol. 21, no. 2, pp. 347–359, 2013.
[5] D. De Roover and O. H. Bosgra, “Synthesis of robust multivariable
iterative learning controllers with application to a wafer stage motion
system,” International Journal of Control, vol. 73, no. 10, pp. 968–
979, 2000.
[6] B. G. Dijkstra, Iterative learning control, with applications to a waferstage. TU Delft, Delft University of Technology, 2004.
[7] T.-C. Tsao and J. Bentsman, “Rejection of unknown periodic load disturbances in continuous steel casting process using learning repetitive
control approach,” Control Systems Technology, IEEE Transactions on,
vol. 4, no. 3, pp. 259–265, 1996.
[8] M. A. McEver, D. G. Cole, and R. L. Clark, “Adaptive feedback control of optical jitter using q-parameterization,” Optical Engineering,
vol. 43, no. 4, pp. 904–910, 2004.
[9] S. Rober and Y. Shin, “Control of cutting force for end milling
processes using an extended model reference adaptive control scheme,”
Journal of Manufacturing Science and Engineering, vol. 118, no. 3,
pp. 339–347, 1996.
[10] T. Tsao and K. Pong, “Control of radial runout in multi-tooth face
milling,” Transactions of the North American Manufacturing Research
Institute of SME, pp. 183–190, 1991.
[11] B. Shahsavari, E. Keikha, F. Zhang, and R. Horowitz, “Adaptive
repetitive control design with online secondary path modeling and
application to bit-patterned media recording,” Magnetics, IEEE Transactions on, vol. 51, no. 4, pp. 1–8, 2015.
[12] B. Shahsavari, E. Keikha, F. Zhang, and R. Horowitz, “Adaptive
repetitive control using a modified filtered-x lms algorithm,” in ASME
2014 Dynamic Systems and Control Conference, pp. V001T13A006–
V001T13A006, American Society of Mechanical Engineers, 2014.
[13] B. A. Francis and W. M. Wonham, “The internal model principle of
control theory,” Automatica, vol. 12, no. 5, pp. 457–465, 1976.
[22] E. W. Kamen and B. S. Heck, Fundamentals of signals and systems:
using the Web and MATLAB. Prentice Hall, 2000.
[23] G. Tao, Adaptive control design and analysis, vol. 37. John Wiley &
Sons, 2003.
[24] L. Ljung, “Analysis of recursive stochastic algorithms,” Automatic
Control, IEEE Transactions on, vol. 22, no. 4, pp. 551–575, 1977.
[25] R. Horowitz, Y. Li, K. Oldham, S. Kon, and X. Huang, “Dual-stage
servo systems and vibration compensation in computer hard disk
drives,” Control Engineering Practice, vol. 15, no. 3, pp. 291–305,
2007.
[26] B. Shahsavari, R. Conway, E. Keikha, F. Zhang, and R. Horowitz,
“Robust track-following controller design for hard disk drives with
irregular sampling,” IEEE TRANSACTIONS ON MAGNETICS, vol. 49,
no. 6, 2013.
[27] B. Shahsavari, R. Conway, E. Keikha, F. Zhang, and R. Horowitz,
“h∞ control design for systems with periodic irregular sampling
using optimal h2 reference controllers,” in ASME 2013 Conference
on Information Storage and Processing Systems, pp. V001T03A008–
V001T03A008, American Society of Mechanical Engineers, 2013.
[28] K. Malang and L. Hutsell, “Method and apparatus for calibrating
piezoelectric driver in dual actuator disk drive,” Dec. 13 2005. US
Patent 6,975,123.
8
| 3cs.SY
|
Understanding Deep Architectures by Interpretable
Visual Summaries
Marco Carletti∗ , Marco Godi∗ , Maedeh Aghaei† and Marco Cristani∗
∗ University
† University
of Verona, Italy; Email: [email protected]
of Barcelona, Spain; Email: [email protected]
Summary 1
Summary 2
Summary 3
robin
arXiv:1801.09103v1 [cs.CV] 27 Jan 2018
bald
eagle
pickup
police
van
Fig. 1: Examples of visual summaries for AlexNet [1]. Each summary contains crisp regions used by the network to classify
objects, exhibiting common semantic parts highlighted in red. Is it easy to see that, in the robin class, the network systematically
considers the head (Summary 1), the body (Summary 2), the legs and the lower body (Summary 3). Best in colors.
Abstract—A consistent body of research investigates the recurrent visual patterns exploited by deep networks for object
classification with the help of diverse visualization techniques.
Unfortunately, no effort has been spent in showing that these
techniques are effective in leading researchers to univocal and
exhaustive explanations. This paper goes in this direction, presenting a visualization framework owing to a group of clusters
or summaries, each one formed by crisp image regions focusing
on a particular part that the network has exploited with high
regularity to classify a given class. In most of the cases, these
parts carry a semantic meaning, making the explanation simple
and universal. For example, the method suggests that AlexNet,
when classifying the ImageNet class “robin”, is very sensible to
the patterns of the head, the body, the legs, the wings and the
tail, providing five summaries where these parts are consistently
highlighted. The approach is composed by a sparse optimization
step providing sharp image masks whose perturbation causes
high loss in the classification. Regions composing the masks
are then clustered together by means of a proposal flow-based
similarity score, that associates visually similar patterns of diverse
objects which are in corresponding positions. The final clusters
are visual summaries easy to be interpreted, as found by the very
first user study of this kind. The summaries can be also used
to compare different architectures: for example, the superiority
of GoogleNet w.r.t. AlexNet is explained by our approach since
the former gives rise to more summaries, indicating its ability in
capturing a higher number of diverse semantic parts.
I. I NTRODUCTION
Distilling the visual information exploited by a deep network for classifying an image is important, since it allows
to understand what makes an object class understandable
by a machine, foreseeing potential failures and highlighting
differences among network architectures [2], [3], [4], [5]. For
such a task, diverse visualization strategies have been designed
so far: starting from the simple visualization of neuron filters
and activations of [3], the works of [2], [6], [7], [3] individuate
those images which activate a certain neuron the most; [8]
and [5] generate dreamlike images bringing the classifier to
high classification scores; image samples are masked in [9],
highlighting more those parts which drive a classifier toward
a class, adding in [4] the parts that are against a certain class.
However, no study has been carried out yet to verify
whether these systems are successful in conveying clear
messages to the researchers, fostering human explanations
that are expressive (much information is disclosed) and
highly shared (diverse researchers give the same explanation).
Notably, most of the previous visualization systems analyze
single images: this brings to cumbersome human interpretation
procedures where hundreds of images per class ought to be
evaluated, undermining the usability of the whole system.
In this work, we provide two main contributions. The first
is a visual understanding system indicating the visual parts a
classifier consistently uses for determining a class; the second
is a user study validating the usability of the system, showing
that the users quickly understand these visual parts with high
inter-rater reliability.
Our understanding system is composed by two phases: in
the first one, crisp image masks are extracted from each image,
indicating the visual patterns more important for a given
class, that is, those patterns that, if perturbed, owe to high
classification loss. The masks are found by an optimization
process borrowed from [10] and made sparse to provide
binary values: in facts, most literature on visualization provide
smooth masks where higher values mean higher importance
[2], [4], [7], [10], [11], [12], [13], [14] On the contrary,
we found that 1) having binary masks brings to the highest
classification loss; 2) smooth masks are more difficult to be
interpreted. Both of these claims are proven in the experiments.
The second phase takes the connected regions of the crisp
mask, grouping them across images by means of the affinity
propagation algorithm [15], in which the similarity measure
is borrowed by the proposal flow algorithm [16]: in practice,
regions around the same object part will be grouped together,
allowing for example to group the lights of the cars under
different poses. The region clusters are opportunely visualized
as mosaics, forming the visual summaries.
In the experiments, we show that our system brings to crisp
masks which cause systematically higher classification loss
than [10]; in the user study, we show that users do explain
the parts considered by a classifier (in term of textual tags)
with a considerable number of tags (much information), in a
shorter time, and with higher inter-rater reliability than what
can be obtained by the visualization method of [10]. Finally,
we show that our system allows to compare different network
architectures in terms of what they capture in the images,
motivating in a new way the superiority of the recent deep
models.
The rest of the paper is organized as follows: Sec.II reviews
the related works; our system is analyzed in Sec.III, with the
experiments following in Sec.IV. Finally, Sec.V concludes the
paper with some observations and future perspectives.
II. R ELATED W ORK
this case the general output is a map showing how crucial is
the covering of a particular area, that can be a pixel [2], [10] or
superpixel-level map [19]. Our approach is also perturbation
based, since it looks for crisp portions of images that, if
perturbed, maximally distract the classifier.
In all of the previous cases, the output are single images
the user has to interpret to explain the behavior of a particular classifier on a particular class. Unfortunately, the visual
semantics span of a visual class cannot be summarized in
a single image, neither a few of them, and the analysis on
hundreds of images becomes necessary. Only in this way one
can actively figure out which would be the critical cases for a
classifier, and the possible corrections a classifier has to take
into account. For example, it would be nice to understand if,
for the class “car”, which parts are recognizable and which one
are not systematically captured. This type of analysis could
be interesting for comparing different classifiers, or to suggest
an enrichment of the training data for letting the classifier
individuate that particular. At the best of our knowledge, our
approach is the first of its kind in doing such kind of analysis,
forging the definition of summarized visualization approach.
III. M ETHOD
Our method is composed by two phases, mask extraction
and clustering. The former captures what visual patterns are
maximally important for the classification, the latter organizes
the visual patterns into summaries, which will ease the understanding of what these patterns do represent.
A. Mask Extraction
We define a classifier as a function y = f (x) where x
is the input image and y is the classification score vector,
in our case the softmax output of the last layer of a deep
network. Our starting point is the gradient-based optimization
of [10]. In that method, the output of the optimization is a
mask m : Λ → [0, 1] with the same resolution of x, in which
higher values mean higher saliency. The original optimization
equation (Eq. (3) of [10]) is
Visualization approaches can be analyzed under different
perspectives: local techniques focus on the understanding of
single neurons by showing the filters or the activations [3];
under this umbrella, input-dependent approaches select the
m = argmin fc (Φ(x; m)) + λ1 k1 − mk1
(1)
images which activate a neuron the most [2], [5], [7]; global
m∈[0,1]Λ
approaches capture some general property of the network, like
the tendency in focusing on some parts of the images for the where Φ(x; m) is a perturbed version of x in correspondence
classification [4], [8], [10], [11], [12], [17]; these approaches of the non-zero pixels of m, in which the perturbation function
are given as input single images, providing smooth saliency Φ does blurring:
maps which highlight more those area which have been
[Φ(x; m)] (u) = ∫ gσ0 m(u) (v − u)x(v)dv
(2)
important for a specific class. Global approaches are mostly
gradient-based, computing the gradient of the class score with with u a pixel location, m(u) the mask value at u and σ0 the
respect to the input image [5], [7], [11], [17]. Our approach maximum isotropic standard deviation of the Gaussian blur
is a global technique. Other gradient-based approaches adds kernel gσ0 , σ0 = 10. The function fc (·) is the classification
activations to the analysis, obtaining edge-based images with score of the model for the class c: the idea is to find a mask
edges highlighted in correspondence of salient parts [11]. that perturbs the original image in a way that the classifier gets
Other gradient-based approaches individuate also the pixels maximally confused, rejecting the sample for that class. The
which are against a certain class [4]. Generative approaches second member of Eq. (1) is a L1-regularizer with strength
generate dreamlike images bringing the classifier to high λ1 , which guides the optimization to minimally perturb the
also
classification scores [8], [18]. Perturbation-based approaches pixels of the input image. The authors
P of [10] suggested
β
edit an input image and observe its effect on the output [4]. In a total variation (TV) regularizer u∈Λ k∇m(u)kβ , in which
r
(i )
j
Object
proposal
Extraction
+
Proposal
flow
P1
Pi
Corr
matrix
creation
PPi1 PPiN
PN
Pruning
Stage +
Corr’’
Matrix
Creation
Corr
P1
PPi1 PPiN
Pi
PN
Clustering
+
Post
processing
Summary k
Corr ' '
Fig. 2: The proposed approach, sketch of the clustering phase of Sec. III-B. Best in colors.
the sum operates on the β-normed partial derivatives on m,
calculated as the difference of the values of two contiguous
pixels according to the direction.
POur contribution consists in adding another regularizer
u∈Λ |1 − m(u)|m(u) enforcing sparsity [20] in the values
of the mask m, making it binary. This regularizer has been
designed to start working after a certain number of iterations,
so that we get a rough version of the mask before start
optimizing its crisp version.
The final version of the optimization is thus:
X
β
m = argmin fc (Φ(x; m))+λ1 k1 − mk1 +λ2
k∇m(u)kβ
m∈[0,1]Λ
u∈Λ
+ λ3
X
|1 − m(u)|m(u) (3)
u∈Λ
All of the λ’s and β have been found by crossvalidation on the
classification drop, so that λ1 = 0.01, λ2 = 0.0001, λ3 = 0
and β = 3 during the first 300 iterations. We then modified the
parameters to λ2 = 1, λ3 = 2 for the next 150 iterations. At
the end of the mask extraction stage, each image xi , i = 1...N
of a given class becomes associated to the corresponding mask
mi .
B. Clustering
Each saliency mask mi can be analyzed by considering its
(i)
connected components {rj }j=1...Ji called here regions. Some
of the regions will be clustered together and will form the
visual summaries. The idea is that each region represents an
articulated visual item composed by parts, and a summary is
an ensemble of regions exhibiting at least a common part. A
graphical sketch of the procedure is shown in Fig. 2.
In our implementation, we model the parts of the regions
using object proposals [21], and we cluster the regions borrowing from the proposal flow technique [16]. Recently, object
proposals have been found well-suited for matching, with
the proposal flow exploiting local and geometrical constraints
to compare structured objects exhibiting sufficiently diverse
poses.
Our procedure start by considering the whole images without resorting to the regions, in order to account as much as
possible of the context where regions are merged. Given a
class, all of its N images are processed; from image xi , the
set of object proposals Pi is extracted. All of the images are
thus matched in a pairwise fashion adopting the proposal flow
algorithm. Each pair of images < xi , xj > will thus produce
a Mi × Mj matrix P Pij , with Mi indicating the number of
object proposals found in image xi . Each entry of the matrix
P Pij (k, l) contains the matching compatibility between the
k-th and the l-th object proposal of the images xi and xj ,
respectively.
After this, all the object proposals of all the pairs of images
are combined
together into a NP × NP matrix Corr, where
P
NP = i=1...N Mi is the total number of object proposals.
A given row of Corr will contain the matching score of
a particular object proposal with all the remaining object
proposals. Corr could be very large but can made easily sparse
by thresholding the minimal admissible matching score.
(i)
At this point, we get back to the image regions {rj },
keeping from Corr all of the object proposals that overlap
sufficiently with a region (overlap ratio higher than 75%) and,
in the case of two overlapping proposals, one of them is
removed if the ratio between the two areas is less than a certain
threshold (set to 2 during the experiments). The pruning stage
leads to the Corr00 matrix.
The matrix Corr00 is considered as a similarity matrix, and
the Affinity Propagation clustering algorithm is applied [15] on
top of it. The resulting clusters are ensembles of parts which,
thanks to the proposal flow algorithm, should consistently
identify a particular portion of an articulated object, thus
carrying a clear visual semantics. Post-processing is carried
out: Structural Similarity Index (SSIM) [22] is applied to all
the pairs of a cluster, discarding it as inconsistent if the median
value of SSIM for that cluster is lower than a threshold based
on the global median of SSIM within the whole class (set
to 90% of the median during the experiments). This has the
purpose of removing obvious mistakes in the clusters, caused
by the variety of different poses that the proposal flow has not
been able to deal with1 .
All the parts of a valid cluster are highlighted in red and
shown surrounded by the regions they belong to; this eases
the human interpretation and provides a summary (see an
excerpt in Fig. 1). Each image class is thus formed by different
summaries, depending on the number of valid clusters that
have been kept.
IV. E XPERIMENTS
Three types of experiments are carried out: masks analysis, investigating how much an image perturbed with our
sparse masks confuses a classifier; user study, checking how
exhaustive, shared and fast is the human understanding of
what the summaries do represent about the network; practical
1 Experimentally we found that in some cases of objects oriented in opposite
directions, like cars towards right and left, proposal flow did not work properly
providing erroneously high matching scores, as for some complex not rigid
objects like animals in drastically different poses.
Fig. 3: Qualitative analysis of the masks. First row, original image from different Imagenet classes. Second line, heatmaps
computed with the method proposed by [10]. Third line, crisp masks computed with our optimization procedure. Best in colors.
application, where the summaries are used to motivate the
superiority of GoogleNet against AlexNet.
A. Masks analysis
We consider 18 classes of Imagenet, that have been randomly selected with the constraint of having instances of
classes that are adjacent in a dense [23] semantic space (See
Table I, adjacent classes are in subsequent rows with same
background color). This gives rise to pairwise fine-grained
classification instances which will be interesting for our user
study, allowing to check how different are the corresponding
summaries. From these classes we uniformly pick a total of
5000 images.
The masks obtained by our approach are compared with
those of the recent IEBB method [10].The protocol is the
same as [10]: given an image, the classification confidence
associated to it with respect to the ground truth class is
measured; in the case of a deep network, this corresponds
to checking the softmax output in the i-th entry for the i-th
object class. Afterwards, the image x is blurred as written in
Sec. III-A by using the corresponding mask m (the one of ours
and the IEBB one), the classification score is re-computed and
the drop w.r.t. the original image is calculated.
We only compare with IEBB, which is shown in [10] to
be the best in terms of classification drop, and for the sake
of space. In addition, we compare with IEBB thresh, which is
the modification presented in [10], in which the smooth mask
generated by IEBB is made crisp by a thresholding operation,
made on the intensity of the mask. This thresholding operation
is done independently on each image, in order to maximize the
drop: this is different and more expensive than our technique,
in which the optimization is done only once on all of the data.
Table II, third column, shows the classification loss of the two
approaches. Notably, we succeed in improving an already high
classification loss, closely reaching the saturation at 100%,
which cannot be reached since the softmax function. Interest-
TABLE I: Classes from ImageNet, classification accuracy of
AlexNet and the set of tags collected during the user study
with our approach.
Class Name
Robin
Acc.
83.8%
Bald eagle
78.6%
Golden retriever
59%
German shepherd
55.4%
Bullet train
91.6%
Steam locomotive
92%
Pick-up
61.4%
Police van
67.2%
Oboe
Saxophone
Crash helmet
Football helmet
Jeans
Miniskirt
Cowboy hat
Windsor tie
55.6%
64.6%
59.6%
78%
59.8%
51.4%
45.2%
57.2%
Sweatshirt
92%
Running shoes
79.8%
Most Proposed Tag per Summary
Head, Body, Legs, Wings, Tail
Head, Neck border, Eye, Beak
Face, Wing
Nose, Eye, Ear, Mouth,
Face, Legs, Head
Eye, Leg, Neck, Body,
Ear, Nose, Face, Feather
Front train, Front glass, Train,
Rails, Lights, Train body
Chimney, Front train, Wheels,
Engine, Side, Window
Mudguard, Step bumpers, Side window,
Windshield, Back, Wheel
Wheel, Police flag, Side window,
Light, Rear window, Vehicle,
Capote, Bumpers, Mudguard,
Body, Buttons
Body, Buttons, Bell
Base, Side, Front, Logo
Front grids, Logo, Side, People
Crotch, Pocket, Legs, Waistband
Face, Waistband, Leg, Head
Ear, Face, Chin
Pattern, Knot, Collar, Neck
Hoodie, Face, Arm, Laces, Wrinkles,
Neck
Laces, Logo, Shoe side
TABLE II: Mask Validation
Method
IEBB
IEBB thresh.
Ours
Ref.
ICCV17[10]
ICCV17[10]
%Drop (Var)
99.738365 (8.13e-4)
97.703865 (5.758e-3)
99.964912(< 10e-6)
ingly, with IEBB thresh, the overall performance diminishes,
with higher variance.
In Fig. 3, examples of the obtained masks using our
approach and IEBB are shown. In general, the sparse optimization finds solutions which are close to the IEBB one: in
facts, IEBB finds already masks which cause a nearly complete
Fig. 4: Coherency in terms of average Jaccard distance (yaxis) among the tags found with the user study, for our
proposed approach (blue), and IEBB approach (red). Lower
is better. The class labels come with two numbers, the tag
dictionary size with our approach (blue) and with IEBB (red),
respectively.
loss. Nonetheless, our improvement gives the same importance
to all of the pixels, facilitating the human interpretation, other
than ensuring a higher drop. In addition, noise is removed,
especially on the background, which is also fruitful for the
interpretation. For example, in the pick-up class there are visible three regions (tire and glasses). The classes of police van,
golden retriever, bald eagle, football helmet the background
cleaning is total. The other classes are basically noisier, since
the masks are usually focusing on small portions of the object
instead of the whole. Interestingly, in the class bullet train,
our approach highlights that the AlexNet focuses steadily on
visual patterns portraying train stations, instead of covering
the train silhouette.
B. User study
Our user study quantifies how much a visualization system
succeeds in communicating to the human users the visual
patterns used by the deep network for the classification, in an
expressive, agreeable and fast way. To validate this, we exploit
textual tags, adopted by the users to state what the visualization
system is communicating. In practice, for each image class, the
user is asked to list a set of tags describing the regions which
have been consistently highlighted by the visualization system.
As in the previous experiment, we compare our approach
against IEBB [10].
A total of 30 users are hired (21 male, 9 female subjects);
the average age is 33 (std:8.4), with all of the subjects being
comfortable with using computer for visualizing images. Half
of the subjects are given the IEBB approach, half our system.
All of them are asked to analyze the visualization results for
the 18 classes, indicating what they are showing. Each class
consists of about 100 images. With IEBB, the user is asked
to observe a single scrollable page in which pairs <original
image, smooth mask overlapping the image> are arranged in
a 4 × 50 item plate (see the first two rows of Fig. 3 to get an
idea of the content). On the interface, a text field is used to
host the tags written by the users. Only nouns are allowed, no
verbs or adjectives. No predefined tags are adopted.
With our approach, for each class we show to a user all of
the summaries, each one of them equipped with 10 regions,
with the object proposal used to build the cluster highlighted
in red (as example, see Fig. 1). Even in this case, the user
is able to scroll the summaries and writing tags in a separate
text field. No time constraints are given to the experiment:
each subject was free to work until she/he was happy with the
list of tags associated to the class. Nonetheless, the timing of
the experiments was collected.
Given a visualization system, the tags collected over all the
users on a single class were used to create a tag dictionary.
Synonimous terms where manually fused, leading to a set of
T different tags. Then, the output of each user was encoded
as a T -dimensional binary vector having 1 at position l in the
case the l-th tag was used. Obviously, different classes lead to
different dictionaries, as so as the same classes analyzed with
different visualization approaches.
In order to compare the tag consistency among users that
use the same visualization system, we compute the Jaccard
distance between all the user descriptions, averaging over
the number of users [24]. Jaccard distance is essentially the
intersection over union usually adopted to compare binary
data.
Fig. 4 shows the Jaccard average distance, as soon as
the number T of tags for each class (on the x-labels). As
visible, our visualization system allows to discover a consistent
number of tags, with a higher coherency. IEBB gives rise to
a higher number of tags, but in most of the cases they are
outliers given by single users with no shared support.
In particular, Table I lists on the right the tags indicating the
patterns individuated in the summaries, in bold representing
the tags which do not overlap with those given by IEBB
(omitted here for the lack of space). Interesting observations
can be assessed: the classes with lower accuracy, cowboy hat,
miniskirt, have visual patterns focusing on people instead of
the object of interest; adjacent classes have many common
visual patterns (german shepard, golden retriever), and usually
highly specific patterns help the classification (the “chimney”
pattern in the Steam locomotive class).
As for the total time that users spent for their job, IEBB
needs ≈ 34.5 minute, ours ≈ 10 minute, showing that expressive and shared descriptions are also obtained very fast.
C. Practical Application: GoogleNet VS AlexNet
The architectural differences between GoogleNet [25] and
AlexNet [1] are widely known in the community, as much as
the differences in the performances (around 10% over the top-5
error rate). We add another facet to this comparison, adopting
our approach and discovering that in general GoogleNet is
capable of capturing systematically more semantic parts than
what AlexNet does, explaining once more its superiority. On
the same classes analyzed above, we visualize the GoogleNet
summaries. In average, our approach finds more summaries for
GoogleNet (7) than AlexNet (5). Looking at the summaries,
Fig. 5: Motivating the superiority of GoogleNet against AlexNet: focusing on the pick-up class (ImageNet), our approach finds
9 summaries for the former architecture, 6 for the latter, showing that GoogleNet is capable of capturing more semantics,
individuating salients parts as the cabin, the body and the front (summary tags suggested by the user study). Best in colors.
one can find that actually GoogleNet individuates systematically more semantic parts than what AlexNet is doing. For
example, in Fig. 5 the pick-up class is taken into account. For
each summary, a single image is shown, together with the most
used textual tags found by the user study (bottom). GoogleNet
has more summaries focusing on important visual parts as the
cabin, the body and the front of the car. Similar reasoning are
possible with the other classes.
V. C ONCLUSION
Our approach is the first visualization system which puts
the emphasis in transmitting to the human user a clear and
concise visual glimpse about what a classifier has captured
from images. The approach uses crisp regions which are
shown to be maximally important for the classification, thanks
to a sparse optimization approach which let them emerge.
Therefore, the regions are organized by visual summaries,
each focusing on precise parts found by a robust clustering
process. The user study promotes the summaries, showing
higher inter rater reliability in having different tags, collected
faster than considering the output of unstructured, per-image
visualization tool as [10]. One may observe that our framework
misses an automatic tagging procedure, and that it would be
nice to let make automatic what the user study has done
manually, that is, giving textual tags to the summaries. This is
an interesting research question: from one side we can confirm
this is a direction worth to be investigated. At the same time, in
some occasion the visual knowledge represents the best way
to understand things, and the use of textual references can
be misleading. For example in fashion, one can immediately
capture the difference between two clothing styles by looking
at them, which will be much simpler than describing them
verbally. Therefore, our next step will be to apply the approach
to clothing data, analyzing what the content of the summaries
would be with classifiers classifying different clothing styles
as [26].
R EFERENCES
[1] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classification
with deep convolutional neural networks,” in Advances in neural information processing systems, pp. 1097–1105, 2012.
[2] B. Zhou, A. Khosla, A. Lapedriza, A. Oliva, and A. Torralba, “Object
detectors emerge in deep scene cnns,” arXiv preprint, 2014.
[3] J. Yosinski, J. Clune, A. Nguyen, T. Fuchs, and H. Lipson, “Understanding neural networks through deep visualization,” arXiv preprint.
[4] L. M. Zintgracef, T. S. Cohen, T. Adel, and M. Welling, “Visualizing
deep neural network decisions: Prediction difference analysis,”
[5] M. D. Zeiler and R. Fergus, “Visualizing and understanding convolutional networks,” in ECCV, pp. 818–833, Springer, 2014.
[6] M. D. Zeiler and R. Fergus, “Visualizing and understanding convolutional networks,” in ECCV, pp. 818–833, Springer, 2014.
[7] A. Dosovitskiy and T. Brox, “Inverting visual representations with
convolutional networks,” in CVPR, pp. 4829–4837, 2016.
[8] K. Simonyan, A. Vedaldi, and A. Zisserman, “Deep inside convolutional
networks: Visualising image classification models and saliency maps,”
[9] D. Erhan, Y. Bengio, A. Courville, and P. Vincent, “Visualizing higherlayer features of a deep network,” UofM, vol. 1341, p. 3, 2009.
[10] R. C. Fong and A. Vedaldi, “Interpretable explanations of black boxes
by meaningful perturbation,” in ICCV, Oct 2017.
[11] R. R. Selvaraju, A. Das, R. Vedantam, M. Cogswell, D. Parikh, and
D. Batra, “Grad-cam: Why did you say that? visual explanations from
deep networks via gradient-based localization,” arXiv, 2016.
[12] B. Zhou, A. Khosla, A. Lapedriza, A. Oliva, and A. Torralba, “Learning
deep features for discriminative localization,” in CVPR, 2016.
[13] A. Mahendran and A. Vedaldi, “Salient deconvolutional networks,” in
Proceedings of the ECCV, 2016.
[14] J. Zhang, Z. Lin, J. Brandt, X. Shen, and S. Sclaroff, “Top-down neural
attention by excitation backprop,” in ECCV, Springer, 2016.
[15] B. J. Frey and D. Dueck, “Clustering by passing messages between data
points,” Science, vol. 315, no. 5814, pp. 972–976, 2007.
[16] B. Ham, M. Cho, , C. Schmid, and J. Ponce, “Proposal flow,” in CVPR.
[17] A. Mahendran and A. Vedaldi, “Understanding deep image representations by inverting them,” in CVPR, pp. 5188–5196, 2015.
[18] A. Nguyen, J. Yosinski, and J. Clune, “Deep neural networks are
easily fooled: High confidence predictions for unrecognizable images,”
in CVPR, pp. 427–436, 2015.
[19] M. T. Ribeiro, S. Singh, and C. Guestrin, “Why should i trust you?:
Explaining the predictions of any classifier,” in ACM SIGKDD.
[20] R. Tibshirani, “Regression shrinkage and selection via the lasso,” Journal of the Royal Statistical Society. Series B (Methodological), 1996.
[21] J. R. Uijlings, K. E. Van De Sande, T. Gevers, and A. W. Smeulders,
“Selective search for object recognition,” IJCV, vol. 104, no. 2, 2013.
[22] Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli, “Image
quality assessment: from error visibility to structural similarity,” IEEE
TIP, vol. 13, no. 4, pp. 600–612, 2004.
[23] J. Deng, A. C. Berg, K. Li, and L. Fei-Fei, “What does classifying more
than 10,000 image categories tell us?,” in ECCV, Springer, 2010.
[24] L. Kaufman and P. J. Rousseeuw, Finding groups in data: an introduction to cluster analysis, vol. 344. John Wiley & Sons, 2009.
[25] C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan,
V. Vanhoucke, and A. Rabinovich, “Going deeper with convolutions,”
in CVPR, pp. 1–9, 2015.
[26] M. H. Kiapour, K. Yamaguchi, A. C. Berg, and T. L. Berg, “Hipster
wars: Discovering elements of fashion styles,” in ECCV, 2014.
| 1cs.CV
|
arXiv:1411.6719v3 [math.ST] 1 May 2016
Asymptotically Optimal Discrete Time
Nonlinear Filters From Stochastically Convergent
State Process Approximations
Dionysios S. Kalogerias and Athina P. Petropulu
May 2016
Abstract
We consider the problem of approximating optimal in the Minimum Mean Squared Error
(MMSE) sense nonlinear filters in a discrete time setting, exploiting properties of stochastically
convergent state process approximations. More specifically, we consider a class of nonlinear,
partially observable stochastic systems, comprised by a (possibly nonstationary) hidden stochastic process (the state), observed through another conditionally Gaussian stochastic process (the
observations). Under general assumptions, we show that, given an approximating process which,
for each time step, is stochastically convergent to the state process, an approximate filtering
operator can be defined, which converges to the true optimal nonlinear filter of the state in a
strong and well defined sense. In particular, the convergence is compact in time and uniform in
a completely characterized measurable set of probability measure almost unity, also providing a
purely quantitative justification of Egoroff’s Theorem for the problem at hand. The results presented in this paper can form a common basis for the analysis and characterization of a number
of heuristic approaches for approximating optimal nonlinear filters, such as approximate grid
based techniques, known to perform well in a variety of applications.
Keywords. Approximate Nonlinear Filtering, Hidden Models, Partially Observable Systems,
Stochastic Processes, C-Weak Convergence, Change of Probability Measures.
1
Introduction
Nonlinear stochastic filtering refers to problems in which a stochastic process, usually called the
state, is partially observed as a result of measuring another stochastic process, usually called the
observations or measurements, and the objective is to estimate the state or some functional of it,
based only on past and present observations. The nonlinearity is due to the general, possibly non
Gaussian nature of the state and observations processes, as well as the fact that, in general, the
state may be partially observed as a nonlinear functional of the observations. Usually, nonlinear
state estimators are designed so as to optimize some performance criterion. Most commonly, this
corresponds to the Minimum Mean Squared Error (MMSE), which is also adopted in this work.
A desirable feature of a nonlinear filter is recursiveness in time, as it greatly reduces computational complexity and allows for real time estimation as new measurements become available.
The Authors are with the Department of Electrical & Computer Engineering, Rutgers, The State University of
New Jersey, 94 Brett Rd, Piscataway, NJ 08854, USA. e-mail: {d.kalogerias, athinap}@rutgers.edu.
This paper constitutes an extended and most up-to-date version of [1]. The work presented herein is supported
by the National Science Foundation (NSF) under Grant CNS-1239188.
1
However, not all nonlinear filters possess this important property [2, 3]. Recursive nonlinear filters
exist for some very special cases, such as those in which the transition model of the state process is linear (Gauss-Markov), or when the state is a Markov chain (discrete state space) [4–7].
In the absence of recursive filter representations, practical filtering schemes have been developed,
which typically approximate the desired quantities of interest, either heuristically (e.g., Gaussian
approximations [8, 9]) or in some more powerful, rigorous sense (e.g., Markov chain approximations
[10, 11]).
In this paper, we follow the latter research direction. Specifically, we consider a partially observable system in discrete time, comprised by a hidden, almost surely compactly bounded state
process, observed through another, conditionally Gaussian measurement process. The mean and
covariance matrix of the measurements both constitute nonlinear, time varying and state dependent
functions, assumed to be known apriori. Employing a change of measure argument and using the
original measurements, an approximate filtering operator can be defined, by replacing the “true”
state process by an appropriate approximation. Our contribution is summarized in showing that if
the approximation converges to the state either in probability or in the C-weak sense (Section II.C),
the resulting filtering operator converges to the true optimal nonlinear filter in a relatively strong
and well defined sense; the convergence is compact in time and uniform in a measurable set of
probability measure almost unity (Theorem 3). The aforementioned set is completely characterized
in terms of a subset of the parameters of the filtering problem of interest. Consequently, our results
provide a purely quantitative justification of Egoroff’s theorem [12] for the problem at hand, which
concerns the equivalence of almost sure convergence and almost uniform convergence of measurable
functions.
To better motivate the reader, let us describe two problems that fit the scenario described
above and can benefit from the contributions of this paper, namely, those of sequential channel
state estimation and (sequential) spatiotemporal channel prediction [13] (see also [14]). The above
problems arise naturally in novel signal processing applications in the emerging area of distributed,
autonomous, physical layer aware mobile networks [15–17]. Such networks usually consist of cooperating mobile sensors, each of them being capable of observing its communication channel (under
a flat fading assumption), relative to a reference point in the space. In most practical scenarios,
the dominant quantities characterizing the wireless links, such as the path loss exponent and the
shadowing power, behave as stochastic processes themselves. For instance, such behavior may be
due to physical changes in the environment and also the inherent randomness of the communication
medium itself. Then, the path loss exponent and the shadowing power can be collectively considered
as the hidden state (suggestively called the channel state) of a partially observable system, where
the channel gains measured at each sensor can be considered as the corresponding observations.
In general, such observations are nonlinear functionals of the state. Assuming additionally that
the channel state is a Markov process, the main results presented herein can essentially provide
strong asymptotic guarantees for approximate sequential nonlinear channel state estimation and
spatiotemporal channel prediction, enabling physical layer aware motion planning and stochastic
control. For more details, the reader is referred to [13].
The idea of replacing the process of interest with some appropriate approximation is borrowed
from [11]. However, [11] deals almost exclusively with continuous time stochastic systems and the
results presented in there do not automatically extend to the discrete time system setting we are
dealing with here. In fact, the continuous time counterparts of the discrete time stochastic processes
considered here are considerably more general than the ones treated in [11]. More specifically,
although some relatively general results are indeed provided for continuous time hidden processes,
2
[11] is primarily focused on the standard hidden diffusion case, which constitutes a Markov process
(and aiming to the development of recursive approximate filters), whereas, in our setting, the hidden
process is initially assumed to be arbitrary (as long as it is confined to a compact set). Also, different
from our formulation (see above), in [11], the covariance matrix of the observation process does not
depend on the hidden state; the state affects only the mean of the observations. Further, the modes
of stochastic convergence considered here are different compared to [11] (in fact, they are stronger),
both regarding convergence of approximations and convergence of approximate filters.
The results presented in this paper provide a framework for analyzing a number of heuristic
techniques for numerically approximating optimal nonlinear filters in discrete time, such as approximate grid based recursive approaches, known to perform well in a wide variety of applications
[13, 18, 19]. Additionally, our results do not refer exclusively to recursive nonlinear filters. The
sufficient conditions which we provide for the convergence of approximate filtering operators are
independent of the way a filter is realized (see Section III). This is useful because, as highlighted
in [20], no one prevents one from designing an efficient (approximate) nonlinear filter which is part
recursive and part nonrecursive, or even possibly trying to combine the best of both worlds, and
there are practical filters designed in this fashion [20].
The paper is organized as follows: In Section II, we introduce the system model, along with some
mild technical assumptions on its structure and also present/develop some preliminary technical
results and definitions, which are important for stating and proving our results. In Section III, we
formulate our problem in detail and present our main results (Theorem (3)), along with a simple
instructive example. Section IV is exclusively devoted to proving the results stated in Section III.
Finally, Section V concludes the paper.
Notation: In the following, the state vector will be represented as Xt , its approximations as
LS
Xt , and all other matrices and vectors, either random or not, will be denoted by boldface letters
(to be clear by the context). Real valued random variables and abstract random elements will be
denoted by uppercase letters. Calligraphic letters and formal script letters will denote sets and
σ-algebras, respectively. The operators (·)T , λmin (·), λmax (·) will denote transposition, minimum
and maximum eigenvalue, respectively. For any random element (same for variable, vector) Y ,
σ {Y } will denote the σ-algebra generated by Y . The ℓp norm of a vector x ∈ Rn is kxkp ,
P
1/p
( ni=1 |xi |p ) , for all naturals p ≥ 1. The spectral
and Frobenius norms of any matrix X ∈ Rn×n
qP
2
n,n
are kXk2 , maxkxk2 ≡1 kXxk2 and kXkF ,
i,j=1 Xij , respectively. Positive definiteness and
semidefiniteness of X will be denoted by X ≻ 0 and X 0, respectively. For any Euclidean space
RN ×1 , IN ×N will denote the respective identity operator. Additionally, throughout the paper, we
employ the identifications R+ ≡ [0, ∞), R++ ≡ (0, ∞), N+ ≡ {1, 2, . . .}, N+
n ≡ {1, 2, . . . , n} and
+
Nn ≡ {0} ∪ Nn , for any positive natural n.
2
Partially Observable System Model & Technical Preliminaries
In this section, we give a detailed description of the partially observable (or hidden) system model of
interest and present our related technical assumptions on its components. Additionally, we present
some essential background on the measure theoretic concept of change of probability measures and
state some definitions and known results regarding specific modes of stochastic convergence, which
will be employed in our subsequent theoretical developments.
3
2.1
Hidden Model: Definitions & Technical Assumptions
First, let us set the basic probabilistic framework, as well as precisely define the hidden system
model considered throughout the paper:
• All stochastic processes considered below are fundamentally generated on a common complete
probability space (the base space), defined by a triplet (Ω, F , P), at each time instant taking
values in a measurable state space, consisting of some Euclidean subspace and the associated
Borel σ-algebra on that subspace. For example, for each t ∈ N, the
Xt
≡ Xt (ω),
state process
M ×1
M ×1
where ω ∈ Ω, takes its values in the measurable state space R
,B R
, where
B RM ×1 constitutes the Borel σ-algebra of measurable subsets of RM ×1 .
• In this work, the evolution mechanism of state process Xt is assumed to be arbitrary. However,
in order to avoid unnecessary technical complications, we assume that, for each t ∈ N, the
induced probability measure of Xt is absolutely continuous with respect to the Lebesgue
measure on its respective state space. Then, by the Radon-Nikodym Theorem, it admits a
density, unique up to sets of zero Lebesgue measure. Also, we will generically assume that
for all t ∈ N, Xt ∈ Z almost surely, where Z constitutes a compact strict subset of RM ×1 .
In what follows, however, in order to lighten the presentation, we will assume that M ≡ 1.
Nevertheless, all stated results hold with the same validity if M > 1 (See also Assumption 2
below).
• The state Xt is partially observed through the observation process
yt , µt (Xt ) + σ t (Xt ) + ξt ∈ RN ×1 ,
∀t ∈ N,
o
n
where, conditioned on Xt and for each t ∈ N, the sequence µt : Z 7→ RN ×1
(1)
t∈N
is known
apriori, the process σ t (Xt ) ∼ N (0, Σt (Xt ) ≻ 0) constitutes Gaussian noise, with the seN ×N
quence {Σ
, also known apriori, and
t : Z 7→ DΣ}t∈N , where DΣ is a bounded subset of R
i.i.d.
ξ t ∼ N 0, σξ2 IN ×N .
As a pair, the state Xt and the observations process described by (1) define a very wide family
of partially observable systems. In particular, any Hidden Markov Model (HMM) of any order,
in which the respective Markov state process is almost surely confined in a compact subset of its
respective Euclidean state space, is indeed a member of this family. More specifically, let us rewrite
(1) in the canonical form
p
(2)
yt ≡ µt (Xt ) + Ct (Xt )ut ∈ RN ×1 , ∀t ∈ N,
where ut ≡ ut (ω) constitutes a standard Gaussian white noise process and, for all x ∈ Z, Ct (x) ,
Σt (x) + σξ2 IN ×N ∈ DC , with DC bounded. Then, for a possibly nonstationary HMM of order m,
assuming the existence of an explicit functional model for describing the temporal evolution of the
state (being a Markov process of order m), we get the system of standardized stochastic difference
equations
Xt ≡ ft {Xt−i }i∈N+ , W t ∈ Z
m
,
p
yt ≡ µt (Xt ) + Ct (Xt )ut
4
∀t ∈ N,
(3)
a.s.
where, for each t, ft : Z m × W 7→ Z (with Z m , ×m times Z) constitutes a measurable nonlinear
state transition mapping and W t ≡ W t (ω) ∈ W ⊆ RMW ×1 , denotes a (discrete time) white noise
process with state space W. For a first order stationary HMM, the above system of equations
reduces to
Xt ≡ f (Xt−1 , W t ) ∈ Z
, ∀t ∈ N,
(4)
p
yt ≡ µt (Xt ) + Ct (Xt )ut
which arguably constitutes the most typical partially observable system model encountered in both
Signal Processing and Control, with plethora of important applications.
Let us also present some more specific assumptions, regarding the nature (boundedness, continuity and expansiveness) of the aforementioned sequences of functions.
Assumption 1: (Boundedness) For later reference, let
λinf , inf inf λmin (Ct (x)) ,
(5)
λsup , sup sup λmax (Ct (x)) ,
(6)
µsup , sup sup kµt (x)k2 ,
(7)
t∈N x∈Z
t∈N x∈Z
t∈N x∈Z
where each quantity of the above is uniformly and finitely bounded for all t ∈ N and for all x ∈ Z.
If x is substituted by the stochastic process Xt (ω), then all the above definitions continue to hold in
the essential sense. For technical reasons related to the bounding-from-above arguments presented
in Section IV, containing the proof of the main result of the paper, it is also assumed that λinf > 1,
a requirement which can always be satisfied by appropriate normalization of the observations.
Assumption
2:o (Continuity & Expansiveness) All members of the functional family
n
N ×1
µt : Z 7→ R
are uniformly Lipschitz continuous, that is, there exists a bounded constant
t∈N
Kµ ∈ R+ , such that, for all t ∈ N,
kµt (x) − µt (y)k2 ≤ Kµ |x − y| , ∀ (x, y) ∈ Z × Z.
o
n
Additionally, all members of the functional family Σt : Z →
7 DΣ ⊂ RN ×N
t∈N
(8)
are elementwise
uniformly Lipschitz continuous, that is, there exists some universal and bounded constant KΣ ∈ R+ ,
+
such that, for all t ∈ N and for all (i, j) ∈ N+
N × NN ,
ij
Σij
t (x) − Σt (y) ≤ KΣ |x − y| ,
∀ (x, y) ∈ Z × Z.
(9)
If x is substituted by the stochastic process Xt (ω), then all the above statements are understood
in the almost sure sense.
Remark 1. As we have already said, for simplicity, we assume that Z ⊂ R, that is, M ≡ 1. In any
other case (when M > 1), we modify the Lipschitz assumptions stated above simply by replacing
|x − y| with kx − yk1 , that is, Lipschitz continuity is meant to be with respect to the ℓ1 norm in
the domain of the respective function. If this holds, everything that follows works also in RM >1 ,
just with some added complexity in the proofs of the results. Also, because kxk2 ≤ kxk1 for any
x ∈ RM , the assumed Lipschitz continuity with respect to ℓ1 norm can be replaced by Lipschitz
continuity with respect to the ℓ2 norm, since the latter implies the former, and again everything
holds. Further, if M > 1, convergence in probability and L1 convergence of random vectors are
both defined by replacing absolute values with the ℓ1 norms of the vectors under consideration.
5
2.2
Conditional Expectations, Change of Measure & Filters
Before proceeding with the general formulation of our estimation problem and for later reference,
let us define the complete natural filtrations of the processes Xt and yt as
oo
n n
and
(10)
{Xt }t∈N , σ {Xi }i∈Nt
oo t∈N
n n
,
(11)
{Yt }t∈N , σ {yi }i∈Nt
t∈N
respectively, and also the complete filtration generated by both Xt and yt as
oo
n n
.
{Ht }t∈N , σ {Xi , yi }i∈Nt
t∈N
(12)
In all the above, σ {Y } denotes the σ-algebra generated by the random variable Y .
In this work, we adopt the MMSE as an optimality criterion. In this case, one would ideally
like to discover a solution to the stochastic optimization problem
2
bt
inf Xb E Xt − X
t
(13)
n
o 2 , ∀t ∈ N,
b
b
subject to E Xt Yt ≡ Xt
bt to the subset of
where the constraint is equivalent to confining the search for possible estimators X
interest, that is, containing the ones which constitute Yt -measurable random variables. Of course,
the solution to the program (13) coincides with the conditional expectation [21]
bt ,
E { Xt | Yt } ≡ X
∀t ∈ N,
(14)
which, in the nonlinear filtering literature, is frequently called a filter. There is also an alternative
bt , using the concept of change of probability
and very useful way of reexpressing the filter process X
measures, which will allow us to stochastically decouple the state and observations of our hidden
system and then let us formulate precisely the approximation problem of interest in this paper.
Change of measure techniques have been extensively used in discrete time nonlinear filtering, mainly
in order to discover recursive representations for various hidden Markov models [3, 6, 7, 22]. In the
following, we provide a brief introduction to these type of techniques (suited to our purposes) which
is also intuitive, simple and technically accessible, including direct proofs of the required results.
Change of Probability Measure in Discrete Time: Demystification & Useful Results
So far, all stochastic processes we have considered are defined on the base space (Ω, F , P). In
fact, it is the structure of the probability measure P that is responsible for the coupling between
the stochastic processes
and yt , being, for each t ∈ N, measurable functions from (Ω, F ) to
Xt
N
N
, respectively. Intuitively, the measure P constitutes our “reference
(R, B (R)) and R , B R
measurement tool” for measuring the events contained in the base σ-algebra F , and any random
variable serves as a “medium” or “channel” for observing these events.
As a result, some very natural questions arise from the above discussion. First, one could ask
if and under what conditions it is possible to change the probability measure P, which constitutes
our fixed way of assigning probabilities to events, to another measure Pe on the same measurable
e
space (Ω, F ), in a way such that there exists some sort of transformation connecting P and P.
6
e could we choose the latter probability
Second, if we can indeed make the transition from P to P,
measure in a way such that the processes Xt and yt behave according to a prespecified statistical
e Xt and yt constitute independent stochastic
model? For instance, we could demand that, under P,
processes. Third and most important, is it possible to derive an expression for the “original” filter
bt ≡ EP { Xt | Yt } under measure P, using only (conditional) expectations under Pe (denoted as
X
EPe { ·| ·})?
The answers to all three questions stated above are affirmative under very mild assumptions
and the key result in order to prove this assertion is the Radon-Nikodym Theorem [23]. However,
assuming that the induced joint probability measure of the processes of interest is absolutely continuous with respect to the Lebesgue measure of the appropriate dimension, in the following we
provide an answer to these questions, employing only elementary probability theory, avoiding the
direct use of the Radon-Nikodym Theorem.
Theorem 1. (Conditional Bayes’ Theorem for Densities) Consider the (possibly vector)
stochastic processes Xt (ω) ∈ RNt ×1 and Yt (ω) ∈ RMt ×1 , both defined on the same measurable space
(Ω, F ), for all t ∈ N. Further, if P and Pe are two probability measures on (Ω, F ), suppose that:
e the process Xt is integrable.
• Under both P and P,
e the induced joint probability measure of
• Under
the base probability
measure P (resp. P),
{Xi }i∈Nt , {Yi }i∈Nt is absolutely continuous with respect to the Lebesgue measure of the
appropriate dimension, implying the existence of a density ft (resp. fet ), with
ft :
× R
i∈Nt
Ni ×1
×
× R
i∈Nt
Mi ×1
7→ R+ .
(15)
• For each set of points, it is true that
fet (· · · ) ≡ 0
⇒
ft (· · · ) ≡ 0,
(16)
or, equivalently, the support of ft is contained in the support of fet .
Also, for all t ∈ N, define the Likelihood Ratio (LR) at t as the {Ht }-adapted, nonnegative stochastic
process1
f (X , X , . . . , Xt , Y0 , Y1 , . . . , Yt )
.
(17)
Λt , t 0 1
fet (X0 , X1 , . . . , Xt , Y0 , Y1 , . . . , Yt )
Then, it is true that
bt ≡ EP { Xt | Yt } ≡
X
EPe { Xt Λt | Yt }
,
EPe { Λt | Yt }
(18)
almost everywhere with respect to P.
Proof of Theorem 1. See the Appendix.
1
With zero probability of confusion, we use {Yt }t∈N and {Ht }t∈N to denote the complete filtrations generated by
Yt and {Xt , Yt }.
7
Remark 2. The {Ht }-adapted LR process
Λt ≡ Λt Xt , {Xi }i∈Nt , Yt , {Yi }i∈Nt ,
t ∈ N,
(19)
as defined in (17), actually coincides with the restriction of the Radon-Nikodym derivative of P
with respect to Pe to the filtration {Ht }t∈N , that is,
dP (ω)
e (ω)
dP
Ht
≡ Λt (Xt (ω) , Yt (ω)) ,
∀t ∈ N,
(20)
a statement which, denoting the collections {xi }i∈Nt and {yi }i∈Nt as xt and yt , respectively, is
rigorously equivalent to
ˆ
e (ω)
Λt (Xt (ω) , Yt (ω)) dP
P (F) ≡
≡
ˆF
B
Λt xt , yt d2t Pe(Xt ,Yt ) xt , yt
≡ P(Xt ,Yt ) (B) ≡ P ((Xt , Yt ) ∈ B) ,
∀F , {ω ∈ Ω |(Xt (ω) , Yt (ω)) ∈ B } ∈ Ht and
Ni ×1
Mi ×1
∀B ∈ ⊗ B R
⊗ ⊗ B R
,
i∈Nt
i∈Nt
(21)
(22)
(23)
∀t ∈ N,
(24)
respectively (in the above, “⊗” denotes the product operator for σ-algebras). Of course, the existence and almost everywhere uniqueness of Λt are guaranteed by the Radon-Nikodym Theorem,
e
provided that the base measure P is absolutely continuous with respect to Pe on Ht (P ≪Ht P).
e
Further, for the case where there exist densities characterizing P and P (as in Theorem 1), demanding that P ≪Ht Pe is precisely equivalent to demanding that (16) is true and, again through
the Radon-Nikodym Theorem, it can be easily shown that the derivative Λt actually coincides with
the likelihood ratio process defined in (17), almost everywhere.
Now, let us apply Theorem 1 for the stochastic processes Xt and yt , comprising our partially
observed system, as defined in Section II.A. In this respect, we present the following result.
Theorem 2. (Change of Measure for the Hidden System under Study) Consider the hidden
stochastic system of Section II.A on the usual base space (Ω, F , P), where Xt ∈ Z and yt ∈ RN ×1 ,
almost surely ∀t ∈ N, constitute the hidden state process and the observation process, respectively.
Then, there exists an alternative, equivalent to P, base measure Pe on (Ω, F ), under which:
• The processes Xt and yt are statistically independent.
• Xt constitutes a stochastic process with exactly the same dynamics as under P.
• yt constitutes a Gaussian vector white noise process with zero mean and covariance matrix
equal to the identity.
bt can be expressed as in (18), where the {Ht }-adapted stochastic process
Additionally, the filter X
Λt , t ∈ N is defined as in (25) (top of next page).
8
−1
1
1
kyi k22 − (yi − µi (Xi ))T Σi (Xi ) + σξ2 IN ×N
(yi − µi (Xi ))
Y
2
2
r
Λt ,
,
λi
2
i∈Nt
i∈N
t
det Σi (Xi ) + σξ IN ×N
−1
1X
kyi k22 − (yi − µi (Xi ))T Σi (Xi ) + σξ2 IN ×N
(yi − µi (Xi ))
exp
2
i∈Nt
r
≡
∈ R++ (25)
Y
2
det Σi (Xi ) + σξ IN ×N
Y exp
i∈Nt
Proof of Theorem 2. Additionally to the similar identifications made above (see (19)) and for later
reference, let
Y t , {yi }i∈Nt and yt , {y i }i∈Nt .
(26)
e this way showing its existence. To accomplish this,
First, we construct the probability measure P,
define, for each t ∈ N, a probability measure PeRt on the measurable space (Rt , B (Rt )), where
Rt ,
× R ×
i∈Nt
× R
i∈Nt
N ×1
,
(27)
being absolutely continuous with respect to the Lebesgue measure on (Rt , B (Rt )) and with density
fet : R 7→ R+ . Since, for each t ∈ N, the processes Xt (ω) and yt (ω) are both, by definition, fixed
and measurable functions from (Ω, Ht ) to (Rt , B (Rt )), with2
(
)
[
Ht ⊆ H∞ , σ
Ht ⊆ F ,
(28)
t∈N
measuring any B ∈ B (Rt ) under PeRt can be replaced by measuring the event (preimage)
e defined collectively for all t ∈ N on
{ ω ∈ Ω| (Xt , Y t ) ∈ B} ∈ Ht under another measure, say P,
the general measurable space (Ω, H∞ ) as
Pe ({ ω ∈ Ω| (Xt , Y t ) ∈ B}) ≡ Pe ((Xt , Y t ) ∈ B) , PeRt (B) ,
∀B ∈ B (Rt ) .
e to the σ-algebra H∞ is induced by the probThat is, the restriction of the probability measure P
eR (also see Kolmogorov’s Extension Theorem [3]). Further, in order to define
ability measure P
∞
e fully on (Ω, F ), we have to extend its behavior on the remaining
the alternative base measure P
events which belong to the potentially finer σ-algebra F but are not included in H∞ . However,
since we are interested in change of measure only for the augmented process (Xt , Y t ), these events
are irrelevant to us. Therefore, Pe can be defined arbitrarily on these events, as long as it remains
a valid and consistent probability measure.
2
H∞ constitutes the join, that is, the smallest σ-algebra generated by the union of all Ht , ∀t ∈ N.
9
Now, to finalize the construction of the restriction of Pe to Ht , ∀t ∈ N, we have to explicitly
specify the density of PeRt , or, equivalently, of the joint density of the random variables (Xt , Y t ), fet ,
for all t ∈ N. According to the statement of Theorem 2, we have to demand that
fet xt , yt ≡ feY t |Xt yt xt feXt (xt )
= feY t yt fXt (xt )
Y
=
feyi (y i ) fXt (xt )
i∈Nt
ky i k22
exp
−
2
Y
q
=
i∈Nt
(2π)N
≡
exp −
q
1X
2
i∈Nt
(2π)
!
ky i k22
N (t+1)
fX (xt )
t
fXt (xt ) .
(29)
Next, by definition, we know that, under P, the joint density of (Xt , Y t ) can be expressed as
ft xt , yt ≡ f Y t |Xt yt xt fXt (xt )
Y
≡
f yi |Xi ( y i | xi ) fXt (xt )
i∈Nt
y Ti C−1
i yi
−2
!
Y exp
q
=
i∈Nt det (Ci ) (2π)N
X y T C−1 y
fX (xt )
t
i
i
−2
i∈Nt
fXt (xt ) ,
≡
q
Yp
N
(t+1)
det (Ci ) (2π)
exp
i
(30)
i∈Nt
where, for all t ∈ N,
y t ≡ y t (xt ) , y t − µt (xt ) ∈ RN ×1
Ct ≡ Ct (xt ) ≡ Σt (xt ) +
σξ2 IN ×N
and
∈ DC ,
(31)
(32)
where DC constitutes a bounded subset of RN ×N . From (29) and (30), it is obvious that the
sufficient condition (16) of Theorem 1 is satisfied (actually, in this case, we have an equivalence; as
10
a result, the change of measure is an invertible transformation). Applying Theorem 1, (18) must
be true by defining the {Ht }-adapted stochastic process
f Y t |Xt ( Y t | Xt )
f Y t |Xt ( Y t | Xt )
ft (Xt , Y t )
≡
≡
fet (Xt , Y t )
feY t |Xt ( Y t | Xt )
feY t (Y t )
X kyi k2 − (yi (Xi ))T (Ci (Xi ))−1 yi (Xi )
2
exp
2
i∈Nt
Yp
≡
,
det (Ci (Xi ))
Λt ≡ Λt (Xt , Y t ) ,
(33)
i∈Nt
or, alternatively,
X kyi k2 − (y (Xi ))T (Ci (Xi ))−1 y (Xi )
i
i
2
exp
2
Y
Y
i∈Nt
p
,
Λt ≡
λi ,
(X
))
det
(C
i
i
i∈N
i∈N
t
t
therefore completing the proof.
2.3
(34)
Weak & C-Weak Convergence of (Random) Probability Measures
In the analysis that will take place in Section IV, we will make use of the notions of weak and conditionally weak (C-weak) convergence of sequences of probability measures. Thus, let us define these
notions of stochastic convergence consistently, suited at least for the purposes of our investigation.
Definition 1. (Weak Convergence [24]) Let S be an arbitrary metric space, let S , B (S) be
the associated Borel σ-algebra and consider a sequence of probability measures {πn }n∈N on S . If
π constitutes another “limit” probability measure on S such that
lim πn (A) = π (A) ,
n→∞
∀A ∈ S such that π (∂A) ≡ 0,
(35)
where ∂A denotes the boundary set of the Borel set A, then we say that the sequence {πn }n∈N
converges to π weakly or in the weak sense and we equivalently write
W
πn −→ π.
n→∞
(36)
Of course, weak convergence of probability measures is equivalent to weak convergence or convergence in distribution, in case we are given sequences of (S, S )-valued random variables whose
induced probability measures converge in the aforementioned sense.
Next, we present a definition for conditionally weak convergence of probability measures. To
avoid possibly complicating technicalities, this definition is not presented in full generality. Rather,
it is presented in an appropriately specialized form, which will be used later on, in the analysis that
follows.
11
Definition 2. (Conditionally Weak
Convergence)
Let (Ω, F , P) be a base probability triplet
and consider the measurable spaces Si , Si , B (Si ) , i = {1, 2}, where S1 and S2 constitute a
complete separable metric (Polish) space and an arbitrary metric space, respectively. Also, let
{X1n : Ω → S1 }n∈N be a sequence of random variables, let X2 : Ω → S2 be another random variable
and consider the sequence of (regular) induced conditional probability distributions (or measures)
P nX n |X : S1 × Ω → [0, 1], such that
1
2
P nX n |X ( A| X2 (ω)) ≡ P ( X1n ∈ A| σ {X2 }) ,
2
1
(37)
P − a.e., for any Borel set A ∈ S1 . If X1 : Ω → S1 constitutes a “limit” random variable, whose
induced conditional measure P X1 |X2 : S1 × Ω → [0, 1] is such that
lim P nX n |X ( A| X2 (ω)) = P X1 |X2 ( A| X2 (ω)) ,
n→∞
(38)
∀A ∈ S1 such that π (∂A) ≡ 0 and P − a.e.,
o
n
converges to P X1 |X2 conditionally weakly ( C-weakly)
then we say that the sequence P nX n |X
2
1
2
1
n∈N
or in the conditionally weak ( C-weak) sense and we equivalently write
W
P nX n |X ( ·| X2 ) −→ P X1 |X2 ( ·| X2 ) .
1
n→∞
2
(39)
Remark 3. Actually, C-weak convergence, as defined above, is strongly related to the more general
concepts of almost sure weak convergence and random probability measures. For instance, the
reader is referred to the related articles [25] and [26].
Further, the following lemma characterizes weak convergence of probability measures (and random variables) [24].
Lemma 1. (Weak Convergence & Expectations) Let S be an arbitrary metric space and let
S , B (S). Suppose we are given a sequence of random variables {X n }n∈N and a “limit” X, all
(S, S )-valued, but possibly defined on different base probability spaces, with {PX n }n∈N and PX being
their induced probability measures on S , respectively. Then,
D
W
n→∞
n→∞
X n −→ X ⇔ PX n −→ PX ,
(40)
if and only if
n
E {f (X )} ≡
ˆ
f dPX −→
n
S
ˆ
n→∞ S
f dPX ≡ E {f (X)} ,
(41)
for all bounded, continuous functions f : S → R.
Of course, if we replace weak convergence by C-weak convergence, Lemma 1 continues to hold,
but, in this case, (41) should be understood in the almost everywhere sense (see, for example, [26]).
More specifically, under the generic notation of Definition 2 and under the appropriate assumptions
according to Lemma 1, it will be true that
E { f (X1n )| X2 } (ω) −→ E { f (X1 )| X2 } (ω) ,
n→∞
for almost all ω ∈ Ω.
12
(42)
3
Problem Formulation & Statement of Main Results
In this section, we formulate the problem of interest, that is, in a nutshell, the problem of approximating a nonlinear MMSE filter by another (asymptotically optimal) filtering operator, defined by
replacing the true process we would like to filter by an appropriate approximation. Although we
do not deal with such a problem here, such an approximation would be chosen in order to yield a
practically realizable approximate filtering scheme. We also present the main result of the paper,
establishing sufficient conditions for convergence of the respective approximate filters, in an indeed
strong sense.
Let us start from the beginning. From Theorem 2, we know that
EP { Xt | Yt } ≡
EPe { Xt Λt | Yt }
,
EPe { Λt | Yt }
∀t ∈ N,
(43)
where the RHS constitutes an alternative representation for the filter on the LHS, which constitutes
the optimal in the MMSE sense estimator of the partially observed process Xt , given the available
observations up to time t. If the numerical evaluation of either of the sides of (43) is difficult
(either we are interested in a recursive realization of the filter or not), one could focus on the
RHS, where the state and the observations constitute independent processes, and, keeping the
same observations, replace Xt by another process XtA , called the approximation, with resolution or
approximation parameter A ∈ N (for simplicity), also independent of the observations (with respect
e for which the evaluation of the resulting “filter” might be easier. Under some appropriate,
to P),
well defined sense, the approximation to the original process improves as A → ∞. This general idea
of replacing the true state process with an approximation is employed in, for instance, [10, 11], and
will be employed here, too.
At this point, a natural question arises: Why are we complicating things with change of measure
arguments and not using XtA directly in the LHS of (43)? Indeed, using classical results such as the
Dominated Convergence Theorem, one could prove at least pointwise convergence of the respective
filter approximations. The main and most important issue with such an approach is that, in order
for such a filter to be realizable in any way, special attention must be paid to the choice of the
approximation, regarding its stochastic dependence on the observations process. This is due to
the original stochastic coupling between the state and the observations of the hidden system of
interest. However, using change of measure, one can find an alternative representation of the filter
process, where, under another probability measure, the state and observations are stochastically
decoupled (independent). This makes the problem much easier, because the approximation can also
be chosen to be independent of the observations. If we especially restrict our attention to recursive
nonlinear filters, change of measure provides a rather versatile means for discovering recursive filter
realizations. See, for example, the detailed treatment presented in [3].
Thus, concentrating on the RHS of (43), we can define an approximate filtering operator of the
process Xt , also with resolution A ∈ N, as
o
n
EPe XtA ΛAt Yt
o , ∀t ∈ N.
n
(44)
E A ( Xt | Yt ) ,
EPe ΛAt Yt
Observe that the above quantity is not a conditional expectation of XtA , because XtA does not follow
the probability law of the true process of interest, Xt [11]. Of course, the question is if and under
13
which sense,
?
E A ( Xt | Yt ) −→ EP { Xt | Yt } ,
A→∞
(45)
that is, if and in which sense our chosen approximate filtering operator is asymptotically optimal,
as the resolution of the approximation increases. In other words, we are looking for a class of
approximations, whose members approximate the process Xt well, in the sense that the resulting
approximate filtering operators converge to the true filter as the resolution parameter increases,
that is, as A → ∞, and under some appropriate notion of convergence. In this respect, below we
formulate and prove the following theorem, which constitutes the main result of this paper (recall
the definition of C-weak convergence given in Section II.C). In the following, 1A : R → {0, 1} denotes
the indicator of the set A. Also, for any Borel set A, 1A (·) constitutes a Dirac (atomic) probability
measure. Equivalently, we write 1A (·) ≡ δ(·) (A).
Theorem 3. (Convergence to the Optimal Filter) Pick any natural T < ∞ and suppose either
of the following:
n o
• For all t ∈ NT , the sequence XtA
is marginally C-weakly convergent to Xt , given Xt , that
A∈N
is,
W
(46)
P AX A |X ( ·| Xt ) −→ δXt (·) , ∀t ∈ NT .
t
A→∞
t
n o
• For all t ∈ NT , the sequence XtA
A∈N
is (marginally) convergent to Xt in probability, that is,
P
XtA −→ Xt ,
A→∞
∀t ∈ NT .
(47)
b T ⊆ Ω with P-measure at least 1−(T + 1)1−CN exp (−CN ),
Then, there exists a measurable subset Ω
such that
(48)
lim sup sup E A ( Xt | Yt ) − EP { Xt | Yt } (ω) ≡ 0,
A→∞ t∈N
T
bT
ω∈Ω
for any free, finite constant C ≥ 1. In other words, the convergence of the respective approximate
filtering operators is compact in t ∈ N and, with probability at least 1 − (T + 1)1−CN exp (−CN ),
uniform in ω.
Interestingly, as noted in the beginning of this section, the mode of convergence of the resulting
approximate filtering operator is particularly strong. In fact, it is interesting that, for fixed T , the
approximate filter E A ( Xt | Yt ) converges to EP { Xt | Yt } (uniformly) in a set that approaches the
certain event, exponentially in N . That is, convergence to the optimal filter tends to be in the
uniformly almost everywhere sense, at an exponential rate (in N ). Consequently, it is revealed that
the dimensionality of the observations process essentially stabilizes the behavior of the approximate
filter, in a stochastic sense. Along the lines of the discussion presented above, it is clear that
Theorem 3 provides a way of quantitatively justifying Egoroff’s theorem [12], which bridges almost
uniform convergence with almost sure convergence, however in an indeed abstract fashion.
Remark 4. The C-weak convergence condition (46) is a rather strong one. In particular, as we
show later in Lemma 8 (see Section IV), it implies L1 convergence, which means that it also
implies (marginal) convergence in probability (which constitutes the alternative sufficient condition
of Theorem 3). In simple words, (46) resembles a situation where, at any time step, one is given
14
or defines an approximation to the original process, in the sense that, conditioned on the original
process at the same time step, the probability of being equal to the latter approaches unity. At
this point, because C-weak convergence is stronger than (and implies) convergence in probability,
one could wonder why we presented both as alternative sufficient conditions for filter convergence
in Theorem 3 (and also in Lemma 10 presented in Section IV). The reason is that, contrary to
convergence in probability, condition (46) provides a nice structural criterion for constructing state
process approximations in a natural way, which is also consistent with our intuition: If, at any time
step, we could observe the value of true state process, then the respective value of the approximation
at that same time step should be “sufficiently close” to the value of the state. Condition (46)
expresses this intuitive idea and provides a version of the required sense of “closeness”.
In order to demonstrate the applicability of Theorem 3, as well as demystify the C-weak convergence condition (46), let us present a simple but illustrative example. The example refers to
a class of approximate grid based filters, based on the so called marginal approximation [14, 18],
according to which the (compactly restricted) state process is fed into a uniform spatial quantizer
of variable resolution. As we will see, this intuitively reasonable approximation idea constitutes a
simple instance of the condition (95).
More specifically, assume that Xt ∈ [a, b] ≡ Z, ∀t ∈ N, almost surely. Let us discretize Z
uniformly into A subintervals, of identical length, calledn cells.
o The l-th cell and its respective center
l
l
+
l
are denoted as ZA and xA , l ∈ NA . Then, letting XA , xA
, the quantizer QA : (Z, B (Z)) 7→
+
l∈NL
S
XA , 2XA is defined as the bijective and measurable function which uniquely maps the l-th cell to
l
l
the respective reconstruction point xlA , ∀l ∈ N+
A . That is, QA (x) , xA if and only if x ∈ ZA [14].
Having defined the quantizer QA (·), the Marginal Quantization of the state is defined as [18]
XtA (ω) , QA (Xt (ω)) ∈ XA , ∀t ∈ N, P − a.s.,
(49)
where A ∈ N is identified as the approximation parameter. That is, Xt is approximated by its nearest
neighbor on the cell grid. That is, the state is represented by a discrete set of reconstruction points,
each one of them uniquely corresponding to a member of a partition of Z.
By construction of marginal state approximations, it can be easily shown that [14]
P−a.s.
XtA (ω) −→ Xt (ω) ,
(50)
A→∞
a fact that will be used in the following. Of course, almost sure convergence implies convergence in
probability and, as we will see, C-weak convergence as well. First, let us determine the conditional
probability measure P AX A |X ( dx| Xt ). Since knowing Xt uniquely determines the value of XtA , it
t
t
must be true that
P AX A |X ( dx| Xt ) ≡ P AQA (Xt )|Xt ( dx| Xt )
t
t
≡ δQA (Xt ) (dx) ,
P − a.s..
(51)
However, from Lemma 1,
n we
know
that
o weak convergence of measures is equivalent to showing
A
that the expectations E f Xt Xt converge to E { f (Xt )| Xt } ≡ f (Xt ), for all bounded and
continuous f (·), almost everywhere. Indeed,
ˆ
n
o
A
f (x) P AX A |X ( dx| Xt (ω))
E f Xt Xt (ω) ≡
t
Z
15
t
≡
ˆ
Z
f (x) δQA (Xt (ω)) (dx)
P−a.s.
≡ f (QA (Xt (ω))) −→ f (Xt (ω)) ,
A→∞
(52)
due to the continuity of f (·). Consequently, we have shown that
W
P AX A |X ( ·| Xt ) ≡ δQA (Xt ) (·) −→ δXt (·) ,
t
A→∞
t
(53)
fulfilling the first requirement of Theorem 3. This very simple example constitutes the basis for constructing more complicated and cleverly designed state approximations (for example, using stochastic quantizers). The challenge here is to come up with such approximations exhibiting nice properties, which would potentially lead to the development of effective approximate recursive or, in
general, sequential filtering schemes, well suited for dynamic inference in complex partially observable stochastic nonlinear systems. As far as grid based approximate recursive filtering is concerned,
a relatively complete discussion of the problem is presented in the recent paper [14], where marginal
state approximations are also treated in full generality.
An important and direct consequence of Theorem 3, also highlighted by the example presented
above, is that, interestingly, the nature of the state process is completely irrelevant when one
is interested in convergence of the respective approximate filters, in the respective sense of the
aforementioned theorem. This fact has the following pleasing and intuitive interpretation: It implies
that if any of the two conditions of Theorem 3 are satisfied, then we should forget about the internal
stochastic structure of the state, and instead focus exclusively on the way the latter is being observed
through time. That is, we do not really care about what we partially observe, but how well we observe
it; and if we observe it well, we can filter it well, too. Essentially, the observations should constitute
a stable functional of the state, of course in some well defined sense. In this work, this notion of
stability is expressed precisely through Assumption 1 and 2, presented earlier in Section II.
Note, however, that the existence of a consistent approximate filter in the sense of Theorem 3
does not automatically imply that this filter will be efficiently implementable; usually, we would like
such a filter to admit a recursive/sequential representation (or possibly a semirecursive one [20]). As
it turns out, this can happen when the chosen state approximation admits a valid semimartingale
type representation (in addition to satisfying one of the sufficient conditions of Theorem 3). For
example, the case where the state is Markovian and the chosen state approximation is of the marginal
type, discussed in the basic example presented above, is treated in detail in [14].
Remark 5. The filter representation (43) coincides with the respective expression employed in ime the observations and state
portance sampling [19, 27]. Since, under the alternative measure P,
constitute statistically independent processes, one can directly sample from the (joint) distribution
of the state, fixing the observations to their respective value at each time t (of course, assuming that
a relevant “sampling device” exists). However, note that that due to the assumptions of Theorem
3, related at least to convergence in probability of the corresponding state approximations, the
aforementioned result cannot be used directly in order to show convergence of importance sampling
or related particle filtering techniques, which are directly related to empirical measures. The possible ways Theorem 3 can be utilized in order to provide asymptotic guarantees for particle filtering
(using additional assumptions) constitutes an interesting open topic for further research.
The rest of the paper is fully devoted in the detailed proof of Theorem 3.
16
4
Proof of Theorem 3
In order to facilitate the presentation, the proof is divided in a number of subsections.
4.1
Two Basic Lemmata, Linear Algebra - Oriented
Parts of the following useful results will be employed several times in the analysis that follows3 .
Lemma 2. Consider arbitrary matrices A ∈ CN1 ×M1 , B ∈ CN1 ×M1 , X ∈ CM2 ×N2 , Y ∈ CM2 ×N2 ,
and let k·kM be any matrix norm. Then, the following hold:
• If either
– N1 ≡ M1 ≡ 1, or
– N1 ≡ N2 ≡ M1 ≡ M2 and k·kM is submultiplicative,
then
kAX − BYkM ≤ kAkM kX − YkM + kYkM kA − BkM .
(54)
• If N2 ≡ 1, M1 ≡ M2 and k·kM constitutes any subordinate matrix norm to the ℓp vector norm,
k·kp , then
kAX − BYkp ≤ kAkM kX − Ykp + kYkp kA − BkM .
(55)
Proof of Lemma 2. We prove the result only for the case where N1 ≡ N2 ≡ M1 ≡ M2 and k·kM is
submultiplicative. By definition of such a matrix norm,
kAX − BYkM ≡ kAX + AY − AY − BYkM
≡ kA (X − Y) + (A − B) YkM
≤ kA (X − Y)kM + k(A − B) YkM
≤ kAkM k(X − Y)kM + kYkM k(A − B)kM ,
(56)
apparently completing the proof. The results for the other two cases considered in Lemma 2 can
be readily shown following similar procedure.
Lemma 3. Consider the collections of arbitrary, square matrices
o
o
n
n
Ai ∈ CN ×N
and
Bi ∈ CN ×N
i∈Nn
Then, for any submultiplicative matrix norm k·kM , it is true that
n
n
n
i−1
n
Y
X
Y
Y
Y
Bi
Ai −
≤
Aj M
Bj
i=0
i=0
M
i=0
j=0
j=i+1
3
i∈Nn
M
.
kAi − Bi k .
M
(57)
In this paper, Lemma 3 presented in this subsection will be applied only for scalars (and where the metric
considered coincides with the absolute value). However, the general version of the result (considering matrices and
submultiplicative norms) is presented for the sake of generality.
17
Proof of Lemma 3. Applying Lemma 2 to the LHS of (57), we get
n
Y
i=0
Ai −
n
Y
i=0
Bi
M
≡ A0
n
Y
i=1
Ai − B0
≤ kA0 kM
n
Y
i=1
n
Y
Bi
i=1
n
Y
Ai −
M
i=1
n
Y
+
Bi
Bi
i=1
M
M
kA0 − B0 kM .
(58)
The repeated application of Lemma 2 to the quantity multiplying kA0 kM on the RHS of the expression above yields
n
Y
i=0
Ai −
n
Y
i=0
Bi
M
≤ kA0 kM kA1 kM
n
Y
i=2
+ kA0 kM
Ai −
n
Y
n
Y
Bi
i=2
Bi
i=2
M
M
n
Y
kA1 − B1 kM +
Bi
i=1
M
kA0 − B0 kM , (59)
where, the “temporal pattern” is apparent. Indeed, iterating (59) and proceeding inductively, we
end up with the bound
n
i−1
n
n
n
X
Y
Y
Y
Y
Aj M
≤
Bi
Bj
kAi − Bi kM
(60)
Ai −
i=0
i=0
M
i=0
j=0
j=i+1
M
and the result readily follows invoking the submultiplicativeness of k·kM .
4.2
Preliminary Results
Here, we present and prove a number of preliminary results, which will help us towards the proof
of an important lemma, which will be the key to showing the validity of Theorem 3.
First, under Assumption 2, stated in Section II.A, the following trivial lemmata hold.
Lemma 4. Each member of the functional family {Σt : Z 7→ DΣ }t∈N is Lipschitz continuous on Z,
in the Euclidean topology induced by the Frobenius norm. That is, ∀t ∈ N,
kΣt (x) − Σt (y)kF ≤ (N KΣ ) |x − y| ,
(61)
∀ (x, y) ∈ Z × Z, for the same constant KΣ ∈ R+ , as defined in Assumption 2. The same also holds
for the family {Ct : Z 7→ DC }t∈N .
Proof of Lemma 4. By definition of the Frobenius norm,
v
2
u
X
ij
kΣt (x) − Σt (y)kF ≡ u
Σij
(x)
−
Σ
(y)
t
t
t
+
≤
≡
s
q
+
(i,j)∈NN ×NN
X
+
+
(i,j)∈NN ×NN
2
KΣ
|x − y|2
2
N 2 KΣ
|x − y|2 ,
∀t ∈ N
(62)
and our first claim follows. The second follows trivially if we recall the definition of each Ct (x).
18
Lemma 5. For each member of the functional family {Ct : Z 7→ DC }t∈N , it is true that, ∀t ∈ N,
|det (Ct (x)) − det (Ct (y))| ≤ (N KDET ) kCt (x) − Ct (y)kF ,
(63)
∀ (x, y) ∈ Z × Z, for some bounded constant KDET ≡ KDET (N ) ∈ R+ , possibly dependent on N
but independent of t.
Proof of Lemma 5. As a consequence of the fact that the determinant of a matrix can be expressed
as a polynomial function in N 2 variables (for example, see the Leibniz formula), it must be true
that, ∀t ∈ N,
X
ij
|det (Ct (x)) − det (Ct (y))| ≤ KDET
Cij
t (x) − Ct (y)
+
+
(i,j)∈NN ×NN
≡ KDET kCt (x) − Ct (y)k1 ,
(64)
where the constant KDET depends on maximized (using the fact that the domain DC is bounded)
(N − 1)-fold products of elements of Ct (x) and Ct (y), with respect to x (resp. y) and t. Consequently, although KDET may depend on N , it certainly does not depend on t. Now, since the ℓ1
entrywise norm of an N × N matrix corresponds to the norm of a vector with N 2 elements, we may
further bound the right had side of the expression above by the Frobenius norm of Ct (x) − Ct (y),
yielding
|det (Ct (x)) − det (Ct (y))| ≤ N KDET kCt (x) − Ct (y)kF ,
(65)
which is what we were set to prove.
Remark 6. The fact that the constant KDET may be a function of the dimension of the observation
vector, N , does not constitute a significant problem throughout our analysis, simply because N
is always considered a finite and fixed parameter of our problem. However, it is true that the
(functional) way N appears in the various constants in our derived expressions can potentially
affect speed of convergence and, for that reason, it constitutes an important analytical aspect.
Therefore, throughout the analysis presented below, a great effort has been made in order to keep
the dependence of our bounds on N within reasonable limits.
We also present
lemma, related to the expansiveness of each member of the
n another useful o
−1
functional family Ct : Z 7→ DC−1
.
t∈N
Lemma 6. Each member of the functional family
n
C−1
t : Z 7→ DC−1
o
t∈N
is Lipschitz continuous
on Z, in the Euclidean topology induced by the Frobenius norm. That is, ∀t ∈ N,
−1
C−1
t (x) − Ct (y)
F
≤ KIN V |x − y| ,
(66)
∀ (x, y) ∈ Z × Z, for some bounded constant KIN V ≡ KIN V (N ) ∈ R+ , possibly dependent on N
but independent of t.
Proof of Lemma 6. As a consequence of Laplace’s formula for the determinant of a matrix and
invoking Lemma 2, it is true that
−1
C−1
t (x) − Ct (y)
F
19
adj (Ct (x)) adj (Ct (y))
−
det (Ct (x)) det (Ct (y)) F
kadj (Ct (x)) − adj (Ct (y))kF
|det (Ct (x)) − det (Ct (y))|
+ kadj (Ct (y))kF
,
≤
det (Ct (x))
det (Ct (x)) det (Ct (y))
≡
(67)
where adj (A) denotes the adjugate of the square matrix A. Since Ct (x) (resp. Ct (y)) is a
symmetric and positive definite matrix, so is its adjugate. Employing one more property regarding
the eigenvalues of the adjugate [28] and the fact that λinf > 1, we can write
kadj (Ct (y))kF ≤
≡
≡
≤
√
√
√
N kadj (Ct (y))k2
N λmax (adj (Ct (y)))
Y
N max
λj (Ct (y))
+
√
i∈NN j6=i
N det (Ct (y)) ,
(68)
and then (67) becomes
−1
C−1
t (x) − Ct (y)
F
≤
≤
kadj (Ct (x)) − adj (Ct (y))kF √ |det (Ct (x)) − det (Ct (y))|
+ N
det (Ct (x))
det (Ct (x))
kadj (Ct (x)) − adj (Ct (y))kF
λN
inf
+
N 3 KDET KΣ
λN
inf
|x − y| .
(69)
Next, the numerator of the first fraction from the left may be expressed as
v
2
u
X
ij
ij
kadj (Ct (x)) − adj (Ct (y))kF ≡ u
adj
(C
(x))
−
adj
(C
(y))
t
t
t
+
v
u
≡u
t
≡
s
+
(i,j)∈NN ×NN
X
+
+
(i,j)∈NN ×NN
X
+
+
(i,j)∈NN ×NN
2
(−1)i+j Mij (Ct (x)) − Mij (Ct (y))
2
Mij (Ct (x)) − Mij (Ct (y)) ,
(70)
where Mij (Ct (x)) denotes the (i, j)-th minor of Ct (x), which constitutes the determinant of the
(N − 1) × (N − 1) matrix formulated by removing the i-th row and the j-th column of Ct (x).
Consequently, from Lemma 5, there exists a constant Kdet , possibly dependent on N , such that,
∀t ∈ N,
s
X
2
2
N 4 Kdet
KΣ
|x − y|2 ,
(71)
kadj (Ct (x)) − adj (Ct (y))kF ≤
+
+
(i,j)∈NN ×NN
or, equivalently,
kadj (Ct (x)) − adj (Ct (y))kF ≤ N 3 Kdet KΣ |x − y| ,
20
(72)
∀ (x, y) ∈ Z × Z. Therefore, combining with (69), we get
C−1
t (x) −
C−1
t (y)
≤
F
N3
λN
inf
(KDET + Kdet ) KΣ |x − y|
−3/ log(λinf )
≤
27λinf
log λinf
3 (KDET + Kdet ) KΣ |x − y|
, KIN V |x − y| ,
(73)
and the proof is complete.
Next, we state the following simple probabilistic result, related to the expansiveness of the norm
e considered
of the observation vector in a stochastic sense, under both base measures P and P
throughout the paper (see Section II.B).
Lemma 7. Consider the random quadratic form
Qt (ω) , kyt (ω)k22 ≡ kyt (Xt (ω)) + µt (Xt (ω))k22 ,
t ∈ N.
(74)
Then, for any fixed t ∈ N and any freely chosen C ≥ 1, there exists a bounded constant γ > 1, such
that the measurable set
)
(
Tt ,
satisfies
ω ∈ Ω sup Qi (ω) < γCN (1 + log (t + 1))
(75)
i∈Nt
n
o
exp (−CN )
,
min P (Tt ) , Pe (Tt ) ≥ 1 −
(t + 1)CN −1
(76)
that is, the sequence of quadratic forms {Qi (ω)}i∈Nt is uniformly bounded with very high probability
e
under both base measures P and P.
Proof of Lemma 7. First, it is true that
kyt (ω)k22 ≡ kyt (Xt (ω)) + µt (Xt (ω))k22
≡ kyt (Xt (ω))k22 + 2ytT (Xt (ω)) µt (Xt (ω)) + kµt (Xt (ω))k22
≤ kyt (Xt (ω))k22 + 2 kyt (Xt (ω))k2 µsup + µ2sup .
(77)
Also, under P, for each t ∈ N, the random variable yt (Xt ) constitutes an N -dimensional, conditionally (on Xt ) Gaussian random variable with zero mean and covariance matrix Ct (Xt ), that
is
(78)
yt |Xt ∼ N 0, Ct (Xt ) ≡ Cyt |Xt .
Then, if Xt is given,
Qt (ω) , kyt (Xt (ω))k22
(79)
can be shown to admit the very useful alternative representation (for instance, see [29], pp. 89 - 90)
X
Qt ≡
λj (Ct (Xt )) Uj2 , ∀t ∈ N, with
(80)
+
j∈NN
21
Uj
i.i.d.
+
j∈NN
∼ N (0, 1) .
(81)
From (80), one can readily observe that the statistical dependence of Qt on Xt concentrates only on
the eigenvalues of the covariance matrix Ct (Xt ), for which we have already assumed the existence
of a finite supremum explicitly (see Assumption 1). Consequently, conditioning on the process Xt ,
we can bound (80) as
X 2
Qt ≤ λsup
Uj , λsup U, with U ∼ χ2 (N ) ,
(82)
+
j∈NN
almost everywhere and everywhere in time, where the RHS is independent of Xt . Next, from ([30],
p. 1325), we know that for any chi squared random variable U with N degrees of freedom,
√
(83)
P U ≥ N + 2 N u + 2u ≤ exp (−u) , ∀u > 0.
Setting u ≡ CN (1 + log (t + 1)) for any C ≥ 1 and any t ∈ N,
P U ≥ N + 2N
p
C (1 + log (t + 1)) + 2CN (1 + log (t + 1))
≤
exp (−CN )
(t + 1)CN
.
(84)
a statement which equivalently means that, with probability at least 1 − (t + 1)−CN exp (−CN ),
p
(85)
U < N + 2N C (1 + log (t + 1)) + 2CN (1 + log (t + 1)) .
However, because the RHS of the above inequality is upper bounded by 5CN (1 + log (t + 1)),
P (U < 5CN (1 + log (t + 1)))
p
exp (−CN )
. (86)
≥ P U < N + 2N C (1 + log (t + 1)) + 2CN (1 + log (t + 1)) ≥ 1 −
(t + 1)CN
Hence, ∀i ∈ Nt ,
P Qi ≥ 5λsup CN (1 + log (t + 1)) Xi ≤ P (U ≥ 5CN (1 + log (t + 1)))
exp (−CN )
≤
,
(t + 1)CN
(87)
and, thus,
P Qi ≥ 5λsup CN (1 + log (t + 1)) Xi dPXi
ˆ
exp (−CN )
exp (−CN )
≤
.
dPXi ≡
CN
(t + 1)
(t + 1)CN
P Qi ≥ 5λsup CN (1 + log (t + 1)) =
ˆ
(88)
However, we would like to produce a bound on the supremum of all the Qi , i ∈ Nt . Indeed, using
the naive union bound,
X
[
P Qi ≥ λsup 5CN (1 + log (t + 1))
Qi ≥ 5λsup CN (1 + log (t + 1)) ≤
P
i∈Nt
i∈Nt
22
≤
(t + 1) exp (−CN )
(t + 1)
CN
≡
exp (−CN )
(t + 1)CN −1
or, equivalently,
P
!
sup Qi < 5λsup CN (1 + log (t + 1))
i∈Nt
≡P
Qi < 5λsup CN (1 + log (t + 1)) , ∀i ∈ Nt
≡P
\
i∈Nt
≥1−
Qi < 5λsup CN (1 + log (t + 1))
exp (−CN )
(t + 1)CN −1
,
(89)
holding true ∀t ∈ N. Consequently, working in the same fashion as above, it is true that, with at
least the same probability of success,
q
sup Qi (ω) < 5λsup CN (1 + log (t + 1)) + 2 5λsup CN (1 + log (t + 1))µsup + µ2sup
i∈Nt
< 5λsup 1 + 2µsup + µ2sup CN (1 + log (t + 1))
or, setting γ1 , 5λsup 1 + µsup
2
(90)
> 1,
sup Qi (ω) < γ1 CN (1 + log (t + 1)) .
(91)
i∈Nt
e yt constitutes a Gaussian vector white noise process
Now, under the alternative base measure P,
with zero mean and covariance matrix the identity, statistically independent of the process Xt (see
Theorem 2). That is, for each t, the elements of yt are themselves independent to each other. Thus,
for all t ∈ N and for all i ∈ Nt and using similar arguments as the ones made above, it should be
true that
e (Qi < 5CN (1 + log (t + 1))) ≥ 1 − exp (−CN )
(92)
P
(t + 1)CN
and taking the union bound, we end up with the inequality
!
exp (−CN )
Pe sup Qi < 5CN (1 + log (t + 1)) ≥ 1 −
.
i∈Nt
(t + 1)CN −1
(93)
Defining γ , max {γ1 , 5} ≡ γ1 , it must be true that, for all t ∈ N,
(
min P
!
sup Qi < γCN (1 + log (t + 1)) , Pe
i∈Nt
sup Qi < γCN (1 + log (t + 1))
i∈Nt
≥1−
!)
exp (−CN )
(t + 1)CN −1
, (94)
therefore completing the proof of the lemma.
23
Continuing our presentation of preliminary results towards the proof of Theorem (3) and leveraging the power of C-weak convergence and Lemma 1, let us present the following lemma, connecting
C-weak convergence of random variables with convergence in the L1 sense.
Lemma 8. (From C-Weak Convergence
to Convergence in L1 ) Consider the sequence
n o
A
of discrete time stochastic processes Xt
, as well as a “limit” process Xt ,t ∈ N, all being
A∈N
R, S , B (R) -valued and all defined on a common base space (Ω, F , P). Further, suppose that
nn o
o
are almost surely bounded in Z ≡ [a, b] (with
all members of the collection
XtA
, Xt
A∈N
−∞ < a < b < ∞) and that
t∈N
W
P AX A |X ( ·| Xt ) −→ δXt (·) ≡ 1(·) (Xt ) ,
t
that is, the sequence
it is true that
n
XtA
o
A∈N
∀t ∈ N,
A→∞
t
(95)
is marginally C-weakly convergent to Xt , given Xt , for all t. Then,
E
L
n
Xt − XtA
1
or, equivalently, XtA −→
Xt , for all t.
o
−→ 0,
A→∞
∀t ∈ N,
(96)
A→∞
Proof of Lemma 8. Let all the hypotheses of Lemma 8 hold true. Then, we know that, ∀t ∈ N,
lim P AX A |X ( A| Xt (ω)) = δXt (ω) (A) ,
n→∞
t
t
P − a.e.,
(97)
for all continuity Borel sets A ∈ S . Using the tower property, it is also true that
n
o
n n
oo
E Xt − XtA ≡ E E Xt − XtA σ {Xt } .
(98)
L
1
Therefore, in order to show that XtA −→
Xt for each t ∈ N, it suffices to show that
A→∞
E
n
o
a.s.
Xt − XtA σ {Xt } (ω) −→ 0,
∀t ∈ N.
A→∞
(99)
Then, the Dominated Convergence Theorem would nn
produce
result.
o the desired
o
A
are almost surely bounded
Of course, because all members of the collection
Xt
, Xt
A∈N
t∈N
nn
o o
in Z, all members of the collection
Xt − XtA
must be bounded almost surely in the
A∈N t∈N
compact set Zb , [0, 2δ] ⊂ R, where δ , max {|a| , |b|}.
Let us define the continuous and bounded function
b
x, if x ∈ Z
f (x) , 2δ, if x > 2δ .
0, if x < 0
(100)
Then, from Lemma 1 and using conditional probability measures it must be true that for each t ∈ N,
a version of the conditional expectation of interest is explicitly given by
ˆ
n
o
A
E f Xt − Xt
σ {Xt } (ω) ≡ f (|Xt (ω) − x|) P AX A |X (dx |Xt (ω) ) −→
t
24
t
A→∞
−→
A→∞
ˆ
f (|Xt (ω) − x|) δXt (ω) (dx) ≡ 0,
P − a.e.,
since, for each ω ∈ Ω, Xt (ω) is constant. Further, by definition of f ,
n
o
n
o
E f Xt − XtA σ {Xt } (ω) ≡ E Xt − XtA 1(|X −X A |) Zb σ {Xt } (ω)
t
t
n
o
A
≡ E Xt − Xt σ {Xt } (ω) , P − a.e.,
and for all t ∈ N, which means that
n
o
a.s.
E Xt − XtA σ {Xt } (ω) −→ 0,
A→∞
∀t ∈ N.
Calling dominated convergence proves the result.
(101)
(102)
(103)
Additionally, the following useful (to us) result is also true. The proof, being elementary, is
omitted.
Lemma 9. (Convergence of the Supremum) Pick any natural T < ∞. If, under any circumstances,
n
o
E Xt − XtA −→ 0, ∀t ∈ NT ,
(104)
A→∞
then
sup E
t∈NT
4.3
n
Xt − XtA
o
−→ 0.
A→∞
(105)
The Key Lemma
We are now ready to present our key lemma, which will play an important role in establishing our
main result (Theorem 3) later on. For proving this result, we make use of all the intermediate ones
presented in the previous subsections.
Lemma 10. (Convergence of the Likelihoods) Consider the stochastic process
X
1
yTi (Xi ) C−1
exp −
i (Xi ) yi (Xi )
2
i∈Nt
N
bt ,
Yp
, t , t ∈ N.
Λ
Dt
det (Ci (Xi ))
(106)
i∈Nt
b At , NAt /DAt , defined exactly the same way as Λ
b t , but replacing Xi with the
Consider also the process Λ
A
approximation Xi , ∀i ∈ Nt . Further, pick any natural T < ∞ and suppose either of the following:
n o
• For all t ∈ NT , the sequence XtA
is marginally C-weakly convergent to Xt , given Xt , that
A∈N
is,
W
(107)
P AX A |X ( ·| Xt ) −→ δXt (·) , ∀t ∈ NT .
t
t
A→∞
25
• For all t ∈ NT , the sequence
n
XtA
o
A∈N
is marginally convergent to Xt in probability, that is,
P
XtA −→ Xt ,
A→∞
∀t ∈ NT .
(108)
b T ⊆ Ω, such that
Then, there exists a measurable subset Ω
o
n
bt − Λ
b At Yt (ω) ≡ 0,
lim sup sup EPe Λ
A→∞ t∈N
T
(109)
bT
ω∈Ω
e
b T satisfy
where the P,P-measures
of Ω
n
o
exp (−CN )
b T , Pe Ω
bT
min P Ω
≥1−
,
(T + 1)CN −1
(110)
for any free but finite constant C ≥ 1.
Proof of Lemma 10. From Lemma 2, it is true that
bt −
Λ
b At
Λ
Nt − NAt
1
Nt NAt
1
+ |Nt |
≤
≡
−
− A
A
Dt DAt
D
Dt
t
Dt
≤
Nt − NAt
DAt
+
1
1
− A .
Dt Dt
(111)
We first concentrate on the determinant part (second term) of the RHS of (111). Directly invoking
Lemma 3, it will be true that
1
1
− A
Dt Dt
t
t
Y
Y
1
p
r
−
≡
det (Ci (Xi )) i=0
i=0
1
det Ci XiA
r
p
det
(C
(X
))−
det
Ci XiA
t
i−1
t
i
i
X
Y
Y
1
1
q
r
≤
r
A
det Cj Xj
i=0 j=0
j=i+1
det Cj Xj
det (Ci (Xi )) det Ci XiA
r
p
det
(C
(X
))−
det
Ci XiA
t
i−1
t
i
i
X Y
Y
1
1
v
v
v
=
uN
uN
uN
uY
j=i+1 u Y
i=0 j=0 u Y
t λ C X
t λ C X A t λ C X λ C X A
n
j
j
j
j
j
j
n
n
j
j
n
n=1
n=1
n=1
26
≤
t
X
1
N i/2 N (t−i)/2
i=0 2λinf λinf
t
X
1
≡ q
N (t+2)
2 λinf
i=0
det (Ci (Xi )) − det Ci XiA
λN
inf
det (Ci (Xi )) − det Ci XiA
.
(112)
From Lemma 5, we can bound the RHS of the above expression as
t
1
1
N KDET X
q
Ci (Xi ) − Ci XiA
− A ≤
N (t+2)
Dt Dt
2 λinf
i=0
t
X
NK
≡ q DET
Σi (Xi ) − Σi XiA
N (t+2)
2 λinf
i=0
F
F
.
(113)
And from Lemma 4, (113) becomes
t
1
1
N 2K
K X
Xi − XiA .
− A ≤ q DET Σ
N (t+2)
Dt Dt
2 λinf
i=0
(114)
We now turn our attention to the “difference of exponentials” part (first term) of the RHS of
(111). First, we know that
t
Y
i=0
t Y
N
Y
N (t+1)
λinf ≡ λinf
,
det Ci XiA
≥
(115)
i=0j=1
yielding the inequality
Nt − NAt
DAt
Nt − NAt
,
≤ q
N (t+1)
λinf
(116)
where λinf > 1 (see Assumption 1). Next, making use of the inequality [11]
|exp (α) − exp (β)| ≤ |α − β| (exp (α) + exp (β)) ,
(117)
∀ (α, β) ∈ R2 , the absolute difference on the numerator of (116) can be upper bounded as
Nt − NAt ≤
≤
t
1 X T
A
−1
A
A
A
T
X
C
X
N
+
N
X
yi (Xi ) C−1
(X
)
y
(X
)
−
y
y
i
i
t
t
i
i
i
i
i
i
i
i
2
i=0
t
X
i=0
A
−1
T
X
C
XiA yi XiA .
yTi (Xi ) C−1
(X
)
y
(X
)
−
y
i
i
i
i
i
i
i
(118)
Concentrating on each member of the series above in the last line of (118) and calling Lemma 2, it
is true that
h
i
h
i
A
−1
T
X
C
XiA yi XiA
yTi (Xi ) C−1
(X
)
y
(X
)
−
y
i
i
i
i
i
i
i
27
−1
+
≤ kyi (Xi )k2 C−1
y
(X
)
−
C
(X
)
XiA yi XiA
i
i
i
i
i
2
A
A
A
+ C−1
X
X
y
y
(X
)
−
y
i
i
i
i
i
i
i Xi
2
2
. (119)
Calling Lemma 2 again for the term multiplying the quantity kyi (Xi )k2 in the RHS of the above
expression, we arrive at the inequalities
A
−1
T
X
C
XiA yi XiA
yTi (Xi ) C−1
(X
)
y
(X
)
−
y
i
i
i
i
i
i
i
A
y
(X
)
−
y
(X
)
≤ kyi (Xi )k2 C−1
i
i
i
i
i Xi
2
2
−1
A
A
+ yi Xi
Ci
Xi
yi (Xi ) − yi XiA
2
2
2
−1
−1
A
Ci (Xi ) − Ci
+ kyi (Xi )k2 yi Xi
XiA
2
2
A
yi Xi
kyi (Xi )k2
2
yi (Xi ) − yi XiA
yi (Xi ) − yi XiA
≤
+
λinf
λinf
2
2
−1
+ kyi (Xi )k2 yi XiA
C−1
XiA
,
(120)
i (Xi ) − Ci
2
2
or, equivalently,
A
−1
T
X
C
XiA yi XiA
yTi (Xi ) C−1
(X
)
y
(X
)
−
y
i
i
i
i
i
i
i
kyi (Xi )k2 + yi XiA
2
≤
yi (Xi ) − yi XiA
+
λinf
2
−1
C−1
(X
)
−
C
XiA
+ kyi (Xi )k2 yi XiA
i
i
i
2
2
. (121)
Now, recalling Assumption 2, the definition of yi (Xi ) (resp. for XiA ) and invoking Lemma 6, it
must be true that
T
−1
A
yTi (Xi ) C−1
(X
)
y
(X
)
−
y
C
XiA yi XiA
X
i
i
i
i
i
i
i
kyi − µi (Xi )k2 + yi − µi XiA
2
≤
µi (Xi ) − µi XiA
λinf
2
−1
C−1
(X
)
−
C
XiA
+ kyi − µi (Xi )k2 yi − µi XiA
i
i
i
2
2
!
2
2 kyi k2 + µsup
(122)
Xi − XiA , Θ (yi ) Xi − XiA .
+ KIN V kyi k2 + µsup
≤ Kµ
λinf
Using the above inequality, the RHS of (116) can be further bounded from above as
Nt − NAt
DAt
sup Θ (yi ) t
X
≤ q
Xi − XiA .
N (t+1)
λinf
i=0
i∈Nt
28
(123)
Therefore, we can bound the RHS of (111) as
sup Θ (yi )
t
2
N KDET KΣ X
i∈Nt
bt − Λ
b At ≤
q
q
+
Xi − XiA .
Λ
N (t+1)
N (t+2)
λinf
2 λinf
i=0
(124)
Taking conditional expectations on both sides of (124), observing that the quantity supi∈Nt Θ (yi )
e (see Theorem 2),
constitutes a {Yt }-adapted process and recalling that under the base measure P
A
the processes yt and Xt (resp. Xt ) are statistically independent, we can write
)
( t
Θ
(y
)
sup
i
2
o
n
X
N
K
K
i∈Nt
A
A
bt − Λ
b t Yt ≤ q
EPe Λ
,
(125)
Xi − Xi
+ q DET Σ EPe
N (t+1)
N (t+2)
λinf
2 λinf
i=0
e − a.e., and, because P ≪H P,
e P − a.e. as well. From the last inequality, we can readily observe
P
t
that in order to be able to talk about any kind of uniform convergence regarding the RHS, it is
vital to ensure that the random variable supi∈Nt Θ (yi ) is bounded from above. However, because
the support of kyi k2 is infinite, it is impossible to bound supi∈Nt Θ (yi ) in the almost sure sense.
b τ ⊆ Ω with
Nevertheless, Lemma 7 immediately implies that there exists a measurable subset Ω
n
o
exp (−CN )
b τ , Pe Ω
bτ
(126)
min P Ω
≥1−
(τ + 1)CN −1
bτ ,
such that, for all ω ∈ Ω
sup kyi (ω)k22 ≡ sup kyi k22 < γCN (1 + log (1 + τ )) ,
(127)
i∈Nτ
i∈Nτ
for some fixed constant γ > 1, for any C ≥ 1 and for any fixed τ ∈ N. Choosing τ ≡ T < ∞, it is
true that
sup Θ (yi ) ≤ sup Θ (yi )
i∈Nt
i∈NT
#
2
2 kyi k2 + µsup
+ KIN V kyi k2 + µsup
≤ sup Kµ
λinf
i∈NT
2e
γ
2
+ KIN V e
γ CN (1 + log (1 + T ))
< Kµ
λinf
"
, Ko CN (1 + log (1 + T )) ,
where γ
e,
√
∀t ∈ NT ,
γ + µsup . Therefore, it will be true that
t
2
n
o
n
o
X
N
K
K
K
CN
(1
+
log
(1
+
T
))
bt − Λ
b At Yt ≤ o
q
EPe Xi − XiA ,
+ q DET Σ
EPe Λ
N (t+1)
N (t+2)
λinf
2 λinf
i=0
for all t ∈ NT , with probability at least
1−
exp (−CN )
(T + 1)CN −1
29
,
(128)
(129)
e Further,
under either P or P.
t
X
i=0
EPe
n
Xi − XiA
o
≤ (t + 1) sup EPe
τ ∈Nt
n
Xτ − XτA
o
.
(130)
Then, with the same probability of success,
n
o
bt − Λ
b At Yt
EPe Λ
2
n
o
K CN (1 + log (1 + T )) (t + 1) KDET KΣ N (t + 1)
q
q
sup EPe Xτ − XτA
+
≤ o
N (t+1)
N (t+2)
τ ∈Nt
λinf
2 λinf
!
2
o
n
K
CN
(1
+
log
(1
+
T
))
(T
+
1)
K
K
N
(T
+
1)
A
DET Σ
X
−
X
E
sup
≤ o
+
e
τ
τ
P
N/2
τ ∈Nt
2λN
λinf
inf
n
o
(131)
, KG (T ) sup EPe Xτ − XτA , ∀t ∈ NT ,
τ ∈Nt
where KG (T ) ≡ O (T log (T )). Alternatively, upper bounding the functions comprised by the
quantities t, N, λinf in the second and third lines of the expressions above as (note that, obviously,
t + 1 ≥ 1, ∀t ∈ R+ )
N (t + 1)
N (t + 1)2
q
≤ max q
N (t+1)
N (t+1)
t∈R+
λinf
λinf
respectively, we can also define
and
(132)
N 2 (t + 1)2
N 2 (t + 1)
q
≤ max q
,
N (t+2)
N (t+2)
t∈R+
λinf
λinf
−2/ log(λinf )
16Ko C (1 + log (1 + T )) λinf
KG (T ) ,
2
N log λinf
(133)
−2/ log(λinf )
8KDET KΣ λ−N
inf λinf
+
2
log λinf
,
(134)
where, in this case, KG (T ) ≡ O (log (T )). Note, however, that although its dependence on T is
logarithmic, KG (T ) may still be large due to the inability to compensate for the size of Ko . In any
bT ,
case, for all ω ∈ Ω
o
n
o
n
bt − Λ
b At Yt (ω) ≤ KG (T ) sup E e Xτ − XτA , ∀t ∈ NT ,
(135)
Ee Λ
P
τ ∈Nt
P
Therefore, we get
sup EPe
bT
ω∈Ω
n
n
o
o
bt − Λ
b At Yt (ω) ≤ KG (T ) sup E e Xτ − XτA ,
Λ
P
τ ∈Nt
∀t ∈ NT
and further taking the supremum over t ∈ NT on both sides, it must be true that
n
o
n
o
bt − Λ
b At Yt (ω) ≤ KG (T ) sup sup E e Xτ − XτA
sup sup EPe Λ
P
t∈NT τ ∈Nt
t∈NT ω∈Ω
bT
30
(136)
≡ KG (T ) sup EPe
t∈NT
n
Xt − XtA
o
.
(137)
Finally, if either
W
P AX A |X ( ·| Xt ) −→ δXt (·) ≡ 1(·) (Xt ) ,
t
or
A→∞
t
∀t ∈ NT ,
(138)
P
XtA −→ Xt , ∀t ∈ NT
(139)
A→∞
n o
and given that since the members of XtA
are almost surely bounded in Z, the aforementioned
A∈N
sequence is also uniformly integrable for all t ∈ N, it must be true that (see Lemma 8)
n
o
EPe Xt − XtA −→ 0, ∀t ∈ NT .
A→∞
Then, Lemma 9 implies that
sup EPe
t∈NT
n
Xt − XtA
o
−→ 0,
(141)
A→∞
which in turn implies the existence of the limit on the LHS of (137). QED.
4.4
(140)
Finishing the Proof of Theorem 3
Considering the absolute difference of the RHSs of (44) and (43), it is true that (see Lemma 2)
n
o
n
o
o
n
A bA
A A
b
X
Λ
E
Y
X
Λ
E
Y
Y
X
Λ
E
e
e
e
t t
t
t t
t
t
t t
EPe { Xt Λt | Yt }
P
P
P
o ≡
o −
o ,
n
n
n
(142)
−
EPe { Λt | Yt }
b Y
bA Y
ΛA Y
Λ
Λ
E
E
E
e
P
t
e
P
t
t
e
P
t
t
t
due to the fact that the increasing stochastic process
Y
i∈Nt
exp
1
kyi k22
2
≡ exp
1X
2
i∈Nt
kyi k22
is {Yt }-adapted. Then, we can write
n
o
n
o
b At Yt
b t Yt
EPe XtA Λ
EPe Xt Λ
n
n
o −
o
b t Yt
b At Yt
EPe Λ
EPe Λ
n
o
n
o
n
o
n
o
b t Yt
b At Yt E e Λ
b At Yt − E e XtA Λ
b t Yt E e Λ
EPe Xt Λ
P
P
P
n
n
o
o
≡
b At Yt
b t Yt E e Λ
EPe Λ
P
n
o
n
o
bt − Λ
b At Yt
b t − XtA Λ
b At Yt
|EP { Xt | Yt }| EPe Λ
EPe Xt Λ
n
n
o
o
≤
+
,
b At Yt
b At Yt
EPe Λ
EPe Λ
(143)
e P − a.e..
P,
(144)
Let us first focus on the difference on the numerator of the second ratio of the RHS of (144).
Recalling that δ ≡ max {|a| , |b|}, we can then write
n
o
n
o
b t − XtA Λ
b At Yt ≤ E e Xt Λ
b t − XtA Λ
b At Yt
EPe Xt Λ
P
31
o
b At Xt − XtA Yt
bt − Λ
b At + Λ
|Xt | Λ
n
o
n
o
bt − Λ
b At Yt + E e Xt − XtA Yt
≤ δEPe Λ
P
n
o
n
o
bt − Λ
b At Yt + E e Xt − XtA ,
≡ δEPe Λ
P
≤ EPe
n
On the other hand, for the denominator for (144), it is true that
X T A −1 A A
1
exp
−
X
C
X
X
y
y
i
i
i
i
i
i
2
n
o
i∈Nt
A
b
r
EPe Λt Yt ≡ EPe
Yt
Y
A
det
C
X
i
i
i∈Nt
X T A −1 A A
1
Xi yi Xi Yt
yi Xi Ci
EPe exp −
2
i∈Nt
q
≥
(t+1)
λN
sup
X
2
1
EPe exp −
kyi k2 + µsup Yt
2λinf
i∈Nt
q
≥
(t+1)
λN
sup
X
1
2
kyi k2 + µsup
exp −
2λinf
i∈Nt
e P − a.e.,
q
, P,
≡
N (t+1)
λsup
since the process
X
i∈Nt
kyi k2 + µsup
2
(146)
is {Yt }-adapted. Now, from Lemma 7, we know that
sup kyi k22 ≤ sup kyi k22 < γCN (1 + log (T + 1)) ,
i∈Nt
(145)
i∈NT
∀t ∈ NT ,
(147)
where the last inequality holds with probability at least 1 − (T + 1)1−CN exp (−CN ), under both
e for any finite constant C ≥ 1. Therefore, it can be trivially shown that
base measures P and P,
p
2
γCN (1 + log (T + 1)) + µsup (T + 1)
exp −
2λinf
n
o
b At Yt ≥
q
EPe Λ
> 0, ∀t ∈ NT ,
(148)
(t+1)
λN
sup
implying that
inf
n
o
b At Yt (ω) > 0,
inf inf EPe Λ
t∈NT ω∈Ω
b T A∈N
32
(149)
b T coincides with the event
where Ω
(
ω∈Ω
sup kyi k22
i∈Nt
< γCN (1 + log (T + 1)) , ∀t ∈ NT
)
e
b T follows
with P,P-measure
at least 1 − (T + 1)1−CN exp (−CN ). Of course, the existence of Ω
from Lemma 7. Putting it altogether, (144) becomes (recall that the base measures P and Pe are
equivalent)
n
o
o
n
bt − Λ
b At Yt + E e Xt − XtA
2δEPe Λ
P
n
o
, P, Pe − a.e.,
(150)
EP { Xt | Yt } − E A ( Xt | Yt ) ≤
A
b
inf EPe Λt Yt
A∈N
b T and t ∈ NT on both
where δ ≡ max {|a| , |b|}. Taking the supremum both with respect to ω ∈ Ω
sides, we get
sup sup EP { Xt | Yt } (ω) − E A ( Xt | Yt ) (ω)
t∈NT ω∈Ω
b
T
sup EPe
≤
t∈NT
n
Xt − XtA
o
+ 2δ sup sup EPe
t∈NT ω∈Ω
b
inf inf inf EPe
t∈NT ω∈Ω
b A∈N
T
with
n
n
o
bt − Λ
b At Yt (ω)
Λ
T
o
b
ΛAt Yt (ω)
n
o
exp (−CN )
b T , Pe Ω
bT
min P Ω
≥1−
.
(T + 1)CN −1
,
(151)
(152)
Finally, calling Lemma 10 and Lemma 9, and since the denominators of the fractions appearing in
(151) are nonzero, its RHS tends to zero as A → ∞, under the respective hypotheses. Consequently,
the LHS will also converge, therefore completing the proof of the Theorem 3.
5
Conclusion
In this paper, we have provided sufficient conditions for convergence of approximate, asymptotically
optimal nonlinear filtering operators, for a general class of hidden stochastic processes, observed in
a conditionally Gaussian noisy environment. In particular, employing a common change of measure
argument, we have shown that using the same measurements, but replacing the “true” state by an
approximation process, which converges to the former either in probability or in the C-weak sense,
one can define an approximate filtering operator, which converges to the optimal filter compactly in
time and uniformly in an event occurring with probability nearly 1, at the same time constituting
a purely quantitative justification of Egoroff’s theorem for the problem of interest. The results
presented in this paper essentially provide a framework for analyzing the convergence properties of
various classes of approximate nonlinear filters (either recursive or nonrecursive), such as existing
grid based approaches, which are known to perform well in various applications.
33
Appendix: Proof of Theorem 1
The proof is astonishingly simple. Let the hypotheses of the statement of Theorem 1 hold true. To
avoid useless notational congestion, let us also make the identifications
Xt , {Xi }i∈Nt
and
Yt , {Yi }i∈Nt .
(153)
Now, by definition of the conditional expectation operator and since we have assumed the existence
of densities, it is true that
bt ≡ EP { Xt | Y0 , Y1 , . . . , Yt }
X
ˆ
= xt f Xt |Yt ( xt | Yt ) dxt
ˆ
xt f(Xt ,Yt ) (xt , Yt ) dxt
=
fYt (Yt )
ˆ
≡ ˆ
ˆ
≡ ˆ
where
xt ft (x0 , x1 , . . . xt , Yt )
t
Y
dxi
i=0
t
Y
ft (x0 , x1 , . . . xt , Yt ) dxi
i=0
t
Y
dxi
xt λt fet {xi }i∈Nt , Yt
i=0
t
Y
λt fet {xi }i∈Nt , Yt
dxi
,
(154)
i=0
ft (x0 , x1 , . . . xt , Y0 (ω) , Y1 (ω) , . . . , Yt (ω))
fet (x0 , x1 , . . . xt , Y0 (ω) , Y1 (ω) , . . . , Yt (ω))
≡ λt {xi }i∈Nt , Yt (ω) ∈ R+ , ∀ω ∈ Ω,
λt ,
(155)
constitutes a “half ordinary function - half random variable” likelihood ratio and where the condition (16) ensures its boundedness. Of course, although the likelihood ratio can be indeterminate
when both densities are zero, the respective points do not contribute in the computation of the
relevant integrals presented above, because these belong to measurable sets corresponding to events
of measure zero. From (154) and by definition of the conditional density of {Xi }i∈Nt given {Yi }i∈Nt ,
we immediately get
bt ≡ EP { Xt | Yt }
X
ˆ
≡ ˆ
t
Y
xt λt feXt |Yt {xi }i∈Nt Yt
dxi
i=0
λt feXt |Yt {xi }i∈Nt Yt
34
t
Y
i=0
dxi
EPe { Xt Λt | Y0 , Y1 , . . . , Yt }
EPe { Λt | Y0 , Y1 , . . . , Yt }
E e { Xt Λt | Yt }
≡ P
,
EPe { Λt | Yt }
=
(156)
which constitutes what we were initially set to show.
References
[1] D. S. Kalogerias and A. P. Petropulu, “Asymptotically optimal discrete time nonlinear filters
from stochastically convergent state process approximations,” IEEE Transactions on Signal
Processing, vol. 63, no. 13, pp. 3522 – 3536, July 2015.
[2] A. Segall, “Stochastic processes in estimation theory,” Information Theory, IEEE Transactions
on, vol. 22, no. 3, pp. 275–286, May 1976.
[3] R. J. Elliott, L. Aggoun, and J. B. Moore, Hidden Markov Models.
Springer, 1994.
[4] A. Segall, “Recursive estimation from discrete-time point processes,” Information Theory,
IEEE Transactions on, vol. 22, no. 4, pp. 422–431, Jul 1976.
[5] S. Marcus, “Optimal nonlinear estimation for a class of discrete-time stochastic systems,”
Automatic Control, IEEE Transactions on, vol. 24, no. 2, pp. 297–302, Apr 1979.
[6] R. J. Elliott, “Exact adaptive filters for markov chains observed in gaussian noise,” Automatica,
vol. 30, no. 9, pp. 1399–1408, 1994.
[7] R. J. Elliott and H. Yang, “How to count and guess well: Discrete adaptive filters,” Applied
Mathematics and Optimization, vol. 30, no. 1, pp. 51–78, 1994.
[8] H. J. Kushner, “Approximations to optimal nonlinear filters,” Automatic Control, IEEE Transactions on, vol. 12, no. 5, pp. 546–556, 1967.
[9] K. Ito and K. Xiong, “Gaussian filters for nonlinear filtering problems,” Automatic Control,
IEEE Transactions on, vol. 45, no. 5, pp. 910–927, May 2000.
[10] H. J. Kushner and P. Dupuis, Numerical methods for stochastic control problems in continuous
time. Springer, 2001, vol. 24.
[11] H.
J.
Kushner,
“Numerical
aproximation
to
optimal
nonlinear
filters,”
http:// www.dam.brown.edu/ lcds/ publications/ documents/ Kusher Pub001.pdf , 2008.
[12] L. F. Richardson, Measure and integration: a concise introduction to real analysis. John Wiley
& Sons, 2009.
[13] D. S. Kalogerias and A. P. Petropulu, “Sequential channel state tracking & spatiotemporal channel prediction in mobile wireless sensor networks,” Available at:
http:// arxiv.org/ pdf/ 1502.01780v1.pdf , 2015.
35
[14] ——, “Grid based filtering of Markov processes revisited: Recursive estimation & asymptotic
optimality,” IEEE Transactions on Signal Processing, to appear, 2015. Extended version available at: http://eceweb1.rutgers.edu/∼ dkalogerias/.
[15] D. S. Kalogerias, N. Chatzipanagiotis, M. M. Zavlanos, and A. P. Petropulu, “Mobile jammers for secrecy rate maximization in cooperative networks,” in Acoustics, Speech and Signal
Processing (ICASSP), 2013 IEEE International Conference on, May 2013, pp. 2901–2905.
[16] D. S. Kalogerias and A. P. Petropulu, “Mobi-cliques for improving ergodic secrecy in fading wiretap channels under power constraints,” in Acoustics, Speech and Signal Processing
(ICASSP), 2014 IEEE International Conference on, May 2014, pp. 1578–1591.
[17] N. Chatzipanagiotis, Y. Liu, A. Petropulu, and M. Zavlanos, “Controlling groups of mobile
beamformers,” in Decision and Control (CDC), 2012 IEEE 51st Annual Conference on, Dec
2012, pp. 1984–1989.
[18] G. Pagès, H. Pham et al., “Optimal quantization methods for nonlinear filtering with discretetime observations,” Bernoulli, vol. 11, no. 5, pp. 893–932, 2005.
[19] M. S. Arulampalam, S. Maskell, N. Gordon, and T. Clapp, “A tutorial on particle filters for
online nonlinear/non-gaussian bayesian tracking,” Signal Processing, IEEE Transactions on,
vol. 50, no. 2, pp. 174–188, 2002.
[20] F. Daum, “Nonlinear filters: Beyond the Kalman Filter,” IEEE Aerospace and Electronic
Systems Magazine, vol. 20, no. 8, pp. 57 – 69, Aug. 2005.
[21] A. Papoulis and S. U. Pillai, Probability, Random Variables and Stochastic Processes, 4th ed.
McGraw-Hill, 2002.
[22] R. Elliott and V. Krishnamurthy, “New finite-dimensional filters for parameter estimation of
discrete-time linear gaussian models,” Automatic Control, IEEE Transactions on, vol. 44, no. 5,
pp. 938–951, May 1999.
[23] R. B. Ash and C. Doleans-Dade, Probability and measure theory.
[24] P. Billingsley, Convergence of Probability Measures.
Academic Press, 2000.
John Wiley & Sons, 2009.
[25] P. Berti, L. Pratelli, and P. Rigo, “Almost sure weak convergence of random probability measures,” Stochastics: An International Journal of Probability and Stochastic Processes, vol. 78,
no. 2, pp. 91 – 97, April 2006.
[26] R. Grubel and Z. Kabluchko, “A functional central limit theorem for branching
random walks, almost sure weak convergence, and applications to random trees,”
http:// arxiv.org/ pdf/ 1410.0469.pdf , 2014.
[27] P. Rebeschini and R. van Handel, “Can local particle filters beat the curse of dimensionality?”
http://arxiv.org/abs/1301.6585, 2013.
[28] Y. Sharifi, “Eigenvalues of adjugate matrix,” http:// ysharifi.wordpress.com/ 2011/ 01/ 22/ eigenvalues-of-adjug
January 2011.
36
[29] A. M. Mathai and S. B. Provost, Quadratic forms in random variables: theory and applications.
M. Dekker New York, 1992.
[30] B. Laurent and P. Massart, “Adaptive estimation of a quadratic functional by model selection,”
Annals of Statistics, vol. 28, no. 5, pp. 1302 – 1338, 2000.
37
| 3cs.SY
|
Experimental Astronomyo - DOI: 10.1007/s10686-011-9241-6 manuscript No.
(will be inserted by the editor)
Using Java for distributed computing in the Gaia satellite
data processing
arXiv:1108.0355v1 [cs.CE] 1 Aug 2011
William O’Mullane1 , Xavier Luri2 , Paul
Parsons3 , Uwe Lammers1 , John Hoar1 , Jose
Hernandez1
the date of receipt and acceptance should be inserted later
Abstract In recent years Java has matured to a stable easy-to-use language with the
flexibility of an interpreter (for reflection etc.) but the performance and type checking of
a compiled language. When we started using Java for astronomical applications around
1999 they were the first of their kind in astronomy. Now a great deal of astronomy
software is written in Java as are many business applications.
We discuss the current environment and trends concerning the language and present
an actual example of scientific use of Java for high-performance distributed computing:
ESA’s mission Gaia. The Gaia scanning satellite will perform a galactic census of
about 1000 million objects in our galaxy. The Gaia community has chosen to write its
processing software in Java. We explore the manifold reasons for choosing Java for this
large science collaboration.
Gaia processing is numerically complex but highly distributable, some parts being
embarrassingly parallel. We describe the Gaia processing architecture and its realisation
in Java. We delve into the astrometric solution which is the most advanced and most
complex part of the processing. The Gaia simulator is also written in Java and is the
most mature code in the system. This has been successfully running since about 2005
on the supercomputer ”Marenostrum” in Barcelona. We relate experiences of using
Java on a large shared machine.
Finally we discuss Java, including some of its problems, for scientific computing.
PACS PACS 07.05.Kf · PACS 95.10.Jk · PACS 07.87.+v
1 Introduction
1.1 The Gaia Mission
The Gaia satellite is destined for the Lagrange point L2 early in 2012 after launch
summer 2012 from French Guiana aboard a Soyuz Fregat Rocket. Gaia is the European
Space Agency’s (ESA) sixth cornerstone mission. Its goal to make a phase space map of
1 European
Space Astronomy Centre
d’astronomia i meteorologia, Universitat of Barcelona, ICC-UB/IEEC
3 The Server Labs, Madrid, Spain
2 Departament
2
our galaxy. Spinning around its own axis in a Lissajous orbit around L2 for five years
gaia will continually scan the sky observing more than one thousand million (109 )
celestial sources, on average eighty times each. The scientific goals of the experiment
are manifold and covered in detail in (Perryman et al, 2001), the data will help key
research in the composition and formation of our galaxy.
Gaia contains two astrometric telescopes at a fixed angle of 106.5◦ as well as a
radial velocity spectrograph and two photometers. The astrometric design allows true
parallaxes (distances to stars) to be obtained (Makarov, 1998) after careful data processing. A least-squares fitting scheme named the Astrometric Global Iterative Solution
(AGIS) (Lindegren et al, 2008) is currently foreseen to perform the astrometric data
reduction. Equally involved processing is required for photometry and spectroscopy.
Gaia processing software is being written by the Gaia Data Processing and Analysis
Consortium (DPAC). DPAC is a pan-European federation of institutes comprising of
over four hundred astronomers, physicists and programmers. The consortium is led by
eleven members of the community who form the DPAC Executive (DPACE). DPAC not
only has the responsibility to write the processing software but also to run and maintain
it until the final Gaia catalogue is produced around the year 2020. The software will be
run in six data processing centres - each responsible for a different facet of the overall
processing. An overview of DPAC and the processing is provided in (O’Mullane et al,
2007).
1.2 Computing
Simulations of Gaia data have been in production at the University of Barcelona (UB)
since about 1998. Some of these simulations require considerable computing power
and are discussed in Section 3. Initial AGIS experiments have been conducted using
simulation data from UB in the past few years. The AGIS efforts are discussed in
Section 2. By nature the data processing must be distributed. If one considers that there
are ≈ 1012 (low resolution cutout) images downlinked from Gaia, at one millisecond
per image that is over 30 years of processor time. Massive distribution is the only
possibility to deal with this data. At this point the Gaia community has quite some
experience with large software and computing efforts, which is not unusual in the
science community. What some consider quite unusual is that all Gaia software is
written in the Java language. We discuss the choice of Java in Section 4.
2 AGIS - a complex computing problem
The Astrometric Global Iterative Solution (AGIS) provides the rigid framework for all
of the Gaia measurements. Gaia is spinning freely and making observations which relate
only to other observations made by Gaia. It is an absolute instrument - there is no list of
input stars such as there was for Hipparcos (ESA, 1997), the predecessor of Gaia. What
this means is that the Gaia data must be reduced in a self-consistent manner such that
all individual observations of celestial sources, the model of each source’s position and
motion, Gaia’s own attitude, orbit and velocity must be in harmony. Later the entire
system will be aligned with the International Celestial Reference System (ICRS).
The complete mathematical formulation for AGIS is presented in (Lindegren et al,
2008) while the computational framework is described in (O’Mullane et al, 2008). Here
3
we recap these aspects briefly in section 2.1 before discussing results (section 2.2) and
performance/manpower trade offs (section 2.3).
2.1 AGIS overview
AGIS is a block-iterative solution for the Gaia astrometry. It consists of at least four
blocks which may be calculated independently. Each block is however dependent on
the result of the other blocks. We term one step through all blocks an outer iteration,
although the term outer is frequently dropped. This is simplistically formulated in the
following equations :
S = A+G+C
(1)
A=S+G+C
(2)
G = S+A+C
(3)
C = S+A+G
(4)
which are discussed further below.
The vast majority of Gaia sources (Equation 1) may be described by a six parameter
astrometric model. These parameters and their derivation are fairly standard in modern
astrometry (Murray, 1983) and describe the position and motion of the source in three
dimensions. The parameters are :
α
δ
π
µα∗
µδ
µr
azimuthal angle of object a.k.a. right ascension
angular distance from the celestial equator (north or south) a.k.a.
declination
annual parallax, the apparent shift in angular position when
viewed from opposite sides of the sun
µα∗ = µα cos(δ) proper motion in α direction
proper motion in δ direction
radial velocity, motion in the line of sight direction
From the ≈ 80 observations of each source made by Gaia a least-squares fitting of
the observations to the model may be made. The fit must include several intricacies
namely:
–
–
–
–
the orientation of Gaia in space (or attitude, Equation 2)
the path of light through the instrument (or calibration, Equation 4)
global parameters such as relativistic numbers (Section 3)
light bending according to general relativity (see (Klioner, 2003))
Current best estimates for attitude, calibration and global parameters are used for
any given fit of the source. This fit must be performed on all sources but to make the
problem tractable a subset of sources may be treated first. Around 108 sources are
needed to make the global reference frame but it could be as many as 5 · 108 . Once
this solution has converged the source calculation (using the final attitude etc.) may
be performed on all remaining sources. This is an important efficiency improvement.
A direct solution for a large number of sources has been shown to be intractable
(Bombrun et al, 2010).
4
The application of the source equation (known as source update) requires the gathering of all observation of the source and the current attitude, calibration and global
parameters. The application of the relativistic model requires precise ephemerides of
the planets and of the satellite. Our current approach is to group observations together
on disk for easy loading, since there is no cross talk between the source equations i.e. all
sources may be processed in parallel if we wish. The other data required is sufficiently
small (order of a few hundred MB) that it may be loaded once in memory and used
to process multiple sources, typically batches of several thousand sources per job on a
processor.
Attitude and calibration on the other hand require all observations in given time
periods. We do not need to hold on to the observations but may add their contributions
to partial equations. Hence as each source is updated the updated parameters and the
observations are used to update the partial solutions for the attitude and calibration.
The system has been run with up to 100 computing nodes and 1400 Java threads. We
constantly encounter bottlenecks usually in the calibration and attitude processes for
the obvious reason that they need to see all sources. The processing of several thousand
sources in a job makes the communication more efficient - we send fewer large messages
rather than millions of small ones.
2.2 Results
AGIS has been running at ESAC since early 2006 with increasing numbers of sources
and observations. This is discussed briefly in (O’Mullane et al, 2008) from which we
reproduce table 1 below. Since the entire system is designed for AGIS and we have few
interconnects between processes we did not find it necessary to use any HPC library nor
GRID software. Effectively jobs are taken from a Whiteboard by numerous DataTrains
which load the appropriate data and process it writing the results back to a database.
The Whiteboard may simply be considered as a table in a database where jobs are
written an updated. DataTrain is a term coined within DPAC for a process which runs
through a set of data passing it to a series of algorithms. It provides an abstraction
layer between algorithm and data access.
The machine used for the Gaia tests so far has been purchased in installments by
ESA. The initial machine for the first tests in 2005 consisted of twelve nodes each
with 6 GB of RAM and two processors (3.6GHz Xeon EM64T). An EMC storage area
network (SAN) with 5 TB of disk is attached to the nodes using fibre optic cable and
the nodes are also connected via Gigabit Ethernet cards in a local area network. AGIS
has also been tested, and shown to run as fast, using a cheap Rack Server Network
Attached Storage device of 6 TB. This cluster was upgraded to 18 nodes in 2006 and
a further 4 nodes but with quad core processors were added in 2007. The quad core
processors functioned very well for us providing the performance of 4 processors and
in table 1 the number of processors counts each core as a processor.
This is not very special machine, it consists of standard Dell power edge blades
with standard Intel Xeon processors running Red Hat Linux. The blades are housed
in a standard Dell rack. There is no special HPC software used, a network of normal
Linux machines could work in the same manner. As such this machine perhaps does
not merit the term HPC but we will move toward 10 Teraflops in the coming years.
Various different types of astronomical test have been performed to show AGIS
produces the correct solution for Gaia. For this paper we are mainly interested in per-
5
Table 1 AGIS run times decrease as more processors are added. Note that the data volume
increased from 2005 to 2006 from 18 months to 5 years, the processor power also increased
but the run time went up. This was dramatically improved in 2007. In 2010 we have implemented a new math frame work around Conjugate gradient - this provides a better solution
at some cost. The normalised column shows throughput per processor in the system (total
observations/processors/hours) e.g. an indication of the real performance.
Date
2005
2006
2007
2008
2009
2010
Observations
1.6x107 src
8x107
8x107
8x107
2.6x108
4x109
Procs
12
36
24
31
50
68
Time
3h
5h
3h
1h
1.8h
9.5
Normalised
0.9x106 obs/hour
0.5x106 obs/hour
1.3x106 obs/hour
3.2x106 obs/hour
2.8x106 obs/hour
6.2x106 obs/hour
formance however, which is best presented by table 1. We see that AGIS has been slow
in 2007 but finally it was made much faster in 2008 with a prospect of more improvement to come. It should be noted that more functionality has also been introduced each
year - a more complete source model, a more complex calibration scheme etc. but the
efficiency has been increased. Some speedups are due to profiling while others are due
to mathematical methods. Between the Lund group and the ESAC group finally the
problem is being tackled from both math and computing side with competence. Early
work on AGIS was not so fruitful as it treated the problem as something to just be
run on a computer rather than a problem for which a system needed to be designed to
make it efficient. In that system math and computing techniques have been questioned
and varied to arrive at the efficient implementation we have today.
2.3 Performance vs Manpower
We know we could write some parts of AGIS to run perhaps two or more times faster
than Java. We could get a little more performance by utilising special machines and
rewriting our code again for example toward a cell processor hybrid such as the Roadrunner. Speed of code is indeed an issue for Gaia in general but manpower is perhaps
a bigger issue - we have a great deal of processing software to write. The estimated
manpower in man years for the Gaia processing is around 2000. Granted this includes
operation but development alone pre launch is estimated at 1000 man years. In the case
of the astrometric solution we have good records. The initial working AGIS in ESAC
was done in about 180 man days so just over half a man year. It was clear to us that
this was only possible in Java - in C or C++ this would have taken far longer. This
claim is of course difficult to quantify, as a programming collective we have decades
of experience with C and C++ projects and it is our opinion that coding in Java is
cleaner and faster. Even in Java the problem is difficult. The previous solution which
AGIS replaced was worked on for four years by various people; to be modest lets say
only 8 man years went into it, but it was more like 12. To date on AGIS we have
spent about 15 man years and we estimate to finish AGIS in Java we need another 10
man years with probably more after launch. There is also post launch maintenance to
consider - we feel Java maintenance will be lower cost than say C, and we have eight
years of operations to consider - we have 27 man years in the plan for this.
6
Of course it is difficult to calculate but we would need far more manpower to
manage code in C and more still if we were to customise for a particular computer
system (especially as it would become obsolete during our operations). Again just for
AGIS we are talking of over 50 man years of effort which in today’s monetary terms
is around 5 million euro. Even 20% more effort to code in another language would
be a considerable sum. We think we can buy the hardware we need to run AGIS for
around a million euro. It has been observed on other space programs than manpower
typically ends up costing more than machines and apparently we are no different. Our
intention is also to buy the most common and cheap processors - these are the the chips
Java is usually running fastest on as well because of their ubiquity. Hence a special
more super computer oriented machine would probably cost more. Finally there is
power consumption if we could make AGIS 4 times faster we could save 25% on power.
Energy costs may well go up but probably not faster than manpower costs since the
saving on energy requires spending on manpower to customise code for our ”hard to
code” problem these will cancel each other out.
3 Gaia Simulations - use of shared computing resource
The development of the Gaia simulator has been an integral part of the Gaia data
processing. This software tool is designed to provide realistic simulations of the Gaia
observations, closely mimicking the format and content of the data that Gaia will send
to ground. The main purpose of this simulated data has been the feeding of the data
processing chain in order to test and validate it, although it has also been used for
other purposes, for instance to evaluate some satellite design options and to prepare
the mission scientific case.
As mentioned above, the simulator has been running at the University of Barcelona
since 1998. Its initial development was closely tied to the first attempts to develop a
viable Global Iterative Solution but now it has become a mature tool able to simulate
a wide variety of celestial objects, physical and instrumental effects and data formats
for the multiple data processing modules developed by the Gaia DPAC.
In this section we will first review our experience with the development in Java,
the advantages and drawbacks we found, and then discuss a specific (and extreme)
example of Java versus C performance found during this development.
3.1 Java, a new kid on the block
When the development of the simulator was started the first choice to be made was the
programming language to use. The team undertaking this task was at the time (1998)
mainly (and almost exclusively) composed of scientists, with few software engineering
expertise. The obvious choice, based on the programming experience of the team, would
have been FORTRAN with perhaps C as a second, but somewhat frowned upon, choice.
Furthermore, astronomical and numerical libraries were available in FORTRAN, and
in most cases also in C, but not in Java.
However, Java was finally chosen for two main reasons:
1. The ongoing Global Iterative Solution development contract was going to be implemented in Java by requirement of ESA. Given that the two projects were closely
tied at the time using a common language was natural.
7
2. The advice of professional software engineers from ESA and some members of the
team with software engineering background pointed towards the use of an Object
Oriented language, strongly advising against FORTRAN for a large collaborative
project. This left the choice between C++ and Java.
At the end Java was chosen, but not without frequent second thoughts in the coming
years. The widening scientific community that was becoming involved in the project
was reticent for some time to adopt the new language, specially given the accumulated
experience in FORTRAN and C and the lack of libraries in Java.
However, eleven years later the landscape has much changed, and now Java has been
even more widely (almost completely) adopted as a viable programming language for
scientific programming in the Gaia simulator community. Furthermore, the project has
become large and complex, involving development teams distributed around Europe.
The management of the project has at this stage adopted many professional software
development tools (UML, Hudson, PMD, Checkstyle, etc.) that would not have been
available had a language like FORTRAN been chosen, making the coding much more
robust and reliable. Also, like the team developing AGIS and based on previous experiences, they feel that the development has been quicker and more seamless than it
would have been using C++. The motives for this feeling are varied, but one of the
big reasons is that thanks to the garbage collection implemented in the JVMs memory
leaks are less of an issue. Since the simulator is quite intensive in memory usage, not
having to worry about this problem is seen as a great advantage.
In short, the initial reticences on the use of Java for scientific programming in Gaia
have vanished, to the point that one of the managers of the simulator development has
changed the teaching of programming for first year physics students from FORTRAN
to Java.
3.2 The blessing of portability
One of the advantages of Java not mentioned in the previous section is its portability.
As said, the simulator has been running for eleven years and during this time it has
been ported to several machines, and in the process Java has shown that its portability
is real and practical.
Due to its intensive computation needs the Gaia simulator has mostly been run at
supercomputing centers, using significant amounts of computation time It was initially
run at the Centre de Supercomputació de Catalunya (CESCA) were several mediumsized clusters were used: an IBM SP2, a Compaq HPC320, an HP RX2600 and an SGI
ALTIX 3700. In all cases the migration from one cluster to the next was seamless, the
only serious complication being the adaptation of the execution to the different queue
systems in each machine. Later on, in 2005, the simulator was moved to the Barcelona
Supercomputing Center (BSC) where the Mare Nostrum supercomputer had been just
installed. Again, the migration was quite simple, with the major complication being
how to adapt the execution to the new distributed environment and queue system.
Today the simulator is still running there, having cruised through a Mare Nostrum
upgrade in 2006 without needing any significant change and having consumed a some
millions of CPU hours and generated many terabytes of simulated data.
In these years the portability of Java has therefore been a major advantage, saving
time and resources that otherwise would have been devoted to adapting the code to the
8
new environment. Furthermore, the portability of Java has also allowed the running
of the simulator on other environments like local clusters for testing and development
and a cluster at CNES for tailored small simulations through a web page manager.
3.3 Performance in scientific computation
In the initial years of the project the performance in numerical computing of Java
was discussed many times. Not having previous experience on its use for scientific
computation there was some worry that the language could prove to be too slow at
some point of the development. However, some initial tests with a set of some often-used
numerical algorithms showed that Java was not much slower that C in solving these
problems, thanks to the Just In Time (JIT) compilation in the Java virtual machines.
In the next years the JIT virtual machines steadily improved, and the first working
code was developed in the project, showing in practice that the implementation in Java
was not producing any of the feared bottlenecks. Nowadays, Java has become part of the
landscape in scientific programming in Gaia, and although worries about performance
are still occasionally an element of resistance, given the accumulated experience and
clear savings in development time performance is no longer an issue.
Furthermore, the development model in DPAC is based on six-month development
cycles; at the end of each cycle an improved version is produced and tested, leading to
further requirements, corrections and improvements for the next cycle. In this framework the optimisation of the system is not as important as the maintainability and
flexibility of the code to allow such a quick production. The robustness and clarity
of Java helps a lot in such a process, and as mentioned before has probably saved a
significant amount of money in programming resources.
A real example encountered during the simulator will illustrate this point. A key
piece of the Gaia astrometry is the calculation of the so-called “relativistic corrections”,
the changes on the apparent position of the objects in the sky deriving from relativistic
effects: aberration, gravitational light bending, etc. This is a complex calculation taking
into account the ephemeris of the major solar system bodies and requiring (for a µas
accuracy) to reach the limit of the numerical precision of variables of the “double” type
(64 bit floating point).
An initial (legacy) implementation of these calculations was available from S.
Klioner in the form of C code and was used in the simulator until 2008 through JNI
calls. At that time, in order to avoid the inconveniences of mixing two languages in
the simulator, the same author developed for DPAC a new implementation coded in
Java. Both implementations were thoroughly compared and results matched at subµas level. However, during the testing it was found that the computation times differed
substantially between the two versions, the Java version being between 4 and 10 times
faster than the C version! Possible external sources of this difference (like overheads
of the JNI calls) were ruled out and it was concluded that the difference was actually
intrinsic to the code.
It is possible, even likely, that an optimisation of the C code would make it much
more efficient, to at least the level of the Java code. However, this example clearly
shows how the same developer did a quicker and better job in Java (a language that,
unlike C, he was unfamiliar with). On the other hand, the difference possibly comes
also from the excellent refactoring afforded by the new JITs virtual machines that
9
now automatically makes many of the performance fixes which previously had to be
manually implemented - in C of course all optimisations must be done by the coder.
Finally, the optimization work on the Java code has continued. The team is now
exploring the increasing possibilities of the options for aggressive optimization and
garbage collection tailoring available in some Java virtual machines (specially the IBM
one, used in the MareNostrum supercomputer) that will possibly lead to further improvements in the performance of the simulator.
4 Java, Gaia and commercial uses
When (O’Mullane et al, 2000) was presented at the Astronomical Data Analysis and
Software Systems (ADASS) conference in 1999 most people in that community did not
know what Java was and this was the only paper mentioning Java. Two years later
about half of the astronomy related projects at ADASS involved Java. Some project
must take the first step, in this case Gaia and Planck were experimenting with Java.
These experiments possibly made it easier for the Integral SOC and the entire Herschel
Science Ground Segment to be written in Java (Wieprecht et al, 2004). Gaia and Java
go back many years, the initial prototype for the Global Iterative Solution was already
in Java in 1999 (O’Mullane and Lindegren, 1999).
Java in the late 90s was not the same as it is today but already then it was seen to
have potential for science development. For a project like Gaia we were faced with a
life span of over 20 years for our software and an entrenched Fortran community. It was
clear already back then, as discussed earlier, that Gaia processing software would not
be written in Fortran but needed to be in a more modern, preferably OO, language,
C++ being the obvious choice. The transition for some scientists from Fortran to C++
was seen as difficult or impossible. Java simply worked more easily and was chosen and
eventually accepted by all.
4.1 Negative Aspects
We do not claim Java is the silver bullet (F.P., 1987) and we have several problems
which we live with especially in numerical coding. As already mentioned especially
when we started, there were no good numerical libraries. The situation improved from
about 2000 onwards when Java started to become really usable for scientific computing in terms of robustness and performance. A number of promising numerical library
developments took place (e.g. JAMA - A Java Matrix Package) but until around 2005
most of them had stopped, leaving the distributions and code available but largely
unmaintained. A positive exception is the Apache Common Math library which is still
in active development and in extensive use in almost all of Gaia’s data processing software. Early versions of Apache Common Math were missing linear algebra functionality
needed by us but that has improved with time.
In parallel with diminishing efforts for developing numerical libraries we observed
a general decline in support for Numerics in Java. For instance the Numerics Working
Group of the Java Grande Forum (1 ), initially a driving force behind many positive developments around Java Numerics, has effectively ceased to exist. This is a bit worrying
1
http://math.nist.gov/javanumerics/
10
along with a general perceived lack of support for Java in the traditional conservative
supercomputing scene that in terms of languages remains to be dominated by Fortran
and C until today.
In 1998 W. Kahan, one of the key persons behind the IEEE 754 Standard for Binary Floating-Point Arithmetic, delivered a keynote “How Java’s Floating-Point Hurts
Everyone Everywhere” Kahan and Darcy (1998) at a Java HPC workshop. In this contribution Java gets harshly criticized for for a number of IEEE non-compliances which
could and should have been avoided by the language designers. In subsequent years Sun
was repeatedly asked to rectify the known deficiencies but chose to ignore all complaints
and even today those issues raised more than 10 years ago are, to our knowledge, still
present in all existing Java implementations. With Java now under Oracle’s control it
seems unlikely that the situation will change in the foreseeable future. Fortunately, in
our view most of the points are fairly subtle in nature and are unlikely to show up as
perceivable flaws in “ordinary” numerical application code. We can confirm this for all
of Gaia’s software and in particular for the AGIS system described here in more detail
(Sect. 2) with one exception: Java does not provide any means to use what IEEE 754
calls ’trap-handlers’ for capturing and dealing with numerical exceptions like division
by zero. Every arithmetic operation in Java always delivers a valid result and that
can make the debugging of numerical code extremely difficult and time consuming. As
an example, there was a coding error in the AGIS attitude update that caused the
introduction of an NaN value into a large matrix of equations which then propagated
and spread through the matrix before eventually leading to NaN in an end result much
later. Without the option to have an exception raised at the first occurrence of an NaN
condition the only way to find the problem is to check explicitly every intermediate
result for NaN (Double.isNaN(x)). For obvious reasons this is not a viable option in a
large numerical code.
4.2 Features
One of the things which made java work well for us was its built in features. We
were especially happy to have multi threading and distributed programming such as
RMI (Remote Method Invocation - allows calling a method on an object on another
machine) built in to the language.
4.3 Rise of Java in industry
At the same time (back in the late 90’s) we saw in industry a surge of Java programming
as it pushed from the the Web language clearly into the back office. Enterprise Java
Beans were starting to appear behind web pages accessing databases and wrapping
legacy applications to make them network available. Java was no longer an ”applet”
language for making more interactive web pages - it was handling credit cards and
transactions and doing serious work. The fact that Java is 100% portable to all O/Ss
where the JVM has been ported to and backward compatible means that using libraries
is no longer a painful issue.
Companies such as IBM, Oracle, BEA, Sun and many open source vendors have
created Java application servers that support the Java standards in JavaEE produced
by Sun.
11
Since then frameworks to further aid productivity in the development of Java software such as Spring and Hibernate have appeared on the scene and these have helped
with the adoption of Java as the technology of choice for developing software. In addition the fact that there are thousands of open source libraries available to use in Java
projects has also helped the spread of Java.
4.4 Maintainability, Robustness and Performance
When developing software there are a number of key issues that have to be addressed
including but not limited to Maintainability, Robustness, Scalability
Addressing maintainability first, Java compared to other languages such as C++
and Fortran offers a number of advantages. The defacto editor for Java, Eclipse, provides a number of features that makes creating and maintaining Java code easier;
Graphical syntax highlighting, Refactoring wizards to improve the design of the application; Plugins to produce documentation. Java was also one of the first languages
to have a really good implementation of the xUnit testing frameworks with the widely
adopted JUnit. With Java 1.5 and the introduction of Generics and Annotations and
with techniques such as Aspect Oriented Programming (AOP - not currently used on
Gaia) we have the possibility of developing software with less code and the less code
there is, the more maintainable it is.
Java is more robust than C++, primarily because the C++ component of all Java
programs, the JVM itself, is the same for all Java programs and therefore tried and
tested millions of times.
Many people believe that Java is not as scalable as other languages because of
the overhead of the Java bytecode interpretation. However the Java Hotspot server
compiler, especially the one in Java 1.6, is incredibly efficient at optimising code that
is called very frequently.
4.5 Portability and the Cloud
Although there are minor problems with Java portability this is usually in Graphical
User Interfaces rather than in the type of code forming the majority of the Gaia
environment. As stated in Section 3 our simulation code has been ported over many
years to many platforms with little effort. Currently development is done on MacOs,
Linux and windows systems without issue. Such portability lead us to consider using
the Amazon E2C for the Astrometric solution (Section 2). We, unlike CERN Shiers
(2009) are not yet tied in to one architecture. Although these days even CERN use
some cloud resource.
With about 2 person weeks of effort part of our team (Parsons and Olias) got AGIS
running on the cloud. Some problems were encountered with the database configuration, although at least Amazon already hand a VM with Oracle on it. Also a scalability
problem in our own code was found and remedied - prior to this we had no opportunity
to run 1400 threads. It worked well, at least as well as our cheap in house cluster. Hence
we agree with the detractors of the Cloud - it is not a supercomputer with super fast
interconnect, but then one is not paying super prices either.
In fact although our intention is/was to build a cheap cluster (around 1MEuro) for
AGIS we estimate all AGIS mission processing for 100 Million Stars could be done for
12
about 400KEuro. When energy is factored in this makes Amazon look very attractive.
It largely depends on idle time however - and we in any case would not go without an in
house cluster for testing and development. It also appears we now may need to process
500 Million sources to counteract possible spatial error correlations (as presented by
Berry Holl in Heidelberg in 2009). This would bring the saving close to zero - but one
must question if the likes of Amazon can do a cheaper job of maintaining a cluster.
The availability of a much larger number of nodes than we can buy is very interesting both for testing and for production. By using Amazon we could perhaps do
our processing faster by using more nodes and still have it cost less than an in house
machine. We shall continue to experiment and evaluate this option - in any case the
final machine at ESAC for the processing will not be purchased until we are a couple
of years into the mission around 2015.
4.6 Virtualisation - non uniqueness
The real power of Java comes from the notion of running on a virtual machine. This
should not be, but often is, confused with interpretation such as done in LISP and
Smalltalk. This virtue is not shared by many new languages or schemes perhaps the
most all encompassing of which is the Microsoft .NET framework in which many languages are compiled to the same virtual machine. We are not trying to say Java is the
only language with this feature. In fact with virtualisation suddenly any particular machine/OS/langugae configuration can be equally portable with its own virtual machine.
Suddenly putting those legacy apps in a cloud may not be such a hard decision.
4.7 Future of Java
The rise of Java in industry and positive experience we had we Java compared to C++
reinforced our choice of Java as the language for Gaia. This choice was reaffirmed within
the Gaia community in 2006 (Hernandez et al, 2006). Indeed at the time of writing all
science development missions at ESAC are using Java as their programming language.
We predict that Java will be the language of choice for the foreseeable future
because of all the advantages outlined earlier, although we believe that within the Java
Virtual Machine, we will see more use of dynamic languages such as Ruby and Groovy
for areas of the application that will need to be changed very frequently.
5 Conclusion
One may argue about the definition of High Performance Computing but within the
Gaia project since the late 90s we have certainly be doing numerically intensive computing and handling increasingly large amounts of data in a highly distributed manner.
All of this is done using Java. We argue that this is a good option for long running
projects where portability and maintenance are possibly more important than squeezing the last FLOP from a specific processor.
The portability of Java has served us well in the last decade allowing code to
move from SPARC station to Linux boxes, WinTel and even the Marenostrum super
computer. Of late we have also leveraged Amazon’s E2C resources.
13
We are not alone in the astronomy world, many other projects are using Java or
other high level languages. The ESA missions Herschel, Planck and many archives are
Java users. Others such as Spitzer and JWST mentioned using Java at least in part.
Other higher level languages such as C# and Python are also in use.
14
A Acronyms used in this paper
The following table has been generated from the on-line Gaia acronym list:
Acronym
ADASS
AGIS
BSC
CERN
CESCA
CNES
CPU
DPAC
DPACE
ESA
ESAC
FLOP
GB
HPC
ICRS
IEEE
JIT
JNI
JVM
JWST
MB
OO
PMD
RAM
RMI
SAN
SOC
TB
UB
UML
VM
Description
Astronomical Data Analysis Software and Systems
Astrometric Global Iterative Solution
Barcelona Supercomputing Centre
Centre Europénne pour la Recherche Nucléaire
CEntre de Supercomputacı́o de CAtalunya
Centre National d’Etudes Spatiales (France)
Central Processing Unit
Data Processing and Analysis Consortium
Data Processing and Analysis Consortium Executive
European Space Agency
European Space Astronomy Centre (VilSpa)
FLoating-point OPeration
GigaByte
High-Performance Computing
International Celestial Reference System
Institute of Electrical and Electronic Engineers
Just In Time Compiler
Java Native Interface
Java Virtual Machine
James Webb Space Telescope (formerly known as NGST)
MegaByte
Object Oriented
Software tool to detect software problems
Random Access Memory
Remote Method Invocation
Storage Area Network
Science Operations Centre
TeraByte
University of Barcelona (Spain)
Unified Modeling Language
Virtual Machine
15
B Bibliography
References
Bombrun A, Lindegren L, Holl B, Jordan S (2010) Complexity of the Gaia astrometric leastsquares problem and the (non-)feasibility of a direct solution method. aap 516:A77, DOI
10.1051/0004-6361/200913503OPEN:
ESA (1997) The Hipparcos and Tycho Catalogues. ESA, eSA SP-1200
FP JB (1987) No silver bullet essence and accidents of software engineering. Computer 20:10–
19, DOI http://doi.ieeecomputersociety.org/10.1109/MC.1987.1663532
Hernandez J, O’Mullane W, Huc C (2006) Language recommendation for the
DPAC. Tech. rep., European Space Astronomy Centre, Villafranca, Madrid, URL
http://www.rssd.esa.int/llink/livelink/open/532588
Kahan W, Darcy J (1998) How java’s floating-point hurts everyone everywhere. Tech. rep.,
Elect. Eng. & Computer Science, University of California at Berkely, keynote presented at
ACM 1998 Workshop on Java for High-Performance Network Computing held at Standford
University, http://www.cs.berkeley.edu/~ wkahan/JAVAhurt.pdf
Klioner SA (2003) A Practical Relativistic Model for Microarcsecond Astrometry in Space.
Astrophysical Journal 125:1580–1597
Lindegren L, O’Mullane W, Lammers U, Hobbs D (2008) The astrometric core solution for
the gaia mission. in prepreation for Astronomy and Astrophysics
Makarov VV (1998) Absolute measurements of trigonometric parallaxes with astrometric satellites. Astronomy and Astrophysics 340:309–314
Murray CA (1983) Vectorial astrometry. Bristol: Adam Hilger, 1983
O’Mullane W, Lindegren L (1999) An Object-Oriented Framework for GAIA Data Processing.
Baltic Astronomy 8:57–72
O’Mullane W, Hazell A, Bennett K, Bartelmann M, Vuerli C (2000) ESA Survey Missions and
Global Processing. In: Manset N, Veillet C, Crabtree D (eds) Astronomical Data Analysis
Software and Systems IX, Astronomical Society of the Pacific Conference Series, vol 216,
pp 419–422
O’Mullane W, Lammers U, Bailer-Jones C, Bastian U, Brown AGA, Drimmel R, Eyer L, Huc
C, Katz D, Lindegren L, Pourbaix D, Luri X, Torra J, Mignard F, van Leeuwen F (2007)
Gaia Data Processing Architecture. In: Shaw RA, Hill F, Bell DJ (eds) Astronomical Data
Analysis Software and Systems XVI, Astronomical Society of the Pacific Conference Series,
vol 376, pp 99–108
O’Mullane W, Lindegren L, Lammers U, Hobbs D (2008) Implementation of the gaia astrometric global iterative solution in java. in prepreation for Experimental Astronomy
Perryman MAC, de Boer KS, Gilmore G, Høg E, Lattanzi MG, Lindegren L, Luri X,
Mignard F, Pace O, de Zeeuw PT (2001) GAIA: Composition, formation and evolution of
the Galaxy. Astronomy and Astrophysics 369:339–363, DOI 10.1051/0004-6361:20010085,
arXiv:astro-ph/0101235
Shiers J (2009) Grid today, clouds on the horizon. Computer Physics Communications
180(4):559
–
563,
DOI
DOI:10.1016/j.cpc.2008.11.027,
URL
http://www.sciencedirect.com/science/article/B6TJ5-4V70NH3-1/2/2fdc77415de62c2e360754c969911850,
special issue based on the Conference on Computational Physics 2008 - CCP 2008
Wieprecht E, Brumfit J, Bakker J, de Candussio N, Guest S, Huygen R, de Jonge A, Matthieu
JJ, Osterhage S, Ott S, Siddiqui H, Vandenbussche B, de Meester W, Wetzstein M, Wiezorrek E, Zaal P (2004) The HERSCHEL/PACS Common Software System as Data Reduction
System. In: Ochsenbein F, Allen MG, Egret D (eds) Astronomical Data Analysis Software
and Systems (ADASS) XIII, Astronomical Society of the Pacific Conference Series, vol 314,
pp 376–379
| 5cs.CE
|
Finite element model based on refined plate theories for
laminated glass units
Alena Zemanová1 , Jan Zeman1,2 , and Michal Šejnoha1
1
arXiv:1410.8674v1 [cs.CE] 31 Oct 2014
Department of Mechanics, Faculty of Civil Engineering, Czech Technical University in
Prague, Thákurova 7, 166 29 Prague 6, Czech Republic
2
Centre of Excellence IT4Innovations, VŠB-TU Ostrava, 17. listopadu 15/2172 708 33
Ostrava-Poruba, Czech Republic
Abstract
Laminated glass units exhibit complex response as a result of different mechanical behavior and properties of glass and polymer foil. We aim to develop a finite element model for
elastic laminated glass plates based on the refined plate theory by Mau. For a geometrically
nonlinear description of the behavior of units, each layer behaves according to the ReissnerMindlin kinematics, complemented with membrane effects and the von Kármán assumptions.
Nodal Lagrange multipliers enforce the compatibility of independent layers in this approach.
We have derived the discretized model by the energy-minimization arguments, assuming that
the unknown fields are approximated by bi-linear functions at the element level, and solved
the resulting system by the Newton method with consistent linearization. We have demonstrated through verification and validation examples that the proposed formulation is reliable
and accurately reproduces the behavior of laminated glass units. This study represents a
first step to the development of a comprehensive, mechanics-based model for laminated glass
systems that is suitable for implementation in common engineering finite element solvers.
Keywords laminated glass; finite element method; refined plate theory; Lagrange multipliers;
Reissner-Mindlin plate theory; von Kármán assumptions
1
Introduction
Laminated glass has been developed to increase safety and security of fragile thin glass sheets.
When the load is transferred in both in-plane directions, the unit cannot be considered as a beam
and a more accurate description is needed. Thus, the goal of this paper is an extension of our
earlier finite element model of laminated glass beams (Zemanová et al., 2014) towards laminated
glass plates. For this reason, the overview of available literature below will focus on laminated
glass plates, whereas the discussion related to beams can be found in Zemanová et al. (2014).
Let us summarize that the laminated glass units behave in an unusual manner due to several
reasons. First, the polymer foil has the elastic modulus which is by orders of magnitude smaller
than the elastic modulus of the glass. Such heterogeneity in elastic properties of individual layers
complicate the modeling of laminated glass units. Second, the polymer foil is highly viscoelastic
1
with a great temperature dependency. Third, geometric nonlinearity should be included in the
laminated glass behavior, which is caused by the small thickness of plates compared to the other
dimensions.
A number of experimental measurements, e.g. (Vallabhan et al., 1992; Behr et al., 1993; Bennison et al., 1999), studied influence of temperature and load duration on a structural behavior
of a laminated glass unit since the polymer foil is a time/temperature-dependent material, whose
performance determines the coupling along glass layers. The basic method to approximate the
mechanical response of laminated glass units involves the utilization of one of the limiting cases:
layered case as an assembly of independent glass layers and monolithic case with thickness equal
to the combined thickness of glass layers and interlayers1 determined under the assumption of
geometric linearity.2 In Behr et al. (1985), the recommendation for the choice of the limiting
case can be found, which takes into account only the temperature dependence, whereas the
extension of the previous experimental stress analysis by load duration influence is summarized
in Behr et al. (1993). Several approaches to the more precise analytical or numerical analysis of
laminated glass structures have been proposed in order to ensure a reliable and efficient design.
Single-layer approaches, useful especially in the preliminary phases of the design procedure,
approximate the behavior of unit by an equivalent homogeneous plate and can be found in a
draft version of European standard for glass CEN/TC 129 (2013a). Benninson et al. (2008)
extended the concept of effective thickness to the case of laminated glass beams and determined
the thickness of equivalent monolithic glass unit as a function of load duration and temperature
history. Galuppi and Royer-Carfagni (2012) proposed new relations for the deflection- and stresseffective thicknesses of geometrically linear laminated glass plates using variational principles and
approximation of shape function for deformation of laminated glass plate. Practical expressions
for the design of laminated glass was summarized in Galuppi et al. (2012). The effects of
geometric nonlinearity are not considered in this approach although they can be significant for
laminated glass units with a small value of the shear modulus of polymer foil.
Analytical methods provide closed-form solutions only for laminated beams with very specific boundary conditions and when excluding the effects of geometric nonlinearity; for plates
only series-based solutions are available. Foraboschi (2012) described a mechanical behavior of
laminated glass plates by a system of differential equations derived under the Kirchhoff assumptions for two thin glass plates with the same thickness. He considered only shear stresses and
strains in the polymer interlayer and solve the resulting system by the Fourier series expansion
for the load and displacements. His approach is valid for a simply supported rectangular laminated glass plate subjected to lateral uniformly distributed static loading. An extension of this
model to other sandwich plates with stiff interlayer can be found in Foraboschi (2013). These
analytical approaches are rather difficult to be generalized to units with multiple layers and are
mainly useful for the verification of numerical models.
We find the numerical analysis of the discretized formulation of the problem as the most
advantageous for the modeling of nonlinear behavior. The most common tool for structural
analysis is the finite element method, nevertheless a few approaches based on finite difference
1
This formulation corresponds to the true monolithic model, whereas the equivalent monolithic model has the
thickness equal only to the total thickness of glass layers. For the upper bound it is more correct to assume the
true monolithic model.
2
Note that the interval between the monolithic and layered limit can be reduced by including the effects of
geometric nonlinearity, see Vallabhan et al. (1987).
2
method can be found in literature, e.g. Vallabhan et al. (1993) and Aşık (2003). The first
possibility for the simulation of behavior of laminated glass units utilizes three-dimensional
solid elements for individual layers and their interaction, cf. (Duser et al., 1999). This approach
requires a large number of solid finite elements and therefore leads to expensive computations,
since the thickness of the laminates, especially the polymer foil, is small compared to the other
dimensions. To avoid fully resolved three-dimensional simulations, we used a refined plate theory
in our approach and considered the layer-wise linear variation of in-plane displacements in the
thickness direction. Most of the commercial finite element codes do not have such a feature, but
rely on classical laminate formulations. In this framework, the laminated glass is modeled as a
homogeneous unit with effective stiffness determined from its layout and standard plate (shell)
elements are employed in the analysis. Because of a great mismatch of the material properties
of PVB and glass, this approach is not appropriate as it is too coarse to correctly represent the
inter-layer interactions.
2
Model assumptions
The proposed numerical model for laminated glass plates is inspired by a specific class of refined
plate theories (Mau, 1973; Šejnoha, 1996; Kruis et al., 2002). In this framework, each layer is
treated as a shear deformable plate with independent kinematics. Interaction between individual
layers is captured by the Lagrange multipliers (with a physical meaning of nodal forces), which
result from the fact that no slippage is supposed to occur between glass plate and polymer foil.
As for constitutive modeling, glass typically responds in a linear elastic manner, whereas
the polymer foil behaves in a viscoelastic manner. However, in many cases even the foil is
modeled as being linear elastic by means of an effective shear modulus related to temperature
and duration of loading. The hypothesis that the viscoelastic material is approximated by the
linear elastic one, assuming a known temperature and duration of loading, may be constraining.
On the other hand, a viscoelastic model of interlayer is more difficult to manage, and will be
discussed independently and in more details in forthcoming publications.
As for the effects of geometric nonlinearity, we have reported that they should not be neglected when modeling laminated glass structures, and also confirmed this statement by numerical examples in Zemanová et al. (2014), based on the fully nonlinear Reissner beam formulation.
Before extending the model to the laminated plates, we compared first the response of the fully
nonlinear formulation with a simpler model based on the von Kármán assumptions, in which
only deflections are large whereas the in-plane displacements and rotations remain small. Since
the response of these two models was found to be very similar, see Zemanová (2014)[Section
7.1], our strategy will be to accommodate the nonlinear effects due to vertical deflections only
in membrane strains.
The following nomenclature is used in the text. Scalar quantities are denoted by lightface
letters, e.g. a, and the bold letters are reserved for matrices, e.g. a or A. AT standardly stands
for the matrix transpose and A−1 for the matrix inverse. The subscript in parentheses, e.g. a(i) ,
is used to emphasize that the variable a is associated with the i-th layer.
3
3
Model formulation
3.1
Kinematics
Being inspired by Pica et al. (1980), each layer is modeled as the Reissner-Mindlin plate with
nonlinear membrane effects, by adopting the following kinematic assumptions:
• a straight line segment, normal to the underformed mid-surface, remains straight but not
necessarily normal to the deformed mid-surface,
• vertical displacements do not vary along the thickness of the plate.
glass
PVB
glass
y
x
x
(1)
y
φx
(1)
φy
(1 )
z
x ,u
1
(1)
z ,w
2
(2)
z ,w
(3)
h
(1)
(2)
y ,v
z
φy
( 2)
x ,u
1
h
2
z ,w
(3)
φy
(3 )
x ,u
(3)
3
(2)
y ,v
(2)
y ,v
(3)
2
h
(3 )
(1)
φx
1
z ,w
(2 )
( 2)
3
z ,w
(1)
(3)
φx
3
z ,w
Figure 1: Kinematics of laminated plate assuming the layer-wise linear variation of in-plane
displacements in the thickness direction
Thus, the nonzero displacement components in each layer u(i) , v (i) , and w(i) are parametrized
as, see Figure 1,
(i)
(i)
u(i) (x, y, z (i) ) = u0 (x, y) + ϕ(i)
y (x, y)z ,
(i)
(i)
(i)
(i)
v (x, y, z ) =
w (x, y, z ) =
(i)
v0 (x, y) −
(i)
w0 (x, y),
(i)
ϕ(i)
x (x, y)z ,
(1a)
(1b)
(1c)
where i = 1, 2, 3 refers to individual layers, u0 , v0 , and w0 denote the displacements of the
mid-surface, ϕx and ϕy represent the straight line segment rotations about the corresponding
axes, and z (i) is measured in the local coordinate system relatively to the mid-plane of the i-th
layer. For the layer surfaces to be connected, the corresponding displacements must satisfy the
compatibility conditions
h(i)
h(i+1)
) − u(i+1) (x, y, −
) = 0,
2
2
h(i+1)
h(i)
v (i) (x, y,
) − v (i+1) (x, y, −
) = 0,
2
2
h(i)
h(i+1)
w(i) (x, y,
) − w(i+1) (x, y, −
) = 0,
2
2
u(i) (x, y,
with i = 1, 2.
4
(2a)
(2b)
(2c)
The nonzero components of the Green-Lagrange strain tensor, e.g. (Jirásek and Bažant, 2002,
Section 24.1), at the i-th layer can be divided into the membrane effects
h
i
(i)
(i)
(i)
(i)
(i)
(i)
(i) T
ε(i)
(3)
m (x, y, z ) = εx (x, y, z ), εy (x, y, z ), γxy (x, y, z )
and the transverse shear strains
h
i
(i)
(i)
γ (i) (x, y) = γxz
(x, y), γyz
(x, y) T .
(4)
Under the von Kármán assumptions (Timoshenko and Woinowsky-Krieger, 1959), the derivatives
of u(i) and v (i) with respect to all coordinates are small and these parts of strain tensor become
(i)
(i)
(i)
(i)
(i)
ε(i)
+ εm0,N L (x, y)
m (x, y, z ) = ∂u0 (x, y) + ∂Sϕ (x, y)z
(i)
(i)
= εm0 (x, y) + κ(i) (x, y)z (i) + εm0,N L (x, y),
γ (i) (x, y) =
(i)
∇w0 (x, y)
+ Sϕ(i) (x, y),
(5a)
(5b)
and are expressed in terms of the mid-plane displacements, rotations, and pseudo-curvatures
as
h
i
(i)
(i)
(i)
u0 (x, y) = u0 (x, y), v0 (x, y) T ,
(6a)
h
i
(i)
T
ϕ(i) (x, y) = ϕ(i)
,
(6b)
x (x, y), ϕy (x, y)
h
i
(i)
(i)
T
κ(i) (x, y) = κ(i)
.
(6c)
x (x, y), κy (x, y), κxy (x, y)
Observe that the shear strain components and pseudo-curvatures are the same as in the geometrically linear case, but the membrane strain components include nonlinear contributions
depending on the deflections of the layer mid-surface
!2
!2
(i)
(i)
(i)
(i)
1 ∂w0
∂w0
∂w
1 ∂w0
(i)
(x, y) ,
(x, y) ,
(x, y) 0 (x, y) T .
(7)
εm0,N L (x, y) =
2
∂x
2
∂y
∂x
∂y
The matrix S and the differential operators ∂ and ∇ are provided by
∂
0
∂x
∂
∂
0 1
0
∂x
∂=
∇ = ∂ ,
S=
.
,
∂y
−1 0
∂
∂
∂y
∂y ∂x
3.2
(8)
Constitutive relations and specific internal forces
Having parametrized the strains with quantities defined on the mid-surface, we proceed to
stresses. We assume that the normal stress in the z direction is negligible compared to the other
stresses. It follows from the generalized Hooke law for isotropic bodies that the membrane stress
components and the shear stress components are given by, cf. (Pica et al., 1980),
(i)
(i) (i)
(i)
σ (i)
m (x, y, z ) = D m εm (x, y, z ),
τ (i) (x, y) = G(i) Iγ (i) (x, y),
5
(9)
where
D (i)
m
0
0
,
(i)
1−ν
2
1 ν (i)
(i)
ν (i) 1
E
=
(i)
2
1 − (ν )
0
0
(10)
is the membrane stiffness matrix, E (i) and G(i) are Young’s modulus of elasticity and shear
modulus of the i-th layer, and I is the identity matrix. Notice that, since the Reissner-Mindlin
kinematics were utilized, the transverse shear stresses are approximated as constant in the z (i)
direction.
The specific internal forces energetically conjugate to the generalized strains (5), are obtained
(i)
by suitable weighted stress averages in the thickness direction. As for the membrane strains εm ,
these are associated with the specific normal forces n(i) in the form
(i)
n (x, y) =
(i)
(i)
T
[n(i)
x (x, y), ny (x, y), nxy (x, y)]
Z
=
h(i)
2
(i)
− h2
(i)
σm
(x, y, z (i) ) dz (i)
(i)
(i)
(i)
(i)
(i)
= h(i) D (i)
ε
(x,
y)
+
ε
(x,
y)
=
D
∂u
(x,
y)
+
ε
(x,
y)
.
m
n
m0
0
m0,N L
m0,N L
(11)
The quantity corresponding to the pseudo-curvature κ(i) is the matrix of specific bending moments m(i) provided by
(i)
m (x, y) =
=
(i)
(i)
T
[m(i)
x (x, y), my (x, y), mxy (x, y)]
Z
=
h(i)
2
(i)
− h2
(i)
σm
(x, y, z (i) )z (i) dz (i)
(h(i) )3 (i) (i)
(i)
D m κ (x, y) = D b ∂Sϕ(i) (x, y).
12
(12)
The specific shear forces q (i) , conjugate to the shear strains γ (i) , are obtained in an analogous
way
(i)
q (x, y) =
[qx(i) (x, y), qy(i) (x, y)]T
= Mk
(i)
Z
h(i)
2
(i)
− h2
τ (i) (x, y) dz (i)
(i)
= k (i) G(i) Ih(i) γ (i) (x, y) = D s(i) ∇w0 (x, y) + Sϕ(i) (x, y) ,
(13)
but need to be corrected due to the assumed constant shear stresses by a constant k (1) = k (3) =
(i)
(i)
5/6 for top and bottom layers and k (2) = 1 for interlayer.3 In Eqs. (11)–(13), D n , D b , and
(i)
D s denote the matrices of normal, bending, and shear stiffness of the i-th layer, respectively.
Thus in the expressions for the specific internal forces, the geometric nonlinearity due to von
Kármán assumptions appears in the normal forces only, and the specific moments and shear
forces remain in the same form as in the geometrically linear case.
3
We assumed the constant value of 5/6 for glass layers, which corresponds to a parabolic distribution in the
thickness direction with almost-zero values on top and bottom surfaces. However interlayer is supposed to exhibit
a state of pure shear with constant distribution and therefore we set the value to 1.
6
3.3
Energy functional
The governing equations of the discretized model are derived by the energy-minimization arguments. To this purpose, we express the internal energy of the i-th layer in the form
Z
1
(i)
(i)
(i)
(i)
(i)
(i)
(i)
(i)
(i)
(εm0 + εm0,N L )T D (i)
(14)
Πint εm0 , εm0,N L , γ , κ
=
n (εm0 + εm0,N L )
2 Ω
(i)
(i)
+(γ (i) )T D (i)
+ (κ(i) )T D b κ(i) dΩ,
s γ
where Ω(i) denotes the mid-surface of the i-th layer. The potential energy of external loading
acting at the i-th layer reads
Z
(i)
(i)
(i)
(i)
(i)
(i) (i)
(i)
Πext u0 , w0 , ϕ
= −
(u0 )T f m + w0 f z + (ϕ(i) )T m(i) dΩ,
(15)
Ω
(i)
(i)
with f m , f z , and m(i) referring to the intensity of the distributed in-plane, out-of-plane and
moment loads. The total energy of the i-th layer is expressed as the sum of internal and external
components
(i)
(i)
(i)
(i)
(i)
(i)
(i)
Π(i) u0 , w0 , ϕ(i) = Πint ∂u0 + εm0,N L (w0 ), ∇w0 + Sϕ(i) , ∂Sϕ(i)
(i)
(i)
(i)
+ Πext u0 , w0 , ϕ(i) ,
(16)
in order to assemble the total energy of the system from the contributions of individual layers
Π
(1)
(1)
(2)
(2)
(3)
(3)
u0 , w0 , ϕ(1) , u0 , w0 , ϕ(2) , u0 , w0 , ϕ(3)
=
3
X
Π
(i)
(i)
(i)
u0 , w0 , ϕ(i)
.
(17)
i=1
3
(i)
(i)
(i)
u0 , w0 , ϕ
i=1
The true fields
then follow from the minimization of functional Π, subject
to the prescribed boundary conditions and the compatibility conditions (2).
4
4.1
Numerical implementation
Discretization
A standard discretization is employed, in which we approximate the unknown fields by bi-linear
functions at the element level, see Appendix A for explicit expressions. As a result of this step,
the internal and external energies are approximated as
Ne
X
(i)
(i)
(i)
(i)
Πint εm0 , εm0,N L , γ (i) , κ(i) ≈
Πint,e r (i)
,
e
(18)
e=1
(i)
(i)
(i)
Πext u0 , w0 , ϕ(i) ≈
Ne
X
e=1
Ne
X
(i)
T (i)
Πext,e r (i)
=
−
r (i)
f ext,e .
e
e
(19)
e=1
(i)
where r e is the vector of generalized nodal displacements of the e-th element and the i-th layer
h
i
h
i
(i)
(i)
(i)
(i) T
(i)
(i)
(i)
(i)
(i)
T
,
(20)
r (i)
,
r (i)
e = r e,1 , r e,2 , r e,3 , r e,4
e,n = ue,n , ve,n , we,n , ϕx,e,n , ϕy,e,n
(i)
and f ext,e stores the generalized nodal forces resulting from the distributed loading in (15), see
again Appendix A for more details.
7
4.2
Resulting system of governing equations
In order to obtain the nodal degrees of freedom corresponding to external loading, we rely again
on variational arguments, and start with the discretized energy functional
Π(r) =
Ne
3 X
X
(i)
(i)
(i)
Πint,e (r (i)
e ) + Πext,e (r e )
(21)
i=1 e=1
to be minimized under the kinematic constraints and compatibility conditions. This produces
the Lagrangian function in the form
L(r, λ) = Π(r) + λT Cr,
(22)
where r stores arbitrary kinematically admissible generalized nodal displacements, and λ is the
vector of admissible Lagrange multipliers (with physical meaning of forces between layers due to
compatibility). The block of matrix C, implementing the tying conditions between two adjacent
layers at the j-th node, now assumes the form (i = 1, 2)
h(i+1)
h(i)
...
−1
0
0
0
1 0 0
0
2
2
(i)
(i+1)
(i,i+1)
(23)
= 0 1 0 − h2
Cj
0 ... 0 −1 0 − h 2
0 ,
0 0 1
0
0 ... 0
0 −1
0
0
so that the condition Cr = 0 enforces the compatibility equations (2) discretely at all interfacial
nodes. Notice that, due to the assumption that the rotations ϕ remain small, the second term
in (22) is linear in r, which significantly simplifies the derivation as well as implementation of
the numerical model, when compared to the formulation presented in Zemanová et al. (2014).
The Karush-Kuhn-Tucker optimality conditions, e.g. (Bonnans et al., 2003, Chapter 14),
associated with (22) read
∇r L (r ∗ , λ∗ ) = ∇Π(r ∗ ) + C T λ∗ = 0,
∗
∗
∗
∇λ L (r , λ ) = Cr = 0,
(24a)
(24b)
where r ∗ and λ∗ are vectors of true nodal degrees of freedom and Lagrange multipliers. Due
to the presence of quadratic terms in the normal strain (5a), the resulting system of nonlinear
equations must be resolved by the Newton scheme, by iteratively searching for the correction
k+1 δr to the known nodal displacements k r in the form
k+1
r = k r + k+1 δr.
(25)
To this purpose, we linearize the term
∇Π(k+1 r) ≈ ∇Π(k r) + ∇2 Π(k r)k+1 δr,
(26)
and plug it into the optimality condition (24a). As a result, we receive a linear system in the
form
k
k
K C T k+1 δr
f int − f ext
,
(27)
k+1 λ = −
C
0
0
in which
k
k
K = ∇2 Π(k r) = K(k r),
k
f int − f ext = ∇Π( r).
8
(28a)
(28b)
Due to the fact that each layer is treated separately, the matrices appearing in Eq. (27) possess
the block structure
(1)
(1)
(1) (1)
f int
f ext
K (r )
(2)
(2)
(2)
(2)
K(r) =
, f int = f int , f ext = f ext ,
(29)
K (r )
(3)
(3)
K (3) (r (3) )
f
f ext
int
and are obtained by the assembly of the contributions of individual elements, cf. (Bittnar and
Šejnoha, 1996)
=
∂Πint,e
∂r (i)
e
(i)
(i)
(i)
(i)
f int,e
K (i)
e
,
=
∂ 2 Πint,e
∂r e(i) 2
=
∂f int,e
∂r (i)
e
.
(30)
The explicit expression for the internal forces and the tangent matrices are elaborated in detail
in Appendix A.
4.3
Implementation issues
The resulting version of the algorithm is shown in Algorithm 1, where the convergence of the
iteration is checked by the equilibrium residual.
k
η=
kk f int − f ext + C Tk λk2
.
max (kf ext k2 , 1)
(31)
Algorithm 1: Conceptual implementation of the Newton method for von Kármán plate
Data: initial displacement 0 r, tolerance
Result: r ∗ , λ∗
k ← 0, 0 λ ← 0, assemble k f int and C
while (k η > ) do
assemble k K
solve for (k+1 δr, k+1 λ) from Eq. (27)
k+1 r ← k r + k+1 δr
assemble k+1 f int
k ←k+1
∗
r ← k r, λ∗ ← k λ
Note that the finite element model of geometrically nonlinear laminated glass plate was
implemented in MATLAB-based application LaPla. In the post-processing step, the stress
components are evaluated from the discretized constitutive equations (9) at the element level
and extrapolated to nodes using standard smoothing procedures, e.g. (Hinton and Campbell,
1974). An interested reader is referred to Zemanová (2014) for a brief description of MATLABbased program LaPla.
5
Examples and discussion
We validated and verified the developed finite element model, Section 5.1, and compared it
against the semi-analytical and effective thickness solutions, Sections 5.2 and 5.3.
9
5.1
Experimental validation and numerical verification
We considered a laminated glass unit simply supported along all four edges as an example
problem for the deflection and stress analysis. Vallabhan et al. (1993) performed a series of
tests on such laminated units. In-plane dimensions of the unit were 1.5 m × 1.5 m and the unit
was subjected to a transverse loading, gradually increasing to 6.9 kPa in 60 s. The temperature
during a series of tests varied between 21 ◦ C and 27 ◦ C. The maximum lateral displacements and
strains at several locations at the top and bottom of the plate were measured and converted to
the principal stresses.
Table 1 summarizes the thicknesses and material properties of individual layers. The material
parameters of glass were taken from Vallabhan et al. (1993), because these values were assumed
for the conversion of measured strains to stresses. The Poisson ratio for interlayer is according
to draft European standard by CEN/TC 129 (2013b). The value of the shear modulus of PVB
was found by matching the measured central deflection of laminated glass units with the values
obtained by the numerical models. Vallabhan et al. (1993) obtained an agreement between
experiment and finite-difference solver using the value of 689.5 kPa in their study, whereas
we matched experimental data with a smaller value of 400 kPa. Our finite element model is
stiffer due to the explicit inclusion of interlayer thickness and this fact is in accordance with the
results of the fully three-dimensional finite element analysis by Duser et al. (1999), and so is the
decreased optimal value of the shear modulus.
Table 1: Properties of glass and PVB layers
Layer/material
Glass (i = 1, 3) PVB (i = 2)
Thickness t(i) [mm]
4.76
1.52
(i)
Young’s modulus of elasticity E [MPa]
68,900
–
Shear modulus of elasticity G(i) [MPa]
–
0.4 / 0.6895
Poisson’s ratio ν (i) [-]
0.22
0.49
Due to the symmetry of the problem, we used a quarter of the plate for computations and
discretized it with 50 × 50 elements per a layer. The dependence of the central deflection
L
L
and principal stresses, at the center [ L2x , 2y ]4 and at the point with coordinates [ L4x , 2y ], on
the increasing intensity of distributed load fz appears in Figures 2–5. We reproduced the
experimental data from Vallabhan et al. (1993) and determined the principal stresses only from
the membrane components in order to compare the principal stresses from the model directly
to experimental values. The response of the finite element model was computed for two values
of interlayer shear modulus, 400 kPa and 689.5 kPa. In addition, we included the prediction
of the monolithic (thickness of two glass panes and interlayer together) and the layered glass
plate (thickness of two independent glass panes), corresponding to the lower and upper bound
for deflections, but not for the principal stresses at the center of the unit. Due to the effects of
geometric nonlinearity, the location of the extreme of principal stresses is not at the center of
the plate, but moves with an increasing load. Such phenomenon is best visible in Figure 3 for
the minimum principal stresses, where the measured values become practically constant at the
load range of 3–5 kPa and then even slightly increase.
4
The origin of the coordinate system is in the corner of the plate. Note that the two extreme principal stresses
should be equal to each other at the center of the unit, because of symmetry. Therefore, two sets of experimental
data appear in Figure 3.
10
Figure 2: Comparison of experimental data for central deflection with the response of nonlinear
finite element model of laminated glass plate with the PVB shear modulus 400 kPa or 689.5 kPa
In Figure 2, we present the comparison of experimental data for the central deflection with
the response of the nonlinear model of laminated glass plates. The proposed nonlinear model
represents the overall behavior of the unit well. We reached the best agreement for the values
of central deflections corresponding to the intensity of the distributed load between 3–5 kPa,
which suggests that the constant value of the shear modulus G(2) is optimally adjusted to this
interval. For lower values of fz , the experimental values of central deflections are smaller and
closer to the monolithic limit, for fz > 5 kPa, the behavior of the unit is closer to the layered
limit.
We attribute this phenomenon to the time-dependent behavior of the shear modulus of the
interlayer under the constant temperature. The load duration, increasing with the value of the
load intensity, corresponds to decreasing values of the effective shear modulus. Therefore, the
effective value of the shear modulus for fz < 3 kPa is higher than the optimized one, which
results in lower values of the central deflections. The effective shear modulus of PVB is lower for
fz > 5 kPa, resulting in higher values of the central deflections. Thus, more accurate results can
be obtained by viscoelastic analysis, but that requires the specification of the shear relaxation
modulus of the interlayer, which is not available in the present case.
Figure 3 shows the comparisons of the results for the maximum principal stresses at the
center of the bottom surface of the bottom layer and for the minimum principal stresses at the
center of the top surface of the top layer. We found excellent agreement between experimental
and computed values for the maximum principal stresses and good agreement for the minimum
principal stresses. Similar conclusions also follow from the comparison of extreme stresses at
the top and bottom layers in a quarter of the plate, Figures 4 and 5, in which we obtained very
good agreement for tensile stresses and good agreement for compression stresses between the
model predictions and experimental values.
In Figure 6, we plotted the distributions of normal stresses through the thickness of the
L
L
previous simply supported laminated glass plate at the center [ L2x , 2y ] and at the point [ L4x , 2y ].
11
Figure 3: Comparison of experimental data for maximum principal stress at the center of the
bottom surface of the bottom layer / for minimum principal stress at the center of the top
surface of the top layer with the response of nonlinear finite element model of laminated glass
plate with the PVB shear modulus 400 kPa or 689.5 kPa
Figure 4: Comparison of experimental data for maximum and minimum principal stress at the
L
point [ L4x , 2y ] of the bottom surface of the bottom layer with the response of nonlinear finite
element model of laminated glass plate with the PVB shear modulus 400 kPa or 689.5 kPa
Figure 5: Comparison of experimental data for maximum and minimum principal stress at the
L
point [ L4x , 2y ] of the top surface of the top layer with the response of nonlinear finite element
model of laminated glass plate with the PVB shear modulus 400 kPa or 689.5 kPa
12
The geometrically nonlinear behavior was assumed and the load intensity was set to 5 kPa. The
normal stresses in the interlayer are negligible when compared with the normal stresses in the
glass layers. The values of normal stresses at the mid-surfaces are generally nonzero, due to the
presence of membrane stresses in the layers, and the load is equally distributed into the two
glass layers, as visible from the identical slopes of the normal stresses.
at the center [ L2x ,
Ly
2 ]
at the point [ L4x ,
Ly
2 ]
at the point [ L4x ,
Ly
2 ]
Figure 6: Distribution of normal stresses through the thickness determined by using the nonlinear
finite element model for the simply supported squared plate uniformly loaded by 5 kPa
After validating the finite element model, we also present its verification against the finite
difference solution by Aşık (2003). The input data for the problem are the same as in Vallabhan
et al. (1993), except for slightly adjusted geometric parameters: the span of the plate is 1.6 m ×
1.6 m and thicknesses of glass units equal to 5 mm. As shown in Figure 8, the distributions of
extreme principal stresses on the top and bottom surfaces of the structure match very well. The
largest difference in maximum values is around 8.5%, which can be caused by only single-digit
precision of the finite difference results for the top surface (the error due to this inaccuracy can
be up to 10%). The differences in maximum values for the bottom surface are under 2.5%. Such
an agreement is reasonable, given that the results were obtained by different numerical method
and that the contours were constructed from different interpolations.
Figure 7: Convergence of nonlinear plate element upon uniform mesh refinement for fz = 5 kPa
Finally, we also checked the convergence of response of the proposed model upon uniform
mesh refinement. The dependence of the central deflections and the principal stress at the center
and at the whole bottom layer appears in Figure 7. Both the central deflections and central
13
fz = 1 kPa
Maximum principal stresses on the bottom surface of the bottom layer
(b) maximum 7.89 MPa
(a) maximum 53.4 MPa
(b) maximum 54.7 MPa
fz = 10 kPa
(a) maximum 7.89 MPa
fz = 1 kPa
Maximum principal stresses on the top surface of the top layer
(b) maximum 5.43 MPa
(a) maximum 50 MPa
(b) maximum 46.7 MPa
fz = 10 kPa
(a) maximum 5 MPa
Figure 8: Contours of maximum principal stresses for distributed load of intensity fz obtained
by (a) Aşık (2003), (b) nonlinear finite element model of the laminated glass plate
14
extreme stresses exhibit a rapid convergence, while the accuracy of the largest principle stress
found in the plate shows significant mesh-dependency. We attribute this phenomenon to the
fact that, as a result of geometric nonlinearity, the extreme stresses become localized to a small
area in the vicinity of the corner, Figure 8, where an adapted finite element mesh would be more
suitable. Nevertheless, the discretization by 50 × 50 elements appears to be sufficiently accurate
for all the three quantities of interest.
5.2
Verification against semi-analytical model
We analyzed again a simply supported square plate loaded with a uniformly distributed pressure
750 Pa. The dimensions of the plate were 3 m × 3 m, Table 2 summarize the thicknesses and
material parameters of layers. The value of the shear modulus of the polymer interlayer was
varied between 10−5 MPa and 105 MPa.5
Table 2: Properties of glass layers and interlayer
Layer/material
Glass (i = 1, 3) Interlayer (i = 2)
Thickness t(i) [mm]
10
1.52
(i)
Young’s modulus of elasticity E [MPa]
70,000
–
Shear modulus of elasticity G(i) [MPa]
–
10−5 – 105
Poisson’s ratio ν (i) [-]
0.22
0.49
The details of the semi-analytical model for a simply-supported rectangular three-layered
plate can be found in Foraboschi (2012). In short, the glass plies were assumed to behave as
thin Kirchhoff plates under small deflections. The interlayer was supposed to be thin, compliant
and to exhibit a state of pure shear. The resulting system was solved by the Fourier series
expansion.
Figure 9: Comparison of maximum deflection and stress at the center of the simply supported
plate obtained by semi-analytical solution (AM) and geometrical linear and nonlinear finite
element model (FEM)
The comparison of maximum deflection and normal stress for the semi-analytical and the
finite element model appears in Figure 9. We achieved an excellent agreement for deflections.
5
Note that the value 105 MPa is not realistic, but it is included in this study to assess the behavior of the
analytical and finite element models in the true monolithic limit, in which the interlayer properties approach the
properties of glass.
15
The differences in values for the semi-analytical solution and geometrical linear finite element
results stay under 2.5% up to the shear modulus of 10 MPa. For larger values of the shear
modulus, the deviations increase up to 10%, because the semi-analytical model in Foraboschi
(2012) is not suitable for such relatively stiff interlayers, for which the model does not approach
the true monolithic response. The results for geometrically linear and nonlinear finite element
model are almost identical for the shear modulus of interlayer greater than 1 MPa, but the
differences increase up to 40% for small values of the shear modulus (close to the layered limit).
The maximum deflection at the center of plate for the geometrically nonlinear prediction ranges
1
1
to 200
of the span, which implies that the effects of geometric nonlinearity can be
from 600
significant in practically relevant cases.
Analogous conclusions follow for the normal stress distributions, Figure 9. The agreement
of solutions is good, the maximum error for geometrically linear model is about 8.5%, which
corresponds with conclusions in Foraboschi (2012). For the geometrically linear and nonlinear
solution, the monolithic limit is almost the same, but the layered limit again shows important
differences.
5.3
Comparison with effective thickness approach
One of the popular simplified approaches to the structural design of laminated glass units are
effective thickness methods, e.g. Benninson et al. (2008) or Galuppi et al. (2012). We compared
the solution of our finite element model for laminated glass plates with the effective thickness
approach using the enhanced effective thickness (EET) method of Galuppi et al. (2012), which
gives more accurate results for deflection and stress calculations for different boundary conditions
than the approach of Benninson et al. (2008). We analyzed two rectangular plates loaded with
a uniformly distributed pressure 750 Pa with the size 3 m × 2 m. Table 3 summarize the
thicknesses and material parameters of layers. The value of shear modulus of the polymeric
interlayer was varied between 0.01 MPa and 10 MPa. Four different boundary conditions were
considered: four, three or two edges simply supported, and two opposite edges simply supported
and one fixed, cf. Figure 10.
Table 3: Properties of glass layers and interlayer
Layer/material
Glass (i = 1, 3) Interlayer (i = 2)
(i)
thickness t [mm]
10
0.76
Young’s modulus of elasticity E (i) [MPa]
70,000
–
(i)
shear modulus of elasticity G [MPa]
–
0.01 – 10
Poisson’s ratio ν (i) [-]
0.22
0.49
We compared the effective thicknesses computed according to Galuppi et al. (2012) with the
values obtained from the finite element modeling, Figure 10. The effective thickness method
provides the deflection- and stress-effective thickness, which is the constant thickness of the
equivalent monolithic homogeneous plate that has the same maximal deflection or maximal
stress as laminated glass unit under the same boundary and load conditions. The effective
thickness approach does not take into account the geometric nonlinearity, but in this example,
the effect of geometric nonlinearity is minor. Therefore, we considered only the geometrically
linear response of the finite element model. We found the effective thicknesses for the finite
element model (FEM) by matching the maximum value of deflections and normal stresses of the
16
(a)
(b)
(c)
(d)
Figure 10: Comparison of the effective thickness obtained with the EET approach and computed
from the response of the proposed model (FEM) for rectangular plates (a) with four edges simply
supported, (b) with three edges simply supported, (c) with two opposite edges simply supported,
(d) with two opposite edges simply supported and one fixed
17
laminated glass unit to the response of the equivalent monolithic homogeneous plate. Due to
the symmetry of the problem, we used only a quarter or a half of the plate for computations
and discretized it with 40 × 60 or 40 × 120 elements per a layer, respectively.
It can be observed that the enhanced effective thickness approach gives a very good approximations for the response of rectangular laminated glass subjected to a uniformly distributed
transverse loading, which in practice coincide with the finite element solution for the plate with
two or three edges simply supported. On the other hand, the enhanced effective method predicts
effective thicknesses slightly greater than these from the numerical simulations for the other two
boundary conditions, so that it underestimates the maximal deflections and stresses.
6
Conclusions
The assemblies of multiple elastic elements bonded by a compliant interlayer are frequently
encountered in modern constructions. The main contribution of this work is the development of
a finite element model describing the laminated glass units based on the Mau refined theory for
plates, which avoids fully resolved three-dimensional simulations, and its detailed verification
and validation. We showed in the presented work that such an approach circumvents the limited
accuracy of similar models available in most finite element systems, which employ a single set
of kinematic variables and average the mechanical response through the thickness of the plate.
Although this paper was focused exclusively on the laminated glass units with two glass plates
and an interlayer, the numerical models for multi-layered sandwich structures developed in this
work are general and may find applications beyond glass structures. As a particular example,
consider structural insulating panels consisting of a layer of polymeric foam sandwiched between
two layers of structural board or laminated hybrid-glass units.
This study represents a first step to the development of a comprehensive, mechanics-based
model for laminated glass systems, which is suitable for implementation in common engineering
finite element solvers. As shown by the presented results, the proposed numerical methods
are well-suited for the modeling of laminated glass beams and plates, mainly because of the
reduction of computational cost and accurate representation of the structural member behavior.
The other advantage of the proposed numerical framework is its conceptual simplicity, which
will enable us to easily include additional refinements of the model, such as a more accurate
description of the time/temperature-dependence of the interlayer or delamination.
Acknowledgments:
This publication was supported by the European social fund within the framework of realizing
the project ”Support of inter-sectoral mobility and quality enhancement of research teams at
Czech Technical University in Prague”, CZ.1.07/2.3.00/30.0034. Period of the project’s realization 1.12.2012 – 30.6.2015. In addition, Jan Zeman acknowledges the partial support of the
European Regional Development Fund under the IT4Innovations Centre of Excellence, project
No. CZ.1.05/1.1.00/02.0070.
18
References
Aşık, M. Z. (2003). Laminated glass plates: revealing of nonlinear behavior. Computers & Structures, 81(28–29):2659–2671.
Behr, R. A., Minor, J. E., Linden, M. P., and Vallabhan, C. V. G. (1985). Laminated glass units
under uniform lateral pressure. Journal of Structural Engineering, 111(5):1037–1050.
Behr, R. A., Minor, J. E., and Norville, H. S. (1993). Structural behavior of architectural
laminated glass. Journal of Structural Engineering, 119(1):202–222.
Benninson, S. J., Qin, M. H. X., and Davis, P. S. (2008). High-performance laminated glass for
structurally efficient glazing. Innovative light-weight structures and sustainable facades, Hong
Kong.
Bennison, S. J., Jagota, A., and Smith, C. A. (1999). Fracture of Glass/Poly(vinyl butyral) (Butacite R ) laminates in biaxial flexure. Journal of the American Ceramic Society,
82(7):1761–1770.
Bittnar, Z. and Šejnoha, J. (1996). Numerical Methods in Structural Mechanics. ASCE Press
and Thomas Telford Ltd, New York and London.
Bonnans, J. F., Gilbert, J. C., Lemaréchal, C., and Sagastizábal, C. A. (2003). Numerical
Optimization: Theoretical and Practical Aspects. Springer.
CEN/TC 129 (2013a). Draft prEN 16612: Glass in building – determination of the load resistance
of glass panes by calculation and testing. Technical report.
CEN/TC 129 (2013b). Draft prEN 16613: Glass in building – laminated glass and laminated
safety glass – determination of interlayer mechanical properties. Technical report.
Duser, A. V., Jagota, A., and Bennison, S. J. (1999). Analysis of glass/polyvinyl butyral laminates subjected to uniform pressure. Journal of Engineering Mechanics, 125(4):435–442.
Foraboschi, P. (2012). Analytical model for laminated-glass plate. Composites: Part B, 43:2094–
2106.
Foraboschi, P. (2013). Three-layered plate: Elasticity solution. Composites: Part B, 60:764–776.
Galuppi, L., Manara, G., and Royer-Carfagni, G. (2012). Practical expression for the design of
laminated glass. Composites: Part B, 45:1677–1688.
Galuppi, L. and Royer-Carfagni, G. (2012). The effective thickness of laminated glass plates.
Journal of Mechanics of Materials and Structures, 7:375–400.
Hinton, E. and Campbell, J. S. (1974). Local and global smoothing of discontinuous finite
element functions using a least squares method. International Journal for Numerical Methods
in Engineering, 8(3):461–480.
Jirásek, M. and Bažant, Z. P. (2002). Inelastic analysis of structures. John Wiley & Sons, Ltd.
19
Kruis, J., Matouš, K., and Dostál, Z. (2002). Solving laminated plates by domain decomposition.
Advances in Engineering Software, 33(7–10):445–452.
Mau, S. T. (1973). A refined laminated plate theory. Journal of Applied Mechanics–Transactions
of the ASME, 40(2):606–607.
Pica, A., Wood, R., and Hinton, E. (1980). Finite element analysis of geometrically nonlinear
plate behaviour using a Mindlin formulation. Computers & Structures, 11(3):203–215.
Timoshenko, S. and Woinowsky-Krieger, S. (1959). Theory of plates and shells. McGraw-Hill
Book Company.
Vallabhan, C. V. G., Das, Y. C., Magdi, M., Asik, M., and Bailey, J. R. (1993). Analysis of
laminated glass units. Journal of Structural Engineering, 119(5):1572–1585.
Vallabhan, C. V. G., Das, Y. C., and Ramasamudra, M. (1992). Properties of PVB interlayer
used in laminated glass. Journal of Materials in Civil Engineering, 4(1):71–76.
Vallabhan, C. V. G., Minor, J. E., and Nagalla, S. R. (1987). Stress in layered glass units and
monolithic glass plates. Journal of Structural Engineering, 113:36–43.
Šejnoha, M. (1996). Micromechanical modeling of unidirectional fibrous plies and laminates. PhD
thesis, Rensselaer Polytechnic Institute, Troy, NY.
Zemanová, A. (2014). Numerical modeling of laminated glass structures. PhD thesis, Faculty of
Civil Engineering, CTU, Prague.
Zemanová, A., Zeman, J., and Šejnoha, M. (2014). Numerical model of elastic laminated
glass beams under finite strain. Archives of Civil and Mechanical Engineering, 14(4):734–744.
1303.6314
A
Sensitivity analysis
We approximated the unknown fields of displacements and rotations by bi-linear functions at
the four-node quadrilateral element
(i)
(i)
u0 (x, y) ≈ N (i)
n,e (x, y)r n,e ,
(i)
(i)
ϕ(i) (x, y) ≈ N b,e (x, y)r b,e ,
(i)
(i)
(i)
w0 (x, y) ≈ N (i)
w,e (x, y)r s,e .
(32)
The vector r e represents generalized nodal displacements of the e-th element at the i-th layer
and it will be convenient for further discussion to organize the degrees of freedom into four
vectors
h
i
(i)
(i)
(i)
(i)
(i)
(i)
(i)
(i)
r (i)
ue,1 , ve,1 , ue,2 , ve,2 , ue,3 , ve,3 , ue,4 , ve,4 T ,
n,e =
h
i
(i)
(i)
(i)
(i)
(i)
(i)
(i)
(i)
(i)
r b,e = ϕx,e,1 , ϕy,e,1 , ϕx,e,2 , ϕy,e,2 , ϕx,e,3 , ϕy,e,3 , ϕx,e,4 , ϕy,e,4 T ,
h
i
(i)
(i)
(i)
(i)
(i)
(i)
(i)
(i)
(i)
(i)
(i)
(i)
(i)
r s,e = we,1 , ϕx,e,1 , ϕy,e,1 , we,2 , ϕx,e,2 , ϕy,e,2 , we,3 , ϕx,e,3 , ϕy,e,3 , we,4 , ϕx,e,4 , ϕy,e,4 T ,
h
i
(i)
(i)
(i)
(i)
(i)
r K,e = we,1 , we,2 , we,3 , we,4 T ,
20
where the first three quantities correspond to the degrees of freedom contributing to the midsurface membrane strains, pseudo-curvatures, and transverse shear strains, respectively, and
(i)
the last matrix r K,e stores the degrees of freedom relevant to the nonlinear membrane strains,
Eq. (7).
The matrices of basis function are formed in the standard way
N1 0 N2 0 N3 0 N4 0
(i)
(i)
N n,e (x, y) = N b,e (x, y) =
,
0 N1 0 N2 0 N3 0 N4
N1 0 0 N2 0 0 N3 0 0 N4 0 0 .
N (i)
w,e (x, y) =
The matrices needed to evaluate the generalized strains then follow directly from Eq. (5) in
the form
(i)
B (i)
n,e (x, y) = ∂N n,e (x, y),
(i)
(i)
B b,e (x, y) = ∂SN b,e (x, y),
(i)
B (i)
w,e (x, y) = ∇N w,e (x, y),
and the matrices for evaluation the von Kármán term in Eq. (5a) read as
∂N1 ∂N2 ∂N3 ∂N4
(i)
,
B K,x,e (x, y) =
∂x
∂x
∂x
∂x
∂N1 ∂N2 ∂N3 ∂N4
(i)
B K,y,e (x, y) =
,
∂y
∂y
∂y
∂y
and the matrix for evaluation the transverse shear strains from Eq. (5b)
0
0
N1 0
0
N2 0
0
N3 0
0
N4
(i)
N s,e (x, y) =
.
0 −N1 0 0 −N2 0 0 −N3 0 0 −N4 0
We used 2 × 2 Gauss quadrature with four integration points [xg , yg ]4g=1 to evaluate the
normal and the bending terms, and 1 × 1 quadrature at the center for the shear terms as a
remedy to shear locking. As a result, we obtained
(i)
f int,n,e
=
4
X
T
(i)
αg B (i)
n,e (xg , yg )D n
(i)
B (i)
n,e (xg , yg )r n,e
+
(i)
εm0,N L,e
g=1
=
4
X
T
(i)
αg B (i)
n,e (xg , yg )ne (xg , yg ),
g=1
(i)
f int,b,e =
4
X
(i)
(i)
(i)
(i)
αg B b,e T (xg , yg )D b B b,e (xg , yg )r b,e =
g=1
4
X
(i)
αg B b,e T (xg , yg )m(i)
e (xg , yg ),
g=1
(i)
(i)
T (i)
(i)
(i)
f int,s,e = α0 B (i)
(x
,
y
)
+
N
(x
,
y
)
D
B
(x
,
y
)
+
N
(x
,
y
)
r (i)
0
0
0
0
0
0
0
0
w,e
s,e
s
w,e
s,e
s,e
(i)
T (i)
= α0 B (i)
q e (xg , yg ),
w,e (x0 , y0 ) + N s,e (x0 , y0 )
(i)
f int,K,e =
4
X
(i)
(i)
(i)
(i)
αg ∇εm0,N L,e (xg , yg ) T D (i)
B
(x
,
y
)r
+
ε
(x
,
y
)
n
n,e g g n,e
m0,N L,e g g
g=1
=
4
X
(i)
αg ∇εm0,N L,e (xg , yg ) T n(i)
e (xg , yg ),
g=1
21
in which the additional terms read
(i)
(i) 2
B K,x,e (xg , yg )r K,e
(i)
(i) 2
1
,
2 B K,y,e (xg , yg )r K,e
(i)
(i)
(i)
(i)
B K,x,e (xg , yg )r nK,e B K,y,e (xg , yg )r K,e
(i)
εm0,N L,e (xg , yg ) =
"
(i)
∇εm0,N L,e (xg , yg )
=
1
2
#
(i)
(i)
(i)
B
(xg , yg )r K,e B K,x,e (xg , yg )
K,x,e
(i)
(i)
(i)
B K,y,e (xg , yg )r K,e B K,y,e (xg , yg )
(i)
(i)
(i)
(i)
(i)
(i)
B K,x,e (xg , yg )r K,e B K,y,e (xg , yg ) + B K,y,e (xg , yg )r K,e B K,x,e (xg , yg )
,
and αg and α0 are weight functions for four- or one-point numerical integration.
(i)
(i)
(i)
(i)
Differentiating the terms f int,n,e , f int,b,e , f int,s,e , and f int,K,e with respect to the corresponding degrees of freedom produces the linear and nonlinear parts of the stiffness matrices. Thus,
the linear contributions to the element stiffness matrix become
4
X
K (i)
n,e =
T
(i) (i)
αg B (i)
n,e (xg , yg )D n B n,e (xg , yg ),
g=1
(i)
K b,e
4
X
=
(i)
(i)
(i)
αg B b,e T (xg , yg )D b B b,e (xg , yg ),
g=1
(i)
(i)
T (i)
(i)
K (i)
D s B w,e
(x0 , y0 ) + N (i)
s,e = α0 B w,e (x0 , y0 ) + N s,e (x0 , y0 )
s,e (x0 , y0 ) .
The treatment of the “von Kármán term” is more involved, due to the nonlinear interaction of
(i)
(i)
quantities r n,e and r K,e . We expressed the additional contributions to the stiffness matrix in
the form
(i)
K nK,e =
4
X
(i)
T
(i)
αg B (i)
n,e (xg , yg )D n ∇εm0,N L,e (xg , yg ),
g=1
(i)
K Kn,e
(i)
= K nK,e T ,
(i)
K K,e =
4
X
(i)
(i)
αg ∇εm0,N L,e (xg , yg ) T D (i)
n ∇εm0,N L,e (xg , yg )
g=1
+
4
X
h
i
(i)
(i)
(i)
αg K (i)
,
Kw,e,1 K Kw,e,2 K Kw,e,3 K Kw,e,4
g=1
with
(i)
K Kw,e,c =
(i)
(i)
(i)
(i)
(i)
ne (xg , yg ).
BK,x,e,c (xg , yg )B K,x,e T (xg , yg )
BK,y,e,c (xg , yg )B K,y,e T (xg , yg )
(i)
(i)
(i)
(i)
BK,x,e,c (xg , yg )B K,y,e T (xg , yg ) + BK,y,e,c (xg , yg )B K,x,e T (xg , yg )
(i)
Note that, for example, the matrix K nK,e is localized to rows corresponding to the degrees of
(i)
(i)
(i)
freedom collected in r n,e and the columns corresponding to r K,e , and BK,x,e,c (xg , yg ) abbreviates
(i)
the c-th entry of matrix B K,x,e .
22
| 5cs.CE
|
arXiv:1708.04419v1 [math.OC] 15 Aug 2017
DISCRETE TIME PONTRYAGIN MAXIMUM PRINCIPLE FOR OPTIMAL
CONTROL PROBLEMS UNDER STATE-ACTION-FREQUENCY
CONSTRAINTS
PRADYUMNA PARUCHURI AND DEBASISH CHATTERJEE
Abstract. We establish a Pontryagin maximum principle for discrete time optimal control
problems under the following three types of constraints: a) constraints on the states pointwise in time, b) constraints on the control actions pointwise in time, and c) constraints on
the frequency spectrum of the optimal control trajectories. While the first two types of constraints are already included in the existing versions of the Pontryagin maximum principle,
it turns out that the third type of constraints cannot be recast in any of the standard forms of
the existing results for the original control system. We provide two different proofs of our
Pontryagin maximum principle in this article, and include several special cases fine-tuned
to control-affine nonlinear and linear system models. In particular, for minimization of
quadratic cost functions and linear time invariant control systems, we provide tight conditions under which the optimal controls under frequency constraints are either normal or
abnormal.
§1. Introduction
As control engineers we encounter various types of constraints in control systems for
a plethora of reasons: limitations on the magnitude of actuator outputs are almost omnipresent; bounds on the state variables of, e.g., robotic arms and chemical plants, should be
ensured for safety considerations; satellites that image particular geographical areas of the
earth must orient themselves and point at precise coordinates at pre-specified instants of
time, etc. While constrained control problems are difficult in general, and this is evidenced
by the fact that the literature on unconstrained control problems by far outweighs that on
constrained problems, control synthesis techniques that account for all possible constraints
are bootstrapped to result in greater accuracy due to increased awareness of the actuator
limitations and foresight. The burgeoning demand for execution of precise control tasks
necessitates the development of tools that permit the inclusion of such constraints at the
synthesis stage, and in this respect, inclusion of control frequency constraints is a natural
direction to pursue.
Optimal control theory provides us with a set of sophisticated and powerful tools to design
controllers under an array of constraints, and also to boost performance by taking account
of such constraints on the states and the control actions in time domain. These techniques
typically rely on the assumption that the values attained by the candidate control functions
Key words and phrases. optimal control, Pontryagin maximum principle, frequency constraints.
The authors are with Systems & Control Engineering, IIT Bombay, Powai, Mumbai 400076, India. They were
supported in part by the grant 17ISROC001 from the Indian Space Research Organization. The authors thank
Masaaki Nagahara for helpful discussions during the early phases of this work, and Navin Khaneja for suggesting
the technique behind the second proof of our main result.
Emails: [email protected], [email protected].
1
2
P. PARUCHURI AND D. CHATTERJEE
can be changed arbitrarily quickly over time, but such an assumption rarely holds true in
practice. In particular, inertial actuators such as robotic arms, rotating machines, etc., cannot
faithfully execute control commands that demand very quick transitions between different
control values. Such issues naturally lead to lacunae between the control commands
received at the actuators and those that are faithfully executed, thereby contributing to loss
of precision and the emergence of differences between desired and observed outputs.
This article addresses a class of optimal control problems that includes constraints on
the frequency of admissible control functions in addition to state and control constraints.
More specifically, we address optimal control problems for discrete-time nonlinear smooth
control systems with the following three important classes of constraints:
(I) constraints on the states at each time instant,
(II) constraints on the control magnitudes at each time instant, and
(III) constraints on the frequency of the control functions.
Constraints on the states (as in (I)) are desirable and/or necessary in most applications; the
class of constraints treated here are capable of describing a general class of path-planning
objectives, and subsumes both ballistic and servomechanism reachability problems. Constraints on the control magnitudes (as in (II)) are typically simpler to deal with compared to
state constraints; in particular, the two general techniques for synthesis of optimal controls,
namely, dynamic programming and the maximum principle [Lib12], are capable of dealing
with these constraints with relative ease.
Constraints on the control frequencies (as in (III)), in contrast to the other two types of
constraints, are rarely encountered in the theory despite the fact that control theory started
off with the so-called frequency-domain techniques. A well-known and widely employed
control strategy that treats frequency-domain properties of control functions is the so-called
H ∞ control [ZDG95], but these techniques can neither treat pre-specified hard bounds on
the frequency components in the control signals, nor are they capable of admitting state or
control constraints at the synthesis stage. Frequency constraints can be indirectly addressed
in H ∞ control through penalization of appropriate H ∞ norms, but such designs rely on
heuristics and many trial-and-error steps. To the best of our knowledge, except for a US
patent [SB95] where frequency constraints were imposed specifically to avoid a resonance
mode in the arm of the read head positioner in a disk drive, there has been no systematic
investigation into control with mixed frequency and time-domain constraints.
The celebrated Pontryagin maximum principle [Bol78], a central tool in optimal control
theory, provides first order necessary conditions for optimal controls. These necessary
conditions, or equivalently, characterizations of optimal controls, serve to narrow the
search space over which algorithms can play and extract optimal controls. The discrete time
Pontryagin maximum principle was developed primarily by Boltyanskii (see [Bol75, Bol78]
and the references therein), with several early refinements reported in [DM65, DP75,
Dub78], and perhaps the most recent extensions appearing in [BT16]; see [Psh71] for a
careful discussion about the differences between continuous and discrete time versions of
the Pontryagin maximum principle. While these versions of the Pontryagin maximum
principle are capable of handling constraints of the form (I) and (II), the new ingredient
in this article is the set of frequency constraints (III). We formulate frequency constraints
on the control functions in terms of the active support set — the set on which the Fourier
transform of the control function is allowed to take non-zero values. We engineer bandlimited controls via appropriately defining the active sets; the constraints may be selected
PMP UNDER FREQUENCY CONSTRAINTS
3
based on specific features or physics of the actuators, thereby ensuring faithful execution of
the control commands. Our main result — Theorem 3.1 in §3 — is a Pontryagin maximum
principle for discrete-time nonlinear control systems with smooth data under all the three
types of constraints (I), (II), and (III).
This maximum principle yields a well-defined two-point boundary value problem, which
may serve as a starting point for algorithms such as shooting techniques that typically employ
variants of Newton methods, to arrive at optimal control functions. If a solution of the twopoint boundary value problem is found, feasibility of the original optimal control problem
is automatically established. However, since the maximum principle provides (local)
necessary conditions for optimality, not all solutions may achieve the minimum cost, and
further analysis may be needed to select the cost-minimizing controls. A number of special
cases of the main result, dealing with control-affine nonlinear systems, time-varying linear
systems, etc., are provided in §3, and the important special case of optimal control of linear
time-invariant control systems under quadratic costs and frequency constraints is treated
in §4. Two different proofs of Theorem 3.1 are provided in Appendix C, and Appendix D
contains the proofs of the various special cases. The necessary prerequisites for the proofs
are reviewed in Appendices A-B.
Notation. We employ standard notation: N denotes the non-negative integers, N∗ the
positive integers, R the real numbers, and C the complex numbers. We denote by the
standard partial order on the set Rn induced by the non-negative orthant: for a, b ∈ Rn ,
a b iff ai 6 bi for every
√ i = 1, . . . , n; we sometimes write b a to express the same
statement. For us i ≔ −1 is the unit complex number, In is the n × n identity matrix.
The vector space Rn is always assumed to be equipped with the standard
⋆inner product
hv, v ′i ≔ v ⊤ v ′ for every v, v ′ ∈ Rn . In the theorem statements we use Rn to denote the
⋆
dual space of Rn for the sake of precision; of course, Rn is isomorphic to Rn in view of
the Riesz representation theorem.
§2. Problem Setup
Consider a discrete time control system described by
xt+1 = ft (xt , ut ) for t = 0, . . . , T − 1,
(2.1)
−1 is a family of maps such that R d × Rm ∋ (ξ, µ) 7−→
where xt ∈ Rd and ut ∈ Rm and ( ft )Tt=0
fs (ξ, µ) ∈ Rd is continuously differentiable for each s = 0, . . . , T − 1.
−1
(k) denote its discrete Fourier
Let u(k) ≔ (ut )Tt=0
denote the k th control sequence, and ud
d
(k)
transform (DFT). The relationship between u(k) and u is given by [SS03, Chapter 7]:
(k)
(2.2)
d
(k) )T −1 =
u(k) ≔ (ud
ξ ξ=0
TÕ
−1
t=0
ut(k) e−i2π ξt/T
T −1
ξ=0
for ξ = 0, . . . , T − 1
and k = 1, . . . , m.
4
P. PARUCHURI AND D. CHATTERJEE
In the context of (2.1), the objective of this article is to characterize solutions of the
finite horizon constrained optimal control problem:
minimize
−1
(ut )Tt=0
(2.3)
subject to
T
−1
Õ
t=0
ct (xt , ut )
dynamics (2.1),
state constraints at each stage t = 0, . . . , T,
control constraints at each stage t = 0, . . . , T − 1,
constraints on frequency components of the control sequence.
where T ∈ N∗ is fixed, and Rd × Rm ∋ (ξ, µ) 7−→ ct (ξ, µ) ∈ R is a continuously differentiable function representing the stage cost at time t, and t = 0, . . . , T − 1.
The three classes of constraints considered in the optimal control problem (2.3) are as
follows:
Control constraints: Ut ⊂ Rm is a given non-empty set for each t = 0, . . . , T . We
impose the constraints that the control action ut at stage t must lie in Ut :
(i)
u t ∈ Ut
(2.4)
for t = 0, . . . , T − 1.
(ii) State constraints: Let St ⊂ Rd be a given non-empty set for each t = 0, . . . , T . We
shall restrict the trajectory of the states (xt )Tt=0 to the tube S0 ×S1 ×· · ·×ST ⊂ (Rd )T +1 ;
xt ∈ S t
(2.5)
for t = 0, . . . , T .
(iii) Frequency constraints: For a control sequence u(k) we define F (k) ⊂ CT to be the set
d
(k) = (u
(k) )T −1 . The set F (k) is constructed
of permissible frequency components ud
ξ ξ=0
such that it allows non-zero components only in the selected frequencies. For a vector
v ∈ CT we define its support as
supp(v) ≔ i ∈ {1, . . . , T } vi , 0 .
We stipulate that
(2.6)
d
u(k) ∈ F (k) ≔ v ∈ CT
supp(v) ⊂ W (k) ,
th
where W (k) ⊂ {1, . . . , T } represents
m the support for the selected frequencies in the k
(k)
control sequence. The sets W k=1 are assumed to be given as part of the problem
specification.
The standard DFT relation in (2.2) can be written in a compact form as:
(2.7)
where
d
u(k) = Fu(k)
© u0 ª
®
(k)
u ≔ ... ® ∈ RT ,
®
(k)
u
« T −1 ¬
1
1
©
1
ω
1
F ≔ √ ..
..
.
T .
T −1
1
ω
«
(k)
for k = 1, . . . , m,
(k)
ud
...
...
..
.
...
d
(k)
© u 0 ª
.. ®
≔ . ® ∈ CT , and
®
d
(k)
u
T
−1
«
¬
1
ª
ωT −1 ®
® ∈ CT ×T ,
..
®
.
®
(T
−1)(T
−1)
ω
¬
PMP UNDER FREQUENCY CONSTRAINTS
5
and ω ≔ e−i2π/T is a primitive T -th root of unity. In order to visualize the frequency
components in all the control inputs, we represent the combined control profile in the
following (stacked) fashion:
c
(1)
©u ª
b ≔ .. ®® ∈ CmT .
(2.8)
and U
. ®
d
(m)
«u ¬
In terms of the representations (2.8), the relation (2.7) can be written in a compact way as:
u(1)
© . ª
U ≔ .. ®® ∈ RmT
(m)
«u ¬
b = F U,
U
(2.9)
with F ≔ blkdiag(F, . . . , F) ∈ CmT ×mT .
Since U is a vector with real entries, the real and imaginary parts of the frequency
components can be separated by considering the real and imaginary parts in the matrix F
individually. To impose the given frequency constraints and yet work with real numbers
b
only, we separate out the real and imaginary parts. We define a band-stop filter BS ≔ SU,
where S ≔ blkdiag(S (1), . . . , S (m) ), with each of the S (k) formed by the rows eξ of IT for
ξ < W (k) . The constraints (2.6) on the frequency components of the control now translate
to:
SFreal
b=0 ⇔
(2.10)
SU
U = 0.
SFimag
SFreal
and let D ∈ RmT ×mT denote the matrix that maps the vector U
SFimag
⊤
uT −1 :
Define F ≔
to u0
...
u0
u(1)
© . ª © . ª
®
.
D . ® = .. ®®
(m)
«u ¬ «uT −1 ¬
(2.11)
Observe that D is non-singular since the transformation representing D is a permutation
matrix, and in particular is a bijection. Then we can write the frequency constraints in (2.6)
as
u0
©
ª
−1 .. ®
F D . ® = 0.
«uT −1 ¬
Eliminating, if necessary, the zero rows of the matrix F , our constraint takes the form
T
−1
Õ
(2.12)
t=0
et ut = 0.
F
et ∈
where F
represents the corresponding columns of F D−1 that multiply ut . In
other words, there exists a linear map F : mT −→ mT that describes the constraints on the
−1
frequency spectrum of the control trajectory (ut )Tt=0
as the following equality constraint:
Rℓ×m
(2.13)
F(u0, . . . , uT −1 ) =
T
−1
Õ
t=0
et ut = 0
F
et
for F
We shall refer to F as our frequency constraint map.
T −1
t=0
⊂ Rℓ×mT as in (2.12).
6
P. PARUCHURI AND D. CHATTERJEE
The abstract optimal control problem (2.3) can now be formally written as:
minimize
−1
(ut )Tt=0
(2.14)
subject to
with the following data:
T
−1
Õ
t=0
ct (xt , ut )
dynamics (2.1),
xt ∈ St for t = 0, . . . , T,
ut ∈ Ut for t = 0, . . . , T − 1,
F(u , . . . , u ) = 0,
0
T −1
(2.14-a) T ∈ N∗ is fixed;
(2.14-b) Rd × Rm ∋ (ξ, µ) 7−→ ct (ξ, µ) ∈ R is a continuosly differentiable function for
each t = 0, . . . , T − 1;
(2.14-c) St is a subset of Rd for each t;
(2.14-d) Ut is a subset of Rm for each t;
(2.14-e) F : RmT −→ Rℓ is a given linear map on the control trajectory u0, . . . , uT −1 for
some ℓ ∈ N∗ .
−1 of (2.14) is a sequence in ÎT −1 U , and it generates its
An optimal solution (ut∗ )Tt=0
i
i=0
−1
corresponding optimal state trajectory (xt∗ )Tt=0 according to (2.1). The pair (xt∗ )Tt=0, (ut∗)Tt=0
is called an optimal state-action trajectory.
Remark 2.1. Constraints on the control frequencies cannot in general be translated into
equivalent constraints on the control actions and/or the states of the original system. Had
that been possible, the standard PMP would have sufficed. To see this negative assertion,
consider the simple case that the system (2.1) is linear and time-invariant, i.e., ft (ξ, µ) =
Aξ + Bµ for all t and for some fixed A ∈ Rd×d and B ∈ Rd×m . Assume further that
the frequency constraint map F is a bijection. Even then the constraint on the control
actions F(u0, . . . , uT −1 ) = 0 cannot in general be transformed into equivalent constraints
on the states of the form (xt )Tt=1 ∈ S ⊂ RdT . Indeed, when T > d, constraints on the
control actions can only be contained in constraints of the form (xt )Tt=1 ∈ S ⊂ RdT since
−1 to state trajectory (x )T
the transformation from the control trajectory (ut )Tt=0
t t=1 is not a
bijection. A fresh investigation is, therefore, needed. The standard PMP [Bol75, Theorem
20] deals with constraints on the states and control actions that are expressed pointwise in
time. Since constraints on the frequency components of the control, by definition, bring in
dependence among the control actions at each time, the standard Hamiltonian maximization
condition [Bol75, Theorem 20 (C)] cannot be used as is.
§3. Main Result
The following theorem provides first order necessary conditions for optimal solutions of
(2.14); it is the main result of this article.
Theorem 3.1 (Pontryagin
maximum principle under state-action-frequency constraints).
−1 be an optimal state-action trajectory for (2.14) with F as defined in
Let (xt∗ )Tt=0, (ut∗ )Tt=0
(2.13). Define the Hamiltonian
⋆
⋆
R × Rℓ × Rd × N × Rd × Rm ∋ (ν, ϑ, ζ, s, ξ, µ) 7−→
D
E
(3.1)
es µ ∈ R.
Hν,ϑ (ζ, s, ξ, µ) ≔ hζ, fs (ξ, µ)i − νcs (ξ, µ) − ϑ, F
PMP UNDER FREQUENCY CONSTRAINTS
7
Then there exist
⋆
T −1
◦ a trajectory ηtf t=0 ⊂ Rd ,
T
⋆
◦ a sequence ηtx t=0 ⊂ Rd , and
⋆
◦ a pair ηC, ηbu ∈ R × Rℓ ,
satisfying the following conditions:
(PMP-i)
non-negativity condition
ηC > 0;
(PMP-ii) non-triviality condition
T −1
the adjoint trajectory ηtf t=0 and the pair ηC, ηbu do not simultaneously vanish;
(PMP-iii) state and adjoint system dynamics
∂ η C, ηcu f
∗
xt+1
=
for t = 0, . . . , T − 1,
H
ηt , t, xt∗, ut∗
∂ζ
∂ η C, ηcu f
f
H
ηt , t, xt∗, ut∗ − ηtx for t = 1, . . . , T − 1,
ηt−1
=
∂ξ
where ηtx lies in the dual cone of a tent qtx (xt∗ ) of St at xt∗ ;
(PMP-iv) transversality conditions
∂ η C, ηcu f
and
ηTf −1 = −ηTx ,
H
η0, 0, x0∗, u0∗ − η0x = 0
∂ξ
where η0x lies in the dual cone of a tent q0x (x0∗ ) of S0 at x0∗ and ηTx lies in the dual
cone of a tent qTx (xT∗ ) of ST at xT∗ ;
(PMP-v) Hamiltonian maximization condition, pointwise in time,
∂ η C, ηcu f
∗ ∗
H
ηt , t, xt , ut , ũt 6 0 whenever ut∗ + ũt ∈ qtu (ut∗),
∂µ
where qtu (ut∗ ) is a local tent at ut∗ of the set Ut of admissible actions;
(PMP-vi) frequency constraints
F u0∗, . . . , uT∗ −1 = 0.
We present a complete proof of Theorem 3.1 in Appendix §C. The rest of this section
is devoted to a scrutiny of various facets of Theorem 3.1 over a sequence of remarks, and
providing a set of corollaries catering to various special cases.
Remark 3.1. It is readily observed that since the scalar ηC and the vectors ηtf, ηtx, ηbu enter linC u
early in the Hamiltonian function H η , ηc , the non-negativity condition
(PMP-i) on ηC can be
T −1
C
C
−1
equivalently posed as the condition that η ∈ {0, 1}. A quintuple η , ηbu, ηtf t=0 , (xt∗ )Tt=0, (ut∗ )Tt=0
that satisfies the PMP is called an extremal lift of the optimal state-action trajectory
−1
. Extremal lifts with ηC = 1 are called normal extremals and the ones
(xt∗ )Tt=0, (ut∗ )Tt=0
C
with η = 0 are caled abnormal extremals
D
E
et µ in the Hamiltonian H η C, ηcu is an additional term compared
Remark 3.2. The term ηbu, F
to the usual Hamiltonian formulation and corresponds to the constraints on the frequency
components of the control sequence. Observe that since this term does not enter the
conditions (PMP-i), (PMP-iii) and (PMP-iv), the state and adjoint dynamics are unaffected.
The element ηbu is a new entity in Theorem 3.1 compared toth e classical PMP in [Bol78].
8
P. PARUCHURI AND D. CHATTERJEE
Remark 3.3. From the definition of the Hamiltonian function H η
∂ η C, ηcu f
H
ηt , t, xt∗, ut∗ = ft (xt∗, ut∗ ).
∂ζ
C, η
cu
we see that
In other words, the state dynamics prescribed
in (PMP-iii) simply states that the optimal
−1 satisfies the system dynamics (2.1).
state-action trajectory (xt∗ )Tt=0, (ut∗ )Tt=0
Remark 3.4. The tents qtx (xt∗ ) and qtu (ut∗ ) mentioned in (PMP-iii) and (PMP-v) are linear
approximations of the sets St and Ut locally at xt∗ and ut∗ respectively. Precise definitions of
these tents will be given in Appendix §B. Intuitively, a tent (to a set at a point) consists of a
set of directions along which it is possible to enter the set from that point. By construction
a tent to a set at a point is a convex cone. The dual cone of a cone C is the convex cone
that consists of all the directions along which one can most efficiently exit/leave the cone
C. The vectors ηtx lying in dual cones of a tent qtx (xt∗ ) of St at xt∗ represent the directions
along which one can leave the set St most efficiently from xt∗ . A detailed exposition of dual
cones and tents is given in Appendix §A and Appendix §B respectively.
Remark 3.5. In simple terms, the condition (PMP-v) means that along the directions
C u
entering the set Ut from ut∗ , the Hamiltonian H η , ηc does not increase locally. We have used
the name "Hamiltonian maximization condition" for this condition; although not entirely
apt, it is borrowed from the continuous time counterpart of the Pontryagin maximum
principle where the optimal control at time t maximizes the Hamiltonian at that instant
t over the admissible action set. At the level of generality of Theorem 3.1, an actual
Hamiltonian maximization does not hold. However, such a maximization condition does
indeed materialize under additional structural assumptions on the sets of admissible actions,
as described in Corollary 3.2.
Remark 3.6. The conditions (PMP-i) - (PMP-vi) together constitute a well-defined two
point boundary value problem with (PMP-iv) giving the entire set of boundary conditions.
Newton-lie methods may be employed to solve this (algebraic) two point boundary value
problem; see, eg., [Tré12, §2.4] for an illuminating discussion in the context of continuoustime problems. Solution techniques for two point boundary value problems is an active
active field of research.
Remark 3.7. Uncertainty principles in time-frequency analysis impose fundamental restrictions on the classes of control magnitude and frequency constraints. For instance, the
Donoho-Stark
uncertainty principle [DS89] shows that every non-zero
p C-valued function
g : 0, 1, . . . , N − 1 7−→ C must satisfy supp(g) + supp(ĝ) > 2 |N |.1 Applied to the
m
control trajectories u(k) k=1 , one immediately finds that imposing certain types of control
magnitude and frequency constraints simultaneously may lead to empty feasible sets of
controls irrespective of the dynamics and other constraints. In other words, sufficient care
needs to be excercised to ensure a well-posed optimal control problem.
We now describe a few special cases of Theorem 3.1 that are fine-tuned to specific
classes of control systems.
Consider a discrete-time control-affine system described by:
(3.2)
xt+1 = ft (xt ) + gt (xt ) ut
for t = 0, . . . , T − 1,
1Further refinements due to Biro-Meshulam-Tao may be found in [Tao05]; see also [MW12] for a recent
generalization.
PMP UNDER FREQUENCY CONSTRAINTS
9
−1 and (g )T −1 are two families of maps such that
where xt ∈ Rd and ut ∈ Rm , and ( ft )Tt=0
t t=0
d
d
d
R ∋ ξ 7−→ fs (ξ) ∈ R and R ∋ ξ 7−→ gs (ξ) ∈ Rd×m are continuously differentiable
for each s = 0, . . . , T − 1. Consider the optimal control problem (2.14) with the dynamics
given by (3.2):
minimize
−1
(ut )Tt=0
T
−1
Õ
t=0
ct (xt , ut )
dynamics (3.2),
xt ∈ St for t = 0, . . . , T,
(3.3)
ut ∈ Ut for t = 0, . . . , T − 1,
subject to
F(u0, . . . , uT −1 ) = 0,
d
c
t (ξ, ·) : Ut −→ R is convex whenever ξ ∈ R , t = 0, . . . , T − 1,
Ut convex, compact, and non-empty for each t = 0, . . . , T − 1.
−1 be an optimal stateCorollary 3.2 (PMP for control-affine systems). Let (xt∗ )Tt=0, (ut∗ )Tt=0
action trajectory for (3.3) with F as defined in (2.13). Define the Hamiltonian
⋆
⋆
R × Rℓ × Rd × N × Rd × Rm ∋ (ν, ϑ, ζ, s, ξ, µ) 7−→
D
E
(3.4)
es µ ∈ R.
Hν,ϑ (ζ, s, ξ, µ) ≔ hζ, fs (ξ) + gs (ξ) µi − νcs (ξ, µ) − ϑ, F
Then there exist
T −1
⋆
◦ a trajectory ηtf t=0 ⊂ Rd ,
⋆
T
◦ a sequence ηtx t=0 ⊂ Rd , and
⋆
◦ a pair ηC, ηbu ∈ R × Rℓ ,
satisfying the following conditions:
(AFF-i)
non-negativity condition
ηC > 0;
(AFF-ii) non-triviality condition
T −1
the adjoint trajectory ηtf t=0 and the pair ηC, ηbu do not simultaneously vanish;
(AFF-iii) state and adjoint system dynamics
∂ η C, ηcu f
for t = 0, . . . , T − 1,
H
ηt , t, xt∗, ut∗
∂ζ
∂ η C, ηcu f
f
ηt−1
=
H
ηt , t, xt∗, ut∗ − ηtx for t = 1, . . . , T − 1,
∂ξ
⋆
x
where ηt ∈ Rd lies in the dual cone of a tent qtx (xt∗ ) of St at xt∗ ;
(AFF-iv) transversality conditions
∗
xt+1
=
∂ η C, ηcu f
H
η0, 0, x0∗, u0∗ − η0x = 0
∂ξ
and
ηTf −1 = −ηTx ,
where η0x lies in the dual cone of a tent q0x (x0∗ ) of S0 at x0∗ and ηTx lies in the dual
cone of a tent qTx (xT∗ ) of ST at xT∗ ;
(AFF-v) Hamiltonian maximization condition, pointwise in time,
C c
C c
u
u
H η , η ηtf , t, xt∗, ut∗ = max H η , η ηtf , t, xt∗, µ for t = 0, . . . , T − 1;
µ ∈Ut
10
P. PARUCHURI AND D. CHATTERJEE
(AFF-vi) frequency constraints
F u0∗, . . . , uT∗ −1 = 0.
−1 be an optimal state-action trajectory for (3.3) with
Corollary 3.3. Let (xt∗ )Tt=0, (ut∗ )Tt=0
F as defined in (2.13). Moreover, suppose that in the optimal control problem (3.3), the
underlying system is linear, state constraints are absent and the end points x0 and xT are
fixed; i.e.,
(3.5)
and
ft (ξ, µ) = At ξ + Bt µ,
S0 ≔ xini ,
St ≔ R d
ST ≔ xfin ,
for t = 1, . . . , T − 1.
With the Hamiltonian as defined in (3.4), the conditions (AFF-i), (AFF-ii), (AFF-v) and
(AFF-vi) hold, the condition (AFF-iv) is trivially satisfied, and the adjoint dynamics in
(AFF-iii) is given by
(3.6)
f
ηt−1
= A⊤t ηtf − ηC
∂
ct (xt∗, ut∗ )
∂ξ
for t = 1, . . . , T − 1.
§4. Linear quadratic optimal control problems
In this section we discuss three special cases of linear quadratic (LQ) optimal control
problems, all under unconstrained control actions. In §4.1 we address the LQ problem
with initial and final state constraints and demonstrate that all extremals are normal; this
material is standard, but we include it only for the sake of easy reference. §4.2 deals with
a variation of the LQ state-transfer problem where frequency components of the control
sequence are constrained, and we provide conditions for normality of LQ extremals under
frequency constraints.
§4.1. Classical LQ problem. Consider a linear time-invariant incarnation of (2.1):
(4.1)
xt+1 = Axt + But,
t = 0, . . . , T − 1,
where xt ∈ Rd is the state, ut ∈ Rm is the control input at time t, and the system matrix
A ∈ Rd×d and the control matrix B ∈ Rm×m are known. Consider the following finite
horizon LQ problem with unconstrained control actions for the system (4.1) given an initial
state x0 = x:
T
−1
Õ
1
1
minimize
hxt , Qxt i + hut , Rut i
−1
2
2
(u)Tt=0
t=0
(4.2)
(
controlled dynamics (4.1),
subject to
x0 = x,
where R ∈ Rm×m is a given positive definite matrix and Q ∈ Rd×d is a given positive
semi-definite matrix.
The solution of the LQ problem (4.2) can be obtained by using Bellman dynamic
programming (DP) principle and algorithm [Ber95, Chapter 1]. This is sketched below in
−1
: The
(4.3), and it gives sufficient conditions for optimality of a control sequence (ut∗ )Tt=0
PMP UNDER FREQUENCY CONSTRAINTS
11
DP algorithm gives us, with Rd ∋ x 7−→ Jt (x) ∈ R denoting the optimal cost-to-go at stage
t,
JT (x) ≔ 0 for x ∈ Rd
and
1
1
Jt (x) = hx, Qxi + minm hu, Rui + Jt+1 (Ax + Bu)
u ∈R
2
2
(4.3)
for t = 0, . . . , T − 1
The fact that the minimum in (4.3) is attained follows from the assumption that R is positive
definite. The following solution of (4.3) can be derived readily: for t = T − 1, . . . , 0,
PT = 0 ∈ Rd×d,
Kt = −(R + B⊤ Pt+1 B)−1 B⊤ Pt+1 A,
Pt = (A + BKt )⊤ Pt+1 (A + BKt ) + Kt⊤ RKt + Q,
u∗ = K x .
t t
t
(4.4)
It is worth noting that the feedback matrix Kt in (4.4) is independent of any state information,
and depends only on how much longer it takes to reach the final stage and the cost-per-stage
matrices Q and R.
We employ the classical PMP [Bol75, Theorem 16] to (4.2): The Hamiltonian function
for (4.2) is
⋆
Rd × N × Rd × Rm ∋ (ζ, s, ξ, µ) 7−→
Hν (ζ, s, ξ, µ) ≔ hζ, Aξ + Bµi − ν
1
1
hξ, Qξi + hµ, Rµi
2
2
for ν ∈ 0, 1 .
−1 is an optimal state-action trajectory, then there exist adjoint sequence
If (xt∗ )Tt=0, (ut∗)Tt=0
T −1
T −1
ηtf t=0 and ηC ∈ 0, 1 , such that ηC and ηtf t=0 are not simultaneously zero, and the
−1 given by the PMP can
necessary conditions of optimality of the trajectory (xt∗ )Tt=0, (ut∗)Tt=0
be written as
(i) the adjoint and state dynamics (PMP-iii):
(4.5)
∗
xt+1
= Axt∗ + But∗,
for t = 0, . . . , T − 1,
f
ηt−1
= A⊤ ηtf − ηC Qxt∗,
for t = 1, . . . , T − 1;
(ii) the Hamiltonian maximization condition (PMP-v): At each stage t,
(4.6)
∂ ηC f
H (ηt , t, xt∗, ut∗ ) = 0
∂µ
⇒
ηC Rut∗ = B⊤ ηtf ;
(iii) boundary conditions for the recursive equations are given by the transversality conditions (PMP-iv):
x0∗ = x
and ηTf −1 = 0.
If ηC = 0, the adjoint dynamics in (4.5) reduces to
f
ηt−1
= A⊤ ηtf
for t = 1, . . . , T − 1.
Since ηTf −1 = 0, this would imply that ηtf = 0 for all t = 0, . . . , T − 1. In other words, ηC and
T −1
ηtf t=0 would simultaneously vanish, contradicting the non-triviality condition. Hence,
there are no abnormal solutions to the PMP in this case. Substituting ηC = 1, we get the
12
P. PARUCHURI AND D. CHATTERJEE
following set of equations characterising the optimal state-action trajectory.
(4.7)
∗ = Ax ∗ + Bu∗
xt+1
t
t
f
η = A⊤ ηtf − Qxt∗
t−1
ut∗ = R−1 B⊤ ηtf
x ∗ = x, ηf = 0.
0
T −1
for t = 0, . . . , T − 1,
for t = 1, . . . , T − 1,
for t = 0, . . . , T − 1,
−1 as a linear
Observe that (4.7) also characterises the optimal control sequence (ut∗ )Tt=0
feedback of the states, which matches with the solution obtained by solving by dynamic
programming as exposed in [Ber95, Chapter 4].
For a certain class of LQ optimal control problems in the absence of state and control
constraints, all the candidates for optimality are characterised by the PMP with ηC = 1, i.e.,
normal extremals.2 One such example is presented next.
Recall that a linear time-invariant
system (4.1) is controllable if rank B . . . Ad−1 B = d.
Consider a variation of the LQ problem (4.2) where the goal is to reach a specified final
state x̂ ∈ Rd at time T :
T
−1
Õ
1
1
hxt , Qxt i + hut , Rut i
minimize
−1
2
2
(u)Tt=0
t=0
(4.8)
(
controlled dynamics (4.1),
subject to
x0 = x, xT = x̂.
Proposition 4.1. If the underlying system (A, B) in (4.8) is controllable and T > d, then
all the optimal state-action trajectories are normal.
§ 4.2. Normality of LQ state transfer under frequency constraints. Let us consider a
third variation of the LQ optimal control problem (4.2) with constraints on the frequency
components of the control sequence but no state and control constraints. We assume that
our frequency constraints stipulate that certain frequency components are set to 0. We
know (cf. §2, (2.2)) that there are T frequency components in a control sequence of length
T , and let us select p of these to be zero. Recall from (2.12) that such constraints can be
written as
T
−1
Õ
et ut = 0,
F
t=0
et are defined appropriately corresponding to the p frequencies chosen to be elimwhere F
inated as discussed in §2.
Consider
minimize
−1
(u)Tt=0
(4.9)
subject to
2See Remark 3.1.
T
−1
Õ
t=0
1
1
hxt , Qxt i + hut , Rut i
2
2
controlled dynamics (4.1),
ÍT −1
e
t=0 Ft ut = 0,
x0 = x, xT = x̂.
PMP UNDER FREQUENCY CONSTRAINTS
13
Applying the PMP
(cf. Theorem 3.1) to get the necessary conditions of optimality of
−1 , we arrive at the following conditions:
(xt∗ )Tt=0, (ut∗ )Tt=0
T −1
There exist ηC ∈ 0, 1 , ηbu ∈ Rℓ , a sequence of adjoint variables ηtf t=0 , such that
T −1
ηC, ηbu , and ηtf t=0 are not simultaneously zero, and
(4.10)
∗ = Ax ∗ + Bu∗
xt+1
t
t
f
ηt−1
= A⊤ ηtf − ηC Qxt∗
et⊤ ηbu
ηC Rut∗ = B⊤ ηtf − F
Í
T −1 e ∗
t=0 Ft ut = 0,
x ∗ = x, and x ∗ = x̂.
T
0
for t = 0, . . . , T − 1,
for t = 1, . . . , T − 1,
for t = 0, . . . , T − 1,
The adjoint variables are free at the boundary, i.e., η0f and ηTf −1 are arbitrary.
Proposition 4.2. If the underlying system (A, B) in (4.9) is controllable, T > d, and the
number of frequency constraints ℓ satisfies ℓ + d > mT , then all the optimal state-action
trajectories are abnormal. Conversely, all the optimal state-action trajectories are normal
when the reachability matrix B . . . AT −1 B and the frequency constraints matrix F D−1
have independent rows.
Appendix A. Convex Cones and Separability
This section deals with defining the basic concepts regarding convex sets used later in
developing the necessary conditions for optimality.
◦ Let n be a positive integer. Recall that a non-empty subset K ⊂ Rn is a cone if for every
y ∈ K and α > 0 we have αy ∈ K. In particular, 0 ∈ Rn belongs to K. A non-empty
subset C ⊂ Rn is convex if for every y, y ′ ∈ C and θ ∈ [0, 1] we have (1 − θ)y + θ y ′ ∈ C.
◦ A hyperplane Γ in Rn is an (n − 1)-dimensional affine subset of Rn . It can be viewed as
the level set of a nontrivial linear function p : Rn −→ R. If p is given by p(x) = ha, xi
for some a(, 0) ∈ Rn , then
Γ ≔ x ∈ Rn ha, xi = α .
◦ We say that a family {K0, K1, . . . , Ks } of convex cones in Rn is separable Ñ
if there
exists a hyperplane Γ and some i ∈ {0, . . . , s} such that the cones Ki and j,i K j
are on two sides of Γ; formally, there exists c ∈ Rn and i ∈ {0, 1, . . . , s} such that
Ñ
Ki ⊂ {y ∈ Rn | hc, yi 6 0} and j,i K j ⊂ {y ∈ Rn | hc, yi > 0}. 3
◦ Let y ∈ Rn . A set K ⊂ Rn is a cone with vertex y if it is expressible as y + K ′ for some
cone K ′ ⊂ Rn . In particular, any cone is a cone with vertex 0 ∈ Rn .
◦ Let Ω be a nonempty set in Rn . By aff Ω we denote the set of all affine combinations of
points in Ω. That is,
Õ
k
k
Õ
∗
θ i = 1, xi ∈ Ω for i = 1, . . . , k, and k ∈ N
aff Ω =
θ i xi
i=1
i=1
In other words, aff Ω is also the smallest affine set containing Ω. The relative interior
ri Ω of Ω denotes the interior of Ω relative to the affine space aff Ω.
3 More information on separability can be obtained in [Gül10]
14
P. PARUCHURI AND D. CHATTERJEE
◦ Let M be a convex set and x0 ∈ M. The union of all the rays emanating from x0 and
passing through points of M other than x0 is a convex cone with vertex at x0 . The closure
of this cone is called the supporting cone of M at x0 .
◦ Let K ⊂ Rn be a convex cone with vertex at x0 . By K ◦ we denote its polar (or dual)
cone defined by
⋆
(A.1)
K ◦ ≔ y ∈ Rn
hy, x − x0 i 6 0 for all x ∈ K .
It is clear that K ◦ is a closed convex cone with vertex at x0 in view of the fact that it is
an intersection of closed half-spaces:
Ù
⋆
K◦ =
z ∈ Rn
hz, y − x0 i 6 0 .
y ∈K
We adopt the contemporary convention of polarity as given in [Cla13, p. 21]. Our
polars are, therefore, negatives of the polars defined in [Bol75, p. 8]; consequently and
in particular, ψ0 in our Theorem B.6 is non-negative while ψ0 in [Bol75, Theorem 16] is
non-positive.
We need a few results from convex analysis, which we quote from various sources below
and for the sake of completeness we provide most of their proofs.
Theorem A.1 ([Bol75, Theorem 4 on p. 8]). Let K1, . . . Ks be closed convex cones in Rn
with vertex at x0 . Then
Ø
◦
Ù
s
s
◦
Ki .
Ki = conv
i=1
i=1
Here S denotes the closure of the set S.
Proof. Let K ≔
(A.2)
Ñs
i=1
Ki . If η ∈ K ◦ , then for every x ∈ K we have
hη, x − x0 i 6 0.
In particular, the relation (A.2) holds for x ∈ Ki for each i = 1, . . . , s. This implies that
η ∈ Ki ◦ for i = 1, . . . , s. Thus,
Ø
Ø
s
s
s
Ù
◦
◦
◦
η∈
Ki ⊂ conv
Ki ⊂ conv
Ki .
This shows that
Ñs
i=1
i=1
Ki
◦
⊂ conv
Ðs
i=1
i=1
◦
i=1
Ki .
Now let us prove the converse inclusion. Let η ∈ conv
Ðs
vectors η1, . . . , ηk ∈ i=1
Ki ◦ such that
Since ηi ∈
Ðs
i=1
◦
η = η1 + · · · + ηk .
Ðs
i=1
Ki ◦ . Then there exist
Ki , for every x ∈ K we have hηi, x − x0 i 6 0 for i = 1, . . . , s. Thus,
hη, x − x0 i = hη1, x − x0 i + · · · + hηs, x − x0 i 6 0.
◦
◦
Ñs
Ñs
Therefore, conv i=1 Ki ◦ ⊂
i=1 Ki . Since the dual cone
i=1 Ki is a closed convex
◦
Ðs
Ñs
cone, the closure conv i=1
Ki ◦ is also a subset of i=1
Ki .
Ðs
Theorem A.2 ([Bol75, Theorem 5 on p. 8]). Let Ω1, . . . , Ωs be convex sets in Rn such that
Ñs
i=1 ri Ωi , . Then
Ñs
Ñs
Ω = i=1
(i)
Ω,
i=1
Ñs i
Ñsi
(ii) aff i=1 Ωi = i=1
aff Ωi ,
PMP UNDER FREQUENCY CONSTRAINTS
(iii) ri
Ñs
i=1
Ñs
Ωi = i=1
ri Ωi .
Proof. Let Ω ≔
(i)
15
Ñs
i=1
Ωi .
If x ∈ Ω, then there exists a sequence xk ∈ Ω such that xk −→ x. But,
x k ∈ Ω ⇔ x k ∈ Ωi
for i = 1, . . . , s.
This means that for each i = 1, . . . , s, there exists a sequence xk ∈ Ωi with xk −→ x,
implying that
i . This proves the condition (i).
Ñs x ∈ Ω
Ñs
(ii) If x ∈ aff i=1
Ωi , then there exist vectors x1, . . . , xk ∈ i=1
Ωi such that
k
Õ
θ j xj = x
with
j=1
k
Õ
θ j = 1.
j=1
Ñs
Since x j ∈ i=1
Ωi if and only if x j ∈ Ωi for each i = 1, . . . , s, we have x ∈ aff Ωi for
each i = 1,
.
.
.
,
s.
Ñs
(iii) If x ∈ ri i=1
Ωi , then there exists an ǫ > 0 such that
Ù
Ù
s
s
Ωi ⊂
Bǫ (x) ∩ aff
Ωi .
i=1
i=1
In view of condition (ii), we have the following.
Ù
Ù
s
s
Bǫ (x) ∩
Ωi
aff Ωi ⊂
i=1
s
Ù
i=1
Ωi ⊂ Ωi
⇔
Bǫ (x) ∩ aff Ωi ⊂
⇔
x ∈ ri Ωi for each i = 1, . . . , s
s
Ù
x∈
ri Ωi .
⇔
i=1
for each i = 1, . . . , s
i=1
Theorem A.3 ([Bol75, Theorem 3 on p. 7]).
Ðs Let K 1, . . . , Ks be closed convex cones in
Rn with vertex at 0. If the cone K = conv i=1
Ks is not closed, then there are vectors
λ1 ∈ K1, . . . , λs ∈ Ks , not all of them zero, such that λ1 + · · · + λs = 0.
Proof. Let x ∈ K \ K. Then there exists a sequence of vectors xk ∈ K such that xk −→ x.
Since xk ∈ K, we can write
xk = xk(1) + · · · + xk(s)
with xk(i) ∈ Ki
Define αk ≔ maxi xk(i) . We may assume that αk > 0 for all k. It can be seen that
(i)
αk −→ ∞ since it would mean that x ∈ K otherwise. Let yk ≔
1
αk
(i)
xk .
Without loss of generality, we may assume that the limits y (i) = limk→∞ yk(i) exist for
i = 1, . . . , s. Since maxi yk(i) = 1 for each k, at least one of the vectors y (1), . . . , y (s) is not
zero. Moreover since Ki is closed, we have y (i) ∈ Ki .
16
P. PARUCHURI AND D. CHATTERJEE
Since limk→∞ xk = x and limk→∞ αk = ∞, we have
y (1) + · · · + y (s) = lim yk(1) + · · · + yk(s)
k→∞
1 (1)
(s)
= lim
xk + · · · + xk
k→∞ αk
1
xk = 0.
= lim
k→∞ αk
Theorem A.4 ([Bol75, Theorem 6 on p. 9]). If a family K1, . . . , Ks of convex cones with a
Ñs
common vertex at x0 is not separable, then i=1
ri Ki , .
Proof. Suppose that
Ñs
i=1
ri Ki = and let m < s be a positive number such that
m
Ù
ri Ki , and
m+1
Ù
ri Ki = .
Ñm
ri Ki = ri
Ki . This implies that
By Theorem A.2 (condition (iii)), i=1
Ù
Ù
m+1
m
m
Ù
ri Ki =
Ki ∩ ri Km+1 = .
ri Ki ∩ ri Km+1 = ri
i=1
i=1
i=1
Ñm
i=1
Ñm
i=1
i=1
Therefore the convex cones Km+1 and i=1 Ki have non-empty
interior and hence are
Ñ
separable. This implies that the convex cones Km+1 and i,m Ki are also separable, which
contradicts the assumption that the family of cones K1, . . . , Ks is not separable.
Theorem A.5 ([Bol75, Theorem 2 on p. 6]). Let s ∈ N∗ and K0, K1, . . . , Ks be a family
of convex cones in Rn with a common vertex x0 . This family is separable if and only if
there exist λi ∈ Ki ◦ for each i = 0, . . . , s, not all zero, that satisfy the condition
(A.3)
λ0 + · · · + λs = 0.
Proof. Let m (6 s) be the least number such that the family of cones K0, . . . , Km is
separable. Renumbering the cones if necessary, let us assume that the cones K0 and
K1 ∩ . . . ∩ Km are separable. This implies that there exists a hyperplane Γ characterised by
a non-zero vector η such that the cones lie in half-spaces given by
H = {x | hη, x − x0 i 6 0},
H ′ = {x | hη, x − x0 i > 0}.
If m = 1, then K0 ⊂ H and K1 ⊂ H ′ . This implies η ∈ K0 ◦ and −η ∈ K1 ◦ . Thus, choosing
the vectors as
λ0 = η, λ1 = −η, λ2 = · · · = λs = 0,
the required condition (A.3) is satisfied. If m > 1, the family of cones K1, . . . , Km is not
Ñm
separable. By Theorem A.4 we have i=1
ri Ki , . By Theorem A.2-(i),
m
Ù
Ki =
i=1
m
Ù
Ki.
i=1
Ñm
Ñm
Since i=1
Ki lies in the closed half-space H ′, its closure i=1
Ki ⊂ H ′ . Therefore,
Ñ
Ñm
◦
m
′
i=1 K i . By Theorem A.1,
i=1 K i ⊂ H , which implies that −η ∈
Ø
Ø
m
m
◦
◦
−η ∈ conv
K i = conv
Ki .
i=1
i=1
PMP UNDER FREQUENCY CONSTRAINTS
17
Ðm
Ðm
If, on the one hand, conv i=1
Ki ◦ is closed, then −η ∈ conv i=1
Ki ◦ , implying that
there exist vectors λi ∈ Ki ◦ for i = 1, . . . , m such that
−η = λ1 + · · · + λm .
Choosing λ0 = η (, 0) ∈ K0 and λ j = 0 for j = m + 1, . . . , s, the required condition (A.3)
Ðm
is satisfied. If, on the other hand, conv i=1
Ki ◦ is not closed, then by Theorem A.3, there
◦
exist vectors λi ∈ Ki for i = 1, . . . , m, not all zero, such that
◦
λ1 + · · · + λm = 0.
Selecting λ0 = λm+1 = · · · = λs = 0 we verify that the condition (A.3) is satisfied.
Conversely, assume that there exist λi ∈ Ki ◦ for i = 0, . . . , s satisfying (A.3) and not all
of them equal to zero (say λ0 , 0). Since λ0 ∈ K0 ◦ , we have hλ0, x − x0 i 6 0 for x ∈ K0 .
This means that K0 is contained in the half-space H = x hλ0, x − x0 i 6 0 . By (A.3),
Ñs
⇒
λ0 = −λ1 − · · · − λs
hλ0, x − x0 i = − hλ1, x − x0 i − · · · − hλs, x − x0 i .
For x ∈ i=1 Ki , we have x ∈ Ki for i = 1, . . . , s. Since λi ∈ Ki ◦ , hλi, x −
Ñxs 0 i 6 0 for each
i = 1, . . . , s. Hence,
hλ
,
x
−
x
i
>
0.
This
implies
that
the
intersection
0
0
i=1 Ki lies in the
half-space H ′ = x hλ0, x − x0 i > 0 . In other words, the family of cones K0, . . . , Ks
is separable.
Theorem A.6 ([Bol75, Theorem 7 on p. 10]). Let s ∈ N∗ , and for each i = 1, . . . , s let
Li ⊂ Rn be a subspace satisfying L1 + · · · + Ls = Rn . For each i = 1, . . . , s let Li∆ denote
the direct sum of all subspaces L1, . . . , Ls except Li , and Ki be a convex cone in Li with a
common vertex x0 ∈ Rn . If Ni ≔ conv Ki ∪ Li∆ for each i, then Ni is a convex cone, and
the family {Ni | i = 1, . . . s} is inseparable in Rn .
Proof. Suppose that the family {Ni | i = 1, . . . , s} is separable Ñ
and let (after renumbering
s
if necessary) N1 is separated in Rn from the intersection Π ≔ i=2
Ni by the hyperplane
Γ characterised by x ha, x − x0 i = 0 . That is,
N1 ⊂ H = x ha, x − x0 i 6 0 , Π ⊂ H ′ = x ha, x − x0 i > 0
Since L1 ⊂ L ∆j ⊂ N j for all j , 1, L1 ⊂ Π ⊂ H ′ . Since K1 ⊂ N1 ⊂ H, we have L1 ⊂ Γ.
For j = 2, . . . , s, L j ⊂ L1∆ ⊂ N∗1 ⊂ H and K j ⊂ Π ⊂ H ′ and this implies that L j ⊂ Γ.
We see that Li ⊂ Γ for all i = 1, . . . , s. This leads to an obvious contradiction as the span
of subspaces
contained in a hyperplane Γ is required to be the full space Rn . Hence, the
family Ni i = 1, . . . , s is not separable in Rn .
Appendix B. Facts about Tents
In this section an outline of the method of tents is provided.
Definition B.1. Let Ω be a subset of Rn and let x0 ∈ Ω. A convex cone Q ⊂ Rn with vertex
x0 is a tent of Ω at x0 if there exists a smooth map ρ defined in a neighbourhood of x0 such
that:4
(1) ρ(x) = x + o(x − x0 ),5 and
4The theory also works for ρ continuous.
5 Recall the Landau notation ϕ(x) = o(x) that stands for a function ϕ(0) = 0 and lim x→0
|ϕ(x)|
|x|
= 0.
18
P. PARUCHURI AND D. CHATTERJEE
(2) there exists ǫ > 0 such that ρ(x) ∈ Ω for x ∈ Q ∩ Bǫ (x0 ).
We say that a convex cone K ⊂ Rn with vertex at x0 is a local tent of Ω at x0 if, for every
x ∈ ri K, there is a convex cone Q ⊂ K with vertex at x0 such that Q is a tent of Ω at x0 ,
x ∈ ri Q, and aff Q = aff K. Observe that if K is a tent of Ω at x0 , then K is a local tent of
Ω at x0 .
We need the following theorems on tents in the formulation of our PMP in the sequel.
Theorem B.1 ([Bol75, Theorem 8 on p. 11]). Let Ω be a smooth manifold in Rn and K
the tangent plane to Ω at x0 ∈ Ω. Then K is a tent of Ω at x0 .
Theorem B.2 ([Bol75, Theorem 9 on p. 12]). Given a smooth function ϕ : Rn −→ R, let
∂
x0 be such that ∂x
ϕ(x0 ) , 0. Define sets Ω, Ω0 ∈ Rn as
Ω ≔ x ∈ Rn ϕ(x) 6 ϕ(x0 ) , Ω0 ≔ x0 ∪ x ∈ Rn ϕ(x) < ϕ(x0 ) .
Then the half-space K given by the inequality
and Ω0 at x0 .
∂
∂x ϕ(x0 ),
x − x0 6 0 is a tent of both Ω
Theorem B.3 ([Bol75, Theorem 10 on p. 12]). Let Ω ∈ Rn be a convex set and let K be
its supporting cone at x0 ∈ Ω. Then K is a local tent of Ω at x0 .
Proof. Let x ∈ ri K, x , x0 . By definition of supporting cone, there exists x ′ ∈ Ω such
that x lies on the ray emanating from x0 and passing through x ′ . Since x ∈ ri K, we also
have that x ′ ∈ ri Ω. Consider a small ball Bδ (x ′ ) choosing around x ′ choosing δ such that
x0 < Bδ (x ′ ) and Bδ (x ′) ∩ aff K ⊂ Ω. Consider a cone Q consisting of rays emanating from
x0 and passing through points in Bδ (x ′) ∩ K. Since Ω is a convex set and the points in
Bδ (x ′ ) ∩ K lie in Ω, there exists an ǫ > 0 such that Bǫ (x0 ) ∩ Q ⊂ Ω. It can be seen that
Q is a tent of Ω at x0 (the tent map can be considered to be the identity map). It is clear
that x ∈ ri Q and aff Q = aff K. Therefore, for every x ∈ ri K, there is a tent Q of Ω with
vertex at x0 containing x in its interior and satisfying aff Q = aff K, indicating that K is a
local tent of Ω at x0 .
Theorem B.4 ([Bol75, Theorem 12 on p. 14]). Let Ω0, Ω1, . . . , Ωs be subsets of Rn with
a common point x0 , and K0, K1, . . . , Ks local tents of these sets at x0 . If the family of cones
K0, K1, . . . , Ks is inseparable and at least one of the cones is not a plane, then there exists
x ′ ∈ Ω0 ∩ Ω1 ∩ . . . ∩ Ωs and x ′ , x0 .
Proposition B.5. A function ϕ(x) considered on the set Σ = Ω1 ∩ . . . ∩ Ωs , attains its
minimum at x0 if and only if
Ω 0 ∩ Ω 1 ∩ . . . ∩ Ω s = x0 ,
where Ω0 ≔ x0 ∪ x ∈ Rn ϕ(x) < ϕ(x0 ) .
Proof. Suppose that there exists a point x ′ ∈ Ω0 ∩Σ, x ′ , x0 . Since x ′ ∈ Ω0 , ϕ(x ′ ) < ϕ(x0 ).
But since x ′ ∈ Σ, x0 is not a minimum point of ϕ(x) on Σ. If x0 is not a minimum of ϕ(x)
on Σ, then there exists a point x ′ ∈ Σ satisfying ϕ(x ′ ) < ϕ(x0 ). This implies that x ′ ∈ Ω0
and the intersection Ω0 ∩ Σ , x0 .
Theorem B.6 ([Bol75, Theorem 16 on p. 20]). Let Ω1, . . Ñ
. , Ωs be subsets of Rn and let
Ñ
s
s
k=1 Ωk ∋ x 7−→ ϕ(x) ∈ R be a smooth function. Let Σ =
k=1 Ωk , let x0 ∈ Σ, and let Ki
PMP UNDER FREQUENCY CONSTRAINTS
19
be a local tent of Ωi at x0 for i = 1, . . . , s. If ϕ attains its minimum relative to Σ at x0 , then
there exist vectors λi ∈ Ki ◦ for i = 1, . . . , s and ψ0 ∈ R satisfying
∂
ϕ(x0 ) + λ1 + · · · + λs = 0
∂x
such that ψ0 > 0, and if ψ0 = 0, then at least one of the vectors λ1, . . . , λs is not zero.
ψ0
∂
ϕ(x0 ) = 0, choosing ψ0 = 1 and λ1 = · · · = λs = 0 will satisfy the given. We
Proof. If ∂x
∂
assume, therefore, that ∂x
ϕ(x0 ) , 0. Consider the set Ω0 ≔ x0 ∪ x ∈ Rn ϕ(x) <
∂
ϕ(x0 ), x − x0 6 0.
ϕ(x0 ) , and let K0 be the half-space in Rn defined by the inequality ∂x
By Theorem B.2 the set K0 is a tent of Ω0 at x0 . Since at least one of the tents K0, . . . , Ks
is not a plane by assumption, Proposition B.5 asserts that if x0 is a minimum of ϕ relative
to Σ, then Ω0 ∩ . . . ∩ Ωs is the singleton set x0 . By Theorem B.4 the tents K0, . . . , Ks
are separable, since otherwise the intersection Ω0 ∩ . . . ∩ Ωs would consist a point x ′ , x0 .
Theorem A.5 now asserts that there exist vectors λ0 ∈ K0 ◦, . . . , λs ∈ Ks ◦ , not all zero, such
that
λ0 + · · · + λs = 0.
∂
ϕ(x0 ) and ψ0 > 0.
The condition follows by noting that λ0 = ψ0 ∂x
Appendix C. Proof of Main Result
§C.1. Version 1. We convert the optimal control problem (2.14) into a relative extremum
problem in a suitable higher-dimensional space. To that end, we define a generic variable
z
}|
{
d
y ≔ (ξ0, . . . , ξT , µ0, . . . , µT −1 ) ∈ R × . . . × Rd × Rm × . . . × Rm,
|
{z
}
T +1 factors
(C.1)
T factors
and let n ≔ d(T + 1) + mT for the rest of this section. We further compress the vector
on the right hand side of (C.1) by writing y ≔ Ξ, M for Ξ ≔ (ξ0, . . . , ξT ) and M ≔
(µ0, . . . , µT −1 ). First, we define the standard projection maps from y ∈ Rn to the individual
factors ξt ∈ Rd and µt ∈ Rm in the following way:
(
πtx (y) ≔ ξt for t = 0, . . . , T,
(C.2)
πtu (y) ≔ µt for t = 0, . . . , T − 1.
In terms of the notations in (C.1) and (C.2), we lift the objective function in (2.14) to a
performance index of the joint variables
Rn ∋ (ξ0, . . . , ξT , µ0, . . . , µT −1 ) ≕ z 7−→
(C.3)
C(z) ≔
T
Õ
t=0
ct (ξt , µt ) =
T
−1
Õ
t=0
ct πtx (z), πtu (z) .
Second, we define constraint sets Ωxt , Ωut ⊂ Rn such that if y = Ξ, M ∈ Ωxt ∩ Ωut in the
notation of (C.1), then the t-th factor ξt of Ξ is constrained to the set St and the t-th factor
µt of M is constrained to the set Ut ; to wit,
for t = 0, . . . , T,
Ωxt ≔ y ∈ Rn πtx (y) ∈ St
(C.4)
u
n
u
for t = 0, . . . , T − 1.
Ωt ≔ y ∈ R πt (y) ∈ Ut
20
P. PARUCHURI AND D. CHATTERJEE
Observe that for y ∈ Ωxs , the coordinates πtx (y) for t , s and all the πτu (y) are arbitrary.
Similarly, for y ∈ Ωus , all the coordinates πtx (y) and πτu (y) for τ , s are arbitrary. We say
that Ωxt and Ωut are lifts of St and Ut , respectively.
Third, we define maps gt : Rn −→ Rd for t = 0, . . . , T − 1, to lift the dynamics of the
system (2.1) to Rn in the following way:
z
}|
{
d
R × . . . × Rd × Rm × . . . × Rm ∋ (ξ0, . . . ξT , µ0, . . . µT −1 ) ≕ y 7−→
|
{z
}
T +1 factors
(C.5)
x
(y) ∈ Rd .
gt (y) ≔ ft (ξt , µt ) − ξt+1 = ft πtx (y), πtu (y) − πt+1
T factors
By definition, therefore, a vector y = (ξ0, . . . , ξT , µ0, . . . , µT −1 ) ∈ Rn satisfies gt (y) = 0
for all t = 0, . . . , T − 1, if and only if ξt+1 = ft (ξt , µt ) for all t = 0, . . . , T − 1. We define a
family of sets
(C.6)
Ωft ≔ y ∈ Rn gt (y) = 0
for t = 0, . . . , T − 1.
Finally, we define the lift of the frequency constraints on the control trajectories:
z
}|
{
Rd × . . . × Rd × Rm × . . . × Rm ∋ (ξ0, . . . ξT , µ0, . . . µT −1 ) ≕ y 7−→
|
{z
}
T +1 factors
(C.7)
b ≔ F π u (y), . . . , π u (y) = F(µ0, . . . , µT −1 ).
F(y)
0
T −1
T factors
We define a process z to be the concatenation of a control trajectory (u0, . . . , uT ) and its
corresponding state trajectory (x0, . . . , xT ) traced by the system according to (2.1) as
z ≔ (x0, . . . , xT , u0, . . . , uT ).
cu defined by
A process z satisfying the frequency constraints belongs to the set Ω
cu ≔ y ∈ Rn F(y)
b =0 .
(C.8)
Ω
Employing the lifts and the notations introduced in (C.3), (C.4), (C.6), and (C.8), we
state the optimal control problem (2.14) equivalently as the following relative extremum
problem:
minimize
n
C(z)
(
z ∈ Σ,
subject to
ÑT −1 f
ÑT
ÑT −1 u cu
x
Σ≔
t=0 Ωt ∩
t=0 Ωt ∩
t=0 Ωt ∩ Ω .
z ∈R
(C.9)
In the sequel z∗ will denote a solution of the relative extremum problem (C.9), comprising
−1 that solves (2.14) and the resulting optimal state
of the optimal control trajectory (ut∗)Tt=0
∗
T
trajectory (xt )t=0 .
Define
ΩC (z∗ ) ≔ {z∗ } ∪ z ∈ Rn C(z) < C(z∗ ) .
By Proposition B.5, z∗ solves (C.9) if and only if Σ ∩ ΩC (z∗ ) = {z∗ }. Let
cu at z∗, respectively.
cu be tents of the sets Ωft , Ωxt , Ωut , and Ω
Qft , Qxt , Qut , and Q
PMP UNDER FREQUENCY CONSTRAINTS
21
cu depend on z∗ , of course, but for notational simplicity we do
(The sets Qft , Qxt , Qut , and Q
not explicitly depict the dependence of these sets on z∗ in what follows.) By Theorem B.2
the half-space given by
∂
C
n
∗
∗
(C.10)
Q ≔ z∈R
C(z ), z − z 6 0
∂z
is a tent of ΩC (z∗ ) at z∗ .
cu ∪ Qft
Proposition C.1. The family of tents QC, Q
T −1
T −1 x T
∪ Qt t=0 ∪ Qut t=0 is separable.
t=0
Proof. The assertion follows from Proposition B.5 and Theorem B.4. Indeed, since the tent
cu ∪ Qft T −1 ∪ Qxt T ∪
QC is a half-space, (and therefore, not a plane,) the family QC, Q
t=0
t=0
u T −1
Qt t=0 of tents satisfies the hypothesis of Theorem B.4. If the family is not separable,
then the intersection Σ ∩ ΩC (z∗ ) contains a point z ′ different from z∗ . This means that
Σ ∩ ΩC (z∗ ) , {z∗ }. But then, this contradicts optimality of z∗ (cf. Proposition B.5).
Proposition C.2. There exist vectors
◦
◦ λC ∈ QC ,
◦
◦ λtf ∈ Qft for t = 0, . . . , T − 1,
◦
x
x
◦ λt ∈ Qt for t = 0, . . . , T ,
◦
◦ λtu ∈ Qut for t = 0, . . . , T − 1, and
cu ◦ ,
◦ λbu ∈ Q
not all zero, such that
(C.11)
λC +
T
−1
Õ
t=0
λtf +
T
Õ
λtx +
t=0
T
−1
Õ
t=0
λtu + λbu = 0.
cu ∪ Qft T −1 ∪ Qxt T ∪ Qut T −1 is separable
Proof. Since the family of cones QC, Q
t=0
t=0
t=0
in view of Proposition C.1, by Theorem A.5 there exist vectors in the dual cones of each of
u T −1
C
f T −1 x T
cu ∪ Qt
∪ Qt t=0 ∪ Qt t=0 that satisfy (C.11).
the cones in Q , Q
t=0
We observe that since z ∈ QC satisfies the inequality
∂
∗
∗
C(z ), z − z 6 0,
∂z
◦
in view of (C.10), every vector in the dual cone QC is of the form
λC = ηC
∂
C(z∗ ),
∂z
where ηC > 0.
Proposition C.3. If z∗ is a solution of the relative extremum problem (C.9), there exist
ηC > 0 and dual vectors
◦
◦ λtf ∈ Qft for t = 0, . . . , T − 1,
◦
◦ λtx ∈ Qxt for t = 0, . . . , T ,
◦
◦ λtu ∈ Qut for t = 0, . . . , T − 1, and
cu ◦ ,
◦ λbu ∈ Q
22
P. PARUCHURI AND D. CHATTERJEE
such that
Õ
Õ
Õ
∂
C(z∗ ) +
λtf +
λtx +
λtu + λbu = 0.
∂z
t=0
t=0
t=0
T −1
C
In particular, if η = 0, then at least one of the vectors λtf t=0 , λtx
not zero.
T −1
ηC
(C.12)
T
T
T
,
t=0
λtu
T −1 u
, λb
t=0
Proof. Follows at once from the arguments in the proof of Theorem B.6.
T
T −1
Proposition C.4. The family of tents Qxt t=0 ∪ Qut t=0 is not separable.
is
Proof. Define the subspaces Lxs , s = 0, . . . , T , and Lus , s = 0, . . . , T − 1, as:
)
(
πtx (z) = 0 for t ∈ {0, . . . , T } \ {s}, and
n
,
Lxs ≔ z ∈ R
πtu (z) = 0 for t = 0, . . . , T − 1
)
(
πtx (z) = 0 for t = 0, . . . , T, and
n
.
Lus ≔ z ∈ R
πtu (z) = 0 for t ∈ {0, . . . , T − 1} \ {s}
Observe that Lx0 + · · · + LxT + Lu0 + · · · + LuT −1 = Rn . Consider the subspaces Lx∆t and
Lu∆t of Rn defined by:
Lx∆t ≔ z ∈ Rn πtx (z) = 0 for t = 0, . . . , T,
Lu∆t ≔ z ∈ Rn πtu (z) = 0 for t = 0, . . . , T − 1.
Lx∆t
z
= Rd × . . . × Rd ×
{ z
}|
{
d
× R . . . × R d × Rm × . . . × . . . × Rm .
T factors
(t+1)-th factor
z
}|
{ z
d
= R × . . . × R d × Rm × . . . × Rm ×
T +1 factors
Lu∆t
}|
{0}
|{z}
T +1 factors
}|
{0}
|{z}
T factors
(t+1)-th factor
{
× Rm × . . . × Rm .
Let qtx (xt∗ ) be a local tent of S t at xt∗ and let qtu (ut∗ ) be a local tent of Ut at ut∗ . Observe
that the inclusions qtx ⊂ πtx Lxt for t = 0, . . . , T , and qtu ⊂ πtu Lut for t = 0, . . . , T − 1,
hold.
We now construct a family of tents Qxt and Qut in the following way:
Qxt ≔ z ∈ Rn πtx (z) ∈ qtx (πtx (z∗ )) for t = 0, . . . , T,
(C.13)
Qut ≔ z ∈ Rn πtu (z) ∈ qtu (πtu (z∗ )) for t = 0, . . . , T − 1.
Let us lift the tents qtx (xt∗ ) as follows:
z
qetx (xt∗ ) ≔ {0} × . . . × {0} ×
}|
T +1 factors
qtx (z∗ )
|{z}
(t+1)-th factor
Observe that qetx (xt∗ ) ∈ Lxt . Therefore,
{ z
}|
{
× {0} × . . . × {0} × {0} × . . . × {0} × . . . × {0}
Qxt = qetx (xt∗ ) ∪ Lx∆t
T factors
PMP UNDER FREQUENCY CONSTRAINTS
23
and similarly for Qut . Since qtx and qtu are convex cones, it follows that the tents Qxt and Qut
satisfy the hypothesis of Theorem A.6.
T
T −1
Theorem A.6 asserts that the family of tents Qxt t=0 ∪ Qut t=0 is inseparable, and this
establishes the claim.
Remark C.1. Note that any sub-family of an inseparable family of cones is also inseparable.
T −1
T
Thus, in addition to the family of tents Qxt t=0 ∪ Qut t=0 being inseparable, we have that
x T u T −1
the families Qt t=0 , Qt t=0 are both individually inseparable.
The following proposition constitutes the keystone of our proof of the main Theorem
3.1.
Proposition C.5. If z∗ is an optimal process of (C.9), then there exist ηC > 0 and dual
vectors
◦
◦ λtf ∈ Qft for t = 0, . . . , T − 1,
◦
◦ λtx ∈ Qxt for t = 0, . . . , T , and
cu ◦ ,
◦ λbu ∈ Q
*
+
T
−1
T
Õ
Õ
∂
∗
f
x
u
b
(C.14)
−η
C(z ) +
λt −
λt − λ , z̃ 6 0,
∂z
t=0
t=0
Ñ −1 u
for every vector z̃ such that z∗ + z̃ ∈ Tt=0
Qt . In particular, if ηC = 0, then at least one
f T −1
of λt t=0 and λbu is not zero.
such that
C
Proof. From Proposition C.3 we infer that there exist ηC > 0 and dual vectors
◦
◦ λtf ∈ Qft for t = 0, . . . , T − 1,
◦
◦ λtx ∈ Qxt for t = 0, . . . , T ,
◦
◦ λtu ∈ Qut for t = 0, . . . , T − 1, and
cu ◦ ,
◦ λbu ∈ Q
satisfying (C.12), such that if ηC = 0, then at least one of the vectors
f T −1 x T u T −1
λt t=0 , λt t=0, λt t=0 , λbu
T −1
T
is not zero. From Proposition C.4 we know that the family of cones Qxt t=0 ∪ Qut t=0
T −1
is inseparable. Observe that if ηC = 0 and all of λtf t=0 and λbu are zero, then the vectors
u T u T −1
T
T
λt t=0, λt t=0 in the dual cones of the family Qxt t=0 ∪ Qut t=0 , not all zero, satisfy
T
Õ
t=0
λtx
+
T
−1
Õ
t=0
λtu = 0.
In view of Theorem A.5, this contradicts the fact that the family of cones Qxt
is inseparable. This establishes the final assertion.
T
T −1
∪ Qut t=0
t=0
We now establish the main assertion. From (C.12) we have, for any z̃ ∈ Rn ,
*
+
T
−1
T
T
−1
Õ
Õ
Õ
C ∂
∗
f
x
u
b
η
C(z ) +
λt +
λtu, z̃ .
λt + λ , z̃ = −
∂z
t=0
t=0
t=0
24
P. PARUCHURI AND D. CHATTERJEE
Ñ −1 u
If z̃ is a vector such that z∗ + z̃ ∈ Tt=0
Qt , then z∗ + z̃ ∈ Qut for each t = 0, . . . , T − 1. Since
◦
u
u
u
λt ∈ Qt , by definition we have λt , z̃ 6 0 for each t = 0, . . . , T − 1, leading to
+
T
−1
T
T
−1
Õ
Õ
Õ
∂
∗
f
x
C(z ) +
λtu, z̃ > 0
λt +
λt + λbu, z̃ = −
η
∂z
t=0
t=0
t=0
*
+
T
−1
T
Õ
Õ x
∂
−ηC C(z∗ ) +
−λtf −
λt − λbu, z̃ 6 0
∂z
t=0
t=0
*
⇒
C
◦
◦
Observe that the dual cones Qft are subspaces and hence, if λtf ∈ Qft , then λetf ≔ −λtf ∈
◦
Qft . And this proves the proposition.
Before we delve into the final result that helps us prove Theorem 3.1, we make some
observations on the dual vectors and the gradient matrices. We have the following charac T −1 T
terisation of the dual vectors λtf t=0 , λtx t=0 and λbu .
• By the construction in (C.13), for z∗ + z̃ ∈ Qxt , the coordinates πsx ( z̃) for s , t are arbitrary
and πτu ( z̃) are arbitrary. The coordinates πtx ( z̃) lie in the cone qtx (z∗ ). Since a dual vector
◦
λtx ∈ Qxt has to satisfy (A.1), for all z∗ + z̃ ∈ Qxt
λtx, z̃ 6 0
But since πsx ( z̃) for s , t and πτu ( z̃) for τ = 0, . . . , T − 1 are arbitrary, it can be seen
that the corresponding coordinates in λtx are zeroes, that is, πsx (λtx ) = 0 for s , t and
πτu (λtx ) = 0 for τ = 0, . . . , T − 1.
Since πtx ( z̃) ∈ q tx (xt∗ ), the corresponding coordinate of λtx , which is πtx (λtx ) lies in the
◦
dual cone qtx (xt∗ ) which we denote by ηtx .
λtx = (0, . . . , ηtx, . . . , 0, 0, . . . , 0)
• By Theorem B.1, the tangent plane of Ωft at z∗ is a tent of Ωft at z∗ . Considering Qft to be
◦
the tangent plane, every vector in the corresponding dual cone Qft is of the form
λtf
⊤
∂
∗
=
gt (z ) ηtf ,
∂z
for t = 0, . . . , T − 1,
where ηtf ∈ Rd .
cu at z∗ is a tent of Ω
cu at z∗ . Considering Q
cu to be the
• Similarly, the tangent plane of Ω
◦
u
c
tanget plane, every vector in the dual cone Q is of the form
where ηbu ∈ Rℓ .
λbu =
⊤
∂
F(z∗ ) ηbu,
∂z
PMP UNDER FREQUENCY CONSTRAINTS
25
From (C.3), (C.5), (C.7), we obtain the components of λC, λtf, λbu as follows:
∂
∂
C(z∗ ) =
ct xt∗, ut∗ ,
∂ξt
∂ξ
∂
∂
∗
∗ ∗
∂ξt gt (z ) = ∂ξ ft xt , ut ,
∂
F(z∗ ) = 0,
∂ξ
t
∂
∂
∗
∗ ∗
∂ µ gt (z ) = ∂ µ ft xt , ut ,
t
t
(C.15)
∂
∂
C(z∗ ) =
ct xt∗, ut∗ ,
∂ µt
∂µ
∂
∗
gt+1 (z ) = −Id,
∂ξt
∂
et ,
F(z∗ ) = F
∂ µt
for t = 0, . . . , T − 1, and Id being the d × d identity matrix.
−1 is an optimal process of the optimal control
Proposition C.6. If z∗ ≔ (xt∗ )Tt=0, (ut∗ )Tt=0
problem (2.14), then there exist ηC > 0 and dual vectors
⋆
◦ ηtf ∈ Rd
for t = 0, . . . , T − 1,
◦
◦ ηtx ∈ qtx (xt∗ )
for t = 0, . . . , T ,
⋆
◦ ηbu ∈ Rℓ ,
such that,
(i)
−ηC
∂
∂
f
ct (xt∗, ut∗ ) +
ft (xt∗, ut∗ )⊤ ηtf − ηt−1
− ηtx = 0
∂ξ
∂ξ
for t = 1, . . . , T − 1;
(ii) while η0f , ηTf −1 satisfy
ηTf −1 = −ηTx ,
− ηC
∂
∂
c0 (x0∗, u0∗ ) +
f0 (x0∗, u0∗ )⊤ η0f − η0x = 0;
∂ξ
∂ξ
and
∂
∂
et⊤ ηbu, ũt 6 0
ct (xt∗, ut∗ ) +
ft (xt∗, ut∗ )⊤ ηtf − F
∂µ
∂µ
∗
for all vectors ũt such that ut + ũt ∈ qtu (ut∗ ), for t = 0, . . . , T − 1.
T −1
In particular, if ηC = 0, then at least one of ηtf t=0 and ηbu is not zero.
(iii)
−ηC
T −1
Proof. By Proposition C.5, there exist vectors ηtf t=0 , ηbu and ηC ∈ R, not all zero satistying
(C.14). By the construction
in (C.13), for z∗ + z̃ ∈ Qut , the coordinates πtx ( z̃) are arbitrary.
So we choose t (∈ 0, . . . , T ) and a z̃ such that πtx ( z̃) ∈ Rd is arbitrary and
πsx ( z̃) = 0 for s = 0, . . . , T,
πτu ( z̃)
=0
s , t,
for τ = 0, . . . , T − 1.
Let x̃t ≔ πtx ( z̃). When we use this particular collection of z̃ in (C.14), only the ξt coordinates
in the dual vectors will survive. And the remaining equation is,
*
+
T
−1
T
Õ
Õ
∂
∂
C ∂
∗
∗
∗
x
−η
C(z ) +
gt (z ) −
F(z ), x̃t 6 0.
λt −
∂ξt
∂ξt
∂ξt
s=0
s=0
26
P. PARUCHURI AND D. CHATTERJEE
Using the fact that x̃t can be positive or negative and the results in (C.15), we have the
following condition for each t = 1, . . . , T − 1:
⊤
T −1
T
−1
Õ
∂ Õ
∂
− ηC
cs (z∗ ) +
gs (z∗ ) ηsf − ηtx = 0
∂ξt s=0
∂ξt
s=0
(C.16)
∂
∂
f
ct (xt∗, ut∗ ) +
ft (xt∗, ut∗ ) − ηt−1
− ηtx = 0
⇒ − ηC
∂ξt
∂ξt
Using x̃T −1 of the same construction, we get the equation ηTf −1 + ηTx = 0 and using x̃0 , we
get,
∂
∂
−ηC
c0 (x0∗, u0∗ ) +
f0 (x0∗, u0∗ ) − η0x = 0
∂ξ0
∂ξ0
This proves the first condition.
If we take z̃ such that its coordinates πsx ( z̃) are all zero and πτu ( z̃) are zero for τ =
0, . . . , T − 1, τ , t. And πtu ( z̃) ≕ ũt is such that ut∗ + ũt ∈ qtu (ut∗ ). It is easy to see that the
Ñ −1 u
vector z̃ thus generated lies in the intersection Tt=0
Qt . Thus, using equation (C.14) and
by the construction, we have
∂
∗ ∗
∗ ∗ ⊤ f
⊤b
C ∂
u
e
ct (xt , ut ) +
ft (xt , ut ) ηt − Ft η , ũt 6 0.
(C.17)
−η
∂ µt
∂ µt
This procedure can be repeated with vectors for each t = 0, . . . , T − 1, and the assertion
follows.
We are finally ready for ready for the proof of our main result.
Proof of Theorem 3.1. Observe that from the definition of the Hamiltonian H η , ηc in (3.1),
we have
∂ η C, ηcu f
∂
∂
H
ft (xt∗, ut∗ )⊤ ηtf
ηt , t, xt∗, ut∗ = −ηC ct (xt∗, ut∗ ) +
∂ξ
∂ξ
∂ξ
for t = 0, . . . , T − 1, and
(C.18)
C c
∂
∂
∂
u
et⊤ ηbu .
ft (xt∗, ut∗ )⊤ ηtf − F
H η , η ηtf , t, xt∗, ut∗ = −ηC ct (xt∗, ut∗ ) +
∂µ
∂µ
∂µ
C
u
The conditions of non-negativity (PMP-i), non-triviality (PMP-ii) follow from the statement
of the Proposition C.6. From (C.18) and (C.16), we get the adjoint dynamics in (PMP-iii).
The transversality conditions follow from Proposition C.6 (ii). The equation (C.17) readily
provides the Hamiltonian maximisation condition (PMP-v).
§ C.2. Alternate Proof. This section provides an alternate approach, suggested to us by
Navin Khaneja, to establish Theorem 3.1; we include it here for its scientific merit and for
completeness.
Alternate Proof of Theorem 3.1. Let us define an auxillary system with the dynamics
(C.19)
where wt ∈ Rℓ .
et ut
wt+1 = ht (wt , ut ) ≔ wt − F
for t = 0, . . . , T − 1,
PMP UNDER FREQUENCY CONSTRAINTS
27
Observe that the frequency constraints in (2.14), in view of (2.13), can now be viewed
as the terminal state constraint on the auxillary system (C.19) as
w0 = 0 and
(C.20)
wT = 0.
We can now rewrite the problem (2.14) into a standard optimal control problem with
constraints on control magnitude and states.
minimize
−1
(ut )Tt=0
(C.21)
subject to
T
−1
Õ
t=0
ct (xt , ut )
dynamics (2.1),
auxillary
dynamics (C.19),
xt ∈ St for t = 0, . . . , T,
w
0 = 0 and wT = 0,
ut ∈ Ut for t = 0, . . . , T − 1,
For the optimal control problem (C.21), using the usual PMP formulation, we can define
the Hamiltonian as
⋆
⋆
R × Rℓ × Rd × N × Rd × Rℓ × Rm ∋ (ν, ϑ, ζ, s, υ, ξ, µ) 7−→
(C.22)
Hν (ϑ, ζ, s, υ, ξ, µ) ≔ hϑ, hs (υ, µ)i + hζ, fs (ξ, µ)i − νcs (ξ, µ) ∈ R.
−1 is an optimal state-action
From the assertions of the usual PMP, if (wt∗ )Tt=0, (xt∗ )Tt=0, (ut∗ )Tt=0
trajectory of (C.21), then there exist
⋆
T −1
◦ a trajectory ηtf t=0 ⊂ Rd ,
⋆
T −1
◦ a trajectory ηth t=0 ⊂ Rℓ ,
⋆
T
◦ a sequence ηtx t=0 ⊂ Rd and
C
◦ η ∈R
satisfying the following conditions:
(N-i)
non-negativity condition
ηC > 0;
(N-ii) non-triviality condition
T −1
the state-adjoint trajectory ηtf t=0 , the auxillary state-adjoint trajectory
T −1
ηth t=0 and ηC do not simultaneously vanish;
(N-iii) state, auxillary state and adjoint system dynamics
∂ ηC h f
∗
xt+1
=
H ηt , ηt , t, wt∗, xt∗, ut∗
for t = 0, . . . , T − 1,
∂ζ
∂ ηC h f
∗
wt+1
=
for t = 0, . . . , T − 1
H ηt , ηt , t, wt∗, xt∗, ut∗
∂ϑ
∂ ηC h f
f
ηt−1
=
H ηt , ηt , t, wt∗, xt∗, ut∗ − ηtx for t = 1, . . . , T − 1,
∂ξ
∂ ηC h f
h
ηt−1
=
for t = 1, . . . , T − 1,
H ηt , ηt , t, wt∗, xt∗, ut∗
∂υ
where ηtx lies in the dual cone of a tent qtx (xt∗ ) of St at xt∗ ;
(N-iv) transversality conditions
∂ ηC h f
and
ηTf −1 = −ηTx ,
H η0, η0, 0, w0∗, x0∗, u0∗ − η0x = 0
∂ξ
28
P. PARUCHURI AND D. CHATTERJEE
w0∗ = 0
and
wT∗ = 0;
where η0x lies in the dual cone of a tent q0x (x0∗ ) of S0 at x0∗ and ηTx lies in the dual
cone of a tent qTx (xT∗ ) of ST at xT∗ ;
(N-v) Hamiltonian maximization condition, pointwise in time,
∂ ηC h f
∗ ∗ ∗
H ηt , ηt , t, wt , xt , ut , ũt 6 0 whenever ut∗ + ũt ∈ qtu (ut∗),
∂µ
where qtu (ut∗ ) is a local tent at ut∗ of the set Ut of admissible actions;
Observe that from the definition of Hamiltonian in (C.22) and from (C.19), the auxillary
state-adjoint dynamics reduces to (for t = 1, . . . , T − 1)
∂ ηC h f
h
ηt−1
=
H ηt , ηt , t, wt∗, xt∗, ut∗
∂υ
∂
h
∗ ∗
f
∗ ∗
C
∗ ∗
ηt , ht (wt , ut ) + ηt , ft (xt , ut ) − η ct (xt , ut )
=
∂υ
∂
ηh, ht (wt∗, ut∗ ) = ηth
=
∂υ t
T −1
and ηTh −1 can be chosen arbitrarily. This implies the trajectory ηth t=0 can be replaced by
a constant vector, say ηbu ∈ Rℓ . That is,
(C.23)
η0h = · · · = ηTh −1 ≕ ηbu .
Similarly, using the definition of the Hamiltonian, the condition (N-v) can be written as
∂ ηC h f
∗ ∗ ∗
H ηt , ηt , t, wt , xt , ut , ũt 6 0
∂µ
∂
ηth, ht (wt∗, ut∗ ) + ηtf, ft (xt∗, ut∗ ) − ηC ct (xt∗, ut∗ ) , ũt 6 0
⇔
∂µ
(C.24)
∂ D h ∗ e ∗E
⇔
ηt , wt + Ft ut + ηtf , ft (xt∗, ut∗ ) − ηC ct (xt∗, ut∗ ) , ũt 6 0
∂µ
D
E
∂
et ut∗ , ũt 6 0
ηtf , ft (xt∗, ut∗ ) − ηC ct (xt∗, ut∗ ) + ηbu, −F
⇔
∂µ
whenever ut∗ + ũt ∈ qtu .
Hence, defining a new Hamiltonian as in (3.1), the conditions (N-i) - (N-v) transform to
the conditions (PMP-i) - (PMP-vi) as shown below.
(i) The non-negativity condition (PMP-i) is same as the condition (N-i)
T −1
T −1
(ii) Since the non-triviality condition (N-ii) asserts that ηC, ηtf t=0 , ηth t=0 do not vanish
simultaneosly, the non-triviality condition (PMP-ii) follows from (C.23).
(iii) It can be observed from the way the Hamiltonian is defined in (3.1), the optimal state
and adjoint dynamics specified in (N-iii) is same as the one in (PMP-iii).
(iv) The transversality conditions in (PMP-iv) also follow from the definition of Hamiltonian in (3.1) and the conditions (N-iv) on states (x) and adjoint (ηtf ).
(v) From (C.24), we can see that (N-v) holds if and only if (PMP-v) holds.
(vi) The condition (PMP-vi) is another way of writing the transversality conditions on
auxillary states and auxillary state-adjoints in (N-iv). The equivalence follows directly
from the dynamics of auxillary states specified by (N-iii) and the equivalence of the
condition (PMP-vi) and the boundary conditions on auxillary states in (N-iv) as shown
in (C.20).
PMP UNDER FREQUENCY CONSTRAINTS
Our proof is now complete.
29
Appendix D. Proofs of Corollaries
Proof of Corollary 3.2. The conditions (AFF-i), (AFF-ii), (AFF-iii), (AFF-iv), and (AFF-vi)
follow directly from Theorem 3.1. The Hamiltonian maximization condition, pointwise in
time, (AFF-v) is proved as follows:
Since Ut is convex, by Theorem B.3, the supporting cone Kt of Ut at ut∗ is a local tent
of Ut at ut∗ . By (PMP-v), for every vector u˜t satisfying ut∗ + u˜t ∈ Kt , the optimal actions
ut∗ , optimal states xt∗ and the adjoint vectors ηtf satisfy
∂ η C, ηcu f
∗ ∗
H
ηt , t, xt , ut , ũt 6 0.
∂µ
Since the supporting cone Kt includes the set Ut , the directions u˜t satisfying ut∗ + u˜t ∈ Kt
include all the directions into the set Ut from ut∗ . This implies that at ut∗ , the directional
C u
∂
H η , ηc ηtf , t, xt∗, ut∗ is non-positive for every direction u˜t into the set Ut , which
derivative ∂µ
C u
is a necessary condition for optimality of H η , ηc ηtf , t, xt∗, µ at ut∗ . Note that since ct (ξ, ·)
is convex, we have
∂2
∂2
cu f
η C, η
∗ ∗
∗ ∗
H
ηt , t, xt , ut
ct (xt , ut )
=−
6 0.
∂ µi ∂ µ j
∂ µi ∂ µ j
i, j
i, j
C u
Thus, the function H η , ηc ηtf, t, xt∗, · : Ut −→ R is concave, and hence the necessary
C u
condition for optimality is also sufficient. The set Ut being compact, the function H η , ηc
attains its maximum.
◦
Proof
of Corollary 3.3. Observe that when St = Rd , the dual cone of qtx (xt∗ ), qtx (xt∗ ) =
◦
0 and when St is a singleton set, the dual cone qtx (xt∗ ) = Rd . Since St = Rd for
t = 1, . . . , T − 1, the vectors ηtx = 0 for t = 1, . . . , T − 1. Thus, the adjoint dynamics in
(AFF-iii) specialises to (3.6). Since S0 and ST are singleton sets, the vectors η0x and ηTx are
arbitrary and thus the transversality conditions in (AFF-iv) are trivially satisfied.
Appendix E. Proofs of LQ Propositions
Proof of Proposition 4.1. Since both the initial and the final states are fixed, from the
transversality conditions we see that η0f and ηTf −1 can be arbitrary. Suppose that the PMP
holds in abnormal form, i.e., ηC = 0. In this case the adjoint dynamics equation reduces to
the following.
f
ηt−1
= A⊤ ηtf
The adjoint variable
(E.1)
ηtf
for t = 1, . . . , T − 1
is given in terms of the ηTf −1 as
T −t−1 f
ηtf = A⊤
ηT −1 for t = 0, . . . , T − 1.
From the Hamiltonian maximization condition (which is uncosntrained optimization with
respect to control variable since there are no control action constraints), we obtain the
following conditions.
ηC Rut∗ = B⊤ ηtf for t = 0, . . . , T − 1.
30
P. PARUCHURI AND D. CHATTERJEE
Since ηC = 0, by assumption, it follows that
B⊤ ηtf = 0 for t = 0, . . . , T − 1,
T −t−1 f
⇒ B⊤ A⊤
ηT −1 = 0 for t = 0, . . . T − 1 in view of (E.1).
⊤
f
This implies that ηT −1 is in the null space of B AB . . . AT −1 B . But since the pair
(A, B) is controllable, the matrix B AB . . . AT −1 B has full column rank and thus,
its range space (image) is Rd . Since the range space (image) of a matrix C is orthogonal
⊤
to the kernel/null space of its transpose C ⊤ , the null space of B AB . . . AT −1 B is
T −1
just the zero vector. This means that ηTf −1 = 0. From (E.1), we see that ηtf t=0 is the zero
sequence. But this contradicts the non-triviality assertion of the PMP.
Proof of Proposition 4.2. If ηC = 0, then from (4.10) we have,
et⊤ ηbu for t = 0, . . . , T − 1, and
B⊤ ηtf = F
f
ηt−1
= A⊤ ηtf
for t = 0, . . . , T − 1.
This means ηtf = (A⊤)T −1−t ηTf −1 for t = 0, . . . , T − 1 and therefore,
et⊤ ηbu
B⊤ (A⊤)T −1−t ηTf −1 = F
Letting
for t = 0, . . . , T − 1.
e⊤
F
B⊤ (A⊤ )T −1
0 ª
©
©
ª
..
® ∈ RmT ×d and F
fT ≔
e ≔ .. ® ∈ RmT ×ℓ,
B
.
. ®
®
e⊤ ¬
B⊤
«F
¬
«
T −1
fT ηf
fT is the transpose of the reachability matrix and
eηbu . Note that B
we have B
=
F
T −1
fT = d.
e = F D−1 ⊤ . By assumption, rank B
F
If the equation
(E.2)
fT
B
η f
T −1 = 0
e
−F
ηbu
T −1
admits a non-trivial solution, then there exist ηC, ηbu and ηbu t=0 , not all zero, satisfying
−1 is independent
(4.10). Since when ηC = the optimal state-action trajectory (xt∗ )Tt , (ut∗ )Tt=0
T −1
of ηbu and ηtf t=0 , every feasible solution of (4.9) is an abnormal solution of PMP.
fT −F
e < d + ℓ.
The equation (E.2) admits a non-trivial solution only when rank B
fT −F
e = min{d + ℓ, mT }, there exist non-trivial solutions to (E.2) when
Since rank B
d + ℓ > mT . And when the rows of the reachability
matrix and the frequency constraints
fT −F
e = d + ℓ and there do not exist any nonmatrix F D−1 are independent, the rank B
trivial solutions to (E.2) and thus, all the optimal state-action trajectories are normal.
References
[Ber95]
[Bol75]
[Bol78]
[BT16]
D. P. Bertsekas. Dynamic Programming and Optimal Control, volume 1. Athena Scientific, 1995.
V. G. Bolt’yanskii. The method of tents in the theory of extremal problems. Russian Mathematical
Surveys, 30(3):1–54, 1975.
V. G. Boltyanskii. Optimal Control of Discrete Systems. John Wiley, 1978.
L. Bourdin and E. Trélat. Optimal sampled-data control, and generalizations on time scales. Mathematical Control and Related Fields, 6(1):53–94, 2016.
PMP UNDER FREQUENCY CONSTRAINTS
[Cla13]
[DM65]
[DP75]
[DS89]
[Dub78]
[Gül10]
[Lib12]
[MW12]
[Psh71]
[SB95]
[SS03]
[Tao05]
[Tré12]
[ZDG95]
31
F. H. Clarke. Functional Analysis, Calculus of Variations and Optimal Control, volume 264 of Graduate
Texts in Mathematics. Springer, London, 2013.
A. Ya. Dubovitskii and A. A. Milyutin. Extremum problems in the presence of restrictions. USSR
Computational Mathematics and Mathematical Physics, 5(3):1–80, 1965.
N. N. Dzjuba and B. N. Pšeničnyı̆. The discrete maximum principle. Otdelenie Matematiki, Mekhaniki
i Kibernetiki Akademii Nauk Ukrainskoı̆ SSR. Kibernetika, (2):46–49, 1975.
D. L. Donoho and P. B. Stark. Uncertainty principles and signal recovery. SIAM Journal on Applied
Mathematics, 49(3):906–931, 1989.
A. Ya. Dubovitskiı̆. The discrete maximum principle (Russian). Akademiya Nauk SSSR. Avtomatika i
Telemekhanika, (10):55–71, 1978.
O. Güler. Foundations of Optimization. Springer, 2010.
D. Liberzon. Calculus of Variations and Optimal Control Theory. Princeton University Press, Princeton,
NJ, 2012. A concise introduction.
M. R. Murty and J. P. Whang. The uncertainty principle and a generalization of a theorem of Tao.
Linear Algebra and its Applications, 437(1):214–220, 2012.
B. N. Pshenichnyi. Necessary Conditions for an Extremum, volume 4 of Translated from the Russian
by Karol Makowski. Translation edited by Lucien W. Neustadt. Pure and Applied Mathematics. Marcel
Dekker, Inc., New York, 1971.
J. A. Scaramuzzo and R. W. Brockett. Band limited control of a flexible structure using piecewise
trigonometric input signals, November 7 1995. US Patent 5,465,035.
E. M. Stein and R. Shakarchi. Fourier Analysis, An Introduction. Princeton University Press, Princeton,
NJ, 2003.
T. Tao. An uncertainty principle for cyclic groups of prime order. Mathematical Research Letters,
12(1):121–127, 2005.
Emmanuel Trélat. Optimal control and applications to aerospace: some results and challenges. Journal
of Optimization Theory and Applications, 154(3):713–758, 2012.
K. Zhou, J. C. Doyle, and K. Glover. Robust and Optimal Control. Pearson, 1995.
| 3cs.SY
|
arXiv:1712.08111v1 [astro-ph.IM] 21 Dec 2017
Improving science yield for NASA Swift with
automated planning technologies
Aaron Tohuvavohu
Swift Mission Operations Center
Department of Astronomy and Astrophysics, The Pennsylvania State University
University Park, PA 16802
E-mail: [email protected]
Abstract. The Swift Gamma-Ray Burst Explorer is a uniquely capable mission, with three
on-board instruments and rapid slewing capabilities. It serves as a fast-response satellite
observatory for everything from gravitational-wave counterpart searches to cometary science.
Swift averages 125 different observations per day, and is consistently over-subscribed, responding
to about one-hundred Target of Oportunity (ToO) requests per month from the general
astrophysics community, as well as co-pointing and follow-up agreements with many other
observatories. Since launch in 2004, the demands put on the spacecraft have grown consistently
in terms of number and type of targets as well as schedule complexity. To facilitate this growth,
various scheduling tools and helper technologies have been built by the Swift team to continue
improving the scientific yield of the Swift mission. However, these tools have been used only to
assist humans in exploring the local pareto surface and for fixing constraint violations. Because
of the computational complexity of the scheduling task, no automation tool has been able
to produce a plan of equal or higher quality than that produced by a well-trained human,
given the necessary time constraints. In this proceeding we formalize the Swift Scheduling
Problem as a dynamic fuzzy Constraint Satisfaction Problem (DF-CSP) and explore the global
solution space. We detail here several approaches towards achieving the goal of surpassing
human quality schedules using classical optimization and algorithmic techniques, as well as
machine learning and recurrent neural network (RNN) methods. We then briefly discuss the
increased scientific yield and benefit to the wider astrophysics community that would result
from the further development and adoption of these technologies.
1. Introduction
The scheduling of astronomical observations has been widely studied, for both space and ground
based observatories. While there now exist a large number of robotic ground based telescopes,
Swift was the first semi-autonomous satellite observatory. Fast slewing and autonomous slew
path determination and constraint checking were key design features of Swift to allow it to
achieve its mission goal of detecting, localizing, and rapidly settling on short Gamma-RayBursts (GRBs) before they fade [1]. Due to these capabilities, Swift was able to detect the
first short GRB afterglow, and Swift science was the major contributor to the contemporary
consensus that the progenitor of short GRBs are compact object mergers (NS-NS or NS-BH).
Swift is a panchromatic mission with three instruments: the Ultraviolet/Optical Telescope,
X-ray Telescope, and Burst Alert Telescope (UVOT, XRT, BAT) that are sensitive in the
ultraviolet/optical, soft/hard X-ray, and gamma-ray regimes, with an energy range (non-
continuous) spanning from 2 eV up to 500 KeV. Due to its multi-wavelength capabilities,
fast slewing, and rapid response time Swift has become an enormously popular observatory for
many different types of astronomical targets. While Swift is nominally a GRB mission, over
80% of observing time is now spent on non-GRB science targets. These include everything
from tracking and observing comets in the solar system, to exoplanet science, to high-energy
astrophysical sources at cosmological distances. Along with normal targets, Swift also engages
in large scale tiling campaigns, made possible by its rapid slewing and low re-pointing overhead.
Swift currently performs weekly tiling of the Small Magellanic Cloud [2] and Galactic Bulge [3],
as well as rapid-response tiling in response to astrophysical neutrino detections from ANTARES
[4] and IceCube [5] and gravitational wave triggers from the LIGO/VIRGO Collaboration [6].
18
16
Awarded Time (Ms)
14
12
10
8
6
4
2
0
2010
2011
2012
2013
Year
2014
2015
2016
Year
Avg. # Obs./day
2005
2010
2016
74
85
112
Table 1:
Average number of
observations-per-day over the lifetime of the mission. In 2017 there
were some days with > 400 individual
observations.
Figure 1: Amount of time awarded, in
megaseconds, through the Swift Target of
Opportunity (TOO) program, from 2010-2016.
The unique capabilities of the spacecraft, and the ever evolving nature of Swift operations,
have put unique demands on the operations team. Along with, and because of, Swift’s abilities, it
also has the most complex, dense, and dynamic observing schedules of any satellite observatory.
The complex scheduling goals (science targets) are accompanied by a host of very particular
constraints, both astronautical and scientific, that these schedules must meet (see [7] for an
example). In this proceeding we give a very brief overview of recent work performed with the
goal of automating the creation of these sets of observing commands called Pre-Planned Science
Timelines (PPSTs). Until now PPSTs have been built by (well trained) human Science Planners
who were found to have a better understanding and command of the myriad constraints, and
built markedly safer, more efficient, and higher quality PPSTs (with the assistance of a set
of visualization tools) than any automation software available. However, building a PPST for
one 24-hour period takes a human between 5-7 hours on average. In order to take maximal
advantage of Swift’s fast-response capabilities, the goal is to build an automated system that
can build PPSTs at greater than human-level quality in under 30 minutes by using state-of-theart optimization and modeling techniques.
This will allow the Swift team to significantly increase the efficiency of the observatory and
take maximal advantage of its fast-response capabilities when responding to ToO requests, while
preserving as much pre-planned science as possible and maintaining spacecraft health and safety.
This work will have a particularly large impact on Swift’s search for electromagnetic (EM)
counterparts to gravitational-wave triggers as in [6] and other rapid follow-up projects.
In section 2 we provide a precise formulation of the Swift Scheduling Problem (SSP), the
solution to which is a PPST. In section 3 we motivate our specific approach through the analysis
of large sets of simulated plans. In Section 4 we briefly motivate and outline some recent work on
constraint modeling. In Section 5 we conclude with next steps and the impact of implementation
of the proposed system. Detailed analysis and results regarding constraint modeling/satisfaction
will be reported elsewhere.
2. Formulation
We state the SSP as a dynamic fuzzy Constraint Satisfaction Problem (DF-CSP). We adopt
this approach as opposed to regular CSP because the SSP is over-constrained. That is, any
complete assignment of the CSP violates some constraint. The fuzzy CSP scheme allows us
to extend the problem to constraints that are not crisp; constraints are assigned satisfaction
degrees on the unit interval. We must also allow for new high priority ToO requests or transient
detections (GRBs, Galactic transients, etc.). In this case, computing a completely new solution
to the F-CSP is possible but has the distinct disadvantages of inefficiency and instability of the
successive solutions, as described in [8]. So we adopt the full DF-CSP framework as a method
of reusing the old solution when constraints are added or updated, by making local changes to
the relevant portion of the schedule.
In general, a CSP problem is defined by a set of variables X = {Xi }ni=1 that take values on
their associated finite domains D = {Di }ni=1 under a set of constraints C = {Ci }ri=1 . We make it
fuzzy, following the formalization in [9], by constructing C as a set of fuzzy constraints Ri with
the membership function µRi indicating to what extent some label ν satisfies the Ri :
(
)
Y
Di → [0, 1] .
(1)
C = R|µRi :
xi ∈var(Ri )
So µRi (ν) gives the degree of satisfaction of the relevant constraint. The ability to treat crisp
constraints is preserved by the extreme cases of µR (ν) = 0 or 1. However, µRi gives only the
local degree of satisfaction; the degree to which the label ν satisfies the particular constraint. To
judge whether a given ν is a solution to the full DF-CSP we need a measure of global constraint
satisfaction:
n
o
α(ν) = ⊕ µR (νvar(R) )|R ∈ C
(2)
where ⊕ is an aggregation operator on the unit interval. With α(ν) one can now define a solution
to the DF-CSP as a compound label νX for which α(νX ) ≥ α0 , where α0 is called the solution
threshold, which we determine based on need for optimality and run-time.
We also attach priorities p to each of the constraints, which allows some constraints to be
treated as more or less important than others. This allows one to be confident that, while
some fuzzy constraints can be violated, no label νX can be found as a solution which completely
violates the most important constraints. This property matches the real-world need to preserve
certain constraints that are critical to spacecraft health and safety. Adding priorities changes
the specific definitions given in Eq. 1 and Eq. 2 slightly, a full formulation of the prioritized
F-CSP can be found in [9]. Either way, we have a formalization of our problem, and a clear
solution to aim for: a label νX whose global constraint satisfaction α(νX ) passes the defined
threshold α0 .
All of the constraints, and associated cost function modeling by which we motivate the choice
of priorities, can not be explicitly defined here, but we list most of the relevant hard and soft
constraints below, with the most difficult to manage in red:
Hard Constraints
• Slew resources capped.
• Slew path constraints.
• Passive maintenance of XRT temperature
between -55 C and -65 C.
• Reaction wheel momentum limit.
• Roll constrained by solar-panel pointing and
star tracker catalog.
Soft Constraints
• Maximize observing/slew time ratio.
• Maximize number of science targets
observed.
• Minimize # of UVOT filter wheel
rotations.
• Maximize time spent observing targets > 8 hours from the sun.
3. Approach
Of course, one desires a strictly optimal solution to the scheduling problem, but combinatorial
runtime is unacceptable: optimal scheduling for Swift is at least NP-hard. As the spacecraft
has no consumables, a sub-optimal (but still highly optimized) plan that satisfies the various
constraints and science priorities is acceptable. Early attempts set the solution threshold α0 quite
high. However, it was found that these extremely optimized plans were too fragile. That is, the
dynamic constraints were balanced so precisely that any disturbance (by ToO upload or GRB
detection) would often break a hard constraint. Based on an analysis of simulation results, the
author has good reason to believe that the effect of adding a constructed metric for ‘robustness’
of a PPST to the list of soft constraints will be similar to that of a ‘good-enough’ plan found by a
heuristic scheduler. To wit, we motivate here the use of a multi-strategy scheduling system that
is comprised of a heuristics-informed seed scheduler, guided by a generational genetic algorithm
[10] which performs a greedy search, a gap-filler using a heuristic-biased stochastic sampling
approach as in [11], and a final-pass constraint fixer based on the contention measure [12].
Large sets of planning simulations were performed for various planning periods over many
seasons to allow for diversity in the set of ephemerides. The results of these simulations confirmed
the experiential wisdom of the Planners, that there is “no one way to build a plan.” That is, it
was rigorously demonstrated that for any set of science targets that reasonably resembled the
historical distribution of priorities for a given planning period, there exist multiple unique tours
which fulfill the science priorities for the period, are safe for the spacecraft to perform (do not
violate any hard constraints), and are reasonably similar in optimization score (soft constraints).
Further it was shown that this holds for all relevant ephemerides, with a few marginal cases that
can generally be alleviated with the assistance of a heuristic long-range planner.
While multiple unique acceptable solution classes exist, they were found to share general
properties, with regard to target placement, that distinguish them from the majority of
unacceptable tours. These properties are highly dependent on the details of the ephemeris
for the given planning period. From the large simulation database, a set of meta-heuristics
were derived for various classes of ephemerides. These meta-heuristics can be used to determine
the appropriate scheduling strategy and partial search algorithm (local heuristic) for a given
planning period. Tests of this approach have shown that once localized in a ‘good’ regime of
the solution space by the meta-heuristic informed seed schedule, it is almost always possible to
quickly converge to an acceptable, and sometimes locally optimal, plan.
In the analysis of the large class of simulated tours, it was found that the success of a given
tour is significantly more dependent on the uncertainty in certain constraints than others. It
was further demonstrated that an increased understanding and better model of those constraints
can have outsize impacts on the success and optimization of a given PPST. In particular, the
rate of convergence to an optimized plan can be significantly hampered by getting ‘stuck’ in
unacceptable regions of the solution space because of an incomplete model of the momentum
constraint. As such, a renewed effort at better understanding some of the constraints is valuable.
A brief description of these efforts is given in Section 4.
4. Constraint modeling
In the course of investigating the solution space to the SSP, the critical constraint of preventing
the saturation of the gyroscopic momentum wheels was found to have both the highest
uncertainty in its modeling and the highest computational cost associated with that uncertainty.
This motivates the construction of a better model. Constraint modeling for Swift is almost
uniquely suited to a machine learning approach, as the spacecraft has performed > 350,000
individual observations so far in its mission. There is plenty of data to train on. In
the investigation of the momentum constraint, previously unknown secular phenomena were
discovered that make model building more difficult. However, over training periods of ∼ 10 days
a recurrent neural network (RNN) with 4 hidden layers and 50 nodes has shown promising
results. This work will continue and results will appear in a future publication.
5. Next steps and impact
In this proceeding we formalized the SSP and motivated a particular approach to its
solution; what remains is to finish building these planning algorithms into the Swift planning
infrastructure. The ability to quickly construct high quality PPSTs in response to new ToOs or
transient events will allow Swift to observe these targets in an optimal manner while minimizing
the impact on pre-planned science priorities. Further, it will reduce inefficiency, continuous
staffing needs, and pressure on members of the Swift Science Operations Team. Finally, these
operational improvements will significantly benefit multi-messenger and localization projects,
and ensure that Swift is able to maximally contribute to the new age of gravitational wave and
multi-messenger astronomy.
Acknowledgments
This work was performed at the Swift Mission Operations Center at Penn State under support
from the NASA Swift contract NAS5-00136. The author thanks his colleagues on the Swift FOT
for their help in better understanding the spacecraft health & safety constraints, Jeffrey Gropp
for ongoing work relevant to constraint modeling and automated plan creation, and Alex Ledger
for useful discussion pertinent to algorithmic optimization.
References
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
Gehrels N et al. 2004 Astrophys. J. 611 1005–1020
Kennea J A et al. 2018, in prep
Shaw A W et al. 2017 (AAS/High Energy Astrophysics Division vol 16) p 400.01
Adrian-Martinez S et al. 2016 JCAP 2 062 (Preprint astro-ph/1508.01180)
Evans P A et al. 2015 MNRAS 448 2210–2223 (Preprint astro-ph/1501.04435)
Evans P A et al. 2017 Science (Preprint astro-ph/1710.05437)
Kennea J A et al. 2005 UV, X-Ray, and Gamma-Ray Space Instrumentation for Astronomy XIV (Proceedings
of SPIE vol 5898) pp 341–351
Verfaillie G and Schiex T 1994 Proceedings of the Twelfth National Conference on Artificial Intelligence (Vol.
1) AAAI ’94 pp 307–312
Luo X, man Lee J H, fung Leung H and Jennings N R 2003 Fuzzy Sets and Systems 136 151 – 188
Ortiz-Bayliss J C, Moreno-Scott J and Terashima-Marn H 2014 512 315–327
Bresina J L 1996 Proceedings of the Thirteenth National Conference on Artificial Intelligence - Volume 1
AAAI’96 pp 271–278
Fraser S 2012 Adaptive optimal telescope scheduling Ph.D. thesis Liverpool John Moores University
This figure "slew.png" is available in "png" format from:
http://arxiv.org/ps/1712.08111v1
This figure "tempmap.png" is available in "png" format from:
http://arxiv.org/ps/1712.08111v1
| 3cs.SY
|
A new hierarchy
for automaton semigroups
Laurent Bartholdi1 , Thibault Godin2 , Ines Klimann3 , and Matthieu Picantin3
arXiv:1803.09991v1 [cs.FL] 27 Mar 2018
1
École Normale Supérieure, Paris, and Georg-August-Universität zu Göttingen,
[email protected],
2
University of Turku, Turku,
[email protected], supported by the Academy of Finland grant 296018
3
IRIF, UMR 8243 CNRS & Université Paris Diderot,
[email protected], [email protected]
Abstract. We define a new strict and computable hierarchy for the
family of automaton semigroups, which reflects the various asymptotic
behaviors of the state-activity growth. This hierarchy happens to be
a nontrivial extension of the hierarchy given by Sidki for automaton
groups [20] and also gives new insights into the latter. Its exponential
part coincides with a notion of entropy for some associated automata.
We prove that the Order Problem is decidable in the first two polynomial
levels of this hierarchy, that is, when the state-activity is bounded. Recently proven to be undecidable for the whole family [9,10,4], the Order
Problem remains open for the third level of this hierarchy, that is, when
the state-activity is linear.
The former results are implemented and will be available in the GAP
package FR developed by the first author [3,8].
Keywords: automaton · semigroup · entropy · hierarchy · decision problem
1
Introduction
The family of automaton groups and semigroups has provided a wide playground
to various algorithmic problems in computational (semi)group theory [1,2,3,4,5,6,9,10,11,12].
While many undecidable questions in the world of (semi)groups remain undecidable for this family, the underlying Mealy automata provide a combinatorial
leverage to solve the Word Problem for this family, and various other problems
in some important subfamilies. Recall that a Mealy automaton is a letter-toletter, complete, deterministic transducer with same input and output alphabet,
so each of its states induces a transformation from the set of words over its
alphabet into itself. Composing these Mealy transformations leads to so-called
automaton (semi)groups, and the Word Problem can be solved using a classical
technique of minimization.
The Order Problem is one of the current challenging problems in computational (semi)group theory. On the one hand, it was proven to be undecidable
2
Laurent Bartholdi, Thibault Godin, Ines Klimann, and Matthieu Picantin
for automaton semigroups by Gillibert [9], who has very recently extended this
undecidability result to the class of automaton groups [10]. On the other hand,
Sidki introduced a polynomial hierarchy for invertible Mealy transformations
in [20] and, with Bondarenko, Bondarenko, and Zapata in [6], solved the Order
Problem for its lowest level (bounded invertible automata).
Our main contributions in this paper are the following: an activity-based hierarchy for possibly non-invertible Mealy transformations (Section 3), restricting to
Sidki’s construction [20] in the case of invertible transformations; and a study of
the algorithmic properties in the lowest level of the hierarchy, namely transducers
with bounded activity. We prove:
Theorem (see Section 5). The Order Problem is decidable for bounded Mealy
transformations; namely, there is an algorithm that, given a bounded initial
Mealy automaton, decides whether the transformation τ that it defines has infinite order, and if not finds the minimal r > s such that τ r = τ s .
2
Notions from automata and graph theory
This section gathers some basics from automata and graph theory, especially
some links between multigraphs, Mealy automata, automaton semigroups, and
finite-state endomorphisms. We refer the reader to handbooks for graph theory [16], automata theory [17], and automaton (semi)groups [5].
By a digraph (Q, ∆), we mean a directed multigraph with vertices in Q and edges
in ∆, i.e., a directed graph where multiple edges are allowed. Such a graph is
labeled on a finite alphabet Σ if a letter of Σ is associated with each of its edges.
We denote it by (Q, ∆, Σ) in this case.
A non-deterministic finite-state automaton (NFA for short) can be represented
by a labeled digraph (Q, ∆, Σ) and two particular subsets of vertices I ⊆ Q
and F ⊆ Q. The vertices of the graph are called states of the automaton and
its edges are called transitions. The elements of I and F are called respectively
initial and final states. A transition from the state p to the state q with label x
x
is denoted by p −−→ q.
A NFA is deterministic—DFA for short—(resp. complete) if for each state q
and each letter x, there exists at most (resp. at least) one transition from q
with label x. Given a word w = w1 w2 · · · wn ∈ Σ ∗ (where the wi are letters), a
run with label w in an automaton (NFA or DFA) is a sequence of consecutive
transitions
wn
w2
w1
→ qn+1 .
→ q3 → · · · → qn −−−
→ q2 −−−
q1 −−−
Such a run is successful whenever q1 is an initial state and qn+1 a final state. A
word is recognized by an automaton if it is the label of at least one successful
run. The language recognized by an automaton is the set of words it recognizes.
A DFA is coaccessible if each state belongs to some run ending at a final state.
Let A be a NFA with statetset Q. The Rabin–Scott powerset construction [15]
returns, in a nutshell, the (co)accessible DFA—denoted by det (A)—with states
A new hierarchy for automaton semigroups
3
corresponding to subsets of Q, whose initial state is the subset of all initial states
of A and whose final states are the subsets containing at least on final state of A;
x
its transition labeled by x from a state S ⊆ 2Q leads to the state {q | ∃p ∈ S, p −
→
q in A}. Notice that the size of the resulting DFA might therefore be exponential
in the size of the original NFA.
In Section 4, we count paths in digraphs. Given a graph, its adjacency matrix
is A = {Ai,j }i,j ∈ Nn×n where Ai,j is the number of edges from i to j. The
powers of the adjacency matrix count paths in the digraph: the entry (Aℓ )i,j is
the number of paths of size ℓ starting from vertex i and ending at j.
We will focus P
on the number of paths of length ℓ starting from some vertex i, i.e.,
the quantity 1≤j≤n (Aℓ )i,j , and, precisely, on the limit (assuming it exists4 )
h(A, i) = lim
ℓ→∞
X
1
log
(Aℓ )i,j .
ℓ
1≤j≤n
These kinds of quantities, often called entropies, are studied in various situations,
for instance in the framework of subshifts [14] or finite-state automata [7]: given
an automaton A on alphabet Σ recognizing the language L, the entropy of L is
h(L) = lim
ℓ→∞
1
log L ∩ Σ ℓ .
ℓ
In order to compute the entropy, we associate a digraph GA to the automaton A,
by keeping its accessible part and by removing the labels. If A = {Ai,j }i,j ∈ Nn×n
denotes the adjacency matrix of GA , we have
h(L) = lim
ℓ→∞
1
log
ℓ
X
X
(Aℓ )i,j .
i initial in A 1≤j≤n
Since the adjacency matrix of a digraph G is non-negative, it admits a positive
real eigenvalue of maximal absolute value, which is called the Perron eigenvalue λ(G). It is well-known ([14,16]) that the Perron eigenvalue of a digraph is
the maximal of the Perron eigenvalue of its strongly connected component.
Moreover, for coaccessible DFA, we have the strong characterization:
Proposition 2.1. [19, Theorem 1.2] Let A be a coaccessible automaton recognizing the language L. Then we have h(L) = log λ(GA ).
A Mealy automaton is a letter-to-letter, complete, deterministic transducer with
same input and output alphabet. Graphically, a Mealy automaton M with stateset Q on the alphabet Σ, where Q and Σ are finite sets, can be represented by a
digraph with vertices the elements of Q and edges labeled in Σ × Σ. A transition
x|y
from the state p to the state q with label (x, y) is denoted by p −−−→ q . The
letter x is the input of the transition and the letter y its output. Such a Mealy
4
If the sum is ultimately zero (i.e. the number of paths is finite) we put h(A, i) = 0.
4
Laurent Bartholdi, Thibault Godin, Ines Klimann, and Matthieu Picantin
automaton M is assumed to be complete and deterministic in its inputs: once
fixed a state p and a letter x, there exists exactly one transition from p with
input letter x, and we denote by xp its output letter and by p@x its final state.
x|xp
p
p@x
In this way, states act on letters and letters on states. Such actions can be
composed in the following way: let q ∈ Q∗ , p ∈ Q, u ∈ Σ ∗ , and x ∈ Σ, we have
p
xqp = (xq )
and
p@(ux) = (p@u)@x .
We extend recursively the actions of states on letters and of letters on states
(see below left). Compositions can be more easily understood via an alternative
representation by a cross-diagram [1] (below right).
For all x ∈ Σ, u ∈ Σ ∗ , p ∈ Q, q ∈ Q∗ , we have:
p
q
q
(ux) = uq xq@u
u
and
uq
p@uq
q@u
(qp)@u = q@u · p@uq .
x
uqp
xq@u
q@ux
By convention, the image of the empty word is itself. The mappings defined
above are length-preserving and prefix-preserving.
From an algebraic point of view, the composition gives a semigroup structure
to the set of transformations u 7→ uq for q ∈ Q∗ . This semigroup is called the
semigroup generated by M and denoted by hMi+ . An automaton semigroup is a
semigroup which can be generated by a Mealy automaton. Any element of such
an automaton semigroup induces a so-called finite-state endomorphism.
Conversely, for any endomorphism t of Σ ∗ and any word u ∈ Σ ∗ , ut denotes the
t
image of u by t, and t@u the unique endomorphism s of Σ ∗ satisfying (uv) =
t s
∗
u v for any v ∈ Σ . Such an endomorphism can be uniquely decomposed into
a wreath recursion
t = (t@x1 , . . . , t@x|Σ| )σt ,
where σt = [x1 t , . . . , x|Σ| t ] denotes the transformation on Σ induces by t. Whenever only a finite number of endomorphisms can appear as components, that is,
whenever Q(t) = {t@u : u ∈ Σ ∗ } is finite, t is said to be finite-state and admits
a unique (minimal) associated Mealy automaton Mt with stateset Q(t).
We consider the semigroup FEnd(Σ ∗ ) of those finite-state endomorphisms of Σ ∗ .
Example 2.2. The endomorphism t0 = (1, t0 )[2, 2] belongs to FEnd({1, 2}∗ )
with Q(t0 ) = {1, t0 }. See Examples 4.4 and 5.3 for further details about t0 .
Example 2.3. The endomorphism p = (q, r) with q = (r, 1) and r = (r, r)[2, 2]
also belongs to FEnd({1, 2}∗) with Q(p) = {1, p, q, r}. See Figure 1a for Mp .
A new hierarchy for automaton semigroups
3
5
An activity-based hierarchy for FEnd(Σ ∗ )
In this section we define a suitable notion of activity for finite-state endomorphisms, together with two norms, from which we build a new hierarchy. We will
prove its strictness and its computability in Section 4.
For any element t ∈ FEnd(Σ ∗ ), we define its activity (see Figure 1) as:
αt : n 7−→ #{v ∈ Σ n : ∃u ∈ Σ n , t@u 6= 1 and ut = v}.
1|2
2|2
r
2|2
p
1|1
1|1
q
p
1|1
2|2
1|1
2|2
q
1
r
1|1
2|2
1|2
2|2
r
1
r
r
2|2
(a)
(b)
Fig. 1: (a) An example of an endomorphism p with αp (1) = 2 and αp (2) = 2.
(b) The endomorphism p induces 3 nontrivial endomorphisms on level 2: the
leftmost one is associated with the output 11, the middle right one with 22 and
the rightmost one with 22, hence nontrivial endomorphisms can be reached by
runs with only two different output words.
∗
We next define two norms on FEnd(Σ
P ). Wheni αt has a polynomial growth,
namely when αt (n) is equivalent to 0≤i≤d at,i n , then the corresponding ordiP
nal is denoted by ktkp = 0≤i≤d ω i at,i . Otherwise, the value of limn→∞ log αnt (n)
is denoted by ktke .
We then define the following classes of finite-state endomorphisms:
SPol(d) = { t ∈ FEnd(Σ ∗ ) : ktkp < ω d+1 }
and SExp(λ) = { t ∈ FEnd(Σ ∗ ) : ktke ≤ λ } .
We shall see in Theorem 3.3 that these yield a strict and computable hierarchy
for FEnd(Σ ∗ ). The following basic lemma is crucial:
Lemma 3.1. For each n ≥ 0, the map t 7→ αt (n) is subadditive.
Proof. Assume s, t ∈ FEnd(Σ ∗ ).
For any u ∈ Σ n with (st)@u 6= 1,
we have either s@u 6= 1 or t@us 6= 1.
We deduce αst (n) ≤ αs (n)+αt (n) for
each n ≥ 0.
s
t
v = us
u
s@u
w = vt
t@v
(st)@u 6= 1
6
Laurent Bartholdi, Thibault Godin, Ines Klimann, and Matthieu Picantin
We obtain that k.kp and k.ke are respectively +- and max-subadditive.
Proposition 3.2. Let Σ be an alphabet. For every integer d ≥ −1, SPol(d) is
a subsemigroup of FEnd(Σ ∗ ). So is SExp(λ) for every 0 ≤ λ ≤ #Σ.
As an easy corollary of Proposition 3.2, the subadditivity property allows us to
compute the hierarchy class of any given Mealy automaton by considering only
its generators.
The proof of the following result is postponed to the end of Section 4 on page 8.
Theorem 3.3. Let Σ be an alphabet. The elements of the semigroup FEnd(Σ ∗ )
can be graded according to the following strict hierarchy: for any d1 , d2 ∈ Z
with −1 < d1 < d2 and any λ1 , λ2 ∈ R with 0 < λ1 < λ2 < #Σ, we have:
SPol(−1) ( · · · ( SPol(d1 ) ( · · · ( SPol(d2 ) ( · · · ( SExp(0)
( SExp(λ1 ) ( · · · ( SExp(λ2 ) ( · · · ( SExp(#Σ) .
Sidki defined in [20] the activity of a finite-state automorphism t ∈ FAut(Σ ∗ ) as
θt : n 7−→ #{u ∈ Σ n : t@u 6= 1},
and the corresponding classes Pol(d). Using this notion of activity θ for endomorphisms leads inevitably to an impasse: the associated classes with fixed
degree polynomial θ-activity would be not closed under composition. However
it is straightforward that our new notion of activity α coincides with Sidki’s
activity θ in the case of automorphisms.
S
The class SExp(0) coincides with the infinite union d≥−1 SPol(d), whose corresponding automorphisms class is denoted by Pol(∞) in [20].
4
Structural characterization of the activity norm
From [6], we know that the finite-state automorphisms which have polynomial
activity are exactly those whose underlying automaton does not contain entangled cycles (except on the trivial state). Moreover, the degree of the polynomial
is given by the number of reachable cycles in the automaton. The first claim remains true for endomorphisms, but things are a bit more involved for the second
one (see Example 4.4).
To any minimal Mealy automaton M with stateset Q and alphabet Σ, we
associate its pruned output automaton Mout defined as the NFA with stateset Q r {1} and alphabet Σ, and whose transitions are given, for p, q ∈ Q r {1},
by
y
p −−→ q ∈ Mout
⇐⇒
x|y
p −−−→ q ∈ M.
According to context, we shall identify an endomorphism t ∈ FEnd(Σ ∗ ), the
state of Mt , and the corresponding state of Mout
t .
A new hierarchy for automaton semigroups
7
Lemma 4.1. The activity of an endomorphism t ∈ FEnd(Σ ∗ ) corresponds to
the number of paths starting from t in the (non-complete) deterministic automaton det (Mout
t ) constructed via the Rabin–Scott construction.
Proof. Let t ∈ FEnd(Σ ∗ ) with Mt its associated automaton. Let us count the
words v ∈ Σ n for which there is a word u ∈ Σ n with t@u 6= 1 and ut = v.
For n = 1, αt (1) is exactly the number of different outputs from the state t that
do not lead to a trivial state of Mt . Now for v ∈ Σ n , if E denotes the set of
those states accessible from t by reading v (this corresponds to the Rabin–Scott
powerset construction) in Mout
t , the number of ways to extend v without getting
into a trivial state in Mt corresponds to the number of outputs of the state E
in det (Mout
t ), whence the result.
Whether the activity of a given t ∈ FEnd(Σ ∗ ) is polynomial or exponential can
be decided by only looking at the cycle structure of Mout
t :
Proposition 4.2. An endomorphism t ∈ FEnd(Σ ∗ ) has an exponential activity
if and only if it can reach two coaccessible cycles with distinct labels in Mout
t .
Proof. (⇐) Assume that t can reach a state s ∈ Mout
that lies on two cycles
t
with distinct labels. There exist a word u ∈ Σ ∗ satisfying t@u = s and two
words v, w ∈ Σ ∗ satisfying s@v = s = s@w and vs 6= ws . We obtain αt (ℓ) ≥
#{xt | x ∈ u(v + w)∗ ∩ Σ ℓ } for ℓ ≥ 0. Therefore αt grows exponentially.
(⇒) Assume that t has an exponential activity and cannot reach two coaccessible cycles in Mout
t . By Lemma 4.1, there exist a subset E ⊂ Q and three
words u, v, w ∈ Σ ∗ such that E is the set of nontrivial states accessible from t
reading u and the paths labeled by v and w are cycles starting from E in Mout
t .
It means that v and w are also cycles starting from t in Mt , contradiction.
Using the subadditivity of the activity (see Lemma 3.1), we get for the polynomial activities:
Corollary 4.3. Let M be a Mealy automaton. The endomorphisms of hMi+
are all of polynomial activity if and only if there are non coaccessible cycles in
the automaton det (Mout ). Moreover the degree of the (polynomial) activity corresponds to the maximum of number of reachable cycles in det (Mout ) minus 1.
Example 4.4. Consider the endomorphism t0 = (1, t0 )[2,2] from
Example 2.2,
:
its square t20 , and the associated automata Mt20 and det Mout
t2
0
t20
2|2
1|2
t0
2|2
1|2
1
1|1
2|2
{t20 }
2
{t20 , t0 }
{t0 }
2
2
Note that, before determinization, two disjoint cycles are accessible from the
state t20 . In the determinized version, {t0 } and {t20 } both access to only one
cycle, and we conclude {t0 , t20 } ⊂ SPol(0). By Proposition 3.2, we actually knew
the full inclusion h t0 i+ ⊂ SPol(0).
8
Laurent Bartholdi, Thibault Godin, Ines Klimann, and Matthieu Picantin
Defining further tk = (tk−1 , tk )[2 − (k mod 2), 2 − (k mod 2)] ∈ FEnd({1, 2}∗ ),
we obtain the family with tk ∈ SPol(k) r SPol(k − 1) for k > 0, that witnesses
the strictness of the polynomial part of the hierarchy from Theorem 3.3.
Using Proposition 2.1, we obtain an explicit formula for the norm k · ke :
Proposition 4.5. Let t be a finite-state endomorphism with associated Mealy
automaton Mt . The norm ktke is the Perron eigenvalue of the digraph Gdet(Mout
)
t
ktke = log λ(Gdet(Mout
)) .
t
Proof. By Lemma 4.1, the activity of t counts the number of paths in det (Mout
t ).
Since all its states are final, this automaton is coaccessible and the cardinality of
the language accepted when putting t as the initial state is exactly the activity
of t. Therefore by Proposition 2.1, we have
n
X
log αt (ℓ)
1
= lim
log
(Aℓ )t,t′ = h(L) = log λ(G),
ℓ→∞
ℓ→∞ ℓ
ℓ
′
ktke = lim
t =1
with G = Gdet(Mout
) and A = (Ai,j )i,j the adjacency matrix of G.
t
Proof of Theorem 3.3. The strictness for the polynomial part is obtained from
Example 4.4. Now, as the norm k.ke is the logarithm of the maximal eigenvalue
of a matrix with integer coefficients, the classes SExp(λ) increase only when eλ
is an algebraic integer that is the largest zero of its minimal polynomial, i.e.,
a root of a Perron number. Furthermore, each of these numbers is the norm of
some finite-state endomorphism, see [13, Theorem 3] for a proof. It is also known
that Perron numbers are dense in [1, ∞[, which gives us the strictness for the
exponential part: λ1 < λ2 implies SExp(λ1 ) ( SExp(λ2 ).
Finally, the growth rate can be computed with any precision 0 < δ < 1 in
time Θ (− log(δn)), where n is the number of states of the automaton [18].
Example 4.6. Consider the endomorphisms r = (s, r)[1, 1] and s = (1, r) with
common associated automata M (on the left) and det (Mout ) (on the right):
1|1
2|1
r
s
2|2
1|1
1
1|1
2|2
{s}
2
{r}
1
{r, s}
1
2
We find that αr (n) and αs (n + 1) correspond to the n-th Fibonacci number. We
deduce krke = kske = log ϕ where ϕ is the golden ratio, hence r, s ∈ SExp(log ϕ).
5
The orbit signalizer graph and the order problem
This section is devoted to the Order Problem: can one decide whether a given
element generates a finite (semi)group? The latter is known to be undecidable
for general automaton (semi)groups [9,10,4] and decidable for Pol(0) [6]. We
present a general construction for a graph which hopefully allows to compute
the (in)finite index and period of the generated semigroup. We show that this
graph has to be finite for elements from SPol(0), what solves the Order Problem.
A new hierarchy for automaton semigroups
9
Let Σ be an alphabet. We define the orbit signalizer graph Φ for FEnd(Σ ∗ ) as the
following (infinite) graph. The vertices are the pairs of elements in FEnd(Σ ∗ ).
For each letter x ∈ Σ, there is an arrow from the source (s, t) with label (x : m , ℓ)
where m and ℓ are the minimal integers (with ℓ > 0) satisfying
m+ℓ
xst
m
= xst ,
and with target (r@x, tℓ @xr ) for r = stm . The parameters m and ℓ correspond
respectively to the index and to the period of the orbit of x under the action
of stω , see Figure 2.
t
r = stm
s
t
xs
x
s@x
xrt
t
.
xst
xst
2
t
t
m−1
t@xs
xrt
xr
t@xst
xrt
tℓ
t
m−1
t
3
xrt
ℓ
xrt
4
Fig. 2: The cross-diagram associated with the orbit of some letter x ∈ Σ under
the action of stω . The index m and period ℓ will complete the label of the x-arrow
away from the vertex (s, t) in the graph Φ. Each of the two gray zones indicates
an entry of the corresponding target vertex (r@x, tℓ @xr ) with r = stm .
The inf-index-cost, sup-index-cost, and the period-cost of a given walk π on Φ
x1 : m1 ,ℓ1
x|π| : m|π| ,ℓ|π|
π : (s, t) −−−−−−→ · · · −−−−−−−−−→ (s′ , t′ )
are respectively defined by
X
Y
(1 − δmk ,0 ) (mk − 1)
i− (π) =
ℓj + 1 ,
1≤k≤h
i+ (π) =
X
1≤j<k
1≤k≤|π|
Y
ℓ j mk ,
and
p(π) =
1≤j<k
Y
ℓi .
1≤i≤|π|
∗
For any t ∈ FEnd(Σ ), we define the orbit signalizer graph Φ(t) as the subgraph
of Φ accessible from the source vertex (1, t). The inf-index-cost, sup-index-cost,
and the period-cost of t ∈ FEnd(Σ ∗ ) are then respectively defined by
i−
t =
max i− (π),
π on Φ(t)
i+
t =
max i+ (π),
π on Φ(t)
and
pt =
lcm p(π) .
π on Φ(t)
10
Laurent Bartholdi, Thibault Godin, Ines Klimann, and Matthieu Picantin
Proposition 5.1. The semigroup generated by an element t ∈ FEnd(Σ ∗ ) is
finite if and only if its index-costs i±
t and its period-cost pt are finite. In that
+
case, we have h t i+ = h t : tit = tit +pt i+ for some index it with i−
t ≤ it ≤ it .
Proof. Let Σ = {x1 , . . . , x|Σ| }. Let (s0 , t0 ) be a vertex in Φ and (sk , tk ) its
successor vertex with arrow xk : mk , ℓk for 1 ≤ k ≤ |Σ|.
(s1 , t1 )
x1 : m1 , ℓ1
(i1 , p1 )
(s0 , t0 )
(i0 , p0 )
x|Σ| : m|Σ| , ℓ|Σ|
(s|Σ| , t|Σ| ) (i|Σ| , p|Σ| )
For 0 ≤ k ≤ |Σ|, let (ik , pk ) ∈ {∞, 0, 1, 2, . . .}×{∞, 1, 2, 3 . . .} denote the possible
minimal pair of integers (with pk > 0) satisfying
sk tikk = sk tikk +pk .
Whenever there is at least one successor with (ik , pk ) = (∞, ∞), (s0 , t0 ) satisfies
also (i0 , p0 ) = (∞, ∞), and so does any of its predecessors. Otherwise, we claim
max mk + max(0, ℓk (ik − 1) + 1) ≤ i0 ≤ max (mk + ℓk ik )
1≤k≤|Σ|
1≤k≤|Σ|
and
p0 =
lcm ℓk pk .
1≤k≤|Σ|
Indeed, for 1 ≤ k ≤ |Σ| and for any u ∈ Σ ∗ , we have
mk +ℓk ik
s0 t0
yk v = (xk u)
mk
with yk = xk s0 t0
ik +pk
and v = usk tk
ℓk p k
t
= (yk v) 0
, as illustrated by the cross-diagram:
pk
ik
k
s 0 tm
0
tℓ0k
yk
xk
sk
yk
tk
·
u
·
tℓ0k
·
tk
·
·
tℓ0k
yk
·
tk
·
·
tℓ0k
·
·
yk
·
tk
tk
·
v
·
tℓ0k
·
·
v
·
We conclude using an induction on the length of the paths.
Theorem 5.2. The Order Problem is decidable for any t ∈ FEnd(Σ ∗ ) with a
finite orbit signalizer graph Φ(t).
Proof. Since Φ(t) is a graph with outdegree kΣk > 0 by construction, its finiteness implies the existence of cycles. Consider the simple cycles (there is only a
finite amount of these). One can compute the index-costs i− (κ) and i+ (κ) and
the period-cost p(κ) of each such cycle κ. Whenever i− (κ) > 0 or p(κ) > 1 for
some cycle κ, then t has infinite order, and finite order otherwise.
A new hierarchy for automaton semigroups
11
Example 5.3. The endomorphisms s = (s, 1)[2, 2] and t0 = (1, t0 )[2, 2] (on the
left) admit respective graphs Φ(s) and Φ(t0 ) (on the right):
1|2
s
2|2
(1, s)
1
2|2
t0
1|1
2|2
1 : 1, 1
1|2
1 : 0, 1
2 : 0, 1
1 : 0, 1
2 : 0, 1
(1, 1)
2 : 0, 1
(s, 1)
1 : 1, 1
2 : 0, 1
(1, t0 )
According to Proposition 5.1, they generate the finite monoid h s : s2 = s i+
and the free monoid h t0 : i+ .
Example 5.4. The endomorphism b = (a, 1, b)[2, 3, 1] from SPol(1) r SPol(0)
with a = (1, 1, a)[1, 1, 2] admits the finite graph Φ(b) displayed on Figure 3,
in which we can read that both ab and ba have period 1, and that b has thus
period pb = 3. According to Proposition 5.1 again, the index of b satisfies 4 ≤
ib ≤ 9, and can be explicitly computed as ib = 8.
1|2
2|2
3|3
ab
3|2
a
1|1
2|2
(1, b)
1|1
2|1
1|1
2|2
3|3
1
1|2
2 : 0, 3
3 : 0, 3
1 : 0, 3
(1, ab)
1 : 1, 1
2 : 0, 1 1 : 0, 1
3 : 0, 1 2 : 0, 1
(1, ba)
3 : 1, 1
2|3
ba
b
3|1
3|1
(a, a)
1 : 0, 1
3 : 0, 1 2 : 1, 1
1 : 0, 1
2 : 0, 1
(1, 1)
1 : 0, 1
2 : 0, 1
3 : 0, 1
(1, a)
1 : 1, 1
2 : 2, 1
3 : 2, 1
1 : 0, 1
2 : 0, 1
(a, 1)
(b, a)
3 : 0, 1
1 : 0, 1
3 : 0, 1
(b, 1)
3 : 0, 1
2 : 0, 1
Fig. 3: The Mealy automaton Mb and the graph Φ(b) from Example 5.4.
Proposition 5.5. Every bounded finite-state endomorphism t ∈ SPol(0) admits
a finite orbit signalizer graph Φ(t).
Proof. The activity αt of t ∈ SPol(0) is uniformly bounded by some constant C:
#{v ∈ Σ n : ∃u ∈ Σ n , t@u 6= 1 and ut = v} ≤ C
for
n≥0.
12
Laurent Bartholdi, Thibault Godin, Ines Klimann, and Matthieu Picantin
Now the vertices of the graph Φ(t) are those pairs (r, s) with rs = tK @u, where K
i
is the greatest integer such that the images ut for i < K are pairwise different.
Hence rs is the product of at most C nontrivial elements. Moreover these elements ti @u for i < K lie in a finite set, as t is a finite-state endomorphism, hence
the vertices belong to a finite set. We conclude that the graph Φ(t) is finite.
Corollary 5.6. The Order Problem is decidable for SPol(0).
References
1. A. Akhavi, I. Klimann, S. Lombardy, J. Mairesse, and M. Picantin, The finiteness
problem for automaton semigroups, Internat. J. Alg. Comput. 22(6) (2012), 1–26.
2. L. Bartholdi, Algorithmic decidability of Engel’s property for automaton groups,
CSR 2016, LNCS, vol. 9691, pp. 29–40.
, FR – GAP package “Computations with functionally recursive groups”, Ver3.
sion 2.4.3, 2017, http://www.gap-system.org/Packages/fr.html.
4. L. Bartholdi and I. Mitrofanov, The word and order problems for self-similar and
automata groups, arXiv:1710.10109.
5. L. Bartholdi and P. V. Silva, Groups defined by automata, AutoMathA Handbook
(J.-É. Pin, ed.), Europ. Math. Soc., 2010, arXiv:1012.1531.
6. I. Bondarenko, N. Bondarenko, S. Sidki, and F. Zapata, On the conjugacy problem
for finite-state automorphisms, Groups Geom. Dyn. 7(2) (2013), 323–355.
7. N. Chomsky and G. A. Miller, Finite state languages, Information and Control
1(2) (1958), 91–112.
8. The GAP Group, GAP – Groups, Algorithms, and Programming, 2018,
http://www.gap-system.org.
9. P. Gillibert, The finiteness problem for automaton semigroups is undecidable, Internat. J. Alg. Comput. 24(1) (2014), 1–9.
10.
, An automaton group with undecidable order and Engel problems, J. Algebra 497 (2018), 363–392.
11. Th. Godin and I. Klimann, On bireversible Mealy automata and the Burnside problem, Theor. Comput. Sci. 707 (2018), 24–35.
12. I. Klimann, J. Mairesse, and M. Picantin, Implementing computations in automaton (semi)groups, CIAA 2012, LNCS, vol. 7381, pp. 240–252.
13. D. A. Lind, The entropies of topological markov shifts and a related class of algebraic integers, Ergodic Theory and Dynamical Systems 4(2) (1984), 283–300.
14. D. A. Lind and B. Marcus, An introduction to symbolic dynamics and coding,
Cambridge University Press, 1995.
15. M. O. Rabin and D. Scott, Finite automata and their decision problems, IBM
Journal of Research and Development 3(2) (1959), 114–125.
16. M. Rigo, Advanced graph theory and combinatorics, Wiley, 2016.
17. J. Sakarovitch, Elements of automata theory, Cambridge University Press, 2009.
18. A. M. Shur, Combinatorial complexity of regular languages, CSR’08, LNCS, vol.
5010, 2008, pp. 289–301.
19.
, Comparing complexity functions of a language and its extendable part,
RAIRO-Theoretical Informatics and Applications 42(3) (2008), 647–655.
20. S. Sidki, Automorphisms of one-rooted trees: growth, circuit structure, and acyclicity, Journal of Mathematical Sciences 100(1) (2000), 1925–1943.
| 4math.GR
|
arXiv:1404.4045v1 [math.AC] 15 Apr 2014
GAUSSIAN PROPERTY IN AMALGAMATED ALGEBRAS ALONG AN IDEAL
NAJIB MAHDOU AND MOUTU ABDOU SALAM MOUTUI
Abstract. Let f : A → B be a ring homomorphism and J be an ideal of B. In this paper,
we investigate the transfer of Gaussian property to the amalgamation of A with B along J
with respect to f (denoted by A ⊲⊳ f J), introduced and studied by D’Anna, Finocchiaro
and Fontana in 2009.
1. Introduction
All rings considered in this paper are commutative with identity elements and all modules are unital. In 1932, Prüfer introduced and studied in [32] integral domains in which
every finitely generated ideal is invertible. In their recent paper devoted to Gaussian properties, Bazzoni and Glaz have proved that a Prüfer ring satisfies any of the other four
Prüfer conditions if and only if its total ring of quotients satisfies that same condition [9,
Theorems 3.3 & 3.6 & 3.7 & 3.12]. Vasconcelos regarding a conjecture of Kaplansky:
The content ideal of a Gaussian polynomial is an invertible (or locally principal) ideal.
The reason behind the conjecture is that the converse holds [33]. Vasconcelos and Glaz
answered the question affirmatively in a large number of cases [24, 25]. The affirmative
answer was later extended by Heinzer and Huneke [27] to include all Noetherian domains.
Recently the question was answered affirmatively for all domains by Loper and Roitman
[29], and finally to non-domains provided the content ideal has zero annihilator by Lucas [30]. The article by Corso and Glaz [14] gives a good account of what was known
about the problem prior to the year 2000 or so. At the October 2003 meeting of the American Mathematical Society in Chapel Hill and North Carolina, Loper presented a proof
that every Gaussian polynomial over an integral domain has invertible content (in fact,
for any ring that is locally an integral domain). The basis of the proof presented in this
paper is highly dependent on the work of Loper and his coauthor Roitman. A flurry of
related research ensued, particularly investigations involving Dedekind-Mertens Lemma
and various extensions of the Gaussian property. [2] and [14] provide a survey of results
obtained up to year 2000 and an extensive bibliography. A related, but different, question
is : How Prüfer-like is a Gaussian ring? Various aspects of the nature of Gaussian rings
were investigated in Tsangs thesis [33], Anderson and Camillo [3], and Glaz [22]. While
all of those works touch indirectly on the mentioned question, it is Glaz [22] that asks
and provides some direct answers. A problem initially associated with Kaplansky and his
student Tsang [8, 24, 30, 33] and also termed as Tsang-Glaz-Vasconcelos conjecture in
[27] sustained that every nonzero Gaussian polynomial over a domain has an invertible (or,
equivalently, locally principal) content ideal.” It is well-known that a polynomial over any
ring is Gaussian if its content ideal is locally principal. The converse is precisely the object of Kaplansky-Tsang-Glaz-Vasconcelos conjecture extended to those rings where every
2000 Mathematics Subject Classification. 16E05, 16E10, 16E30, 16E65.
Key words and phrases. Amalgamated algebra along an ideal, Gaussian rings, arithmetical rings, amalgamated duplication, trivial rings extension.
1
2
NAJIB MAHDOU AND MOUTU ABDOU SALAM MOUTUI
Gaussian polynomial has locally principal content ideal. In [4], the authors examined the
transfer of the Prüfer conditions and obtained further evidence for the validity of BazzoniGlaz conjecture sustaining that ”the weak global dimension of a Gaussian ring is 0, 1, or
∞” [9]. Notice that both conjectures share the common context of rings. They gave new
examples of non-arithmetical Gaussian rings as well as arithmetical rings with weak global
dimension strictly greater than one. Abuihlail, Jarrar and Kabbaj studied in [1] the multiplicative ideal structure of commutative rings in which every finitely generated ideal is
quasi-projective. They provide some preliminaries quasi-projective modules over commutative rings and they investigate the correlation with well-known Prüfer conditions; namely,
they proved that this class of rings stands strictly between the two classes of arithmetical
rings and Gaussian rings. Thereby, they generalized Osofskys theorem on the weak global
dimension of arithmetical rings and partially resolve Bazzoni-Glazs related conjecture on
Gaussian rings. They also established an analogue of Bazzoni-Glaz results on the transfer
of Prüfer conditions between a ring and its total ring of quotients. In [13], the authors
studied the transfer of the notions of local Prüfer ring and total ring of quotients. They examined the arithmetical, Gaussian, and fqp conditions to amalgameted duplication along
an ideal. They also investigated the weak global dimension of an amalgamation and its
possible inheritance of the semihereditary condition. At this point, we make the following
definition:
Definition 1.1. Let R be a commutative ring.
(1) R is called an arithmetical ring if the lattice formed by its ideals is distributive (see
[20]).
(2) R is called a Gaussian ring if for every f, g ∈ R[X], one has the content ideal
equation c( f g) = c( f )c(g) (see [33]).
(3) R is called a Prü f er ring if every finitely generated regular ideal of R is invertible
(see [12, 26]).
In the domain context, all these forms coincide with the definition of a Prü f er domain.
Glaz [23] provides examples which show that all these notions are distinct in the context of
arbitrary rings. The following diagram of implications summarizes the relations between
them [8, 9, 22, 23, 29, 30, 33]:
Arithmetical ⇒ Gaussian ⇒ Prü f er
and examples are given in [23] to show that, in general, the implications cannot be reversed.
Arithmetical and Gaussian notions are local; i.e., a ring is arithmetical (resp., Gaussian) if
and only if its localizations with respect to maximal ideals are arithmetical (resp., Gaussian). We will make frequent use of an important characterization of a local Gaussian ring;
namely, ”for any two elements a, b in the ring, we have < a, b >2 =< a2 > or < b2 >;
moreover, if ab = 0 and, say, < a, b >2 =< a2 >, then b2 = 0” (by [9, Theorem 2.2 ]).
In this paper, we study the transfer of Gaussian property in amalgamation of rings issued
from local rings, introduced and studied by D’Anna, Finocchiaro and Fontana in [15, 16]
and defined as follows :
Definition 1.2. Let A and B be two rings with unity, let J be an ideal of B and let f : A → B
be a ring homomorphism. In this setting, we can consider the following subring of A × B:
A ⊲⊳ f J := {(a, f (a) + j) | a ∈ A, j ∈ J}
GAUSSIAN PROPERTY IN AMALGAMATED ALGEBRAS ALONG AN IDEAL
3
called the amalgamation of A and B along J with respect to f . In particular, they have
studied amalgmations in the frame of pullbacks which allowed them to establish numerous
(prime) ideal and ring-theoretic basic properties for this new construction. This construction is a generalization of the amalgamated duplication of a ring along an ideal (introduced
and studied by D’Anna and Fontana in [17, 18, 19]). The interest of amalgamation resides,
partly, in its ability to cover several basic constructions in commutative algebra, including
pullbacks and trivial ring extensions (also called Nagata’s idealizations) (cf. [31, page 2]).
Moreover, other classical constructions (such as the A + XB[X], A + XB[[X]] and the D + M
constructions) can be studied as particular cases of the amalgamation ([15, Examples 2.5
and 2.6]) and other classical constructions, such as the CPI extensions (in the sense of
Boisen and Sheldon [11]) are strictly related to it ([15, Example 2.7 and Remark 2.8]). In
[15], the authors studied the basic properties of this construction (e.g., characterizations
for A ⊲⊳ f J to be a Noetherian ring, an integral domain, a reduced ring) and they characterized those distinguished pullbacks that can be expressed as an amalgamation. Moreover,
in [16], they pursued the investigation on the structure of the rings of the form A ⊲⊳ f J,
with particular attention to the prime spectrum, to the chain properties and to the Krull
dimension.
2. Transfer of Gaussian property in amalgamated algebras along an ideal
The main Theorem of this paper develops a result on the transfer of the Gaussian property to amalgamation of rings issued from local rings.
Theorem 2.1. Let (A, m) be a local ring, B be a ring, f : A → B be a ring homomorphism
and J be a proper ideal of B such that J ⊆ Rad(B). Then, the following statements hold :
(1) If A ⊲⊳ f J is Gaussian, then so are A and f (A) + J.
(2) Assume that J 2 = 0. Then A ⊲⊳ f J is Gaussian if and only if so is A and f (a)J =
f (a)2 J ∀ a ∈ m.
(3) Assume that f is injective. Then two cases are possible:
Case 1 : f (A) ∩ J = (0). Then A ⊲⊳ f J is Gaussian if and only if so is f (A) + J.
Case 2 : f (A) ∩ J , (0). Assume that A is reduced. Then A ⊲⊳ f J is Gaussian if
and only if so is A, J 2 = 0 and f (a)J = f (a)2 J ∀ a ∈ m.
(4) Assume that f is not injective. Then two cases are possible:
Case 1 : J ∩ Nilp(B) = (0). If A is reduced, then A ⊲⊳ f J is not Gaussian.
Case 2 : J ∩ Nilp(B) , (0). Assume that A is reduced. Then A ⊲⊳ f J is Gaussian
if and only if so is A, J 2 = 0 and f (a)J = f (a)2 J ∀ a ∈ m.
Before proving Theorem 2.1, we establish the following Lemma.
Lemma 2.2. Let (A, B) be a pair of rings, f : A → B be a ring homomorphism and J be a
proper ideal of B. Then, A ⊲⊳ f J is local if and only if so is A and J ⊆ Rad(B).
f
Proof. By [16, Proposition 2.6 (5)], Max(A ⊲⊳ f J) = {m ⊲⊳ f J / m ∈ Max(A)} ∪ {Q } with
f
Q ∈ Max(B) not containing V(J) and Q := {(a, f (a) + j) / a ∈ A, j ∈ J and f (a) + j ∈ Q
f
}. Assume that A ⊲⊳ J is local. It is clear that A is local by the above characterization
of Max(A ⊲⊳ f J). We claim that J ⊆ Rad(B). Deny. Then there exist Q ∈ Max(B) not
containing V(J) and so Max(A ⊲⊳ f J) contains at least two maximal ideals, a contradiction
since A ⊲⊳ f J is local. Hence, J ⊆ Rad(B). Conversely, assume that (A, m) is local and
f
J ⊆ Rad(B). Then J is contained in Q for all Q ∈ Max(B). Consequently, the set {Q }
4
NAJIB MAHDOU AND MOUTU ABDOU SALAM MOUTUI
is empty. And so Max(A ⊲⊳ f J) = {m ⊲⊳ f J/m ∈ Max(A)}. Hence, m ⊲⊳ f J is the only
maximal ideal of A ⊲⊳ f J since (A, m) is local. Thus, (A ⊲⊳ f J, M) is local with M = m ⊲⊳ f J,
as desired.
Proof of Theorem 2.1. By Lemma 2.2, (A ⊲⊳ f J, m ⊲⊳ f J) is local since (A, m) is local and
J ⊆ Rad(B).
(1) If A ⊲⊳ f J is Gaussian, then A and f (A) + J are Gaussian rings since the Gaussian
f
A⊲⊳ f J
J
, by [15,
property is stable under factor ring ( here A {0}×{J}
and f (A) + J f −1A⊲⊳
(J)×{0}
Proposition 5.1 (3)]).
(2) Assume that J 2 = 0.
If A ⊲⊳ f J is Gaussian, then so is A by (1). We claim that f (a)J = f (a)2 J for all a ∈ m.
Indeed, it is clear that f (a)2 J ⊆ f (a)J. Conversely, let x ∈ J and let a ∈ m. Clearly,
(a, f (a)) and (0, x) are elements of A ⊲⊳ f J. And so < (a, f (a)), (0, x) >2 =< (a, f (a))2 >
since J 2 = 0. It follows that x f (a) = j f (a)2 for some j ∈ J. Hence, f (a)J = f (a)2 J.
Conversely, suppose A is Gaussian and f (a)J = f (a)2 J for all a ∈ m. We claim that
A ⊲⊳ f J is Gaussian. Indeed, Let (a, f (a) + i) and (b, f (b) + j) ∈ A ⊲⊳ f J. Then a and b ∈ A.
We may assume that a, b ∈ m and < a, b >2 =< a2 >. Therefore, b2 = a2 x and ab = a2 y
for some x, y ∈ A. Moreover ab = 0 implies that b2 = 0. By assumption, there exist
j1 , i1 , j2 , i2 , i3 ∈ J such that 2 f (b) j = f (a)2 f (x) j1 , 2 f (a)i f (x) = f (a)2 i1 , f (a) j = f (a)2 j2 ,
f (b)i = f (a)2 f (x)i2 and 2 f (a)i f (y) = f (a)2 i3 . In view of the fact J 2 = 0, one can easily
check that (b, f (b) + j)2 = (a, f (a) + i)2(x, f (x) + f (x) j1 − i1 ) and (a, f (a) + i)(b, f (b) + j) =
(a, f (a)+i)2(y, f (y)+ f (x)i2 + j2 −i3 ). Moreover, assume that (a, f (a)+i)(b, f (b)+ j) = (0, 0).
Hence, ab = 0 and so b2 = 0. Consequently, (b, f (b) + j)2 = (0, 0). Finally, A ⊲⊳ f J is
Gaussian.
(3) Assume that f is injective.
Case 1 : Suppose that f (A) ∩ J = (0). If A ⊲⊳ f J is Gaussian, then so is f (A) + J by (1).
Conversely, assume that f (A) + J is Gaussian. We claim that the natural projection :
p : A ⊲⊳ f J → f (A) + J
p((a, f (a) + j)) = f (a) + j is a ring isomorphism. Indeed, it is clear that p is surjective. It
remains to show that p is injective. Let (a, f (a) + j) ∈ Ker(p), it is clear that f (a) + j = 0.
And so f (a) = − j ∈ f (A) ∩ J = (0). Consequently, f (a) = − j = 0 and so a = 0 since f
is injective. It follows that (a, f (a) + j) = (0, 0). Hence, p is injective. Thus, p is a ring
isomorphism. The conclusion is now straightforward.
Case 2 : Assume that f (A) ∩ J , (0) and A is reduced.
By (2) above, it remains to show that if A ⊲⊳ f J is Gaussian, then J 2 = 0. Assume that
A ⊲⊳ f J is Gaussian. We claim that J 2 = 0. Indeed, let 0 , f (a) ∈ f (A) ∩ J and let x, y ∈ J.
Clearly, (0, x) and (a, 0) are elements of A ⊲⊳ f J. So, we have < (a, 0), (0, x) >2 =< (a, 0)2 >
or < (0, x)2 >. It follows that x2 = 0 or a2 = 0. Since A is reduced and 0 , a, then a2 , 0.
Hence, x2 = 0. Likewise y2 = 0. Therefore, xy = 0 since J is an ideal of the local Gaussian
ring f (A) + J and < x, y >2 =< x2 , y2 , xy >=< x2 >= 0 or < y2 >= 0. Hence, J 2 = 0, as
desired.
(4) Assume that f is not injective.
Case 1 : Suppose that J ∩ Nilp(B) = (0) and A is reduced. We claim that A ⊲⊳ f J is not
Gaussian. Deny. Using the fact f is not injective, there is some 0 , a ∈ Ker( f ) and so
(a, f (a)) = (a, 0) ∈ A ⊲⊳ f J. Let x ∈ J, then (0, x) ∈ A ⊲⊳ f J. We have < (a, 0), (0, x) >2 =<
(a, 0)2 > or < (0, x)2 >. And so it follows that x2 = 0 or a2 = 0. Since A is reduced and
0 , a, then a2 , 0. Hence, x2 = 0. Therefore, x ∈ J ∩ Nilp(B) = 0. So, x = 0. Hence, we
obtain J = 0, a contradiction since J is a proper ideal of B. Thus, A ⊲⊳ f J is not Gaussian,
as desired.
GAUSSIAN PROPERTY IN AMALGAMATED ALGEBRAS ALONG AN IDEAL
5
Case 2 : Assume that J ∩ Nilp(B) , (0) and A is reduced.
By (2), it remains to show that if A ⊲⊳ f J is Gaussian, then J 2 = 0. Suppose that A ⊲⊳ f J
is Gaussian. Since f is not injective, then there is some 0 , a ∈ Ker( f ). Let x, y ∈ J,
so (0, x) and (a, f (a)) = (a, 0) are elements of A ⊲⊳ f J. And so, < (a, 0), (0, x) >2 =<
(a, 0)2, (0, x)2 >=< (a, 0)2 > or < (0, x)2 >. It follows that x2 = 0 or a2 = 0. Since A is
reduced and 0 , a, then a2 , 0. Hence, x2 = 0. Likewise y2 = 0. Therefore, xy = 0 since
J is an ideal of f (A) + J which is (local) Gaussian by (1) and < x, y >2 =< x2 , y2 , xy >=<
x2 >= 0 or < y2 >= 0. Hence, J 2 = 0, as desired.
The following corollary is a consequence of Theorem 2.1 and is [13, Theorem 3.2 (2)].
Corollary 2.3. Let (A, m) be a local ring and I a proper ideal of A. Then A ⊲⊳ I is Gaussian
if and only if so is A, I 2 = 0 and aI = a2 I for all a ∈ m.
Proof. It is easy to see that A ⊲⊳ I = A ⊲⊳ f J where f is the identity map of A, B = A
and J = I. By Lemma 2.2, (A ⊲⊳ I, M) is local with M = m ⊲⊳ I since (A, m) is local and
I ⊆ Rad(A) = m. If A is Gaussian, I 2 = 0 and aI = a2 I for all a ∈ m, then A ⊲⊳ I is
Gaussian by (2) of Theorem 2.1. Conversely, assume that A ⊲⊳ I is Gaussian. By (2) of
Theorem 2.1, it remains to show that I 2 = 0. Indeed, let x, y ∈ I. Then, (x, 0), (0, x) ∈ A ⊲⊳ I
and < (x, 0), (0, x) >2 =< (x, 0)2 > or < (0, x)2 >. Hence, it follows that x2 = 0. Likewise
y2 = 0. So, < x, y >2 =< x2 , y2 , xy >=< x2 >= 0 or < y2 >= 0 since A is (local) Gaussian.
Therefore, xy = 0. Thus, I 2 = 0, as desired.
The following example illustrates the statement (3) case 2 of Theorem 2.1.
Example 2.4. Let (A, m) := (Z(2) , 2Z(2) ) be a valuation domain, B := A ∝ A be the trivial
ring extension of A by A. Consider
f :
A
a
֒→
֒→
B
f (a) = (a, 0)
be an injective ring homomorphism and J := 2Z(2) ∝ Z(2) be a proper ideal of B. Then,
A ⊲⊳ f J is not Gaussian.
Proof. By application to statement (3) case 2 of Theorem 2.1, A ⊲⊳ f J is not Gaussian
since A is reduced and J 2 , 0.
Theorem 2.1 enriches the literature with new examples of non-arithmetical Gaussian
rings.
Example 2.5. Let (A, m) be an arithmetical ring which is not a field such that m2 = 0 (for
instance (A = Z/4Z, 2Z/4Z)), E be a non-zero mA −vector space, B := A ∝ E be the trivial
ring extension of A by E. Consider
f :
A
a
֒→
֒→
B
f (a) = (a, 0)
be an injective ring homomorphism and J := I ∝ E be a proper ideal of B with I be a
proper ideal of A. Then :
(1) A ⊲⊳ f J is Gaussian.
(2) A ⊲⊳ f J is not an arithmetical ring.
6
NAJIB MAHDOU AND MOUTU ABDOU SALAM MOUTUI
Proof. (1) It is easy to see that J 2 = 0, f (a)J = f (a)2 J = 0 for all a ∈ m. Hence, by (2) of
Theorem 2.1, A ⊲⊳ f J is Gaussian.
(2) We claim that A ⊲⊳ f J is not an arithmetical ring. Indeed, f (A) + J = (A ∝ 0) + (I ∝
E) = A ∝ E which is not an arithmetical ring (by [4, Theorem 3.1 (3)] since A is not a
field). Hence, A ⊲⊳ f J is not an arithmetical ring since the arithmetical property is stable
under factor rings.
Example 2.6. Let (A0 , m0 ) be a local Gaussian ring, E be a non-zero mA00 −vector space,
(A, m) := (A0 ∝ E, m0 ∝ E) be the trivial ring extension of A0 by E. Let E ′ be a mA −vector
space, B := A ∝ E ′ be the trivial ring extension of A by E ′ . Consider
f :
A
֒→
(a, e) ֒→
B
f ((a, e)) = ((a, e), 0)
be an injective ring homomorphism and J := 0 ∝ E ′ be a proper ideal of B. Then :
(1) A ⊲⊳ f J is Gaussian.
(2) A ⊲⊳ f J is not an arithmetical ring.
Proof. (1) By (3) case 1 of Theorem 2.1, A ⊲⊳ f J is Gaussian since f (A) ∩ J = (0) and
f (A) + J = B which is Gaussian by [4, Theorem 3.1 (2)] since A is Gaussian (since A0 is
Gaussian).
(2) A ⊲⊳ f J is not an arithmetical ring since f (A) + J = B is not an arithmetical ring (by [4,
Theorem 3.1 (3)], A is never a field).
Example 2.7. Let (A0 , m) be a non-arithmetical Gaussian local ring such that m2 , m, for
K[[X]] 2
2
instance (A0 , m) := (K[[X]] ∝ ( K[[X]]
(X) ) , XK[[X]] ∝ ( (X) ) ) (by [4, Theorem 3.1 (2) and
(3) ]). Consider A := A0 ∝ mA02 be the trivial ring extension of A0 by mA02 . Let I := 0 ∝ mm2 be
A
an ideal of A, B :=
A
0∝ m2
m
A0 ∝ 02
A0
m
0∝ m2 A0 ∝ m be a
m
∝ Am0 be a proper ideal
homomorphism and J := 0
(1) A ⊲⊳ f J is Gaussian.
(2) A ⊲⊳ f J is not an arithmetical ring.
ring, f : A → B be a non-injective ring
of B. Then:
Proof. (1) It is easy to check that f (a)J = 0 for all a ∈ m ∝ mA02 which is the maximal
ideal of A and J ⊆ Nilp(B). And so f (a)2 J = f (a)J = 0 for all a ∈ m ∝ mA02 , J 2 = 0 and
by [4, Theorem 3.1 (2)], A is Gaussian since A0 is Gaussian. Hence, by application to the
statement (2) of Theorem 2.1, A ⊲⊳ f J is Gaussian.
(2) A ⊲⊳ f J is not an arithmetical ring since A is not an arithmetical ring (since A0 is not an
arithmetical ring and so by [4, Lemma 2.2], A is not an arithmetical ring).
We need the following result to construct a new class of non-Gaussian Prüfer rings.
Proposition 2.8. Let (A, m) be a local total ring of quotients, B be a ring, f : A → B be
a ring homomorphism and J be a proper ideal of B such that J ⊆ Rad(B) and J ⊆ Z(B).
Then, the following statements hold:
(1) Assume that f is injective and f (A) ∩ J , (0). Then (A ⊲⊳ f J, m ⊲⊳ f J) is a local total
ring of quotients; In particular, A ⊲⊳ f J is a Prüfer ring.
(2) Assume that f is not injective. Then, (A ⊲⊳ f J, m ⊲⊳ f J) is a local total ring of quotients;
In particular, A ⊲⊳ f J is a Prüfer ring.
GAUSSIAN PROPERTY IN AMALGAMATED ALGEBRAS ALONG AN IDEAL
7
Proof. By Lemma 2.2, it is clear that (A ⊲⊳ f J, m ⊲⊳ f J) is local.
(1) Assume that f (A) ∩ J , (0). We claim that A ⊲⊳ f J is a total ring of quotients. Indeed,
let (a, f (a) + j) ∈ A ⊲⊳ f J, we prove that (a, f (a) + j) is invertible or zero-divisor element.
If a < m, then (a, f (a) + j) < m ⊲⊳ f J. And so (a, f (a) + j) is invertible in A ⊲⊳ f J. Assume
that a ∈ m. So, (a, f (a) + j) ∈ m ⊲⊳ f J. Since A is a total ring of quotients, there exists
0 , b ∈ A such that ab = 0. We have (a, f (a) + j)(b, f (b)) = (0, j f (b)). Using the fact
f (A) ∩ J , (0) and J ⊆ Z(B), there exists some 0 , f (c) ∈ J and 0 , k ∈ J such that
jk = 0 and so (c, k) ∈ A ⊲⊳ f J. It follows that (a, f (a) + j)(bc, f (b)k) = (0, 0). Hence,
there exists (0, 0) , (bc, f (b)k) ∈ A ⊲⊳ f J such that (a, f (a) + j)(bc, f (b)k) = (0, 0). Thus,
(A ⊲⊳ f J, m ⊲⊳ f J) is local total ring of quotients.
(2) Assume that f is not injective. Our aim is to show that A ⊲⊳ f J is a total ring of
quotients. We prove that for each element (a, f (a)+ j) of A ⊲⊳ f J is invertible or zero-divisor
element. Indeed, if a < m, then (a, f (a) + j) < m ⊲⊳ f J. And so (a, f (a) + j) is invertible in
A ⊲⊳ f J. Assume that a ∈ m. So, (a, f (a) + j) ∈ m ⊲⊳ f J. Since A is a total ring of quotients,
there exists 0 , b ∈ A such that ab = 0. We have (a, f (a) + j)(b, f (b)) = (0, j f (b)). Using
the fact f is not injective and J ⊆ Z(B), there exist some 0 , c ∈ Ker( f ) and 0 , k ∈ J
such that jk = 0 and (c, k) ∈ A ⊲⊳ f J. It follows that (a, f (a) + j)(bc, f (b)k) = (0, 0). Hence,
there exists (0, 0) , (bc, f (b)k) ∈ A ⊲⊳ f J such that (a, f (a) + j)(bc, f (b)k) = (0, 0). Thus,
(A ⊲⊳ f J, m ⊲⊳ f J) is a local total ring of quotients, completing the proof.
Also, Theorem 2.1 enriches the literature with new class of non-Gaussian Prüfer rings.
Example 2.9. Let (A, m) be a local total ring of quotients and I be a proper ideal of A such
that I 2 , 0. Then :
(1) A ⊲⊳ I is Prüfer.
(2) A ⊲⊳ I is not Gaussian.
Proof. (1) By (1) of Proposition 2.8, A ⊲⊳ I is a local total ring of quotients since A is a
local total ring of quotients, A ∩ I , (0) and I ⊆ m ⊆ Z(A).
(2) By Corollary 2.3, A ⊲⊳ I is not Gaussian since I 2 , 0.
Example 2.10. Let (A0 , m0 ) := (Z/2nZ, 2Z/2nZ) where n ≥ 2 be an integer and let
(A, m) := (A0 ∝ A0 , m0 ∝ A0 ) be the trivial ring extension of A0 by A0 . Consider E be
a non-zero A-module such that mE = 0, B := A ∝ E be the trivial ring extension of A by
E,
f : A
֒→ B
(a, e) ֒→ f ((a, e)) = ((a, e), 0)
be an injective ring homomorphism and J := m ∝ E be a proper ideal of B. Then, the
following statements hold :
(1) A ⊲⊳ f J is Prüfer.
(2) A ⊲⊳ f J is not Gaussian.
Proof. (1) By (1) of Proposition 2.8, A ⊲⊳ f J is a local total ring of quotients since f (A) ∩
J = m ∝ 0 , (0), A is a local total ring of quotients and J ⊆ Z(B). In particular, A ⊲⊳ f J is
a Prüfer ring.
(2) A ⊲⊳ f J is not Gaussian since A is not Gaussian (by [5, Example 3.6]).
Example 2.11. Let K be a field and let (A0 , m) := (K[[X, Y]], < X, Y >) be the ring of formal power series where X and Y are two indeterminate elements. Consider A := A0 ∝ mA02
be the trivial ring extension of A0 by mA02 . Note that I := 0 ∝ mm2 is an ideal of A. Let B := AI
8
NAJIB MAHDOU AND MOUTU ABDOU SALAM MOUTUI
A
be a ring, f : A → B be a non-injective ring homomorphism and J :=
ideal of B. Then:
(1) A ⊲⊳ f J is Prü f er.
(2) A ⊲⊳ f J is not Gaussian.
0∝ m02
I
Proof. (1) One can easily check that A is a local total ring of quotients, B :=
J :=
0∝ A2
m
0∝ m2
m
be a proper
A
0∝ m2
m
0∝
A
m
and J ⊆ Z(B). Moreover, J ⊆ Rad(B) = m ∝
A
m.
f
A
m
A∝
A
m,
since B is local with
f
Hence, by (2) of Proposition 2.8, A ⊲⊳ J is local total ring of
maximal ideal m ∝
quotients. Thus, A ⊲⊳ J is Prü f er.
(2) A ⊲⊳ f J is not Gaussian since A is not Gaussian (since A0 := K[[X, Y]] is a domain such
that w.dim(K[[X, Y]]) = 2).
References
1. J. Abuihlail, M. Jarrar and S. Kabbaj, Commutative rings in which every finitely generated ideal is quasiprojective, J. Pure Appl. Algebra 215 (2011) 2504-2511.
2. D.D. Anderson, GCD domains, Gauss’ Lemma, and contents of polynomials, Non-Noetherian commutative
ring theory, Math. Appl. Kluwer Acad. Publ., Dordrecht 520 (2000) 1-31.
3. D.D. Anderson and V. Camillo, Armendariz rings and Gaussian rings, Comm. Algebra 26 (1998) 2265-2272.
4. C. Bakkari, S. Kabbaj and N. Mahdou, Trivial extensions defined by Pr ü f er conditions, J. of Pure Appl.
Algebra 214 (2010) 53-60.
5. C. Bakkari, N. Mahdou and H. Mouanis, Prüfer-like Conditions in Subrings Retract and Applications,
Comm. Algebra 37 (2009) 47-55.
6. C. Bakkari and N. Mahdou, Gaussian polynomials and content ideal in pullbacks, Comm. Algebra 34 (2006)
2727-2732.
7. C. Bakkari and N. Mahdou, Prüfer-like conditions in pullbacks, Commutative algebra and its applications,
Walter de Gruyter, Berlin, (2009) 41-47.
8. S. Bazzoni and S. Glaz, Prüfer rings, Multiplicative Ideal Theory in Commutative Algebra, Springer, New
York, (2006) 55-72.
9. S. Bazzoni and S. Glaz, Gaussian properties of total rings of quotients, J. Algebra 310 (2007) 180-193.
10. M. Boisen and P. Sheldon, A note on pre-arithmetical rings, Acta. Math. Acad. Sci. Hungar 28 (1976) 257259.
11. M. Boisen and P.B. Sheldon, CPI-extensions : overrings of integral domains with special prime spectrums,
Canad. J. Math. 29 (1977) 722-737.
12. H. S. Butts and W. Smith, Prüfer rings, Math. Z. 95 (1967) 196-211.
13. M. Chhiti, M. Jarrar, S. Kabbaj and N. Mahdou, Pr ü f er conditions in an amalgameted duplication of a ring
along an ideal, Comm. Algebra, Accepted for publication.
14. A. Corso and S. Glaz, Gaussian ideals and the Dedekind-Mertens Lemma, Lecture Notes Pure Appl. Math.
217 (2001) 131-143.
15. M. D’Anna, C. A. Finocchiaro and M. Fontana, Amalgamated algebras along an ideal, Commutative algebra
and its applications, Walter de Gruyter, Berlin, (2009) 241-252.
16. M. D’Anna, C. A. Finocchiaro and M. Fontana, Properties of chains of prime ideals in amalgamated algebras
along an ideal, J. Pure Appl. Algebra 214 (2010) 1633-1641.
17. M. D’Anna, A construction of Gorenstein rings, J. Algebra 306 (2006) 507-519.
18. M. D’Anna and M. Fontana, The amalgamated duplication of a ring along a multiplicative-canonical ideal,
Ark. Mat. 45 (2007) 241-252.
19. M. D’Anna and M. Fontana, An amalgamated duplication of a ring along an ideal : the basic properties, J.
Algebra Appl. 6 (2007) 443-459.
20. L. Fuchs, Uber die Ideale arithmetischer Ringe, Comment. Math. Helv. 23 (1949) 334-341.
21. S. Glaz, Commutative coherent rings, Multiplicative ideal theory in commutative algebra, Springer, New
York, (2006) 55-72.
22. S. Glaz, The weak global dimension of Gaussian rings, Proc. Amer. Math. Soc. 133 (2005) 2507-2513.
GAUSSIAN PROPERTY IN AMALGAMATED ALGEBRAS ALONG AN IDEAL
9
23. S. Glaz, Prüfer conditions in rings with zero-divisors, CRC Press Series of Lectures in Pure Appl. Math. 241
(2005) 272-282.
24. S. Glaz and W. Vasconcelos, The content of Gaussian polynomials, J. Algebra 202 (1998) 1-9.
25. S. Glaz and W. Vasconcelos, Gaussian polynomials, Lecture Notes in Pure and Appl. Math. 185 (1997)
325-337.
26. M. Griffin, Prüfer rings with zero-divisors, J. Reine Angew Math. 239/240 (1969) 55-67.
27. W. Heinzer and C. Huneke, Gaussian polynomials and content ideals, Proc. Amer. Math. Soc. 125 (1997)
739-745.
28. J. A. Huckaba, Commutative Rings with Zero-Divisors, Marcel Dekker, New York, 1988.
29. K. A. Loper and M. Roitman, The content of a Gaussian polynomial is invertible, Proc. Amer. Math. Soc.
133 (2004) 1267-1271.
30. T. G. Lucas, Gaussian polynomials and invertibility, Proc. Amer. Math. Soc. 133 (2005) 1881-1886.
31. M. Nagata, Local Rings, Interscience, New York, 1962.
32. H. Prüfer, Untersuchungen uber teilbarkeitseigenschaften in korpern, J. Reine Angew. Math. 168 (1932)
1-36.
33. H. Tsang, Gauss’ Lemma, Ph.D. thesis, University of Chicago, Chicago, 1965.
Department of Mathematics, Faculty of Science and Technology of Fez, Box 2202, University S. M. Ben
Abdellah Fez, Morocco
E-mail address: [email protected]
E-mail address: moutu [email protected]
| 0math.AC
|
Controllability of networked MIMO systems ⋆
Lin Wang a, Guanrong Chen b, Xiaofan Wang a, Wallace K. S. Tang b
arXiv:1505.01255v4 [cs.SY] 31 Oct 2015
a
Department of Automation, Shanghai Jiao Tong University, and Key Laboratory of System Control and Information
Processing, Ministry of Education, Shanghai, China
b
Department of Electronic Engineering, City University of Hong Kong, Hong Kong, China
Abstract
In this paper, we consider the state controllability of networked systems, where the network topology is directed and weighted
and the nodes are higher-dimensional linear time-invariant (LTI) dynamical systems. We investigate how the network topology,
the node-system dynamics, the external control inputs, and the inner interactions affect the controllability of a networked
system, and show that for a general networked multi-input/multi-output (MIMO) system: 1) the controllability of the overall
network is an integrated result of the aforementioned relevant factors, which cannot be decoupled into the controllability of
individual node-systems and the properties solely determined by the network topology, quite different from the familiar notion
of consensus or formation controllability; 2) if the network topology is uncontrollable by external inputs, then the networked
system with identical nodes will be uncontrollable, even if it is structurally controllable; 3) with a controllable network topology,
controllability and observability of the nodes together are necessary for the controllability of the networked systems under
some mild conditions, but nevertheless they are not sufficient. For a networked system with single-input/single-output (SISO)
LTI nodes, we present precise necessary and sufficient conditions for the controllability of a general network topology.
Key words: Networked system; state controllability; structural controllability; directed network; MIMO LTI system.
1
Introduction
Complex networks of dynamical systems are ubiquitous
in nature and science, as well as in engineering and
technology. When control to a network is taken into consideration, the controllability of the network is essential,
which is a classical concept [Kalman, 1962] applicable
to multi-variable control systems [Gilbert, 1963, Shields
& Pearson, 1976], composite systems [Davison & Wang,
1975] and decentralized control systems [Kobayashi,
Hanafusa, & Yoshikawa, 1978, Tarokh, 1992], etc.
The subject of system controllability has been extensively studied over the last half a century. To date, various criteria have been well developed, including different
kinds of matrix rank conditions and substantial graphic
properties [Gilbert, 1963, Hautus, 1969, Lin, 1974, Davison & Wang, 1975, Glover & Silverman, 1976, Shields
& Pearson, 1976, Kobayashi, Hanafusa, & Yoshikawa,
⋆ Corresponding author X. F. Wang. Tel.: +86-21-54747511. Fax: +86-21-5474-7161.
Email addresses: [email protected] (Lin Wang),
[email protected] (Guanrong Chen),
[email protected] (Xiaofan Wang),
[email protected] (Wallace K. S. Tang).
Preprint submitted to Automatica
1978, Tarokh, 1992, Jarczyk, Svaricek, & Alt, 2011, Liu,
Slotine, & Barabási, 2011, Ruths, & Ruths, 2014]. One
closely-related subject is the controllability of multiagent systems, including the controllability of consensus
systems [Liu et al., 2008, Rahmani et al., 2009, Lou &
Hong, 2012,Ji, Lin, & Yu, 2012,Zhang, Camlibel, & Cao,
2012, Nabi-Abdolyousefi & Mesbahi, 2013, Ni, Wang, &
Xiong, 2013,Xiang et al., 2013], formation [Cai & Zhong,
2010], and pinning strategies [Chen, 2014].
Noticeably, many existing results on controllability are
derived under the assumption that the dimension of
the state of each node is one [Liu, Slotine, & Barabási,
2011, Lou & Hong, 2012, Zhang, Camlibel, & Cao,
2012, Nabi-Abdolyousefi & Mesbahi, 2013, Wang et al.,
2014]. However, most real-world networks of dynamical systems have higher-dimensional node states, and
many multi-input/multi-output (MIMO) nodes are interconnected via multi-dimensional channels. In this
paper we study the controllability of networked higherdimensional systems with higher-dimensional nodes
mainly for the MIMO setting.
In the literature, few results are developed for the controllability of networked higher-dimensional systems.
Controllability and observability of specified cartesian
3 November 2015
real (complex) matrices, IN the N × N identity matrix,
and diag(a1 , · · · , aN ) the N × N diagonal matrix with
diagonal elements a1 , · · · , aN . Denote by σ(A) the set of
all the eigenvalues of matrix A and by ⊗ the Kronecker
product.
product networks are investigated in [Chapman, NabiAbdolyousefi, & Mesbahi, 2014], while a general networked system is considered in [Zhou, 2015] with every
subsystem subject to external control input. It should be
noted that some recent studies have addressed consensus controllability of networks with MIMO nodes [Cai
& Zhong, 2010, Ni, Wang, & Xiong, 2013, Xiang et al.,
2013], where the controllability usually can be decoupled
into two independent parts: one is about the controllability of each individual node and the other is solely
determined by the network topology. A general setting
of complex dynamical networks cannot be decoupled,
however, as will be seen from the discussions below.
In a directed graph, an edge (i, j) is directed from i to
j, where i is the tail and j is the head of the edge. As reviewed for a directed graph in [Liu, Slotine, & Barabási,
2011], a matching is a set of edges that do not share any
common tail or head, and a node being the head of an
edge in the matching is called a matched node; otherwise, it is an unmatched node. A maximum matching
is a matching that contains the largest possible number
of edges in the graph. A perfect matching is a matching
which matches all nodes in the graph. A graph formed
by a sequence of edges {(vi , vi+1 ) | i = 1, · · · , ℓ − 1} with
no repeated node is called a path, denoted as v1 , · · · , vℓ ,
where v1 is the beginning and vℓ is the end of the path,
and vℓ is said to be reachable from v1 . If v1 , · · · , vℓ is a
path, then the graph formed by adding the edge (vℓ , v1 )
is a cycle. A graph without cycles is called a tree. The
node in a tree which can reach every other node is called
the root of the tree. A leaf in a rooted tree is a node of
degree 1 that is not the root.
This paper addresses networked MIMO LTI dynamical node-systems in a directed and weighted topology,
where there is no requirement for every subsystem to
have an external control input. Some controllability
conditions on the network topology, node dynamics,
external control inputs and inner interactions are developed, so that effective criteria can be obtained for determining the large-scale networked system controllability.
On one hand, some necessary and sufficient conditions
on the controllability of a networked system with SISO
higher-dimensional nodes are derived. On the other
hand, some interesting results on the controllability of
a networked system with MIMO nodes are obtained:
1) Under some mild conditions, controllability and observability of the nodes together are necessary for the
controllability of the networked systems but, nevertheless, they are not sufficient; 2) the controllability of the
network topology is necessary for the controllability of
the generated networked system; 3) the controllability
of individual node is necessary for the controllability of
chain-networks, but not necessary for cycle-networks;
4) interactions among the states of different nodes play
an important role in determining the controllability of
a general networked system. Surprisingly, for the same
network topology with the same node systems, the interactions among the states of nodes not only can lead
controllable nodes to form an uncontrollable networked
system, but also can assemble uncontrollable nodes into
a controllable networked system.
Specifically, the notion of system controllability includes
state controllability and structural controllability. For
an n-dimensional system ẋ = Ax + Bu, it is said to
be state controllable, if it can be driven from any initial
state to the origin in finite time by a piecewise continuous control input. (A, B) is state controllable if and only
if the controllability matrix (B, AB, A2 B, · · · , An−1 B)
has a full row rank [Kalman, 1962, Chui & Chen, 1998].
A parameterized system (A, B) (i.e., all of their nonzero
elements are parameters) is said to be structurally controllable, if it is possible to choose a set of nonzero parameter values such that the resulting system (A, B) is
state controllable [Lin, 1974]. In this paper, for brevity,
controllability always means state controllability unless
otherwise specified, e.g., structural controllability.
2.2
The rest of the paper is organized as follows: In Section 2,
some preliminaries and the general model of networked
MIMO LTI systems are presented. Controllability conditions on various networked systems are investigated in
Section 3. Finally, conclusions are drawn with some discussions in Section 4.
2
2.1
The networked system model
Consider a general directed and weighted network consisting of MIMO LTI node-systems in the following form:
N
ẋi = Axi + P βij Hyj ,
y = Cx ,
i
i
Preliminaries and the networked system
j=1
(1)
i = 1, 2, · · · , N,
in which xi ∈ Rn is the state vector and yi ∈ Rm the
output vector of node i, H ∈ Rn×m denotes the inner
coupling matrix, and βij ∈ R represent the communication channels between different nodes. As usual, assume
that βii = 0 and βij 6= 0 if there is an edge from node j
to node i, otherwise βij = 0, for all i, j = 1, 2, · · · , N .
Preliminaries
Throughout, let R and C denote the real and complex
numbers respectively, Rn (Cn ) the vector space of real
(complex) n-vectors, Rn×m (Cn×m ) the set of n × m
2
and 3 can not be matched simultaneously. If each node
has a higher-dimensional state with
When subjected to control inputs, the above networked
system becomes
ẋi = Axi +
N
X
βij HCxj + δi Bui ,
A=
i = 1, 2, · · · , N, (2)
j=1
Here and throughout, for statement simplicity a network
consisting of more than one dynamical node, with or
without control inputs (e.g., (1) and (2)), will be called
a networked system.
,B =
"
1 0
0 1
#
,H =
"
h1 0
0 h2
#
,C =
"
1 0
0 1
Example 2.2 Consider a network of two mutually connected identical nodes, with β12 = β21 = 1. Suppose that
both nodes have external control inputs, i.e. δ1 = δ2 = 1,
and
"
#
" #
" #
10
1
0
A=
,B =
,H =
, C = [1 0] .
11
0
1
(3)
which represent the network topology and the external
input channels of the networked system (2), respectively.
Let X = [xT1 , xT2 , · · · , xTN ]T be the whole state of the
networked system, and U = [uT1 , uT2 , · · · , uTN ]T the total
external control input. Then, this networked system can
be rewritten in a compact form as
Ẋ = ΦX + ΨU ,
a21 a22
#
The following three examples show that, even the network is a cycle having a perfect matching, the controllability of (A, B) is neither necessary nor sufficient for the
controllability of the whole networked system.
Denote
and ∆ = diag(δ1 , · · · , δN ),
a11 a12
where aij 6= 0 and hi 6= 0, i, j = 1, 2, then obviously
(A, B) is controllable and (A, C) is observable. However,
based on the results from Subsection 3.2 below, one knows
that the networked system is uncontrollable for any matrix A, although it is structurally controllable due to the
existence of self-matched cycle in every MIMO node.
where ui ∈ Rp is the external control input to node i,
B ∈ Rn×p , with δi = 1 if node i is under control, but
otherwise δi = 0, for all i = 1, 2, · · · , N . To avoid trivial
situations, always assume that N ≥ 2.
L = [βij ] ∈ RN ×N
"
(4)
It is easy to check that (A, B) is controllable. However,
the networked system (2) is uncontrollable, although each
node has an independent external input.
(5)
Example 2.3 Consider a simple network of two mutually connected identical nodes, with β12 = β21 = 1,
δ1 = 1, δ2 = 0, and
with
Φ = IN ⊗ A + L ⊗ HC,
Ψ = ∆ ⊗ B.
In this paper, the focus is on how the network topology
(described by the matrix L), the node-system (A, B, C),
the external control input (determined by the matrix
∆), and the inner interactions specified by H affect the
controllability of the whole networked system.
2.3
A=
"
10
11
#
,B =
" #
1
0
,H =
" #
0
1
, C = [0 1] .
Then, (A, B) is controllable, (A, C) is observable. However, the networked system (2) is uncontrollable.
Some counter-intuitive examples
Example 2.4 Consider a network of two mutually connected identical nodes, with β12 = β21 = 1, δ1 = 1,
δ2 = 0, and
In [Liu, Slotine, & Barabási, 2011], it is shown that a
network is structurally controllable if and only if there is
an external input on each unmatched node and there are
directed paths from controlled nodes with input signals
to all matched nodes. For the networked system (4)-(5)
formed by nodes with higher-dimensional state vectors,
however, its controllability can be much more complicated, as demonstrated by the following example.
A=
"
10
11
#
,B =
" #
0
1
,H =
" #
1
0
, C = [0 1] .
Then, (A, B) is uncontrollable. However, the networked
system (2) is controllable, although there is only one node
under external control.
Example 2.1 Consider a network of three identical
nodes, with β21 = β31 = 1, and δ1 = 1, δ2 = δ3 = 0. It
is not structurally controllable with one external input
if each node has a one-dimensional state, since nodes 2
Comparing the above three examples, their network
topologies are the same and their node-system matrices
A are identical. However, these networked systems have
3
#
,
very different controllabilities. The interactions among
the states of nodes not only can lead controllable nodes
to form an uncontrollable network, but also can assemble uncontrollable nodes into a controllable network.
Proof: Suppose that (A, C) is unobservable. Then,
there exist an s0 ∈ σ(A) and a nonzero vector ξ ∈ Cn
such that
Cξ = 0 and (s0 I − A)ξ = 0 .
3
Main Results
Consider the matrix Φs0 , s0 I − Φ, and partition
it into N column blocks, [Φ1s0 , Φ2s0 , · · · , ΦN
s0 ], with
Φis0 = [β1i (HC)T , · · · , βi−1,i (HC)T , (sI − A)T , βi+1,i
(HC)T , · · · , βN i (HC)T ]T , which corresponds to node i.
The controllability of the networked system (4)-(5) is
considered in this section, where results on general network topologies are obtained, with specific and precise
conditions obtained for some typical networks including
tress and cycles.
Based on formula (7), one has
First, recall the Popov-Belevitch-Hautus (PBH) rank
condition [Hautus, 1969]: the networked system (4)-(5)
is controllable if and only if
rank(sIN ·n − Φ, Ψ) = N · n
Φis0 ξ = 0,
i = 1, · · · , N,
(8)
which implies that rank(Φis0 ) ≤ n − 1. Therefore,
rank(Φs0 ) ≤ N · (n − 1). In view of rank(Ψ) ≤
m · rank(B) < N , one has rank(s0 I − Φ, Ψ) < N · n,
showing that the networked system (4)-(5) is uncontrollable.
(6)
is satisfied for any complex number s.
3.1
(7)
A general network topology
Theorem 4 If (L, ∆) is uncontrollable, then the networked system (4)-(5) is uncontrollable.
Based on the PBH rank condition (6), one can prove the
following results.
Proof: If (L, ∆) is uncontrollable, then there exist
an s0 ∈ σ(L) and a nonzero vector ξ ∈ C1×N such that
Theorem 1 If there exists one node without incoming
edges, then to reach controllability of the networked system (4)-(5), it is necessary that (A, B) is controllable and
moreover an external control input is applied onto this
node which has no incoming edges.
ξ(s0 I − L) = 0 ,
ξ∆ = 0 .
(9)
Therefore,
(ξ ⊗ I)Ẋ = (ξ ⊗ I) (I ⊗ A + L ⊗ HC)X + (∆ ⊗ B)U
= (ξ ⊗ A + s0 ξ ⊗ HC)X
= ξ ⊗ (A + s0 HC) X ,
Proof: Assume that node i does not have any incoming edge. Then, the ith block row of Φ in (5) becomes
[0, · · · , 0, A, 0, · · · , 0]. If there is no external control input onto node i, that is, δi = 0, then for any s0 ∈ σ(A),
the row rank of [s0 I − Φ, Ψ] will be reduced at least by 1.
If (A, B) is uncontrollable, then there exists an s0 ∈ σ(A)
such that rank(s0 I − A, B) ≤ n − 1, which will also result in the reduction of the rank of [s0 I − Φ, Ψ].
that is,
N
X
i=1
Theorem 2 If there exists one node without external
control inputs, then for networked system (4)-(5) to be
controllable, it is necessary that (A, HC) is controllable.
ξi xi
!′
= (A + s0 HC)
N
X
ξi xi .
(10)
i=1
This implies that the variable
N
P
ξi xi is unaffected by
i=1
the external control input U . For the zero initial state
N
P
ξi xi (t0 ) = 0.
xi (t0 ) = 0, i = 1, · · · , N , one has
Proof: Assume that node i does not have any external input, that is, δi = 0. If (A, HC) is uncontrollable, then there exist an s0 ∈ σ(A) and a nonzero vector ξ ∈ C1×n , such that ξ(s0 I − A) = 0 and ξHC = 0.
Let α = [0, · · · , 0, ξ, 0 · · · , 0] with ξ located at the ith
block. Then, it is easy to verify that α(s0 I − Φ) = 0 and
αΨ = 0.
i=1
Moreover,
N
P
ξi xi (t) = 0 for all t > t0 , because of the
i=1
uniqueness of the solution to the linear equation (10).
e , [e
Consequently, for any state X
xT1 , · · · , x
eTN ]T with
N
P
ξi x
ei 6= 0, there is no external control input U that
Theorem 3 If the number of nodes with external control inputs is m, and N > m · rank(B), then for the networked system (4)-(5) to be controllable, it is necessary
that (A, C) is observable.
i=1
can drive the networked system (4)-(5) to traverse from
e Thus, it is uncontrollable.
state 0 to X.
4
Proof: Necessity. From Theorem 2 and Lemma 3.1,
it follows that condition (i) is necessary. From Theorem
3, it follows that condition (ii) is also necessary.
If the network is not structurally controllable by external inputs, then (L, ∆) is uncontrollable, and thus the
networked system will be uncontrollable. Since a network having more leaf nodes than nodes with external
control input is not structurally controllable [Ruths, &
Ruths, 2014], the following result comes accordingly.
Now, suppose that condition (iii) is not necessary. Then,
there exist an s0 ∈ σ(A) and a nonzero matrix κ ∈ Γ(s0 )
such that
κL = 0 .
Corollary 3.1 If there are more leaf nodes than the
nodes with external control inputs, then the networked
system (4)-(5) is uncontrollable.
For matrix M ∈ Cp×q , denote by vec(M ) ∈ Cpq×1
the vectorization of matrix M formed by stacking the
columns of M into a single column vector. Furthermore,
let α = vec(κ)T . Since κ ∈ Γ(s0 ), it is easy to verify that
αΨ = 0 and
If (L, ∆) is controllable, Examples 3.2, 3.4, and 3.6 discussed bellow for specific network topologies show that,
for the networked system (4)-(5) to be controllable, it
is not sufficient to ensure (A, B) and (A, HC) be both
controllable and (A, C) be observable. Moreover, Example 3.4 shows that even every node has an external input
and (A, B) is controllable, the networked system may
still be uncontrollable.
α(s0 I − Φ) = α(IN ⊗ (s0 I − A) − L ⊗ HC)
= −α(L ⊗ HC)
= −vec(C T H T κL)T = 0 ,
which contradicts the network controllability.
Next, some necessary and sufficient conditions for the
controllability of networked systems with SISO nodes
are developed. First, a lemma is given.
Finally, suppose that condition (iv) is not necessary.
Then, there exists an s0 ∈
/ σ(A) satisfying
Lemma 3.1 Assume that C ∈ R1×n is nonzero. Then,
(A, HC) is controllable if and only if (A, H) is controllable.
rank(I − Lγ0 , ∆η0 ) < N ,
with γ0 = C(s0 I − A)−1 H and η0 = C(s0 I − A)−1 B.
Thus, there exists a nonzero vector ζ = [ζ1 , · · · , ζN ] ∈
C1×N , such that
Proof: Since C ∈ R1×n , one has H ∈ Rn×1
and rank(HC) = 1. Therefore, rank(sI − A, HC) =
rank(sI − A, H), leading to the conclusion.
ζ(I − Lγ0 ) = 0
Before moving on to the theorem, some new notations
are needed. Denote the set of nodes with external control
inputs by
U = {i | δi 6= 0, i = 1, · · · , N } .
Let α = [α1 , · · · , αN ] with αi = ζi C(so I − A)−1 . Then,
since ζ 6= 0, one has α 6= 0. Moreover,
(11)
αΨ = (ζ ⊗ C(s0 I − A)−1 ) · (∆ ⊗ B)
= (ζ∆) ⊗ (C(so I − A)−1 B)
= ζ∆η0 = 0 ,
(12)
and
For any s ∈ σ(A), define a matrix set
Γ(s) =
(
[αT1 , · · · , αTN ]
αi ∈ Γ1 (s) for i ∈
/U
αi ∈ Γ2 (s) for i ∈ U
)
,
α(s0 I − Φ)
= (ζ ⊗ C(so I − A)−1 ) · (IN ⊗ (s0 I − A) − L ⊗ HC)
= ζ ⊗ C − ζL ⊗ (C(so I − A)−1 H)C
= (ζ − ζLγ0 ) ⊗ C = 0 .
where
Γ1 (s) = {ξ ∈ C1×n | ξ(sI − A) = 0},
Γ2 (s) = {ξ ∈ C1×n | ξB = 0, ξ ∈ Γ1 (s)}.
This is also in conflict with the controllability of the
networked system.
Theorem 5 Suppose that |U| < N , B ∈ Rn×1 , and C ∈
R1×n . Then, the networked system (4)-(5) is controllable
if and only if the following hold:
(i)
(ii)
(iii)
(iv)
and ζ∆η0 = 0 .
Sufficiency. For s ∈ C, suppose that there exists a vector
α = [α1 , · · · , αN ], with αi ∈ C1×n , such that α(sI −
Φ) = 0 and αΨ = 0. That is,
(A, H) is controllable;
(A, C) is observable;
for any s ∈ σ(A) and κ ∈ Γ(s), κL 6= 0 if κ 6= 0;
for any s ∈
/ σ(A), rank(I − Lγ, ∆η) = N , with
γ = C(sI − A)−1 H and η = C(sI − A)−1 B.
αi (sI − A) −
X
j6=i
5
βji αj HC = 0, i = 1 · · · , N,
(13)
Then, from (14) and (20), it follows that ζi C(sI −
A)−1 B = 0 for i ∈ U, which is equivalent to
and
αi B = 0 ,
i ∈ U.
(14)
ζ∆η = 0 .
If s ∈ σ(A), then rank(sI − A) < n. From (13), it follows
that, for all i = 1, · · · , N ,
X
βji αj H = 0 .
Consequently, by combining it with (22) and condition
(iv), one has ζ = 0, which together with (20) imply that
α = 0.
(15)
j6=i
"
If not, then rank
It follows from the above analysis that, for any s ∈ C,
the row vectors of matrix [sI − Φ, Ψ] are linearly independent, hence rank(sI − Φ, Ψ) = N · n. Thus, the networked system (4)-(5) is controllable.
#!
C
= rank(sI − A) < n,
sI − A
which contradicts with the observability of (A, C). Moreover, based on (13), one has
αi (sI − A) = 0 ,
i = 1··· ,N .
Next, some typical network structures, i.e., trees and
circles are discussed in details.
(16)
3.2
Therefore, for all i = 1 · · · , N , one has
X
βji αj (sI − A) = 0 .
Corollary 3.2 Consider a tree-network, in which every
node is reachable from the root, and only the root has an
external control input. If there is more than one leaf node
in the tree, then the networked system is uncontrollable.
Consequently, a star networked system with N > 2 is
uncontrollable.
Combining it with (15) and the controllability of (A, H),
one obtains
X
βji αj = 0 , i = 1 · · · , N.
(18)
j6=i
[αT1 , · · ·
A tree with only one leaf is a chain, which could be described by a path 1 → 2 → · · · → n. Based on Theorem
1, node 1 should be under external control.
, αTN ].
In view of (14), (16) and
Next, let κ =
(18), it is easy to verify that κL = 0 with αi (sI − A) = 0
for i = 1, · · · , N , and αi B = 0 for i ∈ U. Therefore, by
condition (iii), one has α = 0.
This chain-networked system (4)-(5) has
If s ∈
/ σ(A), then sI − A is invertible. From (13), one has
αi =
−1
βji αj HC(sI − A)
,
i = 1, · · · , N.
P
(19)
βji αj H. Then, for i = 1, · · · , N ,
j6=i
αi = ζi C(sI − A)−1 ,
···
0
,
..
.
A
B
0
Ψ = .
..
0
, (24)
From Theorems 1 and 2, one obtains the following result.
P
βji αj H =
P
P
βji ζj C(sI − A)−1 H
j6=i
j6=i
=
0
where βi,i−1 6= 0 for i = 2, · · · , N , and βij = 0 for
j 6= i − 1, i = 1, · · · , N .
(20)
and
ζi =
A
β21 HC A
Φ=
..
..
.
.
0
· · · βN,N −1 HC
j6=i
Let ζi =
Trees
In view of Corollary 3.1, the following result comes easily.
(17)
j6=i
X
(23)
βji ζj γ .
Corollary 3.3 A necessary condition for the controllability of the chain networked system (4)-(24) is that
(A, B) and (A, HC) are both controllable.
(21)
j6=i
Remark 3.1 The observability of (A, C) is not necessary for the controllability of the chain networked system
(4)-(24), as shown by the following example.
Let ζ = [ζ1 , · · · , ζN ], and rewrite (21) as
ζ(I − Lγ) = 0 .
(22)
6
Condition (iv) in Theorem 5 is automatically satisfied
for the chain-network, since
Example 3.1 Consider a chain-network of two identical nodes, with β21 = 1 and
A=
"
10
11
#
, B=
"
10
01
#
, H=
" #
1
0
β21
L= .
..
0
, C = [1 0] .
It is easy to check that (A, B) and (A, HC) are both controllable and (A, C) is unobservable. The coupled networked system (24) has rank(Ψ, ΦΨ, Φ2 Ψ, Φ3 Ψ) = 4, indicating that it is controllable. Therefore, the observability of (A, C) is indeed not necessary.
3.3
A=
1 2
5 4
, B=
"
2
−1
#
,H =
"
−1 1
−4 1
#
, C=
"
1 0
0 1
#
···
0
..
.
· · · βN,N −1
0
,
..
.
0
(25)
Cycles
Now, assume that the network topology is a cycle. Since
the cycle has a perfect matching, one external input is
enough for the structural controllability, which can be
added to any node in the cycle. Without loss of generality, assume that node 1 is under external control.
The cycle networked system has
Example 3.2 Consider a chain-network of two nodes,
with β21 = 1 and
#
0
and correspondingly for s ∈
/ σ(A), rank(I − Lγ) = N
with γ = C(sI − A)−1 H.
Remark 3.2 Suppose that (A, B) and (A, HC) are both
controllable and (A, C) is observable. However, these are
not sufficient to guarantee the controllability of the chain
networked system (4)-(24), as shown by the following
example.
"
0
A
0
···
β21 HC A
Φ=
..
..
.
.
0
· · · βN,N −1 HC
h
iT
Ψ = BT 0 · · · 0 ,
.
It is easy to check that (A, B) and (A, HC) are both controllable and (A, C) is observable. However, the coupled
networked system (24) has rank(6 · I − Φ, Ψ) = 3 < 4,
indicating that it is uncontrollable.
β1N HC
..
.
A
,
(26)
where β1N 6= 0, βi,i−1 6= 0 for i = 2, · · · , N , and βij = 0
otherwise.
From Theorem 2, the controllability of (A, HC) is necessary for the controllability of the networked system
(4)-(26).
If the input and output channels are all one-dimensional,
namely, if all the nodes are SISO, then Theorem 5 can
be restated as follows.
Remark 3.3 The controllability of (A, B) and the observability of (A, C) are not necessary for the controllability of the cycle networked system (4)-(26), as can be
seen from the following example.
Corollary 3.4 Assume that B ∈ Rn×1 and C ∈ R1×n .
The chain networked system (4)-(24) is controllable if
and only if (A, B) and (A, H) are both controllable and
(A, C) is observable.
Example 3.3 Consider a cycle-network of three identical nodes, with β13 = β21 = β32 = 1 and
Proof: Construct κ = [αT1 , · · · , αTN ] according to
condition (iii) of Theorem 5, such that α1 ∈ Γ2 satisfies
α1 (sI − A) = 0 and α1 B = 0, and moreover αi ∈ Γ1
satisfies αi (sI − A) = 0 for i = 2, · · · , N . In view of
κL = [β21 αT2 , β32 αT3 , · · · , βN,N −1 αTN , 0], the condition
κL 6= 0 for κ 6= 0 is equivalent to α1 = 0, which implies the equivalence with the controllability of (A, B).
Therefore, condition (iii) in Theorem 5 is equivalent to
the controllability of (A, B).
0 1 00
1 0 0
0 1
0 0 1 0
0 1 0
0 0
A=
, B =
, H =
,
0 0 0 1
0 0 1
0 0
0 0 01
0 0 0
1 0
"
#
0 1 00
C=
.
0 0 10
7
cal nodes, with β13 = −1, β21 = β32 = 1, and
1
1
1 8 7
A = 4 5 6,B = 0,H = 1
, C = [4 3 6] .
1
1
1 2 3
It is easy to check that (A, B) is uncontrollable and (A, C)
is unobservable. However, the coupled system (26) has
rank(Ψ, ΦΨ, Φ2 Ψ, · · · , Φ11 Ψ) = 12, indicating that the
networked system is controllable.
Remark 3.4 Conditions that (A, B) and (A, HC) are
both controllable and (A, C) is observable together are
not sufficient to guarantee the controllability of the cycle
networked system (4)-(26), as shown by the following
example.
It is easy to check that (A, B) and (A, H) are both controllable and (A, C) is observable. However, the coupled
system (26) has rank(Ψ, ΦΨ, Φ2 Ψ, · · · , Φ8 Ψ) = 8 < 9,
showing that the networked system is uncontrollable.
Example 3.4 Consider a cycle-network of three identical nodes, with β13 = β21 = β32 = 1 and
A=
"
1 1
0 1
#
, B=
" #
1
1
, H=
"
0 0
0 1
#
, C=
"
1 0
0 1
#
For circle network with SISO nodes, a new criterion for
the controllability is given as follows.
Theorem 6 Assume that B ∈ Rn×1 and C ∈ R1×n .
The cycle networked system (4)-(26) is controllable if and
only if (A, H) is controllable, (A, C) is observable, and
moreover
rank I − bHC(sI − A)−1 , B = n , ∀s ∈
/ σ(A) , (28)
.
It is easy to check that (A, B) and (A, HC) are both controllable and (A, C) is observable. However, although every node is driven by a control input, the whole networked
system with
A
0 HC
Φ = HC A
0
,
0 HC A
B 0 0
Ψ = 0 B 0
0 0 B
where b = β1N
Proof: For the cycle-network,
(27)
A=
0 0
, B=
" #
1
0
, H=
" #
0
1
0
···
β1,N
..
.
0
,
(29)
which is invertible, therefore condition (iii) in Theorem
5 is automatically satisfied.
In the following, it will be proved that condition (iv) in
Theorem 5 is equivalent to the above rank condition.
Note that the two conditions are both given in terms
of matrix ranks, yet one is about the network topology
which is N -dimensional while the other is about a subsystem which is only n-dimensional.
Example 3.5 Consider a cycle-network of three identical nodes, with β13 = β21 = β32 = 1 and
#
0
β21 0
L= . .
.. . .
0 · · · βN,N −1
Remark 3.5 Even every node is SISO, the controllability of (A, B) is not necessary for the controllability of the
networked system (4)-(26), as shown by the following example.
0 1
βi+1,i γ N −1 , with γ = C(sI −A)−1 H.
i=1
has rank(Ψ, ΦΨ, Φ2 Ψ, · · · , Φ5 Ψ) = 5 < 6, implying that
the networked system is uncontrollable.
"
NQ
−1
, C = [1 0] .
If γ = 0, then the two matrices both have full ranks. In
the following, assume that γ 6= 0.
It is easy to check that (A, H) is controllable, (A, C) is
observable, and (A, B) is uncontrollable. However, the
coupled system (26) has rank(Ψ, ΦΨ, Φ2 Ψ, · · · , Φ5 Ψ) =
6, indicating that the networked system is controllable.
If rank(I − Lγ, ∆η) < N , then there exists a nonzero
vector k = [k1 , · · · , kN ] ∈ C1×N such that
k = kLγ
Remark 3.6 Assume that every node is SISO. The conditions that (A, B) and (A, H) are controllable and (A, C)
is observable together are not sufficient to guarantee the
controllability of the networked system (4)-(26), as shown
by the following example.
and k∆η = 0 ,
that is,
ki
= ki+1 βi+1,i γ,
kN = k1 β1,N γ,
k1 η = 0 .
Example 3.6 Consider a cycle-network of three identi-
8
i = 1, · · · , N − 1,
(30)
From the recursion formula (30), it follows that
k1 6= 0 since k 6= 0. Moreover, k1 = k2 β21 γ =
NQ
−1
βi+1,i γ N −1 = k1 bγ, which implies that
· · · = kN
4
We have investigated a network consisting of MIMO LTI
node-systems (A, B, C), in a topology described by matrix L with inner interactions described by matrix H,
with or without control inputs determined by matrix ∆.
We have studied the integrated effects of the network
topology L, node-system (A, B, C) and inner interactions H on the controllability of the networked system.
i=1
bγ = 1. Choose ξ = k1 C(sI − A)−1 . Then, ξ 6= 0,
ξB = k1 C(sI − A)−1 B = k1 η = 0, and ξ(I − bHC(sI −
A)−1 ) = k1 C(sI − A)−1 − k1 bγC(sI − A)−1 = 0, which
implies that rank(I − bHC(sI − A)−1 , B) < n.
If rank(I − bHC(sI − A)−1 , B) < n, then there exists a
nonzero vector ξ ∈ C1×n , satisfying
ξ = bξHC(sI − A)−1 ,
We have shown that a networked system in the MIMO
setting is uncontrollable if the network topology L is
uncontrollable by external inputs through ∆, e.g., a nontrivial star-network with a single input to its root. For a
networked system to be controllable, the controllability
of (A, B) and (A, HC), as well as the observability of
(A, C), are necessary under some conditions; but they
are not sufficient in general, even for the cycle-network
which has a perfect matching.
ξB = 0 .
Since ξ 6= 0, one has b 6= 0 and ξH 6= 0. Moreover,
ξH = bξHγ, which implies that bγ = 1. Now, define
k1 = bξH,
(31)
kN = β1,N k1 γ,
ki = ki+1 βi+1,i γ ,
For SISO nodes with higher-dimensional state vectors,
we have presented necessary and sufficient conditions for
the controllability of some networked systems, including trees, cycles as well as a general network topology.
These results not only provide precise and efficient criteria for determining the controllability of large-scale networked systems, by means of verifying some properties
of a few matrices of lower dimensions, but also provide
some general guidelines on how to assemble uncontrollable nodes to form a controllable networked system,
which is deemed useful in engineering practice.
i = 2, · · · , N − 1.
One can easily verify that
k1 η = bξHC(sI − A)−1 B = ξB = 0 ,
k2 β21 γ = kN
N
−1
Y
Conclusions
βi+1,i γ N −1 = k1 bγ = k1 .
i=1
Therefore, k = kLγ and k∆η = 0 with k = [k1 , · · · , kN ],
which implies that rank(I − Lγ, ∆η) < N .
If each node-system (described by higher-dimensional
matrices (A, B, H, C)) is viewed as a sub-network, then
the networked system studied in this paper can also be
considered as an interdependent network (or interconnected network, multi-layer network, network of networks, multiplex network, etc. [Boccaletti et al., 2014]);
therefore, the results obtained in this paper should shed
lights onto studying the controllability of such complex
networks.
Looking back to Example 3.5, it can be seen that σ(A) =
{0, 0}. And, for any s 6= 0, one has b = s−4 and
rank I − bHC(sI − A)−1 , B
"
# " #!
1
0
1
= rank
,
= 2.
−5
−6
−s 1 − s
0
Acknowledgements
Moreover, (A, H) is controllable and (A, C) is observable. Therefore, from Theorem 6, it follows that the networked system in Example 3.5 is controllable.
The authors thank Professor Tong Zhou from Tsinghua
University, Beijing for some helpful discussions. This
work was supported by the National Natural Science
Foundation of China under Grant Nos. 61374176 and
61473189, and the Science Fund for Creative Research
Groups of the National Natural Science Foundation of
China (No. 61221003), as well as the Huawei Technologies Co. Ltd.
Looking back to Example 3.6, it can be seen that for
s=2∈
/ σ(A), one has C(2I −A)−1 H = −1, b = −1, and
rank(I − bHC(2I − A)−1 , B)
1
1 −1 0
= rank 0 0 0 , 0
= 2 < 3.
1
0 −1 1
References
[Boccaletti et al., 2014] Boccaletti, S., Bianconi, G., Criado, R.,
et al. (2014). The structure and dynamics of multilayer
networks. Phys. Reports, 544(1), 1-122.
Therefore, from Theorem 6, it follows that the networked
system in Example 3.6 is uncontrollable.
9
[Cai & Zhong, 2010] Cai, N., & Zhong, Y. S. (2010). Formation
controllability of highorder linear time-invariant swarm
systems. IET Control Theory and Applications, 4(4), 646654.
from a graph-theoretic perspective. SIAM J. Contr. Optim.,
48(1), 162–186.
[Ruths, & Ruths, 2014] Ruths, J., & Ruths, D. (2014). Control
profiles of complex networks. Science, 343(6177), 1373–1376.
[Chapman, Nabi-Abdolyousefi, & Mesbahi, 2014] Chapman, A.,
Nabi-Abdolyousefi, M., & Mesbahi M. (2014). Controllability
and observability of network-of-networks via cartesian
products. IEEE Transactions on Automatic Control, 59(10),
2668-2679.
[Shields & Pearson, 1976] Shields, R. W., & Pearson, J. B.
(1976). Structural controllability of multiinput linear
systems. IEEE Transactions on Automatic Control, 21(2),
203–212.
[Tarokh, 1992] Tarokh, M. (1992). Measures for controllability,
observability, and fixed modes. IEEE Transactions on
Automatic Control, 31(8), 1268–1273.
[Chen, 2014] Chen, G. (2014). Pinning control and synchronization on complex dynamical networks. International Journal
of Control, Automation, and Systems, 12(2), 221-230.
[Wang et al., 2014] Wang, B., Gao, L., Gao, Y., Deng, Y., &
Wang, Y. (2014). Controllability and observability analysis
for vertex domination centrality in directed networks.
Scientific Reports, 4, 5399.
[Chui & Chen, 1998] Chui, C. K., & Chen, G. (1998), Linear
Systems and Optimal Control, Springer-Verlasg.
[Davison & Wang, 1975] Davison, E. J., & Wang, S. H. (1975).
New results on the controllability and observability of
composite systems. IEEE Transactions on Automatic
Control, 20(1), 123-128.
[Xiang et al., 2013] Xiang, L., Zhu, J. H., Chen, F., & Chen,
G. (2013). Controllability of weighted and directed networks
with nonidentical node dynamics. Math. Probl. in Eng., ID
405034.
[Gilbert, 1963] Gilbert, E. G. (1963). Controllability and
observability in multivariable control systems. SIAM J.
Contr., 2(1), 128–151.
[Zhang, Camlibel, & Cao, 2012] Zhang, S., Camlibel, M. K., &
Cao, M. (2012). Controllability of diffusively-coupled multiagent systems with general and distance regular coupling
topologies. Proc. IEEE Conf. Decis. Contr., 759-764.
[Glover & Silverman, 1976] Glover, K., & Silverman, L. M.
(1976). Characterization of structural controllability. IEEE
Transactions on Automatic Control, 21(4), 534–537.
[Zhou, 2015] Zhou, T. (2015). On the controllability and
observability of networked dynamic systems. Automatica, 52,
63-75.
[Hautus, 1969] Hautus, M. L. J. (1969). Controllability and
observability conditions of linear autonomous systems, Ned.
Akad. Wetenschappen Proc. Ser. A, 72–443.
[Jarczyk, Svaricek, & Alt, 2011] Jarczyk, J. C., Svaricek, F., &
Alt, B. (2011). Strong structural controllability of linear
systems revisited. IEEE Conf. Decis. Contr., 1213–1218.
[Ji, Lin, & Yu, 2012] Ji, Z., Lin, H., & Yu, H. (2012). Leaders
in multi-agent controllability under consensus algorithm and
tree topology. Systems & Control Letters, 61, 918-925.
[Kalman, 1962] Kalman, R. E. (1962). Canonical structure of
linear dynamical systems. Proc. National Acad. Sci. USA,
48, 596-600.
[Kobayashi, Hanafusa, & Yoshikawa, 1978] Kobayashi, H., Hana
-fusa, H., & Yoshikawa, T. (1978). Controllability under
decentralized information structure. IEEE Transactions on
Automatic Control, 23(1), 182–188.
[Lin, 1974] Lin, C. T. (1974). Structural controllability. IEEE
Transactions on Automatic Control, 19(3), 201–208.
[Liu et al., 2008] Liu, B., Chu, T., Wang, L., & Xie, G. (2008).
Controllability of a leader-follower dynamic network with
switching topology. IEEE Transactions on Automatic
Control, 53(4), 1009-1013.
[Liu, Slotine, & Barabási, 2011] Liu, Y. Y., Slotine, J. J., &
Barabási, A. L. (2011). Controllability of complex networks.
Nature, 473(7346), 167–173.
[Lou & Hong, 2012] Lou, Y., & Hong, Y. (2012). Controllability
analysis of multi-agent systems with directed and weighted
interconnection. Int. J. Contr., 85(10), 1486–1496.
[Nabi-Abdolyousefi & Mesbahi, 2013] Nabi-Abdolyousefi, M., &
Mesbahi, M. (2013). On the controllability properties
of circulant networks. IEEE Transactions on Automatic
Control, 58(12), 3179–3184.
[Ni, Wang, & Xiong, 2013] Ni, W., Wang, X., & Xiong, C. (2013).
Consensus controllability, observability and robust design for
leader-following linear multi-agent systems. Automatica, 49,
2199–2205.
[Rahmani et al., 2009] Rahmani, A., Ji, M., Mesbahi, M., &
Egerstedt, M. (2009). Controllability of multi-agent systems
10
| 3cs.SY
|
arXiv:1209.1711v2 [cs.PL] 9 Jan 2018
Programming Languages for Scientific
Computing
Matthew G. Knepley
Computation Institute
University of Chicago, Chicago, IL
[email protected]
January 10, 2018
1
Introduction
Scientific computation is a discipline that combines numerical analysis, physical understanding, algorithm development, and structured programming.
Several yottacycles per year on the world’s largest computers are spent simulating problems as diverse as weather prediction, the properties of material
composites, the behavior of biomolecules in solution, and the quantum nature of chemical compounds [23]. This article is intended to review specfic
languages features and their use in computational science. We will review
the strengths and weaknesses of different programming styles, with examples
taken from widely used scientific codes.
This article will not cover the broader range of programming languages,
including functional and logic languages, as these have, so far, not made inroads into the scientific computing community. We do not cover systems with
sophisticated runtime requirements, such as Cilk [11], since this is currently
incompatible with high performance on cutting edge hardware. For this reason, we also ignore transactional memory, both software and hardware. We
also will not discuss the particular capabilities of software libraries in detail.
Particular libraries will be used as examples, in order to highlight advantages and disadvantages of the programming paradigm, but no exhaustive
presentations of their capabilities, strengths, or weakenesses will be given.
1
2
Brief Overview of Language Characteristics
We begin our discussion with imperative languages, like C and Fortran, meaning languages where the programmer explicitly tells the computer what to do
at each step. The computation is built from variables, which hold values, and
functions which compute the value of a variable based upon the input values
of other variables. For instance, important functions for scientific computing
are arithmetic functions, like division, and linear algebraic functions, like matrix multiplication. The principal advantage of these imperative languages
over simpler systems, such as Excel, is the ability to flexibly combine these
basic elements.
In C and Fortran 90, groups of related variables can be combined together in a structure, which allows them to be passed as a unit to functions.
This both improves code readability and decreases its conceptual complexity.
For example, a customer structure could store a customer’s name, account
number, and outstanding balance.
struct cu s tom er {
char ∗name ;
int
acct ;
float balance ;
};
Similarly, functions may call other functions, or themselves recursively, in
order to simplify the description of the operation. For example, the merge
sort algorithm works by first sorting each half of an array, and then merging
together these smaller sorted arrays into a completely sorted array.
void m er geS or t ( int a r r a y [ ] , int ar r ayL en gth ) {
int h a l f L e n g t h = ar r ayL en gth / 2 ;
i f ( ar r ayL en gth < 2) return ;
m er geS or t(& a r r a y [ 0 ] , h a l f L e n g t h ) ;
m er geS or t(& a r r a y [ h a l f L e n g t h ] , h a l f L e n g t h ) ;
merge(& a r r a y [ 0 ] , &a r r a y [ h a l f L e n g t h ] ) ;
}
Using these mechanisms, just amounting to the introduction of hierarchical
organization to simple code elements, the complexity of large codes can be
drastically reduced.
2
Object-Oriented languages, such as C++ and Python, allow a further
level of combination. Data can be grouped together with the functions which
operate on it, into a super-structure called an object. This can be useful for
organizing the action on groups of data. For example, we can augment our
customer example with methods which change the account number or debit
the account, where now we declare a class which describes a type of object.
c l a s s cu s tom er {
char ∗name ;
int
acct ;
float balance ;
public :
void d e b i t ( f l o a t amount ) {
this−>b a l a n c e −= amount ;
};
void changeAccount ( int a c c t ) {
this−>a c c t = a c c t ;
};
}
However, this organization can also be accomplished in standard C by passing
the structure as an argument.
void d e b i t ( struct cu s tom er ∗ t h i s , f l o a t amount ) {
t h i s −>b a l a n c e += amount ;
}
Another organizational strategy is to give types to variables. In C and
Fortran, this tells the compiler how much space to use for a variable, such
as 4 bytes for a long int in C. Structures are also types, built out of smaller
types, as are classes. In some languages, such as C, C++, and Fortran, the
type of every variable must be specified before the program is run, which is
called static typing. In contrast, Python, Ruby, and Perl allow the type of a
variable to change at runtime depending on what kind of value is stored in
it, which is called dynamic typing. Dynamic typing makes code smaller and
easier to write, but the code is harder for the compiler to optimize and can
sometimes be harder to understand without types to guide the reader.
Object-oriented languages very often have collections of similar functions
that operate differently depending on the type of argument provided, or the
type of object associated with the function since the object is understood as
a silent first argument. For example,
3
class c i r c l e {
float radius ;
public :
float area ( ) {
return PI∗ this−>r a d i u s ∗ this−>r a d i u s ;
};
}
class t r i a n g l e {
f l o a t base , h e i g h t ;
public :
float area ( ) {
return 0 . 5 ∗ this−>b as e ∗ this−>h e i g h t ;
};
}
the area() function will behave differently when called with a circle object,
rather than a triangle. Choosing a specific function, or method dispatch,
based upon the types of its arguments is called polymorphism. A programmer
might want two classes to share many functions and data, but differ in a few
respects. The inheritance mechanism allows one class to behave exactly as
another, unless that behvior is explicitly redefined.
In languages with static typing, it can be useful to write functions which
have the same form for a range of types, just as they would look in a dynamically typed language. This mechanism is called genericity, and the specific
strategy used in C++ is templating. Templates allow a placeholder, often T,
to be replaced by the specific type of an argument when the code is compiled.
Thus many versions of the function are generated, a process called template
instantiation, one for each different type of argument.
3
3.1
Single language Codes
Imperative Programming
Advantages The still dominant paradigm for both application code and
libraries in scientific computing is a single language code base in a wellestablished imperative language such as C or FORTRAN 77 (F77). These
languages have several notable advantages over more sophisticated alternatives when applied to numerical algorithms. First and foremost, they can be
4
made performant by a mildly proficient user, and the ease of achieving good
performance comes from several language features. Both C and Fortran are
very similar to the underlying assembly code into which they are compiled.
Thus, it is not only obvious to users how a given routine will be executed,
but also obvious to the compiler. This correspondence makes it much easier
to create routines that compilers can optimize well. The simple execution
model for C and F77 also makes inspection of the code by an outside user
possible. More complex constructs, such as templates and deep inheritance
hierarchies, can obscure the actual execution even while making the intent
clearer. Moreover, the state of the computation and data structures can be
easily seen in a debugger, whereas more complex constructs and execution
environments often hide this information.
Simplicity in execution also translates to greater ease in using debugging
and profiling tools. Major debugging tools such as gdb, idb, totalview, and
valgrind [41] have excellent support for C and F77. They do support higher
level features, but there can be inconsistencies, especially with template instantiation, that cause some information to be unavailable. This caveat also
applies to profiling tools. Simplicity in binary interface definition means that
C and F77 are especially easy to interface with other languages and environments. Symbols are not mangled, or made unique using complex names, so
matching ones can be easily created in another system. Function parameter
passing is also unambiguous. This makes C the language of choice when
defining a foreign function interface for a higher level language, that is an
interface which allows functions in one language to be called from another
such as C.
Disadvantages A price is paid, however, for the simplicity of these languages. The size of source code for equivalent tasks is quite often more
than an order of magnitude larger than for object oriented or functional
languages. The user must write code for method dispatch instead of using
polymorphism, write separate routines for many types instead of using templates, produce basic data structures which are not part of core libraries,
and in general reproduce many of the mechanisms built into higher level languages, as described below. In particular, the rich standard libraries of many
OO and functional languages save considerable special purpose code.
One of the most severe omissions in C and F77 is that of flexible namespaces for identifiers, types, and functions. The absence of hierarchical names-
5
paces for symbols, such as namespace in C++ or dot notation in Python,
results in comically long identifier names, and rampant problems with clashing symbol names when linking together different scientific libraries. A second problem is the need for manual memory management of all structures,
or for F77 static declaration of memory up front. In C++, when objects
are declared in an inner scope such as a function body or for loop, they are
automatically created upon entry and destroyed on exit from that scope.
These are called automatic objects, and arrays can also be defined this way.
In C, the creation and destruction must be managed by hand, which may
be complicated when for instance error conditions arise. Lastly, there are
no language facilities for introspection, determination of code structure at
runtime, as there are in C++ or Python. At best, we can use the dynamic
loading infrastructure to search for library symbols, but cannot determine
which types, functions, or structures are defined in a library without making
separate, configuration tests outside the language itself. This usually results
in fantastic complication of the build process.
Example Perhaps the most successful software library written in this paradigm
are the BLAS library [29], dating from 1979, and LAPACK [5] libraries for
linear algebra, first released in February 1992, for linear algebra. They are
both numerically robust and extremely efficient, and used in almost every
serious numerical package. The internals are so easily understood, being
written in simple F77, that they are often copied wholesale into application code without the use of the library itself. However, they suffer from a
classic problem with scientific software of this type, lack of data encapsulation. The data structures upon which the operations, such as matrix-matrix
multiplication, operate are specified directly in the library API. Thus the
layout of dense matrices is given in the interface and cannot be changed by
the implementation. For example, the calling sequence for double precision
matrix-matrix multiplication in BLAS, a workhorse of scientific computing,
is
SUBROUTINE DGEMM(TRANSA, TRANSB, M, N, K, ALPHA, A, LDA,
B, LDB, BETA, C, LDC)
∗
. . S c a l a r Arguments . .
DOUBLE PRECISION ALPHA,BETA
INTEGER K,LDA,LDB,LDC,M,N
CHARACTER TRANSA,TRANSB
∗
..
6
∗
. . Array Arguments . .
DOUBLE PRECISION A(LDA, ∗ ) ,B(LDB, ∗ ) , C(LDC, ∗ )
Here the multiply is prescribed to operate on a dense array of doubles A with
a row stride of LDA. This limitation complicated the implementation of an
efficient distributed memory version of the library, and led to the introduction
of Elemental [38] which uses a more favorable data distribution, especially for
smaller sizes. It has also prevented the fusion of successive operations, which
could result in data reuse or latency hiding, greatly improving the efficiency
of the library.
3.2
Object Orientation
Advantages Object Orientation (OO) is a powerful strategy for data encapsulation. Objects are structures that combine data and functions which
operate on that data. Although this can clearly be accomplished in C using
struct s and function pointers, many languages have builtin support for this,
including Objective C, C++, C#, and Python. This kind of encapsulation
encourages the programmer to produce data structure neutral interfaces [46],
as opposed to those in LAPACK. Combined with polymorphism, or function
dispatch based upon the argument types, we can write a single numerical
code that uses different algorithms and data structures based upon its input
types [44]. This, in a nutshell, is the current strategy for dealing with the
panoply of modern architectures and problem characteristics for scientific
simulation. It should also be noted that all the OO languages mentioned
above provide excellent namespacing facilities, overcoming another obstacle
noted in Section 3.1.
The essential features of OO organization, encapsulation and dynamic
dispatch, can be emulated in C at the cost of many more lines of code. Early
C++ compilers did just this by emitting C rather than object code. Moreover, languages such as C++ and Java have removed some of the dynamism
present in Objective C and C OO frameworks. We will show an example of
this below.
Disadvantages The downsides of object oriented organization have to do
with controlling code complexity, the original motivation for the introduction
of OO structures. The true measure of code complexity is ease of understanding for an outside observer. There can be a temptation to create deep object
7
hierarchies, but this tends to work against both code readability and runtime
flexibility as illustrated below. For numerical code especially, it is common
to introduce operator overloading. This can improve readability, however
transparency of the performance cost is lost, which often results in very slow
application code, unacceptable in most simulation environments.
Examples PETSc [6, 7] and Trilinos [19, 20] are two popular packages
which can solve sparse systems of nonlinear algebraic equations in parallel. A common case for which these libraries use OO techniques to control
complexity is the choice among a dizzying range of iterative solvers and preconditioners.
In PETSc, a Krylov Subspace solver (KSP) object acts as an abstract
base class in C++. However, the key difference is that instantiation of the
subtype is done at runtime,
MPI Comm comm ;
KSP
ksp ;
PC
pc ;
KSPCreate (comm, &ksp ) ;
KSPGetPC( ksp , &pc ) ;
/∗ G e n e r a l l y done w i t h command l i n e o p t i o n s ∗/
KSPSetType ( ksp , " gmres" ) ;
PCSetType ( ksp , " ilu" ) ;
and we see that the Trilinos equivalent in C++ is almost identical.
Teuchos : : RCP<Epetra RowMatrix> A;
E p e t r a V e c t o r LHS , RHS;
E p e t r a L i n e a r P r o b l e m Problem(&∗A,&LHS,&RHS ) ;
I f p a c k Factor y ;
Teuchos : : RCP<I f p a c k P r e c o n d i t i o n e r > Prec =
Teuchos : : r cp ( Factor y . C r eate ( "ILU" , &∗A, 1 ) ) ;
AztecOO S o l v e r ( Problem ) ;
S o l v e r . S etAztecO p tion ( AZ s olver , AZ gmres ) ;
S o l v e r . S e t P r e c O p e r a t o r (&∗ Prec ) ;
Trilinos and PETSc make the same decision to trade language support for
runtime flexibility. In packages like dealII and FEniCS, each linear solver is
instantiated as a separate type which all derive from an abstract base type.
8
Naively, this strategy would force the user to change the application code
in order to try a different solver. The Factory Pattern [15] is often used
to alleviate this difficulty. Both Trilinos and PETSc also use factories to
organize instantiation.
However, two related problems arise. First, if the solver object is defined
by a single concrete type, changing a given solver nested deeply within a
hierarchical solve becomes prohibitively complex. Both solver objects above
can change the concrete type “on the fly”. This ability is key in multiphysics
simulations where already complex solvers are combined and nested. Second,
accessing the concrete solver type would now involve downcasts that may fail,
littering the code with obtrusive checks. In PETSc, each concrete type has
an API which is ignored by other types. Thus,
KSPGMRESSetRestart ( ksp , 4 5 ) ;
KSPChebychevSetEigenvalues ( ksp , 0 . 9 , 0 . 1 ) ;
P C Factor S etL evels ( pc , 1 ) ;
PCASMSetOverlap ( pc , 2 ) ;
will execute without error regardless of the solver type, but will set eigenvalue
bounds if the user selected the Chebychev method. Trilinos uses a bag of
parameters,
Teuchos : : P a r a m e t e r L i s t L i s t ;
L i s t . s e t ( " fact : drop tolerance " , 1 e −9);
L i s t . s e t ( " fact : level -of - fill " , 1 ) ;
L i s t . s e t ( " schwarz: combine mode " , "Add" ) ;
Prec−>S etP ar am eter s ( L i s t ) ;
however this sacrifices type safety for the arguments, and can also result in
aliasing of argument names.
3.3
Code Generation
Advantages Performance has always been a primary concern for numerical
codes. However, the advent of new, massively parallel architectures, such as
the Nvidia Fermi [35] or Intel MIC [32], while providing much more energy
efficient performance, has greatly increased the penalty for suboptimal code.
These chips have vector units accomodating from 4 to 16 double precision
operations, meaning that code without vectorization will achieve no more
than 25% of peak performance, and usually much less. They also increase
9
the imbalance between flop rate and memory bandwidth or latency, so that
thousands of flops can be needed to cover outstanding memory references.
GPUs in particular have very high memory latency coupled with a wide
bus, making the memory access pattern critical for good performance. In
addition, the size of fast cache memory per core has shrunk dramatically, so
that it cannot easily be used to hide irregular memory access.
The strategies for mitigating these problems are familiar, and include
tiling [3, 18], redundant computation, and reordering for spatial and temporal memory locality [16, 48]. The CUDA language incorporates two of
the most important optimizations directly into the language: vectorization
and memory latency hiding through fast context switch [35]. In CUDA, one
writes kernels in a Single Instruction Multiple Thread (SIMT) style, so that
vector operations are simple and explicit, in contrast to the complicated and
non-portable compiler intrinsics for the Intel MIC. These kernel routines may
be swapped onto a processor using an extremely fast context switch, allowing memory latency in one kernel to be hidden by computation in others.
However, in CUDA itself, it is not possible to express dependencies among
kernels. OpenCL [17] has preserved these essential features of CUDA, and
also achieves excellent performance on modern hardware.
It is, however, unlikely that these kernels can be coded by hand for scientific libraries. Even should the model, discretization, coefficient representation, and solver algorithm be fixed, the kernel would still have to take account
of the vector length on the target processor, memory bus width, and available
process local memory. We are not describing merely tuning a small number
of parameters describing the architecture, as for instance in Atlas [49], but
algorithm reorganization at a high level, as shown in the examples.
Disadvantages The pricipal disadvantage of automatically generated code
are weak support in the build toolchain. In contrast to C++ templates, more
exotic methods of code generation require outside tools, usually separate files,
and are not easily incorporated into existing build system, especially for
large projects. A very hopeful development, however, is the incorporation in
OpenCL of compilation as a library call. Thus kernel generation, compilation,
and linking can take place entirely within a running application, much like
the template version.
However code is generated, care must be taken that the final output
can be read by the user, and perhaps improved. A major disadvantage of
10
templates is that it prevents the user from directly inspecting the generated code. Without readable code, the user cannot inspect the high level
transformations which have been used, correct simple errors for new environments, insert specialized code for new problems, and in general understand
the system. Code generators should strive to provide easy access for the
user to generated source, as shown in the FEniCS package, while seamlessly
integrating the result into existing build architectures.
Examples The Thrust [10] package from Nvidia uses the C++ template
mechanism to generate CUDA kernels for common functional operations such
as map, transform, and reduceByKey. Most transformations here amount to
intelligent blocking and tiling, and are well suited to this mechanism. Even
higher level generation is used by both Spiral [39] and FFTW [14]. The
algorithm is broken down into smaller components, for FFTW these are
“codelets” and Spiral produces another low-level language. A particular instantiation of the algorithm can be composed of these pieces in many different
ways. Partial implementations are constructed, run, and timed. This real
time evaluation guides the construction of the final implementation for the
given problem.
3.4
Generiticity and Templating
Advantages By far the most popular type of code generation technique
employed in scientific computing is C++ templates. It gives users the ability
to hardwire constants into a piece of code, allowing the compiler to fold them
and perform loop unrolling optimizations, without sacrificing flexibility in
the code base or using convoluted preprocessing schemes. It is also possible
to write generic operations, independent of the data type on which they
operate, but still have them properly type check. This can make the code
base much smaller, as separate routines for different types are unified, and
is the inspiration behind the Standard Template Library [47, 42]. Moreover,
all this can be done without changing the normal toolchain for C++ use,
including compilation, profiling and debugging.
Disadvantages While templates are integrated into the normal C++ workflow, unfortunately the product of template expansion is not available to the
user. Thus, they cannot inspect the particular optimizations which are per11
formed or specialize it by adding code for a specific instance (although they
can use the template specialization mechanism). Compile time also greatly
increases with templates, becoming problematic for large code bases. In addition, the type safety of templates is enforced at the instantiation point which
can be very far removed from the use location in the code. This very often
results in inpenetrable, voluminous error messages that stretch for hundreds
of thousands of lines. The failure of concepts to enter the C++ standard [43]
seems to indicate that this problem will persist far into the future. The
template mechanism makes language interoperability almost impossible. In
general, one must instantiate all templates to be exposed to another language,
and remove templates from public APIs visible in other languages.
The template mechanism, when used to do simple type naming and constant injection, can be quite effective. However, when used for higher level
logical operations and to execute more complicated code rearrangement,
there are numerous problems. The syntax becomes very cumbersome, as
in the case of optional template arguments. The logic of instantiation (type
resolution) is opaque, and following the process during debugging is nearly
impossible. The gains in source code size and readability are lost when using
templates for more sophisticated code transformation.
Examples The Elemental library [38, 37] exhibits two very common uses
of templates for scientific computing. It templates over basic types, but
it also uses template markers to switch between entirely different routines.
They are both present in the basic distributed matrix class, DistMatrix,
with declaration:
enum D i s t r i b u t i o n {
MC, // Col o f a m a t r i x d i s t r i b u t i o n
MD, // D i ag onal o f a m a t r i x d i s t r i b u t i o n
MR, // Row o f a m a t r i x d i s t r i b u t i o n
VC, // Col−major v e c t o r d i s t r i b u t i o n
VR, // Row−major v e c t o r d i s t r i b u t i o n
STAR // Do not d i s t r i b u t e
};
template<typename T, D i s t r i b u t i o n C olDis t ,
D i s t r i b u t i o n RowDist , typename I n t>
c l a s s DistMatrix ;
The first template argument defines the number field over which the matrix operates. This allows identical source to be used for single precision,
12
double precision, quad precision, and complex matrices, since these types all
respond to the arithmetic operations. At a slightly higher level, search and
sort algorithms in the Standard Template Library rely on the same interface
compatibility to write generic algorithms. This can be extended to very high
level algorithms, such as the Conjugate Gradient solver [40] for sparse linear
systems in the dealII package [8, 9].
template <c l a s s VECTOR>
template <c l a s s MATRIX, c l a s s PRECONDITIONER>
void
SolverCG<VECTOR> : : s o l v e ( const MATRIX
VECTOR
const VECTOR
const PRECONDITIONER
{
if (! x . all zero ()) {
A. vmult ( g , x ) ;
g . add ( − 1. , b ) ;
} else {
g . equ ( − 1. , b ) ;
}
r e s = g . l2 n or m ( ) ;
conv = this−>c o n t r o l ( ) . ch eck ( 0 , r e s ) ;
i f ( conv ) { return ; }
p r e c o n d i t i o n . vmult ( h , g ) ;
d . equ ( − 1. , h ) ;
gh = g ∗h ;
while ( conv == S o l v e r C o n t r o l : : i t e r a t e ) {
i t ++;
A. vmult (Ad , d ) ;
alp h a = d∗Ad ;
alp h a = gh / alp h a ;
g . add ( alpha , Ad ) ;
x . add ( alpha , d ) ;
r e s = g . l2 n or m ( ) ;
conv = this−>c o n t r o l ( ) . ch eck ( i t , r e s ) ;
i f ( conv != S o l v e r C o n t r o l : : i t e r a t e ) break ;
p r e c o n d i t i o n . vmult ( h , g ) ;
b eta = gh ;
13
&A,
&x ,
&b ,
&p r e c o n d i t i o n )
gh
= g ∗h ;
b eta = gh / b eta ;
d . sadd ( beta , − 1 . , h ) ;
}
}
This code is shared among all implementations of VECTOR, MATRIX, and
PRECONDITIONER, in much the same way it is in OO codes using an abstract base class, similar to PETSc.
However, in complicated numeric codes, it is often the case that template instantiation is substituted for dispatch. For example, the AlignWith()
method has different implementations depending on the type of the template
arguments. This evaluation of method displatch at compile time avoids the
overhead of lookup in a virtual table of function pointers, but it sacrifces
flexibility. With types fixed at compile time, we cannot change types in response to different input, or new hardware, or simulation conditions without
recoding and rebuilding the executable. This makes exploration of different
implementations problematic, particularly in the context of solvers. Moreover, more complex block solvers for multiphysics systems [33] are built out
of basic solvers, and runtime type changes allow construction of a range of
powerful solvers [45].
4
4.1
Multi-language Codes
Python and Wrapping
Advantages Multilanguage code allows the designer to combine the strengths
of different approaches to programming. A popular combination in scientific computing is the speed and memory efficiency of languages like C and
Fortran with the flexibility and parsimony of scripting languages such as
Python. Python allows inspection of the full state of the running program,
introspection, and management of both memory and variable typing, speeding development of new code and easing unit testing [28, 34]. Python also
supports generic programming since all variables are dynamically typed and
do not need to be declared when code is written.
Specialized Python tools have been developed for wrapping C libraries,
such as ctypes, SWIG, and Cython. Cython in particular allows C data
structures to be manipulated transparently in Python without copies, Python
14
routines to be called from function pointers in C, and data conversions to be
completely automated. The object structure of C++ can even be mapped
to the object structure in Python. Error and exception handling is also automated. Cython also allows Python routines to be annotated, and then
automatically converted to C and compiled. The numpy library [36] allows
direct manipulation in Python of multi-dimensional arrays, perhaps using
memory allocated in another language. Operations are performed in compiled code, sometimes on the fly, and without copies, making it as efficient
as standard C, and it can leverage system-tuned linear algebra libraries.
Python string processing and easy data structure manipulation are very
useful for managing user input and output. Many libraries, such as PyLith [1,
2], use Python as a top level control language and then dispatch to C/C++/Fortran
for the main numerical processing underneath. Using the tools mentioned
above (PyLith uses SWIG), this process can be almost entirely automated.
Moreover, Python’s ability to easily expose a library API, and the use of
numpy arrays for data interchange, make it an excellent tool for combining
libraries at a higher level. Libraries solving different physical problems or
different models of a given problem can be combined to attack more complex multi-model, multi-physics, multi-scale problems [24, 12]. In addition,
this wrapping capability has been used to great effect on GPU hardware,
incorporating CUDA and OpenCL libraries into both desktop and parallel
computations [27, 25, 26].
Disdvantages The central disadvantage for multi-language codes comes
in debugging. There are certainly hurdles introduced into the build system,
since different compilation and links steps are needed and many more tests
are needed to verify interoperability, but these can alrgely be handled by
standard systems. No tool exists today that can inspect the state of a running program in the style above, for example Python using a C library. Even
the stack trace after an error is routinely unavailable, although it can be
logged by the C library and passed up as is done in petsc4py [13]. However, stepping across language boundaries in a debugger is not yet possible,
and this limitation makes debugging new code extremely difficult. Thus, the
strategy above works best when combining several mature single-language
libraries, so that debugging is focused only on the interactions between libraries, which can be seen in the state of the numpy objects communicated
among them, rather than on library internals. Recent developments, includ-
15
ing the extension support for Python in gdb 7, indicate that this situation
will improve markedly in the new future.
Example The PyClaw package [31, 4] combines the CLAWPACK [30]
library for solving hyperbolic systems of partial differential equations on
mapped Cartesian grids with the PETSc [6] library parallel linear algebra
and scalable solution nonlinear equations. Incorporation of the PETSc library allowed parallelization of the solvers in both Clawpack and SharpClaw [21] in only 300 lines of Python, as detailed in [22]. PETSc parallel
data structures, in particular the DA object for structured grid parallelism,
were exposed to Clawpack using Python numpy arrays as intermediary structures. This allowed no-copy access by both C and Fortran, as well as easy
inspection in the Python code itself. In fact, since numpy structures are
used for both wrappers, any PyClaw script can be run in parallel using the
PETSc extension PetClaw simply by replacing the call to import pyclaw
with import petclaw as pyclaw. The hybrid code showed excellent weak
scaling, when modeling the interaction of a shock with a low-density bubble
in a fluid, on all 65,536 cores of the Shaheen supercomputer at KAUST.
Acknowledgements The author would like to acknowledge the helpful
comments of Wolfgang Bangerth.
References
[1] Brad Aagaard, Sue Kientz, Matthew G. Knepley, Surendra Somala, Leif
Strand, and Charles Williams. PyLith user manual version 2.0.0, 2014.
[2] Brad Aagaard, Matthew G. Knepley, and Charles Williams. PyLith.
http://geodynamics.org/cig/software/pylith/, 2012.
[3] W. Abu-Sufah, D. J. Kuck, and D. H. Lawrie. On the performance
enhancement of paging systems through program analysis and transformations. IEEE Trans. Comput., 30(5):341–356, 1981.
[4] Amal Alghamdi, Aron Ahmadia, David I. Ketcheson, Matthew G. Knepley, Kyle T. Mandli, and Lisandro Dalcin. PetClaw: A scalable parallel nonlinear wave propagation solver for Python. In Proceedings of
SpringSim 2011. ACM, 2011.
16
[5] E. Anderson, Z. Bai, C. Bischof, J. Demmel, J. Dongarra, J. Du Croz,
A. Greenbaum, S. Hammarling, A. McKenney, and D. Sorensen. LAPACK: A portable linear algebra library for high-performance computers. Technical Report CS-90-105, Computer Science Dept., University
of Tennessee, May 1990.
[6] Satish Balay, Shrirang Abhyankar, Mark F. Adams, Jed Brown, Peter
Brune, Kris Buschelman, Lisandro Dalcin, Victor Eijkhout, William D.
Gropp, Dinesh Kaushik, Matthew G. Knepley, Lois Curfman McInnes,
Todd Munson, Karl Rupp, Barry F. Smith, Stefano Zampini, Hong
Zhang, and Hong Zhang. PETSc users manual. Technical Report ANL95/11 - Revision 3.8, Argonne National Laboratory, 2017.
[7] Satish Balay, Shrirang Abhyankar, Mark F. Adams, Jed Brown,
Peter Brune, Kris Buschelman, Lisandro Dalcin, Victor Eijkhout,
William D. Gropp, Dinesh Kaushik, Matthew G. Knepley, Lois Curfman McInnes, Todd Munson, Karl Rupp, Barry F. Smith, Stefano Zampini, Hong Zhang, and Hong Zhang. PETSc Web page.
http://www.mcs.anl.gov/petsc, 2017.
[8] W. Bangerth, R. Hartmann, and G. Kanschat. deal.II – a general purpose object oriented finite element library. ACM Trans. Math. Softw.,
33(4):24/1–24/27, 2007.
[9] W. Bangerth, R. Hartmann,
http://www.dealii.org/, 2012.
and
G.
Kanschat.
[10] N.
Bell
and
J.
Hoberock.
The
http://code.google.com/p/thrust/, 2010.
Thrust
deal.II.
library.
[11] Robert D. Blumofe, Christopher F. Joerg, Bradley C. Kuszmaul,
Charles E. Leiserson, Keith H. Randall, and Yuli Zhou. Cilk: An efficient
multithreaded runtime system. In Proceedings of the Fifth ACM SIGPLAN Symposium on Principles and Practice of Parallel Programming
(PPoPP), pages 207–216, Santa Barbara, California, July 1995.
[12] Jed Brown, Matthew G. Knepley, David A. May, Lois C. McInnes, and
Barry F. Smith. Composable linear solvers for multiphysics. In Proceeedings of the 11th International Symposium on Parallel and Distributed
Computing (ISPDC 2012), pages 55–62. IEEE Computer Society, 2012.
17
[13] Lisandro Dalcin.
PETSc for Python – Python bindings for
PETSc
libraries.
http://code.google.com/p/petsc4py/.
http://code.google.com/p/petsc4py/.
[14] Matteo Frigo and Steven G. Johnson. The design and implementation of
FFTW3. In Proceedings of the IEEE, volume 93, pages 216–231, 2005.
[15] Erich Gamma, Richard Helm, Ralph Johnson, and John Vlissides. Design Patterns. Addison-Wesley Professional, January 1995.
[16] William D. Gropp, Dinesh K. Kaushik, David E. Keyes, and Barry F.
Smith. Towards realistic performance bounds for implicit CFD codes.
In Proceedings of Parallel CFD’99. Elsevier, 1999.
[17] Khronos
Group.
OpenCL
1.2
specification.
http://www.khronos.org/registry/cl/specs/opencl-1.2.pdf,
2011.
[18] Jia Guo, Ganesh Bikshandi, Basilio B. Fraguela, Maria J. Garzaran,
and David Padua. Programming with tiles. In Proceedings of the 13th
ACM SIGPLAN Symposium on Principles and practice of parallel programming, PPoPP ’08, pages 111–122, New York, NY, 2008. ACM.
[19] Michael A. Heroux, Roscoe A. Bartlett, Vicki E. Howle, Robert J. Hoekstra, Jonathan J. Hu, Tamara G. Kolda, Richard B. Lehoucq, Kevin R.
Long, Roger P. Pawlowski, Eric T. Phipps, Andrew G. Salinger, Heidi K.
Thornquist, Ray S. Tuminaro, James M. Willenbring, Alan Williams,
and Kendall S. Stanley. An overview of the Trilinos project. ACM
Transactions on Mathematical Software, 31(3):397–423, 2005.
[20] M. Heroux et al. Trilinos Web page. http://trilinos.sandia.gov/.
[21] D. I. Ketcheson and M. Parsani.
SharpClaw software, 2011.
http://numerics.kaust.edu.sa/sharpclaw.
[22] David I. Ketcheson, Kyle T. Mandli, Aron J. Ahmadia, Amal Alghamdi,
Manuel Quezada de Luna, Matteo Parsani, Matthew G. Knepley, and
Matthew Emmett. PyClaw: Accessible, extensible, scalable tools for
wave propagation problems. SIAM Journal on Scientific Computing,
34(4):C210–C231, 2012. http://arxiv.org/abs/1111.6583.
18
[23] D. E. Keyes, editor. A Science-based Case for Large-scale Simulation.
U.S. Department of Energy, 2004.
[24] D. E. Keyes, L. C. McInnes, C. Woodward, E. Constantinescu, D. Estep,
and B. Sheehan. First steps toward a multiphysics exemplars and benchmarks suite. Technical Report ANL/MCS-TM-330, Argonne National
Laboratory, Oct 2012.
[25] Andreas Klöckner. PyCUDA. http://mathema.tician.de/software/pycuda,
2011.
[26] Andreas Klöckner. PyOpenCL. http://mathema.tician.de/software/pyopencl,
2011.
[27] Andreas Klöckner, Nicolas Pinto, Yunsup Lee, Bryan Catanzaro, Paul
Ivanov, and Ahmed Fasih. PyCUDA and PyOpenCL: A scriptingbased approach to GPU run-time code generation. Parallel Computing,
38(3):157–174, 2012.
[28] Hans Petter Langtangen. Python Scripting for Computational Science.
Texts in Computational Science and Engineering. Springer, 2009.
[29] C. L. Lawson, R. J. Hanson, D. Kincaid, and F. T. Krogh. Basic linear
algebra subprograms for fortran usage. ACM Transactions on Mathematical Software, 5:308–323, 1979.
[30] R. J. LeVeque and M. J. Berger. Clawpack Software version 4.5, 2011.
http://www.clawpack.org.
[31] Kyle T. Mandli, David I. Ketcheson, et al.
PyClaw software.
http://clawpack.github.io/doc/pyclaw/, 2012.
[32] Cambridge
University
Many
Core
Group.
Intel
many
integrated
core
architecture.
http://www.many-core.group.cam.ac.uk/ukgpucc2/talks/Elgar.pdf,
December 2010.
[33] Dave A. May and Louis Moresi. Preconditioned iterative methods for
Stokes flow problems arising in computational geodynamics. Physics of
the Earth and Planetary Interiors, 171(1-4):33–47, 2008. Recent Advances in Computational Geodynamics: Theory, Numerics and Applications.
19
[34] J.K. Nilsen, X. Cai, Bjø rn Høyland, and H. P. Langtangen. Simplifying
the parallelization of scientific codes by a function-centric approach in
Python. Computational Science & Discovery, 3:015003, 2010.
[35] NVIDIA. NVIDIA’s next generation CUDA compute architecture:
Fermi. Technical report, NVIDIA, 2009.
[36] Travis E. Oliphant. Guide to Numpy. Trelgol, 2008.
[37] Jack Poulson. Elemental: Distributed memory dense linear algebra.
http://libelemental.org, 2015.
[38] Jack Poulson, Bryan Marker, Jeff R. Hammond, Nichols A. Romero,
and Robert van de Geijn. Elemental: A new framework for distributed
memory dense matrix computations. ACM Transactions on Mathematical Software, 39(2), 2013.
[39] Markus Püschel, José M. F. Moura, Jeremy Johnson, David Padua,
Manuela Veloso, Bryan W. Singer, Jianxin Xiong, Franz Franchetti,
Aca Gačić, Yevgen Voronenko, Kang Chen, Robert W. Johnson, and
Nick Rizzolo. SPIRAL: Code generation for DSP transforms. Proceedings of the IEEE, 93(2), 2005. special issue on ”Program Generation,
Optimization, and Adaptation”.
[40] Yousef Saad. Iterative Methods for Sparse Linear Systems, 2nd edition.
SIAM, Philadelpha, PA, 2003.
[41] Julian Seward. Valgrind. http://valgrind.org/, 2012.
[42] SGI.
Standard template library
http://www.sgi.com/tech/stl/, 2011.
programmer’s
guide.
[43] Jeremy G. Siek. The C++0x “Concepts” Effort, pages 175–216. Springer
Berlin Heidelberg, 2012. http://arxiv.org/abs/1201.0027.
[44] Barry Smith. The transition of numerical software: From nuts-and-bolts
to abstraction. SIGNUM Newsletter, 1998.
[45] Barry Smith, Lois Curfman McInnes, Emil Constantinescu, Mark
Adams, Satish Balay, Jed Brown, Matthew Knepley, and Hong Zhang.
20
PETSc’s software strategy for the design space of composable extremescale solvers. Preprint ANL/MCS-P2059-0312, Argonne National Laboratory, 2012. DOE Exascale Research Conference, April 16-18, 2012,
Portland, OR.
[46] Barry F. Smith and William D. Gropp. The design of data-structureneutral libraries for the iterative solution of sparse linear systems. Scientific Programming, 5:329–336, 1996.
[47] Alexander Stepanov and Paul McJones.
Addison-Wesley, 2009.
Elements of Programming.
[48] Michelle Mills Strout, Larry Carter, Jeanne Ferrante, and Barbara
Kreaseck. Sparse tiling for stationary iterative methods. International
Journal of High Performance Computing Applications, 18(1):95–114,
February 2004.
[49] R.
Clint
Whaley
et
al.
http://math-atlas.sourceforge.net.
21
ATLAS
Web
Page.
| 5cs.CE
|
Capacity of Molecular Channels with Imperfect
Particle-Intensity Modulation and Detection
Nariman Farsad∗ , Christopher Rose† , Muriel Médard‡ , and Andrea Goldsmith∗
arXiv:1705.08040v1 [cs.IT] 22 May 2017
∗ EE,
Stanford University,
† School
of Engineering, Brown University,
Abstract—This work introduces the particle-intensity channel
(PIC) as a model for molecular communication systems and
characterizes the properties of the optimal input distribution and
the capacity limits for this system. In the PIC, the transmitter
encodes information, in symbols of a given duration, based on
the number of particles released, and the receiver detects and
decodes the message based on the number of particles detected
during the symbol interval. In this channel, the transmitter may
be unable to control precisely the number of particles released,
and the receiver may not detect all the particles that arrive. We
demonstrate that the optimal input distribution for this channel
always has mass points at zero and the maximum number
of particles that can be released. We then consider diffusive
particle transport, derive the capacity expression when the input
distribution is binary, and show conditions under which the
binary input is capacity-achieving. In particular, we demonstrate
that when the transmitter cannot generate particles at a high
rate, the optimal input distribution is binary.
I. I NTRODUCTION
Molecular communication (MC) conveys information using
small particles’ time of release, number of release, and/or
type [1]. These particles travel from the transmitter to the
receiver where they are detected and the message decoded.
The transport process is typically random, and introduces
uncertainty about the time of particle release and even the
number of particles released during a given symbol interval.
One approach to understanding the capacity limits of molecular channels investigated in prior work assumes information
is encoded in the time of particle release. Such channels are
called molecular timing channels (MTCs). In particular, the
additive inverse Gaussian noise channel is presented in [2], [3],
and upper and lower bounds on capacity are derived. These
work assumed a system where information is encoded in the
release time of a single particle, while in [4], molecular timing
channels where information is encoded via the release times of
multiple particles are considered, and upper and lower bounds
on capacity presented. Reference [5] presents a MTC, where
particles decay after a finite interval, and derives upper and
lower bounds on the capacity of this channel.
Another approach to MC encodes information through the
number of particles released at the transmitter and decodes
based on the number of particles that arrive at the receiver
during the symbol interval. We focus on this type of modulation scheme and call it particle-intensity modulation (PIM)1 .
Different channel coding schemes are compared for the MC
This research was supported in part by the NSF Center for Science of
Information (CSoI) under grant CCF-0939370.
1 This has been called the concentration-shift-keying or the amplitudemodulation in previous work. However, we believe PIM captures the physical
properties of this system and its relation to the optical intensity modulation.
‡ EECS,
MIT
systems that employ PIM in [6]. In [7], [8], this concentrationbased channel is considered with a receiver equipped with
ligand receptors. The process of molecule reception of a ligand
receptor is modeled as a Markov chain and the capacity in
bits per channel use is analyzed. The results are extended to
multiple access channels in [9]. In [10], a binomial distribution
is used to model a system where the transmitter can perfectly
control the release of particles and the receiver can perfectly
detect the number of particles that arrive. It is assumed that the
channel has finite memory and particle transport is assisted by
flow. Using this model, bounds on the capacity are derived,
and the capacity for different memory lengths is analyzed.
Reference [11] assumes that the channel input is the rate
of particle release. The channel is represented as a Poisson
channel with finite memory, and upper and lower bounds on
capacity per channel use are presented.
In this paper, molecular channels with imperfect PIM at
the transmitter and imperfect detection at the receiver are
considered. Specifically, we consider the case where the sender
cannot perfectly control the number of particles that are
released, and the destination may not detect all the particles
that arrive. We assume that the duration of the symbol is long
enough such that particles from one symbol do not impinge on
future symbols. This model is reasonable if particles diffuse
beyond the receptor or disappear in some other fashion, for
instance through degradation [12]. Under this assumption, the
system is memoryless. Finally, we assume that the particles
can be generated at a constant fixed rate at the transmitter.
For this model, we show that the system can be represented
with a channel model similar to the binomial channel [13] with
two differences. Firstly, the channel input is discrete instead of
continuous. Second, the size of the symbol set, which depends
on the maximum number of particles that can be released by
the transmitter, changes as a function of symbol duration. This
is because of the assumption that the particles are generated
at a constant rate at the transmitter. In this work we call this
channel the particle-intensity channel (PIC).
We define the capacity of this channel in bits per second and
as a function of symbol duration. It is shown that the optimal
input distribution, for any symbol duration, always has mass
points at zero and the maximum number of particles that can
be released by the transmitter. This maximum number depends
on the rate of particle generation and the symbol duration.
One of the most prominent modes of particle transport in the
literature is diffusive transport [1]. In this scheme, the released
particles would follow a random Brownian path to the receiver.
For diffusive particle transport, the capacity of the binary input
(1)
We assume that particles are released instantly and simultaneously at the beginning of the symbol interval. Let
X ∈ X = {0, 1, · · · , mτ } denote the number of particles the
transmitter intends to release, where X is the input symbol
set. Note that λ constrains the cardinality of the symbol set
|X |. We let Y denote the total number of particles that are
detected at the receiver.
As particles may not dissipate over the symbol duration,
particles left over from prior symbol intervals could interfere
with detection during the current channel use. Such intersymbol interference (ISI) must be incorporated into deriving the
channel capacity, which can be quite challenging, particularly
for MC. Therefore, to make the problem more tractable,
we assume that particles with transit times exceeding τ are
somehow inactivated. That is, particles are assumed to have
a finite lifetime of duration τ . This feature seems reasonable
since particles could be rendered undetectable either naturally
or by design (via denaturing or gettering/enzyme reactions
[12]). With this assumption, channel uses become independent
and the maximum mutual information between input and
output during a single channel use defines the channel capacity.
",
! " ,$
!",$
!","
,$
! " ,$
!"
!","
,$
!
,$
!"
!"
.
.
.
.
.
.
!","
!","
!
!"
!′
Imperfect
Transmitter
,$
,$
,$
.
.
.
!","
m
Fig. 1.
!"
.
.
.
$
! ",$ !",
!","
!"
,$
!"
2
!","
!"
,$
! " ! " ,$
!"
!"
,$
!
1
!","
,$
! " ,$
!","
!"
!","
,$
!","
$
!","
! " ,$
mτ = bλτ c .
0
! " ,$
PIC is derived, where the 0 is represented by sending nothing
and 1 by sending the maximum number of particles. We show
the conditions under which the binary input is a capacityachieving distribution. In particular, for the case where the
transmitter cannot generate particles at a high rate, an optimal
input is binary.
The rest of this paper is organized as follows. In Section II
we present the system model and the capacity expression. Then
in Section III we investigate characteristics of the optimal input
distribution and we derive the capacity of the binary input
diffusion-based PIC. We present numerical results in Section
IV, and in Section V we discuss the results.
II. S YSTEM M ODEL AND C APACITY F ORMULATION
A. The Particle Intensity Channel
We consider an MC channel in which information is modulated through the PIM, i.e., the number of particles simultaneously released by the transmitter. The particles themselves are
assumed to be identical and indistinguishable at the receiver,
and no other properties (such as the time-of-release) are
used for encoding information. The receiver then counts the
number of particles that arrive during the symbol duration for
decoding the information. The particles that are released by
the transmitter travel to the receiver through some random
propagation mechanism (e.g., diffusion). We assume that the
particles travel independently of each other, and are detected
independently of each other. This is a reasonable model that
has been used in many previous works [1].
The channel is used in a time-slotted fashion, where τ is the
symbol duration. We define a parameter λ as a constant fixed
rate at which the transmitter can generate particles. Note that
in this case the maximum number of particles, mτ , that could
be released by the transmitter in one channel use can change
as a function of symbol duration according to
Random
Transport
Imperfect
Receiver
The particle intensity channel with imperfect transmitter.
In practice, particle emission at the transmitter may be
stochastic, as we now describe in more detail. Let X be the
number of particles that the transmitter intends to release, and
X 0 be the number of particles that are actually released into
the medium. Moreover, let px,x0 be the probability that x0
particles are actually released when the receiver intended to
release x particles (as illustrated in the left portion of Fig. 1).
We now consider the stochastic particle transport. Each released particle i will arrive at the receiver at some independent
identically distributed random time Ti ∼ fT (·). Let fT (t)
denote the PDF of the time the particle arrives, and FT (t)
denote its corresponding CDF. Then the probability that a
particle arrives during a symbol duration is given by
ρ = FT (τ ),
(2)
and the probability that it never arrives (and is by assumption
never detected) is 1 − ρ.
If we assume FT (t) is strictly monotone, the symbol
duration can be obtained from ρ by using the inverse CDF
(iCDF) function, i.e., τ = FT−1 (ρ). Using the iCDF, we can
also rewrite mτ as a function of ρ:
mρ = λFT−1 (ρ)
(3)
Let Y 0 denote the total number of particles that arrive at
the receiver during the symbol duration. Since we assume
the particles propagate independently of each other, given x0
particles were actually released by the transmitter, we have
(
0
0
0
0
qx0 ,y0 = xy0 ρy (1 − ρ)x −y , y 0 ≤ x0
0 0
P (y |x ; ρ) =
(4)
qx0 ,y0 = 0,
y 0 > x0
and in general P (Y 0 |X 0 = x0 ; ρ) ∼ B(x0 , ρ), where B(n, q)
indicates the binomial distribution with parameters n and q.
Finally, the receiver may not be able to perfectly detect the
particles that arrive. Let Y 0 be the number of particles that
arrive at the receiver during one symbol duration, and let Y
be the number of particles that are actually detected during the
corresponding symbol duration. Furthermore, let ry,y0 be the
probability that y particles are detected by the receiver, when
y 0 particles arrive at the receiver. Note that the receiver may
fail to detect some of particles that arrive due to sensitivity or
uncertainty in the detection process. The end-to-end channel
is then defined as
mρ mρ
X
X
P (y|x; ρ) =
px,i qi,j ry,j .
(5)
j=0 i=0
mρ
{px,i }i=0
Since
is a PMF, one can think of the channel input
as actually being this PMF.
In the rest of this work we assume that
mρ αx x0
αx mρ −x0
)
,
(6)
( mρ ) (1 − m
px,x0 =
0
ρ
x
where 0 < α ≤ 1. Note that here p0,0 = 1, which means
the transmitter can send zero particles with perfect accuracy.
Moreover, for α = 1, the expected number of particles released
by the transmitter is x, and for mρ large, using the Gaussian
approximation to a binomial, it is normally distributed around
x for x not close to 0 and mρ . The particles that arrive at
the receiver are assumed to be detected independently of each
other with probability β and therefore we have
0
0
y
ry,y0 =
(β)y (1 − β)y −y .
(7)
y
We now present the end-to-end channel characteristic in the
following proposition.
Proposition 1 (Particle-Intensity Channel): For an MC system which consists of an imperfect PI transmitter governed
by (6), the PI propagation channel in (4) and an imperfect
receiver governed by (7), the channel is characterized by
mρ xθρ y
xθ
P (Y = y|X = x; ρ) =
( mρ ) (1 − mρρ )mρ −y , (8)
y
where θρ ≡ αρβ.
xα
Proof: Given X = x, X 0 ∼ B(mρ , m
). We can write
ρ
P
m
ρ
xα
this as X 0 = i=0 Xi0 where Xi0 ∼ B(1, mρ ) is an indicator
that the ith particle is released by the transmitter. Let Zi ∼
B(1, ρ) be an indicator that the ith released particle arrives at
the receiver within the symbol duration, and Yi0 = Xi0 Zi be the
indicator that the ith particle is both released and arrives at the
receiver during the symbol interval. Let Zi0 ∼ B(1, β) be the
indicator that the ith released particle that arrives at the receiver
is detected, and Yi = Yi0 Zi0 be the indicator that the ith particle
is released by the transmitter, arrives at the receiver,
and is
Pm
ρ
detected. Then Yi ∼ B(1, xαρβ
),
and
since
Y
=
Yi ,
i=0
mρ
xαρβ
Y ∼ B(mρ , mρ ).
An important observation here is that as the symbol duration
τ changes, ρ changes, and therefore, the channel changes. In
this work, we incorporate the optimization of the symbol duration into the formulation of capacity, to present the channel
capacity of the memoryless PIC in bits per second. This is
one important distinction between this and previous work [2],
[3], [10], [11] where the channel capacity is typically defined
in bits per channel use.
For the case when mρ is large and θρ is small, the system
can be well approximated by the Poisson distribution [14]
P (y|x; ρ) =
(xθρ )y exθρ
.
y!
(9)
We write this as P (y|x; ρ) ∼ P(xθρ ), where P(a) indicates
the Poisson distribution with parameter a.
Remark 1: Using the Poisson approximation, the PIC in
MC systems can be viewed as a more general formulation
of the discrete-time Poisson channel used to model optical
intensity channels [15]–[17]. The channel input is discrete
here and continuous in the discrete-time Poisson channel.
Another important difference is that the symbol duration is
finite, and the rate of arrival does not scale linearly with the
symbol duration. Note that although here we do not consider
interfering particles, they can be introduced to the Poisson
model in (9) by adding an extra term similar to the dark current
in optical communications [15]–[17].
B. Channel Capacity Formulation
We now characterize the channel capacity of the PIC. Let
PX,ρ = [PX,ρ (0), PX,ρ (1), ..., PX,ρ (mρ )] be the channel input
PMF and let P be the set of all valid PMFs. Then the capacity
of the channel in (8), in bits per second is defined as
C(ρ) = max
PX,ρ ∈P
I(X; Y |ρ)
,
FT−1 (ρ)
(10)
where FT−1 (·) is the iCDF of the particle detection time. Since
the channel changes as a function of the symbol duration, the
fundamental limit of this channel is then
C∗ = max C(ρ).
ρ
(11)
Remark 2: Note we define the capacity in terms of bits per
second, as the capacity or the rate of a particle’s arrival typically does not increase linearly with time. This is in contrast to
the discrete-time Poisson channel in optical communications,
where the rate of photon arrival increases linearly with the
symbol duration [17]. Therefore, in the PIC, symbol duration
can have a significant effect on the information rate. This will
be further demonstrated in Section IV.
III. O PTIMAL I NPUT D ISTRIBUTION
AND B INARY-I NPUT C APACITY
We now investigate the characteristics of the optimal input
distribution, present capacity when the input is binary, and
investigate settings for which a capacity-achieving distribution
is binary.
Let P∗X,ρ be the optimal input PMF that maximizes the
∗
mutual information in (10), for a given ρ, and let PX,ρ
(i) be
th
∗
the i element of PX,ρ . The following theorem shows that the
input distribution P∗X,ρ that maximizes the mutual information
in (10) has mass points at 0 and mρ .
Theorem 1: For a given symbol duration τ , and hence a
given ρ, the mutual information given in (10) is maximized
∗
∗
by a PMF P∗X,ρ , where PX,ρ
(0) > 0 and PX,ρ
(mρ ) > 0.
∗
Proof: First, we prove that PX,ρ (0) > 0 by contradiction.
The mutual information in (10) can be written as
I(X; Y |ρ) = H(X|ρ) − H(X|Y, ρ).
∗
Assume that PX,ρ
(0) = 0 in P∗X,ρ , and let X ∗ be a RV
drawn from this PMF, and Y ∗ the corresponding channel
output. Let 0 < i < mρ be the ith index of P∗X,ρ such
∗
that PX,ρ
(i) > 0. Let P†X,ρ be another input distribution
constructed by swapping the ith element in P∗X,ρ with the
†
†
∗
zeroth element. That is PX,ρ
(0) = PX,ρ
(i), PX,ρ
(i) = 0,
†
and all the other elements of PX,ρ are the same as P∗X,ρ . Let
X † be a RV drawn from this PMF, and Y † the corresponding
channel output. Clearly H(X ∗ |ρ) = H(X † |ρ). Since symbol
0 is always transmitted perfectly, as there is no ISI and
interfering particles, H(X † |Y † , ρ) ≤ H(X ∗ |Y ∗ , ρ). Therefore
we have I(X ∗ ; Y ∗ |ρ) ≤ I(X † ; Y † |ρ), which contradicts our
assumption that P∗X,ρ is optimal. Therefore, the optimal input
∗
distribution always satisfies PX,ρ
(0) > 0. Using the same
reasoning, and the fact that between all the non-zero symbols,
the symbol mρ results in the highest probability that a particle
arrives at the receiver, we can prove that the optimal input
∗
distribution always satisfies PX,ρ
(mρ ) > 0 as well.
Previous work on MC has considered on-off-keying in
diffusive environments [1]. We now derive the capacity in (10)
as a function of ρ for this particular class of MC systems.
We assume that the particles that arrive at the receiver are
either immediately detected by the receiver with probability
β, or they are never detected, perhaps denatured as part of
misdetection. Let ` be the shortest distance between a point
source transmitter and the surface of a spherical receiver, d be
the diffusion coefficient of particles, and r be the radius of a
spherical receiver. If we assume that the particles diffuse from
the transmitter, located at the origin, to the spherical receiver,
the motion of each particle can be represented using a random
Brownian path in 3D space. Since we assume that the particles
are either detected when they arrive at the receiver or they
are never detectable, the time of arrival is given by the first
time the particle reaches the receiver. For Brownian motion
in 3D space, the first arrival time T , to the spherical receiver,
is a scaled Lévy-distributed random variable where the scale
r
constant is η = `+r
[18]. This means that there is a non-zero
probability that a particle never arrives at the receiver. Note
that for Brownian motion in 1D space, η = 1. Using the iCDF
of the scaled Lévy distribution, we obtain
τ = FT−1 (ρ) =
2
c
.
2 erfcinv2 (ρ/η)
(12)
`
and erfcinv(.) is the inverse of the complemenwhere c = 2d
tary error function erfc(.). We call a channel that relies on this
diffusive transport the diffusion-based PIC (DBPIC).
Remark 3: Substituting (12) into (10), we observe that the
diffusion coefficient d has no effect on the optimal input
distribution and the optimal ρ. This is despite the fact that
the capacity increases linearly with d. This means that if the
type of particle is changed, so long as the distance between
the transmitter and the receiver is the same, and the receiver
has the same radius, the optimal distribution and the optimal ρ
values will remain the same. Note that the change in capacity
is due to the fact that a shorter or a longer symbol duration is
required to achieve the same ρ value.
Remark 4: If we consider a 1D environment2 (i.e., η = 1),
we observe that the capacity decreases as l12 , and the distance l
does not affect the optimal input distribution and the optimal
ρ. For a 3D environment however, changing the distance l
and the radius r could affect the optimal ρ and P (x) values
through η.
We now present the channel capacity of the binary-input
DBPIC in the following theorem.
Theorem 2: Let X b ∈ {0, mρ } be the binary input to the
DBPIC in (8), ξρ be the probability that the symbol mρ is
transmitted, and ϕρ = (1 − θρ )mρ . Then an optimal input
distribution ξρ∗ is given by
ξρ∗ =
1
ϕρ
ϕ −1
ϕρ ρ
,
(13)
− ϕρ + 1
and the capacity of (10), in bits per second, is given by
ϕρ !
CbDB (ρ) =
2
c
1−ϕρ
erfcinv2 (ρ/η) log 1 + (1 − ϕρ )ϕρ
.
(14)
Proof: For the binary input DBPIC, the mutual information in (10) can be written as a function of ξρ using
I(X b ; Y |ρ) = I(ξρ ) =ξρ ϕρ log(ϕρ ) − ξρ (1 − ϕρ ) log(ξρ )
− (1 − ξρ + ξρ ϕρ ) log(1 − ξρ + ξρ ϕρ ).
(15)
Taking the derivative of I(ξρ ) with respect to ξρ , setting it
equal to zero, and solving the resulting equation, we obtain
(13). Substituting (13) into (15) and using (12) in (10) we
obtain the capacity expression in (14).
The log expression in the capacity result in Theorem 2 is
similar to the capacity of the z-channel because, for a binary
input, the channel reduces to the z-channel with symbol 0
going to 0 with probability 1, and symbol mρ going to symbol
0 with probability ϕρ [19].
An interesting question that arises here is the following:
when is the binary input optimal for the PIC. In the following
proposition, we provide a guideline for the optimality of the
binary input for a subclass of PICs.
Proposition 2: For the PIC in (8) where mρ is large and θρ
is small such that the Poisson approximation in (9) is accurate,
the binary input distribution given in (13) is optimal if mρ θρ <
3.3679.
Proof: Using the same technique presented in [15] for the
optical channels, the proposition can be proved.
Note that this condition may be satisfied in many practical
systems where the radius of the receiver is much smaller
than the distance between the transmitter and the receiver,
hence the probability of particles arriving is small, and the
rate of particle generation is small. Upper bounds on the total
variation between binomial and Poisson distributions can be
used to show that this variation is small for small θρ [14].
2 Note that a 1D environment is a good approximation if the system is
confined inside a very narrow and long physical channel.
Parameters: c = 1, 2 = 0.2, , = 0.9, - = 0.9
3
Information rate [bits/sec]
2.5
Cap 6 = 1000
Binary Cap 6 = 1000
Cap 6 = 500
Binary Cap 6 = 500
Cap 6 = 100
Binary Cap 6 = 100
2
1.5
Cap Limit
Binary Cap Limit
6=1000
1
0.5
6=100
6=500
0
0
0.01
0.02
0.03
0.04
0.05
0.06
0.07
0.08
0.09
R EFERENCES
Probability that the particle arrives ;
Fig. 2.
The information rate for binary-input (dashed lines) and the
optimal input based on Blahut-Arimoto algorithm (solid lines) for different
λ. The three vertical dashed lines indicate the ρ value after which mρ θρ in
Proposition 2 is greater than 3.3676.
Parameters: 6=1000, c = 1, 2 = 0.2, , = 0.9, - = 0.9
0.45
0.4
Optimal PX(x)
0.35
0.3
0.25
0.2
0.15
0.1
0.05
0
0
50
100
150
200
250
300
350
400
Channel Input x
Fig. 3.
V. C ONCLUSIONS
We introduce the PIC and show that an optimal input
distribution for this channel always has mass points at zero and
the maximum number of particles that can be released by the
transmitter. Interestingly, it was observed that for diffusionbased propagation, the diffusion coefficient, and hence the
type of the particles used, does not affect the optimal input
distribution. We then derive capacity for the binary input
diffusion-based PIC and present conditions under which a
binary input is optimal for this channel. Our numerical results
illustrate that a binary input is optimal for systems where
the transmitter cannot generate particles at rates that satisfy
Proposition 2. This can be thought of as the low SNR regime.
As part of future work, we will introduce ISI into our model.
The optimal input distribution that maximizes C∗ for λ = 1000.
IV. N UMERICAL R ESULTS
We now numerically compare capacity for an optimal input
distribution against capacity under the binary distribution for
DBPIC. For the comparison, the optimal input distribution
and the capacity are calculated using the Blahut-Arimoto
algorithm [20]. Fig. 2 shows the results for three different
particle generation rates, λ. The scaling factor for the Lévy
distribution η = 0.2, which means the distance between the
transmitter and the receiver is four times the radius of the
receiver. The square markers indicated the C∗ in (11) for the
binary input distribution, and the X-markers indicated C∗ for
the optimal input distribution. Based on the results we observe
that it is only for the case of λ = 1000 that the binary input
distribution does not maximize C∗ . The three vertical dashed
lines indicate the ρ value after which mρ θρ in Proposition 2 is
greater than 3.3676. We observe that for the ρ values smaller
than this critical value, the binary input is the optimal input
distribution. Fig. 3 shows the optimal input distribution that
maximizes C∗ in (11) for λ = 1000.
[1] N. Farsad et al., “A comprehensive survey of recent advancements in
molecular communication,” IEEE Commun. Surveys Tuts., vol. 18, no. 3,
pp. 1887–1919, 2016.
[2] K. V. Srinivas et al., “Molecular communication in fluid media: The
additive inverse gaussian noise channel,” IEEE Trans. Inf. Theory,
vol. 58, no. 7, pp. 4678–4692, 2012.
[3] H. Li et al., “Capacity of the memoryless additive inverse gaussian noise
channel,” IEEE J. Sel. Areas Commun., vol. 32, no. 12, pp. 2315–2329,
2014.
[4] C. Rose and I. S. Mian, “Inscribed Matter Communication:
Part I,” arXiv:1606.05023 [cs, math], 2016. [Online]. Available:
http://arxiv.org/abs/1606.05023
[5] N. Farsad et al., “On the capacity of diffusion-based molecular timing
channels,” in Proc. IEEE Int. Symp. on Inf. Theory, 2016.
[6] Y. Lu et al., “Comparison of channel coding schemes for molecular
communications systems,” IEEE Trans. Commun., vol. 63, no. 11, pp.
3991–4001, Nov 2015.
[7] A. Einolghozati et al., “Capacity of discrete molecular diffusion channels,” in Proc. IEEE Int. Symp. on Inf. Theory, 2011.
[8] M. Tahmasbi and F. Fekri, “On the capacity achieving probability
measures for molecular receivers,” in Proc. IEEE Inf. Theory Workshop,
2015.
[9] G. Aminian et al., “On the capacity of point-to-point and multipleaccess molecular communications with ligand-receptors,” IEEE Trans.
Mol. Biol. Multi-Scale Commun., vol. 1, no. 4, pp. 331–346, 2015.
[10] S. Ghavami et al., “Information rates of ask-based molecular communication in fluid media,” IEEE Trans. Mol. Biol. Multi-Scale Commun.,
vol. 1, no. 3, pp. 277–291, 2015.
[11] G. Aminian et al., “Capacity of diffusion-based molecular communication networks over LTI-poisson channels,” IEEE Trans. Mol. Biol.
Multi-Scale Commun., vol. 1, no. 2, pp. 188–201, 2015.
[12] W. Guo et al., “Molecular communications: channel model and physical
layer techniques,” Wireless Commun., vol. 23, no. 4, pp. 120–127, 2016.
[13] C. Komninakis et al., “Capacity of the binomial channel, or minimax
redundancy for memoryless sources,” in Proc. IEEE Int. Symp. on Inf.
Theory, 2001.
[14] W. Vervaat, “Upper bounds for the distance in total variation between the
binomial or negative binomial and the poisson distribution,” Statistica
Neerlandica, vol. 23, no. 1, pp. 79–86, 1969.
[15] S. Shamai, “Capacity of a pulse amplitude modulated direct detection
photon channel,” IEE Proceedings I - Communications, Speech and
Vision, vol. 137, no. 6, pp. 424–430, Dec 1990.
[16] A. Lapidoth and S. Shamai, “The poisson multiple-access channel,”
IEEE Trans. Inf. Theory, vol. 44, no. 2, pp. 488–501, Mar 1998.
[17] J. Cao et al., “Capacity-achieving distributions for the discrete-time
poisson channel–Part I: General properties and numerical techniques,”
IEEE Trans. Commun., vol. 62, no. 1, pp. 194–202, 2014.
[18] H. B. Yilmaz et al., “Three-dimensional channel characteristics for
molecular communications with an absorbing receiver,” IEEE Commun.
Lett., vol. 18, no. 6, pp. 929–932, 2014.
[19] L. G. Tallini et al., “On the capacity and codes for the z-channel,” in
Proc. IEEE Int. Symp. on Inf. Theory, 2002.
[20] R. Blahut, “Computation of channel capacity and rate-distortion functions,” IEEE Trans. Inf. Theory, vol. 18, no. 4, pp. 460–473, 1972.
| 7cs.IT
|
Distributed Nonlinear Observer with Robust Performance
- A Circle Criterion Approach
arXiv:1604.03014v2 [cs.SY] 12 Apr 2016
Jingbo Wu, Frank Allgöwer
Institute for Systems Theory and Automatic Control, University of Stuttgart, 70550 Stuttgart, Germany (e-mail: {jingbo.wu,
allgower}@ ist.uni-stuttgart.de)
Abstract
In this paper, we present a distributed version of the KYP-Lemma with the goal to express the strictly positive real-property
for a class of physically interconnected systems by a set of local LMI-conditions. The resulting conditions are subsequently
used to constructively design distributed circle criterion estimators, which are able to collectively estimate an underlying linear
system with a sector bounded nonlinearity.
1
Introduction
Estimator design has been an essential part of controller
design ever since the development of state-space based
controllers. Milestones were laid by the Luenberger Observer [1], the Kalman Filter [2], and the H∞ -Filter [3].
While in the classical estimator design one estimator is
used for one system, designing distributed estimators
have gained attention since a distributed Kalman Filter
was presented in [4], [5], [6]. In a distributed estimator
setup, multiple estimators create an estimate of the system’s state, while cooperating with each other. In this
setup, even when every single estimator may be able to
obtain an estimate of the state on its own, cooperation
reduces the effects of model and measurement disturbances [7]. Also, the situations are not uncommon where
every single estimator is unable to obtain an estimate of
the state on its own and cooperation becomes an essential prerequisite [8], [9].
Where the literature review above shows that there is
a considerable number of results to address distributed
estimation for linear systems, nonlinear systems have
barely been considered. When looking at existing nonlinear estimation algorithms in literature, one can notice that many of them require some kind of transformation upfront. For instance, the Extended Luenberger
observer [10] and the High-gain Observer [11] require a
transformation to observability normal form. However,
⋆ This work was supported by the German Research Foundation (DFG) through the Cluster of Excellence in Simulation Technology (EXC 310/1) at the University of Stuttgart.
in the case when multiple sensing units cooperate in a
distributed setup, a transformation of coordinate hinders the efficient exchange of information, unless the
transformed coordinates are the same. Restricting the
state transformation to be the same for all sensing units
however requires the measured information to be essentially the same, which is a trivial case. On the other hand,
without coordinate transformation, there are observer
design methods in literature that deal with systems described by a linear state space model with an additive,
sector-bounded nonlinearity [12],[13].
In this paper, we aim at extending LMI-based methods
for distributed estimation such as [8] in order to deal
with linear systems with an additive nonlinearity. Besides for globally Lipschitz nonlinearities, we will mainly
present a design approach for a distributed circle criterion observer. This requires a distributed formulation of
the KYP-Lemma, which shows us that the regular approach of taking the sum-of-squares Lyapunov function
PN
V = k=1 xk Px xk as in [8] is not appropriate for this
case.
The rest of the paper is organized as follows: In Section II, we introduce the notation, some preliminaries
on graph theory, and the respective system class. Then,
in Section III we show an intuitive approach and a motivating example, where the sum-of-squares Lyapunov
PN
function V = k=1 xk Px xk fails. This effect is subsequently discussed by the development of a distributed
version of the KYP-Lemma in Section IV. Section V
then deals with a generalized LMI-based construction
method which overcomes the drawback of the intuitive
approach. A simulation example is shown in Section VI.
Preprint
2
System (1) allows for both Lipschitz-nonlinearities and
monotonous non-Lipschitz-nonlinearities, which together cover a large range of possible nonlinearities,
similar to the incremental quadratic constraint [13]
Preliminaries
Throughout the paper the following notation is used:
Let A be a square matrix. If A is positive definite, it is
denoted A > 0, and we write A < 0, if A is negative
definite. The norm of a matrix kAk is defined as any
induced matrix norm.
2.1
"
Communication graphs
2.3
In this section we summarize some notation from the
graph theory. We use undirected, unweighted graphs
G = (V, E) to describe the communication topology between the individual agents. V = {v1 , ..., vN } is the set
of vertices, where vk ∈ V represents the k-th agent.
E ⊆ V × V are the sets of edges, which model the information flow, i.e. the k-th agent can communicate with
agent j if and only if (vj , vk ) ∈ E. Since the graph is
undirected, (vj , vk ) ∈ E implies that (vk , vj ) ∈ E. The
set of vertices that agent k receives information from is
called the neighbourhood of agent k, which is denoted
by Nk = {j : (vj , vk ) ∈ E}. The degree pk of a vertex k
is defined as the number of vertices in Nk . Assuming the
graph as undirected is restrictive in general, however, we
will later show that it is a sensible assumption for the
problem of constructing the distributed circle criterion
estimator.
2.2
#⊤
θ(a) − θ(b)
M
"
a−b
#
θ(a) − θ(b)
.
Problem statement
The problem considered in this paper is to design N local estimators for (1), where every local estimator i =
1, ..., N relies only on the local measurement yk and communication with the neighboring estimators. In the following, we will denote
C1 x
y1
y2 C2 x
y = . = . .
.. ..
CN x
yN
The vector of local estimates will be denoted x̂k ∈ Rn ,
and the local estimation error vector is defined as ek =
x − x̂k . The aggregated vector for all local estimation er⊤ ⊤
ror vectors is denoted e⊤ = [e⊤
1 , ..., eN ] . Since the separation principle does not hold in general for nonlinear
systems, we need to make a technical assumption on the
closed loop system in order to avoid finite escape time
[12]:
System model
We consider the n-dimensional system
e + g(u) + Bw w
ẋ = Ax + Bφ φ(Hx) + Bθ θ(Hx)
y = Cx
a−b
Assumption 1: Given initial conditions x(0) and a control input g(u), if e(t) ∈ Le∞ , then x(t) ∈ Le∞ .
(1)
where x ∈ Rn is the state variable, u ∈ Rm is the control
input, y ∈ Rq is the output vector, and w(t) ∈ Rl is an
exogenous disturbance in the L2 -space. φ(·) is a known
r-dimensional nonlinearity satisfying
P
n
φ1 ( j=1 H1j xj )
..
φ=
(2)
,
P .
n
φr ( j=1 Hrj xj )
Now, the distributed estimation problem can be expressed as following:
Problem 1 (Distributed estimation): Design a
group of N estimators with respective estimation x̂k (t),
k = 1, . . . , N , such that the following two properties are
satisfied simultaneously:
(i) In the absence of disturbances (i.e., when w = 0),
the estimation errors decay so that ek → 0 exponentially for all k = 1, ..., N .
(ii) The estimators provide guaranteed H∞ performance in the sense that
where every φi (·) is a scalar nondecreasing function and
θ(·) is a known re-dimensional nonlinearity satisfying the
incremental quadratic constraint
N Z
X
(a − b)⊤ (a − b) ≥ τ 2 (θ(a) − θ(b))⊤ (θ(a) − θ(b)), (3)
k=1
r
for any a, b ∈ Re
. In fact, in many practical applications,
the state x(t) will be restricted to a bounded set X . In
this case, it suffices for (2), (3) to hold on this bounded
set X .
0
∞
2
2
e⊤
k Wk ek dt ≤ N γ kwkL2 + I0 .
(4)
Wk is a positive semi-definite weighting matrix and
I0 is the cost due to the estimators’ uncertainty
about the initial conditions of the system.
2
e ⊤ H,
e then
for all k = 1, ..., N , where W k = Wk + H
Problem 1 admits a solution of the form (5), where
In particular, the estimators shall form a distributed
setup in the way that the dynamics of each estimation
x̂k only depends on the local measurement yk and communication with the neighboring estimators j.
3
Lk = Pk−1 Gk
An intuitive approach
Remark 1 In (6), the indexes j1 , ..., jpk enumerate the
(k)
(k)
neighbors of estimator k. Strictly speaking, j1 , ..., jpk
is required as notation, but in this paper, we drop the
superscript (k) to keep the notation simple.
An intuitive approach to solve Problem 1 is by adapting
the method introduced for estimation of linear systems
in [9],[14]: There, for every estimator k, a respective LMI
condition is derived, which allows for distributed calculation of the required filter gains [15].
The proof is omitted here because a more general version
will be introduced and thoroughly proven later. This
approach works in some cases of Problem 1, but however
has significant limitations due to conservativeness of the
approach. One example, where is subsequently fails, is
given in the following.
These LMI-conditions can be extended with respect to
(3) by adding the SPR-condition as done in [12]. The
design conditions, which result from this intuitive approach are shown in the following.
3.1
(8)
Kk = Pk−1 Fk .
Design conditions
3.2
A motivating example
We define the matrices
Consider the six-dimensional oscillator
Qk =Pk A + A⊤ Pk − Gk Ck − (Gk Ck )⊤ − pk Fk − pk Fk⊤
+ αPk + pk πk Pk ,
0
−1
0
ẋ =
−1
0
−1
where Pk ∈ Rσk ×σk is a symmetric, positive definite
matrix. πk and α are positive constants which will later
play the role of design parameters.
Let the estimator dynamics be proposed as
e k (yk − Ck x̂k )) + Bθ θ(H
e x̂k )
x̂˙ k =Ax̂k + Bφ φ(H x̂k + L
X
+ g(u) + Lk (yk − Ck x̂k ) + Kk
(x̂j − x̂k ).
1
0
1
0
1
0
−1 0
1
0 −1 0
−1 0 −1
0 −1 0
0 1
1 0
0 1
x
1 0
0 1
−1 0
(9)
+ Bφ φ(Hx) + Bw w + g(u)
j∈Nk
(5)
with the monotonously increasing nonlinearity φ(·),
h
i⊤
h
i
where Bφ = 1 0 0 −1 0 0 , H = 1 1 1 1 1 1 , and
h
i⊤
Bw = 1 1 1 1 1 1 . The individual measurements
are
Then, we have following design conditions.
Theorem 1 Let a collection of matrices Fk , Gk and Pk ,
k = 1, . . . , N , be a solution of the LMIs
Qk + W k Pk Bθ Pk Bw
Fk
(P B )⊤ −τ 2 I
0
0
k θ
(Pk Bw )⊤ 0 −γ 2 I
0
⊤
Fk
0
0 −πj1Pj1
.
..
0
0
0
Fk⊤
0
0
0
...
Fk
0
0
0
0
0
0
..
.
0
0 −πjpk Pjpk
y1 = x2 −x1 , y2 = x3 −x2 , y3 = x4 −x3 ,
y4 = x5 −x4 , y5 = x6 −x5 , y6 = x1 −x6 ,
<0
where x = [x1 , ..., x6 ]⊤ , and let the estimator be connected by a ring-type communication topology E =
{(vk , vk+1 ), (vk+1 , vk )|k = 1, ..., 5} ∪ {(v6 , v1 ), (v1 , v6 )}.
Applying the solution method (5)-(8) by using a numerical solver like YALMIP/SEDUMI will immediately yield infeasibility. In remainder of the paper, we will present the
reason for this and introduce a more general approach
in order to overcome this problem.
(6)
and the equality constraints
e k Ck
− Pk Bφ = H − L
(10)
(7)
3
4
⊤
1, ..., N , nk × nj matrices Pkj = Pjk
for (k, j) ∈ E, and
constants ǫ, π1 , ..., πN > 0 such that for all k = 1, ..., N ,
it holds that
A distributed KYP-Lemma
In the single-system case, where we have the LTI system
e + Bu
ẋ = Ax
y = Ex,
(11)
the closed loop of (11) with the nonlinear feedback u =
−ψ(y) ∈ Rp satisfying y ⊤ ψ(y) ≥ 0, is globally asymptotically stable, if the p × p transfer function matrix
e −1 B is strictly positive real (SPR).
G(s) = E(sI − A)
Further, from the Kalman-Yakubovich-Popov (KYP),
we have following necessary and sufficient conditions.
e+A
e⊤ P ≤ −ǫI
PA
PB = E
⊤
+
(12)
yk = Ek xk +
X
Ekj xj ,
Qk (j1 , j1 ) Qk (j1 , j2 ) . . .
∗
∗
∗
∗
∗
Qk (j2 , j2 ) . . .
..
.
∗
∗
{z
∗
∗
Qk
pk πk Pk +ǫI +W k
Qk (j1 , jpk )
Qk (j1 , jpk )
..
.
Qk (jpk , jpk )
}
0
−πj1 Pj1
0
−πj2 Pj2
0
0
..
.
−πjpk Pjpk
{z
|
Now, we assume system (11) to be a structured system
in the sense that it is composed out of N subsystems in
the form
X
ẋk = Ak xk +
Akj xj + Bk uk
j∈Nk
∗
|
e B) be controlLemma 1 (Lemma 6.3 in [16]) Let (A,
e E) be observable. Then, the p × p transfer
lable and (A,
e −1 B is strictly posfunction matrix G(s) = E(sI − A)
itive real if and only if there exists a symmetric matrix
P > 0, matrices L, W , and a constant ǫ > 0 such that
Qk (k, k) Qk (k, j1 ) Qk (k, j2 ) . . . Qk (k, jpk )
Sk
with
≤ 0,
}
(14)
Qk (k, k) = Pk Ak + A⊤
k Pk
Qk (k, j) = Pk Akj + A⊤
k Pkj for j ∈ Nk
Qk (j1 , j2 ) =
(13)
j∈Nk
⊤
Akj2
Pkj
1
+
A⊤
kj1 Pkj2
(15)
for j1 , j2 ∈ Nk
and
PN
where xk ∈ Rnk , k=1 nk = n, and uk , yk ∈ Rqk ,
PN
k=1 qk = q. The interconnection topology is repree
sented by the graph G = (V, E). Let G(s)
be defined as
the transfer matrix from U (s) = [u1 (s)⊤ , ..., uN (s)⊤ ]⊤
to Y (s) = [y1 (s)⊤ , ..., yN (s)⊤ ]⊤ .
Pk Bk
Ek⊤
⊤
⊤
Pkj Bk Ekj
1
1
=
.
..
.
.
.
⊤
⊤
Ekj
Pkjp Bk
pk
k
X
−1
kPk Pkj k < 1.
Remark 2 Later, in the design procedure for the distributed estimators, we will refer the estimators to this
class of interconnected systems, and moreover, the design
conditions given by the distributed version of the KYPLemma show the reason for the conservativeness of the
intuitive approach presented above.
(16)
(17)
j∈Nk
W k ≥ 0 in (14) is a positive semi-definite matrix that
can be used as a weighting matrix, e.g. to achieve performance guarantees. For the sake of proving Theorem 2, it
can be assumed the W k = 0. Before the proof, we first
introduce following Lemma on block-diagonal dominant
matrices.
The following theorem delivers a sufficient condition for
e
G(s)
being strictly positive real. In particular, instead
of solving (12) for the global system, the equations can
be decomposed into local subproblems.
Theorem 2 (Distributed KYP-Lemma) The p × p
e
transfer function matrix G(s)
is strictly positive real if
there exist symmetric nk × nk matrices Pk > 0, k =
Lemma 2 ([17]) Let the matrix P be partitioned such
4
that
P1 P12 . . . P1N
..
P21 P2
.
,
P =
.
.
.
. . ..
..
e A
e⊤ P )x =
x⊤ (P A+
(18)
+
N
X
⊤
x⊤
k (Pk Ak + Ak Pk )xk
k=1
N X
X
⊤ ⊤
x⊤
k Pk Akj xj + xj Akj Pk xk
k=1 j∈Nk
PN 1 . . . . . . PN
+
nk ×nk
with Pk ∈ R
, Pk > 0 for all k = 1, ..., N , and
Pkj = 0 if (vk , vj ) 6∈ E. If the reduced matrix R = (r)ij
with the elements rij = 1 for i = j and rij = −kPii−1 Pij k
for i 6= j is strictly diagonal dominant, then for any
eigenvalue λ of P , it holds that λ > 0.
N X
X
⊤
⊤ ⊤
x⊤
j Pkj Ak xk + xk Ak Pkj xj
k=1 j∈Nk
+
N X X
X
⊤
⊤ ⊤
x⊤
j Pkj Aki xi + xi Aki Pkj xj ,
k=1 j∈Nk i=Nk
where the right hand side can be further transformed to
PROOF. [Theorem 2] Let there be matrices Pk and Pkj
satisfying the design conditions of Theorem 2, which are
(14), (16), (17). Now, consider the matrix P as defined
in (18), where Pkj = 0 for (k, j) 6∈ E. With Pk > 0
for k = 1, ..., N and (17), we have that the off-diagonal
elements of P
the reduced matrix R are all negative and it
holds that | j6=i rij | < 1. With the diagonal elements of
R being 1, this implies diagonal-dominance of R. Thus,
we can apply Lemma 2, and obtain P > 0. Now, we need
to show that P is a feasible solution to the centralized
SPR-Lemma (12).
N
X
k=1
+
e =
x⊤ P Ax
N
X
k=1
+
j∈Nk
Ak xk +
x⊤
k Pk
N
X
X
k=1 j∈Nk
X
+
j∈Nk
Aj xj +
x⊤
k Pkj
X
i=Nj
+
+
N
X
X
k=1 j∈Nk
⊤
x⊤
j Pkj
Ak xk +
X
i=Nk
Aki xi
X
N
X
X
Qk (k,j)
⊤
⊤
x⊤
j (Pkj Ak + Akj Pk )xk
{z
}
|
Q⊤
(k,j)
k
⊤
⊤
x⊤
j (Pkj Aki + Akj Pki )xi .
|
{z
}
Qk (j,i)
With (14), we now have
e+A
e⊤ P )x ≤
x⊤ (P A
N
X
X
−x⊤
x⊤
j πj Pj xj
k (pk πk Pk + ǫI + W k )xk +
j∈Nk
k=1
e+A
e⊤ P )x ≤ −ǫx⊤ x −
x⊤ (P A
Aji xi
j∈Nk
N
X
⊤
x⊤
k (Pk Akj + Ak Pkj )xj
|
{z
}
k=1 i,j∈Nk
!
N
X
x⊤
k W k xk
k=1
(20)
and therefore, P satisfies (12).
⊤
Now, from the fact that G is undirected and Pkj = Pjk
we observe that for every (vk , vj ) ∈ E we have both
⊤
⊤
e
x⊤
k Pkj ẋj and xj Pjk ẋk as parts of x P Ax. Therefore,
⊤
⊤
by replacing xk Pkj ẋj with xj Pjk ẋk , we obtain
N
X
X
e =
Ak xk +
Akj xj
x⊤
x⊤ P Ax
k Pk
k=1
X
k=1 j∈Nk
Akj xj
Qk (k,k)
N
X
k=1 j∈Nk
• As B is a block-diagonal matrix with B1 , ..., BN
being the diagonal-blocks, we immediately have
P B = E ⊤ when applying (16)
⊤ ⊤
• Let x = [x⊤
be any global state vector.
1 , ..., xN ]
Then we have
N
X
X
Pk xk +
Pkj xj
x⊤ P x =
x⊤
k
k=1
⊤
x⊤
k (Pk Ak + Ak Pk )xk
{z
}
|
The sufficient conditions derived in Theorem 2 lead to a
set of N coupled LMIs and N equality constraints. For
instance, if 100 subsystems (13) with dimension 10 are
interconnected in a ring-type topology (vi , vi+1 ) ∈ E,
(14) involves 100 LMIs with dimension 30 × 30. In particular, those LMIs are amendable to parallel computing algorithms. Similar technique can be applied as presented in [15].
.
Remark 3 In the design conditions of Theorem 2, (17)
represents a block-diagonal dominance condition, which
is used to ensure positive definiteness of the Lyapunovfunction. This inequality can also be replaced by the ad-
(19)
e⊤ P x results in the comAdding the transposed part x⊤ A
plete equation
5
ditional LMI
1
Pk
1+pk
∗
∗
∗
5
1
2 Pkj1
1
1+pj1
...
0
Pj1
..
0
1
2 Pk,jpk
.
1
1+pjp
k
Pjpk
> 0,
Distributed estimator design
5.1
Estimator setup
The estimator dynamics are proposed as
(21)
e x̂k ) + g(u) + Lk (yk − Ck x̂k )
x̂˙ k =Ax̂k + Bφ φ̂k + Bθ θ(H
X
+
Kkj (x̂j − x̂k )
j∈Nk
which may be easier to implement numerically. In case
there is no exact knowledge about the individual degrees
of the neighbors, pj1 ...pjpk in (21), it suffices to replace
the degrees with upper bounds.
e k (yk − Ck x̂k ) +
φ̂k =φ H x̂k + L
|
Remark 4 The conditions of the Distributed KYP, Theorem 2, are sufficient conditions and thus there is a certain amount of conservativeness. However, conservativeness is expected to be small as it is only introduced by
the coupling terms in the second line of (14) and the assumption of diagonal dominance (21) of P . A numerical
example is shown later in the paper.
{z
X
j∈Nk
vk
e kj (x̂j − x̂k )
K
}
(22)
(k)
with initial condition x̂0 . The filter gains to be designed
e k , Kkj , and K
e kj , which are real matrices of suitare Lk , L
able dimension.
We can now particularize Problem 1 with respect to the
proposed estimator dynamics.
Concerning the interconnection topology G, we can derive the following result for the case of identical Bk .
Problem 1’: For all k = 1, ..., N determine the estimae k , Kkj , and K
e kj in (22) such that the two
tor gains Lk , L
properties of Problem 1 are satisfied simultaneously.
Corollary 1 Suppose system (11) is composed out of
N subsystems (13) where the interconnection topology is
represented by a G. Let the p × p transfer function matrix
e −1 satisfy (12) and let Bk = Bj 6= 0
G(s) = E(sI − A)
for two subsystems k, j. If Ekj 6= 0 and Ejk = 0, then it
holds that Ekj Bk = 0.
5.2
Filter gains design
For the estimator error, we obtain with (1) and (22) that
e − θ(H
e x̂k ))
ėk =Aek + Bφ (φ(Hx) − φ̂k ) + Bθ (θ(Hx)
X
− Lk Ck ek +
Kkj (ej − ek ) + Bw w
PROOF. Let P be partitioned as shown in (18). From
⊤
. Now, let Ejk = 0,
symmetry of P , we have Pkj = Pjk
then with (16) we have
j∈Nk
=(A − Lk Ck −
⊤
⊤
Bk⊤ Ekj
= Bk⊤ Pkj
Bk = Bj⊤ Pjk Bk = Ejk Bk = 0.
X
Kkj )ek +
j∈Nk
X
Kkj ej + Bw w
j∈Nk
e − θ(H
e x̂k )).
+ Bφ (φ(Hx) − φ(vk )) + Bθ (θ(Hx)
(23)
Following the argument from [12], we replace the nonlinearities φ(Hx) − φ(vk ) with the time-varying nonlinearities
This corollary considers a special case of (13), which
applies to the distributed estimator design presented in
the next section. In fact, Ekj is a design parameter for
the distributed estimators, if (vj , vk ) ∈ E. This corollary
shows that in the case of a directed graph, where Ekj
is a design parameter but Ejk = 0, the choice of Ekj is
severely restrained.
ψk (zk , t) = φ(Hx) − φ(vk )
X
e k Ck ek +
e kj (ej − ek )
zk = Hek − L
K
The LMIs (14) give an analysis method for showing the
SPR property for a network of interconnected systems
by solving smaller feasibility problems. In particular, the
individual feasibility problems only take local variables
into account, which is essential for the distributed character of the problem. In the next section, distributed estimators will be designed, but since they are subject to
disturbances, additional rows and columns will be added
to (14).
j∈Nk
e k Ck −
zk = (H − L
X
j∈Nk
e kj )ek +
K
X
j∈Nk
e kj ej
K
(24)
Note that due to the monotonicity of φ(·) (2), ψk (zk , t)
satisfies the sector property
zk⊤ ψk (zk , t) ≥ 0
6
(25)
With this property, we are ready to present the main
result, which delivers a design method for the distributed
filter gains.
e+A
e⊤ P )e =
e⊤ (P A
Akj = Kkj
e k Ck −
Ek = H − L
e kj
Ekj = K
Bk = −Bφ ,
X
Kkj
V̇ (e) =
j∈Nk
X
j∈Nk
e kj
K
Pk Bθ
=
N
X
k=1
+
N
X
−2
(27)
+2
j∈Nk
V̇ (e) <
=
Pk ek +
ė⊤
k
N
X
k=1
N
X
N
X
⊤
e⊤
ėk
j Pkj
X
j∈Nk
e⊤
k Pk +
e⊤
k Pk +
X
j∈Nk
X
j∈Nk
Pkj ej
⊤
Bk ψk (zk , t)
e⊤
j Pkj
⊤
e⊤
(Bw w + Bθ ∆k )
j Pkj
k=1
N
X
−
2 ⊤
2 ⊤
−e⊤
k W k ek + τ ∆k ∆k + γ w w
N
X
k=1
With (23) and (24), the derivatives of ek can be reformulated to
X
Akj ej + Bw w
ėk =Ak ek +
j∈Nk
j∈Nk
e⊤
k Qk ek
k=1
.
e⊤
k Pkj ej
e − θ(H
e x̂k )),
− Bk ψk (zk , t) + Bθ (θ(Hx)
|
{z
}
X
With (27) and the sector property (25) this further simplifies to
We use the Lyapunov function candidate
k=1
k=1
PROOF.
e⊤
k Pk ek +
e⊤
k W k ek .
k=1
⊤
⊤
e⊤
j Pkj ėk + ėk Pkj ej
e⊤
k Pk +
N
X
−
V (e) = e⊤ P e =
X
k=1
Then, the estimators (22) are a solution to Problem 1 in
the sense of (4), with performance parameter γ.
X
k=1
=
the distributed SPR-equations (16), and the LMIs (21)
for k = 1, ..., N , where Qk , Sk are defined in (14), and
e ⊤ H.
e
W k = Wk + H
N
X
k=1 j∈Nk
⊤
⊤
Q k + Sk
Pkj
Bθ Pkj
Bw
1
1
..
..
.
.
< 0,
⊤
⊤
Pkjp Bθ Pkjp Bw
k
k
2
∗
∗
∗ −τ I
0
∗
∗
∗
0
−γ 2 I
N
X
⊤
e⊤
k Pk ėk + ėk Pk ek
+
(26)
k = 1, ..., N
Pk Bw
N
X
k=1
N
X
e k , Kkj , K
e kj , k =
and let the collection of matrices Pk , Pkj , Lk , L
1, ..., N , be a solution to the matrix inequalities
k=1
e⊤
k Qk ek < −
(29)
For the Lie-derivative of V (e), applying (23), (29), and
the same change of index as in (19) leads to
Theorem 3 (H∞ -performance) Consider a nonlinear system (1). Define the following matrices
Ak = A − Lk Ck −
N
X
e⊤
k Pk Bk +
X
j∈Nk
⊤
ψk (zk , t)
e⊤
j Pkj Bk
2 ⊤
2 ⊤
−e⊤
k W k ek + τ ∆k ∆k + γ w w
N
X
k=1
⊤
e⊤
k Ek +
|
X
j∈Nk
{z
⊤
zk
⊤
e⊤
ψk (zk , t)
j Ekj
}
(30)
With the quadratic constraint (3) on the nonlinearity
⊤ e⊤ e
θ, we have τ 2 ∆⊤
k ∆k < ek H Hek . Then, due to the
e ⊤H
e and the sector property of
definition W k = Wk + H
ψk (25), the Lie-derivative of V (e) finally is
(28)
∆k
and in addition, as (14) is satisfied by (27), we know that
(20) holds which ensures that
V̇ (e) <
N
X
k=1
7
2 ⊤
−e⊤
k Wk ek + γ w w .
(31)
Integrating over (0, ∞) now yields the desired H∞ performance
Z
∞
V̇ (e)dt +
0
N Z ∞
X
0
k=1
N Z ∞
X
k=1
0
e⊤
k Wk ek dt <
e⊤
k Wk ek dt
Z
<γ
Qk (k, k) =Pk A − Gk Ck −
N γ 2 w⊤ wdt
⊤
+ λk Pkj2
Qk (j1 , j2 ) =λk Pkj
1
+ I0 ,
N X
X
k=1 j∈Nk
kPk,j k
for j, j1 , j2 ∈ Nk
Step 2: Set Pk , k = 1, ..., N and Pkj , j ∈ Nk as the
results from Step 1 and solve (27) with the remaining
variables. Optionally, γ can be also defined as variable
to be minimized.
Through the off-diagonal blocks, the problem becomes
non-convex as discussed in many papers on decentralized
control, e.g. [18], [19], [20]. Due to this non-convexity,
there is no general solution method available, but instead, alternative design methods are required. While
Youla-Parametrization as in [18], [20] is not suitable for
the class of interconnection graphs under consideration,
and a robustness argument as used in [19] has proven as
too conservative in the present context. In the following we will present a two-step solution strategy that has
proven to be efficient of solving the matrix inequalities
(27).
min
l∈Nk
Through replacing the elements of Qk from (15) with
(34), the matrix inequality (27) is turned into a LMI.
Moreover, the minimization (33) can be executed in parallel fashion. The minimization of the off-diagonal blocks
Pkj ensures that they are only as large as needed for
(16). Then, the exact filter gains need to be calculated
in the second step, where the feasibility is enhanced if
the off-diagonal blocks are small.
Numerical calculation
Step 1: Solve
⊤
Fkl
(34)
and Sk defined in (14). Gk and Fkj are matrix variables
of suitable dimension and λk , πk , k = 1, ..., N are scalar
parameters.
(32)
for I0 = V (e(0)).
Theorem 3 gives us sufficient conditions for designing
distributed estimators that satisfy the distributed KYPLemma from 2 and moreover guarantees robust performance with respect to input disturbances. When substituting (26), however,one can easily see that the matrix
inequalities (2) are not linear in the solution variables
Lk , Kkj , Pk , Pkj for k = 1, ..., N and j ∈ Nk . In the intuitive approach, when there are no off-diagonal blocks
Pkj , a simple substitution Gk = Pk Lk and Fk = Pk Kk
suffices to turn the matrix inequalities into LMIs. Since
in the general case, off-diagonal blocks may be nonzero,
we need to investigate in efficient solution strategies that
can specifically find a suitable solution to the conditions
of Theorem 2.
5.3
X
⊤
Qk (k, j) =Fkj + A⊤ Pkj − λk CK
Ck Pkj − λk pk Pkj
0
kwk2L2
Fkl
l∈Nk
+ A⊤ Pk − Ck⊤ G⊤
k −
∞
2
X
An example where this 2 step approach is used will be
given in the following. This two-step approach for computation has proven capable of solving numerous cases
where the intuitive approach from Section 3 fails.
6
Simulation example
We consider our example from Section 3 (9), (10). Figure
1 and 2 show the simulation results after applying our 2step approach with the parameters πk = 0.1 and λk = 1
for all k = 1, ..., N , and the performance √
parameter γ =
4. The nonlinearity is defined as φ(y) = 3 y.
(33)
40
subject to (16), (21), (27),
20
where (27) is defined with
Qk =
Qk (k, k) Qk (k, j1 ) Qk (k, j2 ) . . . Qk (k, jpk )
∗
∗
∗
∗
0
Qk (j1 , j1 ) Qk (j1 , j2 ) . . . Qk (j1 , jpk )
∗
Qk (j2 , j2 ) . . . Qk (j1 , jpk )
..
..
.
∗
∗
.
∗
∗
∗ Qk (jpk , jpk )
−20
−40
0
5
10
15
t
20
25
30
Fig. 1. Plots of x1 and x2 . Red is the actual state, black are
the estimates.
with
8
60
[7] Maxim V. Subbotin and Roy S. Smith. Design of
distributed decentralized estimators for formations
with fixed and stochastic communication topologies. Automatica, 45(11):2491 – 2501, 2009.
[8] Valery Ugrinovskii and Cédric Langbort. Distributed H∞ consensus-based estimation of uncertain systems via dissipativity theory. IET Control
Theory & Applications, 5(12):1458–1469, 2011.
[9] Valery Ugrinovskii. Distributed robust filtering
with H∞ consensus of estimates. Automatica,
47(1):1–13, 2011.
[10] M. Zeitz. The extended luenberger observer for nonlinear systems. Systems & Control Letters, 9(2):149
– 156, 1987.
[11] Hassan K Khalil and Laurent Praly. High-gain observers in nonlinear feedback control. International
Journal of Robust and Nonlinear Control, 24:993–
1015, 2014.
[12] Murat Arcak and Petar Kokotovic. Nonlinear observers : a circle criterion design and robustness
analysis. Automatica, 37:1923–1930, 2001.
[13] Behcet Acikmese and Milan Mandic. Decentralized observer with a consensus filter for distributed
discrete-time linear systems. In Proc. 2011 American Control Conf., pages 4723–4730, San Francisco,
CA, USA, 2011.
[14] Jingbo Wu, Valery Ugrinovskii, and Frank
Allgöwer. Cooperative estimation for synchronization of heterogeneous multi-agent systems using
relative information. In Proc. 19th IFAC World
Congress, 2014. accepted.
[15] Jingbo Wu, Li Li, Valery Ugrinovskii, and Frank
Allgower. Distributed filter design for cooperative H-infinity-type estimation. In Proc. MultiConference on Systems and Control, page accepted,
2015.
[16] H.K. Khalil. Nonlinear Systems (3rd Edition).
Pearson Education. Prentice Hall, 2001.
[17] Cheng Yi Zhang, Shuanghua Luo, Aiqun Huang,
and Junxiang Lu. The eigenvalue distribution of
block diagonally dominant matrices and block Hmatrices. Electronic Journal of Linear Algebra,
20(September):621–639, 2010.
[18] Carsten W. Scherer. Structured finite-dimensional
controller design by convex optimization. Linear Algebra and its Applications, 351-352:639–669, 2002.
[19] Srdjan S. Stanković, Dušan M. Stipanović, and
Dragoslav D. Šiljak. Decentralized dynamic output
feedback for robust stabilization of a class of nonlinear interconnected systems. Automatica, 43(5):861–
867, 2007.
[20] John Swigart. Optimal Controller Synthesis for Decentralized Systems. PhD thesis, 2010.
40
20
0
−20
−40
−60
0
5
10
15
t
20
25
30
Fig. 2. Plots of the estimation error of all estimators.
7
Conclusion
In this paper, we discussed the extension of results from
distributed estimation to nonlinear systems. While globally Lipschitz nonlinearities pose little problems, the
Circle Criterion approach is far more challenging, requiring us to relax the usual assumption of a sum-ofsquares Lyapunov-function. The new problem is nonconvex, however, we presented an efficient solution algorithm which makes use of the exact structure of the
problem, and is suitable for distributed calculation.
8
Acknowledgment
The authors would like to thank Prof. Hyungbo Shim,
Prof. Valery Ugrinovskii, and Dr. Liron Allerhand for
fruitful discussions.
References
[1] D.G. Luenberger. Observers for multivariable systems. Automatic Control, IEEE Transactions on,
11(2):190–197, Apr 1966.
[2] Rudolph Emil Kalman. A new approach to linear
filtering and prediction problems. Transactions of
the ASME–Journal of Basic Engineering, 82(Series
D):35–45, 1960.
[3] U. Shaked. H∞ -minimum error state estimation
of linear stationary processes. Automatic Control,
IEEE Transactions on, 35(5):554–558, May 1990.
[4] Reza Olfati-Saber. Distributed Kalman Filter with
Embedded Consensus Filters. In Proc. 44th IEEE
Conf. Decision and Control, pages 8179–8184, 2005.
[5] Reza Olfati-Saber, J. Alexander Fax, and
Richard M. Murray. Consensus and cooperation in
networked multi-agent systems. Proceedings of the
IEEE, 95(1):215–233, 2007.
[6] Ruggero Carli, Alessandro Chiuso, Luca Schenato,
and Sandro Zampieri. Distributed Kalman filtering
based on consensus strategies. IEEE J. on Selected
Areas in Comm., 26(4):622–633, 2008.
9
| 3cs.SY
|
arXiv:1312.6184v7 [cs.LG] 11 Oct 2014
Do Deep Nets Really Need to be Deep?
***
Draft for NIPS 2014 (not camera ready copy)
***
Rich Caruana
Microsoft Research
[email protected]
Lei Jimmy Ba
University of Toronto
[email protected]
Abstract
Currently, deep neural networks are the state of the art on problems such as speech
recognition and computer vision. In this paper we empirically demonstrate that
shallow feed-forward nets can learn the complex functions previously learned by
deep nets and achieve accuracies previously only achievable with deep models.
Moreover, in some cases the shallow neural nets can learn these deep functions
using the same number of parameters as the original deep models. On the TIMIT
phoneme recognition and CIFAR-10 image recognition tasks, shallow nets can be
trained that perform similarly to complex, well-engineered, deeper convolutional
architectures.
1
Introduction
You are given a training set with 1M labeled points. When you train a shallow neural net with one
fully-connected feed-forward hidden layer on this data you obtain 86% accuracy on test data. When
you train a deeper neural net as in [1] consisting of a convolutional layer, pooling layer, and three
fully-connected feed-forward layers on the same data you obtain 91% accuracy on the same test set.
What is the source of this improvement? Is the 5% increase in accuracy of the deep net over the
shallow net because: a) the deep net has more parameters; b) the deep net can learn more complex
functions given the same number of parameters; c) the deep net has better bias and learns more
interesting/useful functions (e.g., because the deep net is deeper it learns hierarchical representations
[5]); d) nets without convolution can’t easily learn what nets with convolution can learn; e) current
learning algorithms and regularization methods work better with deep architectures than shallow
architectures[8]; f) all or some of the above; g) none of the above?
There have been attempts to answer the question above. It has been shown that deep nets coupled
with unsupervised layer-by-layer pre-training technique[10] [19] work well. In [8], the authors show
that depth combined with pre-training provides a good prior for model weights, thus improving generalization. There is well-known early theoretical work on the representational capacity of neural
nets. For example, it was proved that a network with a large enough single hidden layer of sigmoid
units can approximate any decision boundary[4]. Empirical work, however, shows that it is difficult
to train shallow nets to be as accurate as deep nets. For vision tasks, a recent study on deep convolutional nets suggests that deeper models are preferred under a parameter budget [7]. In [5], the
authors trained shallow nets on SIFT features to classify a large-scale ImageNet dataset and showed
that it is challenging to train large shallow nets to learn complex functions. And in [17], the authors
show that deeper models are more competitive than shallow models in speech acoustic modeling.
1
In this paper we provide empirical evidence that shallow nets are capable of learning the same
function as deep nets, and in some cases with the same number of parameters as the deep nets. We
do this by first training a state-of-the-art deep model, and then training a shallow model to mimic the
deep model. The mimic model is trained using the model compression scheme described in the next
section. Remarkably, with model compression we are able to train shallow nets to be as accurate
as some deep models, even though we are not able to train these shallow nets to be as accurate as
the deep nets when the shallow nets are trained directly on the original labeled training data. If a
shallow net with the same number of parameters as a deep net can learn to mimic a deep net with
high fidelity, then it is clear that the function learned by that deep net does not really have to be deep.
2
2.1
Training Shallow Nets to Mimic Deep Nets
Model Compression
The main idea behind model compression is to train a compact model to approximate the function
learned by a larger, more complex model. For example, in [3], a single neural net of modest size
could be trained to mimic a much larger ensemble of models — although the small neural nets
contained 1000 times fewer parameters, often they were just as accurate as the ensembles they were
trained to mimic. Model compression works by passing unlabeled data through the large, accurate
model to collect the scores produced by that model. This synthetically labeled data is then used to
train the smaller mimic model. The mimic model is not trained on the original labels—it is trained
to learn the function that was learned by the larger model. If the compressed model learns to mimic
the large model perfectly it makes exactly the same predictions and mistakes as the complex model.
Surprisingly, often it is not (yet) possible to train a small neural net on the original training data to be
as accurate as the complex model, nor as accurate as the mimic model. Compression demonstrates
that a small neural net could, in principle, learn the more accurate function, but current learning
algorithms are unable to train a model with that accuracy from the original training data; instead, we
must train the complex intermediate model first and then train the neural net to mimic it. Clearly,
when it is possible to mimic the function learned by a complex model with a small net, the function
learned by the complex model wasn’t truly too complex to be learned by a small net. This suggests
to us that the complexity of a learned model, and the size of the representation best used to learn that
model, are different things. In this paper we apply model compression to train shallow neural nets
to mimic deeper neural nets, thereby demonstrating that deep neural nets may not need to be deep.
2.2
Mimic Learning via Regressing Logit with L2 Loss
On both TIMIT and CIFAR-10 we train shallow mimic nets using data labeled by either a deep
net, or an ensemble of deep nets, trained on the original TIMIT or CIFAR-10 training data. The
deep models are trained in the usual way using softmax output and cross-entropy cost function. The
shallow mimic models,
however, instead of being trained with cross-entropy on the 183 p values
P
where pk = ezk / j ezj output by the softmax layer from the deep model, are trained directly on
the 183 log probability values z, also called logit, before the softmax activation.
Training on these logit values makes learning easier for the shallow net by placing emphasis on
all prediction targets. Because the logits capture the logarithm relationships between the probability predictions, a student model trained on logits has to learn all of the additional fine detailed
relationships between labels that is not obvious in the probability space yet was learned by the
teacher model. For example, assume there are three targets that the teacher predicts with probability
[2e − 9, 4e − 5, 0.9999]. If we use these probabilities as prediction targets directly to minimize a
cross entropy loss function, the student will focus on the third target and easily ignore the first and
second target. Alternatively, one can extract the logit prediction from the teacher model and obtain
our new targets [10, 20, 30]. The student will learn to regress the third target, yet it still learns the
first and second target along with their relative difference. The logit values provide richer information to student to mimic the exact behaviours of a teach model. Moreover, consider a second training
case where the teacher predicts logits [−10, 0, 10]. After softmax, these logits yield the same predicted probabilities as [10, 20, 30], yet clearly the teacher has learned internally to model these two
cases very differently. By training the student model on the logits directly, the student is better able
2
to learn the internal model learned by the teacher, without suffering from the information loss that
occurs after passing through the logits to probability space.
We formulate the SNN-MIMIC learning objective function as a regression problem given training
data {(x(1) , z (1) ),...,(x(T ) , z (T ) ) }:
L(W, β) =
1 X
||g(x(t) ; W, β) − z (t) ||22 ,
2T t
(1)
where, W is the weight matrix between input features x and hidden layer, β is the weights from
hidden to output units, g(x(t) ; W, β) = βf (W x(t) ) is the model prediction on the tth training data
point and f (·) is the non-linear activation of the hidden units. The parameters W and β are updated
using standard error back-propagation algorithm and stochastic gradient descent with momentum.
We have also experimented with other different mimic loss function, such as minimizing the KL
divergence KL(pteacher kpstudent ) cost function and L2 loss on the probability. Logits regression outperforms all the other loss functions and is one of the key technique for obtaining the results in the
rest of this paper. We found that normalizing the logits from the teacher model, by subtracting the
mean and dividing the standard deviation of each target across the training set, can improve the L2
loss slightly during training. Normalization is not crucial for obtaining a good student model.
2.3
Speeding-up Mimic Learning by Introducing a Linear Layer
To match the number of parameters in a deep net, a shallow net has to have more non-linear hidden
units in a single layer to produce a large weight matrix W . When training a large shallow neural
network with many hidden units, we find it is very slow to learn the large number of parameters in the
weight matrix between input and hidden layers of size O(HD), where D is input feature dimension
and H is the number of hidden units. Because there are many highly correlated parameters in this
large weight matrix gradient descent converges slowly. We also notice that during learning, shallow
nets spend most of the computation in the costly matrix multiplication of the input data vectors and
large weight matrix. The shallow nets eventually learn accurate mimic functions, but training to
convergence is very slow (multiple weeks) even with a GPU.
We found that introducing a bottleneck linear layer with k linear hidden units between the input
and the non-linear hidden layer sped up learning dramatically: we can factorize the weight matrix
W ∈ RH×D into the product of two low rank matrices, U ∈ RH×k and V ∈ Rk×D , where
k << D, H. The new cost function can be written as:
1 X
||βf (U V x(t) ) − z (t) ||22
(2)
L(U, V, β) =
2T t
The weights U and V can be learnt by back-propagating through the linear layer. This reparameterization of weight matrix W not only increases the convergence rate of the shallow mimic
nets, but also reduces memory space from O(HD) to O(k(H + D)).
Factorizing weight matrices has been previously explored in [16] and [20]. While these prior works
focus on using matrix factorization in the last output layer, our method is applied between input and
hidden layer to improve the convergence speed during training.
The reduced memory usage enables us to train large shallow models that were previously infeasible
due to excessive memory usage. The linear bottle neck can only reduce the representational power
of the network, and it can always be absorbed into a signle weight matrix W .
3
TIMIT Phoneme Recognition
The TIMIT speech corpus has 462 speakers in the training set. There is a separate development set
for cross-validation including 50 speakers, and a final test set with 24 speakers. The raw waveform
audio data were pre-processed using 25ms Hamming window shifting by 10ms to extract Fouriertransform-based filter-banks with 40 coefficients (plus energy) distributed on a mel-scale, together
3
with their first and second temporal derivatives. We included +/- 7 nearby frames to formulate the
final 1845 dimension input vector. The data input features were normalized by subtracting the mean
and dividing by the standard deviation on each dimension. All 61 phoneme labels are represented in
tri-state, i.e., 3 states for each of the 61 phonemes, yielding target label vectors with 183 dimensions
for training. At decoding time these are mapped to 39 classes as in [13] for scoring.
3.1
Deep Learning on TIMIT
Deep learning was first successfully applied to speech recognition in [14]. We follow the same
framework and train two deep models on TIMIT, DNN and CNN. DNN is a deep neural net consisting of three fully-connected feedforward hidden layers consisting of 2000 rectified linear units
(ReLU) [15] per layer. CNN is a deep neural net consisting of a convolutional layer and max-pooling
layer followed by three hidden layers containing 2000 ReLU units [2]. The CNN was trained using
the same convolutional architecture as in [6]. We also formed an ensemble of nine CNN models,
ECNN.
The accuracy of DNN, CNN, and ECNN on the final test set are shown in Table 1. The error rate
of the convolutional deep net (CNN) is about 2.1% better than the deep net (DNN). The table also
shows the accuracy of shallow neural nets with 8000, 50,000, and 400,000 hidden units (SNN-8k,
SNN-50k, and SNN-400k) trained on the original training data. Despite having up to 10X as many
parameters as DNN, CNN and ECNN, the shallow models are 1.4% to 2% less accurate than the
DNN, 3.5% to 4.1% less accurate than the CNN, and 4.5% to 5.1% less accurate than the ECNN.
3.2
Learning to Mimic an Ensemble of Deep Convolutional TIMIT Models
The most accurate single model we trained on TIMIT is the deep convolutional architecture in [6].
Because we have no unlabeled data from the TIMIT distribution, we are forced to use the same 1.1M
points in the train set as unlabeled data for compression by throwing away their labels.1 Re-using
the train set reduces the accuracy of the mimic models, increasing the gap between the teacher and
mimic models on test data: model compression works best when the unlabeled set is much larger
than the train set, and when the unlabeled samples do not fall on train points where the teacher model
is more likely to have overfit. To reduce the impact of the gap caused by performing compression
with the original train set, we train the student model to mimic a more accurate ensemble of deep
convolutional models.
We are able to train a more accurate model on TIMIT by forming an ensemble of 9 deep, convolutional neural nets, each trained with somewhat different train sets, and with architectures with
different kernel sizes in the convolutional layers. We used this very accurate model, ECNN, as the
teacher model to label the data used to train the shallow mimic nets. As described in Section 2.2,
the logits (log probability of the predicted values) from each CNN in the ECNN model are averaged
and the average logits are used as final regression targets to train the mimic SNNs.
We trained shallow mimic nets with 8k (SNN-MIMIC-8k) and 400k (SNN-MIMIC-400k) hidden
units on the re-labeled 1.1M training points. As described in Section 2.3, both mimic models have
250 linear units between the input and non-linear hidden layer to speed up learning — preliminary
experiments suggest that for TIMIT there is little benefit from using more than 250 linear units.
3.3
Compression Results For TIMIT
The bottom of Table 1 shows the accuracy of shallow mimic nets with 8000 ReLUs and 400,000
ReLUs (SNN-MIMIC-8k and -400k) trained with model compression to mimic the ECNN. Surprisingly, shallow nets are able to perform as well as their deep counter-parts when trained with model
compression to mimic a more accurate model. A neural net with one hidden layer (SNN-MIMIC-8k)
can be trained to perform as well as a DNN with a similar number of parameters. Furthermore, if we
increase the number of hidden units in the shallow net from 8k to 400k (the largest we could train),
we see that a neural net with one hidden layer (SNN-MIMIC-400k) can be trained to perform comparably to a CNN even though the SNN-MIMIC-400k net has no convolutional or pooling layers.
1
That SNNs can be trained to be as accurate as DNNs using only the original training data data highlights
that it should be possible to train accurate SNNs on the original train data given better learning algorithms.
4
SNN-8k
SNN-50k
SNN-400k
DNN
CNN
ECNN
SNN-MIMIC-8k
SNN-MIMIC-400k
Architecture
# Param.
# Hidden units
PER
8k + dropout
trained on original data
50k + dropout
trained on original data
250L-400k + dropout
trained on original data
2k-2k-2k + dropout
trained on original data
c-p-2k-2k-2k + dropout
trained on original data
∼12M
∼8k
23.1%
∼100M
∼50k
23.0%
∼180M
∼400k
23.6%
∼12M
∼6k
21.9%
∼13M
∼10k
19.5%
ensemble of 9 CNNs
∼125M
∼90k
18.5%
250L-8k
no convolution or pooling layers
250L-400k
no convolution or pooling layers
∼12M
∼8k
21.6%
∼180M
∼400k
20.0%
Table 1: Comparison of shallow and deep models: phone error rate (PER) on TIMIT core test set.
83
82
ShallowNet
DeepNet
ShallowMimicNet
Convolutional Net
Ensemble of CNNs
81
80
79
78
77
76
ShallowNet
DeepNet
ShallowMimicNet
Convolutional Net
Ensemble of CNNs
81
Accuracy on TIMIT Test Set
Accuracy on TIMIT Dev Set
82
80
79
78
77
76
1
10
75
100
Number of Parameters (millions)
1
10
100
Number of Parameters (millions)
Figure 1: Accuracy of SNNs, DNNs, and Mimic SNNs vs. # of parameters on TIMIT Dev (left) and
Test (right) sets. Accuracy of the CNN and target ECNN are shown as horizontal lines for reference.
This is interesting because it suggests that a large single hidden¡ layer without a topology custom
designed for the problem is able to reach the performance of a deep convolutional neural net that
was carefully engineered with prior structure and weight sharing without any increase in the number
of training examples, even though the same architecture trained on the original data could not.
Figure 1 shows the accuracy of shallow nets and deep nets trained on the original TIMIT 1.1M data,
and shallow mimic nets trained on the ECNN targets, as a function of the number of parameters in
the models. The accuracy of the CNN and the teacher ECNN are shown as horizontal lines at the top
of the figures. When the number of parameters is small (about 1 million), the SNN, DNN, and SNNMIMIC models all have similar accuracy. As the size of the hidden layers increases and the number
of parameters increases, the accuracy of a shallow model trained on the original data begins to lag
behind. The accuracy of the shallow mimic model, however, matches the accuracy of the DNN until
about 4 million parameters, when the DNN begins to fall behind the mimic. The DNN asymptotes
at around 10M parameters, while the shallow mimic continues to increase in accuracy. Eventually
the mimic asymptotes at around 100M parameters to an accuracy comparable to that of the CNN.
The shallow mimic never achieves the accuracy of the ECNN it is trying to mimic (because there
is not enough unlabeled data), but it is able to match or exceed the accuracy of deep nets (DNNs)
having the same number of parameters trained on the original data.
5
4
Object Recognition: CIFAR-10
To verify that the results on TIMIT generalize to other learning problems and task domains, we ran
similar experiments on the CIFAR-10 Object Recognition Task[12]. CIFAR-10 consists of a set
of natural images from 10 different object classes: airplane, automobile, bird, cat, deer, dog, frog,
horse, ship, truck. The dataset is a labeled subset of the 80 million tiny images dataset[18] and is
divided into 50,000 train and 10,000 test images. Each image is 32x32 pixels in 3 color channels,
yielding input vectors with 3072 dimensions. We prepared the data by subtracting the mean and
dividing the standard deviation of each image vector to perform global contrast normalization. We
then applied ZCA whitening to the normalized images. This pre-processing is the same used in [9].
4.1
Learning to Mimic a Deep Convolutional Neural Network
Deep learning currently achieves state-of-the-art accuracies on many computer vision problems. The
key to this success is deep convolutional nets with many alternating layers of convolutional, pooling
and non-linear units. Recent advances such as dropout are also important to prevent over-fitting in
these deep nets.
We follow the same approach as with TIMIT: An ensemble of deep CNN models is used to label
CIFAR-10 images for model compression. The logit predictions from this teacher model are used
as regression targets to train a mimic shallow neural net (SNN). CIFAR-10 images have a higher
dimension than TIMIT (3072 vs. 1845), but the size of the CIFAR-10 training set is only 50,000
compared to 1.1 million examples for TIMIT. Fortunately, unlike TIMIT, in CIFAR-10 we have
access to unlabeled data from a similar distribution by using the super set of CIFAR-10: the 80
million tiny images dataset. We add the first 1 million images from the 80 million set to the original
50,000 CIFAR-10 training images to create a 1.05M mimic training (transfer) set.
CIFAR-10 images are raw pixels for objects viewed from many different angles and positions,
whereas TIMIT features are human-designed filter-bank features. In preliminary experiments we
observed that non-convolutional nets do not perform well on CIFAR-10 no matter what their depth.
Instead of raw pixels, the authors in [5] trained their shallow models on the SIFT features. Similarly,
[7] used a base convolution and pooling layer to study different deep architectures. We follow the
approach in [7] to allow our shallow models to benefit from convolution while keeping the models
as shallow as possible, and introduce a single layer of convolution and pooling in our shallow mimic
models to act as a feature extractor to create invariance to small translations in the pixel domain. The
SNN-MIMIC models for CIFAR-10 thus consist of a convolution and max pooling layer followed
by fully connected 1200 linear units and 30k non-linear units. As before, the linear units are there
only to speed learning; they do not increase the model’s representational power and can be absorbed
into the weights in the non-linear layer after learning.
Results on CIFAR-10 are consistent with those from TIMIT. Table 2 shows results for the shallow
mimic models, and for much-deeper convolutional nets. The shallow mimic net trained to mimic
the teacher CNN (SNN-CNN-MIMIC-30k) achieves accuracy comparable to CNNs with multiple
convolutional and pooling layers. And by training the shallow model to mimic the ensemble of
CNNs (SNN-ECNN-MIMIC-30k), accuracy is improved an additional 0.9%. The mimic models
are able to achieve accuracies previously unseen on CIFAR-10 with models with so few layers.
Although the deep convolution nets have more hidden units than the shallow mimic models, because
of weight sharing, the deeper nets with multiple convolution layers have fewer parameters than the
shallow fully-connected mimic models. Still, it is surprising to see how accurate the shallow mimic
models are, and that their performance continues to improve as the performance of the teacher model
improves (see further discussion of this in Section 5.2).
5
5.1
Discussion
Why Mimic Models Can Be More Accurate than Training on Original Labels
It may be surprising that models trained on the prediction targets taken from other models can be
more accurate than models trained on the original labels. There are a variety of reasons why this can
happen:
6
DNN
SNN-30k
single-layer
feature extraction
CNN[11]
(no augmentation)
CNN[21]
(no augmentation)
teacher CNN
(no augmentation)
ECNN
(no augmentation)
SNN-CNN-MIMIC-30k
trained on a single CNN
SNN-CNN-MIMIC-30k
trained on a single CNN
SNN-ECNN-MIMIC-30k
trained on ensemble
Architecture
# Param.
# Hidden units
Err.
2000-2000 + dropout
∼10M
4k
57.8%
128c-p-1200L-30k
+ dropout input&hidden
4000c-p
followed by SVM
64c-p-64c-p-64c-p-16lc
+ dropout on lc
64c-p-64c-p-128c-p-fc
+ dropout on fc
and stochastic pooling
128c-p-128c-p-128c-p-1000fc
+ dropout on fc
and stochastic pooling
∼70M
∼190k
21.8%
∼125M
∼3.7B
18.4%
∼10k
∼110k
15.6%
∼56k
∼120k
15.13%
∼35k
∼210k
12.0%
ensemble of 4 CNNs
∼140k
∼840k
11.0%
64c-p-1200L-30k
with no regularization
128c-p-1200L-30k
with no regularization
128c-p-1200L-30k
with no regularization
∼54M
∼110k
15.4%
∼70M
∼190k
15.1%
∼70M
∼190k
14.2%
Table 2: Comparison of shallow and deep models: classification error rate on CIFAR-10. Key: c,
convolution layer; p, pooling layer; lc, locally connected layer; fc, fully connected layer
• if some labels have errors, the teacher model may eliminate some of these errors (i.e., censor the data), thus making learning easier for the student: on TIMIT, there are mislabeled
frames introduced by the HMM forced-alignment procedure.
• if there are regions in the p(y|X) that are difficult to learn given the features, sample density, and function complexity, the teacher may provide simpler, soft labels to the student.
The complexity in the data set has been washed away by filtering the targets through the
teacher model.
• learning from the original hard 0/1 labels can be more difficult than learning from the
teacher’s conditional probabilities: on TIMIT only one of 183 outputs is non-zero on each
training case, but the mimic model sees non-zero targets for most outputs on most training
cases. Moreover, the teacher model can spread the uncertainty over multiple outputs when
it is not confident of its prediction. Yet, the teacher model can concentrate the probability
mass on one (or few) outputs on easy cases. The uncertainty from the teacher model is far
more informative to guiding the student model than the original 0/1 labels. This benefit
appears to be further enhanced by training on logits.
77.5
77.0
SNN-8k
SNN-8k + dropout
SNN-MIMIC-8k
76.5
Phone Recognition Accuracy
The mechanisms above can be seen as forms
of regularization that help prevent overfitting in
the student model. Shallow models trained on
the original targets are more prone to overfitting than deep models—they begin to overfit
before learning the accurate functions learned
by deeper models even with dropout (see Figure 2). If we had more effective regularization
methods for shallow models, some of the performance gap between shallow and deep models might disappear. Model compression appears to be a form of regularization that is effective at reducing this gap.
76.0
75.5
75.0
74.5
74.00
2
4
6
8
Number of Epoches
10
12
14
Figure 2: Training shallow mimic model prevents
7overfitting.
5.2 The Capacity and Representational
Power of Shallow Models
Accuracy of Mimic Model on Dev Set
Figure 3 shows results of an experiment with
TIMIT where we trained shallow mimic mod83
els of two sizes (SNN-MIMIC-8k and SNNMIMIC-160k) on teacher models of different
82
accuracies. The two shallow mimic models are
trained on the same number of data points. The
81
only difference between them is the size of the
hidden layer. The x-axis shows the accuracy of
80
the teacher model, and the y-axis is the accuracy of the mimic models. Lines parallel to the
79
diagonal suggest that increases in the accuracy
of the teacher models yield similar increases in
78
the accuracy of the mimic models. Although
78
79
80
81
82
83
the data does not fall perfectly on a diagonal,
Accuracy of Teacher Model on Dev Set
there is strong evidence that the accuracy of the
mimic models continues to increase as the ac- Figure 3: Accuracy of student models continues to
curacy of the teacher model improves, suggest- improve as accuracy of teacher models improves.
ing that the mimic models are not (yet) running
out of capacity. When training on the same targets, SNN-MIMIC-8k always perform worse than
SNN-MIMIC-160K that has 10 times more parameters. Although there is a consistent performance
gap between the two models due to the difference in size, the smaller shallow model was eventually able to achieve a performance comparable to the larger shallow net by learning from a better
teacher, and the accuracy of both models continues to increase as teacher accuracy increases. This
suggests that shallow models with a number of parameters comparable to deep models are likely
capable of learning even more accurate functions if a more accurate teacher and/or more unlabeled
data became available. Similarly, on CIFAR-10 we saw that increasing the accuracy of the teacher
model by forming an ensemble of deep CNNs yielded commensurate increase in the accuracy of the
student model. We see little evidence that shallow models have limited capacity or representational
power. Instead, the main limitation appears to be the learning and regularization procedures used to
train the shallow models.
5.3
Mimic with 8k Non-Linear Units
Mimic with 160k Non-Linear Units
y=x (no student-teacher gap)
Parallel Distributed Processing vs. Deep Sequential Processing
Our results show that shallow nets can be competitive with deep models on speech and vision tasks.
One potential benefit of shallow nets is that training them scales well with the modern parallel
hardware. In our experiments the deep models usually required 8–12 hours to train on Nvidia GTX
580 GPUs to reach the state-of-the-art performance on TIMIT and CIFAR-10 datasets. Although
some of the shallow mimic models have more parameters than the deep models, the shallow models
train much faster and reach similar accuracies in only 1–2 hours.
Also, given parallel computational resources, at run-time shallow models can finish computation in
2 or 3 cycles for a given input, whereas a deep architecture has to make sequential inference through
each of its layers, expending a number of cycles proportional to the depth of the model. This benefit
can be important in on-line inference settings where data parallelization is not as easy to achieve
as it is in the batch inference setting. For real-time applications such as surveillance or real-time
speech translation, a model that responds in fewer cycles can be beneficial.
6
Future Work
The tiny images dataset contains 80 millions images. We are currently investigating if by labeling
these 80M images with a teacher, it is possible to train shallow models with no convolutional or
pooling layers to mimic deep convolutional models.
This paper focused on training the shallowest-possible models to mimic deep models in order to
better understand the importance of model depth in learning. As suggested in Section 5.3, there are
practical applications of this work as well: student models of small-to-medium size and depth can be
trained to mimic very large, high accuracy deep models, and ensembles of deep models, thus yielding
8
better accuracy with reduced runtime cost than is currently achievable without model compression.
This approach allows one to adjust flexibly the trade-off between accuracy and computational cost.
In this paper we are able to demonstrate empirically that shallow models can, at least in principle,
learn more accurate functions without a large increase in the number of parameters. The algorithm
we use to do this—training the shallow model to mimic a more accurate deep model, however,
is awkward. It depends on the availability of either a large unlabeled data set (to reduce the gap
between teacher and mimic model) or a teacher model of very high accuracy, or both. Developing
algorithms to train shallow models of high accuracy directly from the original data without going
through the intermediate teacher model would, if possible, be a significant contribution.
7
Conclusions
We demonstrate empirically that shallow neural nets can be trained to achieve performances previously achievable only by deep models on the TIMIT phoneme recognition and CIFAR-10 image
recognition tasks. Single-layer fully-connected feedforward nets trained to mimic deep models can
perform similarly to well-engineered complex deep convolutional architectures. The results suggest
that the strength of deep learning may arise in part from a good match between deep architectures
and current training procedures, and that it may be possible to devise better learning algorithms to
train more accurate shallow feed-forward nets. For a given number of parameters, depth may make
learning easier, but may not always be essential.
Acknowledgements We thank Li Deng for generous help with TIMIT, Li Deng and Ossama AbdelHamid for code for the TIMIT convolutional model, Chris Burges, Li Deng, Ran Gilad-Bachrach,
Tapas Kanungo and John Platt for discussion that significantly improved this work, and Mike Aultman for help with the GPU cluster.
References
[1] Ossama Abdel-Hamid, Abdel-rahman Mohamed, Hui Jiang, and Gerald Penn. Applying convolutional neural networks concepts to hybrid nn-hmm model for speech recognition. In
Acoustics, Speech and Signal Processing (ICASSP), 2012 IEEE International Conference on,
pages 4277–4280. IEEE, 2012.
[2] Ossama Abdel-Hamid, Li Deng, and Dong Yu. Exploring convolutional neural network structures and optimization techniques for speech recognition. Interspeech 2013, 2013.
[3] Cristian Bucilu, Rich Caruana, and Alexandru Niculescu-Mizil. Model compression. In Proceedings of the 12th ACM SIGKDD international conference on Knowledge discovery and data
mining, pages 535–541. ACM, 2006.
[4] George Cybenko. Approximation by superpositions of a sigmoidal function. Mathematics of
control, signals and systems, 2(4):303–314, 1989.
[5] Yann N Dauphin and Yoshua Bengio. Big neural networks waste capacity. arXiv preprint
arXiv:1301.3583, 2013.
[6] Li Deng, Jinyu Li, Jui-Ting Huang, Kaisheng Yao, Dong Yu, Frank Seide, Michael Seltzer,
Geoff Zweig, Xiaodong He, Jason Williams, et al. Recent advances in deep learning for speech
research at microsoft. ICASSP 2013, 2013.
[7] David Eigen, Jason Rolfe, Rob Fergus, and Yann LeCun. Understanding deep architectures
using a recursive convolutional network. arXiv preprint arXiv:1312.1847, 2013.
[8] Dumitru Erhan, Yoshua Bengio, Aaron Courville, Pierre-Antoine Manzagol, Pascal Vincent,
and Samy Bengio. Why does unsupervised pre-training help deep learning? The Journal of
Machine Learning Research, 11:625–660, 2010.
[9] Ian Goodfellow, David Warde-Farley, Mehdi Mirza, Aaron Courville, and Yoshua Bengio.
Maxout networks. In Proceedings of The 30th International Conference on Machine Learning,
pages 1319–1327, 2013.
[10] G.E. Hinton and R.R. Salakhutdinov. Reducing the dimensionality of data with neural networks. Science, 313(5786):504–507, 2006.
9
[11] G.E. Hinton, N. Srivastava, A. Krizhevsky, I. Sutskever, and R.R. Salakhutdinov. Improving neural networks by preventing co-adaptation of feature detectors. arXiv preprint
arXiv:1207.0580, 2012.
[12] Alex Krizhevsky and Geoffrey Hinton. Learning multiple layers of features from tiny images.
Computer Science Department, University of Toronto, Tech. Rep, 2009.
[13] K-F Lee and H-W Hon. Speaker-independent phone recognition using hidden markov models.
Acoustics, Speech and Signal Processing, IEEE Transactions on, 37(11):1641–1648, 1989.
[14] Abdel-rahman Mohamed, George E Dahl, and Geoffrey Hinton. Acoustic modeling using
deep belief networks. Audio, Speech, and Language Processing, IEEE Transactions on, 20(1):
14–22, 2012.
[15] V. Nair and G.E. Hinton. Rectified linear units improve restricted boltzmann machines. In Proc.
27th International Conference on Machine Learning, pages 807–814. Omnipress Madison,
WI, 2010.
[16] Tara N Sainath, Brian Kingsbury, Vikas Sindhwani, Ebru Arisoy, and Bhuvana Ramabhadran.
Low-rank matrix factorization for deep neural network training with high-dimensional output targets. In Acoustics, Speech and Signal Processing (ICASSP), 2013 IEEE International
Conference on, pages 6655–6659. IEEE, 2013.
[17] Frank Seide, Gang Li, and Dong Yu. Conversational speech transcription using contextdependent deep neural networks. In Interspeech, pages 437–440, 2011.
[18] Antonio Torralba, Robert Fergus, and William T Freeman. 80 million tiny images: A large data
set for nonparametric object and scene recognition. Pattern Analysis and Machine Intelligence,
IEEE Transactions on, 30(11):1958–1970, 2008.
[19] P. Vincent, H. Larochelle, I. Lajoie, Y. Bengio, and P.A. Manzagol. Stacked denoising autoencoders: Learning useful representations in a deep network with a local denoising criterion. The
Journal of Machine Learning Research, 11:3371–3408, 2010.
[20] Jian Xue, Jinyu Li, and Yifan Gong. Restructuring of deep neural network acoustic models
with singular value decomposition. Proc. Interspeech, Lyon, France, 2013.
[21] Matthew D Zeiler and Rob Fergus. Stochastic pooling for regularization of deep convolutional
neural networks. arXiv preprint arXiv:1301.3557, 2013.
10
| 9cs.NE
|
arXiv:1403.0541v1 [cs.AI] 3 Mar 2014
Representing, Reasoning and Answering Questions About
Biological Pathways - Various Applications
Saadat Anwar
Arizona State University,
Tempe, AZ 85233
USA
[email protected]
March 4, 2014
Abstract
Biological organisms are made up of cells containing numerous interconnected biochemical processes. Diseases occur when normal functionality of these processes is disrupted, manifesting as
disease symptoms. Thus, understanding these biochemical processes and their interrelationships
is a primary task in biomedical research and a prerequisite for activities including diagnosing
diseases, and drug development. Scientists studying these interconnected processes have identified various pathways involved in drug metabolism, diseases, and signal transduction, etc.
Over the last decade high-throughput technologies, new algorithms and speed improvements
have resulted in deeper knowledge about biological systems and pathways, resulting in more
refined models. These refined models tend to be large and complex, making it difficult for
a person to remember all aspects of it. Thus, computer models are needed to represent and
analyze them. The refinement activity itself requires reasoning with a pathway model by posing
queries against it and comparing the results against a real biological system. We want to model
biological systems and pathways in such a way that we can answer questions about them.
Many existing models focus on structural and/or factoid questions, relying on surface-level
information that does not require understanding the underlying model. We believe these are
not the kind of questions that a biologist may ask someone to test their understanding of the
biological processes. We want our system to be able to answer the kind of questions a biologist
may ask. So, we turned to early college level text books on biology for such questions.
Thus the main goal of our thesis is to develop a system that allows us to encode knowledge
about biological pathways and answer such questions about them that demonstrate understanding of the pathway. To that end, we develop a language that will allow posing such questions
and illustrate the utility of our framework with various applications in the biological domain.
We use some existing tools with modifications to accomplish our goal.
Finally, we use our question answering system in real world applications by extracting pathway knowledge from text and answering questions related to drug development.
Contents
1
2
Introduction
1.1 Choosing the right questions . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.2 Choosing the right tools . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.3 Need for a pathway a specification and a query language . . . . . . . . . . . .
1.4 Text extraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.5 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.6 Specific contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.6.1 General ASP encoding of Petri Net for simulation . . . . . . . . . . . .
1.6.2 Answering simulation based reasoning questions . . . . . . . . . . . .
1.6.3 BioPathQA: a system and a language to represent pathways and query
them . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.6.4 Text Extraction to Answer Questions about Real World Applications . .
1.7 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
13
14
14
Petri Net Encoding in ASP for Biological Domain
2.1 Introduction . . . . . . . . . . . . . . . . . . . . .
2.2 Background . . . . . . . . . . . . . . . . . . . . .
2.2.1 Answer Set Programming . . . . . . . . .
2.2.2 Multiset . . . . . . . . . . . . . . . . . . .
2.2.3 Petri Net . . . . . . . . . . . . . . . . . .
2.3 Translating Basic Petri Net Into ASP . . . . . . . .
2.3.1 An example execution . . . . . . . . . . .
2.4 Changing Firing Semantics . . . . . . . . . . . . .
2.5 Extension - Reset Arcs . . . . . . . . . . . . . . .
2.6 Extension - Inhibitor Arcs . . . . . . . . . . . . .
2.7 Extension - Read Arcs . . . . . . . . . . . . . . .
2.8 Extension - Colored Tokens . . . . . . . . . . . . .
2.9 Translating Petri Nets with Colored Tokens to ASP
2.10 Extension - Priority Transitions . . . . . . . . . . .
2.11 Extension - Timed Transitions . . . . . . . . . . .
2.12 Other Extensions . . . . . . . . . . . . . . . . . .
16
16
17
17
20
20
22
25
25
26
29
30
31
33
36
37
39
1
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
9
9
10
10
11
11
12
12
13
CONTENTS
3
4
2
2.13 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.14 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
39
39
Answering Questions using Petri Nets and ASP
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.2 Comparing Altered Trajectories due to Reset Intervention . . . . . . . . .
3.3 Determining Conditions Leading to an Observation . . . . . . . . . . . .
3.4 Comparing Altered Trajectories due to Accumulation Intervention . . . .
3.5 Comparing Altered Trajectories due to Initial Value Intervention . . . . .
3.6 Comparing Altered Trajectories due to Inhibition Intervention . . . . . .
3.7 Comparing Altered Trajectories due to Gradient Equilization Intervention
3.8 Comparing Altered Trajectories due to Delay Intervention . . . . . . . .
3.9 Comparing Altered Trajectories due to Priority and Read Interventions . .
3.10 Comparing Altered Trajectories due to Automatic Conversion Intervention
3.11 Comparing Altered Trajectories due to Initial Value Intervention . . . . .
3.12 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
40
40
40
44
46
48
51
54
56
58
61
63
66
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
The BioPathQA System
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.2 Description of BioPathQA . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.3 Syntax of Pathway Specification Language (BioPathQA-PL) . . . . . . . . . .
4.4 Semantics of Pathway Specification Language (BioPathQA-PL) . . . . . . . .
4.4.1 Guarded-Arc Petri Net . . . . . . . . . . . . . . . . . . . . . . . . . .
4.4.2 Construction of Guarded-Arc Petri Net from a Pathway Specification .
4.4.3 Guarded-Arc Petri Net with Colored Tokens . . . . . . . . . . . . . . .
4.4.4 Construction of Guarded-Arc Petri Net with Colored Tokens from a
Pathway Specification with Locational Fluents . . . . . . . . . . . . .
4.5 Syntax of Query Language (BioPathQA-QL) . . . . . . . . . . . . . . . . . .
4.6 Semantics of the Query Language (BioPathQA-QL) . . . . . . . . . . . . . . .
4.6.1 An Illustrative Example . . . . . . . . . . . . . . . . . . . . . . . . .
4.6.2 Domain Transformation due to Interventions and Initial Conditions . .
4.6.3 Formula Semantics . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.6.4 Trajectory Filtering due to Internal Observations . . . . . . . . . . . .
4.6.5 Query Description Satisfaction . . . . . . . . . . . . . . . . . . . . . .
4.6.6 Query Statement Satisfaction . . . . . . . . . . . . . . . . . . . . . . .
4.6.7 Example Encodings . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.6.8 Example Encoding with Conditional Actions . . . . . . . . . . . . . .
4.6.9 ASP Program . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.6.10 Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.6.11 Evaluation Methodolgy . . . . . . . . . . . . . . . . . . . . . . . . . .
4.7 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.7.1 Comparison with π-Calculus . . . . . . . . . . . . . . . . . . . . . . .
68
68
68
69
74
75
78
80
83
85
94
94
96
98
109
113
121
122
130
132
132
148
148
148
CONTENTS
4.8
5
6
3
4.7.2 Comparison with Action Language A .
4.7.3 Comparison with Action Language B .
4.7.4 Comparison with Action Language C .
4.7.5 Comparison with Action Language C+
4.7.6 Comparison with BC . . . . . . . . . .
4.7.7 Comparison with ASPMT . . . . . . .
4.7.8 Comparison with BioSigNet-RR . . . .
Conclusion . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Text Extraction for Real World Applications
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.2 Extracting Relationships about Drug Interactions . . . . . . .
5.2.1 Method . . . . . . . . . . . . . . . . . . . . . . . . .
5.2.2 Results . . . . . . . . . . . . . . . . . . . . . . . . .
5.3 Extracting Knowledge About Genetic Variants . . . . . . . . .
5.3.1 Methods . . . . . . . . . . . . . . . . . . . . . . . .
5.3.2 Results . . . . . . . . . . . . . . . . . . . . . . . . .
5.4 Applying BioPathQA to Drug-Drug Interaction Discovery . .
5.4.1 Drug Administration . . . . . . . . . . . . . . . . . .
5.4.2 Drug Development . . . . . . . . . . . . . . . . . . .
5.4.3 Drug Administration in Presence of Genetic Variation
5.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . .
Conclusion and Future Work
6.1 Pathway Extraction . . . . . . . . .
6.2 Pathway Selection . . . . . . . . . .
6.3 Pathway Modeling . . . . . . . . .
6.4 Pathway Simulation . . . . . . . . .
6.5 Extend High Level Language . . . .
6.6 Result Formatting and Visualization
6.7 Summary . . . . . . . . . . . . . .
A Proofs of Various Propositions
A.1 Proof of Proposition 1 . . .
A.2 Proof of Proposition 2 . . .
A.3 Proof of Proposition 3 . . .
A.4 Poof of Proposition 4 . . .
A.5 Proof of Proposition 5 . . .
A.6 Proof of Proposition 6 . . .
A.7 Proof of Proposition 7 . . .
A.8 Proof of Proposition 8 . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
148
149
149
150
151
152
152
153
.
.
.
.
.
.
.
.
.
.
.
.
155
155
156
157
160
160
161
164
164
165
166
166
167
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
169
171
171
171
172
172
173
173
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
174
183
191
202
213
225
237
248
264
CONTENTS
B Drug-Drug Interaction Queries
B.1 Drug Activates Gene . . . . . . . . . . . . . . .
B.2 Gene Induces Gene . . . . . . . . . . . . . . . .
B.3 Gene Inhibits Gene . . . . . . . . . . . . . . . .
B.4 Drug Changes Gene Expression/Activity . . . . .
B.5 Drug Induces/Stimulates Gene . . . . . . . . . .
B.6 Drug Inhibits Gene . . . . . . . . . . . . . . . .
B.7 Gene Metabolized Drug . . . . . . . . . . . . . .
B.8 Gene Regulates Gene . . . . . . . . . . . . . . .
B.9 Gene Regulate Gene (Xenobiotic Metabolism) . .
B.10 Negative Drug Induces/Metabolizes/Inhibits Gene
B.11 Negative Drug Induces Gene . . . . . . . . . . .
B.12 Negative Drug Inhibits Gene . . . . . . . . . . .
B.13 Negative Gene Metabolizes Drug . . . . . . . . .
B.14 Negative Gene Downregulates Gene . . . . . . .
B.15 Negative Gene Upregulates Gene . . . . . . . . .
B.16 Drug Gene Co-Occurrence . . . . . . . . . . . .
4
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
282
282
283
284
285
286
288
296
299
299
299
301
301
302
302
303
303
List of Figures
2.1
2.2
2.3
2.4
2.5
2.6
3.1
3.2
Petri Net graph (of sub-section of glycolysis pathway) showing places as circles,
transitions as boxes and arcs as directed arrows. Places have token count (or
marking) written above them, assumed 0 when missing. Arcs labels represent
arc-weights, assumed 1 when missing. . . . . . . . . . . . . . . . . . . . . . .
Petri Net of Fig 2.1 extended with a reset arc from dhap to tr shown with double
arrowhead. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Petri Net showing feedback inhibition arc from atp to gly1 with a bullet arrowhead. Inhibitor arc weight is assumed 1 when not specified. . . . . . . . . . . .
Petri Net with read arc from h is to syn shown with arrowhead on both ends.
The transition syn will not fire unless there are at least 25 tokens in h is, but
when it executes, it only consumes 3 tokens. . . . . . . . . . . . . . . . . . . .
Petri Net with tokens of colors {e, h, h2o, nadh, nad p, o2}. Circles represent
places, and rectangles represent transitions. Arc weights such as “nadh/2, h/2”,
“h/2, h2o/1” specify the number of tokens consumed and produced during the
execution of their respective transitions, where “nadh/2, h/2” means 2 tokens
of color nadh and 2 tokens of h. Similar notation is used to specify marking on
places, when not present, the place is assumed to be empty of tokens. . . . . . .
An extended version of the Petri Net model from Fig. 2.5. The new transitions
tq, tcytc have a duration of 2 each (shown in square brackets (“[ ]”) next to the
transition). When missing, transition duration is assumed to be 1. . . . . . . . .
Petri Net graph relevant to question 1. “f16bp” is the compound Fructose 1,6biphosphate, “bpg13” is 1,3-Bisphosphoglycerate. Transition tr shown in dotted
lines is added to model the elimination of dhap as soon as it is produced. . . . .
Amount of “bpg13” produced in unique answer-sets produced by a 10 step simulation. The graph shows two situations, without the (dhap, tr) reset arc (normal
situation) and with the reset arc (abnormal situation). The purpose of this graph
is to depict the variation in the amounts of glycolysis produced in various answer
sets. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5
21
26
29
30
33
38
41
42
LIST OF FIGURES
Average amount of “bpg13” produced during the 10-step simulation at various
time steps. The average is over all answer-sets. The graph shows two situations, without the (dhap, tr) reset arc (normal situation) and with the reset arc
(abnormal situation). The divergence in “bpg13” production is clearly shown. .
3.4 Minimal version of the Petri Net graph in Figure 3.1. All reactants that do not
contribute to the estimation of the rate of glycolysis have been removed. . . . .
3.5 Petri Net graph relevant to question 2. “fats” are fats, “prot” are proteins, “fac”
are fatty acids, “sug” are sugars, “amin” are amino acids and “acoa” is ACoA.
Transition “box” is the beta oxidation, “t5” is glycolysis, “t1” is fat digestion
into fatty acids, and “t9” is protein deamination. . . . . . . . . . . . . . . . . .
3.6 Petri Net graph relevant to question 3. “glu” is Glucose, “pyr” is Pyruvate. Transitions “gly1” represents glycolysis and “cw1” is cellular work that consumes
ATP and produces AMP. Transition “gly1” is inhibited only when the number
of atp tokens is greater than 4. . . . . . . . . . . . . . . . . . . . . . . . . . .
3.7 Amount of pyruvate produced from various lengths of runs. . . . . . . . . . . .
3.8 Petri Net graph relevant to question 4. “glu” is Glucose, “pyr” is Pyruvate, “atp”
is ATP, “eth” is ethenol or other products of fermentation, and “o2” is Oxygen.
Transitions “gly1” represents glycolysis, “res1” is respiration in presence of
oxygen, “fer1” is fermentation when no oxygen is present, and “cw1” is cellular
work that consumes ATP. Transition “gly1” is inhibited only when the number
of atp tokens is greater than 4. . . . . . . . . . . . . . . . . . . . . . . . . . .
3.9 Firing counts of glycolysis (gly1) and respiration (res1) for different simulation
lengths for the petri net in Figure 3.8 . . . . . . . . . . . . . . . . . . . . . . .
3.10 Petri Net graph relevant to question 5. “is” is the intermembrane space, “mm”
is mitochondrial matrix, “q” is ubiquinone and “cytc” is cytochrome c. The
inhibition arcs (q, t1), (q, t2) and (cytc, t3) capture the electron carrying capacities of q and cytc. Over capacity will cause backup in electron transport chain.
Tokens are colored, e.g. “h:1, nadh:1” specify one token of h and nadh each.
Token types are “h” for H+, “nadh” for NADH, “nadp” for NADP, “fadh2” for
FADH2, “fad” for FAD, “e” for electrons, “o2” for oxygen and “h2o” for water.
We remove t4 to model non-functioning protein complex IV. . . . . . . . . . .
3.11 Simulation of Petri Net in Figure 3.10. In a complete model of the biological
system, there will be a mechanism that keeps the quantity of H+ in check in the
intermembrane space and will plateau at some point. . . . . . . . . . . . . . .
6
3.3
43
43
44
47
48
49
50
52
53
LIST OF FIGURES
3.12 Petri Net graph relevant to question 6. “is” is the intermembrane space, “mm”
is mitochondrial matrix, “q” is ubiquinone and “cytc” is cytochrome c. Tokens
are colored, e.g. “h:1, nadh:1” specify one token of h and nadh each. Token
types are “h” for H+, “nadh” for NADH, “e” for electrons, “o2” for oxygen,
“h2o” for water, “atp” for ATP and “adp” for ADP. We add t6, t7 to model cross
domain diffusion from intermembrane space to mitochondrial matrix. One or
both of t6, t7 may be enabled at a time to control the degree of permeability.
The text above “t5” is an additional condition which must be satisfied for “t5”
to be enabled. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.13 Quantities of H+ and ATP at various run lengths and permeabilities for the Petri
Net model in Figure 3.12. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.14 Petri Net graph relevant to question 7. “is” is the intermembrane space, “mm”
is mitochondrial matrix, “q” is ubiquinone and “cytc” is cytochrome c. Tokens
are colored, e.g. “h:1, nadh:1” specify one token of h and nadh each. Token
types are “h” for H+, “nadh” for NADH, “e” for electrons. Numbers in square
brackets below the transition represent transition durations with default of one
time unit, if the number is missing. . . . . . . . . . . . . . . . . . . . . . . . .
3.15 Quantities of H+ produced in the intermembrane space at various run lengths
and fluidities for the Petri Net model in Figure 3.14. . . . . . . . . . . . . . . .
3.16 Petri Net graph relevant to question 8. “pfk” is phosphofructokinase, “f6p”
is fructose 6-phosphate, “atp” is ATP and “pfkatp” is the pfk bound with atp
for allosteric regulation. Transition “t3” represents enzymic action of pfk, “t4”
represents the binding of pfk with atp. The double arrowed arc represents a
threshold arc, which enables “t4” when there are at least 4 tokens available at
“atp”. Numbers above transitions in angular brackets represent arc priorities. .
3.17 Petri Net model in Figure 3.16. . . . . . . . . . . . . . . . . . . . . . . . . . .
3.18 Petri Net graph relevant to question 9. “glu” represents glucose, “gly1” represents glycolysis, “pyr” represents pyruvate, “ox1” represents oxidative phosphorylation, “nadh” represents NADH and “nadp” represents NAD+. “ox1” is
removed to model stoppage of oxidation of NADH to NAD+. . . . . . . . . . .
3.19 Amount of “pyr” produced by runs of various lengths of Petri Net in Figure 3.18.
It shows results for both normal situation where “nadh” is recycled to “nadp” as
well as the abnormal situation where this recycling is stopped. . . . . . . . . .
3.20 Petri Net graph relevant to question 10. “fats” are fats, “dig” is digestion of
fats, “gly” is glycerol, “fac” is fatty acid, “g3p” is Glyceraldehyde 3-phosphate,
“pyr” is pyruvate, “o2” is oxygen, “nadh” is NADH, “acoa” is Acyl CoA, “atp”
is ATP, “op1” is oxidative phosphorylation, “cac1” is citric acid cycle, “fer1” is
fermentation, “ox1” is oxidation of pyruvate to Acyl CoA and “box1” is beta
oxidation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.21 Amount of “atp” produced by runs of various lengths of Petri Net in Figure 3.20.
Two situations are shown: when oxygen is in short supply and when it is abundant.
7
55
56
57
58
59
61
62
63
64
66
LIST OF FIGURES
8
4.1
4.2
4.3
4.4
4.5
4.6
4.7
4.8
Example of a guarded-arc Petri Net. . . . . . . . . . . .
Example of a guarded-arc Petri Net with colored tokens.
Petri Net for question 11 . . . . . . . . . . . . . . . . .
Petri Net for question 12 . . . . . . . . . . . . . . . . .
Petri Net for question 13 . . . . . . . . . . . . . . . . .
Petri Net with colored tokens alternate for question 14 .
BioPathQA Implementation System Architecture . . . .
BioPathQA Graphical User Interface . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
75
80
122
124
126
128
133
135
5.1
This figure from Tari et al. (2010b) outlines the effects of drug A on drug B
through (a) direct induction/inhibition of enzymes; (b) indirect induction/inhibition of transportation factors that regulate the drug-metabolizing enzymes. . . 158
Chapter 1
Introduction
Biological organisms are composed of cells that contain numerous interconnected and interacting biochemical processes occurring simultaneously. Disruptions in the normal functionality
of these processes causes diseases, which appear as symptoms (of these diseases). As a result
understanding these processes is a fundamental activity in the biological domain and is prerequisite for activities such as disease diagnosis and drug discovery. One aspect of understanding the
biological systems is the identification of pathways responsible for drug metabolism, diseases,
and signal transduction, etc. The availability of high throughput approaches like micro-arrays,
improvements in algorithms, and hardware that have come online during the last decade has
resulted in significant refinement in these pathways. The pathways have become much larger
in size and complexity to the degree that it is not reasonable for one person to fully retain all
aspects of the pathway. As a result, computer based models of pathways are needed that allow
the biologists to ask questions against them and compare them with real-world knowledge. The
model should be such that it has an understanding of the pathway. Such a system would be
considered intelligent and would assist the biologists in expanding the breadth of their search
for new drugs and diagnoses. Source knowledge for these pathways comes from volumes of
research papers published every year. Though there are a number of curated pathway resources
available, they significantly lag behind the current state of the research in biology. As a result,
we need a way to extract this pathway information from published text.
1.1
Choosing the right questions
A large body of research exists on computer modeling of biological processes and it continues
to be an active area of research. However, many such models focus on surface properties,
like structure; or factoid questions. Though important, we feel these systems do not test the
understanding of the underlying system being modeled. We want to go beyond this surface
level information and answer questions requiring deeper reasoning. We want our system to
answer questions that a biology teacher expects his / her students to answer after reading the
required text. So, we turned to college level biological text books for the questions that we
9
CHAPTER 1. INTRODUCTION
10
feel are more indicative of such understanding. Following questions from Reece et al. (2010)
illustrate the kind of questions we are interested in answering:
• “What would happen to the rate of glycolysis if DHAP were removed from the process of
glycolysis as quickly as it was produced?”
• “A muscle cell had used up its supply of oxygen and ATP. Explain what affect would this
have on the rate of cellular respiration and glycolysis?”
These questions and others like it were the subject of a recent deep knowledge representation
challenge1 . In this thesis, we focus on questions that require reasoning over simulations.
1.2
Choosing the right tools
Data about biological systems can be qualitative or quantitative in nature. The fully quantitative
data about reaction dynamics is based on ordinary differential equations (ODEs) of reaction kinetics, which are often lacking Chaouiya (2007). Qualitative data is more prevalent. It is less
precise, but tends to capture general relationships between various components of a biological
pathway. Adding quantitative information to a qualitative model provides the next step in refinement of the biological pathways Heiner et al. (2004), providing better coverage of biological
systems and processes. We want to use this qualitative+quantitative data for our modeling.
To simulate and reason with the pathways, we need tools that can model a biological pathway that contains qualitative+quantitative information, simulate the pathway and reason with
the results.
1.3
Need for a pathway a specification and a query language
Pathway information comes in various formats, such as cartoon drawings, formal graphical representations like Kohn’s Maps Kohn et al. (2006), curated databases of pathways Kanehisa and
Goto (2000); Karp et al. (2002); Croft et al. (2011) and free text. The depth of this knowledge as
well as its taxonomy varies with the source. Thus, a common specification language is needed.
Such a language must be easy to understand and must have a well defined semantics.
Queries are normally specified in natural language, which is vague. So, a more precise query
language is needed. One could ask queries in one of the existing formal languages Gelfond and
Lifschitz (1998), but that will be burdensome for a user to become fluent. As a result, we need a
language that has a simple English-like syntax, but a well defined semantics, so that it does not
have the vagaries of the natural language.
1
https://sites.google.com/site/2nddeepkrchallenge/
CHAPTER 1. INTRODUCTION
1.4
11
Text extraction
Knowledge about biological pathways is spread over collections of published papers as nuggets
of information, such as relationships between proteins; between proteins and drugs; genetic
variation; and association of population groups with genetic variation; to name a few. Published
research may also contain contradictory information, e.g. an earlier conjecture that was proven
to be untrue in later research, or knowledge with limited amount of certainty. To extract these
nuggets and to assemble them into a coherent pathway requires background knowledge, similar
to other technical fields. Portions of this knowledge are published in books and online repositories. Thus, we need a method of text extraction that allows one to extract nuggets of information,
consult available databases and produce knowledge about pathway that is self-consistent.
1.5
Overview
In this thesis, we propose to build a system, called BioPathQA, to answer deeper reasoning
questions using existing tools with modifications. To that end, we develop a language to specify
pathways and queries. Our system is designed to answer reasoning questions requiring simulation. We demonstrate the applicability of our system with applications to drug development on
knowledge obtained from text extraction.
To implement an answering system that can answer simulation based reasoning questions,
we first looked for available tools that could help in this task and we found Petri Nets as providing the right level of formalism for our application. Petri Nets Peterson (1977) are a popular
representation formalism used to model biological systems and to simulate them. They have
been used to model and analyze the dynamic behavior as well as structural properties of biological systems. However, such analysis is usually limited to invariant determination, liveness,
boundedness and reachability. To our knowledge they have not been used to answer questions
similar to the aforementioned.
In order to represent deeper reasoning questions, we have to make extensions to the Petri Net
model as the basic model lacks sufficient richness. For example, we may want to change the firing semantics to limit the state space or maximize parallel activity. Although numerous Petri Net
modeling, simulation and analysis systems exist Jensen et al. (2007); Rohr et al. (2010); Kounev
et al. (2006); Berthomieu* et al. (2004); Nagasaki et al. (2010); Kummer et al. (2000), we found
certain limitations in the default implementation of these systems as well that prevented us from
using them as is. For example, the Colored Petri Net implementation CPNtools 2 does not allow inhibitor arcs (we use to model protein inhibition); Cell Illustrator Nagasaki et al. (2010) is
closed source and does not support colored tokens (we use to model locations); Snoopy Rohr
et al. (2010) supports a large number of extensions, but it is unclear how one exports the simulation results for further reasoning; and most did not allow exploring all possible state evolutions
of a pathway or using different firing semantics.
2
http://cpntools.org
CHAPTER 1. INTRODUCTION
12
To make these extensions in an easy manner we use Answer Set Programming (ASP) Lifschitz (2008) as the language to represent and simulate Petri Nets. It allows a simple encoding
of the Petri Net and can be easily extended to incorporate extensions 3 . In addition, ASP allows powerful reasoning capability and the possibility of implementing additional constructs
not supported by Petri Nets directly, such as the ability to filter trajectories.
Petri Net to ASP translation has been studied before Behrens and Dix (2007); Heljanko and
Niemelä (2000). However, these implementations have been limited to specific classes of Petri
Nets and have different focus. For example, the Simple Logic Petri Nets Behrens and Dix (2007)
do not allow numerical accumulation of the same tokens from multiple transitions to a single
place and the Binary Petri Nets Heljanko and Niemelä (2000) do not allow more than one tokens
at any place.
1.6
Specific contributions
The research contribution of this thesis can be divided into four major parts. The first part gives
a general encoding of Petri Nets in ASP, which allows easy extension by making local changes.
The second part shows how the ASP encoding of Petri Nets can be used to answer simulation
based reasoning questions. The third part describes the high-level language for pathway and
query specification; and the system that we have developed to answer deep reasoning questions.
The fourth part shows how knowledge is extracted from text of research papers, cleaned and
assembled into a pathway to answer simulation based reasoning questions using our system.
1.6.1
General ASP encoding of Petri Net for simulation
Although previous work on encoding Petri Nets in ASP exists, it is limited to specific classes
of Petri Nets. We present an encoding of a basic Petri Net in ASP to show it is done in an
intuitive yet concise way. The default execution semantics of a Petri Net is the set-semantics,
which allows a subset of transitions to fire simultaneously when ready. This can result in far too
many combinations of transition firing arrangements. A simpler approach is to use the so called
interleaved execution semantics, in which at most one transition fires when ready. This too can
generate many firing arrangements. Biological systems are highly parallel in nature, as a result
it is beneficial to model maximum parallel activity. So, we introduce a new firing semantics,
called the maximal firing set semantics by extending the set semantics. In this semantics, a
maximal subset of non-conflicting enabled transitions fire simultaneously when ready.
Then, we extend the basic ASP encoding to include Petri Net extensions like reset-arcs (to
model immediate consumption of any amount of substrate), inhibit-arcs (to model gene/protein
inhibition), read-arcs (to model additional pre-conditions of a reaction, such as different start
vs. maintenance quantity of a reactant), colored-tokens (to model quantities of different types
3
Certain commercial tools, like Cell Illustrator (http://www.cellillustrator.com) do allow exporting their model
into a high level language, but we believe that a declarative language is more suited to succinctly describe the
problem.
CHAPTER 1. INTRODUCTION
13
of substances at the same location), priority-transitions (to select between alternate metabolic
paths), and timed-transitions (to model slow reactions) that allow modeling of various concepts
in biological systems. We show how ASP allows us to make these extensions with small amount
of local changes.
This component is one of the major focuses of our research. It is described in Chapter 2 and
is the basis for implementation of our system to model pathways and answer questions about
them.
1.6.2
Answering simulation based reasoning questions
We use the encoding developed in Chapter 2 to questions in (Reece et al., 2010, Chapter 9)
that were a part of the Second Deep Knowledge Representation Challenge 4 . These questions
are focused on the mechanism of cellular respiration and test the understanding of the student
studying the material; and appear in two main forms: (i) inquiry about change of rate of a
process due to a change in the modeled system, and (ii) explanation of a change due to a change
in the modeled system.
We built Petri Net models for the situations specified in the questions, encoded them in ASP
and simulated them over a period of time. For change of rate questions, we computed the rate for
both nominal and modified cases and observed that they matched the responses provided with
the challenge questions. For the explanation of change questions, we collected the summary of
firing transitions as well as substance quantities produced at various times. This information
formed the basis of our answer. We compared our results with the answers provided with the
challenge questions.
A novel aspect of our approach is that we apply the initial conditions and interventions
mentioned in the questions as modifications to the pathway representation. These interventions
can be considered as a generalized form of actions.
For certain questions, additional domain knowledge outside the source material was required. We filled this gap in knowledge as necessary. We also kept the models to a subset of the
pathway for performance as well as to reduce clutter in the output that can bury the results with
unnecessary details.
This component of our research is described in Chapter 3.
1.6.3
BioPathQA: a system and a language to represent pathways and query
them
We combined the techniques learned from Chapter 2, action languages, and biological modeling
languages to build a question answering system that takes a pathway and a query as input. Both
the pathway specification language and the query language have strict formal semantics, which
allow them to be free of the vagaries of natural language, the language of the research papers as
well as the query statements.
4
https://sites.google.com/site/2nddeepkrchallenge/
CHAPTER 1. INTRODUCTION
14
Guarded-Arc Petri Net
Since the biological pathways are constructed of biochemical reactions, they are effected by
environmental changes. Mutations within the cell can also result in conditional change in behavior of certain processes. As a result, we needed actions with conditional effects. Our Petri
Net model wasn’t rich enough to model conditional actions, so we extended the Petri Nets with
conditional arcs. We call this extension, the Guarded-Arc Petri Net, where a guard is a condition on an arc, which must be true for that arc to be traversed. With this extension, a Petri Net
transition can have different outcomes for different markings. Our model is similar to the model
in Jensen et al. (2007) in many aspects.
This component of our research is described in Chapter 4.
1.6.4
Text Extraction to Answer Questions about Real World Applications
To apply our system to real world applications, we have to extract pathway knowledge from
published papers, which are published in natural language text. For this, we use a system called
the Parse Tree Query Language (PTQL) Tari et al. (2012) to nuggets of information from the
abstracts published on PubMed 5 . Sentences are parsed using the Link Grammar Sleator and
Temperley (1993) or Stanford Parser de Marneffe and Manning (2008); with various objectclasses identified within the sentence. Unlike Information Retrieval (IR) approaches that tend
to treat documents as unstructured bags-of-words, PTQL treats words (or word-groups) as sentence elements with syntactic as well as dependency relationships between them. PTQL queries
combine lexical, syntactic and semantic features of sentence elements. Thus with PTQL, one
can ask for the first-noun of a noun-phrase that is the direct-object of a verb-phrase for some
specific verb string. To accomplish its task, PTQL performs a number of pre-processing steps
on its input useful for text extraction and leverages on various existing databases. These include
sentence splitting, tokenization, part-of-speech (POS) tagging, named entity recognition, entitymention normalization, cross-linking with concepts from external databases, such as Gene Ontology Camon et al. (2004) and UniProt Consortium (2012). We extract gene-gene, gene-drug,
and gene-disease relationships using PTQL, assemble them into a pathway specification and
reason with the extracted knowledge to determine possible drug interactions.
Facts and relationships extracted using PTQL are further subject to filtering to remove inconsistent information. A pathway specification is then constructed from the extracted facts,
which can be queried using the query specification language. We illustrate the use of our deep
reasoning system by an example from the drug-drug interaction domain.
This component is described in Chapter 5.
1.7
Summary
The main contributions of this thesis can be summarized as follows:
5
http://www.ncbi.nlm.nih.gov/pubmed
CHAPTER 1. INTRODUCTION
15
1. Generalized Petri Net encoding in ASP, including a new maximal firing set semantics
(Chapter 2)
• An easy to extend encoding is developed, that allows adding extensions using local
changes
• A new Petri Net firing semantics, the so called maximal firing set semantics is defined, which ensures maximum possible parallel activity at any given point
2. Answering simulation based deep reasoning questions using our ASP encoding (Chapter 3)
• It is shown, how deep reasoning questions requiring simulation based reasoning can
be answered.
3. Developed a system called BioPathQA and a language to specify biological pathways and
answer deep reasoning questions about it (Chapter 4)
• A pathway specification language is developed, combining concepts from Petri Nets,
Action Languages, and Biological Pathways
• A query specification language is developed, which is english like, with well defined
semantics, avoiding the vagaries of Natural Language
• A description of our implementation using ASP and Python is given; and an execution trace is shown
4. Performed text extraction to extract pathway knowledge (Chapter 5)
• It is shown pathway knowledge is extracted and used to answer questions in the
drug-drug interaction domain
Chapter 2
Petri Net Encoding in ASP for
Biological Domain
2.1
Introduction
Petri Net Peterson (1977) is a graphical modeling language with formal semantics used for
description of distributed systems. It is named after Carl Adam Petri, who formally defined Petri
Nets in his PhD thesis in the 1960’s Brauer and Reisig (2006). Petri nets have been widely used
to model a wide range of systems, from distributed systems to biological pathways. The main
advantages of Petri Net representation include its simplicity and the ability to model concurrent
and asynchronous systems and processes.
A variety of Petri Net extensions have been proposed in the literature, e.g. inhibitor arcs,
reset transitions, timed transitions, stochastic transitions, prioritized transitions, colored petri
nets, logic petri nets, hierarchical petri nets, hybrid petri nets and functional petri nets to a name
a few Behrens and Dix (2007); Music (2012); Hardy and Robillard (2004).
Our interest in Petri Nets is for representing biological pathways and simulating them in order to answer simulation based reasoning questions. We show how Petri nets can be represented
in ASP. We also demonstrate how various extensions of basic Petri nets can be easily expressed
and implemented by making small changes to the initial encoding. During this process we will
relate the extensions to their use in the biological domain. Later chapters will show how this
representation and simulation is used to answer biologically relevant questions.
The rest of this chapter is organized as follows: We present some background material on
Answer Set Programming (ASP) and Petri Nets. Following that, we present the Answer Set
encoding of a basic Petri Net. After that we will introduce various Petri Nets extensions and the
relevant ASP code changes to implement such extensions.
16
CHAPTER 2. PETRI NET ENCODING IN ASP FOR BIOLOGICAL DOMAIN
2.2
2.2.1
17
Background
Answer Set Programming
Answer Set Programming (ASP) is a declarative logic programming language that is based on
the Stable Model Semantics Gelfond and Lifschitz (1988). It has been applied to a problems
ranging from spacecrafts, work flows, natural language processing and biological systems modeling.
Although ASP language is quite general, we limit ourselves to language and extensions
relevant to our work.
Definition 1 (Term) A term is a term in the propositional logic sense.
Definition 2 (Literal) A literal is an atom in the propositional logic sense. A literal prefixed
with not is referred to as a negation-as-failure literal or a naf-literal, with not representing
negation-as-failure.
We will refer to propositional atoms as basic atoms to differentiate them from other atoms,
such as the aggregate atoms defined below.
Definition 3 (Aggregate Atom) A sum aggregate atom is of the form:
L [B0 = w0 , . . . , Bm = wm ] U
(2.1)
where, Bi are basic atoms, wi are positive integer weight terms, L, U are integer terms specifying
the lower and upper limits of aggregate weights. The lower and upper limits are assumed to be
−∞ and ∞, if not specified.
A count aggregate atom is a special case of the sum aggregate atom in which all weights are 1,
i.e. L [B0 = 1, . . . , Bm = 1] U and it is represented by:
L {B0 , . . . , Bm } U
(2.2)
A choice atom is a special case of the count aggregate atom (2.2) in which n = m.
Definition 4 (ASP Program) An ASP program Π is a finite set of rules of the following form:
A0 ← A1 , . . . , Am , not B1 , . . . , not Bn , C1 , . . . , Ck .
(2.3)
where each A0 is either a basic atom or a choice atom, Ai and Bi are basic atoms, Ci are
aggregate atoms and not is negation-as-failure.
In rule (2.3), {A0 } is called the head of the rule, and {A1 , . . . , Am , not B1 , . . . , not Bn , C1 , . . . ,
Ck } is called its tail. A rule in which A0 is a choice atom is called a choice rule. A rule without
CHAPTER 2. PETRI NET ENCODING IN ASP FOR BIOLOGICAL DOMAIN
18
a head is called a constraint. A rule with a basic atom as its head and empty tail is called a fact
in which case the “←” is dropped.
Let R be an ASP rule of the form (2.3) and let pos(R) = {A1 , . . . , Am } represent the positive
atoms, neg(R) = {B1 , . . . , Bn } the negation-as-failure atoms, and agg(R) = {C1 , . . . , Ck } represent
the aggregate atoms in the body of a rule R. Let lit(A) represent the set of basic literals in atom
A, i.e. lit(A) = {A} if A is a basic atom; lit(A) = {B0 , . . . , Bn } if A is an aggregate atom. Let C be
an aggregate atom of the form (2.1) and let pos(C) = {B0 , . . . , Bm } be the sets of basic positive
literals such that lit(C) = pos(C).
S
Let lit(R) = lit(head(R)) ∪ pos(R) ∪ neg(R) ∪ C∈agg(R) lit(C) for a rule R ∈ Π and lit(Π) =
S
R∈Π lit(R) be the set of basic literals in ASP program Π.
Definition 5 (Aggregate Atom Satisfaction) A ground aggregate atom C of the form (2.1) is
P
satisfied by a set of basic ground atoms S , if L ≤ 0≤i≤m,Bi ∈S wi ≤ U and we write S |= C.
Given a set of basic ground literals S and a basic ground atom A, we say S |= A if A ∈ S ,
S |= not A if A < S . For a rule R of the form (2.3) S |= body(R) if ∀A ∈ {A1 , . . . , Am }, S |= A,
∀B ∈ {B1 , . . . , Bn }, S |= not B, and ∀C ∈ {C1 , . . . , Ck }, S |= C; S |= head(R) if S |= A0 .
Definition 6 (Rule Satisfaction) A ground rule R ∈ Π is satisfied by a set of basic ground
atoms S , iff, S |= body(R) implies S |= head(R). A constraint rule R ∈ Π is satisfied by set S if
S 6|= body(R).
We define reduct of an ASP program by treating aggregate atoms in a similar way as
negation-as-failure literals, since our code does not contain recursion through aggregation (which
can yield non-intuitive answer-sets Son and Pontelli (2007)).
Definition 7 (Reduct) Let S be a set of ground basic atoms, the reduct of ground ASP program Π w.r.t. S , written ΠS is the set of rules: {p ← A1 , . . . , Am . | A0 ← A1 , . . . , Am ,
not B1 , . . . , not Bn , C1 , . . . , Ck . ∈ Π, p ∈ lit(A0 ) ∩ S , {A1 , . . . , Am } ⊆ S , {B1 , . . . , Bn } ∩ S = ∅,
@C ∈ {C1 , . . . , Ck }, S 6|= C}.
Intuitively, this definition of reduct removes all rules which contain a naf-literal or an aggregate atom in their bodies that does not hold in S , and it removes aggregate atoms as well as
naf-literals from the body of the remaining rules.
Heads of choice-rules are split into multiple rules containing at most one atom in their heads.
The resulting reduct is a program that does not contain any aggregate atoms or negative literals.
The rules of such a program are monotonic, such that if it satisfied by a set S of atoms, it is also
satisfied by any superset of S .
A deductive closure of such a (positive) monotonic program is defined as the unique smallest set of atoms S such that whenever all body atoms of a rule hold in S , the head also holds in
S . The deductive closure can be iteratively computed by starting with an empty set and adding
heads of rules for which the bodies are satisfied, until a fix point is reached, where no additional
rules can be satisfied. (adopted from Baral (2003))
CHAPTER 2. PETRI NET ENCODING IN ASP FOR BIOLOGICAL DOMAIN
19
Definition 8 (Answer Set) A set of basic ground atoms S is an answer set of a ground ASP
program Π, iff S is equal to the deductive closure of ΠS and S satisfies each rule of Π. (adopted
from Baral (2003))
Clingo Specific Syntactic Elements
The ASP code in this thesis is in the syntax of ASP solver called clingo Gebser et al. (2011).
The “←” in ASP rules is replaced by the symbol “:-”. Though the semantics of ASP are defined
on ground programs, Clingo allows variables and other constructs for compact representation.
We intuitively describe specific syntactic elements and their meanings below:
Comments: Text following “%” to the end of the line is treated as a comment.
Interval: Atoms defined over an contiguous range of integer values can be compactly written as intervals, e.g. p(1 .. 5) represents atoms p(1), p(2), p(3), p(4), p(5).
Pooling: Symbol “;” allows for pooling alternative terms. For example, an atom p(. . . , X, . . . )
and p(. . . , Y, . . . ) can be pooled together into a single atom as p(. . . , X; Y, . . . ).
Aggregate assignment atom: An aggregate assignment atom Q = #sum[A0 = w0 , . . . ,
A = wm , not Am+1 = wm+1 , . . . , not An = wn ] assigns the sum
Pm
P
Ai ∈S ,0≤i≤m wi + A j <S ,m+1≤n w j to Q w.r.t. a consistent set of basic ground atoms S .
Condition: Conditions allow instantiating variables to collections of terms within aggregates, e.g. {p(X) : q(X)} instantiates p(X) for only those X that q(X) satisfies. For example, if
we have p(1..5) but only q(3; 5), then {p(X) : q(X)} is expanded to {p(3), p(5)}.
Grounding
Grounding makes a program variable free by replacing variables with the possible values they
can take. Clingo uses the grounder Gringo ? for “smart” grounding, which results in substantial
reduction in the size of the program. Details of this grounding are implementation specific. We
present the intuitive process of grounding below.
1. A set of ground terms is constructed, where a ground term is a term that contains no
variables.
2. The variables are split into two categories: local and global. Local variables are the ones
that appear only within an aggregate atom (minus the limits) and nowhere else in a rule.
Such variables are considered local writ. the aggregate atom. All other variables are
considered global.
3. First the global variables are eliminated in the rules as follows:
• Each rule r containing an aggregate assignment atom of the form (2.1) is replaced
with set of rules r0 in which the aggregate assignment atom is is replaced with an
aggregate atom with lower and upper bounds of Q for all possible substitutions of
Q. This is generalized to multiple aggregate assignment atoms by repeating this step
CHAPTER 2. PETRI NET ENCODING IN ASP FOR BIOLOGICAL DOMAIN
20
for each such atom, where output of previous iteration forms the input of the next
iteration.
• Each rule r0 , is replaced with the set of all rules r00 obtained by all possible substitutions of ground terms for global variables in r.
4. Then the local variables are eliminated in the rules by expanding conditions, such that
p(. . . , X, . . . ) : d(X) are replaced by p(. . . , d1 , . . . ), . . . , p(. . . , dk , . . . ) for the extent {d1 , . . . , dk }
of d(X). This is generalized to multiple conditions in the obvious way.
Following the convention of the Clingo system, Variables in rules presented in this thesis
start with capital letters while lower-case text and numbers are constants. Italicized text represents a constant term from a definition in context.
A recent work Harrison et al. (2013) gives the semantics of Gringo with ASP Core 2 syntax Calimeri et al. (2013) using Infintary Propositional Formulas, which translate Gringo to
propositional formulas with infinitely long conjunctions and disjunctions. Their approach removes the safety requirement, but the subset of Gringo presented does appear to cover assignments. Although their approach provides a way to improve our ASP encoding by removing the
requirement of specifying the maximum number of tokens or running simulations until a condition holds, our simpler (limited) semantics is sufficient for the limited syntax and semantics we
use.
2.2.2
Multiset
A multiset A over a domain set D is a pair hD, mi, where m : D → N is a function giving
the multiplicity of d ∈ D in A. Given two multsets A = hD, mA i, B = hD, mB i, A B if
∀d ∈ D : mA (d) mB (d), where ∈ {<, >, ≤, ≥, =}, and A , B if ∃d ∈ D : mA (d) , mB (d).
Multiset sum/difference is defined in the usual way. We use the short-hands d ∈ A to represent
mA (d) > 0, A = ∅ to represent ∀d ∈ D, m(d) = 0, A ⊗ n to represent ∀d ∈ D, m(d) ⊗ n, where
n ∈ N, ⊗ ∈ {<, >, ≤, ≥, =, ,}. We use the notation d/n ∈ A to represent that d appears n-times in
A; we drop A when clear from context. The reader is referred to Syropoulos (2001) for details.
2.2.3
Petri Net
A Petri Net is a graph of a finite set of nodes and directed arcs, where nodes are split between
places and transitions, and each arc either connects a place to a transition or a transition to a
place. Each place has a number of tokens (called the its marking) 1 . Collective marking of all
places in a Petri Net is called its marking (or state). Arc labels represent arc weights. When
missing, arc-weight is assumed as one, and place marking is assumed as zero.
The set of place nodes on incoming and outgoing arcs of a transition are called its preset (input place set or input-set) and post-set (output place set or output-set), respectively. A
1
Standard convention is to use dots in place nodes to represent the marking of the place. We use numbers for
compact representation.
CHAPTER 2. PETRI NET ENCODING IN ASP FOR BIOLOGICAL DOMAIN
21
dhap
t5b
0
t3
f16bp
t5a
t4
g3p
t6
2
bpg13
Figure 2.1: Petri Net graph (of sub-section of glycolysis pathway) showing places as circles,
transitions as boxes and arcs as directed arrows. Places have token count (or marking) written
above them, assumed 0 when missing. Arcs labels represent arc-weights, assumed 1 when
missing.
transition t is enabled when each of its pre-set place p has at least the number of tokens equal to
the arc-weight from p to t. An enabled transition may fire, consuming tokens equal to arc-weight
from place p to transition t from each pre-set place p, producing tokens equal to arc-weight from
transition t to place p to each post-set place p.
Multiple transitions may fire as long as they consume no more than the available tokens, with
the assumption that tokens cannot be shared. Fig. 2.1 shows a representation of a portion of the
glycolysis pathway as given in Reece et al. (2010). In this figure, places represent reactants and
products, transitions represent reactions, and arc weights represent reactant quantity consumed
or the product quantity produced by the reaction. When unspecified, arc-weight is assumed to
be 1 and place-marking is assumed to be 0.
Definition 9 (Petri Net) A Petri Net is a tuple PN = (P, T, E, W), where, P = {p1 , . . . , pn } is a
finite set of places; T = {t1 , . . . , tm } is a finite set of transitions, P ∩ T = ∅; E + ⊆ T × P is a
set of arcs from transitions to places; E − ⊆ P × T is a set of arcs from places to transitions;
E = E + ∪ E − ; and W : E → N \ {0} is the arc-weight function
Definition 10 (Marking) A marking M = (M(p1 ), . . . , M(pn )) is the token assignment of each
place node pi ∈ P of PN, where M(pi ) ∈ N. Initial token assignment M0 : P → N is called the
initial marking. Marking at step k is written as Mk .
Definition 11 (Pre-set & post-set of a transition) Pre-set / input-set of a transition t ∈ T of
PN is •t = {p ∈ P : (p, t) ∈ E − }, while the post-set / output-set is t• = {p ∈ P : (t, p) ∈ E + }
Definition 12 (Enabled Transition) A transition t ∈ T of PN is enabled with respect to marking M, enabled M (t), if ∀p ∈ •t, W(p, t) ≤ M(p). An enabled transition may fire.
Definition 13 (Transition Execution) A transition execution is the simulation of change of
marking from Mk to Mk+1 due to firing of a transition t ∈ T of PN. Mk+1 is computed as
follows:
∀pi ∈ •t, Mk+1 (pi ) = Mk (pi ) − W(pi , t)
CHAPTER 2. PETRI NET ENCODING IN ASP FOR BIOLOGICAL DOMAIN
22
∀p j ∈ t•, Mk+1 (p j ) = Mk (p j ) + W(t, p j )
Petri Nets allow simultaneous firing of a set of enabled transitions w.r.t. a marking as long
as they do not conflict.
Definition 14 (Conflicting Transitions) Given PN with marking M. A set of enabled transitions T e = {t ∈ T : enabled M (t)} of PN conflict if their simultaneous firing will consume more
tokens than are available at an input place:
X
∃p ∈ P : M(p) <
W(p, t)
t∈T e ∧p∈•t
Definition 15 (Firing Set) A firing set is a set T k = {t1 , . . . , tm } ⊆ T of simultaneously firing
transitions that are enabled and do not conflict w.r.t. to the current marking Mk of PN.
Definition 16 (Firing Set Execution) Execution of a firing set T k of PN on a marking Mk computes the new marking Mk+1 as follows:
X
X
∀p ∈ P, Mk+1 (p) = Mk (p) −
W(p, t) +
W(t, p)
t∈T k ∧p∈•t
t∈T k ∧p∈t•
P
P
where t∈Tk ∧p∈•t W(p, t) is the total consumption from place p and t∈Tk ∧p∈t• W(t, p) is the total
production at place p.
Definition 17 (Execution Sequence) An execution sequence X = M0 , T 0 , M1 , T 1 , . . . , Mk , T k , Mk+1
of PN is the simulation of a firing set sequence σ = T 1 , T 2 , . . . , T k w.r.t. an initial marking M0 ,
producing the final marking Mk+1 . Mk+1 is the transitive closure of firing set executions, where
subsequent marking become the initial marking for the next firing set.
For an execution sequence X = M0 , T 0 , M1 , T 1 , . . . , Mk , T k , Mk+1 , the firing of T 0 with respect
to marking M0 produces the marking M1 which becomes the initial marking for T 1 , which
produces M2 and so on.
2.3
Translating Basic Petri Net Into ASP
In this section we present ASP encoding of simple Petri Nets. We describe, how a given Petri
Net PN, and an initial marking M0 are encoded into ASP for a simulation length k. Following
sections will show how Petri Net extensions can be easily added to it. We represent a Petri Net
with the following facts:
f1: Facts place(pi ). where pi ∈ P is a place.
f2: Facts trans(t j ). where t j ∈ T is a transition.
f3: Facts ptarc(pi , t j , W(pi , t j )). where (pi , t j ) ∈ E − with weight W(pi , t j ).
f4: Facts tparc(ti , p j , W(ti , p j )). where (ti , p j ) ∈ E + with weight W(ti , p j ).
CHAPTER 2. PETRI NET ENCODING IN ASP FOR BIOLOGICAL DOMAIN
23
Petri Net execution simulation proceeds in discrete time-steps, these time steps are encoded
by the following facts:
f5: Facts time(tsi ) where 0 ≤ tsi ≤ k.
The initial marking (or initial state) of the Petri Net is represented by the following facts:
i1: Facts holds(pi , M0 (pi ), 0) for every place pi ∈ P with initial marking M0 (pi ).
ASP requires all variables in rule bodies be domain restricted. So, we add the following
facts to capture the possible token quantities produced during the simulation 2 :
f6: Facts num(n)., where 0 ≤ n ≤ ntok
A transition ti is enabled if each of its input places p j ∈ •ti has at least arc-weight W(p j , ti )
tokens. Conversely, ti is not enabled if ∃p j ∈ •ti : M(p j ) < W(p j , ti ), and is only enabled when
no such place p j exists. These are captured in e1 and e2 respectively:
e1: notenabled(T,TS):-ptarc(P,T,N),holds(P,Q,TS),Q<N, place(P),
trans(T), time(TS),num(N),num(Q).
e2: enabled(T,TS) :- trans(T), time(TS), not notenabled(T, TS).
Rule e1 encodes notenabled(T,TS) which captures the existence of an input place P of transition T that violates the minimum token requirement N at time-step T S . Where, the predicate
holds(P,Q,TS) encodes the marking Q of place P at T S .
Rule e2 encodes enabled(T,TS) which captures that transition T is enabled at T S since
there is no input place P of transition T that violates the minimum input token requirement at
T S . In biological context, e2 captures the conditions when a reaction (represented by T ) is
ready to proceed. A subset of enabled transitions may fire simultaneously at a given time-step.
This is encoded as:
a1: {fires(T,TS)} :- enabled(T,TS), trans(T), time(TS).
Rule a1 encodes fires(T,TS), which captures the firing of transition T at T S . The rule
is encoded with a count atom as its head, which makes it a choice rule. This rule either picks
the enabled transition T for firing at T S or not, effectively enumerating a subset of enabled
transitions to fire. Whether this set can fire or not in an answer set is subject to conflict checking,
which is done by rules a2, a3, a4 shown later.
In biological context, the selected transition-set models simultaneously occurring reactions
and the conflict models limited reactant supply that cannot be shared. Such a conflict can lead
to multiple choices in parallel reaction evolutions and different outcomes.
The next set of rules captures the consumption and production of tokens due to the firing
of individual transitions in a firing-set as well as their aggregate effect, which computes the
marking for the next time step:
2
Note that ntok can be arbitrarily chosen to be larger than the maximum expected token quantity produced during
the simulation.
CHAPTER 2. PETRI NET ENCODING IN ASP FOR BIOLOGICAL DOMAIN
24
r1: add(P,Q,T,TS) :- fires(T,TS), tparc(T,P,Q), time(TS).
r2: del(P,Q,T,TS) :- fires(T,TS), ptarc(P,T,Q), time(TS).
r3: tot incr(P,QQ,TS) :- QQ=#sum[add(P,Q,T,TS)=Q:num(Q):trans(T)], time(TS),
num(QQ), place(P).
r4: tot decr(P,QQ,TS) :- QQ=#sum[del(P,Q,T,TS)=Q:num(Q):trans(T)], time(TS),
num(QQ), place(P).
r5: holds(P,Q,TS+1) :-holds(P,Q1,TS),tot incr(P,Q2,TS),time(TS+1),
tot decr(P,Q3,TS),Q=Q1+Q2-Q3,place(P),num(Q;Q1;Q2;Q3),time(TS).
Rule r1 encodes add(P,Q,T,TS) and captures the addition of Q tokens to place P due to
firing of transition T at time-step T S . Rule r2 encodes del(P,Q,T,TS) and captures the deletion
of Q tokens from place P due to firing of transition T at T S .
Rules r3 and r4 aggregate all add’s and del’s for place P due to r1 and r2 at time-step T S ,
respectively, by using the QQ=#sum[] construct to sum the Q values into QQ. Rule r5 which
encodes holds(P,Q,TS+1) uses these aggregate adds and removes and updates P’s marking for
the next time-step T S +1. In biological context, these rules capture the effect of a reaction on reactant and product quantities available in the next simulation step. To prevent overconsumption
at a place following rules are added:
a2: consumesmore(P,TS) :- holds(P,Q,TS), tot decr(P,Q1,TS), Q1 > Q.
a3: consumesmore :- consumesmore(P,TS).
a4: :- consumesmore.
Rule a2 encodes consumesmore(P,TS) which captures overconsumption of tokens at input
place P at time T S due to the firing set selected by a1. Overconsumption (and hence conflict)
occurs when tokens Q1 consumed by the firing set are greater than the tokens Q available at P.
Rule a3 generalizes this notion of overconsumption and constraint a4 eliminates answers where
overconsumption is possible.
Definition 18 Given a Petri Net PN, its initial marking M0 and its encoding Π(PN, M0 , k,
ntok) for k-steps and maximum ntok tokens at any place. We say that there is a 1-1 correspondence between the answer sets of Π(PN, M0 , k, ntok) and the execution sequences of PN
iff for each answer set A of Π(PN, M0 , k, ntok), there is a corresponding execution sequence
X = M0 , T 0 , M1 , . . . , Mk , T k , Mk+1 of PN and for each execution sequence X of PN there is an
answer-set A of Π(PN, M0 , k, ntok) such that
{ f ires(t, ts) : t ∈ T ts , 0 ≤ ts ≤ k} = { f ires(t, ts) : f ires(t, ts) ∈ A}
{holds(p, q, ts) : p ∈ P, q = Mts (p), 0 ≤ ts ≤ k + 1}
= {holds(p, q, ts) : holds(p, q, ts) ∈ A}
Proposition 1 There is a 1-1 correspondence between the answer sets of Π0 (PN, M0 , k, ntok)
and the execution sequences of PN.
CHAPTER 2. PETRI NET ENCODING IN ASP FOR BIOLOGICAL DOMAIN
2.3.1
25
An example execution
Next we look at an example execution of the Petri Net shown in Figure 2.1. The Petri Net and
its initial marking are encoded as follows3 :
num(0..60).time(0..5).place(f16bp;dhap;g3p;bpg13).
trans(t3;t4;t5a;t5b;t6).tparc(t3,f16bp,1).ptarc(f16bp,t4,1).
tparc(t4,dhap,1).tparc(t4,g3p,1).ptarc(dhap,t5a,1).
tparc(t5a,g3p,1).ptarc(g3p,t5b,1).tparc(t5b,dhap,1).
ptarc(g3p,t6,1).tparc(t6,bpg13,2).holds(f16bp;dhap;g3p;bgp13,0,0).
we get thousands of answer-sets, for example4 :
holds(bpg13,0,0) holds(dhap,0,0) holds(f16bp,0,0) holds(g3p,0,0)
holds(bpg13,0,1) holds(dhap,0,1) holds(f16bp,1,1) holds(g3p,0,1)
holds(bpg13,0,2) holds(dhap,1,2) holds(f16bp,1,2) holds(g3p,1,2)
holds(bpg13,0,3) holds(dhap,2,3) holds(f16bp,1,3) holds(g3p,2,3)
holds(bpg13,2,4) holds(dhap,3,4) holds(f16bp,1,4) holds(g3p,2,4)
holds(bpg13,4,5) holds(dhap,4,5) holds(f16bp,1,5) holds(g3p,2,5)
fires(t3,0) fires(t3;t4,1) fires(t3;t4;t5a;t5b,2)
fires(t3;t4;t5a;t5b;t6,3) fires(t3;t4;t5a;t5b;t6,4)
fires(t3;t4;t5a;t5b;t6,5)
2.4
Changing Firing Semantics
The ASP code above implements the set firing semantics. It can produce a large number of
answer-sets, since any subset of a firing set will also be fired as a firing set. For our biological
system modeling, it is often beneficial to simulate only the maximum activity at any given timestep. We accomplish this by defining the maximal firing set semantics, which requires that
a maximal subset of non-conflicting transitions fires at a single time step5 . Our semantics is
different from the firing multiplier approach used by Krepska et al. (2008), in which a transition
can fire as many times as allowed by the tokens available in its source places. Their approach
requires an exponential time firing algorithm in the number of transitions. Our maximal firing
set semantics is implemented by adding the following rules to the encoding in Section 2.3:
a5:
could not have(T,TS) :- enabled(T,TS), not fires(T,TS), ptarc(S,T,Q),
holds(S,QQ,TS), tot decr(S,QQQ,TS), Q > QQ - QQQ.
a6: :- not could not have(T,TS), enabled(T,TS), not fires(T,TS), trans(T),
time(TS).
3
{holds(p1,0,0),...,holds(pN,0,0)}, {num(0),...,num(60)}, {time(0),...,time(5)} have been
written as holds(p1;...;pN,0,0), num(0..60), time(0..5), respectively, to save space.
4
{fires(t1,ts1),...,fires(tN,ts1)} have been written as fires(t1;...;tN;ts1) to save space.
5
Such a semantics reduces the reachable markings. See Burkhard (1980) for the analysis of its computational
power.
CHAPTER 2. PETRI NET ENCODING IN ASP FOR BIOLOGICAL DOMAIN
26
Rule a5 encodes could not have(T,TS) which means that an enabled transition T that did
not fire at time T S , could not have fired because its firing would have resulted in overconsumption. Rule a6 eliminates any answer-sets in which an enabled transition did not fire, that could
not have caused overconsumption. Intuitively, these two rules guarantee that the only reason for
an enabled transition to not fire is conflict avoidance (due to overconsumption). With this firing
semantics, the number of answer-sets produced for Petri Net in Figure 2.1 reduces to 2.
Proposition 2 There is 1-1 correspondence between the answer sets of Π1 (PN, M0 , k, ntok)
and the execution sequences of PN.
Other firing semantics can be encoded with similar ease. For example, if interleaved firing
semantics is desired, replace rules a5, a6 with the following:
a5’: more than one fires :- fires(T1,TS), fires(T2, TS), T1 != T2, time(TS).
a6’: :- more than one fires.
We now look at Petri Net extensions and show how they can be easily encoded in ASP.
2.5
Extension - Reset Arcs
Definition 19 (Reset Arc) A Reset Arc in a Petri Net PN R is an arc from place p to transition
t that consumes all tokens from its input place p upon firing of t. A Reset Petri Net is a tuple
PN R = (P, T, E, W, R) where, P, T, E, W are the same as for PN; and R : T → 2P defines reset
arcs
0
t3
f16bp
dhap
tr
t5a
t5b
g3p
t6
t4
2
bpg13
Figure 2.2: Petri Net of Fig 2.1 extended with a reset arc from dhap to tr shown with double
arrowhead.
Figure 2.2 shows an extended version of the Petri Net in Figure 2.1 with a reset arc from
dhap to tr (shown with double arrowhead). In biological context it models the removal of
all quantity of compound dhap. Petri Net execution semantics with reset arcs is modified for
conflict detection and execution as follows:
CHAPTER 2. PETRI NET ENCODING IN ASP FOR BIOLOGICAL DOMAIN
27
Definition 20 (Reset Transition) A transition t ∈ T of PN R is called a reset-transition if it has
a reset arc incident on it, i.e. R(t) , ∅.
Definition 21 (Firing Set) A firing set is a set T k = {t1 , . . . , tm } ⊆ T of simultaneously firing
transitions that are enabled and do not conflict w.r.t. to the current marking Mk of PN R . T k is not
a firing set if there is an enabled reset-transition that is not in T k , i.e. ∃t : enabled Mk (t), R(t) ,
∅, t < T k . 6 .
Definition 22 (Transition Execution in PN R ) A transition execution is the simulation of change
of marking from Mk to Mk+1 due to firing of a transition t ∈ T of PN R . Mk+1 is computed as
follows:
∀pi ∈ •t, Mk+1 (pi ) = Mk (pi ) − W(pi , t)
∀p j ∈ t•, Mk+1 (p j ) = Mk (p j ) + W(t, p j )
∀pr ∈ R(t), Mk+1 (pr ) = Mk (pr ) − Mk (pr )
Definition 23 (Conflicting Transitions in PN R ) A set of enabled transitions conflict in PN R
w.r.t. Mk if firing them simultaneously will consume more tokens than are available at any one
of their common input-places. T e = {t ∈ T : enabled Mk (t)} conflict if:
X
X
∃p ∈ P : Mk (p) < (
W(p, t) +
Mk (p))
t∈T e ∧(p,t)∈E −
t∈T e ∧p∈R(t)
Definition 24 (Firing Set Execution in PN R ) Execution of a transition set T i in PN R has the
following effect:
X
X
∀p ∈ P \ R(T i ), Mk+1 (p) = Mk (p) −
W(p, t) +
W(t, p)
t∈T i ∧p∈•t
∀p ∈ R(T i ), Mk+1 (p) =
X
t∈T i ∧p∈t•
W(t, p)
t∈T i ∧p∈t•
where R(T i ) =
[
R(t) and represents the places emptied by T i due to reset arcs 7 .
t∈T i
Since a reset arc from p to t, p ∈ R(t) consumes current marking dependent tokens, we
extend ptarc to include time-step and replace f 3, f 4, e1, r1, r2, a5 with f 7, f 8, e3, r6, r7, a7,
respectively in the Section 2.4 encoding and add rule f 9 for each reset arc:
6
The reset arc is involved here because we use a modified execution semantics of reset arcs compared to the
standard definition Araki and Kasami (1976). Even though both capture similar operation, our definition allows us
to model elimination of all quantity of a substance as soon as it is produced, even in a maximal firing set semantics.
Our semantics considers reset arc’s token consumption in contention with other arcs, while the standard definition
does not.
7
Our definition of conflicting transitions allows at most one transition with a reset arc from a place to fire,
any more create a conflict. Thus, the new marking computation is equivalent to ∀p ∈ P, Mk+1 (p) = Mk (p) −
P
P
P
( t∈Tk ∧p∈•t W(p, t) + t∈Tk ∧p∈R(t) Mk (p)) + t∈Tk ∧p∈t• W(t, p)
CHAPTER 2. PETRI NET ENCODING IN ASP FOR BIOLOGICAL DOMAIN
28
f7: Rules ptarc(pi , t j , W(pi , t j ), tsk ):-time(tsk ). for each non-reset arc (pi , t j ) ∈ E −
f8: Rules tparc(ti , p j , W(ti , p j ), tsk ):-time(tsk ). for each non-reset arc (ti , p j ) ∈ E +
e3: notenabled(T,TS) :- ptarc(P,T,N,TS), holds(P,Q,TS), Q < N,
place(P), trans(T), time(TS), num(N), num(Q).
r6: add(P,Q,T,TS) :- fires(T,TS), tparc(T,P,Q,TS), time(TS).
r7: del(P,Q,T,TS) :- fires(T,TS), ptarc(P,T,Q,TS), time(TS).
f9: Rules ptarc(pi , t j , X, tsk ) :- holds(pi , X, tsk ), num(X), X > 0. for each reset arc between pi
and t j using X = Mk (pi ) as arc-weight at time step tsk .
f10: Rules :- enabled(t j , tsk ),not fires(t j , tsk ), time(tsk ). for each transition t j with an incoming reset arc, i.e. R(t j ) , ∅.
a7: could not have(T,TS) :- enabled(T,TS), not fires(T,TS),
ptarc(S,T,Q,TS), holds(S,QQ,TS), tot decr(S,QQQ,TS), Q>QQ-QQQ.
Rule f 9 encodes place-transition arc with marking dependent weight to capture the notion
of a reset arc, while rule f 10 ensures that the reset-transition (i.e. the transition on which the
reset arc terminates) always fires when enabled.
Proposition 3 There is 1-1 correspondence between the answer sets of Π2 (PN R , M0 , k, ntok)
and the execution sequences of PN R .
The execution semantics of our definition are slightly different from the standard definition
in Araki and Kasami (1976), even though both capture similar operations. Our implementation
considers token consumption by reset arc in contention with other token consuming arcs from
the same place, while the standard definition considers token consumption as a side effect, not
in contention with other arcs.
We chose our definition to allow modeling of biological process that removes all available
quantity of a substance in a maximal firing set. Consider Figure 2.2, if dhap has 1 or more
tokens, our semantics would only permit either t5a or tr to fire in a single time-step, while
the standard semantics can allow both t5a and tr to fire simultaneously, such that the reset arc
removes left over tokens after (dhap, t5a) consumes one token.
We could have, instead, extended our encoding to include self-modifying nets Valk (1978),
but our modified-definition provides a simpler solution. Standard semantics, however, can be
easily encoded by replacing r5 by r5a0 , r5b0 ; replacing f 9, f 10 with f 90 ; and adding a8 as
follows:
f9’: rptarc(pi ,t j ). - for each reset arc between pi ∈ R(t j ) and t j .
a8: reset(P,TS) :- rptarc(P,T), place(P), trans(T), fires(T,TS), time(TS).
r5a’: holds(P,Q,TS+1) :- holds(P,Q1,TS), tot incr(P,Q2,TS), tot decr(P,Q3,TS),
Q=Q1+Q2-Q3, place(P), num(Q;Q1;Q2;Q3), time(TS), time(TS+1), not
reset(P,TS).
r5b’: holds(P,Q,TS+1) :- tot incr(P,Q,TS), place(P), num(Q), time(TS),
time(TS+1), reset(P,TS).
CHAPTER 2. PETRI NET ENCODING IN ASP FOR BIOLOGICAL DOMAIN
29
pyr
glu
gly1
atp
amp
cw1
Figure 2.3: Petri Net showing feedback inhibition arc from atp to gly1 with a bullet arrowhead.
Inhibitor arc weight is assumed 1 when not specified.
where, the fact f 90 encodes the reset arc; rule a8 encodes if place P will be reset at time T S
due to firing of transition T that has a reset arc on it from P to T ; rule r5a0 computes marking
at T S + 1 when place P is not being reset; and rule r5b0 computes marking at T S + 1 when P is
being reset.
2.6
Extension - Inhibitor Arcs
Definition 25 (Inhibitor Arc) An inhibitor arc Peterson (1977) is a place-transition arc that
inhibits its transition from firing as long as the place has any tokens in it. An inhibitor arc does
not consume any tokens from its input place. A Petri Net with reset and inhibitor arcs is a tuple
PN RI = (P, T, E, W, R, I), where, P, T, E, W, R are the same as for PN R ; and I : T → 2P defines
inhibitor arcs.
Figure 2.3 shows a Petri Net with inhibition arc from atp to gly1 with a bulleted arrowhead. It models biological feedback regulation in simplistic terms, where excess atp downstream causes the upstream atp production by glycolysis gly to be inhibited until the excess
quantity is consumed Reece et al. (2010). Petri Net execution semantics with inhibit arcs is
modified for determining enabled transitions as follows:
Definition 26 (Enabled Transition in PN RI ) A transition t is enabled with respect to marking
M, enabled M (t), if all its input places p have at least the number of tokens as the arc-weight
W(p, t) and all p ∈ I(t) have zero tokens, i.e. (∀p ∈ •t, W(p, t) ≤ M(p)) ∧ (∀p ∈ I(t), M(p) = 0)
We add inhibitor arcs to our encoding in Section 2.5 as follows:
f11: Rules iptarc(pi , t j , 1, tsk ):-time(tsk ). for each inhibitor arc between pi ∈ I(t j ) and t j .
e4: notenabled(T,TS) :- iptarc(P,T,N,TS), holds(P,Q,TS), place(P),
trans(T), time(TS), num(N), num(Q), Q >= N.
The new rule e4 encodes another reason for a transition to be disabled (or not enabled). An
inhibitor arc from p to t with arc weight N will cause its target transition t to not enable when
CHAPTER 2. PETRI NET ENCODING IN ASP FOR BIOLOGICAL DOMAIN
30
the number of tokens at its source place p is greater than or equal to N, where N is always 1 per
rule f 11.
Proposition 4 There is 1-1 correspondence between the answer sets of Π3 (PN RI , M0 , k, ntok)
and the execution sequences of PN.
2.7
Extension - Read Arcs
Definition 27 (Read Arc) A read arc (a test arc or a query arc) Christensen and Hansen (1993)
is an arc from place to transition, which enables its transition only when its source place has at
least the number of tokens as its arc weight. It does not consume any tokens from its input place.
A Petri Net with reset, inhibitor and read arcs is a tuple PN RIQ = (P, T, W, R, I, Q, QW), where,
P, T, E, W, R, I are the same as for PN RI ; Q ⊆ P × T defines read arcs; and QW : Q → N \ {0}
defines read arc weight.
h_is
3
25
3
h_mm
syn
adp
atp
p
Figure 2.4: Petri Net with read arc from h is to syn shown with arrowhead on both ends. The
transition syn will not fire unless there are at least 25 tokens in h is, but when it executes, it only
consumes 3 tokens.
Figure 2.4 shows a Petri Net with read arc from h is to syn shown with arrowhead on both
ends. It models the ATP synthase syn activation requiring a higher concentration of H+ ions
h is in the intermembrane space 8 . The reaction itself consumes a lower quantity of H+ ions
represented by the regular place-transition arc Reece et al. (2010); Berg et al. (2002). Petri Net
execution semantics with read arcs is modified for determining enabled transitions as follows:
Definition 28 (Enabled Transition in PN RIQ ) A transition t is enabled with respect to marking
M, enabled M (t), if all its input places p have at least the number of tokens as the arc-weight
W(p, t), all pi ∈ I(t) have zero tokens and all pq : (pq , t) ∈ Q have at least the number of tokens
as the arc-weight W(p, t), i.e. (∀p ∈ •t, W(p, t) ≤ M(p)) ∧ (∀p ∈ I(t), M(p) = 0) ∧ (∀(p, t) ∈
Q, M(p) ≥ QW(p, t))
8
This is an oversimplified model of syn (ATP synthase) activation, since the actual model requires an H+ concentration differential across membrane.
CHAPTER 2. PETRI NET ENCODING IN ASP FOR BIOLOGICAL DOMAIN
31
We add read arcs to our encoding of Section 2.6 as follows:
f12: Rules tptarc(pi , t j , QW(pi , t j ), tsk ):-time(tsk ). for each read arc (pi , t j ) ∈ Q.
e5: notenabled(T,TS):-tptarc(P,T,N,TS),holds(P,Q,TS),
place(P),trans(T), time(TS), num(N), num(Q), Q < N.
The new rule f 12 captures the read arc and its arc-weight; and the new rule e5 encodes
another reason for a transition to not be enabled. A read arc from p to t with arc weight N will
cause its target transition t to not enable when the number of tokens at its source place p is less
than the arc weight N.
Proposition 5 There is a 1-1 correspondence between the answer sets of Π4 (PN RIQ , M0 , k, ntok)
and the execution sequences of PN RIQ .
2.8
Extension - Colored Tokens
Higher level Petri Nets extend the notion of tokens to typed (or colored) tokens. This allows a
more compact representation of complicated networks Peterson et al. (1980).
Definition 29 (Petri Net with Colored Tokens) A Petri Net with Colored Tokens (with reset,
inhibit and read arcs) is a tuple PN C = (P, T, E, C, W, R, I, Q, QW), where P, T, E, R, I, Q are
the same as for basic Petri Nets, C = {c1 , . . . , cl } is a finite set of colors (or types), and arc
weights W : E → hC, mi, QW : Q → hC, mi are specified as multi-sets of colored tokens over
color set C. The state (or marking) of place nodes M(pi ) = hC, mi, pi ∈ P is specified as a
multiset of colored tokens over set C.
We will now update some definitions related to Petri Nets to include colored tokens.
Definition 30 (Marking) A marking M = (M(p1 ), . . . , M(pn )) assigns a colored multi-set of
tokens over the domain of colors C to each place {p1 , . . . , pn } ∈ P of PN C . The initial marking
is the initial token assignment of place nodes and is represented by M0 . The marking at time-step
k is written as Mk .
Definition 31 (Pre-set and post-set of a transition) The pre-set (or input-set) of a transition t
is •t = {p ∈ P|(p, t) ∈ E − }, while the post-set (or output-set) is t• = {p ∈ P|(t, p) ∈ E + }.
Definition 32 (Enabled Transition) A transition t is enabled with respect to marking M, enabled M (t),
if each of its input places p has at least the number of colored-tokens as the arc-weight W(p, t)9 ,
each of its inhibiting places pi ∈ I(t) have zero tokens and each of its read places pq : (pq , t) ∈
Q have at least the number of colored-tokens as the read-arc-weight QW(pq , t), i.e. (∀p ∈
•t, W(p, t) ≤ M(p)) ∧ (∀p ∈ I(t), M(p) = ∅) ∧ (∀(p, t) ∈ Q, M(p) ≥ QW(p, t)) for a given t.10
9
In the following text, for simplicity, we will use W(p, t) to mean W(hp, ti). We use similar simpler notation for
QW.
10
This is equivalent to ∀c ∈ C, (∀p ∈ •t, mW(p,t) (c) ≤ m M(p) (c))∧(∀p ∈ I(t), m M(p) (c) = 0)∧(∀(p, t) ∈ Q, m M(p) (c) ≥
mQW(p,t) (c)).
CHAPTER 2. PETRI NET ENCODING IN ASP FOR BIOLOGICAL DOMAIN
32
Definition 33 (Transition Execution) Execution of a transition t of PN C on a marking Mk
computes a new marking Mk+1 as:
∀p ∈ •tMk+1 (p) = Mk (p) − W(p, t)
∀p ∈ t • Mk+1 (p) = Mk (p) + W(t, p)
∀p ∈ R(t)Mk+1 (p) = Mk (p) − Mk (p)
Any number of enabled transitions may fire simultaneously as long as they don’t conflict. A
transition when fired consumed tokens from its pre-set places equivalent to the (place,transition)
arc-weight.
Definition 34 (Conflicting Transitions) A set of transitions T c ⊆ {t : enabled Mk (t)} is in conflict in PN C with respect to Mk if firing them will consume more tokens than are available at one
P
P
of their common input places, i.e., ∃p ∈ P : Mk (p) < ( t∈Tc ∧p∈•t W(p, t) + t∈Tc ∧p∈R(t) Mk (p))
Definition 35 (Firing Set) A firing set is a set T k = {tk1 , . . . , tkn } ⊆ T of simultaneously firing
transitions of PN C that are enabled and do not conflict w.r.t. to the current marking Mk of
PN. A set T k is not a firing set if there is an enabled reset-transition that is not in T k , i.e.
∃t ∈ enabled Mk , R(t) , ∅, t < T k . 11
Definition 36 (Firing Set Execution) Execution of a firing set T k of PN C on a marking Mk
computes a new marking Mk+1 as:
X
X
∀p ∈ P \ R(T k ), Mk+1 (p) = Mk (p) −
W(p, t) +
W(t, p)
t∈T k ∧p∈•t
∀p ∈ R(T k ), Mk+1 (p) =
X
t∈T k ∧p∈t•
W(t, p)
t∈T k ∧p∈t•
where R(T k ) =
S
t∈T k
R(t) 12
Definition 37 (Execution Sequence) An execution sequence X = M0 , T 0 , M1 , T 1 , . . . , Mk , T k , Mk+1
of PN is the simulation of a firing set sequence σ = T 1 , T 2 , . . . , T k w.r.t. an initial marking M0 ,
producing the final marking Mk+1 . Mk+1 is the transitive closure of firing set executions, where
subsequent marking become the initial marking for the next firing set.
For an execution sequence X = M0 , T 0 , M1 , T 1 , . . . , Mk , T k , Mk+1 , the firing of T 0 with respect
to marking M0 produces the marking M1 which becomes the initial marking for T 1 , which
produces M2 and so on.
If the Figure 2.5 Petri Net has the marking: M0 (mm) = [nadh/2, h/6], M0 (q) = [e/2],
M0 (cytc) = [e/2], M0 (is) = [o2/1], then transitions t1, t3, t4 are enabled. However, either
{t1, t3} or {t4} can fire simultaneously in a single firing at time 0 due to limited h tokens in mm.
t4 is said to be in conflict with t1, t3.
11
12
See footnote 6
See footnote 7
CHAPTER 2. PETRI NET ENCODING IN ASP FOR BIOLOGICAL DOMAIN
t12
o2/1
h/2
nadh/2
h/2
t10
nadh/2
h/6
t1
nadp/2
mm
h/2
q
h/2
h2o/1
e/2
t3
h/2
o2/1
is
e/2
33
e/2
cytc
e/2
t4
h/6
Figure 2.5: Petri Net with tokens of colors {e, h, h2o, nadh, nad p, o2}. Circles represent places,
and rectangles represent transitions. Arc weights such as “nadh/2, h/2”, “h/2, h2o/1” specify
the number of tokens consumed and produced during the execution of their respective transitions, where “nadh/2, h/2” means 2 tokens of color nadh and 2 tokens of h. Similar notation
is used to specify marking on places, when not present, the place is assumed to be empty of
tokens.
2.9
Translating Petri Nets with Colored Tokens to ASP
In order to represent the Petri Net PN C with colored tokens, initial marking M0 , and simulation
length k, we modify our encoding in Section 2.7 to add a new color parameter to all rules and
facts containing token counts in them. We keep rules f 1, f 2, f 5, f 6, f 10 remain as they were
for basic Petri Nets. We add a new rule f 13 for possible set of token colors and replace rules
f 7, f 8, f 9, f 11, f 12, i1 with f 14, f 15, f 16, f 17, f 18, i2 to add the color parameter as follows:
f13: Facts col(ck ) where ck ∈ C is a color.
f14: Rules ptarc(pi , t j , nc , c, tsk ) :- time(tsk ). for each (pi , t j ) ∈ E − , c ∈ C, nc = mW(pi ,t j ) (c) :
nc > 0.13
f15: Rules tparc(ti , p j , nc , c, tsk ) :- time(tsk ). for each (ti , p j ) ∈ E + , c ∈ C, nc = mW(ti ,p j ) (c) :
nc > 0.
f16: Rules ptarc(pi , t j , nc , c, tsk ) :- holds(pi , nc , c, tsk ), num(nc ), nc >0, time(tsk ). for each
(pi , t j ) : pi ∈ R(t j ), c ∈ C, nc = m Mk (pi ) (c).
f17: Rules iptarc(pi , t j , 1, c, tsk ) :- time(tsk ). for each (pi , t j ) : pi ∈ I(t j ), c ∈ C.
f18: Rules tptarc(pi , t j , nc , c, tsk ) :- time(tsk ).
mQW(pi ,t j ) (c) : nc > 0.
for each (pi , t j ) ∈ Q, c ∈ C, nc =
i2: Facts holds(pi , nc , c, 0). for each place pi ∈ P, c ∈ C, nc = m M0 (pi ) (c).
13
The time parameter tsk allows us to capture reset arcs, which consume tokens equal to the current (time-step
based) marking of their source nodes.
CHAPTER 2. PETRI NET ENCODING IN ASP FOR BIOLOGICAL DOMAIN
34
Next, we encode Petri Net’s execution behavior, which proceeds in discrete time steps. Rules
e3, e4, e5, e2 are replaced by e6, e7, e8, e9. For a transition ti to be enabled, it must satisfy the
following conditions: (i) @p j ∈ •ti : M(p j ) < W(p j , ti ), (ii) @p j ∈ I(ti ) : M(p j ) > 0, and
(iii) @(p j , ti ) ∈ Q : M(p j ) < QW(p j , ti ) . These three conditions are encoded as e6, e7, e8,
respectively and we encode the absence of any of these conditions for a transition as e9:
e6: notenabled(T,TS) :- ptarc(P,T, N,C,TS), holds(P,Q,C,TS), place(P),
trans(T), time(TS), num(N), num(Q), col(C), Q<N.
e7: notenabled(T,TS) :- iptarc(P,T,N,C,TS), holds(P,Q,C,TS), place(P),
trans(T), time(TS), num(N), num(Q), col(C), Q>=N.
e8: notenabled(T,TS) :- tptarc(P,T,N,C,TS), holds(P,Q,C,TS), place(P),
trans(T), time(TS), num(N), num(Q), col(C), Q<N.
e9: enabled(T,TS) :- trans(T), time(TS), not notenabled(T,TS).
Rule e6 captures the existence of an input place P with insufficient number of tokens for
transition T to fire. Rule e7 captures existence of a non-empty source place P of an inhibitor
arc to T preventing T from firing. Rule e8 captures existence of a source place P with less
than arc-weight tokens required by the read arc to transition T for T to be enabled. The,
holds(P,Q,C,TS) predicate captures the marking of place P at time T S as Q tokens of color C.
Rule e9 captures enabling of transition T when no reason for it to be not enabled is determined
by e6, e7, e8. In a biological context, this enabling is equivalent to a reaction’s pre-conditions
being satisfied. A reaction can proceed when its input substances are available in the required
quantities, it is not inhibited, and any required activation quantity of activating substances is
available.
Any subset of enabled transitions can fire simultaneously at a given time-step. We select a
subset of fireable transitions using the choice rule a1 The choice rule a1 either picks an enabled
transition T for firing at time T S or not. The combined effect over all transitions is to pick a
subset of enabled transitions to fire. Rule f 10 ensures that enabled reset-transitions will be a part
of this firing set. Whether these transitions are in conflict are checked by later rules a9, a3, a4. In
a biological context, the multiple firing models parallel processes occurring simultaneously. The
marking is updated according to the firing set using rules r8, r9, r10, r11, r12 which replaced
r6, r7, r3, r4, r5 as follows:
r8: add(P,Q,T,C,TS) :- fires(T,TS), tparc(T,P,Q,C,TS), time(TS).
r9: del(P,Q,T,C,TS) :- fires(T,TS), ptarc(P,T,Q,C,TS), time(TS).
r10: tot incr(P,QQ,C,TS) :- col(C), QQ = #sum[add(P,Q,T,C,TS) = Q : num(Q) :
trans(T)], time(TS), num(QQ), place(P).
r11: tot decr(P,QQ,C,TS) :- col(C), QQ = #sum[del(P,Q,T,C,TS) = Q : num(Q) :
trans(T)], time(TS), num(QQ), place(P).
CHAPTER 2. PETRI NET ENCODING IN ASP FOR BIOLOGICAL DOMAIN
35
r12: holds(P,Q,C,TS+1):-place(P),num(Q;Q1;Q2;Q3),time(TS),time(TS+1),col(C),
holds(P,Q1,C,TS), tot incr(P,Q2,C,TS), tot decr(P,Q3,C,TS), Q=Q1+Q2-Q3.
Rules r8 and r9 capture that Q tokens of color C will be added or removed to/from place
P due to firing of transition T at the respective time-step T S . Rules r10 and r11 aggregate
these tokens for each C for each place P (using aggregate assignment QQ = #sum[...]) at the
respective time-step T S . Rule r12 uses the aggregates to compute the next marking of P for
color C at the time-step (T S + 1) by subtracting removed tokens and adding added tokens to
the current marking. In a biological context, this captures the effect of a process / reaction,
which consumes its inputs and produces outputs for the downstream processes. We capture
token overconsumption using the rules a9, a3, a4 of which a9 is a colored replacement for a2
and is encoded as follows:
a9: consumesmore(P,TS) :- holds(P,Q,C,TS), tot decr(P,Q1,C,TS), Q1 > Q.
Rule a9 determines whether firing set selected by a1 will cause overconsumption of tokens
at P at time T S by comparing available tokens to aggregate tokens removed as determined by
r11. Rule a3 generalizes the notion of overconsumption, while rule a4 eliminates answer with
such overconsumption.
In a biological context, conflict (through overconsumption) models the limitation of input
substances, which dictate which downstream processes can occur simultaneously.
We remove rules a5, a6 from previous encoding to get the set firing semantics. Now, we
extend the definition (18) of 1-1 correspondence between the execution sequence of Petri Net
and the answer-sets of its ASP encoding to Petri Nets with colored tokens as follows.
Definition 38 Given a Petri Net PN with colored tokens, its initial marking M0 and its encoding
Π(PN, M0 , k, ntok) for k-steps and maximum ntok tokens at any place. We say that there is a 1-1
correspondence between the answer sets of Π(PN, M0 , k, ntok) and the execution sequences of
PN iff for each answer set A of Π(PN, M0 , k, ntok), there is a corresponding execution sequence
X = M0 , T 0 , M1 , . . . , Mk , T k , Mk+1 of PN and for each execution sequence X of PN there is an
answer-set A of Π(PN, M0 , k, ntok) such that
{ f ires(t, ts) : t ∈ T ts , 0 ≤ ts ≤ k} = { f ires(t, ts) : f ires(t, ts) ∈ A}
{holds(p, q, c, ts) : p ∈ P, c/q = Mts (p), 0 ≤ ts ≤ k + 1}
= {holds(p, q, c, ts) : holds(p, q, c, ts) ∈ A}
Proposition 6 There is 1-1 correspondence between the answer sets of Π5 (PN C , M0 , k, ntok)
and the execution sequences of PN.
To add maximal firing semantics, we add a6 as it is and replace a5 with a10 as follows:
a10: could not have(T,TS):-enabled(T,TS),not fires(T,TS), ptarc(S,T,Q,C,TS),
holds(S,QQ,C,TS), tot decr(S,QQQ,C,TS), Q > QQ - QQQ.
CHAPTER 2. PETRI NET ENCODING IN ASP FOR BIOLOGICAL DOMAIN
36
Rule a10 captures the fact that transition T , though enabled, could not have fired at T S , as its
firing would have caused overconsumption. Rule a6 eliminates any answers where an enabled
transition could have fired without causing overconsumption but did not. This modification
reduces the number of answers produced for the Petri Net in Figure 2.5 to 4. We can encode
other firing semantics with similar ease14 . We now look at how additional extensions can be
easily encoded by making small code changes.
2.10
Extension - Priority Transitions
Priority transitions enable ordering of Petri Net transitions, favoring high priority transitions
over lower priority ones Best and Koutny (1992). In a biological context, this is used to model
primary (or dominant) vs. secondary pathways / processes in a biological system. This prioritization may be due to an intervention (such as prioritizing elimination of a metabolite over
recycling it).
Definition 39 (Priority Colored Petri Net) A Priority Colored Petri Net with reset, inhibit, and
read arcs is a tuple PN pri = (P, T, E, C, W, R, I, Q, QW, Z), where: P, T, E, C, W, R, I, Q, QW are
the same as for PN C , and Z : T → N is a priority function that assigns priorities to transitions.
Lower number signifies higher priority.
Definition 40 (Enabled Transition) A transition ti is enabled in PN pri w.r.t. a marking M
(prenabled M (t)) if it would be enabled in PN C w.r.t. M and there isn’t another transition t j that
would be enabled in PN C (with respect to M) s.t. Z(t j ) < Z(ti ).
Definition 41 (Firing Set) A firing set is a set T k = {tk1 , . . . , tkn } ⊆ T of simultaneously firing
transitions of PN pri that are priority enabled and do not conflict w.r.t. to the current marking
Mk of PN. A set T k is not a firing set if there is an priority enabled reset-transition that is not in
T k , i.e. ∃t : prenabled Mk (t), R(t) , ∅, t < T k . 15
We add the following facts and rules to encode transition priority and enabled priority transitions:
f19: Facts transpr(ti ,pri ) where pri = Z(ti ) is ti0 s priority.
a11: notprenabled(T,TS) :- enabled(T,TS), transpr(T,P), enabled(TT,TS),
transpr(TT,PP), PP < P.
a12: prenabled(T,TS) :- enabled(T,TS), not notprenabled(T,TS).
14
For example, if interleaved semantics is desired, rules a10, a6 can changed to capture and eliminate answer-sets
in which more than one transition fires in a firing set as:
a10’: more than one fires :- fires(T1,TS),fires(T2,TS),T1!=T2,time(TS).
a6’: :-more than one fires.
15
See footnote 6
CHAPTER 2. PETRI NET ENCODING IN ASP FOR BIOLOGICAL DOMAIN
37
Rule a11 captures that an enabled transition T is not priority-enabled, if there is another
enabled transition with higher priority at T S . Rule a12 captures that transition T is priorityenabled at T S since there is no enabled transition with higher priority. We replace rules a1, f 10, a10, a6
with a13, f 20, a14, a15 respectively to propagate priority as follows:
a13: {fires(T,TS)} :- prenabled(T,TS), trans(T), time(TS).
f20: Rules :- prenabled(t j , tsk ),not fires(t j , tsk ), time(tsk ). for each transition t j with an incoming reset arc.
a14: could not have(T,TS) :- prenabled(T,TS), not fires(T,TS),
ptarc(S,T,Q,C,TS), holds(S,QQ,C,TS), tot decr(S,QQQ,C,TS), Q > QQ - QQQ.
a15: :- not could not have(T,TS), time(TS), prenabled(T,TS), not fires(T,TS),
trans(T).
Rules a13, f 10, a14, a15 perform the same function as a1, f 20, a10, a6, except that they
consider only priority-enabled transitions as compared all enabled transitions.
Proposition 7 There is 1-1 correspondence between the answer sets of Π6 (PN pri , M0 , k, ntok)
and the execution sequences of PN pri .
2.11
Extension - Timed Transitions
Biological processes vary in time required for them to complete. Timed transitions Ramchandani (1974) model this variation of duration. The timed transitions can be reentrant or nonreentrant16 . We extend our encoding to allow reentrant timed transitions.
Definition 42 (Priority Colored Petri Net with Timed Transitions) A Priority Colored Petri
Net with Timed Transitions, reset, inhibit, and query arcs is a tuple PN D = (P, T, E, C, W, R, I, Q, QW, Z, D),
where P, T, E, C, W, R, I, Q, QW, Z are the same as for PN pri , and D : T → N \ {0} is a duration
function that assigns positive integer durations to transitions.
Figure 2.6 shows an extended version of Petri Net model of the Electron Transport Chain Reece
et al. (2010) shown in Figure 2.5. The new transitions tq and tcytc (shown in dotted outline)
are timed transitions modeling the speed of the small carrier molecules, Coenzyme Q (q) and
Cytochrome C (cytc) as an effect of membrane fluidity. Higher numbers for transition duration
represent slower movement of the carrier molecules due to lower fluidity.
Definition 43 (Transition Execution) A transition t in PN D consumes tokens from its input
places and reset places immediately, while it produces tokens in its output places at the end of
transition duration D(t), as follows:
∀p ∈ •t, Mk+1 (p) = Mk (p) − W(p, t)
16
A reentrant transition is like a vehicle assembly line, which accepts new parts while working on multiple
vehicles at various stages of completion; whereas a non-reentrant transition only accepts new input when the current
processing is finished.
CHAPTER 2. PETRI NET ENCODING IN ASP FOR BIOLOGICAL DOMAIN
38
cytc'
e/2
h/2
[2]
q'
e/2
[2]
e/2
cytc
h/2
tq
e/2
t10
nadh/2
h/6
mm
nadp/2
nadh/2
h/2
e/2
t1
e/2
e/2
t3
e/2
tcytc
t12
q
o2/1
is
h/2
h2o/1
o2/1
t4
h/2
h/6
Figure 2.6: An extended version of the Petri Net model from Fig. 2.5. The new transitions
tq, tcytc have a duration of 2 each (shown in square brackets (“[ ]”) next to the transition). When
missing, transition duration is assumed to be 1.
∀p ∈ t•, Mk+D(t) (p) = Mk+D(t)−1 (p) + W(p, t)
∀p ∈ R(t), Mk+1 (p) = Mk (p) − Mk (p)
Execution in PN D changes, since the token update from Mk to Mk+1 can involve transitions
that started at some time l before time k, but finish at k + 1.
Definition 44 (Firing Set Execution) New marking due to firing set execution is computed as
follows:
X
X
∀p ∈ P \ R(T k ), Mk+1 (p) = Mk (p) −
W(p, t) +
W(t, p)
t∈T k ,p∈•t
∀p ∈ R(T k ), Mk+1 (p) =
t∈T l ,p∈t•:0≤l≤k,l+D(t)=k+1
X
W(t, p)
t∈T l ,p∈t•:l≤k,l+D(t)=k+1
where R(T i ) = ∪t∈Ti R(t).
A timed transition t produces its output D(t) time units after being fired. We replace f 15
with f 21 adding transition duration and replace rule r8 with r13 that produces tokens at the end
of transition duration:
f21: Rules tparc(ti , p j , nc , c, tsk , D(ti )):-time(tsk ). for each (ti , p j ) ∈ E + , c ∈ C, nc = mW(ti ,p j ) (c) :
nc > 0.
CHAPTER 2. PETRI NET ENCODING IN ASP FOR BIOLOGICAL DOMAIN
39
r13: add(P,Q,T,C,TS):-fires(T,TS0),time(TS0;TS), tparc(T,P,Q,C,TS0,D),
TS=TS0+D-1.
Proposition 8 There is 1-1 correspondence between the answer sets of Π7 (PN D , M0 , k, ntok)
and the execution sequences of PN D .
Above implementation of timed-transition is reentrant, however, we can easily make these
timed transitions non-reentrant by adding rule e10 that disallows a transition from being enabled
if it is already in progress:
e10: notenabled(T,TS):-fires(T,TS0), num(N), TS>TS0, tparc(T,P,N,C,TS0,D),
col(C), time(TS0), time(TS), TS<(TS0+D).
2.12
Other Extensions
Other Petri Net extensions can be implemented with similar ease. For example, Guard Conditions on transitions can be trivially implemented as a notenabled/2 rules. Self Modifying Petri
Nets Valk (1978), which allow marking-dependent arc-weights can be implemented in a similar
manner as the Reset Arc extension in section 2.5. Object Petri Nets Valk (2004), in which each
token is a Petri Net itself can be implemented (using token reference semantics) by adding an
additional “network-id” parameter to our encoding, where “id=0” is reserved for system net and
higher numbers are used for token nets. Transition coordination between system & token nets is
enforced through constraints on transition labels, where transition labels are added as additional
facts about transitions.
2.13
Related Work
Petri Nets have been previously encoded in ASP, but the previous implementations have been
limited to restricted classes of Petri Nets. For example, 1-safe Petri Net to ASP translation
has been presented in Heljanko and Niemelä (2000), which is limited to binary Petri Nets.
Translation of Logic Petri Nets to ASP has been presented in Behrens and Dix (2007), but their
model cannot handle numerical aggregation of tokens from multiple input transitions to the
same place. Our work focused on problems in the biological domain and is more generalized.
We can represent reset arcs, inhibition arcs, priority arcs as well as durative transitions.
2.14
Conclusion
We have presented an encoding of basic Petri Nets in ASP and showed how it can be easily
extended to include extension to model various biological constructs. Portions of this work
were published in Anwar et al. (2013b) and Anwar et al. (2013a). In the next chapter we will
use Petri Nets and their ASP encoding to model biological pathways to answer questions about
them.
Chapter 3
Answering Questions using Petri Nets
and ASP
3.1
Introduction
In this chapter we use various Petri Net extensions presented in Chapter 2 and their ASP encoding to answer question from Reece et al. (2010) that were a part of the Second Deep Knowledge
Representation Challenge1 .
Definition 45 (Rate) Rate of product P is defined as the quantity of P produced per unit-time.
Rate of an action A is defined as the number of time A occurs per unit-time.
3.2
Comparing Altered Trajectories due to Reset Intervention
Question 1 At one point in the process of glycolysis, both dihydroxyacetone phosphate (DHAP)
and glyceraldehyde 3-phosphate (G3P) are produced. Isomerase catalyzes the reversible conversion between these two isomers. The conversion of DHAP to G3P never reaches equilibrium
and G3P is used in the next step of glycolysis. What would happen to the rate of glycolysis if
DHAP were removed from the process of glycolysis as quickly as it was produced?
Provided Answer:
“Glycolysis is likely to stop, or at least slow it down. The conversion of the two
isomers is reversible, and the removal of DHAP will cause the reaction to shift in
that direction so more G3P is converted to DHAP. If less (or no) G3P were available,
the conversion of G3P into DHAP would slow down (or be unable to occur).”
Solution 1 The process of glycolysis is shown in Fig 9.9 of Campbell’s book. Glycolysis splits
Glucose into Pyruvate. In the process it produces ATP and NADH. Any one of these can be used
1
https://sites.google.com/site/2nddeepkrchallenge/
40
CHAPTER 3. ANSWERING QUESTIONS USING PETRI NETS AND ASP
41
to gauge the glycolysis rate, since they will be produced in proportion to the input Glucose. The
amount of pyruvate produced is the best choice since it is the direct end product of glycolysis.
The ratio of the quantity of pyruvate produced over a representative span of time gives us the
glycolysis rate. We assume a steady supply of Glucose is available and also assume that sufficient quantity of various enzymes used in glycolysis is available, since the question does not
place any restriction on these substances.
We narrow our focus to a subsection from Fructose 1,6-bisphosphate (F16BP) to 1,3-Bisphosphoglycerate
(BPG13) as shown in Figure 3.1 since that is the part the question is concerned with. We can
ignore the linear chain up-stream of F16BP as well as the linear chain down-stream of BPG13
since the amount of F16BP available will be equal to Glucose and the amount of BPG13 will
be equal to the amount of Pyruvate given our steady supply assumption.
ald
t3
f16bp
t4
g3p
t5a
dhap
p
t5b
tr
2
2
2
nadp
iso
t6
2
2
bpg13
nadh
h
Figure 3.1: Petri Net graph relevant to question 1. “f16bp” is the compound Fructose 1,6biphosphate, “bpg13” is 1,3-Bisphosphoglycerate. Transition tr shown in dotted lines is added
to model the elimination of dhap as soon as it is produced.
We fulfill the steady supply requirement of Glucose by a source transition-node t3. We fulfill
sufficient enzyme supply by a fixed quantity for each enzyme such that this quantity is in excess
of what can be consumed during our simulation interval. Where the simulation interval is the
number of time-steps over which we will measure the rate of glycolysis.
We model the elimination of DHAP as soon as it is produced with a reset arc, shown with a
dotted style in Figures 3.1. Such an arc removes all tokens from its source place when it fires.
Since we have added it as an unconditional arc, it is always enabled for firing. We encode both
situations in ASP with the maximal firing set policy. Both situations (without and with reset arc)
CHAPTER 3. ANSWERING QUESTIONS USING PETRI NETS AND ASP
42
are encoded in ASP and run for 10 steps. At the end of those 10 steps the amount of BPG13 is
compared to determine the difference in the rate of glycolysis.
In normal situation (without (dhap, tr) reset arc), unique quantities of “bpg13” from all (2)
answer-sets after 10 steps were as follows:
holds(bpg13,14,10)
holds(bpg13,16,10)
with reset arc tr, unique quantities of “bpg13” from all (512) answer-sets after 10 steps
were as follows:
holds(bpg13,0,10)
holds(bpg13,10,10)
holds(bpg13,12,10)
holds(bpg13,14,10)
holds(bpg13,16,10)
holds(bpg13,2,10)
holds(bpg13,4,10)
holds(bpg13,6,10)
holds(bpg13,8,10)
Unique quantities of bpg13 produced by different answersets of 10 step simulation
30
normal
abnormal
Unique amount of bpg13 produced
25
20
15
10
5
0
0
2
4
6
8
10
12
14
Unique run serial number
Figure 3.2: Amount of “bpg13” produced in unique answer-sets produced by a 10 step simulation. The graph shows two situations, without the (dhap, tr) reset arc (normal situation) and
with the reset arc (abnormal situation). The purpose of this graph is to depict the variation in
the amounts of glycolysis produced in various answer sets.
Note that the rate of glycolysis is generally lower when DHAP is immediately consumed.
It is as low as zero essentially stopping glycolysis. The range of values are due to the choice
between G3P being converted to DHAP or BPG13. If more G3P is converted to DHAP, then less
CHAPTER 3. ANSWERING QUESTIONS USING PETRI NETS AND ASP
43
BPG13 is produced and vice versa. Also, note that if G3P is not converted to BPG13, no NADH
or ATP is produced either due to the liner chain from G3P to Pyruvate. The unique quantities
of BPG13 are shown in a graphical format in Figure 3.2, while a trend of average quantity of
BPG13 produced is shown in Figure 3.3.
Average quantity of bpg13 produced
30
normal
abnormal
25
bpg13 quantity
20
15
10
5
0
0
2
4
6
8
10
12
14
16
Time
Figure 3.3: Average amount of “bpg13” produced during the 10-step simulation at various
time steps. The average is over all answer-sets. The graph shows two situations, without the
(dhap, tr) reset arc (normal situation) and with the reset arc (abnormal situation). The divergence in “bpg13” production is clearly shown.
We created a minimal model of the Petri Net in Figure 3.1 by removing enzymes and reactants that were not relevant to the question and did not contribute to the estimation of glycolysis.
This is shown in Figure 3.4.
tr
t3
f16bp
dhap
t5a
g3p
t5b
t4
t6
2
bpg13
Figure 3.4: Minimal version of the Petri Net graph in Figure 3.1. All reactants that do not
contribute to the estimation of the rate of glycolysis have been removed.
CHAPTER 3. ANSWERING QUESTIONS USING PETRI NETS AND ASP
44
Simulating it for 10 steps with the same initial marking as the Petri Net in Figure 3.1 produced the same results as for Figure 3.1.
3.3
Determining Conditions Leading to an Observation
Question 2 When and how does the body switch to B oxidation versus glycolysis as the major
way of burning fuel?
Provided Answer:
“The relative volumes of the raw materials for B oxidation and glycolysis indicate
which of these two processes will occur. Glycolysis uses the raw material glucose,
and B oxidation uses Acyl CoA from fatty acids. When the blood sugar level decreases below its homeostatic level, then B oxidation will occur with available fatty
acids. If no fatty acids are immediately available, glucagon and other hormones
regulate release of stored sugar and fat, or even catabolism of proteins and nucleic
acids, to be used as energy sources.”
Solution 2 The answer provided requires background knowledge about the mechanism that
regulates which source of energy will be used. This information is not presented in Chapter 9
of Campbell’s book, which is the source material of this exercise. However, we can model it
based on background information combined with Figure 9.19 of Campbell’s book. Our model
is presented in Figure 3.52 .
fats
t1
fac
box
acoa
sug
prot
t5
t9
amin
Figure 3.5: Petri Net graph relevant to question 2. “fats” are fats, “prot” are proteins, “fac” are
fatty acids, “sug” are sugars, “amin” are amino acids and “acoa” is ACoA. Transition “box” is
the beta oxidation, “t5” is glycolysis, “t1” is fat digestion into fatty acids, and “t9” is protein
deamination.
We can test the model by simulating it for a time period and testing whether beta oxidation
(box) is started when sugar (sug) is finished. We do not need a steady supply of sugar in this
2
We can extend this model by adding expressions to inhibition arcs that compare available substances.
CHAPTER 3. ANSWERING QUESTIONS USING PETRI NETS AND ASP
45
case, just enough to be consumed in a few time steps to capture the switch over. Fats and
proteins may or may not modeled as a steady supply, since their bioavailability is dependent
upon a number of external factors. We assume a steady supply of both and model it with large
enough initial quantity that will last beyond the simulation period.
We translate the petri net model into ASP and run it for 10 iterations. Following are the
results:
holds(acoa,0,0) holds(amin,0,0) holds(fac,0,0) holds(fats,5,0)
holds(prot,3,0) holds(sug,4,0)
fires(t1,0) fires(t5,0) fires(t9,0)
holds(acoa,1,1) holds(amin,1,1) holds(fac,1,1) holds(fats,4,1)
holds(prot,3,1) holds(sug,3,1)
fires(t5,1)
holds(acoa,2,2) holds(amin,1,2) holds(fac,1,2) holds(fats,4,2)
holds(prot,3,2) holds(sug,2,2)
fires(t5,2)
holds(acoa,3,3) holds(amin,1,3) holds(fac,1,3) holds(fats,4,3)
holds(prot,3,3) holds(sug,1,3)
fires(t5,3)
holds(acoa,4,4) holds(amin,1,4) holds(fac,1,4) holds(fats,4,4)
holds(prot,3,4) holds(sug,0,4)
fires(box,4)
holds(acoa,5,5) holds(amin,1,5) holds(fac,0,5) holds(fats,4,5)
holds(prot,3,5) holds(sug,0,5)
fires(t1,5) fires(t9,5)
holds(acoa,5,6) holds(amin,2,6) holds(fac,1,6) holds(fats,3,6)
holds(prot,3,6) holds(sug,0,6)
fires(box,6)c
holds(acoa,6,7) holds(amin,2,7) holds(fac,0,7) holds(fats,3,7)
holds(prot,3,7) holds(sug,0,7)
fires(t1,7) fires(t9,7)
holds(acoa,6,8) holds(amin,3,8) holds(fac,1,8) holds(fats,2,8)
holds(prot,3,8) holds(sug,0,8)
fires(box,8)
CHAPTER 3. ANSWERING QUESTIONS USING PETRI NETS AND ASP
46
holds(acoa,7,9) holds(amin,3,9) holds(fac,0,9) holds(fats,2,9)
holds(prot,3,9) holds(sug,0,9)
fires(t1,9) fires(t9,9)
holds(acoa,7,10) holds(amin,4,10) holds(fac,1,10) holds(fats,1,10)
holds(prot,3,10) holds(sug,0,10)
fires(box,10)
We can see that by time-step 4, the sugar supply is depleted and beta oxidation starts occurring.
3.4
Comparing Altered Trajectories due to Accumulation Intervention
Question 3 ATP is accumulating in the cell. What affect would this have on the rate of glycolysis? Explain.
Provided Answer:
“ATP and AMP regulate the activity of phosphofructokinase. When there is an
abundance of AMP in the cell, this indicates that the rate of ATP consumption is
high. The cell is in need for more ATP. If ATP is accumulating in the cell, this
indicates that the cell’s demand for ATP had decreased. The cell can decrease its
production of ATP. Therefore, the rate of glycolysis will decrease.”
Solution 3 Control of cellular respiration is summarized in Fig 9.20 of Campbell’s book. We
can gauge the rate of glycolysis by the amount of Pyruvate produced, which is the end product
of glycolysis. We assume a steady supply of glucose is available. Its availability is not impacted
by any of the feedback mechanism depicted in Fig 9.20 of Campbell’s book or restricted by the
question. We can ignore the respiration steps after glycolysis, since they are directly dependent
upon the end product of glycolysis, i.e. Pyruvate. These steps only reinforce the negative effect
of ATP. The Citrate feed-back shown in Campbell’s Fig 9.20 is also not relevant to the question,
so we can assume a constant level of it and leave it out of the picture. Another simplification
that we do is to treat the inhibition of Phosphofructokinase (PFK) by ATP as the inhibition of
glycolysis itself. This is justified, since PFK is on a linear path from Glucose to Fructose 1,6bisphosphate (F16BP), and all downstream product quantities are directly dependent upon the
amount of F16BP (as shown in Campbell’s Fig 9.9), given steady supply of substances involved
in glycolysis. Our assumption also applies to ATP consumed in Fig 9.9. Our simplified picture
is shown in Figure 3.6 as a Petri Net.
CHAPTER 3. ANSWERING QUESTIONS USING PETRI NETS AND ASP
2
glu
amp
gly1
2
cw1
2
4
47
pyr
atp
2
Figure 3.6: Petri Net graph relevant to question 3. “glu” is Glucose, “pyr” is Pyruvate. Transitions “gly1” represents glycolysis and “cw1” is cellular work that consumes ATP and produces
AMP. Transition “gly1” is inhibited only when the number of atp tokens is greater than 4.
We model cellular work that recycles ATP to AMP (see p/181 of Campbell’s book) by the
cw1 transition, shown in dotted style. In normal circumstances, this arc does not let ATP to
collect. If we reduce the arc-weights incident on cw1 to 1, we get the situation where less work
is being done and some ATP will collect, as a result glycolysis will pause and resume. If we
remove cw1 (representing no cellular work), ATP will start accumulating and glycolysis will
stop. We use an arbitrary arc-weight of 4 on the inhibition arc (atp, gly1) to model an elevated
level of ATP beyond normal that would cause inhibition 3 . We encode all three situations in ASP
with maximal firing set policy. We run them for 10 steps and compare the quantity of pyruvate
produced to determine the difference in the rate of glycolysis.
In normal situation when cellular work is being performed (cw1 arc is present), unique
quantities of “pyr” after 10 step are as follows:
holds(pyr,20,10)
when the cellular work is reduced, i.e. ((atp, cw1), (cw1, amp) arc weights changed to 1),
unique quantities of “pyr” after 10 steps are as follows:
holds(pyr,14,10)
with no cellular work (cw1 arc removed), unique quantities of “pyr” after 10 steps are as
follows:
holds(pyr,6,10)
The results show the rate of glycolysis reducing as the cellular work decreases to the point
where it stops once ATP reaches the inhibition threshold. Higher numbers of ATP produced
3
An alternate modeling would be compare the number of tokens on the amp node and the atp node and set a
level-threshold that inhibits gly1. Such technique is common in colored-peri nets.
CHAPTER 3. ANSWERING QUESTIONS USING PETRI NETS AND ASP
48
in later steps of cellular respiration will reinforce this inhibition even more quickly. Trend of
answers from various runs is shown in Figure 3.7.
Quantity of pyr produced by various run lengths
35
normal
small amount of cellular work
large amount of cellular work
Amount of pyr produced
30
25
20
15
10
5
4
6
8
10
12
14
16
Run length
Figure 3.7: Amount of pyruvate produced from various lengths of runs.
3.5
Comparing Altered Trajectories due to Initial Value Intervention
Question 4 A muscle cell had used up its supply of oxygen and ATP. Explain what affect would
this have on the rate of cellular respiration and glycolysis?
Provided Answer:
“Oxygen is needed for cellular respiration to occur. Therefore, cellular respiration
would stop. The cell would generate ATP by glycolysis only. Decrease in the
concentration of ATP in the cell would stimulate an increased rate of glycolysis in
order to produce more ATP.”
Solution 4 Figure 9.18 of Campbell’s book gives the general idea of what happens when oxygen
is not present. Figure 9.20 of Campbell’s book shows the control of glycolysis by ATP. To
formulate the answer, we need pieces from both.
ATP inhibits Phosphofructokinase (Fig 9.20 of Campbell), which is an enzyme used in glycolysis. No ATP means that enzyme is no longer inhibited and glycolysis can proceed at full
throttle. Pyruvate either goes through aerobic respiration when oxygen is present or it goes
through fermentation when oxygen is absent (Fig 9.18 of Campbell). We can monitor the rate
of glycolysis and cellular respiration by observing these operations occurring (by looking at
corresponding transition firing) over a simulation time period. Our simplified Petri Net model
is shown in Figure 3.8.
CHAPTER 3. ANSWERING QUESTIONS USING PETRI NETS AND ASP
49
We ignore the details of processes following glycolysis, except that these steps produce additional ATP. We do not need an exact number of ATP produced as long as we keep it higher
than the ATP produced by glycolysis. Higher numbers will just have a higher negative feed-back
(or inhibition) effect on glycolysis. We ignore citrate’s inhibition of glycolysis since that is not
relevant to the question and since it gets recycled by the citric acid cycle (see Fig 9.12 of Campbell). We also ignore AMP, since it is not relevant to the question, by assuming sufficient supply
to maintain glycolysis. We also assume continuous cellular work consuming ATP, without that
ATP will accumulate almost immediately and stop glycolysis.
We assume a steady supply of glucose is available to carry out glycolysis and fulfill this
requirement by having a quantity in excess of the consumption during our simulation interval.
All other substances participating in glycolysis are assumed to be available in a steady supply
so that glycolysis can continue.
4
glu
2
gly1
atp
2
cw1
4
o2
res1
pyr
fer1
eth
Figure 3.8: Petri Net graph relevant to question 4. “glu” is Glucose, “pyr” is Pyruvate, “atp”
is ATP, “eth” is ethenol or other products of fermentation, and “o2” is Oxygen. Transitions
“gly1” represents glycolysis, “res1” is respiration in presence of oxygen, “fer1” is fermentation
when no oxygen is present, and “cw1” is cellular work that consumes ATP. Transition “gly1” is
inhibited only when the number of atp tokens is greater than 4.
We then consider two scenarios, one where oxygen is present and where oxygen is absent
and determine the change in rate of glycolysis and respiration by counting the firings of their
respective transitions. We encode both situations in ASP with maximal firing set policy. Both
situations are executed for 10 steps. At the end of those steps the firing count of “gly1” and
“res1” is computed and compared to determine the difference in the rates of glycolysis and
respiration respectively.
In the normal situation (when oxygen is present), we get the following answer sets:
fires(gly1,0)
fires(cw1,1) fires(gly1,1) fires(res1,1)
fires(cw1,2) fires(res1,2)
fires(cw1,3)
fires(cw1,4)
CHAPTER 3. ANSWERING QUESTIONS USING PETRI NETS AND ASP
50
fires(cw1,5) fires(gly1,5)
fires(cw1,6) fires(gly1,6) fires(res1,6)
fires(cw1,7) fires(res1,7)
fires(cw1,8)
fires(cw1,9)
fires(cw1,10)
while in the abnormal situation (when oxygen is absent), we get the following firings:
fires(gly1,0)
fires(cw1,1) fires(fer1,1) fires(gly1,1)
fires(cw1,2) fires(fer1,2) fires(gly1,2)
fires(cw1,3) fires(fer1,3) fires(gly1,3)
fires(cw1,4) fires(fer1,4) fires(gly1,4)
fires(cw1,5) fires(fer1,5) fires(gly1,5)
fires(cw1,6) fires(fer1,6) fires(gly1,6)
fires(cw1,7) fires(fer1,7) fires(gly1,7)
fires(cw1,8) fires(fer1,8) fires(gly1,8)
fires(cw1,9) fires(fer1,9) fires(gly1,9)
fires(cw1,10) fires(fer1,10) fires(gly1,10)
Note that the number of firings of glycolysis for normal situation is lower when oxygen is
present and higher when oxygen is absent. While, the number of firings is zero when no oxygen
is present. Thus, respiration stops when no oxygen is present and the need of ATP by cellular
work is fulfilled by a higher amount of glycolysis. Trend from various runs is shown in Figure
3.9.
Firing counts of gly1 and res1 at various run lengths
o2 present
o2 absent
gly1
res1
Firing count
16
14
14
12
12
10
10
8
8
6
6
4
4
2
2
0
0
4
6
8
10
Run length
12
gly1
res1
16
14
16
4
6
8
10
12
14
16
Run length
Figure 3.9: Firing counts of glycolysis (gly1) and respiration (res1) for different simulation
lengths for the petri net in Figure 3.8
CHAPTER 3. ANSWERING QUESTIONS USING PETRI NETS AND ASP
3.6
51
Comparing Altered Trajectories due to Inhibition Intervention
Question 5 The final protein complex in the electron transport chain of the mitochondria is
non-functional. Explain the effect of this on pH of the intermembrane space of the mitochondria.
Provided Answer:
“The H+ ion gradient would gradually decrease and the pH would gradually increase. The other proteins in the chain are still able to produce the H+ ion gradient.
However, a non-functional, final protein in the electron transport chain would mean
that oxygen is not shuttling electrons away from the electron transport chain. This
would cause a backup in the chain, and the other proteins in the electron transport
chain would no longer be able to accept electrons and pump H+ ions into the intermembrane space. A concentration decrease in the H+ ions means an increase in
the pH.”
Solution 5 The electron transport chain is shown in Fig 9.15 (1) of Campbell’s book. In order
to explain the effect on pH, we will show the change in the execution of the electron transport
chain with both a functioning and non-functioning final protein. Since pH depends upon the
concentration of H+ ions, we will quantify the difference its quantity in the intermembrane
space in both scenarios as well. We assume that a steady input of NADH, FADH2, H+ and
O2 are available in the mitochondrial matrix. We also assume an electron carrying capacity
of 2 for both ubiquinone (Q)4 and cytochrome c (Cyt c). This carrying capacity is background
information not provided in Campbell’s Chapter 9. As with previous questions, we fulfill the
steady supply requirement of substances by having input quantities in excess of what would be
consumed during our simulation interval.
4
http://www.benbest.com/nutrceut/CoEnzymeQ.html
CHAPTER 3. ANSWERING QUESTIONS USING PETRI NETS AND ASP
52
h:2, o2:1
h:1
fadh2:1
t2
fad:1
mm
h:1, nadh:1
nadp:1
e:2
e:2
q
e:2
e:2
t1
t3
e:2
e:2
cytc
h:1
e:2
h:1
t4
e:2
h:1
is
h:1
h2o:1
h:4
adp:1, p:1
t6
h:1, atp:1
Figure 3.10: Petri Net graph relevant to question 5. “is” is the intermembrane space, “mm” is
mitochondrial matrix, “q” is ubiquinone and “cytc” is cytochrome c. The inhibition arcs (q, t1),
(q, t2) and (cytc, t3) capture the electron carrying capacities of q and cytc. Over capacity will
cause backup in electron transport chain. Tokens are colored, e.g. “h:1, nadh:1” specify one
token of h and nadh each. Token types are “h” for H+, “nadh” for NADH, “nadp” for NADP,
“fadh2” for FADH2, “fad” for FAD, “e” for electrons, “o2” for oxygen and “h2o” for water. We
remove t4 to model non-functioning protein complex IV.
We model this problem as a colored petri net shown in Figure 3.10. The normal situation is
made up of the entire graph. The abnormal situation (with non-functional final protein complex)
is modeled by removing transition t4 from the graph5 . We encode both situations in ASP with
maximal firing set policy. Both are run for 10 steps. The amount of h (H+) is compared in the
is (intermembrane space) to determine change in pH and the firing sequence is compared to
explain the effect.
In normal situation (entire graph), we get the following:
fires(t1,0)
fires(t3,1)
fires(t1,2)
fires(t3,3)
fires(t1,4)
fires(t3,5)
fires(t1,6)
fires(t3,7)
fires(t2,0)
fires(t2,2)
fires(t4,3)
fires(t2,4)
fires(t4,5)
fires(t2,6)
fires(t4,7)
fires(t3,2) fires(t4,2)
fires(t6,3)
fires(t3,4) fires(t4,4) fires(t6,4)
fires(t6,5)
fires(t3,6) fires(t4,6) fires(t6,6)
fires(t6,7)
5
Alternatively, we can model a non-functioning transition by attaching an inhibition arc to it with one token at its
source place
CHAPTER 3. ANSWERING QUESTIONS USING PETRI NETS AND ASP
53
fires(t1,8) fires(t2,8) fires(t3,8) fires(t4,8) fires(t6,8)
fires(t3,9) fires(t4,9) fires(t6,9)
fires(t1,10) fires(t2,10) fires(t3,10) fires(t4,10) fires(t6,10)
holds(is,15,h,10)
with t4 removed, we get the following:
fires(t1,0) fires(t2,0)
fires(t3,1)
fires(t1,2) fires(t2,2) fires(t3,2)
fires(t6,3)
fires(t6,4)
holds(is,2,h,10)
Note that the amount of H+ (h) produced in the intermembrane space (is) is much smaller
when the final protein complex is non-functional (t4 removed). Lower H+ translates to higher
pH. Thus, the pH of intermembrane space will increase as a result of nonfunctional final protein.
Also, note that the firing of t3, t1 and t2 responsible for shuttling electrons also stop very quickly
when t4 no longer removes the electrons (e) from Cyt c (cytc) to produce H2 O. This is because
cytc and q are at their capacity on electrons that they can carry and stop the electron transport
chain by inhibiting transitions t3, t2 and t1. Trend for various runs is shown in Figure 3.11.
Quantity of H+ in the intermembrane space
25
normal
abnormal
Quantity of H+
20
15
10
5
0
0
2
4
6
8
10
12
14
16
Time step
Figure 3.11: Simulation of Petri Net in Figure 3.10. In a complete model of the biological
system, there will be a mechanism that keeps the quantity of H+ in check in the intermembrane
space and will plateau at some point.
CHAPTER 3. ANSWERING QUESTIONS USING PETRI NETS AND ASP
3.7
54
Comparing Altered Trajectories due to Gradient Equilization
Intervention
Question 6 Exposure to a toxin caused the membranes to become permeable to ions. In a mitochondrion, how would this affect the pH in the intermembrane space and also ATP production?
Provided Answer:
“The pH of the intermembrane space would decrease as H+ ions diffuse through
the membrane, and the H+ ion gradient is lost. The H+ gradient is essential in
ATP production b/c facilitated diffusion of H+ through ATP synthase drives ATP
synthesis. Decreasing the pH would lead to a decrease in the rate of diffusion
through ATP synthase and therefore a decrease in the production of ATP.”
Solution 6 Oxidative phosphorylation is shown in Fig 9.15 of Campbell’s book. In order to
explain the effect on pH in the intermembrane space and the ATP production we will show
the change in the amount of H+ ions in the intermembrane space as well as the amount of
ATP produced when the inner mitochondrial membrane is impermeable and permeable. Note
that the concentration of H+ determines the pH. we have chosen to simplify the diagram by
not having FADH2 in the picture. Its removal does not change the response, since it provides
an alternate input mechanism to electron transport chain. We will assume that a steady input
of NADH, H+, O2, ADP and P is available in the mitochondrial matrix. We also assume an
electron carrying capacity of 2 for both ubiquinone (Q) and cytochrome c (Cyt c). We fulfill the
steady supply requirement of substances by having input quantities in excess of what would be
consumed during our simulation interval.
We model this problem as a colored petri net shown in Figure 3.12. Transition t6, t7 shown
in dotted style are added to model the abnormal situation6 . They capture the diffusion of H+
ions back from Intermembrane Space to the Mitochondrial matrix. One or both may be enabled
to capture degrees of permeability. we have added a condition on the firing of t5 (ATP Synthase
activation) to enforce gradient to pump ATP Synthase.
6
If reverse permeability is also desired additional arcs may be added from mm to is
CHAPTER 3. ANSWERING QUESTIONS USING PETRI NETS AND ASP
55
h:1
h:1, nadh:1
t1
e:2
q
e:2
t3
h:1
h:1
e:2
cytc
e:2
h:1
mm
(is(h)>=mm(h)+3)
adp:1, p:1
t5
atp:1, h:1
h:1
h:1
t6
t7
t4
h:1
h:1
is
h:1
h:1
Figure 3.12: Petri Net graph relevant to question 6. “is” is the intermembrane space, “mm”
is mitochondrial matrix, “q” is ubiquinone and “cytc” is cytochrome c. Tokens are colored,
e.g. “h:1, nadh:1” specify one token of h and nadh each. Token types are “h” for H+, “nadh”
for NADH, “e” for electrons, “o2” for oxygen, “h2o” for water, “atp” for ATP and “adp” for
ADP. We add t6, t7 to model cross domain diffusion from intermembrane space to mitochondrial
matrix. One or both of t6, t7 may be enabled at a time to control the degree of permeability. The
text above “t5” is an additional condition which must be satisfied for “t5” to be enabled.
We encode both situations in ASP with maximal firing set policy. Both are run for 10 steps
each and the amount of h and atp is compared to determine the effect of pH and ATP production.
We capture the gradient requirement as the following ASP code7 :
notenabled(T,TS) :T==t5, C==h, trans(T), col(C), holds(is,Qis,C,TS),
holds(mm,Qmm,C,TS), Qmm+3 > Qis,
num(Qis;Qmm), time(TS).
In the normal situation, we get the following h token distribution after 10 steps:
holds(is,11,h,10) holds(mm,1,h,10) holds(mm,6,atp,10)
we change the permeability to 1 (t6 enabled), we get the following token distribution instead:
holds(is,10,h,10) holds(mm,2,h,10) holds(mm,5,atp,10)
we change the permeability to 2 (t6, t7 enabled), the distribution changes as follows:
holds(is,8,h,10) holds(mm,4,h,10) holds(mm,2,atp,10)
7
We can alternatively model this by having a threshold arc from “is” to “t5” if only a minimum trigger quantity
is required in the intermembrane space.
CHAPTER 3. ANSWERING QUESTIONS USING PETRI NETS AND ASP
56
Note that as the permeability increases, the amount of H+ (h) in intermembrane space (is)
decreases and so does the amount of ATP (h) in mitochondrial matrix. Thus, an increase in
permeability will increase the pH. If the permeability increases even beyond 2, no ATP will be
produced from ADP due to insufficient H+ gradient. Trend from various runs is shown in Figure
3.13.
Quantity of H+ and ATP produced by various run lengths and permeabilities
normal
mm(h)
is(h)
mm(atp)
14
Amount of substance produced
low permeability
high permeability
mm(h)
is(h)
mm(atp)
14
12
12
12
10
10
10
8
8
8
6
6
6
4
4
4
2
2
2
0
0
0
4
5
6
7
8
Run length
9
10 11
4
5
6
7
8
Run length
mm(h)
is(h)
mm(atp)
14
9
10 11
4
5
6
7
8
9
10 11
Run length
Figure 3.13: Quantities of H+ and ATP at various run lengths and permeabilities for the Petri
Net model in Figure 3.12.
3.8
Comparing Altered Trajectories due to Delay Intervention
Question 7 Membranes must be fluid to function properly. How would decreased fluidity of the
membrane affect the efficiency of the electron transport chain?
Provided Answer:
“Some of the components of the electron transport chain are mobile electron carriers, which means they must be able to move within the membrane. If fluidity decreases, these movable components would be encumbered and move more slowly.
This would cause decreased efficiency of the electron transport chain.”
Solution 7 The answer deals requires background knowledge about fluidity and how it relates
to mobile carriers not presented in the source chapter. From background knowledge we find
that the higher the fluidity, higher the mobility. The electron transport chain is presented in Fig
9.15 of Campbell’s book. From background knowledge, we know that the efficiency of the electron transport chain is measured by the amount of ATP produced per NADH/FADH2. The ATP
CHAPTER 3. ANSWERING QUESTIONS USING PETRI NETS AND ASP
57
production happens due to the gradient of H+ ions across the mitochondrial membrane. The
higher the number of H+ ions in the intermembrane space, the higher would be the gradient and
the resulting efficiency. So we measure the efficiency of the chain by the amount of H+ transported to intermembrane space, assuming all other (fixed) molecules behave normally. This is
a valid assumption since H+ transported from mitochondrial matrix is directly proportional to
the amount of electrons shuttled through the non-mobile complexes and there is a linear chain
from the electron carrier to oxygen.
We model this chain using a Petri Net with durative transitions shown in Figure 3.14. Higher
the duration of transitions, lower the fluidity of the membrane. We assume that a steady supply
of NADH and H+ is available in the mitochondrial membrane. We fulfill this requirement by
having quantities in excess of what will be consumed during the simulation. We ignore FADH2
from the diagram, since it is just an alternate path to the electron chain. Using it by itself will
produce a lower number of H+ transporter to intermembrane space, but it will not change the
result. We compare the amount of H+ transported into the intermembrane space to gauge the
efficiency of the electron transport chain. More efficient the chain is, more H+ will it transport.
We model three scenarios: normal fluidity, low fluidity with transitions t3 and t4 having an
execution time of 2 and an lower fluidity with transitions t3, t4 having execution time of 4.
h:1
h:1, nadh:1
t1
e:2
q
h:1
is
e:2
t3
h:1
mm
[2]
h:1
h:1
e:2
cytc
e:2
t4
[2]
Figure 3.14: Petri Net graph relevant to question 7. “is” is the intermembrane space, “mm” is
mitochondrial matrix, “q” is ubiquinone and “cytc” is cytochrome c. Tokens are colored, e.g.
“h:1, nadh:1” specify one token of h and nadh each. Token types are “h” for H+, “nadh” for
NADH, “e” for electrons. Numbers in square brackets below the transition represent transition
durations with default of one time unit, if the number is missing.
We encode these cases in ASP with maximal firing set semantics and simulate them for 10
time steps. For the normal fluidity we get:
holds(is,27,h,10)
for low fluidity we get:
holds(is,24,h,10)
for lower fluidity we get:
holds(is,18,h,10)
CHAPTER 3. ANSWERING QUESTIONS USING PETRI NETS AND ASP
58
Note that as the fluidity decreases, so does the amount of H+ transported to intermembrane
space, pointing to lower efficiency of electron transport chain. Trend of various runs is shown
in Figure 3.15.
Quantity of H+ produced by various run lengths fluidity (simulated by transition duration)
normal
low fluidity (dur = 2)
30
30
is(h)
Amount of substance produced
lower fluidity (dur = 4)
30
is(h)
is(h)
25
25
25
20
20
20
15
15
15
10
10
10
5
5
4
5
6
7
8
Run length
9
10 11
5
4
5
6
7
8
Run length
9
10 11
4
5
6
7
8
9
10 11
Run length
Figure 3.15: Quantities of H+ produced in the intermembrane space at various run lengths and
fluidities for the Petri Net model in Figure 3.14.
3.9
Comparing Altered Trajectories due to Priority and Read Interventions
Question 8 Phosphofructokinase (PFK) is allosterically regulated by ATP. Considering the result of glycolysis, is the allosteric regulation of PFK likely to increase or decrease the rate of
activity for this enzyme?
Provided Answer:
“Considering that one of the end products of glycolysis is ATP, PFK is inhibited
when ATP is abundant and bound to the enzyme. The inhibition decreases ATP
production along this pathway.”
Solution 8 Regulation of Phosphofructokinase (PFK) is presented in Figure 9.20 of Campbell’s
book. We ignore substances upstream of Fructose 6-phosphate (F6P) by assuming they are
available in abundance. We also ignore AMP by assuming normal supply of it. We also ignore
any output of glycolysis other than ATP production since the downstream processes ultimately
produce additional ATP. Citric acid is also ignored since it is not relevant to the question at
hand. We monitor the rate of activity of PFK by the number of times it gets used for glycolysis.
CHAPTER 3. ANSWERING QUESTIONS USING PETRI NETS AND ASP
59
We model this problem as a Petri Net shown in Figure 3.16. Allosteric regulation of PFK
is modeled by a compound “pfkatp” which represents PFK’s binding with ATP to form a compound. Details of allosteric regulation are not provided in the same chapter, they are background
knowledge from external sources. Higher than normal quantity of ATP is modeled by a threshold
arc (shown with arrow-heads at both ends) with an arbitrary threshold value of 4. This number
can be increased as necessary. The output of glycolysis and down stream processes “t3” has
been set to 2 to run the simulation in a reasonable amount of time. It can be made larger as
necessary. The allosteric regulation transition “t4” has also been given a higher priority than
glycolysis transition “t3”. This way, ATP in excess will cause PFK to be converted to PFK+ATP
compound, reducing action of PFK.
We assume that F6P is available in sufficient quantity and so is PFK. This requirement is fulfilled by having more quantity than can be consumed in the simulation duration. We model both
the normal situation including transition t4 shown in dotted style and the abnormal situation
where t4 is removed.
<2>
f6p
t3
2
<1>
pfk
t4
atp
4
pfkatp
Figure 3.16: Petri Net graph relevant to question 8. “pfk” is phosphofructokinase, “f6p” is fructose 6-phosphate, “atp” is ATP and “pfkatp” is the pfk bound with atp for allosteric regulation.
Transition “t3” represents enzymic action of pfk, “t4” represents the binding of pfk with atp.
The double arrowed arc represents a threshold arc, which enables “t4” when there are at least 4
tokens available at “atp”. Numbers above transitions in angular brackets represent arc priorities.
We encode both situations in ASP with maximal firing set policy and run them for 10 time
steps. At the end of the run we compare the firing count of transition t3 for both cases. For the
normal case (with t4), we get the following results:
holds(atp,0,c,0)
fires(t3,0)
holds(atp,2,c,1)
fires(t3,1)
holds(atp,4,c,2)
fires(t4,2)
holds(atp,3,c,3)
fires(t3,3)
holds(f6p,20,c,0) holds(pfk,20,c,0) holds(pfkatp,0,c,0)
holds(f6p,19,c,1) holds(pfk,19,c,1) holds(pfkatp,0,c,1)
holds(f6p,18,c,2) holds(pfk,18,c,2) holds(pfkatp,0,c,2)
holds(f6p,18,c,3) holds(pfk,17,c,3) holds(pfkatp,1,c,3)
CHAPTER 3. ANSWERING QUESTIONS USING PETRI NETS AND ASP
60
holds(atp,5,c,4) holds(f6p,17,c,4) holds(pfk,16,c,4) holds(pfkatp,1,c,4)
fires(t4,4)
holds(atp,4,c,5) holds(f6p,17,c,5) holds(pfk,15,c,5) holds(pfkatp,2,c,5)
fires(t4,5)
holds(atp,3,c,6) holds(f6p,17,c,6) holds(pfk,14,c,6) holds(pfkatp,3,c,6)
fires(t3,6)
holds(atp,5,c,7) holds(f6p,16,c,7) holds(pfk,13,c,7) holds(pfkatp,3,c,7)
fires(t4,7)
holds(atp,4,c,8) holds(f6p,16,c,8) holds(pfk,12,c,8) holds(pfkatp,4,c,8)
fires(t4,8)
holds(atp,3,c,9) holds(f6p,16,c,9) holds(pfk,11,c,9) holds(pfkatp,5,c,9)
fires(t3,9)
holds(atp,5,c,10) holds(f6p,15,c,10) holds(pfk,10,c,10) holds(pfkatp,5,c,10)
fires(t4,10)
Note that t3 fires only when the ATP falls below our set threshold, above it PFK is converted
to PFK+ATP compound via t4. For the abnormal case (without t4) we get the following results:
holds(atp,0,c,0) holds(f6p,20,c,0) holds(pfk,20,c,0) holds(pfkatp,0,c,0)
fires(t3,0)
holds(atp,2,c,1) holds(f6p,19,c,1) holds(pfk,19,c,1) holds(pfkatp,0,c,1)
fires(t3,1)
holds(atp,4,c,2) holds(f6p,18,c,2) holds(pfk,18,c,2) holds(pfkatp,0,c,2)
fires(t3,2)
holds(atp,6,c,3) holds(f6p,17,c,3) holds(pfk,17,c,3) holds(pfkatp,0,c,3)
fires(t3,3)
holds(atp,8,c,4) holds(f6p,16,c,4) holds(pfk,16,c,4) holds(pfkatp,0,c,4)
fires(t3,4)
holds(atp,10,c,5) holds(f6p,15,c,5) holds(pfk,15,c,5) holds(pfkatp,0,c,5)
fires(t3,5)
holds(atp,12,c,6) holds(f6p,14,c,6) holds(pfk,14,c,6) holds(pfkatp,0,c,6)
fires(t3,6)
holds(atp,14,c,7) holds(f6p,13,c,7) holds(pfk,13,c,7) holds(pfkatp,0,c,7)
fires(t3,7)
holds(atp,16,c,8) holds(f6p,12,c,8) holds(pfk,12,c,8) holds(pfkatp,0,c,8)
fires(t3,8)
holds(atp,18,c,9) holds(f6p,11,c,9) holds(pfk,11,c,9) holds(pfkatp,0,c,9)
fires(t3,9)
holds(atp,20,c,10) holds(f6p,10,c,10) holds(pfk,10,c,10) holds(pfkatp,0,c,10)
fires(t3,10)
Note that when ATP is not abundant, transition t3 fires continuously, which represents the
enzymic activity that converts F6P to downstream substances. Trend of various runs is shown
in Figure 3.17.
CHAPTER 3. ANSWERING QUESTIONS USING PETRI NETS AND ASP
61
Firing count of enzymic reaction transition of PFK
normal
abnormal
14
12
Firing count
10
8
6
4
2
4
6
8
10
12
14
16
Run length
Figure 3.17: Petri Net model in Figure 3.16.
3.10
Comparing Altered Trajectories due to Automatic Conversion
Intervention
Question 9 How does the oxidation of NADH affect the rate of glycolysis?
Provided Answer:
“NADH must be oxidized back to NAD+ in order to be used in glycolysis. Without
this molecule, glycolysis cannot occur.”
Solution 9 Cellular respiration is summarized in Fig 9.6 of Campbell’s book. NAD+ is reduced
to NADH during glycolysis (see Campbell’s Fig 9.9) during the process of converting Glyceraldehyde 3-phosphate (G3P) to 1,3-Bisphosphoglycerate (BPG13). NADH is oxidized back to
NAD+ during oxidative phosphorylation by the electron transport chain (see Campbell’s Fig
9.15). We can gauge the rate of glycolysis by the amount of Pyruvate produced, which is the end
product of glycolysis. We simplify our model by abstracting glycolysis as a black-box that takes
Glucose and NAD+ as input and produces NADH and Pyruvate as output, since there is a linear
chain from Glucose to Pyruvate that depends upon the availability of NAD+. We also abstract
oxidative phosphorylation as a black-box which takes NADH as input and produces NAD+ as
output. None of the other inner workings of oxidative phosphorylation play a role in answering
the question assuming they are functioning normally. We also ignore the pyruvate oxidation and
citric acid cycle stages of cellular respiration since their end products only provide additional
raw material for oxidative phosphorylation and do not add value to answering the question.
CHAPTER 3. ANSWERING QUESTIONS USING PETRI NETS AND ASP
62
We assume a steady supply of Glucose and all other substances used in glycolysis but a
limited supply of NAD+, since it can be recycled from NADH and we want to model its impact.
We fulfill the steady supply requirement of Glucose with sufficient initial quantity in excess of
what will be consumed during our simulation interval. We also ensure that we have sufficient
initial quantity of NAD+ to maintain glycolysis as long as it can be recycled.
2
glu
gly1
2
2
nadh
2
2
nadp
ox1
pyr
Figure 3.18: Petri Net graph relevant to question 9. “glu” represents glucose, “gly1” represents
glycolysis, “pyr” represents pyruvate, “ox1” represents oxidative phosphorylation, “nadh” represents NADH and “nadp” represents NAD+. “ox1” is removed to model stoppage of oxidation
of NADH to NAD+.
Figure 3.18 is a Petri Net representation of our simplified model. Normal situation is modeled by the entire graph, where NADH is recycled back to NAD+, while the abnormal situation
is modeled by the graph with the transition ox1 (shown in dotted style) removed. We encode
both situations in ASP with the maximal firing set policy. Both situations are run for 5 steps and
the amount of pyruvate is compared to determine the difference in the rate of glycolysis.
In normal situation (with ox1 transition), unique quantities of pyruvate (pyr) are as follows:
fires(gly1,0)
fires(gly1,1)
fires(gly1,2)
fires(gly1,3)
fires(gly1,4)
holds(pyr,10,5)
while in abnormal situation (without ox1 transition), unique quantities of pyruvate are as
follows:
fires(gly1,0)
fires(gly1,1)
fires(gly1,2)
holds(pyr,6,5)
Note that the rate of glycolysis is lower when NADH is not recycled back to NAD+, as the
glycolysis stops after the initial quantity of 6 NAD+ is consumed. Also, the gly1 transition does
CHAPTER 3. ANSWERING QUESTIONS USING PETRI NETS AND ASP
63
not fire after time-step 2, indicating glycolysis has stopped. Trend of various runs is shown in
Figure 3.19.
Quantity of pyr produced by various run lengths
normal
without ox1
30
Amount of pyr produced
25
20
15
10
5
4
6
8
10
12
14
16
18
Run length
Figure 3.19: Amount of “pyr” produced by runs of various lengths of Petri Net in Figure 3.18.
It shows results for both normal situation where “nadh” is recycled to “nadp” as well as the
abnormal situation where this recycling is stopped.
3.11
Comparing Altered Trajectories due to Initial Value Intervention
Question 10 During intense exercise, can a muscle cell use fat as a concentrated source of
chemical energy? Explain.
Provided Answer:
“When oxygen is present, the fatty acid chains containing most of the energy of a
fat are oxidized and fed into the citric acid cycle and the electron transport chain.
During intense exercise, however, oxygen is scarce in muscle cells, so ATP must be
generated by glycolysis alone. A very small part of the fat molecule, the glycerol
backbone, can be oxidized via glycolysis, but the amount of energy released by this
portion is insignificant compared to that released by the fatty acid chains. (This
is why moderate exercise, staying below 70% maximum heart rate, is better for
burning fat because enough oxygen remains available to the muscles.)”
Solution 10 The process of fat consumption in glycolysis and citric acid cycle is summarized in
Fig 9.19 of Campbell’s book. Fats are digested into glycerol and fatty acids. Glycerol gets fed
into glycolysis after being converted into Gyceraldehyde 3-phosphate (G3P), while fatty acids
get fed into citric acid cycle after being broken down through beta oxidation and converted into
CHAPTER 3. ANSWERING QUESTIONS USING PETRI NETS AND ASP
64
Acetyl CoA. Campbell’s Fig 9.18 identify a junction in catabolism where aerobic respiration or
fermentation take place depending upon whether oxygen is present or not. Energy produced
at various steps is in terms of ATP produced. In order to explain whether fat can be used as
a concentrated source of chemical energy or not, we have to show the different ways of ATP
production and when they kick in.
We combine the various pieces of information collected from Fig 9.19, second paragraph
on second column of p/180, Fig 9.15, Fig 9.16 and Fig 9.18 of Campbell’s book into Figure
3.20. We model two situations when oxygen is not available in the muscle cells (at the start of a
intense exercise) and when oxygen is available in the muscle cells (after the exercise intensity is
plateaued). We then compare and contrast them on the amount of ATP produced and the reasons
for the firing sequences.
o2
gly
t2
g3p
gly6
pyr
fer1
ox1
nadh
op1
3
fats
dig
fac
box1
acoa
cac1
atp
Figure 3.20: Petri Net graph relevant to question 10. “fats” are fats, “dig” is digestion of fats,
“gly” is glycerol, “fac” is fatty acid, “g3p” is Glyceraldehyde 3-phosphate, “pyr” is pyruvate,
“o2” is oxygen, “nadh” is NADH, “acoa” is Acyl CoA, “atp” is ATP, “op1” is oxidative phosphorylation, “cac1” is citric acid cycle, “fer1” is fermentation, “ox1” is oxidation of pyruvate to
Acyl CoA and “box1” is beta oxidation.
Figure 3.20 is a petri net representation of our simplified model. Our edge labels have lower
numbers on them than the yield in Fig 9.16 of Campbell’s book but they still capture the difference in volume that would be produced due to oxidative phosphorylation vs. glycolysis. Using
exact amounts will only increase the difference of ATP production due to the two mechanisms.
We encode both situations (when oxygen is present and when it is not) in ASP with maximal
firing set policy. We run them for 10 steps. The firing sequence and the resulting yield of ATP
explain what the possible use of fat as a source of chemical energy.
At he start of intense exercise, when oxygen is in short supply:
holds(acoa,0,0) holds(atp,0,0) holds(fac,0,0) holds(fats,5,0)
holds(g3p,0,0) holds(gly,0,0) holds(nadh,0,0) holds(o2,0,0) holds(pyr,0,0)
fires(dig,0)
holds(acoa,0,1) holds(atp,0,1) holds(fac,1,1) holds(fats,4,1)
CHAPTER 3. ANSWERING QUESTIONS USING PETRI NETS AND ASP
holds(g3p,0,1) holds(gly,1,1) holds(nadh,0,1) holds(o2,0,1) holds(pyr,0,1)
fires(box1,1) fires(dig,1) fires(t2,1)
holds(acoa,1,2) holds(atp,0,2) holds(fac,1,2) holds(fats,3,2)
holds(g3p,1,2) holds(gly,1,2) holds(nadh,1,2) holds(o2,0,2) holds(pyr,0,2)
fires(box1,2) fires(cac1,2) fires(dig,2) fires(gly6,2) fires(t2,2)
holds(acoa,1,3) holds(atp,2,3) holds(fac,1,3) holds(fats,2,3)
holds(g3p,1,3) holds(gly,1,3) holds(nadh,3,3) holds(o2,0,3) holds(pyr,1,3)
fires(box1,3) fires(cac1,3) fires(dig,3) fires(fer1,3) fires(gly6,3) fires(t2,3)
holds(acoa,1,4) holds(atp,4,4) holds(fac,1,4) holds(fats,1,4)
holds(g3p,1,4) holds(gly,1,4) holds(nadh,5,4) holds(o2,0,4) holds(pyr,1,4)
fires(box1,4) fires(cac1,4) fires(dig,4) fires(fer1,4) fires(gly6,4) fires(t2,4)
holds(acoa,1,5) holds(atp,6,5) holds(fac,1,5) holds(fats,0,5)
holds(g3p,1,5) holds(gly,1,5) holds(nadh,7,5) holds(o2,0,5) holds(pyr,1,5)
fires(box1,5) fires(cac1,5) fires(fer1,5) fires(gly6,5) fires(t2,5)
holds(acoa,1,6) holds(atp,8,6) holds(fac,0,6) holds(fats,0,6)
holds(g3p,1,6) holds(gly,0,6) holds(nadh,9,6) holds(o2,0,6) holds(pyr,1,6)
fires(cac1,6) fires(fer1,6) fires(gly6,6)
holds(acoa,0,7) holds(atp,10,7) holds(fac,0,7) holds(fats,0,7)
holds(g3p,0,7) holds(gly,0,7) holds(nadh,10,7) holds(o2,0,7) holds(pyr,1,7)
fires(fer1,7)
holds(acoa,0,8) holds(atp,10,8) holds(fac,0,8) holds(fats,0,8)
holds(g3p,0,8) holds(gly,0,8) holds(nadh,10,8) holds(o2,0,8) holds(pyr,0,8)
holds(acoa,0,9) holds(atp,10,9) holds(fac,0,9) holds(fats,0,9)
holds(g3p,0,9) holds(gly,0,9) holds(nadh,10,9) holds(o2,0,9) holds(pyr,0,9)
holds(acoa,0,10) holds(atp,10,10) holds(fac,0,10) holds(fats,0,10)
holds(g3p,0,10) holds(gly,0,10) holds(nadh,10,10) holds(o2,0,10) holds(pyr,0,10)
when the exercise intensity has plateaued and oxygen is no longer in short supply:
holds(acoa,0,0) holds(atp,0,0) holds(fac,0,0) holds(fats,5,0)
holds(g3p,0,0) holds(gly,0,0) holds(nadh,0,0) holds(o2,10,0) holds(pyr,0,0)
fires(dig,0)
holds(acoa,0,1) holds(atp,0,1) holds(fac,1,1) holds(fats,4,1)
holds(g3p,0,1) holds(gly,1,1) holds(nadh,0,1) holds(o2,10,1) holds(pyr,0,1)
fires(box1,1) fires(dig,1) fires(t2,1)
holds(acoa,1,2) holds(atp,0,2) holds(fac,1,2) holds(fats,3,2)
holds(g3p,1,2) holds(gly,1,2) holds(nadh,1,2) holds(o2,10,2) holds(pyr,0,2)
fires(box1,2) fires(cac1,2) fires(dig,2) fires(gly6,2) fires(op1,2) fires(t2,2)
holds(acoa,1,3) holds(atp,5,3) holds(fac,1,3) holds(fats,2,3)
holds(g3p,1,3) holds(gly,1,3) holds(nadh,2,3) holds(o2,9,3) holds(pyr,1,3)
fires(box1,3) fires(cac1,3) fires(dig,3) fires(gly6,3) fires(op1,3)
fires(ox1,3) fires(t2,3)
holds(acoa,2,4) holds(atp,10,4) holds(fac,1,4) holds(fats,1,4)
holds(g3p,1,4) holds(gly,1,4) holds(nadh,4,4) holds(o2,8,4) holds(pyr,1,4)
fires(box1,4) fires(cac1,4) fires(dig,4) fires(gly6,4) fires(op1,4)
fires(ox1,4) fires(t2,4)
holds(acoa,3,5) holds(atp,15,5) holds(fac,1,5) holds(fats,0,5)
holds(g3p,1,5) holds(gly,1,5) holds(nadh,6,5) holds(o2,7,5) holds(pyr,1,5)
fires(box1,5) fires(cac1,5) fires(gly6,5) fires(op1,5) fires(ox1,5) fires(t2,5)
holds(acoa,4,6) holds(atp,20,6) holds(fac,0,6) holds(fats,0,6)
holds(g3p,1,6) holds(gly,0,6) holds(nadh,8,6) holds(o2,6,6) holds(pyr,1,6)
fires(cac1,6) fires(gly6,6) fires(op1,6) fires(ox1,6)
65
CHAPTER 3. ANSWERING QUESTIONS USING PETRI NETS AND ASP
66
holds(acoa,4,7) holds(atp,25,7) holds(fac,0,7) holds(fats,0,7)
holds(g3p,0,7) holds(gly,0,7) holds(nadh,9,7) holds(o2,5,7) holds(pyr,1,7)
fires(cac1,7) fires(op1,7) fires(ox1,7)
holds(acoa,4,8) holds(atp,29,8) holds(fac,0,8) holds(fats,0,8)
holds(g3p,0,8) holds(gly,0,8) holds(nadh,10,8) holds(o2,4,8) holds(pyr,0,8)
fires(cac1,8) fires(op1,8)
holds(acoa,3,9) holds(atp,33,9) holds(fac,0,9) holds(fats,0,9)
holds(g3p,0,9) holds(gly,0,9) holds(nadh,10,9) holds(o2,3,9) holds(pyr,0,9)
fires(cac1,9) fires(op1,9)
holds(acoa,2,10) holds(atp,37,10) holds(fac,0,10) holds(fats,0,10)
holds(g3p,0,10) holds(gly,0,10) holds(nadh,10,10) holds(o2,2,10) holds(pyr,0,10)
fires(cac1,10) fires(op1,10)
We see that more ATP is produced when oxygen is available. Most ATP (energy) is produced
by the oxidative phosphorylation which requires oxygen. When oxygen is not available, small
amount of energy is produced due to glycolysis of glycerol (gly). With oxygen a lot more energy
is produced, most of it due to fatty acids ( f ac). Trend of various runs is shown in Figure 3.21.
Quantity of ATP produced at various run lengths
40
without oxygen
with oxygen
35
Amount of ATP produced
30
25
20
15
10
5
5
6
7
8
9
10
Run length
Figure 3.21: Amount of “atp” produced by runs of various lengths of Petri Net in Figure 3.20.
Two situations are shown: when oxygen is in short supply and when it is abundant.
3.12
Conclusion
In this chapter we presented how to model biological systems as Petri Nets, translated them into
ASP, reasoned with them and answered questions about them. We used diagrams from Campbell’s book, background knowledge and assumptions to facilitate our modeling work. However,
source knowledge for real world applications comes from published papers, magazines and
books. This means that we have to do text extraction. In one of the following chapters we look
at some of the real applications that we have worked on in the past in collaboration with other
researchers to develop models using text extraction. But first, we look at how we use the concept
CHAPTER 3. ANSWERING QUESTIONS USING PETRI NETS AND ASP
67
of answering questions using Petri Nets to build a question answering system. We will extend
the Petri Nets even more for this.
Chapter 4
BioPathQA - A System for Modeling,
Simulating, and Querying Biological
Pathways
4.1
Introduction
In this chapter we combine the methods from Chapter 3, notions from action languages, and
ASP to build a system BioPathQA and a language to specify pathways and query them. We
show how various biological pathways are encoded in BioPathQA and how it computes answers
of queries against them.
4.2
Description of BioPathQA
Our system has the following components: (i) a pathway specification language (ii) a query
language to specify the deep reasoning question, (iii) an ASP program that encodes the pathway
model and its extensions for simulation.
Knowledge about biological pathways comes in many different forms, such as cartoon diagrams, maps with well defined syntax and semantics (e.g. Kohn’s maps Kohn et al. (2006)), and
biological publications. Similar to other technical domains, some amount of domain knowledge
is also required. Users want to collect information from disparate sources and encode it in a
pathway specification. We have developed a language to allow users to describe their pathway.
This description includes describing the substances and actions that make up the pathway, the
initial state of the substances, and how the state of the pathway changes due to the actions. An
evolution of a pathway’s state from the initial state, through a set of actions is called a trajectory.
Being a specification language targeted at biological systems, multiple actions autonomously
execute in parallel as soon as their preconditions are satisfied. The amount of parallelism is
dictated by any resource conflicts between the actions. When that occurs, only one sub-set of
the possible actions can execute, leading to multiple outcomes from that point on.
68
CHAPTER 4. THE BIOPATHQA SYSTEM
69
Questions are usually provided in natural language, which is vague. To avoid the vagaries
of natural language, we developed a language with syntax close to natural language but with
a well defined formal semantics. The query language allows a user to make changes to the
pathway through interventions, and restrict its trajectories through observations and query on
aggregate values in a trajectory, across a set of trajectories and even over two sets of trajectories.
This allows the user to compare a base case of a pathway specification with an alternate case
modified due to interventions and observations. This new feature is a major contribution of our
research.
Inspiration for our high level language comes from action languages and query languages
such as Gelfond and Lifschitz (1993). While action languages generally describe transition
systems Gelfond and Lifschitz (1998), our language describes trajectories. In addition, our language is geared towards modeling natural systems, in which actions occur autonomously Reiter
(1996) when their pre-conditions are satisfied; and we do not allow the quantities to become
negative (as the quantities represent amounts of physical entities).
Next we describe the syntax of our pathway specification language and the query language.
Following that we will describe the syntax of our language and how we encode it in ASP.
4.3
Syntax of Pathway Specification Language (BioPathQA-PL)
The alphabet of pathway specification language P consists of disjoint nonempty domain-dependent
sets A, F, L representing actions, fluents, and locations, respectively; a fixed set S of firing
styles; a fixed set K of keywords providing syntactic sugar (shown in bold face in pathway specification language below); a fixed set of punctuations {‘,0 }; and a fixed set of special constants
{‘10 , ‘∗0 , ‘max0 }; and integers.
Each fluent f ∈ F has a domain dom( f ) which is either integer or binary and specifies
the values f can take. A fluent is either simple, such as f or locational, such as f [l], where
l ∈ L. A state s is an interpretation of F that maps fluents to their values. We write s( f ) = v to
represent “ f has the value v in state s”. States are indexed, such that consecutive states si and
si+1 represent an evolution over one time step from i to i + 1 due to firing of an action set T i in
si .
We illustrate the role of various symbols in the alphabet with examples from the biological
domain. Consider the following example of a hypothetical pathway specification:
domain of sug is integer, f ac is integer, acoa is integer, h2o is integer
(4.1)
gly may execute causing sug change value by − 1, acoa change value by + 1
(4.2)
box may execute causing f ac change value by − 1, acoa change value by + 1
(4.3)
if h2o has value 1 or higher
(4.4)
inhibit box if sug has value 1 or higher
(4.5)
initially sug has value 3, f ac has value 4, acoa has value 0
(4.6)
CHAPTER 4. THE BIOPATHQA SYSTEM
70
It describes two processes glycolysis and beta-oxidation represented by actions ‘gly’ and
‘box’ in lines (4.2) and (4.3)-(4.4) respectively. Substances used by the pathway, i.e. sugar, fattyacids, acetyl-CoA, and water are represented by numeric fluents ‘sug’,‘ f ac’,‘acoa’, and ‘h2o’
respectively in line (4.1). When glycolysis occurs, it consumes 1 unit of sugar and produces 1
unit of acetyl-CoA (line (4.2). When beta-oxidation occurs, it consumes 1 unit of fatty-acids
and produces 1 unit of acetyl-CoA (line (4.3)). The inputs of glycolysis implicitly impose a
requirement that glycolysis can only occur when at least 1 unit of sugar is available. Similarly,
the input of beta-oxidation implicitly a requirement that beta-oxidation can only occur when at
least 1 unit of fatty-acids is available. Beta oxidation has an additional condition imposed on it
in line (4.4) that there must be at least 1 unit of water available. We call this a guard condition
on beta-oxidation. Line (4.5) explictly inhibits beta-oxidation when there is any sugar available;
and line (4.6) sets up the initial conditions of the pathway, i.e. Initially 3 units of each sugar, 4
units of fatty-acids are available and no acetyl-CoA is available. The words ‘{domain, is, may,
execute, causing, change, value, by, has, or, higher, inhibit, i f, initially}’ are keywords.
When locations are involved, locational fluents take place of simple fluents and our representation changes to include locations. For example:
gly may execute causing
sug atloc mm change value by − 1,
acoa atloc mm change value by + 1
represents glycolysis taking 1 unit of sugar from mitochondrial matrix (represented by ‘mm’)
and produces acetyl-CoA in the mitochondrial matrix. Here ‘atloc’ is an additional keyword.
A pathway is composed of a collection of different types of statements and clauses. We first
introduce their syntax, following that we will give intuitive definitions, and following that we
will show how they are combined together to construct a pathway specification.
Definition 46 (Fluent domain declaration statement) A fluent domain declaration statement
declares the values a fluent can take. It has the form:
domain of f is ‘integer0 |‘binary0
(4.7)
domain of f atloc l is ‘integer0 |‘binary0
(4.8)
for simple fluent “ f ”, and locational fluent “ f [l]”. Multiple domain statements are compactly
written as:
domain of f1 is ‘integer0 |‘binary0 , . . . , fn is ‘integer0 |‘binary0
(4.9)
domain of f1 atloc l1 is ‘integer |‘binary , . . . , fn atloc ln is ‘integer |‘binary
(4.10)
0
0
0
0
Binary domain is usually used for representing substances in a signaling pathway, while a
metabolic pathways take positive numeric values. Since the domain is for a physical entity,
we do not allow negative values for fluents.
Definition 47 (Guard condition) A guard condition takes one of the following forms:
f has value w or higher
(4.11)
f atloc l has value w or higher
(4.12)
CHAPTER 4. THE BIOPATHQA SYSTEM
71
f has value lower than w
(4.13)
f atloc l has value lower than w
(4.14)
f has value equal to w
(4.15)
f atloc l has value equal to w
(4.16)
f1 has value higher than f2
(4.17)
f1 atloc l1 has value higher than f2 atloc l2
(4.18)
where, each f in (4.11), (4.13), (4.15), (4.17) is a simple fluent, each f [l] in (4.12), (4.14),
(4.16), (4.18) is a locational fluent with location l, and each w ∈ N+ ∪ {0}.
Definition 48 (Effect clause) An effect clause can take one of the following forms:
f change value by e
(4.19)
f atloc l change value by e
(4.20)
where a is an action, f in (4.19) is a simple fluent, f [l] in (4.20) is a locational fluent with
location l, e ∈ N+ ∪ {∗} for integer fluents or e ∈ {1, −1, ∗} for binary fluents.
Definition 49 (May-execute statement) A may-execute statement captures the conditions for
firing an action a and its impact. It is of the form:
a may execute causing effect1 , . . . , effectm
if guard cond1 , . . . , guard condn
(4.21)
where effecti is an effect clause; and guard cond j is a guard condition clause, m > 0, and
n ≥ 0. If n = 0, the effect statement is unconditional (guarded by >) and the if is dropped.
A single may-execute statement must not have effecti , effect j with ei < 0, e j < 0 for the same
fluent; or ei > 0, e j > 0 for the same fluent.
Definition 50 (Must-execute statement) An must-execute statement captures the impact of firing of an action a that must fire when enabled (as long as it is not inhibited). It is an expression
of the form:
a normally must execute causing effect1 , . . . , effectm
if guard cond1 , . . . , guard condn
(4.22)
where effecti , and guard cond j are as in (4.21) above.
Definition 51 (Inhibit statement) An inhibit statement captures the conditions that inhibit an
action from occurring. It is an expression of the form:
inhibit a if guard cond1 , . . . , guard condn
(4.23)
where a is an action, guard condi is a guard condition clause, and n ≥ 0. if n = 0, the inhibition
of action ‘a’ is unconditional ‘if’ is dropped.
Definition 52 (Initial condition statement) An initial condition statement captures the initial
state of pathway. It is of the form:
initially f has value w
(4.24)
CHAPTER 4. THE BIOPATHQA SYSTEM
initially f atloc l has value w
72
(4.25)
where each f in (4.24) is a simple fluent, f [l] in (4.25) is a locational fluent with location l, and
each w is a non-negative integer. Multiple initial condition statements are written compactly as:
initially f1 has value w1 , . . . , fn has value wn
initially f1 atloc l1 has value w1 , . . . , fn atloc ln has value wn
Definition 53 (Duration Statement) A duration statement represents the duration of an action
that takes longer than a single time unit to execute. It is of the form:
a executes in d time units
(4.26)
where d is a positive integer representing the action duration.
Definition 54 (Stimulate Statement) A stimulate statement changes the rate of an action. It is
an expression of the form:
normally stimulate a by factor n if guard cond1 , . . . , guard condn
(4.27)
where guard condi is a condition, n > 0. When n = 0, the stimulation is unconditional and if
is dropped. A stimulation causes the effect in may-cause and must-fire multiplied by n.
Actions execute automatically when fireable, subject to the available fluent quantities.
Definition 55 (Firing Style Statement) A firing style statement specifies how many actions execute simultaneously (or action parallelism). It is of the form:
firing style S
(4.28)
where, S is either “1”, “∗”, or “max” for serial execution, interleaved execution, and maximum parallelism.
We will now give the intuitive meaning of these statements and put them into context w.r.t.
the biological domain. Though our description below uses simple fluents only, it applies to locational fluents in a obvious manner. The reason for having locational fluents at all is that they
allow a more natural pathway specification when substance locations are involved instead of
devising one’s own encoding scheme. For example, in a mitochondria, hydrogen ions (H+) appear in multiple locations (intermembrane space and mitochondrial matrix), with each location
carrying its distinct quantity separate from other locations.
Intuitively, a may-execute statement (4.21) represents an action a that may fire if all conditions ‘guard cond1 , . . . , guard condn ’ hold in the current state. When it executes, it impacts
the state as specified in effects. In biological context, action a represents a process, such as
a reaction, effects represent the inputs / ingredients of the reaction, and guard cond represent
additional preconditions necessary for the reaction to proceed. Condition (4.11) holds in a state
s if s( f ) ≥ w. It could represent an initiation concentration w of a substance f which is higher
than the quantity consumed by the reaction a. Condition (4.13) holds in a state s if s( f ) < w.
Condition (4.15) holds in a state s if s( f ) = w. Condition (4.17) holds in a state s if s( f1 ) > s( f2 )
capturing a situation where a substance gradient is required for a biological process to occur.
CHAPTER 4. THE BIOPATHQA SYSTEM
73
An example of one such process is the synthesis of ATP by ATP Synthase, which requires a H+
(Hydrogen ion) gradient across the inner mitochondrial matrix (Reece et al., 2010, Figure 9.15).
Intuitively, the effect clause (4.19) of an action describes the impact of an action on a fluent.
When an action a fires in a state s, the value of f changes according to the effect clause for
f . The value of f increases by e if e > 0, decreases by e if e < 0, or decreases by s( f ) if
e = ‘∗0 (where ‘∗0 can be interpreted as −s( f )). For a reaction a, a fluent with e < 0 represents
an ingredient consumed in quantity |e| by the reaction; a fluent with e > 0 represents a product
of the reaction in quantity e; a fluent with e = ‘∗0 represents consuming all quantity of the
substance due to the reaction. Since the fluents represent physical substances, their quantities
cannot become negative. As a result, any action that may cause a fluent quantity to go below
zero is disallowed.
Intuitively, a must-execute statement (4.22) is similar to a may-exec, except that when enabled, it preferentially fires over other actions as long as there isn’t an inhibit proposition that
will cause the action to become inactive. It captures the effect of an action that must happen
whenever enabled.
Intuitively, an inhibit statement (4.23) specifies the conditions that inhibits an action. In
a biological context, it defines inhibition of reactions, e.g., through biological feedback control. Though we could have added these conditions to may-exec, it is more intuitive to keep
them separate as inhibition conditions are usually discovered separately in a biological domain.
Including them as part of may-fire would constitute a surgery of existing knowledge bases.
Intuitively, an initial condition statement (4.24) specifies the initial values of fluents. The
collection of such propositions defines the initial state s0 of the pathway. In a biological context,
this defines the initial distribution of substances in the biological system.
Intuitively, an action duration statement (4.26) represents action durations, especially when
an action takes longer to execute. When an action a with duration d fires in state sk , it immediately decreases the values of fluents with e < 0 and e = ∗ upon execution, however, it does
not increase the value of fluents with e > 0 until time the end of its execution in state sk+d . In
a biological context the action duration captures a reaction’s duration. A reaction consumes its
ingredients immediately on firing, processes them for duration d and generates its products at
the end of this duration.
Intuitively, a stimulate statement (4.27) represents a change in the rate of an action a. The
stimulation causes the action to change its rate of consumption of its ingredients and production
of its products by a factor n. In biological context, this stimulation can be a result of an enzyme
or a stimulant’s availability, causing a reaction that normally proceeds slowly to occur faster.
Intuitively, a firing style statement (4.28) specifies the parallelism of actions. When it is “1”,
at most one action may fire, when it is “max”, the maximum allowable actions must fire, and
when it is “∗”, any subset of fireable actions may fire simultaneously. In a biological domain
the firing style allows one to model serial operations, parallel operations and maximally parallel
operations. The maximum parallelism is also useful in quickly discovering changes that occur
in a biological system.
Definition 56 (Pathway Specification) A pathway specification is composed of one or more
CHAPTER 4. THE BIOPATHQA SYSTEM
74
may-execute, must-execute, effect, inhibit, stimulate, initially, priority, duration statements, and
one firing style statement. When a duration statement is not specified for an action, it is assumed
to be 1. Any fluents for which an initial quantity is not specified are assumed to have a value of
zero.
A pathway specification is consistent if (i) there is at most one firing style, priority, duration
statement for each action a; (ii) the guard cond1 , . . . , guard condn from a may-execute or mustexecute are disjoint from any other may-execute or must-execute 1 ; (iii) locational and nonlocational fluents may not be intermixed; (iv) domain of fluents, effects, conditions and numeric
values are consistent, i.e., effects and conditions on binary fluents must be binary; and (v) the
pathway specification does not cause it to violate fluent domains by producing non-binary values
for binary fluents.
Each pathway specification D represents a collection of trajectories of the form: σ =
s0 , T 0 , s1 , . . . , sk−1 , T k−1 , sk . Each trajectory encodes an evolution of the pathway starting from
an initial state s0 , where si ’s are states, and T i ’s are sets of actions that fired in state si .
Intuitively, a trajectory starts from the initial state s0 . Each si , si+1 pair represents the state
evolution in one time step due to the action set T i . An action set T i is only executable in state si ,
if the sum of changes to fluents due to ei < 0 and ei = ∗ will not result in any of the fluents going
negative. Changes to fluents due to ei > 0 for the action set T i occur over subsequent time-steps
depending upon the durations of actions involved. Thus, the state si ( fi ) is the sum of ei > 0 for
actions of duration d that occurred d time steps before (current time step) i, i.e. a ∈ T i−d , where
the default duration d of an action is 1 if none specified.
Next we describe the semantics of the pathway specification language, which describes how
these trajectories are generated.
4.4
Semantics of Pathway Specification Language (BioPathQA-PL)
The semantics of the pathway specification language are defined in terms of the trajectories of
the domain description D. Since our pathway specification language is inspired by Petri Nets,
we use Petri Nets execution semantics to define its trajectories. However, some constructs in our
pathway language specification are not directly representable in standard Petri Nets, as a result,
we will have to extend them.
Let an arc-guard be a conjunction of guard conditions of the form (4.11)-(4.18), such that it
is wholly constructed of either locational or non-locational fluents, but not both.
We introduce a new type of Guarded-arc Petri Net in which each arc has an arc-guard
expression associated with it. Arcs with the same arc-guard are traversed when a transition
connected to them fires and the arc-guard is found to hold in the current state. The arc-guards of
arcs connected to the same transition form an exclusive set, such that only arcs corresponding
1
Note that ‘ f 1 has value 5 or higher ’ overlaps with ‘ f 1 has value 7 or higher’ and the two conditions are not
considered disjoint.
CHAPTER 4. THE BIOPATHQA SYSTEM
75
to one guard expression may fire (for one transition). This setup can lead to different outcomes
of an action2 .
The transitions in this new type of Petri Net can have the following inscriptions on them:
1. Propositional formula, specifying the executability conditions of the transition.
2. Arc-durations, represented as “dur(n)”, where n ∈ N+
3. A must-execute inscription, “must-execute(guard)”, requires that when the guard holds
in a state where this transition is enabled, it must fire, unless explicitly inhibited. The
guard has the same form as an arc-guard
4. A stimulation inscription, “stimulate(n, guard)”, applies a multiplication factor n ∈ N+
to the input and output quantities consumed and produced by the transition, when guard
hold in the current state, where guard has the same form as an arc-guard.
Certain aspects of our nets are similar to CPNs Jensen et al. (2007). However, the CPNs do
not allow our semantics of the reset arcs, or must-fire guards.
4.4.1
Guarded-Arc Petri Net
(f1<5) ∨ (f1>5)) ∧
(¬ (f1>7) ∧ ¬ (f1<3))
stimulate(3,f1>5)
dur(10)
f1
f1<5:1
f1>5:2
t1
f1<5:1
f1>5:2
f2
Figure 4.1: Example of a guarded-arc Petri Net.
Figure 4.1 shows an example of a guarded-arc Petri Net. There are two arc-guard expressions f 1 < 5 and f 1 > 5. When f 1 < 5, t1 consumes one token from place f 1 and produces
one token in place f 2. When f 1 > 5, t1’s consumption and production of the same tokens doubles. The transition t1 implicitly gets the guards for each arc represented the or-ed conditions
( f 1 < 5) ∨ ( f 1 > 5)). The arc also has two conditions inhibiting it, they are represented by the
and-ed conditions ¬( f 1 > 7) ∧ ¬( f 1 < 3), where ‘¬’ represents logical not. Transition t1 is
stimulated by factor 3 when f 1 > 5 and it has a duration of 10 time units. A transition cannot
fire even though one of its arc-guards is enabled, unless the token requirements on the arc itself
are also fulfilled, e.g. if f 1 has value 0 in the current state, even though f 1 < 5 guard is satisfied,
the transition cannot execute, because the input arc ( f 1, t1) for this guard needs 1 token.
Arcs for different guard expressions emanating / terminating at a place can further be combined into a single
conditional arc with conditional arc-weights. If none of the condition applies then the arc is assumed to be missing.
2
CHAPTER 4. THE BIOPATHQA SYSTEM
76
Definition 57 (Guard) A condition is of the form: ( f < v), ( f ≤ v), ( f > v), ( f ≥ v), ( f = v),
where f is a fluent and v either a fluent or a numeric constant. Then, a guard is a propositional
formula of conditions, with each condition treated as a proposition, subject to the restriction
that all fluents in all conditions in a guard are either locational or simple, but not both.
Definition 58 (Interpretation of a Guard) An interpretation of a guard G is a possible assignment of a value to each fleuent f ∈ G from the domain of f .
Definition 59 (Guard Satisfaction) A guard G with simple fluents is satisfied w.r.t. a state s,
written s |= G iff G has an interpretation in which each of its fluents f has the value s( f ) and G
is true. A guard G with locational fluents is satisfied w.r.r. a state s, written s |= G iff G has an
interpretation in which each of its fluents f [l] has the value m s(l) ( f ) and G is true, where mX ( f )
is the multiplicity of f in X.
Definition 60 (Guarded-Arc Petri Net) A Guarded-Arc Petri Net is a tuple
PN G = (P, T, G, E, R, W, D, B, TG, MF, L), where:
P is a finite set of places
T is a finite set of transitions
G is a set of guards as defined in definition (57)
TG : T → G are the transition guards
E ⊆ (T × P × G) ∪ (P × T × G) are the guarded arcs
R ⊆ P × T × G are the guarded reset arcs
W : E → N+ are arc weights
D : T → N+ are the transition durations
B : T → G × N+ transition stimulation or boost
MF : T → 2G must-fire guards for a transition
L : P → N+ specifies maximum number to tokens for each place
subject to constraints:
1. P ∩ T = ∅
2. R ∩ E = ∅
3. Let t1 ∈ T and t2 ∈ T be any two distinct transitions, then g1 ∈ MF(t1 ) and g2 ∈ MF(t2 )
must not have an interpretation that make both g1 and g2 true.
4. Let t ∈ T be a transition, and ggt = {g : (t, p, g) ∈ E}∪{g : (p, t, g) ∈ E}∪{g : (p, t, g) ∈ R}
be the set of arc-guards for normal and reset arcs connected to it, then g1 ∈ ggt , g2 ∈ ggt
must not have an interpretation that makes both g1 and g2 true.
CHAPTER 4. THE BIOPATHQA SYSTEM
77
5. Let t ∈ T be a transition, and ggt = {g : (g, n) ∈ B(t)} be its stimulation arc-guards, then
g1 ∈ ggt , g2 ∈ ggt must not have an interpretation that makes both g1 and g2 true.
6. Let t ∈ T be a transition, and (g, n) ∈ B(t) : n > 1, then there must not exist a place
p ∈ P : (p, t, g) ∈ E, L(p) = 1. Intuitively, stimulation of binary inputs is not supported.
We will make a simplifying assumption that all places are readable by using their place
names. Execution of the PN G occurs in discrete time steps.
Definition 61 (Marking (or State)) Marking (or State) of a Guarded-Arc Petri Net PN G is the
token assignment of each place pi ∈ P. Initial marking M0 : P → N0 , while the token assignment at step k is written as Mk .
Next we define the execution semantics of PN G . First we introduce some terminology that
will be used below. Let
1. s0 = M0 represent the the initial marking (or state), sk = Mk represent the marking (or
state) at time step k,
2. sk (p) represent the marking of place p at time step k, such that sk = [sk (p0 ), . . . , sk (pn )],
where P = {p0 , . . . , pn }
3. T k be the firing-set that fired in step k,
4. bk (t) be the stimulation value applied to a transition t w.r.t. step k
5. enk be the set of enabled transitions in state sk ,
6. m fk be the set of must-execute transitions in state sk ,
7. consumek (p, {t1 , . . . , tn }) be the sum of tokens that will be consumed from place p if transitions t1 , . . . , tn fired in state sk ,
8. overck ({t1 , . . . , tn }) be the set of places that will have over-consumption of tokens if transitions t1 , . . . , tn were to fire simultaneously in state sk ,
9. selk ( f s) be the set of possible firing-set choices in state sk using fs firing style
10. producek (p) be the total production of tokens in place p (in state sk ) due to actions terminating in state sk ,
11. sk+1 be the next state computed from current state sk due to firing of transition-set T k
12. min(a, b) gives the minimum of numbers a, b, such that min(a, b) = a if a < b or b otherwise .
CHAPTER 4. THE BIOPATHQA SYSTEM
78
Then, the execution semantics of the guarded-arc Petri Net starting from state s0 using firingstyle fs is given as follows:
n
bk (t) =
1
if (g, n) ∈ B(t), sk |= g
otherwise
enk = {t : t ∈ T, sk |= TG(t), ∀(p, t, g) ∈ E,
(sk |= g, sk (p) ≥ W(p, t, g) ∗ bk (t))}
m fk = {t : t ∈ enk , ∃g ∈ MF(t), sk |= g}
X
consumek (p, {t1 , . . . , tn }) =
W(p, ti , g) ∗ bk (t) : (p, ti , g) ∈ E, sk |= g
i=1,...,n
+
X
sk (p) : (p, ti , g) ∈ R, sk |= g
i=1,...,n
overck ({t1 , . . . , tn }) = {p : p ∈ P, sk (p) < consumek (p, {t1 , . . . , tn })}
if |m fk | = 1
m fk
selk (1) =
{{t} : t ∈ en } if |m f | < 1
k
k
selk (∗) = {ss : ss ∈ 2 , m fk ⊆ ss, overck (ss) = ∅}
enk
selk (max) = {ss : ss ∈ 2enk , m fk ⊆ ss, overck (p, ss) = ∅,
(@ss0 ∈ 2enk : ss ⊂ ss0 , m fk ⊆ ss0 , overck (ss0 ) = ∅)}
T k = T k0 : T k0 ∈ selk (fs), (@t ∈ enk \ T k0 , t is a reset transition )
X
producek (p) =
W(ti , p, g) ∗ b j (ti ) : (ti , p, g) ∈ E, ti ∈ T j , D(ti ) + j = k + 1
j=0,...,k
sk+1 (p) = min(sk (p) − consumek (p, T k ) + producek (p), L(p))
(4.29)
Definition 62 (Trajectory) σ = s0 , T 0 , s1 , . . . , sk , T k , sk+1 is a trajectory of PN G iff given s0 =
M0 , each T i is a possible firing-set in si whose firing produces si+1 per PN G ’s execution semantics in (4.29).
4.4.2
Construction of Guarded-Arc Petri Net from a Pathway Specification
Now we describe how to construct such a Petri Net from a given pathway specification D with
locational fluents. We proceed as follows:
1. The set of transitions T = {a : a is an action in D}.
2. The set of places P = { f : f is a fluent in D}.
if ‘domain of f is binary0 ∈ D
1
.
3. The limit relation for places L( f ) =
∞ otherwise
4. An arc-guard expression guard cond1 , . . . , guard condn is translated into the conjunction
(guard1 , . . . , guardn ), where guardi is obtained from guard condi as follows:
CHAPTER 4. THE BIOPATHQA SYSTEM
79
(a) A guard condition (4.11) is translated to f ≥ w
(b) A guard condition (4.13) is translated to f < w
(c) A guard condition (4.15) is translated to f = w
(d) A guard condition (4.17) is translated to f1 > f2
5. A may-execute statement (4.21) is translated into guarded arcs as follows:
(a) Let guard G be the translation of arc-guard conditions guard cond1 , . . . , guard condn
specified in the may-execute proposition.
(b) The effect clause (4.19) are translated into arcs as follows:
i. An effect clause with e < 0 is translated into an input arc ( f, a, G), with arcweight W( f, a, G) = |e|.
ii. An effect clause with e = ‘∗0 is translated into a reset set ( f, a, G) with arcweight W( f, a, G) = ∗.
iii. An effect clause with e > 0 is translated into an output arc (a, f, G), with arcweight W(a, f, G) = e.
6. A must-execute statement (4.22) is translated into guarded arcs in the same way as may
execute. In addition, it adds an arc-inscription must-exec(G), where G is the translation
of the arc-guard.
7. An inhibit statement (4.23) is translated into IG = (guard1 , . . . , guardn ), where (guard1 , . . . ,
guardn ) is the translation of (guard cond 1, . . . , guard condn )
8. An initial condition statement (4.25) sets the initial marking of a specific place p to w, i.e.
M0 (p) = w.
9. An duration statement (4.26) adds a dur(d) inscription to transition a.
10. A stimulate statement (4.27) adds a stimulate(n, G) inscription to transition a, where G is
the translation of the stimulate guard, a conjunction of guard cond1 , . . . , guard condn .
11. A guard (G1 ∨ · · · ∨ Gn ) ∧ (¬IG1 ∧ · · · ∧ ¬IGm ) is added to each transition a, where
Gi , 1 ≤ i ≤ n is a guard for a may-execute or a must-execute statement and IGi , 1 ≤ i ≤ m
is a guard for an inhibit statement.
12. A firing style statement (4.28) does not visibly appear on a Petri Net diagram, but it
specifies the transition firing regime the Petri Net follows.
CHAPTER 4. THE BIOPATHQA SYSTEM
80
Example: Consider the following pathway specification:
domain of
t1 may execute
causing
if
t1 may execute
causing
if
duration of
inhibit t1 if
inhibit t1 if
normally stimulate t1
if
initially
firing style
f1 is integer,
f2 is integer
f1 change value by − 1,
f1 has value lower than 5
f2 change value by + 1
f1 change value by − 2,
f1 has value higher than 5
t1 is 10
f1 has value higher than 7
f1 has value lower than 3
by factor 3
f2 has value higher than 5
f1 has value 0,
max
f2 change value by + 2
(4.30)
f2 has value 0,
This pathway specification is encoded as the Petri Net in figure 4.1.
4.4.3
Guarded-Arc Petri Net with Colored Tokens
Next we extend the Guarded-arc Petri Nets to add Colored tokens. We will use this extension to
model pathways with locational fluents.
f1/0
f2/0
f3/0
p1
((f3[p1]<5) ∨ (f3[p1]>5)) ∧
(¬ (f3[p1]>7) ∧ ¬ (f3[p1]<3))
stimulate(3,f3[p1]>5)
dur(10)
f3[p1]<5:f1/1
f3[p1]>5:f1/2
t1
f3[p1]<5:f2/1
f3[p1]>5:f2/2
f1/0
f2/0
f3/0
p2
Figure 4.2: Example of a guarded-arc Petri Net with colored tokens.
Figure 4.2 shows an example of a guarded-arc Petri Net with colored tokens. There are two
arc-guard expressions f 3[p1] < 5 and f 3[p1] > 5. When f 3[p1] < 5, t1 consumes one token of
color f 1 from place p1 and produces one token of color f 2 in place p2. When f 3[p1] > 5, t1’s
consumption and production of the same colored tokens doubles. The transition t1 implicitly
gets the guards for each arc represented the or-ed conditions (( f 3[p1] < 5) ∨ ( f 3[p1] > 5)).
The arc also has two conditions inhibiting it, they are represented by the and-ed conditions
¬( f 3[p1] > 7) ∧ ¬( f 3[p1] < 3), where ‘¬’ represents logical not. Transition t1 is stimulated by
factor 3 when f 3[p1] > 5 and it has a duration of 10 time units.
Definition 63 (Guarded-Arc Petri Net with Colored Tokens) A Guarded-Arc Petri Net with
Colored Tokens is a tuple PN GC = (P, T, C, G, E, R, W, D, B, TG, MF, L), such that:
P : finite set of places
CHAPTER 4. THE BIOPATHQA SYSTEM
81
T : finite set of transitions
C : finite set of colors
G : set of guards as defined in definition (57) with locational fluents
TG : T → G are the transition guards
E ⊆ (T × P × G) ∪ (P × T × G) are the guarded arcs
R ⊆ P × T × G are the guarded reset arcs
W : E → hC, mi are arc weights; each arc weight is a multiset over C
D : T → N+ are the transition durations
B : T → G × N+ transition stimulation or boost
MF : T → 2G must-fire guards for a transition
L : P × C → N+ specifies maximum number of tokens for each color in each place
subject to constraints:
1. P ∩ T = ∅
2. R ∩ E = ∅
3. Let t1 ∈ T and t2 ∈ T be any two distinct transitions, then g1 ∈ MF(t1 ) and g2 ∈ MF(t2 )
must not have an interpretation that make both g1 and g2 true.
4. Let t ∈ T be a transition, and ggt = {g : (t, p, g) ∈ E}∪{g : (p, t, g) ∈ E}∪{g : (p, t, g) ∈ R}
be the set of arc-guards for normal and reset arcs connected to it, then g1 ∈ ggt , g2 ∈ ggt
must not have an interpretation that makes both g1 and g2 true.
5. Let t ∈ T be a transition, and ggt = {g : (g, n) ∈ B(t)} be its stimulation guards, then
g1 ∈ ggt , g2 ∈ ggt must not have an interpretation that makes both g1 and g2 true.
6. Let t ∈ T be a transition, and (g, n) ∈ B(t) : n > 1, then there must not exist a place p ∈ P
and a color c ∈ C such that (p, t, g) ∈ E, L(p, c) = 1. Intuitively, stimulation of binary
inputs is not supported.
Definition 64 (Marking (or State)) Marking (or State) of a Guarded-Arc Petri Net with Colored Tokens PN GC is the colored token assignment of each place pi ∈ P. Initial marking is
written as M0 : P → hC, mi, while the token assignment at step k is written as Mk .
We make a simplifying assumption that all places are readable by using their place name.
Next we define the execution semantics of the guarded-arc Petri Net. First we introduce some
terminology that will be used below. Let
1. s0 = M0 represent the the initial marking (or state), sk = Mk represent the marking (or
state) at time step k,
CHAPTER 4. THE BIOPATHQA SYSTEM
82
2. sk (p) represent the marking of place p at time step k, such that sk = [sk (p0 ), . . . , sk (pn )],
where P = {p0 , . . . , pn }.
3. T k be the firing-set that fired in state sk ,
4. bk (t) be the stimulation value applied to transition t w.r.t. state sk ,
5. enk be the set of enabled transitions in state sk ,
6. m fk be the set of must-fire transitions in state sk ,
7. consumek (p, {t1 , . . . , tn }) be the sum of colored tokens that will be consumed from place
p if transitions t1 , . . . , tn fired in state sk ,
8. overck ({t1 , . . . , tn }) be the set of places that will have over-consumption of tokens if transitions t1 , . . . , tn were to fire simultaneously in state sk ,
9. selk ( f s) be the set of possible firing-sets in state sk using fs firing style,
10. producek (p) be the total production of tokens in place p in state sk due to actions terminating in state sk ,
11. sk+1 be the next state computed from current state sk and T k ,
12. mX (c) represents the multiplicity of c ∈ C in multiset X = hC, mi,
13. c/n represents repetition of an element c of a multi-set n-times,
14. multiplication of multiset X = hC, mi with a number n be defined in terms of multiplication of element multiplicities by n, i.e. ∀c ∈ C, mX (c) ∗ n, and
15. min(a, b) gives the minimum of numbers a, b, such that min(a, b) = a if a < b or b otherwise .
Then, the execution semantics of the guarded-arc Petri Net starting from state s0 using firingstyle fs is given as follows:
n
bk (t) =
1
if (g, n) ∈ B(t), sk |= g
otherwise
enk = {t ∈ T, sk |= TG(t), ∀(p, t, g) ∈ E,
(sk |= g, sk (p) ≥ W(p, t, g) ∗ bk (t))}
m fk = {t ∈ enk , ∃g ∈ MF(t), sk |= g}
X
consumek (p, {t1 , . . . , tn }) =
W(p, ti , g) ∗ bk (t) : (p, ti , g) ∈ E, sk |= g
i=1,...,n
+
X
sk (p) : (p, ti , g) ∈ R, sk |= g
i=1,...,n
overck ({t1 , . . . , tn }) = {p ∈ P : ∃c ∈ C, m sk (p) (c) < mconsumek (p,{t1 ,...,tn }) (c)}
CHAPTER 4. THE BIOPATHQA SYSTEM
83
if |m fk | = 1
m fk
selk (1) =
{{t} : t ∈ enk } if |m fk | < 1
n
selk (∗) = ss ∈ 2enk : m fk ⊆ ss, overck (ss) = ∅
ss ∈ 2enk : m fk ⊆ ss, overck (p, ss) = ∅,
selk (max) =
(@ss0 ∈ 2enk : ss ⊂ ss0 , m fk ⊆ ss0 , overck (ss0 ) = ∅)
T k = T k0 : T k0 ∈ selk (fs), (@t ∈ enk \ T k0 , t is a reset transition )
X
producek (p) =
W(ti , p, g) ∗ b j (ti ) : (ti , p, g) ∈ E, ti ∈ T j ), D(ti ) + j = k + 1
j=0,...,k
sk+1 = [c/n : c ∈ C,
n = min(m sk (p) (c)
− mconsumek (p,Tk ) (c)
+ m producek (p) (c), L(p, c))]
(4.31)
Definition 65 (Trajectory) σ = s0 , T 0 , s1 , . . . , sk , T k , sk+1 is a trajectory of PN GC iff given
s0 = M0 , each T i is a possible firing-set in si whose firing produces si+1 per PN GC ’s execution semantics in (4.31).
4.4.4
Construction of Guarded-Arc Petri Net with Colored Tokens from a Pathway Specification with Locational Fluents
Now we describe how to construct such a Petri Net from a given pathway specification D with
locational fluents. We proceed as follows:
1. The set of transitions T = {a : a is an action in D}.
2. The set of colors C = { f : f [l] is a fluent in D}.
3. The set of places P = {l : f [l] is a fluent in D}.
4. The limit relation for each colored token in a place
if ‘domain of f atloc l is binary0 ∈ D
1
.
L( f, c) =
∞ otherwise
5. A guard expression guard cond1 , . . . , guard condn is translated into the conjunction (guard1 , . . . ,
guardn ), where guardi is obtained from guard condi as follows:
(a) A guard condition (4.14) is translated to f [l] < w
(b) A guard condition (4.16) is translated to f [l] = w
(c) A guard condition (4.12) is translated to f [l] ≥ w
(d) A guard condition (4.18) is translated to f1 [l1 ] > f2 [l2 ]
CHAPTER 4. THE BIOPATHQA SYSTEM
84
6. A may-execute statement (4.21) is translated into guarded arcs as follows:
(a) Let guard G be the translation of guard conditions guard cond1 , . . . , guard condn
specified in the may-execute proposition.
(b) The effect clauses of the form (4.20) are grouped into input, reset and output effect
sets for an action as follows:
i. The clauses with e < 0 for the same place l are grouped together into an input
set of a requiring input from place l.
ii. The clauses with e = ‘∗0 for the same place l are grouped together into a reset
set of a requiring input from place l.
iii. The clauses with e > 0 for the same place l are grouped together into an output
set of a to place l.
(c) A group of input effect clauses effect1 , . . . , effectm , m > 0 of the form (4.20) is translated into an input arc (l, a, G), with arc-weight W(l, a, G) = w+ , where w+ is the
multi-set union of fi /|ei | in effecti , 1 ≤ i ≤ m.
(d) A group of output effect clauses effect1 , . . . , effectm , m > 0 of the form (4.20) is
translated into an output arc (a, l, G), with arc-weight W(a, l, G) = w− , where w− is
the multi-set union of fi /ei in effecti , 1 ≤ i ≤ m.
(e) A group of reset effect clauses effect1 , . . . , effectm , m > 0 of the form (4.20) is translated into a reset arc (l, a, G) with arc-weight W(l, a, G) = ∗.
7. A must-execute statement (4.22) is translated into guarded arcs in the same way as may
execute. In addition, it adds an arc-inscription must-exec(G), where G is the guard, which
is the translation of guard cond1 , . . . , guard condn .
8. An inhibit statement (4.23) is translated into IG = (guard1 , . . . , guardn ), where (guard1 , . . . , guardn )
is the translation of (guard cond 1, . . . , guard condn )
9. An initial condition statement (4.25) sets the initial marking of a specific place l for a
specific color f to w, i.e. m(M0 (l)) ( f ) = w.
10. An duration statement (4.26) adds a dur(d) inscription to transition a.
11. A stimulate statement (4.27) adds a stimulate(n, G) inscription to transition a, where
guard G is the translation of its guard expression guard cond1 , . . . , guard condn .
12. A guard (G1 ∨ · · · ∨ Gn ) ∧ (¬IG1 ∧ · · · ∧ ¬IGm ) is added to each transition a, where Gi , 1 ≤
i ≤ n is the guard for a may-execute or a must-execute proposition and IGi , 1 ≤ i ≤ m is a
guard for an inhibit proposition.
13. A firing style statement (4.28) does not visibly show on a Petri Net, but it specifies the
transition firing regime the Petri Net follows.
CHAPTER 4. THE BIOPATHQA SYSTEM
85
Example
Consider the following pathway specification:
domain of
f1 atloc l1 is integer, f2 atloc l1 is integer,
f3 atloc l1 is integer, f1 atloc l2 is integer,
f2 atloc l2 is integer, f3 atloc l2 is integer
t1 may execute causing
f1 atloc l1 change value by − 1,
f2 atloc l2 change value by + 1
if f3 atloc l1 has value lower than 5
t1 may execute causing
f1 atloc l1 change value by − 2,
f2 atloc l2 change value by + 2
if f3 atloc l1 has value higher than 5
duration of t1 is 10
inhibit t1 if f3 atloc l1 has value higher than 7
inhibit t1 if f3 atloc l1 has value lower than 3
normally stimulate t1 by factor 3
if f2 atloc l2 has value higher than 5
initially
f1 atloc l1 has value 0, f1 atloc l2 has value 0,
f2 atloc l1 has value 0, f2 atloc l2 has value 0,
f3 atloc l1 has value 0, f3 atloc l2 has value 0
firing style max
4.5
(4.32)
Syntax of Query Language (BioPathQA-QL)
The alphabet of query language Q consists of the same sets A, F, L from P representing actions,
fluents, and locations, respectively; a fixed set of reserved keywords K shown in bold in syntax
below; a fixed set {‘ :0 , ‘;0 , ‘,0 , ‘00 } of punctuations; a fixed set of {‘ <0 , ‘ >0 , ‘ =0 } of directions;
and constants. Our query language asks questions about biological entities and processes in a
biological pathway described through the pathway specification language. This is our domain
description. A query statement is composed of a query description (the quantity, observation, or
condition being sought by the question), interventions (changes to the pathway), observations
(about states and actions of the pathway), and initial setup conditions.
The query statement is evaluated against the trajectories of the pathway, generated by simulating the pathway. These trajectories are modified by the initial setup and interventions. The
CHAPTER 4. THE BIOPATHQA SYSTEM
86
resulting trajectories are then filtered to retain only those which satisfy the observations specified
in the query statement.
A query statement can take various forms: The simplest queries do not modify the pathway
and check if a specific observation is true on a trajectory or not. An observation can be a point
observation or an interval observation depending upon whether they can be evaluated w.r.t. a
point or an interval on a trajectory. More complex queries modify the pathway in various ways
and ask for comparison of an observation before and after such modification.
Following query statements about the rate of production of ‘bpg13’ illustrate the kind of
queries that can be asked from our system about the specified glycolysis pathway as given
in (Reece et al., 2010, Figure 9.9).
Determine if ‘n’ is a possible rate of production of substance ‘bpg13’:
rate of production of 0 bpg130 is n
when observed between time step 0 and time step k;
(4.33)
Determine if ‘n’ is a possible rate of production of substance ‘bpg13’ in a pathway when it
is being supplied with a limited supply of an upstream substance ‘ f 16bp’:
rate of production of 0 bpg130 is n
when observed between time step 0 and time step k;
using initial setup : set value of ‘ f 16bp0 to 5;
(4.34)
Determine if ‘n’ is a possible rate of production of substance ‘bpg13’ in a pathway when it
is being supplied with a steady state supply of an upstream substance ‘ f 16bp’ at the rate of 5
units per time-step:
rate of production of 0 bpg130 is n
when observed between time step 0 and time step k;
using initial setup : continuously supply ‘ f 16bp0 in quantity 5;
(4.35)
Determine if ‘n’ is a possible rate of production of substance ‘bpg13’ in a pathway when it
is being supplied with a steady state supply of an upstream substance ‘ f 16bp’ at the rate of 5
units per time-step and the pathway is modified to remove all quantity of the substance ’dhap’
as soon as it is produced:
rate of production of 0 bpg130 is n
when observed between time step 0 and time step k;
due to interventions : remove ‘dhap0 as soon as produced;
using initial setup : continuously supply ‘ f 16bp0 in quantity 5;
(4.36)
Determine if ‘n’ is a possible rate of production of substance ‘bpg13’ in a pathway when it
is being supplied with a steady state supply of an upstream substance ‘ f 16bp’ at the rate of 5
units per time-step and the pathway is modified to remove all quantity of the substance ’dhap’
as soon as it is produced and a non-functional pathway process / reaction named ‘t5b’:
rate of production of 0 bpg130 is n
CHAPTER 4. THE BIOPATHQA SYSTEM
87
when observed between time step 0 and time step k;
due to interventions : remove ‘dhap0 as soon as produced;
due to observations : ‘t5b0 does not occur;
using initial setup : continuously supply ‘ f 16bp0 in quantity 5;
(4.37)
Determine if ‘n’ is the average rate of production of substance ‘bpg13’ in a pathway when
it is being supplied with a steady state supply of an upstream substance ‘ f 16bp’ at the rate of 5
units per time-step and the pathway is modified to remove all quantity of the substance ’dhap’
as soon as it is produced and a non-functional pathway process / reaction named ‘t5b’:
average rate of production of 0 bpg130 is n
when observed between time step 0 and time step k;
due to interventions : remove ‘dhap0 as soon as produced;
due to observations : ‘t5b0 does not occur;
using initial setup : continuously supply ‘ f 16bp0 in quantity 5;
(4.38)
Determine if ‘d’ is the direction of change in the average rate of production of substance
‘bpg13’ with a steady state supply of an upstream pathway input when compared with a pathway
with the same steady state supply of an upstream pathway input, but in which the substance
‘dhap’ is removed from the pathway as soon as it is produced and pathway process / reaction
called ‘t5b’ is non-functional:
direction of change in average rate of production of 0 bpg130 is d
when observed between time step 0 and time step k;
comparing nominal pathway with modified pathway obtained
due to interventions : remove ‘dhap0 as soon as produced ;
due to observations : ‘t5b0 does not occur;
using initial setup : continuously supply ‘ f 16bp0 in quantity 5;
(4.39)
Queries can also be about actions, as illustrated in the following examples. Determine if
action ‘t5b’ ever occurs when there is a continuous supply of ‘ f 16bp’ is available and ‘t5a’ is
disabled:
‘t5b0 occurs ;
due to interventions : disable ‘t5a0 ;
using initial setup : continuously produce ‘ f 16bp0 in quantity 5;
(4.40)
Determine if glycolysis (‘gly0 ) gets replaced with beta-oxidation (‘box0 ) when sugar (‘sug0 )
is exhausted but fatty acids (‘ f ac0 ) are available, when starting with a fixed initial supply of
sugar and fatty acids in quantity 5:
‘gly0 switches to ‘box0 when
value of 0 sug0 is 0,
value of 0 f ac0 is higher than 0
in all trajectories;
CHAPTER 4. THE BIOPATHQA SYSTEM
88
due to observations :
‘gly0 switches to ‘box0
using initial setup :
set value of ‘sug0 to 5,
set value of ‘ f ac0 to 5;
(4.41)
Next we define various syntactic elements of a query statement, give their intuitive meaning,
and how these components fit together to form a query statement. We will define the formal
semantics in a later section. Note that some of the single-trajectory queries can be represented
as LTL formulas. However, we have chosen to keep the current representation as it is more
intuitive for our biological domain.
In the following description, fi ’s are fluents, li ’s are locations, n’s are numbers, q’s are
positive integer numbers, d is one of the directions from {<, >, =}.
Definition 66 (Point) A point is a time-step on the trajectory. It has the form:
time step ts
(4.42)
Definition 67 (Interval) An interval is a sub-sequence of time-steps on a trajectory. It has the
form:
hpointi and hpointi
(4.43)
Definition 68 (Aggregate Operator (aggop)) An aggregate operator computes an aggregate
quantity over a sequence of values. It can be one of the following:
minimum
(4.44)
maximum
(4.45)
average
(4.46)
Definition 69 (Quantitative Interval Formula) A quantitative interval formula is a formula
that is evaluated w.r.t. an interval on a trajectory for some quantity n.
rate of production of f is n
(4.47)
rate of production of f atloc l is n
(4.48)
rate of firing of a is n
(4.49)
total production of f is n
(4.50)
total production of f atloc l is n
(4.51)
Intuitively, the rate of production of a fluent f in interval si , . . . , s j on a trajectory s0 , T 0 , . . . , T k−1 , sk
is n = (s j ( f ) − si ( f ))/( j − i); rate of firing of an action a in interval si , . . . , s j is n = |{T l : a ∈
T l , i ≤ l ≤ j−1}|/( j−i); and total production of a fluent f in interval si , . . . , s j is n = s j ( f )− si ( f ).
If the given n equals the computed n, then the formula holds. The same intuition extends to locational fluents, except that fluent f is replaced by f [l], e.g. rate of production of fluent f at
location l in interval si , . . . , s j on a trajectory is n = (s j ( f [l]) − si ( f [l]))/( j − i). In biological
context, the actions represent reactions and fluents substances used in these reactions. The rate
and total production formulas are used in aggregate observations to determine if reactions are
slowing down or speeding up during various portions of a simulation.
CHAPTER 4. THE BIOPATHQA SYSTEM
89
Definition 70 (Quantitative Point Formula) A quantitative point formula is a formula that is
evaluated w.r.t. a point on a trajectory for some quantity n.
value of f is higher than n
(4.52)
value of f atloc l is higher than n
(4.53)
value of f is lower than n
(4.54)
value of f atloc l is lower than n
(4.55)
value of f is n
(4.56)
value of f atloc l is n
(4.57)
Definition 71 (Qualitative Interval Formula) A qualitative interval formula is a formula that
is evaluated w.r.t. an interval on a trajectory.
f is accumulating
(4.58)
f is accumulating atloc l
(4.59)
f is decreasing
(4.60)
f is decreasing atloc l
(4.61)
Intuitively, a fluent f is accumulating in interval si , . . . , s j on a trajectory if f ’s value monotonically increases during the interval. A fluent f is decreasing in interval si , . . . , s j on a trajectory if
f ’s value monotonically decreases during the interval. The same intuition extends to locational
fluents by replacing f with f [l].
Definition 72 (Qualitative Point Formula) A qualitative point formula is a formula that is
evaluated w.r.t. a point on a trajectory.
a occurs
(4.62)
a does not occur
(4.63)
a1 switches to a2
(4.64)
(4.65)
Intuitively, an action occurs at a point i on the trajectory if a ∈ T i ; an action does not occur at
point i if a < T i ; an action a1 switches to a2 at point i if a1 ∈ T i−1 , a2 < T i−1 , a1 < T i , a2 ∈ T i .
Definition 73 (Quantitative All Interval Formula) A quantitative all interval formula is a formula that is evaluated w.r.t. an interval on a set of trajectories σ1 , . . . , σm and corresponding
quantities r1 , . . . , rm .
rates of production of f are [r1 , . . . , rm ]
(4.66)
rates of production of f altoc l are [r1 , . . . , rm ]
(4.67)
rates of firing of f are [r1 , . . . , rm ]
(4.68)
totals of production of f are [r1 , . . . , rm ]
(4.69)
totals of production of f altoc l are [r1 , . . . , rm ]
(4.70)
Intuitively, a quantitative all interval formula holds on some interval [i, j] over a set of trajectories σ1 , . . . , σm for values [r1 , . . . , rm ] if for each r x the corresponding quantitative interval formula holds in interval [i, j] in trajectory σ x . For example, rates of production of f are [r1 , . . . , rm ]
in interval [i, j] over a set of trajectories σ1 , . . . , σm if for each x ∈ {1 . . . m}, rate of production of f is r x
in interval [i, j] in trajectory σ x .
CHAPTER 4. THE BIOPATHQA SYSTEM
90
Definition 74 (Quantitative All Point Formula) A quantitative all point formula is a formula
that is evaluated w.r.t. a point on a set of trajectories σ1 , . . . , σm and corresponding quantities
r1 , . . . , rm .
values of f are [r1 , . . . , rm ]
(4.71)
values of f atloc l are [r1 , . . . , rm ]
(4.72)
Intuitively, a quantitative all point formula holds at some point i over a set of trajectories
σ1 , . . . , σm for values [r1 , . . . , rm ] if for each r x the corresponding quantitative point formula
holds at point i in trajectory σ x . For example, values of f are [r1 , . . . , rm ] at point i over a set
of trajectories σ1 , . . . , σm if for each x ∈ {1 . . . m}, value of f is r x at point i in trajectory σ x .
Definition 75 (Quantitative Aggregate Interval Formula) A quantitative aggregate interval
formula is a formula that is evaluated w.r.t. an interval on a set of trajectories σ1 , . . . , σm and
an aggregate value r, where r is the aggregate of [r1 , . . . , rm ] using aggop.
haggopi rate of production of f is n
(4.73)
haggopi rate of production of f atloc l is n
(4.74)
haggopi rate of firing of a is n
(4.75)
haggopi total production of f is n
(4.76)
haggopi total production of f atloc l is n
(4.77)
Intuitively, a quantitative aggregate interval formula holds on some interval [i, j] over a set of
trajectories σ1 , . . . , σm for a value r if there exist [r1 , . . . , rm ] whose aggregate value per aggop
is r, such that for each r x the quantitative interval formula (corresponding to the quantitative aggregate interval formual) holds in interval [i, j] in trajectory σ x . For example, average rate of
production of f is r in interval [i, j] over a set of trajectories σ1 , . . . , σm if r = (r1 + · · · + rm )/m
and for each x ∈ {1 . . . m}, rate of production of f is r x in interval [i, j] in trajectory σ x .
Definition 76 (Quantitative Aggregate Point Formula) A quantitative aggregate point formula
is a formula that is evaluated w.r.t. a point on a set of trajectories σ1 , . . . , σm and an aggregate
value r, where r is the aggregate of [r1 , . . . , rm ] using aggop.
haggopi value of f is r
(4.78)
haggopi value of f atloc l is r
(4.79)
Intuitively, a quantitative aggregate point formula holds at some point i over a set of trajectories
σ1 , . . . , σm for a value r if there exist [r1 , . . . , rm ] whose aggregate value per aggop is r, such
that for each r x the quantitative point formula (corresponding to the quantitative aggregate point
formual) holds at point i in trajectory σ x . For example, average value of f is r at point i over a
set of trajectories σ1 , . . . , σm if r = (r1 + · · · + rm )/m and for each x ∈ {1 . . . m}, value of f is r x
at point i in trajectory σ x .
Definition 77 (Quantitative Comparative Aggregate Interval Formula) A quantitative comparative aggregate interval formula is a formula that is evaluated w.r.t. an interval over two sets
of trajectories and a direction ‘d’ such that ‘d’ relates the two sets of trajectories.
direction of change in haggopi rate of production of f is d
(4.80)
CHAPTER 4. THE BIOPATHQA SYSTEM
91
direction of change in haggopi rate of production of f atloc l is d
(4.81)
direction of change in haggopi rate of firing of a is d
(4.82)
direction of change in haggopi total production of f is d
(4.83)
direction of change in haggopi total production of f atloc l is d
(4.84)
Intuitively, a comparative quantitative aggregate interval formula compares two quantitative interval formulas over using the direction d over a given interval.
Definition 78 (Quantitative Comparative Aggregate Point Formula) A quantitative comparative aggregate point formula is a formula that is evaluated w.r.t. a point over two sets of trajectories and a direction ‘d’ such that ‘d’ relates the two sets of trajectories.
direction of change in haggopi value of f is d
(4.85)
direction of change in haggopi value of f atloc l is d
(4.86)
Intuitively, a comparative quantitative aggregate point formula compares two quantitative point
formulas over using the direction d at a given point.
Definition 79 (Simple Interval Formula) A simple interval formula takes the following forms:
hquantitative interval formulai
(4.87)
hqualitative interval formulai
(4.88)
Definition 80 (Simple Point Formula) A simple interval formula takes the following forms:
hquantitative point formulai
(4.89)
hqualitative point formulai
(4.90)
Definition 81 (Internal Observation Description) An internal observation description takes
the following form:
hsimple point formulai
(4.91)
hsimple point formulai at hpointi
(4.92)
hsimple interval formulai
(4.93)
hsimple interval formulai when observed between hintervali
(4.94)
Definition 82 (Simple Point Formula Cascade) A simple point formula cascade takes the following form:
hsimple point formulai0
after hsimple point formulai1,1 , . . . , hsimple point formulai1,n1
..
.
after hsimple point formulaiu,1 , . . . , hsimple point formulaiu,nu
(4.95)
hsimple point formulai0
when hsimple point formulai1,1 , . . . , hsimple point formulai1,n1
(4.96)
hsimple point formulai0
when hcondi
(4.97)
where u ≥ 1 and ‘cond’ is a conjunction of simple point formulas that is true in the same point
as the simple point formula.
CHAPTER 4. THE BIOPATHQA SYSTEM
92
Intuitively, the simple point formula cascade (4.95) holds if a given sequence of point formulas
hold in order in a trajectory. Intuitively, simple point formula cascade (4.96) holds if a given
point formula occurs at the same point as a set of simple point formulas in a trajectory. Note
that these formulas and many other of our single trajectory formulas can be replaced by an
LTL Manna and Pnueli (1992) formula, but we have kept this syntax as it is more relevant to the
question answering needs in the biological domain.
Definition 83 (Query Description) A query description specifies a non-comparative observation that can be made either on a trajectory or a set of trajectories.
hquantitative aggregate interval formulai when observed between hintervali
hquantitative aggregate point formulai when observed at hpointi
(4.98)
(4.99)
hquantitative all interval formulai when observed between hintervali
(4.100)
hquantitative all point formulai when observed at hpointi
(4.101)
hinternal observation descriptioni
(4.102)
hinternal observation descriptioni in all trajectories
(4.103)
hsimple point formula cascadei
(4.104)
hsimple point formula cascadei in all trajectories
(4.105)
The single trajectory observations are can be represented using LTL formulas, but we have
chosen to keep them in this form for ease of use by users from the biological domain.
Definition 84 (Comparative Query Description) A comparative query description specifies a
comparative observation that can be made w.r.t. two sets of trajectories.
hquantitative comparative aggregate interval formulai
when observed between hintervali
(4.106)
hquantitative comparative aggregate point formulai
when observed at hpointi
(4.107)
Definition 85 (Intervention) Interventions define modifications to domain descriptions.
remove f1 as soon as produced
(4.108)
remove f1 atloc l1 as soon as produced
(4.109)
disable a2
(4.110)
continuously transform f1 in quantity q1 to f2
(4.111)
continuously transform f1 atloc l1 in quantity q1 to f2 atloc l2
(4.112)
make f3 inhibit a3
(4.113)
make f3 atloc l3 inhibit a3
(4.114)
continuously supply f4 in quantity q4
(4.115)
contiunously supply f4 atloc l4 in quantity q4
(4.116)
continuously transfer f1 in quantity q1 across l1 , l2 to lower gradient
(4.117)
add delay of q1 time units in availability of f1
(4.118)
add delay of q1 time units in availability of f1 atloc l1
(4.119)
set value of f4 to q4
(4.120)
set value of f4 atloc l4 to q4
(4.121)
CHAPTER 4. THE BIOPATHQA SYSTEM
93
Intuitively, intervention (4.108) modifies the pathway such that all quantity of f1 is removed
as soon as it is produced; intervention (4.109) modifies the pathway such that all quantity of
f1 [l1 ] is removed as soon as it is produced; intervention (4.110) disables the action a2 ; intervention (4.111) modifies the pathway such that f1 gets converted to f2 at the rate of q1 units per
time-unit; intervention (4.112) modifies the pathway such that f1 [l1 ] gets converted to f2 [l2 ] at
the rate of q1 units per time-unit; intervention (4.113) modifies the pathway such that action a3
is now inhibited each time there is 1 or more units of f3 and sets value of f3 to 1 to initially
inhibit a3 ; intervention (4.114) modifies the pathway such that action a3 is now inhibited each
time there is 1 or more units of f3 [l3 ] and sets value of f3 [l3 ] to 1 to initially inhibit a3 ; intervention (4.115) modifies the pathway to continuously supply f4 at the rate of q4 units per time-unit;
intervention (4.116) modifies the pathway to continuously supply f4 [l4 ] at the rate of q4 units
per time-unit; intervention (4.117) modifies the pathway to transfer f1 [l1 ] to f1 [l2 ] in quantity q1
or back depending upon whether f1 [l1 ] is higher than f1 [l2 ] or lower; intervention (4.118) modifies the pathway to add delay of q1 time units between when f1 is produced to when it is made
available to next action; intervention (4.119) modifies the pathway to add delay of q1 time units
between when f1 [l1 ] is produced to when it is made available to next action; intervention (4.120)
modifies the pathway to set the initial value of f4 to q4 ; and intervention (4.121) modifies the
pathway to set the initial value of f4 [l4 ] to q4 .
Definition 86 (Initial Condition) An initial condition is one of the intervention (4.115), (4.116),
(4.120), (4.121) as given in definition 85.
Intuitively, initial conditions are interventions that setup fixed or continuous supply of substances participating in a pathway.
Definition 87 (Query Statement) A query statement can be of the following forms:
hquery descriptioni;
due to interventions : hinterventioni1 , . . . , hinterventioniN1 ;
due to observations : hinternal observationi1 , . . . , hinternal observationiN2 ;
using initial setup : hinitial conditioni1 , . . . , hinitial conditioniN3 ;
(4.122)
hcomparative query descriptioni;
comparing nominal pathway with modified pathway obtained
due to interventions : hinterventioni1 , . . . , hinterventioniN1 ;
due to observations : hinternal observationi1 , . . . , hinternal observationiN2 ;
using initial setup : hinitial conditioni1 , . . . , hinitial conditioniN3 ;
(4.123)
where interventions, observations, and initial setup are optional.
Intuitively, a query statement asks whether a query description holds in a pathway, perhaps
after modifying it with initial setup, interventions and observations. Intuitively, a comparative
query statement asks whether a comparative query description holds with a nominal pathway
is compared against a modified pathway, where both pathways have the same initial setup, but
only the modified pathway has been modified with interventions and observations.
CHAPTER 4. THE BIOPATHQA SYSTEM
4.6
94
Semantics of the Query Language (BioPathQA-QL)
In this section we give the semantics of our pathway specification language and the query language. The semantics of the query language is in terms of the trajectories of a domain description D that satisfy a query Q. We will present the semantics using LTL-style formulas. First, we
informally define the semantics of the query language as follows.
Let Q be a query statement of the form (4.122) with a query description U, interventions
V1 , . . . , V|V| , internal observations O1 , . . . , O|O| , and initial setup conditions I1 , . . . , I|I| . We construct a modified domain description D1 by applying I1 , . . . , I|I| andV1 , . . . , V|V| to D. We filter
the trajectories of D1 to retain only those trajectories that satisfy the observations O1 , . . . , O|O| .
Then we determine if U holds on any of the retained trajectories. If it does, then we say that D
satisfies Q.
Let Q be a comparative query statement of the form (4.123) with quantitative comparative
aggregate query description U, interventions V1 , . . . , V|V| , internal observations O1 , . . . , O|O| ,
and initial conditions I1 , . . . , I|I| . Then we evaluate Q by deriving two sub-query statements. Q0
is constructed by removing the interventions V1 , . . . , V|V| and observations O1 , . . . , O|O| from Q
and replacing the quantitative comparative aggregate query description U with the corresponding quantitative aggregate query description U 0 , Q1 is constructed by replacing the quantitative
comparative aggregate query description U with the corresponding quantitative aggregate query
description U 0 . Then D satisfies Q iff we can find d ∈ {<, >, =} s.t. n d n0 , where D satisfies Q0
for some value n and D satisfies Q1 for some value n0 .
4.6.1
An Illustrative Example
In this section, we illustrate with an example how we intuitively evaluate a comparative query
statement. In the later sections, we will give the formal semantics of query satisfaction.
Consider the following simple pathway specification:
domain of
t1 may fire causing
initially
firing style
f1 is integer,
f1 change value by − 1,
f1 has value 0,
max
f2 is integer
f2 change value by + 1
f2 has value 0,
(4.124)
Let the following specify a query statement Q:
direction of change in average rate of production of f2 is d
when observed between time step 0 and time step k;
comparing nominal pathway with modified pathway obtained
due to interventions : remove f2 as soon as produced;
using initial setup : continuously supply f1 in quantity 1;
that we want to evaluate against D using a simulation length k with maximum ntok tokens at
any place to determine ‘d’ that satisfies it.
CHAPTER 4. THE BIOPATHQA SYSTEM
95
We construct the baseline query Q0 by removing interventions and observations, and replacing the comparative aggregate quantitative query description with the corresponding aggregate
quantitative query description as follows:
average rate of production of f2 is n
when observed between time step 0 and time step k;
using initial setup : continuously supply f1 in quantity 1;
We construct the alternate query Q1 by replacing the comparative aggregate quantitative
query description with the corresponding aggregate quantitative query description as follows:
average rate of production of f2 is n0
when observed between time step 0 and time step k;
due to interventions : remove f2 as soon as produced;
using initial setup : continuously supply f1 in quantity 1;
We build a modified domain description D0 as D (continuously supply f1 in quantity 1)
based on initial conditions in Q0 . We get:
domain of
t1 may fire causing
t f1 may fire causing
initially
firing style
f1 is integer,
f1 change value by − 1,
f1 change value by + 1
f1 has value 0,
max
f2 is integer
f2 change value by + 1
f2 has value 0,
We evaluate Q0 against D0 . It results in m0 trajectories with rate of productions n j = (sk ( f2 )−
s0 ( f2 ))/k on trajectory τ j = s0 , . . . , sk , 1 ≤ j ≤ m0 using time interval [0, k]. The rate of
productions are averaged to produce n = (n1 + · · · + nm0 )/m0 .
Next, we construct the alternate domain description D1 as D0 (remove f2 as soon as produced )
based on initial conditions and interventions in Q1 . We get:
domain of
t1 may fire causing
t f1 may fire causing
t f2 may fire causing
initially
firing style
f1 is integer,
f1 change value by − 1,
f1 change value by + 1
f2 change value by ∗
f1 has value 0,
max
f2 is integer
f2 change value by + 1
f2 has value 0
We evaluate Q1 against D1 . Since there are no observations, no filtering is required. This
results in m1 trajectories, each with rate of production n0j = (sk ( f2 ) − s0 ( f2 ))/k on trajectory
τ0j = s0 , . . . , sk , 1 ≤ j ≤ m1 using time interval [0, k]. The rate of productions are averaged to
produce n0 = (n01 + · · · + n0m1 )/m1 .
Due to the simple nature of our domain description, it has only one trajectory for each of the
two domains. As a result, for any k > 1, n0 < n. Thus, D satisfies Q iff d = “ <00 .
CHAPTER 4. THE BIOPATHQA SYSTEM
96
We will now define the semantics of how a domain description D is modified according
to the interventions and initial conditions, the semantics of conditions imposed by the internal
observations. We will then formally define how Q is entailed in D.
4.6.2
Domain Transformation due to Interventions and Initial Conditions
An intervention I modifies a given domain description D, potentially resulting in a different
set of trajectories than D. We define a binary operator that transforms D by applying an
intervention I as a set of edits to D using the pathway specification language. The trajectories
of the modified domain description D0 = D I are given by the semantics of the pathway
specification language. Below, we give the intuitive impact and edits required by each of the
interventions.
Domain modification by intervention (4.108) D0 = D (remove f1 as soon as produced)
modifies the pathway by removing all existing quantity of f1 at each time step:
D0 = D +
n
tr may execute causing f1 change value by ∗
o
Domain modification by intervention (4.109) D0 = D (remove f1 atloc l1 ) modifies the
pathway by removing all existing quantity of f1 at each time step:
D0 = D +
n
tr may execute causing f1 atloc l1 change value by ∗
o
Domain modification by intervention (4.110) D0 = D (disable a2 ) modifies the pathway
such that its trajectories have a2 < T i , where i ≥ 0.
D0 = D +
n
inhibit a2
o
Domain modification by intervention (4.111) D0 = D(continuously transform f1 in quantity q1
to f2 ) where si+1 ( f1 ) decreases, and si+1 ( f2 ) increases by q1 at each time step i ≥ 0, when
si ( f1 ) ≥ q1 .
(
D0 = D +
a f1,2 may execute causing
f1 change value by − q1 ,
f2 change value by + q1
)
Domain modification by intervention (4.112) D0 = D(continuously transform f1 atloc l1
in quantity q1 to f2 atloc l2 ) where si+1 ( f1 [l1 ]) decreases, and si+1 ( f2 [l2 ]) increases by q1 at
each time step i ≥ 0, when si ( f1 [l1 ]) ≥ q1 .
(
D0 = D +
a f1,2 may execute causing
f1 atloc l1 change value by − q1 ,
f2 atloc l2 change value by + q1
)
Domain modification by intervention (4.113) D0 = D (make f3 inhibit a3 ) modifies the
pathway such that it has a3 inhibited due to f3 .
D0 = D −
+
n
initially f3 has value q ∈ D
(
inhibit a3 if
initially
o
f3 has value 1 or higher
f3 has value 1
)
CHAPTER 4. THE BIOPATHQA SYSTEM
97
Domain modification by intervention (4.114) D0 = D (make f3 atloc l3 inhibit a3 ) modifies the pathway such that it has a3 inhibited due to f3 [l3 ].
D0 = D −
+
n
initially f3 atloc l3 has value q ∈ D
(
inhibit a3 if
initially
o
f3 atloc l3 has value 1 or higher,
f3 atloc l3 has value 1
)
Domain modification by intervention (4.115) D0 = D( continuously supply f4 in quantity q4 )
modifies the pathway such that a quantity q4 of substance f4 is supplied at each time step.
D0 = D +
n
o
f4 change value by + q4
t f4 may execute causing
Domain modification by intervention (4.116) D0 = D ( continuously supply f4 atloc l4
in quantity q4 ) modifies the pathway such that a quantity q4 of substance f4 at location l4 is
supplied at each time step.
D0 = D +
n
f4 atloc l4 change value by + q4
t f4 may execute causing
o
Domain modification by intervention (4.117) D0 = D(continuously transfer f1 in quantity q1
across l1 , l2 to lower gradient) modifies the pathway such that substance represented by f1 is
transferred from location l1 to l2 or l2 to l1 depending upon whether it is at a higher quantity at
l1 or l2 .
t f1 may execute causing
if
D0 = D +
t0f1 may execute causing
if
f1
f1
f1
f1
f1
f1
atloc l1
atloc l2
atloc l1
atloc l2
atloc l1
atloc l2
change value by − q1 ,
change value by + q1
has higher value than f1 atloc l2 ,
change value by − q1 ,
change value by + q1
has higher value than f1 atloc l1 ,
Domain modification by intervention (4.118) D0 = D(add delay of q1 time units in availability of f1 )
modifies the pathway such that f1 ’s arrival is delayed by q1 time units. We create additional cases
for all actions that produce f1 , such that it goes through an additional delay action.
D =D−
+
a may execute causing
0
a may execute causing
a f1 may execute causing
a f1 executes
f1 change value by + w1 ,
effect1 , . . . , effectn
if cond1 , . . . , condm ∈ D
f10 change value by + w1 ,
effect1 , . . . , effectn ,
if cond1 , . . . , condm
f10 change value by − w1 ,
f1 change value by + w1
in q1 time units
Domain modification by intervention (4.119) D0 = D(add delay ofq1 time units in availability of f1 atloc l1 )
modifies trajectories such that f1 [l1 ]’s arrival is delayed by q1 time units. We create additional
cases for all actions that produce f1 atloc l1 , such that it goes through an additional delay action.
a may execute causing
D0 = D −
f1 atloc l1 change value by + w1 ,
effect1 , . . . , effectn
if cond1 , . . . , condm ∈ D
CHAPTER 4. THE BIOPATHQA SYSTEM
a may execute causing
+
a f1 may execute causing
a executes
f1
98
f1 atloc l10 change value by + w1 ,
effect1 , . . . , effectn
if cond1 , . . . , condn
f1 atloc l10 change value by − w1 ,
f1 atloc l1 change value by + w1
in q1 time units
Domain modification by intervention (4.120) D0 = D ( set value of f4 to q4 ) modifies the
pathway such that its trajectories have s0 ( f4 ) = q4 .
D0 = D −
n
initially f4 has value n ∈ D
n
o
+ initially f4 has value q4
o
Domain modification by intervention (4.121) D0 = D ( set value of f4 atloc l4 to q4 )
modifies the pathway such that its trajectories have s0 ( f4 l4 ) = q4 .
D0 = D −
n
initially f4 atloc l4 has value n ∈ D
n
o
+ initially f4 atloc l4 has value q4
4.6.3
o
Formula Semantics
We will now define the semantics of some common formulas that we will use in the following
sections. First we introduce the LTL-style formulas that we will be using to define the syntax.
A formula hsi , σi |= F represents that F holds at point i.
A formula {hs1i , σ1 i, . . . , hsm
i , σm i} |= F represents that F holds at point i on a set of trajectories σ1 , . . . , σm .
1
m̄
A formula {hs1i , σ1 i, . . . , hsm
i , σm i}, {h s̄i , σ̄1 i, . . . , h s̄i , σ̄m̄ i} |= F represents that F holds at
point i on two sets of trajectories σ1 , . . . , σm and {σ̄1 , . . . , σ̄m̄ }.
A formula (hsi , σi, j) |= F represents that F holds in the interval [i, j] on trajectory σ.
A formula ({hs1i , σ1 i, . . . , hsm
i , σm i}, j) |= F represents that F holds in the interval [i, j] on a
set of trajectories σ1 , . . . , σm .
1
m̄
A formula ( {hs1i , σ1 i, . . . , hsm
i , σm i}, {h s̄i , σ̄1 i, . . . , h s̄i , σ̄m̄ i} , j) |= F represents that F holds
in the interval [i, j] over two sets of trajectories {σ1 , . . . , σm } and {σ̄1 , . . . , σ̄m̄ }.
Given a domain description D with simple fluents represented by a Guarded-Arc Petri Net
as defined in definition 60. Let σ = s0 , T 0 , s1 , . . . , T k−1 , sk be its trajectory as defined in (62),
and si be a state on that trajectory. Let actions T i firing in state si be observable in si such that
T i ⊆ si .
First we define how interval formulas are satisfied on a trajectory σ, starting state si and an
ending point j:
(hsi , σi, j) |= rate of production of f is n
if n = (s j ( f ) − si ( f ))/( j − i)
(hsi , σi, j) |= rate of firing of a is n
(4.125)
CHAPTER 4. THE BIOPATHQA SYSTEM
X
if n =
99
1/( j − i)
(4.126)
i≤k≤ j,hsk ,σi|=a occurs
(hsi , σi, j) |= total production of f is n
if n = (s j ( f ) − si ( f ))
(4.127)
(hsi , σi, j) |= f is accumulating
if (@k, i ≤ k ≤ j : sk+1 ( f ) < sk ( f )) and s j ( f ) > si ( f )
(4.128)
(hsi , σi, j) |= f is decreasing
if (@k, i ≤ k ≤ j : sk+1 ( f ) > sk ( f )) and s j ( f ) < si ( f )
(4.129)
Next we define how a point formula is satisfied on a trajectory σ, in a state si :
hsi , σi |= value of f is higher than n
if si ( f ) > n
(4.130)
hsi , σi |= value of f is lower than n
if si ( f ) < n
(4.131)
hsi , σi |= value of f is n
if si ( f ) = n
(4.132)
hsi , σi |=a occurs
if a ∈ si
(4.133)
hsi , σi |=a does not occur
if a < si
(4.134)
hsi , σi |=a1 switches to a2
if a1 ∈ si−1 and a2 < si−1 and a1 < si and a2 ∈ si
(4.135)
Next we define how a quantitative all interval formula is satisfied on a set of trajectories
σ1 , . . . , σm with starting states s1i , . . . , sm
i and end point j:
({hs1i , σ1 i, . . . , hsmi , σm i}, j) |= rates of production of f are [r1 , . . . , rm ]
if (hs1i , σ1 i, j) |= rate of production of f is r1
..
.
and (hsmi , σm i, j) |= rate of production of f is rm
({hs1i , σ1 i, . . . , hsmi , σm i},
(4.136)
j) |= rates of firing of a are [r1 , . . . , rm ]
if (hs1i , σ1 i, j) |= rate of firing of a is r1
..
.
and (hsmi , σm i, j) |= rate of firing of a is rm
({hs1i , σ1 i, . . . , hsmi , σm i},
(4.137)
j) |= totals of production of f are [r1 , . . . , rm ]
if (hs1i , σ1 i, j) |= total production of f is r1
..
.
and (hsmi , σm i, j) |= total production of f is rm
(4.138)
CHAPTER 4. THE BIOPATHQA SYSTEM
100
Next we define how a quantitative all point formula is satisfied on a set of trajectories
σ1 , . . . , σm in states s1i , . . . , sm
i :
{hs1i , σ1 i, . . . , hsmi , σm i} |= values of f are [r1 , . . . , rm ]
if hs1i , σ1 i |= value of f is r1
..
.
and hsmi , σm i |= value of f is rm
(4.139)
Next we define how a quantitative aggregate interval formula is satisfied on a set of trajectories σ1 , . . . , σm with starting states s1i , . . . , sm
i and end point j:
({hs1i , σ1 i, . . . , hsmi , σm i}, j) |= average rate of production of f is r
if ∃[r1 , . . . , rm ] :
(hs1i , σ1 i, j) |= rate of production of f is r1
..
.
(hsmi , σm i, j) |= rate of production of f is rm
and r = (r1 + · · · + rm )/m
({hs1i , σ1 i, . . . , hsmi , σm i},
(4.140)
j) |= minimum rate of production of f is r
if ∃[r1 , . . . , rm ] :
(hs1i , σ1 i, j) |= rate of production of f is r1
..
.
(hsmi , σm i, j) |= rate of production of f is rm
and ∃k, 1 ≤ k ≤ m : r = rk and ∀x, 1 ≤ x ≤ m, rk ≤ r x
({hs1i , σ1 i, . . . , hsmi , σm i},
(4.141)
j) |= maximum rate of production of f is r
if ∃[r1 , . . . , rm ] :
(hs1i , σ1 i, j) |= rate of production of f is r1
..
.
(hsmi , σm i, j) |= rate of production of f is rm
and ∃k, 1 ≤ k ≤ m : r = rk and ∀x, 1 ≤ x ≤ m, rk ≥ r x
(4.142)
({hs1i , σ1 i, . . . , hsmi , σm i}, j) |= average rate of firing of f is r
iff ∃[r1 , . . . , rm ] :
(hs1i , σ1 i, j) |= rate of firing of f is r1
..
.
(hsmi , σm i, j) |= rate of firing of f is rm
and r = (r1 + · · · + rm )/m
({hs1i , σ1 i, . . . , hsmi , σm i}, j) |= minimum rate of firing of f is r
iff ∃[r1 , . . . , rm ] :
(4.143)
CHAPTER 4. THE BIOPATHQA SYSTEM
101
(hs1i , σ1 i, j) |= rate of firing of f is r1
..
.
(hsmi , σm i, j) |= rate of firing of f is rm
and ∃k, 1 ≤ k ≤ m : r = rk and ∀x, 1 ≤ x ≤ m, rk ≤ r x
({hs1i , σ1 i, . . . , hsmi , σm i},
(4.144)
j) |= maximum rate of firing of f is r
iff ∃[r1 , . . . , rm ] :
(hs1i , σ1 i, j) |= rate of firing of f is r1
..
.
(hsmi , σm i, j) |= rate of firing of f is rm
and ∃k, 1 ≤ k ≤ m : r = rk and ∀x, 1 ≤ x ≤ m, rk ≥ r x
(4.145)
({hs1i , σ1 i, . . . , hsmi , σm i}, j) |= average total production of f is r
if ∃[r1 , . . . , rm ] :
(hs1i , σ1 i, j) |= total production of f is r1
..
.
(hsmi , σm i, j) |= total production of f is rm
and r = (r1 + · · · + rm )/m
(4.146)
({hs1i , σ1 i, . . . , hsmi , σm i}, j) |= minimum total production of f is r
if ∃[r1 , . . . , rm ] :
(hs1i , σ1 i, j) |= total production of f is r1
..
.
(hsmi , σm i, j) |= total production of f is rm
and ∃k, 1 ≤ k ≤ m : r = rk and ∀x, 1 ≤ x ≤ m, rk ≤ r x
({hs1i , σ1 i, . . . , hsmi , σm i},
(4.147)
j) |= maximum total production of f is r
if ∃[r1 , . . . , rm ] :
(hs1i , σ1 i, j) |= total production of f is r1
..
.
(hsmi , σm i, j) |= total production of f is rm
and ∃k, 1 ≤ k ≤ m : r = rk and ∀x, 1 ≤ x ≤ m, rk ≥ r x
(4.148)
Next we define how a quantitative aggregate point formula is satisfied on a set of trajectories
σ1 , . . . , σm in states s1i , . . . , sm
i :
{hs1i , σ1 i, . . . , hsmi , σm i} |= average value of f is r
if ∃[r1 , . . . , rm ] :
hs1i , σ1 i |= value of f is r1
..
.
CHAPTER 4. THE BIOPATHQA SYSTEM
102
hsmi , σm i |= value of f is rm
and r = (r1 + · · · + rm )/m
{hs1i , σ1 i, . . . , hsmi , σm i}
(4.149)
|= minimum value of f is r
if ∃[r1 , . . . , rm ] :
hs1i , σ1 i |= value of f is r1
..
.
hsmi , σm i |= value of f is rm
and ∃k, 1 ≤ k ≤ m : r = rk and ∀x, 1 ≤ x ≤ m, rk ≤ r x
(4.150)
{hs1i , σ1 i, . . . , hsmi , σm i} |= maximum value of f is r
if ∃[r1 , . . . , rm ] :
hs1i , σ1 i |= value of f is r1
..
.
hsmi , σm i |= value of f is rm
and ∃k, 1 ≤ k ≤ m : r = rk and ∀x, 1 ≤ x ≤ m, rk ≥ r x
(4.151)
Next we define how a comparative quantitative aggregate interval formula is satisfied on two
sets of trajectories σ1 , . . . , σm , and σ̄1 , . . . , σ̄m̄ a starting point i and an ending point j:
n
o
{hs1i , σ1 i, . . . , hsmi , σm i}, {h s̄1i , σ̄1 i, . . . , h s̄m̄i , σ̄m̄ i} , j
|= direction of change in average rate of production of f is d
if ∃ n1 : ({hs1i , σ1 i, . . . , hsmi , σm i}, j) |= average rate of production of f is n1
and ∃ n2 : ({h s̄1i , σ̄1 i, . . . , h s̄m̄i , σ̄m̄ i}, j) |= average rate of production of f is n2
and n2 d n1
o
n
1
{hsi , σ1 i, . . . , hsmi , σm i}, {h s̄1i , σ̄1 i, . . . , h s̄m̄i , σ̄m̄ i} , j
(4.152)
|= direction of change in minimum rate of production of f is d
if ∃ n1 : ({hs1i , σ1 i, . . . , hsmi , σm i}, j) |= minimum rate of production of f is n1
and ∃ n2 : ({h s̄1i , σ̄1 i, . . . , h s̄m̄i , σ̄m̄ i}, j) |= minimum rate of production of f is n2
and n2 d n1
o
n
1
{hsi , σ1 i, . . . , hsmi , σm i}, {h s̄1i , σ̄1 i, . . . , h s̄m̄i , σ̄m̄ i} , j
(4.153)
|= direction of change in maximum rate of production of f is d
if ∃ n1 : ({hs1i , σ1 i, . . . , hsmi , σm i}, j) |= maximum rate of production of f is n1
and ∃ n2 : ({h s̄1i , σ̄1 i, . . . , h s̄m̄i , σ̄m̄ i}, j) |= maximum rate of production of f is n2
and n2 d n1
o
n
{hs1i , σ1 i, . . . , hsmi , σm i}, {h s̄1i , σ̄1 i, . . . , h s̄m̄i , σ̄m̄ i} , j
|= direction of change in average rate of firing of f is d
if ∃ n1 : ({hs1i , σ1 i, . . . , hsmi , σm i}, j) |= average rate of firing of f is n1
and ∃ n2 : ({h s̄1i , σ̄1 i, . . . , h s̄m̄i , σ̄m̄ i}, j) |= average rate of firing of f is n2
(4.154)
CHAPTER 4. THE BIOPATHQA SYSTEM
103
and n2 d n1
n
{hs1i , σ1 i, . . . , hsmi , σm i}, {h s̄1i , σ̄1 i, . . . , h s̄m̄i , σ̄m̄ i}
(4.155)
o
,j
|= direction of change in minimum rate of firing of f is d
if ∃ n1 : ({hs1i , σ1 i, . . . , hsmi , σm i}, j) |= minimum rate of firing of f is n1
and ∃ n2 : ({h s̄1i , σ̄1 i, . . . , h s̄m̄i , σ̄m̄ i}, j) |= minimum rate of firing of f is n2
and n2 d n1
o
n
1
{hsi , σ1 i, . . . , hsmi , σm i}, {h s̄1i , σ̄1 i, . . . , h s̄m̄i , σ̄m̄ i} , j
(4.156)
|= direction of change in maximum rate of firing of f is d
if ∃ n1 : ({hs1i , σ1 i, . . . , hsmi , σm i}, j) |= maximum rate of firing of f is n1
and ∃ n2 : ({h s̄1i , σ̄1 i, . . . , h s̄m̄i , σ̄m̄ i}, j) |= maximum rate of firing of f is n2
and n2 d n1
(4.157)
o
n
{hs1i , σ1 i, . . . , hsmi , σm i}, {h s̄1i , σ̄1 i, . . . , h s̄m̄i , σ̄m̄ i} , j
|= direction of change in average total production of f is d
if ∃ n1 : ({hs1i , σ1 i, . . . , hsmi , σm i}, j) |= average total production of f is n1
and ∃ n2 : ({h s̄1i , σ̄1 i, . . . , h s̄m̄i , σ̄m̄ i}, j) |= average total production of f is n2
and n2 d n1
o
n
{hs1i , σ1 i, . . . , hsmi , σm i}, {h s̄1i , σ̄1 i, . . . , h s̄m̄i , σ̄m̄ i} , j
(4.158)
|= direction of change in minimum total production of f is d
if ∃ n1 : ({hs1i , σ1 i, . . . , hsmi , σm i}, j) |= minimum total production of f is n1
and ∃ n2 : ({h s̄1i , σ̄1 i, . . . , h s̄m̄i , σ̄m̄ i}, j) |= minimum total production of f is n2
and n2 d n1
o
n
1
{hsi , σ1 i, . . . , hsmi , σm i}, {h s̄1i , σ̄1 i, . . . , h s̄m̄i , σ̄m̄ i} , j
(4.159)
|= direction of change in maximum total production of f is d
if ∃ n1 : ({hs1i , σ1 i, . . . , hsmi , σm i}, j) |= maximum total production of f is n1
and ∃ n2 : ({h s̄1i , σ̄1 i, . . . , h s̄m̄i , σ̄m̄ i}, j) |= maximum total production of f is n2
and n2 d n1
(4.160)
Next we define how a comparative quantitative aggregate point formula is satisfied on two
sets of trajectories σ1 , . . . , σm , and σ̄1 , . . . , σ̄m̄ at point i:
n
o
{hs1i , σ1 i, . . . , hsmi , σm i}, {h s̄1i , σ̄1 i, . . . , h s̄m̄i , σ̄m̄ i}
|= direction of change in average value of f is d
if ∃ n1 : {hs1i , σ1 i, . . . , hsmi , σm i} |= average value of f is n1
and ∃ n2 : {h s̄1i , σ̄1 i, . . . , h s̄m̄i , σ̄m̄ i} |= average value of f is n2
and n2 d n1
n
o
1
{hsi , σ1 i, . . . , hsmi , σm i}, {h s̄1i , σ̄1 i, . . . , h s̄m̄i , σ̄m̄ i}
|= direction of change in minimum value of f is d
(4.161)
CHAPTER 4. THE BIOPATHQA SYSTEM
104
if ∃ n1 : {hs1i , σ1 i, . . . , hsmi , σm i} |= minimum value of f is n1
and ∃ n2 : {h s̄1i , σ̄1 i, . . . , h s̄m̄i , σ̄m̄ i} |= minimum value of f is n2
and n2 d n1
n
(4.162)
{hs1i , σ1 i, . . . , hsmi , σm i}, {h s̄1i , σ̄1 i, . . . , h s̄m̄i , σ̄m̄ i}
o
|= direction of change in maximum value of f is d
if ∃ n1 : {hs1i , σ1 i, . . . , hsmi , σm i} |= maximum value of f is n1
and ∃ n2 : {h s̄1i , σ̄1 i, . . . , h s̄m̄i , σ̄m̄ i} |= maximum value of f is n2
and n2 d n1
(4.163)
Given a domain description D with simple fluents represented by a Guarded-Arc Petri Net
with Colored tokens as defined in definition 63. Let σ = s0 , T 0 , s1 , . . . , T k−1 , sk be its trajectory
as defined in definition 65, and si be a state on that trajectory. Let actions T i firing in state si be
observable in si such that T i ⊆ si . We define observation semantics using LTL below. We will
use si ( f [l]) to represent m si (l) ( f ) (multiplicity / value of f in location l) in state si .
First we define how interval formulas are satisfied on a trajectory σ, starting state si and an
ending point j:
(hsi , σi, j) |= rate of production of f atloc l is n
if n = (s j ( f [l]) − si ( f [l]))/( j − i)
(4.164)
(hsi , σi, j) |= rate of firing of a is n
X
if n =
1/( j − i)
(4.165)
i≤k≤ j,hsk ,σi|=a occurs
(hsi , σi, j) |= total production of f atloc l is n
if n = (s j ( f [l]) − si ( f [l]))
(4.166)
(hsi , σi, j) |= f is accumulating atloc l
if (@k, i ≤ k ≤ j : sk+1 ( f [l]) < sk ( f [l])) and s j ( f [l]) > si ( f [l])
(4.167)
(hsi , σi, j) |= f is decreasing atloc l
if (@k, i ≤ k ≤ j : sk+1 ( f [l]) > sk ( f [l])) and s j ( f [l]) < si ( f [l])
(4.168)
Next we define how a point formula is satisfied on a trajectory σ, in a state si :
hsi , σi |= value of f atloc l is higher than n
if si ( f [l]) > n
(4.169)
hsi , σi |= value of f atloc l is lower than n
if si ( f [l]) < n
(4.170)
hsi , σi |= value of f atloc l is n
if si ( f [l]) = n
(4.171)
hsi , σi |=a occurs
if a ∈ si
(4.172)
CHAPTER 4. THE BIOPATHQA SYSTEM
105
hsi , σi |=a does not occur
if a < si
(4.173)
hsi , σi |=a1 switches to a2
if a1 ∈ si−1 and a2 < si−1 and a1 < si and a2 ∈ si
(4.174)
Next we define how a quantitative all interval formula is satisfied on a set of trajectories
σ1 , . . . , σm with starting states s1i , . . . , sm
i and end point j:
({hs1i , σ1 i, . . . , hsmi , σm i}, j) |= rates of production of f atloc l are [r1 , . . . , rm ]
if (hs1i , σ1 i, j) |= rate of production of f atloc l is r1
..
.
and (hsmi , σm i, j) |= rate of production of f atloc l is rm
({hs1i , σ1 i, . . . , hsmi , σm i},
(4.175)
j) |= rates of firing of a are [r1 , . . . , rm ]
if (hs1i , σ1 i, j) |= rate of firing of a is r1
..
.
and (hsmi , σm i, j) |= rate of firing of a is rm
(4.176)
({hs1i , σ1 i, . . . , hsmi , σm i}, j) |= totals of production of f atloc l are [r1 , . . . , rm ]
if (hs1i , σ1 i, j) |= total production of f atloc l is r1
..
.
and (hsmi , σm i, j) |= total production of f atloc l is rm
(4.177)
Next we define how a quantitative all point formula is satisfied w.r.t. a set of trajectories
σ1 , . . . , σm in states s1i , . . . , sm
i :
{hs1i , σ1 i, . . . , hsmi , σm i} |= values of f atloc l are [r1 , . . . , rm ]
if hs1i , σ1 i |= value of f atloc l is r1
..
.
and hsmi , σm i |= value of f atloc l is rm
(4.178)
Next we define how a quantitative aggregate interval formula is satisfied w.r.t. a set of
trajectories σ1 , . . . , σm with starting states s1i , . . . , sm
i and end point j:
({hs1i , σ1 i, . . . , hsmi , σm i}, j) |= average rate of production of f atloc l is r
if ∃[r1 , . . . , rm ] :
(hs1i , σ1 i, j) |= rate of production of f atloc l is r1
..
.
(hsmi , σm i, j) |= rate of production of f atloc l is rm
and r = (r1 + · · · + rm )/m
(4.179)
CHAPTER 4. THE BIOPATHQA SYSTEM
106
({hs1i , σ1 i, . . . , hsmi , σm i}, j) |= minimum rate of production of f atloc l is r
if ∃[r1 , . . . , rm ] :
(hs1i , σ1 i, j) |= rate of production of f atloc l is r1
..
.
(hsmi , σm i, j) |= rate of production of f atloc l is rm
and ∃k, 1 ≤ k ≤ m : r = rk and ∀x, 1 ≤ x ≤ m, rk ≤ r x
({hs1i , σ1 i, . . . , hsmi , σm i},
(4.180)
j) |= maximum rate of production of f atloc l is r
if ∃[r1 , . . . , rm ] :
(hs1i , σ1 i, j) |= rate of production of f atloc l is r1
..
.
(hsmi , σm i, j) |= rate of production of f atloc l is rm
and ∃k, 1 ≤ k ≤ m : r = rk and ∀x, 1 ≤ x ≤ m, rk ≥ r x
(4.181)
({hs1i , σ1 i, . . . , hsmi , σm i}, j) |= average rate of firing of f is r
iff ∃[r1 , . . . , rm ] :
(hs1i , σ1 i, j) |= rate of firing of f is r1
..
.
(hsmi , σm i, j) |= rate of firing of f is rm
and r = (r1 + · · · + rm )/m
({hs1i , σ1 i, . . . , hsmi , σm i},
(4.182)
j) |= minimum rate of firing of f is r
iff ∃[r1 , . . . , rm ] :
(hs1i , σ1 i, j) |= rate of firing of f is r1
..
.
(hsmi , σm i, j) |= rate of firing of f is rm
and ∃k, 1 ≤ k ≤ m : r = rk and ∀x, 1 ≤ x ≤ m, rk ≤ r x
({hs1i , σ1 i, . . . , hsmi , σm i},
(4.183)
j) |= maximum rate of firing of f is r
iff ∃[r1 , . . . , rm ] :
(hs1i , σ1 i, j) |= rate of firing of f is r1
..
.
(hsmi , σm i, j) |= rate of firing of f is rm
and ∃k, 1 ≤ k ≤ m : r = rk and ∀x, 1 ≤ x ≤ m, rk ≥ r x
({hs1i , σ1 i, . . . , hsmi , σm i}, j) |= average total production of f atloc l is r
if ∃[r1 , . . . , rm ] :
(hs1i , σ1 i, j) |= total production of f atloc l is r1
..
.
(4.184)
CHAPTER 4. THE BIOPATHQA SYSTEM
107
(hsmi , σm i, j) |= total production of f atloc l is rm
and r = (r1 + · · · + rm )/m
({hs1i , σ1 i, . . . , hsmi , σm i},
(4.185)
j) |= minimum total production of f atloc l is r
if ∃[r1 , . . . , rm ] :
(hs1i , σ1 i, j) |= total production of f atloc l is r1
..
.
(hsmi , σm i, j) |= total production of f atloc l is rm
and ∃k, 1 ≤ k ≤ m : r = rk and ∀x, 1 ≤ x ≤ m, rk ≤ r x
(4.186)
({hs1i , σ1 i, . . . , hsmi , σm i}, j) |= maximum total production of f atloc l is r
if ∃[r1 , . . . , rm ] :
(hs1i , σ1 i, j) |= total production of f atloc l is r1
..
.
(hsmi , σm i, j) |= total production of f atloc l is rm
and ∃k, 1 ≤ k ≤ m : r = rk and ∀x, 1 ≤ x ≤ m, rk ≥ r x
(4.187)
Next we define how a quantitative aggregate point formula is satisfied w.r.t. a set of trajectories σ1 , . . . , σm in states s1i , . . . , sm
i :
{hs1i , σ1 i, . . . , hsmi , σm i} |= average value of f atloc l is r
if ∃[r1 , . . . , rm ] :
hs1i , σ1 i |= value of f atloc l is r1
..
.
hsmi , σm i |= value of f atloc l is rm
and r = (r1 + · · · + rm )/m
{hs1i , σ1 i, . . . , hsmi , σm i}
(4.188)
|= minimum value of f atloc l is r
if ∃[r1 , . . . , rm ] :
hs1i , σ1 i |= value of f atloc l is r1
..
.
hsmi , σm i |= value of f atloc l is rm
and ∃k, 1 ≤ k ≤ m : r = rk and ∀x, 1 ≤ x ≤ m, rk ≤ r x
(4.189)
{hs1i , σ1 i, . . . , hsmi , σm i} |= maximum value of f atloc l is r
if ∃[r1 , . . . , rm ] :
hs1i , σ1 i |= value of f atloc l is r1
..
.
hsmi , σm i |= value of f atloc l is rm
and ∃k, 1 ≤ k ≤ m : r = rk and ∀x, 1 ≤ x ≤ m, rk ≥ r x
(4.190)
CHAPTER 4. THE BIOPATHQA SYSTEM
108
Next we define how a comparative quantitative aggregate interval formula is satisfied w.r.t.
two sets of trajectories σ1 , . . . , σm , and σ̄1 , . . . , σ̄m̄ a starting point i and an ending point j:
o
n
{hs1i , σ1 i, . . . , hsmi , σm i}, {h s̄1i , σ̄1 i, . . . , h s̄m̄i , σ̄m̄ i} , j
|= direction of change in average rate of production of f atloc l is d
if ∃ n1 : ({hs1i , σ1 i, . . . , hsmi , σm i}, j) |= average rate of production of f atloc l is n1
and ∃ n2 : ({h s̄1i , σ̄1 i, . . . , h s̄m̄i , σ̄m̄ i}, j) |= average rate of production of f atloc l is n2
and n2 d n1
n
(4.191)
o
{hs1i , σ1 i, . . . , hsmi , σm i}, {h s̄1i , σ̄1 i, . . . , h s̄m̄i , σ̄m̄ i} , j
|= direction of change in minimum rate of production of f atloc l is d
if ∃ n1 : ({hs1i , σ1 i, . . . , hsmi , σm i}, j) |= minimum rate of production of f atloc l is n1
and ∃ n2 : ({h s̄1i , σ̄1 i, . . . , h s̄m̄i , σ̄m̄ i}, j) |= minimum rate of production of f atloc l is n2
and n2 d n1
n
{hs1i , σ1 i, . . . , hsmi , σm i}, {h s̄1i , σ̄1 i, . . . , h s̄m̄i , σ̄m̄ i}
(4.192)
o
,j
|= direction of change in maximum rate of production of f atloc l is d
if ∃ n1 : ({hs1i , σ1 i, . . . , hsmi , σm i}, j) |= maximum rate of production of f atloc l is n1
and ∃ n2 : ({h s̄1i , σ̄1 i, . . . , h s̄m̄i , σ̄m̄ i}, j) |= maximum rate of production of f atloc l is n2
and n2 d n1
(4.193)
o
n
{hs1i , σ1 i, . . . , hsmi , σm i}, {h s̄1i , σ̄1 i, . . . , h s̄m̄i , σ̄m̄ i} , j
|= direction of change in average rate of firing of f is d
if ∃ n1 : ({hs1i , σ1 i, . . . , hsmi , σm i}, j) |= average rate of firing of f is n1
and ∃ n2 : ({h s̄1i , σ̄1 i, . . . , h s̄m̄i , σ̄m̄ i}, j) |= average rate of firing of f is n2
and n2 d n1
o
n
{hs1i , σ1 i, . . . , hsmi , σm i}, {h s̄1i , σ̄1 i, . . . , h s̄m̄i , σ̄m̄ i} , j
(4.194)
|= direction of change in minimum rate of firing of f is d
if ∃ n1 : ({hs1i , σ1 i, . . . , hsmi , σm i}, j) |= minimum rate of firing of f is n1
and ∃ n2 : ({h s̄1i , σ̄1 i, . . . , h s̄m̄i , σ̄m̄ i}, j) |= minimum rate of firing of f is n2
and n2 d n1
o
n
{hs1i , σ1 i, . . . , hsmi , σm i}, {h s̄1i , σ̄1 i, . . . , h s̄m̄i , σ̄m̄ i} , j
(4.195)
|= direction of change in maximum rate of firing of f is d
if ∃ n1 : ({hs1i , σ1 i, . . . , hsmi , σm i}, j) |= maximum rate of firing of f is n1
and ∃ n2 : ({h s̄1i , σ̄1 i, . . . , h s̄m̄i , σ̄m̄ i}, j) |= maximum rate of firing of f is n2
and n2 d n1
o
n
{hs1i , σ1 i, . . . , hsmi , σm i}, {h s̄1i , σ̄1 i, . . . , h s̄m̄i , σ̄m̄ i} , j
|= direction of change in average total production of f atloc l is d
if ∃ n1 : ({hs1i , σ1 i, . . . , hsmi , σm i}, j) |= average total production of f atloc l is n1
(4.196)
CHAPTER 4. THE BIOPATHQA SYSTEM
109
and ∃ n2 : ({h s̄1i , σ̄1 i, . . . , h s̄m̄i , σ̄m̄ i}, j) |= average total production of f atloc l is n2
and n2 d n1
o
n
1
{hsi , σ1 i, . . . , hsmi , σm i}, {h s̄1i , σ̄1 i, . . . , h s̄m̄i , σ̄m̄ i} , j
(4.197)
|= direction of change in minimum total production of f atloc l is d
if ∃ n1 : ({hs1i , σ1 i, . . . , hsmi , σm i}, j) |= minimum total production of f atloc l is n1
and ∃ n2 : ({h s̄1i , σ̄1 i, . . . , h s̄m̄i , σ̄m̄ i}, j) |= minimum total production of f atloc l is n2
and n2 d n1
n
(4.198)
{hs1i , σ1 i, . . . , hsmi , σm i}, {h s̄1i , σ̄1 i, . . . , h s̄m̄i , σ̄m̄ i}
o
,j
|= direction of change in maximum total production of f atloc l is d
if ∃ n1 : ({hs1i , σ1 i, . . . , hsmi , σm i}, j) |= maximum total production of f atloc l is n1
and ∃ n2 : ({h s̄1i , σ̄1 i, . . . , h s̄m̄i , σ̄m̄ i}, j) |= maximum total production of f atloc l is n2
and n2 d n1
(4.199)
Next we define how a comparative quantitative aggregate point formula is satisfied w.r.t.
two sets of trajectories σ1 , . . . , σm , and σ̄1 , . . . , σ̄m̄ at point i:
n
{hs1i , σ1 i, . . . , hsmi , σm i}, {h s̄1i , σ̄1 i, . . . , h s̄m̄i , σ̄m̄ i}
o
|= direction of change in average value of f atloc l is d
if ∃ n1 : {hs1i , σ1 i, . . . , hsmi , σm i} |= average value of f atloc l is n1
and ∃ n2 : {h s̄1i , σ̄1 i, . . . , h s̄m̄i , σ̄m̄ i} |= average value of f atloc l is n2
and n2 d n1
n
{hs1i , σ1 i, . . . , hsmi , σm i}, {h s̄1i , σ̄1 i, . . . , h s̄m̄i , σ̄m̄ i}
(4.200)
o
|= direction of change in minimum value of f atloc l is d
if ∃ n1 : {hs1i , σ1 i, . . . , hsmi , σm i} |= minimum value of f atloc l is n1
and ∃ n2 : {h s̄1i , σ̄1 i, . . . , h s̄m̄i , σ̄m̄ i} |= minimum value of f atloc l is n2
and n2 d n1
n
o
{hs1i , σ1 i, . . . , hsmi , σm i}, {h s̄1i , σ̄1 i, . . . , h s̄m̄i , σ̄m̄ i}
(4.201)
|= direction of change in maximum value of f atloc l is d
if ∃ n1 : {hs1i , σ1 i, . . . , hsmi , σm i} |= maximum value of f atloc l is n1
and ∃ n2 : {h s̄1i , σ̄1 i, . . . , h s̄m̄i , σ̄m̄ i} |= maximum value of f atloc l is n2
and n2 d n1
4.6.4
(4.202)
Trajectory Filtering due to Internal Observations
The trajectories produced by the Guarded-Arc Petri Net execution are filtered to retain only the
trajectories that satisfy all internal observations in a query statement. Let σ = s0 , . . . , sk be a
trajectory as given in definition 62. Then σ satisfies an observation:
rate of production of f is n
CHAPTER 4. THE BIOPATHQA SYSTEM
if (hs0 , σi, k) |= rate of production of f is n
110
(4.203)
rate of production of f is n
if (hs0 , σi, k) |= rate of firing of a is n
(4.204)
total production of f is n
if (hs0 , σi, k) |= total production of f is n
(4.205)
f is accumulating
if (hs0 , σi, k) |= f is accumulating
(4.206)
f is decreasing
if (hs0 , σi, k) |= f is decreasing
(4.207)
rate of production of f is n
when observed between time step i and time step j
if (hsi , σi, j) |= rate of production of f is n
(4.208)
rate of production of f is n
when observed between time step i and time step j
if (hsi , σi, j) |= rate of firing of a is n
(4.209)
total production of f is n
when observed between time step i and time step j
if (hsi , σi, j) |= total production of f is n
(4.210)
f is accumulating
when observed between time step i and time step j
if (hsi , σi, j) |= f is accumulating
(4.211)
f is decreasing
when observed between time step i and time step j
if (hsi , σi, j) |= f is decreasing
(4.212)
f is higher than n
if ∃i, 0 ≤ i ≤ k : hsi , σi |= value of f is higher than n
(4.213)
f is lower than n
if ∃i, 0 ≤ i ≤ k : hsi , σi |= value of f is lower than n
(4.214)
if ∃i, 0 ≤ i ≤ k : hsi , σi |= value of f is n
(4.215)
f is n
a occurs
if ∃i, 0 ≤ i ≤ k : hsi , σi |= a occurs
(4.216)
a does not occur
if ∃i, 0 ≤ i ≤ k : hsi , σi |= a does not occur
a1 switches to a2
if ∃i, 0 ≤ i ≤ k : hsi , σi |= a1 switches to a2
f is higher than n
at time step i
(4.217)
CHAPTER 4. THE BIOPATHQA SYSTEM
if hsi , σi |= value of f is higher than n
111
(4.218)
f is lower than n
at time step i
if hsi , σi |= value of f is lower than n
(4.219)
f is n
at time step i
if hsi , σi |= value of f is n
(4.220)
a occurs
at time step i
if hsi , σi |= a occurs
(4.221)
a does not occur
at time step i
if hsi , σi |= a does not occur
(4.222)
a1 switches to a2
at time step i
if hsi , σi |= a1 switches to a2
(4.223)
Let σ = s0 , . . . , sk be a trajectory of the form (65). Then σ satisfies an observation:
rate of production of f atloc l is n
if (hs0 , σi, k) |= rate of production of f atloc l is n
(4.224)
rate of production of f atloc l is n
if (hs0 , σi, k) |= rate of firing of a is n
(4.225)
total production of f atloc l is n
if (hs0 , σi, k) |= total production of f atloc l is n
(4.226)
f is accumulating atloc l
if (hs0 , σi, k) |= f is accumulating atloc l
(4.227)
f is decreasing atloc l
if (hs0 , σi, k) |= f is decreasing atloc l
(4.228)
rate of production of f atloc l is n
when observed between time step i and time step j
if (hsi , σi, j) |= rate of production of f atloc l is n
(4.229)
rate of production of f atloc l is n
when observed between time step i and time step j
if (hsi , σi, j) |= rate of firing of a is n
(4.230)
total production of f atloc l is n
when observed between time step i and time step j
if (hsi , σi, j) |= total production of f atloc l is n
f is accumulating atloc l
when observed between time step i and time step j
(4.231)
CHAPTER 4. THE BIOPATHQA SYSTEM
if (hsi , σi, j) |= f is accumulating atloc l
112
(4.232)
f is decreasing atloc l
when observed between time step i and time step j
if (hsi , σi, j) |= f is decreasing atloc l
(4.233)
f atloc l is higher than n
if ∃i, 0 ≤ i ≤ k : hsi , σi |= value of f atloc l is higher than n
(4.234)
f atloc l is lower than n
if ∃i, 0 ≤ i ≤ k : hsi , σi |= value of f atloc l is lower than n
(4.235)
f atloc l is n
if ∃i, 0 ≤ i ≤ k : hsi , σi |= value of f atloc l is n
(4.236)
a occurs
if ∃i, 0 ≤ i ≤ k : hsi , σi |= a occurs
(4.237)
a does not occur
if ∃i, 0 ≤ i ≤ k : hsi , σi |= a does not occur
(4.238)
a1 switches to a2
if ∃i, 0 ≤ i ≤ k : hsi , σi |= a1 switches to a2
f atloc l is higher than n
at time step i
if hsi , σi |= value of f atloc l is higher than n
(4.239)
f atloc l is lower than n
at time step i
if hsi , σi |= value of f atloc l is lower than n
(4.240)
f atloc l is n
at time step i
if hsi , σi |= value of f atloc l is n
(4.241)
a occurs
at time step i
if hsi , σi |= a occurs
(4.242)
a does not occur
at time step i
if hsi , σi |= a does not occur
(4.243)
a1 switches to a2
at time step i
if hsi , σi |= a1 switches to a2
(4.244)
A trajectory σ is kept for further processing w.r.t. a set of internal observations
hinternal observationi1 , . . . , hinternal observationin if σ |= hinternal observationii , 1 ≤ i ≤ n.
CHAPTER 4. THE BIOPATHQA SYSTEM
4.6.5
113
Query Description Satisfaction
Now, we define query statement semantics using LTL syntax. Let D be a domain description
with simple fluents and σ = s0 , . . . , sk be its trajectory of length k as defined in (62). Let
σ1 , . . . , σm represent the set of trajectories of D filtered by observations as necessary, with each
trajectory has the form σi = si0 , . . . , sik 1 ≤ i ≤ m. Let D̄ be a modified domain description
and σ̄1 , . . . , σ̄m̄ be its trajectories of the form σ̄i = s̄i0 , . . . , s̄ik . Then we define query satisfaction
using the formula satisfaction in section 4.6.3 as follows.
Two sets of trajectories σ1 , . . . , σm and σ̄1 , . . . , σ̄m̄ satisfy a comparative query description
based on formula satisfaction of section 4.6.3 as follows:
n
{hs10 , σ1 i, . . . , hsm0 , σm i}, {h s̄10 , σ̄1 i, . . . , h s̄m̄0 , σ̄m̄ i}
o
|= direction of change in aggop rate of production of f is d
when observed between time step i and time step j
o
n
1
if {hsi , σ1 i, . . . , hsmi , σm i}, {h s̄1i , σ̄1 i, . . . , h s̄m̄i , σ̄m̄ i} , j
|= direction of change in aggop rate of production of f is d
n
o
1
{hs0 , σ1 i, . . . , hsm0 , σm i}, {h s̄10 , σ̄1 i, . . . , h s̄m̄0 , σ̄m̄ i}
(4.245)
|= direction of change in aggop rate of firing of f is d
if
n
when observed between time step i and time step j
o
,j
{hs1i , σ1 i, . . . , hsmi , σm i}, {h s̄1i , σ̄1 i, . . . , h s̄m̄i , σ̄m̄ i}
|= direction of change in aggop rate of firing of f is d
n
o
1
{hs0 , σ1 i, . . . , hsm0 , σm i}, {h s̄10 , σ̄1 i, . . . , h s̄m̄0 , σ̄m̄ i}
(4.246)
|= direction of change in aggop total production of f is d
when observed between time step i and time step j
o
if {hs1i , σ1 i, . . . , hsmi , σm i}, {h s̄1i , σ̄1 i, . . . , h s̄m̄i , σ̄m̄ i} , j
n
|= direction of change in aggop total production of f is d
n
o
1
{hs0 , σ1 i, . . . , hsm0 , σm i}, {h s̄10 , σ̄1 i, . . . , h s̄m̄0 , σ̄m̄ i}
(4.247)
|= direction of change in aggop value of f is d
when observed at time step i
if
n
{hs1i , σ1 i, . . . , hsmi , σm i}, {h s̄1i , σ̄1 i, . . . , h s̄m̄i , σ̄m̄ i}
o
|= direction of change in aggop value of f is d
(4.248)
A set of trajectories σ1 , . . . , σm , each of length k satisfies a quantitative or a qualitative
interval query description based on formula satisfaction of section 4.6.3 as follows:
{hs10 , σ1 i, . . . , hsm0 , σm i} |= rate of production of f is n
if ∃x, 1 ≤ x ≤ m : (hs0x , σ x i, k) |= rate of production of f is n
{hs10 , σ1 i, . . . , hsm0 , σm i}
|= rate of firing of a is n
if ∃x, 1 ≤ x ≤ m : (hs0x , σ x i, k) |= rate of firing of a is n
{hs10 , σ1 i, . . . , hsm0 , σm i}
(4.249)
(4.250)
|= total production of f is n
if ∃x, 1 ≤ x ≤ m : (hs0x , σ x i, k) |= total production of f is n
(4.251)
CHAPTER 4. THE BIOPATHQA SYSTEM
114
{hs10 , σ1 i, . . . , hsm0 , σm i} |= rate of production of f is n
when observed between time step i and time step j
if ∃x, 1 ≤ x ≤ m : (hsix , σ x i, j) |= rate of production of f is n
{hs10 , σ1 i, . . . , hsm0 , σm i}
(4.252)
|= rate of firing of a is n
when observed between time step i and time step j
if ∃x, 1 ≤ x ≤ m : (hsix , σ x i, j) |= rate of firing of a is n
{hs10 , σ1 i, . . . , hsm0 , σm i}
(4.253)
|= total production of f is n
when observed between time step i and time step j
if ∃x, 1 ≤ x ≤ m : (hsix , σ x i, j) |= total production of f is n
(4.254)
{hs10 , σ1 i, . . . , hsm0 , σm i} |= f is accumulating
if ∃x, 1 ≤ x ≤ m : (hs0x , σ x i, k) |= f is accumulating
{hs10 , σ1 i, . . . , hsm0 , σm i}
(4.255)
|= f is decreasing
if ∃x, 1 ≤ x ≤ m : (hs0x , σ x i, k) |= f is decreasing
(4.256)
{hs10 , σ1 i, . . . , hsm0 , σm i} |= f is accumulating
when observed between time step i and time step j
if ∃x, 1 ≤ x ≤ m : (hsix , σ x i, j) |= f is accumulating
{hs10 , σ1 i, . . . , hsm0 , σm i}
(4.257)
|= f is decreasing
when observed between time step i and time step j
if ∃x, 1 ≤ x ≤ m : (hsix , σ x i, j) |= f is decreasing
(4.258)
{hs10 , σ1 i, . . . , hsm0 , σm i} |= rate of production of f is n
in all trajectories
if ∀x, 1 ≤ x ≤ m : (hs0x , σ x i, k) |= rate of production of f is n
{hs10 , σ1 i, . . . , hsm0 , σm i}
(4.259)
|= rate of firing of a is n
in all trajectories
if ∀x, 1 ≤ x ≤ m : (hs0x , σ x i, k) |= rate of firing of a is n
{hs10 , σ1 i, . . . , hsm0 , σm i}
(4.260)
|= total production of f is n
in all trajectories
if ∀x, 1 ≤ x ≤ m : (hs0x , σ x i, k) |= total production of f is n
(4.261)
{hs10 , σ1 i, . . . , hsm0 , σm i} |= rate of production of f is n
when observed between time step i and time step j
in all trajectories
if ∀x, 1 ≤ x ≤ m : (hsix , σ x i, j) |= rate of production of f is n
{hs10 , σ1 i, . . . , hsm0 , σm i}
|= rate of firing of a is n
when observed between time step i and time step j
in all trajectories
(4.262)
CHAPTER 4. THE BIOPATHQA SYSTEM
if ∀x, 1 ≤ x ≤ m : (hsix , σ x i, j) |= rate of firing of a is n
{hs10 , σ1 i, . . . , hsm0 , σm i}
115
(4.263)
|= total production of f is n
when observed between time step i and time step j
in all trajectories
if ∀x, 1 ≤ x ≤ m : (hsix , σ x i, j) |= total production of f is n
(4.264)
{hs10 , σ1 i, . . . , hsm0 , σm i} |= f is accumulating
in all trajectories
if ∀x, 1 ≤ x ≤ m : (hs0x , σ x i, k) |= f is accumulating
{hs10 , σ1 i, . . . , hsm0 , σm i}
(4.265)
|= f is decreasing
in all trajectories
if ∀x, 1 ≤ x ≤ m : (hs0x , σ x i, k) |= f is decreasing
(4.266)
{hs10 , σ1 i, . . . , hsm0 , σm i} |= f is accumulating
when observed between time step i and time step j
in all trajectories
if ∀x, 1 ≤ x ≤ m : (hsix , σ x i, j) |= f is accumulating
{hs10 , σ1 i, . . . , hsm0 , σm i}
(4.267)
|= f is decreasing
when observed between time step i and time step j
in all trajectories
if ∀x, 1 ≤ x ≤ m : (hsix , σ x i, j) |= f is decreasing
(4.268)
A set of trajectories σ1 , . . . , σm , each of length k satisfies a quantitative or a qualitative point
query description based on formula satisfaction of section 4.6.3 as follows:
{hs10 , σ1 i, . . . , hsm0 , σm i} |= value of f is higher than n
if ∃x∃i, 1 ≤ x ≤ m, 0 ≤ i ≤ k : hsix , σ x i |= value of f is higher than n
(4.269)
{hs10 , σ1 i, . . . , hsm0 , σm i} |= value of f is lower than n
if ∃x∃i, 1 ≤ x ≤ m, 0 ≤ i ≤ k : hsix , σ x i |= value of f is lower than n
{hs10 , σ1 i, . . . , hsm0 , σm i}
(4.270)
|= value of f is n
if ∃x∃i, 1 ≤ x ≤ m, 0 ≤ i ≤ k : hsix , σ x i |= value of f is n
(4.271)
{hs10 , σ1 i, . . . , hsm0 , σm i} |= value of f is higher than n
at time step i
if ∃x, 1 ≤ x ≤ m : hsix , σ x i |= value of f is higher than n
{hs10 , σ1 i, . . . , hsm0 , σm i}
(4.272)
|= value of f is lower than n
at time step i
if ∃x, 1 ≤ x ≤ m : hsix , σ x i |= value of f is lower than n
{hs10 , σ1 i, . . . , hsm0 , σm i}
(4.273)
|= value of f is n
at time step i
if ∃x, 1 ≤ x ≤ m : hsix , σ x i |= value of f is n
(4.274)
CHAPTER 4. THE BIOPATHQA SYSTEM
116
{hs10 , σ1 i, . . . , hsm0 , σm i} |= value of f is higher than n
in all trajectories
if ∀x∃i, 1 ≤ x ≤ m, 0 ≤ i ≤ k : hsix , σ x i |= value of f is higher than n
{hs10 , σ1 i, . . . , hsm0 , σm i}
(4.275)
|= value of f is lower than n
in all trajectories
if ∀x∃i, 1 ≤ x ≤ m, 0 ≤ i ≤ k : hsix , σ x i |= value of f is lower than n
{hs10 , σ1 i, . . . , hsm0 , σm i}
(4.276)
|= value of f is n
in all trajectories
if ∀x∃i, 1 ≤ x ≤ m, 0 ≤ i ≤ k : hsix , σ x i |= value of f is n
(4.277)
{hs10 , σ1 i, . . . , hsm0 , σm i} |= value of f is higher than n
at time step i
in all trajectories
if ∀x, 1 ≤ x ≤ m : hsix , σ x i |= value of f is higher than n
{hs10 , σ1 i, . . . , hsm0 , σm i}
(4.278)
|= value of f is lower than n
at time step i
in all trajectories
if ∀x, 1 ≤ x ≤ m : hsix , σ x i |= value of f is lower than n
{hs10 , σ1 i, . . . , hsm0 , σm i}
(4.279)
|= value of f is n
at time step i
in all trajectories
if ∀x, 1 ≤ x ≤ m : hsix , σ x i |= value of f is n
(4.280)
Now, we turn our attention to domain descriptions with locational fluents. Let D be a domain
description and σ = s0 , . . . , sk be its trajectory of length k as defined in (65). Let σ1 , . . . , σm
represent the set of trajectories of D filtered by observations as necessary, with each trajectory
has the form σi = si0 , . . . , sik 1 ≤ i ≤ m. Let D̄ be a modified domain description and σ̄1 , . . . , σ̄m̄
be its trajectories of the form σ̄i = s̄i0 , . . . , s̄ik . Then we define query satisfaction using the
formula satisfaction in section 4.6.3 as follows.
Two sets of trajectories σ1 , . . . , σm and σ̄1 , . . . , σ̄m̄ satisfy a comparative query description
based on formula satisfaction of section 4.6.3 as follows:
n
o
{hs10 , σ1 i, . . . , hsm0 , σm i}, {h s̄10 , σ̄1 i, . . . , h s̄m̄0 , σ̄m̄ i}
|= direction of change in haggopi rate of production of f atloc l is d
if
n
when observed between time step i and time step j
o
,j
{hs1i , σ1 i, . . . , hsmi , σm i}, {h s̄1i , σ̄1 i, . . . , h s̄m̄i , σ̄m̄ i}
|= direction of change in haggopi rate of production of f atloc l is d
(4.281)
CHAPTER 4. THE BIOPATHQA SYSTEM
n
{hs10 , σ1 i, . . . , hsm0 , σm i}, {h s̄10 , σ̄1 i, . . . , h s̄m̄0 , σ̄m̄ i}
117
o
|= direction of change in haggopi rate of firing of f is d
if
n
when observed between time step i and time step j
o
,j
{hs1i , σ1 i, . . . , hsmi , σm i}, {h s̄1i , σ̄1 i, . . . , h s̄m̄i , σ̄m̄ i}
|= direction of change in haggopi rate of firing of f is d
n
o
1
{hs0 , σ1 i, . . . , hsm0 , σm i}, {h s̄10 , σ̄1 i, . . . , h s̄m̄0 , σ̄m̄ i}
(4.282)
|= direction of change in haggopi total production of f atloc l is d
if
n
when observed between time step i and time step j
o
,j
{hs1i , σ1 i, . . . , hsmi , σm i}, {h s̄1i , σ̄1 i, . . . , h s̄m̄i , σ̄m̄ i}
|= direction of change in haggopi total production of f atloc l is d
n
o
1
{hs0 , σ1 i, . . . , hsm0 , σm i}, {h s̄10 , σ̄1 i, . . . , h s̄m̄0 , σ̄m̄ i}
(4.283)
|= direction of change in haggopi value of f atloc l is d
when observed at time step i
if
n
{hs1i , σ1 i, . . . , hsmi , σm i}, {h s̄1i , σ̄1 i, . . . , h s̄m̄i , σ̄m̄ i}
o
|= direction of change in haggopi value of f atloc l is d
(4.284)
A set of trajectories σ1 , . . . , σm , each of length k satisfies a quantitative or a qualitative
interval query description based on formula satisfaction of section 4.6.3 as follows:
{hs10 , σ1 i, . . . , hsm0 , σm i} |= rate of production of f atloc l is n
if ∃x, 1 ≤ x ≤ m : (hs0x , σ x i, k) |= rate of production of f atloc l is n
{hs10 , σ1 i, . . . , hsm0 , σm i}
|= rate of firing of a is n
if ∃x, 1 ≤ x ≤ m : (hs0x , σ x i, k) |= rate of firing of a is n
{hs10 , σ1 i, . . . , hsm0 , σm i}
(4.285)
(4.286)
|= total production of f atloc l is n
if ∃x, 1 ≤ x ≤ m : (hs0x , σ x i, k) |= total production of f atloc l is n
(4.287)
{hs10 , σ1 i, . . . , hsm0 , σm i} |= rate of production of f atloc l is n
when observed between time step i and time step j
if ∃x, 1 ≤ x ≤ m : (hsix , σ x i, j) |= rate of production of f atloc l is n
{hs10 , σ1 i, . . . , hsm0 , σm i}
(4.288)
|= rate of firing of a is n
when observed between time step i and time step j
if ∃x, 1 ≤ x ≤ m : (hsix , σ x i, j) |= rate of firing of a is n
{hs10 , σ1 i, . . . , hsm0 , σm i}
(4.289)
|= total production of f atloc l is n
when observed between time step i and time step j
if ∃x, 1 ≤ x ≤ m : (hsix , σ x i, j) |= total production of f atloc l is n
(4.290)
{hs10 , σ1 i, . . . , hsm0 , σm i} |= f is accumulating atloc l
if ∃x, 1 ≤ x ≤ m : (hs0x , σ x i, k) |= f is accumulating atloc l
{hs10 , σ1 i, . . . , hsm0 , σm i} |= f is decreasing atloc l
(4.291)
CHAPTER 4. THE BIOPATHQA SYSTEM
if ∃x, 1 ≤ x ≤ m : (hs0x , σ x i, k) |= f is decreasing atloc l
118
(4.292)
{hs10 , σ1 i, . . . , hsm0 , σm i} |= f is accumulating atloc l
when observed between time step i and time step j
if ∃x, 1 ≤ x ≤ m : (hsix , σ x i, j) |= f is accumulating atloc l
{hs10 , σ1 i, . . . , hsm0 , σm i}
(4.293)
|= f is decreasing atloc l
when observed between time step i and time step j
if ∃x, 1 ≤ x ≤ m : (hsix , σ x i, j) |= f is decreasing atloc l
(4.294)
{hs10 , σ1 i, . . . , hsm0 , σm i} |= rate of production of f atloc l is n
in all trajectories
if ∀x, 1 ≤ x ≤ m : (hs0x , σ x i, k) |= rate of production of f atloc l is n
(4.295)
{hs10 , σ1 i, . . . , hsm0 , σm i} |= rate of firing of a is n
in all trajectories
if ∀x, 1 ≤ x ≤ m : (hs0x , σ x i, k) |= rate of firing of a is n
{hs10 , σ1 i, . . . , hsm0 , σm i}
(4.296)
|= total production of f atloc l is n
in all trajectories
if ∀x, 1 ≤ x ≤ m : (hs0x , σ x i, k) |= total production of f atloc l is n
(4.297)
{hs10 , σ1 i, . . . , hsm0 , σm i} |= rate of production of f atloc l is n
when observed between time step i and time step j
in all trajectories
if ∀x, 1 ≤ x ≤ m : (hsix , σ x i, j) |= rate of production of f atloc l is n
{hs10 , σ1 i, . . . , hsm0 , σm i}
(4.298)
|= rate of firing of a is n
when observed between time step i and time step j
in all trajectories
if ∀x, 1 ≤ x ≤ m : (hsix , σ x i, j) |= rate of firing of a is n
{hs10 , σ1 i, . . . , hsm0 , σm i}
(4.299)
|= total production of f atloc l is n
when observed between time step i and time step j
in all trajectories
if ∀x, 1 ≤ x ≤ m : (hsix , σ x i, j) |= total production of f atloc l is n till j
(4.300)
{hs10 , σ1 i, . . . , hsm0 , σm i} |= f is accumulating atloc l
in all trajectories
if ∀x, 1 ≤ x ≤ m : (hs0x , σ x i, k) |= f is accumulating atloc l
{hs10 , σ1 i, . . . , hsm0 , σm i}
(4.301)
|= f is decreasing atloc l
in all trajectories
if ∀x, 1 ≤ x ≤ m : (hs0x , σ x i, k) |= f is decreasing atloc l
{hs10 , σ1 i, . . . , hsm0 , σm i} |= f is accumulating atloc l
(4.302)
CHAPTER 4. THE BIOPATHQA SYSTEM
119
when observed between time step i and time step j
in all trajectories
if ∀x, 1 ≤ x ≤ m : (hsix , σ x i, j) |= f is accumulating atloc l
{hs10 , σ1 i, . . . , hsm0 , σm i}
(4.303)
|= f is decreasing atloc l
when observed between time step i and time step j
in all trajectories
if ∀x, 1 ≤ x ≤ m : (hsix , σ x i, j) |= f is decreasing atloc l
(4.304)
A set of trajectories σ1 , . . . , σm , each of length k satisfies a quantitative or a qualitative point
query description based on formula satisfaction of section 4.6.3 as follows:
{hs10 , σ1 i, . . . , hsm0 , σm i} |= value of f atloc l is higher than n
if ∃x∃i, 1 ≤ x ≤ m, 0 ≤ i ≤ k : hsix , σ x i |= value of f atloc l is higher than n
{hs10 , σ1 i, . . . , hsm0 , σm i}
|= value of f atloc l is lower than n
if ∃x∃i, 1 ≤ x ≤ m, 0 ≤ i ≤ k : hsix , σ x i |= value of f atloc l is lower than n
{hs10 , σ1 i, . . . , hsm0 , σm i}
(4.305)
(4.306)
|= value of f atloc l is n
if ∃x∃i, 1 ≤ x ≤ m, 0 ≤ i ≤ k : hsix , σ x i |= value of f atloc l is n
(4.307)
{hs10 , σ1 i, . . . , hsm0 , σm i} |= value of f atloc l is higher than n
at time step i
if ∃x, 1 ≤ x ≤ m : hsix , σ x i |= value of f atloc l is higher than n
{hs10 , σ1 i, . . . , hsm0 , σm i}
(4.308)
|= value of f atloc l is lower than n
at time step i
if ∃x, 1 ≤ x ≤ m : hsix , σ x i |= value of f atloc l is lower than n
{hs10 , σ1 i, . . . , hsm0 , σm i}
(4.309)
|= value of f atloc l is n
at time step i
if ∃x, 1 ≤ x ≤ m : hsix , σ x i |= value of f atloc l is n
(4.310)
{hs10 , σ1 i, . . . , hsm0 , σm i} |= value of f atloc l is higher than n
in all trajectories
if ∀x∃i, 1 ≤ x ≤ m, 0 ≤ i ≤ k : hsix , σ x i |= value of f atloc l is higher than n
{hs10 , σ1 i, . . . , hsm0 , σm i}
(4.311)
|= value of f atloc l is lower than n
in all trajectories
if ∀x∃i, 1 ≤ x ≤ m, 0 ≤ i ≤ k : hsix , σ x i |= value of f atloc l is lower than n
{hs10 , σ1 i, . . . , hsm0 , σm i}
(4.312)
|= value of f atloc l is n
in all trajectories
if ∀x∃i, 1 ≤ x ≤ m, 0 ≤ i ≤ k : hsix , σ x i |= value of f atloc l is n
{hs10 , σ1 i, . . . , hsm0 , σm i} |= value of f atloc l is higher than n
at time step i
in all trajectories
(4.313)
CHAPTER 4. THE BIOPATHQA SYSTEM
if ∀x, 1 ≤ x ≤ m : hsix , σ x i |= value of f atloc l is higher than n
{hs10 , σ1 i, . . . , hsm0 , σm i}
120
(4.314)
|= value of f atloc l is lower than n
at time step i
in all trajectories
if ∀x, 1 ≤ x ≤ m : hsix , σ x i |= value of f atloc l is lower than n
{hs10 , σ1 i, . . . , hsm0 , σm i}
(4.315)
|= value of f atloc l is n
at time step i
in all trajectories
if ∀x, 1 ≤ x ≤ m : hsix , σ x i |= value of f atloc l is n
(4.316)
Next, we generically define the satisfaction of a simple point formula cascade query w.r.t.
a set of trajectories σ1 , . . . , σm . The trajectories will either be as defined in definitions (62) or
(65) for simple point formula cascade query statement made up of simple fluents or locational
fluents, respectively. A set of trajectories σ1 , . . . , σm , each of length k satisfies a simple point
formula cascade based on formula satisfaction of section 4.6.3 as follows:
{hs10 , σ1 i, . . . , hsm0 , σm i} |= hsimple point formulai0
after hsimple point formulai1,1 , . . . , hsimple point formulai1,n1
..
.
after hsimple point formulaiu,1 , . . . , hsimple point formulaiu,nu
if ∃x∃i0 ∃i1 . . . ∃iu , 1 ≤ x ≤ m, i1 < i0 ≤ k, . . . , iu < iu−1 ≤ k, 0 ≤ iu ≤ k :
hsix0 , σ x i |= hsimple point formulai0 and
hsix1 , σ x i |= hsimple point formulai1,1 and . . . and hsix1 , σ x i |= hsimple point formulai1,n1 and
..
.
hsixu , σ x i |= hsimple point formulaiu,1 and . . . and hsixu , σ x i |= hsimple point formulaiu,nu
(4.317)
{hs10 , σ1 i, . . . , hsm0 , σm i} |= hsimple point formulai0
after hsimple point formulai1,1 , . . . , hsimple point formulai1,n1
..
.
after hsimple point formulaiu,1 , . . . , hsimple point formulaiu,nu
in all trajectories
if ∀x, 1 ≤ x ≤ m, ∃i0 ∃i1 . . . ∃iu , i1 < i0 ≤ k, . . . , iu < iu−1 ≤ k, 0 ≤ iu ≤ k :
hsix0 , σ x i |= hsimple point formulai0 and
hsix1 , σ x i |= hsimple point formulai1,1 and . . . and hsix1 , σ x i |= hsimple point formulai1,n1 and
..
.
hsixu , σ x i |= hsimple point formulaiu,1 and . . . and hsixu , σ x i |= hsimple point formulaiu,nu
CHAPTER 4. THE BIOPATHQA SYSTEM
121
(4.318)
{hs10 , σ1 i, . . . , hsm0 , σm i} |= hsimple point formulai0
when hsimple point formulai1,1 , . . . , hsimple point formulai1,n1
if ∃x∃i, 1 ≤ x ≤ m, 0 ≤ i ≤ k :
hsix0 , σ x i |= hsimple point formulai0 and
hsix1 , σ x i |= hsimple point formulai1,1 and . . . and hsix1 , σ x i |= hsimple point formulai1,n1
(4.319)
{hs10 , σ1 i, . . . , hsm0 , σm i} |= hsimple point formulai0
when hsimple point formulai1,1 , . . . , hsimple point formulai1,n1
in all trajectories
if ∀x, 1 ≤ x ≤ m, ∃i, 0 ≤ i ≤ k :
hsix0 , σ x i |= hsimple point formulai0 and
hsix1 , σ x i |= hsimple point formulai1,1 and . . . and hsix1 , σ x i |= hsimple point formulai1,n1
(4.320)
4.6.6
Query Statement Satisfaction
Let D as defined in section 4.3 be a domain description and Q be a query statement (4.122) as
defined in section 4.5 with query description U, interventions V1 , . . . , V|V| , internal observations
O1 , . . . , O|O| , and initial conditions I1 , . . . , I|I| . Let D1 ≡ DI1 · · ·I|I| V1 · · ·V|V| be the modified domain description constructed by applying the initial conditions and interventions from
Q as defined in section 4.6.2. Let σ1 , . . . , σm be the trajectories of D1 that satisfy O1 , . . . , O|O|
as given in section 4.6.4. Then, D satisfies Q if {σ1 , . . . , σm } |= U as defined in section 4.6.5.
Let D as defined in section 4.3 be a domain description and Q be a query statement (4.123) as
defined in section 4.5 with query description U, interventions V1 , . . . , V|V| , internal observations
O1 , . . . , O|O| , and initial conditions I1 , . . . , I|I| . Let D) ≡ D I1 · · · I|I| be the nominal domain
description constructed by applying the initial conditions from Q as defined in section 4.6.2. Let
σ1 , . . . , σm be the trajectories of D0 . Let D1 ≡ D I1 · · · I|I| V1 · · · V|V| be the modified
domain description constructed by applying the initial conditions and interventions from Q as
defined in section 4.6.2. Let σ̄1 , . . . , σ̄m̄ be the
O1 , . . . , O|O| as
n trajectories of D1 that satisfy
o
given in section 4.6.4. Then, D satisfies Q if {σ1 , . . . , σm }, {σ̄1 , . . . , σ̄m̄ } |= U as defined in
section 4.6.5.
CHAPTER 4. THE BIOPATHQA SYSTEM
4.6.7
122
Example Encodings
In this section we give some examples of how we will encode queries and pathways related to
these queries. We will also show how the pathway is modified to answer questions 3 .
Question 11 At one point in the process of glycolysis, both DHAP and G3P are produced.
Isomerase catalyzes the reversible conversion between these two isomers. The conversion of
DHAP to G3P never reaches equilibrium and G3P is used in the next step of glycolysis. What
would happen to the rate of glycolysis if DHAP were removed from the process of glycolysis as
quickly as it was produced?
tr
dhap
t3
f16bp
t4
g3p
t5a
t5b
t6
2
bpg13
Figure 4.3: Petri Net for question 11
The question is asking for the direction of change in the rate of glycolysis when the nominal
glycolysis pathway is compared against a modified pathway in which dhap is removed as soon
as it is produced. Since this rate can vary with the trajectory followed by the world evolution, we
consider the average change in rate. From the domain knowledge (Reece et al., 2010, Figure 9.9)
we know that the rate of glycolysis can be measured by the rate of pyruvate (the end product
of glycolysis) and that the rate of pyruvate is equal to the rate of bpg13 (due to linear chain
from bpg13 to pyruvate). Thus, we can monitor the rate of bpg13 instead to determine the rate
of glycolysis. To ensure that our pathway is not starved due to source ingredients, we add a
continuous supply of f 16bp in quantity 1 to the pathway.
3
Some of the same pathways appear in previous chapters, they have been updated here with additional background
knowledge.
CHAPTER 4. THE BIOPATHQA SYSTEM
123
Then, the following pathway specification encodes the domain description D for question 11
and produces the PN in Fig. 4.3 minus the tr, t3 transitions:
domain of
t4 may execute causing
t5a may execute causing
t5b may execute causing
t6 may execute causing
initially
firing style
f 16bp is integer,
g3p is integer,
f 16bp change value by − 1,
g3p change value by + 1
dhap change value by − 1,
g3p change value by − 1,
g3p change value by − 1,
f 16bp has value 0,
g3p has value 0,
max
dhap is integer,
bpg13 is integer
dhap change value by + 1,
g3p change value by + 1
dhap change value by + 1
bpg13 change value by + 2
dhap has value 0,
bpg13 has value 0
(4.321)
And the following query Q for a simulation of length k encodes the question:
direction of change in average rate of production of bpg13 is d
when observed between time step 0 and time step k;
comparing nominal pathway with modified pathway obtained
due to interventions : remove dhap as soon as produced;
using initial setup : continuously supply f 16bp in quantity 1;
(4.322)
Since this is a comparative quantitative query statement, it is decomposed into two subqueries, Q0 capturing the nominal case of average rate of production w.r.t. given initial conditions:
average rate of production of bpg13 is navg
when observed between time step 0 and time step k;
using initial setup : continuously supply f 16bp in quantity 1;
and Q1 capturing the modified case in which the pathway is subject to interventions and observations w.r.t. initial
conditions:
average rate of production of bpg13 is n0avg
when observed between time step 0 and time step k;
due to interventions : remove dhap as soon as produced;
using initial setup : continuously supply f 16bp in quantity 1;
Then the task is to determine d, such that D |= Q0 for some value of navg , D |= Q1 for
some value of n0avg , and n0avg d navg . To answer the sub-queries Q0 and Q1 we build modified
domain descriptions D0 and D1 , where, D0 ≡ D (continuously supply f 16bp in quantity 1)
is the nominal domain description D modified according to Q0 to include the initial conditions;
and D1 ≡ D0 (remove dhap as soon as produced) is the modified domain description D
modified according to Q1 to include the initial conditions as well as the interventions. The
operator modifies the domain description to its left by adding, removing or modifying pathway
specification language statements to add the interventions and initial conditions description to
its right. Thus,
D0 = D +
n
t f f 16bp may execute causing
f 16bp change value by + 1
CHAPTER 4. THE BIOPATHQA SYSTEM
D1 = D0 +
n
tr may execute causing
124
dhap change value by ∗
Performing a simulation of k = 5 steps with ntok = 20 max tokens, we find that the average
rate of bpg13 production decreases from navg = 0.83 to n0avg = 0.5. Thus, D |= Q iff d =0 <0 .
Alternatively, we say that the rate of glycolysis decreases when DHAP is removed as quickly as
it is produced.
Question 12 When and how does the body switch to B-oxidation versus glycolysis as the major
way of burning fuel?
t1
sug
gly
acoa
t2
fac
box
Figure 4.4: Petri Net for question 12
The following pathway specification encodes the domain description D for question 12 and
produces the PN in Fig. 4.4:
domain of
gly may execute causing
box may execute causing
inhibit box if
initially
t1 may execute causing
t2 may execute causing
firing style
gly is integer,
f ac is integer,
sug change value by − 1,
f ac change value by − 1,
sug has value 1 or higher
sug has value 3,
acoa has value 0
sug change value by + 1
f ac change value by + 1
∗
sug is integer,
acoa is integer
acoa change value by + 1
acoa change value by + 1
f ac has value 3
where, f ac represents fatty acids, sug represents sugar, acoa represents acetyl coenzyme-A, gly
represents the process of glycolysis, and box represents the process of beta oxidation.
The question is asking for the general conditions when glycolysis switches to beta-oxidation,
which is some property “p” that holds after which the switch occurs. The query Q is encoded
as:
gly switches to box when p;
due to observations :
gly switches to box
in all trajectories
CHAPTER 4. THE BIOPATHQA SYSTEM
125
where condition ‘p’ is a conjunction of simple point formulas. Then the task is to determine a
minimal such conjunction of formulas that is satisfied in the state where ‘gly0 switches to ‘box0
holds over all trajectories.4
Since there is no change in initial conditions of the pathway and there are no interventions,
the modified domain description D1 ≡ D.
Intuitively, p is the property that holds over fluents of the transitional state s j in which the
switch takes palce, such that gly ∈ T j−1 , box < T j−1 , gly < T j , box ∈ T j and the minimal set of
firings leading up to it. The only trajectories to consider are the ones in which the observation
is true. Thus the condition p is determined as the intersection of sets of fluent based conditions
that were true at the time of the switch, such as:
{sug has value s j (sug), sug has value higher than 0, . . .
sug has value higher than s j (sug) − 1, sug has value lower than s j (sug) + 1,
f ac has value s j ( f ac), f ac has value higher than 0, . . .
f ac has value higher than s j ( f ac) − 1, f ac has value lower than s j ( f ac) + 1,
acoa has value s j (acoa), acoa has value higher than 0, . . .
acoa has value higher than s j (acoa) − 1, acoa has value lower than s j (acoa) + 1}
Simulating it for k = 5 steps with ntok = 20 max tokens, we find the condition p =
acoa has value greater than 0, sug has value 0, sug has value lower than 1, f ac has value higher than 0}.
Thus, the state when this switch occurs must sugar (sug) depleted and available supply of fatty
acids ( f ac).
Question 13 The final protein complex in the electron transport chain of the mitochondria is
non-functional. Explain the effect of this on pH of the intermembrane space of the mitochondria.
The question is asking for the direction of change in the pH of the intermembrane space
when the nominal case is compared against a modified pathway in which the complex 4 (t4)
is defective. Since pH is defined as −log10 (H+), we monitor the total production of H+ ions
to determine the change in pH value. However, since different world evolutions can follow
different trajectories, we consider the average production of H+. Furthermore, we model the
defective t4 as being unable to carry out its reaction, by disabling/inhibiting it.
Then the following pathway specification encodes the domain description D for question 13
and produces the PN in Fig. 4.5 without the f t4 place node. In the pathway description below,
we have included only one domain declaration for all fluents as an example and left the rest out
to save space. Assume integer domain declaration for all fluents.
domain of
nadh atloc mm is integer
t1 may execute causing
nadh atloc mm change value by − 2, h atloc mm change value by − 2,
e atloc q change value by + 2, h atloc is change value by + 2,
4
Note that this could be an LTL formula that must hold in all trajectories, but we did not add it here to keep the
language simple.
CHAPTER 4. THE BIOPATHQA SYSTEM
126
h/2
h/2,
nadh/2
h/2
t1
nadp/2
t10
nadh/2,
h/4,o2/1
mm
fad/2
e/2
e/2
e/2
q
h/2
t3
e/2
t2
is
h/2
ft4
fadh2/2
cytc
e/2
t4
h2o/2
h/6,o2/1
Figure 4.5: Petri Net for question 13
nad p atloc mm change value by + 2
t2 may execute causing
f adh2 atloc mm change value by − 2, e atloc q change value by + 2,
f ad atloc mm change value by + 2
t3 may execute causing
e atloc q change value by − 2, h atloc mm change value by − 2,
e atloc cytc change value by + 2, h atloc is change value by + 2
t4 may execute causing
o2 atloc mm change value by − 1, e atloc cytc change value by − 2,
h atloc mm change value by − 6, h2o atloc mm change value by + 2,
h atloc is change value by + 2
t10 may execute causing
nadh atloc mm change value by + 2,
h atloc change value by + 4, o2 atloc mm change value by + 1
initially
f adh2 atloc mm has value 0, e atloc mm has value 0,
f ad atloc mm has value 0, o2 atloc mm has value 0,
h2o atloc mm has value 0, atp atloc mm has value 0
initially
f adh2 atloc is has value 0, e atloc is has value 0,
f ad atloc is has value 0, o2 atloc is has value 0,
h2o atloc is has value 0, atp atloc is has value 0
initially
f adh2 atloc q has value 0, e atloc q has value 0,
f ad atloc q has value 0, o2 atloc q has value 0,
CHAPTER 4. THE BIOPATHQA SYSTEM
127
h2o atloc q has value 0, atp atloc q has value 0
initially
f adh2 atloc cytc has value 0, e atloc cytc has value 0,
f ad atloc cytc has value 0, o2 atloc cytc has value 0,
h2o atloc cytc has value 0, atp atloc cytc has value 0
firing style max
where mm represents the mitochondrial matrix, is represents the intermembrane space, t1 − t4
represent the reaction of the four complexes making up the electron transport chain, h is the H+
ion, nadh is NADH, f adh2 is FADH2 , f ad is FAD, e is electrons, o2 is oxygen O2 , atp is
AT P, h2o is water H2 O, and t10 is a source transition that supply a continuous supply of source
ingredients for the chain to function, such as nadh, h, o2.
As a result, the query Q asked by the question is encoded as follows:
direction of change in average total production of h atloc is is d
when observed between time step 0 and time step k;
comparing nominal pathway with modified pathway obtained
due to intervention t4 disabled;
Since this is a comparative quantitative query statement, it is decomposed into two queries,
Q0 capturing the nominal case of average production w.r.t. given initial conditions:
average total production of h atloc is is navg
when observed between time step 0 and time step k;
and Q1 the modified case w.r.t. initial conditions, modified to include interventions and subject
to observations:
average total production of h atloc is is n0avg
when observed between time step 0 and time step k;
due to intervention t4 disabled;
The the task is to determine d, such that D |= Q0 for some value of navg , D |= Q1 for some
value of n0avg , and n0avg d navg . To answer the sub-queries Q0 and Q1 we build nominal description D0 and modified pathway D1 , where D0 ≡ D since there are no initial conditions; and
D1 ≡ D0 (t4 disabled) is the domain description D modified according to Q1 to modified to
include the initial conditions as well as interventions. The operator modifies the domain description to its left by adding, removing or modifying pathway specification language statements
to add the intervention and initial conditions to its right. Thus,
D1 = D0 +
n
inhibit t4
Performing a simulation of k = 5 steps with ntok = 20 max tokens, we find that the average
total production of H+ in the intermembrane space (h at location is) reduces from 16 to 14.
Lower quantity of H+ translates to a higer numeric value of −log10 , as a result the pH increases.
Question 14 Membranes must be fluid to function properly. How would decreased fluidity of
the membrane affect the efficiency of the electron transport chain?
CHAPTER 4. THE BIOPATHQA SYSTEM
128
h/2
nadh/2
h/2
t1
nadp/2
t10
nadh/2,
h/6,o2/1
e/2
q_3
e/2
h/2
mm
tq
e/2
h/2
q
is
e/2
t3
[4]
h/2
e/2
cytc_4
o2/1
h/2
e/2
tcytc
e/2
cytc
e/2
[4]
t4
h2o/2
Figure 4.6: Petri Net with colored tokens alternate for question 14
From background knowledge, we know that, “Establishing the H+ gradient is a major function of the electron transport chain” (Reece et al., 2010, Chapter 9), we measure the efficiency
in terms of H+ ions moved to the intermembrane space (is) over time. Thus, we interpret the
question is asking for the direction of change in the production of H+ moved to the intermembrane space when the nominal case is compared against a modified pathway with decreased
fluidity of membrane. Additional background knowledge from Reece et al. (2010) tells us that
the decreased fluidity reduces the speed of mobile carriers such as q and cytc. Fluidity can span
a range of values, but we will consider one such value v per query.
The following pathway specification encodes the domain description D for question 14 and
produces the PN in Fig. 4.6 minus the places q 3, cytc 4 and transitions tq, tcytc. In the pathway
description below, we have included only one domain declaration for all fluents as an example
and left the rest out to save space. Assume integer domain declaration for all fluents.
domain of
nadh atloc mm is integer
t1 may execute causing
nadh atloc mm change value by − 2, h atloc mm change value by − 2,
e atloc q change value by + 2, h atloc is change value by + 2,
nad p atloc mm change value by + 2
t3 may execute causing
e atloc q change value by − 2, h atloc mm change value by − 2,
e atloc cytc change value by + 2, h atloc is change value by + 2
t4 may execute causing
o2 atloc mm change value by − 1, e atloc cytc change value by − 2,
h atloc mm change value by − 2, h2o atloc mm change value by + 2,
h atloc is change value by + 2
t10 may execute causing
nadh atloc mm change value by + 2, h atloc mm change value by + 4,
o2 atloc mm change value by + 1
initially
nadh atloc mm has value 0, h atloc mm has value 0,
nad p atloc mm has value 0, o2 atloc mm has value 0
h2o atloc mm has value 0
CHAPTER 4. THE BIOPATHQA SYSTEM
129
initially
nadh atloc is has value 0, h atloc is has value 0,
nad p atloc is has value 0, o2 atloc is has value 0
h2o atloc is has value 0
initially
nadh atloc q has value 0, h atloc q has value 0,
nad p atloc q has value 0, o2 atloc q has value 0
h2o atloc q has value 0
initially
nadh atloc cytc has value 0, h atloc cytc has value 0,
nad p atloc cytc has value 0, o2 atloc cytc has value 0
h2o atloc cytc has value 0
firing style max
The query Q asked by the question is encoded as follows:
direction of change in average total production of h atloc is is d
when observed between time step 0 and time step k;
comparing nominal pathway with modified pathway obtained
due to interventions :
add delay of v time units in availability of e atloc q,
add delay of v time units in availability of e atloc cytc;
Since this is a comparative quantitative query statement, we decompose it into two queries,
Q0 capturing the nominal case of average rate of production w.r.t. given initial conditions:
average total production of h atloc is is d
when observed between time step 0 and time step k;
and Q1 the modified case w.r.t. initial conditions, modified to include interventions and subject
to observations:
average total production of h atloc is is d
when observed between time step 0 and time step k;
due to interventions :
add delay of v time units in availability of e atloc q,
add delay of v time units in availability of e atloc cytc;
Then the task is to determine d, such that D |= Q0 for some value of navg , D |= Q1 for some
value of n0avg , and n0avg d navg . To answer the sub-queries Q0 and Q1 we build modified domain
descriptions D0 and D1 , where, D0 ≡ D is the nominal domain description D modified according
to Q0 to include the initial conditions; and D1 ≡ D0 (add delay of v time units in availability of
e atloc q) (add delay of v time units in availability of e atloc cytc) is the modified domain
description D modified according to Q1 to include the initial conditions as well as the interventions. The operator modifies the domain description to its left by adding, removing or
CHAPTER 4. THE BIOPATHQA SYSTEM
130
modifying pathway specification language statements to add the interventions and initial conditions description to its right. Thus,
D1 = D −
+
t1 may execute causing
t3 may execute causing
t1 may execute causing
t3 may execute causing
tq may fire causing
tcytc may fire causing
tq executes
tcytc executes
nadh atloc mm change value by − 2,
h atloc mm change value by − 2,
e atloc q 3 change value by + 2,
h atloc is change value by + 2,
nad p atloc mm change value by + 2
e atloc q change value by − 2,
h atloc mm change value by − 2,
e atloc cytc 4 change value by + 2,
h atloc is change value by + 2
e atloc q 3 change value by − 2,
e atloc q change value by + 2
e atloc cytc 4 change value by − 2,
e atloc cytc change value by + 2
in 4 time units
in 4 time units
nadh atloc mm change value by − 2,
h atloc mm change value by − 2,
e atloc q change value by + 2,
h atloc is change value by + 2,
nad p atloc mm change value by + 2
e atloc q change value by − 2,
h atloc mm change value by − 2,
e atloc cytc change value by + 2,
h atloc is change value by + 2
Performing a simulation of k = 5 steps with ntok = 20 max tokens with a fluidity based delay
of 2, we find that the average total production of H+ in the intermembrane space (h at location
is) reduces from 16 to 10. Lower quantity of H+ going into the intermembrane space means
lower efficiency, where we define the efficiency as the total amount of H+ ions transferred to
the intermembrane space over the simulation run.
4.6.8
Example Encoding with Conditional Actions
Next, we illustrate how conditional actions would be encoded in our high-level language with
an example. Consider the pathway from question 11. Say, the reaction step t4 has developed
a fault, in which it has two modes of operation, in the first mode, when f 16bp has less than
3 units available, the reaction proceeds normally, but when f 16bp is available in 3 units or
higher, the reaction continues to produce g3p but not dhap directly. dhap can still be produced
by subsequent step from g3p. The modified pathway is given in our pathway specification
CHAPTER 4. THE BIOPATHQA SYSTEM
131
language below:
t3 may execute causing
t4 may execute causing
if
t4 may execute causing
if
t5a may execute causing
t5b may execute causing
t6 may execute causing
initially
firing style
f 16bp change value by + 1
f 16bp change value by − 1,
g3p change value by + 1
g3p has value lower than 3
f 16bp change value by − 1,
g3p has value 3 or higher
dhap change value by − 1,
g3p change value by − 1,
g3p change value by − 1,
f 16bp has value 0,
g3p has value 0,
max
dhap change value by + 1,
dhap change value by + 1
g3p change value by + 1
dhap change value by + 1
bpg13 change value by + 2
dhap has value 0,
bpg13 has value 0
We ask the same question Q:
direction of change in average rate of production of bpg13 is d
when observed between time step 0 and time step k;
comparing nominal pathway with modified pathway obtained
due to interventions : remove dhap as soon as produced;
using initial setup : continuously supply f 16bp in quantity 1;
Since this is a comparative quantitative query statement, we decompose it into two queries,
Q0 capturing the nominal case of average rate of production w.r.t. given initial conditions:
average rate of production of bpg13 is n
when observed between time step 0 and time step k;
using initial setup : continuously supply f 16bp in quantity 1;
and Q1 the modified case w.r.t. initial conditions, modified to include interventions and subject
to observations:
average rate of production of bpg13 is n
when observed between time step 0 and time step k;
due to interventions : remove dhap as soon as produced;
using initial setup : continuously supply f 16bp in quantity 1;
Then the task is to determine d, such that D |= Q0 for some value of navg , D |= Q1 for some
value of n0avg , and n0avg d navg . To answer the sub-queries Q0 and Q1 we build modified domain
descriptions D0 and D1 , where, D0 ≡ D (continuously supply f 16bp in quantity 1) is the
nominal domain description D modified according to Q0 to include the initial conditions; and
D1 ≡ D0 (remove dhap as soon as produced) is the modified domain description D modified
according to Q1 to include the initial conditions as well as the interventions. The operator
modifies the domain description to its left by adding, removing or modifying pathway specification language statements to add the interventions and initial conditions description to its right.
Thus,
D0 = D +
n
D1 = D0 +
t f f 16bp may execute causing
n
tr may execute causing
f 16bp change value by + 1
dhap change value by ∗
CHAPTER 4. THE BIOPATHQA SYSTEM
4.6.9
132
ASP Program
Next we briefly outline how the pathway specification and the query statement components are
encoded in ASP (using Clingo syntax). In the following section, we will illustrate the process
using an example.
As evident from the previous sections, we need to simulate non-comparative queries only.
Any comparative queries are translated into non-comparative sub-queries, each of which is simulated and their results compared to evaluate the comparative query.
The ASP program is a concatenation of the translation of a pathway specification (domain
description which includes the firing style) and internal observations. Any initial setup conditions and interventions are pre-applied to the pathway specification using intervention semantics
in section 4.6.2 before it is translated to ASP using the translation in chapter 2 as our basis. The
encoded pathway specification has the semantics defined in section 4.4. Internal observations
in the ‘due to observations:’ portion of query statement are translated into ASP constraints using the internal observation semantics defined in section 4.6.4 and added to the encoding of the
pathway specification.
The program if simulated for a specified simulation length k produces all trajectories of the
pathway for the specified firing style, filtered by the internal observations. The query description
specified in the query statement is then evaluated w.r.t. these trajectories. Although this part can
be done in ASP, we have currently implemented it outside ASP in our implementation for ease
of using floating point math.
Next, we describe an implementation of our high level language and illustrate the construction of an ASP program, its simulation, and query statement evaluation.
4.6.10
Implementation
We have developed an implementation 5 of a subset of our high level (Pathway and Query
specification) language in Python. We use the Clingo ASP implementation for our simulation. In
this section we describe various components of this implementation. An architectural overview
of our implementation is shown in figure 4.7.
The Pathway Specification Language (BioPathQA-PL) Parser component is responsible for
parsing the Pathway Specification Language (BioPathQA-PL). It use PLY (Python Lex-Yacc)6
to parse a given pathway specification using grammar based on section 4.3. On a successful
parse, a Guarded-Arc Petri Net pathway model based on section 4.4 is constructed for the pathway specification.
The Query Language Parser component is responsible for parsing the Query Specification Language (BioPathQA-QL). It uses PLY to parse a given query statement using grammar
based on section 4.5. On a successful parse, an internal representation of the query statement
is constructed. Elements of this internal representation include objects representing the query
description, the list of interventions, the list of internal observations, and the list of initial setup
5
6
Implementation available at: https://sites.google.com/site/deepqa2014/
http://www.dabeaz.com/ply
CHAPTER 4. THE BIOPATHQA SYSTEM
Figure 4.7: BioPathQA Implementation System Architecture
133
CHAPTER 4. THE BIOPATHQA SYSTEM
134
conditions. Each intervention and initial setup condition object has logic in it to modify a given
pathway per the intervention semantics described in section 4.6.2. The Query Statement Model
component is also responsible for generating basic queries for aggregate queries and implementing interventions in the Petri Net Pathway Model.
The Dictionary of fluents, locations, and actions is consulted by the ASP code generator
to standardize symbol names in the ASP code produced for the pathway specification and the
internal observations.
The ASP Translator component is responsible for translating the Guarded-Arc Petri Net
model into ASP facts and rules; and the driver needed to simulate the model using the firing
semantics specified in the pathway model. The code generated is based on the ASP translation
of Petri Nets and its various extensions given in chapter 2. To reduce the ASP code and its
complexity, the translator limits the output model to the extensions used in the Petri Net model
to be translated. Thus, the colored tokens extension code is not produced unless colored tokens
have been used. Similarly, guarded-arcs code is not produced if no arc-guards are used in the
model.
The ASP Translator component is also responsible for translating internal observations from
the Query Statement into ASP constraints to filter Petri Net trajectories based on the observation
semantics in section 4.6.4. Following examples illustrate our encoding scheme. The observation
‘a1 switches to a2 ’ is encoded as a constraint using the following rules:
obs_1_occurred(TS+1) :- time(TS;TS+1), trans(a1;a2),
fires(a1,TS), not fires(a2,TS),
not fires(a1,TS+1), fires(a2,TS+1).
obs_1_occurred :- obs_1_occurred(TS), time(TS).
obs_1_had_occurred(TSS) :- obs_1_occurred(TS), TS<=TSS, time(TSS;TS).
:- not obs_1_occurred.
The observation ‘a1 occurs at time step 5’ is encoded as a constraint using the following
rules:
obs_2_occurred(TS) :- fires(a1,TS), trans(a1), time(TS), TS=5.
obs_2_occurred :- obs_2_occurred(TS), time(TS).
obs_2_had_occurred(TSS) :- obs_2_occurred(TS), TS<=TSS, time(TSS;TS).
:- not obs_2_occurred.
The observation ‘s1 is decreasing atloc l1 when observed between time step 0 and time step 5’
is encoded as a constraint using the following rules:
%
obs_3_violated(TS) :- place(l1), col(s1),
holds(l1,Q1,s1,TS), holds(l1,Q2,s1,TS+1),
num(Q1;Q2), Q2 > Q1, time(TS;TS+1), TS=0, TS+1=5.
obs_3_violated :- obs_3_violated(TS), time(TS).
%
obs_3_occurred(TS+1) :- not obs_3_violated,
holds(l1,Q1,s1,TS), holds(l1,Q2,s1,TS+1), time(TS;TS+1), num(Q1;Q2),
Q2<Q1, TS=0, TS+1=5.
obs_3_occurred :- obs_3_occurred(TS), time(TS).
obs_3_had_occurred(TSS) :- obs_3_occurred(TS), TS<=TSS, time(TSS;TS).
:- obs_3_occurred.
CHAPTER 4. THE BIOPATHQA SYSTEM
135
In addition, the translator is also responsible for any rules needed to ease post-processing of
the query description. For example, for qualitative queries, a generic predicate tgt obs occurred(TS)
is generated that is true when the given qualitative description holds in an answer-set at time step
T S . The output of the translator is an ASP program, which when simulated using Clingo produces the (possibly) filtered trajectories of the pathway.
The Post Processor component is responsible for parsing the ASP answer sets, identifying
the correct atoms from it, extracting quantities from atom-bodies as necessary, organizing them
into a matrix form, and aggregating them as needed.
Figure 4.8: BioPathQA Graphical User Interface
The User Interface component is responsible for coordinating the processing of query statement. It presents the user with a graphical user interface shown in figure 4.8. The user types
a Pathway Specification (in BioPathQA-PL syntax), a Query Specification (in BioPathQA-QL
syntax), and simulation parameters. On pressing “Execute Query”, the user interface component
processes the query as prints results in the bottom box. Query evaluation differs by the type of
query. We describe the query evaluation methodology used below.
For non-comparative quantitative queries:
1. Pathway specification is parsed into a Guarded-Arc Petri Net model.
2. Query statement is parsed into an internal form.
3. Initial conditions from the query are applied to the pathway model.
4. Interventions are applied to the pathway model.
5. Modified pathway model is translated to ASP.
CHAPTER 4. THE BIOPATHQA SYSTEM
136
6. Internal observations are added to the ASP code as ASP constraints.
7. Answer sets of the ASP code are computed using Clingo.
8. Relevant atoms are extracted: fires/2 predicate for firing rate, holds/3 (or holds/4 –
colored tokens) predicate for fluent quantity or rate formulas.
9. Fluent value or firing-count values are extracted and organized as matrices with rows
representing answer-sets and columns representing time-steps.
10. Within answer-set interval or point value sub-select is done and the values converted to
rates or totals as needed.
11. If aggregation, such as average, minimum, or maximum is desired, it is performed over
rows of values from the last step.
12. If a value was specified in the query, it is compared against the computed aggregate for
boolean result.
13. If a value was not specified, the computed value is returned as the value satisfying the
query statement.
14. For queries over all trajectories, the same value must hold over all trajectories, otherwise,
only one match is required to satisfy the query.
For non-comparative qualitative queries:
1. Follow steps (1)-(6) of non-comparative quantitative queries.
2. Add rules for the query description for post-processing.
3. Answer sets of the ASP code are computed using Clingo.
4. Relevant atoms are extracted: tgt obs occurred/1 identifying the time step when the
observation within the query description is satisfied.
5. Truth value of the query observation is determined, including determining the truth value
over all trajectories.
For comparative quantitative queries:
1. Query statement is decomposed into two non-comparative quantitative sub-query statements as illustrated in section 4.6.1:
(a) A nominal sub-query which has the same initial conditions as the comparative query,
but none of its interventions or observations
(b) A modified sub-query which has the same initial conditions, interventions, and observations as the comparative query
CHAPTER 4. THE BIOPATHQA SYSTEM
137
both sub-query statements have the same query description, which is the non-aggregate
form of the comparative query description. Thus, a comparative average rate query is
translated to non-comparative average rate sub-queries.
2. Each sub-query statements is evaluated using steps (1)-(11) from the non-comparative
quantitative query processing.
3. A direction of change is computed by comparing the computed aggregate value for the
modified query statement to the nominal query statement.
4. If the comparative quantitative query has a direction specified, it is the compared against
the computed value for a boolean result.
5. If the comparative quantitative query did not have a direction specified, the computed
value is returned as the value satisfying the query statement.
For explanation queries with query description with formula of the form (4.97), it is expected
that the number of answer-sets will be quite large. So, we avoid generating all answer-sets before
processing them, instead we process them in-line as they are generated. It is a bit slower, but
leads to a smaller memory foot print.
1. Follow steps (1)-(6) of non-comparative quantitative queries.
2. Add rules for the query description for post-processing.
3. Compute answer sets of the ASP code using Clingo.
4. Extract relevant atoms:
(a) extract tgt obs occurred/1 identifying the time step when the query description
is satisfied
(b) extract holds/3 (or holds/4 – for colored tokens) at the same time-step as tgt obs occurred/1
to construct fluent-based conditions
5. Construct fluent-based conditions as explanation of the query observation.
6. If the query is over all trajectories, fluent-based conditions for each trajectory are intersected across trajectories to determine the minimum set of conditions explaining the query
observation.
Next we illustrate query processing through an execution trace of question (11). The following shows the encoding of the base case domain, which includes the pathway specification
from (4.321) with initial setup conditions from the query statement (4.322) applied:
%
#const nts=1.
time(0..nts).
CHAPTER 4. THE BIOPATHQA SYSTEM
%
#const ntok=1.
num(0..ntok).
%
place(bpg13).
place(dhap).
place(f16bp).
place(g3p).
%
trans(src_f16bp_1).
trans(t3).
trans(t4).
trans(t5a).
trans(t5b).
trans(t6).
%
tparc(src_f16bp_1,f16bp,1,TS) :- time(TS).
tparc(t3,f16bp,1,TS) :- time(TS).
ptarc(f16bp,t4,1,TS) :- time(TS).
tparc(t4,g3p,1,TS) :- time(TS).
tparc(t4,dhap,1,TS) :- time(TS).
ptarc(dhap,t5a,1,TS) :- time(TS).
tparc(t5a,g3p,1,TS) :- time(TS).
ptarc(g3p,t5b,1,TS) :- time(TS).
tparc(t5b,dhap,1,TS) :- time(TS).
ptarc(g3p,t6,1,TS) :- time(TS).
tparc(t6,bpg13,2,TS) :- time(TS).
%
holds(bpg13,0,0).
holds(dhap,0,0).
holds(f16bp,0,0).
holds(g3p,0,0).
%
%
could_not_have(T,TS) :- enabled(T,TS), not fires(T,TS), ptarc(P,T,Q,TS),
holds(P,QQ,TS), tot_decr(P,QQQ,TS), Q > QQ - QQQ.
:- not could_not_have(T,TS), time(TS), enabled(T,TS), not fires(T,TS), trans(T).
%
138
CHAPTER 4. THE BIOPATHQA SYSTEM
139
%
min(A,B,A) :- A<=B, num(A;B).
min(A,B,B) :- B<=A, num(A;B).
#hide min/3.
%
holdspos(P):- holds(P,N,0), place(P), num(N), N > 0.
holds(P,0,0) :- place(P), not holdspos(P).
%
notenabled(T,TS) :- ptarc(P,T,N,TS), holds(P,Q,TS), Q < N, place(P), trans(T),
time(TS), num(N), num(Q).
%
enabled(T,TS) :- trans(T), time(TS), not notenabled(T, TS).
%
{ fires(T,TS) } :- enabled(T,TS), trans(T), time(TS).
%
add(P,Q,T,TS) :- fires(T,TS), tparc(T,P,Q,TS), time(TS).
del(P,Q,T,TS) :- fires(T,TS), ptarc(P,T,Q,TS), time(TS).
%
tot_incr(P,QQ,TS) :- QQ = #sum[add(P,Q,T,TS) = Q : num(Q) : trans(T)],
time(TS), num(QQ), place(P).
tot_decr(P,QQ,TS) :- QQ = #sum[del(P,Q,T,TS) = Q : num(Q) : trans(T)],
time(TS), num(QQ), place(P).
%
holds(P,Q,TS+1) :- holds(P,Q1,TS), tot_incr(P,Q2,TS), tot_decr(P,Q3,TS),
Q=Q1+Q2-Q3, place(P), num(Q;Q1;Q2;Q3), time(TS), time(TS+1).
%
%
consumesmore(P,TS) :- holds(P,Q,TS), tot_decr(P,Q1,TS), Q1 > Q.
consumesmore :- consumesmore(P,TS).
:- consumesmore.
The following shows encoding of the alternate case domain, which consists of the pathway specification from (4.321) with initial setup conditions and interventions applied; and any
internal observations from the query statement (4.322) added:
%
#const nts=1.
time(0..nts).
%
#const ntok=1.
num(0..ntok).
CHAPTER 4. THE BIOPATHQA SYSTEM
%
place(bpg13).
place(dhap).
place(f16bp).
place(g3p).
%
trans(reset_dhap_1).
trans(src_f16bp_1).
trans(t3).
trans(t4).
trans(t5a).
trans(t5b).
trans(t6).
%
ptarc(dhap,reset_dhap_1,Q,TS) :- holds(dhap,Q,TS), Q>0, time(TS).
:- enabled(reset_dhap_1,TS), not fires(reset_dhap_1,TS), time(TS).
tparc(src_f16bp_1,f16bp,1,TS) :- time(TS).
tparc(t3,f16bp,1,TS) :- time(TS).
ptarc(f16bp,t4,1,TS) :- time(TS).
tparc(t4,g3p,1,TS) :- time(TS).
tparc(t4,dhap,1,TS) :- time(TS).
ptarc(dhap,t5a,1,TS) :- time(TS).
tparc(t5a,g3p,1,TS) :- time(TS).
ptarc(g3p,t5b,1,TS) :- time(TS).
tparc(t5b,dhap,1,TS) :- time(TS).
ptarc(g3p,t6,1,TS) :- time(TS).
tparc(t6,bpg13,2,TS) :- time(TS).
%
holds(bpg13,0,0).
holds(dhap,0,0).
holds(f16bp,0,0).
holds(g3p,0,0).
%
%
could_not_have(T,TS) :- enabled(T,TS), not fires(T,TS), ptarc(P,T,Q,TS),
holds(P,QQ,TS), tot_decr(P,QQQ,TS), Q > QQ - QQQ.
:- not could_not_have(T,TS), time(TS), enabled(T,TS), not fires(T,TS), trans(T).
%
140
CHAPTER 4. THE BIOPATHQA SYSTEM
141
%
min(A,B,A) :- A<=B, num(A;B).
min(A,B,B) :- B<=A, num(A;B).
#hide min/3.
%
holdspos(P):- holds(P,N,0), place(P), num(N), N > 0.
holds(P,0,0) :- place(P), not holdspos(P).
%
notenabled(T,TS) :- ptarc(P,T,N,TS), holds(P,Q,TS), Q < N, place(P), trans(T),
time(TS), num(N), num(Q).
%
enabled(T,TS) :- trans(T), time(TS), not notenabled(T, TS).
%
{ fires(T,TS) } :- enabled(T,TS), trans(T), time(TS).
%
add(P,Q,T,TS) :- fires(T,TS), tparc(T,P,Q,TS), time(TS).
del(P,Q,T,TS) :- fires(T,TS), ptarc(P,T,Q,TS), time(TS).
%
tot_incr(P,QQ,TS) :- QQ = #sum[add(P,Q,T,TS) = Q : num(Q) : trans(T)],
time(TS), num(QQ), place(P).
tot_decr(P,QQ,TS) :- QQ = #sum[del(P,Q,T,TS) = Q : num(Q) : trans(T)],
time(TS), num(QQ), place(P).
%
holds(P,Q,TS+1) :- holds(P,Q1,TS), tot_incr(P,Q2,TS), tot_decr(P,Q3,TS),
Q=Q1+Q2-Q3, place(P), num(Q;Q1;Q2;Q3), time(TS), time(TS+1).
%
%
consumesmore(P,TS) :- holds(P,Q,TS), tot_decr(P,Q1,TS), Q1 > Q.
consumesmore :- consumesmore(P,TS).
:- consumesmore.
Both programs are simulated for 5 time-steps and 20 max tokens using the following Clingo
command:
clingo 0 -cntok=20 -cnts=5 program.lp
Answer sets of the base case are as follows:
Answer: 1
holds(bpg13,0,0) holds(dhap,0,0)
fires(src_f16bp_1,0) fires(t3,0)
holds(bpg13,0,1) holds(dhap,0,1)
fires(src_f16bp_1,1) fires(t3,1)
holds(bpg13,0,2) holds(dhap,1,2)
holds(f16bp,0,0) holds(g3p,0,0)
holds(f16bp,2,1) holds(g3p,0,1)
fires(t4,1)
holds(f16bp,3,2) holds(g3p,1,2)
CHAPTER 4. THE BIOPATHQA SYSTEM
fires(src_f16bp_1,2) fires(t3,2)
holds(bpg13,0,3) holds(dhap,2,3)
fires(src_f16bp_1,3) fires(t3,3)
holds(bpg13,2,4) holds(dhap,3,4)
fires(src_f16bp_1,4) fires(t3,4)
holds(bpg13,4,5) holds(dhap,4,5)
fires(src_f16bp_1,5) fires(t3,5)
Answer: 2
holds(bpg13,0,0) holds(dhap,0,0)
fires(src_f16bp_1,0) fires(t3,0)
holds(bpg13,0,1) holds(dhap,0,1)
fires(src_f16bp_1,1) fires(t3,1)
holds(bpg13,0,2) holds(dhap,1,2)
fires(src_f16bp_1,2) fires(t3,2)
holds(bpg13,2,3) holds(dhap,1,3)
fires(src_f16bp_1,3) fires(t3,3)
holds(bpg13,4,4) holds(dhap,2,4)
fires(src_f16bp_1,4) fires(t3,4)
holds(bpg13,6,5) holds(dhap,3,5)
fires(src_f16bp_1,5) fires(t3,5)
142
fires(t4,2) fires(t5a,2) fires(t5b,2)
holds(f16bp,4,3) holds(g3p,2,3)
fires(t4,3) fires(t5a,3) fires(t5b,3) fires(t6,3)
holds(f16bp,5,4) holds(g3p,2,4)
fires(t4,4) fires(t5a,4) fires(t5b,4) fires(t6,4)
holds(f16bp,6,5) holds(g3p,2,5)
fires(t4,5) fires(t5a,5) fires(t5b,5) fires(t6,5)
holds(f16bp,0,0) holds(g3p,0,0)
holds(f16bp,2,1) holds(g3p,0,1)
fires(t4,1)
holds(f16bp,3,2) holds(g3p,1,2)
fires(t4,2) fires(t5a,2) fires(t6,2)
holds(f16bp,4,3) holds(g3p,2,3)
fires(t4,3) fires(t5a,3) fires(t5b,3) fires(t6,3)
holds(f16bp,5,4) holds(g3p,2,4)
fires(t4,4) fires(t5a,4) fires(t5b,4) fires(t6,4)
holds(f16bp,6,5) holds(g3p,2,5)
fires(t4,5) fires(t5a,5) fires(t5b,5) fires(t6,5)
Answer sets of the alternate case are as follows:
Answer: 1
holds(bpg13,0,0) holds(dhap,0,0) holds(f16bp,0,0) holds(g3p,0,0)
fires(reset_dhap_1,0) fires(src_f16bp_1,0) fires(t3,0)
holds(bpg13,0,1) holds(dhap,0,1) holds(f16bp,2,1) holds(g3p,0,1)
fires(reset_dhap_1,1) fires(src_f16bp_1,1) fires(t3,1) fires(t4,1)
holds(bpg13,0,2) holds(dhap,1,2) holds(f16bp,3,2) holds(g3p,1,2)
fires(reset_dhap_1,2) fires(src_f16bp_1,2) fires(t3,2) fires(t4,2)
holds(bpg13,2,3) holds(dhap,1,3) holds(f16bp,4,3) holds(g3p,1,3)
fires(reset_dhap_1,3) fires(src_f16bp_1,3) fires(t3,3) fires(t4,3)
holds(bpg13,2,4) holds(dhap,2,4) holds(f16bp,5,4) holds(g3p,1,4)
fires(reset_dhap_1,4) fires(src_f16bp_1,4) fires(t3,4) fires(t4,4)
holds(bpg13,2,5) holds(dhap,2,5) holds(f16bp,6,5) holds(g3p,1,5)
fires(reset_dhap_1,5) fires(src_f16bp_1,5) fires(t3,5) fires(t4,5)
Answer: 2
holds(bpg13,0,0) holds(dhap,0,0) holds(f16bp,0,0) holds(g3p,0,0)
fires(reset_dhap_1,0) fires(src_f16bp_1,0) fires(t3,0)
holds(bpg13,0,1) holds(dhap,0,1) holds(f16bp,2,1) holds(g3p,0,1)
fires(reset_dhap_1,1) fires(src_f16bp_1,1) fires(t3,1) fires(t4,1)
holds(bpg13,0,2) holds(dhap,1,2) holds(f16bp,3,2) holds(g3p,1,2)
fires(reset_dhap_1,2) fires(src_f16bp_1,2) fires(t3,2) fires(t4,2)
holds(bpg13,2,3) holds(dhap,1,3) holds(f16bp,4,3) holds(g3p,1,3)
fires(reset_dhap_1,3) fires(src_f16bp_1,3) fires(t3,3) fires(t4,3)
holds(bpg13,2,4) holds(dhap,2,4) holds(f16bp,5,4) holds(g3p,1,4)
fires(reset_dhap_1,4) fires(src_f16bp_1,4) fires(t3,4) fires(t4,4)
holds(bpg13,2,5) holds(dhap,2,5) holds(f16bp,6,5) holds(g3p,1,5)
fires(reset_dhap_1,5) fires(src_f16bp_1,5) fires(t3,5) fires(t4,5)
Answer: 3
holds(bpg13,0,0) holds(dhap,0,0) holds(f16bp,0,0) holds(g3p,0,0)
fires(reset_dhap_1,0) fires(src_f16bp_1,0) fires(t3,0)
fires(t6,2)
fires(t5b,3)
fires(t5b,4)
fires(t5b,5)
fires(t6,2)
fires(t5b,3)
fires(t5b,4)
fires(t6,5)
CHAPTER 4. THE BIOPATHQA SYSTEM
holds(bpg13,0,1) holds(dhap,0,1) holds(f16bp,2,1) holds(g3p,0,1)
fires(reset_dhap_1,1) fires(src_f16bp_1,1) fires(t3,1) fires(t4,1)
holds(bpg13,0,2) holds(dhap,1,2) holds(f16bp,3,2) holds(g3p,1,2)
fires(reset_dhap_1,2) fires(src_f16bp_1,2) fires(t3,2) fires(t4,2)
holds(bpg13,2,3) holds(dhap,1,3) holds(f16bp,4,3) holds(g3p,1,3)
fires(reset_dhap_1,3) fires(src_f16bp_1,3) fires(t3,3) fires(t4,3)
holds(bpg13,2,4) holds(dhap,2,4) holds(f16bp,5,4) holds(g3p,1,4)
fires(reset_dhap_1,4) fires(src_f16bp_1,4) fires(t3,4) fires(t4,4)
holds(bpg13,4,5) holds(dhap,1,5) holds(f16bp,6,5) holds(g3p,1,5)
fires(reset_dhap_1,5) fires(src_f16bp_1,5) fires(t3,5) fires(t4,5)
Answer: 4
holds(bpg13,0,0) holds(dhap,0,0) holds(f16bp,0,0) holds(g3p,0,0)
fires(reset_dhap_1,0) fires(src_f16bp_1,0) fires(t3,0)
holds(bpg13,0,1) holds(dhap,0,1) holds(f16bp,2,1) holds(g3p,0,1)
fires(reset_dhap_1,1) fires(src_f16bp_1,1) fires(t3,1) fires(t4,1)
holds(bpg13,0,2) holds(dhap,1,2) holds(f16bp,3,2) holds(g3p,1,2)
fires(reset_dhap_1,2) fires(src_f16bp_1,2) fires(t3,2) fires(t4,2)
holds(bpg13,2,3) holds(dhap,1,3) holds(f16bp,4,3) holds(g3p,1,3)
fires(reset_dhap_1,3) fires(src_f16bp_1,3) fires(t3,3) fires(t4,3)
holds(bpg13,2,4) holds(dhap,2,4) holds(f16bp,5,4) holds(g3p,1,4)
fires(reset_dhap_1,4) fires(src_f16bp_1,4) fires(t3,4) fires(t4,4)
holds(bpg13,4,5) holds(dhap,1,5) holds(f16bp,6,5) holds(g3p,1,5)
fires(reset_dhap_1,5) fires(src_f16bp_1,5) fires(t3,5) fires(t4,5)
Answer: 5
holds(bpg13,0,0) holds(dhap,0,0) holds(f16bp,0,0) holds(g3p,0,0)
fires(reset_dhap_1,0) fires(src_f16bp_1,0) fires(t3,0)
holds(bpg13,0,1) holds(dhap,0,1) holds(f16bp,2,1) holds(g3p,0,1)
fires(reset_dhap_1,1) fires(src_f16bp_1,1) fires(t3,1) fires(t4,1)
holds(bpg13,0,2) holds(dhap,1,2) holds(f16bp,3,2) holds(g3p,1,2)
fires(reset_dhap_1,2) fires(src_f16bp_1,2) fires(t3,2) fires(t4,2)
holds(bpg13,0,3) holds(dhap,2,3) holds(f16bp,4,3) holds(g3p,1,3)
fires(reset_dhap_1,3) fires(src_f16bp_1,3) fires(t3,3) fires(t4,3)
holds(bpg13,0,4) holds(dhap,2,4) holds(f16bp,5,4) holds(g3p,1,4)
fires(reset_dhap_1,4) fires(src_f16bp_1,4) fires(t3,4) fires(t4,4)
holds(bpg13,0,5) holds(dhap,2,5) holds(f16bp,6,5) holds(g3p,1,5)
fires(reset_dhap_1,5) fires(src_f16bp_1,5) fires(t3,5) fires(t4,5)
Answer: 6
holds(bpg13,0,0) holds(dhap,0,0) holds(f16bp,0,0) holds(g3p,0,0)
fires(reset_dhap_1,0) fires(src_f16bp_1,0) fires(t3,0)
holds(bpg13,0,1) holds(dhap,0,1) holds(f16bp,2,1) holds(g3p,0,1)
fires(reset_dhap_1,1) fires(src_f16bp_1,1) fires(t3,1) fires(t4,1)
holds(bpg13,0,2) holds(dhap,1,2) holds(f16bp,3,2) holds(g3p,1,2)
fires(reset_dhap_1,2) fires(src_f16bp_1,2) fires(t3,2) fires(t4,2)
holds(bpg13,0,3) holds(dhap,2,3) holds(f16bp,4,3) holds(g3p,1,3)
fires(reset_dhap_1,3) fires(src_f16bp_1,3) fires(t3,3) fires(t4,3)
holds(bpg13,0,4) holds(dhap,2,4) holds(f16bp,5,4) holds(g3p,1,4)
fires(reset_dhap_1,4) fires(src_f16bp_1,4) fires(t3,4) fires(t4,4)
holds(bpg13,0,5) holds(dhap,2,5) holds(f16bp,6,5) holds(g3p,1,5)
fires(reset_dhap_1,5) fires(src_f16bp_1,5) fires(t3,5) fires(t4,5)
Answer: 7
holds(bpg13,0,0) holds(dhap,0,0) holds(f16bp,0,0) holds(g3p,0,0)
143
fires(t6,2)
fires(t5b,3)
fires(t6,4)
fires(t5b,5)
fires(t6,2)
fires(t5b,3)
fires(t6,4)
fires(t6,5)
fires(t5b,2)
fires(t5b,3)
fires(t5b,4)
fires(t5b,5)
fires(t5b,2)
fires(t5b,3)
fires(t5b,4)
fires(t6,5)
CHAPTER 4. THE BIOPATHQA SYSTEM
fires(reset_dhap_1,0) fires(src_f16bp_1,0) fires(t3,0)
holds(bpg13,0,1) holds(dhap,0,1) holds(f16bp,2,1) holds(g3p,0,1)
fires(reset_dhap_1,1) fires(src_f16bp_1,1) fires(t3,1) fires(t4,1)
holds(bpg13,0,2) holds(dhap,1,2) holds(f16bp,3,2) holds(g3p,1,2)
fires(reset_dhap_1,2) fires(src_f16bp_1,2) fires(t3,2) fires(t4,2)
holds(bpg13,0,3) holds(dhap,2,3) holds(f16bp,4,3) holds(g3p,1,3)
fires(reset_dhap_1,3) fires(src_f16bp_1,3) fires(t3,3) fires(t4,3)
holds(bpg13,0,4) holds(dhap,2,4) holds(f16bp,5,4) holds(g3p,1,4)
fires(reset_dhap_1,4) fires(src_f16bp_1,4) fires(t3,4) fires(t4,4)
holds(bpg13,2,5) holds(dhap,1,5) holds(f16bp,6,5) holds(g3p,1,5)
fires(reset_dhap_1,5) fires(src_f16bp_1,5) fires(t3,5) fires(t4,5)
Answer: 8
holds(bpg13,0,0) holds(dhap,0,0) holds(f16bp,0,0) holds(g3p,0,0)
fires(reset_dhap_1,0) fires(src_f16bp_1,0) fires(t3,0)
holds(bpg13,0,1) holds(dhap,0,1) holds(f16bp,2,1) holds(g3p,0,1)
fires(reset_dhap_1,1) fires(src_f16bp_1,1) fires(t3,1) fires(t4,1)
holds(bpg13,0,2) holds(dhap,1,2) holds(f16bp,3,2) holds(g3p,1,2)
fires(reset_dhap_1,2) fires(src_f16bp_1,2) fires(t3,2) fires(t4,2)
holds(bpg13,0,3) holds(dhap,2,3) holds(f16bp,4,3) holds(g3p,1,3)
fires(reset_dhap_1,3) fires(src_f16bp_1,3) fires(t3,3) fires(t4,3)
holds(bpg13,0,4) holds(dhap,2,4) holds(f16bp,5,4) holds(g3p,1,4)
fires(reset_dhap_1,4) fires(src_f16bp_1,4) fires(t3,4) fires(t4,4)
holds(bpg13,2,5) holds(dhap,1,5) holds(f16bp,6,5) holds(g3p,1,5)
fires(reset_dhap_1,5) fires(src_f16bp_1,5) fires(t3,5) fires(t4,5)
Answer: 9
holds(bpg13,0,0) holds(dhap,0,0) holds(f16bp,0,0) holds(g3p,0,0)
fires(reset_dhap_1,0) fires(src_f16bp_1,0) fires(t3,0)
holds(bpg13,0,1) holds(dhap,0,1) holds(f16bp,2,1) holds(g3p,0,1)
fires(reset_dhap_1,1) fires(src_f16bp_1,1) fires(t3,1) fires(t4,1)
holds(bpg13,0,2) holds(dhap,1,2) holds(f16bp,3,2) holds(g3p,1,2)
fires(reset_dhap_1,2) fires(src_f16bp_1,2) fires(t3,2) fires(t4,2)
holds(bpg13,0,3) holds(dhap,2,3) holds(f16bp,4,3) holds(g3p,1,3)
fires(reset_dhap_1,3) fires(src_f16bp_1,3) fires(t3,3) fires(t4,3)
holds(bpg13,2,4) holds(dhap,1,4) holds(f16bp,5,4) holds(g3p,1,4)
fires(reset_dhap_1,4) fires(src_f16bp_1,4) fires(t3,4) fires(t4,4)
holds(bpg13,2,5) holds(dhap,2,5) holds(f16bp,6,5) holds(g3p,1,5)
fires(reset_dhap_1,5) fires(src_f16bp_1,5) fires(t3,5) fires(t4,5)
Answer: 10
holds(bpg13,0,0) holds(dhap,0,0) holds(f16bp,0,0) holds(g3p,0,0)
fires(reset_dhap_1,0) fires(src_f16bp_1,0) fires(t3,0)
holds(bpg13,0,1) holds(dhap,0,1) holds(f16bp,2,1) holds(g3p,0,1)
fires(reset_dhap_1,1) fires(src_f16bp_1,1) fires(t3,1) fires(t4,1)
holds(bpg13,0,2) holds(dhap,1,2) holds(f16bp,3,2) holds(g3p,1,2)
fires(reset_dhap_1,2) fires(src_f16bp_1,2) fires(t3,2) fires(t4,2)
holds(bpg13,0,3) holds(dhap,2,3) holds(f16bp,4,3) holds(g3p,1,3)
fires(reset_dhap_1,3) fires(src_f16bp_1,3) fires(t3,3) fires(t4,3)
holds(bpg13,2,4) holds(dhap,1,4) holds(f16bp,5,4) holds(g3p,1,4)
fires(reset_dhap_1,4) fires(src_f16bp_1,4) fires(t3,4) fires(t4,4)
holds(bpg13,2,5) holds(dhap,2,5) holds(f16bp,6,5) holds(g3p,1,5)
fires(reset_dhap_1,5) fires(src_f16bp_1,5) fires(t3,5) fires(t4,5)
Answer: 11
144
fires(t5b,2)
fires(t5b,3)
fires(t6,4)
fires(t5b,5)
fires(t5b,2)
fires(t5b,3)
fires(t6,4)
fires(t6,5)
fires(t5b,2)
fires(t6,3)
fires(t5b,4)
fires(t5b,5)
fires(t5b,2)
fires(t6,3)
fires(t5b,4)
fires(t6,5)
CHAPTER 4. THE BIOPATHQA SYSTEM
holds(bpg13,0,0) holds(dhap,0,0) holds(f16bp,0,0) holds(g3p,0,0)
fires(reset_dhap_1,0) fires(src_f16bp_1,0) fires(t3,0)
holds(bpg13,0,1) holds(dhap,0,1) holds(f16bp,2,1) holds(g3p,0,1)
fires(reset_dhap_1,1) fires(src_f16bp_1,1) fires(t3,1) fires(t4,1)
holds(bpg13,0,2) holds(dhap,1,2) holds(f16bp,3,2) holds(g3p,1,2)
fires(reset_dhap_1,2) fires(src_f16bp_1,2) fires(t3,2) fires(t4,2)
holds(bpg13,0,3) holds(dhap,2,3) holds(f16bp,4,3) holds(g3p,1,3)
fires(reset_dhap_1,3) fires(src_f16bp_1,3) fires(t3,3) fires(t4,3)
holds(bpg13,2,4) holds(dhap,1,4) holds(f16bp,5,4) holds(g3p,1,4)
fires(reset_dhap_1,4) fires(src_f16bp_1,4) fires(t3,4) fires(t4,4)
holds(bpg13,4,5) holds(dhap,1,5) holds(f16bp,6,5) holds(g3p,1,5)
fires(reset_dhap_1,5) fires(src_f16bp_1,5) fires(t3,5) fires(t4,5)
Answer: 12
holds(bpg13,0,0) holds(dhap,0,0) holds(f16bp,0,0) holds(g3p,0,0)
fires(reset_dhap_1,0) fires(src_f16bp_1,0) fires(t3,0)
holds(bpg13,0,1) holds(dhap,0,1) holds(f16bp,2,1) holds(g3p,0,1)
fires(reset_dhap_1,1) fires(src_f16bp_1,1) fires(t3,1) fires(t4,1)
holds(bpg13,0,2) holds(dhap,1,2) holds(f16bp,3,2) holds(g3p,1,2)
fires(reset_dhap_1,2) fires(src_f16bp_1,2) fires(t3,2) fires(t4,2)
holds(bpg13,0,3) holds(dhap,2,3) holds(f16bp,4,3) holds(g3p,1,3)
fires(reset_dhap_1,3) fires(src_f16bp_1,3) fires(t3,3) fires(t4,3)
holds(bpg13,2,4) holds(dhap,1,4) holds(f16bp,5,4) holds(g3p,1,4)
fires(reset_dhap_1,4) fires(src_f16bp_1,4) fires(t3,4) fires(t4,4)
holds(bpg13,4,5) holds(dhap,1,5) holds(f16bp,6,5) holds(g3p,1,5)
fires(reset_dhap_1,5) fires(src_f16bp_1,5) fires(t3,5) fires(t4,5)
Answer: 13
holds(bpg13,0,0) holds(dhap,0,0) holds(f16bp,0,0) holds(g3p,0,0)
fires(reset_dhap_1,0) fires(src_f16bp_1,0) fires(t3,0)
holds(bpg13,0,1) holds(dhap,0,1) holds(f16bp,2,1) holds(g3p,0,1)
fires(reset_dhap_1,1) fires(src_f16bp_1,1) fires(t3,1) fires(t4,1)
holds(bpg13,0,2) holds(dhap,1,2) holds(f16bp,3,2) holds(g3p,1,2)
fires(reset_dhap_1,2) fires(src_f16bp_1,2) fires(t3,2) fires(t4,2)
holds(bpg13,2,3) holds(dhap,1,3) holds(f16bp,4,3) holds(g3p,1,3)
fires(reset_dhap_1,3) fires(src_f16bp_1,3) fires(t3,3) fires(t4,3)
holds(bpg13,4,4) holds(dhap,1,4) holds(f16bp,5,4) holds(g3p,1,4)
fires(reset_dhap_1,4) fires(src_f16bp_1,4) fires(t3,4) fires(t4,4)
holds(bpg13,4,5) holds(dhap,2,5) holds(f16bp,6,5) holds(g3p,1,5)
fires(reset_dhap_1,5) fires(src_f16bp_1,5) fires(t3,5) fires(t4,5)
Answer: 14
holds(bpg13,0,0) holds(dhap,0,0) holds(f16bp,0,0) holds(g3p,0,0)
fires(reset_dhap_1,0) fires(src_f16bp_1,0) fires(t3,0)
holds(bpg13,0,1) holds(dhap,0,1) holds(f16bp,2,1) holds(g3p,0,1)
fires(reset_dhap_1,1) fires(src_f16bp_1,1) fires(t3,1) fires(t4,1)
holds(bpg13,0,2) holds(dhap,1,2) holds(f16bp,3,2) holds(g3p,1,2)
fires(reset_dhap_1,2) fires(src_f16bp_1,2) fires(t3,2) fires(t4,2)
holds(bpg13,2,3) holds(dhap,1,3) holds(f16bp,4,3) holds(g3p,1,3)
fires(reset_dhap_1,3) fires(src_f16bp_1,3) fires(t3,3) fires(t4,3)
holds(bpg13,4,4) holds(dhap,1,4) holds(f16bp,5,4) holds(g3p,1,4)
fires(reset_dhap_1,4) fires(src_f16bp_1,4) fires(t3,4) fires(t4,4)
holds(bpg13,4,5) holds(dhap,2,5) holds(f16bp,6,5) holds(g3p,1,5)
fires(reset_dhap_1,5) fires(src_f16bp_1,5) fires(t3,5) fires(t4,5)
145
fires(t5b,2)
fires(t6,3)
fires(t6,4)
fires(t5b,5)
fires(t5b,2)
fires(t6,3)
fires(t6,4)
fires(t6,5)
fires(t6,2)
fires(t6,3)
fires(t5b,4)
fires(t6,5)
fires(t6,2)
fires(t6,3)
fires(t5b,4)
fires(t5b,5)
CHAPTER 4. THE BIOPATHQA SYSTEM
Answer: 15
holds(bpg13,0,0) holds(dhap,0,0) holds(f16bp,0,0) holds(g3p,0,0)
fires(reset_dhap_1,0) fires(src_f16bp_1,0) fires(t3,0)
holds(bpg13,0,1) holds(dhap,0,1) holds(f16bp,2,1) holds(g3p,0,1)
fires(reset_dhap_1,1) fires(src_f16bp_1,1) fires(t3,1) fires(t4,1)
holds(bpg13,0,2) holds(dhap,1,2) holds(f16bp,3,2) holds(g3p,1,2)
fires(reset_dhap_1,2) fires(src_f16bp_1,2) fires(t3,2) fires(t4,2)
holds(bpg13,2,3) holds(dhap,1,3) holds(f16bp,4,3) holds(g3p,1,3)
fires(reset_dhap_1,3) fires(src_f16bp_1,3) fires(t3,3) fires(t4,3)
holds(bpg13,4,4) holds(dhap,1,4) holds(f16bp,5,4) holds(g3p,1,4)
fires(reset_dhap_1,4) fires(src_f16bp_1,4) fires(t3,4) fires(t4,4)
holds(bpg13,6,5) holds(dhap,1,5) holds(f16bp,6,5) holds(g3p,1,5)
fires(reset_dhap_1,5) fires(src_f16bp_1,5) fires(t3,5) fires(t4,5)
Answer: 16
holds(bpg13,0,0) holds(dhap,0,0) holds(f16bp,0,0) holds(g3p,0,0)
fires(reset_dhap_1,0) fires(src_f16bp_1,0) fires(t3,0)
holds(bpg13,0,1) holds(dhap,0,1) holds(f16bp,2,1) holds(g3p,0,1)
fires(reset_dhap_1,1) fires(src_f16bp_1,1) fires(t3,1) fires(t4,1)
holds(bpg13,0,2) holds(dhap,1,2) holds(f16bp,3,2) holds(g3p,1,2)
fires(reset_dhap_1,2) fires(src_f16bp_1,2) fires(t3,2) fires(t4,2)
holds(bpg13,2,3) holds(dhap,1,3) holds(f16bp,4,3) holds(g3p,1,3)
fires(reset_dhap_1,3) fires(src_f16bp_1,3) fires(t3,3) fires(t4,3)
holds(bpg13,4,4) holds(dhap,1,4) holds(f16bp,5,4) holds(g3p,1,4)
fires(reset_dhap_1,4) fires(src_f16bp_1,4) fires(t3,4) fires(t4,4)
holds(bpg13,6,5) holds(dhap,1,5) holds(f16bp,6,5) holds(g3p,1,5)
fires(reset_dhap_1,5) fires(src_f16bp_1,5) fires(t3,5) fires(t4,5)
Atoms selected for bpg13 quantity extraction for the nominal case:
[holds(bpg13,0,0),holds(bpg13,0,1),holds(bpg13,0,2),
holds(bpg13,0,3),holds(bpg13,2,4),holds(bpg13,4,5)],
[holds(bpg13,0,0),holds(bpg13,0,1),holds(bpg13,0,2),
holds(bpg13,2,3),holds(bpg13,4,4),holds(bpg13,6,5)]
Atoms selected for bpg13 quantity extraction for the modified case:
[holds(bpg13,0,0),holds(bpg13,0,1),holds(bpg13,0,2),
holds(bpg13,2,3),holds(bpg13,2,4),holds(bpg13,2,5)],
[holds(bpg13,0,0),holds(bpg13,0,1),holds(bpg13,0,2),
holds(bpg13,2,3),holds(bpg13,2,4),holds(bpg13,2,5)],
[holds(bpg13,0,0),holds(bpg13,0,1),holds(bpg13,0,2),
holds(bpg13,2,3),holds(bpg13,2,4),holds(bpg13,4,5)],
[holds(bpg13,0,0),holds(bpg13,0,1),holds(bpg13,0,2),
holds(bpg13,2,3),holds(bpg13,2,4),holds(bpg13,4,5)],
[holds(bpg13,0,0),holds(bpg13,0,1),holds(bpg13,0,2),
holds(bpg13,0,3),holds(bpg13,0,4),holds(bpg13,0,5)],
[holds(bpg13,0,0),holds(bpg13,0,1),holds(bpg13,0,2),
holds(bpg13,0,3),holds(bpg13,0,4),holds(bpg13,0,5)],
[holds(bpg13,0,0),holds(bpg13,0,1),holds(bpg13,0,2),
holds(bpg13,0,3),holds(bpg13,0,4),holds(bpg13,2,5)],
[holds(bpg13,0,0),holds(bpg13,0,1),holds(bpg13,0,2),
holds(bpg13,0,3),holds(bpg13,0,4),holds(bpg13,2,5)],
[holds(bpg13,0,0),holds(bpg13,0,1),holds(bpg13,0,2),
146
fires(t6,2)
fires(t6,3)
fires(t6,4)
fires(t5b,5)
fires(t6,2)
fires(t6,3)
fires(t6,4)
fires(t6,5)
CHAPTER 4. THE BIOPATHQA SYSTEM
147
holds(bpg13,0,3),holds(bpg13,2,4),holds(bpg13,2,5)],
[holds(bpg13,0,0),holds(bpg13,0,1),holds(bpg13,0,2),
holds(bpg13,0,3),holds(bpg13,2,4),holds(bpg13,2,5)],
[holds(bpg13,0,0),holds(bpg13,0,1),holds(bpg13,0,2),
holds(bpg13,0,3),holds(bpg13,2,4),holds(bpg13,4,5)],
[holds(bpg13,0,0),holds(bpg13,0,1),holds(bpg13,0,2),
holds(bpg13,0,3),holds(bpg13,2,4),holds(bpg13,4,5)],
[holds(bpg13,0,0),holds(bpg13,0,1),holds(bpg13,0,2),
holds(bpg13,2,3),holds(bpg13,4,4),holds(bpg13,4,5)],
[holds(bpg13,0,0),holds(bpg13,0,1),holds(bpg13,0,2),
holds(bpg13,2,3),holds(bpg13,4,4),holds(bpg13,4,5)],
[holds(bpg13,0,0),holds(bpg13,0,1),holds(bpg13,0,2),
holds(bpg13,2,3),holds(bpg13,4,4),holds(bpg13,6,5)],
[holds(bpg13,0,0),holds(bpg13,0,1),holds(bpg13,0,2),
holds(bpg13,2,3),holds(bpg13,4,4),holds(bpg13,6,5)]
Progression from raw matrix of bpg13 quantities in various answer-sets (rows) at various
simulation
steps (columns)
to the average aggregate for the nominal case:
#
" finally
# to rates and
"
0.8
0 0 0 0 2 4
= average => 1.0
= rate =>
1.2
0 0 0 2 4 6
Progression from raw matrix of bpg13 quantities in various answer-sets (rows) at various
simulation steps (columns) to rates and finally to the average aggregate for the modified case:
0 0 0 2 2 2
0.4
0 0 0 2 2 2
0.4
0 0 0 2 2 4
0.8
0 0 0 2 2 4
0.8
0 0 0 0 0 0
0.
0.
0 0 0 0 0 0
0.4
0 0 0 0 0 2
0 0 0 0 0 2
0.4
0 0 0 0 2 2 = rate => 0.4 = average => 0.6
0 0 0 0 2 2
0.4
0 0 0 0 2 4
0.8
0 0 0 0 2 4
0.8
0 0 0 2 4 4
0.8
0.8
0 0 0 2 4 4
1.2
0 0 0 2 4 6
1.2
0 0 0 2 4 6
We find that d =0 <0 comparing the modified case rate of 0.6 to the nominal case rate of 1.0.
Since the direction d was an unknown in the query statement, our system generates produces
the full query specification with d replaced by 0 <0 as follows:
direction of change in average rate of production of ’bpg13’ is ’<’ (0.6<1)
when observed between time step 0 and time step 5
comparing nominal pathway with modified pathway obtained;
due to interventions:
remove ’dhap’ as soon as produced;
CHAPTER 4. THE BIOPATHQA SYSTEM
148
using initial setup:
continuously supply ’f16bp’ in quantity 1;
4.6.11
Evaluation Methodolgy
A direct comparison against other tools is not possible, since most other programs explore one
state evolution, while we explore all possible state evolutions. In addition ASP has to ground
the program completely, irrespective of whether we are computing one answer or all. So, to
evaluate our system, we compare our results for the questions from the 2nd Deep KR Challenge
against the answers they have provided. Our results in essence match the responses given for
the questions.
4.7
Related Work
In this section, we relate our high level language with other high level action languages.
4.7.1
Comparison with π-Calculus
π-calculus is a formalism that is used to model biological systems and pathways by modeling
biological systems as mobile communication systems. We use the biological model described
by Regev (2001) for comparison against our system. In their model they represent molecules
and their domains as computational processes, interacting elements of molecules as communication channels (two molecules interact if they fit together as in a lock-and-key mechanism),
and reactions as communication through channel transmission.
• π-calculus models have the ability of changing their structure during simulation. Our system on the other hand only allows modification of the pathway at the start of simulation.
• Regular π-calculus models appear qualitative in nature. However, stochastic extensions
allow representation of quantitative data Priami et al. (2001). In contrast, the focus of our
system is on the quantitative+qualitative representation using numeric fluents.
• It is unclear how one can easily implement maximal-parallelism of our system in πcalculus, where a maximum number of simultaneous actions occur such that they do not
cause a conflict. Where, a set of actions is said to be in conflict if their simultaneous
execution will cause a fluent to become negative.
4.7.2
Comparison with Action Language A
Action language A Gelfond and Lifschitz (1993) is a formalism that has been used to model
biological systems and pathways. First we give a brief overview of A in an intuitive manner.
Assume two sets of disjoint symbols containing fluent names and action names, then a fluent
expression is either a fluent name F or ¬F. A domain description is composed of propositions
of the following form:
CHAPTER 4. THE BIOPATHQA SYSTEM
149
value proposition: F after A1 ; . . . ; Am , where (m ≥ 0), F is a fluent and A1 , . . . , Am are fluents.
effect propostion: A causes F if P1 , . . . , Pn , where (n ≥ 0), A is an action, F, P1 , . . . , Pn are
fluent expressions. P1 , . . . , Pn are called preconditions of A and the effect proposition
describes the effect on F.
We relate it to our work:
• Fluents are boolean. We support numeric valued fluents, with binary fluents.
• Fluents are non-inertial, but inertia can be added. Our fluents are always intertial.
• Action description specifies the effect of an action. Our domain description specifies
‘natural’-actions, which execute automatically when their pre-conditions are satisfied,
subject to certain conditions. As a result our domain description represents trajectories.
• No built in support for aggregates exists. We support a selected set of aggregates, on a
single trajectory and over multiple trajectories.
• Value propositions in A are representable as observations in our query language.
4.7.3
Comparison with Action Language B
Action language B extends A by adding static causal laws, which allows one to specify indirect
effects or ramifications of an action Gelfond and Lifschitz (2012). We relate it to our work
below:
• Inertia is built into the semantics of B Gelfond and Lifschitz (2012). Our language also
has intertia built in.
• B supports static causal laws that allow defining a fluent in terms of other fluents. We do
not support static causal laws.
4.7.4
Comparison with Action Language C
Action language C is based on the theory of causal explanation, i.e. a formula is true if there is a
cause for it to be true Giunchiglia and Lifschitz (1998). It has been previously used to represent
and reason about biological pathways Dworschak et al. (2008). We relate it to our work below:
• C supports boolean fluents only. We support numeric valued fluents, and binary fluents.
• C allows both inertial and non-inertial fluents. While our fluents are always inertial.
• C support static causal laws (or ramifications), that allow defining a fluent in terms of
other fluents. We do not support them.
• C describes causal relationships between fluents and actions. Our language on the other
hand describes trajectories.
CHAPTER 4. THE BIOPATHQA SYSTEM
4.7.5
150
Comparison with Action Language C+
First, we give a brief overview of C+ Giunchiglia et al. (2004). Intuitively, atoms C+ are of the
form c = v, where c is either a fluent or an action constant, v belongs to the domain of c, and
fluents and actions form a disjoint set. A formula is a propositional combination of atoms. A
fluent formula is a formula in which all constants are fluent constants; and an action formula
is a formula with one action constant but no fluent constants. An action description in C+ is
composed of causal laws of the following forms:
static law: caused F if G, where F and G are fluent formulas
action dynamic law: caused F if G, where F is an action formula and G is a formula
fluent dynamic law: caused F if G after H, where F and G are fluent formulas and H is a
formula
Concise forms of these laws exist, e.g. ‘intertial f ≡ caused f = v if f = v after f = v, ∀v ∈
domain of f ’ that allow a more intuitive program declaration.
We now relate it to our work.
• Some of the main differences include:
– Fluents are multi-valued, fluent values can be integer, boolean or other types. We
support integer and binary valued fluents only.
– Actions are multi-valued. We do not support multi-valued actions.
– Both inertial and non-inertial fluents are supported. In comparison we allow inertial
fluents only.
– Static causal laws are supported that allow changing the value of a fluent based on
other fluents (ramifications). We do not allow static causal laws.
– Effect of parallel actions on numeric fluents is not additive. However, the additive
fluents extension Lee and Lifschitz (2003) adds the capability of additive fluents
through new rules. The extended language, however, imposes certain restrictions on
additive fluents and also restricts the domain of additive actions to boolean actions
only. Our fluents are always additive.
– Supports defaults. We do not have the same notion as defaults, but allow initial
values for fluents in our domain description.
– Action’s occurrence and its effect are defined in separate statements. In our case, the
action’s occurrence and effect are generally combined in one statement.
• Although parallel actions are supported, it is unclear how one can concisely describe the
condition implicit in our system that simultaneously occurring actions may not conflict.
Two actions conflict if their simultaneous execution will cause a fluent to become negative.
CHAPTER 4. THE BIOPATHQA SYSTEM
151
• Exogenous actions seem the closest match to our may execute actions. However, our actions are ‘natural’, in that they execute automatically when their pre-conditions are satisfied, they are not explicitly inhibited, and they do not conflict. Actions conflict when their
simultaneous execution will cause one of the fluents to become negative. The exogenousstyle character of our actions holds when the firing style is ‘∗’. When the firing style
changes, the execution semantics change as well. Consider the following two may execute statements in our language:
a1 may execute causing f1 change value by − 5 if f2 has value 3 or higher
a2 may execute causing f1 change value by − 3 if f2 has value 2 or higher
and two states: (i) f1 = 10, f2 = 5, (ii) f1 = 6, f2 = 5. In state (i) both a1 , a2 can occur
simultaneously (at one point) resulting in firing-choices {a1 , a2 }, {a1 }, {a2 }, {}, ; whereas,
in state (ii) only one of a1 or a2 can occur at one point resulting in the firing-choices:
{a1 }, {a2 }, {}, because of a conflict due to the limited amount of f1 . These firing choices
apply for firing style ‘*’, which allows any combination of fireable actions to occur. If the
firing style is set to ‘max’, the maximum set of non-conflicting actions may fire, and the
firing choices for state (i) change to {a1 , a2 } and the firing choices for state (ii) change
to {a1 }, {a2 } . If the firing style is set to ‘1’, at most one action may fire at one point,
and the firing choices for both state (i) and state (ii) reduce to {a1 }, {a2 }, {} . So, the case
with ‘*’ firing style can be represented in C+ with exogenous actions a1 , a2 ; the case with
‘1’ firing style can be represented in C+ with exogenous actions a1 , a2 and a constraint
requiring that both a1 , a2 do not occur simultaneously; while the case with ‘max’ firing
style can be represented by exogenous actions a1 , a2 with additional action dynamic laws.
They will still be subject to the conflict checking. Action dynamic laws can be used to
force actions similar to our must execute actions.
• Specification of initial values of fluents seem possible through the query language. The
default statement comes close, but it does not have the same notion as setting a fluent to
a value once. We support specifying initial values both in the domain description as well
as the query.
• There does not appear built-in support for aggregation of fluent values within the same
answer set, such as sum, count, rate, minimum, maximum etc. Although some of it could
be implemented using the additive fluents extension. We support a set of aggregates, such
as total, and rate. Additional aggregates can be easily added.
• We support queries over aggregates (such as minimum, maximum, average) of singletrajectory aggregates (such as total, and rate etc.) over a set of trajectories. We also
support comparative queries over two sets of trajectories. Our queries allow modification
of the domain description as part of query evaluation.
4.7.6
Comparison with BC
Action language BC combines features of B and C+ Lee et al. (2013). First we give a brief
overview of BC.
CHAPTER 4. THE BIOPATHQA SYSTEM
152
Intuitively, BC has actions and valued fluents. A valued fluent, called an atom, is of the
form ‘ f = v’, where f is a fluent constant and v ∈ domain( f ). A fluent can be regular or
statically determined. An action description in BC is composed of static and dynamic laws of
the following form:
static law: A0 if A1 , . . . , Am ifcons Am+1 , . . . , An , where (n ≥ m ≥ 0), and each Ai is an atom.
dynamic law: A0 after A1 , . . . , Am ifcons Am+1 , . . . , An , where (n ≥ m ≥ 0), A0 is a regular
fluent atom, each of A1 , . . . , Am is an atom or an action constant, and Am+1 , . . . , An are
atoms.
Concise forms of these laws exist that allow a more intuitive program declaration.
Now we relate it to our work.
• Some of the main differences include:
– Fluents are multi-valued, fluent values can which can be integer, boolean, or other
types. We only support integer and binary fluents.
– Static causal laws are allowed. We do not support static causal laws.
– Similar to C+ numeric fluent accumulation is not supported. It is supported in our
system.
– It is unclear how aggregate queries within a trajectory can be concisely represented.
Aggregate queries such as rate are supported in our system.
– It does not seem that queries over multiple trajectories or sets of trajectories are
supported. Such queries are supported in our system.
4.7.7
Comparison with ASPMT
ASPMT combines ASP with Satisfiability Modulo Theories. We relate the work in Lee and
Meng (2013) where C+ is extended using ASPMT with our work.
• It adds support for real valued fluents to C+ including additive fluents. Thus, it allows
reasoning with continuous and discrete processes simultaneously. Our language does not
support real numbers directly.
Several systems also exist to model and reason with biological pathway. For example:
4.7.8
Comparison with BioSigNet-RR
BioSigNet-RR Baral et al. (2004) is a system for representing and reasoning with signaling
networks. We relate it to our work in an intuitive manner.
• Fluents are boolean, so qualitative queries are possible. We support both integer and
binary fluents, so quantiative queries are also possible.
CHAPTER 4. THE BIOPATHQA SYSTEM
153
• Indirect effects (or ramifications) are supported. We do not support these.
• Action effects are captured separately in ‘ causes statement’ from action triggering statements ‘ triggers ’ and ‘ n triggers ’. We capture both components in a ‘ may execute causing ’
or ‘ normally must execute causing ’ statement.
• Their action triggering semantics have some similarity to our actions. Just like their actions get triggered when the action’s pre-conditions are satisfied, our actions are also
triggered when their pre-conditions are satisfied. However, the triggering semantics are
different, e.g. their triggers statement causes an action to occur even if it is disabled,
we do not have an equivalent for it; and their n triggers is similar in semantics to normally must execute causing statement.
• It is not clear how loops in biological systems can be modeled in their system. Loops are
possible in our by virtue of the Petri Net semantics.
• Their queries can have time-points and their precedence relations as part of the query.
Though our queries allow the specification of some time points for interval queries, timepoints are not supported in a similar way. However, we do support certain types of observation relative queries.
• The intervention in their planning queries has similarities to interventions in our system.
However, it appears that our intervention descriptions are higher level.
4.8
Conclusion
In this chapter we presented the BioPathQA system and the languages to represent and query
biological pathways. We also presented a new type of Petri Net, the so called Guarded-Arc Petri
Net that is used as a model behind our pathway specification language, which shares certain
aspects with CPNs Jensen et al. (2007), but our semantics for reset arcs is different, and we
allow must-fire actions that prioritize actions for firing over other actions. We also showed
how the system can be applied to questions from college level text books that require deeper
reasoning and cannot be answered by using surface level knowledge. Although our system is
developed with respect to the biological domain, it can be applied to non-biological domain as
well.
Some of the features of our language include: natural-actions that automatically fire when
their prerequisite conditions are met (subject to certain restrictions); an automatic default constraint that ensures fluents do not go negative, since they model natural systems substances;
a more natural representation of locations; and control of the level of parallelism to a certain
degree. Our query language also allows interventions similar to Pearl’s surgeries Pearl (1995),
which are more general than actions.
Next we want to apply BioPathQA to a real world application by doing text extraction.
Knowledge for real world is extracted from research papers. In the next chapter we show how
CHAPTER 4. THE BIOPATHQA SYSTEM
154
such text extraction is done for pathway construction and drug development. We will then
show how we can apply BioPathQA to the extracted knowledge to answer questions about the
extracted knowledge.
Chapter 5
Text Extraction for Real World
Applications
In the previous chapter we looked at the BioPathQA system and how it answers simulation based
reasoning questions about biological pathways, including questions that require comparison of
alternate scenarios through simulation. These so called ‘what-if’ questions arise in biological
activities such as drug development, drug interaction, and personalized medicine. We will now
put our system and language in context of such activities.
Cutting-edge knowledge about pathways for activities such as drug development, drug interaction, and personalized medicine comes in the form of natural language research papers,
thousands of which are published each year. To use this knowledge with our system, we need
to perform extraction. In this chapter we describe techniques we use for such knowledge extraction for discovering drug interactions. We illustrate with an example extraction how we
organize the extracted knowledge into a pathway specification and give examples of relevant
what-if questions that a researcher performing may ask in the drug development domain.
5.1
Introduction
Thousands of research papers are published each year about biological systems and pathways
over a broad spectrum of activities, including interactions between dugs and diseases, the associated pathways, and genetic variation. Thus, one has to perform text extraction to extract
relationships between the biochemical processes, their involvement in diseases, and their interaction with drugs. For personalized medicine, one is also interested in how these interrelationships change in presence of genetic variation. In short, we are looking for relationships between
various components of the biochemical processes and their internal and external stimuli.
Many approaches exist for extracting relationships from a document. Most rely on some
form of co-occurrence, relative distance, or order of words in a single document. Some use
shallow parsing as well. Although these techniques tend to have a higher recall, they focus on
extracting explicit relationships, which are relationships that are fully captured in a sentence
155
CHAPTER 5. TEXT EXTRACTION FOR REAL WORLD APPLICATIONS
156
or a document. These techniques also do not capture implicit relationships that may be spread
across multiple documents. are spread across multiple documents relating to different species.
Additional issues arise from the level of detail from in older vs. newer texts and seemingly
contradictory information due to various levels of confidence in the techniques used. Many do
not handle negative statements.
We primarily use a system called PTQL Tari et al. (2012) to extract these relationships,
which allows combining the syntactic structure (parse tree), semantic elements, and word order
in a relationship query. The sentences are pre-processed by using named-entity recognition, and
entity normalization to allow querying on classes of entity types, such as drugs, and diseases;
and also to allow cross-linking relationships across documents when they refer to the same entity
with a different name. Queries that use such semantic association between words/phrases are
likely to produce higher precision results.
Source knowledge for extraction primarily comes from thousands of biological abstracts
published each year in PubMed 1 .
Next we briefly describe how we extract relationships about drug interactions. Following
that we briefly describe how we extract association of drugs, and diseases with genetic variation.
We conclude this chapter with an illustrative example of how the drug interaction relationships
are used with our system to answer questions about drug interactions and how genetic variation
could be utilized in our system.
5.2
Extracting Relationships about Drug Interactions
We summarize the extraction of relationships for our work on drug-drug interactions from Tari
et al. (2010b).
Studying drug-drug interactions are a major activity in drug development. Drug interactions
occur due to the interactions between the biological processes / pathways that are responsible
metabolizing and transporting drugs. Metabolic processes remove a drug from the system within
a certain time period. For a drug to remain effective, it must be maintained within its therapeutic
window for the period of treatment, requiring taking the drug periodically. Outside the therapeutic window, a drug can become toxic if a quantity greater than the therapeutic window is
retained; or it can become ineffective if a quantity less than the therapeutic window is retained.
Since liver enzymes metabolize most drugs, it is the location where most metabolic-interaction
takes place. Induction or inhibition of these enzymes can affect the bioavailability of a drug
through transcriptional regulation, either directly or indirectly. For example, if drug A inhibits
enzyme E, which metabolizes drug B, then the bioavailability of drug B will be higher than
normal, rendering it toxic. On the other hand, if drug A induces enzyme E, which metabolizes
drug B, then drug B’s bioavailability will be lesser than normal, rendering it ineffective.
Inhibition of enzymes is a common form of drug-drug interactions Boobis (2009). In direct
inhibition, a drug A inhibit enzyme E, which is responsible for metabolism of drug B. Drug
A, leads to a decrease in the level of enzyme E, which in turn can increase bioavailability of
1
http://www.ncbi.nlm.nih.gov/pubmed
CHAPTER 5. TEXT EXTRACTION FOR REAL WORLD APPLICATIONS
157
drug B potentially leading to toxicity. Alternatively, insufficient metabolism of drug B can lead
to smaller amount of drug B’s metabolites being produced, leading to therapeutic failure. An
example of one such direct inhibition is the interaction between CYP2D6 inhibitor quinidine and
CYP2D6 substrates (i.e. substances metabolized by CYP2D6), such as Codeine. The inhibition
of CYP2D6 increases the bioavailability of drugs metabolized by CYP2D6 leading to adverse
side effects.
Another form of drug interactions is through induction of enzymes Boobis (2009). In direct
induction, a drug A induces enzyme E, which is responsible for metabolism of drug B. An
example of such direct induction is between phenobarbital, a CYP2C9 inducer and warfarin
(a CYP2C9 substrate). Phenobarbital leads to increased metabolism of warfarin, decreasing
warfarins bioavailability. Direct interaction due to induction though possible is not as common
as indirect interaction through transcription factors, which regulate the drug metabolizing enzymes. In such an interaction, drug A activates a transcription factor T F, which regulates and
induces enzyme E, where enzyme E metabolizes drug B. Transcription factors are referred to
as regulators of xenobiotic-metabolizing enzymes. Examples of such regulators include aryl hydrocarbon receptor AhR, pregnane X receptor PXR and constitutive androstane receptor CAR.
Drug interactions can also occur due to the induction or inhibition of transporters. Transporters are mainly responsible for cellular uptake or efflux (elimination) of drugs. They play
an important part in drug disposition, by transporting drugs into the liver cells, for example.
Transporter-based drug interactions, however, are not as well studies as metabolism-based interactions Boobis (2009).
5.2.1
Method
Extraction of drug-drug interactions from the text can either be explicit or implicit. Explicit extraction refers to extraction of drug-drug interaction mentioned within a single sentence, while
implicit extraction requires extraction of bio-properties of drug transport, metabolism and elimination that can lead to drug-drug interaction. This type of indirect extraction combines background information about biological processes, identification of protein families and the interactions that are involved in drug metabolism. Our approach is to extract both explicit and implicit
drug interactions as summarized in Fig 5.1 and it builds upon the work done in Tari et al. (2009).
Explicit Drug Interaction Extraction
Explicit extraction mainly extracts drug-drug interactions directly mentioned in PubMed (or
Medline) abstracts. For example, the following sentences each have a metabolic interaction
mentioned within the same sentence:
S1: Ciprofloxacin strongly inhibits clozapine metabolism. (PMID: 19067475)
S2: Enantioselective induction of cyclophosphamide metabolism by phenytoin.
which can be extracted by using the following PTQL query using the underlined keywords from
above sentences:
CHAPTER 5. TEXT EXTRACTION FOR REAL WORLD APPLICATIONS
158
Figure 5.1: This figure from Tari et al. (2010b) outlines the effects of drug A on drug B through
(a) direct induction/inhibition of enzymes; (b) indirect induction/inhibition of transportation
factors that regulate the drug-metabolizing enzymes.
//S{//?[Tag=‘Drug’](d1) =>
//?[Value IN {‘induce’,‘induces’,‘inhibit’,‘inhibits’}](v) =>
//?[Tag=‘Drug’](d2) => //?[Value=‘metabolism’](w)} :::
[d1 v d2 w] 5 : d1.value, v.value, d2.value.
This PTQL query specifies that a drug (denoted by d1) must be followed by one of the keywords
from {‘induce0 , ‘inhibit0 , ‘inhibits0 } (denoted by v), which in turn must be followed by another
drug (denoted by d2) followed the keyword ‘metabolism0 (denoted by w); all found within a
proximity of 5 words of each other. The query produces tripes of hd1, v, d2i values as output.
Thus the results will produce triples hd1, induces, d2i and hd1, inhibits, d2i which mean that the
drug d1 increases the effect of d2 (i.e. hd1, increases, d2i) and decreases the effect of d2 (i.e.
hd1, decreases, d2i) respectively. For example, the sentence S1 above matches this PTQL query
and the query will produce the triplet hciprofloxacin, increases, clozapinei.
Implicit Drug Interaction Extraction
Implicit extraction mainly extracts drug-drug interactions not yet published, but which can be
inferred from published articles and properties of drug metabolism. The metabolic properties
themselves have their origin in various publications. The metabolic interactions extracted from
published articles and the background knowledge of properties of drug metabolism are reasoned
with in an automated fashion to infer drug interactions. The following table outlines the kinds
of interactions extracted from the text:
which require multiple PTQL queries for extraction. As an example, the following PTQL
query is used to extract hprotein, metabolizes, drugi triplets:
CHAPTER 5. TEXT EXTRACTION FOR REAL WORLD APPLICATIONS
hactor1, relation, actor2i
hd, metabolizes, pi
hd, inhibits, pi
hd, induces, pi
hp1, regulates, p2i
159
Description
Drug d is metabolized by protein p
Drug d inhibits the activity of protein p
Drug d induces the activity of protein p
Protein p1 regulates the activity of protein p2
Table 5.1: This table from Tari et al. (2010b) shows the triplets representing various properties
relevant to the extraction of implicit drug interactions and their description.
PMID
8689812
8477556
10678302
Extracted Interaction
hcyp3a4, metabolizes, lovastatini
h f luoxetine, inhibits, cyp2d6i
hphenytoin, induces, cyp2ci
11502872
hpxr, regulates, cyp2b6i
Evidence
Lovastatin is metabolized by CYP3A4
Inhibition by fluoxetine of cytochrome P450 2D6 activity
Phenytoin induces CYP2C and CYP3A4 isoforms, but not
CYP2E1
The CYP2B6 gene is directly regulated by PXR
Table 5.2: This table from Tari et al. (2010b) shows the triplets representing various properties
relevant to the extraction of implicit drug interactions and their description.
//S{/?[Tag=‘Drug’](d1) =>
//VP{//?[Value IN {‘metabolized’,‘metabolised’}](rel) =>
//?[Tag=‘GENE’](g1)}} ::: g1.value, rel.value, d1.value
which specifies that the extracted triplets must have a drug (denoted by d1) followed by a verb
phrase (denoted by VP) with the verb in {‘metabolized0 , ‘metabolised0 }, followed by a gene
(denoted by g1). Table 5.2 shows examples of extracted triplets.
Data Cleaning
The protein-protein and protein-drug relationships extracted from the parse tree database need
an extra step of refinement to ensure that they correspond to the known properties of drug
metabolism. For instance, for a protein to metabolize a drug, the protein must be an enzyme. Similarly, for a protein to regulate an enzyme, the protein must be a transcription
factor. Thus, the hprotein, metabolizes, drugi facts get refined to henzyme, metabolizes, drugi
and hprotein, regulates, proteini gets refined to htranscription f actor, regulates, enzymei respectively. Classification of proteins is done using UniProt, the Gene Ontology (GO) and Entrez
Gene summary by applying rules such as:
• A protein p is an enzyme if it belongs to a known enzyme family, such as CYP, UGT or
SULT gene families; or is annotated under UniProt with the hydrolase, ligase, lyase or
transferase keywords; or is listed under the “metabolic process” GO-term; or its Entrez
Gene summary mentions key phrases like “drug metabolism” or roots for “enzyme” or
“catalyzes”.
• A protein p is considered as a transcription factor if it is annotated with keywords transcription, transcription-regulator or activator under UniProt; or it is listed under the “transcription factor activity” category in GO; or its Entrez Gene summary contains the phrase
“transcription factor”.
CHAPTER 5. TEXT EXTRACTION FOR REAL WORLD APPLICATIONS
Relations
< protein, metabolizes, drug >
< drug, induces, protein >
< drug, inhibits, protein >
< protein, regulates, protein >
negation
Precision (# TP)
93.1% (54)
61.8% (42)
58.6% (99)
68.7% (46)
84.4% (38)
Recall (# FN)
26.7% (148)
30.7% (95)
48.5% (105)
100.0% (0)
–
160
F-measure
41.5%
41.0%
53.1%
81.4%
–
Table 5.3: Performance of interactions extracted from Medline abstracts. TP represents truepositives, while FN represents false-negatives Tari et al. (2010b)
Additional rules are applied to remove conflicting information, such as, favoring negative
extractions (such as ‘P does not metabolize D’) over positive extractions (such as ‘P metabolizes
D’). For details, see Tari et al. (2010b).
5.2.2
Results
The correctness of extracted interactions was determined by manually compiling a gold standard
for each type of interaction using co-occurrence queries. For example, for hprotein, metabolizes, drugi
relations, we examined sentences that contain co-occurrence of protein, drug and one of the
keywords “metabolized”, “metabolize”, “metabolises”, “metabolise”, “substrate” etc. Table 5.3
summarizes the performance of our extraction approach.
5.3
Extracting Knowledge About Genetic Variants
We summarize the relevant portion of our work on associating genetic variants with drugs,
diseases and adverse reactions as presented in Hakenberg et al. (2012).
Incorrect drug dosage is the leading cause of adverse drug reactions. Doctors prescribe
the same amount of medicine to a patient for most drugs using the average drug response, even
though a particular person’s drug response may be higher or lower than the average case. A large
part of the difference in drug response can be attributed to single nucleotide polymorphisms
(SNPs). For example, the enzyme CYP2D6 has 70 known allelic variations, 50 of which are
non-functional H et al. (2002). Patients with poor metabolizer variations may retain higher
concentration of drug for typical dosage, while patients with rapid metabolizers may not even
reach therapeutic level of drug or toxic level of drug metabolites SF (2009). Thus, it is important
to consider the individual’s genetic composition for dosage determination, especially for narrow
therapeutic index drugs.
Scientists studying these variations have grouped metabolizers into categories of poor (PM),
intermediate (IM), rapid (RM) and ultra-rapid metabolizers (UM) and found that for some drugs,
only 20% of usual dosage was required for PM and up to 140% for UM J et al. (2001). Information about SNPs, their frequency in various population groups, their effect on genes (enzymic
activity) and related data is stored in research papers, clinical studies and surveys. However, it
is spread-out among them. Various databases collect this information in different forms. PharmGKB collects information such information and how it related to drug response Thorn et al.
CHAPTER 5. TEXT EXTRACTION FOR REAL WORLD APPLICATIONS
161
(2010). However, it is a small percentage of the total number of articles on PharmGKB, due to
time consuming manual curation of data.
Our work focuses on automatically extracting genetic variations and their impact on drug
responses from PubMed abstracts to catch up with the current state of research in the biological
domain, producing a repository of knowledge to assist in personalized medicine. Our approach
leverages on as many existing tools as possible.
5.3.1
Methods
Next, we describe the methods used in our extraction, including: named entity recognition,
entity normalization, and relation extraction.
Named Entity Recognition
We want to identify entities including genes (also proteins and enzymes), drugs, diseases, ADRs
(adverse drug reactions), SNPs, RefSNPs (rs-numbers), names of alleles, populations and frequencies.
For genes, we use BANNER Leaman et al. (2008) trained on BioCreative II GM training set Krallinger et al. (2008). For genotypes (genetic variations including SNPs) we used a
combination of MutationFinder Caporaso et al. (2007) and custom components. Custom components were targeted mostly on non-SNPs (“c.76 78delACT”, 11MV324KF”) and insertions/deletions (“1707 del T”, “c.76 77insG”), RefSNPs (rs-numbers) and names of alleles/haplotypes (“CYP2D6*4”, “T allele”, “null allele”).
For diseases (and ADRs), we used BANNER trained on a corpus of 3000 sentences with
disease annotations Leaman et al. (2009). An additional 200 random sentences containing no
disease were added from BioCreative II GM data to offset the low percentage (10%) of sentences
without disease in the 3000 sentence corpus. In addition to BANNER, we used a dictionary
extracted from UMLS. This dictionary consisted of 162k terms for 54k concepts from the six
categories “Disease or Syndrome”, ”Neoplastic Process”,“Congenital Abnormality”,“Mental or
Behavioral Dysfunction”,“Experimental Model of Disease” and “Acquired Abnormality”. The
list was filtered to remove unspecific as well as spurious disease names such as “symptoms”,
“disorder”, . . . . A dictionary for adverse drug reactions originated from SIDER Side Effect
Resource Kuhn et al. (2010), which provides a mapping of ADR terminology to UMLS CUIs.
It consisted of 1.6k UMLS concepts with 6.5k terms.
For drugs, we used a dictionary based on DrugBankWishart et al. (2006) containing about
29k different drug names including both generic as well as brand names. We used the crosslinking information from DrugBank to collect additional synonyms and IDs from PharmGKB.
We cross linked to Compound and Substance IDs from PubChem to provide hyperlinks to additional information.
For population, we collected a dictionary of terms referring to countries, regions, regions
inhabitants and their ethnicities from WikiPedia, e.g. “Caucasian”, “Italian”, “North African”,
. . . . We filtered out irrelevant phrases like “Chinese [hamster]”.
CHAPTER 5. TEXT EXTRACTION FOR REAL WORLD APPLICATIONS
162
For frequencies, we extract all numbers and percentages as well as ranges from sentences
that contain the word “allele”, “variant”, “mutation”, or “population”. The output is filtered in
this case as well to remove false positives referring to p-values, odd ratios, confidence intervals
and common trigger words.
Entity Normalization
Genes, diseases and drugs can appear with many different names in the text. For example,
“CYP2D6” can appear as “Cytochrome p450 2D6” or “P450 IID6” among others, but they all
refer to the same enzyme (EntrezGene ID 1565). We use GNAT on recognized genes Hakenberg et al. (2008), but limit them to human, mouse and rat genes. The gene name recognized
by BANNER is filtered by GNAT to remove non-useful modifiers and looked up against EntrezGene restricted to human, mouse and rat genes to find candidate IDs for each gene name.
Ambiguity (multiple matches) is resolved by matching the text surrounding the gene mention
with gene’s annotation from a set of resources, like EntrezGene, UniProt.
Drugs and diseases/ADRs are resolved to their official IDs from DrugBank or UMLS. If
none is found, we choose an ID for it based on its name. Genetic variants
Genetic variations are converted to HGVScite den Dunnen and Antonarakis (200) recommended format. Alleles were converted to the star notation (e.g. “CYP2D6*1”) and the genotype (“TT allele”) or fixed terms such as “null allele” and “variant allele”.
Populations mentions are mapped to controlled vocabulary to remove orthographic and lexical variations.
Relation Extraction
Twelve type of relations were extracted between the detected entities as given in Table 5.4. Different methods were applied to detect different relations depending upon relation type, sentence
structure and whether another method was able to extraction a relation beforehand. Gene-drug,
gene-disease, drug-disease were extracted using sentence based co-occurrence (possibly refined
by using relation-specific keywords) due to its good precision yield of this method for these
relations. For other relations additional extraction methods were implemented. These include:
High-confidence co-occurrence that includes keywords These co-occurences have the relation keyword in them. This method is applied to gene-drug, gene-disease, drug-ADR,
drug-disease and mutation-disease associations. It uses keywords from PolySearch Wishart
et al. (2008) as well as our own.
Co-occurrence without keywords Such co-occurrences do not require any relationship keyword. This method is used for allele-population and variant-population relationships.
This method can misidentify negative relationships. High-confidence relationships, if not
found with a keyword drop down to this method for a lower confidence result.
1:n co-occurrence Relationships where one entity has one instance in a given sentence and the
other occurs one or more times. Single instance entity may have more than one occur-
CHAPTER 5. TEXT EXTRACTION FOR REAL WORLD APPLICATIONS
Gene-drug
Gene-disease
Drug-disease
Drug-adverse effect
Gene-variant
Gene-allele
Gene-RefSNP
Allele-population
Variant-population
Allele frequency
Variant frequency
Population frequency
Drugs-populations
Drug-alleles
Drug-RefSNP
Drug-mutation
Sum
Total
191,054
709,987
117,834
73,696
101,477
65,569
12,881
7181
12,897
6,893
6,646
8404
12,849
6,778
1,161
10,809
1,315,811
Unique
31,593
102,881
26,268
16,569
21,704
6802
5748
1891
6765
279
1654
144
4388
1858
721
5491
233,964
163
PharmGKB
6820/3014
8147/4478
4343/939
645/516
146/99
1820/1125
Table 5.4: Unique binary relations identified between detected entities from Hakenberg et al.
(2012).
rence. This method is useful in identifying gene mutations, where a gene is mentioned in
a sentence along with a number of its mutations. The gene itself may be repeated.
Enumerations with matching counts Captures entities in sentences where a list of entities is
followed by an equal number of counts. This method is useful in capturing alleles and
their associated frequencies, e.g. “The frequencies of CYP1B1*1, *2, *3, and *4 alleles
were 0.087, 0.293, 0.444, and 0.175, respectively.”
Least common ancestor (LCA) sub-tree Assigns associations based on distance in parse tree.
We used Stanford parser Klein and Manning (2003) to get grammatical structure of a sentence as a dependency tree. This allows relating verb to its subject and noun to its modifiers. This method picks the closest pair in the lowest common ancestors (dependency)
sub-tree of the entities. Maximum distance in terms of edges connecting the entity nodes
was set to 10, which was determined empirically to provide the best balance between precision and recall. This method associates frequencies with alleles in the sentence “The
allele frequencies were 18.3% (-24T) and 21.2% (1249A)”.
m:n co-occurrence This method builds associations between all pairs of entities.
Low confidence co-occurrence This acts as the catch-all case if none of the above methods
work.
These methods were applied in order to each sentence, stoping at the first method that extracted the desired relationship. Order of these methods was determined empirically based of
their precision. The order of the method used determines our confidence in the result. If none
of the higher confidence methods are successful, a co-occurrence based method is used for extraction with low confidence.
CHAPTER 5. TEXT EXTRACTION FOR REAL WORLD APPLICATIONS
164
Abstract-level co-occurrence are also extracted to provide hits on potential relations. They
appear in the database only when they appear in more than a pre-set threshold number of abstracts.
5.3.2
Results
Performance was evaluated by evaluating the precision and recall of individual components
and coverage of existing results. Precision and recall were tested by processing 3500 PubMed
abstracts found via PharmGKB relations and manually checking the 2500 predictions. Coverage was tested against DrugBank and PharmGKB. Extracted relations went through manual
evaluation for correctness. Each extraction was also assigned a confidence value based on the
confidence in the method of extraction used. We got a coverage of 91% of data in DrugBank and
94% in PharmGKB. Taking into false positive rates for genes, drugs and gene-drug relations,
SNPshot has more than 10,000 new relations.
5.4
Applying BioPathQA to Drug-Drug Interaction Discovery
Now we use our BioPathQA system from chapter 4 to answer questions about drug-drug interaction using knowledge extracted from research publications using the approach in sections 5.2,5.3. We supplement the extracted knowledge with domain knowledge as needed.
Let the extracted facts be as follows:
• The drug ge f itinib is metabolized by CY P3A4.
• The drug phenytoin induces CY P3A4.
Following additional facts have been provided about a new drug currently in development:
• A new drug being developed test drug is a CYP3A4 inhibitor
We show the pathway specification based on the above facts and background knowledge,
then elaborate on each component:
domain of ge f itinib is integer, cyp3a4 is integer,
phenytoin is integer, test drug is integer
(5.1)
t1 may execute causing
ge f itinib change value by − 1,
cyp3a4 change value by − 1,
cyp3a4 change value by + 1
(5.2)
normally stimulate t1 by factor 2
if phenytoin has value 1 or higher
inhibit t1 if test drug has value 1 or higher
initially ge f itinib has value 20, cyp3a4 has value 60,
(5.3)
(5.4)
CHAPTER 5. TEXT EXTRACTION FOR REAL WORLD APPLICATIONS
phenytoin has value 0, test drug has value 0
firing style max
165
(5.5)
(5.6)
Line 5.1 declares the domain of the fluents as integer numbers. Line 5.2 represents the activity of enzyme cyp3a4 as the action t1. Due to the enzymic action t1, one unit of ge f itinib
is metabolized, and thus converted to various metabolites (not shown here). The enzymic action uses one unit of cyp3a4 as catalyst, which is used in the reaction and released afterwards.
Line 5.3 represents the knowledge that phenytoin induces the activity of cyp3a4. From background knowledge we find out that the stimulation in the activity can be as high as 2-times Luo
et al. (2002). Line 5.4 represents the knowledge that there is a new drug test drug being tested
that is known to inhibit the activity of cyp3a4. Line 5.5 specifies the initial distribution of the
drugs and enzymes in the pathway. Assuming the patient has been given some fixed dose, say
20 units, of the medicine ge f itinib. It also specifies there is a large 60 units quantity of cyp3a4
available to ensure reactions do not slow down due to unavailability of enzyme availability. Additionaly, the drug phenytoin is absent from the system and a new drug test drug to be tested
is not in the system either. This gives us our pathway specification. Now we consider two
application scenarios for drug development.
5.4.1
Drug Administration
A patient is taking 20 units of ge f itinib, and is being prescribed additional drugs to be coadministered. The drug administrator wants to know if there will be an interaction with ge f itinib
if 5 units of phenytoin are co-administered. If there is an interaction, what will be the bioavailability of ge f itinib so that its dosage could be appropriately adjusted.
The first question is asking whether giving the patient 5-units of phenytoin in addition to
the existing ge f itinib dose will cause a drug-interaction. It is encoded as the following query
statement Q for a k-step simulation:
direction of change in average value of ge f itinib is d
when observed at time step k;
comparing nominal pathway with modified pathway obtained
due to interventions : set value of phenytoin to 5;
If the direction of change is “=” then there was no drug-interaction. Otherwise, an interaction
was noticed. For a simulation of length k = 5, we find 15 units of ge f itinib remained at the
end of simulation in the nominal case when no phenytoin is administered. The amount drops to
10 units of ge f itinib when phenytoin is co-administered. The change in direction is “<”. Thus
there is an interaction.
The second question is asking about the bioavailability of the drug ge f itinib after some after
giving phenytoin in 5 units. If this bioavailability falls below the efficacy level of the drug, then
the drug would not treat the disease effectively. It is encoded as the following query statement
CHAPTER 5. TEXT EXTRACTION FOR REAL WORLD APPLICATIONS
166
Q for a k-step simulation:
average value of ge f itinib is n
when observed at time step k;
due to interventions :
set value of phenytoin to 5;
For a simulation of length k = 5, we find 10 units of ge f itinib remain. A drug administrator
(such as a pharmacist) can adjust the drug accordingly.
5.4.2
Drug Development
A drug manufacturer is developing a new drug test drug that is known to inhibit CYP3A4
that will be co-administered with drugs ge f itinib and phenytoin. He wants to determine the
bioavailability of ge f itinib over time to determine the risk of toxicity.
The question is asking about the bioavailability of the drug ge f itinib after 10 time units
after giving phenytoin in 5 units and the new drug test drug in 5 units. If this bioavailability
remains high, there is chance for toxicity due to the drug at the subsequent dosage intervals. It
is encoded as the following query statement Q for a k-step simulation:
average value of ge f itinib is n
when observed at time step k;
due to interventions :
set value of phenytoin to 5,
set value of test drug to 5;
For a simulation of length k = 5, we find all 20 units of ge f itinib remain. This could lead to
toxicity by building high concentration of ge f itinib in the body.
5.4.3
Drug Administration in Presence of Genetic Variation
A drug administrator wants to establish the dosage of morphine for a person based on its genetic
profile using its bioavailability.
Consider the following facts extracted about a simplified morphine pathway:
• codeine is metabolized by CY P2D6 producing morphine
• CY P2D6 has three allelic variations
– “*1” – (EM) effective metabolizer (normal case)
– “*2” – (UM) ultra rapid metabolizer
– “*9” – (PM) poor metabolizer
CHAPTER 5. TEXT EXTRACTION FOR REAL WORLD APPLICATIONS
167
For simplicity, assume UM allele doubles the metabolic rate, while PM allele halves the
metabolic rate of CYP2D6. Then, the resulting pathway is given by:
domain of cyp2d6 allele is integer, cyp2d6 is integer
domain of codeine is integer, morphine is integer
cyp2d6 action may execute causing
codeine change value by − 2, morphine change value by + 2
if cyp2d6 allele has value equal to 1
cyp2d6 action may execute causing
codeine change value by − 4, morphine change value by + 4
if cyp2d6 allele has value equal to 2
cyp2d6 action may execute causing
codeine change value by − 1, morphine change value by + 1
if cyp2d6 allele has value equal to 9
initially
codeine has value 0, morphine has value 0,
cyp2d6 has value 20, cyp2d6 allele has value 1
firing style max
Then, the bioavailability of morphine can be determined by the following query:
average value of morphine is n
when observed at time step k;
due to interventions :
set value of codeine to 20,
set value of cyp2d6 to 20,
set value of cyp2d6 allele to 9;
Simulation for 5 time steps reveal that the average bioavailability of morphine after 5 time-steps
is 5 for PM (down from 10 for EM).
Although this is a toy example, it is easy to see the potential of capturing known genetic
variations in the pathway and setting the complete genetic profile of a person in the intervention
part of the query.
5.5
Conclusion
In this chapter we presented how we extract biological pathway knowledge from text, including
knowledge about drug-gene interactions and their relationship to genetic variation. We showed
CHAPTER 5. TEXT EXTRACTION FOR REAL WORLD APPLICATIONS
168
how the information extracted is used to build pathway specification and illustrated how biologically relevant questions can be answered about drug-drug interaction using the BioPathQA
system developed in chapter 4. Next we look at the future directions in which the research work
done in this thesis can be extended.
Chapter 6
Conclusion and Future Work
The field of knowledge representation and reasoning (KR) is currently one of the most active
research areas. It represents the next step in the evolution of systems that know how to organize
knowledge, and have the ability to intelligently respond to questions about this knowledge. Such
questions could be about static knowledge or the dynamic processes.
Biological systems are uniquely positioned as role models for this next evolutionary step
due to their precise vocabulary and mechanical nature. As a result, a number of recent research challenges in the KR field are focused on it. The biological field itself needs systems
that can intelligently answer questions about such biological processes and systems in an automated fashion, given the large number of research papers published each year. Curating these
publications is time consuming and expensive, as a result the state of over all knowledge about
biological systems lags behind the cutting edge research.
An important class of questions asked about biological systems are the so called “what-if”
questions that compare alternate scenarios of a biological pathway. To answer such questions,
one has to perform simulation on a nominal pathway against a pathway modified due to the
interventions specified for the alternate scenario. Often, this means creating two pathways (for
nominal and alternate cases) and simulate them separately. This opens up the possibility that
the two pathways can become out of synchronization. A better approach is to allow the user to
specify the needed interventions in the query statement itself. In addition, to understand the full
spread of possible outcomes, given the parallel nature of biological pathways, one must consider
all possible pathway evolutions, otherwise, some outcomes may remain hidden.
If a system is to be used by biologists, it must have a simple interface, lowering the barrier of
entry. Since biological pathway knowledge can arrive from different sources, including books,
published articles, and lab experiments, a common input format is desired. Such a format allows
specification of pathways due to automatic extraction, as well as any changes / additions due to
locally available information.
A comprehensive end-to-end system that accomplish all the goals would take a natural language query along with any additional specific knowledge about the pathway as input, extract
the relevant portion of the relevant pathway from published material (and background knowl-
169
CHAPTER 6. CONCLUSION AND FUTURE WORK
170
edge), simulate it based on the query, and generate the results in a visual format. Each of these
tasks comes with its own challenges, some of which have been addressed in this thesis.
In this thesis, we have developed a system and a high level language to specify a biological
pathway and answer simulation based reasoning questions about it. The high level language
uses controlled-English vocabulary to make it more natural for a researcher to use directly.
The high level language has two components: a pathway specification language, and a query
specification language. The pathway specification language allows the user to specify a pathway
in a source independent form, thus locally obtained knowledge (e.g. from lab) can be combined
with automatically extracted knowledge. We believe that our pathway specification language
is easy for a person to understand and encode, lowering the bar to using our system. Our
pathway specification language allows conditional actions, enabling the encoding of alternate
action outcomes due to genetic variation. An important aspect of our pathway specification
language is that it specifies trajectories, which includes specifying the initial configuration of
substances, as well as state evolution style, such as maximal firing of actions, or serialized
actions etc.
Our query specification language provides a bridge between natural language questions and
their formal representation. It is English-like but with precise semantics. A main feature of
our query language is its support for comparative queries over alternate scenarios, which is not
currently supported by any of the query languages (associated with action languages) we have
reviewed. Our specification of alternate scenarios uses interventions (a general form of actions),
that allow the user to modify the pathway as part of the query processing. We believe our query
language is easier for a biologist to understand without requiring formal training.
To model the pathways, we use Petri Nets, which have been used in the past to model and
simulate biological pathways. Petri Nets have a simple visual representation, which closely
matches biological pathways; and they inherently support parallelism. We extended the Petri
Nets to add features that we needed to suit our domain, e.g., reset arcs that remove all quantity of
a substance as soon as it is produced, and conditional arcs that specify the conditional outcome
of an action.
For simulation, we use ASP, which allowed us straight forward way to implement Petri Nets.
It also gave us the ability to add extensions to the Petri Net by making local edits, implement
different firing semantics, filter trajectories based on observations, and reason with the results.
One of the major advantage of using Petri Net based simulation is the ability to generate all
possible state evolutions, enabling us to process queries that determine the conditions when a
certain observation becomes true.
Our post-processing step is done in Python, which allows strong text processing capabilities
using regular expressions, as well as libraries to easy process large matrices of numbers for
summarization of results.
Now we present additional challenges that need to be addressed.
CHAPTER 6. CONCLUSION AND FUTURE WORK
6.1
171
Pathway Extraction
In Chapter 5 we described how we extract facts for drug-drug interaction and gene variation.
This work needs to be extended to include some of the newer databases that have come online
recently. This may provide us with enzyme reaction rates, and substance quantities used in
various reactions. The relation extraction for pathways must also be cognizant of any genetic
variation mentioned in the text.
Since the knowledge about the pathway appears in relationships at varying degree of detail, a
process needs to be devised to assemble the pathway from the same level to granularity together,
while also maintaining pathways at different levels of granularities.
Since pathway extraction is a time consuming task, it would be best to create a catalog of the
pathways. The cataloged pathways could be manually edited by the user as necessary. Storing
pathways in this way means that would have to be updated periodically, requiring merging of
new knowledge into existing pathways. Manual edits would have to be identified, such that the
updated pathway does not overwrite them without the user’s knowledge.
6.2
Pathway Selection
Questions presented in biological texts do not explicitly mention the relevant pathway to use
for answering the question. One way to address this issue is to maintain a catalog of predefined pathways with keywords associated with them. Such keywords can include names of
the substances, names of the processes, and other relevant meta-data about the pathway. The
catalog can be searched to find the closest match to the query being asked.
An additional aspect in proper pathway selection is to use the proper abstraction level. If our
catalog contains a pathway at different abstraction levels, the coarsest pathway that contains the
processes and substances in the query should be selected. Any higher fidelity will increase the
processing time and generate too much irrelevant data. Alternatively, the catalog could contain
the pathway in a hierarchical form, allowing extraction of all elements of a pathway at the same
depth. A common way to hierarchically organize the pathway related to our system is to have
hierarchical actions, which is the approach taken by hierarchical Petri nets.
Lastly, the question may only ask about a small subsection of a much larger pathway. For
better performance, it is beneficial to extract the smallest biological pathway network model that
can answer the question.
6.3
Pathway Modeling
In Chapter 3, we presented our modeling of biological questions using Petri Nets and their
extensions encoded in ASP. We came across concepts like allosteric regulation, inhibition of
inhibition, and inhibition of activation that we currently do not model. In allosteric regulation,
an enzyme is not fully enabled or disabled, the enzyme’s shape changes, making it more or
less active. The level of its activity depends upon concentrations of activators and inhibitors.
CHAPTER 6. CONCLUSION AND FUTURE WORK
172
In inhibition of inhibition, the inhibition of a reaction is itself inhibited by another inhibition;
while in inhibition of activation (or stimulation), a substance inhibits the stimulation produced
by a substance. Both of these appear to be actions on actions, something that Petri Nets do not
allow. An alternate coding for these would have to be devised.
As more detailed information about pathways becomes available, the reactions and processes that we have in current pathways may get replaced with more detailed sub-pathways
themselves. However, such refinement may not come at the same time for separate legs of the
pathway. Just replacing the coarse transition with a refined transition may not be sufficient due to
relative timing constraints. Hence, a hierarchical Petri Net model may need to be implemented
(see ??).
6.4
Pathway Simulation
In Chapter 2 we presented our approach to encode Petri Nets and their extensions. We used a
discrete solver called clingo for our ASP encoding. As the number of simulation length increases
in size or larger quantities are involved, the solver slows down significantly. This is due to an
increased grounding cost of the program. Incremental solving (using iclingo) does not help,
since the program size still increases, and the increments merely delays the slow down but does
not stop it.
Systems such as constraint logic solvers (such as ?) could be used for discrete cases. Alternatively, a system developed on the ASPMT Lee and Meng (2013) approach could be used,
since it can represent longer runs, larger quantities, and real number quantities.
6.5
Extend High Level Language
In Chapter 4 we described the BioPathQA system, the pathway specification and the query
specification high level languages. As we enhance the modeling of the biological pathways, we
will need to improve or extend the system as well as the high level language. We give a few
examples of such extensions.
Our pathway specification language currently does not support continuous quantities (real
numbers). Extending to real numbers will improve the coverage of the pathways that can be
modeled. In addition, derived quantities (fluents) can be added, e.g. pH could be defined as a
formula that is read-only in nature.
Certain observations and queries can be easily specified using a language such as LTL,
especially for questions requiring conditions leading to an action or a state. As a result, it may
be useful to add LTL formulas to the query language. We did not take this approach because it
would have presented an additional non-English style syntax for the biologists.
Our substance production / consumption rates and amounts are currently tied to the fluents.
In certain situations it is desirable to analyze the quantity of a substance produced / consumed
by a specific action, e.g. one is interested in finding the amount of H+ ions produced by a
multi-protein complex IV only.
CHAPTER 6. CONCLUSION AND FUTURE WORK
173
Interventions (that are a part of the query statement) presented in this thesis are applied at
the start of the simulation. Eliminating this restriction would allow applying surgeries to the
pathway mid execution. Thus, instead of specifying the steady state conditions in the query
statement, one could apply the intervention when a certain steady state is reached.
6.6
Result Formatting and Visualization
In Chapter 4 we described our system that answers questions specified in our high level language. At the end of its process, it outputs the final result. This output method can be enhanced
by allowing to look at the progression of results in addition to the final result. This provides the
biologist with the whole spread of possible outcomes. An example of such a spread is shown
in Fig. fig:q1:result for question 1. A graphical representation of the simulation progression is
also beneficial in enhancing the confidence of the biologist. Indeed many existing tools do this.
A similar effect can be achieved by parsing and showing the relevant portion of the answer set.
6.7
Summary
In Chapter 1 we introduced the thesis topic and summarized specific research contributions
In Chapter 2 we introduced the foundational material of this thesis including Petri Nets
and ASP. We showed how ASP could be used to encode basic Petri Nets. We also showed
how ASP’s elaboration tolerance and declarative syntax allows us to encode various Petri Net
extensions with small localized changes. We also introduced a new firing semantics, the so
called maximal firing set semantics to simulate a Petri Net with maximum parallel activity.
In Chapter 3 we showed how the Petri Net extensions and the ASP encoding can be used
to answer simulation based deep reasoning questions. This and the work in Chapter 2 was
published in Anwar et al. (2013a,b).
In Chapter 4 we developed a system called BioPathQA to allow users to specify a pathway
and answer queries against it. We also developed a pathway specification language and a query
language for this system in order to avoid the vagaries of natural language. We introduced a new
type of Guarded-arc Petri Nets to model conditional actions as a model for pathway simulation.
We also described our implementation developed around a subset of the pathway specification
language.
In Chapter 5 we briefly described how text extraction is done to extract real world knowledge
about pathways and drug interactions. We then used the extracted knowledge to answer question
using BioPathQA. The text extraction work was published in Tari et al. (2010b,a); Hakenberg
et al. (2012).
Appendix A
Proofs of Various Propositions
174
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
175
Assumption: The definitions in this section assume the programs Π do not have recursion
through aggregate atoms. Our ASP translation ensures this due to the construction of programs
Π.
First we extend some definitions and properties related to ASP, such that they apply to rules
with aggregate atoms. We will refer to the non-aggregate atoms as basic atoms. Recall the
definitions of an ASP program given in section 2.2.1.
Proposition 9 (Forced Atom Proposition) Let S be an answer set of a ground ASP program
Π as defined in definition 4. For any ground instance of a rule R in Π of the form A0 :-A1 , . . . ,
Am , not B1 , . . . , not Bn , C1 , . . . , Ck . if ∀Ai , 1 ≤ i ≤ m, S |= Ai , and ∀B j , 1 ≤ j ≤ n, S 6|= B j ,
∀Cl , 1 ≤ l ≤ k, S |= Cl then S |= A0 .
Proof: Let S be an answer set of a ground ASP program Π, R ∈ Π be a ground rule such that
∀B j , 1 ≤ j ≤ n, S 6|= B j ; and ∀Cl , 1 ≤ l ≤ k, S |= Cl . Then, the reduct RS ≡ {p1 :-A1 , . . . , Am .; . . . ; ph :-A1 , . . . , Am . | {p1 , .
S ∩ lit(A0 )} . Since S is an answer set of Π, it is a model of ΠS . As a result, whenever,
∀Ai , 1 ≤ i ≤ m, S |= Ai , {p1 , . . . , ph } ⊆ S and S |= A0 .
Proposition 10 (Supporting Rule Proposition) If S is an answer set of a ground ASP program
Π as defined in definition 4 then S is supported by Π. That is, if S |= A0 , then there exists a
ground instance of a rule R in Π of the type A0 :-A1 , . . . , Am , not B1 , . . . , not Bn , C1 , . . . , Ck .
such that ∀Ai , 1 ≤ i ≤ m, S |= Ai , ∀B j , 1 ≤ j ≤ n, S 6|= B j , and ∀Cl , 1 ≤ l ≤ k, S |= Cl .
Proof: For S to be an answer set of Π, it must be the deductive closure of reduct ΠS . The
deductive closure S of ΠS is iteratively built by starting from an empty set S , and adding head
atoms of rules RSh ≡ ph :-A1 , . . . , Am ., RSh ∈ ΠS , whenever, S |= Ai , 1 ≤ i ≤ m, where, RSh is a rule
in the reduct of ground rule R ∈ Π with ph ∈ lit(A0 ) ∩ S . Thus, there is a rule R ≡ A0 :-A1 , . . . ,
Am , not B1 , . . . , not Bn , C1 , . . . , Ck ., R ∈ Π, such that ∀Cl , 1 ≤ l ≤ k and ∀B j , 1 ≤ j ≤ n, S 6|= B j .
Nothing else belongs in S .
Next, we extend the splitting set theorem ? to include aggregate atoms.
Definition 88 (Splitting Set) A Splitting Set for a program Π is any set U of literals such that,
for every rule R ∈ Π, if U |= head(R) then lit(R) ⊂ U. The set U splits Π into upper and
lower parts. The set of rules R ∈ Π s.t. lit(R) ⊂ U is called the bottom of Π w.r.t. U, denoted
by botU (Π). The rest of the rules, i.e. Π \ botU (Π) is called the top of Π w.r.t. U, denoted by
topU (Π).
Proposition 11 Let U be a splitting set of Π with answer set S and let X = S ∩U and Y = S \U.
Then, the reduct of Π w.r.t. S , i.e. ΠS is equal to botU (Π)X ∪ (Π \ botU (Π))X∪Y .
Proof: We can rewrite Π as botU (Π) ∪ (Π \ botU (Π)) using the definition of splitting set. Then
the reduct of Π w.r.t. S can be written in terms of X and Y, since S = X ∪ Y. ΠS = ΠX∪Y =
(botU (Π) ∪ (Π \ botU (Π)))X∪Y = botU (Π)X∪Y ∪ (Π \ botU (Π))X∪Y . Since lit(botU (Π)) ⊆ U and
Y ∩ U = ∅, the reduct of botU (Π)X∪Y = botU (Π)X . Thus, ΠX∪Y = botU (Π)X ∪ (Π \ botU (Π))X∪Y .
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
176
Proposition 12 Let S be an answer set of a program Π, then S ⊆ lit(head(Π)).
Proof: If S is an answer set of a program Π then S is produced by the deductive closure of ΠS
(the reduct of Π w.r.t S ). By definition of the deductive closure, nothing can be in S unless it is
the head of some rule supported by S .
Splitting allow computing the answer set of a program Π in layers. Answer sets of the
bottom layer are first used to partially evaluate the top layer, and then answer sets of the top
layer are computed. Next, we define how a program is partially evaluated.
Intuitively, the partial evaluation of an aggregate atom c given splitting set U w.r.t. a set of
literals X removes all literals that are part of the splitting set U from c and updates c’s lower
and upper bounds based on the literals in X, which usually come from botU of a program. The
set X represents our knowledge about the positive literals, while the set U \ X represents our
knowledge about naf-literals at this stage. We can remove all literals in U from c, since the
literals in U will not appear in the head of any rule in topU .
Definition 89 (Partial Evaluation of Aggregate Atom) The partial evaluation of an aggregate atom c = l [B0 = w0 , . . . , Bm = wm ] u, given splitting set U w.r.t. a set of literals X,
written evalU (c, X) is a new aggregate atom c0 constructed from c as follows:
1. pos(c0 ) = pos(c) \ U
P
2. d = Bi ∈pos(c)∩U∩X wi
3. l0 = l − d, u0 = u − d are the lower and upper limits of c0
Next, we define how a program is partially evaluated given a splitting set U w.r.t. a set of
literals X that form the answer-set of the lower layer. Intuitively, a partial evaluation deletes all
rules from the partial evaluation for which the body of the rule is determined to be not supported
by U w.r.t. X. This includes rules which have an aggregate atom c in their body s.t. lit(c) ⊆ U,
but X 6|= c 1 . In the remaining rules, the positive and negative literals that overlap with U are
deleted, and so are the aggregate atoms that have lit(c) ⊆ U (since such a c can be fully evaluated
w.r.t. X). Each remaining aggregate atom is updated by removing atoms that belong to U 2 , and
updating its limits based on the answer-set X 3 . The head atom is not modified, since evalU (...)
is performed on Π \ botU (Π), which already removes all rules with heads atoms that intersect U.
Definition 90 (Partial Evaluation) The partial evaluation of Π, given splitting set U w.r.t. a set
of literals X is the program evalU (Π, X) composed of rules R0 for each R ∈ Π that satisfies all
the following conditions:
1. pos(R) ∩ U ⊆ X,
1
Note that we can fully evaluate an aggregate atom c w.r.t. answer-set X if lit(c) ⊆ U.
Since the atoms in U will not appear in the head of any atoms in topU and hence will not form a basis in future
evaluations of c.
3
The limit update requires knowledge of the current answer-set to update limit values.
2
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
177
2. ((neg(R) ∩ U) ∩ X) = ∅, and
3. if there is a c ∈ agg(R) s.t. lit(c) ⊆ U, then X |= c
A new rule R0 is constructed from a rule R as follows:
1. head(R0 ) = head(R),
2. pos(R0 ) = pos(R) \ U,
3. neg(R0 ) = neg(R) \ U,
4. agg(R0 ) = {evalU (c, X) : c ∈ agg(R), lit(c) * U}
Proposition 13 Let U be a splitting set for Π, X be an answer set of botU (Π), and Y be an
answer set of evalU (Π \ botU (Π), X). Then, X ⊆ lit(Π) ∩ U and Y ⊆ lit(Π) \ U.
Proof: By proposition 12, X ⊆ lit(head(botU (Π))), and Y ⊆ lit(head(evalU (Π \ botU (Π), X))).
In addition, lit(head(botU (Π))) ⊆ lit(botU (Π)) and lit(botU (Π)) ⊆ lit(Π) ∩ U by definition of
botU (Π). Then X ⊆ lit(Π) ∩ U, and Y ⊆ lit(Π) \ U.
Definition 91 (Solution) Baral (2003) Let U be a splitting set for a program Π. A solution to
Π w.r.t. U is a pair hX, Yi of literals such that:
• X is an answer set for botU (Π)
• Y is an answer set for evalU (topU (Π), X); and
• X ∪ Y is consistent.
Proposition 14 (Splitting Theorem) Baral (2003) Let U be a splitting set for a program Π. A
set S of literals is a consistent answer set for Π iff S = X ∪ Y for some solution hX, Yi of Π w.r.t.
U.
Lemma 1 Let U be a splitting set of Π, C be an aggregate atom in Π, and X and Y be subsets
of lit(Π) s.t. X ⊆ U, and Y ∩ U = ∅. Then, X ∪ Y |= C iff Y |= evalU (C, X).
Proof:
1. Let C 0 = evalU (C, X), then by definition of partial evaluation of aggregate atom, pos(C 0 ) =
pos(C) \ U, with lower limit l0 = l − d, and upper limit u0 = u − d, computed from l, u, the
lower and upper limits of C, where
X
wi
d=
Bi ∈pos(C)∩U∩X
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
2. Y |= C 0 iff
0
l ≤
X
B0i ∈pos(C 0 )∩Y
178
0
wi ≤ u0
– by definition of aggregate atom satisfaction.
3. then Y |= C iff
l ≤
X
X
wi +
B0i ∈(pos(C)\U)∩Y
Bi ∈pos(C)∩U∩X
w0i ≤ u
4. however, (pos(C) ∩ U) ∩ X and (pos(C) \ U) ∩ Y combined represent pos(C) ∩ (X ∪ Y) –
since
pos(C) ∩ (X ∪ Y) = ((pos(C) ∩ U) ∪ (pos(C) \ U)) ∩ (X ∪ Y)
= [((pos(C) ∩ U) ∪ (pos(C) \ U)) ∩ X]
∪ [((pos(C) ∩ U) ∪ (pos(C) \ U)) ∩ Y]
= [(pos(C) ∩ U) ∩ X) ∪ ((pos(C) \ U) ∩ X)]
∪ [(pos(C) ∩ U) ∩ Y) ∪ ((pos(C) \ U) ∩ Y)]
= [((pos(C) ∩ U) ∩ X) ∪ ∅] ∪ [∅ ∪ ((pos(C) \ U) ∩ Y)]
= ((pos(C) ∩ U) ∩ X) ∪ ((pos(C) \ U) ∩ Y)
where X ⊆ U and Y ∩ U = ∅
5. thus, Y |= C iff
l ≤
X
Bi ∈pos(C)∩(X∪Y)
wi ≤ u
6. which is the same as X ∪ Y |= C
Lemma 2 Let U be a splitting set for Π, and X, Y be subsets of lit(Π) s.t. X ⊆ U and Y ∩ U = ∅.
Then the body of a rule R0 ∈ evalU (Π, X) is satisfied by Y iff the body of the rule R ∈ Π it was
constructed from is satisfied by X ∪ Y.
Proof: Y satisfies body(R0 )
iff pos(R0 ) ⊆ Y, neg(R0 ) ∩ Y = ∅, Y |= C 0 for each C 0 ∈ agg(R0 ) – by definition of rule
satisfaction
iff (pos(R) ∩ U) ⊆ X, (pos(R) \ U) ⊆ Y, (neg(R) ∩ U) ∩ X) = ∅, (neg(R) \ U) ∩ Y) = ∅, X
satisfies C for all C ∈ agg(C) in which lit(C) ⊆ U, and Y satisfies evalU (C, X) for all C ∈ agg(C)
in which lit(C) * U – using definition of partial evaluation
iff pos(R) ⊆ X ∪ Y, neg(R) ∩ (X ∪ Y) = ∅, X ∪ Y |= C – using
• (A ∩ U) ∪ (A \ U) = A
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
179
• A ∩ (X ∪ Y) = ((A ∩ U) ∪ (A \ U)) ∩ (X ∪ Y) = ((A ∩ U) ∩ (X ∪ Y)) ∪ ((A \ U) ∩ (X ∪ Y)) =
(A ∩ U) ∩ X) ∪ ((A \ U) ∩ Y) – given X ⊆ U and Y ∩ U = ∅.
• and lemma 1
Proof of Splitting Theorem: Let U be a splitting set of Π, then a consistent set of literals S is an
answer set of Π iff it can be written as S = X ∪ Y, where X is an answer set of botU (Π); and Y
is an answer set of evalU (Π \ botU (Π), Y).
(⇐) Let X is an answer set of botU (Π); and Y is an answer set of evalU (Π \ botU (Π), X); we
show that X ∪ Y is an answer set of Π.
By definition of botU (Π), lit(botU (Π)) ⊆ U. In addition, by proposition 13, Y ∩ U = ∅.
Then, ΠX∪Y = (botU (Π) ∪ (Π \ botU (Π)))X∪Y = botU (Π)X∪Y ∪ (Π \ botU (Π))X∪Y = botU (Π)X ∪
(Π \ botU (Π))X∪Y .
Let r be a rule in ΠX∪Y , s.t. X ∪ Y |= body(r) then we show that X ∪ Y |= head(r). The rule
r either belongs to botU (Π)X or (Π \ botU (Π))X∪Y .
Case 1: say r ∈ botU (Π)X be a rule whose body is satisfied by X ∪ Y
1. then there is a rule R ∈ botU (Π) s.t. r ∈ RX
2. then X |= body(R) – since X ∪ Y |= body(r); lit(botU (Π)) ⊆ U and Y ∩ U = ∅
3. we already have X |= head(R) – since X is an answer set of botU (Π); given
4. then X ∪ Y |= head(R) – because lit(R) ⊆ U and Y ∩ U = ∅
5. consequently, X ∪ Y |= head(r)
Case 2: say r ∈ (Π \ botU (Π))X∪Y be a rule whose body is satisfied by X ∪ Y
1. then there is a rule R ∈ (Π \ botU (Π)) s.t. r ∈ RX∪Y
2. then lit(head(R)) ∩ U = ∅ – otherwise, R would have belonged to botU (Π), by definition
of splitting set
3. then head(r) ∈ Y – since X ⊆ U
4. in addition, pos(R) ⊆ X ∪ Y, neg(R) ∩ (X ∪ Y) = ∅, X ∪ Y |= C for each C ∈ agg(R) –
using definition of reduct
5. then pos(R) ∩ U ⊆ X or pos(R) \ U ⊆ Y; (neg(R) ∩ U) ∩ X = ∅ and (neg(R) \ U) ∩ Y = ∅;
and for each C ∈ agg(R), either X |= C if lit(C) ⊆ U, or Y |= evalU (C, X) if lit(C) * U
– by rearranging, lemma 1, X ⊆ U, Y ∩ U = ∅, and definition of partial evaluation of an
aggregate atom
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
180
6. note that pos(R) ∩ U ⊆ X, (neg(R) ∩ U) ∩ X = ∅, and for each C ∈ agg(R), s.t. lit(C) ⊆ U,
X |= C, represent conditions satisfied by each rule that become part of a partial evaluation
– using definition of partial evaluation
7. and pos(R) \ U, neg(R) \ U, and for each C ∈ agg(R), evalU (C, X) are the modifications
made to the rule during partial evaluation given splitting set U w.r.t. X – using definition
of partial evaluation
8. and pos(R) \ U ⊆ Y, (neg(R) \ U) ∩ Y = ∅, and for each C ∈ agg(R), Y |= evalU (C, X) if
lit(C) * U represent conditions satisfied by rules that become part of the reduct w.r.t Y –
using definition of partial evaluation and reduct
9. then r is a rule in reduct evalU (Π \ botU (Π), X)Y – using (8), (2) above
10. in addition, given that Y satisfies evalU (Π \ botU (Π), X), and head(r) ∩ U = ∅, we have
X ∪ Y |= head(r)
Next we show that X ∪Y satisfies all rules of Π. Say, R is a rule in Π not satisfied by X ∪Y. Then,
either it belongs to botU (Π) or (Π \ botU (Π)). If it belongs to botU (Π), it must not be satisfied by
X, since lit(botU (Π)) ⊆ U and Y ∩U = ∅. However, the contrary is given to be true. On the other
hand if it belongs to (Π \ botU (Π)), then X ∪ Y satisfies body(R) but not head(R). That would
mean that its head(R) is not satisfied by Y, since head(R) ∩ U = ∅ by definition of splitting set.
However, from lemma 2 we know that if body(R) is satisfied by X ∪ Y, body(R0 ) is satisfied by Y
for R0 ∈ evalU (Π \ botU (Π), X). We also know that Y satisfies all rules in evalU (Π \ botU (Π), X).
So, R0 must be satisfied by Y contradicting our assumption. Thus, all rules of Π are satisfied by
X ∪ Y and X ∪ Y is an answer set of Π.
(⇒) Let S be a consistent answer set of Π, we show that S = X ∪ Y for sets X and Y s.t. X is
an answer set of botU (Π) and Y is an answer set of evalU (Π \ botU (Π), X). We take X = S ∩ U,
Y = S \ U, then S = X ∪ Y.
Case 1: We show that X is answer set of botU (Π)
1. Π can be split into botU (Π) ∪ (Π \ botU (Π)) – by definition of splitting
2. then X ∪ Y satisfies botU (Π) – X ∪ Y is an answer set of Π; given
3. however lit(botU (Π)) ⊆ U, Y ∩ U = ∅ – by definition of splitting
4. then X satisfies botU (Π) – since elements of Y do not appear in the rules of botU (Π)
5. then X is an answer set of botU (Π)
Case 2: We show that Y is answer set of evalU (Π \ botU (Π), X)
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
181
1. let r be a rule in evalU (Π \ botU (Π), X)Y , s.t. its body is satisfied by Y
2. then r ∈ RY for an R ∈ evalU (Π \ botU (Π), X) s.t. (i) pos(R) ⊆ Y (ii) neg(R) ∩ Y = ∅
(iii) Y |= C for all C ∈ agg(R) (iv) head(R) ∩ Y , ∅ – using definition of reduct
3. each R is constructed from R0 ∈ Π that satisfies all the following conditions (i) pos(R0 ) ⊆
U ∩ X (ii) (neg(R0 ) ∩ U) ∩ X = ∅ (iii) if there is a C 0 ∈ agg(R0 ) s.t. lit(C 0 ) ⊆ U, then
X |= C 0 ; and each C ∈ agg(R) is a partial evaluation of C 0 ∈ agg(R0 ) s.t. C = evalU (C 0 , X)
– using definition of partial evaluation
4. then the body(R0 ) satisfies all the following conditions:
(a) pos(R0 ) ⊆ X ∪ Y – since X ⊆ U, X ∩ Y = ∅
(b) neg(R0 ) ∩ (X ∪ Y) = ∅ – since X ⊆ U, X ∩ Y = ∅
(c) X ∪ Y |= C 0 for each C 0 ∈ agg(R0 ) – since (i) each C 0 ∈ agg(R0 ) with lit(C 0 ) ⊆ U
satisfied by X is also satisfied by X ∪ Y as lit(Y) ∩ lit(C 0 ) = ∅; and (ii) each C 0 ∈
agg(R0 ) with lit(C 0 ) * U is satisfied by X ∪ Y – using partial evaluation, reduct
construction, and X ∩ Y = ∅
5. then X ∪ Y satisfies body(R0 ) – from previous line
6. in addition, lit(head(R0 )) ∩ U = ∅, otherwise, R0 would have belonged to botU (Π) by
definition of splitting set
7. then R0 is a rule in Π \ botU (Π) – from the last two lines
8. we know that X ∪ Y satisfies every rule in (Π \ botU (Π)) – given; and that elements of U
do not appear in the head of rules in (Π \ botU (Π)) – from definition of splitting; then Y
must satisfy the head of these rules
9. then Y satisfies head(R0 ) – from (8)
10. Next we show that Y satisfies all rules of evalU (Π \ botU (Π), X). Let R0 be a rule in
evalU (Π\botU (Π), X) such that body(R0 ) is satisfied by Y but not head(R0 ). Since head(R0 )∩
Y = ∅, head(R0 ) is not satisfied by X ∪ Y either. Then, there is an R ∈ (Π \ botU (Π)) such
that X ∪ Y satisfies body(R) but not head(R), which contradicts given. Thus, Y satisfies all
rules of evalU (Π \ botU (Π), X).
11. Then Y is an answer set of evalU (Π \ botU (Π), X)
Definition 92 (Splitting Sequence) Baral (2003) A splitting sequence for a program Π is a
S
monotone, continuous sequence hUα iα<µ of splitting sets of Π such that α<µ Uµ = lit(Π).
Definition 93 (Solution) Baral (2003) Let U = hUα iα<µ be a splitting sequence for a program
Π. A solution to Π w.r.t U is a sequence hXα iα<µ of sets of literals such that:
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
182
1. X0 is an answer set for botU0 (Π)
2. for any ordinal α + 1 < µ, Xα+1 is an answer set of the program:
[
evalUα (botUα+1 (Π) \ botUα (Π),
Xν )
ν≤α
3. for any limit ordinal α < µ, Xα = ∅, and
S
4. α≤µ (Xα ) is consistent
Proposition 15 (Splitting Sequence Theorem) Baral (2003) Let U = hUα iα<µ be a splitting
S
sequence for a program Π. A set S of literals is a consistent answer set for Π iff S = α<µ Xα
for some solution hXα iα<µ to Π w.r.t U.
Proof: Let U = hUα iα<µ be a splitting sequence of Π, then a consistent set of literals S =
S
S
α<µ Xα is an answer set of Π iff X0 is an answer set ofSbotU0 (Π) and for any ordinal α + 1 < µ,
Xα+1 is an answer set of evalUα (botUα+1 (Π) \ botUα (Π), ν≤α Xν ).
Note that every literal in botU0 (Π) belongs to lit(Π) ∩ U0 , and every literal occurring in
S
evalUα (botUα+1 (Π) \ botUα (Π), ν≤α Xν ), (α + 1 < µ) belongs to lit(Π) ∩ (Uα+1 \ Uα ). In addition,
X0 , and all Xα+1 are pairwise disjoint.
We prove the theorem by induction over the splitting sequence.
Base case: α = 1. The splitting sequence is U0 ⊆ U1 .
Then the sub-program Π1 = botU1 (Π) contains all literals in U1 ; and U0 splits Π1 into
botU0 (Π1 ) and botU1 (Π1 ) \ botU0 (Π1 ). Then, S 1 = X0 ∪ X1 is a consistent answer set of Π1 iff
X0 = S 1 ∩ U0 is an answer set of botU0 (Π1 ) and X1 = S 1 \ U0 is an answer set of evalU0 (Π1 \
botU0 (Π1 ), X1 ) – by the splitting theorem
Since botU0 (Π1 ) = botU0 (Π) and botU1 (Π1 ) \ botU0 (Π1 ) = botU1 (Π) \ botU0 (Π); S 1 = X0 ∪ X1
is an answer set for Π \ botU1 (Π).
Induction: Assume theorem holds for α = k, show theorem holds for α = k + 1.
The inductive assumption holds for the splitting sequence U0 ⊆ · · · ⊆ Uk . Then the subprogram Πk = botUk (Π) contains all literals in Uk and S k = X0 ∪ · · · ∪ Xk is a consistent
answer set of ΠS k iff X0 is an answer set for botU0 (Πk ) and for any α ≤ k, Xα+1 is answer set of
evalUα (botUα+1 (Π) \ botUα (Π), X0 ∪ · · · ∪ Xα )
We show that the theorem holds for α = k + 1. The splitting sequence is U0 ⊆ Uk+1 .
Then the sub-program Πk+1 = botUk+1 (Π) contains all literals Uk+1 . We have Uk split Πk+1
into botUk (Πk+1 ) and botUk+1 (Πk+1 ) \ botUk (Πk+1 ). Then, S k+1 = X0:k ∪ Xk+1 is a consistent
answer set of Πk+1 iff X0:k = S k+1 ∩ Uk is an answer set of botUk (Πk+1 ) and Xk+1 = S k+1 \ Uk is
an answer set of evalUk (Πk+1 \ botUk (Πk+1 , Xk+1 ) – by the splitting theorem
Since botUk (Πk+1 ) = botUk (Π) and botUk+1 (Πk+1 ) \ botUk (Πk+1 ) = botUk+1 (Π) \ botUk (Π);
S k+1 = X0:k ∪ X1 is an answer set for Π \ botUk+1 (Π).
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
183
From the inductive assumption we know that X0 ∪ · · · ∪ Xk is a consistent answer set of
botUk (Π), X0 is the answer set of botU0 (Π), and for each 0 ≤ α ≤ k, Xα+1 is answer set of
evalUα (botUα+1 (Π) \ botUα (Π), X0 ∪ · · · ∪ Xα ). Thus, X0:k = X0 ∪ · · · ∪ Xk .
Combining above with the inductive assumption, we get S k+1 = X0 ∪· · ·∪Xk+1 is a consistent
answer set of ΠS k+1 iff X0 is an answer set for botU0 (Πk+1 ) and for any α ≤ k + 1, Xα+1 is answer
set of evalUα (botUα+1 (Π) \ botUα (Π), X0 ∪ · · · ∪ Xα ).
In addition, for some α < µ, where µ is the length of the splitting sequence U = hUα iα<µ
of Π, botUα (Π) will be the entire Π, i.e. lit(Π) = Uα . Then the set S of literals is a consistent
S
answer set of Π iff S = α<µ (Xα ) for some solution hXα iα<µ to Π w.r.t U.
A.1
Proof of Proposition 1
Let PN = (P, T, E, W) be a Petri Net, M0 be its initial marking and let Π0 (PN, M0 , k, ntok) be
the ASP encoding of PN and M0 over a simulation length k, with maximum ntok tokens on any
place node, as defined in section 2.3. Then X = M0 , T 0 , M1 , . . . , Mk , T k , Mk+1 is an execution
sequence of a PN (w.r.t. M0 ) iff there is an answer set A of Π0 (PN, M0 , k, ntok) such that:
{ f ires(t, ts) : t ∈ T ts , 0 ≤ ts ≤ k} = { f ires(t, ts) : f ires(t, ts) ∈ A}
{holds(p, q, ts) : p ∈ P, q = Mts (p), 0 ≤ ts ≤ k + 1}
= {holds(p, q, ts) : holds(p, q, ts) ∈ A}
(A.1)
(A.2)
We prove this by showing that:
(I) Given an execution sequence X, we create a set A such that it satisfies (A.1) and (A.2) and
show that A is an answer set of Π0
(II) Given an answer set A of Π0 , we create an execution sequence X such that (A.1) and (A.2)
are satisfied
First we show (I): Given PN and its execution sequence X, we create a set A as a union of the
following sets:
1. A1 = {num(n) : 0 ≤ n ≤ ntok}
2. A2 = {time(ts) : 0 ≤ ts ≤ k}
3. A3 = {place(p) : p ∈ P}
4. A4 = {trans(t) : t ∈ T }
5. A5 = {ptarc(p, t, n) : (p, t) ∈ E − , n = W(p, t)}, where E − ⊆ E
6. A6 = {tparc(t, p, n) : (t, p) ∈ E + , n = W(t, p)}, where E + ⊆ E
7. A7 = {holds(p, q, 0) : p ∈ P, q = M0 (p)}
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
184
8. A8 = {notenabled(t, ts) : t ∈ T, 0 ≤ ts ≤ k, ∃p ∈ •t, Mts (p) < W(p, t)}
per definition 12 (enabled transition)
9. A9 = {enabled(t, ts) : t ∈ T, 0 ≤ ts ≤ k, ∀p ∈ •t, W(p, t) ≤ Mts (p)}
per definition 12 (enabled transition)
10. A10 = { f ires(t, ts) : t ∈ T ts , 0 ≤ ts ≤ k}
from definition 15 (firing set) , only an enabled transition may fire
11. A11 = {add(p, q, t, ts) : t ∈ T ts , p ∈ t•, q = W(t, p), 0 ≤ ts ≤ k}
per definition 13 (transition execution)
12. A12 = {del(p, q, t, ts) : t ∈ T ts , p ∈ •t, q = W(p, t), 0 ≤ ts ≤ k}
per definition 13 (transition execution)
P
13. A13 = {tot incr(p, q, ts) : p ∈ P, q = t∈Tts ,p∈t• W(t, p), 0 ≤ ts ≤ k}
per definition 16 (firing set execution)
P
14. A14 = {tot decr(p, q, ts) : p ∈ P, q = t∈Tts ,p∈•t W(p, t), 0 ≤ ts ≤ k}
per definition 16 (firing set execution)
P
15. A15 = {consumesmore(p, ts) : p ∈ P, q = Mts (p), q1 = t∈Tts ,p∈•t W(p, t), q1 > q, 0 ≤
ts ≤ k}
per definition 14 (conflicting transitions) for enabled transition set T ts
P
16. A16 = {consumesmore : ∃p ∈ P : q = Mts (p), q1 = t∈Tts ,p∈•t W(p, t), q1 > q, 0 ≤ ts ≤ k}
per definition 14 (conflicting transitions)
17. A17 = {holds(p, q, ts + 1) : p ∈ P, q = Mts+1 (p), 0 ≤ ts < k},
P
P
where Mts+1 (p) = Mts (p) − t∈Tts ,p∈•t W(p, t) + t∈Tts ,p∈t• W(t, p)
according to definition 15 (firing set execution)
We show that A satisfies (A.1) and (A.2), and A is an answer set of Π0 .
A satisfies (A.1) and (A.2) by its construction (given above). We show A is an answer set of
Π0 by splitting. We split lit(Π0 ) (literals of Π0 ) into a sequence of 6(k + 1) + 2 sets:
• U0 = head( f 1) ∪ head( f 2) ∪ head( f 3) ∪ head( f 4) ∪ head( f 5) ∪ head( f 6) ∪ head(i1) =
{place(p) : p ∈ P} ∪ {trans(t) : t ∈ T } ∪ {ptarc(p, t, n) : (p, t) ∈ E − , n = W(p, t)} ∪
{tparc(t, p, n) : (t, p) ∈ E + , n = W(t, p)}∪ {time(0), . . . , time(k)}∪{num(0), . . . , num(ntok)}∪
{holds(p, q, 0) : p ∈ P, q = M0 (p)}
• U6k+1 = U6k+0 ∪ head(e1)ts=k = U6k+0 ∪ {notenabled(t, k) : t ∈ T }
• U6k+2 = U6k+1 ∪ head(e2)ts=k = U6k+1 ∪ {enabled(t, k) : t ∈ T }
• U6k+3 = U6k+2 ∪ head(a1)ts=k = U6k+2 ∪ { f ires(t, k) : t ∈ T }
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
185
• U6k+4 = U6k+3 ∪ head(r1)ts=k ∪ head(r2)ts=k = U6k+3 ∪ {add(p, q, t, k) : p ∈ P, t ∈ T, q =
W(t, p)} ∪ {del(p, q, t, k) : p ∈ P, t ∈ T, q = W(p, t)}
• U6k+5 = U6k+4 ∪ head(r3)ts=k ∪ head(r4)ts=k = U6k+4 ∪ {tot incr(p, q, k) : p ∈ P, 0 ≤ q ≤
ntok} ∪ {tot decr(p, q, k) : p ∈ P, 0 ≤ q ≤ ntok}
• U6k+6 = U6k+5 ∪ head(r5)ts=k ∪ head(a2)ts=k = U6k+5 ∪ {holds(p, q, k + 1) : p ∈ P, 0 ≤
q ≤ ntok} ∪ {consumesmore(p, k) : p ∈ P}
• U6k+7 = U6k+6 ∪ head(a3) = U7k+6 ∪ {consumesmore}
where head(ri )ts=k are head atoms of ground rule ri in which ts = k. We write Ats=k
=
i
{a(. . . , ts) : a(. . . , ts) ∈ Ai , ts = k} as short hand for all atoms in Ai with ts = k. Uα , 0 ≤
α ≤ 6k + 7 form a splitting sequence, since each Ui is a splitting set of Π0 , and hUα iα<µ is a
S
monotone continuous sequence, where U0 ⊆ U1 · · · ⊆ U6k+7 and α<µ Uα = lit(Π0 ).
We compute the answer set of Π0 using the splitting sets as follows:
1. botU0 (Π0 ) = f 1 ∪ f 2 ∪ f 3 ∪ f 4 ∪ f 5 ∪ i1 ∪ f 6 and X0 = A1 ∪ · · · ∪ A7 (= U0 ) is its answer
set – using forced atom proposition
2. evalU0 (botU1 (Π0 )\botU0 (Π0 ), X0 ) = {notenabled(t, 0):-.|{trans(t), ptarc(p, t, n), holds(p, q, 0)} ⊆
– using forced atom proposition and conX0 , where q < n}. Its answer set X1 = Ats=0
8
struction of A8 .
(a) where, q = M0 (p), n = W(p, t), and for an arc (p, t) ∈ E − – by construction of i1
and f 3 in Π0 , and
(b) in an arc (p, t) ∈ E − , p ∈ •t (by definition 11 of preset)
(c) thus, notenabled(t, 0) ∈ X1 represents ∃p ∈ •t : M0 (p) < W(p, t).
3. evalU1 (botU2 (Π0 )\botU1 (Π0 ), X0 ∪X1 ) = {enabled(t, 0):-.|trans(t) ∈ X0 ∪X1 , notenabled(t, 0) <
– using forced atom proposition and construction
X0 ∪ X1 }. Its answer set is X2 = Ats=0
9
of A9 .
(a) since an enabled(t, 0) ∈ X2 if @ notenabled(t, 0) ∈ X0 ∪ X1 , which is equivalent to
@p ∈ •t : M0 (p) < W(p, t) ≡ ∀p ∈ •t : M0 (p) ≥ W(p, t).
4. evalU2 (botU3 (Π0 ) \ botU2 (Π0 ), X0 ∪ X1 ∪ X2 ) = {{ f ires(t, 0)}:-.|enabled(t, 0)
holds in X0 ∪ X1 ∪ X2 }. It has multiple answer sets X3.1 , . . . , X3.n , corresponding to elements of power set of f ires(t, 0) atoms in evalU2 (...) – using supported rule proposition.
Since we are showing that the union of answer sets of Π0 determined using splitting is
equal to A, we only consider the set that matches the f ires(t, 0) elements in A and call it
X3 , ignoring the rest. Thus, X3 = Ats=0
10 , representing T 0 .
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
186
5. evalU3 (botU4 (Π0 ) \ botU3 (Π0 ), X0 ∪ · · · ∪ X3 ) = {add(p, n, t, 0):-.|{ f ires(t, 0),
tparc(t, p, n)} ⊆ X0 ∪· · ·∪X3 }∪{del(p, n, t, 0):-.|{ f ires(t, 0), ptarc(p, t, n)} ⊆ X0 ∪· · ·∪X3 }.
ts=0
It’s answer set is X4 = Ats=0
11 ∪ A12 – using forced atom proposition and definitions of
A11 and A12 .
(a) where each add atom is equivalent to n = W(t, p) : p ∈ t•,
(b) and each del atom is equivalent to n = W(p, t) : p ∈ •t,
(c) representing the effect of transitions in T 0 – by construction
6. evalU4 (botU5 (Π0 ) \ botU4 (Π0 ), X0 ∪ · · · ∪ X4 ) = {tot incr(p, qq, 0):-.|
P
P
qq = add(p,q,t,0)∈X0 ∪···∪X4 q}∪ {tot decr(p, qq, 0):-.|qq = del(p,q,t,0)∈X0 ∪···∪X4 q}. It’s ants=0
swer set is X5 = Ats=0
13 ∪ A14 – using forced atom proposition and definitions of A13 and
A14 , ad definition 3 (semantics of aggregate assignment atom).
P
(a) where each tot incr(p, qq, 0), qq = add(p,q,t,0)∈X0 ∪...X4 q
P
≡ qq = t∈X3 ,p∈t• W(p, t),
P
(b) and each tot decr(p, qq, 0), qq = del(p,q,t,0)∈X0 ∪...X4 q
P
≡ qq = t∈X3 ,p∈•t W(t, p),
(c) represent the net effect of transitions in T 0 – by construction
7. evalU5 (botU6 (Π0 ) \ botU5 (Π0 ), X0 ∪ · · · ∪ X5 ) = {consumesmore(p, 0):-.|
{holds(p, q, 0), tot decr(p, q1, 0)} ⊆ X0 ∪· · ·∪X5 , q1 > q}∪ {holds(p, q, 1):-.| {holds(p, q1, 0),
tot incr(p, q2, 0), tot decr(p, q3, 0)} ⊆ X0 ∪ · · · ∪ X5 , q = q1 + q2 − q3}. It’s answer set is
∪ Ats=0
X6 = Ats=0
17 – using forced atom proposition.
15
(a) where, consumesmore(p, 0) represents ∃p : q = M0 (p),
P
q1 = t∈T0 ,p∈•t W(p, t), q1 > q – indicating place p will be overconsumed if T 0 is
fired as defined in definition 14 (conflicting transitions)
(b) and holds(p, q, 1) represents q = M1 (p) – by construction
..
.
8. evalU6k+0 (botU6k+1 (Π0 ) \ botU6k+0 (Π0 ), X0 ∪ · · · ∪ X6k+0 ) =
{notenabled(t, k):-.|{trans(t), ptarc(p, t, n), holds(p, q, k)} ⊆ X0 ∪ · · · ∪ X6k+0 ,
where q < n}. Its answer set X6k+1 = Ats=k
– using forced atom proposition and construc8
tion of A8 .
(a) where, q = Mk (p), and n = W(p, t) for an arc (p, t) ∈ E − – by construction of holds
and ptarc predicates in Π0 , and
(b) in an arc (p, t) ∈ E − , p ∈ •t (by definition 11 of preset)
(c) thus, notenabled(t, k) ∈ X6k+1 represents ∃p ∈ •t : Mk (p) < W(p, t).
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
187
9. evalU6k+1 (botU6k+2 (Π0 ) \ botU6k+1 (Π0 ), X0 ∪ · · · ∪ X6k+1 ) = {enabled(t, k):-.|
trans(t) ∈ X0 ∪ · · · ∪ X6k+1 , notenabled(t, k) < X0 ∪ · · · ∪ X6k+1 }. Its answer set is X6k+2 =
Ats=k
– using forced atom proposition and construction of A9 .
9
(a) since an enabled(t, k) ∈ X6k+2 if @ notenabled(t, k) ∈ X0 ∪ · · · ∪ X6k+1 , which is
equivalent to @p ∈ •t : Mk (p) < W(p, t) ≡ ∀p ∈ •t : Mk (p) ≥ W(p, t).
10. evalU6k+2 (botU6k+3 (Π0 ) \ botU6k+2 (Π0 ), X0 ∪ · · · ∪ X6k+2 ) =
{{ f ires(t, k)}:-.| enabled(t, k) holds in X0 ∪· · ·∪X6k+2 }. It has multiple answer sets X6k+3.1 , . . . , X6k+3.n ,
corresponding to elements of power set of f ires(t, k) atoms in evalU6k+2 (...) – using supported rule proposition. Since we are showing that the union of answer sets of Π0 determined using splitting is equal to A, we only consider the set that matches the f ires(t, k)
elements in A and call it X6k+3 , ignoring the reset. Thus, X6k+3 = Ats=k
10 , representing T k .
11. evalU6k+3 (botU6k+4 (Π0 ) \ botU6k+3 (Π0 ), X0 ∪ · · · ∪ X6k+3 ) = {add(p, n, t, k):-.|
{ f ires(t, k), tparc(t, p, n)} ⊆ X0 ∪· · ·∪X6k+3 }∪ {del(p, n, t, k):-.| { f ires(t, k), ptarc(p, t, n)} ⊆
ts=k
X0 ∪ · · · ∪ X6k+3 }. It’s answer set is X6k+4 = Ats=k
11 ∪ A12 – using forced atom proposition
and definitions of A11 and A12 .
(a) where, each add atom is equivalent to n = W(t, p) : p ∈ t•,
(b) and each del atom is equivalent to n = W(p, t) : p ∈ •t
(c) representing the effect of transitions in T k
12. evalU6k+4 (botU6k+5 (Π0 ) \ botU6k+4 (Π0 ), X0 ∪ · · · ∪ X6k+4 ) = {tot incr(p, qq, k):-.|
P
qq = add(p,q,t,k)∈X0 ∪···∪X6k+4 q}∪ {tot decr(p, qq, k):-.|
P
ts=k
qq = del(p,q,t,k)∈X0 ∪···∪X6k+4 q}. It’s answer set is X5 = Ats=k
13 ∪ A14 – using forced atom
proposition and definitions of A13 and A14 .
P
(a) where, each tot incr(p, qq, k), qq = add(p,q,t,k)∈X0 ∪...X7k+4 q
P
≡ qq = t∈X6k+3 ,p∈t• W(p, t),
P
(b) and each tot decr(p, qq, k), qq = del(p,q,t,k)∈X0 ∪...X7k+4 q
P
≡ qq = t∈X6k+3 ,p∈•t W(t, p),
(c) represent the net effect of transitions in T k
13. evalU6k+5 (botU6k+6 (Π0 )\botU6k+5 (Π0 ), X0 ∪· · ·∪X6k+5 ) = {consumesmore(p, k):-.| {holds(p, q, k),
tot decr(p, q1, k)} ⊆ X0 ∪· · ·∪X6k+5 , q1 > q}∪{holds(p, q, k+1):-., |{holds(p, q1, k), tot incr(p, q2, k),
tot decr(p, q3, k)} ⊆ X0 ∪ · · · ∪ X6k+5 , q = q1 + q2 − q3}. It’s answer set is X6k+6 =
Ats=k
∪ Ats=k
17 – using forced atom proposition.
15
(a) where, holds(p, q, k + 1) represents the marking of place p in the next time step due
to firing T k ,
P
(b) and, consumesmore(p, k) represents ∃p : q = Mk (p), q1 = t∈Tk ,p∈•t W(p, t), q1 > q
indicating place p that will be overconsumed if T k is fired as defined in definition 14
(conflicting transitions)
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
188
14. evalU6k+6 (botU6k+7 (Π0 )\botU6k+6 (Π0 ), X0 ∪· · ·∪X6k+6 ) = {consumesmore:-.| {consumesmore(p, 0),
. . . , consumesmore(p, k)} ∩ (X0 ∪ · · · ∪ X6k+7 ) , ∅}. It’s answer set is X6k+7 = Ats=k
–
16
using forced atom proposition
(a) X6k+7 will be empty since none of consumesmore(p, 0), . . . ,
consumesmore(p, k) hold in X0 ∪ · · · ∪ X6k+6 due to the construction of A and encoding of a2, and it is not eliminated by the constraint a4.
The set X = X0 ∪ · · · ∪ X6k+7 is the answer set of Π0 by the splitting sequence theorem 15.
Each Xi , 0 ≤ i ≤ 6k + 7 matches a distinct partition of A, and X = A, thus A is an answer set of
Π0 .
Next we show (II): Given Π0 be the encoding of a Petri Net PN(P, T, E, W) with initial
marking M0 , and A be an answer set of Π0 that satisfies (A.1) and (A.2), then we can construct
X = M0 , T 0 , . . . , Mk , T k , Mk+1 from A, such that it is an execution sequence of PN.
We construct the X as follows:
1. Mi = (Mi (p0 ), . . . , Mi (pn )), where {holds(p0 , Mi (p0 ), i), . . . holds(pn , Mi (pn ), i)}
⊆ A, for 0 ≤ i ≤ k + 1
2. T i = {t : f ires(t, i) ∈ A}, for 0 ≤ i ≤ k
and show that X is indeed an execution sequence of PN. We show this by induction over k (i.e.
given Mk , T k is a valid firing set and its firing produces marking Mk+1 ).
Base case: Let k = 0, show (1) T 0 is a valid firing set for M0 , and (2) T 0 ’s firing in M0 producing
marking M1 .
1. We show T 0 is a valid firing set w.r.t. M0 . Let { f ires(t0 , 0), . . . , f ires(t x , 0)} be the set of
all f ires(. . . , 0) atoms in A,
(a) Then for each f ires(ti , 0) ∈ A
enabled(ti , 0) ∈ A – from rule a1 and supported rule proposition
Then notenabled(ti , 0) < A – from rule e2 and supported rule proposition
Then body(e1) must not hold in A – from rule e1 and forced atom proposition
Then q ≮ ni ≡ q ≥ ni in e1 for all {holds(p, q, 0), ptarc(p, ti , ni )} ⊆ A – from e1,
forced atom proposition, and the following:
A. holds(p, q, 0) ∈ A represents q = M0 (p) – rule i1 construction
B. ptarc(p, ti , ni ) ∈ A represents ni = W(p, ti ) – rule f 3 construction
v. Then ∀p ∈ •ti , M0 (p) > W(p, ti ) – from definition 11 of preset •ti in PN
i.
ii.
iii.
iv.
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
189
vi. Then ti is enabled and can fire in PN, as a result it can belong to T 0 – from
definition 12 of enabled transition
(b) And consumesmore < A, since A is an answer set of Π0 – from rule a4 and supported
rule proposition
i. Then @consumesmore(p, 0) ∈ A – from rule a3 and supported rule proposition
ii. Then @{holds(p, q, 0), tot decr(p, q1, 0)} ⊆ A : q1 > q in body(a2) – from a2
and forced atom proposition
P
iii. Then @p : ti ∈{t0 ,...,tx },p∈•ti W(p, ti ) > M0 (p) – from the following
A. holds(p, q, 0) represents q = M0 (p) – from rule i1 construction, given
B. tot decr(p, q1, 0) ∈ A if {del(p, q10 , t0 , 0), . . . del(p, q1 x , t x , 0)} ⊆ A, where
q1 = q10 + · · · + q1 x – from r4 and forced atom proposition
C. del(p, q1i , ti , 0) ∈ A if { f ires(ti , 0), ptarc(p, ti , q1i )} ⊆ A – from r2 and
supported rule proposition
D. del(p, q1i , ti , 0) ∈ A represents removal of q1i = W(p, ti ) tokens from p ∈
•ti – from r2, supported rule proposition, and definition 13 of transition
execution in PN
iv. Then the set of transitions in T 0 do not conflict – by the definition 14 of conflicting transitions
(c) Then {t0 , . . . , t x } = T 0 – using 1(a),1(b) above, and definition of firing set
2. We show M1 is produced by firing T 0 in M0 . Let holds(p, q, 1) ∈ A
(a) Then {holds(p, q1, 0), tot incr(p, q2, 0), tot decr(p, q3, 0)} ⊆ A : q = q1 + q2 − q3 –
from rule r5 and supported rule proposition
(b) Then holds(p, q1, 0) ∈ A represents q1 = M0 (p) – given, rule i1 construction; and
{add(p, q0 , t0 , 0), . . . , add(p, q j , t j , 0)} ⊆ A : q0 +· · ·+q j = q2 ; and {del(p, q0 , t0 , 0), . . . ,
del(p, ql , tl , 0)} ⊆ A : q0 + · · · + ql = q3 – rules r3, r4 and supported rule proposition,
respectively
(c) Then { f ires(t0 , 0), . . . , f ires(t j , 0)} ⊆ A and { f ires(t0 , 0), . . . , f ires(tl , 0)} ⊆ A –
rules r1, r2 and supported rule proposition; and the following
i. tparc(ty , p, qy ) ∈ A, 0 ≤ y ≤ j represents qy = W(ty , p) – given
ii. ptarc(p, tz , qz ) ∈ A, 0 ≤ z ≤ l represents qz = W(p, tz ) – given
(d) Then { f ires(t0 , 0), . . . , f ires(t j , 0)} ∪ { f ires(t0 , 0), . . . , f ires(tl , 0)} ⊆ A
= { f ires(t0 , 0), . . . , f ires(t x , 0)} ⊆ A – set union of subsets
(e) Then for each f ires(t x , 0) ∈ A we have t x ∈ T 0 – already shown in item 1 above
P
P
(f) Then q = M0 (p)+ tx ∈T0 ∧p∈tx • W(t x , p)− tx ∈T0 ∧p∈•tx W(p, t x ) – from (2b) above and
the following
i. Each add(p, q j , t j , 0) ∈ A represents q j = W(t j , p) for p ∈ t j • – rule r1 encoding, and definition 13 of transition execution in PN
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
190
ii. Each del(p, ty , qy , 0) ∈ A represents qy = W(p, ty ) for p ∈ •ty – from rule r2
encoding, and definition 13 of transition execution in PN
P
iii. Each tot incr(p, q2, 0) ∈ A represents q2 = tx ∈T0 ∧p∈tx • W(t x , p) – aggregate
assignment atom semantics in rule r3
P
iv. Each tot decr(p, q3, 0) ∈ A represents q3 = tx ∈T0 ∧p∈•tx W(p, t x ) – aggregate
assignment atom semantics in rule r4
(g) Then, M1 (p) = q – since holds(p, q, 1) ∈ A encodes q = M1 (p) – by construction
Inductive Step: Assume Mk is a valid marking in X for PN, show (1) T k is a valid firing set for
Mk , and (2) firing T k in Mk produces marking Mk+1 .
1. We show that T k is a valid firing set for Mk . Let { f ires(t0 , k), . . . , f ires(t x , k)} be the set
of all f ires(. . . , k) atoms in A,
(a) Then for each f ires(ti , k) ∈ A
i. enabled(ti , k) ∈ A – from rule a1 and supported rule proposition
ii. Then notenabled(ti , k) < A – from rule e2 and supported rule proposition
iii. Then body of e1 must hold in A – from rule e1 and forced proposition
iv. Then q ≮ ni ≡ q ≥ ni in e1 for all {holds(p, q, k), ptarc(p, ti , ni )} ⊆ A – from e1
using forced atom proposition, and the following
A. holds(p, q, k) ∈ A represents q = Mk (p) – construction, inductive assumption
B. ptarc(p, t, ni ) ∈ A represents ni = W(p, t) – rule f 4 construction
v. Then ∀p ∈ •ti , Mk (p) > W(p, ti ) – from definition 11 of preset •ti in PN
vi. Then ti is enabled and can fire in PN, as a result it can belong to T k – from
definition 12 of enabled transition
(b) And consumesmore < A, since A is an answer set of Π0 – from rule a4 and supported
rule proposition
i. Then @consumesmore(p, k) ∈ A – from rule a3 and forced atom proposition
ii. Then @{holds(p, q, k), tot decr(p, q1, k)} ⊆ A : q1 > q in body(a2) – from a2
and forced atom proposition
P
iii. Then @p : ti ∈{t0 ,...,tx },p∈•ti W(p, ti ) > Mk (p) – from the following
A. holds(p, q, k) ∈ A represents q = Mk (p) – by construction, and the inductive assumption
B. tot decr(p, q1, k) ∈ A if {del(p, q10 , t0 , k), . . . , del(p, q1 x , t x , k)} ⊆ A, where
q1 = q10 + · · · + q1 x – from r4 and forced atom proposition
C. del(p, q1i , ti , k) ∈ A if { f ires(ti , k), ptarc(p, ti , q1i )} ⊆ A – from r2 and
supported rule proposition
D. del(p, q1i , ti , k) represents removal of q1i = W(p, ti ) tokens from p ∈ •ti –
from construction rule r2, supported rule proposition, and definition 13 of
transition execution in PN
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
191
iv. Then the set of transitions T k does conflict – by the definition 14 of conflicting
transitions
(c) Then {t0 , . . . , t x } = T k – using 1(a),1(b) above
2. We show Mk+1 is produced by firing T k in Mk . Let holds(p, q, k + 1) ∈ A
(a) Then {holds(p, q1, k), tot incr(p, q2, k), tot decr(p, q3, k)} ∈ A : q = q2 + q2 − q3 –
from rule r5 and supported rule proposition
(b) Then holds(p, q1, k) ∈ A represents q1 = Mk (p) – inductive assumption and construction; and {add(p, q20 , t0 , k), . . . , add(p, q2 j , t j , k)} ⊆ A : q20 + · · · + q2 j = q2
and {del(p, q30 , t0 , k), . . . , del(p, q3l , tl , k)} ⊆ A : q30 + · · · + q3l = q3 – from rules
r3, r4 using supported rule proposition, respectively
(c) Then { f ires(t0 , k), . . . , f ires(t j , k)} ⊆ A and { f ires(t0 , k), . . . , f ires(tl , k)} ⊆ A – from
rules r1, r2 using supported rule proposition, respectively
(d) Then { f ires(t0 , k), . . . , f ires(t j , k)} ∪ { f ires(t0 , k), . . . , f ires(tl , k)}
= { f ires(t0 , k), . . . , f ires(t x , k)} ⊆ A – subset union property
(e) Then for each f ires(t x , k) ∈ A we have t x ∈ T k - already shown in item 1 above
P
P
(f) Then q = Mk (p) + tx ∈Tk ∧p∈tx • W(t x , p) − tx ∈Tk ∧p∈•tx W(p, t x – from (2b) above and
the following
i. Each add(p, q j , t j , 0) ∈ A represents q j = W(t j , p) for p ∈ t j • – encoding of r1
and definition 13 of transition execution in PN
ii. Each del(p, ty , qy , 0) ∈ A represents qy = W(p, ty ) for p ∈ •ty – encoding of r2
and definition 13 of transition execution in PN
P
iii. Each tot incr(p, q2, 0) ∈ A represents q2 = tx ∈T k ∧p∈tx • W(t x , p) – aggregate
assignment atom semantics in rule r3
P
iv. Each tot decr(p, q3, 0) ∈ A represents q3 = tx ∈Tk ∧p∈•tx W(p, t x ) – aggregate
assignment atom semantics in rule r4
(g) Then Mk+1 (p) = q – since holds(p, q, k + 1) ∈ A encodes q = Mk+1 (p) by construction
As a result, for any n > k, T n is a valid firing set w.r.t. Mn and its firing produces marking Mn+1 .
Conclusion: Since both I and II hold, X = M0 , T 0 , M1 , . . . , Mk , T k , Mk+1 is an execution sequence of PN(P, T, E, W) (w.r.t. M0 ) iff there is an answer set A of Π0 (PN, M0 , k, ntok) such
that (A.1) and (A.2) hold.
A.2
Proof of Proposition 2
Let PN = (P, T, E, W) be a Petri Net, M0 be its initial marking and let Π1 (PN, M0 , k, ntok) be
the ASP encoding of PN and M0 over a simulation length k, with maximum ntok tokens on any
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
192
place node, as defined in section 2.4. Then X = M0 , T 0 , M1 , . . . , Mk , T k , Mk+1 is an execution
sequence of PN (w.r.t. M0 ) iff there is an answer set A of Π1 (PN, M0 , k, ntok) such that:
{ f ires(t, ts) : t ∈ T ts , 0 ≤ ts ≤ k} = { f ires(t, ts) : f ires(t, ts) ∈ A}
{holds(p, q, ts) : p ∈ P, q = Mts (p), 0 ≤ ts ≤ k + 1}
= {holds(p, q, ts) : holds(p, q, ts) ∈ A}
(A.3)
(A.4)
We prove this by showing that:
(I) Given an execution sequence X, we create a set A such that it satisfies (A.3) and (A.4) and
show that A is an answer set of Π1
(II) Given an answer set A of Π1 , we create an execution sequence X such that (A.3) and (A.4)
are satisfied.
First we show (I): Given a PN and an execution sequence X of PN, we create a set A as a union
of the following sets:
1. A1 = {num(n) : 0 ≤ n ≤ ntok}
2. A2 = {time(ts) : 0 ≤ ts ≤ k}
3. A3 = {place(p) : p ∈ P}
4. A4 = {trans(t) : t ∈ T }
5. A5 = {ptarc(p, t, n) : (p, t) ∈ E − , n = W(p, t)}, where E − ⊆ E
6. A6 = {tparc(t, p, n) : (t, p) ∈ E + , n = W(t, p)}, where E + ⊆ E
7. A7 = {holds(p, q, 0) : p ∈ P, q = M0 (p)}
8. A8 = {notenabled(t, ts) : t ∈ T, 0 ≤ ts ≤ k, ∃p ∈ •t, Mts (p) < W(p, t)}
per definition 12 (enabled transition)
9. A9 = {enabled(t, ts) : t ∈ T, 0 ≤ ts ≤ k, ∀p ∈ •t, W(p, t) ≤ Mts (p)}
per definition 12 (enabled transition)
10. A10 = { f ires(t, ts) : t ∈ T ts , 0 ≤ ts ≤ k}
per definition 15 (firing set), only an enabled transition may fire
11. A11 = {add(p, q, t, ts) : t ∈ T ts , p ∈ t•, q = W(t, p), 0 ≤ ts ≤ k}
per definition 13 (transition execution)
12. A12 = {del(p, q, t, ts) : t ∈ T ts , p ∈ •t, q = W(p, t), 0 ≤ ts ≤ k}
per definition 13 (transition execution)
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
193
P
13. A13 = {tot incr(p, q, ts) : p ∈ P, q = t∈Tts ,p∈t• W(t, p), 0 ≤ ts ≤ k}
per definition 15 (firing set execution)
P
14. A14 = {tot decr(p, q, ts) : p ∈ P, q = t∈Tts ,p∈•t W(p, t), 0 ≤ ts ≤ k}
per definition 15 (firing set execution)
P
15. A15 = {consumesmore(p, ts) : p ∈ P, q = Mts (p), q1 = t∈Tts ,p∈•t W(p, t), q1 > q, 0 ≤
ts ≤ k}
per definition 14 (conflicting transitions) for enabled transition set T ts
P
16. A16 = {consumesmore : ∃p ∈ P : q = Mts (p), q1 = t∈Tts ,p∈•t W(p, t), q1 > q, 0 ≤ ts ≤ k}
per definition 14 (conflicting transitions)
17. A17 = {could not have(t, ts) : t ∈ T, (∀p ∈ •t, W(p, t) ≤ Mts (p)), t < T ts , (∃p ∈ •t :
P
W(p, t) > Mts (p) − t0 ∈Tts ,p∈•t0 W(p, t0 )), 0 ≤ ts ≤ k}
per the maximal firing set semantics
18. A18 = {holds(p, q, ts + 1) : p ∈ P, q = Mts+1 (p), 0 ≤ ts < k},
P
P
where Mts+1 (p) = Mts (p) − t∈Tts ,p∈•t W(p, t) + t∈Tts ,p∈t• W(t, p)
according to definition 16 (firing set execution)
We show that A satisfies (A.3) and (A.4), and A is an answer set of Π1 .
A satisfies (A.3) and (A.4) by its construction above. We show A is an answer set of Π1 by
splitting. We split lit(Π1 ) into a sequence of 6k + 8 sets:
• U0 = head( f 1) ∪ head( f 2) ∪ head( f 3) ∪ head( f 4) ∪ head( f 5) ∪ head( f 6) ∪head(i1) =
{place(p) : p ∈ P} ∪ {trans(t) : t ∈ T } ∪ {ptarc(p, t, n) : (p, t) ∈ E − , n = W(p, t)} ∪
{tparc(t, p, n) : (t, p) ∈ E + , n = W(t, p)}∪ {time(0), . . . , time(k)}∪{num(0), . . . , num(ntok)}∪
{holds(p, q, 0) : p ∈ P, q = M0 (p)}
• U6k+1 = U6k+0 ∪ head(e1)ts=k = U6k+0 ∪ {notenabled(t, k) : t ∈ T }
• U6k+2 = U6k+1 ∪ head(e2)ts=k = U6k+1 ∪ {enabled(t, k) : t ∈ T }
• U6k+3 = U6k+2 ∪ head(a1)ts=k = U6k+2 ∪ { f ires(t, k) : t ∈ T }
• U6k+4 = U6k+3 ∪ head(r1)ts=k ∪ head(r2)ts=k = U6k+3 ∪ {add(p, q, t, k) : p ∈ P, t ∈ T, q =
W(t, p)} ∪ {del(p, q, t, k) : p ∈ P, t ∈ T, q = W(p, t)}
• U6k+5 = U6k+4 ∪ head(r3)ts=k ∪ head(r4)ts=k = U6k+4 ∪ {tot incr(p, q, k) : p ∈ P, 0 ≤ q ≤
ntok} ∪ {tot decr(p, q, k) : p ∈ P, 0 ≤ q ≤ ntok}
• U6k+6 = U6k+5 ∪ head(r5)ts=k ∪ head(a2)ts=k ∪ head(a5)ts=k = U6k+5
∪ {consumesmore(p, k) : p ∈ P} ∪ {holds(p, q, k + 1) : p ∈ P, 0 ≤ q ≤ ntok}∪
{could not have(t, k) : t ∈ T }
• U6k+7 = U6k+6 ∪ head(a3) = U6k+6 ∪ {consumesmore}
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
194
where head(ri )ts=k are head atoms of ground rule ri in which ts = k. We write Ats=k
=
i
{a(. . . , ts) : a(. . . , ts) ∈ Ai , ts = k} as short hand for all atoms in Ai with ts = k. Uα , 0 ≤
α ≤ 6k + 7 form a splitting sequence, since each Ui is a splitting set of Π1 , and hUα iα<µ is a
S
monotone continuous sequence, where U0 ⊆ U1 · · · ⊆ U6k+7 and α<µ Uα = lit(Π1 ).
We compute the answer set of Π1 using the splitting sets as follows:
1. botU0 (Π1 ) = f 1 ∪ f 2 ∪ f 3 ∪ f 4 ∪ f 5 ∪ i1 ∪ f 6 and X0 = A1 ∪ · · · ∪ A7 (= U0 ) is its answer
set – using forced atom proposition
2. evalU0 (botU1 (Π1 ) \ botU0 (Π1 ), X0 ) = {notenabled(t, 0):-.|{trans(t),
ptarc(p, t, n), holds(p, q, 0)} ⊆ X0 , where q < n}. Its answer set X1 = Ats=0
– using forced
8
atom proposition and construction of A8 .
(a) where, q = M0 (p), and n = W(p, t) for an arc (p, t) ∈ E − – by construction of i1 and
f 3 in Π1 , and
(b) in an arc (p, t) ∈ E − , p ∈ •t (by definition 11 of preset)
(c) thus, notenabled(t, 0) ∈ X1 represents ∃p ∈ •t : M0 (p) < W(p, t).
3. evalU1 (botU2 (Π1 )\botU1 (Π1 ), X0 ∪X1 ) = {enabled(t, 0):-.|trans(t) ∈ X0 ∪X1 , notenabled(t, 0) <
– using forced atom proposition and construction
X0 ∪ X1 }. Its answer set is X2 = Ats=0
9
of A9 .
(a) since an enabled(t, 0) ∈ X2 if @ notenabled(t, 0) ∈ X0 ∪ X1 , which is equivalent to
@p ∈ •t : M0 (p) < W(p, t) ≡ ∀p ∈ •t : M0 (p) ≥ W(p, t).
4. evalU2 (botU3 (Π1 ) \ botU2 (Π1 ), X0 ∪ X1 ∪ X2 ) = {{ f ires(t, 0)}:-.|enabled(t, 0)
holds in X0 ∪ X1 ∪ X2 }. It has multiple answer sets X3.1 , . . . , X3.n , corresponding to elements of power set of f ires(t, 0) atoms in evalU2 (...) – using supported rule proposition.
Since we are showing that the union of answer sets of Π1 determined using splitting is
equal to A, we only consider the set that matches the f ires(t, 0) elements in A and call it
X3 , ignoring the rest. Thus, X3 = Ats=0
10 , representing T 0 .
5. evalU3 (botU4 (Π1 ) \ botU3 (Π1 ), X0 ∪ · · · ∪ X3 ) = {add(p, n, t, 0):-.|{ f ires(t, 0),
tparc(t, p, n)} ⊆ X0 ∪· · ·∪X3 }∪{del(p, n, t, 0):-.|{ f ires(t, 0), ptarc(p, t, n)} ⊆ X0 ∪· · ·∪X3 }.
ts=0
It’s answer set is X4 = Ats=0
11 ∪ A12 – using forced atom proposition and definitions of
A11 and A12 .
(a) where, each add atom encodes n = W(t, p) : p ∈ t•,
(b) and each del atom encodes n = W(p, t) : p ∈ •t
(c) representing the effect of transitions in T 0 – by construction
6. evalU4 (botU5 (Π1 ) \ botU4 (Π1 ), X0 ∪ · · · ∪ X4 ) = {tot incr(p, qq, 0):-.|
P
P
qq = add(p,q,t,0)∈X0 ∪···∪X4 q} ∪ {tot decr(p, qq, 0):-.|qq = del(p,q,t,0)∈X0 ∪···∪X4 q}. It’s ants=0
swer set is X5 = Ats=0
13 ∪ A14 – using forced atom proposition and definitions of A13 , A14 ,
and semantics of aggregate assignment atom
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
195
P
(a) where, each tot incr(p, qq, 0), qq = add(p,q,t,0)∈X0 ∪...X4 q
P
≡ qq = t∈X3 ,p∈t• W(p, t),
P
(b) and, each tot decr(p, qq, 0), qq = del(p,q,t,0)∈X0 ∪...X4 q
P
≡ qq = t∈X3 ,p∈•t W(t, p),
(c) represent the net effect of actions in T 0 – by construction
7. evalU5 (botU6 (Π1 ) \ botU5 (Π1 ), X0 ∪ · · · ∪ X5 ) = {consumesmore(p, 0):-.|
{holds(p, q, 0), tot decr(p, q1, 0)} ⊆ X0 ∪ · · · ∪ X5 : q1 > q} ∪ {holds(p, q, 1):-., |
{holds(p, q1, 0), tot incr(p, q2, 0), tot decr(p, q3, 0)} ⊆ X0 ∪ · · · ∪ X5 , q = q1 + q2 − q3} ∪
{could not have(t, 0):-.|{enabled(t, 0), ptarc(s, t, q), holds(s, qq, 0),
tot decr(s, qqq, 0)} ⊆ X0 ∪ · · · ∪ X5 , f ires(t, 0) < (X0 ∪ · · · ∪ X5 ), q > qq − qqq}. It’s
ts=0
answer set is X6 = Ats=0
∪ Ats=0
17 ∪ A18 – using forced atom proposition and definitions
15
of A15 , A17 , A18 .
P
(a) where, consumesmore(p, 0) represents ∃p : q = M0 (p), q1 = t∈T0 ,p∈•t W(p, t),
q1 > q indicating place p will be overconsumed if T 0 is fired, as defined in definition 14 (conflicting transitions)
(b) and, holds(p, q, 1) encodes q = M1 (p) – by construction
(c) and could not have(t, 0) represents an enabled transition t in T 0 that could not fire
due to insufficient tokens
(d) X6 does not contain could not have(t, 0), when enabled(t, 0) ∈ X0 ∪ · · · ∪ X5 and
f ires(t, 0) < X0 ∪ · · · ∪ X5 due to construction of A, encoding of a5 and its body
atoms. As a result it is not eliminated by the constraint a6
..
.
8. evalU6k+0 (botU6k+1 (Π1 ) \ botU6k+0 (Π1 ), X0 ∪ · · · ∪ X6k+0 ) =
{notenabled(t, k):-.|{trans(t), ptarc(p, t, n), holds(p, q, k)} ⊆ X0 ∪ · · · ∪ X6k+0 ,
where q < n}. Its answer set X6k+1 = Ats=k
– using forced atom proposition and construc8
tion of A8 .
(a) where, q = Mk (p), and n = W(p, t) for an arc (p, t) ∈ E − – by construction of holds
and ptarc predicates in Π1 , and
(b) in an arc (p, t) ∈ E − , p ∈ •t (by definition 11 of preset)
(c) thus, notenabled(t, k) ∈ X6k+1 represents ∃p ∈ •t : Mk (p) < W(p, t).
9. evalU6k+1 (botU6k+2 (Π1 ) \ botU6k+1 (Π1 ), X0 ∪ · · · ∪ X6k+1 ) = {enabled(t, k):-.|
trans(t) ∈ X0 ∪ · · · ∪ X6k+1 , notenabled(t, k) < X0 ∪ · · · ∪ X6k+1 }. Its answer set is X6k+2 =
Ats=k
– using forced atom proposition and construction of A9 .
9
(a) since an enabled(t, k) ∈ X6k+2 if @ notenabled(t, k) ∈ X0 ∪ · · · ∪ X6k+1 , which is
equivalent to @p ∈ •t : Mk (p) < W(p, t) ≡ ∀p ∈ •t : Mk (p) ≥ W(p, t).
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
196
10. evalU6k+2 (botU6k+3 (Π1 ) \ botU6k+2 (Π1 ), X0 ∪ · · · ∪ X6k+2 ) =
{{ f ires(t, k)}:-.|enabled(t, k) holds in X0 ∪· · ·∪X6k+2 }. It has multiple answer sets X6k+3.1 , . . . , X6k+3.n ,
corresponding to elements of power set of f ires(t, k) atoms in evalU6k+2 (...) – using supported rule proposition. Since we are showing that the union of answer sets of Π1 determined using splitting is equal to A, we only consider the set that matches the f ires(t, k)
elements in A and call it X6k+3 , ignoring the reset. Thus, X6k+3 = Ats=k
10 , representing T k .
11. evalU6k+3 (botU6k+4 (Π1 ) \ botU6k+3 (Π1 ), X0 ∪ · · · ∪ X6k+3 ) =
{add(p, n, t, k):-.|{ f ires(t, k), tparc(t, p, n)} ⊆ X0 ∪ · · · ∪ X6k+3 } ∪
{del(p, n, t, k):-.|{ f ires(t, k), ptarc(p, t, n)} ⊆ X0 ∪ · · · ∪ X6k+3 }. It’s answer set is X6k+4 =
ts=k
Ats=k
11 ∪ A12 – using forced atom proposition and definitions of A11 and A12 .
(a) where, each add atom is equivalent to n = W(t, p) : p ∈ t•,
(b) and, each del atom is equivalent to n = W(p, t) : p ∈ •t,
(c) representing the effect of transitions in T k
12. evalU6k+4 (botU6k+5 (Π1 ) \ botU6k+4 (Π1 ), X0 ∪ · · · ∪ X6k+4 ) = {tot incr(p, qq, k):-.|
P
qq = add(p,q,t,k)∈X0 ∪···∪X6k+4 q} ∪ {tot decr(p, qq, k):-.|
P
ts=k
qq = del(p,q,t,k)∈X0 ∪···∪X6k+4 q}. It’s answer set is X6k+5 = Ats=k
13 ∪ A14 – using forced atom
proposition and definitions of A13 and A14 .
P
(a) where, each tot incr(p, qq, k), qq = add(p,q,t,k)∈X0 ∪...X6k+4 q ≡
P
qq = t∈X6k+3 ,p∈t• W(p, t),
P
(b) and, each tot decr(p, qq, k), qq = del(p,q,t,k)∈X0 ∪...X6k+4 q ≡
P
qq = t∈X6k+3 ,p∈•t W(t, p),
(c) represent the net effect of transitions in T k
13. evalU6k+5 (botU6k+6 (Π1 ) \ botU6k+5 (Π1 ), X0 ∪ · · · ∪ X6k+5 ) =
{consumesmore(p, k):-.|{holds(p, q, k), tot decr(p, q1, k)} ⊆ X0 ∪ · · · ∪ X6k+5 : q1 > q} ∪
{holds(p, q, k + 1):-., |{holds(p, q1, k), tot incr(p, q2, k),
tot decr(p, q3, k)} ⊆ X0 ∪ · · · ∪ X6k+5 : q = q1 + q2 − q3} ∪ {could not have(t, k):{enabled(t, k), ptarc(s, t, q), holds(s, qq, k), tot decr(s, qqq, k)} ⊆ X0 ∪· · ·∪X6k+5 , f ires(t, k) <
ts=k
(X0 ∪ · · · ∪ X6k+5 ), q > qq − qqq}. It’s answer set is X6k+6 = Ats=k
∪ Ats=k
17 ∪ A18 – using
15
forced atom proposition.
(a) where, consumesmore(p, k) represents ∃p : q = Mk (p),
P
q1 = t∈Tk ,p∈•t W(p, t), q1 > q
(b) holds(p, q, k + 1) represents q = Mk+1 (p) indicating place p that will be over consumed if T k is fired, as defined in definition 14 (conflicting transitions),
(c) holds(p, q, k + 1) represents q = Mk+1 (p) – by construction
(d) and could not have(t, k) represents an enabled transition t in T k that could not fire
due to insufficient tokens
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
197
(e) X6k+6 does not contain could not have(t, k), when enabled(t, k) ∈ X0 ∪ · · · ∪ X6k+5
and f ires(t, k) < X0 ∪ · · · ∪ X6k+5 due to construction of A, encoding of a5 and its
body atoms. As a result it is note eliminated by the constraint a6
14. evalU6k+6 (botU6k+7 (Π1 ) \ botU6k+6 (Π1 ), X0 ∪ · · · ∪ X6k+6 ) = {consumesmore:-.|
{consumesmore(p, 0), . . . , consumesmore(p, k)}∩ (X0 ∪ · · · ∪ X6k+6 ) , ∅}. It’s answer set
is X6k+7 = A16 – using forced atom proposition
(a) X6k+7 will be empty since none of consumesmore(p, 0), . . . ,
consumesmore(p, k) hold in X0 ∪ · · · ∪ X6k+6 due to the construction of A, encoding
of a2 and its body atoms. As a result, it is not eliminated by the constraint a4
The set X = X0 ∪ · · · ∪ X6k+7 is the answer set of Π0 by the splitting sequence theorem 15.
Each Xi , 0 ≤ i ≤ 6k + 7 matches a distinct portion of A, and X = A, thus A is an answer set of
Π1 .
Next we show (II): Given Π1 be the encoding of a Petri Net PN(P, T, E, W) with initial marking
M0 , and A be an answer set of Π1 that satisfies (A.3) and (A.4), then we can construct X =
M0 , T 0 , . . . , Mk , T k , Mk+1 from A, such that it is an execution sequence of PN.
We construct the X as follows:
1. Mi = (Mi (p0 ), . . . , Mi (pn )), where {holds(p0 , Mi (p0 ), i), . . . holds(pn , Mi (pn ), i)}
⊆ A, for 0 ≤ i ≤ k + 1
2. T i = {t : f ires(t, i) ∈ A}, for 0 ≤ i ≤ k
and show that X is indeed an execution sequence of PN. We show this by induction over k (i.e.
given Mk , T k is a valid firing set and its firing produces marking Mk+1 ).
Base case: Let k = 0, and M0 is a valid marking in X for PN, show (1) T 0 is a valid firing set
for M0 , and (2) T 0 ’s firing w.r.t. marking M0 produces M1 .
1. We show T 0 is a valid firing set for M0 . Let { f ires(t0 , 0), . . . , f ires(t x , 0)} be the set of all
f ires(. . . , 0) atoms in A,
(a) Then for each f ires(ti , 0) ∈ A
i.
ii.
iii.
iv.
enabled(ti , 0) ∈ A – from rule a1 and supported rule proposition
Then notenabled(ti , 0) < A – from rule e2 and supported rule proposition
Then body(e1) must not hold in A – from rule e1 and forced atom proposition
Then q ≮ ni ≡ q ≥ ni in e1 for all {holds(p, q, 0), ptarc(p, ti , ni )} ⊆ A – from e1,
forced atom proposition, and the following
A. holds(p, q, 0) ∈ A represents q = M0 (p) – rule i1 construction
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
198
B. ptarc(p, ti , ni ) ∈ A represents ni = W(p, ti ) – rule f 3 construction
v. Then ∀p ∈ •ti , M0 (p) > W(p, ti ) – from definition 11 of preset •ti in PN
vi. Then ti is enabled and can fire in PN, as a result it can belong to T 0 – from
definition 12 of enabled transition
(b) And consumesmore < A, since A is an answer set of Π1 – from rule a4 and supported
rule proposition
i. Then @consumesmore(p, 0) ∈ A – from rule a3 and supported rule proposition
ii. Then @{holds(p, q, 0), tot decr(p, q1, 0)} ⊆ A : q1 > q in body(a2) – from a2
and forced atom proposition
P
iii. Then @p : ti ∈{t0 ,...,tx },p∈•ti W(p, ti ) > M0 (p) – from the following
A. holds(p, q, 0) represents q = M0 (p) – from rule i1 construction, given
B. tot decr(p, q1, 0) ∈ A if {del(p, q10 , t0 , 0), . . . , del(p, q1 x , t x , 0)} ⊆ A, where
q1 = q10 + · · · + q1 x – from r4 and forced atom proposition
C. del(p, q1i , ti , 0) ∈ A if { f ires(ti , 0), ptarc(p, ti , q1i )} ⊆ A – from r2 and
supported rule proposition
D. del(p, q1i , ti , 0) represents removal of q1i = W(p, ti ) tokens from p ∈ •ti
– from rule r2, supported rule proposition, and definition 13 of transition
execution in PN
iv. Then the set of transitions in T 0 do not conflict – by the definition 14 of conflicting transitions
(c) And for each enabled(t j , 0) ∈ A and f ires(t j , 0) < A, could not have(t j , 0) ∈ A,
since A is an answer set of Π1 - from rule a6 and supported rule proposition
i. Then {enabled(t j , 0), holds(s, qq, 0), ptarc(s, t j , q, 0),
tot decr(s, qqq, 0)} ⊆ A, such that q > qq − qqq and f ires(t j , 0) < A - from rule
a5 and supported rule proposition
P
ii. Then for an s ∈ •t j , W(s, t j ) > M0 (s) − ti ∈T0 ,s∈•ti W(s, ti ) - from the following:
A. ptarc(s, ti , q) represents q = W(s, ti ) – from rule f 7 construction
B. holds(s, qq, 0) represents qq = M0 (s) – from i1 construction
C. tot decr(s, qqq, 0) ∈ A if {del(s, qqq0 , t0 , 0), . . . , del(s, qqq x , t x , 0)} ⊆ A –
from rule r4 construction and supported rule proposition
D. del(s, qqqi , ti , 0) ∈ A if { f ires(ti , 0), ptarc(s, ti , qqqi )} ⊆ A – from rule r7
and supported rule proposition
E. del(s, qqqi , ti , 0) represents qqqi = W(s, ti ) : ti ∈ T 0 , (s, ti ) ∈ E − – from rule
f 7 construction
P
F. tot decr(q, qqq, 0) represents ti ∈T0 ,s∈•ti W(s, ti ) – from (C,D,E) above
iii. Then firing T 0 ∪ {t j } would have required more tokens than are present at its
source place s ∈ •t j . Thus, T 0 is a maximal set of transitions that can simultaneously fire.
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
199
(d) Then {t0 , . . . , t x } = T 0 – using 1(a),1(b) above; and using 1(c) it is a maximal firing
set
2. We show M1 is produced by firing T 0 in M0 . Let holds(p, q, 1) ∈ A
(a) Then {holds(p, q1, 0), tot incr(p, q2, 0), tot decr(p, q3, 0)} ⊆ A : q = q1 + q2 − q3 –
from rule r5 and supported rule proposition
(b) Then, holds(p, q1, 0) ∈ A represents q1 = M0 (p) – given, rule i1 construction;
and {add(p, q20 , t0 , 0), . . . , add(p, q2 j , t j , 0)} ⊆ A : q20 + · · · + q2 j = q2 and
{del(p, q30 , t0 , 0), . . . , del(p, q3l , tl , 0)} ⊆ A : q30 + · · · + q3l = q3 – rules r3, r4
using supported rule proposition
(c) Then { f ires(t0 , 0), . . . , f ires(t j , 0)} ⊆ A and { f ires(t0 , 0), . . . , f ires(tl , 0)}
⊆ A – rules r1, r2 and supported rule proposition, respectively
(d) Then { f ires(t0 , 0), . . . , f ires(t j , 0)}∪{ f ires(t0 , 0), . . . , f ires(tl , 0)} ⊆ A = { f ires(t0 , 0), . . . ,
f ires(t x , 0)} ⊆ A – set union of subsets
(e) Then for each f ires(t x , 0) ∈ A we have t x ∈ T 0 – already shown in item 1 above
P
P
(f) Then q = M0 (p)+ tx ∈T0 ∧p∈tx • W(t x , p)− tx ∈T0 ∧p∈•tx W(p, t x ) – from (2b) above and
the following
i. Each add(p, q j , t j , 0) ∈ A represents q j = W(t j , p) for p ∈ t j • – rule r1 encoding, and definition 13 of transition execution in PN
ii. Each del(p, ty , qy , 0) ∈ A represents qy = W(p, ty ) for p ∈ •ty – from rule r2
encoding, and definition 13 of transition execution in PN
P
iii. Each tot incr(p, q2, 0) ∈ A represents q2 = tx ∈T0 ∧p∈tx • W(t x , p) – aggregate
assignment atom semantics in rule r3
P
iv. Each tot decr(p, q3, 0) ∈ A represents q3 = tx ∈T0 ∧p∈•tx W(p, t x ) – aggregate
assignment atom semantics in rule r4
(g) Then, M1 (p) = q – since holds(p, q, 1) ∈ A encodes q = M1 (p) from construction
Inductive Step: Let k > 0, and Mk is a valid marking in X for PN, show (1) T k is a valid firing
set for Mk , and (2) T k ’s firing in Mk produces marking Mk+1 .
1. We show T k is a valid firing set. Let { f ires(t0 , k), . . . , f ires(t x , k)} be the set of all f ires(. . . , k)
atoms in A,
(a) Then for each f ires(ti , k) ∈ A
i.
ii.
iii.
iv.
enabled(ti , k) ∈ A – from rule a1 and supported rule proposition
Then notenabled(ti , k) < A – from rule e2 and supported rule proposition
Then body(e1) must hold in A – from rule e1 and forced atom proposition
Then q ≮ ni ≡ q ≥ ni in e1 for all {holds(p, q, k), ptarc(p, ti , n)} ⊆ A – from e1,
forced atom proposition, and the following
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
200
A. holds(p, q, k) ∈ A represents q = Mk (p) – construction, inductive assumption
B. ptarc(p, t, ni ) ∈ A represents ni = W(p, ti ) – rule f 3 construction
v. Then ∀p ∈ •ti , Mk (p) ≥ W(p, ti ) – from definition 11 of preset •ti in PN
vi. Then ti is enabled and can fire in PN, as a result it can belong to T k – from
definition 12 of enabled transition
(b) And consumesmore < A, since A is an answer set of Π1 – from rule a4 and supported
rule proposition
i. Then @consumesmore(p, k) ∈ A – from rule a3 and forced atom proposition
ii. Then @{holds(p, q, k), tot decr(p, q1, k)} ⊆ A : q1 > q in body(a2) – from a2
and forced atom proposition
P
iii. Then @p : ti ∈{t0 ,...,tx },p∈•ti W(p, ti ) > Mk (p) – from the following
A. holds(p, q, k) ∈ A represents q = Mk (p) – by construction of Π1 , and the
inductive assumption about Mk (p)
B. tot decr(p, q1, k) ∈ A if {del(p, q10 , t0 , k), . . . , del(p, q1 x , t x , k)} ⊆ A, where
q1 = q10 + · · · + q1 x – from r4 and forced atom proposition
C. del(p, q1i , ti , k) ∈ A if { f ires(ti , k), ptarc(p, ti , q1i )} ⊆ A – from r2 and
supported rule proposition
D. del(p, q1i , ti , k) represents removal of q1i = W(p, ti ) tokens from p ∈ •ti –
from construction rule r2, supported rule proposition, and definition 13 of
transition execution in PN
iv. Then the set of transitions T k do not conf – by the definition 14 of conflicting
transitions
(c) And for each enabled(t j , k) ∈ A and f ires(t j , k) < A, could not have(t j , k) ∈ A,
since A is an answer set of Π1 - from rule a6 and supported rule proposition
i. Then {enabled(t j , k), holds(s, qq, k), ptarc(s, t j , q, k),
tot decr(s, qqq, k)} ⊆ A, such that q > qq − qqq and f ires(t j , 0) < A - from rule
a5 and supported rule proposition
P
ii. Then for an s ∈ •t j , W(s, t j ) > Mk (s) − ti ∈Tk ,s∈•ti W(s, ti ) - from the following:
A. ptarc(s, ti , q) represents q = W(s, ti ) – from rule f 7 construction
B. holds(s, qq, k) represents qq = Mk (s) – from i1 construction
C. tot decr(s, qqq, k) ∈ A if {del(s, qqq0 , t0 , k), . . . , del(s, qqq x , t x , k)} ⊆ A –
from rule r4 construction and supported rule proposition
D. del(s, qqqi , ti , k) ∈ A if { f ires(ti , k), ptarc(s, ti , qqqi )} ⊆ A – from rule r7
and supported rule proposition
E. del(s, qqqi , ti , k) represents qqqi = W(s, ti ) : ti ∈ T k , (s, ti ) ∈ E − – from rule
f 7 construction
P
F. tot decr(q, qqq, k) represents ti ∈T k ,s∈•ti W(s, ti ) – from (C,D,E) above
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
201
iii. Then firing T k ∪ {t j } would have required more tokens than are present at its
source place s ∈ •t j . Thus, T k is a maximal set of transitions that can simultaneously fire.
(d) Then {t0 , . . . , t x } = T k – using 1(a),1(b) above; and using 1(c) it is a maximal firing
set
2. We show Mk+1 is produced by firing T k in Mk . Let holds(p, q, k + 1) ∈ A
(a) Then {holds(p, q1, k), tot incr(p, q2, k), tot decr(p, q3, k)} ∈ A : q = q2 + q2 − q3 –
from rule r5 and supported rule proposition
(b) holds(p, q1, k) ∈ A represents q1 = Mk (p) – construction, inductive assumption;
and {add(p, q20 , t0 , k), . . . , add(p, q2 j , t j , k)} ⊆ A : q20 + · · · + q2 j = q2 and
{del(p, q30 , t0 , k), . . . , del(p, q3l , tl , k)} ⊆ A : q30 + · · · + q3l = q3 – from rules
r3, r4 using supported rule proposition, respectively
(c) Then { f ires(t0 , k), . . . , f ires(t j , k)} ⊆ A and { f ires(t0 , k), . . . , f ires(tl , k)}
⊆ A – from rules r1, r2 using supported rule proposition, respectively
(d) Then { f ires(t0 , k), . . . , f ires(t j , k)}∪{ f ires(t0 , k), . . . , f ires(tl , k)} = { f ires(t0 , k), . . . ,
f ires(t x , k)} ⊆ A – subset union property
(e) Then for each f ires(t x , k) ∈ A we have t x ∈ T k - already shown in item (1a) above
P
P
(f) Then q = Mk (p) + tx ∈Tk ∧p∈tx • W(t x , p) − tx ∈T k ∧p∈•tx W(p, t x ) – from (2b) above and
the following
i. Each add(p, q j , t j , k) ∈ A represents q j = W(t j , p) for p ∈ t j • – rule r1 encoding
and definition 13 of transition execution in PN
ii. Each del(p, ty , qy , k) ∈ A represents qy = W(p, ty ) for p ∈ •ty – rule r2 encoding
and definition 13 of transition execution in PN
P
iii. Each tot incr(p, q2, k) ∈ A represents q2 = tx ∈Tk ∧p∈tx • W(t x , p) – aggregate
assignment atom semantics in rule r3
P
iv. Each tot decr(p, q3, k) ∈ A represents q3 = tx ∈Tk ∧p∈•tx W(p, t x ) – aggregate
assignment atom semantics in rule r4
(g) Then Mk+1 (p) = q – since holds(p, q, k + 1) ∈ A encodes q = Mk+1 (p) – from
construction
As a result, for any n > k, T n will be a valid firing set w.r.t. Mn and its firing produces marking
Mn+1 .
Conclusion: Since both I and II hold, X = M0 , T 0 , M1 , . . . , Mk , T k , Mk+1 is an execution sequence of PN(P, T, E, W) (w.r.t. M0 ) iff there is an answer set A of Π1 (PN, M0 , k, ntok) such
that (A.3) and (A.4) hold.
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
A.3
202
Proof of Proposition 3
Let PN = (P, T, E, W, R) be a Petri Net, M0 be its initial marking and let Π2 (PN, M0 , k, ntok) by
the ASP encoding of PN and M0 over a simulation length k, with maximum ntok tokens on any
place node, as defined in section 2.5. Then X = M0 , T 0 , M1 , . . . , Mk , T k , Mk+1 is an execution
sequence of PN (w.r.t M0 ) iff there is an answer set A of Π2 (PN, M0 , k, ntok) such that:
{ f ires(t, ts) : t ∈ T ts , 0 ≤ ts ≤ k} = { f ires(t, ts) : f ires(t, ts) ∈ A}
{holds(p, q, ts) : p ∈ P, q = Mts (p), 0 ≤ ts ≤ k + 1}
= {holds(p, q, ts) : holds(p, q, ts) ∈ A}
(A.5)
(A.6)
We prove this by showing that:
(I) Given an execution sequence X, we create a set A such that it satisfies (A.5) and (A.6) and
show that A is an answer set of Π2
(II) Given an answer set A of Π2 , we create an execution sequence X such that (A.5) and (A.6)
are satisfied.
First we show (I): Given PN and an execution sequence X of PN, we create a set A as a union
of the following sets:
1. A1 = {num(n) : 0 ≤ n ≤ ntok}
2. A2 = {time(ts) : 0 ≤ ts ≤ k}
3. A3 = {place(p) : p ∈ P}
4. A4 = {trans(t) : t ∈ T }
5. A5 = {ptarc(p, t, n, ts) : (p, t) ∈ E − , n = W(p, t), 0 ≤ ts ≤ k}, where E − ⊆ E
6. A6 = {tparc(t, p, n, ts) : (t, p) ∈ E + , n = W(t, p), 0 ≤ ts ≤ k}, where E + ⊆ E
7. A7 = {holds(p, q, 0) : p ∈ P, q = M0 (p)}
8. A8 = {notenabled(t, ts) : t ∈ T, 0 ≤ ts ≤ k, ∃p ∈ •t, Mts (p) < W(p, t)}
per definition 12 (enabled transition)
9. A9 = {enabled(t, ts) : t ∈ T, 0 ≤ ts ≤ k, ∀p ∈ •t, W(p, t) ≤ Mts (p)}
per definition 12 (enabled transition)
10. A10 = { f ires(t, ts) : t ∈ T ts , 0 ≤ ts ≤ k}
per definition 21 (firing set), only an enabled transition may fire
11. A11 = {add(p, q, t, ts) : t ∈ T ts , p ∈ t•, q = W(t, p), 0 ≤ ts ≤ k}
per definition 22 (transition execution)
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
203
12. A12 = {del(p, q, t, ts) : t ∈ T ts , p ∈ •t, q = W(p, t), 0 ≤ ts ≤ k} ∪ {del(p, q, t, ts) : t ∈
T ts , p ∈ R(t), q = Mts (p), 0 ≤ ts ≤ k}
per definition 22 (transition execution)
P
13. A13 = {tot incr(p, q, ts) : p ∈ P, q = t∈Tts ,p∈t• W(t, p), 0 ≤ ts ≤ k}
per definition 24 (execution)
P
P
14. A14 = {tot decr(p, q, ts) : p ∈ P, q = t∈Tts ,p∈•t W(p, t) + t∈Tts ,p∈R(t) Mts (p),
0 ≤ ts ≤ k}
per definition 24 (execution)
P
P
15. A15 = {consumesmore(p, ts) : p ∈ P, q = Mts (p), q1 = t∈Tts ,p∈•t W(p, t)+ t∈T ts ,p∈R(t) Mts (p), q1 >
q, 0 ≤ ts ≤ k}
per definition 23 (conflicting transitions)
P
P
16. A16 = {consumesmore : ∃p ∈ P : q = Mts (p), q1 = t∈T ts ,p∈•t W(p, t)+ t∈Tts ,p∈R(t) Mts (p), q1 >
q, 0 ≤ ts ≤ k}
per definition 23 (conflicting transitions)
17. A17 = {could not have(t, ts) : t ∈ T, (∀p ∈ •t, W(p, t) ≤ Mts (p)), t < T ts , (∃p ∈
P
P
•t ∪ R(t) : q > Mts (p) − ( t0 ∈Tts ,p∈•t0 W(p, t0 ) + t0 ∈Tts ,p∈R(t0 ) Mts (p)), q = W(p, t) if p ∈
•t or Mts (p) otherwise ), 0 ≤ ts ≤ k}
per the maximal firing set semantics
18. A18 = {holds(p, q, ts + 1) : p ∈ P, q = Mts+1 (p), 0 ≤ ts < k}, where Mts+1 (p) = Mts (p) −
P
P
P
( t∈Tts ,p∈•t W(p, t) + t∈Tts ,p∈R(t) Mts (p)) + t∈T ts ,p∈t• W(t, p)
according to definition 24 (firing set execution)
19. A19 = {ptarc(p, t, n, ts) : p ∈ R(t), n = Mts (p), n > 0, 0 ≤ ts ≤ k}
We show that A satisfies (A.5) and (A.6), and A is an answer set of Π1 .
A satisfies (A.5) and (A.6) by its construction above. We show A is an answer set of Π1 by
splitting. We split lit(Π1 ) into a sequence of 7(k + 1) + 2 sets:
• U0 = head( f 1) ∪ head( f 2) ∪ head( f 5) ∪ head( f 6) ∪ head(i1) = {place(p) : p ∈ P} ∪
{trans(t) : t ∈ T } ∪ {time(0), . . . , time(k)} ∪ {num(0), . . . , num(ntok)} ∪ {holds(p, q, 0) : p ∈
P, q = M0 (p)}
• U7k+1 = U7k+0 ∪ head( f 7)ts=k ∪ head( f 8)ts=k ∪ head( f 9)ts=k = U7k+0 ∪
{ptarc(p, t, n, k) : (p, t) ∈ E − , n = W(p, t)} ∪ {tparc(t, p, n, k) : (t, p) ∈ E + , n = W(t, p)} ∪
{ptarc(p, t, n, k) : p ∈ R(t), n = Mk (p), n > 0}
• U7k+2 = U7k+1 ∪ head(e3)ts=k = U7k+1 ∪ {notenabled(t, k) : t ∈ T }
• U7k+3 = U7k+2 ∪ head(e2)ts=k = U7k+2 ∪ {enabled(t, k) : t ∈ T }
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
204
• U7k+4 = U7k+3 ∪ head(a1)ts=k = U7k+3 ∪ { f ires(t, k) : t ∈ T }
• U7k+5 = U7k+4 ∪ head(r6)ts=k ∪ head(r7)ts=k = U7k+4 ∪ {add(p, q, t, k) : p ∈ P, t ∈ T, q =
W(t, p)} ∪ {del(p, q, t, k) : p ∈ P, t ∈ T, q = W(p, t)} ∪ {del(p, q, t, k) : p ∈ P, t ∈ T, q =
Mk (p)}
• U7k+6 = U7k+5 ∪ head(r3)ts=k ∪ head(r4)ts=k = U7k+5 ∪ {tot incr(p, q, k) : p ∈ P, 0 ≤ q ≤
ntok} ∪ {tot decr(p, q, k) : p ∈ P, 0 ≤ q ≤ ntok}
• U7k+7 = U7k+6 ∪ head(r5)ts=k ∪ head(a2)ts=k ∪ head(a7)ts=k = U7k+6 ∪
{consumesmore(p, k) : p ∈ P}∪{holds(p, q, k+1) : p ∈ P, 0 ≤ q ≤ ntok}∪{could not have(t, k) :
t ∈ T}
• U7k+8 = U7k+7 ∪ head(a3) = U7k+7 ∪ {consumesmore}
where head(ri )ts=k are head atoms of ground rule ri in which ts = k. We write Ats=k
=
i
{a(. . . , ts) : a(. . . , ts) ∈ Ai , ts = k} as short hand for all atoms in Ai with ts = k. Uα , 0 ≤
α ≤ 7k + 8 form a splitting sequence, since each Ui is a splitting set of Π1 , and hUα iα<µ is a
S
monotone continuous sequence, where U0 ⊆ U1 · · · ⊆ U7k+8 and α<µ Uα = lit(Π1 ).
We compute the answer set of Π2 using the splitting sets as follows:
1. botU0 (Π2 ) = f 1 ∪ f 2 ∪ f 5 ∪ i1 ∪ f 6 and X0 = A1 ∪ · · · ∪ A4 ∪ A7 (= U0 ) is its answer set
– using forced atom proposition
2. evalU0 (botU1 (Π2 ) \ botU0 (Π2 ), X0 ) = {ptarc(p, t, q, 0):-.|q = W(p, t)} ∪
{tparc(t, p, q, 0):-.| q = W(t, p)} ∪ {ptarc(p, t, q, 0):-.| q = M0 (p)}. Its answer set X1 =
∪ Ats=0
∪ Ats=0
Ats=0
19 – using forced atom proposition and construction of A5 , A6 , A19 .
6
5
3. evalU1 (botU2 (Π2 ) \ botU1 (Π2 ), X0 ∪ X1 ) = {notenabled(t, 0):-.|{trans(t),
– using
ptarc(p, t, n, 0), holds(p, q, 0)} ⊆ X0 ∪ X1 , where q < n}. Its answer set X2 = Ats=0
8
forced atom proposition and construction of A8 .
(a) where, q = M0 (p), and n = W(p, t) for an arc (p, t) ∈ E − – by construction of i1 and
f 7 in Π2 , and
(b) in an arc (p, t) ∈ E − , p ∈ •t (by definition 11 of preset)
(c) thus, notenabled(t, 0) ∈ X1 represents ∃p ∈ •t : M0 (p) < W(p, t).
4. evalU2 (botU3 (Π2 ) \ botU2 (Π2 ), X0 ∪ · · · ∪ X2 ) = {enabled(t, 0):-.|trans(t) ∈ X0 ∪ · · · ∪
X2 , notenabled(t, 0) < X0 ∪ · · · ∪ X2 }. Its answer set is X3 = Ats=0
– using forced atom
9
proposition and construction of A9 .
(a) since an enabled(t, 0) ∈ X3 if @ notenabled(t, 0) ∈ X0 ∪ · · · ∪ X2 ; which is equivalent
to @p ∈ •t : M0 (p) < W(p, t) ≡ ∀p ∈ •t : M0 (p) ≥ W(p, t).
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
205
5. evalU3 (botU4 (Π2 ) \ botU3 (Π2 ), X0 ∪ · · · ∪ X3 ) = {{ f ires(t, 0)}:-.|enabled(t, 0)
holds in X0 ∪ · · · ∪ X3 }. It has multiple answer sets X4.1 , . . . , X4.n , corresponding to elements of power set of f ires(t, 0) atoms in evalU3 (...) – using supported rule proposition.
Since we are showing that the union of answer sets of Π2 determined using splitting is
equal to A, we only consider the set that matches the f ires(t, 0) elements in A and call it
X4 , ignoring the rest. Thus, X4 = Ats=0
10 , representing T 0 .
(a) in addition, for every t such that enabled(t, 0) ∈ X0 ∪ · · · ∪ X3 , R(t) , ∅; f ires(t, 0) ∈
X4 – per definition 21 (firing set); requiring that a reset transition is fired when
enabled
(b) thus, the firing set T 0 will not be eliminated by the constraint f 10
6. evalU4 (botU5 (Π2 ) \ botU4 (Π2 ), X0 ∪ · · · ∪ X4 ) = {add(p, n, t, 0):-.|{ f ires(t, 0),
tparc(t, p, n, 0)} ⊆ X0 ∪· · ·∪X4 }∪{del(p, n, t, 0):-.|{ f ires(t, 0), ptarc(p, t, n, 0)} ⊆ X0 ∪· · ·∪
ts=0
X4 }. It’s answer set is X5 = Ats=0
11 ∪ A12 – using forced atom proposition and definitions
of A11 and A12 .
(a) where, each add atom is equivalent to n = W(t, p) : p ∈ t•,
(b) and each del atom is equivalent to n = W(p, t) : p ∈ •t; or n = Mk (p) : p ∈ R(t),
(c) representing the effect of transitions in T 0 – by construction
7. evalU5 (botU6 (Π2 ) \ botU5 (Π2 ), X0 ∪ · · · ∪ X5 ) = {tot incr(p, qq, 0):-.|
P
P
qq = add(p,q,t,0)∈X0 ∪···∪X5 q} ∪ {tot decr(p, qq, 0):-.|qq = del(p,q,t,0)∈X0 ∪···∪X5 q}. It’s ants=0
swer set is X6 = Ats=0
13 ∪ A14 – using forced atom proposition, definitions of A13 , A14 ,
and definition 5 (semantics of aggregate assignment atom).
P
(a) where, each for tot incr(p, qq, 0), qq = add(p,q,t,0)∈X0 ∪...X5 q
P
≡ qq = t∈X4 ,p∈t• W(p, t),
P
(b) and each tot decr(p, qq, 0), qq = del(p,q,t,0)∈X0 ∪...X5 q
P
P
≡ qq = t∈X4 ,p∈•t W(t, p) + t∈X4 ,p∈R(t) Mk (p),
(c) represent the net effect of transitions in T 0 – by construction
8. evalU6 (botU7 (Π2 ) \ botU6 (Π2 ), X0 ∪ · · · ∪ X6 ) = {consumesmore(p, 0):-.|
{holds(p, q, 0), tot decr(p, q1, 0)} ⊆ X0 ∪ · · · ∪ X6 , q1 > q} ∪ {holds(p, q, 1):-., |
{holds(p, q1, 0), tot incr(p, q2, 0), tot decr(p, q3, 0)} ⊆ X0 ∪ · · · ∪ X6 , q = q1 + q2 − q3} ∪
{could not have(t, 0):-.|{enabled(t, 0), ptarc(s, t, q), holds(s, qq, 0),
tot decr(s, qqq, 0)} ⊆ X0 ∪ · · · ∪ X6 , f ires(t, 0) < (X0 ∪ · · · ∪ X6 ), q > qq − qqq}. It’s
ts=0
answer set is X7 = Ats=0
∪ Ats=0
17 ∪ A18 – using forced atom proposition and definitions
15
of A15 , A17 , A18 .
(a) where, consumesmore(p, 0) represents ∃p : q = M0 (p), q1 =
P
P
t∈T 0 ,p∈•t W(p, t)+
t∈T 0 ,p∈R(t) Mk (p), q1 > q, indicating place p will be overconsumed if T 0 is fired – as defined in definition 23 (conflicting transitions)
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
206
(b) and, holds(p, q, 1) represents q = M1 (p) – by construction of Π2
(c) and could not have(t, 0) represents an enabled transition t in T 0 that could not fire
due to insufficient tokens
(d) X7 does not contain could not have(t, 0), when enabled(t, 0) ∈ X0 ∪ · · · ∪ X6 and
f ires(t, 0) < X0 ∪ · · · ∪ X6 due to construction of A, encoding of a7 and its body
atoms. As a result it is not eliminated by the constraint a6
..
.
9. evalU7k+0 (botU7k+1 (Π2 ) \ botU7k+0 (Π2 ), X0 ∪ · · · ∪ X7k+0 ) = {ptarc(p, t, q, k):-.|q = W(p, t)} ∪
{tparc(t, p, q, k):-.|q = W(t, p)} ∪ {ptarc(p, t, q, k):-.|q = M0 (p)}. Its answer set X7k+1 =
Ats=k
∪ Ats=k
∪ Ats=k
19 – using forced atom proposition and construction of A5 , A6 , A19 .
5
6
10. evalU7k+1 (botU7k+2 (Π2 ) \ botU7k+1 (Π2 ), X0 ∪ X7k+1 ) = {notenabled(t, k):-.|{trans(t),
ptarc(p, t, n, k), holds(p, q, k)} ⊆ X0 ∪ X7k+1 , where q < n}. Its answer set X7k+2 = Ats=k
8
– using forced atom proposition and construction of A8 .
(a) where, q = M0 (p), and n = W(p, t) for an arc (p, t) ∈ E − – by construction of holds
and ptarc predicates in Π2 , and
(b) in an arc (p, t) ∈ E − , p ∈ •t (by definition 11 of preset)
(c) thus, notenabled(t, k) ∈ X7k+1 represents ∃p ∈ •t : M0 (p) < W(p, t).
11. evalU7k+2 (botU7k+3 (Π2 ) \ botU7k+2 (Π2 ), X0 ∪ · · · ∪ X7k+2 ) = {enabled(t, k):-.|trans(t) ∈ X0 ∪
· · · ∪ X7k+2 , notenabled(t, k) < X0 ∪ · · · ∪ X7k+2 }. Its answer set is X7k+3 = Ats=k
– using
9
forced atom proposition and construction of A9 .
(a) since an enabled(t, k) ∈ X7k+3 if @ notenabled(t, k) ∈ X0 ∪ · · · ∪ X7k+2 ; which is
equivalent to @p ∈ •t : M0 (p) < W(p, t) ≡ ∀p ∈ •t : M0 (p) ≥ W(p, t).
12. evalU7k+3 (botU7k+4 (Π2 ) \ botU7k+3 (Π2 ), X0 ∪ · · · ∪ X7k+3 ) = {{ f ires(t, k)}:-.|
enabled(t, k)
holds in X0 ∪ · · · ∪ X7k+3 }. It has multiple answer sets X7k+4.1 , . . . , X1k+4.n , corresponding to elements of power set of f ires(t, k) atoms in evalU7k+3 (...) – using supported rule
proposition. Since we are showing that the union of answer sets of Π2 determined using
splitting is equal to A, we only consider the set that matches the f ires(t, k) elements in A
and call it X7k+4 , ignoring the rest. Thus, X7k+4 = Ats=k
10 , representing T k .
(a) in addition, for every t such that enabled(t, k) ∈ X0 ∪· · ·∪X7k+3 , R(t) , ∅; f ires(t, k) ∈
X7k+4 – per definition 21 (firing set); requiring that a reset transition is fired when
enabled
(b) thus, the firing set T k will not be eliminated by the constraint f 10
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
207
13. evalU7k+4 (botU7k+5 (Π2 ) \ botU7k+4 (Π2 ), X0 ∪ · · · ∪ X7k+4 ) =
{add(p, n, t, k):-.|{ f ires(t, k), tparc(t, p, n, k)} ⊆ X0 ∪ · · · ∪ X7k+4 } ∪
{del(p, n, t, k):-.|{ f ires(t, k), ptarc(p, t, n, k)} ⊆ X0 ∪ · · · ∪ X7k+4 }. It’s answer set is X7k+5 =
ts=k
Ats=k
11 ∪ A12 – using forced atom proposition and definitions of A11 and A12 .
(a) where, each add atom is equivalent to n = W(t, p) : p ∈ •t,
(b) and each del atom is equivalent to n = W(p, t) : p ∈ t•; or n = Mk (p) : p ∈ R(t),
(c) representing the effect of transitions in T k
14. evalU7k+5 (botU7k+6 (Π2 ) \ botU7k+5 (Π2 ), X0 ∪ · · · ∪ X7k+5 ) =
P
{tot incr(p, qq, k):-.|qq = add(p,q,t,k)∈X0 ∪···∪X7k+5 q} ∪
P
ts=k
{tot decr(p, qq, k):-.|qq = del(p,q,t,k)∈X0 ∪···∪X7k+5 q}. It’s answer set is X7k+6 = Ats=k
13 ∪ A14
– using forced atom proposition and definitions of A13 and A14 .
P
(a) where, each tot incr(p, qq, k), qq = add(p,q,t,k)∈X0 ∪...X7k+5 q
P
≡ qq = t∈X7k+4 ,p∈t• W(p, t),
P
(b) and each tot decr(p, qq, k), qq = del(p,q,t,k)∈X0 ∪...X7k+5 q
P
P
≡ qq = t∈X7k+4 ,p∈•t W(t, p) + t∈X7k+4 ,p∈R(t) Mk (p),
(c) represent the net effect of transitions in T k
15. evalU7k+6 (botU7k+7 (Π2 ) \ botU7k+6 (Π2 ), X0 ∪ · · · ∪ X7k+6 ) = {consumesmore(p, k):-.|
{holds(p, q, k), tot decr(p, q1, k)} ⊆ X0 ∪ · · · ∪ X7k+6 , q1 > q} ∪ {holds(p, q, k + 1):-., |
{holds(p, q1, k), tot incr(p, q2, k), tot decr(p, q3, k)} ⊆ X0 ∪ · · · ∪ X7k+6 , q = q1 + q2 −
q3} ∪ {could not have(t, k):-.|{enabled(t, k), ptarc(s, t, q), holds(s, qq, k),
tot decr(s, qqq, k)} ⊆ X0 ∪ · · · ∪ X7k+6 , f ires(t, k) < (X0 ∪ · · · ∪ X7k+6 ), q > qq − qqq}. It’s
ts=k
answer set is X7k+7 = Ats=k
∪ Ats=k
17 ∪ A18 – using forced atom proposition and definitions
15
of A15 , A17 , A18 .
(a) where, consumesmore(p, k) represents ∃p : q = M0 (p), q1 =
P
P
t∈T 0 ,p∈•t W(p, t) + t∈T 0 ,p∈R(t) Mk (p), q1 > q, indicating place p that will be over
consumed if T k is fired, as defined in definition 23 (conflicting transitions)
(b) holds(p, q, k + 1) represents q = Mk+1 (p) – by construction of Π2 ,
(c) and could not have(t, k) represents enabled transition t in T k that could not be fired
due to insufficient tokens
(d) X7k+7 does not contain could not have(t, k), when enabled(t, k) ∈ X0 ∪ · · · ∪ X7k+6
and f ires(t, k) < X0 ∪ · · · ∪ X7k+6 due to the construction of A, encoding of a7 and
its body atoms. As a result it is not eliminated by the constraint a6
16. evalU7k+7 (botU7k+8 (Π2 ) \ botU7k+7 (Π2 ), X0 ∪ · · · ∪ X7k+7 ) = {consumesmore:-.|
{consumesmore(p, 0), . . . , consumesmore(p, k)} ⊆ X0 ∪ · · · ∪ X7k+7 }. It’s answer set is
X7k+8 = A16 – using forced atom proposition
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
208
(a) X7k+8 will be empty since none of consumesmore(p, 0), . . . ,
consumesmore(p, k) hold in X0 ∪ · · · ∪ X7k+7 due to the construction of A, encoding
of a2 and its body atoms. As a result, it is not eliminated by the constraint a4
The set X = X0 ∪ · · · ∪ X7k+8 is the answer set of Π2 by the splitting sequence theorem 15.
Each Xi , 0 ≤ i ≤ 7k + 8 matches a distinct portion of A, and X = A, thus A is an answer set of
Π2 .
Next we show (II): Given Π2 be the encoding of a Petri Net PN(P, T, E, W, R) with initial
marking M0 , and A be an answer set of Π2 that satisfies (A.5) and (A.6), then we can construct
X = M0 , T 0 , . . . , Mk , T k , Mk+1 from A, such that it is an execution sequence of PN.
We construct the X as follows:
1. Mi = (Mi (p0 ), . . . , Mi (pn )), where {holds(p0 , Mi (p0 ), i), . . . holds(pn , Mi (pn ), i)}
⊆ A, for 0 ≤ i ≤ k + 1
2. T i = {t : f ires(t, i) ∈ A}, for 0 ≤ i ≤ k
and show that X is indeed an execution sequence of PN. We show this by induction over k (i.e.
given Mk , T k is a valid firing set and its firing produces marking Mk+1 ).
Base case: Let k = 0, and M0 is a valid marking in X for PN, show (1) T 0 is a valid firing set
for M0 , and (2) T 0 ’s firing in M0 produces marking M1 .
1. We show T 0 is a valid firing set for M0 . Let { f ires(t0 , 0), . . . , f ires(t x , 0)} be the set of all
f ires(. . . , 0) atoms in A,
(a) Then for each f ires(ti , 0) ∈ A
enabled(ti , 0) ∈ A – from rule a1 and supported rule proposition
Then notenabled(ti , 0) < A – from rule e2 and supported rule proposition
Then body(e3) must not hold in A – from rule e3 and forced atom proposition
Then q ≮ ni ≡ q ≥ ni in e3 for all {holds(p, q, 0), ptarc(p, ti , ni , 0)} ⊆ A – from
e3, forced atom proposition, and the following
A. holds(p, q, 0) ∈ A represents q = M0 (p) – rule i1 construction
B. ptarc(p, ti , ni , 0) ∈ A represents ni = W(p, ti ) – rule f 7 construction; or
it represents ni = M0 (p) – rule f 9 construction; the construction of f 9
ensures that notenabled(t, 0) is never true due to the reset arc
v. Then ∀p ∈ •ti , M0 (p) ≥ W(p, ti ) – from definition 11 of preset •ti in PN
vi. Then ti is enabled and can fire in PN, as a result it can belong to T 0 – from
definition 12 of enabled transition
i.
ii.
iii.
iv.
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
209
(b) And consumesmore < A, since A is an answer set of Π2 – from rule a4 and supported
rule proposition
i. Then @consumesmore(p, 0) ∈ A – from rule a3 and supported rule proposition
ii. Then @{holds(p, q, 0), tot decr(p, q1, 0)} ⊆ A : q1 > q in body(a2) – from a2
and forced atom proposition
P
P
iii. Then @p : ( ti ∈{t0 ,...,tx },p∈•ti W(p, ti ) + ti ∈{t0 ,...,tx },p∈R(ti ) M0 (p)) > M0 (p) – from
the following
A. holds(p, q, 0) represents q = M0 (p) – from rule i1 construction, given
B. tot decr(p, q1, 0) ∈ A if {del(p, q10 , t0 , 0), . . . , del(p, q1 x , t x , 0)} ⊆ A, where
q1 = q10 + · · · + q1 x – from r4 and forced atom proposition
C. del(p, q1i , ti , 0) ∈ A if { f ires(ti , 0), ptarc(p, ti , q1i , 0)} ⊆ A – from r7 and
supported rule proposition
D. del(p, q1i , ti , 0) either represents removal of q1i = W(p, ti ) tokens from
p ∈ •ti ; or it represents removal of q1i = M0 (p) tokens from p ∈ R(ti )–
from rule r7, supported rule proposition, and definition 22 of transition
execution in PN
iv. Then the set of transitions in T 0 do not conflict – by the definition 23 of conflicting transitions
(c) And for each enabled(t j , 0) ∈ A and f ires(t j , 0) < A, could not have(t j , 0) ∈ A,
since A is an answer set of Π2 - from rule a6 and supported rule proposition
i. Then {enabled(t j , 0), holds(s, qq, 0), ptarc(s, t j , q, 0),
tot decr(s, qqq, 0)} ⊆ A, such that q > qq − qqq and f ires(t j , 0) < A - from rule
a7 and supported rule proposition
P
P
ii. Then for an s ∈ •t j ∪R(t j ), q > M0 (s)−( ti ∈T0 ,s∈•ti W(s, ti )+ ti ∈T0 ,s∈R(ti ) M0 (s)),
where q = W(s, t j ) if s ∈ •t j , or M0 (s) otherwise - from the following:
A. ptarc(s, ti , q, 0) represents q = W(s, ti ) if (s, ti ) ∈ E − or q = M0 (s) if s ∈
R(ti ) – from rule f 7, f 9 construction
B. holds(s, qq, 0) represents qq = M0 (s) – from i1 construction
C. tot decr(s, qqq, 0) ∈ A if {del(s, qqq0 , t0 , 0), . . . , del(s, qqq x , t x , 0)} ⊆ A –
from rule r4 construction and supported rule proposition
D. del(s, qqqi , ti , 0) ∈ A if { f ires(ti , 0), ptarc(s, ti , qqqi , 0)} ⊆ A – from rule r7
and supported rule proposition
E. del(s, qqqi , ti , 0) represents qqqi = W(s, ti ) : ti ∈ T 0 , (s, ti ) ∈ E − , or qqqi =
M0 (ti ) : ti ∈ T 0 , s ∈ R(ti ) – from rule f 7, f 9 construction
P
F. tot decr(q, qqq, 0) represents ti ∈T0 ,s∈•ti W(s, ti ) +
P
ti ∈T 0 ,s∈R(ti ) M0 (s) – from (C,D,E) above
iii. Then firing T 0 ∪ {t j } would have required more tokens than are present at its
source place s ∈ •t j ∪ R(t j ). Thus, T 0 is a maximal set of transitions that can
simultaneously fire.
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
210
(d) And for each reset transition tr with enabled(tr , 0) ∈ A, f ires(tr , 0) ∈ A, since A is
an answer set of Π2 - from rule f 10 and supported rule proposition
i. Then, the firing set T 0 satisfies the reset-transition requirement of definition 21
(firing set)
(e) Then {t0 , . . . , t x } = T 0 – using 1(a),1(b),1(d) above; and using 1(c) it is a maximal
firing set
2. We show M1 is produced by firing T 0 in M0 . Let holds(p, q, 1) ∈ A
(a) Then {holds(p, q1, 0), tot incr(p, q2, 0), tot decr(p, q3, 0)} ⊆ A : q = q1 + q2 − q3 –
from rule r5 and supported rule proposition
(b) holds(p, q1, 0) ∈ A represents q1 = M0 (p) – given, rule i1 construction; and {add(p, q20 , t0 , 0), . . . ,
add(p, q2 j , t j , 0)} ⊆ A : q20 +· · ·+q2 j = q2 and {del(p, q30 , t0 , 0), . . . , del(p, q3l , tl , 0)} ⊆
A : q30 + · · · + q3l = q3 – rules r3, r4 using supported rule proposition
(c) Then { f ires(t0 , 0), . . . , f ires(t j , 0)} ⊆ A and { f ires(t0 , 0), . . . , f ires(tl , 0)}
⊆ A – rules r6, r7 and supported rule proposition, respectively
(d) Then { f ires(t0 , 0), . . . , f ires(t j , 0)}∪{ f ires(t0 , 0), . . . , f ires(tl , 0)} ⊆ A = { f ires(t0 , 0), . . . , f ires(t x , 0)} ⊆
A – set union of subsets
(e) Then for each f ires(t x , 0) ∈ A we have t x ∈ T 0 – already shown in item 1 above
P
P
(f) Then q = M0 (p) + tx ∈T0 ∧p∈tx • W(t x , p) − ( tx ∈T0 ∧p∈•tx W(p, t x ) +
P
t x ∈T 0 ∧p∈R(t x ) M0 (p)) – from (2a),(2b) (2b) above and the following
i. Each add(p, q j , t j , 0) ∈ A represents q j = W(t j , p) for p ∈ t j • – rule r6 encoding, and definition 11 of postset in PN
ii. Each del(p, ty , qy , 0) ∈ A represents either qy = W(p, ty ) for p ∈ •ty , or qy =
M0 (p) for p ∈ R(ty ) – from rule r7, f 7 encoding and definition 11 of preset in
PN; or from rule r7, f 9 encoding and definition of reset arc in PN
P
iii. Each tot incr(p, q2, 0) ∈ A represents q2 = tx ∈T0 ∧p∈tx • W(t x , p) – aggregate
assignment atom semantics in rule r3
P
P
iv. Each tot decr(p, q3, 0) ∈ A represents q3 = tx ∈T0 ∧p∈•tx W(p, t x )+ tx ∈T0 ∧p∈R(tx ) M0 (p)
– aggregate assignment atom semantics in rule r4
(g) Then M1 (p) = q – since holds(p, q, 1) ∈ A encodes q = M1 (p) – from construction
Inductive Step: Let k > 0, and Mk is a valid marking in X for PN, show (1) T k is a valid firing
set in Mk , and (2) firing T k in Mk produces marking Mk+1 .
1. We show that T k is a valid firing set in Mk . Let { f ires(t0 , k), . . . , f ires(t x , k)} be the set of
all f ires(. . . , k) atoms in A,
(a) Then for each f ires(ti , k) ∈ A
i. enabled(ti , k) ∈ A – from rule a1 and supported rule proposition
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
211
ii. Then notenabled(ti , k) < A – from rule e2 and supported rule proposition
iii. Then body(e3) must not hold in A – from rule e3 and forced atom proposition
iv. Then q ≮ ni ≡ q ≥ ni in e3 for all {holds(p, q, k), ptarc(p, ti , ni , k)} ⊆ A – from
e3, forced atom proposition, and the following
A. holds(p, q, k) ∈ A represents q = Mk (p) – construction, inductive assumption
B. ptarc(p, ti , ni , k) ∈ A represents ni = W(p, ti ) – rule f 7 construction; or
it represents ni = Mk (p) – rule f 9 construction; the construction of f 9
ensures that notenabled(t, 0) is never true due to the reset arc
v. Then ∀p ∈ •ti , Mk (p) ≥ W(p, ti ) – from definition 11 of preset •ti in PN
vi. Then ti is enabled and can fire in PN, as a result it can belong to T k – from
definition 12 of enabled transition
(b) And consumesmore < A, since A is an answer set of Π2 – from rule a4 and supported
rule proposition
i. Then @consumesmore(p, k) ∈ A – from rule a3 and supported rule proposition
ii. Then @{holds(p, q, k), tot decr(p, q1, k)} ⊆ A : q1 > q in body(a2) – from a2
and forced atom proposition
P
P
iii. Then @p : ( ti ∈{t0 ,...,tx },p∈•ti W(p, ti ) + ti ∈{t0 ,...,tx },p∈R(ti ) Mk (p)) > Mk (p) – from
the following
A. holds(p, q, k) represents q = Mk (p) – inductive assumption, given
B. tot decr(p, q1, k) ∈ A if {del(p, q10 , t0 , k), . . . , del(p, q1 x , t x , k)} ⊆ A, where
q1 = q10 + · · · + q1 x – from r4 and forced atom proposition
C. del(p, q1i , ti , k) ∈ A if { f ires(ti , k), ptarc(p, ti , q1i , k)} ⊆ A – from r7 and
supported rule proposition
D. del(p, q1i , ti , k) either represents removal of q1i = W(p, ti ) tokens from
p ∈ •ti ; or it represents removal of q1i = Mk (p) tokens from p ∈ R(ti )–
from rule r7, supported rule proposition, and definition 22 of transition
execution in PN
iv. Then the set of transitions in T k do not conflict – by the definition 23 of conflicting transitions
(c) And for each enabled(t j , k) ∈ A and f ires(t j , k) < A, could not have(t j , k) ∈ A,
since A is an answer set of Π2 - from rule a6 and supported rule proposition
i. Then {enabled(t j , k), holds(s, qq, k), ptarc(s, t j , q, k),
tot decr(s, qqq, k)} ⊆ A, such that q > qq − qqq and f ires(t j , k) < A - from rule
a7 and supported rule proposition
P
P
ii. Then for an s ∈ •t j ∪R(t j ), q > Mk (s)−( ti ∈Tk ,s∈•ti W(s, ti )+ ti ∈Tk ,s∈R(ti ) Mk (s)),
where q = W(s, t j ) if s ∈ •t j , or Mk (s) otherwise - from the following:
A. ptarc(s, ti , q, k) represents q = W(s, ti ) if (s, ti ) ∈ E − or q = Mk (s) if s ∈
R(ti ) – from rule f 7, f 9 construction
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
212
B. holds(s, qq, k) represents qq = Mk (s) – construction
C. tot decr(s, qqq, k) ∈ A if {del(s, qqq0 , t0 , k), . . . , del(s, qqq x , t x , k)} ⊆ A –
from rule r4 construction and supported rule proposition
D. del(s, qqqi , ti , k) ∈ A if { f ires(ti , k), ptarc(s, ti , qqqi , k)} ⊆ A – from rule r7
and supported rule proposition
E. del(s, qqqi , ti , k) represents qqqi = W(s, ti ) : ti ∈ T k , (s, ti ) ∈ E − , or qqqi =
Mk (ti ) : ti ∈ T k , s ∈ R(ti ) – from rule f 7, f 9 construction
P
F. tot decr(q, qqq, k) represents ti ∈T k ,s∈•ti W(s, ti ) +
P
ti ∈T k ,s∈R(ti ) Mk (s) – from (C,D,E) above
iii. Then firing T k ∪ {t j } would have required more tokens than are present at its
source place s ∈ •t j ∪ R(t j ). Thus, T k is a maximal set of transitions that can
simultaneously fire.
(d) And for each reset transition tr with enabled(tr , k) ∈ A, f ires(tr , k) ∈ A, since A is
an answer set of Π2 - from rule f 10 and supported rule proposition
i. Then the firing set T k satisfies the reset transition requirement of definition 21
(firing set)
(e) Then {t0 , . . . , t x } = T k – using 1(a),1(b), 1(d) above; and using 1(c) it is a maximal
firing set
2. We show that Mk+1 is produced by firing T k in Mk . Let holds(p, q, k + 1) ∈ A
(a) Then {holds(p, q1, k), tot incr(p, q2, k), tot decr(p, q3, k)} ⊆ A : q = q1 + q2 − q3 –
from rule r5 and supported rule proposition
(b) holds(p, q1, k) ∈ A represents q1 = Mk (p) – construction, inductive assumption;
and {add(p, q20 , t0 , k), . . . , add(p, q2 j , t j , k)} ⊆ A : q20 + · · · + q2 j = q2 and
{del(p, q30 , t0 , k), . . . , del(p, q3l , tl , k)} ⊆ A : q30 + · · · + q3l = q3 – rules r3, r4
using supported rule proposition
(c) Then { f ires(t0 , k), . . . , f ires(t j , k)} ⊆ A and { f ires(t0 , k), . . . , f ires(tl , k)}
⊆ A – rules r6, r7 and supported rule proposition, respectively
(d) Then { f ires(t0 , k), . . . , f ires(t j , k)}∪{ f ires(t0 , k), . . . , f ires(tl , k)} ⊆ A = { f ires(t0 , k), . . . ,
f ires(t x , k)} ⊆ A – set union of subsets
(e) Then for each f ires(t x , k) ∈ A we have t x ∈ T k – already shown in item 1 above
P
P
(f) Then q = Mk (p) + tx ∈T0 ∧p∈tx • W(t x , p) − ( tx ∈Tk ∧p∈•tx W(p, t x ) +
P
t x ∈T k ∧p∈R(t x ) Mk (p)) – from (2b) above and the following
i. Each add(p, q j , t j , k) ∈ A represents q j = W(t j , p) for p ∈ t j • – rule r6 encoding, and definition 22 of transition execution in PN
ii. Each del(p, ty , qy , k) ∈ A represents either qy = W(p, ty ) for p ∈ •ty , or qy =
Mk (p) for p ∈ R(ty ) – from rule r7, f 7 encoding and definition 22 of transition
execution in PN; or from rule r7, f 9 encoding and definition of reset arc in PN
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
213
P
iii. Each tot incr(p, q2, k) ∈ A represents q2 = tx ∈Tk ∧p∈tx • W(t x , p) – aggregate
assignment atom semantics in rule r3
P
P
iv. Each tot decr(p, q3, 0) ∈ A represents q3 = tx ∈Tk ∧p∈•tx W(p, t x )+ tx ∈Tk ∧p∈R(tx ) M0 (p)
– aggregate assignment atom semantics in rule r4
(g) Then, Mk+1 (p) = q – since holds(p, q, k + 1) ∈ A encodes q = Mk+1 (p) – from
construction
As a result, for any n > k, T n is a valid firing set w.r.t. Mn and its firing produces marking Mn+1 .
Conclusion: Since both (I) and (II) hold, X = M0 , T 0 , M1 , . . . , Mk , T k+1 is an execution sequence of PN(P, T, E, W, R) (w.r.t M0 ) iff there is an answer set A of Π2 (PN, M0 , k, ntok) such
that (A.5) and (A.6) hold.
A.4
Poof of Proposition 4
Let PN = (P, T, E, W, R, I) be a Petri Net, M0 be its initial marking and let Π3 (PN, M0 , k, ntok)
be the ASP encoding of PN and M0 over a simulation length k, with maximum ntok tokens
on any place node, as defined in section 2.6. Then X = M0 , T 0 , M1 , . . . , Mk , T k , Mk+1 is an
execution sequence of PN (w.r.t. M0 ) iff there is an answer set A of Π3 (PN, M0 , k, ntok) such
that:
{ f ires(t, ts) : t ∈ T ts , 0 ≤ ts ≤ k} = { f ires(t, ts) : f ires(t, ts) ∈ A}
(A.7)
{holds(p, q, ts) : p ∈ P, q = Mts (p), 0 ≤ ts ≤ k + 1}
= {holds(p, q, ts) : holds(p, q, ts) ∈ A}
(A.8)
We prove this by showing that:
(I) Given an execution sequence X, we create a set A such that it satisfies (A.7) and (A.4) and
show that A is an answer set of Π3
(II) Given an answer set A of Π3 , we create an execution sequence X such that (A.3) and (A.8)
are satisfied.
First we show (I): Given PN and an execution sequence X of PN, we create a set A as a union
of the following sets:
1. A1 = {num(n) : 0 ≤ n ≤ ntok}
2. A2 = {time(ts) : 0 ≤ ts ≤ k}
3. A3 = {place(p) : p ∈ P}
4. A4 = {trans(t) : t ∈ T }
5. A5 = {ptarc(p, t, n, ts) : (p, t) ∈ E − , n = W(p, t), 0 ≤ ts ≤ k}, where E − ⊆ E
6. A6 = {tparc(t, p, n, ts) : (t, p) ∈ E + , n = W(t, p), 0 ≤ ts ≤ k}, where E + ⊆ E
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
214
7. A7 = {holds(p, q, 0) : p ∈ P, q = M0 (p)}
8. A8 = {notenabled(t, ts) : t ∈ T, 0 ≤ ts ≤ k, (∃p ∈ •t, Mts (p) < W(p, t)) ∨ (∃p ∈
I(t), Mts (p) , 0)}
per definition 26 (enabled transition)
9. A9 = {enabled(t, ts) : t ∈ T, 0 ≤ ts ≤ k, (∀p ∈ •t, W(p, t) ≤ Mts (p)) ∧ (∀p ∈ I(t), Mts (p) =
0)}
per definition 26 (enabled transition)
10. A10 = { f ires(t, ts) : t ∈ T ts , 0 ≤ ts ≤ k}
per definition 21 (firing set), only an enabled transition may fire
11. A11 = {add(p, q, t, ts) : t ∈ T ts , p ∈ t•, q = W(t, p), 0 ≤ ts ≤ k}
per definition 22 (transition execution)
12. A12 = {del(p, q, t, ts) : t ∈ T ts , p ∈ •t, q = W(p, t), 0 ≤ ts ≤ k} ∪ {del(p, q, t, ts) : t ∈
T ts , p ∈ R(t), q = Mts (p), 0 ≤ ts ≤ k}
per definition 24 (transition execution)
P
13. A13 = {tot incr(p, q, ts) : p ∈ P, q = t∈Tts ,p∈t• W(t, p), 0 ≤ ts ≤ k}
per definition 24 (firing set execution)
P
P
14. A14 = {tot decr(p, q, ts) : p ∈ P, q = t∈Tts ,p∈•t W(p, t) + t∈Tts ,p∈R(t) Mts (p), 0 ≤ ts ≤ k}
per definition 24 (firing set execution)
P
P
15. A15 = {consumesmore(p, ts) : p ∈ P, q = Mts (p), q1 = t∈Tts ,p∈•t W(p, t)+ t∈T ts ,p∈R(t) Mts (p), q1 >
q, 0 ≤ ts ≤ k}
per definition 23 (conflicting transitions)
P
P
16. A16 = {consumesmore : ∃p ∈ P : q = Mts (p), q1 = t∈T ts ,p∈•t W(p, t)+ t∈Tts ,p∈R(t) (Mts (p)), q1 >
q, 0 ≤ ts ≤ k}
per definition 23 (conflicting transitions)
17. A17 = {could not have(t, ts) : t ∈ T, (∀p ∈ •t, W(p, t) ≤ Mts (p)), t < T ts , (∃p ∈ •t :
P
P
W(p, t) > Mts (p) − ( t0 ∈Tts ,p∈•t0 W(p, t0 ) + t0 ∈T ts ,p∈R(t0 ) Mts (p)), 0 ≤ ts ≤ k}
per the maximal firing set semantics
18. A18 = {holds(p, q, ts + 1) : p ∈ P, q = Mts+1 (p), 0 ≤ ts < k},
P
P
where Mts+1 (p) = Mts (p) − ( t∈T ts ,p∈•t W(p, t)+ t∈T ts ,p∈R(t) Mts (p)) +
P
t∈T ts ,p∈t• W(t, p)
according to definition 21 (firing set execution)
19. A19 = {ptarc(p, t, n, ts) : p ∈ R(t), n = Mts (p), n > 0, 0 ≤ ts ≤ k}
20. A20 = {iptarc(p, t, 1, ts) : p ∈ P, 0 ≤ ts < k}
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
215
We show that A satisfies (A.7) and (A.8), and A is an answer set of Π3 .
A satisfies (A.7) and (A.8) by its construction above. We show A is an answer set of Π3 by
splitting. We split lit(Π3 ) into a sequence of 7k + 9 sets:
• U0 = head( f 1) ∪ head( f 2) ∪ head( f 5) ∪ head( f 6) ∪ head(i1) = {place(p) : p ∈ P} ∪
{trans(t) : t ∈ T } ∪ {time(0), . . . , time(k)} ∪ {num(0), . . . , num(ntok)} ∪ {holds(p, q, 0) : p ∈
P, q = M0 (p)}
• U7k+1 = U7k+0 ∪ head( f 7)ts=k ∪ head( f 8)ts=k ∪ head( f 9)ts=k ∪ head( f 11)ts=k = U7k+0 ∪
{ptarc(p, t, n, k) : (p, t) ∈ E − , n = W(p, t)} ∪ {tparc(t, p, n, k) : (t, p) ∈ E + , n = W(t, p)} ∪
{ptarc(p, t, n, k) : p ∈ R(t), n = Mk (p), n > 0} ∪
{iptarc(p, t, 1, k) : p ∈ I(t)}
• U7k+2 = U7k+1 ∪ head(e3)ts=k ∪ head(e4)ts=k = U7k+1 ∪ {notenabled(t, k) : t ∈ T }
• U7k+3 = U7k+2 ∪ head(e2)ts=k = U7k+2 ∪ {enabled(t, k) : t ∈ T }
• U7k+4 = U7k+3 ∪ head(a1)ts=k = U7k+3 ∪ { f ires(t, k) : t ∈ T }
• U7k+5 = U7k+4 ∪ head(r6)ts=k ∪ head(r7)ts=k = U7k+4 ∪ {add(p, q, t, k) : p ∈ P, t ∈ T, q =
W(t, p)} ∪ {del(p, q, t, k) : p ∈ P, t ∈ T, q = W(p, t)} ∪ {del(p, q, t, k) : p ∈ P, t ∈ T, q =
Mk (p)}
• U7k+6 = U7k+5 ∪ head(r3)ts=k ∪ head(r4)ts=k = U7k+5 ∪ {tot incr(p, q, k) : p ∈ P, 0 ≤ q ≤
ntok} ∪ {tot decr(p, q, k) : p ∈ P, 0 ≤ q ≤ ntok}
• U7k+7 = U7k+6 ∪ head(r5)ts=k ∪ head(a2)ts=k ∪ head(a7)ts=k = U7k+6 ∪
{consumesmore(p, k) : p ∈ P}∪{holds(p, q, k+1) : p ∈ P, 0 ≤ q ≤ ntok}∪{could not have(t, k) :
t ∈ T}
• U7k+8 = U7k+7 ∪ head(a3)ts=k = U7k+7 ∪ {consumesmore}
where head(ri )ts=k are head atoms of ground rule ri in which ts = k. We write Ats=k
=
i
{a(. . . , ts) : a(. . . , ts) ∈ Ai , ts = k} as short hand for all atoms in Ai with ts = k. Uα , 0 ≤
α ≤ 7k + 8 form a splitting sequence, since each Ui is a splitting set of Π1 , and hUα iα<µ is a
S
monotone continuous sequence, where U0 ⊆ U1 · · · ⊆ U7k+8 and α<µ Uα = lit(Π1 ).
We compute the answer set of Π3 using the splitting sets as follows:
1. botU0 (Π3 ) = f 1 ∪ f 2 ∪ f 5 ∪ i1 ∪ f 6 and X0 = A1 ∪ · · · ∪ A4 ∪ A7 (= U0 ) is its answer set
– using forced atom proposition
2. evalU0 (botU1 (Π3 ) \ botU0 (Π3 ), X0 ) = {ptarc(p, t, q, 0):-.|q = W(p, t)} ∪
{tparc(t, p, q, 0):-.|q = W(t, p)} ∪ {ptarc(p, t, q, 0):-.|q = M0 (p)} ∪
ts=0
{iptarc(p, t, 1, 0):-.}. Its answer set X1 = Ats=0
∪ Ats=0
∪ Ats=0
19 ∪ A20 – using forced atom
5
6
proposition and construction of A5 , A6 , A19 , A20 .
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
216
3. evalU1 (botU2 (Π3 ) \ botU1 (Π3 ), X0 ∪ X1 ) = {notenabled(t, 0):-.|({trans(t),
ptarc(p, t, n, 0), holds(p, q, 0)} ⊆ X0 ∪ X1 , where q < n) or {notenabled(t, 0):-.|
({trans(t), iptarc(p, t, n2, 0), holds(p, q, 0)} ⊆ X0 ∪ X1 , where q ≥ n2}}. Its answer set
X2 = Ats=0
– using forced atom proposition and construction of A8 .
8
(a) where, q = M0 (p), and n = W(p, t) for an arc (p, t) ∈ E − – by construction i1 and
f 7 in Π3 , and
(b) in an arc (p, t) ∈ E − , p ∈ •t (by definition 11 of preset)
(c) n2 = 1 – by construction of iptarc predicates in Π3 , meaning q ≥ n2 ≡ q ≥ 1 ≡ q >
0,
(d) thus, notenabled(t, 0) ∈ X1 represents (∃p ∈ •t : M0 (p) < W(p, t)) ∨ (∃p ∈ I(t) :
M0 (p) > 0).
4. evalU2 (botU3 (Π3 ) \ botU2 (Π3 ), X0 ∪ · · · ∪ X2 ) = {enabled(t, 0):-.|trans(t) ∈ X0 ∪ · · · ∪
– using forced atom
X2 , notenabled(t, 0) < X0 ∪ · · · ∪ X2 }. Its answer set is X3 = Ats=0
9
proposition and construction of A9 .
(a) since an enabled(t, 0) ∈ X3 if @ notenabled(t, 0) ∈ X0 ∪ · · · ∪ X2 ; which is equivalent
to (@p ∈ •t : M0 (p) < W(p, t)) ∧ (@p ∈ I(t) : M0 (p) > 0) ≡ (∀p ∈ •t : M0 (p) ≥
W(p, t)) ∧ (∀p ∈ I(t) : M0 (p) = 0).
5. evalU3 (botU4 (Π3 ) \ botU3 (Π3 ), X0 ∪ · · · ∪ X3 ) = {{ f ires(t, 0)}:-.|enabled(t, 0)
holds in X0 ∪ · · · ∪ X3 }. It has multiple answer sets X4.1 , . . . , X4.n , corresponding to elements of power set of f ires(t, 0) atoms in evalU3 (...) – using supported rule proposition.
Since we are showing that the union of answer sets of Π3 determined using splitting is
equal to A, we only consider the set that matches the f ires(t, 0) elements in A and call it
X4 , ignoring the rest. Thus, X4 = Ats=0
10 , representing T 0 .
(a) in addition, for every t such that enabled(t, 0) ∈ X0 ∪ · · · ∪ X3 , R(t) , ∅; f ires(t, 0) ∈
X4 – per definition 21 (firing set); requiring that a reset transition is fired when
enabled
(b) thus, the firing set T 0 will not be eliminated by the constraint f 10
6. evalU4 (botU5 (Π3 ) \ botU4 (Π3 ), X0 ∪ · · · ∪ X4 ) = {add(p, n, t, 0):-.|{ f ires(t, 0),
tparc(t, p, n, 0)} ⊆ X0 ∪ · · · ∪ X4 } ∪ {del(p, n, t, 0):-.|{ f ires(t, 0), ptarc(p, t, n, 0)}
ts=0
⊆ X0 ∪ · · · ∪ X4 }. It’s answer set is X5 = Ats=0
11 ∪ A12 – using forced atom proposition
and definitions of A11 and A12 .
(a) where, each add atom is equivalent to n = W(t, p) : p ∈ t•,
(b) and each del atom is equivalent to n = W(p, t) : p ∈ •t; or n = Mk (p) : p ∈ R(t),
(c) representing the effect of transitions in T 0 .
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
217
7. evalU5 (botU6 (Π3 ) \ botU5 (Π3 ), X0 ∪ · · · ∪ X5 ) = {tot incr(p, qq, 0):-.|
P
P
qq = add(p,q,t,0)∈X0 ∪···∪X5 q} ∪ {tot decr(p, qq, 0):-.|qq = del(p,q,t,0)∈X0 ∪···∪X5 q}. It’s ants=0
swer set is X6 = Ats=0
13 ∪ A14 – using forced atom proposition and definitions of A13 and
A14 .
P
(a) where, each tot incr(p, qq, 0), qq = add(p,q,t,0)∈X0 ∪...X5 q
P
≡ qq = t∈X4 ,p∈t• W(p, t),
P
(b) and each tot decr(p, qq, 0), qq = del(p,q,t,0)∈X0 ∪...X5 q
P
P
≡ qq = t∈X4 ,p∈•t W(t, p) + t∈X4 ,p∈R(t) Mk (p),
(c) represent the net effect of transitions in T 0 .
8. evalU6 (botU7 (Π3 ) \ botU6 (Π3 ), X0 ∪ · · · ∪ X6 ) = {consumesmore(p, 0):-.|
{holds(p, q, 0), tot decr(p, q1, 0)} ⊆ X0 ∪ · · · ∪ X6 , q1 > q} ∪ {holds(p, q, 1):-., |
{holds(p, q1, 0), tot incr(p, q2, 0), tot decr(p, q3, 0)} ⊆ X0 ∪ · · · ∪ X6 , q = q1 + q2 − q3} ∪
{could not have(t, 0):-.|{enabled(t, 0), ptarc(s, t, q), holds(s, qq, 0),
tot decr(s, qqq, 0)} ⊆ X0 ∪ · · · ∪ X6 , f ires(t, 0) < (X0 ∪ · · · ∪ X6 ), q > qq − qqq}. It’s
ts=0
∪ Ats=0
answer set is X7 = Ats=0
17 ∪ A18 – using forced atom proposition and definitions
15
of A15 , A17 , A18 , A9 .
(a) where, consumesmore(p, 0) represents ∃p : q = M0 (p), q1 =
P
P
t∈T 0 ,p∈•t W(p, t) + t∈T 0 ,p∈R(t) Mk (p), q1 > q, indicating place p will be over consumed if T 0 is fired, as defined in definition 23 (conflicting transitions)
(b) holds(p, q, 1) represents q = M1 (p) – by construction of Π3
(c) and could not have(t, 0) represents enabled transition t ∈ T 0 that could not fire due
to insufficient tokens
(d) X7 does not contain could not have(t, 0), when enabled(t, 0) ∈ X0 ∪ · · · ∪ X5 and
f ires(t, 0) < X0 ∪ · · · ∪ X6 due to construction of A, encoding of a7 and its body
atoms. As a result it is not eliminated by the constraint a6
..
.
9. evalU7k+0 (botU7k+1 (Π3 ) \ botU7k+0 (Π3 ), X0 ∪ · · · ∪ X7k+0 ) = {ptarc(p, t, q, k):-.|q = W(p, t)} ∪
{tparc(t, p, q, k):-.|q = W(t, p)} ∪ {ptarc(p, t, q, k):-.|q = Mk (p)} ∪ {iptarc(p, t, 1, k):-.}. Its
ts=k
answer set X7k+1 = Ats=k
∪ Ats=k
∪ Ats=k
19 ∪ A20 – using forced atom proposition and
5
6
construction of A5 , A6 , A19 , A20 .
10. evalU7k+1 (botU7k+2 (Π3 ) \ botU7k+1 (Π3 ), X0 ∪ · · · ∪ X7k+1 ) = {notenabled(t, k):-.|
({trans(t), ptarc(p, t, n, k), holds(p, q, k)} ⊆ X0 ∪· · ·∪X7k+1 , where q < n)} or {notenabled(t, k):-.|({trans(t), iptar
X0 ∪ · · · ∪ X7k+1 , where q ≥ n2}}. Its answer set X7k+2 = Ats=k
– using forced atom
8
proposition and construction of A8 .
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
218
(a) where, q = Mk (p), and n = W(p, t) for an arc (p, t) ∈ E − – by construction of holds
and ptarc predicates in Π3 , and
(b) in an arc (p, t) ∈ E − , p ∈ •t (by definition 11 of preset)
(c) n2 = 1 – by construction of iptarc predicates in Π3 , meaning q ≥ n2 ≡ q ≥ 1 ≡ q >
0,
(d) thus, notenabled(t, k) ∈ X7k+1 represents (∃p ∈ •t : Mk (p) < W(p, t)) ∨ (∃p ∈ I(t) :
Mk (p) > k).
11. evalU7k+2 (botU7k+3 (Π3 ) \ botU7k+2 (Π3 ), X0 ∪ · · · ∪ X7k+2 ) = {enabled(t, k):-.|
trans(t) ∈ X0 ∪ · · · ∪ X7k+2 ∧ notenabled(t, k) < X0 ∪ · · · ∪ X7k+2 }. Its answer set is
X7k+3 = Ats=k
– using forced atom proposition and construction of A9 .
9
(a) since an enabled(t, k) ∈ X7k+3 if @ notenabled(t, k) ∈ X0 ∪ · · · ∪ X7k+2 ; which is
equivalent to (@p ∈ •t : Mk (p) < W(p, t)) ∧ (@p ∈ I(t) : Mk (p) > k) ≡ (∀p ∈ •t :
Mk (p) ≥ W(p, t)) ∧ (∀p ∈ I(t) : Mk (p) = k).
12. evalU7k+3 (botU7k+4 (Π3 ) \ botU7k+3 (Π3 ), X0 ∪ · · · ∪ X7k+3 ) = {{ f ires(t, k)}:-.|
enabled(t, k)
holds in X0 ∪ · · · ∪ X7k+3 }. It has multiple answer sets X7k+4.1 , . . . , X1k+4.n , corresponding to elements of power set of f ires(t, k) atoms in evalU7k+3 (...) – using supported rule
proposition. Since we are showing that the union of answer sets of Π3 determined using
splitting is equal to A, we only consider the set that matches the f ires(t, k) elements in A
and call it X7k+4 , ignoring the rest. Thus, X7k+4 = Ats=k
10 , representing T k .
(a) in addition, for every t such that enabled(t, k) ∈ X0 ∪· · ·∪X7k+3 , R(t) , ∅; f ires(t, k) ∈
X7k+4 – per definition 21 (firing set); requiring that a reset transition is fired when
enabled
(b) thus, the firing set T k will not be eliminated by the constraint f 10
13. evalU7k+4 (botU7k+5 (Π3 ) \ botU7k+4 (Π3 ), X0 ∪ · · · ∪ X7k+4 ) = {add(p, n, t, k):-.|
{ f ires(t, k), tparc(t, p, n, k)} ⊆ X0 ∪ · · · ∪ X7k+4 } ∪ {del(p, n, t, k):-.|{ f ires(t, k),
ts=k
ptarc(p, t, n, k)} ⊆ X0 ∪ · · · ∪ X7k+4 }. It’s answer set is X7k+5 = Ats=k
11 ∪ A12 – using forced
atom proposition and definitions of A11 and A12 .
(a) where, each add atom is equivalent to n = W(t, p) : p ∈ t•,
(b) and each del atom is equivalent to n = W(p, t) : p ∈ •t; or n = Mk (p) : p ∈ R(t),
(c) representing the effect of transitions in T k
14. evalU7k+5 (botU7k+6 (Π3 ) \ botU7k+5 (Π3 ), X0 ∪ · · · ∪ X7k+5 ) =
P
P
{tot incr(p, qq, k):-.| qq = add(p,q,t,k)∈X0 ∪···∪X7k+5 q}∪ {tot decr(p, qq, k):-.| qq = del(p,q,t,k)∈X0 ∪···∪X7k+5 q}.
ts=k
It’s answer set is X7k+6 = Ats=k
13 ∪ A14 – using forced atom proposition and definitions of
A13 and A14 .
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
219
P
(a) where, each tot incr(p, qq, k), qq = add(p,q,t,k)∈X0 ∪...X7k+5 q
P
≡ qq = t∈X7k+4 ,p∈t• W(p, t),
P
(b) and each tot decr(p, qq, k), qq = del(p,q,t,k)∈X0 ∪...X7k+5 q
P
P
≡ qq = t∈X7k+4 ,p∈•t W(t, p) + t∈X7k+4 ,p∈R(t) Mk (p),
(c) representing the net effect of transitions in T k
15. evalU7k+6 (botU7k+7 (Π3 ) \ botU7k+6 (Π3 ), X0 ∪ · · · ∪ X7k+6 ) = {consumesmore(p, k):-.|
{holds(p, q, k), tot decr(p, q1, k)} ⊆ X0 ∪ · · · ∪ X7k+6 , q1 > q} ∪ {holds(p, q, 1):-., |
{holds(p, q1, k), tot incr(p, q2, k), tot decr(p, q3, k)} ⊆ X0 ∪ · · · ∪ X7k+6 , q = q1 + q2 −
q3} ∪ {could not have(t, k):-.|{enabled(t, k), ptarc(s, t, q), holds(s, qq, k),
tot decr(s, qqq, k)} ⊆ X0 ∪ · · · ∪ X7k+6 , f ires(t, k) < (X0 ∪ · · · ∪ X7k+6 ), q > qq − qqq}. It’s
ts=k
answer set is X7k+7 = Ats=k
∪ Ats=k
17 ∪ A18 – using forced atom proposition and definitions
15
of A15 , A17 , A18 , A9 .
(a) where, consumesmore(p, k) represents ∃p : q = Mk (p),
P
P
q1 = t∈Tk ,p∈•t W(p, t) + t∈Tk ,p∈R(t) Mk (p), q1 > q, indicating place p that will be
over consumed if T k is fired, as defined in definition 23 (conflicting transitions),
(b) holds(p, q, k + 1) represents q = Mk+1 (p) – by construction of Π3
(c) and could not have(t, k) represents enabled transition t in T k that could not be fired
due to insufficient tokens
(d) X7k+7 does not contain could not have(t, k), when enabled(t, k) ∈ X0 ∪ · · · ∪ X7k+6
and f ires(t, k) < X0 ∪ · · · ∪ X7k+6 due to construction of A, encoding of a7 and its
body atoms. As a result it is not eliminated by the constraint a6
16. evalU7k+7 (botU7k+8 (Π3 ) \ botU7k+7 (Π3 ), X0 ∪ · · · ∪ X7k+7 ) = {consumesmore:-.|
{consumesmore(p, 0), . . . , consumesmore(p, k)} ∩ (X0 ∪ · · · ∪ X7k+7 ) , ∅}. It’s answer set
is X7k+8 = A16 – using forced atom proposition
(a) X7k+8 will be empty since none of consumesmore(p, 0), . . . ,
consumesmore(p, k) hold in X0 ∪ · · · ∪ X7k+7 due to the construction of A, encoding
of a2 and its body atoms. As a result, it is not eliminated by the constraint a4
The set X = X0 ∪ · · · ∪ X7k+8 is the answer set of Π3 by the splitting sequence theorem 15.
Each Xi , 0 ≤ i ≤ 7k + 8 matches a distinct portion of A, and X = A, thus A is an answer set of
Π3 .
Next we show (II): Given Π3 be the encoding of a Petri Net PN(P, T, E, W, R, I) with initial
marking M0 , and A be an answer set of Π3 that satisfies (A.7) and (A.8), then we can construct
X = M0 , T 0 , . . . , Mk , T k , Mk+1 from A, such that it is an execution sequence of PN.
We construct the X as follows:
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
220
1. Mi = (Mi (p0 ), . . . , Mi (pn )), where {holds(p0 , Mi (p0 ), i), . . . holds(pn , Mi (pn ), i)}
⊆ A, for 0 ≤ i ≤ k + 1
2. T i = {t : f ires(t, i) ∈ A}, for 0 ≤ i ≤ k
and show that X is indeed an execution sequence of PN. We show this by induction over k (i.e.
given Mk , T k is a valid firing set and its firing produces marking Mk+1 ).
Base case: Let k = 0, and M0 is a valid marking in X for PN, show (1) T 0 is a valid firing set
for M0 , and (2) M1 is T 0 ’s target marking w.r.t. M0 .
1. We show T 0 is a valid firing set for M0 . Let { f ires(t0 , 0), . . . , f ires(t x , 0)} be the set of all
f ires(. . . , 0) atoms in A,
(a) Then for each f ires(ti , 0) ∈ A
i. enabled(ti , 0) ∈ A – from rule a1 and supported rule proposition
ii. Then notenabled(ti , 0) < A – from rule e2 and supported rule proposition
iii. Then body(e3) must not hold in A and body(e4) must not hold in A – from rules
e3, e4 and forced atom proposition
iv. Then q ≮ ni ≡ q ≥ ni in e3 for all {holds(p, q, 0), ptarc(p, ti , ni , 0)} ⊆ A – from
e3, forced atom proposition, and given facts (holds(p, q, 0) ∈ A, ptarc(p, ti , ni , 0) ∈
A)
v. And q ni ≡ q < ni in e4 for all {holds(p, q, 0), iptarc(p, ti , ni , 0)} ⊆ A, ni = 1;
q > ni ≡ q = 0 – from e4, forced atom proposition, given facts (holds(p, q, 0) ∈
A, iptarc(p, ti , 1, 0) ∈ A), and q is a positive integer
vi. Then (∀p ∈ •ti , M0 (p) ≥ W(p, ti )) ∧ (∀p ∈ I(ti ), M0 (p) = 0) – from
A. holds(p, q, 0) ∈ A represents q = M0 (p) – rule i1 construction
B. ptarc(p, ti , ni , 0) ∈ A represents ni = W(p, ti ) – rule f 7 construction; or
it represents ni = M0 (p) – rule f 9 construction; the construction of f 9
ensures that notenabled(t, 0) is never true due to the reset arc
C. definition 11 of preset •ti in PN
vii. Then ti is enabled and can fire in PN, as a result it can belong to T 0 – from
definition 26 of enabled transition
(b) And consumesmore < A, since A is an answer set of Π3 – from rule a4 and supported
rule proposition
i. Then @consumesmore(p, 0) ∈ A – from rule a3 and supported rule proposition
ii. Then @{holds(p, q, 0), tot decr(p, q1, 0)} ⊆ A : q1 > q in body(a2) – from a2
and forced atom proposition
P
P
iii. Then @p : ( ti ∈{t0 ,...,tx },p∈•ti W(p, ti ) + ti ∈{t0 ,...,tx },p∈R(ti ) M0 (p)) > M0 (p) – from
the following
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
221
A. holds(p, q, 0) represents q = M0 (p) – from rule i1 encoding, given
B. tot decr(p, q1, 0) ∈ A if {del(p, q10 , t0 , 0), . . . , del(p, q1 x , t x , 0)} ⊆ A, where
q1 = q10 + · · · + q1 x – from r4 and forced atom proposition
C. del(p, q1i , ti , 0) ∈ A if { f ires(ti , 0), ptarc(p, ti , q1i , 0)} ⊆ A – from r7 and
supported rule proposition
D. del(p, q1i , ti , 0) either represents removal of q1i = W(p, ti ) tokens from
p ∈ •ti ; or it represents removal of q1i = M0 (p) tokens from p ∈ R(ti )–
from rule r7, supported rule proposition, and definition 22 of transition
execution in PN
iv. Then the set of transitions in T 0 do not conflict – by the definition 23 of conflicting transitions
(c) And for each enabled(t j , 0) ∈ A and f ires(t j , 0) < A,
could not have(t j , 0) ∈ A, since A is an answer set of Π3 - from rule a6 and supported rule proposition
i. Then {enabled(t j , 0), holds(s, qq, 0), ptarc(s, t j , q, 0),
tot decr(s, qqq, 0)} ⊆ A, such that q > qq − qqq and f ires(t j , 0) < A - from rule
a7 and supported rule proposition
P
P
ii. Then for an s ∈ •t j ∪R(t j ), q > M0 (s)−( ti ∈T0 ,s∈•ti W(s, ti )+ ti ∈T0 ,s∈R(ti ) M0 (s)),
where q = W(s, t j ) if s ∈ •t j , or M0 (s) otherwise - from the following:
A. ptarc(s, ti , q, 0) represents q = W(s, ti ) if (s, ti ) ∈ E − or q = M0 (s) if s ∈
R(ti ) – from rule f 7, f 9 construction
B. holds(s, qq, 0) represents qq = M0 (s) – from i1 construction
C. tot decr(s, qqq, 0) ∈ A if {del(s, qqq0 , t0 , 0), . . . , del(s, qqq x , t x , 0)} ⊆ A –
from rule r4 construction and supported rule proposition
D. del(s, qqqi , ti , 0) ∈ A if { f ires(ti , 0), ptarc(s, ti , qqqi , 0)} ⊆ A – from rule r7
and supported rule proposition
E. del(s, qqqi , ti , 0) represents qqqi = W(s, ti ) : ti ∈ T 0 , (s, ti ) ∈ E − , or qqqi =
M0 (ti ) : ti ∈ T 0 , s ∈ R(ti ) – from rule f 7, f 9 construction
P
F. tot decr(q, qqq, 0) represents ti ∈T0 ,s∈•ti W(s, ti ) +
P
ti ∈T 0 ,s∈R(ti ) M0 (s) – from (C,D,E) above
iii. Then firing T 0 ∪ {t j } would have required more tokens than are present at its
source place s ∈ •t j ∪ R(t j ). Thus, T 0 is a maximal set of transitions that can
simultaneously fire.
(d) And for each reset transition tr with enabled(tr , 0) ∈ A, f ires(tr , 0) ∈ A, since A is
an answer set of Π2 - from rule f 10 and supported rule proposition
i. Then, the firing set T 0 satisfies the reset-transition requirement of definition 21
(firing set)
(e) Then {t0 , . . . , t x } = T 0 – using 1(a),1(b),1(d) above; and using 1(c) it is a maximal
firing set
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
222
2. We show M1 is produced by firing T 0 in M0 . Let holds(p, q, 1) ∈ A
(a) Then {holds(p, q1, 0), tot incr(p, q2, 0), tot decr(p, q3, 0)} ⊆ A : q = q1 + q2 − q3 –
from rule r5 and supported rule proposition
(b) Then holds(p, q1, 0) ∈ A represents q1 = M0 (p) – given, rule i1 construction;
and {add(p, q20 , t0 , 0), . . . , add(p, q2 j , t j , 0)} ⊆ A : q20 + · · · + q2 j = q2 and
{del(p, q30 , t0 , 0), . . . , del(p, q3l , tl , 0)} ⊆ A : q30 + · · · + q3l = q3 – rules r3, r4
and supported rule proposition, respectively
(c) Then { f ires(t0 , 0), . . . , f ires(t j , 0)} ⊆ A and { f ires(t0 , 0), . . . , f ires(tl , 0)}
⊆ A – rules r6, r7 and supported rule proposition, respectively
(d) Then { f ires(t0 , 0), . . . , f ires(t j , 0)}∪{ f ires(t0 , 0), . . . , f ires(tl , 0)} ⊆ A = { f ires(t0 , 0), . . . ,
f ires(t x , 0)} ⊆ A – set union of subsets
(e) Then for each f ires(t x , 0) ∈ A we have t x ∈ T 0 – already shown in item 1 above
P
P
(f) Then q = M0 (p) + tx ∈T0 ∧p∈tx • W(t x , p) − ( tx ∈T0 ∧p∈•tx W(p, t x ) +
P
t x ∈T 0 ∧p∈R(t x ) M0 (p)) – from (2b) above and the following
i. Each add(p, q j , t j , 0) ∈ A represents q j = W(t j , p) for p ∈ t j • – rule r6 encoding, and definition 22 of transition execution in PN
ii. Each del(p, ty , qy , 0) ∈ A represents either qy = W(p, ty ) for p ∈ •ty , or qy =
M0 (p) for p ∈ R(ty ) – from rule r7, f 7 encoding and definition 22 of transition
execution in PN; or from rule r7, f 9 encoding and definition of reset arc in PN
P
iii. Each tot incr(p, q2, 0) ∈ A represents q2 = tx ∈T0 ∧p∈tx • W(t x , p) – aggregate
assignment atom semantics in rule r3
P
P
iv. Each tot decr(p, q3, 0) ∈ A represents q3 = tx ∈T0 ∧p∈•tx W(p, t x )+ tx ∈T0 ∧p∈R(tx ) M0 (p)
– aggregate assignment atom semantics in rule r4
(g) Then, M1 (p) = q – since holds(p, q, 1) ∈ A encodes q = M1 (p) – from construction
Inductive Step: Let k > 0, and Mk is a valid marking in X for PN, show (1) T k is a valid firing
set for Mk , and (2) firing T k in Mk produces marking Mk+1 .
1. We show that T k is a valid firing set in Mk . Let { f ires(t0 , k), . . . , f ires(t x , k)} be the set of
all f ires(. . . , k) atoms in A,
(a) Then for each f ires(ti , k) ∈ A
i. enabled(ti , k) ∈ A – from rule a1 and supported rule proposition
ii. Then notenabled(ti , k) < A – from rule e2 and supported rule proposition
iii. Then body(e3) must not hold in A and body(e4) must not hold in A – from rule
e3, e4 and forced atom proposition
iv. Then q ≮ ni ≡ q ≥ ni in e3 for all {holds(p, q, k), ptarc(p, ti , ni , k)} ⊆ A – from
e3, forced atom proposition, and given facts (holds(p, q, k) ∈ A, ptarc(p, ti , ni , k) ∈
A)
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
223
v. And q ni ≡ q < ni in e4 for all {holds(p, q, k), iptarc(p, ti , ni , k)} ⊆ A, ni =
1; q > ni ≡ q = 0 – from e4, forced atom proposition, given facts (holds(p, q, k) ∈
A, iptarc(p, ti , 1, k) ∈ A), and q is a positive integer
vi. Then (∀p ∈ •ti , Mk (p) ≥ W(p, ti )) ∧ (∀p ∈ I(ti ), Mk (p) = 0) – from
A. holds(p, q, k) ∈ A represents q = Mk (p) – inductive assumption, given
B. ptarc(p, ti , ni , k) ∈ A represents ni = W(p, ti ) – rule f 7 construction; or
it represents ni = Mk (p) – rule f 9 construction; the construction of f 9
ensures that notenabled(t, 0) is never true due to the reset arc
C. definition 11 of preset •ti in PN
vii. Then ti is enabled and can fire in PN, as a result it can belong to T k – from
definition 26 of enabled transition
(b) And consumesmore < A, since A is an answer set of Π3 – from rule a4 and supported
rule proposition
i. Then @consumesmore(p, k) ∈ A – from rule a3 and supported rule proposition
ii. Then @{holds(p, q, k), tot decr(p, q1, k)} ⊆ A : q1 > q in body(a2) – from a2
and forced atom proposition
P
P
iii. Then @p : ( ti ∈{t0 ,...,tx },p∈•ti W(p, ti ) + ti ∈{t0 ,...,tx },p∈R(ti ) Mk (p)) > Mk (p) – from
the following
A. holds(p, q, k) represents q = Mk (p) – inductive assumption, construction
B. tot decr(p, q1, k) ∈ A if {del(p, q10 , t0 , k), . . . , del(p, q1 x , t x , k)} ⊆ A, where
q1 = q10 + · · · + q1 x – from r4 and forced atom proposition
C. del(p, q1i , ti , k) ∈ A if { f ires(ti , k), ptarc(p, ti , q1i , k)} ⊆ A – from r7 and
supported rule proposition
D. del(p, q1i , ti , k) either represents removal of q1i = W(p, ti ) tokens from
p ∈ •ti ; or it represents removal of q1i = Mk (p) tokens from p ∈ R(ti )–
from rule r7, supported rule proposition, and definition 22 of transition
execution in PN
iv. Then T k does not contain conflicting transitions – by the definition 23 of conflicting transitions
(c) And for each enabled(t j , k) ∈ A and f ires(t j , k) < A, could not have(t j , k) ∈ A,
since A is an answer set of Π3 - from rule a6 and supported rule proposition
i. Then {enabled(t j , k), holds(s, qq, k), ptarc(s, t j , q, k),
tot decr(s, qqq, k)} ⊆ A, such that q > qq − qqq and f ires(t j , k) < A - from rule
a7 and supported rule proposition
P
P
ii. Then for an s ∈ •t j ∪R(t j ), q > Mk (s)−( ti ∈Tk ,s∈•ti W(s, ti )+ ti ∈Tk ,s∈R(ti ) Mk (s)),
where q = W(s, t j ) if s ∈ •t j , or Mk (s) otherwise - from the following:
A. ptarc(s, ti , q, k) represents q = W(s, ti ) if (s, ti ) ∈ E − or q = Mk (s) if s ∈
R(ti ) – from rule f 7, f 9 construction
B. holds(s, qq, k) represents qq = Mk (s) – construction
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
224
C. tot decr(s, qqq, k) ∈ A if {del(s, qqq0 , t0 , k), . . . , del(s, qqq x , t x , k)}
⊆ A – from rule r4 construction and supported rule proposition
D. del(s, qqqi , ti , k) ∈ A if { f ires(ti , k), ptarc(s, ti , qqqi , k)} ⊆ A – from rule r7
and supported rule proposition
E. del(s, qqqi , ti , k) represents qqqi = W(s, ti ) : ti ∈ T k , (s, ti ) ∈ E − , or qqqi =
Mk (ti ) : ti ∈ T k , s ∈ R(ti ) – from rule f 7, f 9 construction
P
F. tot decr(q, qqq, k) represents ti ∈T k ,s∈•ti W(s, ti ) +
P
ti ∈T k ,s∈R(ti ) Mk (s) – from (C,D,E) above
iii. Then firing T k ∪ {t j } would have required more tokens than are present at its
source place s ∈ •t j ∪ R(t j ). Thus, T k is a maximal set of transitions that can
simultaneously fire.
(d) And for each reset transition tr with enabled(tr , k) ∈ A, f ires(tr , k) ∈ A, since A is
an answer set of Π2 - from rule f 10 and supported rule proposition
i. Then the firing set T k satisfies the reset transition requirement of definition 21
(firing set)
(e) Then {t0 , . . . , t x } = T k – using 1(a),1(b), 1(d) above; and using 1(c) it is a maximal
firing set
2. We show that Mk+1 is produced by firing T k in Mk . Let holds(p, q, k + 1) ∈ A
(a) Then {holds(p, q1, k), tot incr(p, q2, k), tot decr(p, q3, k)} ⊆ A : q = q1 + q2 − q3 –
from rule r5 and supported rule proposition
(b) Then, holds(p, q1, k) ∈ A represents q1 = Mk (p) – inductive assumption, construction ; and {add(p, q20 , t0 , k), . . . , add(p, q2 j , t j , k)} ⊆ A : q20 + · · · + q2 j = q2 and
{del(p, q30 , t0 , k), . . . , del(p, q3l , tl , k)} ⊆ A : q30 + · · · + q3l = q3 – rules r3, r4 and
supported rule proposition, respectively
(c) Then { f ires(t0 , k), . . . , f ires(t j , k)} ⊆ A and { f ires(t0 , k), . . . , f ires(tl , k)}
⊆ A – rules r6, r7 and supported rule proposition, respectively
(d) Then { f ires(t0 , k), . . . , f ires(t j , k)}∪{ f ires(t0 , k), . . . , f ires(tl , k)} ⊆ A = { f ires(t0 , k), . . . ,
f ires(t x , k)} ⊆ A – set union of subsets
(e) Then for each f ires(t x , k) ∈ A we have t x ∈ T k – already shown in item 1 above
P
P
(f) Then q = Mk (p) + tx ∈T0 ∧p∈tx • W(t x , p) − ( tx ∈Tk ∧p∈•tx W(p, t x ) +
P
t x ∈T k ∧p∈R(t x ) Mk (p)) – from (2b) above and the following
i. Each add(p, q j , t j , k) ∈ A represents q j = W(t j , p) for p ∈ t j • – rule r6 encoding, and definition 22 of transition execution in PN
ii. Each del(p, ty , qy , k) ∈ A represents either qy = W(p, ty ) for p ∈ •ty , or qy =
Mk (p) for p ∈ R(ty ) – from rule r7, f 7 encoding and definition 22 of transition
execution in PN; or from rule r7, f 9 encoding and definition of reset arc in PN
P
iii. Each tot incr(p, q2, k) ∈ A represents q2 = tx ∈Tk ∧p∈tx • W(t x , p) – aggregate
assignment atom semantics in rule r3
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
225
P
P
iv. Each tot decr(p, q3, 0) ∈ A represents q3 = tx ∈Tk ∧p∈•tx W(p, t x )+ tx ∈Tk ∧p∈R(tx ) Mk (p)
– aggregate assignment atom semantics in rule r4
(g) Then, Mk+1 (p) = q – since holds(p, q, k + 1) ∈ A encodes q = Mk+1 (p) – from
construction
As a result, for any n > k, T n will be a valid firing set for Mn and Mn+1 will be its target marking.
Conclusion: Since both (I) and (II) hold, X = M0 , T 0 , M1 , . . . , Mk , T k+1 is an execution sequence of PN(P, T, E, W, R) (w.r.t M0 ) iff there is an answer set A of Π3 (PN, M0 , k, ntok) such
that (A.5) and (A.6) hold.
A.5
Proof of Proposition 5
Let PN = (P, T, E, W, R, I, Q, QW) be a Petri Net, M0 be its initial marking and let Π4 (PN, M0 , k, ntok)
be the ASP encoding of PN and M0 over a simulation length k, with maximum ntok tokens on
any place node, as defined in section 2.7. Then X = M0 , T 0 , M1 , . . . , Mk , T k , Mk+1 is an execution sequence of PN (w.r.t. M0 ) iff there is an answer set A of Π4 (PN, M0 , k, ntok) such
that:
{ f ires(t, ts) : t ∈ T ts , 0 ≤ ts ≤ k} = { f ires(t, ts) : f ires(t, ts) ∈ A}
(A.9)
{holds(p, q, ts) : p ∈ P, q = Mts (p), 0 ≤ ts ≤ k + 1}
= {holds(p, q, ts) : holds(p, q, ts) ∈ A}
(A.10)
We prove this by showing that:
(I) Given an execution sequence X, we create a set A such that it satisfies (A.9) and (A.10)
and show that A is an answer set of Π4
(II) Given an answer set A of Π4 , we create an execution sequence X such that (A.9) and
(A.10) are satisfied.
First we show (I): Given PN and an execution sequence X of PN, we create a set A as a union
of the following sets:
1. A1 = {num(n) : 0 ≤ n ≤ ntok}
2. A2 = {time(ts) : 0 ≤ ts ≤ k}
3. A3 = {place(p) : p ∈ P}
4. A4 = {trans(t) : t ∈ T }
5. A5 = {ptarc(p, t, n, ts) : (p, t) ∈ E − , n = W(p, t), 0 ≤ ts ≤ k}, where E − ⊆ E
6. A6 = {tparc(t, p, n, ts) : (t, p) ∈ E + , n = W(t, p), 0 ≤ ts ≤ k}, where E + ⊆ E
7. A7 = {holds(p, q, 0) : p ∈ P, q = M0 (p)}
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
226
8. A8 = {notenabled(t, ts) : t ∈ T, 0 ≤ ts ≤ k, (∃p ∈ •t, Mts (p) < W(p, t)) ∨ (∃p ∈
I(t), Mts (p) , 0) ∨ (∃(p, t) ∈ Q, Mts (p) < QW(p, t))}
per definition 28 (enabled transition)
9. A9 = {enabled(t, ts) : t ∈ T, 0 ≤ ts ≤ k, (∀p ∈ •t, W(p, t) ≤ Mts (p)) ∧ (∀p ∈ I(t), Mts (p) =
0) ∧ (∀(p, t) ∈ Q, Mts (p) ≥ QW(p, t))}
per definition 28 (enabled transition)
10. A10 = { f ires(t, ts) : t ∈ T ts , 0 ≤ ts ≤ k}
per definition 28 (enabled transitions), only an enabled transition may fire
11. A11 = {add(p, q, t, ts) : t ∈ T ts , p ∈ t•, q = W(t, p), 0 ≤ ts ≤ k}
per definition 22 (transition execution)
12. A12 = {del(p, q, t, ts) : t ∈ T ts , p ∈ •t, q = W(p, t), 0 ≤ ts ≤ k} ∪ {del(p, q, t, ts) : t ∈
T ts , p ∈ R(t), q = Mts (p), 0 ≤ ts ≤ k}
per definition 22 (transition execution)
P
13. A13 = {tot incr(p, q, ts) : p ∈ P, q = t∈Tts ,p∈t• W(t, p), 0 ≤ ts ≤ k}
per definition 24 (firing set execution)
P
P
14. A14 = {tot decr(p, q, ts) : p ∈ P, q = t∈Tts ,p∈•t W(p, t) + t∈Tts ,p∈R(t) Mts (p), 0 ≤ ts ≤ k}
per definition 24 (firing set execution)
P
P
15. A15 = {consumesmore(p, ts) : p ∈ P, q = Mts (p), q1 = t∈Tts ,p∈•t W(p, t)+ t∈T ts ,p∈R(t) Mts (p), q1 >
q, 0 ≤ ts ≤ k}
per definition 23 (conflicting transitions)
P
P
16. A16 = {consumesmore : ∃p ∈ P : q = Mts (p), q1 = t∈T ts ,p∈•t W(p, t)+ t∈Tts ,p∈R(t) (Mts (p)), q1 >
q, 0 ≤ ts ≤ k}
per definition 23 (conflicting transitions)
17. A17 = {could not have(t, ts) : t ∈ T, (∀p ∈ •t, W(p, t) ≤ Mts (p)), t < T ts , (∃p ∈ •t :
P
P
W(p, t) > Mts (p) − ( t0 ∈Tts ,p∈•t0 W(p, t0 ) + t0 ∈T ts ,p∈R(t0 ) Mts (p)), 0 ≤ ts ≤ k}
per the maximal firing set semantics
18. A18 = {holds(p, q, ts + 1) : p ∈ P, q = Mts+1 (p), 0 ≤ ts < k},
P
where Mts+1 (p) = Mts (p) − ( t∈T ts ,p∈•t W(p, t) +
P
P
t∈T ts ,p∈R(t) Mts (p))+ t∈T ts ,p∈t• W(t, p)
according to definition 21 (firing set execution)
19. A19 = {ptarc(p, t, n, ts) : p ∈ R(t), n = Mts (p), n > 0, 0 ≤ ts ≤ k}
20. A20 = {iptarc(p, t, 1, ts) : p ∈ P, 0 ≤ ts < k}
21. A21 = {tptarc(p, t, n, ts) : (p, t) ∈ Q, n = QW(p, t), 0 ≤ ts ≤ k}
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
227
We show that A satisfies (A.7) and (A.8), and A is an answer set of Π4 .
A satisfies (A.7) and (A.8) by its construction above. We show A is an answer set of Π4 by
splitting. We split lit(Π4 ) into a sequence of 7k + 9 sets:
• U0 = head( f 1) ∪ head( f 2) ∪ head( f 5) ∪ head( f 6) ∪ head(i1) = {place(p) : p ∈ P} ∪
{trans(t) : t ∈ T } ∪ {time(0), . . . , time(k)} ∪ {num(0), . . . , num(ntok)} ∪ {holds(p, q, 0) : p ∈
P, q = M0 (p)}
• U7k+1 = U7k+0 ∪head( f 7)ts=k ∪head( f 8)ts=k ∪head( f 9)ts=k ∪head( f 11)ts=k ∪head( f 12)ts=k =
U7k+0 ∪ {ptarc(p, t, n, k) : (p, t) ∈ E − , n = W(p, t)} ∪
{tparc(t, p, n, k) : (t, p) ∈ E + , n = W(t, p)}∪ {ptarc(p, t, n, k) : p ∈ R(t), n = Mk (p), n >
0}∪ {iptarc(p, t, 1, k) : p ∈ I(t)}∪ {tptarc(p, t, n, k) : (p, t) ∈ Q, n = QW(p, t)}
• U7k+2 = U7k+1 ∪ head(e3)ts=k = U7k+1 ∪ {notenabled(t, k) : t ∈ T }
• U7k+3 = U7k+2 ∪ head(e2)ts=k = U7k+2 ∪ {enabled(t, k) : t ∈ T }
• U7k+4 = U7k+3 ∪ head(a1)ts=k = U7k+3 ∪ { f ires(t, k) : t ∈ T }
• U7k+5 = U7k+4 ∪ head(r6)ts=k ∪ head(r7)ts=k = U7k+4 ∪ {add(p, q, t, k) : p ∈ P, t ∈ T, q =
W(t, p)} ∪ {del(p, q, t, k) : p ∈ P, t ∈ T, q = W(p, t)} ∪ {del(p, q, t, k) : p ∈ P, t ∈ T, q =
Mk (p)}
• U7k+6 = U7k+5 ∪ head(r3)ts=k ∪ head(r4)ts=k = U7k+5 ∪ {tot incr(p, q, k) : p ∈ P, 0 ≤ q ≤
ntok} ∪ {tot decr(p, q, k) : p ∈ P, 0 ≤ q ≤ ntok}
• U7k+7 = U7k+6 ∪ head(r5)ts=k ∪ head(a2)ts=k ∪ head(a7)ts=k = U7k+6 ∪
{consumesmore(p, k) : p ∈ P}∪{holds(p, q, k+1) : p ∈ P, 0 ≤ q ≤ ntok}∪{could not have(t, k) :
t ∈ T}
• U7k+8 = U7k+7 ∪ head(a3) = U7k+7 ∪ {consumesmore}
where head(ri )ts=k are head atoms of ground rule ri in which ts = k. We write Ats=k
=
i
{a(. . . , ts) : a(. . . , ts) ∈ Ai , ts = k} as short hand for all atoms in Ai with ts = k. Uα , 0 ≤
α ≤ 7k + 8 form a splitting sequence, since each Ui is a splitting set of Π4 , and hUα iα<µ is a
S
monotone continuous sequence, where U0 ⊆ U1 · · · ⊆ U8(k+1) and α<µ Uα = lit(Π4 ).
We compute the answer set of Π4 using the splitting sets as follows:
1. botU0 (Π4 ) = f 1 ∪ f 2 ∪ f 5 ∪ i1 ∪ f 6 and X0 = A1 ∪ · · · ∪ A4 ∪ A7 (= U0 ) is its answer set
– using forced atom proposition
2. evalU0 (botU1 (Π4 ) \ botU0 (Π4 ), X0 ) = {ptarc(p, t, q, 0):-.|q = W(p, t)} ∪
{tparc(t, p, q, 0):-.|q = W(t, p)} ∪ {ptarc(p, t, q, 0):-.|q = M0 (p)} ∪
{iptarc(p, t, 1, 0):-.}∪{tptarc(p, t, q, 0):-.|q = QW(p, t)}. Its answer set X1 = Ats=0
∪Ats=0
∪
5
6
ts=0
ts=0
ts=0
A19 ∪A20 ∪A21 – using forced atom proposition and construction of A5 , A6 , A19 , A20 , A21 .
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
228
3. evalU1 (botU2 (Π4 ) \ botU1 (Π4 ), X0 ∪ X1 ) = {notenabled(t, 0):-.|({trans(t),
ptarc(p, t, n, 0), holds(p, q, 0)} ⊆ X0 ∪ X1 , where q < n) or ({notenabled(t, 0):-.|
({trans(t), iptarc(p, t, n2, 0), holds(p, q, 0)} ⊆ X0 ∪ X1 , where q ≥ n2}) or
({trans(t), tptarc(p, t, n3, 0), holds(p, q, 0)} ⊆ X0 ∪ X1 , where q < n3)}. Its answer set
X2 = Ats=0
– using forced atom proposition and construction of A8 .
8
(a) where, q = M0 (p), and n = W(p, t) for an arc (p, t) ∈ E − – by construction of i1 and
f 7 in Π4 , and
(b) in an arc (p, t) ∈ E − , p ∈ •t (by definition 11 of preset)
(c) n2 = 1 – by construction of iptarc predicates in Π4 , meaning q ≥ n2 ≡ q ≥ 1 ≡ q >
0,
(d) tptarc(p, t, n3, 0) represents n3 = QW(p, t), where (p, t) ∈ Q
(e) thus, notenabled(t, 0) ∈ X1 represents (∃p ∈ •t : M0 (p) < W(p, t)) ∨ (∃p ∈ I(t) :
M0 (p) > 0) ∨ (∃(p, t) ∈ Q : Mts (p) < QW(p, t)).
4. evalU2 (botU3 (Π4 ) \ botU2 (Π4 ), X0 ∪ · · · ∪ X2 ) = {enabled(t, 0):-.|trans(t) ∈ X0 ∪ · · · ∪
– using forced atom
X2 , notenabled(t, 0) < X0 ∪ · · · ∪ X2 }. Its answer set is X3 = Ats=0
9
proposition and construction of A9 .
(a) since an enabled(t, 0) ∈ X3 if @ notenabled(t, 0) ∈ X0 ∪ · · · ∪ X2 ; which is equivalent
to (@p ∈ •t : M0 (p) < W(p, t)) ∧ (@p ∈ I(t) : M0 (p) > 0) ∧ (@(p, t) ∈ Q : M0 (p) <
QW(p, t)) ≡ (∀p ∈ •t : M0 (p) ≥ W(p, t)) ∧ (∀p ∈ I(t) : M0 (p) = 0).
5. evalU3 (botU4 (Π4 ) \ botU3 (Π4 ), X0 ∪ · · · ∪ X3 ) = {{ f ires(t, 0)}:-.|enabled(t, 0)
holds in X0 ∪ · · · ∪ X3 }. It has multiple answer sets X4.1 , . . . , X4.n , corresponding to elements of power set of f ires(t, 0) atoms in evalU3 (...) – using supported rule proposition.
Since we are showing that the union of answer sets of Π4 determined using splitting is
equal to A, we only consider the set that matches the f ires(t, 0) elements in A and call it
X4 , ignoring the rest. Thus, X4 = Ats=0
10 , representing T 0 .
(a) in addition, for every t such that enabled(t, 0) ∈ X0 ∪ · · · ∪ X3 , R(t) , ∅; f ires(t, 0) ∈
X4 – per definition 21 (firing set); requiring that a reset transition is fired when
enabled
(b) thus, the firing set T 0 will not be eliminated by the constraint f 10
6. evalU4 (botU5 (Π4 ) \ botU4 (Π4 ), X0 ∪ · · · ∪ X4 ) = {add(p, n, t, 0):-.|{ f ires(t, 0),
tparc(t, p, n, 0)} ⊆ X0 ∪ · · · ∪ X4 } ∪ {del(p, n, t, 0):-.|{ f ires(t, 0), ptarc(p, t, n, 0)}
ts=0
⊆ X0 ∪ · · · ∪ X4 }. It’s answer set is X5 = Ats=0
11 ∪ A12 – using forced atom proposition
and definitions of A11 and A12 .
(a) where, each add atom is equivalent to n = W(t, p), p ∈ t•,
(b) and each del atom is equivalent to n = W(p, t), p ∈ •t; or n = Mk (p), p ∈ R(t),
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
229
(c) representing the effect of transitions in T 0 – by construction
7. evalU5 (botU6 (Π4 ) \ botU5 (Π4 ), X0 ∪ · · · ∪ X5 ) = {tot incr(p, qq, 0):-.|
P
P
qq = add(p,q,t,0)∈X0 ∪···∪X5 q} ∪ {tot decr(p, qq, 0):-.|qq = del(p,q,t,0)∈X0 ∪···∪X5 q}. It’s ants=0
swer set is X6 = Ats=0
13 ∪ A14 – using forced atom proposition and definitions of A13 and
A14 .
P
(a) where, each tot incr(p, qq, 0), qq = add(p,q,t,0)∈X0 ∪...X5 q
P
≡ qq = t∈X4 ,p∈t• W(p, t),
P
(b) and each tot decr(p, qq, 0), qq = del(p,q,t,0)∈X0 ∪...X5 q
P
P
≡ qq = t∈X4 ,p∈•t W(t, p) + t∈X4 ,p∈R(t) Mk (p),
(c) represent the net effect of transitions in T 0 – by construction
8. evalU6 (botU7 (Π4 ) \ botU6 (Π4 ), X0 ∪ · · · ∪ X6 ) = {consumesmore(p, 0):-.|
{holds(p, q, 0), tot decr(p, q1, 0)} ⊆ X0 ∪ · · · ∪ X6 , q1 > q} ∪ {holds(p, q, 1):-., |
{holds(p, q1, 0), tot incr(p, q2, 0), tot decr(p, q3, 0)} ⊆ X0 ∪ · · · ∪ X6 , q = q1 + q2 − q3} ∪
{could not have(t, 0):-.|{enabled(t, 0), ptarc(s, t, q), holds(s, qq, 0),
tot decr(s, qqq, 0)} ⊆ X0 ∪ · · · ∪ X6 , f ires(t, 0) < (X0 ∪ · · · ∪ X6 ), q > qq − qqq}. It’s
ts=0
answer set is X7 = Ats=0
∪ Ats=0
17 ∪ A18 – using forced atom proposition and definitions
15
of A15 , A17 , A18 , A9 .
(a) where, consumesmore(p, 0) represents ∃p : q = M0 (p), q1 =
P
P
t∈T 0 ,p∈•t W(p, t) + t∈T 0 ,p∈R(t) M0 (p), q1 > q, indicating place p will be over consumed if T 0 is fired, as defined in definition 23 (conflicting transitions),
(b) holds(p, q, 1) represents q = M1 (p) – by construction of Π4 ,
(c) and could not have(t, 0) represents enabled transition t ∈ T 0 that could not fire due
to insufficient tokens
(d) X7 does not contain could not have(t, 0), when enabled(t, 0) ∈ X0 ∪ · · · ∪ X6 and
f ires(t, 0) < X0 ∪ · · · ∪ X6 due to construction of A, encoding of a7 and its body
atoms. As a result it is not eliminated by the constraint a6
..
.
9. evalU7k+0 (botU7k+1 (Π4 ) \ botU7k+0 (Π4 ), X0 ∪ · · · ∪ X7k+0 ) = {ptarc(p, t, q, k):-.|q = W(p, t)} ∪
{tparc(t, p, q, k):-.|q = W(t, p)} ∪ {ptarc(p, t, q, k):-.|q = Mk (p)} ∪ {iptarc(p, t, 1, k):-.}. Its
ts=k
answer set X7k+1 = Ats=k
∪ Ats=k
∪ Ats=k
19 ∪ A20 – using forced atom proposition and
5
6
construction of A5 , A6 , A19 , A20 .
10. evalU7k+1 (botU7k+2 (Π4 )\botU7k+1 (Π4 ), X0 ∪· · ·∪X7k+1 ) = {notenabled(t, k):-.| ({trans(t), ptarc(p, t, n, k), holds(p, q, k
X0 ∪ · · · ∪ X7k+1 , where q < n) or
{notenabled(t, k):-.|({trans(t), iptarc(p, t, n2, k), holds(p, q, k)} ⊆ X0 ∪ · · · ∪ X7k+1 ,
where q ≥ n2}}. Its answer set X7k+2 = Ats=k
– using forced atom proposition and con8
struction of A8 .
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
230
(a) where, q = Mk (p), and n = W(p, t) for an arc (p, t) ∈ E − – by construction of holds
and ptarc predicates in Π4 , and
(b) in an arc (p, t) ∈ E − , p ∈ •t (by definition 11 of preset)
(c) n2 = 1 – by construction of iptarc predicates in Π4 , meaning q ≥ n2 ≡ q ≥ 1 ≡ q >
0,
(d) thus, notenabled(t, k) ∈ X7k+1 represents (∃p ∈ •t : Mk (p) < W(p, t)) ∨ (∃p ∈ I(t) :
Mk (p) > k).
11. evalU7k+2 (botU7k+3 (Π4 ) \ botU7k+2 (Π4 ), X0 ∪ · · · ∪ X7k+2 ) = {enabled(t, k):-.|
trans(t) ∈ X0 ∪ · · · ∪ X7k+2 ∧ notenabled(t, k) < X0 ∪ · · · ∪ X7k+2 }. Its answer set is
X7k+3 = Ats=k
– using forced atom proposition and construction of A9 .
9
(a) since an enabled(t, k) ∈ X7k+3 if @ notenabled(t, k) ∈ X0 ∪ · · · ∪ X7k+2 ; which is
equivalent to (@p ∈ •t : Mk (p) < W(p, t)) ∧ (@p ∈ I(t) : Mk (p) > k) ≡ (∀p ∈ •t :
Mk (p) ≥ W(p, t)) ∧ (∀p ∈ I(t) : Mk (p) = k).
12. evalU7k+3 (botU7k+4 (Π4 ) \ botU7k+3 (Π4 ), X0 ∪ · · · ∪ X7k+3 ) = {{ f ires(t, k)}:-.|
enabled(t, k)
holds in X0 ∪ · · · ∪ X7k+3 }. It has multiple answer sets X7k+4.1 , . . . , X7k+4.n , corresponding to elements of power set of f ires(t, k) atoms in evalU7k+3 (...) – using supported rule
proposition. Since we are showing that the union of answer sets of Π4 determined using
splitting is equal to A, we only consider the set that matches the f ires(t, k) elements in A
and call it X7k+4 , ignoring the rest. Thus, X7k+4 = Ats=k
10 , representing T k .
(a) in addition, for every t such that enabled(t, k) ∈ X0 ∪· · ·∪X7k+3 , R(t) , ∅; f ires(t, k) ∈
X7k+4 – per definition 21 (firing set); requiring that a reset transition is fired when
enabled
(b) thus, the firing set T k will not be eliminated by the constraint f 10
13. evalU7k+4 (botU7k+5 (Π4 ) \ botU7k+4 (Π4 ), X0 ∪ · · · ∪ X7k+4 ) = {add(p, n, t, k):-.|
{ f ires(t, k), tparc(t, p, n, k)} ⊆ X0 ∪ · · · ∪ X7k+4 } ∪ {del(p, n, t, k):-.|{ f ires(t, k),
ts=k
ptarc(p, t, n, k)} ⊆ X0 ∪ · · · ∪ X7k+4 }. It’s answer set is X7k+5 = Ats=k
11 ∪ A12 – using forced
atom proposition and definitions of A11 and A12 .
(a) where, each add atom is equivalent to n = W(t, p) : p ∈ t•,
(b) and each del atom is equivalent to n = W(p, t) : p ∈ •t; or n = Mk (p) : p ∈ R(t),
(c) representing the effect of transitions in T k
14. evalU7k+5 (botU7k+6 (Π4 ) \ botU7k+5 (Π4 ), X0 ∪ · · · ∪ X7k+5 ) =
P
{tot incr(p, qq, k):-.|qq = add(p,q,t,k)∈X0 ∪···∪X7k+5 q} ∪
P
ts=k
{tot decr(p, qq, k):-.|qq = del(p,q,t,k)∈X0 ∪···∪X7k+5 q}. It’s answer set is X7k+6 = Ats=k
13 ∪ A14
– using forced atom proposition and definitions of A13 and A14 .
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
231
P
(a) where, each tot incr(p, qq, k), qq = add(p,q,t,k)∈X0 ∪...X7k+5 q
P
≡ qq = t∈X7k+4 ,p∈t• W(p, t),
P
(b) and each tot decr(p, qq, k), qq = del(p,q,t,k)∈X0 ∪...X7k+5 q
P
P
≡ qq = t∈X7k+4 ,p∈•t W(t, p) + t∈X7k+4 ,p∈R(t) Mk (p),
(c) represent the net effect of transition in T k
15. evalU7k+6 (botU7k+7 (Π4 ) \ botU7k+6 (Π4 ), X0 ∪ · · · ∪ X7k+6 ) = {consumesmore(p, k):-.|
{holds(p, q, k), tot decr(p, q1, k)} ⊆ X0 ∪ · · · ∪ X7k+6 , q1 > q} ∪ {holds(p, q, 1):-., |
{holds(p, q1, k), tot incr(p, q2, k), tot decr(p, q3, k)} ⊆ X0 ∪ · · · ∪ X7k+6 , q = q1 + q2 −
q3} ∪ {could not have(t, k):-.|{enabled(t, k), ptarc(s, t, q), holds(s, qq, k),
tot decr(s, qqq, k)} ⊆ X0 ∪ · · · ∪ X7k+6 , f ires(t, k) < (X0 ∪ · · · ∪ X7k+6 ), q > qq − qqq}. It’s
ts=k
answer set is X7k+7 = Ats=k
∪ Ats=k
17 ∪ A18 – using forced atom proposition and definitions
15
of A15 , A17 , A18 , A9 .
(a) where, consumesmore(p, k) represents ∃p : q = Mk (p), q1 =
P
P
t∈T k ,p∈•t W(p, t) + t∈T k ,p∈R(t) Mk (p), q1 > q, indicating place p that will be over
consumed if T k is fired, as defined in definition 23 (conflicting transitions),
(b) holds(p, q, k + 1) represents q = Mk+1 (p) – by construction of Π4 ,
(c) and could not have(t, k) represents enabled transition t in T k that could not be fired
due to insufficient tokens
(d) X7k+7 does not contain could not have(t, k), when enabled(t, k) ∈ X0 ∪ · · · ∪ X7k+6
and f ires(t, k) < X0 ∪ · · · ∪ X7k+6 due to construction of A, encoding of a7 and its
body atoms. As a result it is not eliminated by the constraint a6
16. evalU7k+7 (botU7k+8 (Π4 ) \ botU7k+7 (Π4 ), X0 ∪ · · · ∪ X7k+7 ) = {consumesmore:-.|
{consumesmore(p, 0), . . . , consumesmore(p, k)} ∩ (X0 ∪ · · · ∪ X7k+7 ) , ∅}. It’s answer set
is X7k+8 = A16 – using forced atom proposition
(a) X7k+8 will be empty since none of consumesmore(p, 0), . . . ,
consumesmore(p, k) hold in X0 ∪ · · · ∪ X7k+8 due to the construction of A, encoding
of a2 and its body atoms. As a result, it is not eliminated by the constraint a4
The set X = X0 ∪ · · · ∪ X7k+8 is the answer set of Π4 by the splitting sequence theorem 15.
Each Xi , 0 ≤ i ≤ 7k + 8 matches a distinct portion of A, and X = A, thus A is an answer set of
Π4 .
Next we show (II): Given Π4 be the encoding of a Petri Net PN(P, T, E, W, R, I) with initial
marking M0 , and A be an answer set of Π4 that satisfies (A.7) and (A.8), then we can construct
X = M0 , T 0 , . . . , Mk , T k , Mk+1 from A, such that it is an execution sequence of PN.
We construct the X as follows:
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
232
1. Mi = (Mi (p0 ), . . . , Mi (pn )), where {holds(p0 , Mi (p0 ), i), . . . holds(pn , Mi (pn ), i)}
⊆ A, for 0 ≤ i ≤ k + 1
2. T i = {t : f ires(t, i) ∈ A}, for 0 ≤ i ≤ k
and show that X is indeed an execution sequence of PN. We show this by induction over k (i.e.
given Mk , T k is a valid firing set and its firing produces marking Mk+1 ).
Base case: Let k = 0, and M0 is a valid marking in X for PN, show (1) T 0 is a valid firing set
for M0 , and (2) firing of T 0 in M0 produces marking M1 .
1. We show T 0 is a valid firing set for M0 . Let { f ires(t0 , 0), . . . , f ires(t x , 0)} be the set of all
f ires(. . . , 0) atoms in A,
(a) Then for each f ires(ti , 0) ∈ A
i. enabled(ti , 0) ∈ A – from rule a1 and supported rule proposition
ii. Then notenabled(ti , 0) < A – from rule e2 and supported rule proposition
iii. Then either of body(e3), body(e4), or body(e5) must not hold in A – from rules
e3, e4, e5 and forced atom proposition
iv. Then q ≮ ni ≡ q ≥ ni in e3 for all {holds(p, q, 0), ptarc(p, ti , ni , 0)} ⊆ A – from
e3, forced atom proposition, and given facts (holds(p, q, 0) ∈ A, ptarc(p, ti , ni , 0) ∈
A)
v. And q ni ≡ q < ni in e4 for all {holds(p, q, 0), iptarc(p, ti , ni , 0)} ⊆ A, ni = 1;
q > ni ≡ q = 0 – from e4, forced atom proposition, given facts (holds(p, q, 0) ∈
A, iptarc(p, ti , 1, 0) ∈ A), and q is a positive integer
vi. And q ≮ ni ≡ q ≥ ni in e5 for all {holds(p, q, 0), tptarc(p, ti , ni , 0)} ⊆ A – from
e5, forced atom proposition, and given facts
vii. Then (∀p ∈ •ti , M0 (p) ≥ W(p, ti )) ∧ (∀p ∈ I(ti ), M0 (p) = 0) ∧ (∀(p, ti ) ∈
Q, M0 (p) ≥ QW(p, ti )) – from the following
A. holds(p, q, 0) ∈ A represents q = M0 (p) – rule i1 construction
B. ptarc(p, ti , ni , 0) ∈ A represents ni = W(p, ti ) – rule f 7 construction; or
it represents ni = M0 (p) – rule f 9 construction; the construction of f 9
ensures that notenabled(t, 0) is never true due to the reset arc
C. definition 11 of preset •ti in PN
D. definition 28 of enabled transition in PN
viii. Then ti is enabled and can fire in PN, as a result it can belong to T 0 – from
definition 28 of enabled transition
(b) And consumesmore < A, since A is an answer set of Π4 – from rule a4 and supported
rule proposition
i. Then @consumesmore(p, 0) ∈ A – from rule a3 and supported rule proposition
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
233
ii. Then @{holds(p, q, 0), tot decr(p, q1, 0)} ⊆ A : q1 > q in body(a2) – from a2
and forced atom proposition
P
P
iii. Then @p : ( ti ∈{t0 ,...,tx },p∈•ti W(p, ti ) + ti ∈{t0 ,...,tx },p∈R(ti ) M0 (p)) > M0 (p) – from
the following
A. holds(p, q, 0) represents q = M0 (p) – from rule i1 construction, given
B. tot decr(p, q1, 0) ∈ A if {del(p, q10 , t0 , 0), . . . , del(p, q1 x , t x , 0)} ⊆ A, where
q1 = q10 + · · · + q1 x – from r4 and forced atom proposition
C. del(p, q1i , ti , 0) ∈ A if { f ires(ti , 0), ptarc(p, ti , q1i , 0)} ⊆ A – from r7 and
supported rule proposition
D. del(p, q1i , ti , 0) represents removal of q1i = W(p, ti ) tokens from p ∈ •ti ; or
it represents removal of q1i = M0 (p) tokens from p ∈ R(ti )– from rule r7,
supported rule proposition, and definition 22 of transition execution in PN
iv. Then the set of transitions in T 0 do not conflict – by the definition 23 of conflicting transitions
(c) And for each enabled(t j , 0) ∈ A and f ires(t j , 0) < A,
could not have(t j , 0) ∈ A, since A is an answer set of Π4 - from rule a6 and supported rule proposition
i. Then {enabled(t j , 0), holds(s, qq, 0), ptarc(s, t j , q, 0),
tot decr(s, qqq, 0)} ⊆ A, such that q > qq − qqq and f ires(t j , 0) < A - from rule
a7 and supported rule proposition
P
P
ii. Then for an s ∈ •t j ∪R(t j ), q > M0 (s)−( ti ∈T0 ,s∈•ti W(s, ti )+ ti ∈T0 ,s∈R(ti ) M0 (s)),
where q = W(s, t j ) if s ∈ •t j , or M0 (s) otherwise - from the following:
A. ptarc(s, ti , q, 0) represents q = W(s, ti ) if (s, ti ) ∈ E − or q = M0 (s) if s ∈
R(ti ) – from rule f 7, f 9 construction
B. holds(s, qq, 0) represents qq = M0 (s) – from i1 construction
C. tot decr(s, qqq, 0) ∈ A if {del(s, qqq0 , t0 , 0), . . . , del(s, qqq x , t x , 0)} ⊆ A –
from rule r4 construction and supported rule proposition
D. del(s, qqqi , ti , 0) ∈ A if { f ires(ti , 0), ptarc(s, ti , qqqi , 0)} ⊆ A – from rule r7
and supported rule proposition
E. del(s, qqqi , ti , 0) represents qqqi = W(s, ti ) : ti ∈ T 0 , (s, ti ) ∈ E − , or qqqi =
M0 (ti ) : ti ∈ T 0 , s ∈ R(ti ) – from rule f 7, f 9 construction
P
F. tot decr(q, qqq, 0) represents ti ∈T0 ,s∈•ti W(s, ti ) +
P
ti ∈T 0 ,s∈R(ti ) M0 (s) – from (C,D,E) above
iii. Then firing T 0 ∪ {t j } would have required more tokens than are present at its
source place s ∈ •t j ∪ R(t j ). Thus, T 0 is a maximal set of transitions that can
simultaneously fire.
(d) And for each reset transition tr with enabled(tr , 0) ∈ A, f ires(tr , 0) ∈ A, since A is
an answer set of Π2 - from rule f 10 and supported rule proposition
i. Then, the firing set T 0 satisfies the reset-transition requirement of definition 21
(firing set)
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
234
(e) Then {t0 , . . . , t x } = T 0 – using 1(a),1(b),1(d) above; and using 1(c) it is a maximal
firing set
2. We show M1 is produced by firing T 0 in M0 . Let holds(p, q, 1) ∈ A
(a) Then {holds(p, q1, 0), tot incr(p, q2, 0), tot decr(p, q3, 0)} ⊆ A : q = q1 + q2 − q3 –
from rule r5 and supported rule proposition
(b) Then holds(p, q1, 0) ∈ A represents q1 = M0 (p) – given, rule i1 construction;
(c) Then {add(p, q20 , t0 , 0), . . . , add(p, q2 j , t j , 0)} ⊆ A : q20 + · · · + q2 j = q2 and
{del(p, q30 , t0 , 0), . . . , del(p, q3l , tl , 0)} ⊆ A : q30 + · · · + q3l = q3 – rules r3, r4
and supported rule proposition, respectively
(d) Then { f ires(t0 , 0), . . . , f ires(t j , 0)} ⊆ A and { f ires(t0 , 0), . . . , f ires(tl , 0)}
⊆ A – rules r6, r7 and supported rule proposition, respectively
(e) Then { f ires(t0 , 0), . . . , f ires(t j , 0)}∪{ f ires(t0 , 0), . . . , f ires(tl , 0)} ⊆ A = { f ires(t0 , 0), . . . ,
f ires(t x , 0)} ⊆ A – set union of subsets
(f) Then for each f ires(t x , 0) ∈ A we have t x ∈ T 0 – already shown in item 1 above
P
P
(g) Then q = M0 (p) + tx ∈T0 ∧p∈tx • W(t x , p) − ( tx ∈T0 ∧p∈•tx W(p, t x ) +
P
t x ∈T 0 ∧p∈R(t x ) M0 (p)) – from (2b) above and the following
i. Each add(p, q j , t j , 0) ∈ A represents q j = W(t j , p) for p ∈ t j • – rule r6 encoding, and definition 22 of transition execution in PN
ii. Each del(p, ty , qy , 0) ∈ A represents either qy = W(p, ty ) for p ∈ •ty , or qy =
M0 (p) for p ∈ R(ty ) – from rule r7, f 7 encoding and definition 22 of transition
execution in PN; or from rule r7, f 9 encoding and definition of reset arc in PN
P
iii. Each tot incr(p, q2, 0) ∈ A represents q2 = tx ∈T0 ∧p∈tx • W(t x , p) – aggregate
assignment atom semantics in rule r3
P
P
iv. Each tot decr(p, q3, 0) ∈ A represents q3 = tx ∈T0 ∧p∈•tx W(p, t x )+ tx ∈T0 ∧p∈R(tx ) M0 (p)
– aggregate assignment atom semantics in rule r4
(h) Then, M1 (p) = q – since holds(p, q, 1) ∈ A encodes q = M1 (p) – from construction
Inductive Step: Let k > 0, and Mk is a valid marking in X for PN, show (1) T k is a valid firing
set for Mk , and (2) firing T k in Mk produces marking Mk+1 .
1. We show that T k is a valid firing set in Mk . Let { f ires(t0 , k), . . . , f ires(t x , k)} be the set of
all f ires(. . . , k) atoms in A,
(a) Then for each f ires(ti , k) ∈ A
i. enabled(ti , k) ∈ A – from rule a1 and supported rule proposition
ii. Then notenabled(ti , k) < A – from rule e2 and supported rule proposition
iii. Then either of body(e3), body(e4), or body(e5) must not hold in A – from rule
e3, e4, e5 and forced atom proposition
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
235
iv. Then q ≮ ni ≡ q ≥ ni in e3 for all {holds(p, q, k), ptarc(p, ti , ni , k)} ⊆ A – from
e3, forced atom proposition, and given facts (holds(p, q, k) ∈ A, ptarc(p, ti , ni , k) ∈
A)
v. And q ni ≡ q < ni in e4 for all {holds(p, q, k), iptarc(p, ti , ni , k)} ⊆ A, ni =
1; q > ni ≡ q = 0 – from e4, forced atom proposition, given facts (holds(p, q, k) ∈
A, iptarc(p, ti , 1, k) ∈ A), and q is a positive integer
vi. And q ≮ ni ≡ q ≥ ni in e5 for all {holds(p, q, k), tptarc(p, ti , ni , k)} ⊆ A – from
e5, forced atom proposition, and given facts
vii. Then (∀p ∈ •ti , Mk (p) ≥ W(p, ti )) ∧ (∀p ∈ I(ti ), Mk (p) = 0) ∧ (∀(p, ti ) ∈
Q, Mk (p) ≥ QW(p, ti )) – from
A. holds(p, q, k) ∈ A represents q = Mk (p) – inductive assumption, given
B. ptarc(p, ti , ni , k) ∈ A represents ni = W(p, ti ) – rule f 7 construction; or
it represents ni = Mk (p) – rule f 9 construction; the construction of f 9
ensures that notenabled(t, k) is never true due to the reset arc
C. definition 11 of preset •ti in PN
D. definition 28 of enabled transition in PN
viii. Then ti is enabled and can fire in PN, as a result it can belong to T k – from
definition 28 of enabled transition
(b) And consumesmore < A, since A is an answer set of Π4 – from rule a4 and supported
rule proposition
i. Then @consumesmore(p, k) ∈ A – from rule a3 and supported rule proposition
ii. Then @{holds(p, q, k), tot decr(p, q1, k)} ⊆ A : q1 > q in body(e3) – from a2
and forced atom proposition
P
P
iii. Then @p : ( ti ∈{t0 ,...,tx },p∈•ti W(p, ti ) + ti ∈{t0 ,...,tx },p∈R(ti ) Mk (p)) > Mk (p) – from
the following
A. holds(p, q, k) represents q = Mk (p) – inductive assumption, given
B. tot decr(p, q1, k) ∈ A if {del(p, q10 , t0 , k), . . . , del(p, q1 x , t x , k)} ⊆ A, where
q1 = q10 + · · · + q1 x – from r4 and forced atom proposition
C. del(p, q1i , ti , k) ∈ A if { f ires(ti , k), ptarc(p, ti , q1i , k)} ⊆ A – from r7 and
supported rule proposition
D. del(p, q1i , ti , k) either represents removal of q1i = W(p, ti ) tokens from
p ∈ •ti ; or it represents removal of q1i = Mk (p) tokens from p ∈ R(ti )–
from rule r7, supported rule proposition, and definition 22 of transition
execution in PN
iv. Then T k does not contain conflicting transitions – by the definition 23 of conflicting transitions
(c) And for each enabled(t j , k) ∈ A and f ires(t j , k) < A,
could not have(t j , k) ∈ A, since A is an answer set of Π4 - from rule a6 and supported rule proposition
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
236
i. Then {enabled(t j , k), holds(s, qq, k), ptarc(s, t j , q, k),
tot decr(s, qqq, k)} ⊆ A, such that q > qq − qqq and f ires(t j , k) < A - from rule
a7 and supported rule proposition
P
P
ii. Then for an s ∈ •t j ∪R(t j ), q > Mk (s)−( ti ∈Tk ,s∈•ti W(s, ti )+ ti ∈Tk ,s∈R(ti ) Mk (s)),
where q = W(s, t j ) if s ∈ •t j , or Mk (s) otherwise - from the following:
A. ptarc(s, ti , q, k) represents q = W(s, ti ) if (s, ti ) ∈ E − or q = Mk (s) if s ∈
R(ti ) – from rule f 7, f 9 construction
B. holds(s, qq, k) represents qq = Mk (s) – construction
C. tot decr(s, qqq, k) ∈ A if {del(s, qqq0 , t0 , k), . . . , del(s, qqq x , t x , k)}
⊆ A – from rule r4 construction and supported rule proposition
D. del(s, qqqi , ti , k) ∈ A if { f ires(ti , k), ptarc(s, ti , qqqi , k)} ⊆ A – from rule r7
and supported rule proposition
E. del(s, qqqi , ti , k) represents qqqi = W(s, ti ) : ti ∈ T k , (s, ti ) ∈ E − , or qqqi =
Mk (ti ) : ti ∈ T k , s ∈ R(ti ) – from rule f 7, f 9 construction
P
F. tot decr(q, qqq, k) represents ti ∈T k ,s∈•ti W(s, ti ) +
P
ti ∈T k ,s∈R(ti ) Mk (s) – from (C,D,E) above
iii. Then firing T k ∪ {t j } would have required more tokens than are present at its
source place s ∈ •t j ∪ R(t j ). Thus, T k is a maximal set of transitions that can
simultaneously fire.
(d) And for each reset transition tr with enabled(tr , k) ∈ A, f ires(tr , k) ∈ A, since A is
an answer set of Π2 - from rule f 10 and supported rule proposition
i. Then the firing set T k satisfies the reset transition requirement of definition 21
(firing set)
(e) Then {t0 , . . . , t x } = T k – using 1(a),1(b), 1(d) above; and using 1(c) it is a maximal
firing set
2. We show that Mk+1 is produced by firing T k in Mk . Let holds(p, q, k + 1) ∈ A
(a) Then {holds(p, q1, k), tot incr(p, q2, k), tot decr(p, q3, k)} ⊆ A : q = q1 + q2 − q3 –
from rule r5 and supported rule proposition
(b) Then holds(p, q1, k) ∈ A represents q1 = Mk (p) – construction, inductive assumption; and {add(p, q20 , t0 , k), . . . , add(p, q2 j , t j , k)} ⊆ A : q20 + · · · + q2 j = q2 and
{del(p, q30 , t0 , k), . . . , del(p, q3l , tl , k)} ⊆ A : q30 + · · · + q3l = q3 – rules r3, r4 and
supported rule proposition, respectively
(c) Then { f ires(t0 , k), . . . , f ires(t j , k)} ⊆ A and { f ires(t0 , k), . . . , f ires(tl , k)}
⊆ A – rules r6, r7 and supported rule proposition, respectively
(d) Then { f ires(t0 , k), . . . , f ires(t j , k)}∪{ f ires(t0 , k), . . . , f ires(tl , k)} ⊆ A = { f ires(t0 , k), . . . ,
f ires(t x , k)} ⊆ A – set union of subsets
(e) Then for each f ires(t x , k) ∈ A we have t x ∈ T k – already shown in item 1 above
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
237
P
P
(f) Then q = Mk (p) + tx ∈T0 ∧p∈tx • W(t x , p) − ( tx ∈Tk ∧p∈•tx W(p, t x ) +
P
t x ∈T k ∧p∈R(t x ) Mk (p)) – from (2b) above and the following
i. Each add(p, q j , t j , k) ∈ A represents q j = W(t j , p) for p ∈ t j • – rule r6 encoding, and definition 22 of transition execution in PN
ii. Each del(p, ty , qy , k) ∈ A represents either qy = W(p, ty ) for p ∈ •ty , or qy =
Mk (p) for p ∈ R(ty ) – from rule r7, f 7 encoding and definition 22 of transition
execution in PN; or from rule r7, f 9 encoding and definition of reset arc in PN
P
iii. Each tot incr(p, q2, k) ∈ A represents q2 = tx ∈Tk ∧p∈tx • W(t x , p) – aggregate
assignment atom semantics in rule r3
P
P
iv. Each tot decr(p, q3, 0) ∈ A represents q3 = tx ∈Tk ∧p∈•tx W(p, t x )+ tx ∈T k ∧p∈R(tx ) M0 (p)
– aggregate assignment atom semantics in rule r4
(g) Then, Mk+1 (p) = q – since holds(p, q, k + 1) ∈ A encodes q = Mk+1 (p) – from
construction
As a result, for any n > k, T n will be a valid firing set for Mn and Mn+1 will be its target marking.
Conclusion: Since both (I) and (II) hold, X = M0 , T 0 , M1 , . . . , Mk , T k+1 is an execution sequence of PN(P, T, E, W, R) (w.r.t M0 ) iff there is an answer set A of Π4 (PN, M0 , k, ntok) such
that (A.5) and (A.6) hold.
A.6
Proof of Proposition 6
Let PN = (P, T, E, C, W, R, I, Q, QW) be a Petri Net, M0 be its initial marking and let Π5 (PN, M0 , k, ntok)
be the ASP encoding of PN and M0 over a simulation length k, with maximum ntok tokens on
any place node, as defined in section 2.8. Then X = M0 , T k , M1 , . . . , Mk , T k , Mk+1 is an execution sequence of PN (w.r.t. M0 ) iff there is an answer set A of Π5 (PN, M0 , k, ntok) such
that:
{ f ires(t, ts) : t ∈ T ts , 0 ≤ ts ≤ k} = { f ires(t, ts) : f ires(t, ts) ∈ A}
(A.11)
{holds(p, q, c, ts) : p ∈ P, c/q = Mts (p), 0 ≤ ts ≤ k + 1}
= {holds(p, q, c, ts) : holds(p, q, c, ts) ∈ A}
(A.12)
We prove this by showing that:
(I) Given an execution sequence X, we create a set A such that it satisfies (A.11) and (A.12)
and show that A is an answer set of Π5
(II) Given an answer set A of Π5 , we create an execution sequence X such that (A.11) and
(A.12) are satisfied.
First we show (I): Given PN and an execution sequence X of PN, we create a set A as a union
of the following sets:
1. A1 = {num(n) : 0 ≤ n ≤ ntok}
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
238
2. A2 = {time(ts) : 0 ≤ ts ≤ k}
3. A3 = {place(p) : p ∈ P}
4. A4 = {trans(t) : t ∈ T }
5. A5 = {color(c) : c ∈ C}
6. A6 = {ptarc(p, t, nc , c, ts) : (p, t) ∈ E − , c ∈ C, nc = mW(p,t) (c), nc > 0, 0 ≤ ts ≤ k}, where
E− ⊆ E
7. A7 = {tparc(t, p, nc , c, ts) : (t, p) ∈ E + , c ∈ C, nc = mW(t,p) (c), nc > 0, 0 ≤ ts ≤ k}, where
E+ ⊆ E
8. A8 = {holds(p, qc , c, 0) : p ∈ P, c ∈ C, qc = m M0 (p) (c)}
9. A9 = {ptarc(p, t, nc , c, ts) : p ∈ R(t), c ∈ C, nc = m Mts (p) , nc > 0, 0 ≤ ts ≤ k}
10. A10 = {iptarc(p, t, 1, c, ts) : p ∈ I(t), c ∈ C, 0 ≤ ts < k}
11. A11 = {tptarc(p, t, nc , c, ts) : (p, t) ∈ Q, c ∈ C, nc = mQW(p,t) (c), nc > 0, 0 ≤ ts ≤ k}
12. A12 = {notenabled(t, ts) : t ∈ T, 0 ≤ ts ≤ k, ∃c ∈ C, (∃p ∈ •t, m Mts (p) (c) < mW(p,t) (c)) ∨
(∃p ∈ I(t), m Mts (p) (c) > 0) ∨ (∃(p, t) ∈ Q, m Mts (p) (c) < mQW(p,t) (c))}
per definition 32 (enabled transition)
13. A13 = {enabled(t, ts) : t ∈ T, 0 ≤ ts ≤ k, ∀c ∈ C, (∀p ∈ •t, mW(p,t) (c) ≤ m Mts (p) (c)) ∧ (∀p ∈
I(t), m Mts (p) (c) = 0) ∧ (∀(p, t) ∈ Q, m Mts (p) (c) ≥ mQW(p,t) (c))}
per definition 32 (enabled transition)
14. A14 = { f ires(t, ts) : t ∈ T ts , 0 ≤ ts ≤ k}
per definition 35 (firing set), only an enabled transition may fire
15. A15 = {add(p, qc , t, c, ts) : t ∈ T ts , p ∈ t•, c ∈ C, qc = mW(t,p) (c), 0 ≤ ts ≤ k}
per definition 33 (transition execution)
16. A16 = {del(p, qc , t, c, ts) : t ∈ T ts , p ∈ •t, c ∈ C, qc = mW(p,t) (c), 0 ≤ ts ≤ k} ∪
{del(p, qc , t, c, ts) : t ∈ T ts , p ∈ R(t), c ∈ C, qc = m Mts (p) (c), 0 ≤ ts ≤ k}
per definition 33 (transition execution)
P
17. A17 = {tot incr(p, qc , c, ts) : p ∈ P, c ∈ C, qc = t∈Tts ,p∈t• mW(t,p) (c), 0 ≤ ts ≤ k}
per definition 36 (firing set execution)
P
18. A18 = {tot decr(p, qc , c, ts) : p ∈ P, c ∈ C, qc = t∈T ts ,p∈•t mW(p,t) (c) +
P
t∈T ts ,p∈R(t) m Mts (p) (c), 0 ≤ ts ≤ k}
per definition 36 (firing set execution)
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
239
19. A19 = {consumesmore(p, ts) : p ∈ P, c ∈ C, qc = m Mts (p) (c),
P
P
q1c = t∈Tts ,p∈•t mW(p,t) (c) + t∈Tts ,p∈R(t) m Mts (p) (c), q1 > q, 0 ≤ ts ≤ k}
per definition 34 (conflicting transitions)
20. A20 = {consumesmore : ∃p ∈ P, c ∈ C, qc = m Mts (p) (c),
P
P
q1c = t∈Tts ,p∈•t mW(p,t) (c) + t∈Tts ,p∈R(t) (m Mts (p) (c)), q1 > q, 0 ≤ ts ≤ k}
per definition 34 (conflicting transitions) per the maximal firing set semantics
21. A21 = {holds(p, qc , c, ts + 1) : p ∈ P, c ∈ C, qc = m Mts+1 (p) (c), 0 ≤ ts < k},
P
where Mts+1 (p) = Mts (p) − ( t∈Tts ,p∈•t W(p, t) +
P
P
t∈T ts ,p∈R(t) Mts (p)) + t∈T ts ,p∈t• W(t, p)
according to definition 36 (firing set execution)
We show that A satisfies (A.11) and (A.12), and A is an answer set of Π5 .
A satisfies (A.11) and (A.12) by its construction above. We show A is an answer set of Π5
by splitting. We split lit(Π5 ) into a sequence of 7k + 9 sets:
• U0 = head( f 1)∪head( f 2)∪head( f 13)∪head( f 5)∪head( f 6)∪head(i2) = {place(p) : p ∈
P}∪{trans(t) : t ∈ T }∪{col(c) : c ∈ C}∪{time(0), . . . , time(k)}∪{num(0), . . . , num(ntok)}∪
{holds(p, qc , c, 0) : p ∈ P, c ∈ C, qc = m M0 (p) (c)}
• U7k+1 = U7k+0 ∪head( f 14)ts=k ∪head( f 15)ts=k ∪head( f 16)ts=k ∪head( f 17)ts=k ∪head( f 18)ts=k =
U7k+0 ∪ {ptarc(p, t, nc , c, k) : (p, t) ∈ E − , c ∈ C, nc = mW(p,t) (c)} ∪ {tparc(t, p, nc , c, k) :
(t, p) ∈ E + , c ∈ C, nc = mW(t,p) (c)} ∪ {ptarc(p, t, nc , c, k) : p ∈ R(t), c ∈ C, nc =
m Mk (p) (c), n > 0} ∪ {iptarc(p, t, 1, c, k) : p ∈ I(t), c ∈ C} ∪ {tptarc(p, t, nc , c, k) : (p, t) ∈
Q, c ∈ C, nc = mQW(p,t) (c)}
• U7k+2 = U7k+1 ∪ head(e6)ts=k ∪ head(e7)ts=k ∪ head(e8)ts=k = U7k+1 ∪
{notenabled(t, k) : t ∈ T }
• U7k+3 = U7k+2 ∪ head(e9)ts=k = U7k+2 ∪ {enabled(t, k) : t ∈ T }
• U7k+4 = U7k+3 ∪ head(a1)ts=k = U7k+3 ∪ { f ires(t, k) : t ∈ T }
• U7k+5 = U7k+4 ∪ head(r8)ts=k ∪ head(r9)ts=k = U7k+4 ∪ {add(p, qc , t, c, k) : p ∈ P, t ∈
T, c ∈ C, qc = mW(t,p) (c)} ∪ {del(p, qc , t, c, k) : p ∈ P, t ∈ T, c ∈ C, qc = mW(p,t) (c)} ∪
{del(p, qc , t, c, k) : p ∈ P, t ∈ T, c ∈ C, qc = m Mk (p) (c)}
• U7k+6 = U7k+5 ∪ head(r10)ts=k ∪ head(r11)ts=k = U7k+5 ∪ {tot incr(p, qc , c, k) : p ∈ P, c ∈
C, 0 ≤ qc ≤ ntok} ∪ {tot decr(p, qc , c, k) : p ∈ P, c ∈ C, 0 ≤ qc ≤ ntok}
• U7k+7 = U7k+6 ∪ head(a9)ts=k ∪ head(r12)ts=k = U7k+6 ∪ {consumesmore(p, k) : p ∈
P} ∪ {holds(p, q, k + 1) : p ∈ P, 0 ≤ q ≤ ntok}
• U7k+8 = U7k+7 ∪ head(a3) = U7k+7 ∪ {consumesmore}
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
240
where head(ri )ts=k are head atoms of ground rule ri in which ts = k. We write Ats=k
=
i
{a(. . . , ts) : a(. . . , ts) ∈ Ai , ts = k} as short hand for all atoms in Ai with ts = k. Uα , 0 ≤
α ≤ 7k + 8 form a splitting sequence, since each Ui is a splitting set of Π5 , and hUα iα<µ is a
S
monotone continuous sequence, where U0 ⊆ U1 · · · ⊆ U7k+8 and α<µ Uα = lit(Π5 ).
We compute the answer set of Π5 using the splitting sets as follows:
1. botU0 (Π5 ) = f 1 ∪ f 2 ∪ f 14 ∪ f 9 ∪ f 18 ∪ i2 and X0 = A1 ∪ · · · ∪ A5 ∪ A8 (= U0 ) is its
answer set – using forced atom proposition
2. evalU0 (botU1 (Π5 ) \ botU0 (Π5 ), X0 ) = {ptarc(p, t, qc , c, 0):-.|c ∈ C, qc = mW(p,t) (c), qc >
0} ∪ {tparc(t, p, qc , c, 0):-.|c ∈ C, qc = mW(t,p) (c), qc > 0} ∪ {ptarc(p, t, qc , c, 0):-.|c ∈
C, qc = m M0 (p) (c), qc > 0} ∪ {iptarc(p, t, 1, c, 0):-.|c ∈ C} ∪ {tptarc(p, t, qc , c, 0):-.|c ∈
ts=0
C, qc = mQW(p,t) (c), qc > 0}. Its answer set X1 = Ats=0
∪ Ats=0
∪ Ats=0
∪ Ats=0
7
9
10 ∪ A11 –
6
using forced atom proposition and construction of A6 , A7 , A9 , A10 , A11 .
3. evalU1 (botU2 (Π5 ) \ botU1 (Π5 ), X0 ∪ X1 ) = {notenabled(t, 0):-.|({trans(t),
ptarc(p, t, nc , c, 0), holds(p, qc , c, 0)} ⊆ X0 ∪ X1 , where qc < nc ) or
({notenabled(t, 0):-.|({trans(t), iptarc(p, t, n2c , c, 0), holds(p, qc , c, 0)} ⊆ X0 ∪ X1 ,
where qc ≥ n2c }) or ({trans(t), tptarc(p, t, n3c , c, 0), holds(p, qc , c, 0)} ⊆ X0 ∪X1 , where qc <
n3c )}. Its answer set X2 = Ats=0
12 – using forced atom proposition and construction of A12 .
(a) where, qc = m M0 (p) (c), and nc = mW(p,t) (c) for an arc (p, t) ∈ E − – by construction
of i2 and f 14 in Π5 , and
(b) in an arc (p, t) ∈ E − , p ∈ •t (by definition 31 of preset)
(c) n2c = 1 – by construction of iptarc predicates in Π5 , meaning qc ≥ n2c ≡ qc ≥ 1 ≡
qc > 0,
(d) tptarc(p, t, n3c , c, 0) represents n3c = mQW(p,t) (c), where (p, t) ∈ Q
(e) thus, notenabled(t, 0) ∈ X1 represents ∃c ∈ C, (∃p ∈ •t : m M0 (p) (c) < mW(p,t) (c)) ∨
(∃p ∈ I(t) : m M0 (p) (c) > 0) ∨ (∃(p, t) ∈ Q : m M0 (p) (c) < mQW(p,t) (c)).
4. evalU2 (botU3 (Π5 ) \ botU2 (Π5 ), X0 ∪ · · · ∪ X2 ) = {enabled(t, 0):-.|trans(t) ∈ X0 ∪ · · · ∪
X2 , notenabled(t, 0) < X0 ∪ · · · ∪ X2 }. Its answer set is X3 = Ats=0
13 – using forced atom
proposition and construction of A13 .
(a) where, an enabled(t, 0) ∈ X3 if @ notenabled(t, 0) ∈ X0 ∪· · ·∪X2 ; which is equivalent
to ∀c ∈ C, (@p ∈ •t : m M0 (p) (c) < mW(p,t) (c))∧(@p ∈ I(t) : m M0 (p) (c) > 0)∧(@(p, t) ∈
Q : m M0 (p) (c) < mQW(p,t) (c)) ≡ ∀c ∈ C, (∀p ∈ •t : m M0 (p) (c) ≥ mW(p,t) (c)) ∧ (∀p ∈
I(t) : m M0 (p) (c) = 0).
5. evalU3 (botU4 (Π5 ) \ botU3 (Π5 ), X0 ∪ · · · ∪ X3 ) = {{ f ires(t, 0)}:-.|enabled(t, 0)
holds in X0 ∪ · · · ∪ X3 }. It has multiple answer sets X4.1 , . . . , X4.n , corresponding to elements of power set of f ires(t, 0) atoms in evalU3 (...) – using supported rule proposition.
Since we are showing that the union of answer sets of Π5 determined using splitting is
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
241
equal to A, we only consider the set that matches the f ires(t, 0) elements in A and call it
X4 , ignoring the rest. Thus, X4 = Ats=0
14 , representing T 0 .
(a) in addition, for every t such that enabled(t, 0) ∈ X0 ∪ · · · ∪ X3 , R(t) , ∅; f ires(t, 0) ∈
X4 – per definition 35 (firing set); requiring that a reset transition is fired when
enabled
(b) thus, the firing set T 0 will not be eliminated by the constraint f 10
6. evalU4 (botU5 (Π5 ) \ botU4 (Π5 ), X0 ∪ · · · ∪ X4 ) = {add(p, nc , t, c, 0):-.|{ f ires(t, 0),
tparc(t, p, nc , c, 0)} ⊆ X0 ∪ · · · ∪ X4 } ∪ {del(p, nc , t, c, 0):-.|{ f ires(t, 0),
ptarc(p, t, nc , c, 0)} ⊆ X0 ∪ · · · ∪ X4 }. It’s answer set is X5 = Ats=0
∪ Ats=0
– using forced
15
16
atom proposition and definitions of A15 and A16 .
(a) where, each add atom is equivalent to nc = mW(t,p) (c), c ∈ C, p ∈ t•,
(b) and each del atom is equivalent to nc = mW(p,t) (c), c ∈ C, p ∈ •t; or nc = m M0 (p) (c), c ∈
C, p ∈ R(t),
(c) representing the effect of transitions in T 0
P
7. evalU5 (botU6 (Π5 )\botU5 (Π5 ), X0 ∪· · ·∪X5 ) = {tot incr(p, qqc , c, 0):-.|qqc = add(p,qc ,t,c,0)∈X0 ∪···∪X5 qc }∪
{tot decr(p, qqc , c, 0):-.|qqc =
P
ts=0
ts=0
del(p,qc ,t,c,0)∈X0 ∪···∪X5 qc }. It’s answer set is X6 = A17 ∪ A18 – using forced atom proposition and definitions of A17 and A18 .
P
(a) where, each tot incr(p, qqc , c, 0), qqc = add(p,qc ,t,c,0)∈X0 ∪...X5 qc
P
≡ qqc = t∈X4 ,p∈t• mW(p,t) (c),
P
(b) and each tot decr(p, qqc , c, 0), qqc = del(p,qc ,t,c,0)∈X0 ∪...X5 qc
P
P
≡ qq = t∈X4 ,p∈•t mW(t,p) (c) + t∈X4 ,p∈R(t) m M0 (p) (c),
(c) represent the net effect of transitions in T 0
8. evalU6 (botU7 (Π5 ) \ botU6 (Π5 ), X0 ∪ · · · ∪ X6 ) = {consumesmore(p, 0):-.|
{holds(p, qc , c, 0), tot decr(p, q1c , c, 0)} ⊆ X0 ∪ · · · ∪ X6 , q1c > qc } ∪
{holds(p, qc , c, 1):-., |{holds(p, q1c , c, 0), tot incr(p, q2c , c, 0), tot decr(p, q3c , c, 0)}
ts=0
⊆ X0 ∪ · · · ∪ X6 , qc = q1c + q2c − q3c }. It’s answer set is X7 = Ats=0
19 ∪ A21 – using forced
atom proposition and definitions of A19 , A21 .
(a) where, consumesmore(p, 0) represents ∃p : qc = m M0 (p) (c), q1c =
P
P
t∈T 0 ,p∈•t mW(p,t) (c) + t∈T 0 ,p∈R(t) m M0 (p) (c), q1c > qc , c ∈ C, indicating place p will
be over consumed if T 0 is fired, as defined in definition 34 (conflicting transitions),
(b) holds(p, qc , c, 1) represents qc = m M1 (p) (c) – by construction of Π5
..
.
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
242
9. evalU7k+0 (botU7k+1 (Π5 ) \ botU7k+0 (Π5 ), X0 ∪ · · · ∪ X7k+0 ) = {ptarc(p, t, qc , c, k):-.|c ∈ C, qc =
mW(p, t)(c), qc > 0}∪{tparc(t, p, qc , c, k):-.|c ∈ C, qc = mW(t,p) (c), qc > 0}∪ {ptarc(p, t, qc , c, k):-.|
c ∈ C, qc = m Mk (p) (c), qc > 0} ∪ {iptarc(p, t, 1, c, k):-.|c ∈ C} ∪ {tptarc(p, t, qc , c, k):-.|c ∈
ts=k
C, qc = mQW(p,t) (c), qc > 0}. Its answer set X7k+1 = Ats=k
∪ Ats=k
∪ Ats=k
∪ Ats=k
7
9
10 ∪ A11 –
6
using forced atom proposition and construction of A6 , A7 , A9 , A10 , A11 .
10. evalU7k+1 (botU7k+2 (Π5 ) \ botU7k+1 (Π5 ), X0 ∪ · · · ∪ X7k+1 ) = {notenabled(t, k):-.|
({trans(t), ptarc(p, t, nc , c, k), holds(p, qc , c, k)} ⊆ X0 ∪ · · · ∪ X7k+1 , where qc < nc ) or
({notenabled(t, k):-.|({trans(t), iptarc(p, t, n2c , c, k), holds(p, qc , c, k)} ⊆ X0 ∪· · ·∪X7k+1 , where qc ≥
n2c }) or ({trans(t), tptarc(p, t, n3c , c, k),
holds(p, qc , c, k)} ⊆ X0 ∪ · · · ∪ X7k+1 , where qc < n3c )}. Its answer set X7k+2 = Ats=k
12 –
using forced atom proposition and construction of A12 .
(a) since qc = m Mk (p) (c), and nc = mW(p,t) (c) for an arc (p, t) ∈ E − – by construction of
holds and ptarc predicates in Π5 , and
(b) in an arc (p, t) ∈ E − , p ∈ •t (by definition 31 of preset)
(c) n2c = 1 – by construction of iptarc predicates in Π5 , meaning qc ≥ n2c ≡ qc ≥ 1 ≡
qc > 0,
(d) tptarc(p, t, n3c , c, k) represents n3c = mQW(p,t) (c), where (p, t) ∈ Q
(e) thus, notenabled(t, k) ∈ X8k+1 represents ∃c ∈ C(∃p ∈ •t : m Mk (p) (c) < mW(p,t) (c)) ∨
(∃p ∈ I(t) : m Mk (p) (c) > 0) ∨ (∃(p, t) ∈ Q : m Mts (p) (c) < mQW(p,t) (c)).
11. evalU7k+2 (botU7k+3 (Π5 ) \ botU7k+2 (Π5 ), X0 ∪ · · · ∪ X7k+2 ) = {enabled(t, k):-.|trans(t) ∈ X0 ∪
· · · ∪ X7k+2 ∧ notenabled(t, k) < X0 ∪ · · · ∪ X7k+2 }. Its answer set is X7k+3 = Ats=k
13 – using
forced atom proposition and construction of A13 .
(a) since an enabled(t, k) ∈ X7k+3 if @ notenabled(t, k) ∈ X0 ∪ · · · ∪ X7k+2 ; which is
equivalent to ∀c ∈ C, (@p ∈ •t : m Mk (p) (c) < mW(p,t) (c)) ∧ (@p ∈ I(t) : m Mk (p) (c) >
0) ∧ (@(p, t) ∈ Q : m Mk (p) (c) < mQW(p,t) (c)) ≡ ∀c ∈ C, (∀p ∈ •t : m Mk (p) (c) ≥
mW(p,t) (c)) ∧ (∀p ∈ I(t) : m Mk (p) (c) = 0).
12. evalU7k+3 (botU7k+4 (Π5 ) \ botU7k+3 (Π5 ), X0 ∪ · · · ∪ X7k+3 ) = {{ f ires(t, k)}:-.|
enabled(t, k) holds in X0 ∪ · · · ∪ X7k+3 }. It has multiple answer sets
X7k+ 4.1, . . . , X7k+ 4.n, corresponding to elements of power set of f ires(t, k) atoms in evalU3 (...)
– using supported rule proposition. Since we are showing that the union of answer sets
of Π5 determined using splitting is equal to A, we only consider the set that matches
the f ires(t, k) elements in A and call it X7k+4 , ignoring the rest. Thus, X7k+4 = Ats=k
14 ,
representing T k .
(a) in addition, for every t such that enabled(t, k) ∈ X0 ∪· · ·∪X7k+3 , R(t) , ∅; f ires(t, k) ∈
X7k+4 – per definition 35 (firing set); requiring that a reset transition is fired when
enabled
(b) thus, the firing set T k will not be eliminated by the constraint f 10
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
243
13. evalU7k+4 (botU7k+5 (Π5 ) \ botU7k+4 (Π5 ), X0 ∪ · · · ∪ X7k+4 ) = {add(p, nc , t, c, k):-.| { f ires(t, k),
tparc(t, p, nc , c, k)} ⊆ X0 ∪ · · · ∪ X7k+4 }∪ {del(p, nc , t, c, k):-.| { f ires(t, k),
ptarc(p, t, nc , c, k)} ⊆ X0 ∪ · · · ∪ X7k+4 }. It’s answer set is X7k+5 = Ats=k
∪ Ats=k
– using
15
16
forced atom proposition and definitions of A15 and A16 .
(a) where each add atom is equivalent to nc = mW(t,p) (c), c ∈ C, p ∈ t•,
(b) and each del atom is equivalent to nc = mW(p,t) (c), c ∈ C, p ∈ •t; or nc = m Mk (p) (c), c ∈
C, p ∈ R(t),
(c) representing the effect of transitions in T k
14. evalU7k+5 (botU7k+6 (Π5 ) \ botU7k+5 (Π5 ), X0 ∪ · · · ∪ X7k+5 ) = {tot incr(p, qqc , c, k):-.| qq =
P
q }∪ {tot decr(p, qqc , c, k):-.|
add(p,q
Pc ,t,c,k)∈X0 ∪···∪X7k+5 c
ts=k
qqc = del(p,qc ,t,c,k)∈X0 ∪···∪X7k+5 qc }. It’s answer set is X7k+6 = Ats=k
17 ∪ A18 – using forced
atom proposition and definitions of A17 and A18 .
P
(a) where, each tot incr(p, qqc , c, k), qqc = add(p,qc ,t,c,k)∈X0 ∪...X8k+5 qc
P
≡ qqc = t∈X7k+4 ,p∈t• mW(p,t) (c),
P
(b) and each tot decr(p, qqc , c, k), qqc = del(p,qc ,t,c,k)∈X0 ∪...X8k+5 qc
P
P
≡ qq = t∈X7k+4 ,p∈•t mW(t,p) (c) + t∈X7k+4 ,p∈R(t) m Mk (p) (c),
(c) represent the net effect of transitions in T k
15. evalU7k+6 (botU7k+7 (Π5 ) \ botU7k+6 (Π5 ), X0 ∪ · · · ∪ X7k+6 ) = {consumesmore(p, k):-.|
{holds(p, qc , c, k), tot decr(p, q1c , c, k)} ⊆ X0 ∪ · · · ∪ X7k+6 , q1 > q} ∪
{holds(p, qc , c, k + 1):-.|{holds(p, q1c , c, k), tot incr(p, q2c , c, k),
tot decr(p, q3c , c, k)} ⊆ X0 ∪ · · · ∪ X7k+6 , qc = q1c + q2c − q3c }. It’s answer set is X7k+7 =
ts=k
Ats=k
19 ∪ A21 – using forced atom proposition and definitions of A19 , A21 .
(a) where, consumesmore(p, k) represents ∃p : qc = m Mk (p) (c), q1c =
P
P
t∈T k ,p∈•t mW(p,t) (c) + t∈T k ,p∈R(t) m Mk (p) (c), q1c > qc , c ∈ C, indicating place p that
will be over consumed if T k is fired, as defined in definition 34 (conflicting transitions),
(b) and holds(p, qc , c, k + 1) represents qc = m Mk+1 (p) (c) – by construction of Π5
16. evalU7k+7 (botU7k+8 (Π5 ) \ botU7k+7 (Π5 ), X0 ∪ · · · ∪ X7k+7 ) = {consumesmore:-.|
{consumesmore(p, 0), . . . , consumesmore(p, k)} ∩ (X0 ∪ · · · ∪ X7k+7 ) , ∅}. It’s answer set
is X7k+8 = Ats=k
20 – using forced atom proposition and the definition of A20
(a) X7k+7 will be empty since none of consumesmore(p, 0), . . . ,
consumesmore(p, k) hold in X0 ∪ · · · ∪ X7k+7 due to the construction of A, encoding
of a9 and its body atoms. As a result, it is not eliminated by the constraint a4
The set X = X0 ∪ · · · ∪ X7k+8 is the answer set of Π5 by the splitting sequence theorem 15.
Each Xi , 0 ≤ i ≤ 7k + 8 matches a distinct portion of A, and X = A, thus A is an answer set of
Π5 .
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
244
Next we show (II): Given Π5 be the encoding of a Petri Net PN(P, T, E, C, W, R, I, Q, WQ) with
initial marking M0 , and A be an answer set of Π5 that satisfies (A.11) and (A.12), then we can
construct X = M0 , T k , . . . , Mk , T k , Mk+1 from A, such that it is an execution sequence of PN.
We construct the X as follows:
1. Mi = (Mi (p0 ), . . . , Mi (pn )), where {holds(p0 , m Mi (p0 ) (c), c, i), . . .
holds(pn , m Mi (pn ) (c), c, i)} ⊆ A, for c ∈ C, 0 ≤ i ≤ k + 1
2. T i = {t : f ires(t, i) ∈ A}, for 0 ≤ i ≤ k
and show that X is indeed an execution sequence of PN. We show this by induction over k (i.e.
given Mk , T k is a valid firing set and its firing produces marking Mk+1 ).
Base case: Let k = 0, and M0 is a valid marking in X for PN, show (1) T 0 is a valid firing set
for M0 , and (2) firing T 0 in M0 results in marking M1 .
1. We show T 0 is a valid firing set for M0 . Let { f ires(t0 , 0), . . . , f ires(t x , 0)} be the set of all
f ires(. . . , 0) atoms in A,
(a) Then for each f ires(ti , 0) ∈ A
i. enabled(ti , 0) ∈ A – from rule a1 and supported rule proposition
ii. Then notenabled(ti , 0) < A – from rule e9 and supported rule proposition
iii. Then either of body(e6), body(e7), or body(e8) must not hold in A – from rules
e6, e7, e8 and forced atom proposition
iv. Then qc ≮ nic ≡ qc ≥ nic in e6 for all {holds(p, qc , c, 0),
ptarc(p, ti , nic , c, 0)} ⊆ A – from e6, forced atom proposition, and given facts
(holds(p, qc , c, 0) ∈ A, ptarc(p, ti , nic , 0) ∈ A)
v. And qc nic ≡ qc < nic in e7 for all {holds(p, qc , c, 0),
iptarc(p, ti , nic , c, 0)} ⊆ A, nic = 1; qc > nic ≡ qc = 0 – from e7, forced atom
proposition, given facts (holds(p, qc , c, 0) ∈ A, iptarc(p, ti , 1, c, 0) ∈ A), and qc
is a positive integer
vi. And qc ≮ nic ≡ qc ≥ nic in e8 for all {holds(p, qc , c, 0),
tptarc(p, ti , nic , c, 0)} ⊆ A – from e8, forced atom proposition, and given facts
vii. Then ∀c ∈ C, (∀p ∈ •ti , m M0 (p) (c) ≥ mW(p,ti ) (c)) ∧ (∀p ∈ I(ti ), m M0 (p) (c) =
0) ∧ (∀(p, ti ) ∈ Q, m M0 (p) (c) ≥ mQW(p,ti ) (c)) – from the following
A. holds(p, qc , c, 0) ∈ A represents qc = m M0 (p) (c) – rule i2 construction
B. ptarc(p, ti , nic , 0) ∈ A represents nic = mW(p,ti ) (c) – rule f 14 construction;
or it represents nic = m M0 (p) (c) – rule f 16 construction; the construction of
f 16 ensures that notenabled(t, 0) is never true for a reset arc
C. definition 31 of preset •ti in PN
D. definition 32 of enabled transition in PN
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
245
viii. Then ti is enabled and can fire in PN, as a result it can belong to T 0 – from
definition 32 of enabled transition
(b) And consumesmore < A, since A is an answer set of Π5 – from rule a3 and supported
rule proposition
i. Then @consumesmore(p, 0) ∈ A – from rule a9 and supported rule proposition
ii. Then @{holds(p, qc , c, 0), tot decr(p, q1c , c, 0)} ⊆ A, q1c > qc in body(a9) –
from a9 and forced atom proposition
P
iii. Then @c ∈ C@p ∈ P, ( ti ∈{t0 ,...,tx },p∈•ti mW(p,ti ) (c) +
P
ti ∈{t0 ,...,t x },p∈R(ti ) m M0 (p) (c)) > m M0 (p) (c) – from the following
A. holds(p, qc , c, 0) represents qc = m M0 (p) (c) – from rule i2 construction,
given
B. tot decr(p, q1c , c, 0) ∈ A if {del(p, q10 c , t0 , c, 0), . . . ,
del(p, q1 x c , t x , c, 0)} ⊆ A, where q1c = q10 c + · · · + q1 x c – from r11 and
forced atom proposition
C. del(p, q1i c , ti , c, 0) ∈ A if { f ires(ti , 0), ptarc(p, ti , q1i c , c, 0)} ⊆ A – from r9
and supported rule proposition
D. del(p, q1i c , ti , c, 0) either represents removal of q1i c = mW(p,ti ) (c) tokens
from p ∈ •ti ; or it represents removal of q1i c = m M0 (p) (c) tokens from p ∈
R(ti )– from rules r9, f 14, f 16, supported rule proposition, and definition 33
of transition execution in PN
iv. Then the set of transitions in T 0 do not conflict – by the definition 34 of conflicting transitions
(c) And for each reset transition tr with enabled(tr , 0) ∈ A, f ires(tr , 0) ∈ A, since A is
an answer set of Π5 - from rule f 10 and supported rule proposition
i. Then, the firing set T 0 satisfies the reset-transition requirement of definition 35
(firing set)
(d) Then {t0 , . . . , t x } = T 0 – using 1(a),1(b),1(d) above; and using 1(c) it is a maximal
firing set
2. We show M1 is produced by firing T 0 in M0 . Let holds(p, qc , c, 1) ∈ A
(a) Then {holds(p, q1c , c, 0), tot incr(p, q2c , c, 0), tot decr(p, q3c , c, 0)} ⊆ A : qc = q1c +
q2c − q3c – from rule r12 and supported rule proposition
(b) Then, holds(p, q1c , c, 0) ∈ A represents q1c = m M0 (p) (c) – given, rule i2 construction; and {add(p, q20 c , t0 , c, 0), . . . , add(p, q2 j c , t j , c, 0)} ⊆ A : q20 c +· · ·+q2 j c = q2c
and {del(p, q30 c , t0 , c, 0), . . . , del(p, q3l c , tl , c, 0)} ⊆ A : q30 c +· · ·+q3l c = q3c – rules
r10, r11 and supported rule proposition, respectively
(c) Then { f ires(t0 , 0), . . . , f ires(t j , 0)} ⊆ A and { f ires(t0 , 0), . . . , f ires(tl , 0)} ⊆ A –
rules r8, r9 and supported rule proposition, respectively
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
246
(d) Then { f ires(t0 , 0), . . . , f ires(t j , 0)}∪ { f ires(t0 , 0), . . . , f ires(tl , 0)} ⊆ A = { f ires(t0 , 0), . . . ,
f ires(t x , 0)} ⊆ A – set union of subsets
(e) Then for each f ires(t x , 0) ∈ A we have t x ∈ T 0 – already shown in item 1 above
P
P
P
(f) Then qc = m M0 (p) (c)+ tx ∈T0 ∧p∈tx • mW(tx ,p) (c)−( tx ∈T0 ∧p∈•tx mW(p,tx ) (c)+ tx ∈T0 ∧p∈R(tx ) m M0 (p) (c))
– from (2b) above and the following
i. Each add(p, q j c , t j , c, 0) ∈ A represents q j c = mW(t j ,p) (c) for p ∈ t j • – rule
r8, f 15 encoding, and definition 33 of transition execution in PN
ii. Each del(p, ty , qy c , c, 0) ∈ A represents either qy c = mW(p,ty ) (c) for p ∈ •ty , or
qy c = m M0 (p) (c) for p ∈ R(ty ) – from rule r9, f 14 encoding and definition 33
of transition execution in PN; or from rule r9, f 16 encoding and definition of
reset arc in PN
P
iii. Each tot incr(p, q2c , c, 0) ∈ A represents q2c = tx ∈T0 ∧p∈tx • mW(tx ,p) (c) – aggregate assignment atom semantics in rule r10
P
P
iv. Each tot decr(p, q3c , c, 0) ∈ A represents q3c = tx ∈T0 ∧p∈•tx mW(p,tx ) (c)+ tx ∈T0 ∧p∈R(tx ) m M0 (p) (c)
– aggregate assignment atom semantics in rule r11
(g) Then, m M1 (p) (c) = qc – since holds(p, qc , c, 1) ∈ A encodes qc = m M1 (p) (c) – from
construction
Inductive Step: Let k > 0, and Mk is a valid marking in X for PN, show (1) T k is a valid firing
set for Mk , and (2) firing T k in Mk produces marking Mk+1 .
1. We show that T k is a valid firing set in Mk . Let { f ires(t0 , k), . . . , f ires(t x , k)} be the set of
all f ires(. . . , k) atoms in A,
(a) Then for each f ires(ti , k) ∈ A
i. enabled(ti , k) ∈ A – from rule a1 and supported rule proposition
ii. Then notenabled(ti , k) < A – from rule e9 and supported rule proposition
iii. Then either of body(e6), body(e7), or body(e8) must not hold in A – from rules
e6, e7, e8 and forced atom proposition
iv. Then qc ≮ nic ≡ qc ≥ nic in e6 for all {holds(p, qc , c, k),
ptarc(p, ti , nic , c, k)} ⊆ A – from e6, forced atom proposition, given facts, and
the inductive assumption
v. And qc nic ≡ qc < nic in e7 for all {holds(p, qc , c, k),
iptarc(p, ti , nic , c, k)} ⊆ A, nic = 1; qc > nic ≡ qc = 0 – from e7, forced atom
proposition, given facts (holds(p, qc , c, k) ∈ A, iptarc(p, ti , 1, c, k) ∈ A), and qc
is a positive integer
vi. And qc ≮ nic ≡ qc ≥ nic in e8 for all {holds(p, qc , c, k),
tptarc(p, ti , nic , c, k)} ⊆ A – from e8, forced atom proposition, and inductive
assumption
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
247
vii. Then (∀p ∈ •ti , m Mk (p) (c) ≥ mW(p,ti ) (c))∧(∀p ∈ I(ti ), m Mk (p) (c) = 0)∧(∀(p, ti ) ∈
Q, m Mk (p) (c) ≥ mQW(p,ti ) (c)) – from
A. holds(p, qc , c, k) ∈ A represents qc = m Mk (p) (c) – construction of Π5
B. ptarc(p, ti , nic , k) ∈ A represents nic = mW(p,ti ) (c) – rule f 14 construction;
or it represents nic = m Mk (p) (c) – rule f 16 construction; the construction of
f 16 ensures that notenabled(t, k) is never true for a reset arc
C. definition 31 of preset •ti in PN
D. definition 32 of enabled transition in PN
viii. Then ti is enabled and can fire in PN, as a result it can belong to T k – from
definition 32 of enabled transition
(b) And consumesmore < A, since A is an answer set of Π5 – from rule a3 and supported
rule proposition
i. Then @consumesmore(p, k) ∈ A – from rule a9 and supported rule proposition
ii. Then @{holds(p, qc , c, k), tot decr(p, q1c , c, k)} ⊆ A, q1c > qc in body(a9) –
from a9 and forced atom proposition
P
iii. Then @c ∈ C, @p ∈ P, ( ti ∈{t0 ,...,tx },p∈•ti mW(p,ti ) (c) +
P
ti ∈{t0 ,...,t x },p∈R(ti ) m Mk (p) (c)) > m Mk (p) (c) – from the following
A. holds(p, qc , c, k) represents qc = m Mk (p) (c) – construction of Π5 , inductive
assumption
B. tot decr(p, q1c , c, k) ∈ A if {del(p, q10 c , t0 , c, k), . . . ,
del(p, q1 x c , t x , c, k)} ⊆ A, where q1c = q10 c + · · · + q1 x c – from r11 and
forced atom proposition
C. del(p, q1i c , ti , c, k) ∈ A if { f ires(ti , k), ptarc(p, ti , q1i c , c, k)} ⊆ A – from r9
and supported rule proposition
D. del(p, q1i c , ti , c, k) either represents removal of q1i c = mW(p,ti ) (c) tokens
from p ∈ •ti ; or it represents removal of q1i c = m Mk (p) (c) tokens from p ∈
R(ti )– from rules r9, f 14, f 16, supported rule proposition, and definition 33
of transition execution in PN
iv. Then the set of transitions in T k do not conflict – by the definition 34 of conflicting transitions
(c) And for each reset transition tr with enabled(tr , k) ∈ A, f ires(tr , k) ∈ A, since A is
an answer set of Π5 - from rule f 10 and supported rule proposition
i. Then the firing set T k satisfies the reset transition requirement of definition 35
(firing set)
(d) Then {t0 , . . . , t x } = T k – using 1(a),1(b), 1(d) above; and using 1(c) it is a maximal
firing set
2. We show that Mk+1 is produced by firing T k in Mk . Let holds(p, qc , c, k + 1) ∈ A
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
248
(a) Then {holds(p, q1c , c, k), tot incr(p, q2c , c, k), tot decr(p, q3c , c, k)} ⊆ A : qc = q1c +
q2c − q3c – from rule r12 and supported rule proposition
(b) Then, holds(p, q1c , c, k) ∈ A represents q1c = m Mk (p) (c) – construction, inductive
assumption; and {add(p, q20 c , t0 , c, k), . . . , add(p, q2 j c , t j , c, k)} ⊆ A : q20 c + · · · +
q2 j c = q2c and {del(p, q30 c , t0 , c, k), . . . , del(p, q3l c , tl , c, k)} ⊆ A : q30 c +· · ·+q3l c =
q3c – rules r10, r11 and supported rule proposition, respectively
(c) Then { f ires(t0 , k), . . . , f ires(t j , k)} ⊆ A and { f ires(t0 , k), . . . , f ires(tl , k)}
⊆ A – rules r8, r9 and supported rule proposition, respectively
(d) Then { f ires(t0 , k), . . . , f ires(t j , k)}∪{ f ires(t0 , k), . . . , f ires(tl , k)} ⊆ A = { f ires(t0 , k), . . . ,
f ires(t x , k)} ⊆ A – set union of subsets
(e) Then for each f ires(t x , k) ∈ A we have t x ∈ T k – already shown in item 1 above
P
P
P
(f) Then qc = m Mk (p) (c)+ tx ∈Tk ∧p∈tx • mW(tx ,p) (c)−( tx ∈T k ∧p∈•tx mW(p,tx ) (c)+ tx ∈Tk ∧p∈R(tx ) m Mk (p) (c))
– from (2b) above and the following
i. Each add(p, q j c , t j , c, k) ∈ A represents q j c = mW(t j ,p) (c) for p ∈ t j • – rule
r8, f 15 encoding, and definition 33 of transition execution in PN
ii. Each del(p, ty , qy c , c, k) ∈ A represents either qy c = mW(p,ty ) (c) for p ∈ •ty , or
qy c = m Mk (p) (c) for p ∈ R(ty ) – from rule r9, f 14 encoding and definition 33
of transition execution in PN; or from rule r9, f 16 encoding and definition of
reset arc in PN
P
iii. Each tot incr(p, q2c , c, k) ∈ A represents q2c = tx ∈Tk ∧p∈tx • mW(tx ,p) (c) – aggregate assignment atom semantics in rule r10
P
P
iv. Each tot decr(p, q3c , c, k) ∈ A represents q3c = tx ∈Tk ∧p∈•tx mW(p,tx ) (c)+ tx ∈Tk ∧p∈R(tx ) m Mk (p) (c)
– aggregate assignment atom semantics in rule r11
(g) Then, m Mk+1 (p) (c) = qc – since holds(p, qc , c, k + 1) ∈ A encodes qc = m Mk+1 (p) (c) –
from construction
As a result, for any n > k, T n will be a valid firing set for Mn and Mn+1 will be its target marking.
Conclusion: Since both (I) and (II) hold, X = M0 , T k , M1 , . . . , Mk , T k+1 is an execution sequence of PN(P, T, E, C, W, R, I, Q, QW) (w.r.t M0 ) iff there is an answer set A of Π5 (PN, M0 , k, ntok)
such that (A.11) and (A.12) hold.
A.7
Proof of Proposition 7
Let PN = (P, T, E, C, W, R, I, Q, QW, Z) be a Petri Net, M0 be its initial marking and let Π6 (PN, M0 , k, ntok)
be the ASP encoding of PN and M0 over a simulation length k, with maximum ntok tokens on
any place node, as defined in section 2.10. Then X = M0 , T k , M1 , . . . , Mk , T k , Mk+1 is an execution sequence of PN (w.r.t. M0 ) iff there is an answer set A of Π6 (PN, M0 , k, ntok) such
that:
{ f ires(t, ts) : t ∈ T ts , 0 ≤ ts ≤ k} = { f ires(t, ts) : f ires(t, ts) ∈ A}
(A.13)
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
{holds(p, q, c, ts) : p ∈ P, c/q = Mts (p), 0 ≤ ts ≤ k + 1}
= {holds(p, q, c, ts) : holds(p, q, c, ts) ∈ A}
249
(A.14)
We prove this by showing that:
(I) Given an execution sequence X, we create a set A such that it satisfies (A.13) and (A.14)
and show that A is an answer set of Π6
(II) Given an answer set A of Π6 , we create an execution sequence X such that (A.13) and
(A.14) are satisfied.
First we show (I): We create a set A as a union of the following sets:
1. A1 = {num(n) : 0 ≤ n ≤ ntok}
2. A2 = {time(ts) : 0 ≤ ts ≤ k}
3. A3 = {place(p) : p ∈ P}
4. A4 = {trans(t) : t ∈ T }
5. A5 = {color(c) : c ∈ C}
6. A6 = {ptarc(p, t, nc , c, ts) : (p, t) ∈ E − , c ∈ C, nc = mW(p,t) (c), nc > 0, 0 ≤ ts ≤ k}, where
E− ⊆ E
7. A7 = {tparc(t, p, nc , c, ts) : (t, p) ∈ E + , c ∈ C, nc = mW(t,p) (c), nc > 0, 0 ≤ ts ≤ k}, where
E+ ⊆ E
8. A8 = {holds(p, qc , c, 0) : p ∈ P, c ∈ C, qc = m M0 (p) (c)}
9. A9 = {ptarc(p, t, nc , c, ts) : p ∈ R(t), c ∈ C, nc = m Mts (p) , nc > 0, 0 ≤ ts ≤ k}
10. A10 = {iptarc(p, t, 1, c, ts) : p ∈ I(t), c ∈ C, 0 ≤ ts < k}
11. A11 = {tptarc(p, t, nc , c, ts) : (p, t) ∈ Q, c ∈ C, nc = mQW(p,t) (c), nc > 0, 0 ≤ ts ≤ k}
12. A12 = {notenabled(t, ts) : t ∈ T, 0 ≤ ts ≤ k, ∃c ∈ C, (∃p ∈ •t, m Mts (p) (c) < mW(p,t) (c)) ∨
(∃p ∈ I(t), m Mts (p) (c) > 0) ∨ (∃(p, t) ∈ Q, m Mts (p) (c) < mQW(p,t) (c))}
per definition 40 (enabled transition)
13. A13 = {enabled(t, ts) : t ∈ T, 0 ≤ ts ≤ k, ∀c ∈ C, (∀p ∈ •t, mW(p,t) (c) ≤ m Mts (p) (c)) ∧ (∀p ∈
I(t), m Mts (p) (c) = 0) ∧ (∀(p, t) ∈ Q, m Mts (p) (c) ≥ mQW(p,t) (c))}
per definition 40 (enabled transition)
14. A14 = { f ires(t, ts) : t ∈ T ts , 0 ≤ ts ≤ k}
per definition 41 (firing set), only an enabled transition may fire
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
250
15. A15 = {add(p, qc , t, c, ts) : t ∈ T ts , p ∈ t•, c ∈ C, qc = mW(t,p) (c), 0 ≤ ts ≤ k}
per definition 33 (transition execution)
16. A16 = {del(p, qc , t, c, ts) : t ∈ T ts , p ∈ •t, c ∈ C, qc = mW(p,t) (c), 0 ≤ ts ≤ k} ∪
{del(p, qc , t, c, ts) : t ∈ T ts , p ∈ R(t), c ∈ C, qc = m Mts (p) (c), 0 ≤ ts ≤ k}
per definition 33 (transition execution)
P
17. A17 = {tot incr(p, qc , c, ts) : p ∈ P, c ∈ C, qc = t∈Tts ,p∈t• mW(t,p) (c), 0 ≤ ts ≤ k}
per definition 36 (firing set execution)
P
18. A18 = {tot decr(p, qc , c, ts) : p ∈ P, c ∈ C, qc = t∈T ts ,p∈•t mW(p,t) (c) +
P
t∈T ts ,p∈R(t) m Mts (p) (c), 0 ≤ ts ≤ k}
per definition 36 (firing set execution)
19. A19 = {consumesmore(p, ts) : p ∈ P, c ∈ C, qc = m Mts (p) (c),
P
P
q1c = t∈Tts ,p∈•t mW(p,t) (c) + t∈Tts ,p∈R(t) m Mts (p) (c), q1c > qc , 0 ≤ ts ≤ k}
per definition 34 (conflicting transitions)
20. A20 = {consumesmore : ∃p ∈ P, c ∈ C, qc = m Mts (p) (c),
P
P
q1c = t∈Tts ,p∈•t mW(p,t) (c) + t∈Tts ,p∈R(t) (m Mts (p) (c)), q1c > qc , 0 ≤ ts ≤ k}
per definition 34 (conflicting transitions)
21. A21 = {holds(p, qc , c, ts + 1) : p ∈ P, c ∈ C, qc = m Mts+1 (p) (c), 0 ≤ ts < k},
P
P
where Mts+1 (p) = M j (p) − ( t∈Tts ,p∈•t W(p, t) + t∈Tts ,p∈R(t) Mts (p)) +
P
t∈T ts ,p∈t• W(t, p)
according to definition 36 (firing set execution)
22. A22 = {transpr(t, pr) : t ∈ T, pr = Z(t)}
23. A23 = {notprenabled(t, ts) : t ∈ T, enabled(t, ts) ∈ A13 , (∃tt ∈ T, enabled(tt, ts) ∈
A13 , Z(tt) < Z(t)), 0 ≤ ts ≤ k}
= {notprenabled(t, ts) : t ∈ T, (∀p ∈ •t, W(p, t) ≤ Mts (p)), (∀p ∈ I(t), Mts (p) =
0), (∀(p, t) ∈ Q, M0 (p) ≥ WQ(p, t)), ∃tt ∈ T, (∀pp ∈ •tt, W(pp, tt) ≤ Mts (pp)),
(∀pp ∈ I(tt), Mts (pp) = 0), (∀(pp, tt) ∈ Q, Mts (pp) ≥ QW(pp, tt)), Z(tt) < Z(t), 0 ≤ ts ≤
k}
24. A24 = {prenabled(t, ts) : t ∈ T, enabled(t, ts) ∈ A13 , (@tt ∈ T : enabled(tt, ts) ∈
A13 , Z(tt) < Z(t)), 0 ≤ ts ≤ k} = {prenabled(t, ts) : t ∈ T, (∀p ∈ •t, W(p, t) ≤ Mts (p)), (∀p ∈
I(t), Mts (p) = 0), (∀(p, t) ∈ Q, M0 (p) ≥ WQ(p, t)), @tt ∈ T, (∀pp ∈ •tt, W(pp, tt) ≤
Mts (pp)), (∀pp ∈ I(tt), Mts (pp) = 0), (∀(pp, tt) ∈ Q, Mts (pp) ≥ QW(pp, tt)), Z(tt) <
Z(t), 0 ≤ ts ≤ k}
25. A25 = {could not have(t, ts) : t ∈ T, prenabled(t, ts) ∈ A24 , f ires(t, ts) < A14 , (∃p ∈
P
P
•t ∪ R(t) : q > Mts (p) − ( t0 ∈Tts ,p∈•t0 W(p, t0 ) + t0 ∈Tts ,p∈R(t0 ) Mts (p)), q = W(p, t) if (p, t) ∈
E − or R(t) otherwise), 0 ≤ ts ≤ k} = {could not have(t, ts) : t ∈ T, (∀p ∈ •t, W(p, t) ≤
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
251
Mts (p)), (∀p ∈ I(t), Mts (p) = 0), (∀(p, t) ∈ Q, M0 (p) ≥ QW(p, t)), (@tt ∈ T, (∀pp ∈
•tt, W(pp, tt) ≤ Mts (pp)), (∀pp ∈ I(tt), Mts (pp) = 0), (∀(pp, tt) ∈ Q, Mts (pp) ≥ QW(pp, tt)), Z(tt) <
P
P
Z(t)), t < T ts , (∃p ∈ •t∪R(t) : q > Mts (p)−( t0 ∈T ts ,p∈•t0 W(p, t0 )+ t0 ∈Tts ,p∈R(t0 ) Mts (p)), q =
W(p, t) if
(p, t) ∈ E − or R(t) otherwise), 0 ≤ ts ≤ k}
per the maximal firing set semantics
We show that A satisfies (A.13) and (A.14), and A is an answer set of Π6 .
A satisfies (A.13) and (A.14) by its construction above. We show A is an answer set of Π6
by splitting. We split lit(Π6 ) into a sequence of 9k + 11 sets:
• U0 = head( f 1) ∪ head( f 2) ∪ head( f 13) ∪ head( f 5) ∪ head( f 6) ∪ head(i2) ∪ head( f 19) =
{place(p) : p ∈ P} ∪ {trans(t) : t ∈ T } ∪ {col(c) : c ∈ C} ∪ {time(0), . . . , time(k)} ∪
{num(0), . . . , num(ntok)}∪{holds(p, qc , c, 0) : p ∈ P, c ∈ C, qc = m M0 (p) (c)}∪{transpr(t, pr) :
t ∈ T, pr = Z(t)}
• U9k+1 = U9k+0 ∪head( f 14)ts=k ∪head( f 15)ts=k ∪head( f 16)ts=k ∪head( f 17)ts=k ∪head( f 18)ts=k =
U10k+0 ∪ {ptarc(p, t, nc , c, k) : (p, t) ∈ E − , c ∈ C, nc = mW(p,t) (c)}
∪ {tparc(t, p, nc , c, k) : (t, p) ∈ E + , c ∈ C, nc = mW(t,p) (c)} ∪ {ptarc(p, t, nc , c, k) : p ∈
R(t), c ∈ C, nc = m Mk (p) (c), n > 0}∪{iptarc(p, t, 1, c, k) : p ∈ I(t), c ∈ C}∪{tptarc(p, t, nc , c, k) :
(p, t) ∈ Q, c ∈ C, nc = mQW(p,t) (c)}
• U9k+2 = U9k+1 ∪ head(e6)ts=k ∪ head(e7)ts=k ∪ head(e8)ts=k = U9k+1
∪ {notenabled(t, k) : t ∈ T }
• U9k+3 = U9k+2 ∪ head(e9)ts=k = U9k+2 ∪ {enabled(t, k) : t ∈ T }
• U9k+4 = U9k+3 ∪ head(a11)ts=k = U9k+3 ∪ {notprenabled(t, k) : t ∈ T }
• U9k+5 = U9k+4 ∪ head(a12)ts=k = U9k+4 ∪ {prenabled(t, k) : t ∈ T }
• U9k+6 = U9k+5 ∪ head(a13)ts=k = U9k+5 ∪ { f ires(t, k) : t ∈ T }
• U9k+7 = U9k+6 ∪ head(r8)ts=k ∪ head(r9)ts=k = U9k+6 ∪ {add(p, qc , t, c, k) : p ∈ P, t ∈
T, c ∈ C, qc = mW(t,p) (c)} ∪ {del(p, qc , t, c, k) : p ∈ P, t ∈ T, c ∈ C, qc = mW(p,t) (c)} ∪
{del(p, qc , t, c, k) : p ∈ P, t ∈ T, c ∈ C, qc = m Mk (p) (c)}
• U9k+8 = U9k+7 ∪ head(r10)ts=k ∪ head(r11)ts=k = U9k+7 ∪ {tot incr(p, qc , c, k) : p ∈ P, c ∈
C, 0 ≤ qc ≤ ntok} ∪ {tot decr(p, qc , c, k) : p ∈ P, c ∈ C, 0 ≤ qc ≤ ntok}
• U9k+9 = U9k+8 ∪ head(a9)ts=k ∪ head(r12)ts=k ∪ head(a14)ts=k = U9k+8 ∪
{consumesmore(p, k) : p ∈ P}∪{holds(p, q, k+1) : p ∈ P, 0 ≤ q ≤ ntok}∪{could not have(t, k) :
t ∈ T}
• U9k+10 = U9k+9 ∪ head(a3)ts=k = U9k+9 ∪ {consumesmore}
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
252
where head(ri )ts=k are head atoms of ground rule ri in which ts = k. We write Ats=k
=
i
{a(. . . , ts) : a(. . . , ts) ∈ Ai , ts = k} as short hand for all atoms in Ai with ts = k. Uα , 0 ≤
α ≤ 9k + 10 form a splitting sequence, since each Ui is a splitting set of Π6 , and hUα iα<µ is a
S
monotone continuous sequence, where U0 ⊆ U1 · · · ⊆ U9k+10 and α<µ Uα = lit(Π6 ).
We compute the answer set of Π6 using the splitting sets as follows:
1. botU0 (Π6 ) = f 1 ∪ f 2 ∪ f 13 ∪ f 5 ∪ f 6 ∪ i2 ∪ f 19 and X0 = A1 ∪ · · · ∪ A5 ∪ A8 (= U0 ) is
its answer set – using forced atom proposition
2. evalU0 (botU1 (Π6 ) \ botU0 (Π6 ), X0 ) = {ptarc(p, t, qc , c, 0):-.|c ∈ C, qc = mW(p,t) (c), qc >
0} ∪ {tparc(t, p, qc , c, 0):-.|c ∈ C, qc = mW(t,p) (c), qc > 0} ∪ {ptarc(p, t, qc , c, 0):-.|c ∈ C,
qc = m M0 (p) (c), qc > 0} ∪ {iptarc(p, t, 1, c, 0):-.|c ∈ C} ∪ {tptarc(p, t, qc , c, 0):-.|c ∈ C, qc =
ts=0
mQW(p,t) (c), qc > 0}. Its answer set X1 = Ats=0
∪ Ats=0
∪ Ats=0
∪ Ats=0
7
9
10 ∪ A11 – using
6
forced atom proposition and construction of A6 , A7 , A9 , A10 , A11 .
3. evalU1 (botU2 (Π6 ) \ botU1 (Π6 ), X0 ∪ X1 ) = {notenabled(t, 0):-.|({trans(t),
ptarc(p, t, nc , c, 0), holds(p, qc , c, 0)} ⊆ X0 ∪ X1 , where qc < nc ) or
({notenabled(t, 0):-.|({trans(t), iptarc(p, t, n2c , c, 0), holds(p, qc , c, 0)} ⊆ X0 ∪ X1 ,
where qc ≥ n2c }) or ({trans(t), tptarc(p, t, n3c , c, 0), holds(p, qc , c, 0)} ⊆ X0 ∪ X1 , where
qc < n3c )}. Its answer set X2 = Ats=0
12 – using forced atom proposition and construction of
A12 .
(a) where, qc = m M0 (p) (c), and nc = mW(p,t) (c) for an arc (p, t) ∈ E − – by construction
of i2 and f 14 in Π6 , and
(b) in an arc (p, t) ∈ E − , p ∈ •t (by definition 31 of preset)
(c) n2c = 1 – by construction of iptarc predicates in Π6 , meaning qc ≥ n2c ≡ qc ≥ 1 ≡
qc > 0,
(d) tptarc(p, t, n3c , c, 0) represents n3c = mQW(p,t) (c), where (p, t) ∈ Q
(e) thus, notenabled(t, 0) ∈ X1 represents ∃c ∈ C, (∃p ∈ •t : m M0 (p) (c) < mW(p,t) (c)) ∨
(∃p ∈ I(t) : m M0 (p) (c) > 0) ∨ (∃(p, t) ∈ Q : m M0 (p) (c) < mQW(p,t) (c)).
4. evalU2 (botU3 (Π6 ) \ botU2 (Π6 ), X0 ∪ · · · ∪ X2 ) = {enabled(t, 0):-.|trans(t) ∈ X0 ∪ · · · ∪
X2 , notenabled(t, 0) < X0 ∪ · · · ∪ X2 }. Its answer set is X3 = Ats=0
13 – using forced atom
proposition and construction of A13 .
(a) since an enabled(t, 0) ∈ X3 if @ notenabled(t, 0) ∈ X0 ∪ · · · ∪ X2 ; which is equivalent
to @t, ∀c ∈ C, (@p ∈ •t, m M0 (p) (c) < mW(p,t) (c)), (@p ∈ I(t), m M0 (p) (c) > 0), (@(p, t) ∈
Q : m M0 (p) (c) < mQW(p,t) (c)), ∀c ∈ C, (∀p ∈ •t : m M0 (p) (c) ≥ mW(p,t) (c)), (∀p ∈ I(t) :
m M0 (p) (c) = 0).
5. evalU3 (botU4 (Π6 ) \ botU3 (Π6 ), X0 ∪ · · · ∪ X3 ) = {notprenabled(t, 0):-.|
{enabled(t, 0), transpr(t, p), enabled(tt, 0), transpr(tt, pp)} ⊆ X0 ∪ · · · ∪ X3 , pp < p}. Its
answer set is X4 = Ats=k
23 – using forced atom proposition and construction of A23 .
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
253
(a) enabled(t, 0) represents ∃t ∈ T, ∀c ∈ C, (∀p ∈ •t, m M0 (p) (c) ≥ mW(p,t) (c)), (∀p ∈
I(t), m M0 (p) (c) = 0), (∀(p, t) ∈ Q, m M0 (p) (c) ≥ mQW(p,t) (c))
(b) enabled(tt, 0) represents ∃tt ∈ T, ∀c ∈ C, (∀pp ∈ •tt, m M0 (pp) (c) ≥
mW(pp,tt) (c))∧(∀pp ∈ I(tt), m M0 (pp) (c) = 0), (∀(pp, tt) ∈ Q, m M0 (pp) (c) ≥ mQW(pp,tt) (c))
(c) transpr(t, p) represents p = Z(t) – by construction
(d) transpr(tt, pp) represents pp = Z(tt) – by construction
(e) thus, notprenabled(t, 0) represents ∀c ∈ C, (∀p ∈ •t, m M0 (p) (c) ≥
mW(p,t) (c))∧(∀p ∈ I(t), m M0 (p) (c) = 0), ∃tt ∈ T, (∀pp ∈ •tt, m M0 (pp) (c) ≥ mW(pp,tt) (c))∧
(∀pp ∈ I(tt), m M0 (pp) (c) = 0), Z(tt) < Z(t)
(f) which is equivalent to (∀p ∈ •t : M0 (p) ≥ W(p, t)) ∧ (∀p ∈ I(t), M0 (p) = 0), ∃tt ∈
T, (∀pp ∈ •tt, M0 (pp) ≥ W(pp, tt)), (∀pp ∈ I(tt), M0 (pp) = 0), Z(tt) < Z(t) –
assuming multiset domain C for all operations
6. evalU4 (botU5 (Π6 ) \ botU4 (Π6 ), X0 ∪ · · · ∪ X4 ) = {prenabled(t, 0):-.|enabled(t, 0) ∈ X0 ∪
· · · ∪ X4 , notprenabled(t, 0) < X0 ∪ · · · ∪ X4 }. Its answer set is X5 = Ats=k
24 – using forced
atom proposition and construction of A24
(a) enabled(t, 0) represents ∀c ∈ C, (∀p ∈ •t, m M0 (p) (c) ≥ mW(p,t) (c)), (∀p ∈ I(t), m M0 (p) (c) =
0), (∀(p, t) ∈ Q, m M0 (p) (c) ≥ mQW(p,t) (c)) ≡ (∀p ∈ •t,
M0 (p) ≥ W(p, t)), (∀p ∈ I(t), M0 (p) = 0), (∀(p, t) ∈ Q, M0 (p) ≥ QW(p, t)) – from 4
above and assuming multiset domain C for all operations
(b) notprenabled(t, 0) represents (∀p ∈ •t, M0 (p) ≥ W(p, t)), (∀p ∈ I(t),
M0 (p) = 0), (∀(p, t) ∈ Q, M0 (p) ≥ QW(p, t)), ∃tt ∈ T, (∀pp ∈ •tt, M0 (pp) ≥
W(pp, tt)), (∀pp ∈ I(tt), M0 (pp) = 0), (∀(pp, tt) ∈ Q, M0 (pp) ≥ W(pp, tt)),
Z(tt) < Z(t) – from 5 above and assuming multiset domain C for all operations
(c) then, prenabled(t, 0) represents (∀p ∈ •t, M0 (p) ≥ W(p, t)), (∀p ∈ I(t),
M0 (p) = 0), (∀(p, t) ∈ Q, M0 (p) ≥ QW(p, t)), @tt ∈ T, ((∀pp ∈ •tt, M0 (pp) ≥
W(pp, tt)), (∀pp ∈ I(tt), M0 (pp) = 0), (∀(pp, tt) ∈ Q, M0 (pp) ≥ W(pp, tt)),
Z(tt) < Z(t)) – from (a), (b) and enabled(t, 0) ∈ X0 ∪ · · · ∪ X4
7. evalU5 (botU6 (Π6 ) \ botU5 (Π6 ), X0 ∪ · · · ∪ X5 ) = {{ f ires(t, 0)}:-.|prenabled(t, 0)
holds in X0 ∪ · · · ∪ X5 }. It has multiple answer sets X6.1 , . . . , X6.n , corresponding to elements of power set of f ires(t, 0) atoms in evalU5 (...) – using supported rule proposition.
Since we are showing that the union of answer sets of Π6 determined using splitting is
equal to A, we only consider the set that matches the f ires(t, 0) elements in A and call it
X6 , ignoring the rest. Thus, X6 = Ats=0
14 , representing T 0 .
(a) in addition, for every t such that prenabled(t, 0) ∈ X0 ∪· · ·∪X5 , R(t) , ∅; f ires(t, 0) ∈
X6 – per definition 41 (firing set); requiring that a reset transition is fired when
enabled
(b) thus, the firing set T 0 will not be eliminated by the constraint f 20
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
254
8. evalU6 (botU7 (Π6 ) \ botU6 (Π6 ), X0 ∪ · · · ∪ X6 ) = {add(p, nc , t, c, 0):-.|{ f ires(t, 0),
tparc(t, p, nc , c, 0)} ⊆ X0 ∪ · · · ∪ X6 } ∪ {del(p, nc , t, c, 0):-.|{ f ires(t, 0),
ptarc(p, t, nc , c, 0)} ⊆ X0 ∪ · · · ∪ X6 }. It’s answer set is X7 = Ats=0
∪ Ats=0
– using forced
15
16
atom proposition and definitions of A15 and A16 .
(a) where, each add atom is equivalent to nc = mW(t,p) (c), c ∈ C, p ∈ t•,
(b) and each del atom is equivalent to nc = mW(p,t) (c), c ∈ C, p ∈ •t; or nc = m M0 (p) (c), c ∈
C, p ∈ R(t),
(c) representing the effect of transitions in T 0 – by construction
P
9. evalU7 (botU8 (Π6 )\botU7 (Π6 ), X0 ∪· · ·∪X7 ) = {tot incr(p, qqc , c, 0):-.|qqc = add(p,qc ,t,c,0)∈X0 ∪···∪X7 qc }∪
P
{tot decr(p, qqc , c, 0):-.|qqc = del(p,qc ,t,c,0)∈X0 ∪···∪X7 qc }. It’s answer set is X8 = Ats=0
17 ∪
Ats=0
–
using
forced
atom
proposition
and
definitions
of
A
and
A
.
17
18
18
P
(a) where, each tot incr(p, qqc , c, 0), qqc = add(p,qc ,t,c,0)∈X0 ∪...X7 qc
P
≡ qqc = t∈X6 ,p∈t• mW(p,t) (c),
P
(b) where, each tot decr(p, qqc , c, 0), qqc = del(p,qc ,t,c,0)∈X0 ∪...X7 qc
P
P
≡ qq = t∈X6 ,p∈•t mW(t,p) (c) + t∈X6 ,p∈R(t) m M0 (p) (c),
(c) represent the net effect of transitions in T 0
10. evalU8 (botU9 (Π6 ) \ botU8 (Π6 ), X0 ∪ · · · ∪ X8 ) = {consumesmore(p, 0):-.|
{holds(p, qc , c, 0), tot decr(p, q1c , c, 0)} ⊆ X0 ∪ · · · ∪ X8 , q1c > qc } ∪
{holds(p, qc , c, 1):-.|{holds(p, q1c , c, 0), tot incr(p, q2c , c, 0), tot decr(p, q3c , c, 0)}
⊆ X0 ∪ · · · ∪ X6 , qc = q1c + q2c − q3c } ∪ {could not have(t, 0):-.|{prenabled(t, 0),
ptarc(s, t, q, c, 0), holds(s, qq, c, 0), tot decr(s, qqq, c, 0)} ⊆ X0 ∪ · · · ∪ X8 ,
ts=0
ts=0
f ires(t, 0) < X0 ∪ · · · ∪ X8 , q > qq − qqq}. It’s answer set is X9 = Ats=0
19 ∪ A21 ∪ A25 –
using forced atom proposition and definitions of A19 , A21 , A25 .
(a) where, consumesmore(p, 0) represents ∃p : qc = m M0 (p) (c), q1c =
P
P
t∈T 0 ,p∈•t mW(p,t) (c) + t∈T 0 ,p∈R(t) m M0 (p) (c), q1c > qc , c ∈ C, indicating place p will
be over consumed if T 0 is fired, as defined in definition 34 (conflicting transitions),
P
(b) holds(p, qc , c, 1) if qc = m M0 (p) (c) + t∈T0 ,p∈t• mW(t,p) (c) −
P
P
( t∈T0 ,p∈•t mW(p,t) (c) + t∈T0 ,p∈R(t) m M0 (p) (c)) represented by qc = m M1 (p) (c) for some
c ∈ C – by construction of Π6
(c) could not have(t, 0) if
i. (∀p ∈ •t, W(p, t) ≤ M0 (p)), (∀p ∈ I(t), M0 (p) = 0), (∀(p, t) ∈ Q,
M0 (p) ≥ WQ(p, t)), @tt ∈ T, (∀pp ∈ •tt, W(pp, tt) ≤ Mts (pp)), (∀pp ∈ I(tt), M0 (pp) =
0), (∀(pp, tt) ∈ Q, M0 (pp) ≥ QW(pp, tt)), Z(tt) < Z(t),
P
P
ii. and qc > m M0 (s) (c) − ( t0 ∈T0 ,s∈•t0 mW(s,t0 ) (c) + t0 ∈T 0 ,s∈R(t) m M0 (s) (c)),
qc = mW(s,t) (c) if s ∈ •t or m M0 (s) (c) otherwise for some c ∈ C, which beP
P
comes q > M0 (s) − ( t0 ∈T 0 ,s∈•t0 W(s, t0 ) + t0 ∈T0 ,s∈R(t) M0 (s)), q = W(s, t) if s ∈
•t or M0 (s) otherwise for all c ∈ C
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
255
iii. (i), (ii) above combined match the definition of A25
(d) X9 does not contain could not have(t, 0), when prenabled(t, 0) ∈ X0 ∪ · · · ∪ X8 and
f ires(t, 0) < X0 ∪ · · · ∪ X8 due to construction of A, encoding of a9 and its body
atoms. As a result, it is not eliminated by the constraint a6
..
.
11. evalU9k+0 (botU9k+1 (Π6 ) \ botU9k+0 (Π6 ), X0 ∪ · · · ∪ X9k+0 ) = {ptarc(p, t, qc , c, k):-.|c ∈ C, qc =
mW(p,t) (c), qc > 0}∪{tparc(t, p, qc , c, k):-.|c ∈ C, qc = mW(t,p) (c), qc > 0}∪{ptarc(p, t, qc , c, k):-.|c ∈
C, qc = m M0 (p) (c), qc > 0} ∪ {iptarc(p, t, 1, c, k):-.|c ∈ C} ∪ {tptarc(p, t, qc , c, k):-.|c ∈
ts=k
C, qc = mQW(p,t) (c), qc > 0}. Its answer set X9k+1 = Ats=k
∪ Ats=k
∪ Ats=k
∪ Ats=k
7
9
10 ∪ A11 –
6
using forced atom proposition and construction of A6 , A7 , A9 , A10 , A11 .
12. evalU9k+1 (botU9k+2 (Π6 ) \ botU9k+1 (Π6 ), X0 ∪ · · · ∪ X9k+1 ) = {notenabled(t, k):-.| ({trans(t),
ptarc(p, t, nc , c, k), holds(p, qc , c, k)} ⊆ X0 ∪· · ·∪X10k+1 , where qc < nc ) or ({notenabled(t, k):-.|
({trans(t), iptarc(p, t, n2c , c, k), holds(p, qc , c, k)} ⊆ X0 ∪· · ·∪X10k+1 , where qc ≥ n2c }) or
({trans(t), tptarc(p, t, n3c , c, k),
holds(p, qc , c, k)} ⊆ X0 ∪ X9k+1 , where qc < n3c )}. Its answer set X9k+2 = Ats=k
12 – using
forced atom proposition and construction of A12 .
(a) where, qc = m M0 (p) (c), and nc = mW(p,t) (c) for an arc (p, t) ∈ E − – by construction
of i2 and f 14 predicates in Π6 , and
(b) in an arc (p, t) ∈ E − , p ∈ •t (by definition 31 of preset)
(c) n2c = 1 – by construction of iptarc predicates in Π6 , meaning qc ≥ n2c ≡ qc ≥ 1 ≡
qc > 0,
(d) tptarc(p, t, n3c , c, k) represents n3c = mQW(p,t) (c), where (p, t) ∈ Q
(e) thus, notenabled(t, k) ∈ X9k+1 represents ∃c ∈ C, (∃p ∈ •t : m M0 (p) (c) < mW(p,t) (c))∨
(∃p ∈ I(t) : m M0 (p) (c) > k) ∨ (∃(p, t) ∈ Q : m M0 (p) (c) < mQW(p,t) (c)).
13. evalU9k+2 (botU9k+3 (Π6 ) \ botU9k+2 (Π6 ), X0 ∪ · · · ∪ X9k+2 ) = {enabled(t, k):-.|trans(t) ∈ X0 ∪
· · · ∪ X9k+2 , notenabled(t, k) < X0 ∪ · · · ∪ X9k+2 }. Its answer set is X9k+3 = Ats=k
13 – using
forced atom proposition and construction of A13 .
(a) since an enabled(t, k) ∈ X9k+3 if @ notenabled(t, k) ∈ X0 ∪ · · · ∪ X9k+2 ; which is
equivalent to @t, ∀c ∈ C, (@p ∈ •t, m M0 (p) (c) < mW(p,t) (c)), (@p ∈ I(t), m M0 (p) (c) >
k), (@(p, t) ∈ Q : m M0 (p) (c) < mQW(p,t) (c)), ∀c ∈ C, (∀p ∈ •t : m M0 (p) (c) ≥
mW(p,t) (c)), (∀p ∈ I(t) : m M0 (p) (c) = k).
14. evalU9k+3 (botU9k+4 (Π6 ) \ botU9k+3 (Π6 ), X0 ∪ · · · ∪ X9k+3 ) = {notprenabled(t, k):-.|
{enabled(t, k), transpr(t, p), enabled(tt, k), transpr(tt, pp)} ⊆ X0 ∪ · · · ∪ X9k+3 ,
pp < p}. Its answer set is X9k+4 = Ats=k
23 – using forced atom proposition and construction
of A23 .
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
256
(a) enabled(t, k) represents ∃t ∈ T, ∀c ∈ C, (∀p ∈ •t, m M0 (p) (c) ≥ mW(p,t) (c)),
(∀p ∈ I(t), m M0 (p) (c) = k), (∀(p, t) ∈ Q, m M0 (p) (c) ≥ mQW(p,t) (c))
(b) enabled(tt, k) represents ∃tt ∈ T, ∀c ∈ C, (∀pp ∈ •tt, m M0 (pp) (c) ≥
mW(pp,tt) (c))∧(∀pp ∈ I(tt), m M0 (pp) (c) = k), (∀(pp, tt) ∈ Q, m M0 (pp) (c) ≥ mQW(pp,tt) (c))
(c) transpr(t, p) represents p = Z(t) – by construction
(d) transpr(tt, pp) represents pp = Z(tt) – by construction
(e) thus, notprenabled(t, k) represents ∀c ∈ C, (∀p ∈ •t, m M0 (p) (c) ≥ mW(p,t) (c))
∧ (∀p ∈ I(t), m M0 (p) (c) = k), ∃tt ∈ T, (∀pp ∈ •tt, m M0 (pp) (c) ≥ mW(pp,tt) (c)) ∧ (∀pp ∈
I(tt), m M0 (pp) (c) = k), Z(tt) < Z(t)
(f) which is equivalent to (∀p ∈ •t : M0 (p) ≥ W(p, t)) ∧ (∀p ∈ I(t), M0 (p) = k), ∃tt ∈
T, (∀pp ∈ •tt, M0 (pp) ≥ W(pp, tt)), (∀pp ∈ I(tt), M0 (pp) = k), Z(tt) < Z(t) –
assuming multiset domain C for all operations
15. evalU9k+4 (botU9k+5 (Π6 )\botU9k+4 (Π6 ), X0 ∪· · ·∪ X9k+4 ) = {prenabled(t, k):-.| enabled(t, k) ∈
X0 ∪ · · · ∪ X9k+4 , notprenabled(t, k) < X0 ∪ · · · ∪ X9k+4 }. Its answer set is X9k+5 = Ats=k
24 –
using forced atom proposition and construction of A24
(a) enabled(t, k) represents ∀c ∈ C, (∀p ∈ •t, m M0 (p) (c) ≥ mW(p,t) (c)), (∀p ∈ I(t), m M0 (p) (c) =
k), (∀(p, t) ∈ Q, m M0 (p) (c) ≥ mQW(p,t) (c)) ≡ (∀p ∈ •t, M0 (p) ≥ W(p, t)), (∀p ∈
I(t), M0 (p) = k), (∀(p, t) ∈ Q, M0 (p) ≥ QW(p, t)) – from 13 above and assuming
multiset domain C for all operations
(b) notprenabled(t, k) represents (∀p ∈ •t, M0 (p) ≥ W(p, t)), (∀p ∈ I(t),
M0 (p) = k), (∀(p, t) ∈ Q, M0 (p) ≥ QW(p, t)), ∃tt ∈ T, (∀pp ∈ •tt, M0 (pp) ≥
W(pp, tt)), (∀pp ∈ I(tt), M0 (pp) = k), (∀(pp, tt) ∈ Q, M0 (pp) ≥ W(pp, tt)),
Z(tt) < Z(t) – from 14 above and assuming multiset domain C for all operations
(c) then, prenabled(t, k) represents (∀p ∈ •t, M0 (p) ≥ W(p, t)), (∀p ∈ I(t),
M0 (p) = k), (∀(p, t) ∈ Q, M0 (p) ≥ QW(p, t)), @tt ∈ T, ((∀pp ∈ •tt, M0 (pp) ≥
W(pp, tt)), (∀pp ∈ I(tt), M0 (pp) = k), (∀(pp, tt) ∈ Q, M0 (pp) ≥ W(pp, tt)),
Z(tt) < Z(t)) – from (a), (b) and enabled(t, k) ∈ X0 ∪ · · · ∪ X9k+4
16. evalU9k+5 (botU9k+6 (Π6 ) \ botU9k+5 (Π6 ), X0 ∪ · · · ∪ X9k+5 ) = {{ f ires(t, k)}:-.|
prenabled(t, k) holds in X0 ∪ · · · ∪ X9k+5 }. It has multiple answer sets X9k+6.1 , . . . ,
X9k+6.n , corresponding to elements of power set of f ires(t, k) atoms in evalU9k+5 (...) –
using supported rule proposition. Since we are showing that the union of answer sets
of Π6 determined using splitting is equal to A, we only consider the set that matches
the f ires(t, k) elements in A and call it X9k+6 , ignoring the rest. Thus, X9k+6 = Ats=k
14 ,
representing T k .
(a) in addition, for every t such that prenabled(t, k) ∈ X0 ∪ · · · ∪ X9k+5 , R(t) , ∅;
f ires(t, k) ∈ X9k+6 – per definition 41 (firing set); requiring that a reset transition
is fired when enabled
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
257
(b) thus, the firing set T k will not be eliminated by the constraint f 20
17. evalU9k+6 (botU9k+7 (Π6 ) \ botU9k+6 (Π6 ), X0 ∪ · · · ∪ X9k+6 ) = {add(p, nc , t, c, k):-.|
{ f ires(t, k), tparc(t, p, nc , c, k)} ⊆ X0 ∪ · · · ∪ X9k+6 } ∪ {del(p, nc , t, c, k):-.|
{ f ires(t, k), ptarc(p, t, nc , c, k)} ⊆ X0 ∪ · · · ∪ X9k+6 }. It’s answer set is X9k+7 = Ats=k
∪ Ats=k
15
16
– using forced atom proposition and definitions of A15 and A16 .
(a) where, each add atom is equivalent to nc = mW(t,p) (c), c ∈ C, p ∈ t•,
(b) and each del atom is equivalent to nc = mW(p,t) (c), c ∈ C, p ∈ •t; or nc = m M0 (p) (c), c ∈
C, p ∈ R(t),
(c) representing the effect of transitions in T k .
18. evalU9k+7 (botU9k+8 (Π6 ) \ botU9k+7 (Π6 ), X0 ∪ · · · ∪ X9k+7 ) = {tot incr(p, qqc , c, k):-.| qqc =
P
q } ∪ {tot decr(p, qqc , c, k):-.|qqc =
Padd(p,qc ,t,c,k)∈X0 ∪···∪X10k+7 c
ts=k
ts=k
del(p,qc ,t,c,k)∈X0 ∪···∪X9k+7 qc }. It’s answer set is X10k+8 = A17 ∪ A18 – using forced atom
proposition and definitions of A17 and A18 .
P
(a) where, each tot incr(p, qqc , c, k), qqc = add(p,qc ,t,c,k)∈X0 ∪...X10k+7 qc
P
≡ qqc = t∈X9k+6 ,p∈t• mW(p,t) (c),
P
(b) and each tot decr(p, qqc , c, k), qqc = del(p,qc ,t,c,k)∈X0 ∪...X10k+7 qc
P
P
≡ qq = t∈X10k+6 ,p∈•t mW(t,p) (c) + t∈X9k+6 ,p∈R(t) m M0 (p) (c),
(c) represent the net effect of transitions in T k
19. evalU9k+8 (botU9k+9 (Π6 ) \ botU9k+8 (Π6 ), X0 ∪ · · · ∪ X9k+8 ) = {consumesmore(p, k):-.|
{holds(p, qc , c, k), tot decr(p, q1c , c, k)} ⊆ X0 ∪ · · · ∪ X9k+8 , q1c > qc } ∪
{holds(p, qc , c, 1):-.|{holds(p, q1c , c, k), tot incr(p, q2c , c, k), tot decr(p, q3c , c, k)}
⊆ X0 ∪ · · · ∪ X9k+6 , qc = q1c + q2c − q3c } ∪ {could not have(t, k):-.|{prenabled(t, k),
ptarc(s, t, q, c, k), holds(s, qq, c, k), tot decr(s, qqq, c, k)} ⊆ X0 ∪ · · · ∪ X9k+8 ,
ts=k
ts=k
f ires(t, k) < X0 ∪· · ·∪ X9k+8 , q > qq−qqq}. It’s answer set is X10k+9 = Ats=k
19 ∪ A21 ∪ A25
– using forced atom proposition and definitions of A19 , A21 , A25 .
(a) where, consumesmore(p, k) represents ∃p : qc = m M0 (p) (c),
P
P
q1c = t∈Tk ,p∈•t mW(p,t) (c) + t∈Tk ,p∈R(t) m M0 (p) (c), q1c > qc , c ∈ C, indicating place
p that will be over consumed if T k is fired, as defined in definition 34 (conflicting
transition)
P
(b) holds(p, qc , c, k + 1) if qc = m M0 (p) (c) + t∈Tk ,p∈t• mW(t,p) (c) −
P
P
( t∈Tk ,p∈•t mW(p,t) (c) + t∈Tk ,p∈R(t) m M0 (p) (c)) represented by qc = m M1 (p) (c) for some
c ∈ C – by construction of Π6
(c) could not have(t, k) if
i. (∀p ∈ •t, W(p, t) ≤ M0 (p)), (∀p ∈ I(t), M0 (p) = k), (∀(p, t) ∈ Q, M0 (p) ≥
WQ(p, t)), @tt ∈ T, (∀pp ∈ •tt, W(pp, tt) ≤ Mts (pp)), (∀pp ∈ I(tt), M0 (pp) =
k), (∀(pp, tt) ∈ Q, M0 (pp) ≥ QW(pp, tt)), Z(tt) < Z(t),
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
258
P
P
ii. and qc > m M0 (s) (c)−( t0 ∈Tk ,s∈•t0 mW(s,t0 ) (c)+ t0 ∈Tk ,s∈R(t) m M0 (s) (c)), qc = mW(s,t) (c) if s ∈
P
•t or m M0 (s) (c) otherwise for some c ∈ C, which becomes q > M0 (s)−( t0 ∈T k ,s∈•t0 W(s, t0 )+
P
t0 ∈T k ,s∈R(t) M0 (s)), q = W(s, t) if s ∈ •t or M0 (s) otherwise for all c ∈ C
iii. (i), (ii) above combined match the definition of A25
(d) X9k+9 does not contain could not have(t, k), when prenabled(t, k) ∈ X0 ∪ · · · ∪ X9k+6
and f ires(t, k) < X0 ∪ · · · ∪ X9k+5 due to construction of A, encoding of a10 and its
body atoms. As a result it is not eliminated by the constraint a6
20. evalU9k+9 (botU9k+10 (Π6 ) \ botU9k+9 (Π6 ), X0 ∪ · · · ∪ X9k+9 ) = {consumesmore:-.|
{consumesmore(p, k)} ⊆ A}. It’s answer set is X9k+10 = A20 – using forced atom proposition and the definition of A20
(a) X9k+10 will be empty since none of consumesmore(p, 0), . . . ,
consumesmore(p, k) hold in X0 ∪ · · · ∪ X9k+9 due to the construction of A, encoding
of a9 and its body atoms. As a result, it is not eliminated by the constraint a4
The set X = X0 ∪ · · · ∪ X9k+10 is the answer set of Π6 by the splitting sequence theorem 15.
Each Xi , 0 ≤ i ≤ 9k + 10 matches a distinct portion of A, and X = A, thus A is an answer set of
Π6 .
Next we show (II): Given Π6 be the encoding of a Petri Net PN(P, T, E, C, W, R, I, Q, QW, Z)
with initial marking M0 , and A be an answer set of Π6 that satisfies (A.13) and (A.14), then we
can construct X = M0 , T k , . . . , Mk , T k , Mk+1 from A, such that it is an execution sequence of PN.
We construct the X as follows:
1. Mi = (Mi (p0 ), . . . , Mi (pn )), where {holds(p0 , m Mi (p0 ) (c), c, i), . . . ,
holds(pn , m Mi (pn ) (c), c, i)} ⊆ A, for c ∈ C, 0 ≤ i ≤ k + 1
2. T i = {t : f ires(t, i) ∈ A}, for 0 ≤ i ≤ k
and show that X is indeed an execution sequence of PN. We show this by induction over k (i.e.
given Mk , T k is a valid firing set and its firing produces marking Mk+1 ).
Base case: Let k = 0, and M0 is a valid marking in X for PN, show (1) T 0 is a valid firing set
for M0 , and (2) firing T 0 in M0 produces marking M1 .
1. We show T 0 is a valid firing set for M0 . Let { f ires(t0 , 0), . . . , f ires(t x , 0)} be the set of all
f ires(. . . , 0) atoms in A,
(a) Then for each f ires(ti , 0) ∈ A
i. prenabled(ti , 0) ∈ A – from rule a13 and supported rule proposition
ii. Then enabled(ti , 0) ∈ A – from rule a12 and supported rule proposition
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
259
iii. And notprenabled(ti , 0) < A – from rule a12 and supported rule proposition
iv. For enabled(ti , 0) ∈ A
A. notenabled(ti , 0) < A – from rule e9 and supported rule proposition
B. Then either of body(e6), body(e7), or body(e8) must not hold in A for ti –
from rules e6, e7, e8 and forced atom proposition
C. Then qc ≮ nic ≡ qc ≥ nic in e6 for all {holds(p, qc , c, 0),
ptarc(p, ti , nic , c, 0)} ⊆ A – from e : 6, forced atom proposition, and given
facts (holds(p, qc , c, 0) ∈ A, ptarc(p, ti , nic , 0) ∈ A)
D. And qc nic ≡ qc < nic in e7 for all {holds(p, qc , c, 0),
iptarc(p, ti , nic , c, 0)} ⊆ A, nic = 1; qc > nic ≡ qc = 0 – from e7, forced
atom proposition, given facts (holds(p, qc , c, 0) ∈ A,
iptarc(p, ti , 1, c, 0) ∈ A), and qc is a positive integer
E. And qc ≮ nic ≡ qc ≥ nic in e8 for all {holds(p, qc , c, 0),
tptarc(p, ti , nic , c, 0)} ⊆ A – from e8, forced atom proposition, and given
facts
F. Then ∀c ∈ C, (∀p ∈ •ti , m M0 (p) (c) ≥ mW(p,ti ) (c)) ∧ (∀p ∈ I(ti ),
m M0 (p) (c) = 0)∧(∀(p, ti ) ∈ Q, m M0 (p) (c) ≥ mQW(p,ti ) (c)) – from i2, f 14, f 17, f 18
construction, definition 31 of preset •ti in PN, definition 40 of enabled
transition in PN, and that the construction of reset arcs by f 16 ensures
notenabled(t, 0) is never true for a reset arc, where holds(p, qc , c, 0) ∈ A
represents qc = m M0 (p) (c), ptarc(p, ti , nic , 0) ∈ A represents nic = mW(p,ti ) (c),
nic = m M0 (p) (c).
G. Which is equivalent to (∀p ∈ •ti , M0 (p) ≥ W(p, ti )) ∧ (∀p ∈ I(ti ), M0 (p) =
0) ∧ (∀(p, ti ) ∈ Q, M0 (p) ≥ QW(p, ti )) – assuming multiset domain C
v. For notprenabled(ti , 0) < A
A. Either (@enabled(tt, 0) ∈ A : pp < pi ) or (∀enabled(tt, 0) ∈ A : pp ≮
pi ) where pp = Z(tt), pi = Z(ti ) – from rule a11, f 19 and forced atom
proposition
B. This matches the definition of an enabled priority transition
vi. Then ti is enabled and can fire in PN, as a result it can belong to T 0 – from
definition 40 of enabled transition
(b) And consumesmore < A, since A is an answer set of Π6 – from rule a3 and supported
rule proposition
i. Then @consumesmore(p, 0) ∈ A – from rule a9 and supported rule proposition
ii. Then @{holds(p, qc , c, 0), tot decr(p, q1c , c, 0)} ⊆ A, q1c > qc in body(a9) –
from a9 and forced atom proposition
P
iii. Then @c ∈ C@p ∈ P, ( ti ∈{t0 ,...,tx },p∈•ti mW(p,ti ) (c) +
P
ti ∈{t0 ,...,t x },p∈R(ti ) m M0 (p) (c)) > m M0 (p) (c) – from the following
A. holds(p, qc , c, 0) represents qc = m M0 (p) (c) – from rule i2 encoding, given
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
260
B. tot decr(p, q1c , c, 0) ∈ A if {del(p, q10 c , t0 , c, 0), . . . ,
del(p, q1 x c , t x , c, 0)} ⊆ A, where q1c = q10 c + · · · + q1 x c – from r11 and
forced atom proposition
C. del(p, q1i c , ti , c, 0) ∈ A if { f ires(ti , 0), ptarc(p, ti , q1i c , c, 0)} ⊆ A – from r9
and supported rule proposition
D. del(p, q1i c , ti , c, 0) either represents removal of q1i c = mW(p,ti ) (c) tokens
from p ∈ •ti ; or it represents removal of q1i c = m M0 (p) (c) tokens from p ∈
R(ti )– from rules r9, f 14, f 16, supported rule proposition, and definition 33
of transition execution in PN
iv. Then the set of transitions in T 0 do not conflict – by the definition 34 of conflicting transitions
(c) And for each prenabled(t j , 0) ∈ A and f ires(t j , 0) < A, could not have(t j , 0) ∈ A,
since A is an answer set of Π6 - from rule a15 and supported rule proposition
i. Then {prenabled(t j , 0), holds(s, qqc , c, 0), ptarc(s, t j , qc , c, 0),
tot decr(s, qqqc , c, 0)} ⊆ A, such that qc > qqc − qqqc and f ires(t j , 0) < A from rule a14 and supported rule proposition
P
P
ii. Then for an s ∈ •t j ∪R(t j ), qc > m M0 (s) (c)−( ti ∈T0 ,s∈•ti mW(s,ti ) (c)+ ti ∈T0 ,s∈R(ti ) m M0 (s) (c)),
where qc = mW(s,t j ) (c) if s ∈ •t j , or m M0 (s) (c) otherwise – from the following
A. ptarc(s, ti , qc , c, 0) represents qc = mW(s,ti ) (c) if (s, ti ) ∈ E − or qc = m M0 (s) (c)
if s ∈ R(ti ) – from rule f 14, f 16 construction
B. holds(s, qqc , c, 0) represents qqc = m M0 (s) (c) – from i2 construction
C. tot decr(s, qqqc , c, 0) ∈ A if {del(s, qqq0 c , t0 , c, 0), . . . ,
del(s, qqq x c , t x , c, 0)} ⊆ A – from rule r11 construction and supported rule
proposition
D. del(s, qqqi c , ti , c, 0) ∈ A if { f ires(ti , 0), ptarc(s, ti , qqqi c , c, 0)} ⊆ A – from
rule r9 and supported rule proposition
E. del(s, qqqi c , ti , c, 0) either represents qqqi c = mW(s,ti ) (c) : ti ∈ T 0 , (s, ti ) ∈
E − , or qqqi c = m M0 (ti ) (c) : ti ∈ T 0 , s ∈ R(ti ) – from rule f 14, f 16 construction
P
F. tot decr(q, qqqc , c, 0) represents ti ∈T0 ,s∈•ti mW(s,ti ) (c) +
P
ti ∈T 0 ,s∈R(ti ) m M0 (s) (c) – from (C,D,E) above
iii. Then firing T 0 ∪ {t j } would have required more tokens than are present at its
source place s ∈ •t j ∪ R(t j ). Thus, T 0 is a maximal set of transitions that can
simultaneously fire.
(d) And for each reset transition tr with prenabled(tr , 0) ∈ A, f ires(tr , 0) ∈ A, since A
is an answer set of Π6 - from rule f 20 and supported rule proposition
i. Then, the firing set T 0 satisfies the reset-transition requirement of definition 41
(firing set)
(e) Then {t0 , . . . , t x } = T 0 – using 1(a),1(b),1(d) above; and using 1(c) it is a maximal
firing set
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
261
2. We show M1 is produced by firing T 0 in M0 . Let holds(p, qc , c, 1) ∈ A
(a) Then {holds(p, q1c , c, 0), tot incr(p, q2c , c, 0), tot decr(p, q3c , c, 0)} ⊆ A : qc = q1c +
q2c − q3c – from rule r12 and supported rule proposition
(b) Then, holds(p, q1c , c, 0) ∈ A represents q1c = m M0 (p) (c) – given, construction;
and {add(p, q20 c , t0 , c, 0), . . . , add(p, q2 j c , t j , c, 0)} ⊆ A : q20 c + · · · + q2 j c = q2c and
{del(p, q30 c , t0 , c, 0), . . . , del(p, q3l c , tl , c, 0)} ⊆ A : q30 c + · · · + q3l c = q3c – rules
r10, r11 and supported rule proposition, respectively
(c) Then { f ires(t0 , 0), . . . , f ires(t j , 0)} ⊆ A and { f ires(t0 , 0), . . . , f ires(tl , 0)} ⊆ A –
rules r8, r9 and supported rule proposition, respectively
(d) Then { f ires(t0 , 0), . . . , f ires(t j , 0)}∪ { f ires(t0 , 0), . . . , f ires(tl , 0)} ⊆ A = { f ires(t0 , 0), . . . ,
f ires(t x , 0)} ⊆ A – set union of subsets
(e) Then for each f ires(t x , 0) ∈ A we have t x ∈ T 0 – already shown in item 1 above
P
P
P
(f) Then qc = m M0 (p) (c)+ tx ∈T0 ∧p∈tx • mW(tx ,p) (c)−( tx ∈T0 ∧p∈•tx mW(p,tx ) (c)+ tx ∈T0 ∧p∈R(tx ) m M0 (p) (c))
– from (2b) above and the following
i. Each add(p, q j c , t j , c, 0) ∈ A represents q j c = mW(t j ,p) (c) for p ∈ t j • – rule
r8, f 15 encoding, and definition 33 of transition execution in PN
ii. Each del(p, ty , qy c , c, 0) ∈ A represents either qy c = mW(p,ty ) (c) for p ∈ •ty , or
qy c = m M0 (p) (c) for p ∈ R(ty ) – from rule r9, f 14 encoding and definition 33
of transition execution in PN; or from rule r9, f 16 encoding and definition of
reset arc in PN
P
iii. Each tot incr(p, q2c , c, 0) ∈ A represents q2c = tx ∈T0 ∧p∈tx • mW(tx ,p) (c) – aggregate assignment atom semantics in rule r10
P
P
iv. Each tot decr(p, q3c , c, 0) ∈ A represents q3c = tx ∈T0 ∧p∈•tx mW(p,tx ) (c)+ tx ∈T0 ∧p∈R(tx ) m M0 (p) (c)
– aggregate assignment atom semantics in rule r11
(g) Then, m M1 (p) (c) = qc – since holds(p, qc , c, 1) ∈ A encodes qc = m M1 (p) (c) – from
construction
Inductive Step: Let k > 0, and Mk is a valid marking in X for PN, show (1) T k is a valid firing
set for Mk , and (2) firing T k in Mk produces marking Mk+1 .
1. We show that T k is a valid firing set in Mk . Let { f ires(T k , k), . . . , f ires(t x , k)} be the set of
all f ires(. . . , k) atoms in A,
(a) Then for each f ires(ti , k) ∈ A
i.
ii.
iii.
iv.
prenabled(ti , k) ∈ A – from rule a13 and supported rule proposition
Then enabled(ti , k) ∈ A – from rule a12 and supported rule proposition
And notprenabled(ti , k) < A – from rule a12 and supported rule proposition
For enabled(ti , k) ∈ A
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
262
A. notenabled(ti , k) < A – from rule e9 and supported rule proposition
B. Then either of body(e6), body(e7), or body(e8) must not hold in A for ti –
from rules e6, e7, e8 and forced atom proposition
C. Then qc ≮ nic ≡ qc ≥ nic in e6 for all {holds(p, qc , c, k),
ptarc(p, ti , nic , c, k)} ⊆ A – from e6, forced atom proposition, and given
facts (holds(p, qc , c, k) ∈ A, ptarc(p, ti , nic , k) ∈ A)
D. And qc nic ≡ qc < nic in e7 for all {holds(p, qc , c, k),
iptarc(p, ti , nic , c, k)} ⊆ A, nic = 1; qc > nic ≡ qc = 0 – from e7, forced
atom proposition, given facts (holds(p, qc , c, k) ∈ A,
iptarc(p, ti , 1, c, k) ∈ A), and qc is a positive integer
E. And qc ≮ nic ≡ qc ≥ nic in e8 for all {holds(p, qc , c, k),
tptarc(p, ti , nic , c, k)} ⊆ A – from e8, forced atom proposition, and given
facts
F. Then ∀c ∈ C, (∀p ∈ •ti , m Mk (p) (c) ≥ mW(p,ti ) (c)) ∧ (∀p ∈ I(ti ), m Mk (p) (c) =
k) ∧ (∀(p, ti ) ∈ Q, m Mk (p) (c) ≥ mQW(p,ti ) (c)) – from f 14, f 17, f 18 construction, inductive assumption, definition 31 of preset •ti in PN, definition 40 of enabled transition in PN, and that the construction of reset
arcs by f 16 ensures notenabled(t, k) is never true for a reset arc, where
holds(p, qc , c, k) ∈ A represents qc = m Mk (p) (c), ptarc(p, ti , nic , k) ∈ A represents nic = mW(p,ti ) (c), nic = m Mk (p) (c).
G. Which is equivalent to (∀p ∈ •ti , Mk (p) ≥ W(p, ti )) ∧ (∀p ∈ I(ti ), Mk (p) =
k) ∧ (∀(p, ti ) ∈ Q, Mk (p) ≥ QW(p, ti )) – assuming multiset domain C
v. For notprenabled(ti , k) < A
A. Either (@enabled(tt, k) ∈ A : pp < pi ) or (∀enabled(tt, k) ∈ A : pp ≮
pi ) where pp = Z(tt), pi = Z(ti ) – from rule a11, f 19 and forced atom
proposition
B. This matches the definition of an enabled priority transition
vi. Then ti is enabled and can fire in PN, as a result it can belong to T k – from
definition 40 of enabled transition
(b) And consumesmore < A, since A is an answer set of Π6 – from rule a4 and supported
rule proposition
i. Then @consumesmore(p, k) ∈ A – from rule a4 and supported rule proposition
ii. Then @{holds(p, qc , c, k), tot decr(p, q1c , c, k)} ⊆ A, q1c > qc in body(a9) –
from a9 and forced atom proposition
P
iii. Then @c ∈ C@p ∈ P, ( ti ∈{Tk ,...,tx },p∈•ti mW(p,ti ) (c) +
P
ti ∈{T k ,...,t x },p∈R(ti ) m Mk (p) (c)) > m Mk (p) (c) – from the following
A. holds(p, qc , c, k) represents qc = m Mk (p) (c) – from inductive assumption
and construction, given
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
263
B. tot decr(p, q1c , c, k) ∈ A if {del(p, q10 c , T k , c, k), . . . ,
del(p, q1 x c , t x , c, k)} ⊆ A, where q1c = q10 c + · · · + q1 x c – from r11 and
forced atom proposition
C. del(p, q1i c , ti , c, k) ∈ A if { f ires(ti , k), ptarc(p, ti , q1i c , c, k)} ⊆ A – from r9
and supported rule proposition
D. del(p, q1i c , ti , c, k) either represents removal of q1i c = mW(p,ti ) (c) tokens
from p ∈ •ti ; or it represents removal of q1i c = m Mk (p) (c) tokens from p ∈
R(ti )– from rules r9, f 14, f 16, supported rule proposition, and definition 33
of transition execution in PN
iv. Then the set of transitions in T k do not conflict – by the definition 34 of conflicting transitions
(c) And for each prenabled(t j , k) ∈ A and f ires(t j , k) < A, could not have(t j , k) ∈ A,
since A is an answer set of Π6 - from rule a15 and supported rule proposition
i. Then {prenabled(t j , k), holds(s, qqc , c, k), ptarc(s, t j , qc , c, k),
tot decr(s, qqqc , c, k)} ⊆ A, such that qc > qqc − qqqc and f ires(t j , k) < A from rule a14 and supported rule proposition
P
P
ii. Then for an s ∈ •t j ∪R(t j ), qc > m Mk (s) (c)−( ti ∈Tk ,s∈•ti mW(s,ti ) (c)+ ti ∈Tk ,s∈R(ti ) m Mk (s) (c)),
where qc = mW(s,t j ) (c) if s ∈ •t j , or m Mk (s) (c) otherwise.
A. ptarc(s, ti , qc , c, k) represents qc = mW(s,ti ) (c) if (s, ti ) ∈ E − or qc = m Mk (s) (c)
if s ∈ R(ti ) – from rule f 14, f 16 construction
B. holds(s, qqc , c, k) represents qqc = m Mk (s) (c) – from inductive assumption
and construction
C. tot decr(s, qqqc , c, k) ∈ A if {del(s, qqq0 c , T k , c, k), . . . ,
del(s, qqq x c , t x , c, k)} ⊆ A – from rule r11 construction and supported rule
proposition
D. del(s, qqqi c , ti , c, k) ∈ A if { f ires(ti , k), ptarc(s, ti , qqqi c , c, k)} ⊆ A – from
rule r9 and supported rule proposition
E. del(s, qqqi c , ti , c, k) either represents qqqi c = mW(s,ti ) (c) : ti ∈ T k , (s, ti ) ∈
E − , or qqqi c = m Mk (ti ) (c) : ti ∈ T k , s ∈ R(ti ) – from rule f 14, f 16 construction
P
F. tot decr(q, qqqc , c, k) represents ti ∈Tk ,s∈•ti mW(s,ti ) (c) +
P
ti ∈T k ,s∈R(ti ) m Mk (s) (c) – from (C,D,E) above
iii. Then firing T k ∪ {t j } would have required more tokens than are present at its
source place s ∈ •t j ∪ R(t j ). Thus, T k is a maximal set of transitions that can
simultaneously fire.
(d) And for each reset transition tr with prenabled(tr , k) ∈ A, f ires(tr , k) ∈ A, since A is
an answer set of Π6 - from rule f 20 and supported rule proposition
i. Then the firing set T k satisfies the reset transition requirement of definition 41
(firing set)
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
264
(e) Then {t0 , . . . , t x } = T k – using 1(a),1(b), 1(d) above; and using 1(c) it is a maximal
firing set
2. We show that Mk+1 is produced by firing T k in Mk . Let holds(p, qc , c, k + 1) ∈ A
(a) Then {holds(p, q1c , c, k), tot incr(p, q2c , c, k), tot decr(p, q3c , c, k)} ⊆ A : qc = q1c +
q2c − q3c – from rule r12 and supported rule proposition
(b) Then, holds(p, q1c , c, k) ∈ A represents q1c = m Mk (p) (c) – inductive assumption;
and {add(p, q20 c , T k , c, k), . . . , add(p, q2 j c , t j , c, k)} ⊆ A : q20 c + · · · + q2 j c = q2c
and {del(p, q30 c , T k , c, k), . . . , del(p, q3l c , tl , c, k)} ⊆ A : q30 c + · · · + q3l c = q3c –
rules r10, r11 and supported rule proposition, respectively
(c) Then { f ires(T k , k), . . . , f ires(t j , k)} ⊆ A and { f ires(T k , k), . . . ,
f ires(tl , k)} ⊆ A – rules r8, r9 and supported rule proposition, respectively
(d) Then { f ires(T k , k), . . . , f ires(t j , k)}∪ { f ires(T k , k), . . . , f ires(tl , k)} ⊆ A = { f ires(T k , k), . . . , f ires(t x , k)} ⊆
A – set union of subsets
(e) Then for each f ires(t x , k) ∈ A we have t x ∈ T k – already shown in item 1 above
P
P
P
(f) Then qc = m Mk (p) (c)+ tx ∈Tk ∧p∈tx • mW(tx ,p) (c)−( tx ∈T k ∧p∈•tx mW(p,tx ) (c)+ tx ∈Tk ∧p∈R(tx ) m Mk (p) (c))
– from (2b) above and the following
i. Each add(p, q j c , t j , c, k) ∈ A represents q j c = mW(t j ,p) (c) for p ∈ t j • – rule
r8, f 15 encoding, and definition 33 of transition execution in PN
ii. Each del(p, ty , qy c , c, k) ∈ A represents either qy c = mW(p,ty ) (c) for p ∈ •ty , or
qy c = m Mk (p) (c) for p ∈ R(ty ) – from rule r9, f 14 encoding and definition 33
of transition execution in PN; or from rule r9, f 16 encoding and definition of
reset arc in PN
P
iii. Each tot incr(p, q2c , c, k) ∈ A represents q2c = tx ∈Tk ∧p∈tx • mW(tx ,p) (c) – aggregate assignment atom semantics in rule r10
P
P
iv. Each tot decr(p, q3c , c, k) ∈ A represents q3c = tx ∈Tk ∧p∈•tx mW(p,tx ) (c)+ tx ∈Tk ∧p∈R(tx ) m Mk (p) (c)
– aggregate assignment atom semantics in rule r11
(g) Then, m Mk+1 (p) (c) = qc – since holds(p, qc , c, k + 1) ∈ A encodes qc = m Mk+1 (p) (c) –
from construction
As a result, for any n > k, T n will be a valid firing set for Mn and Mn+1 will be its target marking.
Conclusion: Since both (I) and (II) hold, X = M0 , T k , M1 , . . . , Mk , T k+1 is an execution sequence of PN(P, T, E, C, W, R, I, Q, QW, Z) (w.r.t M0 ) iff there is an answer set A of Π6 (PN, M0 , k, ntok)
such that (A.13) and (A.14) hold.
A.8
Proof of Proposition 8
Let PN = (P, T, E, C, W, R, I, Q, QW, Z, D) be a Petri Net, M0 be its initial marking and let
Π7 (PN, M0 , k, ntok) be the ASP encoding of PN and M0 over a simulation length k, with maximum ntok tokens on any place node, as defined in section 2.11. Then X = M0 , T k , M1 , . . . , Mk , T k , Mk+1
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
265
is an execution sequence of PN (w.r.t. M0 ) iff there is an answer set A of Π7 (PN, M0 , k, ntok)
such that:
{ f ires(t, ts) : t ∈ T ts , 0 ≤ ts ≤ k} = { f ires(t, ts) : f ires(t, ts) ∈ A}
(A.15)
{holds(p, q, c, ts) : p ∈ P, c/q = Mts (p), 0 ≤ ts ≤ k + 1}
= {holds(p, q, c, ts) : holds(p, q, c, ts) ∈ A}
(A.16)
We prove this by showing that:
(I) Given an execution sequence X, we create a set A such that it satisfies (A.15) and (A.16)
and show that A is an answer set of Π7
(II) Given an answer set A of Π7 , we create an execution sequence X such that (A.15) and
(A.16) are satisfied.
First we show (I): We create a set A as a union of the following sets:
1. A1 = {num(n) : 0 ≤ n ≤ ntok}
2. A2 = {time(ts) : 0 ≤ ts ≤ k}
3. A3 = {place(p) : p ∈ P}
4. A4 = {trans(t) : t ∈ T }
5. A5 = {color(c) : c ∈ C}
6. A6 = {ptarc(p, t, nc , c, ts) : (p, t) ∈ E − , c ∈ C, nc = mW(p,t) (c), nc > 0, 0 ≤ ts ≤ k}, where
E− ⊆ E
7. A7 = {tparc(t, p, nc , c, ts, d) : (t, p) ∈ E + , c ∈ C, nc = mW(t,p) (c), nc > 0, d = D(t), 0 ≤ ts ≤
k}, where E + ⊆ E
8. A8 = {holds(p, qc , c, 0) : p ∈ P, c ∈ C, qc = m M0 (p) (c)}
9. A9 = {ptarc(p, t, nc , c, ts) : p ∈ R(t), c ∈ C, nc = m Mts (p) , nc > 0, 0 ≤ ts ≤ k}
10. A10 = {iptarc(p, t, 1, c, ts) : p ∈ I(t), c ∈ C, 0 ≤ ts < k}
11. A11 = {tptarc(p, t, nc , c, ts) : (p, t) ∈ Q, c ∈ C, nc = mQW(p,t) (c), nc > 0, 0 ≤ ts ≤ k}
12. A12 = {notenabled(t, ts) : t ∈ T, 0 ≤ ts ≤ k, ∃c ∈ C, (∃p ∈ •t, m Mts (p) (c) < mW(p,t) (c)) ∨
(∃p ∈ I(t), m Mts (p) (c) > 0) ∨ (∃(p, t) ∈ Q, m Mts (p) (c) < mQW(p,t) (c))}
per definition 40 (enabled transition)
13. A13 = {enabled(t, ts) : t ∈ T, 0 ≤ ts ≤ k, ∀c ∈ C, (∀p ∈ •t, mW(p,t) (c) ≤ m Mts (p) (c)) ∧ (∀p ∈
I(t), m Mts (p) (c) = 0) ∧ (∀(p, t) ∈ Q, m Mts (p) (c) ≥ mQW(p,t) (c))}
per definition 40 (enabled transition)
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
266
14. A14 = { f ires(t, ts) : t ∈ T ts , 0 ≤ ts ≤ k}
per definition 40 (enabled transitions), only an enabled transition may fire
15. A15 = {add(p, qc , t, c, ts + d − 1) : t ∈ T ts , p ∈ t•, c ∈ C, qc = mW(t,p) (c),
d = D(t), 0 ≤ ts ≤ k}
per definition 43 (transition execution)
16. A16 = {del(p, qc , t, c, ts) : t ∈ T ts , p ∈ •t, c ∈ C, qc = mW(p,t) (c), 0 ≤ ts ≤ k} ∪
{del(p, qc , t, c, ts) : t ∈ T ts , p ∈ R(t), c ∈ C, qc = m Mts (p) (c), 0 ≤ ts ≤ k}
per definition 43 (transition execution)
17. A17 = {tot incr(p, qc , c, ts) : p ∈ P, c ∈ C,
P
qc = t∈Tl ,p∈t•,l≤ts,l+D(t)=ts+1 mW(t,p) (c), 0 ≤ ts ≤ k}
per definition 44 (firing set execution)
P
18. A18 = {tot decr(p, qc , c, ts) : p ∈ P, c ∈ C, qc = t∈T ts ,p∈•t mW(p,t) (c) +
P
t∈T ts ,p∈R(t) m Mts (p) (c), 0 ≤ ts ≤ k}
per definition 44 (firing set execution)
19. A19 = {consumesmore(p, ts) : p ∈ P, c ∈ C, qc = m Mts (p) (c),
P
P
q1c = t∈Tts ,p∈•t mW(p,t) (c) + t∈Tts ,p∈R(t) m Mts (p) (c), q1c > qc , 0 ≤ ts ≤ k}
per definition 34 (conflicting transitions)
20. A20 = {consumesmore : ∃p ∈ P, c ∈ C, qc = m Mts (p) (c),
P
P
q1c = t∈Tts ,p∈•t mW(p,t) (c) + t∈Tts ,p∈R(t) (m Mts (p) (c)), q1c > qc , 0 ≤ ts ≤ k}
per definition 34 (conflicting transitions)
21. A21 = {holds(p, qc , c, ts + 1) : p ∈ P, c ∈ C, qc = m Mts+1 (p) (c), 0 ≤ ts < k},
P
P
where Mts+1 (p) = Mts (p) − ( t∈Tts ,p∈•t W(p, t) + t∈T ts ,p∈R(t) Mts (p)) +
P
t∈T l ,p∈t•,l≤ts,l+D(t)−1=ts W(t, p)
according to definition 44 (firing set execution)
22. A22 = {transpr(t, pr) : t ∈ T, pr = Z(t)}
23. A23 = {notprenabled(t, ts) : t ∈ T, enabled(t, ts) ∈ A13 , (∃tt ∈ T, enabled(tt, ts) ∈
A13 , Z(tt) < Z(t)), 0 ≤ ts ≤ k}
= {notprenabled(t, ts) : t ∈ T, (∀p ∈ •t, W(p, t) ≤ Mts (p)), (∀p ∈ I(t), Mts (p) =
0), (∀(p, t) ∈ Q, M0 (p) ≥ WQ(p, t)), ∃tt ∈ T, (∀pp ∈ •tt, W(pp, tt) ≤ Mts (pp)), (∀pp ∈
I(tt), Mts (pp) = 0), (∀(pp, tt) ∈ Q, Mts (pp) ≥ QW(pp, tt)), Z(tt) < Z(t), 0 ≤ ts ≤ k}
24. A24 = {prenabled(t, ts) : t ∈ T, enabled(t, ts) ∈ A13 , (@tt ∈ T : enabled(tt, ts) ∈
A13 , Z(tt) < Z(t)), 0 ≤ ts ≤ k}
= {prenabled(t, ts) : t ∈ T, (∀p ∈ •t, W(p, t) ≤ Mts (p)), (∀p ∈ I(t), Mts (p) = 0), (∀(p, t) ∈
Q, M0 (p) ≥ WQ(p, t)), @tt ∈ T, (∀pp ∈ •tt, W(pp, tt) ≤ Mts (pp)), (∀pp ∈ I(tt), Mts (pp) =
0), (∀(pp, tt) ∈ Q, Mts (pp) ≥ QW(pp, tt)), Z(tt) < Z(t), 0 ≤ ts ≤ k}
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
267
25. A25 = {could not have(t, ts) : t ∈ T, prenabled(t, ts) ∈ A24 , f ires(t, ts) < A14 , (∃p ∈
P
P
•t ∪ R(t) : q > Mts (p) − ( t0 ∈Tts ,p∈•t0 W(p, t0 ) + t0 ∈Tts ,p∈R(t0 ) Mts (p)), q = W(p, t) if (p, t) ∈
E − or R(t) otherwise), 0 ≤ ts ≤ k}
= {could not have(t, ts) : t ∈ T, (∀p ∈ •t, W(p, t) ≤ Mts (p)), (∀p ∈ I(t), Mts (p) =
0), (∀(p, t) ∈ Q, M0 (p) ≥ QW(p, t)), (@tt ∈ T, (∀pp ∈ •tt, W(pp, tt) ≤ Mts (pp)), (∀pp ∈
I(tt), Mts (pp) = 0), (∀(pp, tt) ∈ Q, Mts (pp) ≥ QW(pp, tt)), Z(tt) < Z(t)), t < T ts , (∃p ∈
P
•t ∪ R(t) : q > Mts (p) − ( t0 ∈Tts ,p∈•t0 W(p, t0 ) +
P
−
t0 ∈T ts ,p∈R(t0 ) Mts (p)), q = W(p, t) if (p, t) ∈ E or R(t) otherwise), 0 ≤ ts ≤ k}
per the maximal firing set semantics
We show that A satisfies (A.15) and (A.16), and A is an answer set of Π7 .
A satisfies (A.15) and (A.16) by its construction above. We show A is an answer set of Π7
by splitting. We split lit(Π7 ) into a sequence of 9k + 11 sets:
• U0 = head( f 1) ∪ head( f 2) ∪ head( f 13) ∪ head( f 5) ∪ head( f 6) ∪ head(i2) ∪ head( f 19) =
{place(p) : p ∈ P} ∪ {trans(t) : t ∈ T } ∪ {col(c) : c ∈ C} ∪ {time(0), . . . , time(k)} ∪
{num(0), . . . , num(ntok)}∪{holds(p, qc , c, 0) : p ∈ P, c ∈ C, qc = m M0 (p) (c)}∪{transpr(t, pr) :
t ∈ T, pr = Z(t)}
• U9k+1 = U9k+0 ∪head( f 14)ts=k ∪head( f 15)ts=k ∪head( f 16)ts=k ∪head( f 17)ts=k ∪head( f 18)ts=k =
U9k+0 ∪ {ptarc(p, t, nc , c, k) : (p, t) ∈ E − , c ∈ C, nc = mW(p,t) (c)}
∪{tparc(t, p, nc , c, k, d) : (t, p) ∈ E + , c ∈ C, nc = mW(t,p) (c), d = D(t)}∪ {ptarc(p, t, nc , c, k) :
p ∈ R(t), c ∈ C, nc = m Mk (p) (c), n > 0} ∪ {iptarc(p, t, 1, c, k) :
p ∈ I(t), c ∈ C}∪ {tptarc(p, t, nc , c, k) : (p, t) ∈ Q, c ∈ C, nc = mQW(p,t) (c)}
• U9k+2 = U9k+1 ∪ head(e6)ts=k ∪ head(e7)ts=k ∪ head(e8)ts=k = U10k+1 ∪
{notenabled(t, k) : t ∈ T }
• U9k+3 = U9k+2 ∪ head(e9)ts=k = U9k+2 ∪ {enabled(t, k) : t ∈ T }
• U9k+4 = U9k+3 ∪ head(a11)ts=k = U9k+3 ∪ {notprenabled(t, k) : t ∈ T }
• U9k+5 = U9k+4 ∪ head(a12)ts=k = U9k+4 ∪ {prenabled(t, k) : t ∈ T }
• U9k+6 = U9k+5 ∪ head(a13)ts=k = U9k+5 ∪ { f ires(t, k) : t ∈ T }
• U9k+7 = U9k+6 ∪ head(r13)ts=k ∪ head(r9)ts=k = U9k+6 ∪ {add(p, qc , t, c, k) : p ∈ P, t ∈
T, c ∈ C, qc = mW(t,p) (c)} ∪ {del(p, qc , t, c, k) : p ∈ P, t ∈ T, c ∈ C, qc = mW(p,t) (c)} ∪
{del(p, qc , t, c, k) : p ∈ P, t ∈ T, c ∈ C, qc = m Mk (p) (c)}
• U9k+8 = U9k+7 ∪ head(r10)ts=k ∪ head(r11)ts=k = U9k+7 ∪ {tot incr(p, qc , c, k) : p ∈ P, c ∈
C, 0 ≤ qc ≤ ntok} ∪ {tot decr(p, qc , c, k) : p ∈ P, c ∈ C, 0 ≤ qc ≤ ntok}
• U9k+9 = U9k+8 ∪ head(a9)ts=k ∪ head(r12)ts=k ∪ head(a14)ts=k = U9k+8 ∪
{consumesmore(p, k) : p ∈ P}∪{holds(p, q, k+1) : p ∈ P, 0 ≤ q ≤ ntok}∪{could not have(t, k) :
t ∈ T}
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
268
• U9k+10 = U9k+9 ∪ head(a3) = U9k+9 ∪ {consumesmore}
where head(ri )ts=k are head atoms of ground rule ri in which ts = k. We write Ats=k
=
i
{a(. . . , ts) : a(. . . , ts) ∈ Ai , ts = k} as short hand for all atoms in Ai with ts = k. Uα , 0 ≤
α ≤ 9k + 10 form a splitting sequence, since each Ui is a splitting set of Π7 , and hUα iα<µ is a
S
monotone continuous sequence, where U0 ⊆ U1 · · · ⊆ U9k+10 and α<µ Uα = lit(Π7 ).
We compute the answer set of Π7 using the splitting sets as follows:
1. botU0 (Π7 ) = f 1 ∪ f 2 ∪ f 13 ∪ f 5 ∪ f 6 ∪ i2 ∪ f 19 and X0 = A1 ∪ · · · ∪ A5 ∪ A8 (= U0 ) is
its answer set – using forced atom proposition
2. evalU0 (botU1 (Π7 )\botU0 (Π7 ), X0 ) = {ptarc(p, t, qc , c, 0):-.|c ∈ C, qc = mW(p,t) (c), qc > 0}∪
{tparc(t, p, qc , c, 0, d):-.|c ∈ C, qc = mW(t,p) (c), qc > 0, d = D(t)} ∪
{ptarc(p, t, qc , c, 0):-.|c ∈ C, qc = m M0 (p) (c), qc > 0}∪ {iptarc(p, t, 1, c, 0):-.|c ∈ C}∪
∪
{tptarc(p, t, qc , c, 0):-.|c ∈ C, qc = mQW(p,t) (c), qc > 0}. Its answer set X1 = Ats=0
∪ Ats=0
7
6
ts=0
ts=0
ts=0
A9 ∪A10 ∪A11 – using forced atom proposition and construction of A6 , A7 , A9 , A10 , A11 .
3. evalU1 (botU2 (Π7 ) \ botU1 (Π7 ), X0 ∪ X1 ) = {notenabled(t, 0):-.|({trans(t),
ptarc(p, t, nc , c, 0), holds(p, qc , c, 0)} ⊆ X0 ∪ X1 , where qc < nc ) or
({notenabled(t, 0):-.| ({trans(t), iptarc(p, t, n2c , c, 0), holds(p, qc , c, 0)} ⊆ X0 ∪X1 , where qc ≥
n2c }) or ({trans(t), tptarc(p, t, n3c , c, 0), holds(p, qc , c, 0)} ⊆ X0 ∪ X1 , where qc < n3c )}.
Its answer set X2 = Ats=0
12 – using forced atom proposition and construction of A12 .
(a) where, qc = m M0 (p) (c), and nc = mW(p,t) (c) for an arc (p, t) ∈ E − – by construction
of i2 and f 14 in Π7 , and
(b) in an arc (p, t) ∈ E − , p ∈ •t (by definition 31 of preset)
(c) n2c = 1 – by construction of iptarc predicates in Π7 , meaning qc ≥ n2c ≡ qc ≥ 1 ≡
qc > 0,
(d) tptarc(p, t, n3c , c, 0) represents n3c = mQW(p,t) (c), where (p, t) ∈ Q
(e) thus, notenabled(t, 0) ∈ X1 means ∃c ∈ C, (∃p ∈ •t : m M0 (p) (c) < mW(p,t) (c))
∨ (∃p ∈ I(t) : m M0 (p) (c) > 0) ∨ (∃(p, t) ∈ Q : m M0 (p) (c) < mQW(p,t) (c)).
4. evalU2 (botU3 (Π7 ) \ botU2 (Π7 ), X0 ∪ · · · ∪ X2 ) = {enabled(t, 0):-.|trans(t) ∈ X0 ∪ · · · ∪
X2 , notenabled(t, 0) < X0 ∪ · · · ∪ X2 }. Its answer set is X3 = Ats=0
13 – using forced atom
proposition and construction of A13 .
(a) since an enabled(t, 0) ∈ X3 if @ notenabled(t, 0) ∈ X0 ∪ · · · ∪ X2 ; which is equivalent
to @t, ∀c ∈ C, (@p ∈ •t, m M0 (p) (c) < mW(p,t) (c)), (@p ∈ I(t),
m M0 (p) (c) > 0), (@(p, t) ∈ Q : m M0 (p) (c) < mQW(p,t) (c)), ∀c ∈ C, (∀p ∈ •t :
m M0 (p) (c) ≥ mW(p,t) (c)), (∀p ∈ I(t) : m M0 (p) (c) = 0).
5. evalU3 (botU4 (Π7 ) \ botU3 (Π7 ), X0 ∪ · · · ∪ X3 ) = {notprenabled(t, 0):-.|
{enabled(t, 0), transpr(t, p), enabled(tt, 0), transpr(tt, pp)} ⊆ X0 ∪ · · · ∪ X3 , pp < p}. Its
answer set is X4 = Ats=k
23 – using forced atom proposition and construction of A23 .
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
269
(a) enabled(t, 0) represents ∃t ∈ T, ∀c ∈ C, (∀p ∈ •t, m M0 (p) (c) ≥ mW(p,t) (c)), (∀p ∈
I(t), m M0 (p) (c) = 0), (∀(p, t) ∈ Q, m M0 (p) (c) ≥ mQW(p,t) (c))
(b) enabled(tt, 0) represents ∃tt ∈ T, ∀c ∈ C, (∀pp ∈ •tt, m M0 (pp) (c) ≥
mW(pp,tt) (c))∧ (∀pp ∈ I(tt), m M0 (pp) (c) = 0), (∀(pp, tt) ∈ Q, m M0 (pp) (c) ≥ mQW(pp,tt) (c))
(c) transpr(t, p) represents p = Z(t) – by construction
(d) transpr(tt, pp) represents pp = Z(tt) – by construction
(e) thus, notprenabled(t, 0) represents ∀c ∈ C, (∀p ∈ •t, m M0 (p) (c) ≥ mW(p,t) (c)),
(∀p ∈ I(t), m M0 (p) (c) = 0), (∀(p, t) ∈ Q, m M0 (p) (c) ≥ mQW(p,t) (c)), ∃tt ∈ T, (∀pp ∈
•tt, m M0 (pp) (c) ≥ mW(pp,tt) (c)), (∀pp ∈ I(tt), m M0 (pp) (c) = 0),
(∀(pp, tt) ∈ Q, m M0 (pp) (c) ≥ mQW(pp,tt) (c)), Z(tt) < Z(t)
(f) which is equivalent to (∀p ∈ •t : M0 (p) ≥ W(p, t)), (∀p ∈ I(t), M0 (p) = 0), (∀(p, t) ∈
Q, M0 (p) ≥ QW(p, t)), ∃tt ∈ T, (∀pp ∈ •tt, M0 (pp) ≥
W(pp, tt)), (∀pp ∈ I(tt), M0 (pp) = 0), (∀(pp, tt) ∈ Q, M0 ≥ QW(pp, tt)),
Z(tt) < Z(t) – assuming multiset domain C for all operations
6. evalU4 (botU5 (Π7 ) \ botU4 (Π7 ), X0 ∪ · · · ∪ X4 ) = {prenabled(t, 0):-.|enabled(t, 0) ∈ X0 ∪
· · · ∪ X4 , notprenabled(t, 0) < X0 ∪ · · · ∪ X4 }. Its answer set is X5 = Ats=k
24 – using forced
atom proposition and construction of A24
(a) enabled(t, 0) represents ∀c ∈ C, (∀p ∈ •t, m M0 (p) (c) ≥ mW(p,t) (c)), (∀p ∈ I(t), m M0 (p) (c) =
0), (∀(p, t) ∈ Q, m M0 (p) (c) ≥ mQW(p,t) (c)) ≡ (∀p ∈ •t,
M0 (p) ≥ W(p, t)), (∀p ∈ I(t), M0 (p) = 0), (∀(p, t) ∈ Q, M0 (p) ≥ QW(p, t)) – from 4
above and assuming multiset domain C for all operations
(b) notprenabled(t, 0) represents (∀p ∈ •t, M0 (p) ≥ W(p, t)), (∀p ∈ I(t),
M0 (p) = 0), (∀(p, t) ∈ Q, M0 (p) ≥ QW(p, t)), ∃tt ∈ T, (∀pp ∈ •tt, M0 (pp) ≥
W(pp, tt)), (∀pp ∈ I(tt), M0 (pp) = 0), (∀(pp, tt) ∈ Q, M0 (pp) ≥ QW(pp, tt)),
Z(tt) < Z(t) – from 5 above and assuming multiset domain C for all operations
(c) then, prenabled(t, 0) represents (∀p ∈ •t, M0 (p) ≥ W(p, t)), (∀p ∈ I(t),
M0 (p) = 0), (∀(p, t) ∈ Q, M0 (p) ≥ QW(p, t)), @tt ∈ T, ((∀pp ∈ •tt,
M0 (pp) ≥ W(pp, tt)), (∀pp ∈ I(tt), M0 (pp) = 0), (∀(pp, tt) ∈ Q, M0 (pp) ≥ QW(pp, tt)), Z(tt) <
Z(t)) – from (a), (b) and enabled(t, 0) ∈ X0 ∪ · · · ∪ X4
7. evalU5 (botU6 (Π7 ) \ botU5 (Π7 ), X0 ∪ · · · ∪ X5 ) = {{ f ires(t, 0)}:-.|prenabled(t, 0)
holds in X0 ∪ · · · ∪ X5 }. It has multiple answer sets X6.1 , . . . , X6.n , corresponding to elements of power set of f ires(t, 0) atoms in evalU5 (...) – using supported rule proposition.
Since we are showing that the union of answer sets of Π7 determined using splitting is
equal to A, we only consider the set that matches the f ires(t, 0) elements in A and call it
X6 , ignoring the rest. Thus, X6 = Ats=0
14 , representing T 0 .
(a) in addition, for every t such that prenabled(t, 0) ∈ X0 ∪· · ·∪X5 , R(t) , ∅; f ires(t, 0) ∈
X6 – per definition 41 (firing set); requiring that a reset transition is fired when
enabled
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
270
(b) thus, the firing set T 0 will not be eliminated by the constraint f 20
8. evalU6 (botU7 (Π7 ) \ botU6 (Π7 ), X0 ∪ · · · ∪ X6 ) = {add(p, nc , t, c, 0):-.| { f ires(t, 0 − d + 1),
tparc(t, p, nc , c, 0, d)} ⊆ X0 ∪· · ·∪X6 }∪{del(p, nc , t, c, 0):-.| { f ires(t, 0), ptarc(p, t, nc , c, 0)} ⊆
X0 ∪ · · · ∪ X6 }. It’s answer set is X7 = Ats=0
∪ Ats=0
– using forced atom proposition and
15
16
definitions of A15 and A16 .
(a) where, each add atom is equivalent to nc = mW(t,p) (c), c ∈ C, p ∈ t•,
(b) and each del atom is equivalent to nc = mW(p,t) (c), c ∈ C, p ∈ •t; or nc = m M0 (p) (c), c ∈
C, p ∈ R(t),
(c) representing the effect of transitions in T 0
9. evalU7 (botU8 (Π7 ) \ botU7 (Π7 ), X0 ∪ · · · ∪ X7 ) =
P
{tot incr(p, qqc , c, 0):-.|qqc = add(p,qc ,t,c,0)∈X0 ∪···∪X7 qc } ∪
P
{tot decr(p, qqc , c, 0):-.|qqc = del(p,qc ,t,c,0)∈X0 ∪···∪X7 qc }. It’s answer set is X8 = Ats=0
17 ∪
Ats=0
–
using
forced
atom
proposition
and
definitions
of
A
and
A
.
17
18
18
P
(a) where, each tot incr(p, qqc , c, 0), qqc = add(p,qc ,t,c,0)∈X0 ∪...X7 qc
P
≡ qqc = t∈X6 ,p∈t•,0+D(t)−1=0 mW(p,t) (c),
P
(b) tot decr(p, qqc , c, 0), qqc = del(p,qc ,t,c,0)∈X0 ∪...X7 qc
P
P
≡ qq = t∈X6 ,p∈•t mW(t,p) (c) + t∈X6 ,p∈R(t) m M0 (p) (c),
(c) represent the net effect of transitions in T 0
10. evalU8 (botU9 (Π7 ) \ botU8 (Π7 ), X0 ∪ · · · ∪ X8 ) = {consumesmore(p, 0):-.|
{holds(p, qc , c, 0), tot decr(p, q1c , c, 0)} ⊆ X0 ∪ · · · ∪ X8 , q1c > qc }∪
{holds(p, qc , c, 1):-.|{holds(p, q1c , c, 0), tot incr(p, q2c , c, 0),
tot decr(p, q3c , c, 0)} ⊆ X0 ∪ · · · ∪ X6 , qc = q1c + q2c − q3c } ∪ {could not have(t, 0):-.|
{prenabled(t, 0), ptarc(s, t, q, c, 0), holds(s, qq, c, 0), tot decr(s, qqq, c, 0)} ⊆ X0 ∪ · · · ∪
ts=0
ts=0
X8 , f ires(t, 0) < X0 ∪ · · · ∪ X8 , q > qq − qqq}. It’s answer set is X9 = Ats=0
19 ∪ A21 ∪ A25
– using forced atom proposition and definitions of A19 , A21 , A25 .
(a) where consumesmore(p, 0) represents ∃p : qc = m M0 (p) (c), q1c =
P
P
t∈T 0 ,p∈•t mW(p,t) (c) + t∈T 0 ,p∈R(t) m M0 (p) (c), q1c > qc , c ∈ C, indicating place p will
be over consumed if T 0 is fired, as defined in definition 34 (conflicting transitions),
P
(b) holds(p, qc , c, 1) if qc = m M0 (p) (c) + t∈T0 ,p∈t•,0+D(t)−1=0 mW(t,p) (c) −
P
P
( t∈T0 ,p∈•t mW(p,t) (c)+ t∈T0 ,p∈R(t) m M0 (p) (c)) represented by qc = m M1 (p) (c) for some
c ∈ C – by construction of Π7
P
P
(c) and consumesmore(p, 0) if t∈T0 ,p∈•t mW(p,t) (c)+ t∈T0 ,p∈R(t) m M0 (p) (c) > m M0 (p) (c)
for any c ∈ C
(d) could not have(t, 0) if
i. (∀p ∈ •t, W(p, t) ≤ M0 (p)), (∀p ∈ I(t), M0 (p) = 0), (∀(p, t) ∈ Q,
M0 (p) ≥ WQ(p, t)), @tt ∈ T, (∀pp ∈ •tt, W(pp, tt) ≤ Mts (pp)), (∀pp ∈ I(tt), M0 (pp) =
0), (∀(pp, tt) ∈ Q, M0 (pp) ≥ QW(pp, tt)), Z(tt) < Z(t),
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
271
P
P
ii. and qc > m M0 (s) (c) − ( t0 ∈T0 ,s∈•t0 mW(s,t0 ) (c) + t0 ∈T 0 ,s∈R(t) m M0 (s) (c)),
qc = mW(s,t) (c) if s ∈ •t or m M0 (s) (c) otherwise for some c ∈ C, which beP
P
comes q > M0 (s) − ( t0 ∈T 0 ,s∈•t0 W(s, t0 ) + t0 ∈T0 ,s∈R(t) M0 (s)), q = W(s, t) if s ∈
•t or M0 (s) otherwise for all c ∈ C
iii. (i), (ii) above combined match the definition of A25
(e) X9 does not contain could not have(t, 0), when prenabled(t, 0) ∈ X0 ∪ · · · ∪ X6 and
f ires(t, 0) < X0 ∪ · · · ∪ X5 due to construction of A, encoding of a10 and its body
atoms. As a result, it is not eliminated by the constraint a15
..
.
11. evalU9k+0 (botU9k+1 (Π7 ) \ botU9k+0 (Π7 ), X0 ∪ · · · ∪ X9k+0 ) = {ptarc(p, t, qc , c, k):-.|c ∈ C, qc =
mW(p,t) (c), qc > 0} ∪ {tparc(t, p, qc , c, k, d):-.|c ∈ C, qc = mW(t,p) (c), qc > 0, d = D(t)}∪
{ptarc(p, t, qc , c, k):-.|c ∈ C, qc = m Mk (p) (c), qc > 0} ∪
{iptarc(p, t, 1, c, k):-.|c ∈ C}∪ {tptarc(p, t, qc , c, k):-.|c ∈ C, qc = mQW(p,t) (c), qc > 0}. Its
ts=k
answer set X9k+1 = Ats=k
∪ Ats=k
∪ Ats=k
∪ Ats=k
7
9
10 ∪ A11 – using forced atom proposition
6
and construction of A6 , A7 , A9 , A10 , A11 .
12. evalU9k+1 (botU10k+2 (Π7 ) \ botU9k+1 (Π7 ), X0 ∪ · · · ∪ X9k+1 ) = {notenabled(t, k):-.|
({trans(t), ptarc(p, t, nc , c, k), holds(p, qc , c, k)} ⊆ X0 ∪· · ·∪X9k+1 , where qc < nc ) or ({notenabled(t, k):-.|({trans(t
X0 ∪ · · · ∪ X9k+1 , where qc ≥ n2c }) or ({trans(t), tptarc(p, t, n3c , c, k),
holds(p, qc , c, k)} ⊆ X0 ∪ · · · ∪ X9k+1 , where qc < n3c )}. Its answer set X9k+2 = Ats=k
12 –
using forced atom proposition and construction of A12 .
(a) where, qc = m Mk (p) (c), and nc = mW(p,t) (c) for an arc (p, t) ∈ E − – by construction
of i2 and f 14 predicates in Π7 , and
(b) in an arc (p, t) ∈ E − , p ∈ •t (by definition 31 of preset)
(c) n2c = 1 – by construction of iptarc predicates in Π7 , meaning qc ≥ n2c ≡ qc ≥ 1 ≡
qc > 0,
(d) tptarc(p, t, n3c , c, k) represents n3c = mQW(p,t) (c), where (p, t) ∈ Q
(e) thus, notenabled(t, k) ∈ X9k+1 represents ∃c ∈ C, (∃p ∈ •t : m Mk (p) (c) < mW(p,t) (c))∨
(∃p ∈ I(t) : m Mk (p) (c) > 0) ∨ (∃(p, t) ∈ Q : m Mk (p) (c) < mQW(p,t) (c)).
13. evalU9k+2 (botU9k+3 (Π7 ) \ botU9k+2 (Π7 ), X0 ∪ · · · ∪ X9k+2 ) = {enabled(t, k):-.|trans(t) ∈ X0 ∪
· · · ∪ X9k+2 , notenabled(t, k) < X0 ∪ · · · ∪ X9k+2 }. Its answer set is X9k+3 = Ats=k
13 – using
forced atom proposition and construction of A13 .
(a) Since an enabled(t, k) ∈ X9k+3 if @ notenabled(t, k) ∈ X0 ∪ · · · ∪ X9k+2 ; which is
equivalent to @t, ∀c ∈ C, (@p ∈ •t, m Mk (p) (c) < mW(p,t) (c)), (@p ∈ I(t), m Mk (p) (c) >
0), (@(p, t) ∈ Q : m Mk (p) (c) < mQW(p,t) (c)), ∀c ∈ C, (∀p ∈ •t : m Mk (p) (c) ≥
mW(p,t) (c)), (∀p ∈ I(t) : m Mk (p) (c) = 0).
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
272
14. evalU9k+3 (botU9k+4 (Π7 ) \ botU9k+3 (Π7 ), X0 ∪ · · · ∪ X9k+3 ) = {notprenabled(t, k):-.|
{enabled(t, k), transpr(t, p), enabled(tt, k), transpr(tt, pp)} ⊆ X0 ∪ · · · ∪ X9k+3 ,
pp < p}. Its answer set is X9k+4 = Ats=k
23 – using forced atom proposition and construction
of A23 .
(a) enabled(t, k) represents ∃t ∈ T, ∀c ∈ C, (∀p ∈ •t, m Mk (p) (c) ≥ mW(p,t) (c)), (∀p ∈
I(t), m Mk (p) (c) = 0), (∀(p, t) ∈ Q, m Mk (p) (c) ≥ mQW(p,t) (c))
(b) enabled(tt, k) represents ∃tt ∈ T, ∀c ∈ C, (∀pp ∈ •tt, m Mk (pp) (c) ≥
mW(pp,tt) (c))∧(∀pp ∈ I(tt), m Mk (pp) (c) = 0), (∀(pp, tt) ∈ Q, m Mk (pp) (c) ≥ mQW(pp,tt) (c))
(c) transpr(t, p) represents p = Z(t) – by construction
(d) transpr(tt, pp) represents pp = Z(tt) – by construction
(e) thus, notprenabled(t, k) represents ∀c ∈ C, (∀p ∈ •t, m Mk (p) (c) ≥
mW(p,t) (c)), (∀p ∈ I(t), m Mk (p) (c) = 0), (∀(p, t) ∈ Q, m Mk (p) (c) ≥
mQW(p,t) (c)), ∃tt ∈ T, (∀pp ∈ •tt, m Mk (pp) (c) ≥ mW(pp,tt) (c)), (∀pp ∈ I(tt),
m Mk (pp) (c) = 0), (∀(pp, tt) ∈ Q, m Mk (pp) (c) ≥ mQW(pp,tt) (c)), Z(tt) < Z(t)
(f) which is equivalent to (∀p ∈ •t : Mk (p) ≥ W(p, t)), (∀p ∈ I(t),
Mk (p) = 0), (∀(p, t) ∈ Q, Mk (p) ≥ QW(p, t)), ∃tt ∈ T, (∀pp ∈ •tt,
Mk (pp) ≥ W(pp, tt)), (∀pp ∈ I(tt), Mk (pp) = 0), (∀(pp, tt) ∈ Q, Mk (pp) ≥ QW(pp, tt)), Z(tt) <
Z(t) – assuming multiset domain C for all operations
15. evalU9k+4 (botU9k+5 (Π7 ) \ botU9k+4 (Π7 ), X0 ∪ · · · ∪ X9k+4 ) = {prenabled(t, k):-.|
enabled(t, k) ∈ X0 ∪ · · · ∪ X9k+4 , notprenabled(t, k) < X0 ∪ · · · ∪ X9k+4 }. Its answer set is
X9k+5 = Ats=k
24 – using forced atom proposition and construction of A24
(a) enabled(t, k) represents ∀c ∈ C, (∀p ∈ •t, m Mk (p) (c) ≥ mW(p,t) (c)), (∀p ∈ I(t), m Mk (p) (c) =
0), (∀(p, t) ∈ Q, m Mk (p) (c) ≥ mQW(p,t) (c)) ≡ (∀p ∈ •t, Mk (p) ≥ W(p, t)), (∀p ∈
I(t), Mk (p) = 0), (∀(p, t) ∈ Q, Mk (p) ≥ QW(p, t)) – from 13 above and assuming
multiset domain C for all operations
(b) notprenabled(t, k) represents (∀p ∈ •t, Mk (p) ≥ W(p, t)), (∀p ∈ I(t),
Mk (p) = 0), (∀(p, t) ∈ Q, Mk (p) ≥ QW(p, t)), ∃tt ∈ T, (∀pp ∈ •tt, Mk (pp) ≥
W(pp, tt)), (∀pp ∈ I(tt), Mk (pp) = 0), (∀(pp, tt) ∈ Q, Mk (pp) ≥ QW(pp, tt)),
Z(tt) < Z(t) – from 14 above and assuming multiset domain C for all operations
(c) then, prenabled(t, k) represents (∀p ∈ •t, Mk (p) ≥ W(p, t)), (∀p ∈ I(t),
Mk (p) = 0), (∀(p, t) ∈ Q, Mk (p) ≥ QW(p, t)), @tt ∈ T, ((∀pp ∈ •tt, Mk (pp) ≥
W(pp, tt)), (∀pp ∈ I(tt), Mk (pp) = 0), (∀(pp, tt) ∈ Q,
Mk (pp) ≥ QW(pp, tt)), Z(tt) < Z(t)) – from (a), (b) and enabled(t, k) ∈ X0 ∪ · · · ∪
X9k+4
16. evalU9k+5 (botU9k+6 (Π7 ) \ botU9k+5 (Π7 ), X0 ∪ · · · ∪ X9k+5 ) = {{ f ires(t, k)}:-.|
prenabled(t, k) holds in X0 ∪ · · · ∪ X9k+5 }. It has multiple answer sets
X9k+6.1 , . . . , X9k+6.n , corresponding to elements of power set of f ires(t, k) atoms in evalU9k+5 (...)
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
273
– using supported rule proposition. Since we are showing that the union of answer sets
of Π7 determined using splitting is equal to A, we only consider the set that matches
the f ires(t, k) elements in A and call it X9k+6 , ignoring the rest. Thus, X9k+6 = Ats=k
14 ,
representing T k .
(a) in addition, for every t such that prenabled(t, k) ∈ X0 ∪ · · · ∪ X9k+5 , R(t) , ∅;
f ires(t, k) ∈ X9k+6 – per definition 41 (firing set); requiring that a reset transition
is fired when enabled
(b) thus, the firing set T k will not be eliminated by the constraint f 20
17. evalU9k+6 (botU9k+7 (Π7 ) \ botU9k+6 (Π7 ), X0 ∪ · · · ∪ X9k+6 ) = {add(p, nc , t, c, k):-.| { f ires(t, k −
d + 1), tparc(t, p, nc , c, 0, d)} ⊆ X0 ∪ · · · ∪ X9k+6 } ∪
{del(p, nc , t, c, k):-.|{ f ires(t, k), ptarc(p, t, nc , c, k)} ⊆ X0 ∪ · · · ∪ X9k+6 }. It’s answer set is
X9k+7 = Ats=k
∪ Ats=k
– using forced atom proposition and definitions of A15 and A16 .
15
16
(a) where, each add atom is equivalent to nc = mW(t,p) (c), c ∈ C, p ∈ t•,
(b) and each del atom is equivalent to nc = mW(p,t) (c), c ∈ C, p ∈ •t; or nc = m Mk (p) (c), c ∈
C, p ∈ R(t),
(c) representing the effect of transitions in T k
18. evalU9k+7 (botU9k+8 (Π7 ) \ botU9k+7 (Π7 ), X0 ∪ · · · ∪ X9k+7 ) = {tot incr(p, qqc , c, k):-.|
P
qqc = add(p,qc ,t,c,k)∈X0 ∪···∪X9k+7 qc } ∪ {tot decr(p, qqc , c, k):-.|
P
ts=k
qqc = del(p,qc ,t,c,k)∈X0 ∪···∪X9k+7 qc }. It’s answer set is X9k+8 = Ats=k
17 ∪ A18 – using forced
atom proposition and definitions of A17 and A18 .
P
(a) where, each tot incr(p, qqc , c, k), qqc = add(p,qc ,t,c,k)∈X0 ∪...X9k+7 qc
P
≡ qqc = t∈X9k+6 ,p∈t•,0≤l≤k,l+D(t)−1=k mW(p,t) (c),
P
(b) and each tot decr(p, qqc , c, k), qqc = del(p,qc ,t,c,k)∈X0 ∪...X9k+7 qc
P
P
≡ qq = t∈X9k+6 ,p∈•t mW(t,p) (c) + t∈X9k+6 ,p∈R(t) m Mk (p) (c),
(c) represent the net effect of transitions in T k
19. evalU9k+8 (botU9k+9 (Π7 ) \ botU9k+8 (Π7 ), X0 ∪ · · · ∪ X9k+8 ) = {consumesmore(p, k):-.|
{holds(p, qc , c, k), tot decr(p, q1c , c, k)} ⊆ X0 ∪ · · · ∪ X9k+8 , q1c > qc } ∪
{holds(p, qc , c, k + 1):-., |{holds(p, q1c , c, k), tot incr(p, q2c , c, k),
tot decr(p, q3c , c, k)} ⊆ X0 ∪ · · · ∪ X9k+6 , qc = q1c + q2c − q3c } ∪
{could not have(t, k):-.|{prenabled(t, k),
ptarc(s, t, q, c, k), holds(s, qq, c, k), tot decr(s, qqq, c, k)} ⊆ X0 ∪ · · · ∪ X9k+8 ,
ts=k
ts=k
f ires(t, k) < X0 ∪· · ·∪ X10k+8 , q > qq−qqq}. It’s answer set is X9k+9 = Ats=k
19 ∪ A21 ∪ A25
– using forced atom proposition and definitions of A19 , A21 , A25 .
(a) where, consumesmore(p, k) represents ∃p : qc = m Mk (p) (c),
P
P
q1c = t∈Tk ,p∈•t mW(p,t) (c) + t∈T k ,p∈R(t) m Mk (p) (c), q1c > qc , c ∈ C, indicating place
p that will be over consumed if T k is fired, as defined in definition 34 (conflicting
transitions),
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
274
P
P
(b) holds(p, qc , c, k+1) if qc = m Mk (p) (c)+ t∈Tl ,p∈t•,0≤l≤k,l+D(t)−1=k mW(t,p) (c)−( t∈T k ,p∈•t mW(p,t) (c)+
P
t∈T k ,p∈R(t) m Mk (p) (c)) represented by qc = m M1 (p) (c) for some c ∈ C – by construction of Π7 ,
(c) and could not have(t, k) if
i. (∀p ∈ •t, W(p, t) ≤ Mk (p)), (∀p ∈ I(t), Mk (p) = 0), (∀(p, t) ∈ Q,
Mk (p) ≥ WQ(p, t)), @tt ∈ T, (∀pp ∈ •tt, W(pp, tt) ≤ Mts (pp)), (∀pp ∈ I(tt), Mk (pp) =
0), (∀(pp, tt) ∈ Q, Mk (pp) ≥ QW(pp, tt)), Z(tt) < Z(t),
P
P
ii. and qc > m Mk (s) (c) − ( t0 ∈Tk ,s∈•t0 mW(s,t0 ) (c) + t0 ∈Tk ,s∈R(t) m Mk (s) (c)),
qc = mW(s,t) (c) if s ∈ •t or m Mk (s) (c) otherwise for some c ∈ C, which beP
P
comes q > Mk (s) − ( t0 ∈Tk ,s∈•t0 W(s, t0 ) + t0 ∈Tk ,s∈R(t) Mk (s)), q = W(s, t) if s ∈
•t or Mk (s) otherwise for all c ∈ C
iii. (i), (ii) above combined match the definition of A25
(d) X9k+9 does not contain could not have(t, k), when prenabled(t, k) ∈ X0 ∪ · · · ∪ X9k+6
and f ires(t, k) < X0 ∪ · · · ∪ X9k+5 due to construction of A, encoding of a10 and its
body atoms. As a result it is not eliminated by the constraint a15
20. evalU9k+9 (botU9k+10 (Π7 ) \ botU9k+9 (Π7 ), X0 ∪ · · · ∪ X9k+9 ) = {consumesmore:-.|
{consumesmore(p, 0), . . . , consumesmore(p, k)} ∩ (X0 ∪ · · · ∪ X9k+9 ) , ∅}. It’s answer set
is X9k+10 = A20 – using forced atom proposition and the definition of A20
(a) X9k+10 will be empty since none of consumesmore(p, 0), . . . ,
consumesmore(p, k) hold in X0 ∪ · · · ∪ X9k+10 due to the construction of A, encoding
of a2 and its body atoms. As a result, it is not eliminated by the constraint a4
The set X = X0 ∪ · · · ∪ X9k+10 is the answer set of Π7 by the splitting sequence theorem 15.
Each Xi , 0 ≤ i ≤ 9k + 10 matches a distinct portion of A, and X = A, thus A is an answer set of
Π7 .
Next we show (II): Given Π7 be the encoding of a Petri Net PN(P, T, E, C, W, R, I, Q, QW, Z, D)
with initial marking M0 , and A be an answer set of Π7 that satisfies (A.15) and (A.16), then we
can construct X = M0 , T k , . . . , Mk , T k , Mk+1 from A, such that it is an execution sequence of PN.
We construct the X as follows:
1. Mi = (Mi (p0 ), . . . , Mi (pn )), where {holds(p0 , m Mi (p0 ) (c), c, i), . . . ,
holds(pn , m Mi (pn ) (c), c, i)} ⊆ A, for c ∈ C, 0 ≤ i ≤ k + 1
2. T i = {t : f ires(t, i) ∈ A}, for 0 ≤ i ≤ k
and show that X is indeed an execution sequence of PN. We show this by induction over k (i.e.
given Mk , T k is a valid firing set and its firing produces marking Mk+1 ).
Base case: Let k = 0, and M0 is a valid marking in X for PN, show (1) T 0 is a valid firing set
for M0 , and (2) firing T 0 in M0 produces marking M1 .
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
275
1. We show T 0 is a valid firing set for M0 . Let { f ires(t0 , 0), . . . , f ires(t x , 0)} be the set of all
f ires(. . . , 0) atoms in A,
(a) Then for each f ires(ti , 0) ∈ A
prenabled(ti , 0) ∈ A – from rule a13 and supported rule proposition
Then enabled(ti , 0) ∈ A – from rule a12 and supported rule proposition
And notprenabled(ti , 0) < A – from rule a12 and supported rule proposition
For enabled(ti , 0) ∈ A
A. notenabled(ti , 0) < A – from rule e9 and supported rule proposition
B. Then either of body(e6), body(e7), or body(e8) must not hold in A for ti –
from rules body(e6), body(e7), body(e8) and forced atom proposition
C. Then qc ≮ nic ≡ qc ≥ nic in e6 for all {holds(p, qc , c, 0),
ptarc(p, ti , nic , c, 0)} ⊆ A – from e6, forced atom proposition, and given
facts (holds(p, qc , c, 0) ∈ A, ptarc(p, ti , nic , 0) ∈ A)
D. And qc nic ≡ qc < nic in e7 for all {holds(p, qc , c, 0),
iptarc(p, ti , nic , c, 0)} ⊆ A, nic = 1; qc > nic ≡ qc = 0 – from e7, forced
atom proposition, given facts (holds(p, qc , c, 0) ∈ A,
iptarc(p, ti , 1, c, 0) ∈ A), and qc is a positive integer
E. And qc ≮ nic ≡ qc ≥ nic in e8 for all {holds(p, qc , c, 0),
tptarc(p, ti , nic , c, 0)} ⊆ A – from e8, forced atom proposition, and given
facts
F. Then ∀c ∈ C, (∀p ∈ •ti , m M0 (p) (c) ≥ mW(p,ti ) (c)) ∧ (∀p ∈ I(ti ),
m M0 (p) (c) = 0) ∧ (∀(p, ti ) ∈ Q, m M0 (p) (c) ≥ mQW(p,ti ) (c)) – from i2, f 14, f 16
construction, definition 31 of preset •ti in PN, definition 40 of enabled
transition in PN, and that the construction of reset arcs by f 16 ensures
notenabled(t, 0) is never true for a reset arc, where holds(p, qc , c, 0) ∈ A
represents qc = m M0 (p) (c), ptarc(p, ti , nic , 0) ∈ A represents nic = mW(p,ti ) (c),
nic = m M0 (p) (c).
G. Which is equivalent to (∀p ∈ •ti , M0 (p) ≥ W(p, ti )) ∧ (∀p ∈ I(ti ), M0 (p) =
0) ∧ (∀(p, ti ) ∈ Q, M0 (p) ≥ QW(p, ti )) – assuming multiset domain C
v. For notprenabled(ti , 0) < A
A. Either (@enabled(tt, 0) ∈ A : pp < pi ) or (∀enabled(tt, 0) ∈ A : pp ≮
pi ) where pp = Z(tt), pi = Z(ti ) – from rule a11, f 19 and forced atom
proposition
B. This matches the definition of an enabled priority transition
vi. Then ti is enabled and can fire in PN, as a result it can belong to T 0 – from
definition 40 of enabled transition
i.
ii.
iii.
iv.
(b) And consumesmore < A, since A is an answer set of Π7 – from rule a4 and supported
rule proposition
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
276
i. Then @consumesmore(p, 0) ∈ A – from rule a3 and supported rule proposition
ii. Then @{holds(p, qc , c, 0), tot decr(p, q1c , c, 0)} ⊆ A, q1c > qc in body(a9) –
from a9 and forced atom proposition
P
iii. Then @c ∈ C@p ∈ P, ( ti ∈{t0 ,...,tx },p∈•ti mW(p,ti ) (c) +
P
ti ∈{t0 ,...,t x },p∈R(ti ) m M0 (p) (c)) > m M0 (p) (c) – from the following
A. holds(p, qc , c, 0) represents qc = m M0 (p) (c) – from rule i2 encoding, given
B. tot decr(p, q1c , c, 0) ∈ A if {del(p, q10 c , t0 , c, 0), . . . ,
del(p, q1 x c , t x , c, 0)} ⊆ A, where q1c = q10 c + · · · + q1 x c – from r11 and
forced atom proposition
C. del(p, q1i c , ti , c, 0) ∈ A if { f ires(ti , 0), ptarc(p, ti , q1i c , c, 0)} ⊆ A – from r9
and supported rule proposition
D. del(p, q1i c , ti , c, 0) represents removal of q1i c = mW(p,ti ) (c) tokens from p ∈
•ti ; or it represents removal of q1i c = m M0 (p) (c) tokens from p ∈ R(ti )–
from rules r9, f 14, f 16, supported rule proposition, and definition 31 of
preset in PN
iv. Then the set of transitions in T 0 do not conflict – by the definition 34 of conflicting transitions
(c) And for each prenabled(t j , 0) ∈ A and f ires(t j , 0) < A,
could not have(t j , 0) ∈ A, since A is an answer set of Π7 - from rule a15 and supported rule proposition
i. Then {prenabled(t j , 0), holds(s, qqc , c, 0), ptarc(s, t j , qc , c, 0),
tot decr(s, qqqc , c, 0)} ⊆ A, such that qc > qqc − qqqc and f ires(t j , 0) < A from rule a14 and supported rule proposition
P
P
ii. Then for an s ∈ •t j ∪R(t j ), qc > m M0 (s) (c)−( ti ∈T0 ,s∈•ti mW(s,ti ) (c)+ ti ∈T0 ,s∈R(ti ) m M0 (s) (c)),
where qc = mW(s,t j ) (c) if s ∈ •t j , or m M0 (s) (c) otherwise.
A. ptarc(s, ti , qc , c, 0) represents qc = mW(s,ti ) (c) if (s, ti ) ∈ E − or qc = m M0 (s) (c)
if s ∈ R(ti ) – from rule f 14, f 16 construction
B. holds(s, qqc , c, 0) represents qqc = m M0 (s) (c) – from i2 construction
C. tot decr(s, qqqc , c, 0) ∈ A if {del(s, qqq0 c , t0 , c, 0), . . . ,
del(s, qqq x c , t x , c, 0)} ⊆ A – from rule r11 construction and supported rule
proposition
D. del(s, qqqi c , ti , c, 0) ∈ A if { f ires(ti , 0), ptarc(s, ti , qqqi c , c, 0)} ⊆ A – from
rule r9 and supported rule proposition
E. del(s, qqqi c , ti , c, 0) either represents qqqi c = mW(s,ti ) (c) : ti ∈ T 0 , (s, ti ) ∈
E − , or qqqi c = m M0 (ti ) (c) : ti ∈ T 0 , s ∈ R(ti ) – from rule f 14, f 16 construction
P
F. tot decr(q, qqqc , c, 0) represents ti ∈T0 ,s∈•ti mW(s,ti ) (c) +
P
ti ∈T 0 ,s∈R(ti ) m M0 (s) (c) – from (C,D,E) above
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
277
iii. Then firing T 0 ∪ {t j } would have required more tokens than are present at its
source place s ∈ •t j ∪ R(t j ). Thus, T 0 is a maximal set of transitions that can
simultaneously fire.
(d) And for each reset transition tr with prenabled(tr , 0) ∈ A, f ires(tr , 0) ∈ A, since A
is an answer set of Π7 - from rule f 20 and supported rule proposition
i. Then, the firing set T 0 satisfies the reset-transition requirement of definition 41
(firing set)
(e) Then {t0 , . . . , t x } = T 0 – using 1(a),1(b),1(d) above; and using 1(c) it is a maximal
firing set
2. Let holds(p, qc , c, 1) ∈ A
(a) Then {holds(p, q1c , c, 0), tot incr(p, q2c , c, 0), tot decr(p, q3c , c, 0)} ⊆ A : qc = q1c +
q2c − q3c – from rule r12 and supported rule proposition
(b) Then, holds(p, q1c , c, 0) ∈ A represents q1c = m M0 (p) (c) – given ;
and {add(p, q20 c , t0 , c, 0), . . . , add(p, q2 j c , t j , c, 0)} ⊆ A : q20 c + · · · + q2 j c = q2c and
{del(p, q30 c , t0 , c, 0), . . . , del(p, q3l c , tl , c, 0)} ⊆ A : q30 c + · · · + q3l c = q3c – rules
r10, r11 and supported rule proposition, respectively
(c) Then { f ires(t0 , 0), . . . , f ires(t j , 0)} ⊆ A and { f ires(t0 , 0), . . . , f ires(tl , 0)}
⊆ A – rules r13, r9 and supported rule proposition, respectively
(d) Then { f ires(t0 , 0), . . . , f ires(t j , 0)}∪{ f ires(t0 , 0), . . . , f ires(tl , 0)} ⊆ A = { f ires(t0 , 0), . . . ,
f ires(t x , 0)} ⊆ A – set union of subsets
(e) Then for each f ires(t x , 0) ∈ A we have t x ∈ T 0 – already shown in item 1 above
P
(f) Then qc = m M0 (p) (c) + tx ∈T0 ,p∈tx •,0+D(tx )−1=0 mW(tx ,p) (c) −
P
P
( tx ∈T 0 ∧p∈•tx mW(p,tx ) (c)+ tx ∈T0 ∧p∈R(tx ) m M0 (p) (c)) – from (2b) above and the following
i. Each add(p, q j c , t j , c, 0) ∈ A represents q j c = mW(t j ,p) (c) for p ∈ t j • – rule
r13, f 21 encoding, and definition 43 of transition execution in PN
ii. Each del(p, ty , qy c , c, 0) ∈ A represents either qy c = mW(p,ty ) (c) for p ∈ •ty , or
qy c = m M0 (p) (c) for p ∈ R(ty ) – from rule r9, f 14 encoding and definition 43
of transition execution in PN; or from rule r9, f 16 encoding and definition of
reset arc in PN
iii. Each tot incr(p, q2c , c, 0) ∈ A represents
P
q2c = tx ∈T0 ∧p∈tx •,0+D(tx )−1=0 mW(tx ,p) (c) – aggregate assignment atom semantics in rule r10
P
iv. Each tot decr(p, q3c , c, 0) ∈ A represents q3c = tx ∈T0 ∧p∈•tx mW(p,tx ) (c)
P
+ tx ∈T0 ∧p∈R(tx ) m M0 (p) (c) – aggregate assignment atom semantics in rule r11
(g) Then, m M1 (p) (c) = qc – since holds(p, qc , c, 1) ∈ A encodes qc = m M1 (p) (c) – from
construction
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
278
Inductive Step: Let k > 0, and Mk is a valid marking in X for PN, show (1) T k is a valid firing
set for Mk , and (2) firing T k in Mk produces marking Mk+1 .
1. We show that T k is a valid firing set in Mk . Let { f ires(t0 , k), . . . , f ires(t x , k)} be the set of
all f ires(. . . , k) atoms in A,
(a) We show that T k is a valid firing set in Mk . Then for each f ires(ti , k) ∈ A
prenabled(ti , k) ∈ A – from rule a13 and supported rule proposition
Then enabled(ti , k) ∈ A – from rule a12 and supported rule proposition
And notprenabled(ti , k) < A – from rule a12 and supported rule proposition
For enabled(ti , k) ∈ A
A. notenabled(ti , k) < A – from rule e9 and supported rule proposition
B. Then either of body(e6), body(e7), or body(e8) must not hold in A for ti –
from rules body(e6), body(e7), body(e8) and forced atom proposition
C. Then qc ≮ nic ≡ qc ≥ nic in e6 for all {holds(p, qc , c, k),
ptarc(p, ti , nic , c, k)} ⊆ A – from e6, forced atom proposition, and given
facts (holds(p, qc , c, k) ∈ A, ptarc(p, ti , nic , k) ∈ A)
D. And qc nic ≡ qc < nic in e7 for all {holds(p, qc , c, k),
iptarc(p, ti , nic , c, k)} ⊆ A, nic = 1; qc > nic ≡ qc = 0 – from e7, forced
atom proposition, given facts (holds(p, qc , c, k) ∈ A,
iptarc(p, ti , 1, c, k) ∈ A), and qc is a positive integer
E. And qc ≮ nic ≡ qc ≥ nic in e8 for all {holds(p, qc , c, k),
tptarc(p, ti , nic , c, k)} ⊆ A – from e8, forced atom proposition, and given
facts
F. Then ∀c ∈ C, (∀p ∈ •ti , m Mk (p) (c) ≥ mW(p,ti ) (c)) ∧ (∀p ∈ I(ti ),
m Mk (p) (c) = 0) ∧ (∀(p, ti ) ∈ Q, m Mk (p) (c) ≥ mQW(p,ti ) (c)) – from the inductive assumption, f 14, f 16 construction, definition 31 of preset •ti in PN,
definition 40 of enabled transition in PN, and that the construction of reset arcs by f 16 ensures notenabled(t, k) is never true for a reset arc, where
holds(p, qc , c, k) ∈ A represents qc = m Mk (p) (c), ptarc(p, ti , nic , k) ∈ A represents nic = mW(p,ti ) (c), nic = m Mk (p) (c).
G. Which is equivalent to (∀p ∈ •ti , Mk (p) ≥ W(p, ti )) ∧ (∀p ∈ I(ti ), Mk (p) =
0) ∧ (∀(p, ti ) ∈ Q, Mk (p) ≥ QW(p, ti )) – assuming multiset domain C
v. For notprenabled(ti , k) < A
A. Either (@enabled(tt, k) ∈ A : pp < pi ) or (∀enabled(tt, k) ∈ A : pp ≮
pi ) where pp = Z(tt), pi = Z(ti ) – from rule a11, f 19 and forced atom
proposition
B. This matches the definition of an enabled priority transition
vi. Then ti is enabled and can fire in PN, as a result it can belong to T k – from
definition 40 of enabled transition
i.
ii.
iii.
iv.
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
279
(b) And consumesmore < A, since A is an answer set of Π7 – from rule a4 and supported
rule proposition
i. Then @consumesmore(p, k) ∈ A – from rule a3 and supported rule proposition
ii. Then @{holds(p, qc , c, k), tot decr(p, q1c , c, k)} ⊆ A, q1c > qc in body(a9) –
from a9 and forced atom proposition
P
iii. Then @c ∈ C@p ∈ P, ( ti ∈{t0 ,...,tx },p∈•ti mW(p,ti ) (c) +
P
ti ∈{t0 ,...,t x },p∈R(ti ) m Mk (p) (c)) > m Mk (p) (c) – from the following
A. holds(p, qc , c, k) represents qc = m Mk (p) (c) – from rule PN encoding, given
B. tot decr(p, q1c , c, k) ∈ A if {del(p, q10 c , t0 , c, k), . . . ,
del(p, q1 x c , t x , c, k)} ⊆ A, where q1c = q10 c + · · · + q1 x c – from r11 and
forced atom proposition
C. del(p, q1i c , ti , c, k) ∈ A if { f ires(ti , k), ptarc(p, ti , q1i c , c, k)} ⊆ A – from r9
and supported rule proposition
D. del(p, q1i c , ti , c, k) either represents removal of q1i c = mW(p,ti ) (c) tokens
from p ∈ •ti ; or it represents removal of q1i c = m Mk (p) (c) tokens from p ∈
R(ti )– from rules r9, f 14, f 16, supported rule proposition, and definition 43
of transition execution in PN
iv. Then the set of transitions in T k do not conflict – by the definition 34 of conflicting transitions
(c) And for each prenabled(t j , k) ∈ A and f ires(t j , k) < A,
could not have(t j , k) ∈ A, since A is an answer set of Π7 - from rule a15 and supported rule proposition
i. Then {prenabled(t j , k), holds(s, qqc , c, k), ptarc(s, t j , qc , c, k),
tot decr(s, qqqc , c, k)} ⊆ A, such that qc > qqc − qqqc and f ires(t j , k) < A from rule a14 and supported rule proposition
P
P
ii. Then for an s ∈ •t j ∪R(t j ), qc > m Mk (s) (c)−( ti ∈Tk ,s∈•ti mW(s,ti ) (c)+ ti ∈Tk ,s∈R(ti ) m Mk (s) (c)),
where qc = mW(s,t j ) (c) if s ∈ •t j , or m Mk (s) (c) otherwise.
A. ptarc(s, ti , qc , c, k) represents qc = mW(s,ti ) (c) if (s, ti ) ∈ E − or qc = m Mk (s) (c)
if s ∈ R(ti ) – from rule f 14, f 16 construction
B. holds(s, qqc , c, k) represents qqc = m Mk (s) (c) – from i2 construction
C. tot decr(s, qqqc , c, k) ∈ A if {del(s, qqq0 c , t0 , c, k), . . . ,
del(s, qqq x c , t x , c, k)} ⊆ A – from rule r11 construction and supported rule
proposition
D. del(s, qqqi c , ti , c, k) ∈ A if { f ires(ti , k), ptarc(s, ti , qqqi c , c, k)} ⊆ A – from
rule r9 and supported rule proposition
E. del(s, qqqi c , ti , c, k) represents qqqi c = mW(s,ti ) (c) : ti ∈ T k , (s, ti ) ∈ E − , or
qqqi c = m Mk (ti ) (c) : ti ∈ T k , s ∈ R(ti ) – from rule f 14, f 16 construction
P
F. tot decr(q, qqqc , c, k) represents ti ∈Tk ,s∈•ti mW(s,ti ) (c) +
P
ti ∈T k ,s∈R(ti ) m Mk (s) (c) – from (C,D,E) above
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
280
iii. Then firing T k ∪ {t j } would have required more tokens than are present at its
source place s ∈ •t j ∪ R(t j ). Thus, T k is a maximal set of transitions that can
simultaneously fire.
(d) And for each reset transition tr with enabled(tr , k) ∈ A, f ires(tr , k) ∈ A, since A is
an answer set of Π7 - from rule f 20 and supported rule proposition
i. Then the firing set T k satisfies the reset transition requirement of definition 41
(firing set)
(e) Then {t0 , . . . , t x } = T k – using 1(a),1(b), 1(d) above; and using 1(c) it is a maximal
firing set
2. We show that Mk+1 is produced by firing T k in Mk . Let holds(p, qc , c, k + 1) ∈ A
(a) Then {holds(p, q1c , c, k), tot incr(p, q2c , c, k), tot decr(p, q3c , c, k)} ⊆ A : qc = q1c +
q2c − q3c – from rule r12 and supported rule proposition
(b) Then holds(p, q1c , c, k) ∈ A represents q1c = m Mk (p) (c) – inductive assumption; and
{add(p, q20 c , t0 , c, k), . . . , add(p, q2 j c , t j , c, k)} ⊆ A : q20 c + · · · + q2 j c = q2c and
{del(p, q30 c , t0 , c, k), . . . , del(p, q3l c , tl , c, k)} ⊆ A : q30 c + · · · + q3l c = q3c – rules
r10, r11 and supported rule proposition, respectively
(c) Then { f ires(t0 , k), . . . , f ires(t j , k)} ⊆ A and { f ires(t0 , k), . . . , f ires(tl , k)}
⊆ A – rules r13, r9 and supported rule proposition, respectively
(d) Then { f ires(t0 , k), . . . , f ires(t j , k)}∪{ f ires(t0 , k), . . . , f ires(tl , k)} ⊆ A = { f ires(t0 , k), . . . , f ires(t x , k)} ⊆
A – set union of subsets
(e) Then for each f ires(t x , k) ∈ A we have t x ∈ T k – already shown in item 1 above
P
(f) Then qc = m Mk (p) (c) + tx ∈Tl ,p∈tx •,0≤l≤k,l+D(tx )−1=k mW(tx ,p) (c) −
P
P
( tx ∈T k ∧p∈•tx mW(p,tx ) (c) + tx ∈T k ∧p∈R(tx ) m Mk (p) (c)) – from (2b) above and the following
i. Each add(p, q j c , t j , c, k) ∈ A represents q j c = mW(t j ,p) (c) for p ∈ t j • – rule
r13, f 21 encoding, and definition 31 of postset in PN
ii. Each del(p, ty , qy c , c, k) ∈ A represents either qy c = mW(p,ty ) (c) for p ∈ •ty , or
qy c = m Mk (p) (c) for p ∈ R(ty ) – from rule r9, f 14 encoding and definition 31 of
preset in PN; or from rule r9, f 16 encoding and definition of reset arc in PN
iii. Each tot incr(p, q2c , c, k) ∈ A represents
P
q2c = tx ∈Tl ,p∈tx •,0≤l≤k,l+D(tx )−1=k mW(tx ,p) (c) – aggregate assignment atom semantics in rule r10
P
iv. Each tot decr(p, q3c , c, k) ∈ A represents q3c = tx ∈Tk ∧p∈•tx mW(p,tx ) (c)
P
+ tx ∈Tk ∧p∈R(tx ) m Mk (p) (c) – aggregate assignment atom semantics in rule r11
(g) Then, m Mk+1 (p) (c) = qc – since holds(p, qc , c, k + 1) ∈ A encodes qc = m Mk+1 (p) (c) –
from construction
APPENDIX A. PROOFS OF VARIOUS PROPOSITIONS
281
As a result, for any n > k, T n will be a valid firing set for Mn and Mn+1 will be its target marking.
Conclusion: Since both (I) and (II) hold, X = M0 , T k , M1 , . . . , Mk , T k+1 is an execution sequence of PN(P, T, E, C, W, R, I, Q, QW, Z, D) (w.r.t M0 ) iff there is an answer set A of Π7 (PN, M0 , k, ntok)
such that (A.15) and (A.16) hold.
Appendix B
Complete Set of Queries Used for
Drug-Drug Interaction
B.1
1
2
3
4
5
6
7
8
9
10
Drug Activates Gene
/ / S { / NP { / / ? [ V a l u e = ’ a c t i v a t i o n ’ ] ( kw1 ) = > / / ? [ V a l u e = ’ of ’ ] ( kw2 ) = > / / ? [ Tag = ’GENE ’ ] (
kw0 ) = > / / ? [ V a l u e = ’ by ’ ] ( kw4 ) = > / / ? [ Tag = ’DRUG’ ] ( kw3 ) } } : : : d i s t i n c t s e n t . c i d ,
kw3 . v a l u e , kw1 . v a l u e , kw0 . v a l u e , s e n t . v a l u e
/ / NP { / NP { / ? [ Tag = ’DRUG’ ] ( kw3 ) = >/?[ V a l u e = ’ a c t i v a t i o n ’ ] ( kw1 ) }= >/ PP { / ? [ V a l u e = ’ of
’ ] ( kw2 ) = > / / ? [ Tag = ’GENE ’ ] ( kw0 ) } } : : : d i s t i n c t s e n t . c i d , kw3 . v a l u e , kw1 .
v a l u e , kw0 . v a l u e , s e n t . v a l u e
/ / NP { / ? [ V a l u e = ’ a c t i v a t i o n ’ ] ( kw1 ) = >/PP { / ? [ V a l u e = ’ of ’ ] ( kw2 ) = > / / ? [ Tag = ’GENE ’ ] (
kw0 ) = > / / ? [ V a l u e = ’ by ’ ] ( kw4 ) = > / / ? [ Tag = ’DRUG’ ] ( kw3 ) } } : : : d i s t i n c t s e n t . c i d ,
kw3 . v a l u e , kw1 . v a l u e , kw0 . v a l u e , s e n t . v a l u e
/ / S { / NP { / / ? [ Tag = ’GENE ’ ] ( kw2 ) }= >/VP { / / ? [ V a l u e = ’ a c t i v a t i o n ’ ] ( kw1 ) = > / / ? [ Tag = ’
DRUG’ ] ( kw0 ) } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e ,
sent . value
/ / S { / NP { / / ? [ Tag = ’DRUG’ ] ( kw0 ) }= >/VP { / / ? [ V a l u e = ’ a c t i v a t i o n ’ ] ( kw1 ) = > / / ? [ Tag = ’
GENE ’ ] ( kw2 ) } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e ,
sent . value
/ / NP { / NP { / ? [ Tag = ’DRUG’ ] ( kw0 ) = >/?[ V a l u e = ’ a c t i v a t i o n ’ ] ( kw1 ) }= >/ PP { / / ? [ Tag = ’GENE
’ ] ( kw2 ) } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t .
value
/ / S { / NP { / / ? [ V a l u e = ’ a c t i v a t i o n ’ ] ( kw1 ) = > / / ? [ Tag = ’GENE ’ ] ( kw2 ) }= >/VP { / / ? [ Tag = ’
DRUG’ ] ( kw0 ) } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e ,
sent . value
/ / S { / NP { / / ? [ V a l u e = ’ a c t i v a t i o n ’ ] ( kw1 ) = > / / ? [ V a l u e = ’ of ’ ] ( kw2 ) = > / / ? [ Tag = ’GENE ’ ] (
kw0 ) = > / / ? [ V a l u e = ’ by ’ ] ( kw4 ) = > / / ? [ Tag = ’DRUG’ ] ( kw3 ) } } : : : d i s t i n c t s e n t . c i d ,
kw3 . v a l u e , kw1 . v a l u e , kw0 . v a l u e , s e n t . v a l u e
/ / NP { / NP { / ? [ Tag = ’DRUG’ ] ( kw3 ) = >/?[ V a l u e = ’ a c t i v a t i o n ’ ] ( kw1 ) }= >/ PP { / ? [ V a l u e = ’ of
’ ] ( kw2 ) = > / / ? [ Tag = ’GENE ’ ] ( kw0 ) } } : : : d i s t i n c t s e n t . c i d , kw3 . v a l u e , kw1 .
v a l u e , kw0 . v a l u e , s e n t . v a l u e
/ / NP { / ? [ V a l u e = ’ a c t i v a t i o n ’ ] ( kw1 ) = >/PP { / ? [ V a l u e = ’ of ’ ] ( kw2 ) = > / / ? [ Tag = ’GENE ’ ] (
kw0 ) = > / / ? [ V a l u e = ’ by ’ ] ( kw4 ) = > / / ? [ Tag = ’DRUG’ ] ( kw3 ) } } : : : d i s t i n c t s e n t . c i d ,
kw3 . v a l u e , kw1 . v a l u e , kw0 . v a l u e , s e n t . v a l u e
282
APPENDIX B. DRUG-DRUG INTERACTION QUERIES
B.2
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
283
Gene Induces Gene
/ / S { / NP { / / ? [ Tag = ’GENE ’ ] ( kw0 ) }= >/VP { / / ? [ V a l u e = ’ i n d u c e d ’ ] ( kw1 ) = > / / ? [ V a l u e = ’ by
’ ] ( kw3 ) = > / / ? [ Tag = ’GENE ’ ] ( kw2 ) } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 .
v a l u e , kw2 . v a l u e , s e n t . v a l u e
/ / S { / NP { / ? [ Tag = ’GENE ’ ] ( kw0 ) }= >/VP { / / ? [ V a l u e = ’ i n d u c e d ’ ] ( kw1 ) = > / / ? [ V a l u e = ’ by ’ ] (
kw3 ) = > / / ? [ Tag = ’GENE ’ ] ( kw2 ) } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e ,
kw2 . v a l u e , s e n t . v a l u e
/ / S { / NP { / / ? [ Tag = ’GENE ’ ] ( kw0 ) }= >/VP { / / ? [ V a l u e IN { ’ i n c r e a s e ’ , ’ i n c r e a s e d ’ } ] ( kw1
) = > / / ? [ Tag = ’GENE ’ ] ( kw2 ) = > / / ? [ V a l u e = ’ a c t i v i t y ’ ] ( kw3 ) } } : : : d i s t i n c t s e n t .
c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t . v a l u e
/ / S { / NP { / / ? [ Tag = ’GENE ’ ] ( kw0 ) = > / / ? [ V a l u e = ’ a c t i v i t y ’ ] ( kw3 ) }= >/VP { / / ? [ V a l u e = ’
i n c r e a s e ’ ] ( kw1 ) = > / / ? [ Tag = ’GENE ’ ] ( kw2 ) } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e ,
kw1 . v a l u e , kw2 . v a l u e , s e n t . v a l u e
/ / S { / NP { / / ? [ Tag = ’GENE ’ ] ( kw0 ) }= >/VP { / / ? [ V a l u e IN { ’ s t i m u l a t e s ’ , ’ s t i m u l a t e ’ , ’
s t i m u l a t e d ’ } ] ( kw1 ) = > / / ? [ Tag = ’GENE ’ ] ( kw2 ) } } : : : d i s t i n c t s e n t . c i d , kw0 .
v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t . v a l u e
/ / S { / NP { / / ? [ Tag = ’GENE ’ ] ( kw0 ) }= >/VP { / ? [ V a l u e IN { ’ s t i m u l a t e s ’ , ’ s t i m u l a t e ’ } ] (
kw1 ) = > / / ? [ Tag = ’GENE ’ ] ( kw2 ) } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e ,
kw2 . v a l u e , s e n t . v a l u e
/ / S { / NP { / ? [ Tag = ’GENE ’ ] ( kw0 ) }= >/VP { / ? [ V a l u e = ’ s t i m u l a t e s ’ ] ( kw1 ) = > / / ? [ Tag = ’GENE
’ ] ( kw2 ) } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t .
value
/ / S { / NP { / / ? [ Tag = ’GENE ’ ] ( kw0 ) }= >/VP { / ? [ V a l u e = ’ a c t i v a t e d ’ ] ( kw1 ) = > / / ? [ Tag = ’GENE
’ ] ( kw2 ) } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t .
value
/ / NP { / NP { / ? [ V a l u e = ’ i n d u c t i o n ’ ] ( kw1 ) }= >/ PP { / ? [ V a l u e = ’ of ’ ] ( kw3 ) = >/NP { / / ? [ Tag = ’
GENE ’ ] ( kw0 ) = > / / ? [ V a l u e = ’ by ’ ] ( kw4 ) = > / / ? [ Tag = ’GENE ’ ] ( kw2 ) } } } : : : d i s t i n c t
s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t . v a l u e
/ / NP { / NP { / ? [ V a l u e = ’ s t i m u l a t i o n ’ ] ( kw1 ) }= >/ PP { / ? [ V a l u e = ’ of ’ ] ( kw3 ) = >/NP { / / ? [ Tag
= ’GENE ’ ] ( kw0 ) = > / / ? [ V a l u e = ’ by ’ ] ( kw4 ) = > / / ? [ Tag = ’GENE ’ ] ( kw2 ) } } } : : : d i s t i n c t
s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t . v a l u e
/ / NP { / NP { / ? [ V a l u e = ’ a c t i v a t i o n ’ ] ( kw1 ) }= >/ PP { / ? [ V a l u e IN { ’ of ’ } ] ( kw3 ) = >/NP { / / ? [
Tag = ’GENE ’ ] ( kw0 ) = > / / ? [ V a l u e = ’ by ’ ] ( kw4 ) = > / / ? [ Tag = ’GENE ’ ] ( kw2 ) } } } : : :
d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t . v a l u e
/ / S { / NP { / / ? [ Tag = ’GENE ’ ] ( kw0 ) }= >/VP { / / ? [ V a l u e = ’ i n d u c i b l e ’ ] ( kw1 ) = > / / ? [ Tag = ’GENE
’ ] ( kw2 ) } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t .
value
/ / S { / NP { / / ? [ Tag = ’GENE ’ ] ( kw0 ) }= >/VP { / / ? [ V a l u e = ’ i n d u c e d ’ ] ( kw1 ) = > / / ? [ V a l u e = ’ by
’ ] ( kw3 ) = > / / ? [ Tag = ’GENE ’ ] ( kw2 ) } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 .
v a l u e , kw2 . v a l u e , s e n t . v a l u e
/ / S { / NP { / ? [ Tag = ’GENE ’ ] ( kw0 ) }= >/VP { / / ? [ V a l u e = ’ i n d u c e d ’ ] ( kw1 ) = > / / ? [ V a l u e = ’ by ’ ] (
kw3 ) = > / / ? [ Tag = ’GENE ’ ] ( kw2 ) } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e ,
kw2 . v a l u e , s e n t . v a l u e
/ / S { / NP { / / ? [ Tag = ’GENE ’ ] ( kw0 ) }= >/VP { / / ? [ V a l u e IN { ’ i n c r e a s e ’ , ’ i n c r e a s e d ’ } ] ( kw1
) = > / / ? [ Tag = ’GENE ’ ] ( kw2 ) = > / / ? [ V a l u e = ’ a c t i v i t y ’ ] ( kw3 ) } } : : : d i s t i n c t s e n t .
c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t . v a l u e
/ / S { / NP { / / ? [ Tag = ’GENE ’ ] ( kw0 ) = > / / ? [ V a l u e = ’ a c t i v i t y ’ ] ( kw3 ) }= >/VP { / / ? [ V a l u e = ’
i n c r e a s e ’ ] ( kw1 ) = > / / ? [ Tag = ’GENE ’ ] ( kw2 ) } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e ,
kw1 . v a l u e , kw2 . v a l u e , s e n t . v a l u e
/ / S { / NP { / / ? [ Tag = ’GENE ’ ] ( kw0 ) }= >/VP { / / ? [ V a l u e = ’ s t i m u l a t e d ’ ] ( kw1 ) = > / / ? [ Tag = ’
APPENDIX B. DRUG-DRUG INTERACTION QUERIES
18
19
20
21
22
23
24
25
GENE ’ ] ( kw2 ) } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e ,
sent . value
/ / S { / NP { / / ? [ Tag = ’GENE ’ ] ( kw0 ) }= >/VP { / ? [ V a l u e IN { ’ s t i m u l a t e s ’ , ’ s t i m u l a t e ’ } ] (
kw1 ) = > / / ? [ Tag = ’GENE ’ ] ( kw2 ) } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e ,
kw2 . v a l u e , s e n t . v a l u e
/ / S { / NP { / ? [ Tag = ’GENE ’ ] ( kw0 ) }= >/VP { / ? [ V a l u e = ’ s t i m u l a t e s ’ ] ( kw1 ) = > / / ? [ Tag = ’GENE
’ ] ( kw2 ) } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t .
value
/ / S { / NP { / / ? [ Tag = ’GENE ’ ] ( kw0 ) }= >/VP { / ? [ V a l u e = ’ a c t i v a t e d ’ ] ( kw1 ) = > / / ? [ Tag = ’GENE
’ ] ( kw2 ) } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t .
value
/ / VP { / ? [ V a l u e = ’ a c t i v a t e d ’ ] ( kw1 ) = >/PP { / / ? [ Tag = ’GENE ’ ] ( kw0 ) = > / / ? [ Tag = ’GENE ’ ] (
kw2 ) } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t . v a l u e
/ / NP { / NP { / ? [ V a l u e = ’ i n d u c t i o n ’ ] ( kw1 ) }= >/ PP { / ? [ V a l u e = ’ of ’ ] ( kw3 ) = >/NP { / / ? [ Tag = ’
GENE ’ ] ( kw0 ) = > / / ? [ V a l u e = ’ by ’ ] ( kw4 ) = > / / ? [ Tag = ’GENE ’ ] ( kw2 ) } } } : : : d i s t i n c t
s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t . v a l u e
/ / NP { / NP { / ? [ V a l u e = ’ s t i m u l a t i o n ’ ] ( kw1 ) }= >/ PP { / ? [ V a l u e = ’ of ’ ] ( kw3 ) = >/NP { / / ? [ Tag
= ’GENE ’ ] ( kw0 ) = > / / ? [ V a l u e = ’ by ’ ] ( kw4 ) = > / / ? [ Tag = ’GENE ’ ] ( kw2 ) } } } : : : d i s t i n c t
s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t . v a l u e
/ / NP { / NP { / ? [ V a l u e = ’ a c t i v a t i o n ’ ] ( kw1 ) }= >/ PP { / ? [ V a l u e IN { ’ of ’ } ] ( kw3 ) = >/NP { / / ? [
Tag = ’GENE ’ ] ( kw0 ) = > / / ? [ V a l u e = ’ by ’ ] ( kw4 ) = > / / ? [ Tag = ’GENE ’ ] ( kw2 ) } } } : : :
d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t . v a l u e
/ / S { / NP { / / ? [ Tag = ’GENE ’ ] ( kw0 ) }= >/VP { / / ? [ V a l u e = ’ i n d u c i b l e ’ ] ( kw1 ) = > / / ? [ Tag = ’GENE
’ ] ( kw2 ) } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t .
value
B.3
1
2
3
4
5
6
7
8
284
Gene Inhibits Gene
/ / S { / NP { / / ? [ Tag = ’GENE ’ ] ( kw0 ) }= >/VP { / / ? [ V a l u e = ’ i n d u c e d ’ ] ( kw1 ) = > / / ? [ V a l u e = ’ by
’ ] ( kw3 ) = > / / ? [ Tag = ’GENE ’ ] ( kw2 ) } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 .
v a l u e , kw2 . v a l u e , s e n t . v a l u e
/ / S { / NP { / ? [ Tag = ’GENE ’ ] ( kw0 ) }= >/VP { / / ? [ V a l u e = ’ i n d u c e d ’ ] ( kw1 ) = > / / ? [ V a l u e = ’ by ’ ] (
kw3 ) = > / / ? [ Tag = ’GENE ’ ] ( kw2 ) } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e ,
kw2 . v a l u e , s e n t . v a l u e
/ / S { / NP { / / ? [ Tag = ’GENE ’ ] ( kw0 ) }= >/VP { / / ? [ V a l u e IN { ’ i n c r e a s e ’ , ’ i n c r e a s e d ’ } ] ( kw1
) = > / / ? [ Tag = ’GENE ’ ] ( kw2 ) = > / / ? [ V a l u e = ’ a c t i v i t y ’ ] ( kw3 ) } } : : : d i s t i n c t s e n t .
c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t . v a l u e
/ / S { / NP { / / ? [ Tag = ’GENE ’ ] ( kw0 ) = > / / ? [ V a l u e = ’ a c t i v i t y ’ ] ( kw3 ) }= >/VP { / / ? [ V a l u e = ’
i n c r e a s e ’ ] ( kw1 ) = > / / ? [ Tag = ’GENE ’ ] ( kw2 ) } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e ,
kw1 . v a l u e , kw2 . v a l u e , s e n t . v a l u e
/ / S { / NP { / / ? [ Tag = ’GENE ’ ] ( kw0 ) }= >/VP { / / ? [ V a l u e IN { ’ s t i m u l a t e s ’ , ’ s t i m u l a t e ’ , ’
s t i m u l a t e d ’ } ] ( kw1 ) = > / / ? [ Tag = ’GENE ’ ] ( kw2 ) } } : : : d i s t i n c t s e n t . c i d , kw0 .
v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t . v a l u e
/ / S { / NP { / / ? [ Tag = ’GENE ’ ] ( kw0 ) }= >/VP { / ? [ V a l u e IN { ’ s t i m u l a t e s ’ , ’ s t i m u l a t e ’ } ] (
kw1 ) = > / / ? [ Tag = ’GENE ’ ] ( kw2 ) } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e ,
kw2 . v a l u e , s e n t . v a l u e
/ / S { / NP { / ? [ Tag = ’GENE ’ ] ( kw0 ) }= >/VP { / ? [ V a l u e = ’ s t i m u l a t e s ’ ] ( kw1 ) = > / / ? [ Tag = ’GENE
’ ] ( kw2 ) } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t .
value
/ / S { / NP { / / ? [ Tag = ’GENE ’ ] ( kw0 ) }= >/VP { / ? [ V a l u e = ’ a c t i v a t e d ’ ] ( kw1 ) = > / / ? [ Tag = ’GENE
’ ] ( kw2 ) } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t .
APPENDIX B. DRUG-DRUG INTERACTION QUERIES
9
10
11
12
value
/ / NP { / NP { / ? [ V a l u e = ’ i n d u c t i o n ’ ] ( kw1 ) }= >/ PP { / ? [ V a l u e = ’ of ’ ] ( kw3 ) = >/NP { / / ? [ Tag = ’
GENE ’ ] ( kw0 ) = > / / ? [ V a l u e = ’ by ’ ] ( kw4 ) = > / / ? [ Tag = ’GENE ’ ] ( kw2 ) } } } : : : d i s t i n c t
s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t . v a l u e
/ / NP { / NP { / ? [ V a l u e = ’ s t i m u l a t i o n ’ ] ( kw1 ) }= >/ PP { / ? [ V a l u e = ’ of ’ ] ( kw3 ) = >/NP { / / ? [ Tag
= ’GENE ’ ] ( kw0 ) = > / / ? [ V a l u e = ’ by ’ ] ( kw4 ) = > / / ? [ Tag = ’GENE ’ ] ( kw2 ) } } } : : : d i s t i n c t
s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t . v a l u e
/ / NP { / NP { / ? [ V a l u e = ’ a c t i v a t i o n ’ ] ( kw1 ) }= >/ PP { / ? [ V a l u e IN { ’ of ’ } ] ( kw3 ) = >/NP { / / ? [
Tag = ’GENE ’ ] ( kw0 ) = > / / ? [ V a l u e = ’ by ’ ] ( kw4 ) = > / / ? [ Tag = ’GENE ’ ] ( kw2 ) } } } : : :
d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t . v a l u e
/ / S { / NP { / / ? [ Tag = ’GENE ’ ] ( kw0 ) }= >/VP { / / ? [ V a l u e = ’ i n d u c i b l e ’ ] ( kw1 ) = > / / ? [ Tag = ’GENE
’ ] ( kw2 ) } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t .
value
B.4
1
2
3
4
5
6
7
8
9
10
11
285
Drug Changes Gene Expression/Activity
/ / S { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/?[ V a l u e IN { ’ i n c r e a s e d ’ , ’ i n c r e a s e ’ , ’ i n c r e a s e s ’ } ] (
kw1 ) = >/?[ V a l u e IN { ’ l e v e l s ’ , ’ l e v e l ’ } ] ( kw3 ) = >/?[ Tag = ’GENE ’ ] ( kw0 ) } : : :
d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t . v a l u e
/ / S { / NP { / ? [ Tag = ’DRUG’ ] ( kw2 ) }= >/VP { / ? [ V a l u e IN { ’ i n c r e a s e d ’ , ’ i n c r e a s e s ’ } ] ( kw1 )
= >/NP { / / ? [ Tag = ’GENE ’ ] ( kw0 ) = > / / ? [ V a l u e IN { ’ e x p r e s s i o n ’ , ’ l e v e l ’ , ’ a c t i v i t y
’ , ’ a c t i v i t i e s ’ , ’ l e v e l s ’ } ] ( kw3 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 .
v a l u e , kw2 . v a l u e , s e n t . v a l u e
/ / S { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/VP { / NP { / / ? [ V a l u e IN { ’ i n c r e a s e s ’ , ’ i n c r e a s e ’ } ] ( kw1 )
= > / / ? [ Tag = ’GENE ’ ] ( kw0 ) = > / / ? [ V a l u e IN { ’ l e v e l s ’ , ’ a c t i v i t y ’ } ] ( kw3 ) } } } : : :
d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t . v a l u e
/ / S { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/VP { / ? [ V a l u e IN { ’ i n c r e a s e d ’ , ’ i n c r e a s e s ’ } ] ( kw1 ) = >/NP
{ / / ? [ Tag = ’GENE ’ ] ( kw0 ) = > / / ? [ V a l u e IN { ’ a c t i v i t y ’ , ’ a c t i v i t i e s ’ , ’ l e v e l s ’ } ] (
kw3 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t .
value
/ / S { / NP { / ? [ Tag = ’DRUG’ ] ( kw2 ) }= >/VP { / ? [ V a l u e = ’ i n c r e a s e d ’ ] ( kw1 ) = >/NP { / / ? [ V a l u e
IN { ’ a c t i v i t y ’ , ’ l e v e l s ’ , ’ e x p r e s s i o n ’ } ] ( kw3 ) = > / / ? [ Tag = ’GENE ’ ] ( kw0 ) } } } : : :
d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t . v a l u e
/ / S { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/VP { / VP { / / ? [ V a l u e IN { ’ i n c r e a s e d ’ , ’ i n c r e a s e s ’ , ’
i n c r e a s e ’ } ] ( kw1 ) = > / / ? [ Tag = ’GENE ’ ] ( kw0 ) = > / / ? [ V a l u e IN { ’ a c t i v i t y ’ , ’ l e v e l s
’ } ] ( kw3 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t .
value
/ / S { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/VP { / VP { / ? [ V a l u e = ’ i n c r e a s e d ’ ] ( kw1 ) = > / / ? [ Tag = ’GENE ’ ] (
kw0 ) = > / / ? [ V a l u e IN { ’ e x p r e s s i o n ’ , ’ a c t i v i t y ’ } ] ( kw3 ) } } } : : : d i s t i n c t s e n t .
c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t . v a l u e
/ / S { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/VP { / ? [ V a l u e IN { ’ i n c r e a s e d ’ , ’ i n c r e a s e s ’ } ] ( kw1 ) = >/NP
{ / / ? [ V a l u e IN { ’ e x p r e s s i o n ’ , ’ a c t i v i t y ’ , ’ a c t i v i t i e s ’ , ’ l e v e l ’ , ’ l e v e l s ’ } ] (
kw3 ) = > / / ? [ Tag = ’GENE ’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e
, kw2 . v a l u e , s e n t . v a l u e
/ / S { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/VP { / ? [ V a l u e IN { ’ d e c r e a s e d ’ , ’ d e c r e a s e s ’ } ] ( kw1 ) = >/NP
{ / / ? [ Tag = ’GENE ’ ] ( kw0 ) = > / / ? [ V a l u e IN { ’ a c t i v i t y ’ , ’ e x p r e s s i o n ’ } ] ( kw3 ) } } }
: : : d i s t i n c t s e n t . c i d , kw2 . v a l u e , kw1 , v a l u e , kw0 . v a l u e , s e n t . v a l u e
/ / S { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/?[ V a l u e IN { ’ i n c r e a s e d ’ , ’ i n c r e a s e ’ , ’ i n c r e a s e s ’ } ] (
kw1 ) = >/?[ V a l u e IN { ’ l e v e l s ’ , ’ l e v e l ’ } ] ( kw3 ) = >/?[ Tag = ’GENE ’ ] ( kw0 ) } : : :
d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , kw3 . v a l u e , s e n t . v a l u e
/ / S { / NP { / ? [ Tag = ’DRUG’ ] ( kw2 ) }= >/VP { / ? [ V a l u e IN { ’ i n c r e a s e d ’ , ’ i n c r e a s e s ’ } ] ( kw1 )
APPENDIX B. DRUG-DRUG INTERACTION QUERIES
12
13
14
15
16
17
18
= >/NP { / / ? [ Tag = ’GENE ’ ] ( kw0 ) = > / / ? [ V a l u e IN { ’ e x p r e s s i o n ’ , ’ l e v e l ’ , ’ a c t i v i t y
’ , ’ a c t i v i t i e s ’ , ’ l e v e l s ’ } ] ( kw3 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 .
v a l u e , kw2 . v a l u e , kw3 . v a l u e , s e n t . v a l u e
/ / S { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/VP { / NP { / / ? [ V a l u e IN { ’ i n c r e a s e s ’ , ’ i n c r e a s e ’ } ] ( kw1 )
= > / / ? [ Tag = ’GENE ’ ] ( kw0 ) = > / / ? [ V a l u e IN { ’ l e v e l s ’ , ’ a c t i v i t y ’ } ] ( kw3 ) } } } : : :
d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , kw3 . v a l u e , s e n t . v a l u e
/ / S { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/VP { / ? [ V a l u e IN { ’ i n c r e a s e d ’ , ’ i n c r e a s e s ’ } ] ( kw1 ) = >/NP
{ / / ? [ Tag = ’GENE ’ ] ( kw0 ) = > / / ? [ V a l u e IN { ’ a c t i v i t y ’ , ’ a c t i v i t i e s ’ , ’ l e v e l s ’ } ] (
kw3 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , kw3 . v a l u e
, sent . value
/ / S { / NP { / ? [ Tag = ’DRUG’ ] ( kw2 ) }= >/VP { / ? [ V a l u e = ’ i n c r e a s e d ’ ] ( kw1 ) = >/NP { / / ? [ V a l u e
IN { ’ a c t i v i t y ’ , ’ l e v e l s ’ , ’ e x p r e s s i o n ’ } ] ( kw3 ) = > / / ? [ Tag = ’GENE ’ ] ( kw0 ) } } } : : :
d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , kw3 . v a l u e , s e n t . v a l u e
/ / S { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/VP { / VP { / / ? [ V a l u e IN { ’ i n c r e a s e d ’ , ’ i n c r e a s e s ’ , ’
i n c r e a s e ’ } ] ( kw1 ) = > / / ? [ Tag = ’GENE ’ ] ( kw0 ) = > / / ? [ V a l u e IN { ’ a c t i v i t y ’ , ’ l e v e l s
’ } ] ( kw3 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , kw3 .
value , s e n t . value
/ / S { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/VP { / VP { / ? [ V a l u e = ’ i n c r e a s e d ’ ] ( kw1 ) = > / / ? [ Tag = ’GENE ’ ] (
kw0 ) = > / / ? [ V a l u e IN { ’ e x p r e s s i o n ’ , ’ a c t i v i t y ’ } ] ( kw3 ) } } } : : : d i s t i n c t s e n t .
c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , kw3 . v a l u e , s e n t . v a l u e
/ / S { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/VP { / ? [ V a l u e IN { ’ i n c r e a s e d ’ , ’ i n c r e a s e s ’ } ] ( kw1 ) = >/NP
{ / / ? [ V a l u e IN { ’ e x p r e s s i o n ’ , ’ a c t i v i t y ’ , ’ a c t i v i t i e s ’ , ’ l e v e l ’ , ’ l e v e l s ’ } ] (
kw3 ) = > / / ? [ Tag = ’GENE ’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e
, kw2 . v a l u e , kw3 . v a l u e , s e n t . v a l u e
/ / S { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/VP { / ? [ V a l u e IN { ’ d e c r e a s e d ’ , ’ d e c r e a s e s ’ } ] ( kw1 ) = >/NP
{ / / ? [ Tag = ’GENE ’ ] ( kw0 ) = > / / ? [ V a l u e IN { ’ a c t i v i t y ’ , ’ e x p r e s s i o n ’ } ] ( kw3 ) } } }
: : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , kw3 . v a l u e , s e n t .
value
B.5
1
2
3
4
5
6
7
286
Drug Induces/Stimulates Gene
/ / S { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/VP { / ? [ V a l u e IN { ’ s t i m u l a t e d ’ , ’ i n d u c e d ’ } ] ( kw1 ) = >/NP
{ / / ? [ Tag = ’GENE ’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw2 . v a l u e , kw1 . v a l u e , kw0
. value , s e n t . value
/ / VP { / ? [ V a l u e IN { ’ s t i m u l a t e d ’ , ’ i n d u c e d ’ } ] ( kw1 ) = >/PP { / NP { / / ? [ Tag = ’DRUG’ ] ( kw2 )
= > / / ? [ Tag = ’GENE ’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw2 . v a l u e , kw1 . v a l u e ,
kw0 . v a l u e , s e n t . v a l u e
/ / S { / NP { / ? [ Tag = ’DRUG’ ] ( kw2 ) }= >/VP { / ? [ V a l u e IN { ’ s t i m u l a t e d ’ , ’ i n d u c e d ’ } ] ( kw1 )
= >/NP { / / ? [ Tag = ’GENE ’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw2 . v a l u e , kw1 . v a l u e
, kw0 . v a l u e , s e n t . v a l u e
/ / S { / NP { / ? [ Tag = ’GENE ’ ] ( kw0 ) }= >/VP { / ? [ V a l u e IN { ’ s t i m u l a t e d ’ , ’ i n d u c e d ’ } ] ( kw1 )
= >/PP { / / ? [ Tag = ’DRUG’ ] ( kw2 ) } } } : : : d i s t i n c t s e n t . c i d , kw2 . v a l u e , kw1 . v a l u e
, kw0 . v a l u e , s e n t . v a l u e
/ / S { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/VP { / NP { / / ? [ Tag = ’GENE ’ ] ( kw0 ) = > / / ? [ V a l u e IN { ’
s t i m u l a t e d ’ , ’ i n d u c e d ’ } ] ( kw1 ) } } } : : : d i s t i n c t s e n t . c i d , kw2 . v a l u e , kw1 .
v a l u e , kw0 . v a l u e , s e n t . v a l u e
/ / S { / NP { / PP { / / ? [ Tag = ’DRUG’ ] ( kw2 ) } } = > /VP { / ? [ V a l u e IN { ’ s t i m u l a t e d ’ , ’ i n d u c e d
’ } ] ( kw1 ) = >/NP { / / ? [ Tag = ’GENE ’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw2 . v a l u e ,
kw1 . v a l u e , kw0 . v a l u e , s e n t . v a l u e
/ / S { / NP { / ? [ Tag = ’GENE ’ ] ( kw0 ) }= >/VP { / VP { / ? [ V a l u e IN { ’ s t i m u l a t e d ’ , ’ i n d u c e d ’ } ] (
kw1 ) = > / / ? [ Tag = ’DRUG’ ] ( kw2 ) } } } : : : d i s t i n c t s e n t . c i d , kw2 . v a l u e , kw1 . v a l u e
APPENDIX B. DRUG-DRUG INTERACTION QUERIES
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
287
, kw0 . v a l u e , s e n t . v a l u e
/ / S { / NP { / ? [ Tag = ’GENE ’ ] ( kw0 ) }= >/VP { / VP { / / ? [ V a l u e = ’ i n d u c e d ’ ] ( kw1 ) = > / / ? [ Tag = ’
DRUG’ ] ( kw2 ) } } } : : : d i s t i n c t s e n t . c i d , kw2 . v a l u e , kw1 . v a l u e , kw0 . v a l u e ,
sent . value
/ / S { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/VP { / ? [ V a l u e IN { ’ s t i m u l a t e d ’ , ’ i n d u c e d ’ } ] ( kw1 ) = >/NP
{ / ? [ Tag = ’GENE ’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw2 . v a l u e , kw1 . v a l u e , kw0 .
value , s e n t . value
/ / S { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/VP { / VP { / / ? [ V a l u e IN { ’ s t i m u l a t e ’ , ’ i n d u c e ’ } ] ( kw1 )
= > / / ? [ Tag = ’GENE ’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw2 . v a l u e , kw1 . v a l u e ,
kw0 . v a l u e , s e n t . v a l u e
/ / NP { / NP { / ? [ V a l u e IN { ’ i n d u c t i o n ’ , ’ s t i m u l a t i o n ’ } ] ( kw1 ) }= >/ PP { / NP { / / ? [ Tag = ’
GENE ’ ] ( kw0 ) = > / / ? [ Tag = ’DRUG’ ] ( kw2 ) } } } : : : d i s t i n c t s e n t . c i d , kw2 . v a l u e ,
kw1 . v a l u e , kw0 . v a l u e , s e n t . v a l u e
/ / S { / NP { / ? [ V a l u e IN { ’ s t i m u l a t i o n ’ , ’ i n d u c t i o n ’ } ] ( kw1 ) = >/PP { / / ? [ Tag = ’GENE ’ ] (
kw0 ) } } = > /VP { / VP { / / ? [ Tag = ’DRUG’ ] ( kw2 ) } } } : : : d i s t i n c t s e n t . c i d , kw2 . v a l u e ,
kw1 . v a l u e , kw0 . v a l u e , s e n t . v a l u e
/ / S { / NP { / ? [ Tag = ’DRUG’ ] ( kw2 ) }= >/VP { / NP { / / ? [ V a l u e IN { ’ s t i m u l a t i o n ’ , ’ i n d u c t i o n
’ } ] ( kw1 ) = > / / ? [ Tag = ’GENE ’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw2 . v a l u e , kw1 .
v a l u e , kw0 . v a l u e , s e n t . v a l u e
/ / NP { / NP { / NP { / ? [ V a l u e = ’ i n d u c t i o n ’ ] ( kw1 ) }= >/ PP { / / ? [ Tag = ’GENE ’ ] ( kw0 ) } } = > / PP { / NP
{ / ? [ Tag = ’DRUG’ ] ( kw2 ) } } } : : : d i s t i n c t s e n t . c i d , kw2 . v a l u e , kw1 . v a l u e , kw0 .
value , s e n t . value
/ / NP { / NP { / ? [ Tag = ’GENE ’ ] ( kw0 ) = >/?[ V a l u e IN { ’ s t i m u l a t i o n ’ , ’ i n d u c t i o n ’ } ] ( kw1 )
}= >/ PP { / NP { / ? [ Tag = ’DRUG’ ] ( kw2 ) } } } : : : d i s t i n c t s e n t . c i d , kw2 . v a l u e , kw1 .
v a l u e , kw0 . v a l u e , s e n t . v a l u e
/ / NP { / ? [ V a l u e IN { ’ s t i m u l a t i o n ’ , ’ i n d u c t i o n ’ } ] ( kw1 ) = >/PP { / NP { / / ? [ Tag = ’GENE ’ ] (
kw0 ) = > / / ? [ Tag = ’DRUG’ ] ( kw2 ) } } } : : : d i s t i n c t s e n t . c i d , kw2 . v a l u e , kw1 . v a l u e
, kw0 . v a l u e , s e n t . v a l u e
/ / S { / NP { / ? [ Tag = ’DRUG’ ] ( kw2 ) }= >/VP { / ? [ V a l u e IN { ’ s t i m u l a t e s ’ , ’ i n d u c e s ’ } ] ( kw1 )
= >/NP { / / ? [ Tag = ’GENE ’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw2 . v a l u e , kw1 . v a l u e
, kw0 . v a l u e , s e n t . v a l u e
/ / S { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/VP { / ? [ V a l u e IN { ’ s t i m u l a t e s ’ , ’ i n d u c e s ’ } ] ( kw1 ) = >/NP
{ / / ? [ Tag = ’GENE ’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw2 . v a l u e , kw1 . v a l u e , kw0
. value , s e n t . value
/ / S { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/VP { / ? [ V a l u e IN { ’ s t i m u l a t e d ’ , ’ i n d u c e d ’ } ] ( kw1 ) = >/NP
{ / / ? [ Tag = ’GENE ’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2
. value , s e n t . value
/ / VP { / ? [ V a l u e IN { ’ s t i m u l a t e d ’ , ’ i n d u c e d ’ } ] ( kw1 ) = >/PP { / NP { / / ? [ Tag = ’DRUG’ ] ( kw2 )
= > / / ? [ Tag = ’GENE ’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e ,
kw2 . v a l u e , s e n t . v a l u e
/ / S { / NP { / ? [ Tag = ’DRUG’ ] ( kw2 ) }= >/VP { / ? [ V a l u e IN { ’ s t i m u l a t e d ’ , ’ i n d u c e d ’ } ] ( kw1 )
= >/NP { / / ? [ Tag = ’GENE ’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e
, kw2 . v a l u e , s e n t . v a l u e
/ / S { / NP { / ? [ Tag = ’GENE ’ ] ( kw0 ) }= >/VP { / ? [ V a l u e IN { ’ s t i m u l a t e d ’ , ’ i n d u c e d ’ } ] ( kw1 )
= >/PP { / / ? [ Tag = ’DRUG’ ] ( kw2 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e
, kw2 . v a l u e , s e n t . v a l u e
/ / S { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/VP { / NP { / / ? [ Tag = ’GENE ’ ] ( kw0 ) = > / / ? [ V a l u e IN { ’
s t i m u l a t e d ’ , ’ i n d u c e d ’ } ] ( kw1 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 .
v a l u e , kw2 . v a l u e , s e n t . v a l u e
/ / S { / NP { / PP { / / ? [ Tag = ’DRUG’ ] ( kw2 ) } } = > /VP { / ? [ V a l u e IN { ’ s t i m u l a t e d ’ , ’ i n d u c e d
’ } ] ( kw1 ) = >/NP { / / ? [ Tag = ’GENE ’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e ,
APPENDIX B. DRUG-DRUG INTERACTION QUERIES
25
26
27
28
29
30
31
32
33
34
35
36
kw1 . v a l u e , kw2 . v a l u e , s e n t . v a l u e
/ / S { / NP { / ? [ Tag = ’GENE ’ ] ( kw0 ) }= >/VP { / VP { / ? [ V a l u e IN { ’ s t i m u l a t e d ’ , ’ i n d u c e d ’ } ] (
kw1 ) = > / / ? [ Tag = ’DRUG’ ] ( kw2 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e
, kw2 . v a l u e , s e n t . v a l u e
/ / S { / NP { / ? [ Tag = ’GENE ’ ] ( kw0 ) }= >/VP { / VP { / / ? [ V a l u e = ’ i n d u c e d ’ ] ( kw1 ) = > / / ? [ Tag = ’
DRUG’ ] ( kw2 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e ,
sent . value
/ / S { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/VP { / ? [ V a l u e IN { ’ s t i m u l a t e d ’ , ’ i n d u c e d ’ } ] ( kw1 ) = >/NP
{ / ? [ Tag = ’GENE ’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 .
value , s e n t . value
/ / S { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/VP { / VP { / / ? [ V a l u e IN { ’ s t i m u l a t e ’ , ’ i n d u c e ’ } ] ( kw1 )
= > / / ? [ Tag = ’GENE ’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e ,
kw2 . v a l u e , s e n t . v a l u e
/ / NP { / NP { / ? [ V a l u e IN { ’ i n d u c t i o n ’ , ’ s t i m u l a t i o n ’ } ] ( kw1 ) }= >/ PP { / NP { / / ? [ Tag = ’
GENE ’ ] ( kw0 ) = > / / ? [ Tag = ’DRUG’ ] ( kw2 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e ,
kw1 . v a l u e , kw2 . v a l u e , s e n t . v a l u e
/ / S { / NP { / ? [ V a l u e IN { ’ s t i m u l a t i o n ’ , ’ i n d u c t i o n ’ } ] ( kw1 ) = >/PP { / / ? [ Tag = ’GENE ’ ] (
kw0 ) } } = > /VP { / VP { / / ? [ Tag = ’DRUG’ ] ( kw2 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e ,
kw1 . v a l u e , kw2 . v a l u e , s e n t . v a l u e
/ / S { / NP { / ? [ Tag = ’DRUG’ ] ( kw2 ) }= >/VP { / NP { / / ? [ V a l u e IN { ’ s t i m u l a t i o n ’ , ’ i n d u c t i o n
’ } ] ( kw1 ) = > / / ? [ Tag = ’GENE ’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 .
v a l u e , kw2 . v a l u e , s e n t . v a l u e
/ / NP { / NP { / NP { / ? [ V a l u e = ’ i n d u c t i o n ’ ] ( kw1 ) }= >/ PP { / / ? [ Tag = ’GENE ’ ] ( kw0 ) } } = > / PP { / NP
{ / ? [ Tag = ’DRUG’ ] ( kw2 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 .
value , s e n t . value
/ / NP { / NP { / ? [ Tag = ’GENE ’ ] ( kw0 ) = >/?[ V a l u e IN { ’ s t i m u l a t i o n ’ , ’ i n d u c t i o n ’ } ] ( kw1 )
}= >/ PP { / NP { / ? [ Tag = ’DRUG’ ] ( kw2 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 .
v a l u e , kw2 . v a l u e , s e n t . v a l u e
/ / NP { / ? [ V a l u e IN { ’ s t i m u l a t i o n ’ , ’ i n d u c t i o n ’ } ] ( kw1 ) = >/PP { / NP { / / ? [ Tag = ’GENE ’ ] (
kw0 ) = > / / ? [ Tag = ’DRUG’ ] ( kw2 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e
, kw2 . v a l u e , s e n t . v a l u e
/ / S { / NP { / ? [ Tag = ’DRUG’ ] ( kw2 ) }= >/VP { / ? [ V a l u e IN { ’ s t i m u l a t e s ’ , ’ i n d u c e s ’ } ] ( kw1 )
= >/NP { / / ? [ Tag = ’GENE ’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e
, kw2 . v a l u e , s e n t . v a l u e
/ / S { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/VP { / ? [ V a l u e IN { ’ s t i m u l a t e s ’ , ’ i n d u c e s ’ } ] ( kw1 ) = >/NP
{ / / ? [ Tag = ’GENE ’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2
. value , s e n t . value
B.6
1
2
3
4
5
288
Drug Inhibits Gene
/ / NP { / NP { / NP { / ? [ V a l u e = ’ i n h i b i t o r y ’ ] ( kw1 ) }= >/ PP { / / ? [ Tag = ’DRUG’ ] ( kw2 ) } } = > / PP { /
NP { / / ? [ Tag = ’GENE ’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e ,
kw2 . v a l u e , s e n t . v a l u e
/ / S { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/VP { / NP { / ? [ Tag = ’GENE ’ ] ( kw0 ) = >/?[ V a l u e = ’ i n h i b i t o r ’ ] (
kw1 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t .
value
/ / NP { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/NP { / ? [ Tag = ’GENE ’ ] ( kw0 ) = >/?[ V a l u e = ’ i n h i b i t o r ’ ] ( kw1 )
} } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t . v a l u e
/ / NP { / NP { / ? [ Tag = ’GENE ’ ] ( kw0 ) = >/?[ V a l u e = ’ i n h i b i t o r ’ ] ( kw1 ) } = > / ? [ Tag = ’DRUG’ ] ( kw2
) } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t . v a l u e
/ / NP { / NP { / NP { / ? [ Tag = ’GENE ’ ] ( kw0 ) = >/?[ Tag = ’DRUG’ ] ( kw2 ) } } = > / ? [ V a l u e = ’ i n h i b i t o r
APPENDIX B. DRUG-DRUG INTERACTION QUERIES
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
289
’ ] ( kw1 ) } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t .
value
/ / NP { / ? [ Tag = ’GENE ’ ] ( kw0 ) = >/?[ V a l u e = ’ i n h i b i t o r ’ ] ( kw1 ) = >/?[ Tag = ’DRUG’ ] ( kw2 ) }
: : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t . v a l u e
/ / NP { / NP { / ? [ V a l u e = ’ i n h i b i t o r ’ ] ( kw1 ) }= >/ PP { / NP { / / ? [ Tag = ’GENE ’ ] ( kw0 ) = > / / ? [ Tag = ’
DRUG’ ] ( kw2 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e ,
sent . value
/ / S { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/VP { / VP { / / ? [ Tag = ’GENE ’ ] ( kw0 ) = > / / ? [ V a l u e = ’ i n h i b i t o r
’ ] ( kw1 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t .
value
/ / S { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/VP { / NP { / / ? [ V a l u e = ’ i n h i b i t o r ’ ] ( kw1 ) = > / / ? [ Tag = ’GENE
’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t .
value
/ / NP { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/NP { / NP { / ? [ Tag = ’GENE ’ ] ( kw0 ) } } = > / ? [ V a l u e = ’ i n h i b i t o r
’ ] ( kw1 ) } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t .
value
/ / S { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/?[ V a l u e = ’ i n h i b i t o r ’ ] ( kw1 ) = >/?[ Tag = ’GENE ’ ] ( kw0 ) } : : :
d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t . v a l u e
/ / NP { / NP { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/NP { / ? [ V a l u e = ’ i n h i b i t o r ’ ] ( kw1 ) }= >/ PP { / / ? [ Tag = ’
GENE ’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e ,
sent . value
/ / S { / ? [ Tag = ’GENE ’ ] ( kw0 ) = >/?[ V a l u e = ’ i n h i b i t o r ’ ] ( kw1 ) = >/?[ Tag = ’DRUG’ ] ( kw2 ) } : : :
d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t . v a l u e
/ / NP { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/NP { / NP { / ? [ V a l u e = ’ i n h i b i t o r ’ ] ( kw1 ) }= >/ PP { / / ? [ Tag = ’
GENE ’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e ,
sent . value
/ / S { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/VP { / VP { / / ? [ V a l u e = ’ i n h i b i t o r ’ ] ( kw1 ) = > / / ? [ Tag = ’GENE
’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t .
value
/ / S { / NP { / ? [ Tag = ’DRUG’ ] ( kw2 ) }= >/VP { / NP { / / ? [ V a l u e = ’ i n h i b i t o r ’ ] ( kw1 ) = > / / ? [ Tag = ’
GENE ’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e ,
sent . value
/ / NP { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/NP { / NP { / ? [ Tag = ’GENE ’ ] ( kw0 ) = >/?[ V a l u e = ’ i n h i b i t o r ’ ] (
kw1 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t .
value
/ / NP { / NP { / PP { / / ? [ Tag = ’DRUG’ ] ( kw2 ) = > / / ? [ Tag = ’GENE ’ ] ( kw0 ) }= >/NP { / ? [ V a l u e = ’
i n h i b i t o r ’ ] ( kw1 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 .
value , s e n t . value
/ / NP { / NP { / ? [ Tag = ’GENE ’ ] ( kw0 ) = >/?[ V a l u e = ’ i n h i b i t o r ’ ] ( kw1 ) = >/?[ Tag = ’DRUG’ ] ( kw2 )
} } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t . v a l u e
/ / S { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/?[ Tag = ’GENE ’ ] ( kw0 ) = >/?[ V a l u e = ’ i n h i b i t o r ’ ] ( kw1 ) } : : :
d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t . v a l u e
/ / NP { / NP { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/?[ Tag = ’GENE ’ ] ( kw0 ) = >/?[ V a l u e = ’ i n h i b i t o r ’ ] ( kw1 )
} } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t . v a l u e
/ / NP { / NP { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/NP { / ? [ Tag = ’GENE ’ ] ( kw0 ) = >/?[ V a l u e = ’ i n h i b i t o r ’ ] (
kw1 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t .
value
/ / NP { / NP { / PP { / / ? [ Tag = ’DRUG’ ] ( kw2 ) } = > / ? [ Tag = ’GENE ’ ] ( kw0 ) = >/?[ V a l u e = ’ i n h i b i t o r
’ ] ( kw1 ) } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t .
value
/ / S { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/VP { / ? [ V a l u e = ’ i n h i b i t e d ’ ] ( kw1 ) = >/NP { / / ? [ Tag = ’GENE ’ ] (
APPENDIX B. DRUG-DRUG INTERACTION QUERIES
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
290
kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw2 . v a l u e , kw1 . v a l u e , kw0 . v a l u e , s e n t .
value
/ / S { / NP { / NP { / ? [ Tag = ’DRUG’ ] ( kw2 ) } } = > /VP { / ? [ V a l u e = ’ i n h i b i t e d ’ ] ( kw1 ) = >/NP { / / ? [
Tag = ’GENE ’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw2 . v a l u e , kw1 . v a l u e , kw0 .
value , s e n t . value
/ / S { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/?[ V a l u e = ’ i n h i b i t e d ’ ] ( kw1 ) = >/?[ Tag = ’GENE ’ ] ( kw0 ) } : : :
d i s t i n c t s e n t . c i d , kw2 . v a l u e , kw1 . v a l u e , kw0 . v a l u e , s e n t . v a l u e
/ / S { / NP { / ? [ Tag = ’DRUG’ ] ( kw2 ) }= >/VP { / ? [ V a l u e = ’ i n h i b i t e d ’ ] ( kw1 ) = >/NP { / / ? [ Tag = ’
GENE ’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw2 . v a l u e , kw1 . v a l u e , kw0 . v a l u e ,
sent . value
/ / S { / NP { / NP { / / ? [ Tag = ’DRUG’ ] ( kw2 ) } } = > /VP { / ? [ V a l u e = ’ i n h i b i t e d ’ ] ( kw1 ) = >/NP { / ? [
Tag = ’GENE ’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw2 . v a l u e , kw1 . v a l u e , kw0 .
value , s e n t . value
/ / S { / NP { / PP { / / ? [ Tag = ’GENE ’ ] ( kw0 ) } } = > /VP { / VP { / ? [ V a l u e = ’ i n h i b i t e d ’ ] ( kw1 ) = > / / ? [
Tag = ’DRUG’ ] ( kw2 ) } } } : : : d i s t i n c t s e n t . c i d , kw2 . v a l u e , kw1 . v a l u e , kw0 .
value , s e n t . value
/ / S { / NP { / NP { / / ? [ Tag = ’DRUG’ ] ( kw2 ) } } = > /VP { / ? [ V a l u e = ’ i n h i b i t e d ’ ] ( kw1 ) = >/NP { / / ? [
Tag = ’GENE ’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw2 . v a l u e , kw1 . v a l u e , kw0 .
value , s e n t . value
/ / S { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/VP { / VP { / ? [ V a l u e = ’ i n h i b i t e d ’ ] ( kw1 ) = > / / ? [ Tag = ’GENE ’ ] (
kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw2 . v a l u e , kw1 . v a l u e , kw0 . v a l u e , s e n t .
value
/ / S { / NP { / PP { / / ? [ Tag = ’DRUG’ ] ( kw2 ) } } = > /VP { / ? [ V a l u e = ’ i n h i b i t e d ’ ] ( kw1 ) = >/NP { / ? [
Tag = ’GENE ’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw2 . v a l u e , kw1 . v a l u e , kw0 .
value , s e n t . value
/ / S { / NP { / NP { / ? [ Tag = ’GENE ’ ] ( kw0 ) } } = > /VP { / VP { / ? [ V a l u e = ’ i n h i b i t e d ’ ] ( kw1 ) = > / / ? [
Tag = ’DRUG’ ] ( kw2 ) } } } : : : d i s t i n c t s e n t . c i d , kw2 . v a l u e , kw1 . v a l u e , kw0 .
value , s e n t . value
/ / S { / NP { / NP { / ? [ Tag = ’GENE ’ ] ( kw0 ) } } = > /VP { / ? [ V a l u e = ’ i n h i b i t e d ’ ] ( kw1 ) = >/PP { / / ? [
Tag = ’DRUG’ ] ( kw2 ) } } } : : : d i s t i n c t s e n t . c i d , kw2 . v a l u e , kw1 . v a l u e , kw0 .
value , s e n t . value
/ / S { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/VP { / ? [ V a l u e = ’ i n h i b i t e d ’ ] ( kw1 ) = >/PP { / / ? [ Tag = ’GENE ’ ] (
kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw2 . v a l u e , kw1 . v a l u e , kw0 . v a l u e , s e n t .
value
/ / S { / NP { / ? [ Tag = ’DRUG’ ] ( kw2 ) }= >/VP { / VP { / ? [ V a l u e = ’ i n h i b i t e d ’ ] ( kw1 ) = > / / ? [ Tag = ’
GENE ’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw2 . v a l u e , kw1 . v a l u e , kw0 . v a l u e ,
sent . value
/ / S { / NP { / NP { / ? [ Tag = ’DRUG’ ] ( kw2 ) } } = > /VP { / ? [ V a l u e = ’ i n h i b i t e d ’ ] ( kw1 ) = >/NP { / ? [ Tag
= ’GENE ’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw2 . v a l u e , kw1 . v a l u e , kw0 . v a l u e ,
sent . value
/ / S { / NP { / ? [ Tag = ’GENE ’ ] ( kw0 ) }= >/VP { / VP { / ? [ V a l u e = ’ i n h i b i t e d ’ ] ( kw1 ) = > / / ? [ Tag = ’
DRUG’ ] ( kw2 ) } } } : : : d i s t i n c t s e n t . c i d , kw2 . v a l u e , kw1 . v a l u e , kw0 . v a l u e ,
sent . value
/ / S { / NP { / ? [ Tag = ’GENE ’ ] ( kw0 ) }= >/VP { / ? [ V a l u e = ’ i n h i b i t e d ’ ] ( kw1 ) = >/NP { / ? [ Tag = ’
DRUG’ ] ( kw2 ) } } } : : : d i s t i n c t s e n t . c i d , kw2 . v a l u e , kw1 . v a l u e , kw0 . v a l u e ,
sent . value
/ / S { / NP { / ? [ Tag = ’GENE ’ ] ( kw0 ) }= >/VP { / VP { / / ? [ V a l u e = ’ i n h i b i t e d ’ ] ( kw1 ) = > / / ? [ Tag = ’
DRUG’ ] ( kw2 ) } } } : : : d i s t i n c t s e n t . c i d , kw2 . v a l u e , kw1 . v a l u e , kw0 . v a l u e ,
sent . value
/ / S { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/VP { / ? [ V a l u e = ’ i n h i b i t e d ’ ] ( kw1 ) = >/NP { / ? [ Tag = ’GENE ’ ] (
kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw2 . v a l u e , kw1 . v a l u e , kw0 . v a l u e , s e n t .
APPENDIX B. DRUG-DRUG INTERACTION QUERIES
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
291
value
/ / S { / NP { / NP { / ? [ Tag = ’GENE ’ ] ( kw0 ) = >/?[ Tag = ’DRUG’ ] ( kw2 ) } } = > /VP { / ? [ V a l u e = ’ i n h i b i t
’ ] ( kw1 ) } } : : : d i s t i n c t s e n t . c i d , kw2 . v a l u e , kw1 . v a l u e , kw0 . v a l u e , s e n t .
value
/ / S { / NP { / ? [ Tag = ’DRUG’ ] ( kw2 ) }= >/VP { / VP { / / ? [ V a l u e = ’ i n h i b i t ’ ] ( kw1 ) = > / / ? [ Tag = ’
GENE ’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw2 . v a l u e , kw1 . v a l u e , kw0 . v a l u e ,
sent . value
/ / S { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/VP { / VP { / / ? [ V a l u e = ’ i n h i b i t ’ ] ( kw1 ) = > / / ? [ Tag = ’GENE ’ ] (
kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw2 . v a l u e , kw1 . v a l u e , kw0 . v a l u e , s e n t .
value
/ / NP { / NP { / ? [ Tag = ’GENE ’ ] ( kw0 ) = >/?[ V a l u e = ’ i n h i b i t i o n ’ ] ( kw1 ) }= >/ PP { / NP { / ? [ Tag = ’
DRUG’ ] ( kw2 ) } } } : : : d i s t i n c t s e n t . c i d , kw2 . v a l u e , kw1 . v a l u e , kw0 . v a l u e ,
sent . value
/ / S { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/VP { / PP { / / ? [ V a l u e = ’ i n h i b i t i o n ’ ] ( kw1 ) = > / / ? [ Tag = ’GENE
’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw2 . v a l u e , kw1 . v a l u e , kw0 . v a l u e , s e n t .
value
/ / S { / NP { / PP { / / ? [ Tag = ’DRUG’ ] ( kw2 ) } } = > /VP { / NP { / / ? [ V a l u e = ’ i n h i b i t i o n ’ ] ( kw1 )
= > / / ? [ Tag = ’GENE ’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw2 . v a l u e , kw1 . v a l u e ,
kw0 . v a l u e , s e n t . v a l u e
/ / NP { / ? [ V a l u e = ’ i n h i b i t i o n ’ ] ( kw1 ) = >/PP { / NP { / / ? [ Tag = ’GENE ’ ] ( kw0 ) = > / / ? [ Tag = ’DRUG
’ ] ( kw2 ) } } } : : : d i s t i n c t s e n t . c i d , kw2 . v a l u e , kw1 . v a l u e , kw0 . v a l u e , s e n t .
value
/ / S { / ? [ V a l u e = ’ i n h i b i t i o n ’ ] ( kw1 ) = >/PP { / NP { / / ? [ Tag = ’GENE ’ ] ( kw0 ) = > / / ? [ Tag = ’DRUG
’ ] ( kw2 ) } } } : : : d i s t i n c t s e n t . c i d , kw2 . v a l u e , kw1 . v a l u e , kw0 . v a l u e , s e n t .
value
/ / NP { / NP { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/?[ V a l u e = ’ i n h i b i t i o n ’ ] ( kw1 ) }= >/ PP { / NP { / ? [ Tag = ’
GENE ’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw2 . v a l u e , kw1 . v a l u e , kw0 . v a l u e ,
sent . value
/ / NP { / NP { / ? [ V a l u e = ’ i n h i b i t i o n ’ ] ( kw1 ) }= >/ PP { / NP { / ? [ Tag = ’GENE ’ ] ( kw0 ) = >/?[ Tag = ’
DRUG’ ] ( kw2 ) } } } : : : d i s t i n c t s e n t . c i d , kw2 . v a l u e , kw1 . v a l u e , kw0 . v a l u e ,
sent . value
/ / S { / NP { / ? [ V a l u e = ’ i n h i b i t i o n ’ ] ( kw1 ) = >/PP { / / ? [ Tag = ’GENE ’ ] ( kw0 ) } } = > /VP { / PP { / / ? [
Tag = ’DRUG’ ] ( kw2 ) } } } : : : d i s t i n c t s e n t . c i d , kw2 . v a l u e , kw1 . v a l u e , kw0 .
value , s e n t . value
/ / S { / NP { / NP { / ? [ V a l u e = ’ i n h i b i t i o n ’ ] ( kw1 ) }= >/ PP { / / ? [ Tag = ’GENE ’ ] ( kw0 ) } } = > /VP { / NP
{ / / ? [ Tag = ’DRUG’ ] ( kw2 ) } } } : : : d i s t i n c t s e n t . c i d , kw2 . v a l u e , kw1 . v a l u e , kw0
. value , s e n t . value
/ / S { / ? [ V a l u e IN { ’ i n h i b i t i o n ’ , ’ I n h i b i t i o n ’ } ] ( kw1 ) = >/?[ Tag = ’GENE ’ ] ( kw0 ) = >/?[
Tag = ’DRUG’ ] ( kw2 ) } : : : d i s t i n c t s e n t . c i d , kw2 . v a l u e , kw1 . v a l u e , kw0 . v a l u e ,
sent . value
/ / NP { / NP { / NP { / ? [ V a l u e = ’ i n h i b i t i o n ’ ] ( kw1 ) }= >/ PP { / / ? [ Tag = ’GENE ’ ] ( kw0 ) } } = > / PP { /
NP { / / ? [ Tag = ’DRUG’ ] ( kw2 ) } } } : : : d i s t i n c t s e n t . c i d , kw2 . v a l u e , kw1 . v a l u e ,
kw0 . v a l u e , s e n t . v a l u e
/ / NP { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/NP { / ? [ V a l u e = ’ i n h i b i t i o n ’ ] ( kw1 ) }= >/ PP { / NP { / ? [ Tag = ’
GENE ’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw2 . v a l u e , kw1 . v a l u e , kw0 . v a l u e ,
sent . value
/ / NP { / NP { / ? [ V a l u e = ’ i n h i b i t i o n ’ ] ( kw1 ) }= >/ PP { / NP { / / ? [ Tag = ’GENE ’ ] ( kw0 ) = > / / ? [ Tag
= ’DRUG’ ] ( kw2 ) } } } : : : d i s t i n c t s e n t . c i d , kw2 . v a l u e , kw1 . v a l u e , kw0 . v a l u e ,
sent . value
/ / S { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/VP { / NP { / / ? [ V a l u e = ’ i n h i b i t i o n ’ ] ( kw1 ) = > / / ? [ Tag = ’GENE
’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw2 . v a l u e , kw1 . v a l u e , kw0 . v a l u e , s e n t .
APPENDIX B. DRUG-DRUG INTERACTION QUERIES
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
292
value
/ / S { / NP { / NP { / ? [ V a l u e = ’ i n h i b i t i o n ’ ] ( kw1 ) }= >/ PP { / / ? [ Tag = ’DRUG’ ] ( kw2 ) } } = > /VP { / NP
{ / / ? [ Tag = ’GENE ’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw2 . v a l u e , kw1 . v a l u e , kw0
. value , s e n t . value
/ / NP { / NP { / ? [ V a l u e = ’ i n h i b i t i o n ’ ] ( kw1 ) = >/PP { / / ? [ Tag = ’GENE ’ ] ( kw0 ) } } = > / PP { / NP
{ / / ? [ Tag = ’DRUG’ ] ( kw2 ) } } } : : : d i s t i n c t s e n t . c i d , kw2 . v a l u e , kw1 . v a l u e , kw0
. value , s e n t . value
/ / NP { / NP { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/?[ V a l u e = ’ i n h i b i t i o n ’ ] ( kw1 ) }= >/ PP { / NP { / / ? [ Tag = ’
GENE ’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw2 . v a l u e , kw1 . v a l u e , kw0 . v a l u e ,
sent . value
/ / NP { / NP { / ? [ V a l u e = ’ i n h i b i t i o n ’ ] ( kw1 ) = >/PP { / / ? [ Tag = ’GENE ’ ] ( kw0 ) } } = > / PP { / NP { / ? [
Tag = ’DRUG’ ] ( kw2 ) } } } : : : d i s t i n c t s e n t . c i d , kw2 . v a l u e , kw1 . v a l u e , kw0 .
value , s e n t . value
/ / NP { / NP { / NP { / ? [ V a l u e = ’ i n h i b i t i o n ’ ] ( kw1 ) }= >/ PP { / / ? [ Tag = ’GENE ’ ] ( kw0 ) } } = > / PP { /
NP { / ? [ Tag = ’DRUG’ ] ( kw2 ) } } } : : : d i s t i n c t s e n t . c i d , kw2 . v a l u e , kw1 . v a l u e ,
kw0 . v a l u e , s e n t . v a l u e
/ / S { / NP { / ? [ Tag = ’DRUG’ ] ( kw2 ) }= >/VP { / ? [ V a l u e = ’ i n h i b i t s ’ ] ( kw1 ) = >/NP { / / ? [ Tag = ’
GENE ’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw2 . v a l u e , kw1 . v a l u e , kw0 . v a l u e ,
sent . value
/ S { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/VP { / ? [ V a l u e = ’ i n h i b i t s ’ ] ( kw1 ) = >/NP { / ? [ Tag = ’GENE ’ ] ( kw0
) } } } : : : d i s t i n c t s e n t . c i d , kw2 . v a l u e , kw1 . v a l u e , kw0 . v a l u e , s e n t . v a l u e
/ / S { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/VP { / VP { / ? [ V a l u e = ’ i n h i b i t s ’ ] ( kw1 ) = > / / ? [ Tag = ’GENE ’ ] (
kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw2 . v a l u e , kw1 . v a l u e , kw0 . v a l u e , s e n t .
value
/ / S { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/VP { / ? [ V a l u e = ’ i n h i b i t s ’ ] ( kw1 ) = >/NP { / / ? [ Tag = ’GENE ’ ] (
kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw2 . v a l u e , kw1 . v a l u e , kw0 . v a l u e , s e n t .
value
/ / NP { / NP { / NP { / ? [ V a l u e = ’ i n h i b i t o r y ’ ] ( kw1 ) }= >/ PP { / / ? [ Tag = ’DRUG’ ] ( kw2 ) } } = > / PP { /
NP { / / ? [ Tag = ’GENE ’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e ,
kw2 . v a l u e , s e n t . v a l u e
/ / S { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/VP { / NP { / ? [ Tag = ’GENE ’ ] ( kw0 ) = >/?[ V a l u e = ’ i n h i b i t o r ’ ] (
kw1 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t .
value
/ / NP { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/NP { / ? [ Tag = ’GENE ’ ] ( kw0 ) = >/?[ V a l u e = ’ i n h i b i t o r ’ ] ( kw1 )
} } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t . v a l u e
/ / NP { / NP { / ? [ Tag = ’GENE ’ ] ( kw0 ) = >/?[ V a l u e = ’ i n h i b i t o r ’ ] ( kw1 ) } = > / ? [ Tag = ’DRUG’ ] ( kw2
) } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t . v a l u e
/ / NP { / NP { / NP { / ? [ Tag = ’GENE ’ ] ( kw0 ) = >/?[ Tag = ’DRUG’ ] ( kw2 ) } } = > / ? [ V a l u e = ’ i n h i b i t o r
’ ] ( kw1 ) } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t .
value
/ / NP { / ? [ Tag = ’GENE ’ ] ( kw0 ) = >/?[ V a l u e = ’ i n h i b i t o r ’ ] ( kw1 ) = >/?[ Tag = ’DRUG’ ] ( kw2 ) }
: : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t . v a l u e
/ / NP { / NP { / ? [ V a l u e = ’ i n h i b i t o r ’ ] ( kw1 ) }= >/ PP { / NP { / / ? [ Tag = ’GENE ’ ] ( kw0 ) = > / / ? [ Tag = ’
DRUG’ ] ( kw2 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e ,
sent . value
/ / S { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/VP { / VP { / / ? [ Tag = ’GENE ’ ] ( kw0 ) = > / / ? [ V a l u e = ’ i n h i b i t o r
’ ] ( kw1 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t .
value
/ / S { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/VP { / NP { / / ? [ V a l u e = ’ i n h i b i t o r ’ ] ( kw1 ) = > / / ? [ Tag = ’GENE
’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t .
value
APPENDIX B. DRUG-DRUG INTERACTION QUERIES
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
293
/ / NP { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/NP { / NP { / ? [ Tag = ’GENE ’ ] ( kw0 ) } } = > / ? [ V a l u e = ’ i n h i b i t o r
’ ] ( kw1 ) } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t .
value
/ / S { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/?[ V a l u e = ’ i n h i b i t o r ’ ] ( kw1 ) = >/?[ Tag = ’GENE ’ ] ( kw0 ) } : : :
d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t . v a l u e
/ / NP { / NP { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/NP { / ? [ V a l u e = ’ i n h i b i t o r ’ ] ( kw1 ) }= >/ PP { / / ? [ Tag = ’
GENE ’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e ,
sent . value
/ / S { / ? [ Tag = ’GENE ’ ] ( kw0 ) = >/?[ V a l u e = ’ i n h i b i t o r ’ ] ( kw1 ) = >/?[ Tag = ’DRUG’ ] ( kw2 ) } : : :
d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t . v a l u e
/ / NP { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/NP { / NP { / ? [ V a l u e = ’ i n h i b i t o r ’ ] ( kw1 ) }= >/ PP { / / ? [ Tag = ’
GENE ’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e ,
sent . value
/ / S { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/VP { / VP { / / ? [ V a l u e = ’ i n h i b i t o r ’ ] ( kw1 ) = > / / ? [ Tag = ’GENE
’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t .
value
/ / S { / NP { / ? [ Tag = ’DRUG’ ] ( kw2 ) }= >/VP { / NP { / / ? [ V a l u e = ’ i n h i b i t o r ’ ] ( kw1 ) = > / / ? [ Tag = ’
GENE ’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e ,
sent . value
/ / NP { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/NP { / NP { / ? [ Tag = ’GENE ’ ] ( kw0 ) = >/?[ V a l u e = ’ i n h i b i t o r ’ ] (
kw1 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t .
value
/ / NP { / NP { / PP { / / ? [ Tag = ’DRUG’ ] ( kw2 ) = > / / ? [ Tag = ’GENE ’ ] ( kw0 ) }= >/NP { / ? [ V a l u e = ’
i n h i b i t o r ’ ] ( kw1 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 .
value , s e n t . value
/ / NP { / NP { / ? [ Tag = ’GENE ’ ] ( kw0 ) = >/?[ V a l u e = ’ i n h i b i t o r ’ ] ( kw1 ) = >/?[ Tag = ’DRUG’ ] ( kw2 )
} } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t . v a l u e
/ / S { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/?[ Tag = ’GENE ’ ] ( kw0 ) = >/?[ V a l u e = ’ i n h i b i t o r ’ ] ( kw1 ) } : : :
d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t . v a l u e
/ / NP { / NP { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/?[ Tag = ’GENE ’ ] ( kw0 ) = >/?[ V a l u e = ’ i n h i b i t o r ’ ] ( kw1 )
} } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t . v a l u e
/ / NP { / NP { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/NP { / ? [ Tag = ’GENE ’ ] ( kw0 ) = >/?[ V a l u e = ’ i n h i b i t o r ’ ] (
kw1 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t .
value
/ / NP { / NP { / PP { / / ? [ Tag = ’DRUG’ ] ( kw2 ) } = > / ? [ Tag = ’GENE ’ ] ( kw0 ) = >/?[ V a l u e = ’ i n h i b i t o r
’ ] ( kw1 ) } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t .
value
/ / S { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/VP { / ? [ V a l u e = ’ i n h i b i t e d ’ ] ( kw1 ) = >/NP { / / ? [ Tag = ’GENE ’ ] (
kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t .
value
/ / S { / NP { / NP { / ? [ Tag = ’DRUG’ ] ( kw2 ) } } = > /VP { / ? [ V a l u e = ’ i n h i b i t e d ’ ] ( kw1 ) = >/NP { / / ? [
Tag = ’GENE ’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 .
value , s e n t . value
/ / S { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/?[ V a l u e = ’ i n h i b i t e d ’ ] ( kw1 ) = >/?[ Tag = ’GENE ’ ] ( kw0 ) } : : :
d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t . v a l u e
/ / S { / NP { / ? [ Tag = ’DRUG’ ] ( kw2 ) }= >/VP { / ? [ V a l u e = ’ i n h i b i t e d ’ ] ( kw1 ) = >/NP { / / ? [ Tag = ’
GENE ’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e ,
sent . value
/ / S { / NP { / NP { / / ? [ Tag = ’DRUG’ ] ( kw2 ) } } = > /VP { / ? [ V a l u e = ’ i n h i b i t e d ’ ] ( kw1 ) = >/NP { / ? [
Tag = ’GENE ’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 .
value , s e n t . value
APPENDIX B. DRUG-DRUG INTERACTION QUERIES
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
294
/ / S { / NP { / PP { / / ? [ Tag = ’GENE ’ ] ( kw0 ) } } = > /VP { / VP { / ? [ V a l u e = ’ i n h i b i t e d ’ ] ( kw1 ) = > / / ? [
Tag = ’DRUG’ ] ( kw2 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 .
value , s e n t . value
/ / S { / NP { / NP { / / ? [ Tag = ’DRUG’ ] ( kw2 ) } } = > /VP { / ? [ V a l u e = ’ i n h i b i t e d ’ ] ( kw1 ) = >/NP { / / ? [
Tag = ’GENE ’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 .
value , s e n t . value
/ / S { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/VP { / VP { / ? [ V a l u e = ’ i n h i b i t e d ’ ] ( kw1 ) = > / / ? [ Tag = ’GENE ’ ] (
kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t .
value
/ / S { / NP { / PP { / / ? [ Tag = ’DRUG’ ] ( kw2 ) } } = > /VP { / ? [ V a l u e = ’ i n h i b i t e d ’ ] ( kw1 ) = >/NP { / ? [
Tag = ’GENE ’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 .
value , s e n t . value
/ / S { / NP { / NP { / ? [ Tag = ’GENE ’ ] ( kw0 ) } } = > /VP { / VP { / ? [ V a l u e = ’ i n h i b i t e d ’ ] ( kw1 ) = > / / ? [
Tag = ’DRUG’ ] ( kw2 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 .
value , s e n t . value
/ / S { / NP { / NP { / ? [ Tag = ’GENE ’ ] ( kw0 ) } } = > /VP { / ? [ V a l u e = ’ i n h i b i t e d ’ ] ( kw1 ) = >/PP { / / ? [
Tag = ’DRUG’ ] ( kw2 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 .
value , s e n t . value
/ / S { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/VP { / ? [ V a l u e = ’ i n h i b i t e d ’ ] ( kw1 ) = >/PP { / / ? [ Tag = ’GENE ’ ] (
kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t .
value
/ / S { / NP { / ? [ Tag = ’DRUG’ ] ( kw2 ) }= >/VP { / VP { / ? [ V a l u e = ’ i n h i b i t e d ’ ] ( kw1 ) = > / / ? [ Tag = ’
GENE ’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e ,
sent . value
/ / S { / NP { / NP { / ? [ Tag = ’DRUG’ ] ( kw2 ) } } = > /VP { / ? [ V a l u e = ’ i n h i b i t e d ’ ] ( kw1 ) = >/NP { / ? [ Tag
= ’GENE ’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e ,
sent . value
/ / S { / NP { / ? [ Tag = ’GENE ’ ] ( kw0 ) }= >/VP { / VP { / ? [ V a l u e = ’ i n h i b i t e d ’ ] ( kw1 ) = > / / ? [ Tag = ’
DRUG’ ] ( kw2 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e ,
sent . value
/ / S { / NP { / ? [ Tag = ’GENE ’ ] ( kw0 ) }= >/VP { / ? [ V a l u e = ’ i n h i b i t e d ’ ] ( kw1 ) = >/NP { / ? [ Tag = ’
DRUG’ ] ( kw2 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e ,
sent . value
/ / S { / NP { / ? [ Tag = ’GENE ’ ] ( kw0 ) }= >/VP { / VP { / / ? [ V a l u e = ’ i n h i b i t e d ’ ] ( kw1 ) = > / / ? [ Tag = ’
DRUG’ ] ( kw2 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e ,
sent . value
/ / S { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/VP { / ? [ V a l u e = ’ i n h i b i t e d ’ ] ( kw1 ) = >/NP { / ? [ Tag = ’GENE ’ ] (
kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t .
value
/ / S { / NP { / NP { / ? [ Tag = ’GENE ’ ] ( kw0 ) = >/?[ Tag = ’DRUG’ ] ( kw2 ) } } = > /VP { / ? [ V a l u e = ’ i n h i b i t
’ ] ( kw1 ) } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t .
value
/ / S { / NP { / ? [ Tag = ’DRUG’ ] ( kw2 ) }= >/VP { / VP { / / ? [ V a l u e = ’ i n h i b i t ’ ] ( kw1 ) = > / / ? [ Tag = ’
GENE ’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e ,
sent . value
/ / S { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/VP { / VP { / / ? [ V a l u e = ’ i n h i b i t ’ ] ( kw1 ) = > / / ? [ Tag = ’GENE ’ ] (
kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t .
value
/ / NP { / NP { / ? [ Tag = ’GENE ’ ] ( kw0 ) = >/?[ V a l u e = ’ i n h i b i t i o n ’ ] ( kw1 ) }= >/ PP { / NP { / ? [ Tag = ’
DRUG’ ] ( kw2 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e ,
sent . value
APPENDIX B. DRUG-DRUG INTERACTION QUERIES
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
295
/ / S { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/VP { / PP { / / ? [ V a l u e = ’ i n h i b i t i o n ’ ] ( kw1 ) = > / / ? [ Tag = ’GENE
’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t .
value
/ / S { / NP { / PP { / / ? [ Tag = ’DRUG’ ] ( kw2 ) } } = > /VP { / NP { / / ? [ V a l u e = ’ i n h i b i t i o n ’ ] ( kw1 )
= > / / ? [ Tag = ’GENE ’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e ,
kw2 . v a l u e , s e n t . v a l u e
/ / NP { / ? [ V a l u e = ’ i n h i b i t i o n ’ ] ( kw1 ) = >/PP { / NP { / / ? [ Tag = ’GENE ’ ] ( kw0 ) = > / / ? [ Tag = ’DRUG
’ ] ( kw2 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t .
value
/ / S { / ? [ V a l u e = ’ i n h i b i t i o n ’ ] ( kw1 ) = >/PP { / NP { / / ? [ Tag = ’GENE ’ ] ( kw0 ) = > / / ? [ Tag = ’DRUG
’ ] ( kw2 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t .
value
/ / NP { / NP { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/?[ V a l u e = ’ i n h i b i t i o n ’ ] ( kw1 ) }= >/ PP { / NP { / ? [ Tag = ’
GENE ’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e ,
sent . value
/ / NP { / NP { / ? [ V a l u e = ’ i n h i b i t i o n ’ ] ( kw1 ) }= >/ PP { / NP { / ? [ Tag = ’GENE ’ ] ( kw0 ) = >/?[ Tag = ’
DRUG’ ] ( kw2 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e ,
sent . value
/ / S { / NP { / ? [ V a l u e = ’ i n h i b i t i o n ’ ] ( kw1 ) = >/PP { / / ? [ Tag = ’GENE ’ ] ( kw0 ) } } = > /VP { / PP { / / ? [
Tag = ’DRUG’ ] ( kw2 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 .
value , s e n t . value
/ / S { / NP { / NP { / ? [ V a l u e = ’ i n h i b i t i o n ’ ] ( kw1 ) }= >/ PP { / / ? [ Tag = ’GENE ’ ] ( kw0 ) } } = > /VP { / NP
{ / / ? [ Tag = ’DRUG’ ] ( kw2 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2
. value , s e n t . value
/ / S { / ? [ V a l u e IN { ’ i n h i b i t i o n ’ , ’ I n h i b i t i o n ’ } ] ( kw1 ) = >/?[ Tag = ’GENE ’ ] ( kw0 ) = >/?[
Tag = ’DRUG’ ] ( kw2 ) } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e ,
sent . value
/ / NP { / NP { / NP { / ? [ V a l u e = ’ i n h i b i t i o n ’ ] ( kw1 ) }= >/ PP { / / ? [ Tag = ’GENE ’ ] ( kw0 ) } } = > / PP { /
NP { / / ? [ Tag = ’DRUG’ ] ( kw2 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e ,
kw2 . v a l u e , s e n t . v a l u e
/ / NP { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/NP { / ? [ V a l u e = ’ i n h i b i t i o n ’ ] ( kw1 ) }= >/ PP { / NP { / ? [ Tag = ’
GENE ’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e ,
sent . value
/ / NP { / NP { / ? [ V a l u e = ’ i n h i b i t i o n ’ ] ( kw1 ) }= >/ PP { / NP { / / ? [ Tag = ’GENE ’ ] ( kw0 ) = > / / ? [ Tag
= ’DRUG’ ] ( kw2 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e ,
sent . value
/ / S { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/VP { / NP { / / ? [ V a l u e = ’ i n h i b i t i o n ’ ] ( kw1 ) = > / / ? [ Tag = ’GENE
’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t .
value
/ / S { / NP { / NP { / ? [ V a l u e = ’ i n h i b i t i o n ’ ] ( kw1 ) }= >/ PP { / / ? [ Tag = ’DRUG’ ] ( kw2 ) } } = > /VP { / NP
{ / / ? [ Tag = ’GENE ’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2
. value , s e n t . value
/ / NP { / NP { / ? [ V a l u e = ’ i n h i b i t i o n ’ ] ( kw1 ) = >/PP { / / ? [ Tag = ’GENE ’ ] ( kw0 ) } } = > / PP { / NP
{ / / ? [ Tag = ’DRUG’ ] ( kw2 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2
. value , s e n t . value
/ / NP { / NP { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/?[ V a l u e = ’ i n h i b i t i o n ’ ] ( kw1 ) }= >/ PP { / NP { / / ? [ Tag = ’
GENE ’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e ,
sent . value
/ / NP { / NP { / ? [ V a l u e = ’ i n h i b i t i o n ’ ] ( kw1 ) = >/PP { / / ? [ Tag = ’GENE ’ ] ( kw0 ) } } = > / PP { / NP { / ? [
Tag = ’DRUG’ ] ( kw2 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 .
value , s e n t . value
APPENDIX B. DRUG-DRUG INTERACTION QUERIES
130
131
132
133
134
/ / NP { / NP { / NP { / ? [ V a l u e = ’ i n h i b i t i o n ’ ] ( kw1 ) }= >/ PP { / / ? [ Tag = ’GENE ’ ] ( kw0 ) } } = > / PP { /
NP { / ? [ Tag = ’DRUG’ ] ( kw2 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e ,
kw2 . v a l u e , s e n t . v a l u e
/ / S { / NP { / ? [ Tag = ’DRUG’ ] ( kw2 ) }= >/VP { / ? [ V a l u e = ’ i n h i b i t s ’ ] ( kw1 ) = >/NP { / / ? [ Tag = ’
GENE ’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e ,
sent . value
/ S { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/VP { / ? [ V a l u e = ’ i n h i b i t s ’ ] ( kw1 ) = >/NP { / ? [ Tag = ’GENE ’ ] ( kw0
) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t . v a l u e
/ / S { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/VP { / VP { / ? [ V a l u e = ’ i n h i b i t s ’ ] ( kw1 ) = > / / ? [ Tag = ’GENE ’ ] (
kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t .
value
/ / S { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/VP { / ? [ V a l u e = ’ i n h i b i t s ’ ] ( kw1 ) = >/NP { / / ? [ Tag = ’GENE ’ ] (
kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t .
value
B.7
1
2
3
4
5
6
7
8
9
10
11
12
296
Gene Metabolized Drug
/ / S { / NP { / NP { / ? [ Tag = ’DRUG’ ] ( kw2 ) } } = > /VP { / VP { / ? [ V a l u e IN { ’ m e t a b o l i s e d ’ , ’
m e t a b o l i z e d ’ } ] ( kw1 ) = > / / ? [ Tag = ’GENE ’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 .
v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t . v a l u e
/ / S { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/VP { / VP { / / ? [ V a l u e IN { ’ m e t a b o l i s e d ’ , ’ m e t a b o l i z e d ’ } ] (
kw1 ) = > / / ? [ Tag = ’GENE ’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e
, kw2 . v a l u e , s e n t . v a l u e
/ / S { / NP { / SBAR { / / ? [ Tag = ’DRUG’ ] ( kw2 ) } } = > /VP { / ? [ V a l u e IN { ’ m e t a b o l i s e d ’ , ’
m e t a b o l i z e d ’ } ] ( kw1 ) = >/PP { / / ? [ Tag = ’GENE ’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d ,
kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t . v a l u e
/ / S { / NP { / ? [ Tag = ’DRUG’ ] ( kw2 ) }= >/VP { / VP { / ? [ V a l u e IN { ’ m e t a b o l i s e d ’ , ’ m e t a b o l i z e d
’ } ] ( kw1 ) = > / / ? [ Tag = ’GENE ’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 .
v a l u e , kw2 . v a l u e , s e n t . v a l u e
/ / S { / NP { / ? [ Tag = ’DRUG’ ] ( kw2 ) }= >/VP { / VP { / / ? [ V a l u e IN { ’ m e t a b o l i z e d ’ , ’
m e t a b o l i s e d ’ } ] ( kw1 ) = > / / ? [ Tag = ’GENE ’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 .
v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t . v a l u e
/ / S { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/VP { / VP { / ? [ V a l u e IN { ’ m e t a b o l i s e d ’ , ’ m e t a b o l i z e d ’ } ] (
kw1 ) = > / / ? [ Tag = ’GENE ’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e
, kw2 . v a l u e , s e n t . v a l u e
/ / S { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/VP { / ? [ V a l u e = ’ m e t a b o l i z e d ’ ] ( kw1 ) = >/PP { / / ? [ Tag = ’GENE
’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t .
value
/ / S { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/NP { / ? [ V a l u e = ’ m e t a b o l i s m ’ ] ( kw1 ) }= >/VP { / VP { / / ? [ Tag = ’
GENE ’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e ,
sent . value
/ / S { / NP { / ? [ Tag = ’GENE ’ ] ( kw0 ) }= >/VP { / NP { / / ? [ Tag = ’DRUG’ ] ( kw2 ) = > / / ? [ V a l u e = ’
m e t a b o l i s m ’ ] ( kw1 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 .
value , s e n t . value
/ / NP { / NP { / ? [ Tag = ’GENE ’ ] ( kw0 ) = >/?[ V a l u e = ’ m e t a b o l i s m ’ ] ( kw1 ) }= >/ PP { / NP { / ? [ Tag = ’
DRUG’ ] ( kw2 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e ,
sent . value
/ / NP { / ? [ V a l u e = ’ m e t a b o l i s m ’ ] ( kw1 ) = >/PP { / NP { / / ? [ Tag = ’DRUG’ ] ( kw2 ) = > / / ? [ Tag = ’GENE
’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t .
value
/ / S { / NP { / ? [ Tag = ’GENE ’ ] ( kw0 ) }= >/VP { / VP { / / ? [ V a l u e = ’ m e t a b o l i s m ’ ] ( kw1 ) = > / / ? [ Tag = ’
APPENDIX B. DRUG-DRUG INTERACTION QUERIES
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
297
DRUG’ ] ( kw2 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e ,
sent . value
/ / S { / NP { / NP { / ? [ Tag = ’GENE ’ ] ( kw0 ) } } = > /VP { / NP { / / ? [ V a l u e = ’ m e t a b o l i s m ’ ] ( kw1 ) = > / / ? [
Tag = ’DRUG’ ] ( kw2 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 .
value , s e n t . value
/ / NP { / NP { / PP { / / ? [ Tag = ’GENE ’ ] ( kw0 ) } } = > / PP { / NP { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/?[ V a l u e = ’
m e t a b o l i s m ’ ] ( kw1 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 .
value , s e n t . value
/ / S { / NP { / ? [ Tag = ’GENE ’ ] ( kw0 ) }= >/VP { / PP { / / ? [ V a l u e = ’ m e t a b o l i s m ’ ] ( kw1 ) = > / / ? [ Tag = ’
DRUG’ ] ( kw2 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e ,
sent . value
/ / NP { / NP { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/?[ V a l u e = ’ m e t a b o l i s m ’ ] ( kw1 ) }= >/ PP { / NP { / / ? [ Tag = ’
GENE ’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e ,
sent . value
/ / NP { / NP { / ? [ Tag = ’GENE ’ ] ( kw0 ) = >/?[ V a l u e = ’ m e t a b o l i s m ’ ] ( kw1 ) }= >/ PP { / NP { / / ? [ Tag = ’
DRUG’ ] ( kw2 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e ,
sent . value
/ / S { / NP { / NP { / ? [ V a l u e = ’ m e t a b o l i s m ’ ] ( kw1 ) }= >/ PP { / / ? [ Tag = ’DRUG’ ] ( kw2 ) } } = > /VP { / VP
{ / / ? [ Tag = ’GENE ’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2
. value , s e n t . value
/ / S { / NP { / ? [ Tag = ’GENE ’ ] ( kw0 ) }= >/VP { / PP { / / ? [ Tag = ’DRUG’ ] ( kw2 ) = > / / ? [ V a l u e = ’
m e t a b o l i s m ’ ] ( kw1 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 .
value , s e n t . value
/ / NP { / NP { / PP { / / ? [ Tag = ’GENE ’ ] ( kw0 ) } } = > / PP { / NP { / / ? [ V a l u e = ’ m e t a b o l i s m ’ ] ( kw1 )
= > / / ? [ Tag = ’DRUG’ ] ( kw2 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e ,
kw2 . v a l u e , s e n t . v a l u e
/ / S { / NP { / ? [ Tag = ’DRUG’ ] ( kw2 ) }= >/VP { / NP { / / ? [ V a l u e = ’ s u b s t r a t e ’ ] ( kw1 ) = > / / ? [ Tag = ’
GENE ’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e ,
sent . value
/ / NP { / NP { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/NP { / ? [ V a l u e = ’ s u b s t r a t e ’ ] ( kw1 ) }= >/ PP { / / ? [ Tag = ’
GENE ’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e ,
sent . value
/ / S { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/VP { / VP { / / ? [ V a l u e = ’ s u b s t r a t e ’ ] ( kw1 ) = > / / ? [ Tag = ’GENE
’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t .
value
/ / NP { / ? [ Tag = ’GENE ’ ] ( kw0 ) = >/?[ V a l u e = ’ s u b s t r a t e ’ ] ( kw1 ) = >/?[ Tag = ’DRUG’ ] ( kw2 ) }
: : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t . v a l u e
/ / NP { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/NP { / NP { / ? [ V a l u e = ’ s u b s t r a t e ’ ] ( kw1 ) }= >/ PP { / / ? [ Tag = ’
GENE ’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e ,
sent . value
/ / S { / NP { / NP { / ? [ Tag = ’DRUG’ ] ( kw2 ) } } = > /VP { / VP { / ? [ V a l u e IN { ’ m e t a b o l i s e d ’ , ’
m e t a b o l i z e d ’ } ] ( kw1 ) = > / / ? [ Tag = ’GENE ’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 .
v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t . v a l u e
/ / S { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/VP { / VP { / / ? [ V a l u e IN { ’ m e t a b o l i s e d ’ , ’ m e t a b o l i z e d ’ } ] (
kw1 ) = > / / ? [ Tag = ’GENE ’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e
, kw2 . v a l u e , s e n t . v a l u e
/ / S { / NP { / SBAR { / / ? [ Tag = ’DRUG’ ] ( kw2 ) } } = > /VP { / ? [ V a l u e IN { ’ m e t a b o l i s e d ’ , ’
m e t a b o l i z e d ’ } ] ( kw1 ) = >/PP { / / ? [ Tag = ’GENE ’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d ,
kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t . v a l u e
/ / S { / NP { / ? [ Tag = ’DRUG’ ] ( kw2 ) }= >/VP { / VP { / ? [ V a l u e IN { ’ m e t a b o l i s e d ’ , ’ m e t a b o l i z e d
’ } ] ( kw1 ) = > / / ? [ Tag = ’GENE ’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 .
APPENDIX B. DRUG-DRUG INTERACTION QUERIES
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
298
v a l u e , kw2 . v a l u e , s e n t . v a l u e
/ / S { / NP { / ? [ Tag = ’DRUG’ ] ( kw2 ) }= >/VP { / VP { / / ? [ V a l u e IN { ’ m e t a b o l i z e d ’ , ’
m e t a b o l i s e d ’ } ] ( kw1 ) = > / / ? [ Tag = ’GENE ’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 .
v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t . v a l u e
/ / S { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/VP { / VP { / ? [ V a l u e IN { ’ m e t a b o l i s e d ’ , ’ m e t a b o l i z e d ’ } ] (
kw1 ) = > / / ? [ Tag = ’GENE ’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e
, kw2 . v a l u e , s e n t . v a l u e
/ / S { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/VP { / ? [ V a l u e = ’ m e t a b o l i z e d ’ ] ( kw1 ) = >/PP { / / ? [ Tag = ’GENE
’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t .
value
/ / S { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/NP { / ? [ V a l u e = ’ m e t a b o l i s m ’ ] ( kw1 ) }= >/VP { / VP { / / ? [ Tag = ’
GENE ’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e ,
sent . value
/ / S { / NP { / ? [ Tag = ’GENE ’ ] ( kw0 ) }= >/VP { / NP { / / ? [ Tag = ’DRUG’ ] ( kw2 ) = > / / ? [ V a l u e = ’
m e t a b o l i s m ’ ] ( kw1 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 .
value , s e n t . value
/ / NP { / NP { / ? [ Tag = ’GENE ’ ] ( kw0 ) = >/?[ V a l u e = ’ m e t a b o l i s m ’ ] ( kw1 ) }= >/ PP { / NP { / ? [ Tag = ’
DRUG’ ] ( kw2 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e ,
sent . value
/ / NP { / ? [ V a l u e = ’ m e t a b o l i s m ’ ] ( kw1 ) = >/PP { / NP { / / ? [ Tag = ’DRUG’ ] ( kw2 ) = > / / ? [ Tag = ’GENE
’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t .
value
/ / S { / NP { / ? [ Tag = ’GENE ’ ] ( kw0 ) }= >/VP { / VP { / / ? [ V a l u e = ’ m e t a b o l i s m ’ ] ( kw1 ) = > / / ? [ Tag = ’
DRUG’ ] ( kw2 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e ,
sent . value
/ / S { / NP { / NP { / ? [ Tag = ’GENE ’ ] ( kw0 ) } } = > /VP { / NP { / / ? [ V a l u e = ’ m e t a b o l i s m ’ ] ( kw1 ) = > / / ? [
Tag = ’DRUG’ ] ( kw2 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 .
value , s e n t . value
/ / NP { / NP { / PP { / / ? [ Tag = ’GENE ’ ] ( kw0 ) } } = > / PP { / NP { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/?[ V a l u e = ’
m e t a b o l i s m ’ ] ( kw1 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 .
value , s e n t . value
/ / S { / NP { / ? [ Tag = ’GENE ’ ] ( kw0 ) }= >/VP { / PP { / / ? [ V a l u e = ’ m e t a b o l i s m ’ ] ( kw1 ) = > / / ? [ Tag = ’
DRUG’ ] ( kw2 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e ,
sent . value
/ / NP { / NP { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/?[ V a l u e = ’ m e t a b o l i s m ’ ] ( kw1 ) }= >/ PP { / NP { / / ? [ Tag = ’
GENE ’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e ,
sent . value
/ / NP { / NP { / ? [ Tag = ’GENE ’ ] ( kw0 ) = >/?[ V a l u e = ’ m e t a b o l i s m ’ ] ( kw1 ) }= >/ PP { / NP { / / ? [ Tag = ’
DRUG’ ] ( kw2 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e ,
sent . value
/ / S { / NP { / NP { / ? [ V a l u e = ’ m e t a b o l i s m ’ ] ( kw1 ) }= >/ PP { / / ? [ Tag = ’DRUG’ ] ( kw2 ) } } = > /VP { / VP
{ / / ? [ Tag = ’GENE ’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2
. value , s e n t . value
/ / S { / NP { / ? [ Tag = ’GENE ’ ] ( kw0 ) }= >/VP { / PP { / / ? [ Tag = ’DRUG’ ] ( kw2 ) = > / / ? [ V a l u e = ’
m e t a b o l i s m ’ ] ( kw1 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 .
value , s e n t . value
/ / NP { / NP { / PP { / / ? [ Tag = ’GENE ’ ] ( kw0 ) } } = > / PP { / NP { / / ? [ V a l u e = ’ m e t a b o l i s m ’ ] ( kw1 )
= > / / ? [ Tag = ’DRUG’ ] ( kw2 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e ,
kw2 . v a l u e , s e n t . v a l u e
/ / S { / NP { / ? [ Tag = ’DRUG’ ] ( kw2 ) }= >/VP { / NP { / / ? [ V a l u e = ’ s u b s t r a t e ’ ] ( kw1 ) = > / / ? [ Tag = ’
GENE ’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e ,
APPENDIX B. DRUG-DRUG INTERACTION QUERIES
47
48
49
50
sent . value
/ / NP { / NP { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/NP { / ? [ V a l u e = ’ s u b s t r a t e ’ ] ( kw1 ) }= >/ PP { / / ? [ Tag = ’
GENE ’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e ,
sent . value
/ / S { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/VP { / VP { / / ? [ V a l u e = ’ s u b s t r a t e ’ ] ( kw1 ) = > / / ? [ Tag = ’GENE
’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t .
value
/ / NP { / ? [ Tag = ’GENE ’ ] ( kw0 ) = >/?[ V a l u e = ’ s u b s t r a t e ’ ] ( kw1 ) = >/?[ Tag = ’DRUG’ ] ( kw2 ) }
: : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , s e n t . v a l u e
/ / NP { / ? [ Tag = ’DRUG’ ] ( kw2 ) = >/NP { / NP { / ? [ V a l u e = ’ s u b s t r a t e ’ ] ( kw1 ) }= >/ PP { / / ? [ Tag = ’
GENE ’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e ,
sent . value
B.8
1
2
3
4
5
6
2
Gene Regulates Gene
/ / S { / NP { / ? [ Tag = ’GENE ’ ] ( kw0 ) }= >/VP { / VP { / / ? [ V a l u e IN { ’ r e g u l a t e d ’ , ’ u p r e g u l a t e d
’ , ’ d o w n r e g u l a t e d ’ , ’ up− r e g u l a t e d ’ , ’ down− r e g u l a t e d ’ } ] ( kw1 ) = > / / ? [ V a l u e = ’
by ’ ] ( kw3 ) = > / / ? [ Tag = ’GENE ’ ] ( kw2 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 .
v a l u e , kw2 . v a l u e , kw3 . v a l u e , s e n t . v a l u e
/ / S { / NP { / ? [ Tag = ’GENE ’ ] ( kw0 ) }= >/VP { / VP { / ? [ V a l u e IN { ’ r e g u l a t e d ’ , ’ u p r e g u l a t e d
’ , ’ d o w n r e g u l a t e d ’ , ’ up− r e g u l a t e d ’ , ’ down− r e g u l a t e d ’ } ] ( kw1 ) = > / / ? [ V a l u e = ’
by ’ ] ( kw3 ) = > / / ? [ Tag = ’GENE ’ ] ( kw2 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 .
v a l u e , kw2 . v a l u e , kw3 . v a l u e , s e n t . v a l u e
/ / S { / NP { / ? [ Tag = ’GENE ’ ] ( kw0 ) }= >/VP { / ? [ V a l u e IN { ’ r e g u l a t e d ’ , ’ down− r e g u l a t e d
’ } ] ( kw1 ) = >/NP { / / ? [ Tag = ’GENE ’ ] ( kw2 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e ,
kw1 . v a l u e , kw2 . v a l u e , s e n t . v a l u e
/ / NP { / NP { / ? [ V a l u e IN { ’ r e g u l a t i o n ’ , ’ u p r e g u l a t i o n ’ , ’ d o w n r e g u l a t i o n ’ , ’ up−
r e g u l a t i o n ’ , ’ down− r e g u l a t i o n ’ } ] ( kw1 ) }= >/ PP { / NP { / / ? [ Tag = ’GENE ’ ] ( kw0 )
= > / / ? [ V a l u e = ’ by ’ ] ( kw3 ) = > / / ? [ Tag = ’GENE ’ ] ( kw2 ) } } } : : : d i s t i n c t s e n t . c i d ,
kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , kw3 . v a l u e , s e n t . v a l u e
/ / S { / NP { / ? [ Tag = ’GENE ’ ] ( kw0 ) }= >/VP { / ? [ V a l u e IN { ’ r e g u l a t e s ’ , ’ u p r e g u l a t e s ’ , ’
d o w n r e g u l a t e s ’ , ’ up− r e g u l a t e s ’ , ’ down− r e g u l a t e s ’ } ] ( kw1 ) = >/NP { / / ? [ Tag = ’
GENE ’ ] ( kw2 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e ,
sent . value
/ / S { / NP { / ? [ Tag = ’GENE ’ ] ( kw0 ) }= >/VP { / VP { / / ? [ V a l u e = ’ i n ’ ] ( kw3 ) = > / / ? [ V a l u e = ’
r e g u l a t i n g ’ ] ( kw1 ) = > / / ? [ Tag = ’GENE ’ ] ( kw2 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 .
v a l u e , kw1 . v a l u e , kw2 . v a l u e , kw3 . v a l u e , s e n t . v a l u e
B.9
1
299
Gene Regulate Gene (Xenobiotic Metabolism)
/ / S { / / ? [ Tag = ’GENE’ AND C a n o n i c a l LIKE ’CYP\ % ’ ] ( kw0 ) <= >//?[ Tag = ’GENE’ AND
C a n o n i c a l IN { ’ AhR ’ , ’CASR’ , ’CAR’ , ’PXR’ , ’ NR1I2 ’ , ’ NR1I3 ’ } ] ( kw1 ) } : : :
d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , s e n t . v a l u e
/ / S { / / ? [ Tag = ’GENE’ AND V a l u e LIKE ’ c y t o c h r o m e \ % ’ ] ( kw0 ) <= >//?[ Tag = ’GENE’ AND
C a n o n i c a l IN { ’ AhR ’ , ’CASR’ , ’CAR’ , ’PXR’ , ’ NR1I2 ’ , ’ NR1I3 ’ } ] ( kw1 ) } : : :
d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , s e n t . v a l u e
B.10
Negative Drug Induces/Metabolizes/Inhibits Gene
APPENDIX B. DRUG-DRUG INTERACTION QUERIES
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
300
/ / S { / ? [ Tag = ’DRUG’ ] ( kw0 ) = >/VP { / VP { / ? [ V a l u e IN { ’ i n d u c e d ’ , ’ i n h i b i t e d ’ , ’
m e t a b o l i z e d ’ , ’ m e t a b o l i s e d ’ } ] ( kw1 ) = > / / ? [ Tag = ’GENE ’ ] ( kw2 ) = > / / ? [ V a l u e = ’ n o t
’ ] ( kw3 ) = > / / ? [ Tag = ’GENE ’ ] ( kw4 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 .
v a l u e , kw4 . v a l u e , kw3 . v a l u e , s e n t . v a l u e
/ / S { / ? [ Tag = ’DRUG’ ] ( kw0 ) = >/?[ V a l u e = ’ n o t ’ ] ( kw3 ) = >/?[ Tag = ’DRUG’ ] ( kw4 ) = >/?[ V a l u e
= ’ i n h i b i t e d ’ ] ( kw1 ) = >/?[ Tag = ’GENE ’ ] ( kw2 ) } : : : d i s t i n c t s e n t . c i d , kw4 . v a l u e
, kw1 . v a l u e , kw2 . v a l u e , kw3 . v a l u e , s e n t . v a l u e
/ / S { / SBAR { / S { / / ? [ Tag = ’DRUG’ ] ( kw0 ) } } = > / S { / S { / / ? [ V a l u e = ’ m e t a b o l i z e d ’ ] ( kw1 )
= > / / ? [ Tag = ’GENE ’ ] ( kw2 ) = > / / ? [ V a l u e = ’ n o t ’ ] ( kw3 ) = > / / ? [ Tag = ’GENE ’ ] ( kw4 ) } } }
: : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw4 . v a l u e , kw3 . v a l u e , s e n t .
value
/ / S { / NP { / PP { / / ? [ Tag = ’GENE ’ ] ( kw2 ) } } = > /VP { / VP { / ? [ V a l u e IN { ’ i n d u c e d ’ , ’
i n h i b i t e d ’ , ’ m e t a b o l i z e d ’ , ’ m e t a b o l i s e d ’ } ] ( kw1 ) = > / / ? [ Tag = ’DRUG’ ] ( kw0 )
= > / / ? [ V a l u e = ’ n o t ’ ] ( kw3 ) = > / / ? [ Tag = ’DRUG’ ] ( kw4 ) } } } : : : d i s t i n c t s e n t . c i d ,
kw4 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , kw3 . v a l u e , s e n t . v a l u e
/ / S { / NP { / PP { / / ? [ Tag = ’DRUG’ ] ( kw0 ) } } = > /VP { / VP { / ? [ V a l u e IN { ’ i n d u c e d ’ , ’
i n h i b i t e d ’ , ’ m e t a b o l i z e d ’ , ’ m e t a b o l i s e d ’ } ] ( kw1 ) = > / / ? [ Tag = ’GENE ’ ] ( kw2 )
= > / / ? [ V a l u e IN { ’ n o t ’ , ’ no ’ } ] ( kw3 ) = > / / ? [ Tag = ’GENE ’ ] ( kw4 ) } } } : : : d i s t i n c t
s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw4 . v a l u e , kw3 . v a l u e , s e n t . v a l u e
/ / S { / ? [ Tag = ’DRUG’ ] ( kw0 ) = >/VP { / ? [ V a l u e = ’ n o t ’ ] ( kw3 ) = >/VP { / / ? [ V a l u e IN { ’ i n d u c e d
’ , ’ i n h i b i t e d ’ , ’ m e t a b o l i z e d ’ , ’ m e t a b o l i s e d ’ } ] ( kw1 ) = > / / ? [ Tag = ’GENE ’ ] ( kw2 )
} } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , kw3 . v a l u e ,
sent . value
/ / S { / NP { / ? [ Tag = ’GENE ’ ] ( kw2 ) }= >/VP { / VP { / ? [ V a l u e IN { ’ i n d u c e d ’ , ’ i n h i b i t e d ’ , ’
m e t a b o l i z e d ’ , ’ m e t a b o l i s e d ’ } ] ( kw1 ) = > / / ? [ Tag = ’DRUG’ ] ( kw0 ) = > / / ? [ V a l u e = ’ n o t
’ ] ( kw3 ) = > / / ? [ Tag = ’DRUG’ ] ( kw4 ) } } } : : : d i s t i n c t s e n t . c i d , kw4 . v a l u e , kw1 .
v a l u e , kw2 . v a l u e , kw3 . v a l u e , s e n t . v a l u e
/ / S { / NP { / NP { / ? [ V a l u e = ’ n o t ’ ] ( kw3 ) = >/?[ Tag = ’DRUG’ ] ( kw0 ) } } = > /VP { / ? [ V a l u e IN { ’
i n h i b i t e d ’ , ’ i n d u c e d ’ } ] ( kw1 ) = >/NP { / / ? [ Tag = ’GENE ’ ] ( kw2 ) } } } : : : d i s t i n c t
s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , kw3 . v a l u e , s e n t . v a l u e
/ / S { / NP { / NP { / ? [ V a l u e IN { ’ no ’ , ’ n o t ’ } ] ( kw3 ) = >/?[ Tag = ’DRUG’ ] ( kw0 ) } } = > /VP { / ? [
V a l u e IN { ’ i n h i b i t e d ’ , ’ i n d u c e d ’ } ] ( kw1 ) = >/NP { / ? [ Tag = ’GENE ’ ] ( kw2 ) } } } : : :
d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , kw3 . v a l u e , s e n t . v a l u e
/ / S { / ? [ Tag = ’DRUG’ ] ( kw0 ) = >/S { / S { / / ? [ V a l u e = ’ n o t ’ ] ( kw3 ) = > / / ? [ V a l u e IN { ’ i n d u c e ’ ,
’ i n h i b i t ’ } ] ( kw1 ) = > / / ? [ Tag = ’GENE ’ ] ( kw2 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 .
v a l u e , kw1 . v a l u e , kw2 . v a l u e , kw3 . v a l u e , s e n t . v a l u e
/ / S { / ? [ Tag = ’DRUG’ ] ( kw0 ) = >/?[ Tag = ’GENE ’ ] ( kw2 ) = >/?[ V a l u e IN { ’ n o t ’ } ] ( kw3 ) = >/?[
V a l u e IN { ’ i n h i b i t ’ , ’ i n d u c e ’ } ] ( kw1 ) } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e ,
kw1 . v a l u e , kw2 . v a l u e , kw3 . v a l u e , s e n t . v a l u e
/ / S { / ? [ Tag = ’DRUG’ ] ( kw0 ) = >/VP { / ? [ V a l u e = ’ n o t ’ ] ( kw3 ) = >/VP { / ? [ V a l u e IN { ’ i n h i b i t
’ , ’ i n d u c e ’ } ] ( kw1 ) = > / / ? [ Tag = ’GENE ’ ] ( kw2 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 .
v a l u e , kw1 . v a l u e , kw2 . v a l u e , kw3 . v a l u e , s e n t . v a l u e
/ / S { / NP { / ? [ Tag = ’DRUG’ ] ( kw0 ) }= >/VP { / VP { / ? [ V a l u e = ’ n o t ’ ] ( kw3 ) = > / / ? [ V a l u e = ’
i n h i b i t ’ ] ( kw1 ) = > / / ? [ Tag = ’GENE ’ ] ( kw2 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e ,
kw1 . v a l u e , kw2 . v a l u e , kw3 . v a l u e , s e n t . v a l u e
/ / S { / NP { / NP { / ? [ Tag = ’GENE ’ ] ( kw2 ) } } = > /VP { / ? [ V a l u e = ’ n o t ’ ] ( kw3 ) = >/VP { / ? [ V a l u e = ’
m e t a b o l i z e ’ ] ( kw1 ) = > / / ? [ Tag = ’DRUG’ ] ( kw0 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 .
v a l u e , kw1 . v a l u e , kw2 . v a l u e , kw3 . v a l u e , s e n t . v a l u e
/ / S { / NP { / NP { / / ? [ Tag = ’DRUG’ ] ( kw0 ) } } = > /VP { / ? [ V a l u e = ’ n o t ’ ] ( kw3 ) = >/VP { / ? [ V a l u e IN
{ ’ i n h i b i t ’ , ’ i n d u c e ’ } ] ( kw1 ) = > / / ? [ Tag = ’GENE ’ ] ( kw2 ) } } } : : : d i s t i n c t s e n t .
c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , kw3 . v a l u e , s e n t . v a l u e
APPENDIX B. DRUG-DRUG INTERACTION QUERIES
16
17
18
19
20
21
22
/ / S { / NP { / ? [ Tag = ’DRUG’ ] ( kw0 ) }= >/VP { / ? [ V a l u e = ’ n o t ’ ] ( kw3 ) = >/VP { / ? [ V a l u e = ’ i n h i b i t
’ ] ( kw1 ) = > / / ? [ Tag = ’GENE ’ ] ( kw2 ) } } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 .
v a l u e , kw2 . v a l u e , kw3 . v a l u e , s e n t . v a l u e
/ / S { / NP { / ? [ Tag = ’DRUG’ ] ( kw0 ) }= >/VP { / ? [ V a l u e IN { ’ i n d u c e s ’ , ’ i n h i b i t s ’ } ] ( kw1 )
= >/NP { / / ? [ Tag = ’GENE ’ ] ( kw2 ) = > / / ? [ V a l u e = ’ n o t ’ ] ( kw3 ) = > / / ? [ Tag = ’GENE ’ ] ( kw4 )
} } } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw4 . v a l u e , kw3 . v a l u e ,
sent . value
/ / S { / ? [ Tag = ’DRUG’ ] ( kw0 ) = >/VP { / ? [ V a l u e = ’ i n h i b i t s ’ ] ( kw1 ) = >/NP { / / ? [ Tag = ’GENE ’ ] (
kw2 ) = > / / ? [ V a l u e = ’ n o t ’ ] ( kw3 ) = > / / ? [ Tag = ’GENE ’ ] ( kw4 ) } } } : : : d i s t i n c t s e n t .
c i d , kw0 . v a l u e , kw1 . v a l u e , kw4 . v a l u e , kw3 . v a l u e , s e n t . v a l u e
/ / S { / ? [ Tag = ’DRUG’ ] ( kw0 ) = >/VP { / NP { / / ? [ V a l u e = ’ no ’ ] ( kw3 ) = > / / ? [ V a l u e IN { ’
i n d u c t i o n ’ , ’ m e t a b o l i s m ’ , ’ i n h i b i t i o n ’ } ] ( kw1 ) = > / / ? [ Tag = ’GENE ’ ] ( kw2 ) } } } : : :
d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , kw3 . v a l u e , s e n t . v a l u e
/ / S { / NP { / NP { / / ? [ Tag = ’DRUG’ ] ( kw0 ) = > / / ? [ Tag = ’GENE ’ ] ( kw2 ) = > / / ? [ V a l u e IN { ’
i n d u c t i o n ’ , ’ m e t a b o l i s m ’ , ’ i n h i b i t i o n ’ } ] ( kw1 ) } } = > /VP { / ? [ V a l u e = ’ n o t ’ ] ( kw3 ) } }
: : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , kw3 . v a l u e , s e n t .
value
/ / S { / ? [ Tag = ’DRUG’ ] ( kw0 ) = >/VP { / ? [ V a l u e = ’ n o t ’ ] ( kw3 ) = >/VP { / / ? [ V a l u e IN { ’
i n d u c t i o n ’ , ’ m e t a b o l i s m ’ , ’ i n h i b i t i o n ’ } ] ( kw1 ) = > / / ? [ Tag = ’GENE ’ ] ( kw2 ) } } } : : :
d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , kw3 . v a l u e , s e n t . v a l u e
/ / S { / NP { / PP { / / ? [ Tag = ’DRUG’ ] ( kw0 ) } } = > /VP { / ? [ V a l u e = ’ n o t ’ ] ( kw3 ) = >/VP { / / ? [ V a l u e
IN { ’ i n d u c t i o n ’ , ’ m e t a b o l i s m ’ , ’ i n h i b i t i o n ’ } ] ( kw1 ) = > / / ? [ Tag = ’GENE ’ ] ( kw2 ) } } }
: : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , kw3 . v a l u e , s e n t .
value
B.11
1
2
3
4
5
301
Negative Drug Induces Gene
/ / S { / / ? [ Tag = ’GENE ’ ] ( kw0 ) = > / / ? [ V a l u e IN { ’ i n d u c e d ’ , ’ i n c r e a s e d ’ , ’ s t i m u l a t e d ’ } ] (
kw1 ) = > / / ? [ V a l u e = ’ by ’ ] ( kw3 ) = > / / ? [ Tag = ’DRUG’ ] ( kw2 ) = > / / ? [ V a l u e = ’ n o t ’ ] ( kw4 )
= > / / ? [ Tag = ’DRUG’ ] ( kw5 ) } : : : d i s t i n c t s e n t . c i d , kw5 . v a l u e , kw1 . v a l u e , kw0 .
v a l u e , kw4 . v a l u e , s e n t . v a l u e
/ / S { / / ? [ Tag = ’GENE ’ ] ( kw0 ) = > / / ? [ V a l u e = ’ n o t ’ ] ( kw4 ) = > / / ? [ Tag = ’GENE ’ ] ( kw5 ) = > / / ? [
V a l u e IN { ’ i n d u c e d ’ , ’ i n c r e a s e d ’ , ’ s t i m u l a t e d ’ } ] ( kw1 ) = > / / ? [ V a l u e = ’ by ’ ] ( kw3 )
= > / / ? [ Tag = ’DRUG’ ] ( kw2 ) } : : : d i s t i n c t s e n t . c i d , kw2 . v a l u e , kw1 . v a l u e , kw5 .
v a l u e , kw4 . v a l u e , s e n t . v a l u e
/ / S { / / ? [ Tag = ’DRUG’ ] ( kw0 ) = > / / ? [ V a l u e = ’ n o t ’ ] ( kw4 ) = > / / ? [ V a l u e IN { ’ i n d u c e ’ , ’
i n d u c e d ’ , ’ i n c r e a s e ’ , ’ i n c r e a s e d ’ , ’ s t i m u l a t e ’ , ’ s t i m u l a t e d ’ } ] ( kw1 ) = > / / ? [ Tag
= ’GENE ’ ] ( kw2 ) } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e ,
kw4 . v a l u e , s e n t . v a l u e
/ / S { / / ? [ Tag = ’DRUG’ ] ( kw0 ) = > / / ? [ V a l u e = ’ n o t ’ ] ( kw4 ) = > / / ? [ Tag = ’DRUG’ ] ( kw5 ) = > / / ? [
V a l u e IN { ’ i n d u c e s ’ , ’ i n c r e a s e s ’ , ’ s t i m u l a t e s ’ , ’ i n d u c e d ’ , ’ i n c r e a s e d ’ , ’
s t i m u l a t e d ’ } ] ( kw1 ) = > / / ? [ Tag = ’GENE ’ ] ( kw2 ) } : : : d i s t i n c t s e n t . c i d , kw5 .
v a l u e , kw1 . v a l u e , kw2 . v a l u e , kw4 . v a l u e , s e n t . v a l u e
/ / S { / / ? [ V a l u e = ’ no ’ ] ( kw5 ) = > / / ? [ v a l u e IN { ’ i n d u c t i o n ’ , ’ s t i m u l a t i o n ’ } ] ( kw1 )
= > / / ? [ V a l u e IN { ’ of ’ } ] ( kw2 ) = > / / ? [ Tag = ’GENE ’ ] ( kw0 ) = > / / ? [ v a l u e IN { ’ by ’ } ] (
kw3 ) = > / / ? [ Tag = ’DRUG’ ] ( kw4 ) } : : : d i s t i n c t s e n t . c i d , kw4 . v a l u e , kw1 . v a l u e ,
kw0 . v a l u e , kw5 . v a l u e , s e n t . v a l u e
B.12
Negative Drug Inhibits Gene
APPENDIX B. DRUG-DRUG INTERACTION QUERIES
1
2
3
4
5
/ / S { / / ? [ Tag = ’GENE ’ ] ( kw0 ) = > / / ? [ V a l u e IN { ’ i n h i b i t e d ’ , ’ d e c r e a s e d ’ } ] ( kw1 ) = > / / ? [
V a l u e = ’ by ’ ] ( kw3 ) = > / / ? [ Tag = ’DRUG’ ] ( kw2 ) = > / / ? [ V a l u e = ’ n o t ’ ] ( kw4 ) = > / / ? [ Tag = ’
DRUG’ ] ( kw5 ) } : : : d i s t i n c t s e n t . c i d , kw5 . v a l u e , kw1 . v a l u e , kw0 . v a l u e , kw4 .
value , s e n t . value
/ / S { / / ? [ Tag = ’GENE ’ ] ( kw0 ) = > / / ? [ V a l u e = ’ n o t ’ ] ( kw4 ) = > / / ? [ Tag = ’GENE ’ ] ( kw5 ) = > / / ? [
V a l u e IN { ’ i n h i b i t e d ’ , ’ d e c r e a s e d ’ } ] ( kw1 ) = > / / ? [ V a l u e = ’ by ’ ] ( kw3 ) = > / / ? [ Tag = ’
DRUG’ ] ( kw2 ) } : : : d i s t i n c t s e n t . c i d , kw2 . v a l u e , kw1 . v a l u e , kw5 . v a l u e , kw4 .
value , s e n t . value
/ / S { / / ? [ Tag = ’DRUG’ ] ( kw0 ) = > / / ? [ V a l u e = ’ n o t ’ ] ( kw4 ) = > / / ? [ V a l u e IN { ’ i n h i b i t ’ , ’
i n h i b i t e d ’ , ’ d e c r e a s e ’ , ’ d e c r e a s e d ’ } ] ( kw1 ) = > / / ? [ Tag = ’GENE ’ ] ( kw2 ) } : : :
d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , kw4 . v a l u e , s e n t . v a l u e
/ / S { / / ? [ Tag = ’DRUG’ ] ( kw0 ) = > / / ? [ V a l u e = ’ n o t ’ ] ( kw4 ) = > / / ? [ Tag = ’DRUG’ ] ( kw5 ) = > / / ? [
V a l u e IN { ’ i n h i b i t s ’ , ’ d e c r e a s e s ’ , ’ i n h i b i t e d ’ , ’ d e c r e a s e d ’ } ] ( kw1 ) = > / / ? [ Tag
= ’GENE ’ ] ( kw2 ) } : : : d i s t i n c t s e n t . c i d , kw5 . v a l u e , kw1 . v a l u e , kw2 . v a l u e ,
kw4 . v a l u e , s e n t . v a l u e
/ / S { / / ? [ V a l u e = ’ no ’ ] ( kw5 ) = > / / ? [ v a l u e IN { ’ i n h i b i t i o n ’ } ] ( kw1 ) = > / / ? [ V a l u e IN { ’
of ’ } ] ( kw2 ) = > / / ? [ Tag = ’GENE ’ ] ( kw0 ) = > / / ? [ v a l u e IN { ’ by ’ } ] ( kw3 ) = > / / ? [ Tag = ’
DRUG’ ] ( kw4 ) } : : : d i s t i n c t s e n t . c i d , kw4 . v a l u e , kw1 . v a l u e , kw0 . v a l u e , kw5 .
value , s e n t . value
B.13
1
2
3
4
2
Negative Gene Metabolizes Drug
/ / S { / / ? [ Tag = ’DRUG’ ] ( kw0 ) = > / / ? [ V a l u e IN { ’ m e t a b o l i z e d ’ , ’ m e t a b o l i s e d ’ } ] ( kw1 )
= > / / ? [ V a l u e = ’ by ’ ] ( kw3 ) = > / / ? [ Tag = ’GENE ’ ] ( kw2 ) = > / / ? [ V a l u e = ’ n o t ’ ] ( kw4 ) = > / / ? [
Tag = ’GENE ’ ] ( kw5 ) } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw5 . v a l u e ,
kw4 . v a l u e , s e n t . v a l u e
/ / S { / / ? [ Tag = ’DRUG’ ] ( kw0 ) = > / / ? [ V a l u e = ’ n o t ’ ] ( kw4 ) = > / / ? [ Tag = ’DRUG’ ] ( kw5 ) = > / / ? [
V a l u e IN { ’ m e t a b o l i z e d ’ , ’ m e t a b o l i s e d ’ } ] ( kw1 ) = > / / ? [ V a l u e = ’ by ’ ] ( kw3 ) = > / / ? [
Tag = ’GENE ’ ] ( kw2 ) } : : : d i s t i n c t s e n t . c i d , kw5 . v a l u e , kw1 . v a l u e , kw2 . v a l u e ,
kw4 . v a l u e , s e n t . v a l u e
/ / S { / / ? [ Tag = ’GENE ’ ] ( kw0 ) = > / / ? [ V a l u e = ’ n o t ’ ] ( kw4 ) = > / / ? [ V a l u e IN { ’ m e t a b o l i z e ’ , ’
m e t a b o l i s e ’ } ] ( kw1 ) = > / / ? [ Tag = ’DRUG’ ] ( kw2 ) } : : : d i s t i n c t s e n t . c i d , kw2 .
v a l u e , kw1 . v a l u e , kw0 . v a l u e , kw4 . v a l u e , s e n t . v a l u e
/ / S { / / ? [ Tag = ’GENE ’ ] ( kw0 ) = > / / ? [ V a l u e = ’ n o t ’ ] ( kw4 ) = > / / ? [ Tag = ’GENE ’ ] ( kw5 ) = > / / ? [
V a l u e IN { ’ m e t a b o l i z e ’ , ’ m e t a b o l i s e ’ , ’ m e t a b o l i z e s ’ , ’ m e t a b o l i s e s ’ } ] ( kw1 )
= > / / ? [ Tag = ’DRUG’ ] ( kw2 ) } : : : d i s t i n c t s e n t . c i d , kw2 . v a l u e , kw1 . v a l u e , kw5 .
v a l u e , kw4 . v a l u e , s e n t . v a l u e
B.14
1
302
Negative Gene Downregulates Gene
/ / S { / / ? [ Tag = ’GENE ’ ] ( kw0 ) = > / / ? [ V a l u e IN { ’ s u p p r e s s e d ’ , ’ d o w n r e g u l a t e d ’ , ’
i n h i b i t e d ’ , ’ down− r e g u l a t e d ’ , ’ r e p r e s s e d ’ , ’ d i s r u p t e d ’ } ] ( kw1 ) = > / / ? [ V a l u e = ’ by
’ ] ( kw3 ) = > / / ? [ Tag = ’GENE ’ ] ( kw2 ) = > / / ? [ V a l u e = ’ n o t ’ ] ( kw4 ) = > / / ? [ Tag = ’GENE ’ ] ( kw5
) } : : : d i s t i n c t s e n t . c i d , kw5 . v a l u e , kw1 . v a l u e , kw0 . v a l u e , kw4 . v a l u e ,
sent . value
/ / S { / / ? [ Tag = ’GENE ’ ] ( kw0 ) = > / / ? [ V a l u e = ’ n o t ’ ] ( kw4 ) = > / / ? [ Tag = ’GENE ’ ] ( kw5 ) = > / / ? [
V a l u e IN { ’ s u p p r e s s e d ’ , ’ d o w n r e g u l a t e d ’ , ’ i n h i b i t e d ’ , ’ down− r e g u l a t e d ’ , ’
r e p r e s s e d ’ , ’ d i s r u p t e d ’ } ] ( kw1 ) = > / / ? [ V a l u e = ’ by ’ ] ( kw3 ) = > / / ? [ Tag = ’GENE ’ ] ( kw2 )
} : : : d i s t i n c t s e n t . c i d , kw2 . v a l u e , kw1 . v a l u e , kw5 . v a l u e , kw4 . v a l u e , s e n t
. value
APPENDIX B. DRUG-DRUG INTERACTION QUERIES
3
4
5
/ / S { / / ? [ Tag = ’GENE ’ ] ( kw0 ) = > / / ? [ V a l u e = ’ n o t ’ ] ( kw4 ) = > / / ? [ V a l u e IN { ’ s u p p r e s s e d ’ , ’
s u p p r e s s ’ , ’ d o w n r e g u l a t e d ’ , ’ d o w n r e g u l a t e ’ , ’ i n h i b i t e d ’ , ’ i n h i b i t ’ , ’ down−
r e g u l a t e d ’ , ’ down− r e g u l a t e ’ , ’ r e p r e s s e d ’ , ’ r e p r e s s ’ , ’ d i s r u p t e d ’ , ’ d i s r u p t ’ } ] (
kw1 ) = > / / ? [ Tag = ’GENE ’ ] ( kw2 ) } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e ,
kw2 . v a l u e , kw4 . v a l u e , s e n t . v a l u e
/ / S { / / ? [ Tag = ’GENE ’ ] ( kw0 ) = > / / ? [ V a l u e = ’ n o t ’ ] ( kw4 ) = > / / ? [ Tag = ’GENE ’ ] ( kw5 ) = > / / ? [
V a l u e IN { ’ s u p p r e s s e s ’ , ’ d o w n r e g u l a t e s ’ , ’ i n h i b i t s ’ , ’ down− r e g u l a t e s ’ , ’
r e p r e s s e s ’ , ’ d i s r u p t s ’ , ’ s u p p r e s s e d ’ , ’ d o w n r e g u l a t e d ’ , ’ i n h i b i t e d ’ , ’ down−
r e g u l a t e d ’ , ’ r e p r e s s e d ’ , ’ d i s r u p t e d ’ } ] ( kw1 ) = > / / ? [ Tag = ’GENE ’ ] ( kw2 ) } : : :
d i s t i n c t s e n t . c i d , kw5 . v a l u e , kw1 . v a l u e , kw2 . v a l u e , kw4 . v a l u e , s e n t . v a l u e
/ / S { / / ? [ V a l u e = ’ no ’ ] ( kw5 ) = > / / ? [ v a l u e IN { ’ i n h i b i t i o n ’ , ’ d o w n r e g u l a t i o n ’ , ’ down−
r e g u l a t i o n ’ } ] ( kw1 ) = > / / ? [ V a l u e IN { ’ of ’ } ] ( kw2 ) = > / / ? [ Tag = ’GENE ’ ] ( kw0 ) = > / / ? [
v a l u e IN { ’ on ’ } ] ( kw3 ) = > / / ? [ Tag = ’GENE ’ ] ( kw4 ) } : : : d i s t i n c t s e n t . c i d , kw0 .
v a l u e , kw1 . v a l u e , kw4 . v a l u e , kw5 . v a l u e , s e n t . v a l u e
B.15
1
2
3
4
5
Negative Gene Upregulates Gene
/ / S { / / ? [ Tag = ’GENE ’ ] ( kw0 ) = > / / ? [ V a l u e IN { ’ a c t i v a t e d ’ , ’ i n d u c e d ’ , ’ s t i m u l a t e d ’ , ’
r e g u l a t e d ’ , ’ u p r e g u l a t e d ’ , ’ up− r e g u l a t e d ’ } ] ( kw1 ) = > / / ? [ V a l u e = ’ by ’ ] ( kw3 )
= > / / ? [ Tag = ’GENE ’ ] ( kw2 ) = > / / ? [ V a l u e = ’ n o t ’ ] ( kw4 ) = > / / ? [ Tag = ’GENE ’ ] ( kw5 ) } : : :
d i s t i n c t s e n t . c i d , kw5 . v a l u e , kw1 . v a l u e , kw0 . v a l u e , kw4 . v a l u e , s e n t . v a l u e
/ / S { / / ? [ Tag = ’GENE ’ ] ( kw0 ) = > / / ? [ V a l u e = ’ n o t ’ ] ( kw4 ) = > / / ? [ Tag = ’GENE ’ ] ( kw5 ) = > / / ? [
V a l u e IN { ’ a c t i v a t e d ’ , ’ i n d u c e d ’ , ’ s t i m u l a t e d ’ , ’ r e g u l a t e d ’ , ’ u p r e g u l a t e d ’ , ’
up− r e g u l a t e d ’ } ] ( kw1 ) : : : d i s t i n c t s e n t . c i d , kw2 . v a l u e , kw1 . v a l u e , kw5 .
v a l u e , kw4 . v a l u e , s e n t . v a l u e
/ / S { / / ? [ Tag = ’GENE ’ ] ( kw0 ) = > / / ? [ V a l u e = ’ n o t ’ ] ( kw4 ) = > / / ? [ V a l u e IN { ’ a c t i v a t e d ’ , ’
i n d u c e d ’ , ’ s t i m u l a t e d ’ , ’ r e g u l a t e d ’ , ’ u p r e g u l a t e d ’ , ’ up− r e g u l a t e d ’ , ’ a c t i v a t e
’ , ’ i n d u c e ’ , ’ s t i m u l a t e ’ , ’ r e g u l a t e ’ , ’ u p r e g u l a t e ’ , ’ up− r e g u l a t e ’ } ] ( kw1 ) = > / / ? [
Tag = ’GENE ’ ] ( kw2 ) } : : : d i s t i n c t s e n t . c i d , kw0 . v a l u e , kw1 . v a l u e , kw2 . v a l u e ,
kw4 . v a l u e , s e n t . v a l u e
/ / S { / / ? [ Tag = ’GENE ’ ] ( kw0 ) = > / / ? [ V a l u e = ’ n o t ’ ] ( kw4 ) = > / / ? [ Tag = ’GENE ’ ] ( kw5 ) = > / / ? [
V a l u e IN { ’ a c t i v a t e s ’ , ’ i n d u c e s ’ , ’ s t i m u l a t e s ’ , ’ r e g u l a t e s ’ , ’ u p r e g u l a t e s ’ , ’
up− r e g u l a t e s ’ , ’ a c t i v a t e ’ , ’ i n d u c e ’ , ’ s t i m u l a t e ’ , ’ r e g u l a t e ’ , ’ u p r e g u l a t e ’ , ’ up
− r e g u l a t e ’ } ] ( kw1 ) = > / / ? [ Tag = ’GENE ’ ] ( kw2 ) } : : : d i s t i n c t s e n t . c i d , kw5 . v a l u e
, kw1 . v a l u e , kw2 . v a l u e , kw4 . v a l u e , s e n t . v a l u e
/ / S { / / ? [ V a l u e = ’ no ’ ] ( kw5 ) = > / / ? [ v a l u e IN { ’ i n d u c t i o n ’ , ’ a c t i v a t i o n ’ , ’ s t i m u l a t i o n
’ , ’ r e g u l a t i o n ’ , ’ u p r e g u l a t i o n ’ , ’ up− r e g u l a t i o n ’ } ] ( kw1 ) = > / / ? [ V a l u e IN { ’ of
’ } ] ( kw2 ) = > / / ? [ Tag = ’GENE ’ ] ( kw0 ) = > / / ? [ v a l u e IN { ’ by ’ } ] ( kw3 ) = > / / ? [ Tag = ’GENE
’ ] ( kw4 ) } : : : d i s t i n c t s e n t . c i d , kw4 . v a l u e , kw1 . v a l u e , kw0 . v a l u e , kw5 .
value , s e n t . value
B.16
1
303
Drug Gene Co-Occurrence
/ / S { / / ? [ Tag = ’DRUG’ ] ( kw0 ) <= >//?[ Tag = ’GENE ’ ] ( kw1 ) } : : :
v a l u e , kw1 . v a l u e , kw0 . t y p e , kw1 . t y p e , s e n t . v a l u e
d i s t i n c t s e n t . c i d , kw0 .
Bibliography
Anwar, S., C. Baral and K. Inoue, “Encoding higher level extensions of petri nets in answer set
programming”, arXiv preprint arXiv:1306.3548 (2013a).
Anwar, S., C. Baral and K. Inoue, “Encoding petri nets in answer set programming for simulation based reasoning”, arXiv preprint arXiv:1306.3542 (2013b).
Araki, T. and T. Kasami, “Some decision problems related to the reachability problem for Petri
nets”, Theoretical Computer Science 3, 1, 85–104 (1976).
Baral, C., Knowledge Representation, Reasoning and Declarative Problem Solving (Cambridge
University Press, 2003).
Baral, C., K. Chancellor, N. Tran, N. Tran, A. Joy and M. Berens, “A knowledge based approach
for representing and reasoning about signaling networks”, Bioinformatics 20, suppl 1, i15–
i22 (2004).
Behrens, T. M. and J. Dix, “Model checking with logic based petri nets”, Pre-Proceedings of
CLIMA 7, 6–21 (2007).
Berg, J. M., J. L. Tymoczko and L. Stryer, Biochemistry. 5th edition. (New York: WH Freeman,
2002).
Berthomieu*, B., P.-O. Ribet and F. Vernadat, “The tool tina–construction of abstract state
spaces for petri nets and time petri nets”, International Journal of Production Research 42,
14, 2741–2756 (2004).
Best, E. and M. Koutny, “Petri net semantics of priority systems”, Theoretical Computer Science
96, 1, 175–215 (1992).
Boobis, A., “Drug interactions”, Drug Metabolism Review 41, 3, 486–527 (2009).
Brauer, W. and W. Reisig, “Carl Adam Petri and “Petri Nets””, Informatik-Spektrum 29, 5,
369–374 (2006).
Burkhard, H.-D., “Ordered firing in petri nets”, Journal of Information Processing and Cybernetics 17, 2-3, 71–86 (1980).
304
BIBLIOGRAPHY
305
Calimeri, F., W. Faber, M. Gebser, G. Ianni, R. Kaminski, T. Krennwallner, N. Leone, F. Ricca
and T. Schaub, “Asp-core-2 input language format”, (2013).
Camon, E., D. Barrell and V. Lee, “The gene ontology annotation (goa) database - an integrated
resource of go annotations to the uniprot knowledgebase”, Silico Biology 4, 1, 5–6 (2004).
Caporaso, J. G., W. A. B. Jr, D. A. Randolph, K. B. Cohen and L. Hunter, “Mutationfinder: A
high-performance system for extracting point mutation mentions from text”, Bioinformatics
23, 14, 1862–1865 (2007).
Chaouiya, C., “Petri net modelling of biological networks”, Briefings in bioinformatics 8, 4,
210–219 (2007).
Christensen, S. and N. D. Hansen, Coloured Petri nets extended with place capacities, test arcs
and inhibitor arcs (Springer, 1993).
Consortium, T. U., “Reorganizing the protein space at the universal protein resource (uniprot)”,
Nucleic Acids research 40 (2012).
Croft, D., G. O’Kelly, G. Wu, R. Haw, M. Gillespie, L. Matthews, M. Caudy, P. Garapati,
G. Gopinath, B. Jassal et al., “Reactome: a database of reactions, pathways and biological
processes”, Nucleic acids research 39, suppl 1, D691–D697 (2011).
de Marneffe, M.-C. and C. D. Manning, “The stanford typed dependencies representation”,
Coling 2008: Proceedings of the Workshop on Cross-Framework and Cross-Domain Parser
Evaluation (2008).
den Dunnen, J. and S. Antonarakis, “Mutation nomenclature extensions and suggestions to describe complex mutations: a discussion”, Human Mutation 15, 1, 7–12 (200).
Dimroth, P., C. von Ballmoos, T. Meier and G. Kaim, “Electrical power fuels rotary atp synthase”, Structure 11, 12, 1469–1473 (2003).
Dworschak, S., T. Grote, A. König, T. Schaub and P. Veber, “Tools for representing and reasoning about biological models in action language c”, in “Proceedings of the Twelfth International Workshop on Nonmonotonic Reasoning (NMR 2008). The University of New South
Wales, Technical Report Series”, pp. 94–102 (2008).
Gebser, M., R. Kaminski, B. Kaufmann, M. Ostrowski, T. Schaub and M. Schneider, “Potassco:
The Potsdam answer set solving collection”, aicom 24, 2, 105–124 (2011).
Gelfond, M. and V. Lifschitz, “The stable model semantics for logic programming”, Logic
Programming: Proceedings of the Fifth International Conference and Symposium pp. 1070–
1080 (1988).
Gelfond, M. and V. Lifschitz, “Representing action and change by logic programs”, The Journal
of Logic Programming 17, 2, 301–321 (1993).
BIBLIOGRAPHY
306
Gelfond, M. and V. Lifschitz, “Action languages”, Electronic Transactions on AI 3, 16 (1998).
Gelfond, M. and V. Lifschitz, “The common core of action languages B and C”, in “Proceedings
of the 14th International Workshop on Non-Monotonic Reasoning (NMR 2012)”, (2012).
Giunchiglia, E., J. Lee, V. Lifschitz, N. McCain and H. Turner, “Nonmonotonic causal theories”,
Artificial Intelligence 153, 1, 49–104 (2004).
Giunchiglia, E. and V. Lifschitz, “An action language based on causal explanation: Preliminary
report”, in “AAAI/IAAI”, pp. 623–630 (Citeseer, 1998).
H, W., R. T, H. R, B. K, B. M and W. J, “Increased frequency of cytochrome p450 2d6 poor metabolizers among patients with metoprolol-associcated adverse effects”, Pharmacogenomics
in Clinical Therapeutics 72, 429–437 (2002).
Hakenberg, J., C. Plake, R. Leaman, M. Schroeder and G. Gonzalez, “Inter-species normalization of gene mentions”, Bioinformatics 24, 16, 126–132 (2008).
Hakenberg, J., D. Voronov, V. H. Nguyen, S. Liang, B. Lumpkin, S. Anwar, R. Leaman, L. Tari
and C. Baral, “Taking a snpshot of pubmed - a repository of genetic variants and their drug
response phenotypes”, Journal of Biomedical Informatics (2012).
Hardy, S. and P. N. Robillard, “Modeling and simulation of molecular biology systems using
petri nets: Modeling goals of various approaches”, Journal of Bioinformatics and Computational Biology 2, 4, 619–637 (2004).
Harrison, A., V. Lifschitz and F. Yang, “On the semantics of gringo”, arXiv preprint
arXiv:1312.6149 (2013).
Heiner, M., I. Koch and J. Will, “Model validation of biological pathways using petri nets—
demonstrated for apoptosis”, Biosystems 75, 1, 15–28 (2004).
Heljanko, K. and I. Niemelä, “Petri net analysis and nonmonotomic reasoning”, Leksa Notes in
Computer Science pp. 7–19 (2000).
J, K., B. K, D. ML, G. LF, K. S, R. I, S. F, S. E and B. J, “Cyp2d6 and cyp2c19 genotypebased dose recommendations for antidepressants: a first step towards subpopulation-specific
dosages”, Acta Psychiatr Scand 104, 3, 173–192 (2001).
Jensen, K., L. M. Kristensen and L. Wells, “Coloured petri nets and cpn tools for modelling and
validation of concurrent systems”, International Journal on Software Tools for Technology
Transfer 9, 3-4, 213–254 (2007).
Kanehisa, M. and S. Goto, “Kegg: kyoto encyclopedia of genes and genomes”, Nucleic acids
research 28, 1, 27–30 (2000).
BIBLIOGRAPHY
307
Karp, P. D., M. Riley, S. M. Paley and A. Pellegrini-Toole, “The metacyc database”, Nucleic
acids research 30, 1, 59–61 (2002).
Klein, D. and C. Manning, “Accurate unlexicalized parsing”, Proceedings of the 41st meeting
of the association for computation linguistics pp. 423–430 (2003).
Kohn, K. W., M. I. Aladjem, J. N. Weinstein and Y. Pommier, “Molecular interaction maps of
bioregulatory networks: a general rubric for systems biology”, Molecular biology of the cell
17, 1, 1–13 (2006).
Kounev, S., C. Dutz and A. Buchmann, “Qpme-queueing petri net modeling environment”, in
“Quantitative Evaluation of Systems, 2006. QEST 2006. Third International Conference on”,
pp. 115–116 (IEEE, 2006).
Krallinger, M., A. Morgan, L. Smith, F. Leitner, L. Tanabe, J. Wilbur, L. Hirschman and A. Valencia, “Evaluation of text-mining systems for biology: overview of the second biocreative
community challenge”, Gnome Biology 9, 2 (2008).
Krepska, E., N. Bonzanni, A. Feenstra, W. Fokkink, T. Kielmann, H. Bal and J. Heringa, “Design issues for qualitative modelling of biological cells with petri nets”, Formal Methods in
Systems Biology pp. 48–62 (2008).
Kuhn, M., M. Campillos, I. Letunic, L. J. Jensen and P. Bork, “A side effect resource to capture
phenotypic effects of drugs”, Molecular Systems Biology 6, 343 (2010).
Kummer, O., F. Wienberg, M. Duvigneau, M. Köhler, D. Moldt and H. Rölke, “Renew–the
reference net workshop”, in “Tool Demonstrations, 21st International Conference on Application and Theory of Petri Nets, Computer Science Department, Aarhus University, Aarhus,
Denmark”, pp. 87–89 (2000).
Leaman, R., G. Gonzalez et al., “Banner: An executable survey of advances in biomedical
named entity recognition”, in “Pacific Symposium on Biocomputing”, vol. 13, pp. 652–663
(2008).
Leaman, R., C. Miller and G. Gonzalez, “Enabling recognition of diseases in biomedical text
with machine learning: Corpus and benchmark”, International Symposium on Languages in
Biology and Medicine pp. 82–89 (2009).
Lee, J. and V. Lifschitz, “Describing additive fluents in action language C+”, in “Proc. of IJCAI
2003”, (2003).
Lee, J., V. Lifschitz and F. Yang, “Action language bc: Preliminary report”, in “Proceedings
of the Twenty-Third international joint conference on Artificial Intelligence”, pp. 983–989
(AAAI Press, 2013).
BIBLIOGRAPHY
308
Lee, J. and Y. Meng, “Answer set programming modulo theories and reasoning about continuous
changes”, in “Proceedings of the Twenty-Third international joint conference on Artificial
Intelligence”, pp. 990–996 (AAAI Press, 2013).
Lifschitz, V., “What is answer set programming?”, Proceedings of the Twenty-Third AAAI
Conferece on Artificial Intelligence pp. 1594–1597 (2008).
Luo, G., M. Cunningham, S. Kim, T. Burn, J. Lin, M. Sinz, G. Hamilton, C. Rizzo, S. Jolley,
D. Gilbert et al., “Cyp3a4 induction by drugs: correlation between a pregnane x receptor
reporter gene assay and cyp3a4 expression in human hepatocytes”, Drug Metabolism and
Disposition 30, 7, 795–804 (2002).
Manna, Z. and A. Pnueli, The temporal logic of reactive and concurrent systems: specifications,
vol. 1 (springer, 1992).
Music, G., “Schedule optimization based on coloured petri nets and local search”, in “Mathematical Modelling”, vol. 7, pp. 352–357 (2012).
Nagasaki, M., A. Saito, E. Jeong, C. Li, K. Kojima, E. Ikeda and S. Miyano, “Cell illustrator
4.0: A computational platform for systems biology”, In silico biology 10, 1, 5–26 (2010).
Pearl, J., “Action as local surgery”, in “Working Notes of AAAI Spring Symposium Series,
Extending Theories of Actions”, pp. 157–162 (1995).
Peterson, J. et al., “A note on colored Petri nets”, Information Processing Letters 11, 1, 40–43
(1980).
Peterson, J. L., “Petri nets”, Computing Surveys 9, 3, 223–252 (1977).
Priami, C., A. Regev, E. Shapiro and W. Silverman, “Application of a stochastic name-passing
calculus to representation and simulation of molecular processes”, Information processing
letters 80, 1, 25–31 (2001).
Ramchandani, C., “Analysis of asynchronous concurrent systems by Petri nets.”, Tech. rep.,
DTIC Document (1974).
Reece, J., M. Cain, L. Urry, P. Minorsky and S. Wasserman, Campbell Biology (Pearson Benjamin Cummings, 2010), URL http://books.google.com/books?id=39vMSgAACAAJ.
Regev, A., “Representation and simulation of molecular pathways in the stochastic pi-calculus”,
in “Proceedings of the 2nd workshop on Computation of Biochemical Pathways and Genetic
Networks”, vol. 88 (2001).
Reiter, R., “Natural actions, concurrency and continuous time in the situation calculus”, KR 96,
2–13 (1996).
BIBLIOGRAPHY
309
Rohr, C., W. Marwan and M. Heiner, “Snoopy—a unifying petri net framework to investigate
biomolecular networks”, Bioinformatics 26, 7, 974–975 (2010).
SF, Z., “Polymorphism of human cytochrome p450 2d6 and its clinical significance: Part i”,
Clinical Pharmacoknetics 48, 11, 689–723 (2009).
Sleator, D. and D. Temperley, “Parsing english with a link grammar”, Third International Workshop on Parsing Technologies (1993).
Son, T. C. and E. Pontelli, “A constructive semantic characterization of aggregates in answer set
programming”, Theory and Practice of Logic Programming 7, 3, 355 (2007).
Syropoulos, A., “Mathematics of multisets”, Multiset Processing pp. 347–358 (2001).
Tari, L., C. Baral, S. Anwar, S. Liang and J. Hakenberg, “Synthesis of pharmacokinetic pathways through knowledge acquisition and automated reasoning”, Pacific Symposium on Biocomputing 15, 465–476 (2010a).
Tari, L., J. Cai, C. Baral, S. Liang and S. Anwar, “Discovering drug interactions: a text-mining
and reasoning approach based on properties of drug metabolism.”, European Conference on
Computational Biology (2010b).
Tari, L., J. Hakenberg, G. Gonzalez and C. Baral, “Querying a parse tree database of medline
text to synthesize user-specific biomolecular networks”, Proc Pac Symp Biocomput 14, 87–98
(2009).
Tari, L., P. H. Tu, J. Hakenberg, Y. Chen, T. C. Son, G. Gonzalez and C. Baral, “Incremental
information extraction using relational databases”, IEEE Transactions on Knowledge and
Data Engineering 24, 1, 86–99 (2012).
Thorn, C. F., T. E. Klein and R. B. Altman, “Pharmacogenomics and bioinformatics: Pharmgkb”, Pharmacogenomics 11, 4, 501–505 (2010).
Valk, R., “Self-modifying nets, a natural extension of Petri nets”, in “Automata, Languages and
Programming”, vol. 62 of Lecture Notes in Computer Science, pp. 464–476 (Springer, 1978).
Valk, R., “Object Petri nets: Using the nets-within-nets paradigm”, in “Advances in Petri Nets:
Lectures on Concurrency and Petri Nets”, edited by J. Desel, W. Reisig and G. Rozenberg,
vol. 3098 of Lecture Notes in Computer Science, pp. 819–848 (Springer-Verlag, 2004).
Wishart, D., C. Knox, A. Guo, D. Cheng, S. Shrivastava, D. Tzur, B. Gautam and M. Hassanali,
“Drugbank: a knowledgebase for drugs, drug actions and drug targets.”, Nucleic Acids research 36, 901–906 (2008).
Wishart, D. S., C. Knox and A. C. Guo, “Drugbank: a comprehensive resource for in silico drug
discovery and exploration”, Nucleic Acids research 34, 1, 668–672 (2006).
| 5cs.CE
|
arXiv:1709.08853v4 [cs.LG] 8 Oct 2017
Object-oriented Neural Programming (OONP)
for Document Understanding
Zhengdong Lu1 Haotian Cui2 * Xianggen Liu2 * Yukun Yan2 * Daqi Zheng1
1
DeeplyCurious.ai
{luz,da}@deeplycurious.ai
2
Department of Bio-medical Engineering, Tsinghua University
{cht15, liuxg16, yanyk13}@mails.tsinghua.edu.cn
Abstract
We propose Object-oriented Neural Programming (OONP), a framework for semantically
parsing documents in specific domains. Basically, OONP reads a document and parses it
into a predesigned object-oriented data structure (referred to as ontology in this paper)
that reflects the domain-specific semantics of the document. An OONP parser models semantic parsing as a decision process: a neural net-based Reader sequentially goes through
the document, and during the process it builds and updates an intermediate ontology to
summarize its partial understanding of the text it covers. OONP supports a rich family
of operations (both symbolic and differentiable) for composing the ontology, and a big
variety of forms (both symbolic and differentiable) for representing the state and the document. An OONP parser can be trained with supervision of different forms and strength,
including supervised learning (SL) , reinforcement learning (RL) and hybrid of the two.
Our experiments on both synthetic and real-world document parsing tasks have shown
that OONP can learn to handle fairly complicated ontology with training data of modest
sizes.
1
Introduction
Mapping a document into a structured “machine readable” form is a canonical and probably
the most effective way for document understanding. There are quite some recent efforts on
designing neural net-based learning machines for this purpose, which can be roughly categorized into two groups: 1) sequence-to-sequence model with the neural net as the the black
box [Dong and Lapata, 2016, Liang et al., 2017], and 2) neural net as a component in a predesigned statistical model [Zeng et al., 2014]. We however argue that both approaches have
their own serious problems and cannot be used on document with relatively complicated structures. Towards solving this problem, we proposed Object-oriented Neural Programming (OONP
), a framework for semantically parsing in-domain documents. OONP is neural net-based, but
it also has sophisticated architecture and mechanism designed for taking and outputting discrete structures, hence nicely combining symbolism (for interpretability and formal reasoning)
and connectionism (for flexibility and learnability). This ability, as we argue in this paper, is
critical to document understanding.
OONP seeks to map a document to a graph structure with each node being an object, as illustrated in Figure 1. We borrow the name from Object-oriented Programming [Mitchell, 2003]
* The work is done when the authors worked as interns at DeeplyCurious.ai.
1
to emphasize the central position of “objects” in our parsing model: indeed, the representation of objects in OONP allows neural and symbolic reasoning over complex structures and
hence it make it possible to represent much richer semantics. Similar to Object-oriented Programming, OONP has the concept of “class” and “objects” with the following analogousness:
1) each class defines the types and organization of information it contains, and we can define
inheritance for class with different abstract levels as needed; 2) each object is an instance
of a certain class, encapsulating a number of properties and operations; 3) objects can be
connected with relations (called links) with pre-determined type. Based on objects, we can
define ontology and operations that reflect the intrinsic structure of the parsing task.
For parsing, OONP reads a document and parses it into this object-oriented data structure
through a series of discrete actions along reading the document sequentially. OONP supports
a rich family of operations for composing the ontology, and flexible hybrid forms for knowledge
representation. An OONP parser can be trained with supervised learning (SL) , reinforcement
learning (RL) and hybrid of the two. Our experiments on one synthetic dataset and two realworld datasets have shown the efficacy of OONP on document understanding tasks with a
variety of characteristics.
Figure 1: Illustration of OONP on a parsing task.
In addition to the work on semantic parsing mentioned above, OONP is also related to
multiple threads of work in natural language processing and machine learning. It is inspired
by [Daumé III et al., 2009] on modeling parsing as a decision process, and also state-tracking
models in dialogue system [Henderson et al., 2014] for the mixture of symbolic and probabilistic representations of dialogue state. OONP is also related to [Johnson, 2017] for modeling
the transition of symbolic state and [Henaff et al., 2016] on having explicit (although not
thorough) modeling on entities. OONP is also obviously related to the the recent work on
neural-symbolism [Mou et al., 2017, Liang et al., 2017].
1.1
Overview of OONP
An OONP parser (as illustrated through the diagram in Figure 2) consists of a Reader equipped
with read/write heads, Inline Memory that represents the document, and Carry-on Memory that
summarizes the current understanding of the document at each time step. For each document
to parse, OONP first preprocesses it and puts it into the Inline Memory , and then Reader
controls the read-heads to sequentially go through the Inline Memory (for possibly multiple
times, see Section 6.3 for an example) and at the same time update the Carry-on Memory.
2
Figure 2: The overall digram of OONP, where S stands for symbolic representation, D stands
for distributed representation, and S+D stands for a hybrid representation with both symbolic
and distributed parts.
The major components of OONP are described in the following:
• Memory: we have two types of memory, Carry-on Memory and Inline Memory. Carry-on
Memory is designed to save the state∗ in the decision process and summarize current
understanding of the document based on the text that has been ‘read”. Carry-on Memory
has three compartments:
– Object Memory: denoted as Mobj , the object-based ontology constructed during the
parsing process, see Section 2.1 for details;
– Matrix Memory: denoted as Mmat , a matrix-type memory with fixed size, for differentiable read/write by the controlling neural net [Graves et al., 2014]. In the
simplest case, it could be just a vector as the hidden state of conventional Recurrent Neural Netwokr (RNN);
– Action History: denoted as Mact , saving the entire history of actions made during
the parsing process.
Intuitively, Mobj stores the extracted knowledge with defined structure and strong evidence, while Mmat keeps the knowledge that is fuzzy, uncertain or incomplete, waiting
for future information to confirm, complete and clarify. Inline Memory , denoted Minl ,
is designed to save location-specific information about the document. In a sense, the
information in Inline Memory is low level and unstructured, waiting for Reader to fuse
and integrate for more structured representation.
• Reader: Reader is the control center of OONP, coordinating and managing all the
operations of OONP. More specifically, it takes the input of different forms (reading),
processes it (thinking), and updates the memory (writing). As shown in Figure 3,
Reader contains Neural Net Controller (NNC) and multiple symbolic processors, and
Neural Net Controller also has Policy-net as its sub-component. Similar to the controller
in Neural Turing Machine [Graves et al., 2014], Neural Net Controller is equipped with
∗
It is not entirely accurate, since the Inline Memory can be modified during the reading process it also
records some of the state information.
3
multiple read-heads and write-heads for differentiable read/write over Matrix Memory
and (the distributed part of) Inline Memory, with possibly a variety of addressing strategies [Graves et al., 2014]. Policy-net however issues discrete outputs (i.e., actions), which
gradually builds and updates the Object Memory in time (see Section 2.1 for more details). The actions could also updates the symbolic part of Inline Memory if needed. The
symbolic processors are designed to handle information in symbolic form from Object
Memory, Inline Memory, Action History, and Policy-net, while that from Inline Memory
and Action History is eventually generated by Policy-net.
Figure 3: The overall digram of OONP
We can show how the major components of OONP collaborate to make it work through
the following sketchy example. In reading the following text
OONP has reached the underlined word “BMW” in Inline Memory. At this moment, OONP
has two objects (I01 and I02) for Audi-06 and BMW respectively in Object Memory. Reader
determines that the information it is currently holding is about I02 (after comparing it with
both objects) and updates its status property to sold, along with other update on both Matrix
Memory and Action History.
OONP in a nutshell: The key properties of OONP can be summarized as follows
1. OONP models parsing as a decision process: as the “reading and comprehension” agent
goes through the text it gradually forms the ontology as the representation of the text
through its action;
2. OONP uses a symbolic memory with graph structure as part of the state of the parsing
process. This memory will be created and updated through the sequential actions of
the decision process, and will be used as the semantic representation of the text at the
end;
3. OONP can blend supervised learning (SL) and reinforcement learning (RL) in tuning
its parameters to suit the supervision signal in different forms and strength;
4. OONP allows different ways to add symbolic knowledge into the raw representation of
the text (Inline Memory) and its policy net in forming the final structured representation
of the text.
4
RoadMap of the paper: The rest of the paper is organized as follows. We will elaborate
on the components of OONP in Section 2 and actions of OONP in Section 3. After that we
will give a detailed analysis on the neural-symbolism in OONP in Section 4. Then in Section
5 we will discuss the learning for OONP , which is followed by experiments on three datasets
in Section 6. Finally we conclude the paper in Section 7.
2
OONP: Components
In this section we will discuss the major components in OONP, namely Object Memory , Inline
Memory and Reader. We omit the discussion on Matrix Memory and Action History since they
are straightforward given the description in Section 1.1.
2.1
Object Memory
Object Memory stores an object-oriented representation of document, as illustrated in Figure 4.
Each object is an instance of a particular class† , which specifies the internal structure of the
object, including internal properties, operations, and how this object can be connected with
others. The internal properties can be of different types, for example string or category, which
usually correspond to different actions in composing them: the string-type property is usually
“copied” from the original text in Inline Memory, while the category properties usually needs
to be rendered by a classifier. The links are by nature bi-directional, meaning that it can be
added from both ends (e.g., in the experiment in Section 6.1), but for modeling convenience,
we might choose to let it to be one directional (e.g., in the experiments in Section 6.2 and
6.3). In Figure 4, there are six “linked” objects of three classes (namely, Person, Event,
and Item) . Taking Item-object I02 for example, it has five internal properties (Type, Model,
Color, Value, Status), and is linked with two Event-objects through stolen and disposed
link respectively.
In addition to the symbolic part, each object had also its own distributed presentation
(named object-embedding), which serves as its interface with other distributed representations
in Reader (e.g., those from the Matrix Memory or the distributed part of Inline Memory).
For description simplicity, we will refer to the symbolic part of this hybrid representation
of objects as ontology, with some slight abuse of this word. Object-embedding serves as a
dual representation to the symbolic part of a object, recording all the relevant information
associated with it but not represented in the ontology, e.g., the context of text when the
object is created.
The representations in Object Memory, including the ontology and object embeddings,
will be updated in time by the operations defined for the corresponding classes. Usually, the
actions are the driving force in those operations, which not only initiate and grow the ontology,
but also coordinate other differentiable operations. For example, object-embedding associated
with a certain object changes with any non-trivial action concerning this object, e.g., any
update on the internal properties or the external links, or even a mention (corresponding to
an Assign action described in Section 3) without any update.
†
In this paper, we limit ourselves to a flat structure of classes, but it is possible and even beneficial to have
a hierarchy of classes. In other words, we can have classes with different levels of abstractness, and allow an
object to go from abstract class to its child class during the parsing process, with more and more information
is obtained.
5
Figure 4: An example of the objects from three classes.
According to the way the ontology evolves with time, the parsing task can be roughly
classified into two categories
• Stationary: there is a final ground truth that does not change with time. So with
any partial history of the text, the corresponding ontology is always part of the final
one, while the missing part is due to the lack of information. See task in Section 6.2
and 6.3 for example.
• Dynamical: the truth changes with time, so the ontology corresponding to partial
history of text may be different from that of the final state. See task in Section 6.1 for
example.
It is important to notice that this categorization depends not only on the text but also heavily
on the definition of ontology. Taking the text in Figure 1 for example: if we define ownership
relation between a Person-object and Item-object, the ontology becomes dynamical, since
ownership of the BMW changed from Tom to John.
2.2
Inline Memory
Inline Memory stores the relatively raw representation of the document that follows the temporal structure of the text, as illustrated through Figure 2. Basically, Inline Memory is an array
of memory cells, each corresponding to a pre-defined language unit (e.g., word) in the same
order as they are in the original text. Each cell can have distributed part and symbolic part,
designed to save 1) the result of preprocessing of text from different models, and 2) certain
output from Reader, for example from previous reading rounds. Following are a few examples
for preprocessing
• Word embedding: context-independent vectorial representation of words
6
• Hidden states of NNs: we can put the context in local representation of words through
gated RNN like LSTM [Greff et al., 2015] or GRU [Cho et al., 2014], or particular design
of convolutional neural nets (CNN) [Yu and Koltun, 2015].
• Symbolic preprocessing: this refer to a big family of methods that yield symbolic
result, including various sequential labeling models and rule-based methods. As the
result we may have tag on words, extracted sub-sequences, or even relations on two
pieces.
During the parsing process, Reader can write to Inline Memory with its discrete or continuous
outputs, a process we named “notes-taking”. When the output is continuous, the notes-taking
process is similar to the interactive attention in machine translation [Meng et al., 2016], which
is from a NTM-style write-head [Graves et al., 2014] on Neural Net Controller. When the
output is discrete, the notes-taking is essentially an action issued by Policy-net.
Inline Memory provides a way to represent locally encoded “low level” knowledge of the
text, which will be read, evaluated and combined with the global semantic representation in
Carry-on Memory by Reader. One particular advantage of this setting is that it allows us to
incorporate the local decisions of some other models, including “higher order” ones like local
relations across two language units, as illustrated in the left panel of Figure 5. We can also
have a rather “nonlinear” representation of the document in Inline Memory. As a particular
example [Yan et al., 2017], at each location, we can have the representation of the current
word, the representation of the rest of the sentence, and the representation of the rest of the
current paragraph, which enables Reader to see information of history and future at different
scales, as illustrated in the right panel of Figure 5.
Figure 5: Left panel: Inline Memory with symbolic knowledge; Right panel: one choice of
nonlinear representation of the distributed part onf Inline Memory used in [Yan et al., 2017].
2.3
Reader
Reader is the control center of OONP , which manages all the (continuous and discrete) operations in the OONP parsing process. Reader has three symbolic processors (namely, Symbolic
Matching, Symbolic Reasoner, Symbolic Analyzer) and a Neural Net Controller (with Policy-net
as the sub-component). All the components in Reader are coupled through intensive exchange
of information as shown in Figure 6. Below is a snapshot of the information processing at
time t in Reader
• STEP-1: let the processor Symbolic Analyzer to check the Action History (Mtact ) to construct
some symbolic features for the trajectory of actions;
• STEP-2: access Matrix Memory (Mtmat ) to get an vectorial representation for time t, denoted
as st ;
7
(s)
• STEP-3: access Inline Memory (Mtinl ) to get the symbolic representation xt (through location(d)
based addressing) and distributed representation xt (through location-based addressing and/or
content-based addressing);
(d)
• STEP-4: feed xt
(s)
and the embedding of xt
to Neural Net Controller to fuse with st ;
(s)
• STEP-5: get the candidate objects (some may have been eliminated by xt ) and let them meet
(d)
xt through the processor Symbolic Matching for the matching of them on symbolic aspect;
(s)
• STEP-6: get the candidate objects (some may have been eliminated by xt ) and let them meet
the result of STEP-4 in Neural Net Controller ;
• STEP-7: Policy-net combines the result of STEP-6 and STEP-5, to issue actions;
• STEP-8: update Mtobj , Mtmat and Mtinl with actions on both symbolic and distributed representations;
• STEP-9: put Mtobj through the processor Symbolic Reasoner for some high-level reasoning and
logic consistency.
Note that we consider only single action for simplicity, while in practice it is common to
have multiple actions at one time step, which requires a slightly more complicated design of
the policy as well as the processing pipeline.
Figure 6: A particular implementation of Reader in a closer look, which reveals some details
about the entanglement of neural and symbolic components. Dashed lines stand for continuous
signal and the solid lines for discrete signal
3
OONP: Actions
The actions issued by Policy-net can be generally categorized as the following
• New-Assign : determine whether to create an new object (a “New ” operation) for the
information at hand or assign it to a certain existed object
8
• Update.X : determine which internal property or external link of the selected object to
update;
• Update2what : determine the content of the updating, which could be about string,
category or links.
The typical order of actions is New-Assign → Update.X → Update2what, but it is very
common to have New-Assign action followed by nothing, when, for example, an object is
mentioned but no substantial information is provided,
3.1
New-Assign
With any information at hand (denoted as St ) at time t, the choices of New-Assign typically
include the following three categories of actions: 1) creating (New) an object of a certain
type, 2) assigning St to an existed object, and 3) doing nothing for St and moving on. For
Policy-net, the stochastic policy is to determine the following probabilities:
prob(c, new|St ),
c = 1, 2, · · · , |C|
prob(c, k|St ),
for Otc,k ∈ Mtobj
prob(none|St )
where |C| stands for the number of classes, Otc,k stands for the k th object of class c at time t.
Determining whether to new objects always relies on the following two signals
1. The information at hand cannot be contained by any existed objects;
2. Linguistic hints that suggests whether a new object is introduced.
Based on those intuitions, we takes a score-based approach to determine the above-mentioned
probability. More specifically, for a given St , Reader forms a “temporary” object with its
own structure (denoted Ôt ), including symbolic and distributed sections. In addition, we also
have a virtual object for the New action for each class c, denoted Otc,new , which is typically a
time-dependent vector formed by Reader based on information in Mtmat . For a given Ôt , we
can then define the following 2|C| + 1 types of score functions, namely
New an object of class c:
Assign to existed objects:
Do nothing:
(c)
(c)
scorenew (Otc,new , Ôt ; θnew ),
c = 1, 2, · · · , |C|
(c)
(c)
scoreassign (Otc,k , Ôt ; θassign ),
for Otc,k ∈ Mtobj
scorenone (Ôt ; θnone ).
to measure the level of matching between the information at hand and existed objects, as well
as the likeliness for creating an object or doing nothing. This process is pictorially illustrated
in Figure 7. We therefore can define the following probability for the stochastic policy
prob(c, new|St ) =
prob(c, k|St ) =
prob(none|St ) =
(c)
c,new
(c)
(c)
c,k
(c)
escorenew (Ot ,Ôt ;θnew )
Z(t)
escoreassign (Ot ,Ôt ;θassign )
Z(t)
escorenone (Ôt ;θnone )
Z(t)
9
00
(c00 )
(c00 )
(c0 )
(c0 )
P
P
c0 ,new
scoreassign (Otc ,k ,Ôt ;θassign )
,Ôt ;θnew ) +
where Z(t) = c0 ∈C escorenew (Ot
e
+escorenone (Ôt ;θnone )
t
00
0
(c ,k )∈idx(Mobj )
is the normalizing factor.
Figure 7: A pictorial illustration of what the Reader sees in determining whether to New an
object and the relevant object when the read-head on Inline Memory reaches the last word in
the sentence in Figure 2. The color of the arrow line stands for different matching functions
for object classes, where the dashed lines is for the new object.
Many actions are essentially trivial on the symbolic part, for example, when Policy-net
chooses none in New-Assign, or assigns the information at hand to an existed object but
choose to update nothing in Update.X, but this action will affect the distributed operations
in Reader. This distributed operation will affect the representation in Matrix Memory or the
object-embedding in Object Memory.
3.2
Updating objects: Update.X and Update2what
In Update.X step, Policy-net needs to choose the property or external link (or none) to update for the selected object determined by New-Assign step. If Update.X chooses to update
an external link, Policy-net needs to further determine which object it links to. After that
Update2what updates the chosen property or links. In task with static ontology, most internal
properties and links will be “locked” after they are updated for the first time, with some exception on a few semi-structured property (e.g., the Description property in the experiment
in Section 6.2). For dynamical ontology, on contrary, many important properties and links are
always subject to changes. A link can often be determined from both ends, e.g., the link that
states the fact that “Tina (a Person-object ) carries apple (an Item-object )” can be either
specified from from Tina(through adding the link “carry” to apple) or from apple (through
adding the link “iscarriedby” to Tina ), as in the experiment in Section 6.1. In practice, it
is often more convenient to make it asymmetrical to reduce the size of action space.
In practice, for a particular type of ontology, both Update.X and Update2what can often
be greatly simplified: for example,
10
• when the selected object (in New-Assign step) has only one property “unlocked”, the
Update.X step will be trivial;
• in St , there is often information from Inline Memory that tells us the basic type of the
current information, which can often automatically decide the property or link.
3.3
An example
In Figure 8, we give an example of the entire episode of OONP parsing on the short text given
in the example in Figure 1. Note that different from our late treatment of actions, we let
some selection actions (e.g., the Assign) be absorbed into the updating actions to simplify
the illustration.
Figure 8: A pictorial illustration of a full episode of OONP parsing, where we assume the
description of cars (highlighted with shadow) are segmented in preprocessing.
11
4
OONP: Neural-Symbolism
OONP offers a way to parse a document that imitates the cognitive process of human when
reading and comprehending a document: OONP maintains a partial understanding of document as a mixture of symbolic (representing clearly inferred structural knowledge) and distributed (representing knowledge without complete structure or with great uncertainty). As
shown in Figure 2, Reader is taking and issuing both symbolic signals and continuous signals,
and they are entangled through Neural Net Controller.
OONP has plenty space for symbolic processing: in the implementation in Figure 6, it is
carried out by the three symbolic processors. For each of the symbolic processors, the input
symbolic representation could be rendered partially by neural models, therefore providing an
intriguing way to entangle neural and symbolic components. Here are three examples we
implemented for two different tasks
1. Symbolic analysis in Action History: There are many symbolic summary of history we
can extracted or constructed from the sequence of actions, e.g., “The system just New an
object with Person-class five words ago” or “The system just put a paragraph starting
with ‘(2)’ into event-3”. In the implementation of Reader shown in Figure 6, this
analysis is carried out with the component called Symbolic Analyzer. Based on those
more structured representation of history, Reader might be able to make a informed
guess like “If the coming paragraph starts with ‘(3)’, we might want to put it to
event-2” based on symbolic reasoning. This kind of guess can be directly translated
into feature to assist Reader’s decisions, resembling what we do with high-order features
in CRF [Lafferty et al., 2001], but the sequential decision makes it possible to construct
a much richer class of features from symbolic reasoning, including those with recursive
structure. One example of this can be found in [Yan et al., 2017], as a special case of
OONP on event identification.
2. Symbolic reasoning on Object Memory: we can use an extra Symbolic Reasoner to
take care of the high-order logic reasoning after each update of the Object Memory
caused by the actions. This can illustrated through the following example. Tina (a
Person-object) carries an apple (an Item-object), and Tina moves from kitchen (a
Location-object) to garden (Location-object) at time t. Supposing we have both
Tina-carry-apple and Tina-islocatedat-kitchen relation kept in Object Memory at
time t, and OONP updates the Tina -islocatedat-kitchen to Tina -islocatedatgarden at time t+1, the Symbolic Reasoner can help to update the relation apple
-islocatedat-kitchen to apple -islocatedat-garden . This is feasible since the
Object Memory is supposed to be logically consistent. This external logic-based update
is often necessary since it is hard to let the Neural Net Controller see the entire Object
Memory due to the difficulty to find a distributed representation of the dynamic structure
there. Please see Section 6.1 for experiments.
3. Symbolic prior in New-Assign : When Reader determines an New-Assign action, it needs
to match the information about the information at hand (St ) and existed objects. There
is a rich set of symbolic prior that can be added to this matching process in Symbolic
Matching component. For example, if St contains a string labeled as entity name (in
preprocessing), we can use some simple rules (part of the Symbolic Matching component)
to determine whether it is compatible with an object with the internal property Name.
12
5
Learning
The parameters of OONP models (denoted Θ) include that for all operations and that for
composing the distributed sections in Inline Memory. They can be trained with different learning paradigms: it takes both supervised learning (SL) and reinforcement learning (RL) while
allowing different ways to mix the two. Basically, with supervised learning, the oracle gives
the ground truth about the “right action” at each time step during the entire decision process,
with which the parameter can be tuned to maximize the likelihood of the truth. In a sense,
SL represents rather strong supervision which is related to imitation learning [Stefan, 1999]
and often requires the labeler (expert) to give not only the final truth but also when and
where a decision is made. For supervised learning, the objective function is given as
JSL (Θ) = −
N Ti
1 XX
(i)
log(πt [a?t ])
N
i
(1)
t=1
where N stands for the number of instances, Ti stands for the number of steps in decision
(i)
process for the ith instance, πt [·] stands for the probabilities of the feasible actions at t from
the stochastic policy, and a?t stands fro the ground truth action in step t.
With reinforcement learning, the supervision is given as rewards during the decision process, for which an extreme case is to give the final reward at the end of the decision process
by comparing the generated ontology and the ground truth, e.g.,
(
0, if t 6= Ti
(i)
(2)
rt =
i
, Gi ), if t = Ti
match(MTobj
i
i
and the
, Gi ) measures the consistency between the ontology of in MTobj
where the match(MTobj
?
ground truth G . We can use any policy search algorithm to maximize the expected total
reward. With the commonly used REINFORCE [Williams, 1992] for training, the gradient is
given by
N T
h
(i) i
(i)
1 XX
∇Θ JRL (Θ) = Eπθ ∇Θ log πΘ ait |sit rt:T ≈ −
∇Θ log πΘ ait |sit rt:Ti .
N Ti
i
(3)
t=1
When OONP is applied to real-world tasks, there is often quite natural SL and RL. More
specifically, for “static ontology” one can often infer some of the right actions at certain time
steps by observing the final ontology based on some basic assumption, e.g.,
• the system should New an object the first time it is mentioned,
• the system should put an extracted string (say, that for Name ) into the right property
of right object at the end of the string.
For those that can not be fully reverse-engineered, say the categorical properties of an object
(e.g., Type for event objects), we have to resort to RL to determine the time of decision, while
we also need SL to train Policy-net on the content of the decision. Fortunately it is quite
straightforward to combine the two learning paradigms in optimization. More specifically, we
maximize this combined objective
J (Θ) = JSL (Θ) + λJRL (Θ),
(4)
where JSL and JRL are over the parameters within their own supervision mode and λ coordinates the weight of the two learning mode on the parameters they share. Equation 4
13
actually indicates a deep coupling of supervised learning and reinforcement learning, since for
any episode the samples of actions related to RL might affect the inputs to the models under
supervised learning.
For dynamical ontology (see Section 6.1 for example), it is impossible to derive most of
the decisions from the final ontology since they may change over time. For those, we have
to rely mostly on the supervision at the time step to train the action (supervised mode) or
count on the model to learn the dynamics of the ontology evolution by fitting the final ground
truth. Both scenarios are discussed in Section 6.1 on a synthetic task.
6
Experiments
We applied OONP on three document parsing tasks, to verify its efficacy on parsing documents
with different characteristics and investigate different components of OONP.
6.1
6.1.1
Task-I: bAbI Task
Data and task
We implemented OONP an enriched version of bAbI tasks [Johnson, 2017] with intermediate
representation for history of arbitrary length. In this experiment, we considered only the
original bAbi task-2 [Weston et al., 2015], with an instance shown in the left panel Figure 9.
The ontology has three types of objects: Person-object, Item-object, and Location-object,
and three types of links:
1. is-located-atA : between a Person-object and a Location-object,
2. is-located-atB : between a Item-object and a Location-object;
3. carry: between a Person-object and Item-object;
which could be rendered by description of different ways. All three types of objects have Name
as the only internal property.
Figure 9: One instance of bAbI (6-sentence episode) and the ontology of two snapshots.
14
The task for OONP is to read an episode of story and recover the trajectory of the evolving
ontology. We choose this synthetic dataset because it has dynamical ontology that evolves
with time and ground truth given for each snapshot, as illustrated in Figure 9. Comparing
with the real-world tasks we will present later, bAbi has almost trivial internal property but
relatively rich opportunities for links, considering any two objects of different types could
potentially have a link.
6.1.2
Implementation details
For preprocessing, we have a trivial NER to find the names of people, items and locations
(saved in the symbolic part of Inline Memory) and word-level bi-directional GRU for the
distributed representations of Inline Memory. In the parsing process, Reader goes through the
inline word-by-word in the temporal order of the original text, makes New-Assign action at
every word, leaving Update.X and Update2what actions to the time steps when the read-head
on Inline Memory reaches a punctuation (see more details of actions in Table 1). For this
simple task, we use an almost fully neural Reader (with MLPs for Policy-net) and a vector for
Matrix Memory, with however a Symbolic Reasoner for some logic reasoning after each update
of the links, as illustrated through the following example. Suppose at time t, the ontology in
Mtobj contains the following three facts (among others)
• fact-1: John (a Person-object) is in kichten (a Location-object);
• fact-2: John carries apple (an Item-object);
• fact-3: John drops apple;
where fact-3 is just established by Policy-net at t. Symbolic Reasoner will add a new is-located-atB
link between apple and kitchen based on domain logic‡ .
Action
NewObject(c)
AssignObject(c, k)
Update(c, k).AddLink(c0 , k 0 , `)
Update(c, k).DelLink(c0 , k 0 , `)
Description
New an object of class-c.
Assign the current information to existed object (c, k)
Add an link of type-` from object-(c, k) to object-(c0 , k 0 )
Delete the link of type-` from object-(c, k) to object-(c0 , k 0 )
Table 1: Actions for bAbI.
6.1.3
Results and Analysis
For training, we use 1,000 episodes with length evenly distributed from one to six. We use
just REINFORCE with only the final reward defined as the overlap between the generated
ontology and the ground truth, while step-by-step supervision on actions yields almost perfect
result (result omitted). For evaluation, we use the following two metrics:
• the Rand index [Rand, 1971] between the generated set of objects and the ground truth,
which counts both the duplicate objects and missing ones, averaged over all snapshots
of all test instances;
• the F1 [Rijsbergen, 1979] between the generated links and the ground truth averaged
over all snapshots of all test instances, since the links are typically sparse compared
with all the possible pairwise relations between objects.
‡
The logic says, an item is not “in” a location if it is held by a person.
15
with results summarized in Table 2. OONP can learn fairly well on recovering the evolving
ontology with such a small training set and weak supervision (RL with the final reward),
which clearly shows that the credit assignment over to earlier snapshots does not cause much
difficulty in the learning of OONP even with a generic policy search algorithm. It is not so
surprising to observe that Symbolic Reasoner helps to improve the results on discovering the
links, while it does not improves the performance on identifying the objects although it is
taken within the learning. It is quite interesting to observe that OONP achieves rather high
accuracy on discovering the links while it performs relatively poorly on specifying the objects.
It is probably due to the fact that the rewards does not penalizes the objects.
model
OONP (without S.R.)
OONP (with S.R. )
F1 (for links) %
RandIndex (for objects)%
94.80
95.30
87.48
87.48
Table 2: The performance a implementation of OONP on bAbI task 2.
6.2
6.2.1
Task-II: Parsing Police Report
Data & task
We implement OONP for parsing Chinese police report (brief description of criminal cases
written by policeman), as illustrated in the left panel of Figure 10. We consider a corpus of
5,500 cases with a variety of crime categories, including theft, robbery, drug dealing and others.
The ontology we designed for this task mainly consists of a number of Person-objects and
Item-objects connected through a Event-object with several types of relations, as illustrated
in the right panel of Figure 10. A Person-object has three internal properties: Name (string),
Gender (categorical) and Age (number), and two types of external links (suspect and victim)
to an Event-object. An Item-object has three internal properties: Name (string), Quantity
(string) and Value (string), and six types of external links (stolen, drug, robbed, swindled,
damaged, and other) to an Event-object. Compared with bAbI in Section 6.1, the police
report ontology has less pairwise links but much richer internal properties for objects of all
three objects. Although the language in this dataset is reasonably formal, the corpus coverages
a big variety of topics and language styles, and has a high proportion of typos. The average
length of a document is 95 Chinese characters, with digit string (say, ID number) counted as
one character.
Figure 10: An example of police report and its ontology.
16
6.2.2
Implementation details
The OONP model is designed to generate ontology as illustrated in Figure 10 through a
decision process with actions in Table 3. As pre-processing, we performed regular NER
with third party algorithm (therefore not part of the learning) and simple rule-based extraction to yield the symbolic part of Inline Memory as shown in Figure 11. For the distributed part of Inline Memory, we used dilated CNN with different choices of depth and
kernel size [Yu and Koltun, 2015], all of which will be jointly learned during training. In
making the New-Assign decision, Reader considers the matching between two structured objects, as well as the hints from the symbolic part of Inline Memory as features, as pictorially
illustrated in Figure 7. In updating objects with its string-type properties (e.g., Name for a
Person-object ), we use Copy-Paste strategy for extracted string (whose NER tag already
specifies which property in an object it goes to) as Reader sees it. For undetermined category
properties in existed objects, Policy-net will determine the object to update (an New-Assign
action without New option), its property to update (an Update.X action), and the updating
operation (an Update2what action) at milestones of the decision process , e.g., when reaching an punctuation. For this task, since all the relations are between the single by-default
Event-object and other objects, the relations can be reduced to category-type properties
of the corresponding objects in practice. For category-type properties, we cannot recover
New-Assign and Update.X actions from the label (the final ontology), so we resort RL for
learning to determine that part, which is mixed with the supervised learning for Update2what
and other actions for string-type properties.
Action
NewObject(c)
AssignObject(c, k)
UpdateObject(c, k).Name
UpdateObject(Person, k).Gender
UpdateObject(Item, k).Quantity
UpdateObject(Item, k).Value
UpdateObject(Event, 1).Items.x
UpdateObject(Event, 1).Persons.x
Description
New an object of class-c.
Assign the current information to existed object (c, k)
Set the name of object-(c, k) with the extracted string.
Set the name of a Person-object indexed k with the extracted string.
Set the quantity of an Item-object indexed k with the extracted string.
Set the value of an Item-object indexed k with the extracted string.
Set the link between the Event-object and an Item-object, where x ∈{stolen,
drug, robbed, swindled, damaged, other}
Set the link between the Event-object and an Person-object, and x ∈{victim,
suspect}
Table 3: Actions for parsing police report.
6.2.3
Results & discussion
We use 4,250 cases for training, 750 for validation an held-out 750 for test. We consider the
following four metrics in comparing the performance of different models:
Assignment Accuracy
Category Accuracy
Ontology Accuracy
Ontology Accuracy-95
the accuracy on New-Assign actions made by the model
the accuracy of predicting the category properties of all the objects
the proportion of instances for which the generated ontology is exactly
the same as the ground truth
the proportion of instances for which the generated ontology achieves
95% consistency with the ground truth
which measures the accuracy of the model in making discrete decisions as well as generating
the final ontology. We empirically examined several OONP implementations and compared
them with a Bi-LSTM baseline, with results given in Table 4.
17
Model
Bi-LSTM (baseline)
OONP (neural)
OONP (structured)
OONP (RL)
Assign Acc. (%)
73.2 ± 0.58
88.5 ± 0.44
91.2 ± 0.62
91.4 ± 0.38
Type Acc. (%)
84.3 ± 0.58
87.0 ± 0.40
87.8 ± 0.75
Ont. Acc. (%)
36.4± 1.56
61.4 ± 1.26
65.4 ± 1.42
66.7 ± 0.95
Ont. Acc-95 (%)
59.8 ± 0.83
75.2 ± 1.35
79.9 ± 1.28
80.7 ± 0.82
Table 4: OONP on parsing police reports.
The Bi-LSTM is essentially a simple version of OONP without a structured Carry-on
Memory and designed operations (sophisticated matching function in New-Assign ). Basically
it consists of a Bi-LSTM Inline Memory encoder and a two-layer MLP on top of that acting as
a simple Policy-net for prediction actions. Since this baseline does not has an explicit object
representation, it does not support category type of prediction. We hence only train this
baseline model to perform New-Assign actions, and evaluate with the Assignment Accuracy
(first metric) and a modified version of Ontology Accuracy (third and fourth metric) that
counts only the properties that can be predicted by Bi-LSTM, hence in favor of Bi-LSTM.
We consider three OONP variants:
• OONP (neural): simple version of OONP with only distributed representation in Reader
in determining all actions;
• OONP (structured): OONP that considers the matching between two structured objects
in New-Assign actions, with symbolic prior encoded in Symbolic Matching and other
features for Policy-net;
• OONP (RL): another version of OONP (structured) that uses RL to determine the time
for predicting the category properties, while OONP (neural) and OONP (neural) use a
rule-based approach to determine the time.
As shown in Table 4, Bi-LSTM baseline struggles to achieve around 73% Assignment
Accuracy on test set, while OONP (neural) can boost the performance to 88.5%. Arguably,
this difference in performance is due to the fact that Bi-LSTM lacks Object Memory, so all
relevant information has to be stored in the Bi-LSTM hidden states along the reading process.
When we start putting symbolic representation and operation into Reader , as shown in the
result of OONP (structure), the performance is again significantly improved on all four metrics.
More specifically, we have the following two observations (not shown in the table),
• Adding inline symbolic features as in Figure 11 improves around 0.5% in New-Assign
action prediction, and 2% in category property prediction. The features we use include
the type of the candidate strings and the relative distance to the maker character we
chose.
Figure 11: Information in distributed and symbolic forms in Inline Memory.
18
• Using a matching function that can take advantage of the structures in objects helps
better generalization. Since the objects in this task has multiple property slots like
Name, Gender, Quantity, Value. We tried adding both the original text (e.g., Name,
Gender, Quantity, Value ) string of an property slot and the embedding of that as
additional features, e.g., the length the longest common string between the candidate
string and a relevant property of the object.
When using REINFORCE to determine when to make prediction for category property,
as shown in the result of OONP (RL), the prediction accuracy for category property and
the overall ontology accuracy is improved. It is quite interesting that it has some positive
impact on the supervised learning task (i.e., learning the New-Assign actions) through shared
parameters. The entanglement of the two learning paradigms in OONP is one topic for future
research, e.g., the effect of predicting the right category property on the New-Assign actions if
the predicted category property is among the features of the matching function for New-Assign
actions.
6.3
6.3.1
Task-III: Parsing court judgment documents
Data and task
We also implement OONP for parsing court judgement on theft. Unlike the two previous
tasks, court judgements are typically much longer, containing multiple events of different
types as well as bulks of irrelevant text, as illustrated in the left panel of Figure 10. The
dataset contains 1961 Chinese judgement documents, divided into training/dev/testing set
with 1561/200/200 texts respectively. The ontology we designed for this task mainly consists
of a number of Person-objects and Item-objects connected through a number Event-object
with several types of links. A Event-object has three internal properties: Time (string),
Location (string), and Type (category, ∈{theft, restitutionx, disposal}), four types
of external links to Person-objects (namely, principal, companion, buyer, victim) and
four types of external links to Item-objects (stolen, damaged, restituted, disposed). In
addition to the external links to Event-objects , a Person-object has only the Name (string)
as the internal property. An Item-object has three internal properties: Description (array
of strings), Value (string) and Returned(binary) in addition to its external links to Eventobjects , where Description consists of the words describing the corresponding item, which
could come from multiple segments across the document. A Person-object or an Itemobject could be linked to more than one Event-object, for example a person could be the
principal suspect in event A and also a companion in event B. An illustration of the judgement
document and the corresponding ontology can be found in Figure 12.
6.3.2
Implementation details
We use a model configuration similar to that in Section 6.2, with however the following important difference. In this experiment, OONP performs a 2-round reading on the text. In the first
round, OONP identifies the relevant events, creates empty Event-objects, and does NotesTaking on Inline Memory to save the information about event segmentation (see [Yan et al., 2017]
for more details). In the second round, OONP read the updated Inline Memory, fills the Eventobjects, creates and fills Person-objects and Item-objects, and specifies the links between
them. When an object is created during a certain event, it will be given an extra feature (not
19
Figure 12: Left panel: the judgement document with highlighted part being the description
the facts of crime, right panel: the corresponding ontology
an internal propoerty) indicating this connection, which will be used in deciding links between
this object and event object, as well as in determining the future New-Assign actions. The
actions of the two round reading are summarized in Table 5.
Action for 1st-round
NewObject(c)
NotesTaking(Event, k).word
NotesTaking(Event, k).sentence
NotesTaking(Event, k).paragraph
Skip.word
Skip.sentence
Skip.paragraph
Action for 2nd-round
NewObject(c)
AssignObject(c, k)
UpdateObject(Person, k).Name
UpdateObject(Item, k).Description
UpdateObject(Item, k).Value
UpdateObject(Event, k).Time
UpdateObject(Event, k).Location
UpdateObject(Event, k).Type
UpdateObject(Event, k).Items.x
UpdateObject(Event, k).Persons.x
Description
New an Event-object, with c =Event.
Put indicator of event-k on the current word.
Put indicator of event-k on the rest of sentence, and move the read-head to the
first word of next sentence
Put indicator of event-k on the rest of the paragraph, and move the read-head to
the first word of next paragraph.
Move the read-head to next word
Move the read-head to the first word of next sentence
Move the read-head to the first word of next paragraph.
Description
New an object of class-c.
Assign the current information to existed object (c, k)
Set the name of the k th Person-object with the extracted string.
Add to the description of an k th Item-object with the extracted string.
Set the value of an k th Item-object with the extracted string.
Set the time of an k th Event-object with the extracted string.
Set the location of an k th Event-object with the extracted string.
Set the type of the k th Event-object among {theft, disposal, restitution}
Set the link between the k th Event-object and an Item-object, where x ∈
{stolen, damaged, restituted, disposed }
Set the link between the k th Event-object and an Person-object, and x ∈
{principal, companion, buyer, victim}
Table 5: Actions for parsing court judgements.
6.3.3
Results and Analysis
We use the same metric as in Section 6.2, and compare two OONP variants, OONP (neural)
and OONP (structured), with Bi-LSTM. The Bi-LSTM will be tested only on the secondround reading, while both OONP variant are tested on a two-round reading. The results are
shown in Table 6. OONP parsers attain accuracy significantly higher than Bi-LSTM models.
Among, OONP (structure) achieves over 64% accuracy on getting the entire ontology right
and over 78% accuracy on getting 95% consistency with the ground truth.
20
Model
Bi-LSTM (baseline)
OONP (neural)
OONP (structured)
Assign Acc. (%)
84.66 ± 0.20
94.50 ± 0.24
97.49 ± 0.43
Type Acc. (%)
97.73 ± 0.12
97.43 ± 0.07
Ont. Acc. (%)
18.20 ± 1.01
53.29 ± 0.26
64.51 ± 0.99
Ont. Acc-95 (%)
36.88 ± 0.01
72.22 ± 1.01
78.61 ± 0.95
Table 6: OONP on judgement documents with the four metrics defined in Section 6.2.3.
7
Conclusion
We proposed Object-oriented Neural Programming (OONP), a framework for semantically
parsing in-domain documents. OONP is neural net-based, but equipped with sophisticated
architecture and mechanism for document understanding, therefore nicely combining interpretability and learnability. Our experiments on both synthetic and real-world document
parsing tasks have shown that OONP can learn to handle fairly complicated ontology with
training data of modest sizes.
References
[Cho et al., 2014] Cho, K., van Merrienboer, B., Gulcehre, C., Bougares, F., Schwenk, H.,
and Bengio, Y. (2014). Learning phrase representations using rnn encoder-decoder for
statistical machine translation. Proceedings of EMNLP, pages 17241734.
[Daumé III et al., 2009] Daumé III, H., Langford, J., and Marcu, D. (2009). Search-based
structured prediction.
[Dong and Lapata, 2016] Dong, L. and Lapata, M. (2016). Language to logical form with
neural attention. ACL, pages 33–43.
[Graves et al., 2014] Graves, A., Wayne, G., and Danihelka, I. (2014). Neural turing machines.
CoRR, abs/1410.5401.
[Greff et al., 2015] Greff, K., Srivastava, R. K., Koutnı́k, J., Steunebrink, B. R., and Schmidhuber, J. (2015). LSTM: A search space odyssey. CoRR, abs/1503.04069.
[Henaff et al., 2016] Henaff, M., Weston, J., Szlam, A., Bordes, A., and LeCun, Y. (2016).
Tracking the world state with recurrent entity networks. CoRR, abs/1612.03969.
[Henderson et al., 2014] Henderson, Matthew, Thomson, B., , and Young, S. (2014). Wordbased dialog state tracking with recurrent neural networks. Proceedings of the 15th Annual
Meeting of the Special Interest Group on Discourse and Dialogue (SIGDIAL), pp. 292-299.
[Johnson, 2017] Johnson, D. D. (2017). Learning graphical state transitions. the international
conference on learning representations.
[Lafferty et al., 2001] Lafferty, D., Mccallum, A., and Pereira, F. (2001). Conditional random fields: Probabilistic models for segmenting and labeling sequence data. international
conference on machine learning, pages 282–289.
[Liang et al., 2017] Liang, C., Berant, J., Le, Q., Forbus, K. D., and Lao, N. (2017). Neural
symbolic machines: Learning semantic parsers on freebase with weak supervision. ACL.
21
[Meng et al., 2016] Meng, F., Lu, Z., Li, H., and Liu, Q. (2016). Interactive attention for
neural machine translation. CoRR, abs/1610.05011.
[Mitchell, 2003] Mitchell, J. C. (2003). Concepts in programming languages. Cambridge
University Press.
[Mou et al., 2017] Mou, L., Lu, Z., Li, H., and Jin, Z. (2017). Coupling distributed and
symbolic execution for natural language queries. In Proceedings of the 34th International
Conference on Machine Learning (ICML), pages 2518–2526.
[Rand, 1971] Rand, W. M. (1971). Objective criteria for the evaluation of clustering methods.
Journal of the American Statistical Association, 66(336):846–850.
[Rijsbergen, 1979] Rijsbergen, C. J. V. (1979).
Heinemann, Newton, MA, USA, 2nd edition.
Information Retrieval.
Butterworth-
[Stefan, 1999] Stefan, S. (1999). Is imitation learning the route to humanoid robots? Trends
in cognitive sciences, 3.6:233–242.
[Weston et al., 2015] Weston, J., Bordes, A., Chopra, S., and Mikolov, T. (2015). Towards
ai-complete question answering: A set of prerequisite toy tasks. CoRR, abs/1502.05698.
[Williams, 1992] Williams, R. J. (1992). Simple statistical gradient-following algorithms for
connectionist reinforcement learning. Machine Learning, 8:229–256.
[Yan et al., 2017] Yan, Y., Zheng, D., Lu, Z., and Song, S. (2017). Event identification as a
decision process with non-linear representation of text. CoRR, abs/1710.00969.
[Yu and Koltun, 2015] Yu, F. and Koltun, V. (2015). Multi-scale context aggregation by
dilated convolutions. CoRR, abs/1511.07122.
[Zeng et al., 2014] Zeng, D., Liu, K., Lai, S., Zhou, G., and Zhao, J. (2014). Relation classification via convolutional deep neural network. In Proceedings of COLING.
22
| 9cs.NE
|
1
Distributed Rate and Power Control in Vehicular
Networks
arXiv:1511.01535v1 [cs.SY] 4 Nov 2015
Jubin Jose, Chong Li, Xinzhou Wu, Lei Ying and Kai Zhu
Abstract—The focus of this paper is on the rate and power
control algorithms in Dedicated Short Range Communication
(DSRC) for vehicular networks. We first propose a utility maximization framework by leveraging the well-developed network
congestion control, and formulate two subproblems, one on rate
control with fixed transmit powers and the other on power
control with fixed rates. Distributed rate control and power
control algorithms are developed to solve these two subproblems,
respectively, and are proved to be asymptotically optimal. Joint
rate and power control can be done by using the two algorithms
in an alternating fashion. The performance enhancement of our
algorithms compared with a recent rate control algorithm, called
EMBARC [1], is evaluated by using the network simulator ns2.
I. I NTRODUCTION
Dedicated Short Range Communication (DSRC) service [2]
is for vehicle-to-vehicle and vehicle-to-infrastructure communication in the 5.9 GHz band. Among the 75 MHz allocated to
DSRC, the channel 172 (5.855 GHz – 5.865 GHz) is assigned
for critical safety operations, which allows vehicles to periodically exchange Basic Safety Messages (BSM) to maximize
the mutual awareness to prevent collisions. Such messages
typically include the GPS position, velocity of the vehicle.
By receiving these BSM messages from surrounding vehicles,
all participating vehicles in the DSRC safety system can assess
the threat of potential collisions and provide warnings to the
driver, if necessary. United States Department of Transportation (USDOT) reported that the DSRC safety system based
on this simple mechanism can address 80% of the traffic
accidents on the road today and thus has potentially huge
societal benefit. On the other hand, such benefit is possible
only when timely and reliable information exchange among
vehicles using DSRC can be guaranteed in all deployment
scenarios.
DSRC is based on IEEE 802.11p standards [3] at PHY and
MAC layer. It has been well known that DSRC vehicular networks exhibit degraded performance in congested scenarios. In
particular, excessive packet loss can be observed at high node
density even between vehicles which are in close proximity
to each other [4], which can severely undermine the safety
benefit targeted by deploying such networks. Performance
improvement in such scenarios has been one of the key
challenging issues for the success of DSRC. The industry and
academics have contributed various solutions to this issue in
a collaborative way over the last decade, e.g., [5], [6], [7].
J.Jose, C. Li and X. Wu are with Qualcomm Research; L,Ying and K. Zhu
are with School of Electrical, Computer and Energy Engineering, Arizona
State University. The authors are listed in alphabetical order
The key system parameters one may control at each vehicle
are the transmit power and transmit rate, i.e. the periodicity
of the BSM messages, to alleviate the system congestion, i.e.
lower transmit rate and power reduces the footprint and also
the number of messages a DSRC device may generate and
thus reduce the congestion level in the critical safety channel.
On the other hand, both rate control and power control are
critical for system performance as the transmit power of a
vehicle determines the number of surrounding vehicles which
can discover the vehicle and higher transmit rate of the BSM
message can improve the accuracy the of collision estimation
between two vehicles by having more message exchanges.
Thus, a key problem to be addressed in the DSRC system is:
How to choose the transmit rate and power for each vehicle in
a distributed manner such that the overall system performance
is maximized without creating excessive network congestion,
i.e. observing very high channel load in some locations of the
network?
Both rate control and power control have been studied in
the literature (e.g. [8], [9]). However, most of these works
propose heuristic methods to adjust the rate and (or) power
in simplistic scenarios, e.g. single bottleneck scenarios (i.e.,
there is only one congested channel in the network). Further,
some of the methods [6], [1] rely on the existence of global
parameters for algorithm convergence, which leads to system
resource under-utilization in some scenarios.
The focus of this paper1 is to propose a network resource
allocation framework for rate and power control in vehicular
network, by leveraging existing network congestion control
framework [11] established in the context of wireline and
wireless networks, and then develop optimal distributed rate
and power control algorithms to achieve such a goal. The main
contributions of this paper are summarized below:
• We propose a utility maximization framework for rate and
power control in DSRC. In general, the utility maximization is a non-convex optimization problem with integer
constraints. We separate the problem to two subproblems:
rate control problem and power control problem, where
the rate control problem is to find the optimal broadcast
rates when the transmit power (i.e., the transmit ranges)
of vehicles are fixed and the power control problem is to
find the optimal transmit power (or transmission range)
when the broadcast rates are fixed.
• We develop a distributed rate control algorithm which
is similar to the dual algorithm for the Internet congestion control and prove that the time-average total utility
1 Partial
version of this paper has appeared in [10].
2
obtained under the proposed rate control algorithm can
be arbitrarily close to the optimal total utility with fixed
transmission power.
• The power control problem is a non-convex optimization
problem. We reformulate the problem as an integer programming problem. After relaxing the integer constraints,
we develop a distributed power control algorithm based
on the dual formulation. Interestingly, it can be shown that
one of the optimal solutions to the Lagrangian dual, when
fixing the dual variables, is always an integer solution.
That implies that the distributed algorithm derived from
the relaxed optimization problem produces a valid power
control decision and the relaxation is performed without
loss of optimality (more details can be found in Section
IV). Based on that, we prove that the time-average
total utility obtained under the proposed power control
algorithm can be arbitrarily close to the optimal total
utility with fixed broadcast rates.
The paper is organized as follows. In Section II, a utility
maximization framework is provided for congestion control in
DSRC. Then asymptotically optimal distributed rate control
algorithm and power control algorithm are derived respectively
in Section III and IV. In the end, simulation results are
presented in Section V to verify our algorithms. The proofs in
the paper are provided in the Appendix.
A. Discussion on Related Work
The design of rate and power control algorithms in DSRC
is one of most critical problems in ITS. Error Model Based
Adaptive Rate Control (EMBARC) [1] is a recent rate control protocol which integrates several existing rate control
algorithms including the Linear Integrated Message Rate
Control (LIMERIC) [6], Periodically Updated Load Sensitive
Adaptive Rate control (PULSAR) [12], and the InterVechicle
Transmission Rate Control (IVTRC) [13]. LIMERIC allocates
the wireless channel equally among all vehicles that share
the same bottleneck link while guaranteeing the channel
load is below a given threshold. IVTRC generates messages
and adapts transmission probabilities based on the Suspected
Tracking Error (STE) calculated based on vehicle dynamics to
avoid collisions. In EMBARC, the message rates are controlled
by LIMERIC and are further modified to satisfy the STE
requirement.
A parallel work [14] introduced a network utility maximization (NUM) formulation on the rate control problem when
specified to safety-awareness. A distributed algorithm was
proposed to adjust the rate with the objective to maximize
the utility function. Similarly, [15] also provided a NUM
formulation on the rate control problem and proposed a fair
adaptive beaconing rate for intervehicular communications
(FABRIC) algorithm, which essentially is a particular scaled
gradient projection algorithm to solve the dual of the NUM
problem.
Other related work includes the database approach proposed
in [9], where the optimal broadcast rates and transmission
power are calculated offline based on the network configurations. Also, [16] proposed an environment and context-aware
distributed congestion control (DCC) algorithm, which jointly
control the rate and power to improve cooperative awareness
by adapting to both specific propagation environments (such
as urban intersections, open highways, suburban roads) as well
as application requirements (e.g., different target cooperative
awareness range). However, the stability and convergence of
the algorithm are not proved mathematically. Besides the rate
control algorithm IVTRC, the authors also proposed range
control algorithms in [17], [18], [8] where the objective is to
adapt the transmission ranges to achieve a specific threshold.
The motivation of limiting channel loads below the threshold
is to control channel congestion to maximize effective channel
throughput. However, fair resource allocation among vehicles
to increase the safety awareness of all vehicles are not considered, and the stability of the algorithms is subject to certain
conditions [8].
II. P ROBLEM F ORMULATION
In this section, we formally define the utility maximization
framework for the DSRC congestion control problem. We first
introduce the set of notations used throughout this paper.
• µi : the message broadcast rate of vehicle i;
• pi : the transmit power of vehicle i;
• αij : the minimum transmit power required for node i’s
message to be decoded by node j;
• βij : the minimum transmit power required for node i’s
message to be sensed by node j, i.e. the received energy
is above the carrier sensing energy detection threshold;
• I : indicator function.
Note αij is not necessarily the same as βij , as in IEEE802.11
standards, packet header decoding happens at a much lower
energy level than energy based detection in carrier sensing.
From the definition of αij and βij , it is easy to see that
• Vehicle j can receive the message from vehicle i if pi ≥
αij ;
• Vehicle j can detect (but not necessarily decode) the
message from vehicle i if pi ≥ βij .
We assume αij and βij are constants within the time frame
for the distributed rate and power control algorithm, which is
reasonable as the nominal BSM update rate is 10Hz, i.e. 10
transmissions in every second.
In DSRC, a vehicle can control rate µi and power pi . We
consider the following utility maximization problem for rate
and power control:
P P
General − OPT maxµ,p i j Ipi ≥αij Uij (µi ) (1)
P
∀j
(2)
subject to:
i µi Ipi ≥βij ≤ γ
µi ≥ 0, pi ≥ 0
∀i.
(3)
Now we explain the particular choice of the objective
function and constraints above. In the objective function (1),
X
Ipi ≥αij Uij (µi )
j
is the total utility associated with vehicle i, which depends on
the number of vehicles who can receive the transmissions of
vehicle i, i.e., the size of the set
(4)
j : Ipi ≥αij = 1
3
and the utility function Uij (µi ) associated with each ordered
pair (i, j), which is a concave function and can be interpreted
as the level of j ′ s awareness of i when j receives messages
from i with rate µi . Obviously, higher transmission rate µi
should lead to higher value of Uij in DSRC. The neighborhood
size (4) is controlled by the transmit power pi and the value
of utility Uij (µi ) is determined by rate µi .
Remark 1. A widely used utility function [19] is called the
α-fair utility function which includes proportional fairness,
minimum potential-delay fairness as special cases and is given
by
µ1−αi
Ui (µi ) = wi i
, αi > 0,
(5)
1 − αi
where wi represents the weight of node i, determined by
its local information such as relative velocity, instantaneous
location in the application of vehicular network. Notice that
this utility function, given in a generic form, is independent
of communication links (from j). In other words, each vehicle
only knows its own utility function. As will be seen later, a
choice of such a form of utility function further simplifies the
proposed distributed algorithms because there is no need of
obtaining neighbors’ utility functions.
i
For αi = 2, the utility function turns to be Ui (µi ) = − w
µi
which implies weighted minimum potential delay fairness in
network’s resource allocation. For αi = 1, the utility function
behaves as Ui (µi ) = wi log(µi ) which leads to weighted
proportional fairness. We refer interested readers to [20] for
details.
The constraint (2) states that the channel load at any vehicle
j should be below a target channel load γ. In CSMA based
systems, high γ value indicates channel congestion, which
implies high packet collision rate and low effective channel
throughout [17], [9]. In [17], the authors have observed that the
curve of information dissemination rate versus channel load
remains almost the same under different configurations. In [9],
the authors also found that the effective channel throughput is
maximized when the channel load is around 0.91 under various
settings. Thus, it is natural to impose such a constraint (2) to
limit the congestion level in the system.
A. Problem decomposition
General-OPT is difficult to solve because the objective
function (1) is not jointly convex in (µ, p). We therefore
separate the problem into rate control problem and power
control problem as defined below.
• Assume the transmit power is fixed at each vehicle. Then
we can define
Ri = {j : pi ≥ αij },
i.e., the set of vehicles who can receive the messages
from vehicle i, and
Ii = {j : pi ≥ βij },
i.e., the set of vehicles whose channel load can be affected
by vehicle i′ s transmissions. When transmit power pi is
•
fixed, both Ri and Ii are fixed. In this case, general-OPT
becomes the following Rate-OPT
P P
Rate − OPT : ρ = maxµ i j∈Ri Uij (µi ) (6)
P
subject to:
∀ j. (7)
i:j∈Ii µi ≤ γ
Assuming the broadcast rates are fixed, i.e., µi ’s are fixed,
OPT becomes the following Power-OPT:
P
Power − OPT : ρ = maxp i,j Ipi ≥αij Uij (µi )(8)
P
(9)
subject to:
i µi Ipi ≥βij ≤ γ.
B. Iterative joint rate and power control
In light of the above decompositions, a (suboptimal) solution of General-OPT can be obtained by iterating Rate-OPT
and Power-OPT. The initial set of rate or power parameters for
the iterative algorithm can be appropriately chosen according
to certain practical constraints. The stopping criterion at step k
is typically set to be ρ(k + 1) − ρ(k) ≤ ǫ for ǫ > 0. It is worth
noting that in each step of iterations the utility value ρ(k) is
non-decreasing and ρ(k) is bounded above for all ∀k, given
a well-defined utility function. Therefore, the convergence of
the iterative algorithm is guaranteed.
In the following sections, we will develop distributed algorithms to solve Rate-OPT and Power-OPT separately. The
optimal rate control algorithm directly follows from the welldeveloped network congestion control while the optimal power
control algorithm is innovative and rather technical.
III. R ATE C ONTROL A LGORITHM
In what follows, we study the rate control problem and
develop a distributed rate control algorithm that solves (6).
Note that Rate-OPT is similar to the network utility maximization (NUM) problem for the Internet congestion control
(see [11] for a comprehensive introduction of the NUM
problem for the Internet congestion control). Each vehicle i
may represent both a flow and a link onPthe Internet, and
µi is the data rate of flow i. Regarding j∈Ri Uij (µi ) as
the utility function of vehicle i, the objective is to maximize
the sum of user utilities. We may further say that flow i
uses link j when j ∈ Ii . Then constraint (7) is equivalent
to the link capacity constraint that requires the total data
rate on link j to be no more than the link capacity γ.
To this end, Rate-OPT can be viewed as a standard NUM
problem for the Internet congestion control. The distributed
rate control algorithm below is based on the dual congestion
control algorithm for the Internet congestion control [11],
which consists of rate control and congestion price update. The
congestion price update monitors the channel load of vehicle
j. The congestion price λj increases when the channel load
at vehicle j exceeds the threshold γ and decreases otherwise.
The rate control algorithm adapts the broadcast rate µi based
on the aggregated congestion price from all vehicles who can
sense the transmissions from vehicle i, i.e., the vehicles whose
channel loads are affected by vehicle i.
Rate Control Algorithm
4
1) Rate control algorithm at vehilce i : At time slot t, vehicle
i broadcasts with rate µi [t] such that
µi [t]
= min
arg max
µ
X
X
Uij (µ) − ǫµ
j∈Ri
λj [t − 1], µmax
j∈Ri
(10)
where ǫ ∈ (0, 1] is a tuning parameter.
2) Congestion price update at vehicle j : At time slot t,
vehicle j updates its congestion price λj to be
+
X
λj [t] = λj [t − 1] +
µi [t − 1] − γ .
(11)
i:j∈Ii
This rate control algorithm is developed based on the dual
decomposition approach [11]. Specifically, the Lagrangian of
optimization (6) is
L(µi , λ)
=
X X
i
=
X
i
X
λj
Uij (µi ) − ǫµi
X
Uij (µi ) − ǫ
j
j∈Ri
X
j∈Ri
j∈Ii
X
i:j∈Ii
µi − γ
λj − γ
X
λj ,
j
min g(λ) = min max L(µi , λ)
λ
µi
When λ is fixed, the µi should maximize
X
X
Uij (µi ) − ǫµi
λj ,
j∈Ri
j∈Ii
which motivates the rate control algorithm (10). The congestion price update (11) is designed by taking derivative of g(λ)
over λ and then the optimal λ, as a mean to optimize the dual
problem, can be achieved by using a gradient search in (11).
The next theorem shows the rate control algorithm is
asymptotically optimal.
Theorem 2. Denote by µ∗i the optimal solution to problem (6)
and assume µmax > µ∗i for all i. Then there exists a constant
B > 0, independent of ǫ, such that under the proposed rate
control algorithm
lim inf
T →∞
1
T
T
−1 X
X
t=0
i
X
j∈Ii
Uij (µi [t]) ≥
XX
i
xij = Ipi ≥αij
∀ βik ≤ αij (15)
∀i, j
(16)
yij ∈ {0, 1}
∀i, j.
xij ≥ xik
yik ≥ xij
(17)
∀ αij ≤ αik (20)
∀ βik ≤ αij (21)
0 ≤ xij ≤ 1
0 ≤ yij ≤ 1
∀i, j
∀i, j.
(22)
(23)
Now by including constraint (19) in the Lagrangian, we
obtain
P
P
P
maxx,y i,j xij Uij (µi ) − ǫ j λj ( i yij µi − γ)
s.t.:
xij ≥ xik ∀ αij ≤ αik
yik ≥ xij ∀ βik ≤ αij
0 ≤ xij ≤ 1 ∀i, j
0 ≤ yij ≤ 1
j∈Ii
The proof of the theorem is similar to the proof of Theorem
6.1.1 in [11], and is omitted in thisP
paper. Remark that if the
objective function utility function j∈Ri Uij (µi ) is strictly
concave, the optimal solution of Rate-OPT is unique since the
search space is convex. As a consequence, the above algorithm
converges to the unique optimal solution.
yik ≥ xij
xij ∈ {0, 1}
Recall αij is the minimum transmit power for vehicle j to
receive messages from vehicle i. So constraint (14) states
that if vehicle j requires a smaller minimum transmit power
of vehicle i than vehicle k, then vehicle j can receive from
vehicle i if vehicle k can do so. Constraint (15) is similarly
defined.
Next, we relax the integer constraints (16) and (17) to obtain
the following linear programming problem.
P
maxx,y i,j xij Uij (µi )
(18)
P
∀j
(19)
subject to:
i yij µi ≤ γ
Uij (µ∗i ) − Bǫ.
and yij = Ipi ≥βij .
The Power-OPT problem is equivalent to the following integer
programming problem:
P
(12)
maxx,y i,j xij Uij (µi )
P
∀j
(13)
subject to:
i yij µi ≤ γ
xij ≥ xik
∀ αij ≤ αik (14)
where ǫ is a tuning parameter. Then the dual problem is
λ
IV. P OWER C ONTROL A LGORITHM
In this section, we develop a distributed power control algorithm that solves (8). The power control problem is developed
by formulating the Power-OPT as an integer programming
problem. After relaxing the integer constraint, we develop a
distributed power control algorithm using the dual approach.
Interestingly, it turns out the solution obtained from the
Lagrangian dual is always an integer solution. In other words,
the power control algorithm based on the linear approximation
always gives a valid power control solution and is proved to
be asymptotically optimal for Power-OPT.
We first introduce new variables x and y such that
∀i, j.
where ǫ is a tuning parameter. Note that constraints (20)
and (21) impose conditions on x and y related to the same
transmitter i. Therefore, given λ, the Lagrangian dual problem
above can be decomposed into the sub-problems for each given
i:
P
(24)
maxx,y j xij Uij (µi ) − ǫλj µi yij
subject to:
xij ≥ xik
∀ αij ≤ αik
5
yik ≥ xij
∀ βik ≤ αij
0 ≤ xij ≤ 1 ∀j
0 ≤ yij ≤ 1 ∀j.
Next we will show that one of the optimizers to the problem
(24) is an integer solution. For a fixed vehicle i, we sort the
vehicles in a descendent order according to αij and divide
them into groups, called G-groups and denoted by Gg , such
that αij = αik if j, k ∈ Gg , and αij < αik if j ∈ Gg and
k ∈ Gg+1 . Associated with each group Gg , we define α̃g to
be common α in the group. We further define H-groups
Hg = {m : α̃g−1 < βim ≤ α̃g }.
This is the set of vehicles that can sense the transmission
of vehicle i when the transmit power is α̃g and cannot if
the power is α̃g−1 . Furthermore, let g(j) denote the G-group
vehicle j is in and h(j) the H-group vehicle j is in. The
following lemma proves that one of the optimal solution to
(24) is an integer solution. The proof is presented in the
Appendix.
1. Note that the optimization problem can be further simplified
for specific utility functions, e.g., Ui,j (µi ) = wi log(µi ).
Based on the discussion and lemma above, we develop the
following power control algorithm, which consists of congestion price update and power update. The congestion price
update monitors the channel load and the power update adapts
the transmission power pi based on the aggregated congestion
price from all vehicles who can sense the transmissions from
vehicle i.
Power Control Algorithm
1) Power control at vehicle i : Vehicle i chooses the transmission power to be
pi [t + 1] = α̃gi′ ,
where gi′ is defined in Lemma 3 with λ = λ[t].
2) Congestion price update at vehicle j :
+
X
λj [t + 1] = λj [t] +
µi − γ .
(25)
(26)
i:j∈Ii
Lemma 3. Given λ, one of optimizers to optimization problem
(24) for given vehicle i is the following integer solution
1, if g(j) ≤ gi′
1, if h(j) ≤ gi′ Remark 4. Notice that the second step of the power control,
xij =
and yij =
, congestion price update, is identical to that in the rate control.
0, otherwise.
0, otherwise.
P
In practice, the value of i:j∈Ii µi can be approximated by
where
measured/sensed channel load of individual vehicle2 . Further
more, as shown in Lemma 3, the congestion prices of vehicles
X
Uij (µi )
gi′ = max g :
in the sensing range Ii are required in the power control (25)
j∈∪q:k≤q≤g Gq
while
only the prices of vehicles in the receiving range Ri are
needed
in the rate control. Unlike the price acquisition in the
X
receiving range, which can be piggybacked in the broadcasted
λj µi > 0 ∀0 ≤ k ≤ g .
−ǫ
BSM, the price information of vehicles in the sensing range
j∈∪q:k≤q≤g Hq
cannot be decoded. The approach of obtaining congestion
prices in the sensing range is not discussed in this paper since
it is rather implementation-specific and out of the scope of this
Algorithm 1 Sample algorithm for Lemma 3
paper.
Input: gmax , λj .
The next theorem shows the asymptotic optimality of the
Output: gi′
P
P
proposed
distributed power control algorithm.
1: Define fp =
j∈Hp λj µi , p =
j∈Gp Uij (µi ) − ǫ
1, 2, · · · , gmax .
Theorem 5. Denote by p∗ the optimal solution to Power-OPT.
2: for all g ∈ {gmax , gmax − 1, · · · , 1} do
There exists a constant M > 0, independent of ǫ, such that
3:
k ← g and f lag ← 1.
under the proposed power control algorithm,
4:
while f lag
=
1
and
k
≥
0
do
Pg
T −1
X
1 XX
5:
Fg ← p=k fp ;
lim inf
Ip∗i ≥αij Uij (µi ) − ǫM.
Ipi [t]≥αij Uij (µi ) ≥
6:
if Fg > 0 then
T →∞ T
t=0 i,j
i,j
7:
k ←k−1
8:
else
9:
f lag = 0
10:
end if
V. P ERFORMANCE E VALUATION USING NS 2
11:
end while
In this section, we evaluate the performance of the dis12:
if f lag = 1 then
tributed
rate and power control algorithm developed in this
13:
gi′ ← g; break;
paper,
and
compare the performance with EMBARC [1]. We
14:
end if
used
the
ns2
platform to simulate the asynchronous IEEE
15: end for
802.11p media access control algorithm with the associated
The optimization in Lemma 3 can be solved by low complexity algorithms. A sample algorithm is given in Algorithm
2 The channel load of DSRC is measured by carrier sensing technique which
is widely implemented in CSMA network
6
Fig. 1: Deployment of 6-lane highway vehicles
Number of packet delivered per second
7
5
4
3
2
1
0
0
1
EMBARC
Joint rate&power control
0.8
Channel Load
EMBARC
Rate control
Joint rate&power control
6
50
100
Distance (m)
150
200
Fig. 3: Number of successful received packets per second v.s.
distance between transmitter and receiver
0.6
12
10
0.4
Rate
8
6
4
0.2
0
2
4
6
8
Time (s)
10
12
2
14
0
0
200
400
600
800
1000 1200
Location (m)
1400
1600
1800
2000
0
200
400
600
800
1000 1200
Location (m)
1400
1600
1800
2000
Fig. 2: Convergence of channel load
lower layer functions. The 802.11MacExt package in ns2 is
adopted.
To simulate congestion at high densities, we constructed a
6-lane scenario where each lane is 4 meters wide and 2000
meters long. We use a wrap-around model of a network along
the length of the road (see Figure 1). In each lane, 300 vehicles
are deployed in a dense-sparse-dense-sparse fashion as a grid.
Specifically, the first 120 vehicles are spaced with either 4 or
5 meters distance between any adjacent vehicles. Similarly,
the next 30 vehicles are spaced with either 16 or 17 meters
distance. The last 150 vehicles are deployed by following the
same rule. A comprehensive list of simulation parameters is
summarized in Table I.
number of vehicles
packet size
carrier frequency
noise floor
carrier sense threshold
contention window
transmission rate
carrier sensing period
1800
300 Byte
5.9 GHz
-96 dBm
-76 dBm
15
6 Mbps
0.25 s
TABLE I: Simulation Parameters
We now briefly review the EMBRAC algorithm, of which
the transmission rate is a function of both channel load
(using LIMERIC component) and vehicle dynamics (using the
suspected tracking error component). In our ns2 simulations,
we did not consider vehicle dynamics and assumed that the
relative positions of the vehicles are static, which can be
justified using a time-scale separation assumption under which
the dynamics of the rate and power control algorithms are at
Radius (m)
300
200
100
0
Fig. 4: Broadcast rates and transmission ranges of the vehicles
in the first lane under the joint rate and power control
algorithm
a much faster time scale than the relative dynamics of the
vehicles. Therefore, the suspected tracking error component
of EMBARC was not simulated and EMBARC turns to be
LIMERIC. According to [1], LIMERIC is a distributed and
linear rate-control algorithm and the rate of vehicle i is
evolving as follows,
ri (t) = (1 − α)ri (t − 1) + β(rg − rc (t − 1)),
(27)
where rc is the total rate of all the K vehicles and rg is the
target total rate. α and β are parameters that tunes stability,
fairness and steady state convergence. In EMBARC, however,
rc is defined to be the maximum channel load reported by all
the 2-hop neighbors in order to achieve global fairness [1].
For the implementation of our rate and power control
algorithm, the sum rate from the interfering vehicles in the
congestion price update equations (11) and (26) can be replaced by the measured channel load at vehicle j. Therefore,
each vehicle only needs to piggyback its congestion price
in the safety message in broadcasting. Further, we chose the
7
following utility function for evaluation
Uij (µi ) =
max{vij , α}
log µi .
dij
Awareness
0.7
(28)
Joint rate&power control
EMBARC
0.6
0.5
This specific choice of utility functions is motivated from
the collision avoidance perspective, which we explain in
Appendix. In simulations, the target channel load is set to
be 0.6.
0.4
0.3
0.2
0.1
A. Convergence to Target Channel Load
0
The evolving equation (27) shows that in steady state
LIMERIC converges to a value strictly smaller than rg [6].
In other words, the target channel load can not be achieved in
steady state. However, our algorithm leads to full utilization
of the target channel load. See Figure 2. Furthermore, our
algorithm converges less than 4 seconds while in EMBARC
oscillations still occur after 10 seconds.
0
50
100
150
200
Number of nodes
250
300
Fig. 5: Awareness distribution of joint rate and power control
Coverage
0.7
0.6
Joint rate&power control
EMBARC
0.5
B. Packet Reception Rate
0.4
We compare the number of successful received packets per
second between EMBARC (with α = 0.1 and β = 0.001)
and our joint congestion control algorithm, which motivates
the need of congestion control algorithms in DSRC. To be
fair with (rate-control only) EMBARC, we simulated our
standalone rate control algorithm as well. Figure 3 shows that:
0.3
1) our rate control algorithm performs uniformly better
than EMBARC because of full utilization of the target
channel load. Specifically, our rate control guarantees the
convergence of measured channel load of each vehicle to
the target channel load while EMBARC is proved to have
inevitable gap in its steady state (27);
2) the joint congestion control algorithm provides significant
gain in short distance regime (safety-sensitive zone). This
is because both rate and transmission range are adjusted
according to the deployment topology, as shown in Figure
4. Specifically, the transmission range increases in the
sparse segments and achieves maximum at the center
vehicle, while the range is constantly short in the dense
segments. Note that 80% vehicles have short range, e.g.,
50m, which leads to the performance gain in the shortrange regime.
C. Coverage and Awareness
Figures 5 shows the distribution of the number of vehicles
that a vehicle can receive messages from, called awareness.
Figure 6 shows the distribution of the number of vehicles
within a vehicle’s transmission range, called coverage. Under
EMBARC, there are two peaks at 35 and 145 in both coverage
and awareness, respectively associated with two different
densities in the network. The joint algorithm only has one
peak since the algorithm dynamically allocates the resources
based on the network topology, achieving fairness in terms of
both coverage and awareness.
0.2
0.1
0
0
50
100
150
200
Number of nodes
250
300
350
Fig. 6: Coverage distribution of joint rate and power control
VI. C ONCLUSIONS
In this paper, we proposed a utility maximization framework
for joint rate and power control in DSRC, and formulated
two optimization problems, named Rate-OPT and PowerOPT, where Rate-OPT deals with rate control with fixed
transmit power and Power-OPT deals with power control
with fixed rates. We developed both distributed rate control
and power control algorithms and proved the algorithms are
asymptotically optimal. Evaluations using ns2 showed that our
algorithms outperform EMBARC at several relevant aspects
including channel utilization, packet reception rate, coverage
and awareness.
R EFERENCES
[1] G. Bansal, H. Lu, J. B. Kenney, and C. Poellabauer, “EMBARC: error
model based adaptive rate control for vehicle-to-vehicle communications,” in Proc. ACM Int. Workshop on Vehicular Inter-Networking,
Systems, and Applications (VANET), Taipei, Taiwan, 2013, pp. 41–50.
[2] J. B. Kenney, “Dedicated short-range communications (DSRC) standards
in the United States,” in Proc. IEEE, vol. 99, no. 7, pp. 1162 –1182,
Jul. 2011.
[3] IEEE 802.11p Part 11: Wireless LAN Medium Access Control (MAC)
and Physical Layer (PHY) Specifications, Amendment 6: Wireless Access
in Vehicular Environments, 2007.
[4] T. V. Nguyen, F. Baccelli, K. Zhu, S. Subramanian, and X. Wu, “A
performance analysis of csma based broadcast protocol in vanets,” in
in Proc. of IEEE Conference on Computer Communications, 2013, pp.
2805–2813.
8
[5] ETSI TS 102 687: Decentralized Congestion Control Mechanisms for
Intelligent Transport Systems operating in the 5 GHz range; Access
layer part, 2011.
[6] G. Bansal, J. Kenney, and C. Rohrs, “Limeric: A linear adaptive message
rate algorithm for dsrc congestion control,” IEEE Trans. Veh. Technol.,
vol. 62, no. 9, pp. 4182–4197, 2013.
[7] C.-L. Huang, Y. Fallah, R. Sengupta, and H. Krishnan, “Intervehicle
transmission rate control for cooperative active safety system,” IEEE
Trans. on Intelligent Transportation Systems, vol. 12, no. 3, pp. 645
–658, Sep. 2011.
[8] N. Nasiriani, Y. P. Fallah, and H. Krishnan, “Stability analysis of
congestion control schemes in vehicular ad-hoc networks,” in IEEE
Consumer Communications and Networking Conference (CCNC), Las
Vegas, NV, 2013, pp. 358–363.
[9] T. Tielert, D. Jiang, H. Hartenstein, and L. Delgrossi, “Joint power/rate
congestion control optimizing packet reception in vehicle safety communications,” in Proc. ACM Int. Workshop on Vehicular Inter-Networking,
Systems, and Applications (VANET), Taipei, Taiwan, 2013, pp. 51–60.
[10] J. Jose, C. Li, X. Wu, L. Ying, and K. Zhu, “Distributed rate and
power control in dsrc,” in Proc. of IEEE International Symposium on
Information Theory (ISIT), HongKong, China, 2015, pp. 2822–2826.
[11] R. Srikant and L. Ying, Communication Networks: An Optimization,
Control and Stochastic Networks Perspective. Cambridge University
Press, 2014.
[12] T. Tielert, D. Jiang, Q. Chen, L. Delgrossi, and H. Hartenstein, “Design
methodology and evaluation of rate adaptation based congestion control
for vehicle safety communications,” in Proc. IEEE Vehicular Networking
Conf. (VNC), Amsterdam, The Netherlands, 2011, pp. 116–123.
[13] C.-L. Huang, Y. P. Fallah, R. Sengupta, and H. Krishnan, “Intervehicle
transmission rate control for cooperative active safety system,” IEEE
Trans. on Intelligent Transportation Systems, vol. 12, no. 3, pp. 645–
658, 2011.
[14] L. Zhang and S. Valaee, “Safety context-aware congestion control
for vehicular broadcast networks,” in Signal Processing Advances in
Wireless Communications (SPAWC), 2014, pp. 399–403.
[15] Egea-Lopez, Esteban, and P. Pavon-Mariño, “Distributed and fair beaconing congestion control schemes for vehicular networks,” arXiv
preprint arXiv:1407.6965, 2014.
[16] B. Aygün, M. Boban, and A. M. Wyglinski, “Ecpr: Environment- and
context-aware combined power and rate distributed congestion control
for vehicular communications,” arXiv preprint arXiv:1502.00054, 2015.
[17] Y. P. Fallah, C.-L. Huang, R. Sengupta, and H. Krishnan, “Congestion
control based on channel occupancy in vehicular broadcast networks,”
in Proc. IEEE Vehicular Technology Conf., Ottawa, Canada, 2010, pp.
1–5.
[18] C.-L. Huang, Y. P. Fallah, R. Sengupta, and H. Krishnan, “Adaptive
intervehicle communication control for cooperative safety systems,”
IEEE Network, vol. 24, no. 1, pp. 6–13, 2010.
[19] J. Mo and J. Walrand, “Fair end-to-end window-based congestion
control,” IEEE/ACM Transactions on Networking, vol. 8, no. 5, pp. 556–
567, 2000.
[20] R. Srikant, The mathematics of Internet congestion control. Springer
Science & Business Media, 2012.
A PPENDIX
A. Proof of Lemma 3
Recall that the optimization problem (24) is for fixed i and
λ. Define
X
xij Uij (µi ) − ǫλj µi yij .
Fi (x, y) =
j
Let (x̂, ŷ) denote an optimal solution. We first have the
following two observations:
• According to constraint (20), x̂ij = x̂ik if dij = dik ,
which implies that
x̂ij = x̂ik
•
j, k ∈ Gg .
if
(29)
To maximize the objective (24), y should be chosen as
small as possible. So
ŷik =
max
j:βik ≤αij
x̂ij .
(30)
Since x̂ij ≥ x̂ik when αij < αik , (30) is equivalent to
ŷik = x̂ij ′
where j ′ = arg
min
j:βik ≤αij
αij ,
which further implies that
ŷik = x̂ij
if k ∈ Hg and j ∈ Gg .
(31)
In other words, x̂ij and ŷik are equal if j ∈ Gg and
k ∈ Hg for the same g. This is easy to understand because
the define of H-group Hg is the set of vehicles that can
sense the transmission of vehicle i when the vehicles in
G-group Gg can receive messages from vehicle i.
Now suppose that x̂ij is not an integer solution. Initially,
˜ = x̂. Identify vehicle j ′ is the vehicle has the maximum
let x̂
αij among all vehicles such that 0 < x̂ij < 1, i.e.,
j ′ = arg
max
j:0<x̂ij <1
αij .
According to observations (29) and (31), we have x̂ij = ŷik =
x̂ij ′ for j ∈ Gg (j ′ ) and k ∈ Hg(j′) . Therefore,
X
X
x̂ij Uij (µi ) − ǫ
λk µi ŷik
j∈Gg(j′ )
If
=
k∈Hg(j′ )
X
Uij (µi ) − ǫ
j∈Gg(j′ )
X
Uij (µi ) −
j∈Gg(j′ )
X
k∈Hg(j′ )
X
λk µi x̂ij ′ .
λk µi ≤ 0,
k∈Hg(j′ )
then we define x̂˜ij = ŷ˜ik = 0 for j ∈ Gg (j ′ ) and k ∈ Hg(j′) .
˜ij = ŷ˜ik = x̂ib for j ∈ Gg (j ′ ) and
Otherwise, we define x̂
k ∈ Hg(j′) where b ∈ Gg(j ′ )−1 . It is easy to see that the
following inequality holds:
˜
˜ ŷ).
Fi (x̂, ŷ) ≤ Fi (x̂,
Now for the second scenario discussed above, we have
X
X
˜ij − ǫ
Uij (µi )x̂
λk µi ŷ˜ik
j∈Gg(j′ ) ∪Gg(j′ )−1
k∈Hg(j′ ) ∪Hg(j′ )−1
9
P
2
Note that ( i yij [t]µi − γ) ≤ (nµmax + γmax )2 , where n is
=
Uij (µi ) − ǫ
λk µi x̂ibthe
. maximum number of vehicles whose transmissions can be
j∈Gg(j′ ) ∪Gg(j′ )−1
k∈Hg(j′ ) ∪Hg(j′ )−1
sensed by vehicle j, P
since 0 ≤ yij [t] ≤ 1 and 0 ≤ γ ≤ γmax .
2
′
Defining
M
=
˜
˜
i (γmax + nµmax ) and denoting by
Similarly, we define x̂ij = ŷik = 0 for j ∈ Gg(j ′ ) ∪ Gg(j ′ )−1
∗ ∗
(x̃ , ỹ ) the optimal solution to problem (18), we have
and k ∈ H
∪H
if
X
X
g(j′)
g(j′)−1
X
X
Uij (µi ) − ǫ
j∈Gg(j′ ) ∪Gg(j′ )−1
∆V [t] =
λk µi ≤ 0;
k∈Hg(j′ ) ∪Hg(j′ )−1
≤
˜ij = ŷ˜ik = x̂ic for j ∈ Gg(j ′ ) ∪ Gg(j ′ )−1
and otherwise define x̂
and k ∈ Hg(j′) ∪ Hg(j′)−1 , where c ∈ Gg(j′)−2 . Similarly, we
also have
˜
˜ ŷ).
Fi (x̂, ŷ) ≤ Fi (x̂,
=
j∈∪g:g≤g′ Gg
−ǫ
Uij (µi )
λj µi > 0 ∀k < g
B. Proof of Theorem 5
P
Defining V [t] = j λ2j [t], we have
.
≤
X
λj [t] +
=
λj [t] +
=
X
2λj [t] +
X
j
X
!2
yij [t]µi − γ
i
j
=
yij [t]µi − γ
i
j
X
I(pi [t]≥βij µi − γ
i
j
X
X
2λj [t]
X
i
yij [t]µi − γ
!
!
+
−
−
X
λ2j [t]
j
X
Note that
λ2j [t]
X
yij [t]µi − γ
!
X
yij [t]µi − γ
!2
i
PT −1
≤
∆V [t] = V [T ] − V [0], so
P
′
∗
(V [T ] − V [0]) − ǫM
i,j Uij (µi )x̃ij
2 +
P
P
T −1
1
t=0
i,j Uij (µi )xij [t],
T
t=0
ǫ
2T
j
i
!
which implies
X
X
ǫ
Uij (µi )xij [t],
Uij (µi )x̃∗ij ≤
(∆V [t] − M ′ ) +
2
i,j
i,j
V [t + 1] − V [t]
!2
.
(34)
(35)
(36)
According to Lemma 3, (33) + (34) ≤ 0. Further, since
(x̃∗ , ỹ ∗ ) is a feasible solution to problem (18), we have (36) ≤
0. Hence, we conclude
2X
2X
Uij (µi )x̃∗ij +
Uij (µi )xij [t],
∆V [t] ≤ M ′ −
ǫ i,j
ǫ i,j
∆V [t]
=
−γ
i
!
i
j
j∈∪q:k≤q≤g Hg
λj [t]
∗
ỹij
µi
X
X
2X
∗
ỹij
µi − γ
λi [t]
Uij (µi )x̃∗ij − 2
ǫ i,j
i
j
2X
2X
−
Uij (µi )x̃∗ij +
Uij (µi )xij [t]
ǫ i,j
ǫ i,j
!
X
X
∗
ỹij
µi − γ .
λi [t]
+2
+
j∈∪q:k≤q≤g Gg
X
X
(33)
where
X
X
∆V [t] ≤ M ′
(32)
!
X
X
2X
yij [t]µi − γ
λi [t]
Uij (µi )xij [t] + 2
−
ǫ i,j
i
j
From the discussion above, the integer optimizer is
1, if g(j) ≤ g ′
1, if h(j) ≤ g ′
xij =
and yij =
,
0, otherwise.
0, otherwise.
g:
i
By rearranging the terms above, we obtain
˜ ŷ).
˜
Fi (x̂, ŷ) ≤ Fi (x̂,
g ′ = min
yij [t]µi − γ
i
j
j
and x = 0 otherwise. Therefore, from any optimizer (x̂, ŷ) we
˜ ŷ)
˜ such that
can construct an integer solution (x̂,
λj [t]
!
2X
2X
Uij (µi )x̃∗ij −
Uij (µi )x̃∗ij
ǫ i,j
ǫ i,j
2X
2X
Uij (µi )xij [t] −
Uij (µi )xij [t]
+
ǫ i,j
ǫ i,j
!
X
X
∗
ỹij µi − γ
λj [t]
+2
−2
j∈∪g:g≤g′ Hg
i
X
+
j∈∪g:g≤g′ Hg
It is easy to see that we should choose x = 1 if
X
X
Uij (µi ) − ǫ
λj µi > 0.
M′ + 2
j
X
j
Repeating the same argument, we can conclude that there
˜ij = x if
exists g ′ such that x̂˜ij = 0 if g(j) > g ′ and x̂
′
g(j) ≤ g . Therefore,
X
X
˜ ŷ)
˜ =
Fi (x̂,
Uij (µi ) − ǫ
λj µi x.
j∈∪g:g≤g′ Gg
V [t + 1] − V [t]
!
X
X
′
yij [t]µi − γ
λj [t]
M +2
which implies
.
−
T −1
1 X X ηij
ǫV [0] ǫM ′ X
Uij (µi )x̃∗ij ≤
xij [t].
−
+
2T
2
T t=0 i,j dij
i,j
10
The theorem holds by choosing M = M ′ /2 and because
X
X
Uij (µi )x∗ij .
Uij (µi )x̃∗ij ≥
i,j
i,j
C. Selection of Utility Function (28)
Note that given distance dij and relative speed vij , vehicles
dij
i and j on a line would collide after vij
units of time if they
d
ij
do not change their speeds. We therefore call vij
reaction time
of pair (i, j). To prevent the collision, vehicles i and j need to
communicate at least once during this reaction time. Assume
each message is reliably received with probability p, then the
probability that vehicle j receives at least one message from
vehicle i during the reaction time is
dij
1 − (1 − p) vij
µi
.
Imposing a lower bound pmin on this probability is equivalent
to requiring that
dij
log(1 − pmin )
µi ≥
.
vij
log(1 − p)
(37)
These pairwise safety constraints may not always be feasible
depending on the network density and the geographical distribution of the vehicles.
Fig. 7: Vehicle 1 and vehicle 2 are both approaching vehicle
3 with different speeds
Therefore, we consider a different requirement also from the
collision avoidance perspective. Consider a scenario shown in
Figure 7, where both vehicle 1 and vehicle 2 are approaching
vehicle 3 with different speeds. A fair resource allocation
should equalize the reaction time to avoid collisions, i.e., to
have
d23
d13
µ1 =
µ2 .
(38)
v13
v23
In a general scenario, this objective could also be difficult
to achieve. We note that if we assume all vehicles share a
single-bottleneck link, e.g., in the scenario in Figure 7 where
all vehicles can hear each other, then solving the following
optimization problem
v23
v13
log µ1 +
log µ2
(39)
max
µ d13
d23
results in the solution
v23 1
v13 1
=
,
(40)
d13 µ1
d23 µ2
which is equivalent to (38). This motivated the objective function in (28), where each link dij is associated with a weighted
max{vij ,α}
log-utility
log µi . Since vij may be negative, we
dij
define the weight to be max{vij , α} for some α > 0.
| 3cs.SY
|
1
StackGAN++: Realistic Image Synthesis
with Stacked Generative Adversarial Networks
arXiv:1710.10916v2 [cs.CV] 25 Dec 2017
Han Zhang∗ , Tao Xu, Hongsheng Li, Shaoting Zhang, Senior Member, IEEE,
Xiaogang Wang, Member, IEEE, Xiaolei Huang, Member, IEEE, Dimitris N. Metaxas, Fellow, IEEE
Abstract—Although Generative Adversarial Networks (GANs) have shown remarkable success in various tasks, they still face
challenges in generating high quality images. In this paper, we propose Stacked Generative Adversarial Networks (StackGAN)
aiming at generating high-resolution photo-realistic images. First, we propose a two-stage generative adversarial network architecture,
StackGAN-v1, for text-to-image synthesis. The Stage-I GAN sketches the primitive shape and colors of the object based on given text
description, yielding low-resolution images. The Stage-II GAN takes Stage-I results and text descriptions as inputs, and generates highresolution images with photo-realistic details. Second, an advanced multi-stage generative adversarial network architecture, StackGANv2, is proposed for both conditional and unconditional generative tasks. Our StackGAN-v2 consists of multiple generators and
discriminators in a tree-like structure; images at multiple scales corresponding to the same scene are generated from different branches
of the tree. StackGAN-v2 shows more stable training behavior than StackGAN-v1 by jointly approximating multiple distributions.
Extensive experiments demonstrate that the proposed stacked generative adversarial networks significantly outperform other stateof-the-art methods in generating photo-realistic images.
Index Terms—Generative models, Generative Adversarial Networks (GANs), multi-stage GANs, multi-distribution approximation,
photo-realistic image generation, text-to-image synthesis.
F
1
I NTRODUCTION
Generative Adversarial Network (GAN) is a generative model
proposed by Goodfellow et al. [11]. In the original setting,
GAN is composed of a generator and a discriminator that
are trained with competing goals. The generator is trained to
generate samples towards the true data distribution to fool the
discriminator, while the discriminator is optimized to distinguish between real samples from the true data distribution and
fake samples produced by the generator. Recently, GAN has
shown great potential in simulating complex data distributions,
such as those of texts [5], images [28] and videos [44].
Despite the success, GAN models are known to be difficult
to train. The training process is usually unstable and sensitive
to the choices of hyper-parameters. Several papers argued that
the instability is partially due to the disjoint supports of the
data distribution and the implied model distribution [38], [1].
• H. Zhang is with the Department of Computer Science, Rutgers University,
Piscataway, NJ, 08854. E-mail: [email protected]
• T. Xu is with the Department of Computer Science and Engineering, Lehigh
University, Bethlehem, PA, 18015. E-mail: [email protected]
• H. Li is with the Department of Electronic Engineering, The Chinese University of Hong Kong, Shatin, N. T., Hong Kong. E-mail:
[email protected]
• S. Zhang is with Baidu Research, Beijing, China. E-mail: [email protected]
• X. Wang is with the Department of Electronic Engineering, The Chinese University of Hong Kong, Shatin, N. T., Hong Kong. E-mail: [email protected]
• X. Huang is with the Department of Computer Science and Engineering,
Lehigh University, Bethlehem, PA, 18015. E-mail: [email protected]
• D. N. Metaxas is with the Department of Computer Science, Rutgers
University, Piscataway, NJ, 08854. E-mail: [email protected]
The first two authors contributed equally to this work. Asterisk indicates the
corresponding author.
This problem is more severe when training GAN to generate
high-resolution (e.g., 256×256) images because the chance is
very rare for the high-resolution image distribution and the
model distribution to share supports in a high-dimensional
space. Moreover, a common failure phenomenon for GAN
training is mode collapse: many of the generated samples
contain the same color or texture pattern.
In order to stabilize the GAN training process and improve sample diversity, several methods tried to address
the challenges by proposing new network architectures [28],
introducing heuristic tricks [36] or modifying the learning objectives [2], [4]. But most of the previous methods are designed
to approximate a single data distribution (e.g., images of the
same size). Due to the difficulty in directly approximating
the high-resolution image data distribution because of rare
overlap between model and data distributions in the very highdimensional space, most previous methods are limited to generating low-resolution images. In this work, we observe that,
real world data, especially natural images, can be modeled at
different scales [34]. One can view multi-resolution digitized
images as samplings from the same continuous image signal
with different sampling rates. Henceforth, the distributions
of images at multiple discrete scales are related. Apart from
multiple distributions of different scales, images coupled with
or without auxiliary conditioning variables (e.g., class labels
or text descriptions) can be viewed as conditional distributions or unconditional distributions, which are also related
distributions. Motivated by these observations, we argue that
GAN can be stably trained to generate high resolution images
by breaking the difficult generative task into sub-problems
with progressive goals, i.e., we propose Stacked Generative
Adversarial Networks (StackGAN) to model a series of low-
2
to-high-dimensional data distributions.
First, we propose a two-stage generative adversarial network, StackGAN-v1, to generate images from text descriptions through a sketch-refinement process [50]. Low-resolution
images are first generated by our Stage-I GAN. On top of
the Stage-I GAN, we stack Stage-II GAN to generate highresolution (e.g., 256×256) images. By conditioning on the
Stage-I result and the text again, Stage-II GAN learns to capture the text information that is omitted by Stage-I GAN and
draws more details. Further, we propose a novel Conditioning
Augmentation (CA) technique to encourage smoothness in
the latent conditioning manifold [50]. It allows small random
perturbations in the conditioning manifold and increases the
diversity of synthesized images.
Second, we propose an advanced multi-stage generative
adversarial network architecture, StackGAN-v2, for both conditional and unconditional generative tasks. StackGAN-v2 has
multiple generators that share most of their parameters in a
tree-like structure. As shown in Figure 2, the input to the
network can be viewed as the root of the tree, and multiscale images are generated from different branches of the
tree. The generator at the deepest branch has the final goal of
generating photo-realistic high-resolution images. Generators
at intermediate branches have progressive goals of generating
small to large images to help accomplish the final goal. The
whole network is jointly trained to approximate different but
highly related image distributions at different branches. The
positive feedback from modeling one distribution can improve
the learning of others. For conditional image generation tasks,
our proposed StackGAN-v2 simultaneously approximates the
unconditional image-only distribution and the image distribution conditioned on text descriptions. Those two types of
distributions are complementary to each other. Moreover, we
propose a color-consistency regularization term to guide our
generators to generate more coherent samples across different
scales. The regularization provides additional constraints to
facilitate multi-distribution approximation, which is especially
useful in the unconditional setting where there is no instancewise supervision between the image and the input noise
vector.
In summary, the proposed Stacked Generative Adversarial
Networks have three major contributions. (i) Our StackGANv1 for the first time generates images of 256×256 resolution with photo-realistic details from text descriptions.
(ii) A new Conditioning Augmentation technique is proposed to stabilize the conditional GANs’ training and also
improves the diversity of the generated samples. (iii) Our
StackGAN-v2 further improves the quality of generated
images and stabilizes the GANs’ training by jointly approximating multiple distributions. In the remainder of this
paper, we first discuss related work and preliminaries in
section 2 and section 3, respectively. We then introduce
our StackGAN-v1 [50] in section 4 and StackGAN-v2 in
section 5. In section 6, extensive experiments are conducted to evaluate the proposed methods. Finally, we make
conclusions in section 7. The source code for StackGANv1 is available at https://github.com/hanzhanggit/StackGAN,
and the source code for StackGAN-v2 is available at
https://github.com/hanzhanggit/StackGAN-v2.
2
R ELATED W ORK
Generative image modeling is a fundamental problem in
computer vision. There has been remarkable progress in this
direction with the emergence of deep learning techniques.
Variational Autoencoders (VAEs) [17], [33] formulated the
problem with probabilistic graphical models whose goal was to
maximize the lower bound of data likelihood. Autoregressive
models (e.g., PixelRNN) [40] that utilized neural networks
to model the conditional distribution of the pixel space
have also generated appealing synthetic images. Recently,
Generative Adversarial Networks (GANs) [11] have shown
promising performance for generating sharper images. But
training instability makes it hard for GAN models to generate
high-resolution (e.g., 256×256) images. A lot of works have
been proposed to stabilize the training and improve the image
qualities [28], [36], [23], [51], [4], [25].
Built upon these generative models, conditional image
generation has also been studied. Most methods utilized simple
conditioning variables such as attributes or class labels [47],
[41], [6], [27]. There is also work conditioned on images to
generate images, including photo editing [3], [52], domain
transfer [39], [15] and super-resolution [38], [19]. However,
super-resolution methods [38], [19] can only add limited details to low-resolution images and can not correct large defects
as our proposed StackGAN does. Recently, several methods
have been developed to generate images from unstructured
text. Mansimov et al. [21] built an AlignDRAW model by
learning to estimate alignment between text and the generating canvas. Reed et al. [32] used conditional PixelCNN to
generate images using the text descriptions and object location
constraints. Nguyen et al. [25] used an approximate Langevin
sampling approach to generate images conditioned on text.
However, their sampling approach requires an inefficient
iterative optimization process. With conditional GAN, Reed
et al. [31] successfully generated plausible 64×64 images for
birds and flowers based on text descriptions. Their follow-up
work [29] was able to generate 128×128 images by utilizing
additional annotations on object part locations.
Given the difficulties of modeling details of natural images,
many works have been proposed to use multiple GANs to
improve the sample quality. Denton et al. [7] built a series of
GANs within a Laplacian pyramid framework. At each level
of the pyramid, a residual image was generated conditioned
on the image of the previous stage and then added back
to the input image to produce the input for the next stage.
Wang et al. [46] utilized a structure GAN and a style GAN
to synthesize images of the indoor scenes. Yang et al. [48]
factorized image generation into foreground and background
generation with layered recursive GANs. Concurrent to our
work, Huang et al. [13] added several GANs to reconstruct
the multi-level representations of a pre-trained discriminative
model. But they were still unable to generate high resolution
image with photo-realistic details. Durugkar et al. [9] used
multiple discriminators along with one generator to increase
the chance of the generator receiving effective feedback.
3
However, all discriminators in their framework are trained
to approximate a single data distribution instead of multiple
distributions.
3
P RELIMINARIES
Generative Adversarial Network (GAN) [11] is composed of
two models that are alternatively trained to compete with each
other. The generator G is optimized to reproduce the true data
distribution pdata by generating images that are difficult for the
discriminator D to differentiate from real images. Meanwhile,
D is optimized to distinguish real images and synthetic images
generated by G. Overall, the training procedure is similar
to a two-player min-max game with the following objective
function,
min max V (D, G) = Ex∼pdata [log D(x)] +
G
D
Ez∼pz [log(1 − D(G(z)))],
(1)
where x is a real image from the true data distribution pdata ,
and z is a noise vector sampled from distribution pz (e.g.,
uniform or Gaussian distribution).
Conditional GAN [10], [24] is an extension of GAN
where both the generator and discriminator receive additional
conditioning variables c, yielding G(z, c) and D(x, c). This
formulation allows G to generate images conditioned on
variables c.
4 S TACK GAN- V 1: T WO - STAGE G ENERATIVE
A DVERSARIAL N ETWORK
To generate high-resolution images with photo-realistic details,
we propose a simple yet effective two-stage generative adversarial network, StackGAN-v1. As shown in Figure 1, it decomposes the text-to-image generative process into two stages.
Stage-I GAN sketches the primitive shape and basic colors of
the object conditioned on the given text description, and draws
the background layout from a random noise vector, yielding
a low-resolution image. Stage-II GAN corrects defects in the
low-resolution image from Stage-I and completes details of
the object by reading the text description again, producing a
high-resolution photo-realistic image.
4.1
Conditioning Augmentation
As shown in Figure 1, the text description t is first encoded
by an encoder, yielding a text embedding ϕt . In previous
works [31], [29], the text embedding is nonlinearly transformed to generate conditioning latent variables as the input
of the generator. However, latent space for the text embedding
is usually high dimensional (> 100 dimensions). With limited
amount of data, it usually causes discontinuity in the latent
data manifold, which is not desirable for learning the generator. To mitigate this problem, we introduce a Conditioning
Augmentation technique to produce additional conditioning
variables ĉ. In contrast to the fixed conditioning text variable
c in [31], [29], we randomly sample the latent variables ĉ
from an independent Gaussian distribution N (µ(ϕt ), Σ(ϕt )),
where the mean µ(ϕt ) and diagonal covariance matrix Σ(ϕt )
are functions of the text embedding ϕt . The proposed Conditioning Augmentation yields more training pairs given a small
number of image-text pairs, and thus encourages robustness to
small perturbations along the conditioning manifold. To further
enforce the smoothness over the conditioning manifold and
avoid overfitting [8], [18], we add the following regularization
term to the objective of the generator during training,
DKL (N (µ(ϕt ), Σ(ϕt )) || N (0, I)) ,
(2)
which is the Kullback-Leibler divergence (KL divergence)
between the standard Gaussian distribution and the conditioning Gaussian distribution. The randomness introduced in the
Conditioning Augmentation is beneficial for modeling text to
image translation as the same sentence usually corresponds to
objects with various poses and appearances.
4.2
Stage-I GAN
Instead of directly generating a high-resolution image conditioned on the text description, we simplify the task to first
generate a low-resolution image with our Stage-I GAN, which
focuses on drawing only rough shape and correct colors for
the object.
Let ϕt be the text embedding of the given description, which
is generated by a pre-trained encoder [30] in this paper. The
Gaussian conditioning variables ĉ0 for text embedding are
sampled from N (µ0 (ϕt ), Σ0 (ϕt )) to capture the meaning of
ϕt with variations. Conditioned on ĉ0 and random variable z,
Stage-I GAN trains the discriminator D0 and the generator G0
by alternatively maximizing LD0 in Eq. (3) and minimizing
LG0 in Eq. (4),
LD0 = E(I0 ,t)∼pdata [log D0 (I0 , ϕt )] +
Ez∼pz ,t∼pdata [log(1 − D0 (G0 (z, ĉ0 ), ϕt ))],
LG0 = Ez∼pz ,t∼pdata [log(1 − D0 (G0 (z, ĉ0 ), ϕt ))] +
λDKL (N (µ0 (ϕt ), Σ0 (ϕt )) || N (0, I)),
(3)
(4)
where the real image I0 and the text description t are from
the true data distribution pdata . z is a noise vector randomly
sampled from a given distribution pz (Gaussian distribution in
this paper). λ is a regularization parameter that balances the
two terms in Eq. (4). We set λ = 1 for all our experiments.
Using the reparameterization trick introduced in [17], both
µ0 (ϕt ) and Σ0 (ϕt ) are learned jointly with the rest of the
network.
Model Architecture. For the generator G0 , to obtain text
conditioning variable ĉ0 , the text embedding ϕt is first fed
into a fully connected layer to generate µ0 and σ0 (σ0 are
the values in the diagonal of Σ0 ) for the Gaussian distribution
N (µ0 (ϕt ), Σ0 (ϕt )). ĉ0 are then sampled from the Gaussian
distribution. Our Ng dimensional conditioning vector cˆ0 is
computed by cˆ0 = µ0 + σ0 (where
is the elementwise multiplication, ∼ N (0, I)). Then, cˆ0 is concatenated
with a Nz dimensional noise vector to generate a W0 × H0
image by a series of up-sampling blocks.
For the discriminator D0 , the text embedding ϕt is first
compressed to Nd dimensions using a fully-connected layer
and then spatially replicated to form a Md × Md × Nd
4
Conditioning
Augmentation (CA)
Text description t
Stage-I Generator G0
for sketch
μ0
Embedding ϕt
64 x 64
results
128
512
ĉ0
This bird is grey with
white on its chest and
has a very short beak
Downsampling
Upsampling
z ~ N(0, I)
σ0
Embedding ϕt
{0, 1}
4
4
Compression and
Spatial Replication
64 x 64
real images
ε ~ N(0, I)
Embedding ϕt
ĉ
Compression and
Spatial Replication
256 x 256
real images
Conditioning
Augmentation
64 x 64
Stage-I results
Stage-I Discriminator D0
Downsampling
128
512
16
Residual
blocks
Downsampling
Upsampling
128
512
{0, 1}
4
4
16
256 x 256
results
Stage-II Generator G for refinement
Stage-II Discriminator D
Fig. 1: The architecture of the proposed StackGAN-v1. The Stage-I generator draws a low-resolution image by sketching rough shape and
basic colors of the object from the given text and painting the background from a random noise vector. Conditioned on Stage-I results, the
Stage-II generator corrects defects and adds compelling details into Stage-I results, yielding a more realistic high-resolution image.
tensor. Meanwhile, the image is fed through a series of downsampling blocks until it has Md × Md spatial dimension.
Then, the image filter map is concatenated along the channel
dimension with the text tensor. The resulting tensor is further
fed to a 1×1 convolutional layer to jointly learn features across
the image and the text. Finally, a fully-connected layer with
one node is used to produce the decision score.
4.3
Stage-II GAN
Low-resolution images generated by Stage-I GAN usually lack
vivid object parts and might contain shape distortions. Some
details in the text might also be omitted in the first stage,
which is vital for generating photo-realistic images. Our StageII GAN is built upon Stage-I GAN results to generate highresolution images. It is conditioned on low-resolution images
and also the text embedding again to correct defects in Stage-I
results. The Stage-II GAN completes previously ignored text
information to generate more photo-realistic details.
Conditioning on the low-resolution result s0 = G0 (z, ĉ0 )
and Gaussian latent variables ĉ, the discriminator D and
generator G in Stage-II GAN are trained by alternatively
maximizing LD in Eq. (5) and minimizing LG in Eq. (6),
LD = E(I,t)∼pdata [log D(I, ϕt )] +
Es0 ∼pG0 ,t∼pdata [log(1 − D(G(s0 , ĉ), ϕt ))],
LG = Es0 ∼pG0 ,t∼pdata [log(1 − D(G(s0 , ĉ), ϕt ))] +
λDKL (N (µ(ϕt ), Σ(ϕt )) || N (0, I)) ,
(5)
(6)
Different from the original GAN formulation, the random
noise z is not used in this stage with the assumption that
the randomness has already been preserved by s0 . Gaussian
conditioning variables ĉ used in this stage and ĉ0 used
in Stage-I GAN share the same pre-trained text encoder,
generating the same text embedding ϕt . However, Stage-I
and Stage-II Conditioning Augmentation have different fully
connected layers for generating different means and standard
deviations. In this way, Stage-II GAN learns to capture useful
information in the text embedding that is omitted by Stage-I
GAN.
Model Architecture. We design Stage-II generator as an
encoder-decoder network with residual blocks [12]. Similar
to the previous stage, the text embedding ϕt is used to
generate the Ng dimensional text conditioning vector ĉ, which
is spatially replicated to form a Mg × Mg × Ng tensor.
Meanwhile, the Stage-I result s0 generated by Stage-I GAN
is fed into several down-sampling blocks (i.e., encoder) until
it has a spatial size of Mg × Mg . The image features and the
text features are concatenated along the channel dimension.
The encoded image features coupled with text features are
fed into several residual blocks, which are designed to learn
multi-modal representations across image and text features.
Finally, a series of up-sampling layers (i.e., decoder) are used
to generate a W × H high-resolution image. Such a generator
is able to help rectify defects in the input image while add
more details to generate the realistic high-resolution image.
For the discriminator, its structure is similar to that of StageI discriminator with only extra down-sampling blocks since
the image size is larger in this stage. To explicitly enforce
GAN to learn better alignment between the image and the
conditioning text, rather than using the vanilla discriminator,
we adopt the matching-aware discriminator proposed by Reed
et al. [31] for both stages. During training, the discriminator
takes real images and their corresponding text descriptions as
positive sample pairs, whereas negative sample pairs consist
of two groups. The first is real images with mismatched text
embeddings, while the second is synthetic images with their
corresponding text embeddings.
4.4
Implementation details
The up-sampling blocks consist of the nearest-neighbor upsampling followed by a 3×3 stride 1 convolution. Batch
normalization [14] and ReLU activation are applied after
every convolution except the last one. The residual blocks
consist of 3×3 stride 1 convolutions, Batch normalization and
5
ReLU. Two residual blocks are used in 128×128 StackGANv1 models while four are used in 256×256 models. The downsampling blocks consist of 4×4 stride 2 convolutions, Batch
normalization and LeakyReLU, except that the first one does
not have Batch normalization.
By default, Ng = 128, Nz = 100, Mg = 16, Md = 4,
Nd = 128, W0 = H0 = 64 and W = H = 256. For training,
we first iteratively train D0 and G0 of Stage-I GAN for 600
epochs by fixing Stage-II GAN. Then we iteratively train D
and G of Stage-II GAN for another 600 epochs by fixing
Stage-I GAN. All networks are trained using ADAM [16]
solver with batch size 64 and an initial learning rate of
0.0002. The learning rate is decayed to 1/2 of its previous
value every 100 epochs. The source code for StackGAN-v1
is available at https://github.com/hanzhanggit/StackGAN for
more implementation details.
5 S TACK GAN- V 2:
M ULTI - DISTRIBUTION
G ENERATIVE A DVERSARIAL N ETWORK
As discussed above, our StackGAN-v1 has two separate
networks, Stage-I GAN and Stage-II GAN, to model low-tohigh resolution image distributions. To make the framework
more general, in this paper, we propose a new end-to-end
network, StackGAN-v2, to model a series of multi-scale image
distributions. As shown in Figure 2, StackGAN-v2 consists of
multiple generators (Gs) and discriminators (Ds) in a tree-like
structure. Images from low-resolution to high-resolution are
generated from different branches of the tree. At each branch,
the generator captures the image distribution at that scale and
the discriminator estimates the probability that a sample came
from training images of that scale rather than the generator.
The generators are jointly trained to approximate the multiple
distributions, and the generators and discriminators are trained
in an alternating fashion. In this section, we explore two types
of multi-distributions: (1) multi-scale image distributions; and
(2) joint conditional and unconditional image distributions.
5.1
Multi-scale image distributions approximation
Our StackGAN-v2 framework has a tree-like structure, which
takes a noise vector z ∼ pnoise as the input and has multiple
generators to produce images of different scales. The pnoise
is a prior distribution, which is usually chosen as the standard
normal distribution. The latent variables z are transformed to
hidden features layer by layer. We compute the hidden features
hi for each generator Gi by a non-linear transformation,
h0 = F0 (z);
hi = Fi (hi−1 , z), i = 1, 2, ..., m − 1,
(7)
where hi represents hidden features for the ith branch, m is
the total number of branches, and Fi are modeled as neural
networks (see Figure 2 for illustration). In order to capture
information omitted in preceding branches, the noise vector z
is concatenated to the hidden features hi−1 as the inputs of Fi
for calculating hi . Based on hidden features at different layers
(h0 , h1 , ..., hm−1 ), generators produce samples of small-tolarge scales (s0 , s1 , ..., sm−1 ),
si = Gi (hi ), i = 0, 1, ..., m − 1,
(8)
where Gi is the generator for the ith branch.
Following each generator Gi , a discriminator Di , which
takes a real image xi or a fake sample si as input, is trained
to classify inputs into two classes (real or fake) by minimizing
the following the cross-entropy loss,
1
1
LDi = − Exi ∼pdatai [log Di (xi )] − Esi ∼pGi [log(1 − Di (si )],
2
2
(9)
where xi is from the true image distribution pdatai at the ith
scale, and si is from the model distribution pGi at the same
scale. The multiple discriminators are trained in parallel, and
each of them focuses on a single image scale.
Guided by the trained discriminators, the generators are optimized to jointly approximate multi-scale image distributions
(pdata0 , pdata1 , ..., pdatam−1 ) by minimizing the following loss
function,
LG =
m
X
i=1
LGi ,
LGi =
1
Es ∼p [log(1 − Di (si )] ,
2 i Gi
(10)
where LGi is the loss function for approximating the image
distribution at the ith scale (i.e., pdatai ). In practice, the
generator Gi is modified to maximize log(Di (si )) instead of
minimizing log(1−Di (si )) to mitigate the problem of gradient
vanishing [11]. During the training process, the discriminators
Di and the generators Gi are alternately optimized till convergence.
The motivation of the proposed StackGAN-v2 is that
jointly approximating multiple image distributions increases
the chance of these data distributions sharing supports with
model distributions. Adding auxiliary generation tasks to intermediate branches provides more gradient signals for training
the whole network. For instance, approximating the lowresolution image distribution at the first branch results in
images with basic color and structures. Then the generators
at the subsequent branches can focus on completing details
for generating higher resolution images.
5.2 Joint conditional and unconditional distribution
approximation
For unconditional image generation, discriminators in
StackGAN-v2 are trained to distinguish real images from fake
ones. To handle conditional image generation, conventionally,
images and their corresponding conditioning variables are
input into the discriminator to determine whether an imagecondition pair matches or not, which guides the generator to
approximate the conditional image distribution. We propose
conditional StackGAN-v2 that jointly approximates conditional and unconditional image distributions.
For the generator of our conditional StackGAN-v2, F0 and
Fi are converted to take the conditioning vector c as input,
such that h0 = F0 (c, z) and hi = Fi (hi−1 , c). For Fi , the
conditioning vector c replaces the noise vector z to encourage
the generators to draw images with more details according to
the conditioning variables. Consequently, multi-scale samples
are now generated by si = Gi (hi ). The objective function
of training the discriminator Di for conditional StackGANv2 now consists of two terms, the unconditional loss and the
6
FC with reshape
Upsampling
Joining
Residual
Conv3x3
Generators in a tree-like structure
JCU Discriminator
real
fake
OR
c
4x4
x64Ng
z~N(0,I)
64x64
x4Ng
128x128
x2Ng
G0
G1
64x64x3
256x256
xNg
G2
256x256x3
Downsampling
c
128x128x3
JCU
D0
JCU
D1
JCU
D2
Unconditional
loss
Conditional
loss
Fig. 2: The overall framework of our proposed StackGAN-v2 for the conditional image synthesis task. c is the vector of conditioning variables
which can be computed from the class label, the text description, etc.. Ng and Nd are the numbers of channels of a tensor.
conditional loss,
LDi
1
1
= − Exi ∼pdata [log Di (xi )] − Esi ∼pG [log(1 − Di (si )] +
i
i
2
2
|
{z
}
unconditional loss
1
1
− Exi ∼pdata [log Di (xi , c)] − Esi ∼pG [log(1 − Di (si , c)] .
i
i
2
2
|
{z
}
conditional loss
(11)
The unconditional loss determines whether the image is
real or fake while the conditional one determines whether the
image and the condition match or not. Accordingly, the loss
function for each generator Gi is converted to
LGi =
1
1
Es ∼p
[log(1 − Di (si )] + Esi ∼pG [log(1 − Di (si , c)] .
i
2 i Gi
2
|
{z
}
|
{z
}
unconditional loss
Empirically, we set λ1 = 1 and λ2 = 5 by default. For
the j th input vector, multi-scale samples sj1 , sj2 , ..., sjm are
generated from m generators of StackGAN-v2. LCi can be
added to the loss function of the ith generator defined in
Eq. (10) or Eq. (12), where i = 2, 3, ..., m. Therefore,
the final loss for training the ith generator is defined as
L0Gi = LGi + α ∗ LCi . Experimental results indicate that
the color-consistency regularization is very important (e.g.,
α = 50.0 in this paper) for the unconditional task, while
it is not needed (α = 0.0) for the text-to-image synthesis
task which has a stronger constraint, i.e., the instance-wise
correspondence between images and text descriptions.
conditional loss
(12)
The generator Gi at each scale therefore jointly approximates
unconditional and conditional image distributions. The final
loss for jointly training generators of conditional StackGANv2 is computed by substituting Eq. (12) into Eq. (10).
5.3 Color-consistency regularization
As we increase the image resolution at different generators,
the generated images at different scales should share similar
basic structure and colors. A color-consistency regularization
term is introduced to keep samples generated from the same
input at different generators more consistent in color and thus
to improve the quality of the generated images.
Let xk = (R, G, B)T represent a pixel in a generated
image, then the mean and covariance
of pixels of P
the given
P
image can be defined by µ = k xk /N and Σ = k (xk −
µ)(xk − µ)T /N , where N is the number of pixels in
the image. The color-consistency regularization term aims at
minimizing the differences of µ and σ between different scales
to encourage the consistency, which is defined as
LCi =
n
1 X
λ1 kµsj − µsj k22 + λ2 kΣsj − Σsj k2F , (13)
i
i
i−1
i−1
n j=1
where n is the batch size, µsj and Σsj are mean and
i
i
covariance for the j th sample generated by the ith generator.
5.4
Implementation details
As shown in Figure 2, our StackGAN-v2 models are designed to generate 256×256 images. The input vector (i.e.,
z for unconditional StackGAN-v2, or the concatenated z
and c for conditional StackGAN-v2) is first transformed
to a 4×4×64Ng feature tensor, where Ng is the number
of channels in the tensor. Then, this 4×4×64Ng tensor
is gradually transformed to 64×64×4Ng , 128×128×2Ng ,
and eventually 256×256×1Ng tensors at different layers of
the network by six up-sampling blocks. The intermediate
64×64×4Ng , 128×128×2Ng , and 256×256×1Ng features
are used to generate images of corresponding scales with
3×3 convolutions. Conditioning variables c or unconditional
variables z are also directly fed into intermediate layers of
the network to ensure encoded information in c or z is not
omitted. All the discriminators Di have down-sampling blocks
and 3×3 convolutions to transform the input image to a
4×4×8Nd tensor, and eventually the sigmoid function is used
for outputting probabilities. For all datasets, we set Ng = 32,
Nd = 64 and use two residual blocks between every two
generators. ADAM [16] solver with a learning rate of 0.0002
is used for all models. The source code for StackGAN-v2 is
available at https://github.com/hanzhanggit/StackGAN-v2 for
more implementation details.
7
Dataset
#Samples
CUB [45]
train
test
8,855
2,933
Oxford-102 [26]
train
test
7,034
1,155
COCO [20]
train
test
80,000
40,000
LSUN [49]
bedroom
church
3,033,042
126,227
ImageNet [35]
dog
cat
147,873
6,500
TABLE 1: Statistics of datasets. We do not split LSUN or ImageNet because they are utilized for the unconditional tasks.
6
E XPERIMENTS
We conduct extensive experiments to evaluate the proposed
methods. In section 6.1, several state-of-the-art methods on
text-to-image synthesis [31], [29] and on unconditional image
synthesis [28], [51], [2], [22] are compared with the proposed
methods. We first evaluate the effectiveness of our StackGANv1 by comparing it with GAWWN [29] and GAN-INTCLS [31] for text-to-image synthesis. And then, StackGANv2 is compared with StackGAN-v1 to show its advantages
and limitations. Moreover, StackGAN-v2 as a more general
framework can also solve unconditional image synthesis tasks
and on such tasks, and it is compared with several state-ofthe-art methods [28], [51], [2], [22]. In section 6.2, several
baseline models are designed to investigate the overall design
and important components of our StackGAN-v1. For the first
baseline, we directly train Stage-I GAN for generating 64×64
and 256×256 images to investigate whether the proposed twostage stacked structure and the Conditioning Augmentation
are beneficial. Then we modify our StackGAN-v1 to generate
128×128 and 256×256 images to investigate whether larger
images by our method can result in higher image quality.
We also investigate whether inputting text at both stages of
StackGAN-v1 is useful. In section 6.3, experiments are designed to validate important components of our StackGAN-v2,
including designs with fewer multi-scale image distributions,
the effect of jointly approximating conditional and unconditional distributions, and the effectiveness of the proposed
color-consistency regularization.
Datasets. We evaluate our conditional StackGAN for textto-image synthesis on the CUB [45], Oxford-102 [26] and
MS COCO [20] datasets. CUB [45] contains 200 bird species
with 11,788 images. Since 80% of birds in this dataset have
object-image size ratios of less than 0.5 [45], as a preprocessing step, we crop all images to ensure that bounding
boxes of birds have greater-than-0.75 object-image size ratios.
Oxford-102 [26] contains 8,189 images of flowers from 102
different categories. To show the generalization capability of
our approach, a more challenging dataset, MS COCO [20] is
also utilized for evaluation. Different from CUB and Oxford102, the MS COCO dataset contains images with multiple
objects and various backgrounds. Each image in COCO has
5 descriptions, while 10 descriptions are provided by [30] for
every image in CUB and Oxford-102 datasets. Following the
experimental setup in [31], we directly use the training and
validation sets provided by COCO, meanwhile we split CUB
and Oxford-102 into class-disjoint training and test sets. Our
unconditional StackGAN-v2 utilizes bedroom and church subsets of LSUN [49], a dog-breed 1 and a cat-breed 2 sub-sets of
ImageNet [35] to synthesize different types of images. Since
1. Using the wordNet IDs provided by Vinyals et al., [43]
2. Using the wordNet IDs provided in our supplementary materials
the dog dataset is more complex than others and covers 118
classes of ImageNet, we will also report quantitative results on
this dataset. The statistics of datasets are presented in Table 1.
Evaluation metrics. It is difficult to evaluate the performance of generative models (e.g., GANs). We choose a
recently proposed numerical assessment approach “inception
score” [36] for quantitative evaluation,
I = exp (Ex DKL (p (y|x) || p (y))) ,
(14)
where x denotes one generated sample, and y is the label
predicted by the Inception model [37]. The intuition behind
this metric is that good models should generate diverse but
meaningful images. Therefore, the KL divergence between
the marginal distribution p(y) and the conditional distribution
p(y|x) should be large. In our experiments, we directly use
the pre-trained Inception model for COCO and ImageNet Dog
datasets. For fine-grained datasets, CUB and Oxford-102, we
fine-tune an Inception model for each of them. As suggested
in [36], we evaluate this metric on a large number of samples
(i.e., 30k randomly selected samples) for each model.
Although the inception score has shown to well correlate
with human perception on visual quality of samples [36],
it cannot reflect whether the generated images are well
conditioned on the given text descriptions. Therefore, we
also conduct human evaluation. We randomly select 50 text
descriptions for each class of CUB and Oxford-102 test sets.
For COCO dataset, 4k text descriptions are randomly selected
from its validation set. For each sentence, 5 images are
generated by each model. Given the same text descriptions,
10 users (not including any of the authors) are asked to rank
the results by different methods. The average ranks by human
users are calculated to evaluate all compared methods.
In addition, we use t-SNE [42] embedding method to
visualize a large number (e.g., 30,000 on the CUB test set)
of high-dimensional images in a two-dimensional map. We
observe that t-SNE is a good tool to examine the distribution
of synthesized images and identify collapsed modes.
6.1
Comparison with state-of-the-art GAN models
To demonstrate the effectiveness of the proposed method, we
compare it with state-of-the-art GAN models on text-to-image
synthesis [31], [29] and unconditional image synthesis [28],
[51], [2], [22].
Text-to-image synthesis. We first compare our StackGANv1 with the state-of-the-art text-to-image methods [29], [31] on
CUB, Oxford-102 and COCO datasets to demonstrate its effectiveness. We then compare StackGAN-v2 with StackGAN-v1
to show its advantages and limitations. The inception scores
and average human ranks for the proposed StackGAN models
and compared methods are reported in Table 2. Representative
examples are compared in Figure 3, Figure 4 and Figure 5.
8
Metric
Dataset
CUB
Inception
Oxford
score
COCO
CUB
Human
Oxford
rank
COCO
GAN-INT-CLS
2.88 ± .04
2.66 ± .03
7.88 ± .07
2.81 ± .03
1.87 ± .03
1.89 ± .04
GAWWN
3.62 ± .07
/
/
1.99 ± .04
/
/
Our StackGAN-v1
3.70 ± .04
3.20 ± .01
8.45 ± .03
1.37 ± .02
1.13 ± .03
1.11 ± .03
Our StackGAN-v2
4.04 ± .05
/
/
/
/
/
TABLE 2: Inception scores and average human ranks of our StackGAN-v1, StackGAN-v2, GAWWN [29], and GAN-INT-CLS [31] on CUB,
Oxford-102, and MS-COCO datasets.
Text
description
This bird is red
and brown in
color, with a
stubby beak
The bird is
short and
stubby with
yellow on its
body
A bird with a
medium orange
bill white body
gray wings and
webbed feet
This small
black bird has
a short, slightly
curved bill and
long legs
A small bird
with varying
shades of
brown with
white under the
eyes
A small yellow
bird with a
black crown
and a short
black pointed
beak
This small bird
has a white
breast, light
grey head, and
black wings
and tail
64x64
GAN-INT-CLS
128x128
GAWWN
256x256
StackGAN-v1
Fig. 3: Example results by our StackGAN-v1, GAWWN [29], and GAN-INT-CLS [31] conditioned on text descriptions from CUB test set.
Compared with previous GAN models [31], [29], on
the text-to-image synthesis task, our StackGAN-v1 model
achieves the best inception score and average human rank
on all three datasets. For example, compared with GANINT-CLS [31], StackGAN-v1 achieves 28.47% improvement
in terms of inception score on CUB dataset (from 2.88 to
3.70), and 20.30% improvement on Oxford-102 (from 2.66
to 3.20). The better average human rank of our StackGANv1 also indicates our proposed method is able to generate
more realistic samples conditioned on text descriptions. As
shown in Figure 3, the 64×64 samples generated by GANINT-CLS [31] can only reflect the general shape and color of
the birds. Their results lack vivid parts (e.g., beak and legs)
and convincing details in most cases, which make them neither
realistic enough nor have sufficiently high resolution. By
using additional conditioning variables on location constraints,
GAWWN [29] obtains a better inception score on CUB
dataset, which is still slightly lower than ours. It generates
higher resolution images with more details than GAN-INTCLS, as shown in Figure 3. However, as mentioned by its
authors, GAWWN fails to generate any plausible images when
it is only conditioned on text descriptions [29]. In comparison,
our StackGAN-v1 for the first time generates images of
256×256 resolution with photo-realistic details from only text
descriptions.
Compared with StackGAN-v1, our StackGAN-v2 further
improves the inception score from 3.70 to 4.04 on the CUB
dataset. For visual comparison of the results by the two
models, we utilize the t-SNE [42] algorithm. For each model,
a large number of images are generated and embedded into
the 2D plane. We first extract a 2048d CNN feature from
each generated image using a pre-trained Inception model.
Then, t-SNE algorithm is applied to embed the CNN features
into a 2D plane, resulting a location for each image in the
2D plane. Due to page limits, Figure 5 only shows a 50×50
grid with compressed images for each dataset, where each
generated image is mapped to its nearest grid location. As
shown in Figure 5, StackGAN-v1 contains two collapsed
modes (nonsensical images). StackGAN-v1 generates failurecase images if the Stage-I images are around a collapsed mode.
For StackGAN-v2, since all the generators are trained jointly,
they receive more regularization and thus result in more stable
training behavior and less chance of mode collapse. We did not
observe any collapsed nonsensical mode in the visualization
of StackGAN-v2-generated images. While StackGAN-v2 is
more advanced than StackGAN-v1 in many aspects (including
end-to-end training, more stable training behavior and higher
inception score), StackGAN-v1 has the advantage that it can
be trained stage by stage, requiring less GPU memory.
Unconditional image synthesis. We evaluate the effective-
9
Text
description
This flower has
a lot of small
purple petals in
a dome-like
configuration
This flower is
pink, white,
and yellow in
color, and has
petals that are
striped
This flower has
petals that are
dark pink with
white edges
and pink
stamen
This flower is
white and
yellow in color,
with petals that
are wavy and
smooth
A picture of a
very clean
living room
A group of
people on skis
stand in the
snow
Eggs fruit
candy nuts
and meat
served on
white dish
A street sign
on a stoplight
pole in the
middle of a
day
64x64
GAN-INT-CLS
256x256
StackGAN-v1
Fig. 4: Example results by our StackGAN-v1 and GAN-INT-CLS [31] conditioned on text descriptions from Oxford-102 test set (leftmost
four columns) and COCO validation set (rightmost four columns).
(a) StackGAN-v1 has two collapsed modes (in red rectangles). (b) StackGAN-v2 contains no collapsed nonsensical mode.
Fig. 5: Utilizing t-SNE to embed images generated by our StackGAN-v1 and StackGAN-v2 on the CUB test set.
ness of StackGAN-v2 for the unconditional image generation
task by comparing it with DCGAN [28], WGAN [2], EBGANPT [51], LSGAN [22] on the LSUN bedroom dataset. As
shown in Figure 6, our StackGAN-v2 is able to generate
256×256 images with more photo-realistic details. In Figure 7, we also compare the 256×256 samples generated
by StackGAN-v2 and EBGAN-PT. As shown in the figure,
the samples generated by the two methods have the same
resolution, but StackGAN-v2 generates more realistic ones
(e.g., more recognizable dog faces with eyes and noses). While
only qualitative results are reported in [28], [2], [51], [22],
a DCGAN model [28] is trained for quantitative comparison
using the public available source code 3 on the ImageNet Dog
dataset. The inception score of DCGAN is 8.19 ± 0.11 which
3. https://github.com/carpedm20/DCGAN-tensorflow
is much lower than the inception achieved by our StackGANv2 (9.55 ± 0.11). These experiments demonstrate that our
StackGAN-v2 outperforms the state-of-the-art methods for
unconditional image generation. Example images generated by
StackGAN-v2 on LSUN church and ImageNet cat datasets are
presented in Figure 8.
6.2
The component analysis of StackGAN-v1
In this section, we analyze different components of
StackGAN-v1 on CUB dataset with baseline models.
The design of StackGAN-v1. As shown in the first
four rows of Table 3, if Stage-I GAN is directly used to
generate images, the inception scores decrease significantly.
Such performance drop can be well illustrated by results in
Figure 10. As shown in the first row of Figure 10, Stage-I GAN
10
Samples by DCGAN (Reported in [28])
Samples by WGAN (Reported in [2])
Samples by EBGAN-PT (Reported in [51])
Samples by LSGAN (Reported in [22])
256×256 Samples by our StackGAN-v2
Fig. 6: Comparison of samples generated by models trained on LSUN bedroom [49] dataset.
Fig. 7: 256×256 samples generated by EBGAN-PT [51] (top) and our StackGAN-v2 (bottom) on ImageNet dog dataset [35].
Fig. 8: 256×256 samples generated by our StackGAN-v2 on LSUN church [49] (top) and ImageNet cat [35] (bottom) datasets.
Method
64×64 Stage-I GAN
256×256 Stage-I GAN
128×128 StackGAN-v1
256×256 StackGAN-v1
CA
no
yes
no
yes
yes
no
yes
yes
no
yes
Text twice
/
/
/
/
no
yes
yes
no
yes
yes
Inception
2.66 ±
2.95 ±
2.48 ±
3.02 ±
3.13 ±
3.20 ±
3.35 ±
3.45 ±
3.31 ±
3.70 ±
score
.03
.02
.00
.01
.03
.03
.02
.02
.03
.04
TABLE 3: Inception scores calculated with 30,000 samples generated
by different baseline models of our StackGAN-v1.
fails to generate any plausible 256×256 samples without using
Conditioning Augmentation (CA). Although Stage-I GAN
with CA is able to generate more diverse 256×256 samples,
those samples are not as realistic as samples generated by
StackGAN-v1. It demonstrates the necessity of the proposed
stacked structure. In addition, by decreasing the output resolution from 256×256 to 128×128, the inception score decreases
from 3.70 to 3.35. Note that all images are scaled to 299
× 299 before calculating the inception score. Thus, if our
StackGAN-v1 just increases the image size without adding
more information, the inception score would remain the same
for samples of different resolutions. Therefore, the decrease
in inception score by 128×128 StackGAN-v1 demonstrates
that our 256×256 StackGAN-v1 does add more details into
the larger images. For the 256×256 StackGAN-v1, if the
text is only input to Stage-I (denoted as “no Text twice”),
the inception score decreases from 3.70 to 3.45. It indicates
that processing text descriptions again at Stage-II helps refine
Stage-I results. The same conclusion can be drawn from the
results of 128×128 StackGAN-v1 models.
11
This bird is
Text
blue with white
description and has a very
short beak
This bird has
wings that are
brown and has
a yellow belly
A white bird
with a black
crown and
yellow beak
This bird is
white, black,
and brown in
color, with a
brown beak
The bird has
small beak,
with reddish
brown crown
and gray belly
This is a small,
black bird with
a white breast
and white on
the wingbars.
This bird is
white black and
yellow in color,
with a short
black beak
Stage-I
images
Stage-II
images
Fig. 9: Samples generated by our StackGAN-v1 from unseen texts in CUB test set. Each column lists the text description, images generated
from the text by Stage-I and Stage-II of StackGAN-v1.
A small bird with a black head and
wings and features grey wings
This bird is completely red with black
wings and pointy beak
The bird is completely red → The bird is completely yellow
256x256
Stage-I GAN
without CA
This bird is completely red with black wings and pointy beak →
this small blue bird has a short pointy beak and brown on its wings
256x256
Stage-I GAN
with CA
256x256
StackGAN
with CA,
Text twice
Fig. 10: Conditioning Augmentation (CA) helps stabilize the training
of conditional GAN and improves the diversity of the generated
samples. (Row 1) without CA, Stage-I GAN fails to generate
plausible 256×256 samples. Although different noise vector z is used
for each column, the generated samples collapse to be the same for
each input text description. (Row 2-3) with CA but fixing the noise
vectors z, methods are still able to generate birds with different poses
and viewpoints.
Images
generated from
text in test sets
Five nearest neighbors from training sets
Fig. 11: For generated images (column 1), retrieving their nearest
training images (columns 2-6) by utilizing Stage-II discriminator of
StackGAN-v1 to extract visual features. The L2 distances between
features are calculated for nearest-neighbor retrieval.
Figure 9 illustrates some examples of the Stage-I and
Stage-II images generated by our StackGAN-v1. As shown
in the first row of Figure 9, in most cases, Stage-I GAN is
able to draw rough shapes and colors of objects given text
descriptions. However, Stage-I images are usually blurry with
various defects and missing details, especially for foreground
objects. As shown in the second row, Stage-II GAN generates
Fig. 12: (Left to right) Images generated by interpolating two sentence
embeddings. Gradual appearance changes from the first sentence’s
meaning to that of the second sentence can be observed. The noise
vector z is fixed to be zeros for each row.
4× higher resolution images with more convincing details
to better reflect corresponding text descriptions. For cases
where Stage-I GAN has generated plausible shapes and colors,
Stage-II GAN completes the details. For instance, in the 1st
column of Figure 9, with a satisfactory Stage-I result, StageII GAN focuses on drawing the short beak and white color
described in the text as well as details for the tail and legs.
In all other examples, different degrees of details are added
to Stage-II images. In many other cases, Stage-II GAN is
able to correct the defects of Stage-I results by processing the
text description again. For example, while the Stage-I image
in the 5th column has a blue crown rather than the reddish
brown crown described in the text, the defect is corrected by
Stage-II GAN. In some extreme cases (e.g., the 7th column of
Figure 9), even when Stage-I GAN fails to draw a plausible
shape, Stage-II GAN is able to generate reasonable objects.
We also observe that StackGAN-v1 has the ability to transfer
background from Stage-I images and fine-tune them to be
more realistic with higher resolution at Stage-II.
Importantly, the StackGAN-v1 does not achieve good results
by simply memorizing training samples but by capturing
the complex underlying language-image relations. We extract
visual features from our generated images and all training
images by the Stage-II discriminator D of our StackGAN-
12
v1. For each generated image, its nearest neighbors from
the training set can be retrieved. By visually inspecting
the retrieved images (see Figure 11), we conclude that the
generated images have some similar characteristics with the
training samples but are essentially different.
Conditioning Augmentation. We also investigate the efficacy of the proposed Conditioning Augmentation (CA). By
removing it from StackGAN-v1 256×256 (denoted as “no
CA” in Table 3), the inception score decreases from 3.70
to 3.31. Figure 10 also shows that 256×256 Stage-I GAN
(and StackGAN-v1) with CA can generate birds with different
poses and viewpoints from the same text embedding. In
contrast, without using CA, samples generated by 256×256
Stage-I GAN collapse to nonsensical images due to the unstable training dynamics of GANs. Consequently, the proposed
Conditioning Augmentation helps stabilize the conditional
GAN training and improves the diversity of the generated
samples because of its ability to encourage robustness to small
perturbations along the latent manifold.
Sentence embedding interpolation. To further demonstrate that our StackGAN-v1 learns a smooth latent data manifold, we use it to generate images from linearly interpolated
sentence embeddings, as shown in Figure 12. We fix the noise
vector z, so the generated image is inferred from the given
text description only. Images in the first row are generated
by simple sentences made up by us. Those sentences contain
only simple color descriptions. The results show that the
generated images from interpolated embeddings can accurately
reflect color changes and generate plausible bird shapes. The
second row illustrates samples generated from more complex
sentences, which contain more details on bird appearances.
The generated images change their primary color from red to
blue, and change the wing color from black to brown.
6.3
The component analysis of StackGAN-v2
In this section, we analyze important components of the
proposed StackGAN-v2. Table 4 lists models with different
settings and their inception scores on the CUB test set. Figure 13 shows example images generated by different baseline
models.
Our baseline models are built by removing or changing a certain component of the StackGAN-v2 model. By
only approximating a single image distribution, the inception
scores on CUB dramatically decrease from 4.04 to 3.49 for
“StackGAN-v2-G3 ” and to 2.89 for “StackGAN-v2-all256”
(See Figures 13 (e-f)). Inspired by [9], we also build a baseline model with multiple discriminators, namely “StackGANv2-3G3 ”. Those discriminators have the same structure but
different initializations. However, the results do not show
improvement over “StackGAN-v2-G3 ”. Similar comparisons
have also been done for the unconditional task on the LSUN
bedroom dataset. As shown in Figures 13(a-c), those baseline
models fail to generate realistic images because they suffer
from severe mode collapses.
To further demonstrate the effectiveness of jointly approximating conditional and unconditional distributions,
“StackGAN-v2-no-JCU” replaces the joint conditional and
unconditional discriminators with the conventional ones, resulting in much lower inception score than that of StackGANv2. Another baseline model does not use the color-consistency
regularization term. Results on various datasets (see Figure 14)
show that the color-consistency regularization has significant
positive effects for the unconditional image synthesis task.
Quantitatively, removing the color-consistency regularization
decreases the inception score from 9.55 ± 0.11 to 9.02 ± 0.14
on the ImageNet dog dataset. It demonstrates that the additional constraint provided by the color-consistency regularization is able to facilitate multi-distribution approximation and
help generators at different branches produce more coherent
samples from the same noise input. It is worth mentioning
that there is no need to utilize the color-consistency regularization for the text-to-image synthesis task because of the
instance-wise constraint between images and corresponding
text descriptions. Experimentally, adding the color-consistency
regularization did not improve the inception score on CUB
dataset.
7
C ONCLUSIONS
In this paper, Stacked Generative Adversarial Networks,
StackGAN-v1 and StackGAN-v2, are proposed to decompose
the difficult problem of generating realistic high-resolution
images into more manageable sub-problems. The StackGANv1 with Conditioning Augmentation is first proposed for textto-image synthesis through a novel sketch-refinement process.
It succeeds in generating images of 256×256 resolution
with photo-realistic details from text descriptions. To further improve the quality of generated samples and stabilize GANs’ training, the StackGAN-v2 jointly approximates
multiple related distributions, including (1) multi-scale image
distributions and (2) joint conditional and unconditional image
distributions. In addition, a color-consistency regularization is
proposed to facilitate multi-distribution approximation. Extensive quantitative and qualitative results demonstrate that our
proposed methods significantly improve the state of the art in
both conditional and unconditional image generation tasks.
R EFERENCES
[1]
M. Arjovsky and L. Bottou. Towards principled methods for training
generative adversarial networks. In ICLR, 2017. 1
[2] M. Arjovsky, S. Chintala, and L. Bottou.
Wasserstein GAN.
arXiv:1701.07875, 2017. 1, 7, 9, 10
[3] A. Brock, T. Lim, J. M. Ritchie, and N. Weston. Neural photo editing
with introspective adversarial networks. In ICLR, 2017. 2
[4] T. Che, Y. Li, A. P. Jacob, Y. Bengio, and W. Li. Mode regularized
generative adversarial networks. In ICLR, 2017. 1, 2
[5] T. Che, Y. Li, R. Zhang, R. D. Hjelm, W. Li, Y. Song, and Y. Bengio. Maximum-likelihood augmented discrete generative adversarial
networks. arXiv:1702.07983, 2017. 1
[6] X. Chen, Y. Duan, R. Houthooft, J. Schulman, I. Sutskever, and
P. Abbeel. InfoGAN: Interpretable representation learning by information maximizing generative adversarial nets. In NIPS, 2016. 2
[7] E. L. Denton, S. Chintala, A. Szlam, and R. Fergus. Deep generative
image models using a laplacian pyramid of adversarial networks. In
NIPS, 2015. 2
[8] C. Doersch. Tutorial on variational autoencoders. arXiv:1606.05908,
2016. 3
[9] I. P. Durugkar, I. Gemp, and S. Mahadevan. Generative multi-adversarial
networks. In ICLR, 2017. 2, 12
[10] J. Gauthier. Conditional generative adversarial networks for convolutional face generation. Technical report, 2015. 3
13
Model
StackGAN-v2
StackGAN-v2-no-JCU
StackGAN-v2-G3
StackGAN-v2-3G3
StackGAN-v2-all256
branch G1
64×64
64×64
removed
removed
256×256
branch G2
128×128
128×128
removed
removed
256×256
branch G3
256×256
256×256
256×256
three 256×256
256×256
JCU
yes
no
yes
yes
yes
inception
4.04 ±
3.77 ±
3.49 ±
3.22 ±
2.89 ±
score
.05
.04
.04
.02
.02
TABLE 4: Inception scores by our StackGAN-v2 and its baseline models on CUB test set. “JCU” means using the proposed discriminator
that jointly approximates conditional and unconditional distributions.
(a) StackGAN-v2-all256
(b) StackGAN-v2-G3
(d) StackGAN-v2
(c) StackGAN-v2-3G3
This black and white and grey bird has a black bandit marking around it’s eyes
(e) StackGAN-v2-all256
(f) StackGAN-v2-G3
(g) StackGAN-v2-no-JCU
(h) StackGAN-v2
Fig. 13: Example images generated by the StackGAN-v2 and its baseline models on LSUN bedroom (top) and CUB (bottom) datasets.
ImageNet dog [35]
ImageNet cat [35]
LSUN church [49]
Fig. 14: Example images generated without (top) and with (bottom) the proposed color-consistency regularization for our StackGAN-v2 on
ImageNet dog, ImageNet cat and LSUN church datasets. (Left to right) 64×64, 128×128, and 256×256 images by G1 , G2 , G3 , respectively.
[11] I. J. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley,
S. Ozair, A. C. Courville, and Y. Bengio. Generative adversarial nets.
In NIPS, 2014. 1, 2, 3, 5
[12] K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image
recognition. In CVPR, 2016. 4
[13] X. Huang, Y. Li, O. Poursaeed, J. Hopcroft, and S. Belongie. Stacked
generative adversarial networks. In CVPR, 2017. 2
[14] S. Ioffe and C. Szegedy. Batch normalization: Accelerating deep network
training by reducing internal covariate shift. In ICML, 2015. 4
[15] P. Isola, J.-Y. Zhu, T. Zhou, and A. A. Efros. Image-to-image translation
with conditional adversarial networks. In CVPR, 2017. 2
[16] D. P. Kingma and J. Ba. Adam: A method for stochastic optimization.
In ICLR, 2015. 5, 6
[17] D. P. Kingma and M. Welling. Auto-encoding variational bayes. In
ICLR, 2014. 2, 3
[18] A. B. L. Larsen, S. K. Sønderby, H. Larochelle, and O. Winther.
Autoencoding beyond pixels using a learned similarity metric. In ICML,
2016. 3
[19] C. Ledig, L. Theis, F. Huszar, J. Caballero, A. Aitken, A. Tejani, J. Totz,
Z. Wang, and W. Shi. Photo-realistic single image super-resolution using
a generative adversarial network. In CVPR, 2017. 2
[20] T.-Y. Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ramanan,
P. Dollr, and C. L. Zitnick. Microsoft coco: Common objects in context.
In ECCV, 2014. 7
[21] E. Mansimov, E. Parisotto, L. J. Ba, and R. Salakhutdinov. Generating
images from captions with attention. In ICLR, 2016. 2
[22] X. Mao, Q. Li, H. Xie, R. Y. K. Lau, Z. Wang, and S. P. Smolley. Least
squares generative adversarial networks. arXiv:1611.04076, 2016. 7, 9,
10
[23] L. Metz, B. Poole, D. Pfau, and J. Sohl-Dickstein. Unrolled generative
adversarial networks. In ICLR, 2017. 2
[24] M. Mirza and S. Osindero. Conditional generative adversarial nets.
arXiv:1411.1784, 2014. 3
[25] A. Nguyen, J. Yosinski, Y. Bengio, A. Dosovitskiy, and J. Clune. Plug
& play generative networks: Conditional iterative generation of images
in latent space. In CVPR, 2017. 2
[26] M.-E. Nilsback and A. Zisserman. Automated flower classification over
a large number of classes. In ICCVGIP, 2008. 7
[27] A. Odena, C. Olah, and J. Shlens. Conditional image synthesis with
auxiliary classifier GANs. In ICML, 2017. 2
[28] A. Radford, L. Metz, and S. Chintala. Unsupervised representation
learning with deep convolutional generative adversarial networks. In
ICLR, 2016. 1, 2, 7, 9, 10
[29] S. Reed, Z. Akata, S. Mohan, S. Tenka, B. Schiele, and H. Lee. Learning
what and where to draw. In NIPS, 2016. 2, 3, 7, 8
[30] S. Reed, Z. Akata, B. Schiele, and H. Lee. Learning deep representations
of fine-grained visual descriptions. In CVPR, 2016. 3, 7
[31] S. Reed, Z. Akata, X. Yan, L. Logeswaran, B. Schiele, and H. Lee.
Generative adversarial text-to-image synthesis. In ICML, 2016. 2, 3, 4,
7, 8, 9
[32] S. Reed, A. van den Oord, N. Kalchbrenner, V. Bapst, M. Botvinick,
and N. de Freitas. Generating interpretable images with controllable
structure. Technical report, 2016. 2
[33] D. J. Rezende, S. Mohamed, and D. Wierstra. Stochastic backpropagation and approximate inference in deep generative models. In ICML,
2014. 2
[34] D. L. Ruderman. The statistics of natural images. Network: Computation
In Neural Systems, 5(4):517–548, 1994. 1
14
[35] O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma,
Z. Huang, A. Karpathy, A. Khosla, M. Bernstein, A. C. Berg, and L. FeiFei. ImageNet Large Scale Visual Recognition Challenge. International
Journal of Computer Vision (IJCV), 115(3):211–252, 2015. 7, 10, 13
[36] T. Salimans, I. J. Goodfellow, W. Zaremba, V. Cheung, A. Radford, and
X. Chen. Improved techniques for training GANs. In NIPS, 2016. 1,
2, 7
[37] C. Szegedy, V. Vanhoucke, S. Ioffe, J. Shlens, and Z. Wojna. Rethinking
the inception architecture for computer vision. In CVPR, 2016. 7
[38] C. K. Snderby, J. Caballero, L. Theis, W. Shi, and F. Huszar. Amortised
map inference for image super-resolution. In ICLR, 2017. 1, 2
[39] Y. Taigman, A. Polyak, and L. Wolf. Unsupervised cross-domain image
generation. In ICLR, 2017. 2
[40] A. van den Oord, N. Kalchbrenner, and K. Kavukcuoglu. Pixel recurrent
neural networks. In ICML, 2016. 2
[41] A. van den Oord, N. Kalchbrenner, O. Vinyals, L. Espeholt, A. Graves,
and K. Kavukcuoglu. Conditional image generation with pixelcnn
decoders. In NIPS, 2016. 2
[42] L. van der Maaten and G. Hinton. Visualizing high-dimensional data
using t-sne. Journal of Machine Learning Research, 9: 25792605, Nov
2008. 7, 8
[43] O. Vinyals, C. Blundell, T. Lillicrap, K. Kavukcuoglu, and D. Wierstra.
Matching networks for one shot learning. In NIPS, pages 3630–3638,
2016. 7
[44] C. Vondrick, H. Pirsiavash, and A. Torralba. Generating videos with
scene dynamics. In NIPS, 2016. 1
[45] C. Wah, S. Branson, P. Welinder, P. Perona, and S. Belongie. The
Caltech-UCSD Birds-200-2011 Dataset. Technical Report CNS-TR2011-001, California Institute of Technology, 2011. 7
[46] X. Wang and A. Gupta. Generative image modeling using style and
structure adversarial networks. In ECCV, 2016. 2
[47] X. Yan, J. Yang, K. Sohn, and H. Lee. Attribute2image: Conditional
image generation from visual attributes. In ECCV, 2016. 2
[48] J. Yang, A. Kannan, D. Batra, and D. Parikh. LR-GAN: layered recursive
generative adversarial networks for image generation. In ICLR, 2017. 2
[49] F. Yu, Y. Zhang, S. Song, A. Seff, and J. Xiao. LSUN: Construction
of a large-scale image dataset using deep learning with humans in the
loop. arXiv preprint arXiv:1506.03365, 2015. 7, 10, 13
[50] H. Zhang, T. Xu, H. Li, S. Zhang, X. Wang, X. Huang, and D. Metaxas.
StackGAN: Text to photo-realistic image synthesis with stacked generative adversarial networks. In ICCV, 2017. 2
[51] J. Zhao, M. Mathieu, and Y. LeCun. Energy-based generative adversarial
network. In ICLR, 2017. 2, 7, 9, 10
[52] J. Zhu, P. Krähenbühl, E. Shechtman, and A. A. Efros. Generative visual
manipulation on the natural image manifold. In ECCV, 2016. 2
Han Zhang received his B.S. in information
science from China Agricultural University, Beijing,
China, in 2009 and M.E. in communication and
information systems from Beijing University of
Posts and Telecommunications, Beijing, China, in
2012. He is currently a Ph.D. student of Department
of Computer Science at Rutgers University, Piscataway, NJ. His current research interests include
computer vision, deep learning and medical image
processing.
Tao Xu received her B.E. in agricultural mechanization and automatization from China Agricultural
University, Beijing, China, in 2010, and M.S. in
computer science from the Institute of Computing
Technology, Chinese Academy of Science, Beijing,
China, in 2013. She is currently a Ph.D. student of
Department of Computer Science and Engineering
at Lehigh University, Bethlehem, PA. Her current
research interests include deep learning, computer
vision, and medical image processing.
Hongsheng Li received the bachelors degree
in automation from the East China University
of Science and Technology, and the masters and
doctorate degrees in computer science from Lehigh
University, Pennsylvania, in 2006, 2010, and 2012,
respectively. He is currently with the Department
of Electronic Engineering at the Chinese University
of Hong Kong. His research interests include
computer vision, medical image analysis, and
machine learning.
Shaoting Zhang received the BE degree
from Zhejiang University in 2005, the
MS
degree
from
Shanghai
Jiao
Tong
University in 2007, and the PhD degree in
computer science from Rutgers in January
2012. His research is on the interface
of
medical
imaging
informatics,
largescale visual understanding, and machine
learning. He is a senior member of the
IEEE.
Xiaogang Wang received the BS degree from the
University of Science and Technology of China in
2001, the MS degree from the Chinese University
of Hong Kong in 2003, and the PhD degree from
the Computer Science and Artificial Intelligence
Laboratory, Massachusetts Institute of Technology
in 2009. He is currently an associate professor in the
Department of Electronic Engineering, The Chinese
University of Hong Kong. His research interests
include computer vision and machine learning. He
is a member of the IEEE.
Xiaolei Huang received her doctorate degree
in computer science from Rutgers University–New
Brunswick, and her bachelor’s degree in computer
science from Tsinghua University (China). She is
currently an Associate Professor in the Computer
Science and Engineering Department at Lehigh
University, Bethlehem, PA. Her research interests
are in the areas of Computer Vision, Biomedical
Image Analysis, Computer Graphics, and Machine
Learning. In these areas she has published articles in
journals such as TPAMI, MedIA, TMI, TOG, and Scientific Reports.
She also regularly contributes research papers to conferences such
as CVPR, MICCAI, and ICCV. She serves as an Associate Editor
for the Computer Vision and Image Understanding journal. She is a
member of the IEEE.
Dimitris N. Metaxas received the BE degree from
the National Technical University of Athens Greece
in 1986, the MS degree from the University of
Maryland in 1988, and the PhD degree from the
University of Toronto in 1992. He is a professor in
the Computer Science Department, Rutgers University. He is directing the Computational Biomedicine
Imaging and Modeling Center (CBIM). He has
been conducting research toward the development
of formal methods upon which computer vision,
computer graphics, and medical imaging can advance synergistically.
He is a fellow of the IEEE.
| 1cs.CV
|
arXiv:1604.06359v1 [math.GR] 21 Apr 2016
p-quotients of the G.Higman group
L. Glebsky
February 19, 2018
Abstract
These notes are based on the mini-course “On the Graham Higman group”,
given at the Erwin Schrödinger Institute in Vienna, January 20, 22, 27 and 29,
2016, as a part of the Measured Group Theory program.1 The main purpose is
to describe p-quotients of the Higman group H(k) for p|(k−1). (One may check
that the condition p|(k − 1) is necessary for the existence of such quotients.)
1
Higman group
k
Consider the Higman group H(k) = ha0 , . . . , a3 | {a−1
i ai+1 ai = ai+1 , i = 0, ..., 3}i,
i + 1 is taking mod 4 here. It may be constructed as successive amalgamated free
k
products, starting from Baumslag-Solitar group BS(1,k)=ha0 , a1 | a−1
0 a1 a0 = a1 i :
k
−1
k
B3 ha0 , a1 , a2 i = ha0 , a1 | a−1
0 a1 a0 = a1 i ∗ ha1 , a2 | a1 a2 a1 = a2 i.
a1 ↔a1
Similarly, one may construct B3 ha2 , a3 , a0 i and
Hk = B3 ha0 , a1 , a2 i
∗
a0 ↔a0 ,a2 ↔a2
B3 ha2 , a3 , a0 i
The group H(2) was introduced by Graham Higman in [5] as an example of group
without finite quotients. Still H(2) has a lot of quotients, moreover, it is SQ-universal,
[9]. Actually the proof of [9] works for H(k), k ≥ 2, so, H(k) is SQ-universal for any
k ≥ 2. Some other techniques that were used for H(2) seem to be applicable for
H(k), see [4, 8]. But H(k) for k > 2 have another, compared with H(2), behavior
with respect to finite quotients. Particularly, H(k) has an arbitrary large p-quotient
for p|(k −1).Moreover, the intersection of the kernels of these quotient maps intersects
trivially with the Baumslag-Solitar subgroups B(1, k) = hai , ai+1 i < H(k). Using [4]
it implies the following statement:
1
This work was partially supported by the European Research Council (ERC) grant no. 259527
of G. Arzhantseva, part of this work was done in the Nizhny Nivgorod University and supported by
the RSF (Russia) grant 14-41-00044. The stay in Vienna was supported by ERC grant no. 259527
of G. Arzhantseva.
1
Proposition. Let p|(k − 1) be a prime. Then for any ε > 0 there is an n ∈ Z and
a bijection f : Z/pn Z → Z/pn Z such that f (x + 1) = kf (x) for at least (1 − ε)pn
elements x of Z/pn Z and f (f (f (f (x)))) = x for all x ∈ Z/pn Z.
The interesting property of this f is that it behaves “almost like” a modular
exponent (x → ak x mod pn ), but all it’s cycles are of the length 4. Precisely,
f (x) = a(x)k x mod pn where a(x) = a(x + 1) for almost all x (a(x) is “almost a
constant”). Maybe, the existence of such functions explains the difficulty in proving
estimates for the number of small cycles in repeated modular exponentiation, [2]. The
case k = 2 is not follows from this notes. So, the existence of f for k = 2 is an open
question.
Let X be a group (or other algebraic system), x ∈ X and φ : X → Y be a
homomorphism. We systematically, abusing notations, will write x to denote φ(x) if
it is clear from the context that we are dealing with an element of Y . If there are
φj : X → Yj we may say “x of Yj ” to denote φj (x).
2
p-quotients of a group and it’s p-central series.
Let G be a group. For S ⊆ G let hSi denote the subgroup of G generated by S. For
g, h ∈ G let [g, h] = g −1 h−1 gh denote the commutator of g and h. For H1 , H2 < G
let [H1 , H2 ] = h{[g, h] | g ∈ H1 , h ∈ H2 }i < G denote the commutator subgroup of
H1 and H2 . The p-central series G1 , G2 , . . . of a group G is defined as
G1 = G, Gi+1 = Gpi [Gi , G].
It is clear by the definition that G/Gi+1 is the maximal p-quotient of G of p-class at
most i. There is another, equivalent, definition of Gk . Let G[i] be the lower central
series for G:
G[1] = G, G[i+1] = [G[i] , G].
j
Exercise 1. Show that Gn = hGp[i] , i + j = ni. Hint. Using the commutator
identities (Theorem 5.1 (Witt-Hall identities) of [7]) show that
[u, v p ] = [u, v]p · [[u, v], v] · [[u, v 2], v] · · · · · [[u, v p−1], v].
k−1
k
k−1
p
p
]. Show that
]p [G[i+j] , G[j]
Particularly, this implies that [G[i] , Gp[j] ] ⊆ [G[i] , G[j]
k
j
[G, Gp[i] ] ⊆ hGp[r] , r + j = i + k + 1i. Then apply induction on n.
3
Calculating of p-quotients.
Let Zpn = Z/pn Z. Consider the non-commutative ring Zpn [x̄] of polynomials with
non-commutative (but associative) variables x̄ = (x0 , . . . , xm ) over Zpn . The ring
Zpn [x̄] contains finite subring Zpn [px̄] of polynomials f (px̄). It is clear that each
2
monomial of f (px̄) of order k is divisible by pk . Inside of Zpn [px̄] there is a group
Γ = h(1 + px0 ), . . . , (1 + pxm )i, generated by (1 + pxi ). Notice, that
(1 + px)
−1
=
n−1
X
(−p)j xj
j=0
n
Lemma 1. (1 + pxi )p = 1 in Zpn [x̄]. In other words j → (1 + pxi )j is a function
Zpn → Zpn [x̄].
Lemma 2 (Jacobson). Γ is isomorphic to F/Fn where F is a free group of rank
m + 1.
We prove this lemma in Section 8. Let I be a (two-sided) ideal in Zpn [x̄] and
φ : Zpn [x̄] → Zpn [x̄]/I be a natural map. Let ΓI = φ(Γ). We are going to use ΓI for
calculating G/Gn as follows. Let
G = ha0 , . . . , am | ui (a0 , . . . , am ) = wi (a0 , . . . , am ), i = 0, . . . , ki.
Let gi = p−αi (ui (1 + px0 , . . . , 1 + pxk ) − wi (1 + px0 , . . . , 1 + pxk )), where pαi devides all
coefficients of ui (1 + px0 , . . . , 1 + pxk ) − wi (1 + px0 , . . . , 1 + pxk ) and being maximal
with this property. Consider I = I(gi , i = 0, . . . , k), an ideal in Zpn [x̄] generated by
gi .
Lemma 3. ΓI is a homomorphic image of G/Gn .
Proof. By construction ΓI is a homomorphic image of G under φ : ai → (1 + pxi ). It
is easy to check that φ(Gn ) = {1}.
When ΓI is isomorphic to G/Gn ? Probably the answer is the following: if p 6= 2
then ΓI is isomorphic to G/Gn ; for p = 2 there are G such that ΓI is not isomorphic
to G/Gn . Probably, it is well known. Otherwise, one may try to use similar technique
as in the proof of Lemma 2 (see, Section 8) taking into account the solution of the
dimension subgroup problem. See [3] and the bibliography therein for the dimension
subgroup problem.
4
p-quotients of H(k), p|(k − 1).
H(k) = ha0 , . . . , a3 | {ai+1 ai = ai aki+1 , i = 0, ..., 3}i, here i + 1 is taken mod 4.
This is a presentation of the Higman group without inversion. Let p|(k − 1) then
substituting ai = (1 + pxi ) leads
gi =
1
(ai+1 ai − ai aki+1 ) = xi+1 xi − xi xi+1 + Q0 (xi+1 ) + pQ1 (xi , xi+1 ).
p2
(1)
Our aim is to study I = I(g0 , g1 , g2 , g3 ) in Zpn [x0 , . . . , x3 ], or precisely, Zpn [x0 , . . . , x3 ]/I
and ΓI . To this end we introduce some notions.
3
Definition 1. A (non-commutative) ring A is called to be an algebra over Zpn if
• A has the unity 1 ∈ A.
• Multiplication is Zpn -bilineal.
An algebra A is tame if A is a free Zpn -modul with a free basis A ∋ 1.
All algebras over Zpn we deal with are tame. So, in what follows we use just
“algebra” to denote “tame algebra”.
Definition 2. Let A (resp. B) be a Zpn -algebra and A ∋ 1 (resp. B ∋ 1) be a set
of it’s free generators as a Zpn -modul. A Zappa-Szep product C = A ⊲⊳ B is a Zpn
algebra such that
• C contains isomorphic copies of A and B such that A ∩ B = Zpn · 1 (we have
fixed such a copies of A and B and denote them by the same letters).
• The set {ab | a ∈ A, b ∈ B}, as well as the set {ba | a ∈ A, b ∈ B}, forms a
free basis of C as a Zpn -modul.
Exercise 2. It looks that by definition one should say that C = A ⊲⊳ B with
respect to A and B. Show that C = A ⊲⊳ B with respect to any free bases A′ ∋ 1
and B′ ∋ 1.
Definition 3. Let K, L be a groups. A Zappa-Szep product G = K ⊲⊳ L of groups K
and L is a group such that
• G contains isomorphic copies of K and L. We fix such copies and assume that
K, L < G.
• K ∩ L = {1} in G.
• G = KL. (This easily implies that G = LK.)
Remark. Of course, the definitions do not imply that a Zappa-Szep product is
uniquely defined by a pair of algebras o groups. Really, in order to define X ⊲⊳ Y
uniquely (up to isomorphism) one needs a function com : X × Y → Y × X which
describe how the elements of X commute with elements of Y . So, in some sense,
Z = X ⊲⊳ Y is an abuse of notation. In any case, when we use Z = X ⊲⊳ Y , the
structure of Z will be described.
Theorem 1. Let I = I(g0 , g1, g2 , g3 ) be an ideal in Zpn [x̄] generated by gi of Eq.(1).
Then Zpn [x̄]/I = Zpn [x0 , x2 ] ⊲⊳ Zpn [x1 , x3 ].
Recall, that Gi denotes the i-th term of the p-central series of group G = G0 . Also
we suppose that p|(k − 1).
Corollary 1. There is a surjective homomorphism H(k)/Hn (k) → F/Fn ⊲⊳ F/Fn ,
where F is a free group of rank 2.
4
Proof. By Lemma 3 H(k)/Hn (k) surjects into ΓI . By Theorem 1 and Lemma 2 there
are S, S̃ < ΓI such that
• S = ha0 = (1 + px0 ), a2 = (1 + px2 )i, S̃ = ha1 = (1 + px1 ), a3 = (1 + px3 )i;
• S and S̃ are isomorphic to F/Fn ;
• S ∩ S̃ = {1};
It follows that |S S̃| = |S| · |S̃| and S S̃ ≤ ΓI . So, we have to prove that ΓI ⊂ S S̃.
To this end it suffices to show that in ΓI there exist relations removing appearance
r
m r
m r
m r
of am
1 a0 , a1 a2 , a3 a0 and a3 a2 . By Lemma 1 we may assume that m, r ∈ Zpn .
k
r
r mk r
The relation a1 a0 = a0 a1 implies am
. The relation a2 a1 = a1 ak2 implies
1 a0 = a0 a1
−m
r
rk
am
am
1 a2 = a2
1 . Considerations for the other indexes are the same. Notice here, that
the condition p|(k − 1) implies that r → k r is a well definite function Zpn → Zpn .
5
What is going on
Consider w ∈ Hk as a word ani11 ani22 . . . but now suppose that nj are in some commutative ring, in our example, nj ∈ Zpn . We get a new group H̃k wich is a homomorphic image of Hk . Now, the relation ak1 = a−1
0 a1 a0 implies as well the relation
1/k
−1
j
a1 = a0 a1 a0 . We also need that j → k is well defined mod pn . After that any
mj
1
element of H̃k may be written as wu, where w = an0 1 am
2 . . . a2 , ni , mi ∈ Zpn and,
similarly, u = u(a1 , a3 ). So, we have H̃k = (Zpn ∗ Zpn ) ⊲⊳ (Zpn ∗ Zpn ) The whole
problem is to show that the natural homomorphism Zpn ∗ Zpn → F/Fn is compatible
with the ⊲⊳ structure. ( Here F is a rank 2 free group). Which, probably, may be
done another way as well...
Let restate all it more formally. On can easy to check that if G1 = H1 ⊲⊳ K and
G2 = H2 ⊲⊳ K then G1 ∗ G2 = (H1 ∗ H2 ) ⊲⊳ K. For Hk , p|(k − 1) we have the
K=K
following:
• BS(1, k) → BSpn (1, k) = Zpn ⋉k Zpn ,
• B3 → BSpn ∗ BSpn = (Zpn ∗ Zpn ) ⊲⊳ Zpn , where we amalgamate the different
Z pn
factors of BSpn .
• Finally we get Hk → (Zpn ∗ Zpn ) ⊲⊳ (Zpn ∗ Zpn ).
For the case k = 2 we may do similar things. We need a unitary commutative ring
R such that function r → 2r is defined in R. Such a ring exists, for example the real
numbers R. So, we may embed BS(1, 2) ֒→ R ⋉2 R, where for (αi , βi ) ∈ R ⋉2 R the
multiplication is defined as (α1 , β1 )(α2 , β2 ) = (α1 + α2 , 2α2 β1 + β2 ). Similarly, there is
a homorphism B3 → (R ∗ R) ⊲⊳ R, which is not injective, but nontrivial. And finally
we obtain H2 → HR = (R ∗ R) ⊲⊳ (R ∗ R). (Any element of (R ∗ R) ⊲⊳ (R ∗ R) is of
the form wu, w = aα0 1 aβ2 1 . . . aβ2 k , αi βi ∈ R and, similarly, u = u(a1 , a3 ).) Actually, Hk
has a nontrivial homomorphic image in HR for any k ∈ Z.
5
What is the structure of HR ? Changing ai → aδi we may write
β exp(α)
β
α
HR = haαi , i = 0, . . . , 3, α ∈ R | a−α
i ai+1 ai = ai+1
i.
It is not hard to show that haα0 , aα1 , aα2 , aα3 i < HR is residually finite for all α 6∈ Y for
a countable set Y ⊂ R.
6
Proof Theorem 1
The algebra Zpn [x̄]/I may be constructed as amalgamated free products, similar to
the construction of the Higman group itself.
Definition 4. Let A, B be Zpn -algebras, A ∋ 1, B ∋ 1 their corresponding free
generators as Zpn -modules. A free product A ∗ B is a Zpn -algebra that
• Generated by 1 and alternating words of letters from A′ = A \ {1} and B′ = B \
{1} as a free Zpn -module. A word w = w1 w2 . . . wr is alternating if (wi , wi+1 ) ∈
A′ × B ′ ∪ B ′ × A′ .
• It suffices to define product for two alternating words w = w1 . . . wr and u =
u1 . . . uk . So, wu = w1 . . . wr u1 . . . uk if (wr , u1 ) alternating and wu = w1 . . . wr−1(wr ·
u1 )u2 . . . ur otherwise. Here wr · u1 is the product in A or B.
For example, Zpn [x] ∗ Zpn [x] = Zpn [x0 , x1 ].
Exercise 3. Check that A ∗ B is well defined, independent of A and B, contains
isomorphic copies of A and B with intersection equals to Zpn · 1 and satisfies the
universal property of free product of algebras.
Definition 5. Let C1 = A ⊲⊳ V and C2 = B ⊲⊳ V . We define C = C1 ∗ C2 as
V
C = (A ∗ B) ⊲⊳ V . Let A, B and V be free bases of A, B, and V , correspondingly.
Any element is a Zpn -combination of wv, where w is alternating word of letters from
A′ , B′ and v ∈ V. In order to define multiplication on C it suffices to define vw as a
combination of wi vi . It could be done using Zappa-Zsep product structure of C1 and
C2 . Suppose, for example that w = a1 b1 . . . ak bk then
X
X
a1i1 b1i1 ,j1 vi1 ,j1 . . . ak bk =
a1i1 vi1 b1 . . . ak bk =
va1 b1 . . . ak bk =
i1 ,j1
i1
X
a1i1 b1i1 ,j1 . . . aki1 ,...jk−1 ,ik bki1 ,...,jk vi1 ,...,jk
i1 ,j1 ,...,ik ,jk
Exercise 4. Check that C1 ∗ C2 is well defined, contains isomorphic copies of C1
V
and C2 such that C1 ∩ C2 = V . Check that C1 ∗ C2 satisfies the universal property of
V
amalgamated free products of algebras.
Consider A0 = Zpn [x0 , x1 ]/I(g0 ), where I(g0 ) is the ideal generated by g0 of Eq.(1).
Proposition 1. A0 = Zpn [x0 ] ⊲⊳ Zpn [x1 ]
6
We prove the proposition in Section 7. Now let as show how Theorem 1 follows
from Proposition 1. Let A1 = Z[x1 , x2 ]/I(g1 ). Clearly, A1 is isomorphic to A0 , so
A1 = Zpn [x1 ] ⊲⊳ Zpn [x2 ]. Consider A01 = A0 ∗Zpn [x1 ] A1 . Notice, that the roles of
Zpn [x1 ] in A0 and A1 are different, precisely the isomorphism from A0 to A1 maps
Zpn [x1 ] to Zpn [x2 ] not to Zpn [x1 ] of A1 .
Lemma 4. A01 = Zpn [x0 , x2 ] ⊲⊳ Zpn [x1 ] is isomorphic to Zpn [x0 , x1 , x2 ]/I(g0, g1 ).
Proof. By construction g0 = g1 = 0 in A01 . So, the map xi → xi prolongs to a
surjective homomorphism φ : Zpn [x0 , x1 , x2 ]/I(g0 , g1 ) → A01 . Using the universal
property of A01 define ψ : A01 → Zpn [x0 , x1 , x2 ]/I(g0 , g1 ). Notice, that ψ is surjective
as well (Zpn [x0 , x1 , x2 ]/I(g0 , g1 ) is generated by xi ). Check that ψφ = id : A01 → A01
by the universal property.
Similarly, the algebras A2 = Zpn [x2 , x3 ]/I(g2 ), A3 = Zpn [x3 , x0 ]/I(g3 ), and A23 =
A2 ∗ A3 = Z[x2 , x0 ] ⊲⊳ Z[x3 ] may be constructed. Notice, that there exist isomorZpn [x3 ]
phism A01 → A23 that sends x0 → x2 , x2 → x0 , x1 → x3 . Now, we may construct
A = A01 ∗Zpn [x0 ,x2 ] A23 (we make isomorphism Z[x0 , x2 ] of A01 with Z[x2 , x0 ] of A23
sending x0 → x0 and x2 → x2 ). Theorem 1 follows from the lemma.
Lemma 5. A = Zpn [x0 , x2 ] ⊲⊳ Zpn [x1 , x3 ] is isomorphic to Zpn [x0 , x1 , x2 , x3 ]/I(g0 , . . . , g3 ).
7
Proof of Proposition 1
The proof is based on the fact that a polynomial g0 forms a kind of a Grobner basis
(over non-commutative polynomials). We will not define what a Grobner basis is for
non-commutative polynomials. Instead, we directly apply a Knuth-Bendix algorithm
[1] to {g0 }.
Let m(y1 , . . . , yk ) be a non-commutative monomial or, the same, a word in alphabet
{y1 , . . . , yk }, that is, m(y1 , . . . , yk ) = z1 z2 . . . zr , zi ∈ {y1 , . . . , yk }. The product of two
monomials is just the concatenation. A non-commutative
polynomial f over Zpn is
P
a “linear” combination of monomials f =
ai mi , ai ∈ Zpn . The product f1 f2 of
polynomials is defined using the product of monomials by linearity.
Let us return to the study of A0 = Zpn [x0 , xP
1 ]/I(g0 ). We call a polynomial
P ′ ifj∈
ai,j x1 x0 ).
ai,j xi0 xj1 (resp. f =
Zpn [x0 , x1 ] left (resp. right) reduced if f =
Proposition 1 is equivalent to the claim.
Claim 1. For any f ∈ Zpn [x0 , x1 ] there exists a left (resp. right) reduced polynomial
f˜ such that f − f˜ ∈ I. If f ∈ I is left (right) reduced then f = 0.
From this point we restrict ourselves to the left case. The right case may be
considered similarly. Define the one step reduction based on equalities g0 = 0:
ρ
x1 x0 −→ x0 x1 − Q0 (x1 ) − pQ1 (x0 , x1 )
Let f, f˜ ∈ Zpn [x0 , x1 ].
7
ρ
• f˜ is a one step reduction of f (f −→ f˜) if an appearance of x1 x0 , in some monomial of f is changed according to the above described rule; f˜ is the resulting
polynomial (after applying associativity and linearity).
• f is said to be terminal if no monomial of f contains x1 x0 . It means that we
ρ
are unable to apply −→ to f .
ρ
• We write f =⇒ f˜ (f˜ is a reduction of f ) if there is a sequence f0 = f, f1 , . . . , fk =
ρ
f˜ such that fi −→ fi+1 for i = 0, . . . , k − 1.
ρ
ρ
• We write f =⇒ f˜+ if f =⇒ f˜ and f˜ is terminal.
Proposition 2.
• f is terminal if and only if f is left reduced.
ρ
ρ
• There is no infinite sequence f0 −→ f1 −→ f2 . . .
ρ
• For any non-terminal f there exist a unique f˜ such that f =⇒ f˜+.
ρ
ρ
Two last items mean that any sequence f −→ f1 −→ . . . of one step reduction terminates and the terminal polynomial depends only on f .
We prove the proposition in Subsection 7.1. Now let us show how Proposition 2
ρ
ρ
ρ
implies Claim 1. We use notation f = −→ f ′ (resp. f = =⇒ f ′ ) to denote f −→ f ′
ρ
or f = f ′ (resp. f =⇒ f ′ or f = f ′ ).
ρ
ρ
ρ
Lemma 6. The map f = =⇒ f˜+ is linear, that is, if f1 = =⇒ f˜1 + and f2 = =⇒ f˜2 +
ρ
then a1 f1 + a2 f2 = =⇒ a1 f˜1 + a2 f˜2 , where a1 , a2 ∈ Zpn .
ρ
Proof. Let a1 f1 + a2 f2 −→ f ′ . It means that we apply reduction to a monomial m of
a1 f1 + a2 f2 . The monomial m may appear in f1 , f2 , or in the both polynomials. In
ρ
ρ
any case there exist f1′ and f2′ such that f1 = −→ f1′ , f2 = −→ f2′ and f ′ = a1 f1′ + a2 f2′ .
ρ
It follows by induction that if a1 f1 + a2 f2 =⇒ f ′ then f ′ = a1 f1′ + a2 f2′ for some f1′ , f2′
ρ
ρ
such that f1 = =⇒ f1′ and f2 = =⇒ f2′ . Now suppose that a1 f1′ + a2 f2′ is terminal but,
say, f1′ is not terminal. There are two possibilities:
1. a1 f1′ is terminal. In this case, collecting terminal monomials, we may write
f1′ = α + β with α terminal and a1 β = 0. In this case further reduction of f1′
does not change a1 f1′ . So, w.l.g. we may assume f1′ to be terminal.
2. a1 f1′ is not terminal. In this case we may write fi′ = αi + βi with αi terminal
and a1 β1 + a2 β2 = 0. Now, apply the same reduction to β1 and β2 , keeping the
sum a1 f1′ + a2 f2′ unchanged.
We are done by Proposition 2.
ρ
Lemma 7. f =⇒ 0+ if and only if f ∈ I(g0 , . . . , g3 ).
ρ
Proof. Only if. It is clear by construction
that f =⇒ f˜ implies that f − f˜ ∈ I.
P
If. Let f ∈ I. It means that f = αi g0 βi . Applying associativity we may write f =
P
ai mi g0 m′i , where mi and m′i are monomials and ai ∈ Zpn . By construction, there is
ρ
a one step reduction mi g0 m′i −→ 0. We are done by Lemma 6 and Propostion 2.
8
7.1
Proof of Proposition 2.
The first item of Proposition 2 is straightforward. So we start with the proof of the
second item of the proposition. It is the most difficult part of the proposition and
will be used for the proof of the third item.
7.1.1
Proof of the second item.
Let t = aj m(x0 , x1 ) be a term (a monomial with a coefficient). We are going to
measure how an application of one step reduction makes a term more close to a left
reduced polynomial. To this end we define:
• |t| = min{k ∈ N | pk t = 0}.
• n0 (t) – number of x0 in t, for example, n0 (xi1 xj0 ) = j.
• def (t) – the defect of t, the total number of pairs where x1 appears before x0 ,
for example, def (xj1 xk0 xr1 xm
0 ) = jk + jm + rm.
To each term we associate the ordered triple (|t|, n0 (t), def (t)). On the set of triple
we consider lexicographical order: (α, β, γ) < (α′ , β ′, γ ′ ) iff
• α < α′ ; or
• α = α′ and β < β ′ ; or
• α = α′ , β = β ′ , and γ < γ ′ .
ρ
Now one may check that (|t|, n0 (t), def (t)) > (|tj |, n0 (tj ), def (tj )) if t =⇒
need the following result.
P
j tj .
We
Lemma 8 (Dickson). Any decreasing (with respect to lexicographic order) sequences
in N3 is finite.
Consider now the reduction process of t as a tree: To each vertex we associate a
term in such a way that in any reduction step the resulting polynomial is a sum of
terms of the leafs of the tree. With root we associate t. For each reduction of term
t′ in a leaf l we connect the leaf l with new leafs with all terms appearing in the
reduction. Any descending path in this tree is finite by Dickson Lemma. This tree is
k-regular by construction, so the tree is finite and the reduction process terminates.
7.1.2
Proof of the third item.
This uses the Newman’s lemma, or Diamond lemma for reduction processes.
∗
Suppose that on a set X a reduction process · −→ · (just a relation on X) is defined.
∗
Denote by =⇒ it’s transitive closure. We say that x is terminal if there are no y ∈ X
∗
∗
∗
such that x =⇒ y As before, let x =⇒ y+ denotes x =⇒ y and y is terminal.
∗
Lemma 9 (Diamond lemma). Let −→ satisfies the following properties:
9
∗
∗
• Any sequence x1 −→ x2 −→ ... is finite.
∗
∗
• −→ is locally confluent, that is, for any x, y1 and y2 such that x −→ y1 and
∗
∗
∗
x −→ y2 there exists z ∈ X such that y1 =⇒ z and y2 =⇒ z+.
∗
Then =⇒ is globally confluent, that is, for any non-terminal x there exists unique y
∗
such that x =⇒ y.
So, in order to show Proposition 1 it suffices to check the second condition of
ρ
ρ
ρ
the Diamond Lemma for −→. Let f −→ f1 and f −→ f2 . If the reduction apρ
ρ
plies to a different terms then existence of f3 , f1 −→ f3 and f2 −→ f3 is trivial. It
ρ
suffices to consider f = axi1 ...xim . Suppose w.l.g., that f −→ f1 is an applicaρ
tion of reduction to xij xij+1 = x1 x0 and f −→ f2 xik xik+1 = x1 x0 for k > j + 1.
Then f1 = axi1 . . . xij−1 qxij+2 . . . xim and f2 = axi1 . . . xik−1 qxik+2 . . . xim , where q =
ρ
ρ
x0 x1 − Q0 (x1 ) − pQ1 (x0 , x1 ). One may check that f1 =⇒ f3 and f2 =⇒ f3 for f3 =
axi1 . . . xij−1 qxij+2 . . . xik−1 qxik+2 . . . xim .
8
Proof of Lemma 2
Let Z(x̄) be an algebra of power series with noncomutative (but associative) variables
x̄ = x0 , x1 , . . . , xm over Z. For a, b ∈ Z(x̄) let ⌈a, b⌉ = ab − ba and Λ[x̄] be a
submodule of Z[x̄] generated by ⌈·, ·⌉ starting from x̄. Let Λj ⊂ Λ[x̄] consist of
uniform polynomials of order j. So,
j
Λ[x̄] = ∪∞
j=0 Λ .
Let I = I(x̄) be a (two-sided) ideal in Z(x̄) generated by x̄. Clearly, this ideal consists
of polynomials without constant term. Let G be a group. Notations Gn and G[n] are
defined in Section 2.
Theorem 2 (Magnus’ theorem).
group, freely generated by ai .
• Let ai = 1 + xi . The group F = hai i is a free
• (1 + I n ) ∩ F = F[n] .
• If w ∈ F[n] then w = 1 + d + z, where d ∈ Λn and z does not contain terms of
order ≤ n.
• For any d ∈ Λn there exists z ∈ Z[x̄] without terms of order ≤ n such that
1 + d + z ∈ F[n] .
Consider homomorphism π : Z(x̄) → Z(x̄), defined by π(xi ) = pxi . Clearly,
π(Z(x̄)) = Z(px̄). Also, π(F ) is an inclusion of a free group F into Z(px̄). For a
two sided ideal J of Z(px̄) let Nj = {w ∈ π(F ) | w − 1 ∈ J}.
Lemma 10. Nj ⊳ π(F ).
10
Clearly, Z(x̄)/pn Z(x̄) ≡ Zpn (x̄). Denote (pn ) = pn Z(x̄) ∩ Z(p̄x). Notice, that
Zpn (px̄) = Zpn [px̄]. Now, Lemma 2 is a consequence of the following theorem.
Theorem 3 (Jacobson, [6]). N(pn ) = π(Fn )
Proof. We present here the Jacobson proof (see [6]) which is a reduction to Theorem 2.
In [6] the definition of (pn ) is different and not equivalent of ours. But the proof in
[6] is, actually, for our definition of (pn ).
Notice, that (1 + pm w . . . )p = 1 + pm+1 . . . and [(1 + pm w . . . ), (1 + pk u . . . )] =
1+pm+k (wu−uw) . . . where the omitted terms are of higher p-order. This implies that
p
N(p
m ) ⊆ N(pm+1 ) and [N(pm ) , N(pk ) ] ⊆ N(pm+1 ) . Consequently, we have π(Fn ) ⊆ N(pn )
and N(pn+1 ) ⊆ N(pn ) .
According to [6] we show the equality N(pn ) = π(Fn ) by induction. By definition,
N(p1 ) = π(F1 ) = π(F ). Suppose, that N(pn ) = π(Fn ). Then we know that π(Fn+1 ) ⊆
N(pn+1 ) ⊆ π(Fn ). So, it suffices to show that π(Fn )∩π(Fn+1 ) ⊇ N(pn+1 ) ∩π(Fn ), or, the
same, to prove that if w ∈ π(Fn )\π(Fn+1 ) then w 6∈ N(pn+1 ) . Let w ∈ π(Fn )\π(Fn+1 ).
There exists a unique i such that w ∈ π(F[i] ) \ π(F[i+1] ). By Theorem 2 w = 1 +
pj di (px̄) + z where di ∈ Λi , z has x̄-order more than i. Also we have that i + j ≥ n
(as w ∈ N(pn ) ). Applying once again Theorem 2 we find u = 1 + di (px̄) + z ′ ∈ π(F[i] ),
j
where the x̄-order of z ′ is more than i. So, w = up w1 , where w1 ∈ π(F[i′ ] ) for i′ > i.
jk
j1
j2
Repeating this procedure one gets w = up1 up2 . . . upk wk , where wk ∈ π(F[n+1] ) and
ur = (1 + dir (px̄) . . . ) with dir ∈ Λir (term of higher order in x̄ are omitted) and
jr
jr
ir + jr ≥ n. Now, ur ∈ π(F[ir ] ) and, consequently, upr ∈ π(F[ipr ] ⊆ Fir +jr ). If
∀r ir + jr > n then w ∈ π(Fn+1 ), so, by our assumptions, ir + jr = n for some r. It
jk
j1
implies that w = up1 . . . upk wk = (1 + pj1 di1 (px̄) + pj2 di2 (px̄) . . . ) 6∈ N(pn+1 ) .
Let F = ha0 , . . . , am i be a free group on {a0 , . . . , am }. Let Zpn [F/Fn ] be a group
algebra of F/Fn over Zpn . Theorem 3 implies that there exists unique homomorphism
φ : Zpn [F/Fn ] → Zpn [px̄] such that φ(ai ) = 1+pxi . (Here we, abusing notation, denote
by the same symbol ai its image in F/Fn .) Moreover, (ker(φ) + 1) ∩ F/Fn = {1}. Still
j
ker(φ) is not trivial. For example, if w ∈ F/Fn and w p = 1 then pj (w − 1) ∈ ker(φ).
What is the structure of ker(φ)? For example, is it true that ker(φ) is generated by
j
{pj (w − 1) | w ∈ F/Fn , w p = 1}?
References
[1] Franz Baader, Tobias Nipkov. Term Rewriting and All That, Cambridge University Press 1998. xii+301 pp. ISBN: 0-521-45520-0; 0-521-77920-0 .
[2] Lev Glebsky and Igor E. Shparlinski. Short cycles in repeated exponentiation
modulo a prime. Des. Codes Cryptogr., 56(1):3542, 2010.
[3] Gupta, N. Lectures on Dimension Subgroups. Resenhas IME-USP, 1996, Vol.2,
No.3, 263-273
11
[4] Harald A. Helfgott and Kate Juschenko. Soficity, short cycles and the Higman
group. Preprint arXiv:1512.02135
[5] Graham Higman. A finitely generated infinite simple group. J. London Math.
Soc., 26:6164, 1951.
[6] Jacobson, N.(1-YALE) Magnus’ method in the theory of free groups. Ulam
Quart. 1 (1992), no. 1,
[7] Magnus, Wilhelm; Karrass, Abraham; Solitar, Donald Combinatorial group theory. Presentations of groups in terms of generators and relations. Dover Publications, Inc., Mineola, NY, 2004. xii+444 pp. ISBN: 0-486-43830-9 Reprint of
the 1976 second edition.
[8] Alexandre Martin,
arXiv:1506.02837
On the cubical geometry of Higman’s group.
Preprint
[9] Schupp, Paul E. Small cancellation theory over free products with amalgamation.
Math. Ann. 193 (1971), 255264
12
| 4math.GR
|
arXiv:1510.03611v2 [math.RT] 15 Feb 2018
Restriction of representations of GL(n + 1, C) to
GL(n, C) and action of the Lie overalgebra
Yury A. Neretin
1
1
The statement
Abstract. Consider a restriction of an irreducible finite dimensional holomorphic
representation of GL(n + 1, C) to the subgroup GL(n, C). We write explicitly formulas
for generators of the Lie algebra gl(n + 1) in the direct sum of representations of
GL(n, C). Nontrivial generators act as differential-difference operators, the differential
part has order n − 1, the difference part acts on the space of parameters (highest
weights) of representations. We also formulate a conjecture about unitary principal
series of GL(n, C).
1.1. The Gelfand–Tsetlin formulas. It is well known that restrictions
of finite dimensional holomorphic representations of the general linear group
GL(n, C) to the subgroup GL(n − 1, C) is multiplicity free. Considering a chain
of restrictions
GL(n, C) ⊃ GL(n − 1, C) ⊃ GL(n − 2, C) ⊃ · · · ⊃ GL(1, C)
we get a canonical decomposition of our representation into a direct sum of
one-dimensional subspaces. Taking a vector in each line we get a basis of the
representation. In [7] Gelfand and Tsetlin announced formulas for action of
generators of the Lie algebra gl(n) in this basis. It turns out that the Lie
algebra gl(n) acts by difference operators in the space of functions on a certain
convex polygon in the lattice Zn(n−1)/2 . In particular, this gives an explicit
realization of representations of the Lie algebra gl(n).
For various proofs of the Gelfand–Tsetlin formulas and constructions of the
bases, see [3], [30], [2], [11], [31], [19], [17], [18] (see more references in [17]), for
other classical groups, see [8], [17], [28], for applications to infinite-dimensional
representations, see [27], [13], [9], for formulas on the group level, see [5], [9].
There are many other continuations of this story. However, our standpoint is
slightly different.
1.2. Actions of overalgebras in the spectral decompositions. We
can write images of many operators under the classical Fourier transform. It
is commonly accepted that Plancherel decompositions of representations are
higher analogs of the Fourier transform.
Consider a group G and its subgroup H. Restrict an irreducible unitary
representation of G to H. Generally, an explicit spectral decomposition of the
restriction seems hopeless problem. However, there is a collection of explicitly
1 Supported
by the grant FWF, P25142, P28421
1
solvable problems of this type2 . In [20] it was conjectured that in such cases
the action of the Lie algebra g can be written explicitly as differential-difference
operators. In fact, in [20] there was considered the tensor product Vs ⊗ Vs∗ of
a highest weight and a lowest weight unitary representations of SL(2, R) (i.e.,
G ≃ SL(2, R) × SL(2, R) and H ≃ SL(2, R) is the diagonal). This representation
is a multiplicity free integral over the principal series of SL(2, R). It appears that
the action of all generators of the Lie algebra sl(2, R) ⊕ sl(2, R) in the spectral
decomposition can be written explicitly as differential-difference operators. The
differential part of these operators has order two and the difference operators
are difference operators in the imaginary direction3 .
Molchanov [12]–[15] solved several problems of this type related to rank-one
symmetric spaces4 . In all his cases the Lie overalgebra acts by differentialdifference operators; difference operators act in the imaginary dimension. However, the order of the differential part in some cases is 4. In [25], [26] the author
considered the Fourier transform on the group GL(2, R). There were evaluated
the operators in the target space corresponding to multiplications by matrix elements and partial derivatives with respect to matrix elements. Formulas have
similar structure, but in [26] they are simpler than in previous cases.
In the present paper, we consider restrictions of holomorphic finite-dimensional
representations GL(n + 1, C) to GL(n, C) and write explicit formulas for the action of the overalgebra in the spectral decomposition. Also, we formulate a
conjecture concerning restrictions of unitary principal series.
1.3. Notation. The group GL(n, C) and its subgroups. Denote by
GL(n, C) the group of invertible complex matrices of size n. By gl(n) we denote
its Lie algebra, i.e. the Lie algebra of all matrices of size n. Let U(n) ⊂ GL(n, C)
be the subgroup of unitary matrices.
By Eij we denote the matrix whose ij’s entryPis 1 and other entries are 0.
The unit matrix is denoted by 1 or 1n , i.e. 1 = j Ejj . Matrices Eij can be
regarded as generators of the Lie algebra gln .
Denote by N + = Nn+ ⊂ GL(n, C) the subgroup of all strictly upper trian2 Also,
some explicitly solvable spectral problems in representation theory can be regarded
as special cases of the restriction problem. In particular, a decomposition of L2 on a classical
pseudo-Riemannian symmetric space G/H can be regarded as a special case of the restriction
e to the symmetric subgroup G, [21]. So,
of a Stein type principal series of a certain overgroup G
for L2 on symmetric spaces the problem of action of the overalgebra discussed below makes
sense.
3 On Sturm–Liouville in imaginary direction, see [23] and further references in that paper,
see also [10]. The most of known appearances of such operators are related to representation
theory and spectral decompositions of unitary representations.
4 In particular, he examined the restrictions of maximally degenerate principal series for
the cases GL(n + 1, R) ⊃ O(n, 1), O(p, q) ⊃ O(p − 1, q).
2
gular matrices of size n, i.e.,
1
0
Z = ...
0
0
matrices of the form
z12
1
..
.
...
...
..
.
z1(n−1)
z2(n−1)
..
.
0
0
...
...
1
0
z1n
z2n
..
.
.
z(n−1)n
1
By Nn− we denote the group of strictly lower triangular matrices.
By Bn− ⊂ GL(n, C) we denote the subgroup of lower triangular matrices,
g ∈ Bn− if gij = 0 if i < j. Let ∆ be the subgroup of diagonal matrices, we
denote its elements as δ = diag(δ1 , . . . , δn ).
Next, we need a notation for sub-matrices. For a matrix X denote by
[X]αβ
the left upper corner of X of size α × β. Let I = {i1 , . . . , iα }, J = {j1 , . . . , jβ }
be collections of integers, we assume that their elements are ordered as 0 < i1 <
i2 < . . . iα , J : 1 6 j1 6 . . . 6 jβ . Denote by
I
i1 , . . . , iα
X
= X
(1.1)
J
j1 , . . . , jβ
the matrix composed of entries xiµ ,jν . By
[X(I)] = X(ii , . . . , iα )
(1.2)
we denote matrix composed of ii ,. . . , iα -th rows of X (the order is not necessary
increasing, also we allow coinciding rows).
1.4. Holomorphic representations of GL(n, C). Recall that irreducible
finite dimensional holomorphic representations ρ of the group GL(n, C) are enumerated by collections of integers (signatures)
p : p1 > . . . > pn .
This means that there is a cyclic vector v (a highest weight vector) such that
ρ(Z)v = v
for each Z ∈ Nn+ ,
n
Y
p
where δ = diag(δ1 , . . . , δn ).
ρ(δ)v =
δj j v,
j=1
Denote such representation by ρp = ρnp . Denote the set of all signatures by Λn .
The dual (contragradient) representation to ρp has the signature
p∗ := (−pn , . . . , −p1 ).
(1.3)
1.5. Realizations of representations by differential operators. Recall a model for irreducible finite dimensional holomorphic representations for
3
details, see Section 2). We consider the space Pol(Nn+ ) of polynomials in the
variables zij . The generators of the Lie algebra gl(n) act in this space via
differential operators
X
X
Ekk = −
zik ∂ik + pk +
zkj ∂kj ,
(1.4)
i<k
Ek(k+1) = ∂k(k+1) +
E(k+1)k =
X
X
j>k
zik ∂i(k+1) ,
(1.5)
i<k
zi(k+1) ∂ik + (pk+1 − pk )zk(k+1) − zk(k+1)
i<k
+
X
m>k+1
z
det k(k+1)
1
zkm
z(k+1)m
X
zkj ∂kj +
j>k
∂(k+1)m ,
(1.6)
where
∂
.
∂zkl
All other generators can be expressed in the terms of Ekk , Ek(k+1) , E(k+1)k
(also it is easy to write explicit formulas as it is explained in Section 2). In this
way, we get a representation of the Lie algebra gl(n).
There exists a unique finite dimensional subspace Vp invariant with respect
to such operators. The representation ρp is realized in this space. The highest
weight vector is f (Z) = 1. In the next subsection we describe this space more
explicitly.
Remark. This approach arises in [6]. Of course, this model is a coordinatization of the constructions of representations of GL(n, C) in sections of line
bundles over the flag space Bn− \ GL(n, C) (as in the Borel–Weil–Bott theorem).
The space Nn+ ≃ Cn(n−1)/2 is an open dense chart on this space, the terms
with ∂ in (1.4)–(1.6) correspond to vector fields on the flag space, the zero order
terms are corrections corresponding to the action in the bundle. Elements of the
space Vpn are precisely polynomials, which are holomorphic as section of bundle
on the whole flag space. However, we need explicit formulas and prefer a purely
coordinate language.
⊠
1.6. Descriptions of the space Vp . Zhelobenko operators.
∂kl :=
a) Description-1. Denote by dŻ the standard Lebesgue measure on Nn+
Y
dŻ =
d Re zkl d Im zkl .
k<l
Denote by dµ(Z) the measure on Nn+ given by the formula
dµp (Z) = dµnp (Z) =
n−1
Y
det [Z]jn ([Z]jn )∗
j=1
−(pj −pj−1 )−2
Proposition 1.1 a) Vpn = L2 (Nn+ , µnp ) ∩ Pol(Nn+ ).
b) The L2 -inner product in Vpn is U(n)-invariant.
4
dŻ.
(1.7)
Proof is given in Section 2.
Description-2. We define the Zhelobenko operators by
X
zmj ∂kj ,
m > k.
Rkm := ∂km +
(1.8)
j>m
Theorem 1.2 The space Vpn consists of polynomials satisfying the conditions
n
p −p
j
j+1
Rj(j+1)
+1
f (Z) = 0.
See Zhelobenko, [30], [29], §6, Theorem 2, [32], Theorem 48.7.
Description-3. There is one more description of the space Vpn , which can
be used with coordinate language.
Proposition 1.3 The space Vpn coincides with the space determined by the reproducing kernel
n−1
Y
pj −pj−1
.
det [Z]jn [Z]∗jn
j=1
On Hilbert spaces determined by reproducing kernels, see, e.g., [22], Sect.
7.1. The proposition is more-or-equivalent to the Borel–Weil theorem.
1.7. The restriction of representations GL(n+1, C) to GL(n, C). Consider the representation ρn+1
of GL(n+1, C) with a signature r = (r1 , . . . , rn+1 ).
r
It is well-known (see Gelfand, Tsetlin [7]) that the restriction of ρn+1
to GL(n, C)
r
is multiplicity free and is a direct sum of all representations ρq of GL(n, C) with
signatures satisfying the interlacing conditions
r1 > q1 > r2 > q2 > r3 > . . . > qn > rn+1 .
(1.9)
Our purpose is to write explicitly the action of Lie algebra gl(n + 1) in the
space ⊕Vqn .
1.8. Normalization. First, we intend to write a GL(n, C)-invariant pairing
Jp,q : Vpn+1 × Vqn → C.
as an integral
Jp,q (ϕ, ψ) =
Z
+
+
Nn+1
×Nn
Lp,q (U, Z)ϕ(Z) ψ(U ) dµn+1
(Z) dµnq (U ),
p
+
where Z ranges in Nn+1
, U ranges in Nn+ , and the kernel
Lp,q (U, Z) ∈ Vpn+1 ⊗ Vqn
is a polynomial in the variables zij , ukl .
5
(1.10)
Denote Z cut := [Z](n+1)n , this matrix is obtained from Z by cutting of the
last column. Denote by U ext the n × (n + 1)-matrix obtained from U by adding
the zero last column,
1 u12 . . . u1n 0
0 1 . . . u2n 0
U ext = .
(1.11)
.
..
..
..
..
.
.
.
0 0 ...
1
0
Consider the (n + 1) × (n + 1)-matrix composed from the first (n + 1 − α) rows
of the matrix Z and the first α rows of the matrix U ext . Denote by Φα the
determinant of this matrix:
[Z](n+1−α)(n+1)
Φα (Z, U ) = det
,
(1.12)
[U ext ]α(n+1)
Consider the n × n-matrix composed of the first n − α rows of the matrix
Z cut and the first α rows of the matrix U , denote by Ψα its determinant:
cut
[Z ](n−α)n
.
(1.13)
Ψα (Z, U ) = det
[U ]αn
Proposition 1.4 Consider signatures p = (p1 , . . . , pn+1 ), q = (q1 , . . . , qn )
such that p and q∗ are interlacing. Then the expression (1.10) with the kernel
p
Lp,q (U, Z) = Φp1n +q1 Ψ1−pn −q2 Φ2n−1
+q2
−pn−1 −q3
Ψ2
. . . Φpn1 +qn .
(1.14)
determines a GL(n, C)-invariant nonzero pairing between Vpn+1 and Vqn .
For instance, for n = 3 we get
p +q
−p3 −q2
1 z12 z13 z14 3 1
1 z12 z13
0
1
z23 z24
Lp,q (U, Z) = det
×
det 0
1
z23
0
0
1
z34
1 u12 u13
1 u12 u13
0
p +q
p2 +q2
−p2 −q3
1 z12 z13 z14 1 3
1 z12 z13 z14
1 z12 z13
1 u12 u13
0
0
1
z23 z24
det 1 u12 u13
.
det
det
1 u12 u13
0
1
u23
0
0
0
1
u23
0
0
1
0
0
1
u23
0
Next we pass to the dual signature
r = p∗
(below p and q are rigidly linked by this restraint), and represent Lp,q in the
form
Lrq (U, Z) := Lp,q (U, Z) = Φq11 −r2 Ψ1r2 −q2 Φq22 −r3 Ψ2r3 −q3 . . . Φqnn −rn+1 .
6
(1.15)
1.9. Action of the overalgebra. Fix a signature r = p∗ ∈ Λn+1 as in the
previous subsection. Consider the space
M
Vqn .
(1.16)
Vr :=
q1 ,...,qn : rj >qj >rj−1
We can regard elements of this space as ’expressions’
F (U, q)
of n(n − 1)/2 continuous variables ukl , where 1 6 k < l 6 n, and of integer
variables q1 , . . . , qn . Integer variables range in the domain rj > qj > rj+1 . More
precisely, for a fixed q the expression F (U, q) is a polynomial in the variables
ukl , moreover, this polynomial is contained in Vq .
Our family of forms Lp,q determines a duality between Vr∗ and the space
Vr , hence we get a canonical identification of Vr and Vr . Therefore, we have
a canonically defined action of the Lie algebra gl(n + 1) in Vr . We preserve
the notation Ekl for operators in Vp and denote operators in Vr by Fkl . For
n
1 6 k, l 6 n the operators Fkl
act in the space Vr by the first order differential
operators in U with coefficients depending on q according the standard formulas
(see (1.4)–(1.6)). For instance,
X
X
Fkk = −
uik ∂ik + qk +
ukj ∂kj ,
(1.17)
i<k
Fk(k+1) = ∂k(k+1) +
F(k+1)k =
X
X
j>k
uik ∂i(k+1) ,
(1.18)
i<k
X
ui(k+1) ∂ik + (qk+1 − qk )uk(k+1) − uk(k+1)
i<k
+
X
m>k+1
u
det k(k+1)
1
ukm
u(k+1)m
ukj ∂kj +
(1.19)
j>k
∂(k+1)m ,
(1.20)
where ∂kl := ∂u∂kl .
The purpose of this work is to present formulas for the generators F1(n+1) ,
F(n+1)n . Together with (1.17)–(1.20) they generate the Lie algebra gl(n+1), formulas for the remaining generators Fj(n+1) , F(n+1)j consist of similar aggregates
as below, but are longer.
Denote by Tj± the following difference operators
Tj± F (. . . , qj , . . . ) = F (. . . , qj ± 1, . . . )
(the remaining variables do not change). We will write expressions, which are
polynomial in ukl , ∂kl , linear in Tj± and rational in qj . These expressions satisfy
the commutation relations in gl(n) and preserve the space Vr .
For 1 6 k < l 6 n denote by [k, l] the set
[k, l] := {k, k + 1, . . . , l}.
7
For a set
I : 1 6 i1 < i2 < · · · < im
we write
I ⊳ [k, l]
if i1 = k, im = l.
(1.21)
More generally, if I ⊂ J ⊂ [1, n], we write I ⊳ J if the minimal (resp. maximal)
element of I coincides with the minimal (resp. maximal) element of J.
For any I ⊳ [α, β] we define the operator
RI := Ri1 i2 Ri2 i3 . . . Rim−1 im = Rαi2 Ri2 i3 . . . Rim−1 β ,
where Rij are the Zhelobenko operators. We also set
Rkk := 1.
Theorem 1.5 a) The generator F1(n+1) acts by the formula
F1(n+1) =
n
X
Am
m=1
X
Y
I⊳[1,m] l∈[1,m]\I
−
(qm − ql + l − m) · RI Tm
,
(1.22)
where coefficients Am are given by
Am =
Qn+1
j=m+1 (qm
− rj + j − m − 1)
Q
α6=m (qm
− qα + α − m)
.
b) The generator F(n+1)n is given by the formula
F(n+1)n = −
n
X
m=1
where
Bm
X
Y
I⊳[m,n] l∈[m,n]\I
+
,
(qm − ql − m + l + 1) · RI Tm
Qm
Bm = Q
j=1 (qm
α6=m (qm
− rj + j − m)
− qα − m + α)
(1.23)
.
1.10. Further structure of the paper. Section 2 contains preliminaries
on holomorphic representations of GL(n, C). In Section 3, we verify the formula
for the kernel L(Z, U ). Our main statement is equivalent to a verification of
differential-difference equations for the kernel L(Z, U ), this is done in Section
4. In Section 5, we formulate a conjecture about analog of our statement for
unitary representations.
Acknowledgements. Fifteen years ago the topic of the paper was one of
aims of a joint project with M. I. Graev, which was not realized in that time
(I would like to emphasis his nice paper [9]). I am grateful to him and also to
V. F. Molchanov for discussions of the problem.
8
2
Holomorphic representations of GL(n, C).
2.1. Realization in the space of functions on GL(n, C). For details
and proof, see [30]. We say that a function f (g) on GL(n, C) is a polynomial
function if it can be expressed as a polynomial expression in matrix elements gij
and det(g)−1 . Denote the space of polynomial functions by C[GL(n, C)]. The
group GL(n, C) × GL(n, C) acts in C[GL(n, C)] by the left and right shifts‘:
λlef t−right (h1 , h2 )f (g) = f (h−1
1 gh2 ).
This representation is a direct sum
λ ≃ ⊕p∈Λn ρp∗ ⊗ ρp .
A summand Vp∗ ⊗ Vp is the GL(n, C) × GL(n, C)-cyclic span of the vector
∆[p] =
n−1
Y
p −pj+1
det[g]jjj
· (det g)pn .
j=1
If h1 ∈ Nn− and h2 ∈ Nn+ , then λ(h1 , h2 )∆[p] = ∆[p].
Next, consider the space of polynomial functions invariant with respect to
left shifts on elements of Nn−
f (h−1 g) = f (g).
(2.1)
Equivalently, we consider the space of polynomial functions on C[Nn− \GL(n, C)].
Each ρp∗ has a unique Nn− -invariant vector, therefore C[Nn− \ GL(n, C)] is a
multiplicity free direct sum ⊕p ρp .
Fix a signature p and consider the space of Nn− -invariant functions Hp such
that for any diagonal matrix δ, we have
f (δg) = f (g)
n
Y
p
δj j .
j=1
The group GL(n, C) acts in Hp by the right shifts,
λright (h)f (g) = f (gh).
(2.2)
This representation is irreducible and equivalent to ρp . The vector ∆p is its
highest weight vector.
2.2. Realization in the space of functions on Nn+ . An element g ∈
GL(n, C) satisfying the condition
det[g]jj 6= 0
for all j
admits a unique Gauss decomposition
g = bZ,
where b ∈ Bn− , Z ∈ Nn+ .
9
Notice that
det[g]jj
,
det[g](j−1)(j−1)
1 2 ... i − 1 j
det g
1 2 ...
i
j
.
zij =
1 2 ... i j
det g
1 2 ... i j
bjj =
(2.3)
(2.4)
Restrict a function f ∈ Hp to the subgroup Nn+ . We get a polynomial in
the variables zij , where i < j. By the Gauss decomposition, f is uniquely
determined by this restriction. Therefore the space Vp can be regarded as a
subspace of the space of polynomial in zij . The description of this space is
given by the Zhelobenko Theorem 1.2.
Denote the factors in the Gauss decomposition of Zg in the following way
Zg =: b(Z, g) · Z [g] ,
where b(Z, g) ∈ Bn− , Z [g] ∈ Nn+ ,
Then
ρp (g) f (Z) = f (Z [g] )
n
Y
bjj (Z, g)pj ,
(2.5)
(2.6)
j=1
where bjj (Z, g) are the diagonal matrix elements of the lower triangular matrix
b(Z, g).
We present formulas for transformations Z 7→ Z [g(t)] , where g(t) are standard
one-parametric subgroups in GL(n, C).
1) Let g(t) = exp(1 + tEkl ), where k < l. Then we have transformation
zkl 7→ zkl + t;
zml 7→ zml + tzkl ,
zij 7→ zij
m < k;
for other pairs i, j.
2) Let g(t) = exp(tEkk ). We get
zkj 7→ et zkj ,
k < j;
−t
zik 7→ e zik ,
i < k;
zij 7→ zij
for other pairs i, j.
3) Let g(t) = exp(tE(k+1)k ). Then
zik 7→ zik + tzi(k+1) ,
zkj
zkj 7→
,
1 + tzk(k+1)
z(k+1)m 7→ z(k+1)m + t det
zij 7→ zij
i < k;
j > k;
zk(k+1)
1
zkm
z(k+1)m
for other pairs i, j.
10
,
m > k + 1;
4) Let g(t) = exp(tE1n ), then
z1n 7→ z1n + t,
zij 7→ zij
for other pairs i, j.
Expressions (1.4)–(1.6) for the action of the Lie algebra gl(n) easily follows
from these formulas. Also, we get
E1n = ∂1n .
(2.7)
2.3. The Jacobian.
Lemma 2.1 The complex Jacobian J(Z, g) of a transformation Z 7→ Z [g] is
J(Z, g) =
n−1
Y
det [Zg]jj
j=1
−2
=
n−1
Y
det [Z]jn [Z]nj
j=1
−2
.
(2.8)
Proof. We say that a function
c : Nn+ × GL(n, C) → C
satisfies the chain identity if
c(Z, gh) = c(Z, g)c(Z g , h).
(2.9)
The Jacobian satisfies this identity. Let us verify that the right hand side of
−2
(2.8) also satisfies it. It suffices to consider one factor det[Zg]jj
.
[g]
Let us evaluate det[Z h]jj . Represent Z, g, h etc. as block j + (n − j)matrices,
Z11 Z12
G11 G12
H11 H12
Z=
,
g :=
,
h :=
. (2.10)
0
Z22
G21 G22
H21 G22
Represent Zg as C −1 U , where U ∈ Nn+ , C ∈ Bn− . Then
Z11 G11 + Z12 G21 Z11 G12 + Z12 G22
U11
C11
0
=
C21 C22
Z22 G21
Z22 G22
0
U12
.
U22
Since det U11 = 1, we have
det C11 = det(Z11 G11 + Z12 G21 )−1 .
Next,
det[Z [g] h]jj = det[U h]jj = det[U11 H11 + U12 H21 ] =
= det C11 · det (Z11 G11 + Z12 G21 )H11 + (Z11 G12 + Z12 G22 )H21 =
det Z11 (G11 H11 + G12 H21 ) + Z12 (G21 H21 + G)
det[Zgh]jj
=
,
=
det(Z11 G11 + Z12 G21 )
det[Zg]jj
11
this proves the desired statement.
Since the both sides of (2.8) satisfy the chain identity (2.9), it suffices to verify the identities for a system of generators of GL(n, C). Thus we can verify (2.8)
for elements one-parametric subgroups exp(tEk(k+1) ), exp(tEkk ), exp(tE(k+1)k ).
Formulas for the corresponding transformations are present in the previous subsection. Only the case exp(tE(k+1)k ) requires a calculation. In this case, the
Jacobi matrix is triangular. Its diagonal values are
— (1 + tzk(k+1) )−2 for zk(k+1) ;
— (1 + tzk(k+1) )−1 for zkm , where m > k + 1;
— (1 + tzk(k+1) ) for z(k+1)m , where m > k + 1.
The Jacobian (1 + tzk(k+1) )−2 .
On the other hand, in the product (2.8) the j-th factor is (1 + tzk(k+1) )−2 ,
other factors are 1.
2.4. Proof of Proposition 1.1.
The U(n)-invariance of the L2 (Nn+ , µp )-inner product. We must check the
identity
Z
Z
|ϕ(Z)|2 dµp (Z).
(2.11)
|(ρp (g)ϕ)(Z)|2 dµp (Z) =
+
Nn
+
Nn
[g]
Fix g ∈ U(n). Substituting Z = U to the right-hand side, we get
Z
n−1
∗ −(pj −pj+1 )−2
2 Y
d U ˙[g] .
det [U [g] ]jn U [g] ]jn
ϕ(U [g] )
+
Nn
j=1
In notation (2.10) we have
U11 G11 + U12 G21
Ug =
∗
U11 G12 + U12 G22
.
∗
Therefore [U [g] ]jn has the form
C11 (U11 G11 + U12 G21 ) C11 (U11 G12 + U12 G22 )
where C11 is a lower-triangular matrix, and
Hence
det C11 (U11 G11 + U12 G21 ) = 1.
∗
=
det [U [g] ]jn U [g] ]jn
∗
= det [U g]jn [U g]jn
det(U11 G11 +U12 G21 )−1 det(U11 G11 +U12 G21 )∗−1 .
Next,
∗
[U ]jn [U ]jn =
∗
∗
∗
∗
= (U11 G11 +U12 G21 )(G∗11 U11
+G∗21 U12
)+(U11 G12 +U12 G22 )(G∗12 U11
+G∗22 U12
)=
∗
∗
= U11 (G11 G∗11 + G12 G∗12 )U11
+ U11 (G11 G∗21 + G12 G∗22 )U12
+
∗
∗
∗
∗
+ U12 (G21 G11 + G22 G12 )U11 + U12 (G21 G21 + G22 G∗22 )U22
12
The matrix g is unitary, gg ∗ = 1. Therefore the last expression equals to
∗
∗
∗
∗
U11 · 1 · U11
+ U11 · 0 · U12
+ U12 · 0 · U11
+ U12 · 1 · U12
.
Thus
∗
∗
= det [U ]jn [U ]jn
det(U11 G11 + U12 G21 )
det [U [g] ]jn U [g] ]jn
−2
.
Keeping in mind
we get
d U˙[g] = det(U11 G11 + U12 G21 )
n−1
Y
∗ −(pj −pj+1 )−2
d U ˙[g] =
det [U [g] ]jn U [g] ]jn
j=1
= σ(U, g)
n−1
Y
j=1
where
σ(U, g) =
n−1
Y
−4
dŻ,
∗
dU̇ ,
det [U ]jn U ]jn
2
det([U g]jj )(pj −pj+1 ) .
j=1
Let us represent U g in the form DW , where D ∈ Bn− , g ∈ Nn+ . Then
det([U g]kk ) =
det(g) =
k
Y
i=1
n
Y
bii (U, g),
k 6 n − 1,
bii (U, g),
(2.12)
(2.13)
i=1
here bii (U, g) are the diagonal elements of U g. Therefore,
σ(U, g) =
n−1
Y
2
bjj (U, g)pj −pj+1
j=1
Since g ∈ U(n), we have | det(g)| = 1. Keeping in mind (2.13), we come to
σ(U, g) =
n
Y
2
bjj (U, g)pj .
j=1
Therefore the right-hand side of (2.11) equals
Z
+
Nn
f (U [g] )
n
Y
2
bjj (U, g)pj
j=1
13
dU̇ ,
i.e., coincides with the left-hand side.
Thus the group U(n) acts in L2 (Nn+ , dµp ) by the unitary operators (2.6).
Intersection of the space of L2 and the space of polynomials. Denote this
intersection by W .
First, W is U(n)-invariant. Indeed, for ψ ∈ W we have ρp (g)ψ ∈ L2 .
By definition, ρp (g)ψ is a rational holomorphic function. Represent it as an
irreducible fraction α(Z)/β(Z). Let Z0 be a non-singular point of the manifold
β(Z) = 0, let O be neighborhood of Z0 . It is easy to see that
Z
|ρp (g)ψ|2 dŻ = ∞.
O
Therefore ρp (g)ψ is a polynomial.
Second, the
R space W is non-zero. For instance it contains a constant function. Indeed, Nn+ dµp was evaluated in [24] and it is finite.
The third, W is finite-dimensional. Indeed, our measure has the form
|r(Z)|−2 dŻ, where r(Z, Z) is a polynomial in Z, Z. A polynomial ϕ ∈ W
satisfies the condition
Z
2
ϕ(Z)
dŻ < ∞.
+ r(Z, Z)
Nn
Clearly, degree of ϕ is uniformly bounded.
Next, the operators (2.6) determine a unitary representation in L2 (Nn+ ) and
this representation is an element of the principal non-unitary series (see, e.g., [6]
or [30], Addendum). But a representation of the principal series can not have
more than one finite-dimensional subrepresentation.
3
The formula for the kernel
In this section we prove Proposition 1.4.
3.1. Functions Φ[. . . ], Ψ[. . . ]. Consider the matrices Z, Z cut , U , U ext as
in Subs. 1.8. We use notation (1.2) for their sub-matrices. Let α = 1, . . . , n − 1.
Let I, J be sets of integers,
I : 0 < i1 < · · · < in+1−α 6 n + 1,
J : 0 < j1 < · · · < jα 6 n.
Denote
Φα [I; J] = Φα [i1 , . . . , in+1−α ; j1 , . . . jα ] := det
[Z(I)]
.
[U ext (J)]
Let I, J be sets of integers,
I : 0 < i1 < · · · < in−α 6 n + 1,
14
J : 0 < j1 < · · · < jα 6 n.
(3.1)
Denote
Ψα [I; J] = Ψα [i1 , . . . , in−α ; j1 , . . . jα ] := det
cut
[Z (I)]
.
[U (J)]
(3.2)
We have two collections of variables zij , ui′ j ′ . Zhelobenko operators Rkl in
z and u we denote by
z
u
Rkl
,
Rkl
.
Lemma 3.1
0,
z
Rkl Φα [I, J] = 0,
(−1)θ(I,k,l) Φα [I ◦ ; J],
if k ∈
/ I;
if k ∈ I, l ∈ I;
if k ∈ I, l ∈
/ I.
(3.3)
where I ◦ is obtained from I by replacing of k by l, and the corresponding change
of order. If k = is and it < l < it+1 , then
I ◦ = (i1 , . . . , is−1 , is+1 , . . . , it−1 , l, it+1 , . . . , in+1−α )
and
θ(I, k, l) = t − s.
The same properties (with obvious modifications) hold for
u
Rkl
Φα [I; J],
z
Rkl
Ψα [I; J],
Proof. Recall that
z
Rkl
= ∂kl +
X
u
Rkl
Ψα [I; J].
zlj ∂kj .
j>l
If k ∈
/ I, then Φα [I, J] does not depend on variables zkm . Therefore we get
z
0. If k ∈ I, these variables are present only in one row. Application of Rkl
is
equivalent to a change of the row
0 . . . 0 1 zk(k+1) . . . zkl zk(l+1) . . . zk(n+1)
[Z(I)]
of the matrix
by the row
[U ext (J)]
0 . . . 0 0 0 . . . 1 zl(l+1) . . . zk(n+1) .
If l ∈ I, then this row is present in the initial matrix, and again det = 0.
Otherwise, we come to Φα [I ◦ ; J] up to the order of rows.
Corollary 3.2 For any k, l, α,
z 2
(Rkl
) Φα [I; J] = 0,
u 2
(Rkl
) Φα [I; J] = 0,
z 2
(Rkl
) Ψα [I; J] = 0,
u 2
(Rkl
) Ψα [I; J] = 0.
15
Corollary 3.3
z
z
Rkl
Rlm
Φα [I; J] =
(
z
−Rkm
Φα [I; J],
0,
if k ∈ I, l ∈
/ I, m ∈
/ I;
otherwise.
(3.4)
The same property holds for operators Ru and for Ψα [I; J].
3.2. Verification of the Zhelobenko conditions.
Lemma 3.4 Lp,q ∈ Vp ⊗ Vq .
Proof. By the interlacing conditions, Lp,q is a polynomial. We must verify
the identities
z
(Rk(k+1)
)pk −pk+1 +1 L = 0,
u
(Rk(k+1)
)qk −qk+1 +1 L = 0.
(3.5)
To be definite verify the first equality,
z
)pk −pk+1 +1 Lp,q =
(Rk(k+1)
−pk+1 −qn+1−k pk +qn+1−k
z
pk −pk+1 +1
Ψn−k
Φn+1−k
· remaining factors
=
= (Rk(k+1) )
−pk+1 −qn+1−k pk +qn+1−k
z
Φn+1−k
· remaining factors
)pk −pk+1 +1 Ψn−k
= (Rk(k+1)
The sum of exponents (the both exponents are positive) is
−pk+1 − qn+1−k + pk + qn+1−k = pk − pk+1 ,
and we obtain the desired 0. If k = n the factor Ψn−k is absent, we have
pn +q1
z
pn −pn+1 +1
· remaining factors .
Φ1
(Rn(n+1) )
By the interlacing condition −pn+1 > q1 and we again obtain 0.
3.3. Invariance of the kernel. Consider matrices Φα , Ψα defined by
(1.13).
g
Lemma 3.5 Let g ∈ GL(n, C). Denote e
g=
0
0
∈ GL(n + 1, C). Then
1
Φ(Z [eg] , U [g] ) =
= Φ(Z, U ) det(g) det [Ze
g](n+1−α)(n+1−α)
−1
det [U g]αα
−1
. (3.6)
Ψ(Z [eg] , U [g] ) =
= Ψ(Z, U ) det(g) det [Ze
g ](n−α)(n−α)
16
−1
det [U g]αα
−1
. (3.7)
Proof. We write the Gauss decompositions of Ze
g, U g,
Ze
g = BP,
U g = CQ.
Then the Gauss decomposition of U ext (see (1.11)) is
C 0
Q 0
U ext e
g=
,
0 1
0 1
e Q.
e Then
denote factors in the right-hand side by C,
[eg]
[P ](n+1−α)(n+1)
[Z ](n+1−α)(n+1)
=
=
e α(n+1)
[(U ext )[eg] ]α(n+1)
[Q]
=
!
[Z](n+1−α)(n+1)
·g
e ]α(n+1)
e αα )−1
[U
([C]
e (n+1−α)(n+1−α) )−1
([B]
0
0
We pass to determinants in the left-hand side and the right-hand side. Keeping
in mind
e (n+1−α)(n+1−α) = det[Ze
eg](n+1−α)(n+1−α) ,
det[B]
e αα = det[U ext ge]αα = det[U g]αα ,
det[C]
we come to (3.6). Proof of (3.7) is similar.
Proof of Proposition 1.4. Applying the lemma and (2.12), we get
Lp,q (Z [eg] , U [g] ) = Lp,q(Z, U )
n
Y
bjj (Z, g)−q1 −pj
j=1
Y
bjj (U, g)−pn −qj det(g)p1 +qn .
Since
b(n+1)(n+1) (Z, ge) = 1,
det(g) =
n
Y
bjj (Z, g) =
n
Y
bjj (U, g),
j=1
j=1
we come to
Lp,q (Z [eg] , U [g] ) = Lp,q (Z, U )
n
Y
bjj (Z, g)−pj
j=1
Y
bjj (U, g)−qj .
Therefore,
(ρp (e
g ) ⊗ ρq (g))Lp,q = Lp,q.
Thus Lp,q is a non-zero GL(n, C)-invariant vector in Vp ⊗ Vq . We also know
that such a vector is unique up to a scalar factor.
17
4
The calculation
Below q ∈ Λn , r = p∗ ∈ Λn+1 are signatures. The signatures r and q are
interlacing. The kernel L is the same as above (1.14),
L := Lrq (Z, U ) = Φq11 −r2 Ψ1r2 −q2 Φ2q2 −r3 Ψ2r3 −q3 . . . Φqnn −rn+1 .
Below
u
Rkl := Rkl
.
We must verify the differential-difference equations
E1(n+1) L = F1(n+1) L,
E(n+1)n L = F(n+1)n L,
F1(n+1) , F(n+1)n are given by (1.22)–(1.22). Evaluations of the left-hand sides
is easy, it is contained in the next subsection. Evaluation of the right-hand sides
is tricky and occupies the remaining part of the section.
4.1. Evaluation of E1(n+1) L.
Lemma 4.1
n
q −r
X
(qα − rα+1 )R1α Ψα−1
1
2
E1(n+1) L = −
· L.
+
Φ1
Φα
α=2
(4.1)
1
= −1. On the other hand,
Remark. For α = 1 we immediately get ∂z∂Φ
1(n+1)
taking formally α = 1 in the sum in (4.1) we get R11 Ψ0 /Φ1 . At first glance, we
must assume that Ψ0 = 1 and that R11 is the identical operator. Under this
assumption, R11 Ψ0 /Φ1 = 1/Φ1 . However, this gives an incorrect sign.
⊠
In the following two proofs we need manipulations with determinants. To
avoid huge matrices or compact notation, which are difficult for reading, we
expose calculations for matrices having a minimal size that allows to visualize
picture.
Proof. We have (see (2.7))
E1(n+1) L =
∂L
∂z1(n+1)
=
n
X
(qα − rα+1 )
α=1
∂Φα
· L.
· Φ−1
α
∂z1(n+1)
Take n = 3 and α = 2,
1
0
Φ2 = det
1
0
Then
0
∂Φ2
= − det 1
∂z14
0
1
u12
1
z12
1
u12
1
z23
1
u13 = det
1
u23
18
z13
z23
u13
u23
z23
u23
z14
z24
.
0
0
1
= det 0
0
(4.2)
z12
1
1
z13
z23 = R12 Ψ1 .
u23
Lemma 4.2
n−1
X (rα+1 − qα+1 )R(α+1)n Φα+1
E(n+1)n L = zn(n+1) (r1 − q1 ) −
· L.
Ψα
j=1
Proof. Split E(n+1)n as a sum
E(n+1)n = D + (r1 − r2 )zn(n+1) ,
D :=
n−1
X
2
zj(n+1) ∂jn − zn(n+1)
∂n(n+1) .
j=1
Obviously,
E(n+1)n L =
n−1
n
X
(qα − rα+1 )DΦα X (rβ+1 − qβ+1 )DΨβ
+
= (r1 − r2 )zn(n+1) +
·L
Φα
Ψβ
α=1
β=1
Let us evaluate all DΦα , DΨα .
1) DΦα = 0 for all α 6= 1. Take
1 z12 z13
0 1
z23
DΦ2 = D det
1 u12 u13
0 1 u23
n = 3, α = 2,
z14
1 z12
0 1
z24
= det
1 u12
0
0
0 1
z14
z24
0
0
z14
z24
= 0.
0
0
Variables zjn (in our case zj4 ) are present only in the next-to-last column. We
apply the operator D to this column and come to a matrix with coinciding
columns.
19
2) DΦ1 = −zn(n+1) Φ1 . Take n = 3,
1 z12 z13 z14
0 1
z23 z24
=
DΦ1 = D det
0 0
1 z34
1 u12 u13 0
1 z12 z14 z14
1
0 1 z24 z24
+ z 2 det 0
= det
34
0 0
0 z34
1
1 u12 0
0
1 z12 z14
1
2
= −z34 det 0 1 z24 + z34
det 0
1 u12 0
1
1 z12 z13 z14
1
0 1
0
z
z
23
24
− z34 det
= −z34 det
0 0
0
1
0
1 u12 u13 0
1
1 z12
0 1
= −z34 det
0 0
1 u12
3) DΨα = −R(α+1)n Φα+1 . Take n = 4, α = 1.
1 z12 z13 z14
1 z12
0 1
0 1
z
z
24
24
= det
DΨ1 = D det
0 0
0 0
1
z34
1 u12 u13 u14
1 u12
1 z12 z13
0 1
z24
0
0
1
= − det
1 u12 u13
0 0
0
z12
1
u12
z12
1
u12
z12
1
0
u12
z13
z23
1
u13
z13
z24
1
u13
z14
z24
z25
u14
1
z13
z23 =
u13
z13
z23 =
u13
z13
z23
1
u13
0
0
=
z34
0
z14
z24
= −z34 Φ1 .
z34
0
z15
z25
=
z35
0
z15
z25
z35
= −R24 Φ2 .
0
0
Thus we know all DΦα , DΨα . This gives us the desired statement.
4.2. Quadratic relations. Now we wish to write a family of quadratic
relations between different functions of type Rkl Φα , Rkl Ψβ . We introduce notation
Rkk Φk−1 := +Φk−1 ,
Rkk Ψk−1 = −Ψk−1 ,
R11 Ψ0 = −Ψ0 = 1 .
(4.3)
We do not introduce operators Rkk and Rkk Φk−1 , Rkk Ψk−1 are only symbols
used in formulas. In particular, formula (4.1) now can be written the form
E1(n+1) L = L ·
n
X
(qα − rα+1 )R1α Ψα−1
Φα
α=1
20
without a term with abnormal sign. Below it allows to avoid numerous anomalies
in the formulas, duplications of formulas and branchings of calculations.
Lemma 4.3 a) Let m < α 6 β,
β+1
X
Rmj Φα · Rj(β+1) Ψβ = −Φα · Rm(β+1) Ψβ + Φβ+1 · Rmα Ψα−1 .
(4.4)
j=α+1
b) Let m < α < β. Then
β+1
X
Rmj Ψα · Rj(β+1) Ψβ = −Ψα · Rm(β+1) Ψβ .
(4.5)
j=α+1
c) Let m < α < β. Then
β
X
Rmj Φα · Rj(β+1) Φβ = −Φα · Rm(β+1) − Rm(β+1) Φα · Φβ .
(4.6)
j=α+1
Proof. Denote
ξ the row 0 0 . . . 1 of the length (n+ 1). We write
by
Z
a matrix ∆ := U ext . Let us enumerate z-rows of these matrix by 1, 2, 3,
ξ
. . . and u-rows by marks 1, 2, 3, . . . . Consider the following minors of ∆:
Φα = det ∆[1, . . . , n + 1 − α; 1, . . . , α];
b . . . , β + 1; ξ];
Rm(β+1) Ψβ = (−1)β−m det ∆[1, . . . , n − β; 1, . . . , m,
Notice that we ξ in the last row allows to replace the last determinant by
h
i
b . . . , β + 1; ξ]
.
det ∆[1, . . . , n − β; 1, . . . , m,
(4.7)
nn
and this proves (4.7).
We wish to apply one of the Plücker identities to the product of these minors (see, e.g., [4], Section 9.1). We take m from the collection (1, . . . , n + 1 −
b . . . , β+
α; 1, . . . , α), exchange it with an element of the collection (1, . . . , n−β; 1, . . . , m,
1; ξ) and consider the product of the corresponding minors. Next, we take the
21
sum of all such products of ∆. In this way, we obtain
b . . . , β + 1; ξ] = (4.8)
det ∆[1, . . . , n + 1 − α; 1, . . . , α] · det ∆[1, . . . , n − β; 1, . . . , m,
= det ∆[1, . . . , n + 1 − α; 1, . . . , m − 1, 1, m + 1, . . . , α]×
(4.9)
b . . . , β + 1; ξ] + · · · +
× det ∆[m, 2, . . . , n − β; 1, . . . , m,
(4.10)
b . . . , β + 1; ξ] + · · · +
× det ∆[1, . . . , n − β; m, 2, . . . , m,
(4.12)
+ det ∆[1, . . . , n + 1 − α; 1, . . . , m − 1, 1, m + 1, . . . , α]×
+ det ∆[1, . . . , n + 1 − α; 1, . . . , m − 1, α + 1, m + 1, . . . , α]×
(4.11)
(4.13)
b . . . , α, m, α + 2, . . . , β + 1; ξ] + · · · +
× det ∆[1, . . . , n − β; 1, . . . , m,
(4.15)
b . . . , β − 1, m, β + 1; ξ]
× det ∆[1, . . . , n − β; 1, . . . , m,
(4.17)
+ det ∆[1, . . . , n + 1 − α; 1, . . . , m − 1, β, m + 1, . . . , α]×
+ det ∆[1, . . . , n + 1 − α; 1, . . . , m − 1, β + 1, m + 1, . . . , α]×
b . . . , β, m; ξ]+
det ∆[1, . . . , n − β; 1, . . . , m,
+ det ∆[1, . . . , n + 1 − α; 1, . . . , m − 1, ξ, m + 1, . . . , α]×
b . . . , β + 1; m].
det ∆[1, . . . , n − β; 1, . . . , m,
(4.14)
(4.16)
(4.18)
(4.19)
(4.20)
Look to the summands of our expression up to signs.
1) After exchanging of m with 1, . . . , n − β we get a matrix ∆[. . . ] with
coinciding rows and therefore we obtain 0, see (4.9).
2) After exchanging of m with 1, . . . , α we again obtain 0, see (4.11).
3) The sum (4.13)–(4.16) corresponds to the sum in the left hand side of the
the desired identity (4.4) with j < β + 1.
4) The summand (4.17)–(4.18) corresponds to the last term of the sum.
5) The summand (4.19)–(4.20) corresponds to the expression Φβ+1 ·Rmα Ψα−1 .
6) The term (4.8) is Φα · Rm(β+1) Ψβ .
Next, let us watch signs. Denote by (i1 i2 . . . ik ) a cycle in a substitution.
Denote by σ(·) the parity of a substitution.
Examine summands (4.13)–(4.16). For j = α + 1, . . . , β + 1, we have
det ∆[1, . . . , n + 1 − α; 1, . . . , m − 1, j, m + 1, . . . , α] =
= σ j (m + 1) . . . α Rmj Φα = (−1)α−m Rmj Φα
For j = α + 1, . . . , β we have
b . . . , j − 1, m, j + 1, . . . , β + 1; ξ] =
det ∆[1, . . . , n − β; 1, . . . , m,
= σ m . . . (j − 1) · σ (j + 1) . . . (β + 1) Rj(β+1) Ψβ ] = (−1)β−m+1 Rj(β+1) Ψβ ].
Next, in (4.17) we have
b . . . , β, m; ξ] = (−1)β−m Ψβ
det ∆[1, . . . , n − β; 1, . . . , m,
and this gives the anomaly of a sign mentioned above.
22
Examine the last summand (4.19)–(4.20). We get
det ∆[1, . . . , n + 1 − α; 1, . . . , m − 1, ξ, m + 1, . . . , α − 1, α] =
= − det ∆[1, . . . , n + 1 − α; 1, . . . , m − 1, α, m + 1, . . . , α − 1, ξ] =
i
h
= −Rmα Ψα−1
− det ∆[1, . . . , n + 1 − α; 1, . . . , m − 1, α, m + 1, . . . , α − 1]
nn
and
b . . . , β + 1; m] =
det ∆[1, . . . , n − β; 1, . . . , m,
= σ m . . . (β + 1) Φβ+1 = (−1)β+1−m Φβ+1 .
Combining all signs we get the desired identity.
cut
Z
Proof of the statement b. We compose a new matrix ∆ =
,
U
consider the following minors of ∆,
Ψα = det ∆[1, . . . , n − α; 1, . . . , α],
b . . . , β, β + 1],
Rm(β+1) Ψβ = (−1)β−m det ∆[1, . . . , n − β; 1, . . . , m,
and write the Plücker identity exchanging the m-th row of Ψα with rows of
Rm(β+1) Ψβ .
Z
Proof of the statement c. We compose a matrix ∆ :=
, take
U ext
the minors
Φα = det ∆[1, . . . , n + 1 − α; 1, . . . , α],
b . . . , β, β + 1].
Rm(β+1) Φβ = (−1)β−m det ∆[1, . . . , n + 1 − β; 1, . . . , m,
and write the Plücker identities exchanging m-th of Φα with rows of the second
minor.
4.3. A recurrence formula. Our next purpose is to derive a formula for
F1(n+1) L. Let us evaluate one of summands of this expression,
X
Y
(4.21)
(l − n + ql − qn ) · RI Tn− L.
Θ1n :=
I⊳[1,n] l∈[1,n]\I
Recall that the notation J⊳{1, 2, . . . n} means that J is a subset in {1, 2, . . . n}
containing 1 and n, see Subsection 1.9.
Remark. By definition, the right-hand side is a rational expression of the
form
F
Q sα Q tβ .
Φα
Ψβ
It is easy to observe (see below) that all the exponents sα , tβ are equal 1. In a
calculation below we decompose this function as a sum of ’prime fractions’ with
denominators Φα , Ψβ and finally get an unexpectedly simple expression.
⊠
23
Let 1 6 k < l 6 n, µ < n. Denote
µ
Wkl
:=
µ
wkl
(qµ − rµ+1 )Rkl Φµ
(rµ+1 − qµ+1 )Rkl Ψµ
+
;
Φµ
Ψµ
l−k−1
:= (−1)
µ
Y
(qn − qj + j − n − 1)
l−1
X
(qn − qj + j − n);
(4.23)
j=µ+1
j=k+1
ζkl =
l−1
Y
(4.22)
µ
µ
n−1
wkl
Wkl
+ wk(n−1)
µ=k
Rkl Ψn−1
.
Ψn−1
(4.24)
Notice that the additional term in the last row is 0 if l 6= n.
Lemma 4.4
Rkl L = L ·
X
µ
Wkl
;
µ:k6µ<l
Rkl Tn− L = L ·
X
µ:k6µ<n
Rkn Tn− L = L ·
X
(4.25)
Ψ
n−1
µ
Wkl
·
,
Φn
µ
Wkl
+
µ:k6µ<l
for l < n;
Rkn Ψn−1 Ψn−1
·
.
Ψn−1
Φn
(4.26)
(4.27)
Proof. Obviously,
Rkl L = L
n
X
(qµ − rµ+1 )RklΦµ
Φµ
j=1
+
n
X
(rµ+1 − qµ+1 )RklΨµ
.
Ψµ
j=1
By Lemma 3.1, only terms with k 6 µ < l give a nonzero contribution. This
gives the first row. Next,
Ψ
Ψn−1
Ψn−1
n−1
Tn− L = L ·
,
Rkl Tn− L = (Rkl L) ·
+ L · Rkl
Φn
Φn
Φn
The second term in the right-hand side is 0 for l < n. For l = n this term equals
Rkn Ψn−1 /Φn .
Lemma 4.5
Θ1n =
X
J={j1 <j2 <···<js }⊳{1,2,...n}
Ψn−1
.
ζj1 j2 ζj2 j3 . . . ζjs−1 js L ·
Φn
(4.28)
Proof. First, we evaluate
RI (Tn− L) · L−1 .
Each RI is a product Ri1 i2 . . . Ris−1 is . Each Rit it+1 is a first order differential
operator without term of order zero, therefore we can expand RI L according
the Leibniz rule.
24
Many summands of this expansion are zero by a priory reasons. Indeed,
Rkl Φµ , Rkl Ψµ are nonzero only if k 6 µ < l. Also, Rkl Rab Φµ = 0 if l < a. In
the case l = a we have Rkl Rlb Φµ = −Rkb Φµ .
Therefor the expression (RI Tn− L)/L is a sum of products of the following
type
Ψn−1
A[m1 , m2 ] A[m2 , m3 ] . . . A[mt−1 , mt ] ·
,
(4.29)
Φn
where M = {m1 , . . . , mt } ⊳ I. and each factor A[mτ , mτ +1 ] = A[ia , ic ] has a
form
(qµ − rµ+1 ) Ria ia+1 Ria+1 ia+2 . . . Ric−1 ic Φµ
Φµ
(rµ+1 − qµ+1 ) Ria ia+1 Ria+1 ia+2 . . . Ric−1 ic Ψµ
or
,
Ψµ
where ic−1 6 µ < ic . These expressions are equal correspondingly
(−1)c−a+1 (qµ − rµ+1 ) Ria ic Φµ
Φµ
(−1)c−a+1 (rµ+1 − qµ+1 ) Ria ic Ψµ
Ψµ
and
Now fix M and consider the sum S[I, M ] of all summands (4.29) with fixed M .
It is easy to see that
c−a+1
S[I, M ] = (−1)
t−1
Y
X
X
γ=1 ia ∈I:mγ 6ia <mγ+1 µ:ia 6µ<mc
Wiµa mγ+1
· L.
Next, we represent Θ1n as
Θ1n =
X
Y
I⊳[1,n] l∈[1,n]\I
=
Ψn−1
S[I, M ] · L ·
=
Φn
M⊳I
Y
Ψn−1
.
(l − n + ql − qn ) · S[I, M ] · L ·
Φn
(l − n + ql − qn ) ·
X X
X
M⊳[1,n] I⊃M l∈[1,n]\I
Fix M = {m1 < · · · < mt }. Each summand in the big brackets splits into a
product of the form
t−1
Y
H[mγ , mγ+1 ],
γ=1
where H[mγ , mγ+1 ] is an expression, which depend on mγ , mγ+1 and does not
depend on other mi . Since the coefficients also are multiplicative in the same
sense, the whole sum in the big brackets also splits in a product
t−1
Y
Z[mγ , mγ1 ],
γ=1
25
where each factor Z[mγ , mγ+1 ] depends only on mγ , mγ+1 and not on the
remaining elements of a set M .
It remains to evaluate factors Z[·],
Z[k, l] =
l
X
µ
µ
υkl
Wkl
,
µ=k
where
µ
υkl
=
X
−
J
=
l−1
Y
Y
(−1)(m − n + qn − qm ) ·
m=µ+1
m∈[k,µ]\J
m=µ+1
l−1
Y
X
(m − n + qn − qm ) · −
J
Y
m∈[k,µ]\J
(m − n + qn − qm ) =
(−1)(m − n + qn − qm ) ,
the summation here is taken other all subsets J ⊂ [k, µ], containing k. In the
brackets we get
µ
Y
(−1) 1 − (m − n + qn − qm )
m=k+1
Thus Zkl equals to ζkl .
Let ζkl be as above (4.24). Denote
X
Θmn :=
J={j1 <j2 <···<js }⊳{m,2,...n}
Ψn−1
ζj1 j2 ζj2 j3 . . . ζjs−1 js L ·
.
Φn
By the previous lemma, this notation is compatible with the earlier notation
Θ1n . Also, Θnn = ΨΦn−1
.
n
Lemma 4.6 The Θmn satisfies the following recurrence relation,
Θmn = ζm(m+1) Θ(m+1)n + ζm(m+2) Θ(m+2)n + · · · + ζmn Θnn .
This statement is obvious.
4.4. Evaluation of Θ1n . Denote
sµmτ =
µ−1
Y
(qτ − qj + j − n)
j=m
στµ = (qµ − rµ+1 )
τ
Y
(qτ − ri − n − 1 + i).
i=µ+2
Lemma 4.7
Θmn L−1 = −
n
X
sµmn σnµ
µ=m
26
Rmµ Ψµ−1
.
Φµ
In particular, this gives an explicit expression for Θ1n .
Proof. We prove our statement by induction. Assume that for Θnn ,
Θ(n−1)n , . . . , Θ(m+1)n the formula is correct. We must derive the equality
Θmn · L−1 =
n
X
ζmγ Θ(m+1)n =
γ=m+1
=
n
γ−1
X
X
µ
µ
wmγ
Wmγ
n
X
sνγn σ ν
ν=γ
γ=m+1 µ=m
Rγν Ψν−1
.
Φν
We have
µ
υmγ
snγn =
µ
Y
(qn −qj +j−n+1)·
ν−1
Y
(qn −qj +j−n+1)
(qn − qj + j − n + 1) ·
ν−1
Y
(qn −qj +j−n) =
j=γ
j=µ+1
j=m+1
=
γ−1
Y
γ−1
Y
µ
(qn − qj + j − n + 1) =: ξmν
.
j=µ+1
j=m+1
µ
We stress that the expression ξmν
does not depend on γ. Our sum transforms
to
X
ν
X
Ψn−1
µ Rγν Ψν−1
µ
Wmγ
ξmν
σnν
+
.
Φ
Φn
ν
µ,ν:µ<ν
γ=µ+1
µ
Denote by Bmν
the expression in the square brackets and write it explicitly:
µ
Bmν
=
ν
(r
X
(qµ − rµ+1 )Rmγ Φµ Rγν Ψν−1
µ+1 − qµ+1 )Rmγ Ψµ
+
=
Ψµ
Φµ
Φν
γ=µ+1
ν
X
Rmγ Ψµ Rγν Ψν−1
= (rµ+1 − qµ+1 )
+
Ψµ
Φν
γ=µ+1
+ (qµ − rµ+1 )
ν
X
Rmγ Φµ Rγν Ψν−1
. (4.30)
Ψµ
Φν
γ=µ+1
Next, we apply the quadratic relations (4.4) and (4.5) and transform the last
expression to
(rµ+1 − qµ+1 )
R Ψ
Rmν Ψµ−1
−Rmν Ψν−1
mν ν−1
=
+ (qµ − rµ+1 ) −
+
Φν
Φν
Φµ
Rmν Ψµ−1
Rmν Ψν−1
+ (qµ − rµ+1 )
.
= (qµ+1 − qµ )
Φν
Φµ
Thus,
Θmn :=
Rmν Ψµ−1
Rmν Ψν−1
µ
. (4.31)
+ (qµ − rµ+1 )
ξmν
σnν (qµ+1 − qµ )
Φν
Φµ
µ,ν:µ<ν
X
27
We collect similar terms and get
Θmn :=
X
S1 + S2
ν
R
mν Ψν−1
Φν
,
where
S1 : =
S2 : =
ν−1
X
ξµν σnν (qµ+1 − qµ ),
µ=1
n
X
(4.32)
ξνκ σnκ (qν − rν+1 ).
(4.33)
κ=ν+1
First, we transform S2 ,
S2 = (qν − rν+1 ) ·
ν
Y
(qn − qj − n + j − 1)×
j=m+1
×
n
κ−1
Y
X
(qn − qj + j − n)(qκ − rκ+1 ) ·
n−κ
Y
j=2
κ=ν+1 j=ν+1
(qn − rn+2−j − j) . (4.34)
The sum in the second line is evaluated in the following lemma.
Lemma 4.8 Let n − k > ν + 1. Then
n
X
κ−1
Y
(qn − qj + j − n) · (qκ − rκ+1 ) ·
n−κ
Y
(qn − rn+2−j − j) =
j=2
κ=n−k+1 j=ν+1
=
n−k
Y
(qn − qj + j − n) ·
ν+1
k
Y
(qn − rn+2−j − k + 1). (4.35)
j=2
In particular,
n
X
(. . . ) =
κ=ν+1
n−ν
Y
(qn − rn+2−j − k + 1),
j=2
and
(qν − rν+1 )
n
X
(. . . ) = σnν .
(4.36)
κ=ν+1
Proof of Lemma 4.8. We prove the statement by induction. Let for a
28
given k the statement hold. Then
n
X
n
X
=
κ=n−k
κ=n−k+1
=
n−k
Y
n
o
+ n − k-th term =
(qn − qj + j − n) ·
ν+1
+
n−k−1
Y
k
Y
(qn − rn+2−j − k + 1)+
j=2
(qn − qj + j − n) · (qn−k − rn−k+1 ) ·
n−k−1
Y
(qn − rn+2−j − j) =
j=2
j=ν+1
=
k
Y
(qn −qj +j−n)
k
Y
(qn −rn+2−j −k+1)· (qn −qn−k −k)+(qn−k −rn−k+1 )
j=2
j=ν+1
and the big bracket joins to the product
Qk
j=2 .
Let us return to our calculation. We must evaluate S1 + S2 ,
ν−1
X
S1 + S2 = σnν
ξµν (qµ+1 − qµ ) + S2 =
µ=1
σν
ν−1
Xn
µ
Y
(qn − qj + j − n − 1) ·
µ=1 j=m+1
ν−1
Y
j=µ+1
+ σnν
o
(qn − qj + j − n) · (qµ − qµ+1 ) +
ν
Y
(qn − qj + j − n − 1). (4.37)
j=m+1
An evaluation of this sum is similar to the proof of Lemma 4.8. We verify the
following identity by induction,
σnν
ν−1
X
ξµν (qµ+1 − qµ ) + S2 =
ν−k+1
Y
(qn − qj + j − 1 − n)
j=m+1
ν−k
ν−1
Y
(qn − qj + j − n).
j=ν−k+1
Substituting k = ν − 1 we get the desired coefficient in (4.31).
It remains to notice that formulas for the coefficient in the front of Φ1n are
slightly different. In this case, the sum S2 is absent, but ζmn has an extra term
1
Φn . Starting this place, the calculation is the same, the extra term replaces the
additional term in (4.37).
This completes a proof of Lemma 4.7.
4.5. The extension of calculations. Next, we must evaluate other summands in F1(n+1) L. Denote
Θ1τ :=
X
Y
I⊳[1,τ ] l∈[1,τ ]\I
−
L.
(l − τ + ql − qτ ) · RI Tm
29
Lemma 4.9
Θ1τ = −
m
X
sµ1τ στµ
µ=1
R1µ Ψµ−1
.
Φµ
(4.38)
Proof. We decompose Tτ− L as
r2 −q2
rτ −qτ +1 −rτ +1 +qτ −1
2 +q1
Tτ− L = Φ−r
Ψ
.
.
.
Ψ
Φ
×
τ
1
1
τ −1
rm −qm −rτ +1 +qτ
× Ψm−1
. . . Φn−rn+1 +qn .
Φτ
Factors Rkl of operators RI act as zero on the factors of the second bracket.
Therefore this bracket can be regarded as a constant. After this, we get the
same calculations as for Θ1n but n is replaced by τ .
4.6. The summation.
Lemma 4.10
F1(n+1) L + E1(n+1) L = 0.
Proof.
F1(n+1) L =
n
X
Aτ
τ =1
τ
X
α
sα
1τ στ
α=1
n
n X
X
R1α Ψα−1
α R1α Ψα−1
Aτ sα
=
.
1τ στ
Φα
Φα
σ=1 τ =α
We must verify the following identities
n
X
α
Aτ sα
1τ στ = −(qα − rα+1 ).
(4.39)
τ =α
A summand in the left-hand side equals
α
Aτ sα
1τ στ
=
× (qα − rα+1 )
Qn−τ +1
j=1
α−1
Y
(qm − rn+2−j + n − τ + 1 − j)
Q
×
j6=τ (qτ − qj + j − m)
(qm − qi + i − m) ·
i=1
n−α
Y
(qm − rn+2−k + n − α − k + 1).
k=n−m+2
Qn−α
Qn−τ +1
Qα−1
and k=n−m+2
The factor i=1 (qm − qi + i − m) cancels, the factors j=1
join together and we come to
Qn−σ
j=1 (qm − rn+2−j + n − τ + 1 − j
α
Q
Aτ sα
σ
=
(q
−
r
)
.
α
α+1
1τ τ
j6=τ,j>σ (qτ − qj + j − m)
Next, we write a rational function
Qn−σ
Vα (x) :=
j=1
(x + rn+2−j + n − j + 1)
Qn
j=σ (x − qj + j)
30
and evaluate the sum of its residues,
m
X
Resx=qj −j Vα (x) = −Resx=∞ Vα (x)
(4.40)
j=σ
Multiplying this identity by (qα − rα+1 ) we get (4.39).
4.7. Invariance of the space Vr .
Lemma 4.11 The space Vr (see (1.16)) is invariant with respect to the operator
F1(n+1) .
+
Prove. Define elements ℓZ (U ) ∈ Vr , where Z ranges in Tn+1
, by
ℓZ (U ) = L(Z, U ).
First, functions ℓZ (U ) generate the space Vr (because the pairing Vp × Vr → C
determined by L is nondegenerate). Next,
F1(n+1) ℓZ (U ) = F1(n+1) L(Z, U ) = −E1(n+1) L(Z, U ) = −
∂
∂z1(n+1)
ℓZ (U ) .
Differentiating a family of elements of Vr with respect to a parameter we get
elements of the same space Vr .
4.8. Formula for F(n+1)n . Here a calculation is more-or-less the same, we
omit details (in a critical moment we use the quadratic identities (4.4), (4.6)).
5
Infinite dimensional case. A conjecture
5.1. Principal series. Now consider two collections of complex numbers
(p1 , . . . , pn ) and (p◦1 , . . . , p◦n ) such that
pj − p◦j ∈ Z,
Re(pj + p◦j ) = −2(j − 1).
Consider a representation ρnp|p◦ of the group GL(n, C) in the space L2 (Nn+ )
determined by the formula
ρp|p◦ (g)ϕ(g) = ϕ(Z [g] ) ·
n
Y
bjj (Z, g)pj bjj (Z, g)
p◦
j
.
(5.1)
j=1
We get the unitary (nondegenerate) principal series of representations of GL(n, C)
(see [6]). Denote by Λunitary
the space of all parameters p, p◦ .
n
Remark. In formula (5.1), we have complex numbers in complex powers.
We understand them in the following way:
p
p◦
◦
bjjj bjjj = |bjj |2pj (bjj /bjj )pj −pj .
31
In the right hand side, the first factor has a positive base of the power, the
second factor has an integer exponent. Hence the product is well defined.
⊠
◦
Remark. Formula (5.1) makes sense if pj − pj ∈ Z, and gives a nonunitary principal series of representations. The construction of holomorphic
representation discussed above corresponds to p ∈ Λn , p◦ = 0. If both
p, p◦ ∈ Λn
(5.2)
then our representation contains a finite-dimensional (nonholomorphic) representation ρp ⊗ ρp◦ , where ρ denotes the complex conjugate representation. ⊠
5.2. Restriction to the smaller subgroup. Consider a representation
◦
ρn+1
p|p◦ of the group GL(n + 1, C). According [1], the restriction of ρp|p to the
subgroup GL(n, C) is a multiplicity free integral of all representations of ρnp|p◦
of unitary nondegenerate principal series of GL(n, C). Moreover, the restriction
has Lebesgue spectrum. Thus the restriction Up|p◦ can be realized in the space
L2 (Nn+ × Λunitary
) by the formula
n
Up|p◦ (g)ψ(U, q|q◦ ) = ψ(Z [g] , q|q◦ ) ·
n
Y
qj◦
bjj (Z, g)qj bjj (Z, g) .
j=1
5.3. Intertwining operator. Next, we write an integral operator
+
L2 (Nn+1
) → L2 (Nn+ × Λunitary
)
n
by the formula
Jϕ(U, q|q◦ ) =
Z
+
Nn+1
ϕ(Z)Lp,q (Z, U ) Lp◦ ,q◦ (Z, U ) dŻ.
(5.3)
p
5.4. Additional operators. Denote the operators (1.22), (1.22) by F1(n+1)
,
p◦
p◦
p
p
F(n+1)n
. Denote by F 1(n+1) , F (n+1)n the operators obtained from F1(n+1)
,
p
F(n+1)n by replacing
∂
∂
7→
,
∂zkl
∂z kl
pj 7→ p◦j ,
qk 7→ qk◦ .
Define the operators
p◦
p|p◦
p
+ F 1(n+1) ,
F1(n+1) = F1(n+1)
p◦
p|p◦
p
+ F (n+1)n .
F(n+1)n = F(n+1)n
p|p◦
p|p◦
Conjecture 5.1 The operators F1(n+1) , F(n+1)n are images of the operators
E1(n+1) , E(n+1)n under the integral transform (5.3).
The statement seems doubtless, since we have an analytic continuation of
our finite-dimensional formulas from the set (5.2). However, this is not an
automatic corollary of our result. In particular, it is necessary to find the
Plancherel measure on Λunitary
(i.e. a measure making the operator J unitary).
n
32
References
[1] Anh, N.H. Restriction of the principal series of SL(n, C) to some reductive
subgroups. Pacific J. Math. 38 (1971), 295-314.
[2] Asherova R. M., Smirnov Yu. F., Tolstoi B. N. Projection operators for
simple lie groups II. General scheme for constructing lowering operators.
The groups SU(n). Theor. Math. Phys., 1973, V. 15, 1, pp. 392-401.
[3] Baird G. E. and Biedenharn L. C. On the Representations of the Semisimple
Lie Groups. II J. Math. Phys. 4, 1449 (1963), 1449–1466.
[4] Fulton, W. Young tableaux. With applications to representation theory and
geometry. Cambridge University Press, Cambridge, 1997.
[5] Gelfand, I. M.; Graev, M. I. Finite-dimensional irreducible representations
of the unitary and complete linear group and special functions associated
with them. (Russian) Izv. Akad. Nauk SSSR Ser. Mat. 29 1965 1329-1356.
English transl. in Amer. Math. Soc. Transl. Series 2 64 (1965), 116–146.
[6] Gelfand, I. M.; Naimark, M. A. Unitary representations of the classical groups. Trudy Mat. Inst. Steklov., vol. 36, Izdat. Akad. Nauk SSSR,
Moscow-Leningrad, 1950. German transl.: Gelfand, I. M.; Neumark, M. A.
Unitäre Darstellungen der klassischen Gruppen. Akademie-Verlag, Berlin,
1957.
[7] Gelfand, I. M.; Tsetlin, M. L. Finite-dimensional representations of the
group of unimodular matrices. (Russian) Dokl. Akad. Nauk SSSR (N.S.) 71
(1950), 825-828. English transl. in I. M. Gelfand, Collected Papers, Vol. II,
Springer, Berlin, 1988, 653-656.
[8] Gelfand, I. M.; Tsetlin, M. L. Finite-dimensional representations of groups
of orthogonal matrices. (Russian) Dokl. Akad. Nauk SSSR (N.S.) 71 (1950),
1017-1020. English transl. I. M. Gelfand, Collected Papers, Vol. II, Springer,
Berlin, 1988, 657-661.
[9] Graev, M. I. A continuous analogue of Gelfand-Tsetlin schemes and a realization of the principal series of irreducible unitary representations of the
group GL(n, C) in the space of functions on the manifold of these schemes.
Dokl. Math. 75 (2007), no. 1, 31-35.
[10] Groenevelt, W. The Wilson function transform. Int. Math. Res. Not. 2003,
no. 52, 2779-2817.
[11] Mickelsson, J. Step algebras of semi-simple subalgebras of Lie algebras. Rep.
Math. Phys. 4 (1973), 307-318.
[12] Molchanov, V. F. Canonical representations and overgroups for hyperboloids of one sheet and Lobachevsky spaces. Acta Appl. Math. 86 (2005),
no. 1-2, 115-129.
33
[13] Molchanov, V. F. Canonical representations and overgroups for hyperboloids. Funct. Anal. Appl. 39 (2005), no. 4, 284-295
[14] Molchanov, V. F. Canonical representations on Lobachevsky spaces: an
interaction with an overalgebra. Acta Appl. Math. 99 (2007), no. 3, 321337.
[15] Molchanov, V. F. Fourier and Poisson transforms for tensor products.
Funct. Anal. Appl., 49 (2015), no. 4, 279–288.
[16] Molev, A. I. Unitarizability of some Enright-Varadarajan u(p, q)-modules.
in Topics in representation theory, 199-219, Adv. Soviet Math., 2, Amer.
Math. Soc., Providence, RI, 1991.
[17] Molev, A. I. A basis for representations of symplectic Lie algebras. Comm.
Math. Phys. 201 (1999), no. 3, 591-618.
[18] Molev, A. Yangians and classical Lie algebras. Amer. Math. Soc., Providence, RI, 2007.
[19] Nazarov, Ma.; Tarasov, V. Yangians and Gelfand-Zetlin bases. Publ. Res.
Inst. Math. Sci. 30 (1994), no. 3, 459-478.
[20] Neretin, Yu. A. The action of an overalgebra in the Plancherel decomposition and shift operators in an imaginary direction. Izv. Math. 66 (2002),
no. 5, 1035-1046.
[21] Neretin, Yu. A. Notes on Stein-Sahi representations and some problems of
non-L2 -harmonic analysis. Zap. Nauchn. Sem. S.-Peterburg. Otdel. Mat.
Inst. Steklov. (POMI) 331 (2006), 14, 125–169, 224; translation in J. Math.
Sci. (N. Y.) 141 (2007), no. 4, 1452–1478.
[22] Neretin, Yu. A. Lectures on Gaussian integral operators and classical
groups. Europ. Math. Soc, 2011.
[23] Neretin, Yu. A. Difference Sturm-Liouville problems in the imaginary direction. J. Spectr. Theory 3 (2013), no. 3, 237-269.
[24] Neretin, Yu. A. Hua-Type Beta-Integrals and Projective Systems of Measures on Flag Spaces. IMRN, v. 2015, no. 21, 1128911301.
[25] Neretin, Yu. A. The Fourier transform on the group GL2 (R) and the action
of the overalgebra gl4 . J. Fourier Analysis, DOI: 10.1007/s00041-017-9589-8.
[26] Neretin, Yu. A. Operational calculus for Fourier transform on the group
GL(2, R). Preprint, arXiv:1801.09398.
[27] Olshanski, G. I. Description of unitary representations with highest weight
for the groups U(p, q)∼ . Funct. Anal. Appl. 14 (1980), no. 3, 190–200 (1981).
34
[28] Shtepin, V. V. Intermediate Lie algebras and their finite-dimensional representations. Russian Acad. Sci. Izv. Math. 43 (1994), no. 3, 559–579.
[29] Zhelobenko D.P. Classical groups. Spectral analysis of finite-dimensional
representations. Russ. Math. Surveys, 1962, 17:1, 1-94.
[30] Zhelobenko, D. P. Compact Lie groups and their representations. Amer.
Math. Soc., Providence, R.I., 1973.
[31] Zhelobenko, D. P. Representations of reductive Lie algebras. (Russian) VO
“Nauka”, Moscow, 1994.
[32] Zhelobenko, D. Principal structures and methods of representation theory.
Amer. Math. Soc., Providence, RI, 2006.
Math.Dept., University of Vienna,
Oskar-Morgenstern-Platz 1, 1090 Wien;
& Institute for Theoretical and Experimental Physics (Moscow);
& Mech.Math.Dept., Moscow State University.
& Institute for Information Transmission (Moscow)
e-mail: hepetuh(at) yandex.ru
URL:www.mat.univie.ac.at/∼neretin
35
| 4math.GR
|
arXiv:1511.00824v3 [math.CT] 30 Dec 2016
CENTRAL REFLECTIONS AND NILPOTENCY IN EXACT
MAL’TSEV CATEGORIES
CLEMENS BERGER AND DOMINIQUE BOURN
Abstract. We study nilpotency in the context of exact Mal’tsev categories taking central extensions as the primitive notion. This yields a nilpotency tower
which is analysed from the perspective of Goodwillie’s functor calculus.
We show in particular that the reflection into the subcategory of n-nilpotent
objects is the universal endofunctor of degree n if and only if every n-nilpotent
object is n-folded. In the special context of a semi-abelian category, an object is
n-folded precisely when its Higgins commutator of length n + 1 vanishes.
Contents
Introduction
1. Central extensions and regular pushouts
2. Affine objects and nilpotency
3. Affine morphisms and central reflections
4. Aspects of nilpotency
5. Quadratic identity functors
6. Identity functors with bounded degree
Acknowledgements
References
1
6
17
23
32
36
46
67
67
Introduction
This text investigates nilpotency in the context of exact Mal’tsev categories. Our
purpose is twofold: basic phenomena of nilpotency are treated through universal
properties rather than through commutator calculus, emphasising the fundamental
role played by central extensions; nilpotency is then linked to an algebraic form of
Goodwillie’s functor calculus [29]. This leads to a global understanding of nilpotency
in terms of functors with bounded degree.
A Mal’tsev category is a finitely complete category in which reflexive relations are
equivalence relations [17, 19]. Important examples of exact Mal’tsev categories are
Date: November 25, 2016.
1991 Mathematics Subject Classification. 17B30, 18D35, 18E10, 18G50, 20F18.
Key words and phrases. Nilpotency, Mal’tsev category, central extension, Goodwillie functor
calculus.
1
2
CLEMENS BERGER AND DOMINIQUE BOURN
Mal’tsev varieties [53] and semi-abelian categories [47]. The simplicial objects of an
exact Mal’tsev category are “internal” Kan complexes (cf. [17, 62]).
Nilpotency is classically understood via the vanishing of iterated commutators: in
a Mal’tsev variety by means of so-called Smith commutators [60, 27], in a semi-abelian
category by means of so-called Huq commutators [43, 25]. The first aim of this text
is to promote another point of view which seems more intrinsic to us and is based on
the notion of central extension, by which we mean a regular epimorphism with central
kernel relation. The n-nilpotent objects are defined to be those which can be linked
to a terminal object by a chain of n consecutive central extensions. This notion of
nilpotency is equivalent to the two aforementioned notions in their respective contexts
(cf. Proposition 2.14). In particular, we get the usual notions of n-nilpotent group,
n-nilpotent Lie algebra and n-nilpotent loop [15]. A category is called n-nilpotent if
all its objects are n-nilpotent.
For any exact Mal’tsev category with binary sums, the full subcategory spanned
by the n-nilpotent objects is a reflective Birkhoff subcategory (cf. Theorem 2.12).
This generalises the analogous known results for Mal’tsev varieties [60, 27] and semiabelian categories [25]. We denote the reflection into the subcategory of n-nilpotent
n
objects by I n and the unit of the adjunction at an object X by ηX
: X ։ I n (X).
Since an n-nilpotent object is a fortiori (n + 1)-nilpotent, the different reflections
assemble into the following nilpotency tower
X ❄PP
qq⑧⑧q⑧ ❄❄❄PPPPP
q
q
q ⑧
❄❄ PPP
qqq ⑧⑧⑧
❄❄ n PPP n+1
q
2
q
PPPηX
⑧
η
❄❄ηX
1
X
qq ⑧⑧ ηX
q
PP' '
❄
q
⑧
q
xqx o q 1
2
n
o
o
o
o
o
o
o
⋆
I (X)
I (X)
I (X)
I n+1 (X) o o
in which the successive quotient maps I n+1 (X) ։ I n (X) are central extensions.
Pointed categories with binary sums will be ubiquitous throughout the text; we call
them σ-pointed for short. Among σ-pointed exact Mal’tsev categories we characterise
the n-nilpotent ones as those for which the comparison maps
θX,Y : X + Y ։ X × Y
are (n − 1)-fold central extensions (cf. Theorem 4.3). The nilpotency class of a σpointed exact Mal’tsev category measures thus the discrepancy between binary sum
and binary product. If n = 1, binary sum and binary product coincide, and all objects
are abelian group objects. A σ-pointed exact Mal’tsev category is 1-nilpotent if and
1
: X ։ I 1 (X)
only if it is an abelian category (cf. Corollary 4.4). The unit ηX
of the first Birkhoff reflection is abelianisation (cf. Proposition 4.2). Moreover, the
successive kernels of the nilpotency tower are abelian group objects as well. This
situation is reminiscent of what happens in Goodwillie’s functor calculus [29] where
“infinite loop spaces” play the role of abelian group objects. The second aim of our
study of nilpotency was to get a deeper understanding of this analogy.
Goodwillie’s notions [29] of cross-effect and degree of a functor translate well into
our algebraic setting: for each (n + 1)-tuple (X1 , . . . , Xn+1 ) of objects of a σ-pointed
CENTRAL REFLECTIONS AND NILPOTENCY IN EXACT MAL’TSEV CATEGORIES
3
category and each based endofunctor F we define an (n + 1)-cube ΞF
X1 ,...,Xn+1 consisting of the images F (Xi1 + · · · + Xik ) for all subsequences of (X1 , . . . , Xn+1 ) together
with the obvious contraction maps. We say that a functor F is of degree ≤ n if these
(n + 1)-cubes are limit-cubes for all choices of (X1 , . . . , Xn+1 ).
F
F
We denote by θX
: F (X1 + · · · + Xn+1 ) → PX
the comparison map
1 ,...,Xn+1
1 ,...,Xn+1
towards the limit of the punctured (n + 1)-cube so that F is of degree ≤ n if and only
F
F
if θX
is invertible for each choice of (n + 1)-tuple. The kernel of θX
1 ,...,Xn+1
1 ,...,Xn+1
F
is a (n + 1)-st cross-effect of F , denoted crn+1 (X1 , . . . , Xn+1 ).
A based endofunctor F is linear, i.e. of degree ≤ 1, if and only if F takes binary sums to binary products. In a semi-abelian category, the second cross-effects
cr2F (X, Y ) measure thus the failure of linearity of F . If F is the identity functor, we
drop F from the notation so that cr2 (X, Y ) denotes the kernel of the comparison map
θX,Y : X + Y → X × Y . This kernel is often denoted X ⋄ Y and called the co-smash
product of X and Y (cf. [16] and Remarks 3.10 and 6.2).
An endofunctor of a semi-abelian (or homological [4]) category is of degree ≤ n if
and only if all its cross-effects of order n + 1 vanish. For functors taking values in
abelian categories, our cross-effects agree with the original cross-effects of EilenbergMac Lane [23] (cf. Remark 6.2). For functors taking values in σ-pointed categories
with pullbacks, our cross-effects agree with those of Hartl-Loiseau [38] and Hartl-Van
der Linden [39], defined as kernel intersections (cf. Definition 5.1).
A Goodwillie type characterisation of the nilpotency tower amounts to the property
that for each n, the reflection I n into the Birkhoff subcategory of n-nilpotent objects
is the universal endofunctor of degree ≤ n. In fact, every endofunctor of degree
≤ n of a σ-pointed exact Mal’tsev category takes values in n-nilpotent objects (cf.
Proposition 6.5). The reflection I n is of degree ≤ n if and only if the identity functor
of the Birkhoff subcategory of n-nilpotent objects itself is of degree ≤ n. In the
present article we have mainly been investigating this last property.
The property holds for n = 1 because the identity functor of an abelian category
is linear. However, already for n = 2, there are examples of 2-nilpotent semi-abelian
categories which are not quadratic, i.e. do not have an identity functor of degree ≤ 2
(cf. Section 6.5). We show that a σ-pointed exact Mal’tsev category is quadratic if
and only if the category is 2-nilpotent and algebraically distributive, i.e. endowed
with isomorphisms (X × Z) +Z (Y × Z) ∼
= (X + Y ) × Z for all objects X, Y, Z (cf.
Corollary 5.16). Since algebraic distributivity is preserved under Birkhoff reflection,
the subcategory of 2-nilpotent objects of an algebraically distributive exact Mal’tsev
category is always quadratic (cf. Theorem 5.18).
Algebraic distributivity is a consequence of the existence of centralisers for subobjects as shown by Gray and the second author [13]. For pointed Mal’tsev categories,
it also follows from algebraic coherence in the sense of Cigoli-Gray-Van der Linden
[20]. Our quadraticity result implies that iterated Huq commutator [X, [X, X]] and
ternary Higgins commutator [X, X, X] coincide for each object X of an algebraically
distributive semi-abelian category (cf. Corollary 5.19 and [20, Corollary 7.2]).
There is a remarkable duality for σ-pointed 2-nilpotent exact Mal’tsev categories:
algebraic distributivity amounts to algebraic codistributivity, i.e. to isomorphisms
4
CLEMENS BERGER AND DOMINIQUE BOURN
(X × Y ) + Z ∼
= (X + Z) ×Z (Y + Z) for all X, Y, Z (cf. Proposition 5.15). Indeed, the
difference between 2-nilpotency and quadraticity is precisely algebraic codistributivity
(cf. Theorem 5.5). An extension of this duality to all n ≥ 2 is crucial in relating
general nilpotency to identity functors with bounded degree.
The following characterisation is very useful: The identity functor of a σ-pointed
exact Mal’tsev category E is of degree ≤ n if and only if all its objects are n-folded
(cf. Proposition 6.5). An object is n-folded (cf. Definition 6.3) if the (n + 1)-st
X
folding map δn+1
: X + · · · + X → X factors through the comparison map θX,...,X :
X + · · · + X ։ PX,...,X . In a varietal context this can be expressed in combinatorial
terms (cf. Remark 6.4). The full subcategory Fldn (E) spanned by the n-folded objects
is a reflective Birkhoff subcategory of E, and the reflection J n : E → Fldn (E) is the
universal endofunctor of degree ≤ n (cf. Theorem 6.8). Every n-folded object is
n-nilpotent (cf. Proposition 6.13) while the converse holds if and only if the other
Birkhoff reflection I n : E → Niln (E) is also of degree ≤ n.
In the context of semi-abelian categories, closely related results appear in the work
of Hartl and his coauthors [38, 39, 40], although formulated slightly differently. In a
semi-abelian category, an object X is n-folded if and only if its Higgins commutator
of length n + 1 vanishes (cf. Remark 6.4), where the latter is defined as the image
of the composite map crn+1 (X, . . . , X) → X + · · · + X → X, cf. [38, 39, 54]. The
universal n-folded quotient J n (X) may then be identified with the quotient of X by
the Higgins commutator of length n + 1 in much the same way as the universal nnilpotent quotient I n (X) may be identified with the quotient of X by the iterated
Huq commutator of length n + 1. It was Hartl’s insight that Higgins commutators
are convenient for extending the “polynomial functors” of Eilenberg-Mac Lane [23]
to a semi-abelian context. Our treatment in the broader context of exact Mal’tsev
categories follows more closely Goodwillie’s functor calculus [29].
In a σ-pointed exact Mal’cev category, abelianisation I 1 is the universal endofunctor J 1 of degree ≤ 1 (cf. Mantovani-Metere [54]). For n > 1 however, the universal
endofunctor J n of degree ≤ n is in general a proper quotient of the n-th Birkhoff
reflection I n (cf. Corollary 6.12). In order to show that even in a semi-abelian variety
the two endofunctors may disagree, we exhibit a Moufang loop of order 16 (subloop
of Cayley’s octonions) which is 2-nilpotent but not 2-folded (cf. Section 6.5). Alternatively, Mostovoy’s modified lower central series of a loop [55] yields other examples
of a similar kind provided the latter agrees with the successive Higgins commutators
of the loop (cf. [39, Example 2.15] and [61]).
We did not find a simple categorical structure that would entail the equivalence
between n-nilpotency and n-foldedness for all n. As a first step in this direction
we show that an n-nilpotent semi-abelian category has an identity functor of degree
≤ n if and only if its n-th cross-effect is multilinear (cf. Theorem 6.22). We also
show that the nilpotency tower has the desired universal property if and only if it is
homogeneous, i.e. for each n, the n-th kernel functor is of degree ≤ n (cf. Theorem
6.23). This is preserved under Birkhoff reflection (cf. Theorem 6.24). The categories
of groups and of Lie algebras have homogeneous nilpotency towers so that a group,
resp. Lie algebra is n-nilpotent if and only if it is n-folded, and the Birkhoff reflection
CENTRAL REFLECTIONS AND NILPOTENCY IN EXACT MAL’TSEV CATEGORIES
5
I n is here indeed the universal endofunctor of degree ≤ n. The category of triality
groups [22, 28, 37] also has a homogeneous nilpotency tower although it contains the
category of Moufang loops as a full coreflective subcategory, and the latter has an
inhomogeneous nilpotency tower (cf. Section 6.5).
There are several further ideas closely related to the contents of this article which
we hope to address in future work. Let us mention two of them:
The associated graded object of the nilpotency tower ⊕n≥1 K[I n (X) ։ I n−1 (X)]
is a functor in X taking values in graded abelian group objects. For the category of
groups this functor actually takes values in graded Lie rings and as such preserves
n-nilpotent objects and free objects, cf. Lazard [51]. It is likely that for a large class of
semi-abelian categories, the associated graded object of the nilpotency tower carries
a similar algebraic structure. It would be interesting to establish the relationship
between this algebraic structure and the cross-effects of the identity functor.
It follows from [17, Theorem 4.2] and [59, Theorem IV.4] that the simplicial objects
of a pointed Mal’tsev variety carry a Quillen model structure in which the weak
equivalences are the maps inducing a quasi-isomorphism on Moore complexes. Such
a model structure also exists for the simplicial objects of a semi-abelian category with
enough projectives, cf. [59, 62]. In both cases, regular epimorphisms are fibrations,
and the trivial fibrations are precisely the regular epimorphisms for which the kernel
is homotopically trivial. This implies that Goodwillie’s homotopical cross-effects [29]
agree here with our algebraic cross-effects.
Several notions of homotopical nilpotency are now available. The first is the least
n
: X• ։ I n (X• ) is a trivial fibration, the second (resp.
integer n for which the unit ηX
•
third) is the least integer n for which X• is homotopically n-folded (resp. the value
of an n-excisive approximation of the identity). The first is a lower bound for the
second, and the second is a lower bound for the third invariant. For simplicial groups
the first invariant recovers the Berstein-Ganea nilpotency for loop spaces [2], the
second the cocategory of Hovey [42], and the third the Biedermann-Dwyer nilpotency
for homotopy nilpotent groups [3]. Similar chains of inequalities have recently been
studied by Eldred [24] and Costoya-Scherer-Viruel [21].
The plan of this article is as follows.
Section 1 reviews the notions of central extension and regular pushout. At the
end an algebraic Beck-Chevalley condition for pushouts of regular epimorphisms in
an exact Mal’tsev category is established.
Section 2 presents our definition of nilpotency and studies under which conditions
the n-nilpotent objects form a reflective Birkhoff subcategory.
Section 3 investigates central reflections, the motivating example being the reflection of the category of (n + 1)-nilpotent objects into the category of n-nilpotent
objects. The unit of these central reflections is shown to be pointwise affine.
Section 4 establishes first aspects of nilpotency. The nilpotency class of a σ-pointed
exact Mal’cev category is related to universal properties of the comparison map θX,Y :
X + Y → X × Y . This leads to a new family of binary tensor products interpolating
between binary sum and binary product.
6
CLEMENS BERGER AND DOMINIQUE BOURN
Section 5 studies the σ-pointed exact Mal’tsev categories with quadratic identity
functor. They are characterised among the 2-nilpotent ones as those which are algebraically distributive, resp. algebraically codistributive.
Section 6 studies the σ-pointed exact Mal’tsev categories with an identity functor
of degree ≤ n. They are characterised as those in which all objects are n-folded.
Every n-folded object is shown to be n-nilpotent. Several sufficient criteria for the
converse are given. The semi-abelian varieties of groups, Lie algebras, Moufang loops
and triality groups are discussed.
1. Central extensions and regular pushouts
In this introductory section we review the notion of central equivalence relation
and study basic properties of the associated class of central extensions, needed for
our treatment of nilpotency. By central extension we mean a regular epimorphism
with central kernel relation [8, 11, 12]. This algebraic concept of central extension
has to be distinguished from the axiomatic concept of Janelidze-Kelly [44] which is
based on a previously chosen admissible Birkhoff subcategory. Nevertheless, it is
known that with respect to the Birkhoff subcategory of abelian group objects, the
two approaches yield the same class of central extensions in any congruence modular
variety (cf. [45, 46]) as well as in any exact Mal’tsev category (cf. [11, 26, 31]).
We assume throughout that our ambient category is a Mal’tsev category, i.e. a
finitely complete category in which every reflexive relation is an equivalence relation,
cf. [4, 6, 17, 19]. Most of the material of this section is well-known to the expert, and
treated in some detail here mainly to fix notation and terminology.
One exception is Section 1.6 which establishes an “algebraic” Beck-Chevalley condition for pushouts of regular epimorphisms in exact Mal’tsev categories, dual to
the familiar Beck-Chevalley condition for pullbacks of monomorphisms in elementary
toposes. In recent and independent work, Gran-Rodelo [32] consider a weaker form
of this condition and show that it characterises regular Goursat categories.
1.1. Smith commutator of equivalence relations. –
An equivalence relation R on X will be denoted as a reflexive graph (p0 , p1 ) :
R ⇒ X with section s0 : X → R, but whenever convenient we shall consider R as a
subobject of X × X. By a fibrant map of equivalence relations (X, R) → (Y, S) we
mean a natural transformation of the underlying reflexive graphs such that the three
naturality squares are pullback squares.
A particularly important equivalence relation is the kernel relation R[f ] of a morphism f : X → Y which is part of the following diagram:
p1
R[f ] o
p0
/
/X
f
/ Y.
The discrete equivalence relation ∆X on X is the kernel relation R[1X ] of the
identity map 1X : X → X. The indiscrete equivalence relation ∇X on X is the kernel
relation R[ωX ] of the unique map ωX from X to a terminal object.
CENTRAL REFLECTIONS AND NILPOTENCY IN EXACT MAL’TSEV CATEGORIES
7
Two equivalence relations R, S on the same object X are said to centralise each
other if the square
R
(s0 ,1S )
R ×O X S o
S
(1R ,sS
0)
R
p
pR
0
pS
1
#
/X
admits a (necessarily unique) filler which makes the diagram commute, cf. [12, 57].
In set-theoretical terms such a filler amounts to the existence of a “partial Mal’tsev
operation” on X, namely (considering R ×X S as a subobject of X × X × X) a ternary
operation p : R ×X S → X such that x = p(x, y, y) and p(x, x, y) = y. We shall follow
Marino Gran and the second author in calling p : R ×X S → X a connector between
R and S, cf. [11, 12].
In a finitely cocomplete regular Mal’tsev category, there exists for each pair (R, S)
of equivalence relations on X a smallest effective equivalence relation [R, S] on X
such that R and S centralise each other in the quotient X/[R, S]. This equivalence
relation is the so-called Smith commutator of R and S, cf. [8, 12, 57, 60].
In these terms R and S centralise each other precisely when [R, S] = ∆X . The
Smith commutator is monotone in each variable and satisfies
[R, S] = [S, R] and f ([R, S]) ⊂ [f (R), f (S)]
for each regular epimorphism f : X → Y , where f (R) denotes the direct image of the
subobject R ⊂ X × X under the regular epimorphism f × f : X × X → Y × Y . The
Mal’tsev condition implies that this direct image represents an equivalence relation
on Y . Note that equality f ([R, S]) = [f (R), f (S)] holds if and only if the direct image
f ([R, S]) is an effective equivalence relation on Y , which is always the case in an exact
Mal’tsev category.
1.2. Central equivalence relations and central extensions. An equivalence relation R on X is said to be central if [R, ∇X ] = ∆X . A central extension is by
definition a regular epimorphism with central kernel relation. An n-fold central extension is the composite of n central extensions. An n-fold centrally decomposable
morphism is the composite of n morphisms with central kernel relation.
The indiscrete equivalence relation ∇X is a central equivalence relation precisely
when X admits an internal Mal’tsev operation p : X × X × X → X. In pointed
Mal’tsev categories such a Mal’tsev operation amounts to an abelian group structure
on X, cf. [4, Proposition 2.3.8]. An object X of a pointed Mal’tsev category (D, ⋆D )
is thus an abelian group object if and only if the map X → ⋆D is a central extension.
Central equivalence relations are closed under binary products and inverse image
along monomorphisms. In a regular Mal’tsev category, central equivalence relations
are closed under direct images, cf. [12, Proposition 4.2] and [4, Proposition 2.6.15].
Lemma 1.1. In a regular Mal’tsev category, an n-fold centrally decomposable morphism can be written as an n-fold central extension followed by a monomorphism.
8
CLEMENS BERGER AND DOMINIQUE BOURN
Proof. It suffices to show that a monomorphism ψ followed by a central extension φ
can be rewritten as a central extension φ′ followed by a monomorphism ψ ′ . Indeed,
the kernel relation R[φψ] is central, being the restriction ψ −1 (R[φ]) of the central
equivalence relation R[φ] along the monomorphism ψ; therefore, by regularity, one
obtains φψ = ψ ′ φ′ where φ′ is quotienting by the kernel relation R[φψ].
Lemma 1.2. In a Mal’tsev category, morphisms with central kernel relation are closed
under pullback. In a regular Mal’tsev category, central extensions are closed under
pullback.
Proof. It suffices to show the first statement. In the following diagram,
p′1
R[f ′ ] o
′
/X
p′0
R(x,y)
/ Y′
y
x
p1
R[f ] o
f′
/
/
/X
p0
f
/Y
if the right square is a pullback, then the left square is a fibrant map of equivalence
relations. This permits to lift the connector p : R[f ] ×X ∇X → X so as to obtain a
connector p′ : R[f ′ ] ×X ′ ∇X ′ → X ′ .
f
g
Lemma 1.3. Let X → Y → Z be morphisms in a Mal’tsev category.
If gf is a morphism with central kernel relation then so is f . More generally, if
gf is n-fold centrally decomposable then so is f .
Proof. Since R[f ] ⊂ R[gf ], the commutation relation [R[gf ], ∇X ] = ∆X implies the
commutation relation [R[f ], ∇X ] = ∆X . Assume now gf = kn · · · k1 where each ki is
a morphism with central kernel relation. In the following pullback
P o
γ
/X
φ
gf
ψ
Y
g
/Z
φ is the unique map such that γφ = 1X and ψφ = f . If we denote by hi the
morphism with central kernel relation obtained by pulling back ki along g, we get f =
hn · · · h2 (h1 φ). Since φ is a monomorphism, the kernel relation R[h1 φ] = φ−1 (R[h1 ]) is
central, and hence f is the composite of n morphisms with central kernel relation.
Proposition 1.4 (Corollary 3.3 in [8]). In a finitely cocomplete regular Mal’tsev
category, each morphism f : X → Y factors canonically as in
ηf
/ / X/[∇X , R[f ]]
❧❧
f
❧❧❧
v❧❧❧❧❧❧
o X/R[f ]
Y o
X
ζf
CENTRAL REFLECTIONS AND NILPOTENCY IN EXACT MAL’TSEV CATEGORIES
9
where ηf is a regular epimorphism and ζf has a central kernel relation. If f is a
regular epimorphism then ζf is a central extension.
This factorisation has the following universal property. Any commutative diagram
of undotted arrows as below (with Zf = X/[∇X , R[f ]]), such that ζ ′ has a central
kernel relation, produces a unique dotted map
x
X❅
❅❅ ηf
❅❅
❅❅
/ X′
❆❆ ′
❆❆ η
❆❆
❆
z
/ Z′
⑥
⑥
⑥
⑥⑥
~⑥⑥ ζ ′
/ Y′
Zf
f
⑦
⑦⑦
⑦
~⑦⑦ ζf
Y
y
rendering the whole diagram commutative.
Proposition 1.5. Let n be a positive integer. In a finitely cocomplete regular Mal’tsev
category, each morphism has a universal factorisation into a regular epimorphism
followed by an n-fold centrally decomposable morphism. Each n-fold central extension
has an initial factorisation into n central extensions.
Proof. We proceed by induction, the case n = 1 being treated in Proposition 1.4.
Suppose the assertion holds up to level n − 1 with f = ζfn−1 ζfn−2 · · · ζf1 ηf1 and ηf1 a
regular epimorphism. Take the universal factorisation of ηf1 . The universality of this
new factorisation is then a straightforward consequence of the induction hypothesis
and the universal property stated in Proposition 1.4.
Starting with an n-fold central extension f , its universal factorisation through a
composite of n − 1 morphisms with central kernel relation makes the regular epimorphism ηf1 a central extension by Lemma 1.3, and therefore produces a factorisation
into n central extensions which is easily seen to be the initial one.
1.3. Regular pushouts. In a regular category, any pullback square of regular epimorphisms is also a pushout square, as follows from the pullback stability of regular
epimorphisms. In particular, a commuting square of regular epimorphisms
X
f
x
X′
f′
//Y
y
/ / Y′
is a pushout whenever the comparison map (x, f ) : X → X ′ ×Y ′ Y to the pullback of
y along f ′ is a regular epimorphism. Such pushouts will be called regular, cf. [7]. A
regular pushout induces in particular regular epimorphisms on kernel relations which
we shall denote R(x, y) : R[f ] ։ R[f ′ ] and R(f, f ′ ) : R[x] ։ R[y].
For a regular Mal’tsev category the following more precise result holds.
10
CLEMENS BERGER AND DOMINIQUE BOURN
Proposition 1.6 (cf. Proposition 3.3 in [7]). In a regular Mal’tsev category, a commuting square of regular epimorphisms like in (1.3) is a regular pushout if and only
if one of the following three equivalent conditions holds:
(a) the comparison map X → X ′ ×Y ′ Y is a regular epimorphism;
(b) the induced map R(x, y) : R[f ] → R[f ′ ] is a regular epimorphism;
(c) the induced map R(f, f ′ ) : R[x] → R[y] is a regular epimorphism.
Accordingly, central extensions are closed under regular pushouts.
Proof. We already mentioned that (a) implies (b) and (c) in any regular category. It
remains to be shown that in a regular Mal’tsev category (b) or (c) implies (a).
For this, it is useful to notice that in a regular category condition (a) holds if and
only if the composite relation R[f ] ◦ R[x] equals the kernel relation of the diagonal of
the square by Theorem 5.2 of Carboni-Kelly-Pedicchio [17]. Since this kernel relation
is given by x−1 (R[f ′ ]), and condition (b) just means that x(R[f ]) = R[f ′ ], it suffices
to establish the identity R[f ] ◦ R[x] = x−1 (x(R[f ])). In a regular Mal’tsev category,
the composition of equivalence relations is symmetric and coincides with their join.
The join R[f ] ∨ R[x] is easily identified with x−1 (x([R[f ])).
The second assertion follows from (b) resp. (c) and the closure of central kernel
relations under direct image in regular Mal’tsev categories.
Corollary 1.7. In a regular Mal’tsev category, any commutative square
f
Xo
s
x
X′ o
f′
//Y
y
/ / Y′
s′
with a parallel pair of regular epimorphisms and a parallel pair of split epimorphisms
is a regular pushout.
Proof. The induced map R(f, f ′ ) : R[x] → R[y] is a split and hence regular epimorphism so that the pushout is regular by Proposition 1.6.
Corollary 1.8. In an exact Mal’tsev category, pushouts of regular epimorphisms
along regular epimorphims exist and are regular pushouts.
Proof. Given a pair (f, x) of regular epimorphisms with common domain, consider
the following diagram
p1
R[f ] o
x̃
So
p0
/
/X
f
y
x
q1
q0
/ ′
X
/
//Y
f′
/ / Y′
in which S denotes the direct image x(R[f ]). By exactness, this equivalence relation
on X ′ has a quotient Y ′ . The induced right square is then a regular pushout.
CENTRAL REFLECTIONS AND NILPOTENCY IN EXACT MAL’TSEV CATEGORIES
11
Remark 1.9. It follows from [6] that Corollary 1.7 characterises regular Mal’tsev categories among regular categories, while [17, Theorem 5.7] shows that Corollary 1.8
characterises exact Mal’tsev categories among regular categories.
Remark 1.10. It is worthwhile noting that in any category a commuting square of
epimorphisms in which one parallel pair admits compatible sections is automatically a
pushout square. Dually, a commuting square of monomorphisms in which one parallel
pair admits compatible retractions is automatically a pullback square.
Lemma 1.11 (cf. [11], Lemma 1.1). In a pointed regular category, a regular pushout
induces a regular epimorphism on parallel kernels
K[f ] /
K(x,y)
K[f ′ ] /
/X
f
y
x
/ X′
//Y
f′
// Y′
so that the kernel K[f ′ ] of f ′ is the image under x of the kernel K[f ] of f .
Proof. This follows from the fact that, in any pointed category, the induced map on
kernels factors as a pullback of the comparison map (x, f ) : X ։ X ′ ×Y ′ Y followed
by an isomorphism.
Proposition 1.12. In a finitely cocomplete regular Mal’tsev category, consider the
following diagram of pushouts
X
ηf
x
X′
/ / Zf
ζf
z
ηf ′
/ / Zf ′
ζf ′
//Y
y
// Y′
in which the upper row represents the universal factorisation of the regular epimorphism f into a regular epimorphism ηf followed by a central extension ζf .
If the outer rectangle is a regular pushout then the right square as well, and the
lower row represents the universal factorisation 1.4 of its composite f ′ = ζf ′ ηf ′ .
Proof. Since on vertical kernel relations R[x] → R[z] → R[y] we get a regular epimorphism, the second one R[z] → R[y] is a regular epimorphism as well, and the right
square is a regular pushout by Proposition 1.6. Therefore, ζf ′ is a central extension
by Corollary 1.7. It remains to be shown that the lower row fulfills the universal
property of factorisation 1.4. Consider the following diagram
X
ηf
/ / Zf
ζf
//Y
z
y
z′′ ζ ′
ηf ′
f
// Y′
/ / Zf ′
X ′ ◆◆
qq8 8
◆◆◆
q
q
◆
qq
η ◆◆◆
& ′ qqq ζ
Z
x
12
CLEMENS BERGER AND DOMINIQUE BOURN
with central extension ζ. According to Proposition 1.4, there is a unique dotted
factorisation z ′′ making the diagram commute. Since the left square is a pushout, z ′′
factors uniquely and consistently through z ′ : Zf ′ → Z ′ , showing that the lower row
has indeed the required universal property.
Proposition 1.13. In a finitely cocomplete exact Mal’tsev category, the universal
factorisation of a regular epimorphism through an n-fold central extension is preserved
under pushouts along regular epimorphisms.
Proof. Let us consider the following diagram of pushouts
η
X
x
X′
η
/ / Zn
ζn
/ / Zn−1
zn
zn
′
/ / Zn′
/ / Zn−1
′
′
ζn
ζ1
...
Z1
...
z1
Z1′
//Y
y
// Y′
′
ζ1
in which the upper row is the universal factorisation 1.5 of a regular epimorphism
f : X → Y through an n-fold central extension. By Corollary 1.8, all pushouts are
′
regular. Therefore, the morphisms ζk′ : Zk′ → Zk−1
are central extensions for all
k. It remains to be shown that the lower row satisfies the universal property of the
factorisation 1.5 of f ′ : X ′ → Y ′ through an n-fold central extension. This follows
induction on n beginning with the case n = 1 proved in Proposition 1.12.
Proposition 1.14. Let D be an exact Mal’tsev category. Consider the following
diagram of pushouts
X
xn
X′
/ / Xn−1 fn−1/ / Xn−2
fn
xn−1
′
/ / Xn−1
′
fn
...
X1
...
x1
X1′
xn−2
′
/ / Xn−2
′
fn−1
f1
/ / X0
x0
/ / X0′
′
f1
f0
//Y
y
/ / Y′
′
f0
in which xn : X ։ X ′ is a regular epimorphism.
If the upper row represents an n-fold central extension of the regular epimorphism
f0 : X0 ։ Y in the slice category D/Y then the lower row represents an n-fold central
extension of f0′ : X0′ ։ Y ′ in the slice category D/Y ′ .
Proof. Let us set φi = f0 · f1 · · · fi and φ′i = f0′ · f1′ · · · fi′ . Since the indiscrete equivalence relation ∇f0 on the object f0 : X0 ։ Y of the slice category D/Y is given by
R[f0 ], our assumption on the upper row translates into the conditions
[R[fi ], R[φi ]] = ∆Xi for 1 ≤ i ≤ n.
Since any of the rectangles is a regular pushout by Corollary 1.8, we get xi (R[fi ]) =
R[fi′ ] and xi (R[φi ]) = R[φ′i ], and consequently [R[fi′ ], R[φ′i ]] = ∆Xi′ for all i.
CENTRAL REFLECTIONS AND NILPOTENCY IN EXACT MAL’TSEV CATEGORIES
13
1.4. Regular pushouts in pointed Mal’tsev categories with binary sums.
In a pointed category with binary sums and binary products, each pair of objects
(X1 , X2 ) defines a canonical comparison map θX1 ,X2 : X1 + X2 → X1 × X2 , uniquely
determined by the requirement that the composite morphism
Xi /
θX1 ,X2
/ X1 + X2
/ X1 × X2
/ / Xj
is the identity (resp. the null morphism) if i = j (resp. i 6= j), where i, j ∈ {1, 2}.
Recall that θX1 ,X2 is a strong epimorphism for all objects X1 , X2 precisely when the
category is unital in the sense of the second author, and that every pointed Mal’tsev
category is unital, cf. [4, 6]. In a regular category strong and regular epimorphisms
coincide.
Note also that an exact Mal’tsev category has coequalisers for reflexive pairs, so
that an exact Mal’tsev category with binary sums has all finite colimits. In order to
shorten terminology, we call σ-pointed any pointed category with binary sums.
Later we shall need the following two examples of regular pushouts.
Proposition 1.15. For any regular epimorphism f : X ։ Y and any object Z of a
σ-pointed regular Mal’tsev category, the following square
X+Z
θX,Z
/ / X ×Z
f +Z
Y +Z
θY,Z
f ×Z
/ / Y ×Z
is a regular pushout.
Proof. The regular epimorphism θR[f ],Z : R[f ] + Z ։ R[f ] × Z factors as below
θ
R[f ],Z
/ / R[f ] × Z = R[f × Z]
R[f ] + Z
▼▼▼
5
❧❧❧5
▼▼▼
❧❧❧
❧
▼▼▼
❧
❧❧❧
&
R[f + Z]
inducing a regular epimorphism R[f + Z] → R[f × Z] on the vertical kernel relations
of the square above. Proposition 1.6 allows us to conclude.
Corollary 1.16. For any objects X, Y, Z of a σ-pointed regular Mal’tsev category,
the following square
(X + Y ) + Z
θX+Y,Z
θX,Y +Z
(X × Y ) + Z
is a regular pushout.
θX×Y,Z
/ / (X + Y ) × Z
θX,Y ×Z
/ / (X × Y ) × Z
14
CLEMENS BERGER AND DOMINIQUE BOURN
1.5. Central subobjects, centres and centralisers. In a pointed Mal’tsev category (D, ⋆D ), two morphisms with common codomain f : X → Z and g : Y → Z are
said to commute [9, 43] if the square
(0Y X ,1Y )
X ×O Y o
Y
φf,g
(1X ,0XY )
X
f
g
#
/Z
admits a (necessarily unique) filler φf,g : X × Y → Z making the whole diagram
commute, where 0XY : X → ⋆D → Y denotes the zero morphism. A monomorphism
Z X which commutes with the identity 1X : X → X is called central, and the
corresponding subobject is called a central subobject of X.
Every regular epimorphism f : X ։ Y with central kernel relation R[f ] has a
central kernel K[f ]. In pointed protomodular categories, the converse is true: the
centrality of K[f ] implies the centrality of R[f ], so that central extensions are precisely
the regular epimorphisms with central kernel, cf. [33, Proposition 2.2].
Recall [4, 5] that a pointed category is protomodular precisely when the category
has pullbacks of split epimorphisms, and for each split epimorphism, section and
kernel-inclusion form a strongly epimorphic cospan. Every finitely complete protomodular category is a Mal’tsev category [4, Proposition 3.1.19]. The categories of
groups and of Lie algebras are pointed protomodular. Moreover, in both categories,
each object possesses a centre, i.e. a maximal central subobject. Central group (resp.
Lie algebra) extensions are thus precisely regular epimorphisms f : X ։ Y with
kernel K[f ] contained in the centre of X. This is of course the classical definition of
a central extension in group (resp. Lie) theory.
In these categories, there exists more generally, for each subobject N of X, a
so-called centraliser, i.e. a subobject Z(N X) of X which is maximal among
subobjects commuting with N X. The existence of centralisers has far-reaching
consequences, as shown by James Gray and the second author, cf. [13, 35, 36]. Since
they are useful for our study of nilpotency, we discuss some of them here.
Following [6], we denote by PtZ (D) the category of split epimorphisms (with chosen
section) in D over a fixed codomain Z, cf. Section 3.2. For each f : Z → Z ′ pulling
back along f defines a functor f ∗ : PtZ ′ (D) → PtZ (D) which we call pointed basechange along f . In particular, the terminal map ωZ : Z → 1D defines a functor
(ωZ )∗ : D → PtZ (D). Since in a pointed regular Mal’tsev category D, morphisms
commute if and only if their images commute, morphisms in PtZ (D) of the form
φf,f ′
/
X ×Z
⑦? Y
❍c❍❍❍❍
❍❍❍❍ ⑦s⑦⑦⑦⑦⑦
pZ ❍❍❍# ⑦⑦⑦r
Z
f
correspond bijectively to morphisms f : X → K[r] such that X −→ K[r] Y
commutes with s : Z Y in D. Therefore, if split subobjects have centralisers in D,
CENTRAL REFLECTIONS AND NILPOTENCY IN EXACT MAL’TSEV CATEGORIES
15
then for each object Z, the functor (ωZ )∗ : D → PtZ (D) : X 7→ X × Z admits a right
adjoint (ωZ )∗ : PtZ (D) → D : (r, s) 7→ K[r] ∩ Z(s).
A category with the property that for each object Z, the functor (ωZ )∗ has a right
adjoint is called algebraically cartesian closed [13]. Algebraic cartesian closedness
implies canonical isomorphisms (X × Z) +Z (Y × Z) ∼
= (X + Y ) × Z for all objects
X, Y, Z, a property we shall call algebraic distributivity, cf. Section 5.3.
1.6. An algebraic Beck-Chevalley condition. –
The dual of an elementary topos is an exact Mal’tsev category, cf. [17, Remark
5.8]. This suggests that certain diagram lemmas for elementary toposes admit a
dual version in our algebraic setting. Supporting this analogy we establish here an
“algebraic dual” of the well-known Beck-Chevalley condition. As a corollary we get a
diagram lemma which will be used several times in Section 6.
Another instance of the same phenomenon is the cogluing lemma for regular epimorphisms in exact Mal’tsev categories (cf. proof of Theorem 6.23a and Corollary
1.8) which is dual to a gluing lemma for monomorphisms in an elementary topos.
Lemma 1.17 (cf. Lemma 1.1 in [30]). Consider a commutative diagram
X
Y
x
y
/ / X′
/ X ′′
/ /Y′
/ Y ′′
in a regular category.
If the outer rectangle is a pullback and the left square is a regular pushout (1.3)
then left and right squares are pullbacks.
Proof. The whole diagram contains three comparison maps: one for the outer rectangle, denoted φ : X → Y ×Y ′′ X ′′ , one for the left and one for the right square, denoted
respectively φl : X → Y ×Y ′ X ′ and φr : X ′ → Y ′ ×Y ′′ X ′′ . We get the identity
φ = y ∗ (φr ) ◦ φl where y ∗ denotes base-change along y. Since the outer rectangle is a
pullback, φ is invertible so that φl is a section and y ∗ (φr ) a retraction.
Since the left square is a regular pushout, the comparison map φl is a regular
epimorphism and hence φl and y ∗ (φr ) are both invertible. Since y is a regular epimorphism in a regular category, base-change y ∗ is conservative so that φr is invertible
as well, i.e. both squares are pullbacks.
Proposition 1.18 (Algebraic Beck-Chevalley). Let D be an exact Mal’tsev category
with pushouts of split monomorphisms along regular epimorphisms.
Any pushout of regular epimorphisms
Ū
ḡ
u
U
g
/ / V̄
v
//V
16
CLEMENS BERGER AND DOMINIQUE BOURN
yields a functor isomorphism ḡ! u∗ ∼
= v ∗ g! from the fibre PtU (D) to the fibre PtV (D).
Proof. We have to show that for any point (r, s) over U , the following diagram
Ū ′ ❇ ′
G ❇❇u
r̄ ✍✍✍✍
❇! !
✍✍ ✍
✍✍ ✍✍
U′
✍
F
r ✌✌ ✌
✍✍ ✍✍✍
✌
✌
✍
✌
✌✌
Ū ❆
✌✌✌✌ ḡ
❆❆
✌
✌
u ❆ ✌✌ ✌
✌
U
ḡ′
/ / V̄ ′ ′
G ❇❇v❇
r̄ ✍✍✍✍
❇! !
✍
′
✍
g
/ /V′
✍✍ ✍✍✍
✍
F
✍✍ ✍✍✍
✌✌✌✌
✌
✍
✌
r ′ ✌✌ ✌
/ / V̄
✌
❆❆
✌✌✌✌
❆
v ❆ ✌✌ ✌
✌
//V
g
′
in which (r̄, s̄) = u∗ (r, s) and g! (r, s) = (r′ , s′ ) and ḡ! (r̄, s̄) = (r̄′ , s̄′ ), has a right face
which is a downward-oriented pullback; indeed, this amounts to the required identity
v ∗ (r′ , s′ ) = (r̄′ , s̄′ ).
Since bottom face and the upward-oriented front and back faces are pushouts, the
top face is a pushout as well, which is regular by Corollary 1.8. Taking pullbacks
in top and bottom faces induces a split epimorphism U ′ ×V ′ V̄ ′ ։ U ×V V̄ through
which the left face of the cube factors as in the following commutative diagram
Ū ′
r̄
Ū
/ / U ′ ×V ′ V̄ ′
/ / U′
r
/ / U ×V V̄
//U
in which the left square is a regular pushout by Corollary 1.7. Lemma 1.17 shows
then that the right square is a pullback. Therefore, we get the following cube
U ′ ×V ′ V̄▲ ′
▲▲▲
⑥⑥⑥>
▲▲& &
⑥⑥⑥⑥⑥
⑥
⑥⑥⑥
⑥
U′
⑥⑥⑥⑥⑥
r ☎☎☎A
⑥
☎☎
~⑥⑥⑥⑥⑥
☎☎☎☎☎
☎
U ×V V̄P
ḡ
☎☎☎☎
PPP
PPP
☎☎☎☎☎
☎
P( (
☎
U
g
/ / V̄ ′ ′
G ❇❇v❇
r̄ ✎✎✎✎
❇
✎✎ ✎
// V′
✎✎ ✎✎
✎✎ ✎✎
✍ ✍F
✍✍ ✍✍✍
✎ ✎✎
r ′ ✍ ✍✍
/ V̄
✍✍
❅❅
❅
✍✍ ✍
v ❅❅ ✍✍ ✍✍
✍
//V
′
in which the downward-oriented left face and the bottom face are pullbacks. Therefore, the composite of the top face followed by the downward-oriented right face is a
pullback. Moreover, as above, the top face is a regular pushout. It follows then from
Lemma 1.17 that the downward-oriented right face is a pullback as required.
Corollary 1.19. In an exact Mal’tsev category with pushouts of split monomorphisms
along regular epimorphisms, each commuting square of natural transformations of split
CENTRAL REFLECTIONS AND NILPOTENCY IN EXACT MAL’TSEV CATEGORIES
17
epimorphisms
ŪG ′ ❇ ′
u
f¯ ✍✍ ✍ ❇❇❇
!!
✍✍ ✍✍
✍✍ ✍✍
U′
F
✍✍ ✍✍
✌
f
✌ ✌✌
✍ ✍✍
✌
✌ ✌✌
Ū
✌✌✌✌ ḡ
❆❆
✌
✌
✌✌✌✌
u ❆❆
U
ḡ′
/ / V̄ ′ ′
✍ ✍G ❇❇v❇
✍
❇! !
✍ ✍✍
g′
/ /V′
✍✍ ✍✍✍
F
✍✍ ✍✍
✌✌✌✌
✍ ✍✍
✌
✌
f ′ ✌✌ ✌
/ / V̄
✌ ✌✌
❆❆v
✌
❆❆ ✌✌✌✌✌
//V
f¯′
g
such that all horizontal arrows are regular epimorphisms, and front and back faces are
upward-oriented pushouts, induces the following upper pushout square
ḡ′
Ū✮ T✮′ ❑
✮✮ ✮✮ ❑❑❑❑(f¯,u′ )
❑❑
✮✮ ✮✮
❑❑
✮✮ ✮✮
❑% %
✮✮ ✮✮
Ū
×U U ′
f¯ ✮ ✮
✮✮ ✮✮
✁✁✁@
✮✮ ✮✮
✁✁✁✁✁
✁
✮
✮
✁✁
✁✁✁✁
Ū ✁
ḡ
/ / V̄ ′
✮✮ T✮✮ ❑❑❑
❑❑(f¯′ ,v′ )
✮✮
❑❑
f¯′ ✮ ✮
❑❑
✮✮ ✮✮
❑% %
✮
′✮
ḡ×g ✮ ✮
/ / V̄ ×V V ′
✮✮ ✮✮
✮✮ ✮✮
@
✮✮ ✮✮
✮
✮
/ / V̄
in which the kernel relation of the regular epimorphism (f¯′ , v ′ ) : V̄ ′ → V̄ ×V V ′ may
be identified with the intersection R[f¯′ ] ∩ R[v ′ ].
Proof. Taking downward-oriented pullbacks in left and right face of the first diagram
yields precisely a diagram as studied in the proof of Proposition 1.18. This implies
that the front face of the second diagram is an upward-oriented pushout. Since
the back face of the second diagram is also an upward-oriented pushout, the upper
square is a pushout as well, as asserted. The kernel relation of the comparison map
(f¯′ , v ′ ) : V̄ ′ → V̄ ×V V ′ is the intersection of the kernel relations of f¯′ and of v ′ .
2. Affine objects and nilpotency
2.1. Affine objects.
Definition 2.1. Let C be a full subcategory of a Mal’tsev category D.
An object X of D is said to be C-affine if there exists a morphism f : X → Y in
D with central kernel relation R[f ] and with codomain Y in C.
The morphism f is called a C-nilindex for X.
We shall write Aff C (D) for the full replete subcategory of D spanned by the C-affine
objects of D. Clearly Aff C (D) contains C.
When C consists only of a terminal object 1D of D, we call the C-affine objects
simply the affine objects of D and write Aff C (D) = Aff(D). Recall that the unique
morphism X → 1D has a central kernel relation precisely when the indiscrete equivalence relation on X centralises itself, which amounts to the existence of a (necessarily
unique associative and commutative) Mal’tsev operation on X.
18
CLEMENS BERGER AND DOMINIQUE BOURN
When D is pointed, such a Mal’tsev operation on X induces (and is induced by)
an abelian group structure on X. For a pointed Mal’tsev category D, the category
Aff(D) of affine objects is thus the category Ab(D) of abelian group objects of D.
Remark 2.2. When D is a regular Mal’tsev category, any nilindex f : X → Y factors
as a regular epimorphism f˜ : X ։ f (X) followed by a monomorphism f (X) Y
with codomain in C; therefore, if C is closed under taking subobjects in D, this defines
a strongly epimorphic nilindex f˜ for X with same central kernel relation as f . In
other words, for regular Mal’tsev categories D and subcategories C which are closed
under taking subobjects in D, the C-affine objects of D are precisely the objects which
are obtained as central extensions of objects of C.
Proposition 2.3. For any full subcategory C of a Mal’tsev category D, the subcategory
Aff C (D) is closed under taking subobjects in D. If C is closed under taking binary
products in D then Aff C (D) as well, so that Aff C (D) is finitely complete.
Proof. Let m : X X ′ be a monomorphism with C-affine codomain X ′ . If f ′ :
X ′ → Y ′ is a nilindex for X ′ , then f m : X → Y ′ is a nilindex for X, since central
equivalence relations are stable under pointed base-change along monomorphisms and
we have R[f m] = m−1 (R[f ]).
If X and Y are C-affine with nilindices f and g then f × g is a nilindex for X × Y
since maps with central kernel relations are stable under products.
A Birkhoff subcategory [44] of a regular category D is a subcategory C which is
closed under taking subobjects, products and quotients in D. A Birkhoff subcategory of an exact (resp. Mal’tsev) category is exact (resp. Mal’tsev), and regular
epimorphisms in C are those morphisms in C which are regular epimorphisms in D.
If D is a variety (in the single-sorted monadic sense) then Birkhoff subcategories
of D are precisely subvarieties of D, cf. the proof of Lemma 5.11 below.
Proposition 2.4. Let C be a full subcategory of an exact Mal’tsev category D.
If C is closed under taking subobjects and quotients in D then Aff C (D) as well. In
particular, if C is a Birkhoff subcategory of D, then Aff C (D) as well.
Proof. Let X be a C-affine object of D with nilindex f : X → Y . We can suppose f
is a regular epimorphism, cf. Remark 2.2. Thanks to Proposition 2.3 it remains to
establish closure under quotients. Let g : X ։ X ′ be a regular epimorphism in D.
Since D is exact, the pushout of f along g exists in D
X
f
g
X′
//Y
h
f′
// Y′
and f ′ is a central extension since f is, cf. Corollary 1.8. By hypothesis C is stable
under quotients. Therefore the quotient Y ′ belongs to C, and f ′ is a nilindex for X ′
so that X ′ is C-affine as required.
CENTRAL REFLECTIONS AND NILPOTENCY IN EXACT MAL’TSEV CATEGORIES
19
2.2. The C-lower central sequence. –
Definition 2.1 is clearly the beginning of an iterative process. We write C = Nil0C (D)
and define inductively NilnC (D) to be the category Aff Niln−1 (D) (D). The objects of this
C
category NilnC (D) are called the C-nilpotent objects of order n of D, and we get the
following diagram
♦7 = DO a❇h◗◗◗
♦♦⑤♦⑤⑤ ❇❇❇◗◗◗◗◗
♦
♦
❇❇ ◗◗◗
♦♦ ⑤⑤
◗◗◗
❇❇
♦♦♦ ⑤⑤⑤
♦
◗◗◗
❇❇
♦
♦
⑤
◗◗◗
♦
❇
❇a
♦♦
=⑤
O
⑤
h
♦
♦
7/♦ /
n
2
1
NilC (D) / / Niln+1
C
NilC (D) / / NilC (D)
(D)
C
which we call the C-lower central sequence of D.
If C = {1D }, we obtain the (absolute) lower central sequence of D:
7 D h◗
♥♥⑤♥⑤> O a❇❇◗❇◗◗◗◗◗
♥
♥
❇❇ ◗◗
♥♥ ⑤⑤
◗◗◗
❇❇
♥♥♥ ⑤⑤⑤
◗◗◗
♥
❇❇
♥
♥
⑤
◗◗◗
♥
❇
⑤
♥
❇
♥
◗◗h
⑤
❇a
>⑤
O
7♥♥♥
2
n
1
/ Nil (D) / / Nil (D)
Nil (D) / / Niln+1 (D)
{1D } /
Remark 2.5. It follows from Remark 2.2 and an iterative application of Proposition
2.4 that for an exact Mal’tsev category D, the nilpotent objects of order n are precisely
those which can be obtained as an n-fold central extension of the terminal object 1D
and that moreover Niln (D) is a Birkhoff subcategory of D.
If D is the category of groups (resp. Lie algebras) then Niln (D) is precisely the full
subcategory spanned by nilpotent groups (resp. Lie algebras) of class ≤ n. Indeed,
it is well-known that a group (resp. Lie algebra) is nilpotent of class ≤ n precisely
when it can be obtained as an n-fold “central extension” of the trivial group (resp.
Lie algebra), and we have seen in Section 1.5 that the group (resp. Lie) theorist’s
definition of central extension agrees with ours. We will see in Proposition 2.14
below that the equivalence between the central extension definition and the iterated
commutator definition of nilpotency carries over to our general context of finitely
cocomplete exact Mal’tsev categories. Huq [43] had foreseen a long time ago that a
categorical approach to nilpotency was possible. Everaert-Van der Linden [25] recast
Huq’s approach in modern language in the context of semi-abelian categories.
Definition 2.6. A Mal’tsev category D with full subcategory C is called C-nilpotent
of order n (resp. of class n) if D = NilnC (D) (resp. if n is the least such integer).
When C = {1D } the prefix C will be dropped, and instead of “nilpotent of order
n” we also just say “n-nilpotent”.
Proposition 2.7. A Mal’tsev category is n-nilpotent if and only if each morphism is
n-fold centrally decomposable.
A regular Mal’tsev category is n-nilpotent if and only if each morphism factors as
an n-fold central extension followed by a monomorphism.
20
CLEMENS BERGER AND DOMINIQUE BOURN
Proof. The second statement follows from the first by Lemma 1.1. If each morphism
is n-fold centrally decomposable, then this holds for terminal maps ωX : X → 1D , so
that all objects are n-nilpotent. Conversely, assume that all objects are n-nilpotent,
i.e. that for all objects X, the terminal map ωX is n-fold centrally decomposable.
Then, for each morphism f : X → Y , the identity ωX = ωY f together with Lemma
1.3 imply that f is n-fold centrally decomposable as well.
2.3. Epireflections, Birkhoff reflections and central reflections. –
We shall see that if C is a reflective subcategory of D, then the categories NilnC (D)
are again reflective subcategories of D, provided D and the reflection fulfill suitable
conditions. In order to give precise statements we need to fix some terminology.
A full replete subcategory C of D is called reflective if the inclusion C ֒→ D admits
a left adjoint functor I : D → C, called reflection. The unit of the adjunction at
an object X of D will be denoted by ηX : X → I(X). Reflective subcategories C
are stable under formation of limits in D. In particular, reflective subcategories of
Mal’tsev categories are Mal’tsev categories.
A reflective subcategory C of D is called strongly epireflective and the reflection
I is called a strong epireflection if the unit ηX : X → I(X) is pointwise a strong
epimorphism. Strongly epireflective subcategories are characterised by the property
that C is closed under taking subobjects in D. In particular, strongly epireflective
subcategories of regular categories are regular categories.
A Birkhoff reflection (cf. [14]) is a strong epireflection I : D → C such that for
each regular epimorphism f : X → Y in D, the following naturality square
X
f
Y
ηX
/ / I(X)
I(f )
/ / I(Y )
ηY
is a regular pushout (see Section 1.3 and Proposition 1.6).
A subcategory of D defined by a Birkhoff reflection is a Birkhoff subcategory of
D, and is thus exact whenever D is. It follows from Corollary 1.8 that a reflective
subcategory of an exact Mal’tsev category is a Birkhoff subcategory if and only if the
reflection is a Birkhoff reflection.
A central reflection is a strong epireflection I : D → C with the property that the
unit ηX : X ։ I(X) is pointwise a central extension.
The following exactness result will be used at several places. In the stated generality, it is due to Diana Rodelo and the second author [14], but the interested reader
can as well consult [44, 31, 32] for closely related statements.
Proposition 2.8. In a regular Mal’tsev category, strong epireflections preserve pullback squares of split epimorphisms, and Birkhoff reflections preserve pullbacks of split
epimorphisms along regular epimorphisms.
Proof. See Proposition 3.4 and Theorem 3.16 in [14].
CENTRAL REFLECTIONS AND NILPOTENCY IN EXACT MAL’TSEV CATEGORIES
21
Lemma 2.9. Let C be a reflective subcategory of D with reflection I, and assume that
ηX : X → I(X) factors through an epimorphism f : X ։ Y as in:
ηX
/ I(X)
♦7
♦
♦
♦
f
♦♦♦♦♦ η
Y
X
Then I(f ) is an isomorphism and we have η = I(f )−1 ηY .
Proof. Consider the following diagram
ηX
/ I(X)
♣8
♣
♣
f
I(f )
♣♣
♣♣♣♣ η
/
Y
I(Y )
ηY
X
where the lower triangle commutes because f is an epimorphism. If we apply the
reflection I to the whole diagram we get two horizontal isomorphisms I(ηX ) and I(ηY ).
It follows that I(η) is an isomorphism as well, hence so is I(f ), and η = I(f )−1 ηY .
Lemma 2.10. For any reflective subcategory C of a Mal’tsev category D the C-affine
objects of D are those X for which the unit ηX has a central kernel relation.
Proof. If ηX : X → I(X) has a central kernel relation then X is C-affine. Conversely,
let X be C-affine with nilindex f : X → Y . Then Y is an object of the reflective
subcategory C so that f factors through ηX : X → I(X). Accordingly, we get
R[ηX ] ⊂ R[f ], and hence R[ηX ] is central because R[f ] is.
Corollary 2.11. A reflection I : D → C of a regular Mal’tsev category D is central
if and only if D is C-nilpotent of order 1 (i.e. all objects of D are C-affine).
Theorem 2.12 (cf. [8], Sect. 4.3 in [43], Prp. 7.8 in [25] and Thm. 3.6 in [49]). –
For a reflective subcategory C of a finitely cocomplete regular Mal’tsev category D,
the category Aff C (D) is a strongly epireflective subcategory of D.
The associated strong epireflection IC1 : D → Aff C (D) is obtained by factoring the
unit ηX : X → I(X) universally through a map with central kernel relation.
If C is a reflective Birkhoff subcategory of a finitely cocomplete exact Mal’tsev
category D, then the reflection IC1 : D → Aff C (D) is a Birkhoff reflection.
Proof. Proposition 1.4 yields the following factorisation of the unit:
X
1
ηX
IC1 (X)
ηX
/ I(X)
♥7
♥
♥
♥♥
♥♥♥ η̄X
Since η̄X has a central kernel relation and I(X) is an object of C, the object IC1 (X)
1
belongs to Aff C (D). We claim that the maps ηX
: X ։ IC1 (X) have the universal
1
property of the unit of an epireflection IC : D → Aff C (D).
22
CLEMENS BERGER AND DOMINIQUE BOURN
Let f : X → T be a map with C-affine codomain T which means that ηT has a
central kernel relation. Then consider the following diagram:
f
X■
■■ η1
■■ X
■■
■$ $
ηX
IC1 (X)
✈
✈✈
✈✈
{✈✈ η̄X
I(X)
I(f )
/T ❈
❈❈❈❈
❈❈❈❈
❈❈❈❈
❈❈❈
f¯
/T
⑤
⑤
⑤
⑤⑤
~⑤⑤ ηT
/ I(T )
According to Proposition 1.4, there is a unique factorisation f¯ making the diagram
commute. If D is exact and C a Birkhoff subcategory, the subcategory Aff C (D) is
closed under taking subobjects and quotients by Proposition 2.4. The reflection IC1 is
thus a Birkhoff reflection in this case.
Remark 2.13. A reflective Birkhoff subcategory C of a semi-abelian category D satisfies all hypotheses of the preceding theorem. In this special case, the Birkhoff reflection
IC1 : D → Aff C (D) is given by the formula
IC1 (X) = X/[X, K[ηX ]]
where [X, K[ηX ]] is the Huq commutator of X and K[ηX ], cf. Section 1.5.
Indeed, the pointed protomodularity of D implies (cf. [33, Proposition 2.2]) that
the kernel of the quotient map X → X/[∇X , R[ηX ]] is canonically isomorphic to the
Huq commutator [X, K[ηX ]] so that the formula follows from Proposition 1.4.
2.4. The Birkhoff nilpotency tower. –
According to Theorem 2.12, any reflective Birkhoff subcategory C of a finitely
cocomplete exact Mal’tsev category D produces iteratively the following commutative
diagram of Birkhoff reflections:
D ◗◗
♦♦⑤♦⑤ ❇❇❇◗❇◗◗◗◗
♦
♦
❇❇ ◗◗◗◗
♦♦ ⑤⑤
❇❇ n ◗◗◗
♦1♦♦ ⑤⑤⑤
♦
♦
◗◗◗ I n+1
❇❇ IC
I
♦ C ⑤⑤
IC2
◗◗C◗
❇❇
I ♦♦♦♦
⑤
}
(
⑤
♦
!
♦
w♦
1
2
n
NilC (D) o
NilC (D)
NilC (D) o
Niln+1
(D)
C o
C
A Birkhoff subcategory of an exact Mal’tsev category is an exact Mal’tsev category
so that the subcategories NilnC (D) are all exact Mal’tsev categories, and the horizontal
n
reflections Niln+1
C (D) → NilC (D) are central reflections by Corollary 2.11.
In the special case C = {1D } we get the following commutative diagram of Birkhoff
subcategories and Birkhoff reflections:
CENTRAL REFLECTIONS AND NILPOTENCY IN EXACT MAL’TSEV CATEGORIES
23
D◗
♥♥⑤♥⑤ ❇❇◗❇◗◗◗◗◗
♥
♥
❇❇ ◗◗
♥ ⑤
◗◗
❇❇
♥♥♥ ⑤⑤⑤
♥
♥
❇❇ n ◗◗◗◗◗ n+1
♥ I 1 ⑤⑤
♥
2
♥
❇
◗◗I◗
I
⑤
I
I ♥♥
❇❇
◗(
~⑤⑤
!
w♥♥♥
Nil1 (D) o
Nil2 (D)
Niln (D) o
Niln+1 (D)
{1D } o
1
If D is pointed, then the first Birkhoff reflection I 1 = I{⋆
: D → Nil1 (D) can be
D}
identified with the classical abelianisation functor D → Ab(D). In particular, the
abelian group objects of D are precisely the nilpotent objects of order 1.
When C is a reflective Birkhoff subcategory of a finitely cocomplete exact Mal’tsev
category D, then D is C-nilpotent of class n if and only if n is the least integer such
that either the unit of the n-th Birkhoff reflection ICn is invertible, or equivalently, the
(n − 1)st Birkhoff reflection ICn−1 is a central reflection, see Corollary 2.11.
Proposition 2.14. For an exact Mal’tsev category D with binary sums, the unit of
n
the n-th Birkhoff reflection ηX
: X → I n (X) is given by quotienting out the iterated
Smith commutator [∇X , [∇X , [∇X , · · · , ∇X ]]] of length n + 1.
If D is semi-abelian, this unit is also given as the quotient of X by the iterated
Huq commutator [X, [X, [X, . . . , X]]] of length n + 1.
Proof. The second statement follows from the first by Remark 2.13. The first staten
ment follows from the inductive construction of ηX
in the proof of Theorem 2.12
together with Proposition 1.4.
A finite limit and finite colimit preserving functor is called exact. A functor between
exact Mal’tsev categories with binary sums is exact if and only if it preserves finite
limits, regular epimorphisms and binary sums, cf. Section 1.4.
Lemma 2.15. Any exact functor F : D → E between exact Mal’tsev categories with
binary sums commutes with the n-th Birkhoff reflections, i.e. IEn ◦ F ∼
= F|Niln (D) ◦ IDn .
Proof. According to Theorem 2.12 and Proposition 1.4 it suffices to show that F
takes the canonical factorisation of f : X ։ Y through the central extension ζf :
X/[∇X , R[f ]] ։ Y to the factorisation of F (f ) : F (X) ։ F (Y ) through the central
extension ζF (f ) : F (X)/[∇F (X) , R[F (f )]] ։ F (Y ). Since F is left exact, we have
F (∇X ) = ∇F (X) and F (R[f ]) = R[F (f )], and since F preserves regular epimorphisms, we have F (X/[∇X , R[f ]]) = F (X)/F ([∇X , R[f ]]). It remains to be shown
that F preserves Smith commutators. This follows from exactness of F and the fact
that in a finitely cocomplete exact Mal’tsev category the Smith commutator is given
by an explicit formula involving only finite limits and finite colimits.
3. Affine morphisms and central reflections
We have seen that the nilpotency tower of a σ-pointed exact Mal’tsev category is
a tower of central reflections. In this section we establish a useful general property
of central reflections in exact Mal’tsev categories, namely that the unit of a central
24
CLEMENS BERGER AND DOMINIQUE BOURN
reflection is pointwise affine. Since this property might be useful in other contexts as
well, we first discuss possible weakenings of the notion of exactness.
3.1. Quasi-exact and efficiently regular Mal’tsev categories. –
An exact category is a regular category in which equivalence relations are effective,
i.e. arise as kernel relations of some morphism. In general, effective equivalence
relations R on X have the property that the inclusion R X × X is a strong
monomorphism. Equivalence relations with this property are called strong. A regular
category in which strong equivalence relations are effective is called quasi-exact.
Any quasi-topos (cf. Penon [58]) is quasi-exact so that there are plenty examples
of quasi-exact categories which are not exact. There are also quasi-exact Mal’tsev
categories which are not exact, as for instance the category of topological groups
and continuous group homomorphisms. Further weakenings of exactness occur quite
naturally as shown in the following chain of implications:
exact
+3 quasi-exact
+3 efficiently regular
+3 fibrational kernel relations
A category is efficiently regular [14] if every equivalence relation (X, S), which
is a regular refinement of an effective equivalence relation (X, R), is itself effective.
By regular refinement we mean any map of equivalence relations (X, S) → (X, R)
inducing the identity on X and a regular monomorphism S → R.
We call a kernel relation (X, R) fibrational if for each fibrant map of equivalence
relations (Y, S) → (X, R) the domain is an effective equivalence relation as well.
According to Janelidze-Sobral-Tholen [48] a kernel relation (X, R) is fibrational precisely when its quotient map X ։ X/R has effective descent, i.e. base-change along
X ։ X/R is a monadic functor. A regular category has thus fibrational kernel relations precisely when all regular epimorphisms have effective descent. The careful
reader will observe that in all proofs of this section where we invoke efficient regularity
we actually just need that the considered kernel relations are fibrational.
The second implication above follows from the fact that any regular monomorphism
is a strong monomorphism, while the third implication follows from the facts that for
any fibrant map of equivalence relations f : (Y, S) → (X, R) the induced map on
relations S → f ∗ (R) is a regular (even split) monomorphism, and that in any regular
category, effective equivalence relations are closed under inverse image.
3.2. Fibration of points and essentially affine categories. –
Recall [6] that for any category D, we denote by Pt(D) the category whose objects
are split epimorphisms with chosen section (“genereralised points”) of D and whose
morphisms are natural transformations between such (compatible with the chosen
sections), and that ¶D : Pt(D) → D denotes the functor associating to a split epimorphism its codomain.
The functor ¶D : Pt(D) → D is a fibration (the so-called fibration of points) whenever D has pullbacks of split epimorphisms. The ¶D -cartesian maps are precisely
pullbacks of split epimorphisms. Given any morphism f : X → Y in D, base-change
along f with respect to the fibration ¶D is denoted by f ∗ : PtY (D) → PtX (D), and will
CENTRAL REFLECTIONS AND NILPOTENCY IN EXACT MAL’TSEV CATEGORIES
25
be called pointed base-change in order to distinguish it from the classical base-change
D/Y → D/X on slices.
Pointed base-change f ∗ has a left adjoint pointed cobase-change f! if and only if
pushouts along f of split monomorphisms with domain X exist in D. In this case
pointed cobase-change along f is given by precisely this pushout, cf. [5]. Accordingly,
the unit (resp. counit) of the (f! , f ∗ )-adjunction is an isomorphism precisely when for
each natural transformation of split epimorphisms
′
X
O
s
f′
r
X
/ Y′
O
s′
f
/Y
r′
the downward square is a pullback as soon as the upward square is a pushout (resp.
upward square is a pushout as soon as the downward square is a pullback).
It is important that in a regular Mal’tsev category pointed base-change along a
regular epimorphism is fully faithful : in a pullback square like above with regular
epimorphism f : X ։ Y , the upward-oriented square is automatically a pushout. This
follows from fact that the induced morphism on kernel relations R(s, s′ ) : R[f ] → R[f ′ ]
together with the diagonal X ′ → R[f ′ ] forms a strongly epimorphic cospan because
the kernel relation R[f ′ ] is the product of R[f ] and X ′ in the fibre PtX (D), and all
fibres are unital in virtue of the Mal’tsev condition, cf. [4, 6].
Recall [6] that a category is called essentially affine if pushouts of split monomorphisms and pullbacks of split epimorphisms exist, and moreover for any morphism
f : X → Y the pointed base-change adjunction (f! , f ∗ ) is an adjoint equivalence.
Additive categories with pullbacks of split epimorphisms are essentially affine. This
follows from the fact that in this kind of category every split epimorphism is a projection, and every split monomorphism is a coprojection. Conversely, a pointed essentially affine category is an additive category with pullbacks of split epimorphisms.
Any slice or coslice category of an additive category with pullbacks of split epimorphisms is an example of an essentially affine category that is not pointed, and hence
not additive. Therefore, the property of a morphism f : X → Y in D to induce an
adjoint equivalence f ∗ : PtY (D) → PtX (D) expresses somehow a “relative additivity”
of f . This motivates the following definition:
Definition 3.1. In a category D with pullbacks of split epimorphisms, a morphism
f : X → Y will be called ¶D -affine when the induced pointed base-change functor
f ∗ : PtD (Y ) → PtD (X) is an equivalence of categories.
An affine extension in D is any regular epimorphism which is ¶D -affine.
Clearly any isomorphism is ¶D -affine. It follows from the analogous property of
equivalences of categories that for any composable morphisms f, g in D, if two among
f , g and gf are ¶D -affine then so is the third, i.e. ¶D -affine morphisms fulfill the
so-called two-out-of-three property.
It might be confusing that we use the term “affine” in two different contexts, namely
as well for objects as well for morphisms. Although their respective definitions seem
26
CLEMENS BERGER AND DOMINIQUE BOURN
unrelated at first sight, this isn’t the case. We will see in Proposition 3.2 that every
affine extension is a central extension, and it will follow from Theorem 3.3 that for a
reflective subcategory C of an efficiently regular Mal’tsev category D, every object of
D is C-affine if and only if every object of D is an affine extension of an object of C.
We hope that this justifies our double use of the term “affine”.
Proposition 3.2. In any Mal’tsev category D, the kernel relation of a ¶D -affine
morphism is central. In particular, each affine extension is a central extension.
Proof. Let f : X → Y be an ¶D -affine morphism with kernel relation
(p0 , p1 ) : R[f ] ⇒ X
and let
X
(pX
0 , p1 ) : X × X ⇒ X
be the indiscrete equivalence relation on X with section sX
0 : X → X × X. Since
pointed base-change f ∗ is an equivalence of categories, there is a split epimorphism
X
(r, s) : Y ′ ⇄ Y such that f ∗ (r, s) = (pX
0 , s0 ) and we get the right hand pullback of
diagram
R[fˇ] o
O
R(pX
0 ,r)
p̌1
p̌0
R(sX
0 ,s)
R[f ] o
p1
p0
/
/ X ×O X
pX
0
fˇ
pX
1
/
/X
/ Y′
O
s
r
f
/Y
in which the left hand side are the respective kernel relations. Therefore the left hand
side consists of two pullbacks, and the map pX
1 p̌0 produces the required connector
between R[f ] and the indiscrete equivalence relation ∇X on X.
In particular, if an epireflection I : D → C of a Mal’tsev category D has a pointwise
¶D -affine unit ηX : X → I(X), then it is a central reflection. The following converse
will be essential in understanding nilpotency.
Theorem 3.3. A central reflection I of an efficiently regular Mal’tsev category D has
a unit which is pointwise an affine extension. In particular, morphisms f : X → Y
with invertible image I(f ) : I(X) → I(Y ) are necessarily ¶D -affine.
Proof. The second assertion follows from the first since ¶D -affine morphisms fulfill the
two-out-of-three property and isomorphisms are ¶D -affine.
For the first assertion note that in a regular Mal’tsev category pointed base-change
along a regular epimorphism is fully faithful, and hence ηY∗ : PtI(Y ) (D) → PtY (D) is
a fully faithful functor. Corollary 3.6 below shows that ηY∗ is essentially surjective,
hence ηY∗ is an equivalence of categories for all objects Y in D.
CENTRAL REFLECTIONS AND NILPOTENCY IN EXACT MAL’TSEV CATEGORIES
27
3.3. Centralising double relations. Given a pair (R, S) of equivalence relations
on Y , we denote by RS the inverse image of the reflexive relation S × S under
R
(pR
0 , p1 ) : R Y × Y . This defines a double relation
q1S
o
RS
O
/
/ SO
q0S
q1R
q0R
Ro
pS
1
pS
0
pR
1
/
/Y
pR
0
actually the largest double relation relating R and S. In set-theoretical terms, this
double relation RS corresponds to the subset of elements (u, v, u′ , v ′ ) of Y 4 such
that the relations uRu′ , vRv ′ , uSv, u′ Sv ′ hold.
Lemma 3.4. Any split epimorphism (r, s) : X ⇄ Y of a regular Mal’tsev category
with epireflection I and unit η induces the following diagram
R[ηR[r] ] o
O
R(pr0 ,Ipr0 )
R(pr1 ,Ipr1 )
R[ηX ] o
O
R(r,Ir)
R(s,Is)
R[ηY ] o
/
/ R[r]
O
ηR[r]
pr1
pr0
/
/ XO
r
/
/Y
/ / I(R[r])
O
Ipr0
ηX
s
/ / IX
O
Ir
ηY
Ipr1
Is
/ / IY
in which the rows and the two right vertical columns represent kernel relations. The
left most column represents then the kernel relation of the induced map R(r, Ir) :
R[ηX ] → R[ηY ], and we have R[ηR[r] ] = R[ηX ]R[r] = R[R(r, Ir)].
Proof. This follows from Proposition 2.8 which shows that I(R[r]) may be identified
with the kernel relation R[I(r)] of the split epimorphism Ir : IX → IY .
For sake of simplicity a split epimorphism (r, s) : X ⇄ Y is called a C-affine point
over Y whenever its domain X is C-affine.
Proposition 3.5. Let C be a reflective subcategory of an efficiently regular Mal’tsev
category D with reflection I and unit η.
Any C-affine point (r, s) over Y is the image under ηY∗ of a C-affine point (r̄, s̄)
over IY such that both points have isomorphic reflections in C.
28
CLEMENS BERGER AND DOMINIQUE BOURN
Proof. Since by Lemma 2.10 ηX has a central kernel relation, the kernel relations
R[ηX ] and R[r] centralise each other. This induces a centralising double relation
/
R[ηX ] ×X R[r] o
O
/ R[r]
O
pr0
R[ηX ] o
pr1
/
/X
which we consider as a fibrant split epimorphism of equivalence relations (disregarding
the dotted arrows). Pulling back along the monomorphism
/
R[ηX ] o
O
/ XO
s
R(s,Is)
/
/Y
R[ηY ] o
yields on the left hand side of the following diagram
/
RI [(r, s)] o
O
/ XO
s
r
R[ηY ] o
q
/
/Y
/ / X̄
O
r̄
ηY
s̄
/ / IY
another fibrant split epimorphism of equivalence relations. Since D is efficiently regular, the upper equivalence relation is effective with quotient q : X ։ X̄. We claim
that the induced point (r̄, s̄) : X̄ ⇄ IY has the required properties.
Indeed, the right square is a pullback by a well-known result of Barr and Kock,
cf. [4, Lemma A.5.8], so that ηY∗ (r̄, s̄) = (r, s). The centralising double relation
R[ηX ] ×X R[r] is coherently embedded in the double relation R[ηX ]R[r] of Lemma
3.4, cf. [4, Proposition 2.6.13]. This induces an inclusion RI [(r, s)] R[ηX ] and
hence a morphism φ : X̄ ։ IX such that φq = ηX .
According to Lemma 2.9, we get an isomorphism Iq : IX ∼
= I X̄ compatible with
the units. The kernel relation R[ηX̄ ] is thus the direct image of the central kernel
relation R[ηX ] under the regular epimorphism q : X → X̄ and as such central as well.
In particular, (r̄, s̄) is a C-affine point with same reflection in C as (r, s).
Corollary 3.6. For an efficiently regular Mal’tsev category D with central reflection
I and unit η, pointed base-change ηY∗ : PtI(Y ) (D) → PtY (D) is essentially surjective.
Proof. Since the reflection is central, Corollary 2.11 shows that Proposition 3.5 applies
to the whole fibre PtY (D) whence essential surjectivity of ηY∗ .
CENTRAL REFLECTIONS AND NILPOTENCY IN EXACT MAL’TSEV CATEGORIES
29
3.4. Affine extensions in efficiently regular Mal’tsev categories. –
A functor G : E → E′ is called saturated on quotients if for each object A in E and
each strong epimorphism g ′ : G(A) → B ′ in E′ , there exists a strong epimorphism
g : A → B in E such that G(g) and g ′ are isomorphic under G(A). Note that a
right adjoint functor G : E → E′ is essentially surjective whenever it is saturated on
quotients and each object B ′ of E′ is the quotient of an object B ′′ for which the unit
ηB ′′ : B ′′ → GF (B ′′ ) is invertible.
Lemma 3.7. In an efficiently regular Mal’tsev category, pointed base-change along a
regular epimorphism is saturated on quotients.
Proof. Let f : X ։ Y be a regular epimorphism, let (r, s) be a point over Y , and
l : f ∗ ((r, s)) ։ (r′ , s′ ) be a quotient map over X. Consider the following diagram
pf1
′
′
/
f′
′
/ X O ❇❇
❇❇ l
❇❇
❇❇!
/ ! ′′
/X
⑤⑤
⑤
∗
∗
⑤
f (r) f (s)
⑤⑤
}⑤} ⑤⑤ r′
/
/X
R[f ] o
O ❈❈
❈❈ pf0 ′
❈❈
❈❈!
!
So
④
④
④
④④
}④} ④④
pf1
o
R[f ]
// Y′
O
l̄
f ′′
r
s
~ ~
//Y
f
pf0
/ / Y ′′
ρ
in which the right hand side square is a pullback, and the left hand side is defined by
factoring the induced map on kernel relations R[f ′ ] → R[f ] through the direct image
S = l(R[f ′ ]) under l. Since the right square is a pullback, the left square represents
a fibrant split epimorphism of equivalence relations. The factorisation of this fibrant
morphism induced by l yields two fibrant maps of equivalence relations, cf. Lemma
1.17. Note that the second (X ′′ , S) → (X, R[f ]) is a fibrant split epimorphism. Efficient regularity implies then that the equivalence relation S is effective with quotient
f ′′ : U ′′ ։ V ′′ , defining a point (ρ, σ) over Y .
This induces (by a well-known result of Barr and Kock, cf. [4, Lemma A.5.8]) a
decomposition of the right pullback into two pullbacks. The induced regular epimorphism l̄ : (r, s) ։ (ρ, σ) has the required properties, namely f ∗ (l̄) = l.
Proposition 3.8. In an efficiently regular Mal’tsev category with binary sums, a
regular epimorphism f : X ։ Y is an affine extension if and only if for each object
Z either of the following two diagrams
X +O Z
Z
πX
f +Z
ιZ
X
X
//Y +Z
O
Z
πY
f
ιZ
Y
//Y
is a downward-oriented pullback square.
X +Z
f +Z
/ / Y +Z
θX,Z
X ×Z
θY,Z
f ×Z
/ / Y ×Z
30
CLEMENS BERGER AND DOMINIQUE BOURN
Proof. If f is an affine extension then the downward-oriented left square is a pullback
because the upward-oriented left square is a pushout, cf. Section 3.2. Moreover, the
outer rectangle of the following diagram
f +Z
X +Z
/ Y +Z
θY,Z
θX,Z
X ×Z
/ Y ×Z
f ×Z
pY
pX
X
/Y
f
is a pullback if and only if the upper square is a pullback, because the lower square
is always a pullback.
Assume conversely that the downward oriented left square is a pullback. In a
regular Mal’tsev category, pointed base-change along a regular epimorphism is fully
faithful so that f is affine whenever f ∗ is essentially surjective. Lemma 3.7 shows that
in an efficiently regular Mal’tsev category f ∗ is saturated on quotients. It suffices thus
to show that in the fibre over X each point is the quotient of a point for which the
unit of the pointed base-change adjunction is invertible.
Since for each object Z, the undotted downward-oriented square
X +O Z
Z
πX
f +Z
Z
πY
h1X ,ri
X
/ / Y +Z
O
h1Y ,f ri
//Y
f
is a pullback, the dotted downward-oriented square (which is induced by an arbitrary
morphism r : Z → X) is a pullback as well. This holds in any regular Mal’tsev
category, since the whole diagram represents a natural transformation of reflexive
graphs, cf. [6]. It follows that the point (h1X , ri, ιZ
X ) : X + Z ⇄ X has an invertible
unit with respect to the pointed base-change adjunction (f! , f ∗ ).
Now, an arbitrary point (r, s) : Z ⇄ X can be realised as a quotient
Z ❅o o
_❅❅❅❅
❅❅❅❅r
❅❅❅❅
s ❅❅❅
hs,1Z i
X; + Z
①①①①
①
①①①①
①①①①
①{①①①① h1X ,ri
ιZ
X
X
of the latter point (h1X , ri, ιZ
X ) : X + Y ⇄ X with invertible unit.
We end this section with several properties of affine extensions in semi-abelian
categories. They will only be used in Section 6.
CENTRAL REFLECTIONS AND NILPOTENCY IN EXACT MAL’TSEV CATEGORIES
31
Proposition 3.9. In a semi-abelian category, a regular epimorphism f : X ։ Y is
an affine extension if and only if either of the following conditions is satisfied:
(a) for each object Z, the induced map f ⋄ Z : X ⋄ Z → Y ⋄ Z is invertible, where
X ⋄ Z stands for the kernel of θX,Z : X + Z ։ X × Z;
(b) every pushout of f along a split monomorphism is a pullback.
Proof. That condition (a) characterises affine extensions follows from Proposition 3.8
and protomodularity. The necessity of condition (b) follows from Section 3.2. The
sufficiency of condition (b) follows from the “pullback cancellation property” in semi
abelian categories, cf. [4, Proposition 4.1.4].
Remark 3.10. This product X ⋄ Z is often called the co-smash product of X and Z,
since it is the dual of the smash product as investigated by Carboni-Janelidze [16] in
the context of lextensive categories. The co-smash product X ⋄ Z coincides in semiabelian categories with the second cross-effect cr2 (X, Z) of the identity functor, cf.
Definition 5.1 and [54, 38, 39]. Since the co-smash product is in general not associative
(cf. [16]), parentheses should be used with care.
Proposition 3.11. Let Y −→ W ←− Z and Ȳ −→ W̄ ←− Z̄ be cospans in the fibre
PtX (D) of a semi-abelian category D. Let f : Y ։ Ȳ , g : Z ։ Z̄, h : W ։ W̄ be
affine extensions in D inducing a map of cospans in PtX (D). Assume furthermore
that the first cospan realises W as the binary sum of Y and Z in PtX (D).
Then the second cospan realises W̄ as the binary sum of Ȳ and Z̄ in PtX (D) if
and only if the kernel cospan K[f ] −→ K[h] ←− K[g] is strongly epimorphic in D.
Proof. Let us consider the following commutative diagram
X /
i
/Y
f
/ / Ȳ
j
Z /
g
Z̄ /
/ W h1 / / W1
❈❈
❈❈h
❈❈
h2
❈! !
/ W2
/ / W̄
in which i (resp. j) denotes the section of the point Y (resp. Z) over X, and all
little squares except the lower right one are pushouts. It follows that the outer square
is a pushout (i.e. W̄ = Ȳ +X Z̄) if and only if the lower right square is a pushout.
According to Carboni-Kelly-Pedicchio [17, Theorem 5.2] this happens if and only if
the kernel relation R[h] is the join of the kernel relations R[h1 ] and R[h2 ]. In a semiabelian category this is the case if and only if the kernel K[h] is generated as normal
subobject of W by the kernels K[h1 ] and K[h2 ], resp. (since h1 and h2 are affine
extensions) by the kernels K[f ] and K[g], cf. Proposition 3.9b.
Now, h is also an affine extension so that by Proposition 3.2, the kernel K[h] is a
central subobject of W . In particular, any subobject of K[h] is central and normal
in W (cf. the characterisation of normal subobjects in semi-abelian categories by
Mantovani-Metere [54, Theorem 6.3]). Therefore, generating K[h] as normal subobject of W amounts to the same as generating K[h] as subobject of W .
32
CLEMENS BERGER AND DOMINIQUE BOURN
4. Aspects of nilpotency
Recall that a morphism is called n-fold centrally decomposable if it is the composite
of n morphisms with central kernel relation.
For consistency, a monomorphism is called 0-fold centrally decomposable, and an
isomorphism a 0-fold central extension.
Proposition 4.1. For all objects X, Y of a σ-pointed n-nilpotent Mal’tsev category,
the comparison map θX,Y : X + Y → X × Y is (n − 1)-fold centrally decomposable.
Proof. In a pointed n-nilpotent Mal’tsev category, each object maps to an abelian
group object through an (n − 1)-fold centrally decomposable morphism, cf. Proposition 2.7. Since the codomain of such a morphism φX,Y : X + Y → A is an abelian
group object, the restrictions to the two summands commute and φX,Y factors
θX,Y
X + Y❍
❍❍
❍❍
φX,Y ❍$
A
/ / X ×Y
✈✈
✈
✈ψ
✈
z✈ X,Y
so that θX,Y is (n − 1)-fold centrally decomposable by Lemma 1.3.
Proposition 4.2. For any finitely cocomplete regular pointed Mal’tsev category, the
following pushout square
X +X
h1X ,1X i
X
θX,X
/ / X ×X
/ / A(X)
defines the abelianisation A(X) of X. In particular, the lower row can be identified
1
with the unit ηX
: X → I 1 (X) of the strong epireflection of Theorem 2.12.
Proof. The first assertion follows by combining [4, Proposition 1.7.5, Theorems 1.9.5
and 1.9.11] with the fact that pointed Mal’tsev categories are strongly unital in the
sense of the second author, cf. [4, Corollary 2.2.10]. The second assertion expresses
the fact that X ։ A(X) and X ։ I 1 (X) share the same universal property.
Theorem 4.3. A σ-pointed exact Mal’tsev category is n-nilpotent if and only if for
all objects X, Y the comparison map θX,Y : X + Y → X × Y is an (n − 1)-fold central
extension.
Proof. By Proposition 4.1 n-nilpotency implies that θX,Y is an (n − 1)-fold central
extension. For the converse, consider the pushout square of Proposition 4.2, which is
1
regular by Corollary 1.8. The unit ηX
: X → I 1 (X) is thus an (n − 1)-fold central
extension by Proposition 1.13 so that all objects are n-nilpotent.
Corollary 4.4. For a σ-pointed exact Mal’tsev category, the following three properties
are equivalent:
(a) the category is 1-nilpotent;
(b) the category is linear (cf. Definition 5.1);
CENTRAL REFLECTIONS AND NILPOTENCY IN EXACT MAL’TSEV CATEGORIES
33
(c) the category is abelian.
Proof. The equivalence of (a) and (b) follows from Theorem 4.3. The equivalence of
(b) and (c) follows from the fact that a σ-pointed Mal’tsev category is additive if and
only if it is linear (cf. [4, Theorem 1.10.14]) together with the well-known fact (due
to Miles Tierney) that abelian categories are precisely the additive categories among
exact categories.
Theorem 4.5. For a σ-pointed exact Mal’tsev category, the following five properties
are equivalent:
(a)
(b)
(b′ )
(c)
(c′ )
all objects are 2-nilpotent;
1
for all X, abelianisation ηX
: X → I 1 (X) is a central extension;
1
for all X, abelianisation ηX : X → I 1 (X) is an affine extension;
for all X, Y , the map θX,Y : X + Y → X × Y is a central extension;
for all X, Y , the map θX,Y : X + Y → X × Y is an affine extension.
Proof. Properties (a) and (b) are equivalent by definition of 2-nilpotency. Theorem
4.3 shows that (b) and (c) are equivalent. Theorem 3.3 and Proposition 3.2 imply
that (b) and (b′ ) are equivalent.
Finally, since the first Birkhoff reflection I 1 preserves binary sums and binary
products (cf. Proposition 2.8), we get I 1 (θX,Y ) = θI 1 (X),I 1 (Y ) which is invertible
in the subcategory of 1-nilpotent objects by Proposition 4.1. It follows that under
assumption (a), the map θX,Y is an affine extension by Theorem 3.3, which is property
(c′ ). Conversely, (c′ ) implies (c) by Proposition 3.2.
4.1. Niltensor products. In order to extend Proposition 4.5 to higher n we introduce here a new family of binary tensor products, called niltensor products.
For any finitely cocomplete pointed regular Mal’tsev category (D, ⋆D ) the n-th niltensor product X ⊗n Y is defined by factorizing the comparison map θX,Y universally
n
into a regular epimorphism θX,Y
: X + Y → X ⊗n Y followed by an (n − 1)-fold
central extension
ω n−1
ω2
X,Y
X,Y
/ / X ⊗n−1 Y
X ⊗3 Y❞❞❞❞❞/2 /2 X ⊗2 Y◆
X; ⊗n Y
❞
4
❞
4
◆◆◆ ω1
✇;
n
θX,Y
❞❞❞❞❞❞❞
❥❥❥❥
✇✇
◆◆◆X,Y
❞❞❞❞❞❞❞
n−1
❞
❞
✇✇ ❥❥❥❥❥❥❥θX,Y
❞
❞
◆◆◆
❞
✇
❞
❞
❞
2
❞
✇
❥
❞
❞
❥
❞
θX,Y
◆& &
❥
❞❞❞
✇✇❥
❞❥❞❞❞❞❞❞
/ /X ×Y
X +Y
θX,Y
as provided by Proposition 1.5. This n-th niltensor product is symmetric and has ⋆D
as unit, but it does not seem to be associative in general.
34
CLEMENS BERGER AND DOMINIQUE BOURN
Proposition 4.6. In a σ-pointed exact Mal’tsev category, the following diagram is
an iterated pushout diagram
X: ⊗n X
n
θX,X
✉✉:
✉
✉
✉✉
✉✉
X +X
X●
●●
●●
●●
n
●●#
ηX
#
I n (X)
n−1
ωX,X
/ / X ⊗n−1 X
X ⊗3 X
θX,X
ηX
/ / I n−1 (X)
I (X)
3
2
ωX,X
/ / X ⊗2 X
◆◆◆ ω1
◆◆◆X,X
◆◆◆
◆' '
/ / X ×X
/ / I 1 (X)
q8 8
qqq
q
q
qqq
/ / I 2 (X)
where the left vertical map is the folding map h1X , 1X i : X + X → X.
Proof. This follows from Corollary 1.8 and Propositions 1.13 and 4.2.
Theorem 4.7. For a σ-pointed exact Mal’tsev category, the following five properties
are equivalent:
(a)
(b)
(b′ )
(c)
(c′ )
all objects are n-nilpotent;
n−1
for all X, the (n − 1)th unit ηX
: X ։ I n−1 (X) is a central extension;
n−1
for all X, the (n − 1)th unit ηX : X ։ I n−1 (X) is an affine extension;
n−1
for all X, Y , the map θX,Y
: X + Y → X ⊗n−1 Y is a central extension;
n−1
for all X, Y , the map θX,Y : X + Y → X ⊗n−1 Y is an affine extension.
Proof. Properties (a) and (b) are equivalent by definition of n-nilpotency. Theorem
4.3 shows that (a) implies (c) while Proposition 4.6 shows that (c) implies (b). Therefore, (a), (b) and (c) are equivalent. Proposition 3.2 and Theorem 3.3 imply that (b)
and (b′ ) are equivalent.
Finally, Proposition 1.13 implies that the Birkhoff reflection I n−1 takes the comn−1
parison map θX,Y
to the corresponding map θIn−1
n−1 (X),I n−1 (Y ) for the (n − 1)-nilpotent
n−1
n−1
objects I
(X) and I
(Y ). Since the (n − 1)-nilpotent objects form an (n − 1)nilpotent Birkhoff subcategory, Theorem 4.3 shows that the latter map must be invertible; therefore, (a) implies (c′ ) by Theorem 3.3. Conversely, (c′ ) implies (c) by
Proposition 3.2.
Definition 4.8. A σ-pointed Mal’tsev category is said to be pseudo-additive (resp.
n
pseudo-n-additive) if for all X, Y, the map θX,Y : X + Y ։ X × Y (resp. θX,Y
:
X + Y ։ X ⊗n Y ) is an affine extension.
CENTRAL REFLECTIONS AND NILPOTENCY IN EXACT MAL’TSEV CATEGORIES
35
Proposition 4.9. A σ-pointed exact Mal’tsev category is pseudo-additive (i.e. 2nilpotent) if and only if the following diagram
(X + Y ) + Z
θX,Y +Z
/ / (X × Y ) + Z
θX×Y,Z
θX+Y,Z
(X + Y ) × Z
/ / (X × Y ) × Z
θX,Y ×Z
is a pullback for all objects X, Y, Z.
Proof. This follows from Theorem 4.5 and Proposition 3.8.
Proposition 4.10. A σ-pointed exact Mal’tsev category is pseudo-(n − 1)-additive
(i.e. n-nilpotent) if and only if the following diagram
(X + Y ) + Z
n−1
θX,Y
+Z
θX+Y,Z
(X ⊗n−1 Y ) + Z
/ / (X + Y ) × Z
θX⊗n−1 Y,Z
n−1
θX,Y
×Z
/ / (X ⊗n−1 Y ) × Z
is a pullback for all objects X, Y, Z.
Proof. This follows from Theorem 4.7 and Proposition 3.8.
We end this section with a general remark about the behaviour of n-nilpotency
under slicing and passage to the fibres. Note that any left exact functor between
Mal’tsev categories preserves central equivalence relations, morphisms with central
kernel relation, and consequently n-nilpotent objects.
Proposition 4.11. If D is an n-nilpotent Mal’tsev category, then so are any of its
slice categories D/Y and of its fibres PtY (D).
Proof. The slices D/Y of a Mal’tsev category D are again Mal’tsev categories. Moreover, base-change ωY∗ : D → D/Y is a left exact functor so that the objects of D/Y
of the form ωY∗ (X) = pY : Y × X → Y are n-nilpotent provided D is an n-nilpotent
Mal’tsev category. We can conclude with Proposition 2.3 by observing that any object
f : X → Y of D/Y may be considered as a subobject
(f,1X )
/ Y ×X
X ❆/
❆❆
✇✇
❆❆
✇✇
✇
❆
❆ {✇✇✇ pY
f
Y
36
CLEMENS BERGER AND DOMINIQUE BOURN
of ωY∗ (X) in D/Y . The proof for the fibres is the same as for the slices, since any
object (r, s) of PtY (D) may be considered as a subobject
(r,1X )
/ Y ×X
X `❆/
❆❆❆❆ s
①①①①<
①
p
❆❆❆❆
Y ①① ①
①
❆❆❆❆
①①①①①
①
❆
❆
①
①
(1
❆
❆
Y ,s)
r
❆❆❆ |①①①①①
Y
of the projection pY : Y × X → Y splitted by (1Y , s) : Y → Y × X.
5. Quadratic identity functors
We have seen that 1-nilpotency has much to do with linear identity functors (cf.
Corollary 4.4). We now investigate the relationship between 2-nilpotency and quadratic identity functors, and below in Section 6, the relationship between n-nilpotency
and identity functors of degree n. While a linear functor takes binary sums to binary
products, a quadratic functor takes certain cubes constructed out of triple sums to
limit cubes. This is the beginning of a whole hierarchy assigning degree ≤ n to a
functor whenever the functor takes certain (n + 1)-dimensional cubes constructed out
of iterated sums to limit cubes.
This definition of degree of a functor is much inspired by Goodwillie [29] who described polynomial approximations of a homotopy functor in terms of their behaviour
on certain cubical diagrams. Eilenberg-Mac Lane [23] defined the degree of a functor
with values in an abelian category by a vanishing condition of so-called cross-effects.
Our definition of degree does not need cross-effects. Yet, a functor with values in
a semi-abelian (or homological [4]) category is of degree ≤ n precisely when all its
cross-effects of order n + 1 vanish, cf. Corollary 6.17. Our cubical cross-effects agree
up to isomorphism with those of Hartl-Loiseau [38] and Hartl-Van der Linden [39],
which are defined as kernel intersections.
There are several other places in literature where degree n functors, especially quadratic functors, are studied in a non-additive context, most notably Baues-Pirashvili
[1], Johnson-McCarthy [50] and Hartl-Vespa [40]. In all these places, the definition of
a degree n functor is based on a vanishing condition of cross-effects, closely following
the original approach of Eilenberg-Mac Lane. It turned out that for us Goodwillie’s
cubical approach to functor calculus was more convenient. Our Definition 6.3 of an
n-folded object and the resulting characterisation of degree n identity functors in
terms of n-folded objects (cf. Proposition 6.5) rely in an essential way on cubical
combinatorics.
5.1. Degree and cross-effects of a functor. –
An n-cube in a category E is given by a functor Ξ : [0, 1]n → E with domain the
n-fold cartesian product [0, 1]n of the arrow category [0, 1].
The category [0, 1] has two objects 0,1 and exactly one non-identity arrow 0 → 1.
Thus, an n-cube in E is given by objects Ξ(ǫ1 , . . . , ǫn ) in E with ǫi ∈ {0, 1}, and
CENTRAL REFLECTIONS AND NILPOTENCY IN EXACT MAL’TSEV CATEGORIES
37
ǫ′ ,...,ǫ′
arrows ξǫ11,...,ǫnn : Ξ(ǫ1 , . . . , ǫn ) → Ξ(ǫ′1 , . . . , ǫ′n ) in E, one for each arrow in [0, 1]n ,
which compose in an obvious way.
To each n-cube Ξ we associate a punctured n-cube Ξ̌ obtained by restriction of Ξ
to the full subcategory of [0, 1]n spanned by the objects (ǫ1 , . . . , ǫn ) 6= (0, . . . , 0).
Definition 5.1. Let (E, ⋆E ) be a σ-pointed category. For each n-tuple of objects
(X1 , . . . , Xn ) of E we denote by ΞX1 ,...,Xn the following n-cube:
• ΞX1 ,...,Xn (ǫ1 , . . . , ǫn ) = X1 (ǫ1 ) + · · · + Xn (ǫn )
with X(0) = X and X(1) = ⋆E ;
′n
ǫ′
ǫ′ ,...,ǫ′
• ξǫ11,...,ǫnn = jǫ11 + · · · + jǫǫn
′
where jǫǫ is the identity if ǫ = ǫ′ , resp. the null morphism if ǫ 6= ǫ′ .
A functor of σ-pointed categories F : (E, ⋆E ) → (E′ , ⋆E′ ) is called of degree ≤ n if
• F (⋆E ) ∼
= ⋆ E′ ;
• for each (n + 1)-cube ΞX1 ,...,Xn+1 in E, the image-cube F ◦ ΞX1 ,...,Xn+1 is a
limit-cube in E′ , i.e. F (X1 + · · · + Xn+1 ) may be identified with the limit of
the punctured image-cube F ◦ Ξ̌X1 ,...,Xn+1 .
A functor of degree ≤ 1 (resp. ≤ 2) is called linear (resp. quadratic).
A σ-pointed category is called linear (resp. quadratic) if its identity functor is.
If E′ has pullbacks, the limit over the punctured image-cube is denoted
F
=
PX
1 ,...,Xn+1
lim
←−
F ◦ Ξ̌X1 ,...,Xn+1
[0,1]n+1 −(0,...,0)
and the associated comparison map is denoted
F
F
.
: F (X1 + · · · + Xn+1 ) → PX
θX
1 ,...,Xn+1
1 ,...,Xn+1
The (n + 1)-st cross-effects of the functor F : E → E′ are the total kernels of the
F
image-cubes, i.e. the kernels of the comparison maps θX
:
1 ,...,Xn+1
F
F
].
crn+1
(X1 , . . . , Xn+1 ) = K[θX
1 ,...,Xn+1
If F is the identity functor, the symbol F will be dropped from the notation. A
functor F : E → E′ has degree ≤ n if and only if for all (n + 1)-tuples (X1 , . . . , Xn+1 )
F
of objects of E, the comparison maps θX
are invertible.
1 ,...,Xn+1
F
Our total-kernel definition of the cross-effect crn+1
(X1 , . . . , Xn+1 ) is directly inspired by Goodwillie [29, pg. 676] but agrees up to isomorphism with the kernel
intersection definition of Hartl-Loiseau [38] and Hartl-Van der Linden [39]. Their
kernel intersection is dual to the (n + 1)-fold smash product of Carboni-Janelidze
[16], cf. Remark 3.10 and also Remark 6.2, where the duality between cross-effects
and smash-products is discussed in more detail.
Indeed, each of the n + 1 “contraction morphisms”
F
ci + · · · + Xn+1 ), 1 ≤ i ≤ n + 1,
πX̂
: F (X1 + · · · + Xn+1 ) → F (X1 + · · · + X
i
38
CLEMENS BERGER AND DOMINIQUE BOURN
F
factors through θX
so that we get a composite morphism
1 ,...,Xn+1
rF : F (X1 + · · · + Xn+1 )
F
θX
1 ,...,X
−→
n+1
F
PX
1 ,...,Xn+1
n+1
Y
i=1
ci + · · · + Xn+1 )
F (X1 + · · · + X
F
embedding the limit construction PX
into an (n + 1)-fold cartesian product.
1 ,...,Xn+1
F
Therefore, the kernel of θX1 ,...,Xn+1 coincides with the kernel of rF
K[rF ] =
n+1
\
F
K[πX̂
],
i
i=1
which is precisely the kernel intersection of [38, 39] serving as their definition for the
F
(n + 1)st cross-effect crn+1
(X1 , . . . , Xn+1 ) of the functor F .
For n = 1 and F = idE we get the following 2-cube ΞX,Y
X +Y
πŶ
X
πX̂
/Y
/ ⋆E
so that the limit PX,Y of the punctured 2-cube is X × Y and the comparison map
θX,Y : X + Y → X × Y
is the one already used before. In particular, the just introduced notion of linear
category is the usual one.
For n = 2 and F = idE we get the following 3-cube ΞX,Y,Z
π
Ŷ
/: Z
Y
7 +O Z
♥
πX̂ ✈✈ O
♥
♥
✈
♥
✈
✈✈
♥♥♥
πŶ
/
X + YO + Z
X +O Z
π
πX̂
Ẑ
πẐ
X +Y
Y
7
♦
πX̂ ♦♦♦
♦♦
♦
♦
♦
πŶ
/ ⋆E
:
✈
πẐ
✈✈✈
✈✈✈
/X
which induces a split natural transformation of 2-cubes:
ΞX,Y + Z ⇄ ΞX,Y
For sake of simplicity, we denote by +Z the functor E → PtZ (E) which takes an
object X to X + Z → Z with obvious section, and similarly, we denote by ×Z the
functor E → PtZ (E), which takes an object X to X × Z → Z with obvious section.
CENTRAL REFLECTIONS AND NILPOTENCY IN EXACT MAL’TSEV CATEGORIES
39
The previous split natural transformation of 2-cubes induces a natural transformation of split epimorphisms
X +YO +Z
πẐ
X +Y
+Z
θX,Y
'&%$
!"#
2
θX,Y
/ (X + Z) ×Z (Y + Z)
O
/ X ×Y
the comparison map of which may be identified with the comparison map
θX,Y,Z : X + Y + Z → PX,Y,Z
of ΞX,Y,Z . In particular, the category E is quadratic if and only if square (2) is a
pullback square. Notice that in a regular Mal’tsev category, the downward-oriented
square (2) is necessarily a regular pushout by Corollary 1.7.
Proposition 5.2. In a σ-pointed regular Mal’tsev category, the comparison map
θX,Y,Z is a regular epimorphism with kernel relation R[πX̂ ] ∩ R[πŶ ] ∩ R[πẐ ].
Proof. The first assertion expresses the regularity of pushout (2), the second follows
+Z
+Z
from identities R[θX,Y
] ∩ R[πẐ ] which hold
] = R[πX̂ ] ∩ R[πŶ ] and R[θX,Y,Z ] = R[θX,Y
+Z
because both, θX,Y and θX,Y,Z , are comparison maps.
Lemma 5.3. A σ-pointed category with pullbacks is quadratic if and only if square
(2 ′ ) of the following diagram
X +Y +Z
+Z
θX,Y
(X + Z) ×Z (Y + Z)
θX+Y,Z
/.-,
()*+
2′
/ (X + Y ) × Z
/ X +Y
×Z
θX,Y
θX,Z ×Z θY,Z
/ (X × Z) ×Z (Y × Z)
7654
0123
2′′
θX,Y
/ X ×Y
is a pullback square.
Proof. Composing squares (2′ ) and (2′′ ) yields square (2) above. Square (2′′ ) is a
pullback since (X × Z) ×Z (Y × Z) is canonically isomorphic to (X × Y ) × Z.
5.2. The main diagram. We shall now give several criteria for quadraticity. For
this we consider the following diagram
(X + Y ) ⋄ Z /
θX,Y ⋄Z
(X × Y ) ⋄ Z /
ϕZ
X,Y
(X ⋄ Z) × (Y ⋄ Z) /
θX+Y,Z
/ (X + Y ) + Z
θX,Y +Z
/ (X × Y ) + Z
φZ
/ / (X + Y ) × Z
'&%$
!"#
a
θX,Y ×Z
/ / (X × Y ) × Z
θX×Y,Z
X,Y
/ (X + Z) ×Z (Y + Z)
'&%$
!"#
b
µZ
X,Y
/ / (X × Z) ×Z (Y × Z)
θX,Z ×Z θY,Z
+Z
×Z
⋄Z
in which the vertical composite morphisms from left to right are θX,Y
, θX,Y
, θX,Y
, the
horizontal morphisms on the left are the kernel-inclusions of the horizontal regular
epimorphisms on their right, and µZ
X,Y is the canonical isomorphism.
40
CLEMENS BERGER AND DOMINIQUE BOURN
Observe that square (b) is a pullback if and only if the canonical map
φZ
X,Y : (X × Y ) + Z → (X + Z) ×Z (Y + Z)
is invertible. This is the case if and only if for each Z pointed cobase-change
(αZ )! : D → PtZ (D)
along the initial map αZ : ⋆D → Z preserves binary products, cf. Section 3.2.
Proposition 5.4. A σ-pointed regular Mal’tsev category is quadratic if and only if
squares (a) and (b) of the main diagram are pullback squares.
Proof. Since composing squares (a) and (b) yields square (2′ ) of Lemma 5.3, the
condition is sufficient. Lemma 1.17 and Corollary 1.16 imply that the condition is
necessary as well.
Theorem 5.5. A σ-pointed exact Mal’tsev category is quadratic if and only if it is
2-nilpotent and pointed cobase-change along initial maps preserves binary products.
Proof. By Proposition 5.4, the category is quadratic if and only if the squares (a)
and (b) are pullback squares, i.e. the category is 2-nilpotent by Proposition 4.9, and
pointed cobase-change along initial maps preserves binary products.
Corollary 5.6. A semi-abelian category is quadratic1 if and only if either of the
following three equivalent conditions is satisfied:
(a) all objects are 2-nilpotent, and the comparison maps
ϕZ
X,Y : (X × Y ) ⋄ Z → (X ⋄ Z) × (Y ⋄ Z)
are invertible for all objects X, Y, Z;
(b) the third cross-effects of the identity functor cr3 (X, Y, Z) = K[θX,Y,Z ] vanish
for all objects X, Y, Z;
(c) the co-smash product is linear, i.e. the canonical comparison maps
⋄Z
θX,Y
: (X + Y ) ⋄ Z → (X ⋄ Z) × (Y ⋄ Z)
are invertible for all objects X, Y, Z.
Proof. Theorem 5.5 shows that condition (a) amounts to quadraticity.
For condition (b) note that by protomodularity the cross-effect K[θX,Y,Z ] vanishes
if and only if the regular epimorphism θX,Y,Z is invertible.
The equivalence of conditions (b) and (c) follows from the isomorphism of kernels
⋄Z
K[θX,Y,Z ] ∼
]. The latter is a consequence of the 3 × 3-lemma which, applied
= K[θX,Y
to main diagram 5.2 and square (2), yields the chain of isomorphisms
K[θ⋄Z ] ∼
= K[K[θ+Z ] ։ K[θ×Z ]] ∼
= K[θX,Y,Z ].
X,Y
X,Y
X,Y
1The authors of [20, 38, 39] call a semi-abelian category two-nilpotent if each object has a vanishing
ternary Higgins commutator, cf. Remark 6.4. By Proposition 6.5 this means that the identity functor
is quadratic. Corollaries 5.6 and 5.16 describe how to enhance a 2-nilpotent semi-abelian category
in our sense so as to get a two-nilpotent semi-abelian category in their sense.
CENTRAL REFLECTIONS AND NILPOTENCY IN EXACT MAL’TSEV CATEGORIES
41
5.3. Algebraic distributivity and algebraic extensivity. –
We shall see that in a pseudo-additive regular Mal’tsev category (D, ⋆D ), pointed
cobase-change along initial maps αZ : ⋆D → Z preserves binary products if and only
if pointed base-change along terminal maps ωZ : Z → ⋆D preserves binary sums. The
latter condition means that for all objects X, Y, Z, the following square
Z
/ Y ×Z
X ×Z
/ (X + Y ) × Z
is a pushout, inducing thus for all objects X, Y, Z an isomorphism
(X × Z) +Z (Y × Z) ∼
= (X + Y ) × Z
which can be considered as an algebraic distributivity law. This suggests the following definitions, where the added adjective “algebraic” means here that the familiar
definition has to be modified by replacing the slices of the category with the fibres of
the fibration of points, cf. Section 3.2 and Carboni-Lack-Walters [18].
Definition 5.7. A category with pullbacks of split epimorphisms is algebraically distributive if pointed base-change along terminal maps preserves binary sums.
A category with pullbacks of split epimorphisms and pushouts of split monomorphisms is algebraically extensive if any pointed base-change preserves binary sums.
We get the following implications between several in literature studied “algebraic”
notions, where we assume that pullbacks of split epimorphisms and (whenever needed)
pushouts of split monomorphisms exist:
(5.11)
local alg. cartesian closure
+3 alg. extensivity
(5.8)
+3 alg. coherence
(5.9)
(5.9)
qy
+3 alg. distributivity
protomodularity
alg. cartesian
T\ closure
(5.11)
The existence of centralisers implies algebraic cartesian closure [13] and hence
algebraic distributivity, cf. Section 1.5. The categories of groups and of Lie algebras
are not only algebraically cartesian closed, but also locally algebraically cartesian
closed [35, 36], which means that any pointed base-change admits a right adjoint.
Algebraic coherence, cf. Cigoli-Gray-Van der Linden [20], requires any pointed basechange to be coherent, i.e. to preserve strongly epimorphic cospans.
Lemma 5.8. An algebraically extensive regular category is algebraically coherent.
Proof. In a regular category, pointed base-change preserves regular epimorphisms.
Henceforth, if the fibres have binary sums and pointed base-change preserves them,
pointed base-change also preserves strongly epimorphic cospans.
42
CLEMENS BERGER AND DOMINIQUE BOURN
Lemma 5.9 (cf. [10], Theorem 3.10 and [20], Theorem 6.1). An algebraically coherent
pointed Mal’tsev category is protomodular and algebraically distributive.
Proof. To any split epimorphism (r, s) : Y ⇄ X we associate the split epimorphism
(r̄ = r × 1X , s̄ = s × 1X ) : Y × X ⇄ X × X
r̄
Y × Xc● o
●●●●
●●●●(s,1X )
●●●●
●●
p2 ●●●●●
●●
#
/
X; × X
①①①①
①
①
①①①①①
①①①①①p2
①
①{①①①①
s̄
(1X ,1X )
X
in the fibre over X. The kernel of (r̄, s̄) may be identified with the given point
(r, s) over X where the kernel-inclusion is defined by (1Y , r) : Y Y × X. Kernelinclusion and section strongly generate the point Y × X over X, cf. [10, Proposition
3.7]. Pointed base-change along αX : ⋆ → X takes (r̄, s̄) back to (r, s), so that by
algebraic coherence, section and kernel-inclusion of (r, s) strongly generate Y . In a
pointed category this amounts to protomodularity.
For the second assertion observe that if F and G are composable coherent functors
such that G is conservative and GF preserves binary sums, then F preserves binary
sums as well; indeed, the isomorphism GF (X) + GF (Y ) → GF (X + Y ) decomposes
into two isomorphisms GF (X) + GF (Y ) → G(F (X) + F (Y )) → GF (X + Y ). This
∗
applies to F = ωZ
and G = α∗Z (where αZ : ⋆ → Z and ωZ : Z → ⋆) because
∗
∗
preserves binary sums for all Z.
ωZ αZ = idZ and αZ is conservative, so that ωZ
Lemma 5.10. A σ-pointed exact Mal’tsev category is algebraically extensive if and
only if it is a semi-abelian category with exact pointed base-change functors.
Proof. This follows from Lemmas 5.8 and 5.9 and the fact that a left exact and regular
epimorphism preserving functor between semi-abelian categories is right exact if and
only if it preserves binary sums, cf. Section 1.4.
For the following lemma a variety means a category equipped with a forgetful
functor to sets which is monadic and preserves filtered colimits. Every variety is
bicomplete, cowellpowered, and has finite limits commuting with filtered colimits.
Lemma 5.11 (cf. [35], Theorem 2.9). A semi-abelian variety is (locally) algebraically
cartesian closed if and only if it is algebraically distributive (extensive).
Proof. Since the fibres of a semi-abelian category are semi-abelian, the pointed basechange functors preserve binary sums if and only if they preserve finite colimits, cf.
Section 1.4. Since any colimit is a filtered colimit of finite colimits, and pointed basechange functors of a variety preserve filtered colimits, they preserve binary sums if and
only if they preserve all colimits. It follows then from Freyd’s special adjoint functor
theorem that a pointed base-change functor of a semi-abelian variety preserves binary
sums if and only if it has a right adjoint.
A pointed category D with pullbacks is called fibrewise algebraically cartesian
closed (resp. distributive) if for all objects Z of D the fibres PtZ (D) are algebraically
CENTRAL REFLECTIONS AND NILPOTENCY IN EXACT MAL’TSEV CATEGORIES
43
cartesian closed (resp. distributive). This is the case if and only if pointed base-change
along every split epimorphism has a right adjoint (resp. preserves binary sums). Any
algebraically coherent pointed Mal’tsev category is fibrewise algebraically distributive,
cf. the proof of Lemma 5.9.
Proposition 5.12. For a pointed regular Mal’tsev category, fibrewise algebraic cartesian closure is preserved under strong epireflections.
Proof. Let (r, s) : X ⇄ Y be a point in a strongly epireflective subcategory C of a
fibrewise algebraically cartesian closed regular Mal’tsev category D. Let f : Y → Z
be a split epimorphism in C. We shall denote f∗ : PtY (D) → PtZ (D) the right adjoint
of pointed base-change f ∗ : PtZ (D) → PtY (D). Consider the following diagram
φ̄
f¯
¯
/ X̄
X̄
O ❈❈
O ❈o ❈
❈❈ ηX̄¯
❈❈ ηX̄
❈❈
❈❈
r̄
s̄
r̄¯ s̄¯
❈! !
❈! !
¯)
I(ε)
I(
f
¯
o
/
X
I(X̄ )
I(X̄)
③
③
⑧⑧
③
③
⑧
③③ I(r̄¯)
③③
⑧⑧
}③③③ I(r̄)
}③③③
⑧⑧ r
f
/Z
Y
Y o
XO ❂❂o
❂❂❂❂
❂
r
s ❂❂❂❂
❂❂❂
ε
where (r̄, s̄) = f∗ (r, s), the downward-oriented right square is a pullback, and ε :
(r̄¯, s̄¯) = f ∗ f∗ (r, s) → (r, s) is the counit at (r, s). Since D is a regular Mal’tsev
category, the strong epireflection I preserves this pullback of split epimorphisms (cf.
¯ ) is isomorphic to f ∗ (I(X̄)). Since by adjunction, maps
Proposition 2.8) so that I(X̄
¯ ) = f ∗ (I(X̄)) → X there is a
I(X̄) → f∗ (X) = X̄ correspond bijectively to maps I(X̄
∗
unique dotted map φ̄ : I(X̄) → X̄ such that ε ◦ f (φ̄) = I(ε).
Accordingly we get φ̄ηX̄ = 1X̄ so that ηX̄ is invertible and hence X̄ belongs to C.
This shows that the right adjoint f∗ : PtY (D) → PtZ (D) restricts to a right adjoint
f∗ : PtY (C) → PtZ (C) so that C is fibrewise algebraically cartesian closed.
For regular Mal’tsev categories, algebraic cartesian closure amounts to the existence
of centralisers for all (split) subobjects, see Section 1.5. Part of Proposition 5.12 could
thus be reformulated by saying that in this context the existence of centralisers is
preserved under strong epireflections, which can also be proved directly. In a varietal
context, Proposition 5.12 also follows from Lemmas 5.11 and 5.17.
Lemma 5.13. If an algebraically extensive semi-abelian (or homological [4]) category
D has an identity functor of degree ≤ n, then all its fibres PtZ (D) as well.
Proof. The kernel functors (αZ )∗ : PtZ (D) → D are conservative and preserve binary
Pt (D)
sums. Therefore, the kernel functors preserve the limits PX1Z,...,Xn+1 and the comparPt (D)
ison maps θX1Z,...,Xn+1 . Accordingly, if the identity functor of D is of degree ≤ n, then
Pt (D)
α∗Z (θX1 ,...,Xn+1 ) is invertible, hence so is θX1Z,...,Xn+1 , for all objects X1 , . . . , Xn+1 of
the fibre PtZ (D).
44
CLEMENS BERGER AND DOMINIQUE BOURN
It should be noted that in general neither algebraic extensivity nor local algebraic
cartesian closure is preserved under Birkhoff reflection. This is in neat contrast to
(fibrewise) algebraic distributivity and algebraic coherence, which are preserved under
strong epireflections, cf. Proposition 5.12 and [20, Proposition 3.7].
5.4. Duality for pseudo-additive regular Mal’tsev categories.
Lemma 5.14. For any pointed category D with binary sums and binary products
consider the following commutative diagram
(X × Y ) + Z
O
Y
jX
+Z
ρX,Y,Z
pY
X +Z
X +Z
/ X × (Y + Z)
O
X×ιY
Z
Y
X×πZ
/ X ×Z
θX,Z
Y
in which ρX,Y,Z is induced by the pair X × ιZ
Y : X × Y → X × (Y + Z) and αX × ιZ :
Z → X × (Y + Z).
(1) pointed base-change (ωX )∗ : D → PtX (D) preserves binary sums if and only
if the upward-oriented square is a pushout for all objects Y, Z;
(2) pointed cobase-change (αZ )! : PtZ (D) → D preserves binary products if and
only if the downward-oriented square is a pullback for all objects X, Y .
Proof. The left upward-oriented square of the following diagram
X×ιZ
Y
Y
jX
/
/ (X × Y ) + Z
O
ιZ
X×Y
X O× Y
Y
jX
+Z
pY
X
X
ιZ
X
ρX,Y,Z
pY
X +Z
/ X × (Y + Z)
O
X×ιY
Z
/X +Z
θX,Z
/
Z
jX
Y
X×πZ
/ X ×Z
is a pushout so that the whole upward-oriented rectangle is a pushout (i.e. (ωX )∗
preserves binary sums) if and only if the right upward-oriented square is a pushout.
The right downward-oriented square of the following diagram
pX
Y +Z
(X × Y ) + Z
O
Y
jX
+Z
/
+ Z) X
ρX,Y,Z X × (Y
O
pY +Z
pY
X +Z
X +Z
X×ιY
Z
θX,Z
Y
X×πZ
/ X ×Z
X
πZ
/
/ Y +Z
O
ιY
Z
pX
Z
Y
πZ
/Z
/
is a pullback so that the whole downward-oriented rectangle is a pullback (i.e. (αZ )!
preserves binary products) if and only if the left downward-oriented square is a pullback.
CENTRAL REFLECTIONS AND NILPOTENCY IN EXACT MAL’TSEV CATEGORIES
45
Proposition 5.15. In a σ-pointed regular Mal’tsev category D, pointed base-change
(ωZ )∗ : D → PtZ (D) preserves binary sums for all objects Z as soon as pointed cobasechange (αZ )! : D → PtZ (D) preserves binary products for all objects Z. The converse
implication holds if D is pseudo-additive (cf. Definition 4.8).
Proof. According to the previous lemma pointed cobase-change (αX )! preserves binary products if and only if the downward-oriented square is a pullback which implies
that the upward-oriented square is a pushout, and hence pointed base-change (ωZ )∗
preserves binary sums. If D is pseudo-additive, the comparison map θX,Z is an affine
extension. Therefore, the downward-oriented square is a pullback if and only if the
upward-oriented square is a pushout, whence the converse.
Corollary 5.16. A σ-pointed exact Mal’tsev category is quadratic if and only if it is
2-nilpotent and algebraically distributive.
Lemma 5.17. In a σ-pointed regular Mal’tsev category, (fibrewise) algebraic distributivity is preserved under strong epireflections.
Proof. This follows from Proposition 2.8 which shows that strong epireflections preserve besides pushouts and binary sums also binary products (in the fibres).
Theorem 5.18. For any algebraically distributive, σ-pointed exact Mal’tsev category,
the Birkhoff subcategory of 2-nilpotent objects is quadratic.
Proof. The Birkhoff subcategory is pointed, exact, 2-nilpotent, and algebraically distributive by Lemma 5.17, and hence quadratic by Corollary 5.16.
Corollary 5.19 (cf. Cigoli-Gray-Van der Linden [20], Corollary 7.2). For each object
X of an algebraically distributive semi-abelian category, the iterated Huq commutator
[X, [X, X]] coincides with the ternary Higgins commutator [X, X, X].
Proof. Recall (cf. [38, 39]) that [X, X, X] is the direct image of K[θX,X,X ] under the
ternary folding map X + X + X → X. In general, the iterated Huq commutator
[X, [X, X]] is contained in [X, X, X], cf. Corollary 6.12. In a semi-abelian category,
2
the unit of second Birkhoff reflection I 2 takes the form ηX
: X → X/[X, [X, X]],
cf. Remark 2.13. Since in the algebraically distributive case, the subcategory of 2nilpotent objects is quadratic by Theorem 5.18, the image of [X, X, X] in X/[X, [X, X]]
is trivial by Corollaries 5.6b and 6.20, whence [X, [X, X]] = [X, X, X].
Remark 5.20. The category of groups (resp. Lie algebras) has centralisers for subobjects and is thus algebraically distributive. Therefore, the category of 2-nilpotent
groups (resp. Lie algebras) is a quadratic semi-abelian variety.
The reader should observe that although on the level of 2-nilpotent objects there is a
perfect symmetry between the property that pointed base-change along terminal maps
preserves binary sums and the property that pointed cobase-change along initial maps
preserves binary products (cf. Proposition 5.15), only the algebraic distributivity
carries over to the category of all groups (resp. Lie algebras) while the algebraic
“codistributivity” fails in either of these categories. “Codistributivity” is a quite
restrictive property, which is rarely satisfied without assuming 2-nilpotency.
46
CLEMENS BERGER AND DOMINIQUE BOURN
6. Identity functors with bounded degree
In the previous section we have seen that quadraticity is a slightly stronger property
than 2-nilpotency, insofar as it also requires a certain compatibility between binary
sum and binary product (cf. Theorem 5.5 and Proposition 5.15). In this last section,
we relate n-nilpotency to identity functors of degree ≤ n.
6.1. Degree n functors and n-folded objects. –
Any (n + 1)-cube ΞX1 ,...,Xn+1 (cf. Definition 5.1) defines a split natural transformation of n-cubes inducing a natural transformation of split epimorphisms
X1 + · · · + Xn+1
O
πX̂
n+1
X1 + · · · + Xn
+Xn+1
1 ,...,Xn
θX
/ P +Xn+1
X1 ,...,Xn
O
+ωX
n+1
1 ,...,Xn
'&%$
!"#
n
PX
θX1 ,...,Xn
+αX
n+1
1 ,...,Xn
PX
/ PX1 ,...,Xn
the comparison map of which may be identified with the comparison map
θX1 ,...,Xn+1 : X1 + · · · + Xn+1 → PX1 ,...,Xn+1
of the given (n + 1)-cube. In particular, our category has an identity functor of degree
≤ n if and only if square (n) is a pullback square for all objects X1 , . . . , Xn+1 .
Proposition 6.1. In a σ-pointed regular Mal’tsev category, the comparison map
θX1 ,...,Xn+1 is a regular epimorphism with kernel relation R[πX̂1 ] ∩ · · · ∩ R[πX̂n+1 ].
Proof. This follows by induction on n like in the proof of Proposition 5.2.
Remark 6.2. The previous proposition shows that in a σ-pointed regular Mal’tsev
category, the intersection of the kernel relations of the contraction maps may be considered as the “total kernel relation” of the cube. This parallels the more elementary
fact that the total-kernel definition of the cross-effects crn+1 (X, . . . , Xn+1 ) coincides
with the kernel-intersection definition of Hartl-Loiseau [38] and Hartl-Van der Linden
[39]. In particular, in any σ-pointed regular Mal’tsev category, the image of the morQn+1
ci + · · · + Xn+1 coincides with the
phism rid : X1 + · · · + Xn+1 → i=1 X1 + · · · + X
limit PX1 ,...,Xn+1 of the punctured (n + 1)-cube.
We already mentioned that these kernel intersections are dual to the (n + 1)fold smash products of Carboni-Janelidze [16]. An alternative way to describe the
duality between cross-effects and smash-products is to consider the limit construction
PX1 ,...,Xn as the dual of the so-called fat wedge T X1 ,...,Xn , cf. Hovey [42]. Settheoretically, the fat wedge is the subobject of the product X1 × · · · × Xn formed by
the n-tuples having at least one coordinate at a base-point. If base-point inclusions
behave “well” with respect to cartesian product, the fat wedge is given by a colimit
construction, strictly dual to the limit construction defining PX1 ,...,Xn . The n-fold
smash-product X1 ∧ · · · ∧ Xn is then the cokernel of the monomorphism T X1 ,...,Xn
X1 × · · · × Xn while the n-th cross-effect crn (X1 , . . . , Xn ) is the kernel of the regular
epimorphism X1 + · · · + Xn ։ PX1 ,...,Xn .
CENTRAL REFLECTIONS AND NILPOTENCY IN EXACT MAL’TSEV CATEGORIES
47
The cubical cross-effects are just the algebraic version of Goodwillie’s homotopical cross-effects [29, pg. 676]. Nevertheless, for functors taking values in abelian
categories, the cubical cross-effects agree with the original cross-effects of EilenbergMac Lane [23, pg. 77]. Indeed, by [23, Theorems 9.1 and 9.6], for a based functor
F : D → E from a σ-pointed category (D, +, ⋆D ) to an abelian category (E, ⊕, 0E ), the
latter are completely determined by the following decomposition formula
M
M
F (X1 + · · · + Xn ) ∼
crkF (Xi1 , . . . , Xik )
=
1≤k≤n 1≤i1 <···<ik ≤n
for any objects X1 , . . . , Xn in D. It suffices thus to show that the cubical cross-effects
satisfy the decomposition formula if values are taken in an abelian category.
F
F
For n = 2 we get PX
= F (X1 ) ⊕ F (X2 ) from which it follows that θX
:
1 ,X2
1 ,X2
F (X1 +X2 ) ։ F (X1 )⊕F (X2 ) is a split epimorphism. Henceforth, we get the asserted
isomorphism F (X1 + X2 ) ∼
= F (X1 ) ⊕ F (X2 ) ⊕ cr2F (X1 , X2 ).
The 3-cube F (ΞX1 ,X2 ,X3 ) induces a natural transformation of split epimorphisms
F (X1 + X2 + X3 )
O
/ P3
O
F (X1 + X2 )
/ / F (X1 ) ⊕ F (X2 )
F
θX
1 ,X2
in which P3 is isomorphic to F (X1 ) ⊕ F (X2 ) ⊕ F (X3 ) ⊕ cr2F (X1 , X3 ) ⊕ cr2F (X2 , X3 ).
F
From this, we get for PX
= F (X1 + X2 ) ×F (X1 )⊕F (X2 ) P3 the formula
1 ,X2 ,X3
F
∼
PX
= F (X1 ) ⊕ F (X2 ) ⊕ F (X3 ) ⊕ cr2F (X1 , X2 ) ⊕ cr2F (X1 , X3 ) ⊕ cr2F (X2 , X3 )
1 ,X2 ,X3
F
so that θX
is again a split epimorphism inducing the asserted decomposition
1 ,X2 ,X3
of F (X1 + X2 + X3 ). The same scheme keeps on for all positive integers n.
Definition 6.3. An object X of a σ-pointed regular Mal’tsev category will be called
X
n-folded if the folding map δn+1
factors through the comparison map θX,...,X
θX,...,X
X + · · · ❉+ X
❉❉
❉❉
❉❉
X
δn+1 ❉❉!
!
/ / PX,...,X
mX
X
X
i.e. if the folding map δn+1
annihilates the kernel relation R[θX,...,X ].
An object X is 1-folded if and only if the identity of X commutes with itself. In
a σ-pointed regular Mal’tsev category this is the case if and only if X is an abelian
group object, cf. the proof of Proposition 4.2.
Remark 6.4. In a semi-abelian (or homological [4]) category, an object X is n-folded
X
if and only if the image of the kernel K[θX,...,X ] under the folding map δn+1
:X+
· · · + X ։ X is trivial. Recall [38, 39, 41, 54] that this image is by definition the
so-called Higgins commutator [X, . . . , X] of length n + 1. Therefore, an object of a
48
CLEMENS BERGER AND DOMINIQUE BOURN
semi-abelian category is n-folded precisely when its Higgins commutator of length
n + 1 vanishes. Under this form n-folded objects have already been studied by Hartl
emphasising their role in his theory of polynomial approximation.
In a varietal context, n-foldedness can be expressed in more combinatorial terms.
For instance, a group X is n-folded if and only if (n + 1)-reducible elements of X are
trivial. An element w ∈ X is called (n + 1)-reducible if there is an element v in the
free group F (X ⊔ · · · ⊔ X) on n + 1 copies of X (viewed as a set) such that
(a) w is the image of v under the composite map
n+1
n+1
n+1
X
}|
{
z
z
z
}|
{
}|
{ δn+1
F (X ⊔ · · · ⊔ X) ∼
= F (X) + · · · + F (X) ։ X + · · · + X ։ X
F (X)
: F (X)+(n+1) ։ F (X)+n , cf. Section
(b) for each of the n + 1 contractions πi
F (X)
(v) maps to the neutral element of X under
6.2, the image πi
n
n
z
}|
{
}|
{ δnX
z
F (X) + · · · + F (X) ։ X + · · · + X ։ X.
Indeed, since the evaluation map F (X) ։ X is a regular epimorphism, and the evaluation maps in (a) and (b) are compatible with the contraction maps πi : X +(n+1) ։
X +n , Proposition 6.1, Section 6.2 and Corollary 6.20 imply that we get in this way
X
the image of the kernel K[θX,...,X ] under the folding map δn+1
.
Qk
Qk
−1 −1
Any product of commutators i=1 [xi , yi ] = i=1 xi yi xi yi in X is 2-reducible
by letting the xi (resp. yi ) belong to the first (resp. second) copy of X. Conversely,
a direct computation shows that any 2-reducible element of X can be rewritten as a
product of commutators of X. This recovers in a combinatorial way the aforementioned fact that X is abelian (i.e. 1-nilpotent) if and only if X is 1-folded.
The relationship between n-nilpotency and n-foldedness is more subtle, closely
related to the cross-effects of the identity functor (cf. Theorem 6.8). For groups and
Lie algebras the two concepts coincide (cf. Theorem 6.23c) but, though any n-folded
object is n-nilpotent, the converse is wrong in general (cf. Section 6.5).
Proposition 6.5. Let F : D → E be a based functor between σ-pointed categories and
assume that E is a regular Mal’tsev category.
(a) If F is of degree ≤ n then F takes values in n-folded objects of E;
(b) If F preserves binary sums and takes values in n-folded objects of E then F
is of degree ≤ n;
(c) The identity functor of E is of degree ≤ n if and only if all objects of E are
n-folded.
F (X)
Proof. Clearly, (c) follows from (a) and (b). For (a) note that δn+1 factors through
X
F (δn+1
), and that by definition of a functor of degree ≤ n, the comparison map
F
X
θX,...,X is invertible so that F (δn+1
) gets identified with mF (X) .
For (b) observe first that preservation of binary sums yields the isomorphisms
F
F
∼
∼
PX
= θF (X1 ),...,F (Xn+1 ) . We shall show
= PF (X1 ),...,F (Xn+1 ) and θX
1 ,...,Xn+1
1 ,...,Xn+1
F
that if moreover F takes values in n-folded objects of E then θX
is invertible
1 ,...,Xn+1
for all (n + 1)-tuples (X1 , . . . , Xn+1 ) of objects of D.
CENTRAL REFLECTIONS AND NILPOTENCY IN EXACT MAL’TSEV CATEGORIES
49
T
Consider any family (fi : Xi → T )1≤i≤n+1 of morphisms in E, and let φ = δn+1
◦
(f1 + · · · + fn+1 ) : X1 + · · · + Xn+1 → T be the induced map. We have the following
F
factorisation of F (φ) through θX
:
1 ,...,Xn+1
θF (X1 ),...,F (Xn+1 )
F (X1 ) + · · · + F (Xn+1 )
/ / PF (X ),...,F (X )
1
n+1
PF (f1 ),...,F (fn+1 )
F (f1 )+···+F (fn+1 )
θF (T ),...,F (T )
/ / PF (T ),...,F (T )
F (T ) + · · · + F (T )
◆◆◆
t
◆◆◆
tt
◆◆◆
tt
t
T
◆◆◆
tt mF (T )
δn+1
ztt
&
F (T )
In particular, if T = X1 + · · · + Xn+1 and fi is the inclusion of the ith summand,
we get a retraction of θF (X1 ),...,F (Xn+1 ) which accordingly is a monomorphism. Since
θF (X1 ),...,F (Xn+1 ) is also a regular epimorphism, it is invertible.
Proposition 6.6. The full subcategory Fldn (E) of n-folded objects of a σ-pointed
regular Mal’tsev category E is closed under products, subobjects and quotients.
Proof. For any two n-folded objects X and Y the following diagram
(X × Y ) + · · · + (X × Y )
θX×Y,...,X×Y
/ / PX×Y,...,X×Y
θX,...,X ×θY,...,Y
/ / PX,...,X × PY,...,Y
(X + · · · + X) × (Y + · · · + Y )
❘❘❘
r
❘❘❘
rr
❘❘❘
rr
r
❘
r
❘
X
Y
❘❘❘
r m ×m
δn+1
×δn+1
(
yrr X Y
X ×Y
X×Y
induces the required factorisation of δn+1
through θX×Y,...,X×Y .
For a subobject n : U X of an n-folded object X consider the diagram
U + · · · + U ❩❩❩❩❩
❩❩❩❩❩θ❩U,···
❩❩❩,U❩❩❩❩❩
θ
❩❩❩❩❩❩❩- '
ν
/ PU,...,U
W /
U
δn+1
U /
Pn,...,n
n+···+n
n
(
θX,...,X
/ / PX,...,X
X + ···+ X
X
✐✐✐✐
δn+1
t✐✐✐✐✐✐✐mX
/X
in which the dotted quadrangle is a pullback. The commutatitvity of the diagram
induces a morphism θ such that νθ = θU,...,U . Since θU,...,U is a regular epimorphism,
the monomorphism ν is invertible, whence the desired factorisation.
50
CLEMENS BERGER AND DOMINIQUE BOURN
Finally, for a regular epimorphism f : X ։ Y with n-folded domain X consider
the following diagram
f +···+f
X + · · · +❖ X
❖❖❖θX,...,X
❖❖❖
❖❖❖
''
X
δn+1
PX,...,X
♦
♦
♦
♦♦♦
w♦♦♦♦♦ mX
X
f
/ / Y + ··· + Y
❖❖❖
❖❖θ❖Y,...,Y
❖❖❖
Y
δn+1
❖' '
Pf,...,f
/ / PY,...,Y
//Y w
mY
in which the existence of the dotted arrow has to be shown. According to Lemma 6.7
the induced morphism on kernel relations
R(f + · · · + f, Pf,...,f ) : R[θX,...,X ] → R[θY,...,Y ]
Y
is a regular epimorphism. A diagram chase shows then that δn+1
annihilates R[θY,...,Y ]
Y
whence the required factorisation of δn+1 through θY,...,Y .
Lemma 6.7. In a σ-pointed regular Mal’tsev category, any finite family of regular
epimorphisms fi : Xi ։ Yi (i = 1, . . . , n) induces a regular epimorphism on kernel
relations R(f1 + · · · + fn , Pf1 ,...,fn ) : R[θX1 ,...,Xn ] ։ R[θY1 ,...,Yn ].
Proof. Since regular epimorphisms compose (in any regular category) it suffices to
establish the assertion under the assumption fi = 1Xi for i = 2, . . . , n. Moreover we
can argue by induction on n since for n = 1 the comparison map is the terminal map
θX : X ։ ⋆ and a binary product of regular epimorphisms is a regular epimorphism.
Assume now that the statement is proved for n−1 morphisms. Using the isomorphism
of kernel relations
∼ R[R[θ+Xn
R[θX ,...,X ,X ] =
] ։ R[θX ,...,X ]]
1
n−1
n
X1 ,...,Xn−1
1
n−1
and Propositon 1.6 it suffices then to show that the following by f1 : X1 ։ Y1 induced
commutative square
+Xn
R[θX
]
1 ,X2 ,...,Xn−1
O
/ / R[θ+Xn
Y1 ,X2 ,...,Xn−1 ]
O
R[θX1 ,X2 ,...,Xn−1 ]
/ / R[θY1 ,X2 ,...,Xn−1 ]
is a downward-oriented regular pushout. This in turn follows from Corollary 1.7 since
the vertical arrows above are split epimorphisms by construction and the horizontal
arrows are regular epimorphisms by induction hypothesis.
Special instances of the following theorem have been considered in literature: if E
is the category of groups and n = 2, the result can be deduced from Baues-Pirashvili
[1] and Hartl-Vespa [40]; if E is a semi-abelian category, under the identification of
n-folded objects given in Remark 6.4 and with the kernel intersection definition of
degree, the result has been announced by Manfred Hartl in several of his talks.
CENTRAL REFLECTIONS AND NILPOTENCY IN EXACT MAL’TSEV CATEGORIES
51
Theorem 6.8. The full subcategory Fldn (E) of n-folded objects of a σ-pointed exact Mal’tsev category E is a reflective Birkhoff subcategory. The associated Birkhoff
reflection J n : E → Fldn (E) is the universal endofunctor of E of degree ≤ n.
Proof. Observe that the second assertion is a consequence of the first and of Proposition 6.5a-b. In virtue of Proposition 6.6 it suffices thus to construct the reflection
J n : E → Fldn (E). The latter is obtained by the following pushout
θX,··· ,X
X + ···+ X
X
δn+1
/ / PX,...,X
µn X
X
ǫn X
/ / J n (X)
which is regular by Corollary 1.8 so that J n (X) = X/Hn+1 [X] where Hn+1 [X] is the
X
direct image of R[θX,...,X ] under the folding map δn+1
. We will show that J n (X) is
n-folded and that any morphism X → T with n-folded codomain T factors uniquely
through J n (X). For this, consider the following diagram
θH
[X],...,H
[X]
n+1
n+1
/ / PH [X],...,H [X]
Hn+1 [X] + · · · + H❚n+1
n+1
n+1
❚❚❚[X]
❚❚❚❚ ❚❚p❚1❚+···+p
Pp1 ,··· ,p1
Pp0 ,··· ,p0
❚❚❚❚ ❚❚) 1
θX,··· ,X
p0 +···+p0 ❚❚)
/ / PX,...,X
X + ···+ X
Hn+1 [X]
❯❯❯P❯ǫn X,...,ǫn X
δn+1
❯❯❯* *
X
P n
δn+1
p1
Hn+1 [X]
p0
µn X
/
/X
ǫn X
J (X),...,J n (X)
/ / J n (X) s
in which the existence of the dotted arrow has to be shown. By Lemma 6.9 PJ n (X),...,J n (X)
is the coequaliser of the reflexive pair (Pp0 ,...,p0 , Pp1 ,...,p1 ). It suffices thus to check
that µn X coequalises the same pair. This follows by precomposition with the regular
epimorphism θHn+1 [X],...,Hn+1 [X] using the commutativity of the previous diagram.
For the universal property of ǫn X : X ։ J n (X) let us consider a morphism
f : X → T with n-folded codomain T . By construction of J n (X), the following
commutative diagram
θX,...,X
/ / PX,...,X
X + · · · + ❙X
❙❙f ❙+···+f
◆◆P◆f,...,f
❙❙❙
◆◆&
)
θ
T ,...,T
X
/ / PT,...,T
δn+1
T + · · · +❘T δT
❘❘❘n+1
♦
♦
❘
♦
❘
❘❘( w♦♦♦ mT
/T
X
f
induces the desired factorisation.
Lemma 6.9. In a σ-pointed exact Mal’tsev category, the functor PX1 ,...,Xn+1 preserves reflexive coequalisers in each of its n + 1 variables.
52
CLEMENS BERGER AND DOMINIQUE BOURN
Proof. By exactness, it suffices to show that P preserves regular epimorphisms in each
variable, and that for a regular epimorphism fi : Xi ։ Xi′ the induced map on kernel
relations PX1 ,...,R[fi ],...,Xn+1 → R[PX1 ,...,fi ,...,Xn+1 ] is a regular epimorphism as well.
By symmetry, it is even sufficient to do so for the first variable.
We shall argue by induction on n (since for n = 1 there is nothing to prove) and
consider the following downward-oriented pullback diagram
PX1 ,...,Xn ,Xn+1
O
/ / P +Xn+1
X1 ,...,Xn
O
X1 + · · · + Xn
/ / PX1 ,...,Xn
θX1 ,...,Xn
which derives from square (n) of the beginning of this section. By induction hypothesis, the two lower corners and the upper right corner are functors preserving regular
epimorphisms in the first variable. It follows then from the cogluing lemma (cf. the
proof of Theorem 6.23a) and Corollary 1.7 that the upper left corner also preserves
regular epimorphisms in the first variable.
It remains to be shown that for f : X1 ։ X1′ we get an induced regular epimorphism on kernel relations. For this we denote by F, G, H the functors induced on the
lower left, lower right and upper right corners, and consider the following commutative
diagram
/ H(R[f ])
P (R[f ])
◆◆ρ◆H
tR[f ]
ρP
@
✁✁@
◆◆◆'
✁
✁
'
✁
'
'
✁✁✁
✁
/ R[H(f )]
✁✁✁
R[P (f )]
✁
@
✁@
✁✁✁✁
✁✁✁✁✁
✁✁✁✁✁
✁
R(t
)
X
✁✁
/ G(R[f ])
F (R[f ])
✁✁✁✁✁
θR[f ]
◆
◆◆◆
✁
◆
◆
✁✁
ρG ◆◆◆' '
ρF ◆◆◆' '
✁✁✁✁✁
/ R[G(f )]
R[F (f )]
R(θX )
in which the back vertical square is a downward-oriented pullback by definition of P .
By commutation of limits the front vertical square is a downward oriented pullback
as well. Again, according to the cogluing lemma and Corollary 1.7, the induced
arrow ρP is then a regular epimorphisms, since ρF , ρG and ρH are so by induction
hypothesis.
6.2. Higgins commutator relations and their normalisation. –
We shall now concentrate on the case X = X1 = X2 = · · · = Xn+1 . Accordingly,
we abbreviate the n + 1 “contractions” as follows:
n+1
πX̂i
n
z
}|
{
}|
{
z
= πi : X + · · · + X → X + · · · + X,
i = 1, . . . , n + 1.
Proposition 6.1 reads then R[θX,...,X ] = R[π1 ] ∩ · · · ∩ R[πn+1 ].
CENTRAL REFLECTIONS AND NILPOTENCY IN EXACT MAL’TSEV CATEGORIES
53
We denote the direct image of the kernel relation R[θX,...,X ] under the folding map
X
δn+1
: X + · · · + X → X by a single bracket [∇X , . . . , ∇X ] of length n + 1 and call it
the (n + 1)-ary Higgins commutator relation on X.
The proof of Theorem 6.8 shows that in σ-pointed exact Mal’tsev categories the
universal n-folded quotient J n (X) of an object X is obtained by quotienting out the
(n + 1)-ary Higgins commutator relation. The binary Higgins commutator relation
coincides with the Smith commutator [∇X , ∇X ] (cf. Section 1.1, Corollary 1.8 and
Proposition 4.2) which ensures consistency of our notation.
Recall that in a pointed category the normalisation of an effective equivalence
relation R on X is the kernel of its quotient map X ։ X/R. In σ-pointed exact
Mal’tsev categories normalisation commutes with direct image, cf. Corollary 1.8. In
particular, the normalisation of the Higgins commutator relation yields precisely the
Higgins commutator of same length, cf. Remark 6.4.
Proposition 6.10. In a σ-pointed exact Mal’tsev category, the image of R[θX,X,X ]
under δ2X + 1X is the kernel relation of the pushout of θX,X,X along δ2X + 1X
X +X +X
θX,X,X
δ2X +1X
X +X
X
ζX,X
/ / PX,X,X
/ / JX
X,X
X
which may be computed as an intersection: R[ζX,X
] = [R[π1 ], R[π1 ]] ∩ R[π2 ].
In particular, we get the inclusion [∇X , [∇X , ∇X ]] ⊂ [∇X , ∇X , ∇X ].
Proof. By Corollary 1.8, the pushout is regular so that the first assertion follows from
Proposition 1.6. Consider the following diagram
δ2X +1X
X + X + X❱
❱❱❱❱ θ+X
⑤>
❱❱❱X,X
⑤⑤⑤⑤⑤
❱❱* *
⑤
π3 ⑤⑤⑤⑤
(X
+ X) ×X (X + X)
⑤
⑤
⑤⑤⑤⑤⑤
r8
⑤
rrrrrrr δX
~⑤⑤⑤⑤⑤
r
r
2
r rr
X +X
rrrrrrr
r
r
PPP
r
rr
P
xrrrrrrr
θX,X PPP( (
X ×X
/ / X +X
✠✠D ▼▼▼η▼1 (π1 )
✠
▼& &
✠
✠✠✠✠
✠
/
/ I 1 (π1 )
✠
✠✠
π2 ✠✠✠
✠
✄✄✄A
✠✠✠✠
′
✄✄✄✄✄
f
✄
//X
✄✄
❍❍ η1 (X) ✄✄✄✄✄✄
❍❍
✄
✄
✄✄✄
❍# #
/ / I 1 (X)
in which top and bottom are regular pushouts by Corollary 1.8. The bottom square
constructs the associated abelian object I 1 (X) of X, while the top square constructs
the associated abelian object I 1 (π1 ) of π1 : X + X ⇄ X in the fibre over X. The
upward oriented back and front faces are pushouts of split monomorphisms. The left
face is a specialisation of square (2) just before Proposition 5.2. We can therefore
54
CLEMENS BERGER AND DOMINIQUE BOURN
apply Corollary 1.19 and we get diagram
δ X +1
X
2
/ / X +X
X + X Z✺+ ❘X
✲✲V✲✲ ❑❑❑ X
✺✺✺ ❘❘❘❘
✺
❘❘θ❘X,X,X
✺✺✺
✲✲✲✲ ❑❑❑ζ❑X,X
❘
✺
❘❘❘
❑❑❑
✺✺✺
✲✲✲✲
❘❘❘
❑% %
✺✺✺✺
❘) )
✲✲✲✲
✺
✺✺✺
✲✲
/ / JX
✲
P
✺
X,X,X
X,X
✲✲ ✲
π3 ✺✺✺
✺
✺✺✺
π2 ✲ ✲✲
⑤⑤=
✞✞C
⑤
✲
✺
✞
✺✺✺
✲✲✲✲
⑤⑤⑤⑤
✞✞✞✞
✺✺✺✺
⑤⑤⑤⑤⑤
✲ ✲ ✞✞✞✞✞
⑤
}⑤ ⑤
/ / X ✞✞
X + X⑤
X
δ2
X
in which the kernel relation of the regular epimorphism ζX,X
is given by
R[η 1 (π1 )] ∩ R[π2 ] = [R[π1 ], R[π1 ]] ∩ R[π2 ].
For the second assertion, observe first that the ternary folding map δ3 may be
identified with the composition δ2X ◦ (δ2X + 1X ). Therefore, the ternary Higgins commutator relation [∇X , ∇X , ∇X ] is the direct image under δ2X : X + X → X of the
X
kernel relation of ζX,X
. Now we have the following chain of inclusions, where for
shortness we write R1 = R[π1 ], R2 = R[π2 ], R12 = R1 ∩ R2 :
[R12 , [R12 , R12 ]] ⊂ R12 ∩ [R12 , R12 ] ⊂ R2 ∩ [R1 , R1 ].
By exactness, the direct image of the leftmost relation is [∇X , [∇X , ∇X ]], while
the direct image of the right most relation is [∇X , ∇X , ∇X ].
Proposition 6.11. In a σ-pointed exact Mal’tsev category, the image of R[θX,...,X ]
X
X
+ 1X
+ 1X is the kernel relation of the pushout of θX,...,X along δn−1
under δn−1
X + ···+ X
θX,...,X
X
δn−1
+1X
X +X
X
ζX,...,X
/ / PX,...,X
/ / JX
X,...,X
n−1
which may be computed as an intersection
X
R[ζX,...,X
]
n−1
z
}|
{
= [R[π1 ], . . . , R[π1 ]] ∩ R[π2 ].
n
z
}|
{
}|
{
z
In particular, we get the inclusion [∇X , [∇X , . . . , ∇X ]] ⊂ [∇X , . . . , ∇X ].
CENTRAL REFLECTIONS AND NILPOTENCY IN EXACT MAL’TSEV CATEGORIES
55
Proof. The first assertion follows from Proposition 1.6 and the following diagram
X
δn−1
+1X
X + · · · +PX +X
PPθPX,...,X
①①①;
PP( (
①①①①
πn ①①①①
+X
①
①
①①
PX,...,X
①①①①①
①
=
①①
④④④X
{①①①①①
④④④④④δn−1
④
X + ···+ X
④④④④
❚❚❚❚
④④④④④
④
❚
}④④④
θX,...,X ❚❚❚) )
PX,...,X
// X +X
❚❚❚❚ q
❚❚❚❚)
⑦⑦⑦⑦>
⑦
)
⑦⑦⑦⑦⑦
X+X
⑦
//
⑦
⑦
[∇π1 ,...,∇π1 ]
⑦⑦⑦
π2 ⑦⑦⑦
✉
✉✉ ✉:
~⑦⑦⑦⑦
✉✉✉✉✉✉
//X ⑦
✉
✉✉ ✉
◗◗◗
✉✉✉✉✉✉
◗◗◗
✉
z
✉
◗( (
✉✉
/ / X/[∇
X , . . . , ∇X ]
in which top and bottom are regular pushouts by Corollary 1.8. The bottom square
constructs the quotient of X by the (n−1)-ary Higgins commutator relation [∇X , . . . , ∇X ].
The top square constructs the quotient of π1 : X + X ⇄ X by the (n − 1)-ary Higgins
commutator relation [∇π1 , . . . , ∇π1 ] in the fibre over X. The upward oriented back
and front faces are pushouts of split monomorphisms. The left face is a specialisation
of square (n) of the beginning of this section. We can therefore apply Corollary 1.19
and we get the following diagram
X
δn−1
+1X
X + · · ·\✿+ X
✿✿✿✿ ❯❯❯❯❯❯
❯❯❯θ❯X,...,X
✿✿✿✿
❯❯❯❯
✿✿✿✿
❯❯❯❯
✿✿✿✿
❯❯* *
✿✿✿✿
PX,...,X
✿
πn ✿✿✿
✿✿✿✿
✈
✿✿✿✿
✈✈ ✈:
✈✈✈✈✈✈
✿✿✿✿
✈
✈✈ ✈
✿
z✈✈✈✈✈✈
X + ···+ X
X
δn−1
/ / X +X
✲✲V✲✲ ▲▲▲ X
✲ ▲▲▲ζX,...,X
π2 ✲✲ ✲
▲▲▲
✲✲✲✲
▲▲& &
✲
✲✲ ✲
✲
/ / JX
✲✲ ✲
X,...,X
✲✲✲✲
✄A
✲✲✲✲
✄
✄✄✄✄
✄
✲✲✲✲
✄✄✄✄✄
/ / X ✄✄
X
in which the kernel relation of the regular epimorphism ζX,...,X
is given by
R[q] ∩ R[π2 ] = [∇π1 , . . . , ∇π1 ] ∩ R[π2 ] = [R[π1 ], . . . , R[π1 ]] ∩ R[π2 ].
X
Since δnX = δ2X ◦ (δn−1
+ 1X ), the proof of the second assertion is completely
analogous to the proof of the corresponding part of Proposition 6.10.
The semi-abelian part of the following corollary can also be derived from a direct
analysis of “iterated” Higgins commutators, cf. [39, Proposition 2.21(iv)].
Corollary 6.12. For any object X of a σ-pointed exact Mal’tsev category, the iterated
Smith commutator [∇X , [∇X , [∇X , · · · , [∇X , ∇X ] · · · ]]] is a subobject of the Higgins
commutator relation [∇X , . . . , ∇X ] of same length.
In a semi-abelian category, the iterated Huq commutator [X, [X, · · · , [X, X] · · · ]] is
a subobject of the Higgins commutator [X, . . . , X] of same length.
Proof. The first statement follows inductively from Propositions 6.10 and 6.11.
The second statement follows from the first and the fact that in a semi-abelian
category, the iterated Huq commutator is the normalisation of the iterated Smith
commutator by Remark 2.13 and [33, Proposition 2.2].
56
CLEMENS BERGER AND DOMINIQUE BOURN
Proposition 6.13. In a σ-pointed exact Mal’tsev category E, each n-folded object is
an n-nilpotent object, i.e. Fldn (E) ⊂ Niln (E). In particular, endofunctors of degree
≤ n take values in n-nilpotent objects.
Proof. The second assertion follows from the first and from Proposition 6.5. For any
n-folded object X, the (n + 1)-ary Higgins commutator relation of X is discrete and
hence, by Corollary 6.12, the iterated Smith commutator of same length is discrete as
well. By an iterated application of Theorem 2.12 and Proposition 1.4, this iterated
n
Smith commutator is the kernel relation of ηX
: X ։ I n (X), and hence X ∼
= I n (X),
i.e. X is n-nilpotent.
The following theorem generalises Theorem 4.5 to all positive integers.
Theorem 6.14. For a σ-pointed exact Mal’tsev category D such that the identity
functor of Niln−1 (D) is of degree ≤ n − 1, the following properties are equivalent:
(a) all objects are n-nilpotent;
(b) for all objects X1 , . . . , Xn , the map θX1 ,...,Xn is an affine extension;
(c) for all objects X1 , . . . , Xn , the map θX1 ,...,Xn is a central extension.
Proof. For an n-nilpotent category D, the Birkhoff reflection I n−1 : D → Niln−1 (D)
is a central reflection. Since all limits involved in the construction of PX1 ,...,Xn
are preserved under I n−1 by an iterative application of Proposition 2.8, we get
I n−1 (θX1 ,...,Xn ) = θI n−1 (X1 ),...,I n−1 (Xn ) . Since by assumption the identity functor
of Niln−1 (D) is of degree ≤ n − 1, the latter comparison map is invertible so that by
Theorem 3.3, the comparison map θX1 ,...,Xn is an affine extension, i.e. (a) implies (b).
By Proposition 3.2, (b) implies (c).
Specializing (c) to the case X = X1 = X2 = · · · = Xn we get the following
commutative diagram
R[θX,...,X ]
[∇X , . . . , ∇X ]
/
/ X + ··· + X
X
δn
/
/X
θX,...,X
/ / PX,...,X
/ / X/[∇X , . . . , ∇X ]
in which is the right square is a regular pushout by Corollary 1.8 so that the lower
row represents the kernel relation of a central extension. We have already seen that
the iterated Smith commutator [∇X , [∇X , [∇X , · · · , [∇X , ∇X ] · · · ]]] of length n is the
kernel relation of the unit η n−1 (X) : X ։ I n−1 (X) of the (n−1)st Birkhoff reflection.
Corollary 6.12 implies thus that this unit is a central extension as well so that D is
n-nilpotent, i.e. (c) implies (a).
Definition 6.15. A σ-pointed category (D, ⋆D ) with pullbacks is said to satisfy condition Pn if for all X1 , . . . , Xn , Z, pointed cobase-change (αZ )! : D → PtZ (D) takes
+Z
.
the object PX1 ,...,Xn to the object PX
1 ,...,Xn
In particular, since PX = ⋆ condition P1 is void and just expresses that (αZ )!
preserves the null-object. Since PX,Y = X × Y condition P2 expresses that (αZ )!
CENTRAL REFLECTIONS AND NILPOTENCY IN EXACT MAL’TSEV CATEGORIES
57
preserves binary products. Therefore, the following result extends Corollary 4.4 (n =
1) and Theorem 5.5 (n = 2) to all positive integers.
Proposition 6.16. The identity functor of a σ-pointed exact Mal’tsev category D is
of degree ≤ n if and only if all objects are n-nilpotent and the Birkhoff subcategories
Nilk (D) satisfy condition Pk for 1 ≤ k ≤ n.
Proof. Since the statement is true for n = 1 by Corollary 4.4 we can argue by induction
on n and assume that the statement is true up to level n − 1. In particular, we can
assume that Niln−1 (D) has an identity functor of degree ≤ n − 1. Let us then consider
the following substitute of the main diagram 5.2:
(X1 + · · · + Xn ) ⋄ Z /
θX1 ,...,Xn ⋄Z
PX1 ,...,Xn ⋄ Z /
ϕZ
X1 ,...,Xn
⋄Z
/
PX
,...,X
1
n
θ
/ (X1 + · · · + Xn ) + ZX1 +···+Xn/ / ,Z
(X1 + · · · + Xn ) × Z
!"#
θX1 ,...,Xn +Z '&%$
θ
×Z
a
X1 ,...Xn
θPX ,...,X ,Z
n
1
/ / PX1 ,...,Xn × Z
/ PX1 ,...,Xn + Z
Z
Z
'&%$
!"#
b
φX1 ,...,Xn
µX1 ,...,Xn
/ / P ×Z
/ P +Z
X1 ,...,Xn
X1 ,...,Xn
in which the composite vertical morphisms from left to right are respectively
+Z
×Z
⋄Z
and θX
θX
and θX
,
1 ,...,Xn
1 ,...,Xn
1 ,...,Xn
and the morphism µZ
X1 ,...,Xn is the canonical isomorphism. Exactly as in the proof of
Proposition 5.4 it follows that the identity functor of D is of degree ≤ n if and only
if squares (a) and (b) are pullback squares.
Square (b) is a pullback if and only if φZ
X1 ,...,Xn is invertible which is the case
precisely when condition Pn holds. By Proposition 3.8 and Theorem 6.14, square (a)
is a pullback if and only if θX1 ,...,Xn is an affine (resp. central) extension, which is
the case for all objects X1 , . . . , Xn precisely when D is n-nilpotent.
Corollary 6.17. A semi-abelian category has an identity functor of degree ≤ n if
and only if either of the following three equivalent conditions is satisfied:
(a) all objects are n-nilpotent, and the comparison maps
⋄Z
ϕZ
X1 ,...,Xn : PX1 ,...,Xn ⋄ Z → PX1 ,...,Xn
are invertible for all objects X1 , . . . , Xn , Z;
(b) the (n + 1)st cross-effects of the identity functor
crn+1 (X1 , . . . , Xn , Z) = K[θX1 ,...,Xn ,Z ]
vanish for all objects X1 , . . . , Xn , Z;
(c) the co-smash product is of degree ≤ n − 1, i.e. the comparison maps
⋄Z
⋄Z
: (X1 + · · · + Xn ) ⋄ Z → PX
θX
1 ,...,Xn
1 ,...,Xn
are invertible for all objects X1 , . . . , Xn , Z.
58
CLEMENS BERGER AND DOMINIQUE BOURN
Proof. Condition (a) expresses that squares (a) and (b) of the main diagram are
pullbacks. By Proposition 6.16 this amounts to an identity functor of degree ≤ n.
For (b) note that by protomodularity the cross-effect crn+1 (X1 , . . . , Xn , Z) is trivial
if and only if the regular epimorphism θX1 ,...,Xn ,Z is invertible.
The equivalence of conditions (b) and (c) follows from the isomorphism of kernels
⋄Z
K[θX1 ,...,Xn ,Z ] ∼
]. The latter is a consequence of the 3 × 3-lemma which,
= K[θX
1 ,...,Xn
applied to main diagram 6.16 and to square (n), yields a chain of isomorphisms:
×Z
+Z
⋄Z
]] ∼
] ։ K[θX
]∼
K[θX
= K[θX1 ,...,Xn ,Z ].
= K[K[θX
1 ,...,Xn
1 ,...,Xn
1 ,...,Xn
6.3. Higher duality and multilinear cross-effects. –
In Section 5 we obtained a precise criterion for when 2-nilpotency implies quadraticity, namely algebraic distributivity, cf. Corollary 5.16. We now look for a similar
criterion for when n-nilpotency implies an identity functor of degree ≤ n. Proposition
6.16 gives us an explicit exactness condition in terms of certain limit-preservation
properties (called Pn ) of pointed cobase-change along initial maps. In order to exploit
the latter we first need to dualise condition Pn into a colimit-preservation property,
extending Proposition 5.15. Surprisingly, this dualisation process yields the simple
condition that in each variable, the functor PX1 ,...,−,...,Xn takes binary sums to binary
sums in the fibre over PX1 ,...,⋆,...,Xn . For n-nilpotent semi-abelian categories, this
in turn amounts to the condition that the n-th cross-effect of the identity functor
crn (X1 , . . . , Xn ) = K[θX1 ,...,Xn ] is multilinear.
Such a characterisation of degree n functors in terms of the multilinearity of their nth cross-effect is already present in the original treatment of Eilenberg-Mac Lane [23]
for functors between abelian categories. It plays also an important role in Goodwillie’s
[29] homotopical context (where however linearity has a slightly different meaning).
The following lemma is known in contexts close to our’s.
Lemma 6.18. Let D be a σ-pointed category and let E be an abelian category. Any
∆n
F
D
multilinear functor F : Dn → E has a diagonal G : D −→
Dn −→ E of degree ≤ n.
Proof. This is a consequence of the decomposition formula of Eilenberg-Mac Lane for
functors taking values in abelian categories, cf. Remark 6.2. Indeed, an induction on
k shows that the k-th cross-effect of the diagonal crkG (X1 , . . . , Xk ) is the direct sum of
all terms F (Xj1 , . . . , Xjn ) such that the sequence (j1 , . . . , jn ) contains only integers
1, 2, . . . , k, but each of them at least once. In particular,
M
crnG (X1 , . . . , Xn ) ∼
F (Xσ(1) , . . . , Xσ(n) )
=
σ∈Σn
and the cross-effects of G of order > n vanish, whence G is of degree ≤ n.
Lemma 6.19 (cf. Proposition 2.9 in [39]). For each n ≥ 1, the n-th cross-effect of
the identity functor of a semi-abelian category preserves regular epimorphisms in each
variable.
CENTRAL REFLECTIONS AND NILPOTENCY IN EXACT MAL’TSEV CATEGORIES
59
Proof. The first cross-effect is the identity functor and the second cross-effect is the
co-smash product. Proposition 1.15 and Lemma 1.11 imply that the co-smash product
preserves regular epimorphisms in both variables.
The general case n + 1 follows from the already treated case n = 1. By symmetry
it suffices to establish the preservation property for the last variable which we shall
denote Z. We have the following formula:
crn+1 (X1 , . . . , Xn , Z) = crn⋄Z (X1 , . . . , Xn )
(n ≥ 1)
⋄Z
where crn⋄Z (X1 , . . . , Xn ) = K[θX
] denotes the n-th cross-effect of the functor
1 ,...,Xn
− ⋄ Z. Indeed, this kernel has already been identified with K[θX1 ,...,Xn ,Z ] in the
proofs of Corollaries 5.6 and 6.17. It is now straightforward to deduce preservation
of regular epimorphisms in Z using that (−) ⋄ (−) preserves regular epimorphisms in
both variables.
Corollary 6.20 (cf. Proposition 2.21 in [39]). In a semi-abelian category, the image
of a Higgins commutator [X, . . . , X] of X under a regular epimorphism f : X ։ Y is
the corresponding Higgins commutator [Y, . . . , Y ] of Y .
Proof. The Higgins commutator of length n is the image of the diagonal n-th crosseffect K[θX,...,X ] under the folding map δnX : X + · · · + X → X, cf. Section 6.2. By
Lemma 6.19, any regular epimorphism f : X ։ Y induces a regular epimorphism
K[θX,...,X ] ։ K[θY,...,Y ] on diagonal cross-effects, whence the result.
Note that the commutative square (n) of the beginning of this section induces the
following pullback square
χX1 ,...,Xn+1
PX1 ,...,Xn+1
O
PX1 ,...,Xn ,ωX
n+1
PX1 ,...,Xn ,⋆
/ P +Xn+1
X1 ,...,Xn
O
+ωX
n+1
1 ,...,Xn
PX1 ,...,Xn ,αX
PX
n+1
X1 + · · · + Xn
+αX
n+1
1 ,...,Xn
PX
/ PX1 ,...,Xn
θX1 ,...,Xn
in which the identification PX1 ,...,Xn−1 ,⋆ = X1 + · · · + Xn−1 is exploited to give the
left vertical morphisms names. Recall that αX : ⋆ → X and ωX : X → ⋆ denote the
initial and terminal maps.
Proposition 6.21. For any objects X1 , . . . , Xn−1 , Y, Z of a σ-pointed category (D, ⋆D )
with pullbacks consider the following diagram
ρX1 ,...,Xn−1 ,Y,Z
/ PX1 ,...,Xn−1 ,Y +Z
PX1 ,...,Xn−1 ,Y + Z
O
O
PX
PX1 ,...,Xn−1 ,ωY +Z
X1 + · · · + Xn−1 + Z
Y
1 ,...,Xn−1 ,πZ
PX
Y
1 ,...,Xn−1 ,ιZ
/ PX1 ,...,Xn−1 ,Z
θX1 ,...,Xn−1 ,Z
in which the horizontal map ρX1 ,...,Xn−1 ,Y,Z is induced by the pair PX1 ,...,Xn−1 ,ιZY :
PX1 ,...,Xn−1 ,Y → PX1 ,...,Xn−1 ,Y +Z and PαX1 ,...,αXn−1 ,ιZY : Z → PX1 ,...,Xn−1 ,Y +Z ;
60
CLEMENS BERGER AND DOMINIQUE BOURN
(1) The functor PX1 ,...,Xn−1 ,− : D → PtX1 +···+Xn−1 (D) preserves binary sums if
and only if the upward-oriented square is a pushout for all objects Y, Z;
(2) the category D satisfies condition Pn (cf. Definition 6.15) if and only if the
downward-oriented square is a pullback for all objects X1 , . . . , Xn−1 , Y, Z.
In particular, (1) and (2) hold simultaneously whenever θX1 ,...,Xn−1 ,Z is an affine
extension for all objects X1 , . . . , Xn−1 , Z.
Proof. The second assertion follows from the discussion in Section 3.2. For (1), observe
that the left upward-oriented square of the following diagram
PX
Z
1 ,...,Xn−1 ,ιY
/
/ PX1 ,...,Xn−1 ,Y +Z
/ PX1 ,...,Xn−1 ,Y + Z
ρ
X1 ,...,Xn−1 ,Y,Z
O
O
PX1 ,...,Xn−1 ,Y
O
PX
PX1 ,...,Xn−1 ,αY
X1 + · · · + Xn−1
PX
Y
1 ,...,Xn−1 ,πZ
θ
/ X1 + · · · + Xn−1 + ZX1 ,...,Z
/
PX1 ,...,Xn−1 ,αZ
Y
1 ,...,Xn−1 ,ιZ
/ PX1 ,...,Xn−1 ,Z
is a pushout so that the whole upward-oriented rectangle is a pushout if and only if
the right upward-oriented square is a pushout, establishing (1).
For (2) observe that the right downward-oriented square of the following diagram
ρX1 ,...,Xn−1 ,Y,Z
χX1 ,...,Xn−1 ,Y +Z
/ PX1 ,...,Xn−1 ,Y +Z
/ P +Y +Z
X1 ,...,Xn−1
PX1 ,...Xn−1 ,Y + Z
O
O
Y
+πZ
1 ,...,Xn−1
PX
O
+ιY
Z
1 ,...,Xn−1
PX
θX1 ,...,Xn−1 ,Z
χX ,...,Xn−1 ,Z
/ PX1 ,...,Xn−1 ,Z 1
/ P +Z
X1 + · · · + Xn−1 + Z
/ X1 ,...,Xn−1
+Z
θX
,··· ,X
1
n−1
is a pullback (see below) so that the whole downward-oriented rectangle is a pullback
if and only if the left downward-oriented square is a pullback. The whole downwardoriented rectangle is a pullback if and only if the comparaison map
+Z
PX1 ,...,Xn−1 ,Y + Z → PX
1 ,...,Xn−1 ,Y
is invertible (i.e. if and only if condition Pn holds) since the following square is by
definition a pullback in the fibre PtZ (D):
P +Z
+Z
PX
1 ,...,Xn−1 ,Y
O
χ+Z
X ,...,X
1
n−1 ,Y
Y
+πZ
1 ,...,Xn−1
PX
X1 ,...,Xn−1 ,π Y
Z
X1 + · · · + Xn−1 + Z
/ P +Y +Z
X1 ,...,Xn−1
O
+Z
θX
,...,X
1
n−1
/ P +Z
X1 ,...,Xn−1
CENTRAL REFLECTIONS AND NILPOTENCY IN EXACT MAL’TSEV CATEGORIES
61
Thus (2) is established. The pullback property of the right square above follows finally
from considering the following diagram
χX1 ,...,Xn−1 ,Y +Z
/ P +Y +Z
X1 ,...,Xn−1
PX1 ,...,Xn−1 ,Y +Z
O
Y
+πZ
PX ,...,X
1
n−1
/ P +Z
O
+ιY
Z
1 ,...,Xn−1
PX
PX1 ,...,Xn−1 ,Z
χX1 ,...,Xn−1 ,Z X1 ,...,Xn−1
O
O
X1 + · · · + Xn−1
/ PX1 ,...,Xn−1
θX1 ,...,Xn−1
in which whole rectangle and lower square are downward-oriented pullbacks.
Theorem 6.22. Let D be an n-nilpotent σ-pointed exact Mal’tsev category such that
the identity functor of Niln−1 (D) is of degree ≤ n − 1.
The following properties are equivalent:
(a) the identity functor of D is of degree ≤ n;
(b) the category D satisfies condition Pn (cf. Definition 6.15);
(c) the functor PX1 ,...,Xn−1 ,− : D → PtX1 +···+Xn−1 (D) preserves binary sums for
all objects X1 , . . . , Xn−1 .
If D is semi-abelian then the former properties are also equivalent to:
(d) the n-th cross-effect of the identity is coherent in each variable;
(e) the n-th cross-effect of the identity is linear in each variable;
(f) the diagonal n-th cross-effect of the identity is a functor of degree ≤ n.
Proof. It follows from Proposition 6.16 that properties (a) and (b) are equivalent,
while properties (b) and (c) are equivalent by Theorem 6.14 and Proposition 6.21.
For the equivalence between (c) and (d), note first that the n-th cross-effect preserves
regular epimorphisms in each variable by Lemma 6.19 so that coherence (in the last
variable) amounts to the property that the canonical map
crn (X1 , . . . , Xn−1 , Y ) + crn (X1 , . . . , Xn−1 , Z) → crn (X1 , . . . , Xn−1 , Y + Z)
is a regular epimorphism. Since by Theorem 6.14 for W = Y, Z, Y + Z the regular
epimorphism X1 +· · ·+Xn−1 +W ։ PX1 ,...,Xn−1 ,W is an affine extension, Proposition
3.11 establishes the equivalence between (c) and (d). Finally, consider the following
commutative diagram in Nil1 (D) = Ab(D)
I 1 (crn (X1...n−1 , Y ) + crn (X1...n−1 , Z))
crn (X1...n−1 , Y + Z)
∼
=
/ crn (X1...n−1 , Y ) × crn (X1...n−1 , Z)
O
crn (X1 ,...,Xn−1 ,θY,Z )
/ crn (X1...n−1 , Y × Z)
62
CLEMENS BERGER AND DOMINIQUE BOURN
in which the upper horizontal map is invertible because the n-th cross-effect takes
values in abelian group objects. It follows that the left vertical map is a section so
that property (d) is equivalent to the invertibility of this left vertical map. Therefore,
(d) is equivalent to the invertibility of the diagonal map
crn (X1 , . . . , Xn−1 , Y + Z) → crn (X1 , . . . , Xn−1 , Y ) × crn (X1 , . . . , Xn−1 , Z)
which expresses linearity in the last variable, i.e. property (e).
Property (e) implies property (f) by Lemma 6.18. It suffices now to prove that (f)
implies (a). The Higgins commutator [X, . . . , X] of length n is the image of diagonal
n-th cross-effect crn (X, . . . , X) under the n-th folding map δnX : X + · · · + X → X.
The Higgins commutator of length n is thus a quotient-functor of the diagonal n-th
cross-effect and as such a functor of degree ≤ n by Theorem 6.23a. Corollary 6.12
n−1
and Remark 2.13 imply that the kernel K[ηX
: X ։ I n−1 (X)] (considered as a
functor in X) is a subfunctor of the Higgins commutator of length n and hence, again
by Theorem 6.23a, a functor of degree ≤ n. It follows then from the short exact
sequence of endofunctors
⋆ −→ K[η n−1 ] −→ idD −→ I n−1 −→ ⋆
(by a third application of Theorem 6.23a) that the identity functor of D is also of
degree ≤ n, whence (f) implies (a).
6.4. Homogeneous nilpotency towers. –
One of the starting points of this article was the existence of a functorial nilpotency
tower for any σ-pointed exact Mal’tsev category E, cf. Section 2.4. It is not surprising
that for a semi-abelian category E the successive kernels of the nilpotency tower
capture the essence of the whole tower.
To make this more precise, we denote by
M
M
LE (X) =
LnE (X) =
K[I n (X) ։ I n−1 (X)] ∈ Ab(E)
n≥1
n≥1
the graded abelian group object defined by the successive kernels. This construction
is a functor in X. The nilpotency tower of E is said to be homogeneous if for each n,
the n-th kernel functor LnE : E → Ab(E) is a functor of degree ≤ n.
The degree of a functor does not change under composition with conservative left
exact functors. We can therefore consider LnE as an endofunctor of E. Observe also
that the binary sum in Niln (E) is obtained as the reflection of the binary sum in E.
This implies that the degree of LnE is the same as the degree of LnNiln (E) . We get the
following short exact sequence of endofunctors of Niln (E)
⋆ −→ LnNiln (E) −→ idNiln (E) −→ IEn,n−1 −→ ⋆
where the last term is the relative Birkhoff reflection IEn,n−1 : Niln (E) → Niln−1 (E).
A more familiar way to express the successive kernels LnE (X) of the nilpotency
tower of X is to realise them as subquotients of the lower central series of X. Indeed,
CENTRAL REFLECTIONS AND NILPOTENCY IN EXACT MAL’TSEV CATEGORIES
63
the 3 × 3-lemma implies that there is a short exact sequence
⋆ −→ LnE (X) = γn (X)/γn+1 (X) −→ X/γn+1 (X) −→ X/γn (X) −→ ⋆
where γn+1 (X) denotes the iterated Huq commutator of X of length n + 1, i.e. the
n
kernel of the n-th Birkhoff reflection ηX
: X ։ I n (X), cf. Remark 2.13.
The conclusion of the following theorem is folklore among those who are familiar
with Goodwillie calculus in homotopy theory (cf. [3, 29]). Ideally, we would have liked
to establish Theorem 6.23c by checking inductively one of the conditions of Theorem
6.22 without using any computation involving elements.
Theorem 6.23. Let E be a semi-abelian category.
(a) For any short exact sequence ⋆ −→ F1 −→ F −→ F2 −→ ⋆ of endofunctors
of E, F is of degree ≤ n if and only if F1 and F2 are both of degree ≤ n;
(b) The nilpotency tower of E is homogeneous if and only if the identity functors
of Niln (E) are of degree ≤ n for all n;
(c) The category of groups and the category of Lie algebras have homogeneous
nilpotency towers.
Proof. For (a) we need the following cogluing lemma for regular epimorphisms in
regular categories: for any quotient-map of cospans
X
f
X′
//Zo
h
Y
g
Y′
/ / Z′ o
in which the left naturality square is a regular pushout (cf. Section 1.3), the induced
map on pullbacks f ×h g : X ×Z Y → X ′ ×Z ′ Y ′ is again a regular epimorphism.
Indeed, a diagram chase shows that the following square
Xo
X ×Z ′ Z o
′
X ×Z Y
′
(X ×
Z′
Y ′ ) ×Y ′ Y
is a pullback. The left vertical map is a regular epimorphism by assumption so that the
right vertical map is a regular epimorphism as well. Since g is a regular epimorphism,
the projection (X ′ ×Z ′ Y ′ ) ×Y ′ Y → X ′ ×Z ′ Y ′ is again a regular epimorphism so that
f ×h g is the composite of two regular epimorphisms.
F
The limit construction PX
is an iterated pullback along split epimorphisms.
1 ,...,Xn+1
Therefore, Corollary 1.7 and the cogluing lemma show inductively that the morphism
64
CLEMENS BERGER AND DOMINIQUE BOURN
F2
F
PX
→ PX
induced by the quotient-map F ։ F2 is a regular epimor1 ,...,Xn+1
1 ,...,Xn+1
phism. The 3 × 3-lemma yields then the exact 3 × 3-square
⋆
⋆
⋆
⋆
/ crF1 (X1 , . . . , Xn+1 )
n+1
F
/ crn+1
(X1 , . . . , Xn+1 )
/ crF2 (X1 , . . . , Xn+1 )
n+1
/⋆
⋆
/ F1 (X1 + · · · + Xn+1 )
/ F (X1 + · · · + Xn+1 )
/ F2 (X1 + · · · + Xn+1 )
/⋆
⋆
/ P F1
X1 ,...,Xn+1
/ PF
X1 ,...,Xn+1
/ P F2
X1 ,...,Xn+1
/⋆
⋆
⋆
⋆
from which (a) immediately follows, cf. Corollary 6.17b.
For (b) we can assume inductively that Niln−1 (E) has an identity functor of degree
≤ n − 1 so that the Birkhoff reflection IEn,n−1 : Niln (E) → Niln−1 (E) is of degree
≤ n−1, and finally IEn,n−1 is also of degree ≤ n−1 when considered as an endofunctor
of Niln (E). Statement (b) follows then from (a) by induction on n.
For (c) we treat the group case, the Lie algebra case being very similar. In the
category of groups, the graded object LGrp (X) is a graded Lie ring with Lie bracket
m+n
n
[−, −] : Lm
Grp ⊗ LGrp (X) → LGrp (X) induced by the commutator map (x, y) 7→
−1 −1
xyx y
in X. This graded Lie ring is generated by its elements of degree 1, cf.
Lazard [51, Section I.2]. In particular, there is a regular epimorphism of abelian
groups L1Grp (X)⊗n → LnGrp (X) which is natural in X. The functor which assigns
to X the tensor power L1Grp (X)⊗n is the diagonal of a multilinear abelian-groupvalued functor in n variables, and hence a functor of degree ≤ n by Lemma 6.18. It
follows from (a) that its quotient-functor LnGrp is of degree ≤ n as well, whence the
homogeneity of the nilpotency tower in the category of groups.
Theorem 6.24. Let E be a semi-abelian category.
The following conditions are equivalent:
(a)
(b)
(c)
(d)
The nilpotency tower of E is homogeneous;
For each n, the n-th Birkhoff reflection I n : E → Niln (E) is of degree ≤ n;
For each n, an object of E is n-nilpotent if and only if it is n-folded;
For each object X of E, iterated Huq commutator [X, [X, · · · , X] · · · ]] and
Higgins commutator [X, X, . . . , X] of same length coincide.
If E satisfies one and hence all of these conditions, then so does any reflective
Birkhoff subcategory of E. If E is algebraically extensive and satisfies one and hence
all of these conditions then so does the fibre PtX (E) over any object X of E.
Proof. We have already seen that (b) is equivalent to condition (b) of Theorem 6.23,
which implies the equivalence between (a) and (b). Propositions 6.5 and 6.13 show
CENTRAL REFLECTIONS AND NILPOTENCY IN EXACT MAL’TSEV CATEGORIES
65
that (b) implies (c), while Theorem 6.8 shows that (c) implies (b). The equivalence
between (c) and (d) is proved in exactly the same way as Corollary 5.19.
Let D be a reflective Birkhoff subcategory of E. We shall show that D inherits (c)
from E. By Proposition 6.13, it suffices to show that in D each n-nilpotent object X
is n-folded. Since the inclusion D ֒→ E is left exact, it preserves n-nilpotent objects
so that X is n-nilpotent in E, and hence by assumption n-folded in E. The Birkhoff
reflection E → D preserves sums and the limit construction PX1 ,...,Xn+1 by an iterated
application of Proposition 2.8. Therefore, X is indeed n-folded in D.
By Lemma 5.10 algebraic extensivity implies that all pointed base-change functors
are exact. By Lemma 2.15 this implies that the following square of functors
PtX (E)
n
IPt
X (E)
/ Niln (PtX (E))
∗
ωX
∗
ωX
E
IEn
/ Niln (E)
commutes up to isomorphism. The vertical functors are exact and conservative.
n
Therefore, if IEn is of degree ≤ n then IPt
is of degree ≤ n as well.
X (E)
6.5. On Moufang loops and triality groups. –
We end this article by giving an example of a semi-abelian category in which
2-foldedness is not equivalent to 2-nilpotency, namely the semi-abelian variety of
Moufang loops. In particular, the semi-abelian subvariety of 2-nilpotent Moufang
loops is neither quadratic (cf. Proposition 6.5) nor algebraically distributive (cf.
Corollaries 5.16 and 5.19). The nilpotency tower of the semi-abelian category of
Moufang loops is thus inhomogeneous (cf. Theorem 6.24). Nevertheless, the category
of Moufang loops fully embeds into the category of triality groups [22, 28, 37] which,
as we will see, is a semi-abelian category with homogeneous nilpotency tower.
Recall [15] that a loop is a unital magma (L, ·, 1) such that left and right translation
by any element z ∈ L are bijective. A Moufang loop [56] is a loop L such that
(z(xy))z = (zx)(yz) = z((xy)z) for all x, y, z ∈ L. Moufang loops form a semiabelian variety which contains the variety of groups as a reflective Birkhoff subvariety.
Moufang loops share many properties of groups, but the lack of a full associative law
complicates the situation. The main example of a non-associative Moufang loop is the
set of invertible elements of a non-associative alternative algebra (i.e. in characteristic
6= 2 a unital algebra in which the difference (xy)z − x(yz) alternates in sign whenever
two variables are permuted). In particular, the set O∗ of non-zero octonions forms
a Moufang loop. Taking the standard real basis of the octonions together with their
additive inverses yields a Moufang subloop
O16 = {±1, ±e1, . . . , ±e7 }
with sixteen elements. We will see that O16 is 2-nilpotent, but not 2-folded.
66
CLEMENS BERGER AND DOMINIQUE BOURN
By Moufang’s theorem [56], any Moufang loop, which can be generated by two
elements, is associative and hence a group. In particular, for any element of a Moufang loop, left and right inverse coincide. The kernel of the reflection of a Moofang
loop L into the category of groups is the so-called associator subloop [L, L, L]ass of
L. For a Moufang loop L, the associator subloop is generated by the elements of the
form [x, y, z] = ((xy)z)(x(yz))−1 . Such an “associator” satisfies [1, y, z] = [x, 1, z] =
[x, y, 1] = 1 and is thus 3-reducible, cf. Remark 6.4. This implies that for a Moufang
loop L, the associator subloop [L, L, L]ass is contained in the ternary Higgins commutator [L, L, L], cf. Proposition 6.1 and Section 6.2. In conclusion, any 2-folded
Moufang loop has a trivial associator subloop and is therefore a 2-folded group. In
particular, O16 cannot be 2-folded since O16 is not a group. One can actually show
that [O16 , O16 , O16 ] = {±1}. On the other hand, the centre of O16 is also {±1}, and
the quotient by the centre O16 /{±1} is isomorphic to (Z/2Z)3 . This implies that O16
is 2-nilpotent, i.e. [O16 , [O16 , O16 ]] = {1}.
The variety of Moufang loops is interesting with respect to the existence of centralisers. Since algebraic distributivity fails, such centralisers do not exist for general
subloops, cf. [13]. Nevertheless, each Moufang loop L has a centre Z(L) in the sense
of Section 1.5, i.e. a centraliser Z(1L ) for its identity 1L : L → L. This centre
Z(L) is a normal subloop of L, and is the intersection Z(L) = M (L) ∩ N (L) of
the Moufang centre M (L) = {z ∈ L | zx = xz ∀x ∈ L} with the so-called nucleus
N (L) = {z ∈ L | [z, x, y] = [x, z, y] = [x, y, z] = 1 ∀x, y ∈ L}, cf. Bruck [15].
Groups with triality have been introduced in the context of Moufang loops by
Glauberman [28] and Doro [22]. A triality on a group G0 is an action (by automorphisms) of the symmetric group S3 on three letters such that for all g ∈ G0 and
σ ∈ S3 (resp. ρ ∈ S3 ) of order 2 (resp. 3), the identity
[σ, g](ρ.[σ, g])(ρ2 .[σ, g]) = 1
holds where [σ, g] = (σ.g)g −1 . We denote the split epimorphism associated to the
group action by p : G0 ⋊ S3 ⇆ S3 : i and call it the associated triality group. The
defining relations for a group with triality are equivalent to the following condition
on the associated triality group p : G ⇆ S3 : i (cf. Liebeck [52] and Hall [37]):
for any two special elements g, h ∈ G such that p(g) 6= p(h) one has (gh)3 = 1
where g ∈ G is called special if g is conjugate in G to some element of order 2 in
i(S3 ). For the obvious notion of morphism, the category TriGrp⋆ of triality groups
is a full subcategory of the fibre PtS3 (Grp) over the symmetric group S3 .
The category TriGrp⋆ is closed under taking subobjects, products and quotients
in PtS3 (Grp). Moreover, quotienting out the normal subgroup generated by the
products (gh)3 for all pairs of special elements (g, h) such that p(g) 6= p(h) defines
a reflection PtS3 (Grp) → TriGrp⋆ . Therefore, TriGrp⋆ is a reflective Birkhoff subcategory of PtS3 (Grp). Since the category of groups is an algebraically extensive
semi-abelian category (cf. Section 5.3) with homogeneous nilpotency tower (cf. Theorem 6.23), so is its fibre PtS3 (Grp) by Lemma 5.13 and Theorem 6.24. The reflective
Birkhoff subcategory TriGrp⋆ formed by the triality groups is thus also a semi-abelian
category with homogeneous nilpotency tower, again by Theorem 6.24.
CENTRAL REFLECTIONS AND NILPOTENCY IN EXACT MAL’TSEV CATEGORIES
67
This result is remarkable because the category of triality groups contains the category of Moufang loops as a full coreflective subcategory, and the latter has an inhomogeneous nilpotency tower. The embedding of Moufang loops and its right adjoint
have been described by Doro [22] for groups with triality, and by Hall [37] for the
associated triality groups, see also Grishkov-Zavarnitsine [34]. Moufang loops can
thus up to equivalence of categories be identified with triality groups for which the
counit of the adjunction is invertible. Considering them inside the category of triality
groups permits the construction of a homogeneous nilpotency tower.
Acknowledgements
We are grateful to Georg Biedermann, Rosona Eldred, Marino Gran, James Gray,
Jonathan Hall, George Janelidze, Daniel Tanré and Tim Van der Linden for helpful
discussions. We are also grateful to the referee for his careful reading of our manuscript. Special thanks are due to Manfred Hartl whose seminar talk in September
2013 in Nice was the starting point for this work. The first author acknowledges
financial support of the French ANR grant HOGT.
References
[1] H.-J. Baues and T. Pirashvili – Quadratic endofunctors of the category of groups, Adv. Math.
141 (1999), 167–206. 36, 50
[2] I. Berstein and and T. Ganea – Homotopical nilpotency, Illinois J. Math. 5 (1961), 99–130. 5
[3] G. Biedermann and B. Dwyer – Homotopy nilpotent groups, Algebr. Geom. Topol. 10 (2010),
33–61. 5, 63
[4] F. Borceux and D. Bourn – Mal’cev, protomodular, homological and semi-abelian categories,
Math. Appl. 566, Kluwer Acad. Publ., 2004. 3, 6, 7, 13, 14, 25, 28, 29, 31, 32, 33, 36, 43, 47
[5] D. Bourn – Normalization equivalence, kernel equivalence and affine categories, Lect. Notes
Math. 1488, Springer Verlag 1991, 43–62. 14, 25
[6] D. Bourn – Mal’tsev categories and fibration of pointed objects, Appl. Categ. Struct. 4 (1996),
307–327. 6, 11, 13, 14, 24, 25, 30
[7] D. Bourn – The denormalized 3 × 3 lemma, J. Pure Appl. Algebra 177 (2003), 113–129. 9, 10
[8] D. Bourn – Commutator theory in regular Mal’tsev categories, AMS Fields Inst. Commun. 43
(2004), 61–75. 6, 7, 8, 21
[9] D. Bourn – Commutator theory in strongly protomodular categories, Theory Appl. Categ. 13
(2004), 27–40. 14
[10] D. Bourn – On the monad of internal groupoids, Theory Appl.Categ. 28 (2013), 150–165. 42
[11] D. Bourn and M. Gran – Central extensions in semi-abelian categories, J. Pure Appl. Algebra
175 (2002), 31–44. 6, 7, 11
[12] D. Bourn and M. Gran – Centrality and connectors in Maltsev categories, Algebra Universalis
48 (2002), 309–331. 6, 7
[13] D. Bourn and J.R.A. Gray – Aspects of algebraic exponentiation, Bull. Belg. Math. Soc. 19
(2012), 823–846. 3, 14, 15, 41, 66
[14] D. Bourn and D. Rodelo – Comprehensive factorization and I-central extensions, J. Pure Appl.
Algebra 216 (2012), 598–617. 20, 24
[15] R. H. Bruck – A survey of binary systems, Ergebnisse der Mathematik und ihrer Grenzgebiete
20, Springer Verlag 1958. 2, 65, 66
[16] A. Carboni and G. Janelidze – Smash product of pointed objects in lextensive categories, J. Pure
Appl. Algebra 183 (2003), 27–43. 3, 31, 37, 46
[17] A. Carboni, G. M. Kelly and M. C. Pedicchio – Some remarks on Mal’tsev and Goursat categories, Appl. Categ. Struct. 1 (1993), 385–421. 1, 2, 5, 6, 10, 11, 15, 31
68
CLEMENS BERGER AND DOMINIQUE BOURN
[18] A. Carboni, S. Lack and R. F. C. Walters – Introduction to extensive and distributive categories,
J. Pure Appl. Algebra 84 (1993), 145–158. 41
[19] A. Carboni, J. Lambek, M. C. Pedicchio, – Diagram chasing in Malcev categories, J. Pure Appl.
Algebra 69 (1991), 271–284. 1, 6
[20] A. Cigoli, J. R. A. Gray, T. Van der Linden – Algebraically coherent categories, Theory Appl.
Categ. 30 (2015), 1864–1905. 3, 40, 41, 42, 44, 45
[21] C. Costoya, J. Scherer, A. Viruel – A torus theorem for homotopy nilpotent groups,
arXiv:1504.06100. 5
[22] S. Doro – Simple Moufang loops, Math. Proc. Cambridge Philos. Soc. 83 (1978), 377–392. 5,
65, 66, 67
[23] S. Eilenberg and S. Mac Lane – On the groups H(π,n). II. Methods of computation, Ann. of
Math. (2) 60 (1954), 49–139. 3, 4, 36, 47, 58
[24] R. Eldred – Goodwillie calculus via adjunction and LS cocategory, arXiv:1209.2384. 5
[25] T. Everaert and T. Van der Linden – Baer invariants in semi-abelian categories I: general
theory, Theory Appl. Categ. 12 (2004), 1–33. 2, 19, 21
[26] T. Everaert and T. Van der Linden – A note on double central extensions in exact Mal’tsev
categories, Cah. Topol. Géom. Differ. Catég. 51 (2010), 143–153. 6
[27] R. S. Freese and R. N. McKenzie – Commutator theory for congruence modular varieties, London Math. Soc. Lect. Note Series 125, Cambridge Univ. Press, Cambridge, 1987. 2
[28] G. Glauberman – On loops of odd order II, J. Algebra 8 (1968), 383–414. 5, 65, 66
[29] T. G. Goodwillie – Calculus III. Taylor series, Geom. Topol. 7 (2003), 645–711. 1, 2, 4, 5, 36,
37, 47, 58, 63
[30] M. Gran – Central extensions and internal groupoids in Maltsev categories, J. Pure Appl. Alg.
155 (2001), 139–166. 15
[31] M. Gran – Applications of categorical Galois theory in universal algebra, AMS Fields Inst.
Commun. 43 (2004), 243–280. 6, 20
[32] M. Gran and D. Rodelo – Beck-Chevalley condition and Goursat categories, arXiv:1512.04066.
6, 20
[33] M. Gran and T. Van der Linden – On the second cohomology group in semi-abelian categories,
J. Pure Appl. Algebra 212 (2008), 636–651. 14, 22, 55
[34] A. N. Grishkov and A. V. Zavarnitsine – Groups with triality, J. Algebra Appl. 5 (2006), 441–
463. 67
[35] J. R. A. Gray – Algebraic exponentiation in general categories, Appl. Categ. Struct. 20 (2012),
543–567. 14, 41, 42
[36] J. R. A. Gray – Algebraic exponentiation for categories of Lie algebras, J. Pure Appl. Algebra
216 (2012), 1964–1967. 14, 41
[37] J. I. Hall – Central automorphisms, Z ∗ -theorems, and loop structures, Quasigroups and Related
Systems 19 (2011), 69–108. 5, 65, 66, 67
[38] M. Hartl and B. Loiseau – On actions and strict actions in homological categories, Theory
Appl. Categ. 27 (2013), 347–392. 3, 4, 31, 36, 37, 38, 40, 45, 46, 47
[39] M. Hartl and T. Van der Linden – The ternary commutator obstruction for internal crossed
modules, Adv. Math. 232 (2013), 571–607. 3, 4, 31, 36, 37, 38, 40, 45, 46, 47, 55, 58, 59
[40] M. Hartl and C. Vespa – Quadratic functors on pointed categories, Adv. Math. 226 (2011),
3927–4010. 4, 36, 50
[41] P. J. Higgins – Groups with multiple operators, Proc. London Math. Soc. 6 (1956), 366–416. 47
[42] M. Hovey – Lusternik-Schnirelmann cocategory, Illinois J. Math. 37 (1993), 224–239. 5, 46
[43] S. A. Huq – Commutator, nilpotency and solvability in categories, Quart. J. Math. Oxford 19
(1968), 363–389. 2, 14, 19, 21
[44] G. Janelidze and G. M. Kelly – Galois theory and a general notion of central extension, J. Pure
Appl. Algebra 97 (1994), 135–161. 6, 18, 20
[45] G. Janelidze and G. M. Kelly – Central extensions in universal algebra: a unification of three
notions, Algebra Universalis 44 (2000), 123–128. 6
CENTRAL REFLECTIONS AND NILPOTENCY IN EXACT MAL’TSEV CATEGORIES
69
[46] G. Janelidze and G. M. Kelly – Central extensions in Mal’tsev varieties, Theory Appl. Categ.
7 (2000), 219–226. 6
[47] G. Janelidze, L. Márki and W. Tholen – Semi-abelian categories, J. Pure Appl. Algebra 168
(2002), 367–386. 2
[48] G. Janelidze, M. Sobral and W. Tholen – Beyond Barr exactness: effective descent morphisms,
Cambridge Univ. Press, Encycl. Math. Appl. 97 (2004), 359–405. 24
[49] M. Jibladze and T. Pirashvili – Linear extensions and nilpotence of Maltsev theories, Contributions to Algebra and Geometry 46 (2005), 71–102. 21
[50] B. Johnson and R. McCarthy – A classification of degree n functors I/II, Cah. Topol. Géom.
Différ. Catég. 44 (2003), 2–38, 163–216. 36
[51] M. Lazard – Sur les groupes nilpotents et les anneaux de Lie, Ann. Sci. E. N. S. 71 (1954),
101–190. 5, 64
[52] M. W. Liebeck – The classification of finite simple Moufang loops, Math. Proc. Cambridge
Philos. Soc. 102 (1987), 33–47. 66
[53] A. I. Mal’cev – On the general theory of algebraic systems, Mat. Sbornik N. S. 35 (1954), 3–20.
2
[54] S. Mantovani and G. Metere – Normalities and commutators, J. of Algebra 324 (2010), 2568–
2588. 4, 31, 47
[55] J. Mostovoy – Nilpotency and dimension series for loops, Comm. Algebra 36 (2008), 1565–1579.
4
[56] R. Moufang – Zur Struktur von Alternativkörpern, Math. Ann. 110 (1935), 416–430. 65, 66
[57] M. C. Pedicchio – A categorical approach to commutator theory, J. of Algebra 177 (1995),
647–657. 7
[58] J. Penon – Sur les quasi-topos, Cah. Topol. Géom. Diff. 18 (1977), 181–218. 24
[59] D. Quillen – Homotopical algebra, Lect. Notes Math. 43, Springer Verlag 1967. 5
[60] J. D. H. Smith – Mal’cev varieties, Lect. Notes Math. 554, Springer Verlag 1976. 2, 7
[61] D. Stanovsky and P. Vojtěchovskỳ, Commutator theory for loops, J. of Algebra 399 (2014),
290–322. 4
[62] T. Van der Linden – Simplicial homotopy in semi-abelian categories, J. K-theory 4 (2009),
379–390. 2, 5
Université de Nice, Lab. J. A. Dieudonné, Parc Valrose, 06108 Nice Cedex, France
E-mail address: [email protected]
Lab. J. Liouville, CNRS Fed. Rech. 2956, Calais, France
E-mail address: [email protected]
| 4math.GR
|
Compressed Anomaly Detection with Multiple
Mixed Observations
arXiv:1801.10264v1 [cs.IT] 31 Jan 2018
Natalie Durgin, Rachel Grotheer, Chenxi Huang, Shuang Li, Anna Ma,
Deanna Needell and Jing Qin
Abstract We consider a collection of independent random variables that are identically distributed, except for a small subset which follows a different, anomalous
distribution. We study the problem of detecting which random variables in the collection are governed by the anomalous distribution. Recent work proposes to solve
this problem by conducting hypothesis tests based on mixed observations (e.g. linear
combinations) of the random variables. Recognizing the connection between taking
mixed observations and compressed sensing, we view the problem as recovering
the “support” (index set) of the anomalous random variables from multiple meaNatalie Durgin
Spiceworks
Austin, TX, 78746.
e-mail: [email protected]
Rachel Grotheer
Goucher College
Baltimore, MD 21204.
Chenxi Huang
Yale University
New Haven, CT 06511.
e-mail: [email protected]
Shuang Li
Colorado School of Mines
Golden, CO 80401.
Anna Ma
Claremont Graduate University
Claremont, CA 91711.
Deanna Needell
University of California, Los Angeles
Los Angeles, CA 90095.
Jing Qin
Montana State University
Bozeman, MT 59717.
1
2
Durgin, Grotheer, Huang, Li, Ma, Needell, Qin
surement vectors (MMVs). Many algorithms have been developed for recovering
jointly sparse signals and their support from MMVs. We establish the theoretical
and empirical effectiveness of these algorithms at detecting anomalies. We also extend the LASSO algorithm to an MMV version for our purpose. Further, we perform
experiments on synthetic data, consisting of samples from the random variables, to
explore the trade-off between the number of mixed observations per sample and the
number of samples required to detect anomalies.
1 Introduction
The problem of anomaly detection has been the focus of interest in many fields
of science and engineering, including network tomography, cognitive radio, and
radar [36, 40, 27, 4]. In this paper, we study the problem of identifying a small
number of anomalously distributed random variables within a much larger collection of independent and otherwise identically distributed random variables. We call
the random variables following the anomalous distribution anomalous random variables. A conventional approach to detecting these anomalous random variables is
to sample from each random variable individually and then apply hypothesis testing
techniques [29, 30, 31, 31].
A recent paper [13] proposes to perform hypothesis testing on mixed observations (e.g. linear combinations) of random variables instead of on samples from individual random variables. They call this technique compressed hypothesis testing.
Such an approach is motivated by the recent development of compressed sensing
[7, 15, 23, 21], a signal processing paradigm that shows a small number of random
linear measurements of a signal is sufficient for accurate reconstruction. Now a large
body of work in this area shows that optimization-based [17, 8, 16, 10, 37] and iterative [38, 33, 6] methods can reconstruct the signal accurately and efficiently when
the samples are taken via a sensing matrix satisfying certain incoherence properties
[8, 9]. Compressed sensing is also studied in a Bayesian framework, where signals
are assumed to obey some prior distribution [25, 41, 3].
The results presented in [13] show that the “mixed” measurement approach
achieves better detection accuracy from fewer samples when compared to the conventional “un-mixed” approach. However, compressed hypothesis testing requires
that the distributions of the random variables are known a priori, which may not be
available in practice. Further, as the authors pointed out, their proposed approach
requires conducting a large number of hypothesis tests, especially when the number
of random variables in the collection is large, rendering such an approach computationally prohibitive. Two efficient algorithms are proposed as alternatives in [13],
but no analytical study of their performance is provided.
We propose new methods for detecting anomalous random variables that require
minimal knowledge of the distributions, are computationally efficient, and whose
performance is easy to characterize. We begin by generalizing the compressed hypothesis testing method and posing our problem as a multiple measurement vector
Compressed Anomaly Detection with Multiple Mixed Observations
3
(MMV) problem [24, 2, 19, 18, 11, 14, 32, 5]. In the MMV compressed sensing setting, a collection of signals are recovered simultaneously, under the assumption that
they have some commonalities, such as sharing the same support. A related vein of
work involves signals that are smoothly varying, where the support may not be consistent but changes slowly over time [1, 22, 35]. While the compressed hypothesis
testing in [13] is certainly motivated by compressed sensing techniques, the authors
do not formally frame the anomaly detection problem in the compressed sensing setting. Also, they do not focus on compressed sensing algorithms that might eliminate
the need for prior knowledge of the distributions, and might lead to more efficient
detection for large collections of random variables.
In the following, we view the collection of random variables as a random vector
and aim to identify the indices of the anomalous random variables within the random
vector. We also draw an analogy between the collection of independent samples
from the random vector and an ensemble of signals where in practice these signals
often become available over time. More specifically, we consider a random vector,
X = (X1 , . . . , XN ), where the Xn ’s are independent random variables. We assume
that each Xn follows one of two distributions, D1 , D2 . We call D1 the prevalent
distribution, and D2 the anomalous distribution. We let N = {n ∈ N : 1 ≤ n ≤ N}
denote the index set of the random variables, Xn , and let K denote the index set of the
K random variables that follow the anomalous distribution. Let x(,t) ∈ RN denote
the independent realization of the random vector at time t. At each time-step t, we
obtain M mixed observations by applying the sensing matrix φt ∈ RM×N ,
yt = φt x(,t) , 1 ≤ t ≤ T,
with yt ∈ RM . Thus the goal of the anomaly detection problem in this setting is to
recover the index set K from the MMVs yt , t = 1, · · · , T .
The signals x(,t) in our formulation are not necessarily sparse and may have different supports since they are samples from a random vector and are changing over
time. Nevertheless, there is still a close connection between our formulation and that
for recovering the common sparse support of a collection of signals from MMVs.
The index set of the anomalous random variables, which corresponds to the index set
of the anomalies (realizations of anomalous random variables) in the signals x(,t) ,
is shared by all signals. This index set can thus be viewed as the common “support”
of the anomalies in the signals, which motivates us to consider the applicability of
many MMV algorithms designed for signal reconstruction. Further, the analytical
studies of many of these algorithms are readily available. We therefore investigate
which of these MMV algorithms can be applied or adapted to the anomaly detection
problem under consideration and analyze their performance in detection accuracy in
theory and through numerical experiments. We focus on algorithms presented in [2].
4
Durgin, Grotheer, Huang, Li, Ma, Needell, Qin
1.1 Contributions
In this paper, by extending the definitions of two so-called joint sparsity models
(JSMs) from [2], we introduce two new signal models, JSM-2R and JSM-3R, for
the problem of anomaly detection. For JSM-2R and JSM-3R signals, we adapt several MMV signal reconstruction algorithms to anomaly detection. Additionally, we
develop a new algorithm for the JSM-2R model that extends the Least Absolute
Shrinkage and Selection Operator (LASSO) algorithm [12] to the MMV framework. We show theoretically and numerically that these algorithms accurately detect
the anomalous random variables. We also provide numerical results which demonstrate the trade-off between the number of time-steps, and the number of mixed
observations per time-step needed to detect the anomalies.
1.2 Organization
In Section 2, we introduce the models JSM-2R, JSM-3R and the four algorithms we
have repurposed from MMV signal recovery into MMV anomaly detection, as well
as our new LASSO algorithm. We also provide theoretical guarantees in this section. In Section 3, we explore the performance of these algorithms by conducting
numerical experiments for some strategic choices of the parameters involved. Finally, we conclude in Section 4. To help keep track of notation, we provide a handy
reference table in Section 4. We adopt the convention that random variables will be
upper case and their realizations will be lower case. All matrix entries will have two,
subscripted indices. The first index will indicate the row position, the second will
indicate the column position.
2 Method
In this section, we introduce two new signal models for the anomaly detection problem and describe five algorithms for detecting anomalous random variables under
these signal models. We also provide theoretical guarantees for the algorithms.
Recall that we consider the problem of detecting K anomalous random variables
from a collection of N random variables where K N. The anomalous random
variables have a different probability distribution from that of the remaining N − K
random variables. We seek to identify the K anomalous random variables, from T
independent realizations of the N random variables. To emphasize our framing of
this random variable problem as a compressed sensing problem, we refer to the independent realizations as signals. These T signals have an important commonality:
they share the same indices of anomalous entries (realizations of anomalous random variables). Commonality among signals has already been explored in the field
of distributed compressed sensing for recovering signals that have specific corre-
Compressed Anomaly Detection with Multiple Mixed Observations
5
lation among them. Three joint sparsity models (JSMs) were introduced in [2] to
characterize different correlation structures. To utilize the commonality of the signals for anomaly detection, we propose two new signal models that are motivated
by two of the JSMs defined in [2], namely, JSM-2 and JSM-3. Since the signals
under consideration are realizations of random variables, we term the new models
JSM-2R and JSM-3R, respectively, where the appended “R” indicates the “random
variable” version of the existing JSMs.
Before we define the new models, We first briefly describe JSM-2 and JSM-3.
The JSM-2 signals are jointly sparse signals that share the same support (the indices
of non-zero entries). The JSM-3 signals consist of two components: a non-sparse
“common component” shared by all signals and a sparse “innovation component”
that is different for each signal. But the innovation components of the JSM-3 signals share the same support. We next extend these definitions to the signals in the
anomaly detection setting. The new JSM-2R and JSM-3R models are defined as
follows:
Definition 1 (JSM-2R and JSM-3R). Let the random variable Xn ∼ D1 if n ∈
/K
and Xn ∼ D2 if n ∈ K where K is the set of the anomalous indices. For a signal
ensemble x ∈ RN×T where each of its entries x(n,t) denotes the realization of Xn at
time t,
1. x is a JSM-2R signal ensemble when: |x(n,t) | is small if n ∈
/ K and |x(n,t) | is large
if n ∈ K;
I
I
2. x is a JSM-3R signal ensemble when: x(n,t) = xCn + x(n,t)
such that |x(n,t)
| is small
I
C
if n ∈
/ K and |x(n,t) | is large if n ∈ K. xn is a common component shared by all t,
I
and x(n,t)
is an innovation component that is different for different t.
The JSM-2R signal model assumes a small amplitude for variables generated
from the prevalent distribution and a large amplitude for variables generated from
the anomalous distribution. Such a model characterizes a scenario where anomalies
exhibit large spikes. This model relates to a sparse signal model where the support
of the sparse signal corresponds to the set of indices of the anomalous random variables. In fact, when D1 = N (0, σ 2 ) and D2 = N (µ, σ 2 ) with µ 6= 0, the JSM-2R
signal is a sparse signal with additive Gaussian noise.
Different from the JSM-2R signals, the JSM-3R signal model introduced above
does not have constraints on the amplitude of the signal entries x(n,t) . Rather the
signals at different time-steps are assumed to share an unknown common compoI
for signals at different
nent xCn while having a different innovation component x(n,t)
C
time-steps. Of note, the common component xn from the prevalent distribution may
or may not be the same as that from the anomalous distribution. Further, the innovaI
tion component x(n,t)
is assumed to follow the JSM-2R signal model. Such a model
characterizes a scenario where there exists a background signal that does not change
over time and the anomalies exhibit large spikes on top of the background signal.
Because of the common component, the JSM-3R signals no longer correspond to a
sparse signal model. Fig. 1 provides a visual illustration of the model nuances.
Durgin, Grotheer, Huang, Li, Ma, Needell, Qin
1 2 3 4 5 6 7 8 9 10
t
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
JSM-2R
1 2 3 4 5 6 7 8 9 10
t
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
JSM-3
n
JSM-2
n
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
n
n
6
1 2 3 4 5 6 7 8 9 10
t
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
JSM-3R
10.0
7.5
5.0
2.5
0.0
2.5
1 2 3 4 5 6 7 8 9 10
t
Fig. 1 Depiction of the existing joint sparsity models (JSM-2 and JSM-3) and the new models
developed for anomaly detection (JSM-2R and JSM-3R). The distributions used to generate this
example are the same as the ones used for the numerical experiments in Section 3, see Table 1. The
index set of the anomalies is K = {6, 10}.
2.1 Algorithms
We briefly describe the five algorithms we study in this paper, among which three
are for JSM-2R signals and two are for JSM-3R signals. Two of the algorithms for
JSM-2R signals were originally proposed for recovering JSM-2 signals, including
the one-step greedy algorithm (OSGA) and the multiple measurement vector simultaneous orthogonal matching pursuit (MMV-SOMP) algorithm. We further propose
a new MMV version of the LASSO algorithm for detecting anomalies for JSM-2R
signals and investigate its performance via numerical experiments. The two algorithms for JSM-3R were also proposed in [2] for recovering JSM-3 signals, including the Transpose Estimation of Common Component (TECC) algorithm and the
Alternating Common and Innovation Estimation (ACIE) algorithm.
For each of the presented algorithms, the goal is to identify the indices of
the anomalous random variables from the mixed measurements yt = φt x(,t) for
t = 1, 2, . . . , T. The number of anomalies K is assumed to be known a priori. We first describe three algorithms we applied to anomaly detection for JSM-2R
signals.
2.1.1 OSGA
The OSGA algorithm is a non-iterative greedy algorithm introduced in [2] to recover the support of JSM-2 signals based on inner products of the measurement
and columns of the sensing matrix (Algorithm 1). We show in Theorem 1 that with
some condition on the prevalent and anomalous distributions, the OSGA algorithm
is able to recover the anomaly indices under the JSM-2R model, using a small number of measurements per time-step. Although the OSGA algorithm is shown to work
Compressed Anomaly Detection with Multiple Mixed Observations
7
asymptotically, it may not perform well when only a small number of time-steps are
available. Empirical evidence has confirmed this conjecture when the OSGA algorithm is used to reconstruct JSM-2 signals [2]. Thus we further consider approaches
like matching pursuit [28, 34] for our problem. Next, we describe the MMV version
of orthogonal matching pursuit algorithm proposed in [2].
Algorithm 1 OSGA
1:
2:
3:
4:
5:
6:
Input: yt , φt , K.
b
Output: K.
for n = 1, 2, . . . , N do
T
Compute ξn = T1 ∑t=1
hyt , φt(,n) i2
end for
b = {n, for the K largest ξn }
return K
2.1.2 MMV-SOMP
The MMV-SOMP algorithm is an iterative greedy pursuit algorithm for recovery
of jointly sparse signals. SOMP was first proposed in [39] and was adapted to the
MMV framework in [2]. Since our focus is not on signal recovery but on detecting anomalous entries, we adapt this algorithm for our JSM-2R signal model. The
adapted algorithm is presented in Algorithm 2, which identifies the anomaly indices
one at a time. In each iteration, the column index of the sensing matrices that accounts for the largest residual across signals of all time-steps is selected. Then the
remaining columns of each sensing matrix (for each time-step) are orthogonalized.
The algorithm stops after K iterations where K is the number of anomalous random
variables. We show through numerical experiments in Section 3 that the adapted
MMV-SOMP algorithm performs better than the OSGA algorithm for a small number of time-steps.
2.1.3 MMV-LASSO
The LASSO algorithm aims to find a sparse solution to the regression problem by
constraining the L1 norm of the solution [12]. The LASSO algorithm was also considered in [13] as an efficient algorithm for anomaly detection from mixed observations. However, the authors of [13] considered the LASSO algorithm when using only one measurement at each time-step. In this paper, we further extend the
LASSO algorithm to a more general setting for MMV and term it the MMV-LASSO
algorithm. The MMV-LASSO algorithm is described in Algorithm 3. The measurements yt ∈ RM up to T time-steps are concatenated vertically to become a vector
y ∈ R(MT )×1 ; the sensing matrices φt ∈ RM×N are also concatenated vertically to become φ ∈ R(MT )×N . The concatenated measurements and sensing matrices are then
8
Durgin, Grotheer, Huang, Li, Ma, Needell, Qin
Algorithm 2 MMV-SOMP
1:
2:
3:
4:
5:
Input: yt , φt , K.
b
Output: K.
b = 0,
Initialize: K
/ residual rt0 = yt .
for k = 1, . . . , K do.
Select
|hrtk−1 , φt(,n) i|
.
kφt(,n) k2
t=1
T
nk = arg max ∑
n
6:
7:
b = [K,
b nk ].
Update K
Orthogonalize selected basis vector against previously selected vectors for all t, 1 ≤ t ≤ T :
γt0 = φt(,k) ,
if k = 1,
hφt(,nk ) , γtl i l
γt ,
γtk = φt(,nk ) −
kγtl k22
l=0
k−1
∑
8:
if k > 1.
Update the residual for all t, 1 ≤ t ≤ T ,
rtk = rtk−1 −
hrtk−1 , γtk i k
γt .
kγtk k22
9: end for
b
10: return K
fed to the regular LASSO algorithm, where the anomaly indices are found by taking
indices corresponding to the K largest amplitudes of the estimate. The LASSO problem, that is, Step 4 in Algorithm 3, can be tackled by various approaches [20, 26],
which is out of scope of this paper.
Algorithm 3 MMV-LASSO
Input: yt , φt , K.
b
Output: K.
T T
T
T T
Let y = [yT
1 , . . . , yT ] and φ = [φ1 , . . . , φT ]
Solve
1
x̂ = arg min ky − φ xk22 +λ kxk1
2
x
5: Let x̂n denote the n-th element of x̂
b = {n, for the K largest |x̂n |}
6: return K
1:
2:
3:
4:
We next describe two algorithms for anomaly detection for JSM-3R signals.
Compressed Anomaly Detection with Multiple Mixed Observations
9
2.1.4 TECC
The key difference between JSM-2R and JSM-3R signals is that JSM-3R signals
share a common component that is unknown. Thus the two algorithms for the JSM3R signals aim to first estimate the common component from the mixed measurement and subtract the contribution of this component from the measurement. The
TECC algorithm was proposed in [2] for recovering JSM-3 signals. We also adapt
the algorithm to focus only on detecting the anomalous indices of JSM-3R signals,
and the adapted algorithm can be found in Algorithm 4. The first step of the TECC
algorithm estimates the common component of the JSM-3R signals. Using this estimate, the contribution of the remaining innovation component to the measurement
can be estimated. Then algorithms for JSM-2R signals can be applied to identify the
anomaly indices. We show in Theorem 2 that the TECC algorithm is able to identify the anomalous variables under some conditions on the prevalent and anomalous
distributions. Similar to the OSGA algorithm, while Theorem 2 guarantees the success of the TECC algorithm in the asymptotic case as T goes to infinity, it may not
perform well for a small T . Next we describe an alternative algorithm also proposed
in [2] for cases with a small T .
Algorithm 4 TECC
Input: yt , φt , K.
b
Output: K.
T T
T
T T
Let y = [yT
1 , . . . , yT ] , and φ = [φ1 , . . . , φT ]
1
T
b
C
Calculate x = T M φ y.
5: Calculate ybt = yt − φt xbC .
b from ybt by Algorithm 1, 2 or 3
6: Estimate K
b
7: return K
1:
2:
3:
4:
2.1.5 ACIE
The ACIE algorithm is an extension of the TECC algorithm, also introduced in [2],
based on the observation that the initial estimate of the common component may not
be sufficiently accurate for subsequent steps. Instead of one-time estimation in the
TECC algorithm, the ACIE algorithm iteratively refines the estimates of the common component and the innovation components. The ACIE algorithm can also be
easily adapted for the JSM-3R signals for anomaly detection. In the ACIE algorithm
described in Algorithm 5, we first obtain an initial estimate of the anomaly index
b using the TECC algorithm. Then for each iteration, we build a basis Bt for RM
set K
where M is the number of measurements at each time-step: Bt = [φt,Kb , qt ], where
b and qt
φ b is the subset of the basis vectors in φt corresponding to the indices in K
t,K
has orthonormal columns that spans the orthogonal complement of φt,Kb . Then we
10
Durgin, Grotheer, Huang, Li, Ma, Needell, Qin
can project the measurements onto qt to obtain the part of the measurement caused
by signals not in K:
yet = qt T yt ,
φet = qt T φt .
(1)
(2)
Then yet and φet are used to refine the estimate of the common component. After
subtracting the contribution of this estimated common component, algorithms such
as OSGA and MMV-SOMP described above can be applied to detect the anomalies.
Algorithm 5 ACIE
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
Input: yt , φt , K, L (iteration counter).
b
Output: K.
T T
Let y = [yT
1 , . . . , yT ]
b from Algorithm 4
Obtain an initial estimate of K
for l = 1, 2, . . . , L do
Update yet and φet according to Equations (1) and (2) for all t, 1 ≤ t ≤ T
T e
eT
eT
e†
eT e −1 eT
eT
Update xeC = φe† ye, where ye = [e
yT
T ] , φ = [φ1 , . . . , φT ] and φ = (φ φ ) φ
1 ,···,y
end for
Calculate ybt = yt − φt xeC
b from ybt by Algorithm 1, 2 or 3
Estimate K
b
return K
2.2 Theoretical Guarantees
In this section we show theoretically that Algorithm 1 and Algorithm 4 (coupled
with Algorithm 1 in step 6) can detect anomalies for the JSM-2R and JSM-3R settings, respectively.
Recall that Algorithm 1 is designed for JSM-2R signals where variables generated from the prevalent distribution are much smaller in amplitude than those
from the anomalous distribution. The following theorem shows that for JSM-2R
signals, the OSGA algorithm is able to identify the indices of the anomalous variables asymptotically, with very few measurements at each time-step.
Theorem 1. [Adapted from [2] Theorem 8] Let the M × N sensing matrix, φt , contain entries that are i.i.d ∼ N (0, 1) at each time-step t. Suppose the random variables, Xn , are distributed with D1 = N (0, σ12 ) if n ∈
/ K and D2 = N (µ2 , σ22 ) if
2
2
2
n ∈ K. Assuming µ2 +σ2 > σ1 , then with M ≥ 1 measurements per time-step, OSGA
recovers K with probability approaching one as T → ∞.
Proof. Before diving into the proof of Theorem 1, we first observe that the signals
correspond to the JSM-2R signals: with a zero mean and a potentially small variance
Compressed Anomaly Detection with Multiple Mixed Observations
11
σ12 for the prevalent distribution D1 , the signal entry x(n,t) , n ∈
/ K (i.e. the realization
of Xn at the time-step t) is expected to have small amplitude. In contrast, with a
non-zero mean µ2 and a similar or possibly larger variance σ22 for the anomalous
distribution D2 , the amplitude of x(n,t) , n ∈ K is expected to be much larger.
We assume, for convenience and without loss of generality, that the anomalous random variables are indexed by, K = {1, 2, . . . , K}, and the prevalent random variables are indexed by N\K = {K + 1, . . . , N}. Consider that the test statistic
T
hyt , φt(,n) i2 is the sample mean of the random variable hY, Φ(,n) i2 , so
ξn = T1 ∑t=1
by the Law of Large Numbers,
lim ξn = E[hY, Φ(,n) i2 ].
T →∞
We select an arbitrary index n from each of the anomalous random variable index
set and the prevalent random variable index set, and compute E[hY, Φ(,n) i2 ] in each
case. As the final step, we compare the expected values of the two ξn and establish
that they are distinguishable under very general conditions. Without loss of generality, we select n = K + 1 for the “prevalent case” and n = 1 for the “anomalous
case”. Note that [2] refers to these cases respectively as the “bad statistics” and the
“good statistics” in their setting. For them, “bad” reflects an incorrect estimate of
the sparse support and “good” reflects a correct estimate of the sparse support.
Prevalent Case: Substituting ΦX for Y in hY, Φ(,K+1) i and rearranging we obtain
hY, Φ(,K+1) i = ∑Nn=1 Xn hΦ(,n) , Φ(,K+1) i. We can then write,
!
2
N
2
E[hY, Φ(,K+1) i ] = E
∑ Xn hΦ(,n) , Φ(,K+1) i
n=1
"
#
N
2
2
∑ (Xn ) hΦ(,n) , Φ(,K+1) i
=E
n=1
N
+ E∑
N
∑ Xn Xl hΦ(,l) , Φ(,K+1) ihΦ(,n) , Φ(,K+1) i
n=1 l=1
l6=n
N
=
∑ E[(Xn )2 ]E[hΦ(,n) , Φ(,K+1) i2 ]
n=1
N
+
N
∑ ∑ E[Xn ]E[Xl ]E[hΦ(,l) , Φ(,K+1) ihΦ(,n) , Φ(,K+1) i].
n=1 l=1
l6=n
The last step follows from the independence of Φ and X and the independence of
the Xn ’s from each other. We claim that the cross-terms above sum to zero. To see
this, we set Φ(,l) = a, Φ(,K+1) = b and Φ(,n) = c, where the entries of the vectors
a, b, c are all i.i.d. N (0, 1). We note that if l, K + 1, and n are mutually distinct, then
a, b, c are mutually independent. In this case we have,
12
Durgin, Grotheer, Huang, Li, Ma, Needell, Qin
E[ha, bihc, bi] = E[aT bcT b]
= E[aT ]E[bcT b]
= 0.
Since the cross-terms assume l 6= n, we consider the cases when either n = K + 1
or l = K + 1. In the case where n = K + 1 we have,
E[ha, bihb, bi] = E[aT bbT b]
= E[aT ]E[bbT b]
= 0.
Similarly, in the case where l = K + 1 we have,
E[hb, bihc, bi] = E[bT bcT b]
= E[cT bbT b]
= E[cT ]E[bbT b]
= 0.
Thus, all cross-terms vanish so returning to our original goal we may claim,
N
E[hY, Φ(,K+1) i2 ] =
∑ E[(Xn )2 ]E[hΦ(,n) , Φ(,K+1) i2 ]
n=1
K
=
∑ E[(Xn )2 ]E[hΦ(,n) , Φ(,K+1) i2 ] + E[(XK+1 )2 ]E[kΦ(,K+1) k4 ]
n=1
N
+
∑
E[(Xn )2 ]E[hΦ(,n) , Φ(,K+1) i2 ].
n=K+2
Examining each expected value individually, we recall that for n ∈ {1, . . . , K} = K
the Xn were distributed with D2 and thus E[(Xn )2 ] = E[Xn ]2 + Var(Xn ) = µ22 +
σ22 . Recalling that the rest of the Xn are distributed with D1 which has µ = 0,
we have that E[(Xn )2 ] = σ12 in the subsequent cases. In [2] they establish that
E[kΦ(,K+1) k4 ] = M(M + 2) and E[hΦ(,n) , Φ(,K+1) i2 ] = M, and we may use these
results without further argument because we make the same assumptions about Φ.
Finally, substituting the expected values we have just calculated, we have that as T
grows large, the statistic ξn when n ∈
/ K converges to
2
2
2
E[hY, Φ(,K+1) i ] = K(µ2 + σ2 )M + σ12 M(M + 2) + (N − K − 1)σ12 M
= M[K(µ22 + σ22 ) + (M + 1 + N − K)σ12 ].
(3)
Compressed Anomaly Detection with Multiple Mixed Observations
13
Anomalous Case: With n = 1, we proceed as in the previous case,
!
2
N
2
E[hY, Φ(,1) i ] = E
∑ Xn hΦ(,n) , Φ(,1) i
n=1
N
=
∑ E[(Xn )2 ]E[hΦ(,n) , Φ(,1) i2 ],
n=1
K
= E[(X(1,t) )2 ]E[kΦ(,1) k4 ] +
∑ E[(Xn )2 ]E[hΦ(,n) , Φ(,1) i2 ]
n=2
N
+
∑
E[(Xn )2 ]E[hΦ(,n) , Φ(,1) i2 ]
n=K+1
= (µ22 + σ22 )M(M + 2) + (K − 2)(µ22 + σ22 )M + (N − K)σ12 M
= M[(M + 1 + K)(µ22 + σ22 ) + (N − K)σ12 ].
Combining the results of (3) and (4), we have
(
M[(M + 1 + K)(µ22 + σ22 ) + (N − K)σ12 ]
lim ξn =
T →∞
M[K(µ22 + σ22 ) + (M + 1 + N − K)σ12 ]
(4)
n∈K
n∈
/ K.
The difference in the two expectations is thus,
M(M + 1)(µ22 + σ22 − σ12 ).
For any M ≥ 1 and µ22 + σ22 > σ12 , the expected value of ξn in the “anomalous case”
is strictly larger than the expected value of ξn in the “prevalent case”. Therefore,
as T increases, OSGA can distinguish between the two expected values of ξn with
overwhelming probability.
t
u
The next theorem shows that asymptotically, Algorithm 4 is able to detect anomalous variables with very few measurements at each time-step, for JSM-3R signals.
Recall that JSM-3R signals have an unknown common component shared by signals at all time-steps, while each signal has a different innovation component that
follows the JSM-2R model. The following theorem and proof assume that Algorithm 1 is implemented for step 6 of Algorithm 4. Once results like Theorem 1 exist
for Algorithms 2 and 3, then any JSM-2R algorithm could be used in step 6, and
Theorem 2 would still hold.
Theorem 2 (Adapted from [2] Theorem 10). Let the M × N sensing matrix φt at
each time-step t contain entries that are i.i.d. ∼ N (0, 1). For random variables Xn
that are distributed with D1 = N (µ1 , σ12 ) if n ∈
/ K and D2 = N (µ2 , σ22 ) if n ∈ K, if
2
2
σ2 > σ1 and with M ≥ 1, TECC algorithm (with OSGA) recovers K with probability
approaching one as T → ∞.
14
Durgin, Grotheer, Huang, Li, Ma, Needell, Qin
Proof. We first note that the signals in Theorem 2 correspond to the JSM-3R signals:
I where xI are
for n ∈
/ K, the signal entries x(n,t) can be written as x(n,t) = µ1 + xn,t
n,t
2
i.i.d. ∼ N (0, σ1 ). With zero-mean and a potentially small variance, the amplitude
I ,n∈
of xn,t
/ K is expected to be small. For n ∈ K, the signal entries x(n,t) can be written
I where xI are i.i.d. ∼ N (0, σ 2 ). With a larger variance σ 2 , the
as x(n,t) = µ2 + xn,t
n,t
2
2
I
amplitude of xn,t , n ∈ K is expected to be much larger.
By the common component estimation from Algorithm 4, we have:
1 T
φ y
xbC =
TM
1 1 T T
=
∑ φt yt
M T t=1
1
=
M
1
T
T
∑ φtT φt x(,t)
!
.
t=1
Note that this is 1/M times the sample mean of the random variable ΦT ΦX. Letting
IN denote the N × N identity matrix, we note that since Φ has independent N (0, 1)
entries then E[ΦT Φ] = MIN . Since Φ is fully independent of X,
1
1
E[ΦT ΦX] = E[ΦT Φ]E[X] = IN E[X] = E[X].
M
M
Invoking the Law of Large Numbers, we have
lim xbC = E[X].
T →∞
Let Xb = X − xbC , then as T → ∞, Xbn is distributed as N (0, σ12 ) if n ∈
/ K and
b it follows from Theorem
N (0, σ22 ) if n ∈ K. Since Yb = Y −ΦxbC = Φ(X − xbC ) = ΦX,
1 that with M ≥ 1 and σ22 > σ12 , the TECC with OSGA algorithm recovers K with
probability approaching one as T → ∞.
t
u
3 Experiments
In this section, we evaluate numerically the performance of Algorithms 1, 2, 3,
4 and 5 for anomaly detection. More specifically, we examine the success rate of
determining the anomalous index set K from the signal matrix x ∈ RN×T , whose
columns are signals obtained at each time-step and share the same anomalous indices. The performance is assessed under various settings, by varying the number
of anomalies, the number of columns in x (i.e. the time-steps) and the number of
mixed measurement M at each time-step. Our focus is on the trade-off between the
number of measurements M and the number of time-steps T required to identify K
for varying numbers of anomalies.
Compressed Anomaly Detection with Multiple Mixed Observations
15
In all experiments, the measurement matrices φt ∈ RM×N comprise independent,
N (0, 1) entries and the measurement vectors yt ∈ RM are calculated by yt = φt x(,t)
for t = 1, . . . , T . To obtain an estimate of an algorithm’s recovery success rate with
high confidence, instead of using a fixed number of random trials across the different parameter combinations, we adaptively determine the necessary number of
trials with a Jeffreys interval, a Bayesian two-tailed binomial proportion confidence
interval. When the 95% confidence interval around the true success rate shrinks to a
width smaller than 0.1, we report the current proportion of successes as the recovery
accuracy for the algorithm. The signals (i.e. x(,t) ) are generated under two models
corresponding to the JSM-2R and JSM-3R signal definitions introduced in Section
2. Algorithms 1, 2 and 3 are applied to the JSM-2R signals while Algorithms 4 and
5 are applied to the JSM-3R signals.
The experiments are summarized in Table 1. The JSM-2R experiments assume a
mean zero for the prevalent distribution and a much larger mean for the anomalous
distribution while letting the variance be small. As shown in the previous section,
signals generated from these distributions satisfy the definitions of JSM-2R. For
JSM-3R experiments, we explore two settings: First, the prevalent and anomalous
distributions are assumed to have different means; second, the two distributions have
the same mean. Recall from the previous section, we show that the means of the
distributions are the common components for the JSM-3 signals generated from
these distributions. Note that the algorithms for JSM-3R signals have no knowledge
of the mean of the prevalent or anomalous distributions.
Table 1
Signal model
JSM-2R
JSM-3R
JSM-3R
D1
D2
Algorithms
OSGA, MMV-SOMP,
N (0, 1) N (7, 1)
MMV-LASSO
N (7, 1) N (0, 10)
TECC, ACIE
N (7, 1) N (7, 10)
TECC, ACIE
We chose the distributions in Table 1 for our numerical simulations to remain
consistent with [13]. We observe, in the JSM-2R experiments, that the distributions
N (0, 1) and N (7, 1) have their means separated by three standard deviations each,
with one additional standard deviation in between for good measure. This ensures
that the distributions are statistically distinct from each other. We have not explored
how the detection accuracy is affected as we vary the proportion of overlap in the
two distributions.
3.1 JSM-2R
We now present the results of recovering the anomalous index set for the JSM-2R
signals. The signal length is fixed at N = 100 and results of K = 1, 5, and 10 anoma-
16
Durgin, Grotheer, Huang, Li, Ma, Needell, Qin
lies are presented. For each K value, K random variables follow the distribution
N (7, 1) and the other N − K random variables follow another distribution N (0, 1).
The goal is to recover the index set K of these K random variables. Fig. 2 shows
the success rate of identifying K for the three K values using the OSGA Algorithm.
Each dot in the figure denotes the success rate for a specific M (number of measurements per time-step) and a specific T (number of time-steps) estimated from a
number of trials, and the value is indicated by the color (see the colorbar). Both M
and T take values from 1 to 100. Fig. 3 and 4 plot the success rate for MMV-SOMP
and MMV-LASSO algorithms respectively.
For all three algorithms, the success rate of anomaly identification increases as
the number of measurements M increases and/or as the number of time-steps T
increases. A 100% success of identification is obtained with a sufficiently large
number of measurements and time-steps. There are some important differences in
performance among the three algorithms.
Firstly, for the OSGA and the MMV-SOMP algorithms, with a sufficiently large
number of time-steps, the minimum number of measurements at each time-step required for anomaly detection increases with the number of anomalies present. The
MMV-LASSO performance seems less affected by varying the number of anomalies than the performance of the other two algorithms. Secondly, comparing Fig. 2
and 3 reveals that MMV-SOMP requires fewer time-steps than the OSGA algorithm to reach 100% success for a given number of measurements. Thirdly, the
MMV-LASSO algorithm requires significantly fewer measurements and time-steps
for 100% success compared with OSGA and MMV-SOMP. Finally, there is asymmetry between the effect of increasing the number of measurements versus that of
increasing the number of time-steps on the performance of OSGA and MMV-SOMP.
For these two algorithms, increasing the number of measurements is more effective
than increasing the number of time-steps for improving the performance. No obvious asymmetry of recovery performance is found for the MMV-LASSO algorithm.
The near symmetry phenomenon of the MMV-LASSO is expected since doubling
either M or T doubles the number of rows in the matrix φ in Algorithm 3, providing
similar amounts of information for the algorithm.
K=5
K=10
10
20
20
20
30
30
30
40
40
40
50
50
50
m
10
m
m
K=1
10
60
60
60
70
70
70
80
80
80
90
90
90
100
10 20 30 40 50 60 70 80 90 100
t
100
10 20 30 40 50 60 70 80 90 100
t
100
1.0
0.8
0.6
0.4
0.2
10 20 30 40 50 60 70 80 90 100
0.0
t
Fig. 2 The recovery phase transition for the OSGA algorithm with K = 1, 5, and 10 anomalous
random variables.
Compressed Anomaly Detection with Multiple Mixed Observations
K=5
K=10
10
10
20
20
20
30
30
30
40
40
40
50
50
50
m
10
m
m
K=1
60
60
60
70
70
70
80
80
80
90
90
90
100
10 20 30 40 50 60 70 80 90 100
100
t
17
10 20 30 40 50 60 70 80 90 100
100
t
1.0
0.8
0.6
0.4
0.2
10 20 30 40 50 60 70 80 90 100
0.0
t
Fig. 3 The recovery phase transition for the MMV-SOMP algorithm with K = 1, 5, and 10 anomalous random variables.
K=5
K=10
10
20
20
20
30
30
30
40
40
40
50
50
50
m
10
m
m
K=1
10
60
60
60
70
70
70
80
80
80
90
90
90
100
10 20 30 40 50 60 70 80 90 100
t
100
10 20 30 40 50 60 70 80 90 100
t
100
1.0
0.8
0.6
0.4
0.2
10 20 30 40 50 60 70 80 90 100
0.0
t
Fig. 4 The recovery phase transition for the MMV-LASSO algorithm with K = 1, 5, and 10 anomalous random variables.
3.2 JSM-3R
We next present the results of recovering the anomalous index set for the JSM3R signals. Similar to JSM-2R signals, the length of the signal is set to N = 100
and the number of anomalies takes values of K = 1, 5, and 10. Unlike the JSM2R signals, the N − K random variables now follow the distribution N (7, 1) while
the K anomalous random variables follow the distribution N (0, 10) or N (7, 10).
In order for a fair comparison between the algorithms, we implement the OSGA
algorithm for both step 6 of the TECC algorithm and step 10 of the ACIE algorithm.
The iteration L in the ACIE algorithm is set to L = 5. The performance of the TECC
and ACIE algorithms for varying numbers of measurements M and time-steps T
when the anomalous distribution follows N (0, 10) is presented in Figs. 5 and 6,
where both M and T range from 1 to 100. The performance for the setting where
the anomalous variables are distributed as N (7, 10) is similar to Figs. 5 and 6 and
is thus omitted.
With a sufficiently large number of measurements and time-steps, both algorithms are able to achieve 100% success in recovery of the anomalous index set. For
a fixed number of time-steps, the minimum number of measurements required for
identification increases as the number of anomalies increases for both algorithms.
There is improvement in performance of the ACIE algorithm over the TECC al-
18
Durgin, Grotheer, Huang, Li, Ma, Needell, Qin
gorithm. The ACIE algorithm requires fewer time-steps to reach 100% recovery
success, for a given number of measurements; similarly, it requires fewer measurements for 100% recovery success with a given number of time-steps.
K=5
K=10
10
20
20
20
30
30
30
40
40
40
50
50
50
m
10
m
m
K=1
10
60
60
60
70
70
70
80
80
80
90
90
90
100
10 20 30 40 50 60 70 80 90 100
100
t
10 20 30 40 50 60 70 80 90 100
100
t
1.0
0.8
0.6
0.4
0.2
10 20 30 40 50 60 70 80 90 100
0.0
t
Fig. 5 The recovery phase transition for the TECC algorithm with K = 1, 5, and 10 anomalous
random variables. Here the prevalent distribution is N (7, 1) and the anomalous distribution is
N (0, 10).
K=5
K=10
10
20
20
20
30
30
30
40
40
40
50
50
50
m
10
m
m
K=1
10
60
60
60
70
70
70
80
80
80
90
90
90
100
10 20 30 40 50 60 70 80 90 100
t
100
10 20 30 40 50 60 70 80 90 100
t
100
1.0
0.8
0.6
0.4
0.2
10 20 30 40 50 60 70 80 90 100
0.0
t
Fig. 6 The recovery phase transition for the ACIE algorithm with K = 1, 5, and 10 anomalous
random variables. Here the prevalent distribution is N (7, 1) and the anomalous distribution is
N (0, 10).
4 Conclusion
In this paper, we formally posed the problem of detecting anomalously distributed
random variables as an MMV problem, by drawing an analogy between samples
of the random variables and ensembles of signals. We further established two signal models characterizing possible correlation structures among signals that contain
anomalous entries. Based on the new signal models, we showed through theoretical
and numerical analysis that many of the MMV algorithms for sparse signal recovery can be adapted to the anomaly detection problem. For two of the algorithms, we
Compressed Anomaly Detection with Multiple Mixed Observations
19
provided theoretical guarantees of anomaly detection in the asymptotic case. Our
experimental results on synthetic data show good performance for signals conforming to either model, when a sufficiently large number of time-steps is available.
While these algorithms succeed in detecting anomalies, there is still room for
optimizing performance. Currently these algorithms require storing the sensing matrices at each time-step in memory. In future work, we would like to explore optimal
ways to design sensing matrices to reduce the memory burden. Having provided
asymptotic anomaly detection guarantees for two algorithms, we are further interested in providing such guarantees for all the algorithms presented. Additionally, we
are interested in characterizing the performance bounds for each algorithm’s finite
sample case.
Acknowledgements The initial research for this effort was conducted at the Research Collaboration Workshop for Women in Data Science and Mathematics, July 17-21 held at ICERM. Funding
for the workshop was provided by ICERM, AWM and DIMACS (NSF grant CCF-1144502). SL
was supported by NSF CAREER grant CCF−1149225. DN was partially supported by the Alfred
P. Sloan Foundation, NSF CAREER #1348721, and NSF BIGDATA #1740325. JQ was supported
by the faculty start-up fund of Montana State University.
20
Durgin, Grotheer, Huang, Li, Ma, Needell, Qin
Appendix
Here we provide a summary of notation for reference.
N Number of random variables
N Set of random variables indices, {n ∈ N : 1 ≤ n ≤ N}
K Number of anomalous random variables
K Set of anomalous random variable indices, K ⊂ N, |K|= K
M Number of measurements per time-step
m Measurement index, 1 ≤ m ≤ M
T Number of time-steps measured
t Time-step index, 1 ≤ t ≤ T
D1 Prevalent distribution
D2 Anomalous distribution
X Random vector comprising independent random variables X1 , . . . , XN
x N × T -dimensional matrix of independent realizations of X for all T time-steps
Φ M × N-dimensional sensing matrix, i.i.d. ∼ N (0, 1) entries
φt M × N-dimensional realization of Φ at time t
φ (M · T ) × N-dimensional vertical concatenation of the φt , [φ1T , . . . , φTT ]T
yt M-dimensional result of measuring the signal, φt · x(,t) , at time t
T T
y (M · T )-dimensional vertical concatenation of the yt , [yT
1 , . . . , yT ]
Y M-dimensional random vector defined by ΦX
JSM Joint Sparsity Model, introduced in [2]
JSM-2 Signals are nonzero only on a common set of indices
JSM-3 Signals consist of common non-sparse component and a sparse innovation
JSM-2R “Random variable” version of JSM-2
JSM-3R “Random variable” version of JSM-3
OSGA One-step greedy algorithm
MMV Multiple Measurement Vector
MMV-LASSO MMV Least Absolute Shrinkage and Selection Operator
MMV-SOMP MMV Simultaneous Orthogonal Matching Pursuit
TECC Transpose Estimation of Common Component
ACIE Alternating Common and Innovation Estimation
We adopt the convention that random variables will be upper case and their realizations will be lower case. All matrix entries will have two, subscripted indices. The
first index will indicate the row position, the second will indicate the column position. We will indicate row and column vectors by substituting for the respective
index.
Compressed Anomaly Detection with Multiple Mixed Observations
21
References
[1] Angelosante D, Giannakis GB, Grossi E (2009) Compressed sensing of timevarying signals. In: Digital Signal Processing, 2009 16th International Conference on, IEEE, pp 1–8
[2] Baron D, Wakin MB, Duarte MF, Sarvotham S, Baraniuk RG (2006) Distributed compressed sensing. IEEE Trans Info Theory 52(12):5406–5425
[3] Baron D, Sarvotham S, Baraniuk RG (2010) Bayesian compressive sensing
via belief propagation. IEEE Trans Signal Process 58(1):269–280
[4] Basseville M, Nikiforov IV, et al (1993) Detection of abrupt changes: theory
and application, vol 104. Prentice Hall Englewood Cliffs
[5] Berg E, Friedlander MP (2009) Joint-sparse recovery from multiple measurements. Arxiv preprint arXiv:09042051
[6] Blumensath T, Davies ME (2009) Iterative hard thresholding for compressed
sensing. Appl Comput Harmon Anal 27(3):265–274
[7] Candès EJ (2006) Compressive sampling. In: Proc. Int. Congress of Mathematicians, Madrid, Spain, vol 3, pp 1433–1452
[8] Candès EJ, Tao T (2005) Decoding by linear programming. IEEE Trans Info
Theory 51:4203–4215
[9] Candès EJ, Tao T (2006) Near optimal signal recovery from random projections: Universal encoding strategies? IEEE Trans Info Theory 52(12):5406–
5425
[10] Candès EJ, Romberg J, Tao T (2006) Stable signal recovery from incomplete
and inaccurate measurements. Commun Pur Appl Math 59(8):1207–1223
[11] Chen J, Huo X (2006) Theoretical results on sparse representations of multiple
measurement vectors. IEEE Trans Signal Process 54(12):4634–4643
[12] Chen SS, Donoho DL, Saunders MA (2001) Atomic decomposition by basis
pursuit. SIAM Rev 43(1):129–159
[13] Cho M, Xu W, Lai L (2016) Compressed Hypothesis Testing: To Mix or Not
to Mix? arXiv preprint arXiv:160907528
[14] Cotter S, Rao B, Engan K, Kreutz-Delgado K (2005) Sparse solutions to linear inverse problems with multiple measurement vectors. IEEE Trans Signal
Process 53(7):2477–2488
[15] Donoho D (2006) Compressed sensing. IEEE Trans Info Theory 52(4):1289–
1306
[16] Donoho DL (2006) Compressed sensing. IEEE Trans Info Theory 52(4):1289–
1306
[17] Donoho DL, Huo X (2001) Uncertainty principles and ideal atomic decomposition. IEEE Trans Info Theory 47:2845–2862
[18] Duarte M, Wakin M, Baron D, Baraniuk R (2006) Universal distributed sensing via random projections. In: Proc. Inf. Process. Sensor Networks (IPSN)
[19] Duarte MF, Wakin MB, Baron D, Sarvotham S, Baraniuk RG (2013) Measurement bounds for sparse signal ensembles via graphical models. IEEE Trans
Info Theory 59(7):4280–4289
22
Durgin, Grotheer, Huang, Li, Ma, Needell, Qin
[20] Efron B, Hastie T, Johnstone I, Tibshirani R (2004) Least angle regression.
Annals of Statistics 32:407–499
[21] Eldar YC, Kutyniok G (2012) Compressed sensing: theory and applications.
Cambridge University Press
[22] Filos J, Karseras E, Dai W, Yan S (2013) Tracking dynamic sparse signals with
hierarchical kalman filters: a case study. In: Digital Signal Processing (DSP),
2013 18th International Conference on, IEEE, pp 1–6
[23] Foucart S, Rauhut H (2013) A mathematical introduction to compressive sensing. Birkhäuser Basel
[24] Haupt J, Nowak R (2006) Signal reconstruction from noisy random projections. IEEE Trans Info Theory 52(9):4036–4048
[25] Ji S, Xue Y, Carin L (2008) Bayesian compressive sensing. IEEE Trans Signal
Process 56(6):2346–2356
[26] Keerthi SS, Shevade S (2007) A fast tracking algorithm for generalized
lars/lasso. IEEE Transactions on Neural Networks 18(6):1826–1830
[27] Lai L, Fan Y, Poor HV (2008) Quickest detection in cognitive radio: A sequential change detection framework. In: Global Telecommunications Conference,
2008. IEEE GLOBECOM 2008. IEEE, IEEE, pp 1–5
[28] Mallat SG, Zhang Z (1993) Matching pursuits with time-frequency dictionaries. IEEE Transactions on Signal Processing 41(12):3397–3415
[29] Malloy M, Nowak R (2011) On the limits of sequential testing in high dimensions. In: 2011 Conference Record of the Forty Fifth Asilomar Conference on
Signals, Systems and Computers (ASILOMAR), pp 1245–1249
[30] Malloy M, Nowak R (2011) Sequential analysis in high-dimensional multiple testing and sparse recovery. In: 2011 IEEE International Symposium on
Information Theory Proceedings, pp 2661–2665
[31] Malloy ML, Tang G, Nowak RD (2012) Quickest search for a rare distribution. In: 2012 46th Annual Conference on Information Sciences and Systems
(CISS), pp 1–6
[32] Mishali M, Eldar YC (2009) Reduce and boost: recovering arbitrary sets of
jointly sparse vectors. IEEE Trans Signal Process 56(10):4692–4702
[33] Needell D, Vershynin R (2007) Signal recovery from incomplete and inaccurate measurements via Regularized Orthogonal Matching Pursuit. IEEE J Sel
Top Signa 4:310–316
[34] Pati YC, Rezaiifar R, Krishnaprasad PS (1993) Orthogonal matching pursuit:
recursive function approximation with applications to wavelet decomposition.
In: Proceedings of 27th Asilomar Conference on Signals, Systems and Computers, pp 40–44 vol.1
[35] Patterson S, Eldar YC, Keidar I (2014) Distributed compressed sensing for
static and time-varying networks. IEEE Trans Signal Process 62(19):4931–
4946
[36] Poor HV, Hadjiliadis O (2009) Quickest detection, vol 40. Cambridge University Press Cambridge
[37] Tropp JA (2006) Just relax: Convex programming methods for subset selection
and sparse approximation. IEEE Trans Info Theory 52(3):1030–1051
Compressed Anomaly Detection with Multiple Mixed Observations
23
[38] Tropp JA, Gilbert AC (2007) Signal recovery from random measurements via
Orthogonal Matching Pursuit. IEEE Trans Info Theory 53(12):4655–4666
[39] Tropp JA, Gilbert AC, Strauss M (2005) Simultaneous sparse approximation
via greedy pursuit. In: ICASSP
[40] Xia Y, Tse D (2006) Inference of link delay in communication networks. IEEE
J Sel Areas in Commun 24(12):2235–2248
[41] Yu G, Sapiro G (2011) Statistical compressed sensing of gaussian mixture
models. IEEE Trans Signal Process 59(12):5842–5858
| 7cs.IT
|
MIT CSAIL Technical Report MIT-LCS-TR-997
On Algorithms and Complexity for Sets with Cardinality
Constraints
PSPACE and PTIME Logics for Program Analysis
arXiv:cs/0508123v1 [cs.PL] 28 Aug 2005
Bruno Marnette
ENS de Cachan, MPRI
Paris, France
[email protected]
Abstract
Typestate systems ensure many desirable properties of imperative
programs, including initialization of object fields and correct use of
stateful library interfaces. Abstract sets with cardinality constraints
naturally generalize typestate properties: relationships between the
typestates of objects can be expressed as subset and disjointness
relations on sets, and elements of sets can be represented as sets
of cardinality one. In addition, sets with cardinality constraints
provide a natural language for specifying operations and invariants
of data structures.
Motivated by these program analysis applications, this paper
presents new algorithms and new complexity results for constraints
on sets and their cardinalities. We study several classes of constraints and demonstrate a trade-off between their expressive power
and their complexity.
Our first result concerns a quantifier-free fragment of Boolean
Algebra with Presburger Arithmetic. We give a nondeterministic
polynomial-time algorithm for reducing the satisfiability of sets
with symbolic cardinalities to constraints on constant cardinalities,
and give a polynomial-space algorithm for the resulting problem.
The best previously existing algorithm runs in exponential space
and nondeterministic exponential time.
In a quest for more efficient fragments, we identify several
subclasses of sets with cardinality constraints whose satisfiability is NP-hard. Finally, we identify a class of constraints that has
polynomial-time satisfiability and entailment problems and can
serve as a foundation for efficient program analysis. We give a system of rewriting rules for enforcing certain consistency properties
of these constraints and show how to extract complete information
from constraints in normal form. This result implies the soundness
and completeness of our algorithms.
1. Introduction
Program analyses that reason about deep semantic properties are of
great value for software development; the value of such analyses
is growing with the adoption of language constructs that eliminate
low-level program errors. Many deep semantic properties are naturally expressible in fragments of set theory, so constraint solving for
such fragments is of interest. This paper presents new algorithms
and improved complexity bounds for fragments of set theory. The
starting point of our constraints is the boolean algebra of finite (but
unbounded) sets.
Sets in program analysis. The boolean algebra of finite sets
is a fragment of set theory that allows the basic set operations
of intersection, union, and complement on sets of uninterpreted
elements. Although simple, it turns out that this fragment can
On Algorithms and Complexity for Sets with Cardinality Constraints
Viktor Kuncak
Martin Rinard
MIT Computer Science and Artificial Intelligence Lab
Cambridge, USA
{vkuncak,rinard}@csail.mit.edu
express many properties of interest in program analysis. Examples
include typestate properties and public interfaces of data structures.
Set specifications generalize typestate properties [29, 26]: the
fact that an object o is in the typestate t is represented as the set
membership of o in t. Through inclusion and disjointness constraints, sets can also express relationships (such as hierarchy or
orthogonality) between different typestates. Objects can be represented as sets of cardinality one using a cardinality constraint
|o| = 1, so set membership reduces to subset. Multiple set memberships can then encode constraints such as |t| ≥ k for any constant
k.
Sets can also provide natural abstractions of container data
structures. When a content of a data structure is represented as an
abstract set s, an operation such as insertion can be characterized
by a postcondition s′ = s ∪ e where e is the set corresponding
to the element being inserted. By expressing both typestates and
data structure abstractions, sets can be used to combine the results
of different analyses operating on the same program. Such an
approach allows us to combine the scalability of typestate analysis
with the precision of shape analysis and theorem proving [30, 28,
27, 46].
Sets with cardinality constraints. The use of the cardinality operator on sets leads to a connection between set algebra operations and integer linear arithmetic, as evidenced, for example, in
the condition |a ∪ b| = |a| + |b| for disjoint sets a and b. It is
therefore natural to consider constraints that combine integer linear
arithmetic with set algebra operations. These constraints constitute
the Quantifier-Free Boolean Algebra with Presburger Arithmetic,
or QFBAPA for short — they are the quantifier-free fragment of
BAPA constraints whose decision procedure and complexity we
have studied in [23, 22]. QFBAPA constraints can be used to verify an invariant such as |a| = |b| which allows us to conclude that
if a is nonempty, so is b, and therefore it is possible to call an operation that removes an element from b. Similarly, if i is an integer
variable and s is a set, it is possible to verify an invariant |s| = i
stating that an integer i correctly maintains the size of the set s.
In our experience, specialized decision procedures such as [22] are
the only automated technique for deciding with non-trivial cardinality constraints. Currently, however, the complexity of these decision procedures limits their applicability. In this paper we give
new algorithms for solving set cardinality constraints; these algorithms provide exponential improvements over existing approaches
and make the checking of cardinality constraints in larger formulas
more feasible.
Our paper provides a systematic study of constraints on sets in
the presence of cardinalities. We study both more expressive and
less expressive fragments and demonstrate a trade-off between the
1
2018/2/26
F
::=
A | F1 ∧ F2 | F1 ∨ F2 | ¬F
A
::=
B1 = B2 | B1 ⊆ B2 | T1 = T2 | T1 ≤ T2 | K dvd T
B
::=
s | 0 | 1 | B1 ∪ B2 | B1 ∩ B2 | B c
T
::=
i | K | T1 + T2 | K · T | |B|
K
::=
. . . | −2 | −1 | 0 | 1 | 2 | . . .
Figure 1. Quantifier-Free Formulas of Boolean Algebra with Presburger Arithmetic (QFBAPA)
expressive power and the efficiency of the algorithms. The main
contributions of our paper are the following:
• PSPACE algorithm for QFBAPA. The best previously
known algorithms for QFBAPA [23, 22, 45] execute in nondeterministic exponential time, and involve searching for an exponentially large object. In this paper we first give a form of
bounded model property that shows that it is possible to replace
reasoning about symbolic cardinalities such as |a| = i ∧ |b| = i
where i is an integer variable, with guessing sufficiently large
constant cardinalities, such as |a| = 1000 ∧ |b| = 1000. Moreover, we give a space-efficient algorithm for solving the resulting constraints on sets with large constant cardinalities. This
gives a PSPACE decision procedure for QFBAPA and is the
first contribution of this paper.
• A Polynomial-Time Class. Given that QFBAPA constraints
are NP-hard, the question remains whether there are interesting fragments of sets with cardinalities which can be reasoned
about in polynomial time. In a quest for such fragments, we
identify several features of constraints, each of which leads to
NP-hardness. By eliminating these features we have discovered
a class (called i-trees) that has a polynomial-time satisfiability
and entailment (subsumption) problems, while still supporting
subset, union, disjointness, and arbitrarily large cardinality constraints. This class can therefore express generalized typestate
constraints such as multiple orthogonal classifications into independent or disjoint sets. The identification of this polynomialtime class, and the development of algorithms for testing the
satisfiability and subsumption of constraints in this class is the
second contribution of this paper. While the resulting algorithms are efficient, the proof of their completeness is somewhat lengthy, and involves characterizations of normal forms
of i-trees and the construction of models for i-trees in normal
form. We therefore only summarize the main ideas; we refer
the reader to the full version of the paper [32] for details. Additional proofs are also included in the Appendix.
We proceed by defining the fragment QFBAPA in Section 2. We
present a PSPACE algorithm for QFBAPA in Section 3, defining
the simpler CBAC constraints and identifying their NP-complete
fragment, CBASC constraints.
2. Constraints on Sets with Cardinalities
Boolean Algebra with Presburger Arithmetic. Figure 1 presents
the syntax of the constraints studied in this paper, we call these formulas Quantifier-Free Boolean Algebra with Presburger Arithmetic
(QFBAPA). QFBAPA constraints contain two kinds of values: integers and sets, each with corresponding applicable operations. The
sets are interpreted as subsets of some arbitrarily large finite set. s
denotes a set variable, i denotes an integer variable. The symbol
|B| denotes the cardinality of the set B and establishes the connection between set and integer terms. MAXC is a special free variable
denoting the size of the universal set. If b is a set, bc denotes its
complement. K dvd T denotes that K divides T . K denotes constants, encoded in binary: a constant k is encoded using O(log k)
On Algorithms and Complexity for Sets with Cardinality Constraints
bits. The symbol A in Figure 1 denotes atomic formulas; a literal is
an atomic formula or its negation.
A quantified version of this language (BAPA) is studied in
[23, 22]; where we give an algorithm that establishes a doubly exponential space upper bound on the complexity. Because quantified BAPA subsumes Presburger arithmetic, the doubly exponential
nondeterministic time lower bound [15] applies to BAPA as well.
Preliminaries. If S is a finite set, |S| denotes the number of
elements in S. A literal is an atomic formula or its negation. Z =
{. . . , −1, 0, 1, . . .} is the set of integers, N = {0, 1, . . .} is the
set of natural numbers. [a..b] denotes the set of integers {a, a +
1, . . . , b}. If f : A → B is a function and S ⊆ A, we define
f [S] = {f (a) | a ∈ S}.
If A is a set, the notation Ay has several potential meanings;
the specific meaning should be clear from the context. An for
n ∈ {1, 2, . . . , } is the set of vectors (a1 , . . . , an ) where aj ∈ A
for 1 ≤ j ≤ n, and Am,n is the set of matrices [apq ] with m
rows and n columns with elements apq for 1 ≤ p ≤ m and
1 ≤ q ≤ n. The expression Ac denotes the complement of the
set A. If α ∈ {0, 1}, then Aα denotes A for α = 1 and Ac for
α = 0.
The relation ≡ denotes the equality of the values of metavariables denoting syntactic objects, so if f1 and f2 are formulas, then
f1 ≡ f2 means that they are the same formula. In the context of
inclusion diagrams (Section 4), ≡ will denote the semantic equivalence of diagrams (we use = to denote the equality of diagrams).
3. A PSPACE Algorithm for QFBAPA
Verification conditions arising in program verification can often
be expressed using quantifier-free formulas, so it is natural to examine whether more efficient algorithms exist for QFBAPA constraints. When applied to QFBAPA formulas, existing algorithms
run in non-deterministic exponential time (NEXPTIME): the algorithm [45] requires nondeterministically guessing an exponentially large object, whereas the algorithm α from [22] produces an
exponentially large quantifier-free Presburger arithmetic formula.
The question arises whether there exist algorithms that avoid nondeterministically guessing exponentially large objects. We show
that this is indeed the case. Namely, we first show that Presburger
arithmetic formulas generated by the algorithm α from [22] can in
fact be solved in deterministic exponential time. Our result reduces
QFBAPA to a simpler system of CBAC constraints (shown in Figure 3), then applies a theorem by Papadimitriou [36] in a novel
way. This leads to a deterministic EXPTIME decision procedure
for QFBAPA satisfiability, which is an improvement on previously
existing algorithms. Nevertheless, the question arises whether it is
possible to avoid the construction of a non-deterministically large
system of equations. It turns out that this is indeed possible: we
present an alternating polynomial-time (and therefore, PSPACE)
algorithm for QFBAPA. Therefore, it is possible to solve QFBAPA
using solvers for quantified boolean formulas [9, 48, 37].
Figures 2 and 4 present our PSPACE algorithm for QFBAPA.
The algorithm has two phases.
In the first phase, the non-deterministic polynomial-time algorithm in Figure 2 reduces QFBAPA constraints to a simpler class
of constraints. We call these simpler constraints Conjunctions of
Boolean Algebra expressions with Cardinalities (CBAC). CBAC
constraints have a very simple syntactic structure (see Figure 3),
but capture the key difficulty in solving QFBAPA: the need to consider exponentially large cardinalities on exponentially many set
partitions.
In the second phase, the algorithm in Figure 4 checks the satisfiability of CBAC in alternating polynomial time and therefore in
polynomial space. The key insight behind our algorithm is that it
is possible to use a divide and conquer approach to avoid explicitly
representing all possible regions in the Venn diagram.
2
2018/2/26
Let f be the input QFBAPA formula.
1. Replace each Z-variable with a difference of two Nvariables:
C[i1 , . . . , in ] → C[i′1 − i′′1 , . . . , i′n − i′′n ]
i′1 , i′′1 , . . . , i′n , i′′n
are fresh N-variables
2. Ensure that all set algebra expressions appear within
cardinality constraints by normalizing with the following
rules:
C[b1 = b2 ] → C[b1 ⊆ b2 ∧ b2 ⊆ b1 ]
C[b1 ⊆ b2 ] → C[|b1 ∩
bc2 |
= 0]
3. Eliminate divisibility constraints:
C[k dvd t] → C[ki = t], i is fresh N-variable
4. Move all cardinality constraints to top level:
C[|b1 |, . . . , |bw |] → f1 ∧ f2
def
where f1 ≡ C[i1 , . . . , iw ]
def
f2 ≡ |1|=MAXC ∧
w
V
|bj |=ij
j=1
and i1 , . . . , iw are fresh N-variables; let m1 = w + 1;
5. Let p be a propositional formula such that
p(a1 , . . . , am0 ) ≡ f1 for atomic formulas a1 , . . . , am0 .
Nondeterministically select the truth value αj ∈ {0, 1}
for each atomic formula aj , so that p(a1 , . . . , am0 ) is
V0 αj
def m
true. Let f11 ≡
aj .
F
::=
|B|=K | F1 ∧ F2
B
::=
s | 0 | 1 | B1 ∪ B2 | B1 ∩ B2 | B c
K
::=
0 | 1 | 2 | ...
Figure 3. Conjunctions of Boolean Algebra expressions with Cardinalities (CBAC)
Given a CBAC constraint
m1
X
|bj | = kj
j=1
where the free set variables of b1 , . . . , bm1 are among s1 , . . . , sS ,
run CBAC-check([], d) with d = (k1 , . . . , km1 ).
proc CBAC-check([v1 , . . . , vn ], d) returns result
where v1 , . . . , vn , result ∈ {0, 1}; d ∈ Nm1
if (n < S) then
existentially choose d0 , d1 ∈ Nm1 such that d0 + d1 = d;
universally do
r1 = CBAC-check([v1 , . . . , vn , 0], d0 ) and
r2 = CBAC-check([v1 , . . . , vn , 1], d1 );
return r1 ∧ r2 ;
else
let pj = eval(bj , [s1 7→ v1 , . . . , sS 7→ vS ])
for all (1 ≤ j ≤ m1 );
J0 = {dj | pj = 0};
J1 = {dj | pj = 1};
return J0 ⊆ {0} ∧ |J1 | ≤ 1.
j=1
6. For each conjunct ¬(t1 =t2 ) in f11 , nondeterministically replace the conjunct with one of
the conjuncts (t1 + 1 ≤ t2 ) or (t2 + 1 ≤ t1 ).
7. Transform linear integer constraints to normal form:
C[¬(t1 ≤ t2 )] → C[t2 + 1 ≤ t1 ]
C[t1 ≤ t2 ]
C[t1 = t2 ]
→ C[t1 − t2 + i = 0]
P
→ C[ n
j=1 cj ij = k]
proc eval(b, α) returns result
where b : Boolean Algebra formula
α : {s1 , . . . , sS } → {0, 1}
result ∈ {0, 1}
treating b as a propositional formula,
return the value of b under assignment α.
Figure 4. An Alternating Polynomial-Time (and PSPACE) Algorithm for Checking the Satisfiability of CBAC Constraints
8. Let n0 be the number of N-variables in the entire formula. The resulting system is of the form:
V 1
Av = d ∧ m
j=1 |bj | = ipj
where A ∈ Zm0 ,n0 , d ∈ Zm0 , and v = (i1 , . . . , in0 )
where each ij is an N-variable and 1 ≤ p1 , . . . , pm1 ≤
m1 are variables denoting cardinalities of sets. Let S
be the total number of set variables in b1 , . . . , bm1 . Let
m = m0 + m1 , n = max(n0 , 2S ),
a = max({1} ∪ {|apq | | 1 ≤ p ≤ n0 , 1 ≤ q ≤ m0 }
∪ {|dq | | 1 ≤ q ≤ m0 })
where A = [apq ]pq , and d = (d1 , . . . , dm0 ), and let
M = n(ma)2m+1 .
9. Non-deterministically select a vector k = (k1 , . . . , kn0 )
where kj ∈ {0, 1, . . . , M } for 1 ≤ j ≤ n0 , such that
Ak = d.
m
V1
|bj | = kpj . If
10. Call CBAC decision procedure on
j=1
there exists a solution, then report the formula satisfiable.
Figure 2. An NP Algorithm for Reducing QFBAPA Constraints
to CBAC constraints of Figure 3
On Algorithms and Complexity for Sets with Cardinality Constraints
We next discuss our algorithm in more detail and argue that it is
correct. We begin with the description of the steps of the algorithm
in Figure 2, which reduces symbolic cardinalities to large constant
cardinalities.
1. Non-negative integers. To simplify the later steps, the first step
makes all integer variables range over non-negative integers N, by
replacing each integer variable i with a difference i1 − i2 of fresh
non-negative integer variables i1 , i2 .
2,3. Eliminating set equality and subset, and integer divisibility.
The next step converts set equality and set subset into cardinality
constraints. This step helps the later separation between the boolean
algebra part and the integer linear arithmetic part. We then eliminate any divisibility relations using multiplication and a fresh variable.
4. Flattening. The next step separates the formula into the
boolean algebra part, denoted f1 and the integer linear arithmetic
part, denoted f2 . This step simply amounts to naming the cardinality of each set by a fresh integer variable.
5,6. From quantifier-free formulas to conjunctions. An obvious
source of NP-completeness of QFBAPA is the presence of arbitrary
propositional combinations of atomic formulas. An effective way
of dealing with propositional combinations is to enumerate the
3
2018/2/26
satisfying assignments of the propositional formula using a SAT
solver, and then solve the conjunctions of literals [16, 17]. Steps 5
and 6 of the non-deterministic algorithm in Figure 2 are an abstract
description of such procedure. The goal of step 6 is to eliminate
disequalities, which involve non-deterministic choice between the
two inequalities.
7. Normal form for integer constraints. The algorithm eliminates the remaining negations of atomic formulas and transforms
linear constraints into normal form Av = d.
8,9,10. Estimating sizes of integer variables. The
P resulting system contains linear integer equations of the form n
j=1 cj ij = k,
and set cardinality constraints of the form |b| = i. The algorithm
computes an upper bound M on integer variables in any potential solution of the system, using several parameters: the number
of conjuncts n, the number of integer variables n0 and the number
of set variables S. The computation of the upper bound is based
on an observation that the satisfiability of the conjunction of constraints |b| =
P i can be reduced to the satisfiability of equations
of the form pj=1 lj = i, where variables lj denote sizes of set
partitions (regions in Venn diagram) whose union is the set b; this
is a specialization of the idea in [22] to the case of quantifier-free
formulas.
Let s1 , . . . , sS be all set variables appearing in formula
and
T
αj
consider a constraint |b| = i. Consider all partitions n
j=1 sj
for αj ∈ {0, 1}. For each such partition bp , introduce a fresh Nvariable lp , which denotes the cardinality of cube bp . Then consider
a constraint of the form |b| = i. Each set is a union of regions in the
Venn diagram (by the disjunctive normal form theorem) so suppose
that
Pa b = bp1 ∪ . . . ∪ bpa . Then replace the term |b| = i with the
q=1 lpq = i. We use the term “CBAC linear equations” to denote
a system of linear equations resulting from the constraints |b| = i
as described above.
As a result, we obtain a system of m0 + m1 linear equations
over non-negative integers, where m0 equations have a polynomial
number of variables, and m1 equations (CBAC linear equations)
have exponentially many variables. It is easy to see that there exists
a surjective mapping of solutions of the original constraints on
sets onto solutions of the resulting linear equations (the mapping
computes the cardinality of each Venn diagram). Therefore, the
original system is satisfiable if and only if the resulting equations
are satisfiable. Moreover, we have the following fact.
FACT 1 (Papadimitriou [36]). Let A be an m × n integer matrix
and b an m-vector, both with entries from [−a..a]. Then the system
Ax = b has a solution in Nm if and only if it has a solution in
[0..M ]m where M = n(ma)2m+1 .
Fact 1 implies that the estimate M computed in step 8 of the algorithm in Figure 2 is a correct upper bound. Using this estimate, step
9 of the algorithm non-deterministically guesses the values of all
integer variables such that the original linear equations Ax = d are
satisfied. All this computation can be performed in nondeterministic polynomial time, and (unlike [22]), does not involve constructing explicitly a system with exponentially many equations. Having
picked the values of integer variables, including the variables i on
the right hand side of constraints |b| = i, we obtain a conjunction
of constraints of the form |b| = k where k is a constant whose
binary representation has polynomially many bits—these are precisely the CBAC constraints in Figure 3. We have therefore shown
the following.
L EMMA 1. The algorithm in Figure 2 reduces in non-deterministic
polynomial time the satisfiability of a QFBAPA formula to the
satisfiability of CBAC formulas.
It remains to find an algorithm for CBAC constraints.
On Algorithms and Complexity for Sets with Cardinality Constraints
A PSPACE algorithm for CBAC. One correct way to solve
CBAC constraints is to solve the associated CBAC linear equations. This system has exponentially many variables, each of which
can take any value from [0..M ]. Therefore, guessing the values of
each of these variables can be done in non-deterministic exponential time; similar approaches not based on equations also require
guessing exponentially large objects [45]. Note, however, that there
are only polynomially many CBAC linear equations. Using the idea
of the proof [36, Corollary 1], we can therefore show that a dynamic
programming algorithm can be used to solve the system in polynomial time. In fact, we can use the dynamic programming algorithm
from the proof of [36, Corollary 1]. Instead of fixing the size of the
equations m1 to be constant, we simply observe that m1 is polynomial in the size of the input, whereas the number of variables
is singly exponential. The bound M therefore yields a singly exponential deterministic time dynamic programming algorithm for
CBAC. While this is better than existing results, we show that an
even better result is achievable.
Clearly, any algorithm that explicitly constructs CBAC equations will require at least exponential time and space. Our solution
is therefore to adapt the dynamic programming algorithm to a divide and conquer approach that always represents the equations in
terms of their original, polynomially sized, boolean algebra expression. Such an algorithm runs in alternating polynomial time, consuming polynomial space, and is presented in Figure 4. To see the
idea of our PSPACE algorithm, consider the CBAC linear system
P p
of equations written in the vector form: 2j=1 aj lj = d where d,
aj are vectors and lj are the variables for 1 ≤ j ≤ 2p . The algorithm guesses the vectors d0 , d1 ∈ Nm such that d0 + d1 = d, and
recursively solves two equations:
2p−1
X−1
p
a j lj = d 0
j=1
∧
2
X
a j lj = d 1
j=2p−1
This algorithm creates an OR-AND tree whose search gives the
answer to the original problem. A position in the tree is given by the
propositional assignment [v1 , . . . , vn ] to boolean variables. Each
leaf in the tree is given by a complete assignment [v1 , . . . , vS ]
to set variables. Note that we never need to explicitly maintain
the system during the divide phase of the algorithm, it suffices to
determine in the leaf case p = 0 whether the coefficient aj is 0
or 1. The algorithm does this by simply evaluating each Boolean
algebra expression b for the assignment [v1 , . . . , vS ].
T HEOREM 1. The algorithm in Figure 4 checks the satisfiability of
CBAC constraints in PSPACE. The algorithm given by Figures 2
and 4 checks the satisfiability of QFBAPA constraints in PSPACE.
Theorem 1 improves the existing algorithms for QFBAPA from
both a complexity theoretic and an implementation viewpoint. A
deterministic realization of previous NEXPTIME algorithms runs
in doubly exponential worst-case time and requires exponential
space; a deterministic realization of our new algorithm runs in
singly exponential time and consumes polynomial space. Previous
algorithms would require running a constraint solver such as a SAT
solver [47] on an exponentially large constraint; the new algorithm
can be solved by running a quantified boolean algebra solver [48]
on a polynomially large constraint.
NP fragments of CBAC. We have seen that both CBAC and
QFBAPA constraints are in PSPACE. Both of these classes of constraints are NP-hard, because the constraint |b| = 1 is satisfiable iff
b is corresponds to a satisfiable propositional formula. Moreover,
Lemma 1 shows that QFBAPA constraints are in NP iff CBAC constraints are in NP. For some subclasses of CBAC constraints we can
indeed show membership in NP. Define conjunctions of boolean algebra expressions with small cardinalities, denoted CBASC, to be
the same as CBAC but with constant integers encoded in unary
4
2018/2/26
notation, where an integer x is represented in space O(x) as opposed to O(log x); such encoding can therefore be exponentially
less compact.
L EMMA 2. The satisfiability of CBASC constraints is NPcomplete.
CBASC solutions are NP-hard because |b| = 1 is a CBASC constraint. One way to prove membership in NP is to observe that
CBASC is subsumed by the language of set-valued fields which
was proven to be in NP [24, 25] by reduction to the universal class
of first-order logic formulas, which has the small model property
[7, Page 258]. Another way is to consider the notion of sparse solutions of CBAC linear equations. An M -sparse solution is a solution
to CBAC linear constraints with at most M non-zero elements. An
M -sparse solution to CBAC linear constraints with 2S variables
can be encoded as an M -tuple of pairs ([v1 , . . . , vS ], k) where the
propositional assignment [v1 , . . . , vS ]; encodes one of the 2S integer variables, and k specifies the value of that integer variable.
This encoding is polynomial in M Sw where w is the number of
bits for representing the V
largest component of the solution. For any
CBAC linear constraint m
j=1 |bj | = kj , each solution is M -sparse
where M = max(k1 , . . . , km ). For CBASC solutions, M is polynomial in the size of the CBASC representation because each ki is
encoded in unary, so sparse solutions can be guessed in polynomial
time. This proves that CBASC constraints are in NP.1
4. Inclusion Diagrams
This section introduces inclusion diagrams (i-diagrams), a graph
representation of CBAC constraints. Figure 5 shows a formula with
sets and cardinalities and an equivalent i-diagram. I-diagrams allow
us to naturally describe fragments of CBAC constraints and the algorithms for checking satisfiability and subsumption of these fragments. The basic idea of i-diagrams is to represent the subset partial
order using a graph where sets are annotated with cardinalities, and
then indicate the disjointness and union relations by constraints on
direct subsets of a set. To efficiently represent equal sets, the nodes
in the i-diagram stand not for set names, but for collections of set
names that are guaranteed to be equal. Finally, we associate uninterpreted predicates with collections of nodes, representing the
fact that elements of given sets satisfy the properties given by the
predicate. The uninterpreted predicates illustrate a way to combine
i-diagram representations with other constraints.
D EFINITION 1 (i-diagrams). We fix a finite set SN of Set-Names,
and a finite set PN of predicate names. We denote by PN± the set
of atoms {+P, −P |P ∈ PN}.
An i-diagram (Inclusion-Diagram) is either the null-diagram ⊥d or
a tuple (S, ∅d , Sons, Split, Comp, CInf, CSup, Φ) such that:
• S ⊆ P(SN) is a partition of SN containing (nonempty) equiva-
lence classes of set names that are guaranteed to be equal, with
∅d ∈ S the equivalence class corresponding to names of sets
whose interpretation is the empty set ∅;
• Sons : S → P(S) represents subset relation;
def
we define S
S ′ ⇐⇒ S ∈ Sons(S ′ ); then (S, ) is a
graph, so we call elements of S nodes, and the elements of
∗
edges; we write
for the transitive closure of ;
• Split, Comp : S → P(P(S)) represent disjointness and completeness of set inclusions; if S is a node, then Split(S) is a set
of split views, where each view is a nonempty set of sons that
1 Sparse solutions are interesting for general CBAC constraints as well. As
of yet we have no example of a CBAC constraint whose associated CBAC
equation system is satisfiable but has no sparse solutions; moreover, we can
generalize the notion of sparse solutions to solutions representable using
binary decision diagrams [8] while preserving polynomial-time verifiability.
On Algorithms and Complexity for Sets with Cardinality Constraints
{s1}
[1..5]
+P
d{s2}
[0..0]
{s5,s6}
[0..3]
{s4}
[1..1]
{s3}
[0..2]
−Q
D is such that CInf({s1 }) = 1 , CSup({s1 }) = 5, Sons({s1 }) =
{{s5 , s6 }, {s4 }, {s3 }}, Comp({s1 }) = {{{s5 , s6 }, {s4 }, {s3 }}}
Split({s1 }) = {{{s5 , s6 }}, {{s4 }, {s3 }}}, Φ({s1 }) = {+P }
and is equivalent to
s2 = ∅ ∧ s5 = s6 ∧
s2 ∪ s3 ∪ s4 ∪ s5 ⊆ s1 ∧ s1 ⊆ s5 ∧ s3 ⊆ s2 ∧ s5 ⊆ s4
s3 ∩ s4 = ∅ ∧ s1 ⊆ s3 ∪ s4 ∪ s5 ∧ s4 ⊆ s5
1 ≤ |s1 | ≤ 5 ∧ |s4 | = 1 ∧ |s5 | ≤ 3 ∧ |s3 | ≤ 2 ∧
∀x ∈ s1 . P (x) ∧ ∀x ∈ s3 . ¬Q(x)
Figure 5. An example i-diagram D and an equivalent formula
represent pairwise disjoint sets, and Comp(S) is a set of complete views each of which is a set of nodes that represent sets
whose union is equal to the father; we require
S
S Split(S) = Sons(S)
Comp(S) ⊆ Sons(S)
for all S ∈ S;
• CInf, CSup : S → N specify lower and upper bounds on the
cardinality of sets;
• Φ : S → P(PN± ) maps nodes to the uninterpreted unary
predicates and their negations that are true for all sets of a
node.
To avoid confusion between set names, nodes (sets of set names),
and views (sets of nodes), we use lowercase letters s, si , s′ to
denote set names, uppercase letters S, Si , S ′ to denote nodes,
and letters Q, C to denote views and sets of nodes in general. When D 6=⊥d is a diagram, unless otherwise stated, we
name its components S, ∅d , Sons, Split, Comp, CInf, CSup, Φ,
and similarly we name the components of D′ as
S′ , ∅′d , Sons′ , Split′ , Comp′ , CInf ′ , CSup′ , Φ′ .
In a graphical representation of an i-diagram, we represent
each element S ∈ S where S = {s1 , . . . , sn } using underlying sets {s1 , . . . , sn }. We represent inclusion S1
S2 by an
arrow from S1 to S2 . We represent a split view Q ∈ Split(S)
where Q = {S1 , . . . , Sn } with a circle connected with undirected
edges to S1 , . . . , Sn and an arrow leading to S. We represent a
complete view similarly, using a filled square instead of a circle.
For each node S ∈ S we indicate its cardinality bounds by annotating the node with [a..b] where a = CInf(S), b = CSup(S).
We represent Ξ(S) = {±P1 , . . . , ±Pn } by annotating S with
±P1 , . . . , ±Pn . We represent ∅d = {s1 , . . . , sn } by annotating
the node {s1 , . . . , sn } with ∅d .
D EFINITION 2 (Semantics of i-diagrams). An interpretation of SN
and PN is a triple (∆, α, Ξ) where
• ∆ is a finite set (the universe);
• α : SN → P(∆) specifies the values of sets;
• Ξ : PN → P(∆) specifies the values of unary predicates;
An interpretation I is a model for an i-diagram D, denoted I |= D,
iff ∀s ∈ ∅d .α(s) = ∅, and for all S ∈ S where S = {s1 , . . . , sn },
the following conditions hold:
5
2018/2/26
• α(s1 ) = . . . = α(sn );
def
•
•
•
•
•
•
accordingly, define α(S) = α(s1 ) = . . . = α(sn )
CInf(S) ≤ |α(S)| ≤ CSup(S)
∀P. (+P ) ∈ Φ(S) ⇒ α(S) ⊆ Ξ(P )
∀P. (−P ) ∈ Φ(S) ⇒ α(S) ⊆ Ξ(P )c
∀S ′ ∈ Sons(S). α(S ′ ) ⊆ α(S)
∀Q ∈ Split(S). ∀S1 , S2 ∈ Q. S1 6= S2 ⇒ α(S1 )∩α(S2 ) = ∅
S
∀Q ∈ Comp(S). α(S) ⊆ S1 ∈Q α(S1 )
We use the standard notions of satisfiability, subsumption (entailment), and equivalence:
D is satisfiable
D′ |= D
D′ ≡ D
⇐⇒
⇐⇒
⇐⇒
∃I. I |= D
∀I. I |= D′ ⇒ I |= D
D′ |= D ∧ D |= D′
D EFINITION 3 (Explicit Disjointness).
We write disjD,S0 (S1 , S2 ) as a shorthand for
S1 6= S2 ∧ ∃Q ∈ Sons(S0 ). S1 , S2 ∈ Q
and we say that S1 , S2 are explicitly disjoint, and we write
disj∗D (S1 , S2 ) iff
∃S1′ , S2′ , S0 ∈ S, S1
∗
S1′ ∧ S2
∗
S2′ ∧ disjD,S0 (S1′ , S2′ )
L EMMA 3. I-diagrams have the same expressive power as CBAC
constraints.
By “same expressive power” we here mean that there is a natural
pair of mappings between the models of i-diagrams and solutions
to CBAC constraints.
Because nodes in i-diagrams are collections of set names, we
can define the following operations.
D EFINITION 4 (Factor-i-diagram). Let ρ ⊆ S × S be an equivalence relation on nodes. We define D/ρ as follows. Define
⊥d /ρ =⊥d . Let D = (S, ∅d , Sons, Split, Comp, CInf, CSup, Φ).
We define D/ρ = D′ = (S′ , Sons′ , Split′ , Comp′ , CInf ′ , CSup′ ,
Φ′ ) as follows. Define h so that if {S1 , . . . , Sn } is the equivalence
class of S under ρ, then h(S) = S1 ∪ . . . ∪ Sn . If Q ⊆ S, define
h[Q] = {h(S) | S ∈ Q}. Then let S′ = h[S]. Consider S ′ ∈ S′ .
Both S and S′ are partitions, and given S ′ ∈ S′ there is a unique
set {S1 , . . . , Sn } ⊆ S such that S ′ = S1 ∪ . . . ∪ Sn . Then define:
CInf ′ (S ′ ) = max(CInf(S1 ), . . . , CInf(Sn ))
CSup′ (S ′ ) = min(CSup(S1 ), . . . , CSup(Sn ))
Sons′ (S ′ ) = h[Sons(S1 ) ∪ . . . ∪ Sons(Sn )]
Φ′ (S ′ ) = Φ(S1 ) ∪ . . . ∪ Φ(Sn )
Split′ (S ′ ) = {h[Q] | Q ∈ Split(S1 ) ∪ . . . ∪ Split(Sn )}
Comp′ (S ′ ) = {h[Q] | Q ∈ Comp(S1 ) ∪ . . . ∪ Comp(Sn )}
D EFINITION 5 (Merge). For any i-diagram D we define the idef
diagram D[Merge(Q)] = D/ρ for the equivalence relation ρ =
{(S1 , S2 ) | S1 , S2 ∈ Q} ∪ {(S, S) | S ∈ S}
In the sequel we impose the following restrictions on the form
of i-diagrams.
D EFINITION 6 (Simple Diagrams). A diagram is D is simple iff
D = ∅d or all of the following conditions hold for all S ∈ S:
a)
b)
c)
d)
e)
f)
(S, ) has no cycles, in particular S 6∈ Sons(S)
∅D 6∈ Sons(S)
∅ 6∈ Split(S) ∧ ∅ 6∈ Comp(S)
∀Q, Q′ . Q ∈ Split(S) ∧ Q′ ( Q ⇒ Q′ 6∈ Split(S)
∀Q, Q′ . Q ∈ Comp(S) ∧ Q′ ) Q ⇒ Q′ 6∈ Comp(S)
CSup(∅d ) = 0, Sons(∅d )=Φ(∅d ) = ∅
On Algorithms and Complexity for Sets with Cardinality Constraints
proc Simplify(D) :
1. use fixpoint iteration to compute ρ as
the smallest equivalence relation such that:
∗
∗
1.1. S1 S2 ∧ S2 S1 ⇒ (S1 , S2 ) ∈ ρ
1.2. (S, ∅d ) ∈ ρ ∧ S1
∗
S ⇒ (S1 , ∅d ) ∈ ρ
1.3. ∅ ∈ Comp(S) ⇒ (S, ∅d ) ∈ ρ
1.4. disjD,S0 (S1 , S2 ) ∧ (S1 , S2 ) ∈ ρ ⇒ (S2 , ∅d ) ∈ ρ
1.5. disjD,S0 (S1 , S2 ) ∧ (S0 , S1 ) ∈ ρ ⇒ (S2 , ∅d ) ∈ ρ
1.6. Q∈Comp(S1 ) ∧ (∀S∈Q.(S, ∅d )∈ρ) ⇒ (S1 , ∅d ) ∈ ρ
2. D := D/ρ
Split(S) ←{Q − {∅d }|Q ∈ Split(S), S ∈
/ Q}
Comp(S)←{Q
−
{∅
}|Q
∈
Comp(S),
S
∈
/
Q}
d
3.
Sons(S) ←Sons(S) − {∅d , S}
S∈S
Split(S) ← Split(S) − {∅}
−{Q | ∃Q′ ∈ Split(S), Q′ ) Q}
4.
Comp(S) ← ′Comp(S) − {∅}′
−{Q | ∃Q ∈ Comp(S), Q ( Q}
S∈S
CSup(∅d ) ← 0
Φ(∅d )
←∅
5. Sons(∅d ) ← ∅
Comp(∅d ) ← ∅
Split(∅d ) ← ∅
6. return D
Where [a ← b] denotes the result of updating the component a of
i-diagram D with value b.
Figure 6. Polynomial-time algorithm Simplify to compute an
equivalent simple i-diagram
Simplicity eliminates redundancy from diagrams, but does not restrict their expressive power, as the following lemma shows.
L EMMA 4. For every i-diagram D we can obtain an equivalent
simple i-diagram using the polynomial-time algorithm Simplify
in Figure 6.
5. Sources of NP Hardness and Definition of
I-Trees
The satisfiability of i-diagrams is NP-hard because i-diagrams have
the same expressive power as CBAC constraints. We have observed
that the general directed acyclic graph structure of i-diagrams allows us to encode NP-complete problems; this motives the following two restrictions.
D EFINITION 7.
An i-diagram D is tree shaped iff
(S, ) is a tree (with an additional isolated node ∅d )
An i-diagram D has independent views iff
for all Q1 , Q2 ∈ Split(S) ∪ Comp(S) at least one of the following
two conditions holds:
• Q1 ∩ Q2 = ∅
• Q1 ∈ Split(S) ∧ Q2 ∈ Comp(S) ∧ Q1 ⊆ Q2 .
Recall that, by Lemma 4, it suffices to consider i-diagrams with
acyclic graphs of the subset relation. The tree shape condition is
then a natural next restriction on the structure of i-diagrams. However, due to the presence of Split and Comp, the tree shape condition by itself does not reduce the expressive power of i-diagrams,
and further restrictions are necessary. The independent views con-
6
2018/2/26
dition extends the tree condition to the entire graphical representation of i-diagrams, including the circles and squares that represent
Split and Comp views. The conjunction of these two conditions
can be expressed by saying that the graphical representation of idiagram is a tree.
R EMARK 1. We can express the combination of the conditions:
being simple, being tree shaped, and having independent views by
saying that there are only four kinds of edges in the corresponding
graphical representation:2
• from an element S ∈ S−{∅d } to a circle
• from a circle to a square, indicating that all nodes of a split view
belong to a complete view
• from a circle to an element S ∈ S−{∅d },
• from a square to an element S ∈ S−{∅d }.
Unfortunately, the restrictions on tree shape and independent views
are not sufficient to guarantee a polynomial-time decision procedure in the presence of predicates associated with nodes. The reason is that the ability to encode disjointness of arbitrary sets leads to
NP-hardness, yet even with tree structure and independent views it
is possible to assert that two arbitrary sets S1 and S2 are disjoint by
letting (+P ) ∈ Φ(S1 ) and (−P ) ∈ Φ(S2 ) for some uninterpreted
predicate P . A simple way to avoid this problem is to require that
Φ contains only positive atoms (+P ). A more flexible restriction
is the following.
D EFINITION 8. An i-diagram D has independent signatures iff
for every pair of distinct nodes S1 , S2 such that (−P ) ∈ Φ(S1 )
and (+P ) ∈ Φ(S2 ) for some P ∈ PN, at least one of the following
two conditions holds:
1. S1 and S2 are explicitly disjoint, that is, disj∗D (S1 , S2 )
2. S1 and S2 have compatible signatures, that is, there exists a
node S such that
∗
∗
S1 S ∧ S2 S ∧
Sig(S1 ) ∩ Sig(S2 ) ⊆ Sig(S)
where Sig(S) = {P | (+P ) ∈ Φ(S) ∨ (−P ) ∈ Φ(S)}.
The independent signatures condition ensures that any disjointness
conditions are either 1) a result of the fact that the ancestors of
S1 and S2 are explicitly stated as disjoint, or 2) a result of a
contradictory predicate assignment (the case when S1 and S2 have
compatible signatures, so there exists a parent that resolves which
of (+P ) or (−P ) hold for both S1 and S2 ).
The discussion above leads to the definition of i-trees, for which
we will give polynomial-time algorithms for satisfiability and subsumption in Sections 6 and 7.
D EFINITION 9 (i-trees, iT). An i-tree T is a simple i-diagram
such that T =⊥d or such that all of the following three conditions
hold:
1. T is tree shaped
2. T has independent views
3. T has independent signatures.
We denote by iT the set of i-trees.
The following theorem justifies why all three conditions in our definition of i-trees are necessary. Its proof is based on a reduction
from graph 3-colorability, which can be encoded using slightly different i-diagrams for each of the three cases. The common property
of these diagrams is that they can encode disjointness of arbitrary
pairs of nodes.
2 As a result, we can recognize this structure in linear time using, for
example, a tree-automaton [12].
On Algorithms and Complexity for Sets with Cardinality Constraints
T HEOREM 2. Omitting any one out of three conditions from Definition 9 yields a class of diagrams whose satisfiability is NP-hard.
We note that in addition to NP-hardness, the omission of tree
shaped or independent views properties in fact retains the full
expressive power of CBAC constraints, using a similar argument
as in Lemma 3.
Our ability to specify i-trees as a natural subclass of i-diagrams
justifies the definition of i-diagrams themselves. For example, the
definition of i-trees would have been more complex had we chosen
to represent disjointness using a binary relation s1 ∩ s2 = ∅.
Let us also observe that, despite the imposed restrictions, itrees are fairly expressive. In particular they can express hierarchical decomposition of a set given by a node S into disjoint sets
S1 , . . . , Sn , by letting {S1 , . . . , Sn } ∈ Split(S) ∩ Comp(S). Despite the independent view condition, we can have multiple orthog′
} ∈ Split(S) ∩ Comp(S) for
onal decompositions, so {S1′ , . . . , Sm
′
′
{S1 , . . . , Sm }∩{S1 , . . . , Sn } = ∅. This allows i-trees to naturally
express generalized typestate constraints.
6. Deciding the Satisfiability of I-Trees
In this section we prove that the satisfiability of i-trees is decidable
in polynomial time. For this purpose we introduce a set of weak
consistency conditions Ci (Definition 10) such that:
(6.1) We can enforce weak consistency for any satisfiable i-tree using a rewriting system Rw (Definition 11) with the following
properties (Lemma 5):
• Rw is semantic-preserving;
• if a non-⊥d i-tree is in Rw normal form, then it satisfies
weak consistency conditions;
• for a particular strategy (Figure 9) the system Rw termi-
nates in polynomial time.
(6.2) Every i-tree that satisfies weak consistency conditions is satisfiable; Lemma 6 gives an algorithm for constructing a model for
any i-tree that satisfies weak consistency conditions.
Figure 9 summarizes the polynomial-time satisfiability decision
procedure whose correctness (Theorem 3) follows from the results
of this section.
D EFINITION 10 (Weak Consistency). An i-tree satisfies weak
consistency iff T =
6 ⊥d and T satisfies the following conditions
for all S ∈ S:
∀S ′ ∈ Sons(S). Φ(S ′ ) ⊇ Φ(S)
(C1 )
CSup(S) > 0 ⇒ ∀P ∈ PN. {+P, −P } 6⊆ Φ(S) (C2 )
∀Q ∈ Comp(S). CSup(S) ≤ Σ(CSup[Q])
(C3 )
∀Q ∈ Split(S). CInf(S) ≥ Σ(CInf[Q])
(C4 )
CInf(S) ≤ CSup(S)
(C5 )
w
6.1 A Rewriting System R for Enforcing Weak Consistency
We introduce the following rewriting system to enforce weak consistency properties when possible.
D EFINITION 11 (System Rw ). For each tuple (k, name,
condition, effect) in Figure 7, we define a rewriting rule on
i-diagrams by
spot
def
D −→ D′ ⇐⇒ (D =
6 ⊥d ∧ condition ∧ D′ = D[effect])
name
for each assignment spot of the free variables appearing in the
condition column. We define Rk by
def
spot
D −→ D′ ⇐⇒ ∃spot. D −→ D′
name
Rk
We define Rw as union of −→ for 1 ≤ j ≤ 5.
Rj
7
2018/2/26
k
name
1
DnPhi
2
Unsat
3
UpSup
4
UpInf
5
Error
conditions
a1 ) S ∈ Sons(S ′ )
.
b1 ) φn = Φ(S) ∪ Φ(S ′ )
c1 ) φn 6⊆ Φ(S)
a2 ) {+P, −P } ⊆ Φ(S)
.
b2 ) n = 0
c2 ) CSup(S) > n
a3 ) Q ∈ Comp(S)
.
b3 ) n = Σ(CSup[Q])
c3 ) CSup(S) > n
a4 ) Q ∈ Split(S)
.
b4 ) n = Σ(CInf[Q])
c4 ) CInf(S) < n
a5 ) CInf(S) > CSup(S)
proc RwNF (T )
1. for every S ∈ S from the root to the leaves
for every Q ∈ Comp(S)
try to apply DnPhi(S, Q) to T
2.for every S ∈ S
for every P ∈ PN
try to apply Unsat(S, P ) to T
3. for every S ∈ S from the leaves to the root
for every Q ∈ Comp(S)
try to apply UpSup(S, Q) to T
4. for every S ∈ S from the leaves to the root
for every Q ∈ Split(S)
try to apply UpInf (S, Q) to T
5. for every S ∈ S
try to apply Error(S) to T
return T
effect
Φ(S) ← φn
CSup(S) ← n
CSup(S) ← n
CInf(S) ← n
D ←⊥d
Figure 7. System Rw for ensuring weak consistency
A
[3..2]
+P
A
[3..5]
+P
A
[3..5]
+P
A
[3..5]
+P
Figure 9. Polynomial-time algorithms RwNF and ItreeSAT to
compute Rw normal form and check satisfiability of i-trees
B
[4..5]
B
[4..5]
B
[4..5]
B
[4..5]
C
[2..2]
C
[2..2]
C
[2..2]
C
[2..2]
T0
D,A
−→
DnPhi
T1
D
−→
Unsat
T2
d
fied direction (from the root to the leaves or from the leaves
to the root), enforces Ck everywhere, and when Ck holds, the
rule is not applicable anymore. Finally, we prove that each
rule
V Rk for k = 1..5 preserves the conjunction of properties
j=1..(k−1) Rk , and as a consequence, we never need to reapply any of the rules Rj for j < k.
D
[0..0]
+P−P
D
[0..0]
+P−P
D
[0..6]
+P−P
D
[0..6]
−P
proc ItreeSAT(T )
if (RwNF (T ) =⊥d ) return satisfiable
else return unsatisfiable
A,{C,D}
−→
UpSup
T3
A
−→
Error
⊥d
Figure 8. An example sequence of rewriting steps for Rw
Figure 8 shows an example sequence of rewriting steps applied to
an i-tree.
L EMMA 5 (Properties of Rw ).
1. Rw is iT-stable, that is
T ∈ iT ∧ T −→
T ′ ⇒ T ′ ∈ iT
w
R
2. Rw preserves the semantics, that is
D −→
D′ ⇒ D ≡ D′
w
R
3. Rw enforces weak consistency when possible, that is, a diagram D in Rw normal form is either equal to ⊥d or it is weakly
consistent
4. Rw terminates in polynomial time for the strategy corresponding to the algorithm RwNF in Figure 9.
Proof sketch.
1. Follows easily from the fact that Rw rules do not modify Sons,
Split, Comp.
2. Follows by construction of Rw rules. Suppose D −→
D′ . Then
w
R
D |= D′ follows from conditions ai (1≤i≤5), and D′ |= D
follows from conditions ci (1≤i≤4).
3. For every k = 1..5, the condition of application of the rule Rk
corresponds to the negation of Ck . When a diagram is in normal
form for the rule Rk , it either satisfies Ck , or is ⊥d .
4. To prove that RwNF corresponds to a polynomial strategy, we
prove by induction that applying the rule Rk in the speci-
On Algorithms and Complexity for Sets with Cardinality Constraints
6.2 Constructing Models for Weakly Consistent I-Trees
The following Lemma 6 is crucial for the completeness of our
algorithm, and justifies the definition of weak consistency.
L EMMA 6 (Model Construction). If an i-tree T is weakly consistent, then we can construct a model for T .
The high-level idea of the proof of Lemma 6 is to first build the first
two components (∆, α) of the model, and then extend the model
with Ξ using the independent signatures condition for i-trees. We
build the (∆, α) part of the model by building a model for each
subtree using an induction on the height of the i-tree. To construct
models that satisfy Split and Comp constraints in the inductive
step, we use a stronger induction hypothesis: we show that there
exists a model (∆, α) for a tree rooted in node S with |∆| = k
for all CInf(S) ≤ k ≤ CSup(S), and we rely on the properties
of weak consistency to prove the inductive step. The proof of this
lemma is interesting because similar ideas are used when building
example models that show the completeness in Section 7.
Putting all results in this section together using the argument at
the beginning of the section, we obtain the following theorem.
T HEOREM 3 (ItreeSAT Correctness). T is satisfiable if and only
if RwNF (T ) 6=⊥d . Therefore, the algorithm ItreeSAT in Figure 9 is
a sound and complete polynomial-time decision procedure for the
satisfiability of i-trees.
7. Deciding Subsumption of I-Trees
The goal of this section is to prove that we can decide the subsumption of i-trees in polynomial time. Note that the subclass of i-trees
is not closed under negation or implication, so we cannot decide
T |= T ′ by checking the satisfiability of ¬(T ⇒ T ′ ). Instead,
our approach is to bring T into a form where the properties of the
models of T are easy to read from T . We then check that T entails each of the conditions that correspond to the semantics of T ′ .
8
2018/2/26
k
condition
a6 ) ({S} ⊎ Q0 ) ∈ Comp(S ′ )
.
b6 ) n = CInf(S ′ )
6 DnInf
−Σ(CSup[Q0 ])
c6 ) n > CInf(S)
a7 ) ({S} ⊎ Q0 ) ∈ Split(S ′ )
.
b ) n = CSup(S ′ )
7 DnSup 7
−Σ(CInf[Q0 ])
c7 ) n < CSup(S)
a8 ) Q ∈ Split(S) ∧
CSup(S) ≤ Σ(CInf[Q])
∗
.
8 CCmp
b8 ) Cn = Comp(S)∪{Q}
c8 ) Cn 6⊆ Comp(S)
a9 ) Q ∈ Comp(S) ∧
CInf(S) ≥ Σ(CSup[Q])
∗
.
9 CSplit
b9 ) Cn = Split(S)∪{Q}
c9 ) Cn 6⊆ Split(S)
a10 ) Q ∈ Comp(S)T
.
10 UpPhi b10 ) φn = Φ(S) ∪ Φ[Q]
c10 ) φn 6⊆ Φ(S)
a11 ) S 6= ∅d ∧ CSup(S) = 0
11 Void∗
12 Equal∗ a12 ) {S ′ } ∈ Comp(S)
∗
name
effect
CInf(S) ← n
CSup(S) ← n
Comp(S) ← Cn
Split(S) ← Cn
Φ(S) ← φn
Merge({S, ∅d })
Merge({S, S ′ })
Follow the application of these rules by Simplify.
proc RNF (T )
1 . . . 5. T ← RwNF (T )
6. for each S ′ ∈ S from the root to the leaves
for each Q ∈ Comp(S)
for every S ∈ Q
try DnInf (S, S ′ , Q)
7. for each S ′ ∈ S from the root to the leaves
for each Q ∈ Split(S)
for each S ∈ Q
try DnSup(S, S ′ , Q)
8. for each S ∈ S
for each Q ∈ Split(S)
try CCmp(S, Q)
9. for each S ∈ S
for each Q ∈ Comp(S)
try CSplit(S, Q)
10. for each S ∈ S from the leaves to the root
for each Q ∈ Comp(S)
try UpPhi(S, Q)
11. for each S ∈ S from the leaves to the root
try Void(S)
12. for each S ∈ S
for each Q ∈ Comp(S)
try Equal(S, Q)
return T
Figure 10. Rules for System R
We formalize the intuitive condition of being easy to read in the
notion of strong consistency. We build on the system Rw from the
previous section to create a larger rewriting system R for ensuring
strong consistency. We introduce a polynomial-time strategy for R
that transforms every i-tree into ⊥d or into an i-tree that is strongly
consistent, and we give polynomial-time algorithms for extracting
the information from strongly consistent i-trees.
D EFINITION 12 (Strong Consistency). An i-tree T is strongly
consistent iff it is weakly consistent and satisfies all of the following
properties:
∀Q ∈ Comp(S). ∀S0 ∈ Q.
CInf(S0 ) ≥ CInf(S) − Σ(CSup[Q − {S0 }])
(C6 )
∀Q ∈ Split(S). ∀S0 ∈ Q.
CSup(S0 ) ≤ CSup(S) − Σ(CInf[Q − {S0 }])
(C7 )
∀Q ∈ Split(S). Q 6∈ Comp(S) ⇒
CSup(S) > Σ(CInf[Q])
(C8 )
∀Q ∈ Comp(S). Q 6∈ Split(S) ⇒
CInf(S) < Σ(CSup[Q])
T
∀Q ∈ Comp(S). (Φ[Q]) ⊆ Φ(S)
(C9 )
(C10 )
S 6= ∅D ⇒ CSup(S) > 0
(C11 )
Q ∈ Comp(S) ⇒ |Q| > 1
(C12 )
7.1 A rewriting system R to enforce strong consistency
This section follows the development of Section 6.1.
D EFINITION 13 (System R). The system R extends Rw with the
additional rules of Figure 10, analogously to Definition 11.
L EMMA 7 (Properties of R). 1. R is iT-stable, that is
T ∈ iT ∧ T −→ T ′ ⇒ T ′ ∈ iT
R
2. R preserves the semantics, that is
D −→ D′ ⇒ D ≡ D′
R
On Algorithms and Complexity for Sets with Cardinality Constraints
Figure 11. Polynomial-time algorithm RNF (T ) to compute R
normal form
3. R enforces strong consistency when possible, that is, a diagram
D in R normal form is either equal to ⊥d or it is strongly
consistent.
4. R terminates in polynomial time for the strategy corresponding
to the algorithm RNF described in Figure 11.
Proof sketch.
1. The iT-stability is trivial for the rules DnInf , DnSup,
UpPhi. The other rules are marked with a star and we use the
algorithm Simplify. In fact, we can show that it is not necessary
to apply Simplify in its full generality, but only to remove any
redundant views introduced by CCmp and CSplit, remove
any self edges introduced by the operation Merge used in the
rules Equal and Void, and to remove the edges going to ∅d
that can be introduced by the rule Void.
2,3. Follow by construction as in the previous section.
4. This part is significantly more difficult than for system Rw ,
because the interactions between the rules are more complex,
but follows the same structure as the proof for Rw .
7.2 Extracting Information from Strongly Consistent I-Trees
In this section we start from a strongly consistent i-tree T and consider the problem of checking T |= D′ . Analyzing Definition 2,
we observe that a diagram corresponds to a conjunction of constraints. Therefore, the subsumption problem T |= D′ corresponds
to the problem of verifying that T entails atomic formulas of the
form s = ∅, s1 = s2 ,Ss1 ⊆ s2 , a ≤ |s| ≤ b, s ⊆ P , s ⊆ P c ,
s1 ∩ s2 = ∅ and s ⊆ {s1 , . . . , sn }. Without the danger of confusion, we write T |= A when the atomic formula A holds in all
models for T .
T HEOREM 4. Let T be a strongly consistent i-tree and let HTA for
atomic formula A be as defined in Figure 12. Then T |= A if and
only if HTA .
9
2018/2/26
proc Subsumes(T , D′ )
T := RNF (T )
let f : SN → S such that ∀s ∈ SN. s ∈ f (s)
let h′ : S′ → SN be any function such that ∀S ′ ∈ S. h′ (S ′ ) ∈ S ′
check all
conditions:
V of the
V following
1.
HTs1 =s2
S∈S′ s1 ,s2 ∈S
def
2.
3.
where HTs1 =s2 ⇐⇒ f (s1 ) = f (s2 )
HTh(∅′ )=∅
d
def
T
where
V THs=∅ ⇐⇒ f (s) = ∅d
HCInf ′ (S)≤|h(S)|≤CSup′ (S)
S∈S′
S∈S′ (−P )∈Φ′ (S)
L EMMA 9 (Parallel Bounds Refinement). Let T be a strongly consistent i-tree, and (Q0 , ) a subtree of T such that
def
6.
where HT+P (s) ⇐⇒ (+P ) ∈ Φ(f (s))
V
V
HT−P (h(S))
where
V
HT−P (s)
V
S∈S′ S ′ ∈Sons′ (S)
7.
def
⇐⇒ (−P ) ∈ Φ(f (s))
HTh(S)⊆h(S ′ )
def
• The nodes of Q0 are pairwise independent, that is,
∗
T
where
⇐⇒Vf (s1 ) = ∅d ∨ f (s1 ) f (s2 )
1 ⊆s2
V HsV
HTh(S1 )∩h(S2 )=∅
S∈S′ Q∈Split′ (S)
S1 ,S2 ∈Q
S1 6=S2
def
where HTs1 ∩s2 =∅ ⇐⇒
8.
2. Follows easily from the hypothesis CInf(S) ≤ i ≤ s ≤
CSup(S) and the fact that RwNF is semantics preserving.
3. It is enough to notice that only rules UpInf and UpSup are used
when applying RwNF , and these rules are applied in the bottomup direction.
′
The fact that the resulting i-tree TNF
is not strongly consistent anymore prevents us to apply this lemma twice from a given strongly
consistent i-tree. To enforce more than one restriction, we need to
refine simultaneously the bounds of several nodes. For this purpose,
we use the following lemma.
S∈S′ (+P )∈Φ′ (S)
5.
(a) both UpSup and UpInf are applicable; then the result follows from C5 ;
(b) only UpSup is applicable; then the result follows from C6 ;
(c) only UpInf is applicable; then the result follows from C7 .
where HTa≤|s|≤b ⇐⇒ CInf(f (s)) ≤ a ≤ b ≤ CSup(f (s))
V
V
HT+P (h(S))
def
4.
and only decreases CSup. At each step we distinguish three
cases:
V
V
f (s1 ) = ∅d ∨ f (s2 ) = ∅d ∨
disj∗T (S1 , S2 )
HTh(S)⊆∪h[Q]
S∈S′ Q∈Comp′ (S)
def
where HTs⊆∪Z ⇐⇒ f (s) = ∅d ∨ Included(f (s), f [Z], T )
where proc Included(S
W 0 , C, T )
return
Incl(S)
S0
∗
S
proc Incl(S)
if S ∈ C then return true
V
W
Incl(S ′ )
else return
Q∈Comp(S)
S ′ ∈Q
Figure 12. An Algorithm for Computing T |= D′ for a an i-tree
T and an arbitrary diagram D′ .
It is easy to verify that HTA implies T |= A. The proof of the
converse is based on the following two lemmas, which provide a
link between strong and weak consistency.
L EMMA 8 (Bounds Refinement). Let T be a strongly consistent itree, S ∈ S, i, s such that CInf(S) ≤ i ≤ s ≤ CSup(S), let
′
T ′ = T [CInf(S) ← i, CSup(S) ← s] and TNF
= RwNF (T ′ ). Then
∗
′
′
1) TNF 6=⊥d , 2) TNF |= T , and 3) if ¬(S S0 ), then
(CInf(S0 ), CSup(S0 )) = (CInf ′NF (S0 ), CSup′NF (S0 )).
Proof sketch.
1. We prove this result by induction on the depth of S in the tree
(S, ). The key step of this proof is to show that the application
of UpSup and/or UpInf to the father S ′ of S does not produce
a situation where a5 holds in the resulting diagram T ′′ (and
therefore the rule Error is not applicable in T ′′ ). We use the fact
that RwNF applies the rules UpInf and UpSup bottom up, and
prove that each application preserves C5 , only increases CInf
On Algorithms and Complexity for Sets with Cardinality Constraints
∀S1 , S2 ∈ Q0 . ¬(disj∗T (S1 , S2 ))
• (Q0 , ) has the same root as T .
Then the i-tree T ′ defined by the simultaneous update
def
T ′ = T [ ∀S ∈ Q0 :CInf(S)←CSup(S) ]
def
′
is such that its Rw normal form TNF
= RwNF (T ′ ) satisfies
′
1. TNF
6 ⊥d
=
′
2. TNF
|= T
Lemmas 8 and 9 are the basic tools we need to show that
the information syntactically computed from an i-tree is the most
precise information computable from the semantics of the i-tree.
We prove this property for each of the atomic formulas A.
L EMMA 10. If an i-tree T is strongly consistent, then for all S ∈ S
we have
S 6= ∅d ⇒ ∃M. M |= T ∧ αM (S) 6= ∅
Proof. If S 6= ∅d , we have CSup(S) > 0 by C11 and therefore
def
the i-tree T ′ = T [CInf(S)← max(1, CInf(S))] subsumes T . By
′
Lemma 8, T is satisfiable, and we can take any model of T ′ as a
model of T .
L EMMA 11. If an i-tree T is strongly consistent, then for all S ∈ S
we have
∃M. M |= T ∧ |αM (S)| = CInf(S)
∃M. M |= T ∧ |αM (S)| = CSup(S)
Proof. According
to
Lemma
8,
the
two
i-trees
′ def
′ def
T1 = T [CSup(S)←CInf(S)] and T2 = T [CInf(S)←CSup(S)]
are satisfiable, and both T1′ and T2′ trivially subsume T . Any model
M1 of T1′ is such that |αM1 (S)| = CInf(S), and any model M2
of T2′ is such that |αM2 (S)| = CSup(S).
L EMMA 12. If an i-tree T is strongly consistent, S0 ∈ S, C ∈
P(S), and Included(S0 , C, T ) returns false, then
[
αM [C]
∃M. M |= T ∧ αM (S0 ) 6⊆
Proof sketch. Assume that Included returns false. We argue
that the model M exists in several steps. Let Q0 be the smallest
10
2018/2/26
set of nodes such that:
∗
S0 S ⇒ S ∈ Q0
S ∈ Q0 ∧ C1 ∈ Comp(S)
⇒ S1 ∈ Q 0
∧ S1 ∈ C1 ∧ ¬Incl(S1 )
By definition of Incl, we have Q0 ∩ C = ∅.
Q0 is tree-shaped by construction, but may contain two nodes
which are explicitly disjoint. We therefore compute a subtree Q1 of
Q0 , by starting from the root and keeping at most one son for each
complete view. In this process we ensure that Q1 contains S0 , by
avoiding to cut the branch which leads to S0 .
We then apply Lemma 9 to Q1 and construct a model of the
resulting i-tree while enforcing that a certain element x ∈ α(S) is
such that x ∈ α(S ′ ) ⇔ S ′ ∈ Q1 for all S ′ ∈ S. More precisely,
we prove by induction on n that for each node S1 of Q1 of depth n
in the tree Q1 we can construct a model (∆, α, Ξ) for the sub-i-tree
of T with root S1 such that
∀S ′ . S ′
∗
S1 ⇒ (x ∈ α(S ′ ) ⇔ S ′ ∈ Q1 )
If n = 0 then S1 is a leaf of (Q1 , ) and has no complete view by
construction of Q1 . Then using C8 we show that we can construct
a model of the sub-i-tree with root S1 containing a fresh element
(not included in any of the sons of S1 ).
If n > 0, we can deal with the split views in the same way, but
this time S1 can have some complete views. If this complete view
contains a unique split view, we avoid merging x in a son of S1 with
elements in the other sons of S1 . If there exist more than one split
view, we can use C8 and construct the model using a refinement of
the ideas of Lemma 6.
Finally, since S0 ∈ Q
S1 , we have x ∈ α(S0 ), and since
C ∩ Q1 = ∅ we have x ∈
/ α[C].
L EMMA 13. If an i-tree T is strongly consistent, for all S1 , S2 ∈ S
such that S1 6= ∅ and S2 6= ∅ we have
¬(S1
∗
S2 ) ⇒ ∃M. M |= T ∧ αM (S1 ) 6⊆ αM (S2 )
Proof. The property ∃M. M |= T ∧ αM (S1 ) 6⊆ αM (S2 ) can
be checked using Included(S1 , {S2 }, T ). Using C12 we show
∗
that this test is equivalent to test S1 S2 .
L EMMA 14. If an i-tree T is strongly consistent, for all S1 , S2 ∈ S
we have
S1 6= S2 ⇒ ∃M. M |= T ∧ αM (S1 ) 6= αM (S2 )
∗
Proof. If S1 6= S2 , then ¬(S1 S2 ) or ¬(S2
case the result follows from Lemma 13.
∗
S1 ). In either
L EMMA 15. If an i-tree T is strongly consistent, then for all S ∈ S
and P ∈ PN we have
(+P ) 6∈ Φ(S) ⇒ ∃M. M |= T ∧ αM (S) 6⊆ Ξ(P )
(−P ) 6∈ Φ(S) ⇒ ∃M. M |= T ∧ αM (S) 6⊆ Ξ(P )c
Proof sketch. Let S ∈ S be such that (+P ) 6∈ Φ(S). We define
def
QP = {S ′ ∈ S|(+P ) ∈ Φ(S ′ )}. Using C10 and C1 we show that
Included(S, QP , T ) returns
S false. By Lemma 12, there exists a
We then change the model by
model such that α(S) 6⊆ α[QP ].S
redefining Ξ′ on PN as Ξ′ (P ) =
α[QP ], so α(S) 6⊆ Ξ′ (P ).
The case (−P ) 6∈ Φ(S) is dual and follows from the previous case
by swapping (+P ) and (−P ) in the i-tree and taking complements
of Ξ(P ).
L EMMA 16. If an i-tree T is strongly consistent, then for all
S1 , S2 ∈ S such that S1 6= ∅d , S2 6= ∅d we have
¬(disj∗T (S1 , S2 )) ⇒ ∃M. M |= T ∧ αM (S1 ) ∩ αM (S2 ) 6= ∅.
On Algorithms and Complexity for Sets with Cardinality Constraints
From the previously stated lemmas, we can prove Theorem 4.
From Theorem 4 and Lemma 7 we conclude that the algorithm in
Figure 12 is a correct and complete test for subsumption, not only
of between trees, but also between a tree and an arbitrary diagram.
8. Related Work
Boolean algebras with cardinalities. Quantifier-free formulas of
boolean algebra are NP-complete [33]. Quantified formulas of
boolean algebra are in alternating exponential space with a linear number of alternations [21]. Cardinality constraints naturally
arise in quantifier elimination for boolean algebras [31, 42, 43].
Quantifier elimination implies that each first-order formula of the
language of boolean algebras is equivalent to some quantifier-free
formula with constant cardinalities; however, quantifier elimination may introduce an exponential blowup. The first-order theory
of boolean algebras of finite sets with symbolic cardinalities, or,
equivalently, boolean algebras of sets with equicardinality operator
is shown decidable in [14]. These results are repeated, motivated
by constraint solving applications, in [23, 39] and a special case
with quantification over elements only is presented in [44]. Upper
and lower bounds on the complexity of this problem were shown in
[22] which also introduces the name BAPA, for Boolean Algebra
with Presburger Arithmetic. The quantifier-free case of BAPA was
studied in [45] with an NEXPTIME decision procedure, which is
also achieved as a special case of [23, 22]. The new decision procedure in the present paper improves this bound to PSPACE and
gives insight into the problem by reducing it to boolean algebras
with binary-encoded large cardinalities, and showing that it is not
necessary to explicitly construct all set partitions.
Several decidable fragments of set theory are studied in [10].
Cardinality constraints also occur in description logics [5] and
two-variable logic with counting [35, 19, 38]. However, all logics
of counting that we are aware of have complexity that is beyond
PSPACE.
We are not aware of any previously known fragments of boolean
algebras of sets with cardinality constraints that have polynomialtime satisfiability or subsumption algorithms. Our polynomial-time
result for i-trees is even more interesting in the light of the fact that
our constraints can express some “disjunction-like” properties such
as A = B ∪ C.
Set constraints. Set constraints [1, 3, 2, 6] are incomparable to
the constraints studied in our paper. On the one hand, set constraints
are interpreted over ground terms and contain operations that apply a given free function symbol to each element of the set, which
makes them suitable for encoding type inference [4] and interprocedural analysis [20, 34]. Researchers have also explored the efficient
computation of the subset relation for set constraints [13]. On the
other hand, set constraints do not support cardinality operators that
are useful in modelling databases [40, 11] and analysis of the sizes
of data structures [29]. Tarskian constraints use uninterpreted function symbols instead of free function symbols and have very high
complexity [18].
9. Conclusions
Constraints on sets and relations are very useful for analysis of software artifacts and their abstractions. Reasoning about sets and relations often involves reasoning about their sizes. For example, an
integer field may be used to track the size of the set of objects stored
in a data structure. In this paper, we have presented new complexity
results and algorithms for solving constraints on boolean algebra
of sets with symbolic and constant cardinality constraints. We have
presented symbolic constraints and large constant constraints, gave
more efficient algorithm for quantifier-free symbolic constraints,
identified several sources of NP-hardness of constraints, and presented a new class of constraints for which satisfiability and en-
11
2018/2/26
tailment are solvable in polynomial time. We hope that our results
will serve as concrete recipes and general guidance in the design of
algorithms for constraint solving and program analysis.
References
[1] A. Aiken. Introduction to set constraint-based program analysis.
Science of Computer Programming, 35:79–111, 1999.
[2] A. Aiken, D. Kozen, M. Vardi, and E. Wimmers. The complexity
of set constraints. In Proceedings of Computer Science Logic 1993,
pages 1–17, September 1993.
[3] A. Aiken, D. Kozen, and E. Wimmers. Decidability of systems of set
constraints with negative constraints. Information and Computation,
122, 1995.
[4] A. Aiken and E. L. Wimmers. Type inclusion constraints and type
inference. In Proc. 7th ACM Conference on Functional Programming
and Computer Architecture, Copenhagen, Denmark, 1993.
[5] F. Baader, D. Calvanese, D. McGuinness, D. Nardi, and P. PatelSchneider, editors. The Description Logic Handbook: Theory,
Implementation and Applications. CUP, 2003.
[6] L. Bachmair, H. Ganzinger, and U. Waldmann. Set constraints are the
monadic class. In Logic in Computer Science, pages 75–83, 1993.
[7] E. Börger, E. Grädel, and Y. Gurevich. The Classical Decision
Problem. Springer-Verlag, 1997.
[8] R. E. Bryant. Graph-based algorithms for boolean function
manipulation. IEEE Transactions on Computers, C-35(8):677–691,
August 1986.
[9] M. Cadoli, M. Schaerf, A. Giovanardi, and M. Giovanardi. An algorithm to evaluate quantified boolean formulae and its experimental
evaluation. J. Autom. Reasoning, 28(2):101–142, 2002.
[10] D. Cantone, E. Omodeo, and A. Policriti. Set Theory for Computing.
Springer, 2001.
[11] P. P.-S. Chen. The entity-relationship model–toward a unified view
of data. ACM Transactions on Database Systems (TODS), 1(1):9–36,
1976.
[12] H. Comon, M. Dauchet, R. Gilleron, F. Jacquemard, D. Lugiez, S. Tison, and M. Tommasi. Tree automata techniques and applications.
Available on: http://www.grappa.univ-lille3.fr/tata,
1997. release 1999.
[13] M. Fähndrich, J. Foster, Z. Su, and A. Aiken. Partial online cycle
elimination in inclusion constraint graphs. In PLDI’98, 1998.
[14] S. Feferman and R. L. Vaught. The first order properties of products
of algebraic systems. Fundamenta Mathematicae, 47:57–103, 1959.
[15] M. J. Fisher and M. O. Rabin. Super-exponential complexity of
presburger arithmetic. SIAM-AMS Proceedings, 7, 1974.
[16] C. Flanagan, R. Joshi, X. Ou, and J. B. Saxe. Theorem proving using
lazy proof explication. In CAV, pages 355–367, 2003.
[17] H. Ganzinger, G. Hagen, R. Nieuwenhuis, A. Oliveras, and C. Tinelli.
DPLL(T): Fast decision procedures. In R. Alur and D. Peled, editors,
16th CAV, volume 3114 of LNCS, pages 175–188. Springer, 2004.
[18] R. Givan, D. McAllester, C. Witty, and D. Kozen. Tarskian set
constraints. Inf. Comput., 174(2):105–131, 2002.
[19] E. Grädel, M. Otto, and E. Rosen. Two-variable logic with counting
is decidable. In Proceedings of 12th IEEE Symposium on Logic in
Computer Science LICS ‘97, Warschau, 1997.
[20] J. Kodumal and A. Aiken. The set constraint/CFL reachability
connection in practice. In Proc. ACM PLDI, June 2004.
[21] D. Kozen. Complexity of boolean algebras. Theoretical Computer
Science, 10:221–247, 1980.
[22] V. Kuncak, H. H. Nguyen, and M. Rinard. An algorithm for
deciding BAPA: Boolean Algebra with Presburger Arithmetic. In
20th International Conference on Automated Deduction, CADE-20,
Tallinn, Estonia, July 2005.
[23] V. Kuncak and M. Rinard. The first-order theory of sets with
cardinality constraints is decidable. Technical Report 958, MIT
CSAIL, July 2004.
[24] V. Kuncak and M. Rinard. On decision procedures for set-valued
fields. Technical Report 975, MIT CSAIL, November 2004.
On Algorithms and Complexity for Sets with Cardinality Constraints
[25] V. Kuncak and M. Rinard. Decision procedures for set-valued fields.
In 1st International Workshop on Abstract Interpretation of ObjectOriented Languages (AIOOL 2005), 2005.
[26] P. Lam, V. Kuncak, and M. Rinard. Generalized typestate checking
using set interfaces and pluggable analyses. SIGPLAN Notices,
39:46–55, March 2004.
[27] P. Lam, V. Kuncak, and M. Rinard. On our experience with modular
pluggable analyses. Technical Report 965, MIT CSAIL, September
2004.
[28] P. Lam, V. Kuncak, and M. Rinard. Cross-cutting techniques in
program specification and analysis. In 4th International Conference
on Aspect-Oriented Software Development (AOSD’05), 2005.
[29] P. Lam, V. Kuncak, and M. Rinard. Generalized typestate checking
for data structure consistency. In 6th International Conference on
Verification, Model Checking and Abstract Interpretation, 2005.
[30] P. Lam, V. Kuncak, and M. Rinard. Hob: A tool for verifying data
structure consistency. In 14th International Conference on Compiler
Construction (tool demo), April 2005.
[31] L. Loewenheim. Über mögligkeiten im relativkalkül. Math. Annalen,
76:228–251, 1915.
[32] B. Marnette, V. Kuncak, and M. Rinard. On algorithms and
complexity for sets with cardinality constraints. Technical report,
MIT CSAIL, August 2005.
[33] K. Marriott and M. Odersky. Negative boolean constraints. Technical
Report 94/203, Monash University, August 1994.
[34] D. Melski and T. Reps. Interconvertibility of a class of set constraints
and context-free language reachability. TCS, 248:29–98, November
2000.
[35] L. Pacholski, W. Szwast, and L. Tendera. Complexity results for
first-order two-variable logic with counting. SIAM J. on Computing,
29(4):1083–1117, 2000.
[36] C. H. Papadimitriou. On the complexity of integer programming. J.
ACM, 28(4):765–768, 1981.
[37] D. A. Plaisted, A. Biere, and Y. Zhu. A satisfiability procedure for
quantified boolean formulae. Discrete Appl. Math., 130(2):291–328,
2003.
[38] I. Pratt-Hartmann. Counting quantifiers and the stellar fragment.
http://www.cs.man.ac.uk/˜ipratt/papers/logic/f2.ps, 2004.
[39] P. Revesz. Quantifier-elimination for the first-order theory of boolean
algebras with linear cardinality constraints. In Proc. Advances in
Databases and Information Systems (ADBIS’04), volume 3255 of
LNCS, September 2004.
[40] J. Rumbaugh, I. Jacobson, and G. Booch. The Unified Modelling
Language Reference Manual. Addison-Wesley, Reading, Mass.,
1999.
[41] M. Sipser. Introduction to the Theory of Computation. PWS
Publishing Company, 1997.
[42] T. Skolem. Untersuchungen über die Axiome des Klassenkalküls and
über “Produktations- und Summationsprobleme”, welche gewisse
Klassen von Aussagen betreffen. Skrifter utgit av Vidnskapsselskapet
i Kristiania, I. klasse, no. 3, Oslo, 1919.
[43] A. Tarski. Arithmetical classes and types of boolean algebras. Bull.
Amer. Math. Soc., 55, 64, 1192, 1949.
[44] C. G. Zarba. A quantifier elimination algorithm for a fragment of
set theory involving the cardinality operator. In 18th International
Workshop on Unification, 2004.
[45] C. G. Zarba. Combining sets with cardinals. Journal of Automated
Reasoning, 34(1), 2005.
[46] K. Zee, P. Lam, V. Kuncak, and M. Rinard. Combining theorem
proving with static analysis for data structure consistency. In
International Workshop on Software Verification and Validation (SVV
2004), Seattle, November 2004.
[47] L. Zhang, C. F. Madigan, M. W. Moskewicz, and S. Malik. Efficient
conflict driven learning in boolean satisfiability solver. In ICCAD,
pages 279–285, 2001.
[48] L. Zhang and S. Malik. Conflict driven learning in a quantified
boolean satisfiability solver. In ICCAD ’02: Proceedings of the 2002
IEEE/ACM international conference on Computer-aided design,
pages 442–449, New York, NY, USA, 2002. ACM Press.
12
2018/2/26
A. Proofs
A.1 I-Diagrams
Lemma 3 I-diagrams have the same expressive power as CBAC
constraints.
Proof. We translate an i-diagram into a CBAC constraint as follows. As in Figure 2, we note that b1 = b2 , b1 ⊆ b2 can be expressed in the form |b| = 0, so we may assume that they are part of
CBAC. Similarly, |b| ≤ k can be expressed as b ⊆ s ∧ |s| = k
for a fresh variable s, and |b| ≥ k can be expressed as s ⊆
b ∧ |s| = k. We translate ⊥d into e.g. |0| = 1. Next consider D =
(S, ∅d , Sons, Split, Comp, CInf, CSup, Φ). For each S ∈ S, let
η(S) ∈ S be a representative set name. For each S1 ∈ S \ {η(S)}
introduce conjunct S1 = η(S). Next, for each S1 ∈ Sons(S), introduce a conjunct S1 ⊆ S. For each +P ∈ Φ(S), introduce conjunct S ⊆ P , and for each −P ∈ Φ(S) conjunct S ⊆ P c . Express
the bounds using conjuncts |S| ≤ CSup(S) and |S| ≥ CInf(S).
For each Q ∈ Split(S) and S1 , S2 ∈ Q where S1 6= S2 , introduce conjunct |S1 ∩ S2 | = 0. For each {S1 , . . . , Sn } ∈ Comp(S)
introduce conjunct S = S1 ∪ . . . ∪ Sn .
We translate a CBAC constraint into i-diagram using the following observations. It is sufficient to translate the following boolean
algebra expressions: s0 = s1 ∪ s2 , s0 = sc1 , and |s| = k. We
construct an i-diagram whose nodes are singletons. We pick one
set variable u to act as a universal set and put {s} ∈ Sons({u})
for every set variable s in the i-diagram. We translate s0 = s1 ∪
s2 as {{s1 }, {s2 }} ∈ Comp({s0 }) and translate s0 = sc1 as
{{s0 }, {s1 }} ∈ Split({u}), {{s0 }, {s1 }} ∈ Comp({u}). We
translate |s| = k as CInf({s}) = k and CSup({s}) = k. Then
for each satisfiable assignment of CBAC there is a model for the
constructed i-diagram where {u} is interpreted as a universal set.
Conversely, for a model of i-diagram where ᾱ({s}) = A, we let
[s 7→ A ∩ ᾱ(u)]. The result is an assignment that satisfies the original CBAC formula.
Lemma 4 For every i-diagram D we can obtain an equivalent
simple i-diagram using the polynomial-time algorithm in Figure 6.
Proof. We argue that algorithm in Figure 6 produces diagram that
is i) well-formed, ii) simple, iii) equivalent to the original diagram.
We first observe that after step 2, the following two conditions
hold:
C) if Q ∈ Split(S) and S ∈ Q, then Q ⊆ {S, ∅d }. This condition
holds because the step 1.5 of the algorithm merges all nodes
Q \ {S} with ∅d when S ∈ Q ∈ Split(S).
D) if Q ∈ Comp(S), then Q 6= ∅ (by step 1.3), if Q = {∅d } then
S = ∅d (by step 1.6).
i) To see that the resulting
i-diagram is well-formed,
S
S it suffices
to check the conditions Split(S) = Sons(S) and Comp(S) ⊆
Sons(S). This condition is preserved by factor-diagram construction (for any equivalence relation). It is preserved by step 3 for the
following reason. The only nodes removed
S from Sons(S) are ∅d
and S. These nodes do not appear in Comp(S) because ∅d is
removed from each view Q, and views
Swith S ∈ Q are removed.
It remains to check that Sons(S) ⊆ Split(S) after step 3, and
this holds because our condition
(C) implies that no element other
S
than S, ∅d is lost from Split(S) in step 3. The well-formedness
condition is preserved
by step 4 because this step does not change
S
S
Comp(S) or Split(S). Step 5 does not violate this condition
either because it sets the components of ∅d to ∅.
ii) To see that the resulting diagram is simple, we show that
it satisfies conditions a),. . . ,f) of Definition 6. After step 2 of the
algorithm, the resulting factor-diagram has no cycles of length 2
or more, there are only potentially some self-cycles. These are
eliminated in step 3 and no further edges are introduced. Hence,
a) holds. For each of the following condition, they are enforced in
On Algorithms and Complexity for Sets with Cardinality Constraints
certain step and not violated afterwards, according to the following
table:
b) c) d) e) f )
3. 4. 4. 4. 5.
iii) We show that semantics is preserved when executing each
sequence of steps 1, . . . , k for 2 ≤ k ≤ 5, that is, each step preserves the semantics provided that it is executed after the previous
steps.
k = 2. Each equality introduced into ρ is a semantic consequence of the diagram, because
α(S1 ) ⊆ S2 and α(S2 ) ⊆ α(S1 ),
α(S1 ) ⊆ α(∅d ) = ∅,
S
α(S1 ) ⊆ ∅ = ∅,
α(S1 ) ∩ α(S2 ) = ∅ for α(S1 ) = α(S2 ) so α(S2 ) = ∅, or
α(S1 ) ∩ α(S2 ) = ∅, for α(S1 ) = α(S0 ), and α(S2 ) ⊆ α(S0 ),
so again α(S2 ) = ∅,
S
1.6 α(S1 ) ⊆ S and α(S) ⊆ {α(S1 )} = α(S1 ).
1.1
1.2
1.3
1.4
1.5
It follows that the condition on equality of sets, as well as the conditions on CInf, CSup, Sons, Φ, Comp are all semantically equivalent when applied to the original and the factor diagram. The only
semantic condition which can be lost in factor-diagram construction is disjD,S0 (S1 , S2 ) when S1 and S2 nodes are merged, that is,
when (S1 , S2 ) ∈ ρ. However, in this case the disjointness condition follows from α(S1 ) = ∅, which is enforced in 1.3. Therefore,
for the particular relation constructed in step 1, factor-diagram is
an equivalence preserving transformation.
k = 3. We need to show that no information is lost by removing
∅d and S from the sons, as well as split and complete views of
S. Clearly, removing S and ∅d from Sons(S) does not change
the subset conditions because ∅ ⊆ α(S) and α(S) ⊆ α(S).
Eliminating ∅d from Q ∈ Comp(S) is justified because the view
has the same semantics with or without ∅d . Dropping a view Q ∈
Comp(S) for S ∈ Q is justified because in that case α(S) ⊆
S
S1 ∈Q α(S1 ) holds trivially. Eliminating ∅d from Q ∈ Split(S)
is justified because intersection with empty set is always empty,
so this condition does not bring any new information. Finally,
dropping a Q ∈ Split(S) with S ∈ Q is justified because condition
(C) implies that in such case Q ⊆ {∅d , S} so the Split condition is
trivial.
k = 4. Removing {∅d } from Split(S) preserves semantics
because such view carries no information. Similarly, because all
maximal views are preserved, removing their subsets does not
change the semantics. For Comp(S), we consider two cases. In first
case ∅ ∈
/ Comp(S). In this case, removing ∅ does not have any
effect, and it is sound to remove all non-minimal views because
they are implied by the minimal views. The second case is ∅ ∈
Comp(S). By condition D) on the step 1, we know that Q 6= ∅ after
the step 2, and the only node removed in step 3 is ∅d , so it must have
been the case that Q = {∅d } after step 2. By condition D), we then
have S = ∅d . Because the semantic condition on Comp for Q = ∅
reduces to α(S) = ∅, this condition brings no new information, so
we can remove it.
k = 5. Because α(∅d ) =, CSup(∅d ) = 0 does not change
semantics, similarly for Φ(∅d ) = ∅. We also know that Sons(S) ⊆
{∅d } because this condition is ensured by step 2 and is not violated
afterwards. Because we have already observed that the diagram
is well-formed, we conclude Comp(S) ⊆ {∅d } and Split(S) ⊆
{∅d }, so setting these values to ∅ does not change the semantics.
Theorem 2 Omitting any one out of three conditions from Definition 9 (1. being tree-shaped, 2. having independent views, and 3.
having independent signatures) yields a class of diagrams whose
satisfiability is NP-hard.
Proof. Suppose that at least one of the three conditions does not
apply to a class of i-diagrams. We then give a reduction from the
13
2018/2/26
problem 3COL to the satisfiability of i-diagrams in this class. Here
3COL denotes the NP-complete problem of deciding, given an
indirected graph, whether the graph can be colored using 3 colors
such that adjacent nodes have different colors [41, Page 275].
Given a graph (N, E) where E ⊆ N × N is a symmetric irreflexive relation, we first build the i-diagram D defined as follows:
S = N ∪ {U }, CInf(U ) = CSup(U ) = 3, Sons(U ) = N ,
Split(U ) = {{n}|n ∈ N }, and Comp(U ) = Φ(U ) = ∅. For
all n ∈ N we let CInf(n) = CSup(n) = 1 and let Sons(n) =
Split(n) = Comp(n) = Φ(n) = ∅.
Each model of this diagram is a triple (∆, α, Ξ) such that
|α(U )| = 3 and for all n, α(n) is a singleton included in α(U ).
If we consider α(U ) as a set of three colors, then in each model
with this property, α(n) indicates the color of node n.
Then for each edge (n1 , n2 ) ∈ E we encode the fact that
n1 and n2 must have different colors by enforcing the property
α(n1 ) ∩ α(n2 ) = ∅ on the models of D. We encode this constraint
in different ways depending on the class of i-diagrams:
• If D allows dependent signatures, we introduce a fresh pred-
icate symbol Pn1 ,n2 , and add (+Pn1 ,n2 ) to Φ(n1 ), and
(−Pn1 ,n2 ) to Φ(n2 ).
• If D allows dependent views, we add {n1 , n2 } to Split(U ).
• If D allows multiple fathers, then we simulate dependent views by introducing a new node m(n1 , n2 ).
We let CInf(m(n1 , n2 )) = CSup(m(n1 , n2 )) = 2,
Sons(m(n1 , n2 ))
=
{n1 , n2 }, Split(m(n1 , n2 ))
=
Comp(m(n1 , n2 )) = {{n1 , n2 }}, and Φ(m(n1 , n2 )) = ∅.
We then remove n1 , n2 from Sons(U ), and add m(n1 , n2 ) to
Sons(U ) instead.
None of these constructions violates more than one of the three considered restrictions. It is straightforward to verify that the diagram
is satisfiable iff the graph is colorable, and that the construction of
D can be done in polynomial time. This proves that the satisfiability of i-diagrams with any of the three restrictions removed is
NP-hard.
A.2 Termination of System R
′
T and T such
Proof. We analyze each rule Rk , for two i-trees
V
that T −→ T ′ , assuming that T satisfies j=1..(k−1) Cj , and,
Rk
′
more precisely, T −→ T , where spot are variable names as they
Rk
appear in the definition of R.
1. (DnPhi). Trivial.
2. (Unsat). (C1 ) does not depend on CSup.
3. (UpSup). If n = 0, (C2 ) is trivialy true for S in D′ . If
n > 0, by (c3 ) CSup(S) > 0 and we have already ∀P ∈
PN.{+P, −P } 6⊆ Φ(S) and since Φ′ = Φ, (C2 ) holds in T ′ .
4. (UpInf ). Neither (C1 ),(C2 ) nor (C3 ) depend on CInf.
5. (Error). T ′ =⊥d
6. (DnInf ). Only (C4 ) and (C5 ) depends on CInf.
• (C5 ) is maintained for S in T ′ because, noticing that (C5 ) is
maintained for S ′ 6= S, we have
CInf ′ (S) = CInf(S ′ ) − Σ(CSup(Q0 ))
≤ CSup(S ′ ) − Σ(CSup(Q0 ))
≤ CSup(S)
• To prove that (C4 ) is maintained in T
ΣCInf ′ (Q) = CInf ′ (S)+ΣCInf(Q′0 )
= (CInf(S ′ )−ΣCSup(Q0 ))+ΣCInf(S0′ )
= CInf(S ′ )−ΣCSup(Q0 − Q′0 )
+ΣS0′ ∈Q′0 (CInf(S0′ ) − CSup(S0′ ))
≤ CInf(S ′ )−ΣCSup(Q0 − Q′0 ) (by C5 )
≤ CInf(S ′ )
= CInf ′ (S ′ )
7. (DnSup). Only (C2 ), (C3 ) and (C5 ) depend on CSup. (C2 ) is
maintained thanks to (c7 ) as for the case of rule UpSup.
• (C5 ) is maintained for S in T ′ because, noticing that (C5 ) is
maintained for S ′ 6= S we have
CSup′ (S) = CSup(S ′ ) − Σ(CInf(Q0 ))
≥ CInf(S ′ ) − Σ(CInf(Q0 ))
≥ CInf(S)
(by C5 )
(by C4 )
• To prove that (C3 ) is maintained in T ′ we need to check
that (C3 ) is maintained for S, which is trivial by (c7 ), and
that (C3 ) is maintained for the father S ′ of S and the views
Q ∈ Comp(S ′ ) containing S. By property of independent
views there exists only one such view Q = ({S} ⊎ Q0 ) and
ΣCSup′ (Q) = CSup′ (S)+ΣCSup(Q0 )
= (CSup(S ′ )−ΣCInf(Q0 ))+ΣCSup(S0 )
= CInf(S ′ )+ΣS0 ∈Q0 (CSup(S0 ) − CInf(S0 ))
≥ CInf(S ′ ) (by C5 )
= CInf ′ (S ′ )
8. (CCmp). Only (C3 ) and (C6 ) depend on Comp.
• (C3 ) is maintained for S, Q because
CSup′ (S) = CSup(S)
≤ Σ(CInf(Q))
≤ Σ(CSup(Q))
= Σ(CSup′ (Q))
(by a8 )
(by C5 )
• (C6 ) is maintained for S, Q and all S0 ∈ Q because
L EMMA 17
V (Invariants of R). For every k ∈ [1..12] the rule Rk
preserves j=1..(k−1) Cj or returns ⊥d .
spot
views there exists only one such view Q = ({S} ⊎ Q′0 )
such that Q′0 ⊆ Q0 and
′
(by C5 )
(by a8 )
(by C5 )
(Remark). If ever we use a simplification afterwards, as indicated by the star in figure 10, it can only consists in removing
a completeVview Q′ such that Q′ ( Q. This operation trivially
maintains j=1..(7) Cj because in every properties of consistency where complete views appear they are universally quantified.
9. (CSplit). Only (C4 ) and (C7 ) depend on Split.
• (C4 ) is maintained for S, Q because
CInf ′ (S) = CInf(S)
≥ Σ(CSup(Q))
≥ Σ(CInf(Q))
= Σ(CInf ′ (Q))
(by a9 )
(by C5 )
• (C7 ) is maintained for S, Q and all S0 ∈ Q because
(by C5 )
(by C3 )
we need to check
that (C4 ) is maintained for S, which is trivial by (c6 ), and
that (C4 ) is maintained for the father S ′ of S and the views
Q ∈ Split(S ′ ) containing S. By property of independent
On Algorithms and Complexity for Sets with Cardinality Constraints
CInf ′ (S) = CInf(S)
≤ CSup(S)
≤ Σ(CInf(Q))
= CInf(S0 ) + Σ(CInf(Q − {S0 }))
≤ CInf(S0 ) + Σ(CSup(Q − {S0 }))
= CInf ′ (S0 ) + Σ(CSup′ (Q − {S0 }))
CSup′ (S) = CSup(S)
≥ CInf(S)
≥ Σ(CSup(Q))
= CSup(S0 ) + Σ(CSup(Q − {S0 }))
≥ CSup(S0 ) + Σ(CInf(Q − {S0 }))
= CSup′ (S0 ) + Σ(CInf ′ (Q − {S0 }))
14
(by C5 )
(by a9 )
(by C5 )
2018/2/26
(Remark). If ever we use a simplification afterwards, as indicated by the star in figure 10, it can only consists in removing a split V
view Q′ such that Q′ ( Q. This operation trivially
maintains j=1..8 Cj because in every properties of consistency
where split views appear they are universally quantified.
10. (UpPhi). Only (C1 ) and (C2 ) depend on Φ
• (C1 ) is maintained for S and all S ′ ∈ Q because
′
T
Φ (S) = Φ(S) ∪ Φ(Q)T
= Φ(S) ∪ (Φ(S ′ ) Φ(Q − {S ′ })
⊆ Φ(S) ∪ Φ(S ′ )
⊆ Φ(S ′ )
= Φ′ (S ′ )
(by b10 )
(by C1 )
• We can also prove that (C2 ) is maintained for S. Suppose
there exists P in PN such that {+P, −P } ∈ Φ′ (S). Then
we distinguish three cases.
if {+P, −P } ⊆ Φ(S), by C2 , CSup(S) = 0
T
if {+P, −P } ∩ Φ(S) = ∅, then {+P, −P } ∈ Φ(Q).
Then for all S ′ ∈ Q, CSup(S ′ ) = 0 by C2 . Then by C3 ,
CSup(S) = 0
if {+P, −P } ∩ Φ(S) = ±P for one atom ±P ∈
{+P,
−P }. Then the oppositeTatom ∓P belong
T
Φ(Q) and by C1 , ±P belong to Φ(Q). By C2 , each
node S ′ ∈ Q is such that CSup(S ′ ) = 0 and by C3 ,
CSup(S) = 0
In the three cases CSup′ (S) = CSup(S) = 0.
11. (Void). If S is the root,the resulting i-tree T ′ is such that
S′ = {SN} = {∅d } and Ci is trivial for all i ∈ [1..11].
Otherwize, we have to check that removing the node S from
the sons of the father of S (asVindicated in the step 3 of the
procedure simplify) maintains j=1..10 Cj . We denote by S0
the father of S and by Q0 the split view of S0 containing S. If
S is also contained in a complete view of S0 we denote by C0
this complete view.
• (C1 ) and (C2 ) are trivially maintained.
• (C3 ) is maintained because, if C0 exists and C0′ = C0 −{S}
is not empty
CSup′ (S0 ) = CSup(S0 )
≤ ΣCSup(C0 )
= ΣCSup(C0′ )
= ΣCSup′ (C0′ )
( by C3 )
( by a12 )
• (C4 ) is maintained because, if Q′0 = Q0 − {S} is not empty
CInf ′ (S0 ) = CInf(S0 )
≥ ΣCInf(Q0 )
≥ ΣCInf(Q′0 )
= ΣCInf ′ (Q′0 )
( by C4 )
• (C5 ) is maintained for the node ∅′d = S ∪ ∅d of T ′ because
by (a12 ) and (C5 ), CInf(S) ≤ CSup(S) = 0, by simplicity and (C5 ), CInf(∅d ) ≤ CSup(∅d ) = 0, and therefore
CInf ′ (∅′d ) = M ax(CInf(S), CInf(∅d )) = 0 ≤ CSup′ (∅′d ).
• (C6 ) is maintained for S0 , C0 and every S1 ∈ C0 such that
S1 6= S because
′
CInf (S1 ) = CInf(S1 )
≥ CInf(S0 )−Σ(CSup(C0 −{S1 }))
(by C6 )
= CInf(S0 )−Σ(CSup(C0 −{S}−{S1 })) (by a12 )
= CInf ′ (S0 )−Σ(CSup′ (C0′ −{S1 }))
On Algorithms and Complexity for Sets with Cardinality Constraints
• (C7 ) is maintained for S0 , Q0 and every S1 ∈ Q0 such that
S1 6= S because
CSup′ (S1 ) = CSup(S1 )
≤ CSup(S0 ) − Σ(CInf(C0 − {S1 })) (by C7 )
≤ CSup(S0 ) − Σ(CInf(C0 − {S} − {S1 }))
= CSup′ (S0 ) − Σ(CInf ′ (C0′ − {S1 }))
• (C8 ) is maintained for Q0 because
CSup′ (S0 ) = CSup(S0 )
> Σ(CInf(Q0 ))
≥ Σ(CInf(Q0 − {S0 }))
= Σ(CInf ′ (Q′0 ))
( by C8 )
• (C9 ) is maintained for C0 (if exists and C0′ = C0 − {S0 } =
6
∅) because
CInf ′ (S0 ) = CInf(S0 )
< Σ(CSup(Q0 ))
= Σ(CSup(Q0 − {S0 }))
= Σ(CSup′ (Q′0 ))
( by C9 )
( by a11 )
• (C10 ) is maintained for C0 (if exists and C0′ = C0 −{S0 } =
6
∅) because
T
T
Φ′ (C0′ ) = T Φ(C0 − {S0 })
⊆ Φ(C0 )
⊆ Φ(S0 )
= Φ′ (S0 )
( by C10 )
12. (Equal). Before to merge S and S ′ for {S ′ } ∈ Comp(S) we
have
• CInf(S) = CInf(S ′ ) by C4 and C6
• CSup(S) = CSup(S ′ ) by C3 and C7
• Φ(S) = Φ(S ′ ) by C1 and C10
Therefore all the properties Ci for i = 1..10 are trivially
maintained.
A.3 Model Construction
Lemma 6 (Model Construction) If an i-tree T is weakly consistent, then we can construct a model for T .
Proof. Let T be a weakly consistent i-tree.
We construct a model (∆, α, Ξ) by first constructing a partial
model (∆, α) for all parts of T except Φ, and then extending
(∆, α) with Ξ to satisfy Φ.
Constructing (∆, α). We write (∆, α) |= T to denote that ∆ and
α satisfy those conditions on S, Sons, Split, Comp, CSup, CInf that
do not mention Φ in Definition 2. To show we can construct (∆, α)
such that (∆, α) |= T we prove by induction on n the following
more general claim.
C LAIM 1. For every i-tree T of height n with root SR :
∀k ∈ [CInf(SR ), CSup(SR )].
∃(∆, α). (∆, α) |= D ∧ |α(SR )| = |∆| = k
If n = 1, the claim holds by C5 taking
∆ = α(SR ) = {1, . . . , k}
For n>1, consider an i-tree T with root SR and k ∈
[CInf(SR ), CSup(SR )]. By examining the constraints in T , we
choose the cardinalities for subtrees of T , use the induction hypothesis to construct models for subtrees, and paste the models for
subtrees into a model for T . We decompose this process into three
steps:
15
2018/2/26
1. For each C ∈ Comp(SR ), consider
a subtree TC built from T
S
by removing all sons outside C. We construct a model MC
for TC of cardinality k.
S
2. For each remaining Q ∈ Split(SR ) where Q 6⊆ Comp(SR ),
consider a subtree
TQ with the same root SR but without the
S
sons outside Q. We construct a model MQ for TC of cardinality k.
3. Because the constructed models have the same cardinality, we
can easily merge them to obtain a model for T .
Step 1. Let C ∈ Comp(SR ), and C ⊆ Split(SR ) such that
def
C = ∪C. For each Q ∈ C, let tQ = min(k, Σ(CSup[Q])). Using
k ≥ CInf(SR ) and C4 we can show
Σ(CInf[Q]) ≤ tQ ≤ Σ(CSup[Q])
(H1)
To each node S ∈ Q we can therefore assign an integer K(S) ∈
[CInf(S), CSup(S)] such that tQ = Σ(K[Q]). Let TS be the subi-tree of T rooted at S. By induction hypothesis, let MS be the
model of TS of cardinality K(S). We can then take the disjoint
union of these
S models to construct a model MMQ of size tQ for
the forest S∈Q TS .
For all Q ∈ C, we have tQ ≤ k by definition of tQ and k.
We can also prove that k ≤ ΣQ∈C tQ . Indeed, if there exists a
Q0 ∈ C such that tQ0 = k, this is trivial. Otherwise, because T
is weakly consistent, from C3 we can show that k ≤ CSup(S) ≤
Σ(CSup[C]) because
Σ(CSup[C]) = Σ (Σ(CSup[Q])) = Σ (tQ ).
Q∈C
Q∈C
We finally obtain
max tQ ≤ k ≤ Σ tQ
Q∈C
Q∈C
(H2)
Thanks to (H2), we can build a model for the i-tree TC , as follows.
We start with the disjoint union of models MQ for TQ for Q ∈ C.
This model has cardinality ΣQ∈C tQ . Then, we rename elements
from different models to be identical to elements from other models. Such merging is possible as long as there is no model whose
domain contains the domains of all others, so we can reach any
cardinality k for maxQ∈C ≤ k.
R EMARK 2. (Freedom in the choice of {ti }i∈[1..n] ) In Section 7
we enforce some additional properties on models using a different choice of tQ . Such construction is possible whenever ti satisfy
(H1) and (H2). Moreover, if K(S) denotes some chosen cardinality for each node S, and the values K(S) satisfy certain assumptions, then we can enforce additional properties when merging the
models MQ corresponding for Q ∈ Split(S). The following two
cases are of interest.
P
tQ > K(S), we can chose any pair of different split
1. If
Q∈C
views Q1 , Q2 ∈ C, and two elements x1 from MQ1 and x2
from MQ2 and decide to merge them.
2. If C = {Q0 } ⊎ C0 and max tQ < K(S), we can chose any
Q∈C0
element in the model MQ0 and decide not to merge it with any
of the elements of the models MQ′ for Q′ ∈ C0 .
Step 2. Let Q ∈ Split(SR ), such that Q is not included in any
complete view. We construct a model MQ of size k for the i-forest
TT by first building a model of size K(S ′ ) = CInf(S ′ ) for each
S ′ ∈ Q. Because k ≥ CInf(S) ≥ Σ(CInf[Q]), by C4 , the disjoint
union of these models has cardinality smaller than k. By adding
the correct number of fresh elements to ∆, we obtain a model of
cardinality k.
R EMARK 3. (Existence of fresh elements) In Section 7 we use the
following property: For all Q ∈ Split(S) and Q 6∈ ∪Comp(S),
On Algorithms and Complexity for Sets with Cardinality Constraints
if Σ(CInf[Q]) < K(S), then there exists a model such that α(S)
contains an element which does not belong to any α(S ′ ) for any of
the sons of S.
Step 3. We can apply an arbitrary bijection σC : ∆C → [1..n]
to each model MC constructed as previously described before to
build a model for the entire i-tree. We let α(SR ) = [1..n] and for
all S 6= SR , α(S) = αC (S) where C is the view containing an
ancestor of S in Split(SR ) or Comp(SR ).
R EMARK 4. (Freedom in the choice of σ.) If we know that
K(S) > 0, for any pair S1 , S2 of sons of S such that S1 , S2 belong
neither to the same split view nor to the same complete view, for
each choice of elements x1 , x2 in the models TS1 and TS2 we can
def
choose σ1 , σ2 such that σ1 (x1 ) = σ2 (x2 ) = x and the resulting
i-tree will be such that x ∈ α(S1 ) ∩ α(S2 ).
Extending the model with Ξ. Let (∆, α) |= T . Then for each
P ∈ PN, define
[
Ξ(P ) = {α(S) | S ∈ S ∧ P ∈ Φ(S)}
Then (+P ) ∈ Φ(S) ⇒ α(S) ⊆ Ξ(P ) holds by construction,
it remains to show (−P ) ∈ Φ(S) ⇒ α(S) ⊆ Ξ(P )c for every
node S. Consider S1 ∈ S such that (−P ) ∈ Φ(S1 ). If S1 = ∅d ,
then α(S1 ) = ∅, so the condition trivially holds. Similarly, if
(+P ) ∈ Φ(S1 ), then by C2 , CSup(S1 ) = 0 so α(S1 ) = ∅ and the
condition holds. Otherwise, assume (+P ) ∈
/ Φ(S1 ). For the sake
of contradiction suppose that there exists an element x ∈ α(S1 ),
x ∈ Ξ(P ). By definition of Ξ(P ), there exists a node S2 6= S1
such that x ∈ α(S2 ) and (+P ) ∈ Φ(S2 ). By the condition on
independent signatures, one of the followig two cases applies.
1. disj∗T (S1 , S2 ). Then α(S1 ) ∩ α(S2 ) = ∅ by the semantics
of i-diagrams, which is a contradiction with x ∈ α(S1 ) and
x ∈ α(S2 ).
2. S1 and S2 have compatiable signatures. Then there exists a
∗
∗
node S such that S1 S, S2 S and Sig(S1 ) ∩ Sig(S2 ) ⊆
Sig(S). Because (−P ) ∈ Φ(S1 ), (+P ) ∈ Sig(S1 ), and
because (−P ) ∈ Φ(S2 ), P ∈ Sig(S2 ). Therefore P ∈ Sig(S).
We have two cases:
(a) (+P ) ∈ Φ(S). By C1 , then (+P ) ∈ Φ(S1 ), a contradiction.
(b) (−P ) ∈ Φ(S). By C2 , then (−P ) ∈ Φ(S2 ). By C2 then
CSup(S2 ) = 0, so α(S2 ) = ∅, a contradiction with x ∈
α(S2 ).
We have reached the contradiction in each case, so we conclude
α(S1 ) ⊆ Ξ(P )c .
A.4 Details of the Proofs for Subsumption Completeness
A.4.1 Refinemenents of Lemma 6
According to the remarks in the proof of Lemma 6, if an i-tree T is
weakly consistent, there exists a choice of cardinalities K : S → N,
such that we can build a model (∆, α, Ξ) for T with the property
|α(S)| = K(S) for all S ∈ S. For a fixed choice of cardinalities
K, we can, in certain cases, enforce some additional properties
by choosing which element we merge in the steps 1 and 3 of the
construction. The three following Lemmas are based on this idea.
L EMMA 18 (Non-empty intersection (1)). Let S1 and S2 be nodes
in a weakly consistent i-tree T is such that
CInf(S1 ) > 0 ∧ S1
∗
S1′ ∧ S1′ ∈ Q1 ∧
CInf(S2 ) > 0 ∧ S2
∗
S2′ ∧ S2′ ∈ Q2
for some S, S1′ , S2′ ∈ S, Q1 , Q2 ∈ Split(S) where Q1 6= Q2 and
¬(∃C ∈ Comp(S). Q1 ⊆ C ∧ Q2 ⊆ C). Then there exists a
16
2018/2/26
model (∆, α, Ξ) for T such that
α(S1 ) ∩ α(S2 ) 6= ∅.
Proof. We use the construction described in Lemma 6, with the
exception of step 3 of the construction of the model MS for the
subtree TS with root S, where we do the following:
• We choose an element x1 in α1 (S1 ) in the model MS ′ built for
1
TS1′ (we know there exists one such x1 because CInf(S1 ) > 0)
• Analogously, we choose an element x2 in α2 (S2 ) in the model
MS2′ build for TS2′
• We choose the bijections σQ1 and σQ2 in step 3 such that
σQ1 (x1 ) = σQ2 (x2 ).
L EMMA 19 (Non-empty intersection (2)). Let S1 and S2 be nodes
in a weakly consistent i-tree T is such that
CInf(S1 ) > 0 ∧
CInf(S2 ) > 0 ∧
∗
S1 S1′
∗
S2 S2′
∧
∧
S1′
S2′
∈ Q1 ∧
∈ Q2
for some S, S1′ , S2′ ∈ S, Q1 , Q2 ∈ Split(S) where Q1 6= Q2 , and
Q1 ⊆ C, Q2 ⊆ C for some C ∈ Comp(S) with the property
CSup(S) < Σ(CSup[C]).
Then there exists a model (∆, α, Ξ) for T such that
α(S1 ) ∩ α(S2 ) 6= ∅.
Proof. We use the construction described in the proof of
Lemma 6, except for the step 1 of the construction of the model
MS for the subtree TS with root S, for which we do the following.
From K(S) ≤ CSup(S) < Σ(CSup[Q]), we conclude that the
choice of cardinalities tQ for Q ∈ C in the proof of Lemma 6 is
such that ΣQ∈C tQ > K(S), by considering two cases.
1. There exists Q ∈ C such that
Proof. We use a variation of the construction in the proof of
Lemma 6. We apply the assumptions about the subtree Q1 to show
that we always have enough “slack” to avoid merging one specific
element from Q1 with the elements of neighbors. We being by
describing a slightly modified Step 1 of the proof of Lemma 6.
Step 1’ (for nodes of Q1 ). Consider the root SR ∈ Q1 . Let
C ∈ Comp(SR ), and C ⊆ Split(SR ) such that C = ∪C. By
construction, there exists a unique son S ′ ∈ Q1 of SR and a
corresponding split view Q′ such that S ′ ∈ Q′ and C = {Q′ }⊎C0 .
We define tQ′ = min(k, Σ(CSup[Q′ ])) and for each Q0 ∈ C0 , we
def
define tQ0 = min(k−1, Σ(CSup[Q0 ])). For each Q0 ∈ C0 , since
k ≥ CInf(SR ) by choice of k and CInf(SR ) > Σ(CInf[Q0 ]) by
hypothesis 3 on T , we have Σ(CInf[Q0 ]) ≤ k−1, so
Σ(CInf[Q0 ]) ≤ tQ0 ≤ Σ(CSup[Q0 ])
(H1)
and the property (H1) also holds for tQ′ , because tQ′ is defined as
in Lemma 6.
By definition of t we clearly have maxQ∈C tQ ≤ k. We next
show Σ tQ ≥ k by considering the following cases.
Q∈C
• tQ′ = k. Then the claim is obvious.
• For all Q ∈ C we have tQ = Σ(CSup[Q]). The claim follows
from C3 and the choice of k because ΣtQ ≥ Σ(CSup[C]) ≥
CSup(SR ) ≥ k.
• There exists Q0 ∈ C0 such that tQ0 = k − 1. Using
Σ(CSup[Q′ ]) ≥ CSup(S ′ ) > 0 and k ≥ CInf(SR ) > 0,
we obtain tQ′ > 0, so
Σ tQ ≥ tQ0 + tQ′ ≥ (k − 1) + 1 ≥ k.
Q∈C
We finally obtain
max tQ ≤ k ≤ Σ tQ
Q∈C
(H2)
Q∈C
By definition of all tQ0 we then have
tQ = min(K(S), Σ(CSup[Q])) = K(S).
∀Q0 ∈ C0 . tQ0 < k
Then we choose Qi ∈ {Q1 , Q2 } such that Qi 6= Q. Since
∗
CInf(S1 ) > 0 ∧ S1 S1′
S, repeatedly applying C4 and
using C5 , we have
According to Remark 2, H3 allows us to choose an element of the
model MS ′ constructed for the subtree TS ′ and decide not to merge
it with any other element. This observation allows us to recursively
enforce x ∈ α(S) ⇐⇒ S ∈ Q1 .
p
1
Indeed, consider a node SR ∈ Q1 and let {SR
, . . . , SR
} =
Sons(SR ) ∩ Q1 be its sons in Q1 . For each i, we can then recuri
sively ensure xi ∈ α(S) ⇐⇒ S ∈ Q1 for each S in the SR
∗
i
subtree i.e. for each S for which S SR . By definition of Q1 ,
i
each SR
is in a different complete view, so we can apply bijection
to the submodels (Remark 4) and let σi (x1 ) = . . . = σp (xp ) = x.
We ensure that x does not belong to any subtree rooted at a node
S0 ∈ Sons(SR ) \ Q1 , using Remark 2 to make sure that x is not
merged with any of the elements of α(S0 ), which is possible thanks
to H3. Finally, for the base case, when S has no sons, we pick x to
be a fresh element, which is possible by assumption 3 on the subtree Q1 , as noted in Remark 3.
• Σ(CSup[Qi ]) ≥ CSup(Si′ ) ≥ CInf(Si′ ) ≥ CInf(Si ) > 0
• K(S) ≥ CInf(S) ≥ CInf(Si ) > 0
As a consequence tQi = min(K(S), Σ(CSup[Qi ])) > 0 and
ΣQ∈C tQ ≥ tQ + tQi > K(S).
2. For all Q ∈ C, tQ = Σ(CSup[Q])). Then ΣQ∈C tQ =
Σ(CSup[C]) > CSup(S) ≥ K(S).
Because ΣQ∈C tQ > K(S), we can apply Remark 2 and choose
one element x1 in α1 (S1 ) in the model MS1′ built for TS1′ , and
an element x2 in α2 (S2 ) in the model built for TS2′ , and decide
to merge these elements in the step 1 of the construction. We
know that such elements x1 , x2 exist because CInf(S1 ) > 0 and
CInf(S2 ) > 0.
L EMMA 20 (Isolated element). Let T be a weakly consistent itree and (Q1 , ) a subtree of (S, ) with the same root SR as
T , such that for all S ∈ Q1 all the following conditions hold:
1. CInf(S) > 0
2. ∀C ∈ Comp(S). ∃=1 Q ∈ Split(S). Q ⊆ C ∧ |Q ∩ Q1 | = 1
3. ∀Q ∈ Split(S). |Q ∩ Q1 | =
6 1 ⇒
(|Q ∩ Q1 | = 0 ∧ Σ(CInf[Q]) < CInf(S)).
Then we can construct a model for T such that
∃x ∈ α(R). ∀S ∈ S. (x ∈ α(S) ⇔ S ∈ Q1 )
On Algorithms and Complexity for Sets with Cardinality Constraints
(H3)
A.4.2 Links between weak and strong consistency
Lemma 8 (Bounds Refinement) Let T be a strongly consistent
i-tree, S ∈ S, i, s such that CInf(S) ≤ i ≤ s ≤ CSup(S), let
′
T ′ = T [CInf(S) ← i, CSup(S) ← s] and TNF
= RwNF (T ′ ). Then
∗
′
′
1) TNF 6=⊥d , 2) TNF |= T , and 3) if ¬(S S0 ), then
(CInf(S0 ), CSup(S0 )) = (CInf ′NF (S0 ), CSup′NF (S0 )).
Proof.
1. We prove this result by induction on the depth of S in the tree
(S, ). The key step of this proof is to show that the application
of UpSup and/or UpInf to the father S ′ of S do not produce
17
2018/2/26
a situation where a5 holds in the resulting diagram T ′′ (and
therefore the rule Error is not applicable in T ′′ ). We distinguish
three different cases :
• When UpSup and UpInf are both applicable to this node
′
Therefore, T ′ is already in normal form, so TNF
is identical to T ′
and is clearly distinct from ⊥d , proving condition 1. Condition 2
holds because T ′ |= T because the cardinality bounds in T ′ are at
least as strong as in T . Condition 3 holds because
S ′ we have CInf ′′ (S ′ ) ≤ CSup′′ (S ′ ) because for C ∈
Comp(S ′ ), Q ∈ Split(S ′ ) such that Q ⊆ C, Q = {S} ⊎
S ′ ,C
Q0 , C = {S} ⊎ C0 and T ′ −→
Σ(CInf ′ [Q]) = Σ(CInf[Q])
< CSup(S)
= CInf ′ (S)
S ′ ,Q
−→ T ′′ we have
(because Q ∩ Q0 = ∅)
(by C8 )
(by definition of T ′ )
UpSup UpInf
CInf ′′ (S ′ ) = Σ(CInf ′ [Q0 ]) + CInf ′ (S)
= Σ(CInf ′ [Q0 ]) + CSup′ (S)
≤ Σ(CSup′ [Q0 ]) + CSup′ (S)
≤ Σ(CSup′ [C0 ]) + CSup′ (S)
= CSup′′ (S ′ )
(by b4 )
(i ≤ s)
(by C5 )
(Q0 ⊆ C0 )
(by b3 )
′
• When only UpSup is applicable to this node S we have
CInf ′′ (S ′ ) ≤ CSup′′ (S ′ ) because for C ∈ Comp(S ′ ),
S ′ ,C
C = {S} ⊎ C0 and T ′ −→ T ′′ we have
′
′
′
CSup (S ) = Σ(CSup [C0 ]) + CSup (S)
≥ Σ(CSup′ [Q0 ]) + CInf ′ (S)
≥ CInf ′ (S ′ )
= CInf ′′ (S ′ )
(by b3 )
(i ≤ s)
(by C6 )
CInf ′′ (S ′ ) ≤ CSup′′ (S ′ ) because for Q ∈ Split(S ′ ),
Q = {S} ⊎ Q0 and T
′′
′
−→ T
′′
UpInf
′
we have
′
CInf (S ) = Σ(CInf [Q0 ]) + CInf (S)
≤ Σ(CInf ′ [Q0 ]) + CSup′ (S)
≤ CSup′ (S ′ )
= CSup′′ (S ′ )
∗
def
T ′ = T [∀S ′ , S1
[∀S ′ , S2
∗
∗
S′
S′
∗
∗
S1′ : CInf(S ′ ) ← CSup(S ′ )]
S2′ : CInf(S ′ ) ← CSup(S ′ )]
def
′
TNF
= RwNF (T ′ )
def
• When only UpInf is applicable to this node S ′ we have
′
′ S ,Q
C ∈ Comp(S), Q1 , Q2 ∈ Split(S)
S1 S1′ ∧ S1′ ∈ Q1 ∧ Q1 ⊆ C
∗
S2 S2′ ∧ S2′ ∈ Q2 ∧ Q2 ⊆ C
Q1 =
6 Q2
Define
UpSup
′′
L EMMA 21 (Parallel Bounds Refinement (2)). Let T be a strongly
consistent i-tree and S1 , S2 , S1′ , S2′ , S ∈ S and C, Q such that
(by b4 )
(i ≤ s)
(by C7 )
2. Follows easily from the hypothesis CInf(S) ≤ i ≤ s ≤
CSup(S) and the fact that RwNF is semantics preserving.
3. It is enough to notice that only rules UpInf and UpSup are used
when applying RwNF , and these rules are applied in the bottomup direction.
Lemma 9 (Parallel Bounds Refinement (1)) Let T be a strongly
consistent i-tree, and (Q0 , ) a subtree of T which has the same
root as T and is such that
• The nodes of Q0 are pairwise independent, that is
∀S1 , S2 ∈ Q0 .¬(disj∗T (S1 , S2 ))
′
T ′′ = RwNF (TNF
[CSup′′ (S) ← CInf ′NF (S)])
Then T ′′ 6=⊥d , T ′′ |= T , and CSup′′ (S) < Σ(CSup′′ [C]).
Proof. As in the proof of Lemma 9, weak consistency con∗
ditions hold in T ′ for all nodes in S ′ such that S1 S1′ or
∗
′
S2 S2 . Therefore, the only rewrite step that mat be applicable
in T ′ is the application of UpInf to S. This application may
lead to applications of other instances of UpInf , but the proof
of Lemma 8 shows that this process will result in a weakly con′
sistent i-tree, so TNF
6=⊥d . Moreover, the process of computw
′
′′
ing RNF (TNF [CSup (S) ← CInf ′NF (S)]) is identical to applying
Lemma 8 to T ′ with bounds i = s = CInf ′NF (S), and therefore
leads to a weakly consistent i-tree T ′′ , so T ′′ 6=⊥d .
The condition T ′′ |= T follows because RwNF is semanticspreserving, and the updates of trees only shrink the bounds on
nodes, so they convert a diagram into a stronger one.
To prove CSup′′ (S) < Σ(CSup′′ [C]), observe first
that CSup′′ (S) = CInf ′NF (S) by definition of T ′′ , and
Σ(CSup′′ [C]) = Σ(CSup[C]) because CSup does not change for
any ancestors of S. Therefore, it suffices to show
CInf ′NF (S) < Σ(CSup[C])
′
Then the i-tree T defined by
We prove this condition by distinguishing two cases.
def
T ′ = T [ ∀S ∈ Q0 :CInf(S)←CSup(S) ]
def
′
is such that his Rw normal form TNF
= RwNF (T ′ ) satisfies
′
1. TNF
6=⊥d
′
2. TNF
|= T
3. ∀S ∈ Q0 . ∀Q ∈ Split′NF (S). Q ∩ Q0 = ∅ ⇒
Σ(CInf ′NF [Q]) < CInf ′NF (S)
CInf ′NF (S)
Proof. If we apply [CInf(S ′ )←CSup(S ′ )] to every node S ′ of Q0
starting from the root to the leaves, we always maintain C1 , C2 , C3
because Φ and CSup are never modified. We also maintain C4
because for each S ∈ Q0 and each view Q ∈ Split(S) such that
there exists S ′ ∈ Q ∩ Q0 , by ¬disj∗T (S1 , S2 ), we know that S ′ is
the only modified node, and
CInf ′ (S) = CSup(S)
≥ Σ(CInf[Q − {S ′ }]) + CSup(S ′ )
= Σ(CInf ′ [Q])
On Algorithms and Complexity for Sets with Cardinality Constraints
1. UpInf is not applicable to S. Then CInf ′NF (S) = CInf(S) and
the condition follows by C9 .
2. UpInf is applicable to S. Then for some a, b where {a, b} =
{1, 2} we have
(by C7 )
=
<
≤
≤
Σ(CInf ′ [Qa ])
Σ(CInf ′ [Qa ]) + Σ(CInf ′ [Qb ])
Σ(CInf ′ [C])
Σ(CSup[C])
A.4.3 Completeness of the algorithm Subsumes
Theorem 4 Let T be a strongly consistent i-tree and let HTA for
atomic formula A be as defined in Figure 12. Then HTA if and only
if T |= A.
The (⇒) direction of Theorem 4 is trivial by the semantics of idiagrams. For (⇐) direction we prove the following characteriza-
18
2018/2/26
tions:
S 6= ∅d
k ∈ [CInf(S), CSup(S)]
¬(Included(S0 , C, T ))
∗
S1 6= ∅d ∧ ¬(S1 S2 )
S1 6= S2
∅d 6∈ {S1 , S2 } ∧ ¬disj∗T (S1 , S2 )
+P 6∈ Φ(S)
−P 6∈ Φ(S)
⇒ ∃M.
⇒ ∃M.
⇒ ∃M.
⇒ ∃M.
⇒ ∃M.
⇒ ∃M.
⇒ ∃M.
⇒ ∃M.
α(S) 6= ∅
k
|α(S)| = S
α(S0 ) 6⊆ α[C]
α(S1 ) 6⊆ α(S2 )
α(S1 ) 6= α(S2 )
α(S1 ) ∩ α(S2 ) 6= ∅
α(S) 6⊆ Ξ(P )
α(S) 6⊆ Ξ(P )c
where M denotes a model M = (∆, α, Ξ) of T . We next present
the remaining lemmas that prove these characterizations (see also
Section 7).
Lemma 12 If an i-tree T is strongly consistent, S0 , ∈ S, C ∈
P(S), and Included(S0 , C, T ) returns false, then
[
∃M. α(S0 ) 6⊆
α[C]
Proof.
Let Q0 be the smallest set of nodes such that:
∗
S0 S
⇒S ∈ Q0
S ∈ Q0 ∧ Q1 ∈ Comp(S)∧
⇒S1 ∈ Q0
S1 ∈ Q1 ∧ ¬Incl(S1 , C)
By definition of Incl we know that Q0 ∩ C = ∅.
Q0 is tree-shaped by construction, but may contain two nodes
which are explicitly disjoint. We compute a subtree Q1 of Q0 ,
by starting from the root and keeping each time at most one son
for each complete view. We also impose that Q1 contains S, by
avoiding to cut the branch which leads to S.
We then define
Proof. Let S1 , S2 ∈ S \ {∅d } such that ¬(disj∗D (S1 , S2 )). If
S1 = S2 we can find a model M where α(S1 ) = α(S2 ) 6= ∅ by
Lemma 10, in this model α(S1 ) ∩ α(S2 ) = α(S2 ) 6= ∅. Suppose
S1 6= S2 . Define S1′ , S2′ , S0 as the unique nodes such that S0 is the
least common ancestor of S1 and S2 in T , and S1′ ,S2′ ∈ Sons(S0 )
are the ancestors of S1 and S2 , respectively. We distinguish two
cases:
• S1′ and S2′ do not belong to a same complete view of S0 . Then
apply Lemma 9 to the subtree
def
Q0 = {S ∈ S | S1
∗
S ∨ S2
∗
S}
whose nodes are pairwise independent by the hypothesis
′
disj∗T (S1 , S2 ). The resulting tree TNF
satisfies the hypothesis
′
of Lemma 18, so there exists a model M = (∆, α, Ξ) for TNF
such that
α(S1 ) ∩ α(S2 ) 6= ∅.
′
M is also a model of T because TNF
|= T .
′
′
• S1 and S2 belong to a same complete view C. Define
def
T ′ = RwNF (T [∀S ′ , S1
[∀S ′ , S2
∗
∗
S′
S′
∗
∗
S1′ : CInf(S ′ ) ← CSup(S ′ )]
S2′ : CInf(S ′ )←CSup(S ′ )])
def
T ′′ = RwNF (T ′ [CSup′′ (S) ← CInf ′ (S)]
By Lemma 21, then T ′′ |= T and CSup′′ (S) < Σ(CSup′′ [C]).
This last property allows us to apply Lemma 19 and prove
the existence of a model (∆, α, Ξ) for T such that α(S1 ) ∩
α(S2 ) 6= ∅.
T ′ = RwNF (T [∀S ∈ Q1 : CInf(S)←CSup(S)])
According to Lemma 9 (Parallel Bounds Refinement) we have
T ′ 6=⊥d .
We then apply Lemma 20 to construct a model for the weakly
consistent i-tree T ′ such that
We illustrate the idea of the algorithm in Figure 2 through an
example of checking the validity of the formula
∀S ′ ∈ S. (x ∈ α(S ′ ) ⇔ S ′ ∈ Q1 )
|A ∪ B| = |A| + |B| − |A ∩ B|
for some element x ∈ α(S). BecauseTS ∈ Q1 , we have x ∈ α(S).
Because C ∩ Q1 = ∅, we have x ∈
/ α[C].
Lemma 15 If an i-tree T is strongly consistent, then for all S ∈ S
and P ∈ PN we have
(+P ) 6∈ Φ(S) ⇒ ∃M. M |= T ∧ αM (S) 6⊆ Ξ(P )
(−P ) 6∈ Φ(S) ⇒ ∃M. M |= T ∧ αM (S) 6⊆ Ξ(P )c
Proof. Let S ∈ S be such that (+P ) 6∈ Φ(S). We define
def
Q+P = {S ′ ∈ S|(+P ) ∈ Φ(S ′ )}. Using C10 we show that
Included(S, Q+P , T ) cannot
return true. Then, there exists a
S
model such that α(S) 6⊆ ( α[Q+P ]). We then change the model
by redefining Ξ′ by:
[
∀P ∈ PN. Ξ′ (P ) = {α(S ′ )|S ′ ∈ Q+P }
to ensure that Ξ′ (P ) 6⊆ α(S). The case of (−P
S ) 6∈ Φ(S) is
analogous by taking a model such that α(S) 6⊆ ( α[Q−P ]) and
redefining Ξ′ by:
[
∀P ∈ PN. Ξ′ (P ) = ∆ − {α(S ′ )|S ′ ∈ Q−P }
Lemma 16 If an i-tree T is strongly consistent, then for all
S1 , S2 ∈ S such that S1 6= ∅, S2 6= ∅ we have
¬(disj∗T (S1 , S2 )) ⇒ ∃M. α(S1 ) ∩ α(S2 ) 6= ∅.
On Algorithms and Complexity for Sets with Cardinality Constraints
B. Example transformation to CBAC constraints
that is, checking the unsatisfiability of the formula
|A ∪ B| 6= |A| + |B| − |A ∩ B|
This formula has no integer variables initially, so the first step does
not apply. Furthermore, all set algebra expressions appear already
within cardinality constraints, so step 2 is unnecessary as well, as
is step 3 because there are no divisibility constraints.
In step 4, we introduce a non-negative integer variable for each
cardinality term, yielding the system:
iA∪B 6= iA + iB − iA∩B
|1| = MAXC
|A ∪ B| = iA∪B
|A| = iA
|B| = iB
|A ∩ B| = iA∩B
Because the formula is already in the form of a conjunction, there
is no need to non-deterministically guess the conjunction in step 5,
and we continue with the current constraints.
In step 6, we non-deterministically replace iA∪B 6= iA + iB −
iA∩B with iA∪B + 1 ≤ iA + iB − iA∩B or iA + iB − iA∩B + 1 ≤
iA∪B , and we illustrate the first case (the other case needs to be
checked as well unless the satisfying assignment is found in the
19
2018/2/26
and M ′ = 16, so it suffices to use only 4 bits to represent the
constants in the resulting CBAC constraints generated in step 10.
first case):
iA∪B + 1 ≤ iA + iB − iA∩B
|1| = MAXC
|A ∪ B| = iA∪B
|A| = iA
|B| = iB
|A ∩ B| = iA∩B
In step 7, we introduce a slack variable i0 to eliminate the first
inequation, and transform the equation to normal form:
iA∪B − iA − iB + iA∩B + i0 = −1
|1| = MAXC
|A ∪ B| = iA∪B
|A| = iA
|B| = iB
|A ∩ B| = iA∩B
In step 8, we identify v = (MAXC, iA∪B , iA , iB , iA∩B , i0 ) and
have A = [0, 1, −1, −1, 1, 1] and d = (−1). We have m0 = 1,
m1 = 5, n0 = 6, and S = 2. Therefore, m = 6, n = 6, and a = 1.
From these values we compute M = 6 · (6 · 1)11 = 612 < 232 .
Therefore, if there exists a solution to the system of equations, there
exists a solution that can be represented by at most 32 bits in binary.
In step 9, we guess non-deterministically a solution vector
bounded by M that satisfies the first equation (that is, Ak = −1).
One such solution is k = (101, 100, 77, 40, 10, 6) (written in decimal notation). For this guessed solution, we generate the CBAC
constraint:
|1| = 101
|A ∪ B| = 100
|A| = 77
|B| = 40
|A ∩ B| = 10
This example CBAC constraint does not have a solution, and neither do the remaining examples generated by the non-deterministic
algorithm. The search tree corresponding to the non-deterministic
algorithm returns false in all branches, so the formula is unsatisfiable, and its negation is valid.
We note that, in this case, the dimensions of the system
of equations are such that the estimate M can be improved
if we consider the system of equations where the variables
MAXC, iA∪B , iA , iB , iA∩B are all substituted into the the original
integer part of the QFBAPA problem. In general, this alternative
′
estimate is given by M ′ = n′ (m′ a′ )2m +1 where
m′ = m0
S
n′ = max(n
0 − m1 , 2 )
a′ = max
max |dq |,
1≤q≤m0
m0
max max |apq |,
p=1 q∈R
m0
max max(
p=1
P
apq ,
P
q∈Q
q∈Q
apq >0
apq <0
(−apq ))
where Q = {p1 , . . . , pm1 } are indices of variables denoting cardinalities (and that are being substituted), and R = {1, . . . , n0 } \ R.
In our example,
m′ = 1
n′ = 4
a′ = max(1, 1, 2)
On Algorithms and Complexity for Sets with Cardinality Constraints
20
2018/2/26
| 6cs.PL
|
Reinforced stochastic gradient descent for deep neural network learning
Haiping Huang∗ and Taro Toyoizumi†
arXiv:1701.07974v5 [cs.LG] 22 Nov 2017
RIKEN Brain Science Institute, Wako-shi, Saitama 351-0198, Japan
(Dated: November 23, 2017)
Stochastic gradient descent (SGD) is a standard optimization method to minimize a training error
with respect to network parameters in modern neural network learning. However, it typically suffers
from proliferation of saddle points in the high-dimensional parameter space. Therefore, it is highly
desirable to design an efficient algorithm to escape from these saddle points and reach a parameter
region of better generalization capabilities. Here, we propose a simple extension of SGD, namely
reinforced SGD, which simply adds previous first-order gradients in a stochastic manner with a
probability that increases with learning time. As verified in a simple synthetic dataset, this method
significantly accelerates learning compared with the original SGD. Surprisingly, it dramatically
reduces over-fitting effects, even compared with state-of-the-art adaptive learning algorithm—Adam.
For a benchmark handwritten digits dataset, the learning performance is comparable to Adam,
yet with an extra advantage of requiring one-fold less computer memory. The reinforced SGD is
also compared with SGD with fixed or adaptive momentum parameter and Nesterov’s momentum,
which shows that the proposed framework is able to reach a similar generalization accuracy with
less computational costs. Overall, our method introduces stochastic memory into gradients, which
plays an important role in understanding how gradient-based training algorithms can work and its
relationship with generalization abilities of deep networks.
Keywords: Neuronal networks, Machine learning, Backpropagation
I.
INTRODUCTION
Multilayer neural networks have achieved state-of-the-art performances in image recognition [1], speech recognition,
and even natural language processing [2]. This impressive success is based on a simple powerful stochastic gradient
descent (SGD) algorithm [3], and its variants. This algorithm estimates gradients of an error function based on
mini-batches of an entire dataset. Gradient noise caused by mini-batches helps exploration of parameter space to
some extent. However, the parameter space is highly non-convex for a typical deep network training, and finding a
good path for SGD to improve generalization ability of deep neural networks is thus challenging [4].
As found in standard spin glass models of neural networks [5, 6], a non-convex error surface is accompanied by
exponentially many local minima, which hides the (isolated) global minima and thus makes any local search algorithms
easily get trapped. In addition, the error surface structure of deep networks might behave similarly to random Gaussian
error surface [7, 8], which demonstrates that critical points (defined as zero-gradient points) of high error have a large
number of negative eigenvalues of the corresponding Hessian matrix. Consistent with this theoretical study, empirical
studies on deep network training [9, 10] showed that SGD is slowed down by a proliferation of saddle points with
many negative curvatures and even plateaus (eigenvalues close to zero in many directions). The prevalence of saddle
points poses an obstacle to attain better generalization properties for a deep network, especially for SGD based on
first-order optimization, while second-order optimization relying on Hessian-vector products is more computationally
expensive [11]. The second order method that relies on positive-definite curvature approximations, can not follow
directions with negative curvature, and is easily trapped by saddle points [12].
In this paper, we show a heuristic strategy to overcome the plateaus problem for SGD learning. We call this strategy
reinforced SGD (R-SGD), which provides a new effective strategy to use the gradient information, i.e., to update one
network parameter, an instantaneous gradient is reinforced by (accumulated) previous gradients with an increasing
reinforcement probability that grows with learning time steps. In other words, the reinforcement may be turned off,
and then only the instantaneous gradient is used for learning. This kind of stochastic reinforcement enhances the
exploration of parameter space. The excellent performance of R-SGD is verified first on training a toy fully-connected
deep network model to learn a simple non-linear mapping generated by a two-layer feedforward network, and then
on a benchmark handwritten digits dataset [13], in comparison to both vanilla backpropagation (BackProp) [14] and
state-of-the-art Adam algorithm [15]. In the benchmark dataset, we also clarify the performance difference between
∗ Electronic
† Electronic
address: [email protected]
address: [email protected]
2
R-SGD and SGD with fixed or adaptive momentum parameter [16] and Nesterov’s momentum [17].
II.
FULLY-CONNECTED DEEP NETWORKS
We consider a toy deep network model with L layers of fully-connected feedforward architecture. Each layer has
nk neurons (so-called width of layer k). We define the input as n1 -dimensional vector v, and the weight matrix Wk
specifies the symmetric connections between layer k and layer k − 1. The symmetry means that the same connections
are used to backpropagate the error during training. A bias parameter can also be incorporated into the weight matrix
by assuming an additional constant input. The output at the final layer is expressed as:
y = fL WL fL−1 (WL−1 · · · f2 (W2 v)) ,
(1)
where fk (·) is an element-wise sigmoid function for neurons at layer k, defined as f (x) = 1+e1−x , unless otherwise
specified (e.g., ReLU activation funtion). The network is trained to learn the target mapping generated randomly as
{vm , y∗m }M
m=1 , where the input is generated from a standard normal distribution with zero mean and unit variance,
and the target label y∗ is generated according to the non-linear mapping y∗ = f (Wg v), in which each entry of
the data-generating matrix Wg follows independently a standard normal distribution as well. The deep network is
trained to learn this non-linear mapping (continuous target labels) from a set of examples. We generate a total of
2M examples, in which the first M examples are used for training and the last M examples are used for testing to
evaluate the generalization ability of the learned model.
In simulations, we use deep network architecture of L = 4 layers to learn the target non-linear mapping, in which the
network is thus specified by n1 -n2 -n3 -n4 , with n1 indicating the dimension of the input data and nL the dimension of
the output. We use this simple toy setting to test our idea first, and then the idea is further verified in the handwritten
digits dataset.
III.
BACKPROPAGATION AND ITS VARIANTS
We first introduce the vanilla BackProp [14] for training the deep network defined in Sec. II. We use quadratic
loss (error) function defined as E = 21 T , where T denotes a vector (matrix) transpose operation, and defines the
difference between the target and actual outputs as = y∗ − y. To backpropagate the error, we also define two
associated quantities: one is the state of neurons at k-th layer defined by sk (e.g., sL = y, s1 = v), and the other
is the weighted-sum input to neurons at k-th layer defined by hk ≡ Wk sk−1 . Accordingly, we define two related
gradient vectors:
∂E
,
∂sk
∂E
κk ≡
,
∂hk
δk ≡
(2a)
(2b)
which will be used to derive the propagation equation based on the chain rule. It is straightforward to derive
κL = − ◦ y0 , where ◦ indicates the element-wise multiplication, and y0 is the derivative of the non-linear transfer
function with respect to its argument. By applying the chain rule, we obtain the weight update equation for the top
layer as
∆WL = −ηκL (sL−1 )T ,
(3)
where η is the learning rate, and the remaining part is the gradient information, which indicates how a small perturbation to the weight affects the change of the error computed at the top (output) layer.
To update the weight parameters at lower layers, we first derive the propagating equations for gradient vectors as
follows:
δ k = (Wk+1 )T κk+1 ,
k
k
0
κ = δ ◦ (fk ) ,
(4a)
(4b)
where k ≤ L − 1. Using the above backpropagation equation, the weight at lower layers is updated as:
∆Wk = −ηκk (sk−1 )T ,
(5)
3
where k ≤ L − 1. The neural state used to update the weight parameters comes from a forward pass from the input
vector to the output vector at the top layer. A forward pass combined with a backward propagation of the error
forms the vanilla BackProp widely used in training deep networks given the labeled data [18]. To improve the training
efficiency, one usually divides the entire large dataset into a set of mini-batches, each of which is used to get the
average gradients across the examples within that mini-batch. One epoch corresponds to a sweep of the full dataset.
The learning time is thus measured in units of epochs. For one epoch, the weight is actually updated for M/B times
(B is the size of a mini-batch). This process is usually termed SGD.
Here, we briefly introduce two kinds of SGD with momentum techniques. The first one is the SGD with momentum
(SGDM). The learning equation is revised as
νt = ρt νt−1 + gt ,
∆Wt = −ηνt ,
(6a)
(6b)
where gt ≡ ∇W E(Wt−1 ) denotes the gradient estimated from the average over examples within the current minibatch, and ρt is the momentum parameter, which can be either prefixed (ρt = ρ in the classical SGDM) or varied
over learning steps (t).
The second one is Nesterov’s accelerated gradient (NAG) [17], which first implements a partial update to Wt , and
then uses the updated Wt to evaluate gradients, i.e.,
νt = ρt νt−1 − ηgt0 ,
∆Wt = νt ,
(7a)
(7b)
where gt0 ≡ ∇W E(Wt−1 + ρt νt−1 ). Note that the partial update takes an extra computational cost of Tmax M
B |W|,
where Tmax denotes the maximal number of epochs, and |W| denotes the total amount of network parameters.
IV.
REINFORCED STOCHASTIC GRADIENT DESCENT
In the above vanilla BackProp, only current gradients are used to update the weight matrix. Therefore in a nonconvex optimization, the backpropation gets easily stalled by the plateaus or saddle points on the error surface, and it
is hard to escape from these regions. During training, gradients may be very noisy with large fluctuations. If update
directions along some weight components are stochastically allowed to accumulate the history of gradient information,
while other directions still follow the current gradients, the learning performance may be boosted. This stochastic
rule of turning on accumulation may help SGD to handle the uncertainty of updating the weights. We will test this
idea in the following deep neural network learning.
To enable SGD to use previous gradient information, we define a stochastic process for updating modified gradient
g̃t used at each learning step as follows:
(
(gt )i ,
with prob. 1 − Γ(t),
(g̃t )i ←
(8)
(gt )i + (g̃t−1 )i , with prob. Γ(t).
where the stochastic reinforcement is independently applied to each weight component, and g̃t−1 contains information
about the history of the evolving gradients, and the current gradients are reinforced by the previous accumulated
gradients with a reinforcement probability defined by Γ(t). The stochastic rule in Eq. (8) is a switch-like (all-ornone) event; its smooth averaged version given g̃t−1 is E[g̃t |g̃t−1 ] = gt + Γ(t)g̃t−1 , where Γ(t) is equivalent to ρt in
SGDM with time-dependent momentum parameter. Using adaptive momentum parameter is important in boosting
the learning performance of SGDM. However, the switch-like property is able to reach a better or equivalent test
accuracy with fewer training steps. Comparisons will be made on the handwritten digits dataset.
Eq. (8) is a very simple way to re-use the previous gradient information, and forms the key component of R-SGD.
We first choose Γ(t) = 1 − γ t , where γ = γ0 e−λtep . γ0 and λ are prefixed constants, and tep refers to the learning
1
time in units of epochs. Γ(t) can be rewritten as 1 − e−t/τR , where τR ≡ − ln γ0 −λt
setting the time scale of the
ep
dynamics of the reinforcement probability. γ0 is usually fixed to a value very close to one, and λ takes a small value.
Therefore the current gradient has an increasing probability to be reinforced by the previous gradients, and retains its
instantaneous value otherwise. This reinforcement probability is not the unique choice, e.g., 1 − a0 /(t + 1)b0 (a0 and
b0 are constants) is also a candidate (discussed in Sec. V B). We show a typical trace of the reinforcement probability
and γ in Fig. 1 (a), and will test effects of hyper-parameters (γ0 , λ) on training dynamics in Sec. V A. Note that
by setting (γ0 , λ) = (1, 0), one recovers the vanilla BackProp. In all simulations, we use an exponentially-decaying
4
Algorithm 1 R-SGD (η0 , β, γ0 , λ)
m
, y∗m }M
m=1 ,
Input: data {v
Tmax = 100, mini-batch size B
t←0
γ ← γ0
η ← η0
for tep = 1 to Tmax do
for l = 1 to M/B do
PB
m
m
gt ← B1
m=1 ∇Wt−1 E(v , y∗ )
g̃t ← Eq. (8) using γ
Wt ← Wt−1 − ηg̃t
t←t+1
end for
γ ← γ0 e−λtep
η ← ηβ tep
end for
learning rate ηtep = ηtep −1 β tep , where η0 = 0.8 and β = 0.999, with a minimal learning rate of 0.02, unless otherwise
specified. R-SGD is summarized in algorithm 1.
The gradient used in R-SGD may be the accumulated one (over an unfixed or stochastic number of consecutive
steps), which contains short or long-term memory of previous gradient information (Fig. 1 (b)). Hence, the step-size
is a sum of previous gradients over a memory length L, which follows a probability Pt (L) decaying with L (Fig. 1
(b)). This stochastic process is summarized by
∆WtR−SGD = −ηt
t
X
gl ,
(9a)
l=t−L
L ∼ Pt (L) = (1 − Γ(t − L)) exp
t
X
!
ln Γ(l) ,
(9b)
l=t−L+1
where the prefactor P
indicates the probability
Qt that the memory is cleaned before accumulation. This probability
t
is normalized since
P
(L)
=
1
−
L=0 t
l=0 Γ(l) = 1, where Γ(0) = 0. In the SGDM, the momentum term is
deterministically added to the learning step size with coefficient ρt (Eq. (6)). Unfolding this process, we obtain the
step-size as
∆WtSGDM = −ηt
t
X
t
Y
ρl0 + δl,t gl ,
(10)
l=1 l0 =l+1:l6=t
where δl,t is a Kronecker delta function, and each gradient is weighted by a value smaller than one.
To show the efficiency of R-SGD, we also compare its performance with that of a state-of-the-art stochastic optimization algorithm, namely adaptive moment estimation (Adam) [15]. Adam performs a running average of gradients
and their second raw moments, which are used to adaptively change the learning step-size. We use heuristic parameters of Adam given in [15], except that η0 = 0.01 with the lowest value set to 0.001. Large η0 as we use in R-SGD
does not work in our simulations for Adam.
V.
A.
RESULTS AND DISCUSSION
Learning performance in simple synthetic dataset
We first test our idea in the simple synthetic dataset described in Sec. II. We use a 4-layer deep network architecture
as 100-400-200-10. Training examples are divided into mini-batches of size B = 100. In simulations, we use the
parameters (γ0 , λ) = (0.9995, 0.0001), unless otherwise specified. Although the chosen parameters are not optimal
to achieve the best performance, we still observe the outstanding performance of R-SGD. In Fig. 2 (a), we compare
the vanilla BackProp with R-SGD. We clearly see that the test performance is finally improved at 100-th epoch by
a significant amount (about 77.5%). Meanwhile, the training error is also significantly lower than that of BackProp.
A salient feature of R-SGD is that, at the intermediate stage, the reinforcement strategy guides SGD to escape from
possible plateau regions of high error surrounding saddle points, and finally reach a region of very nice generalization
5
(a)
(b)
0.4
0.8
1
0.6
0.95
0.4
0.2
0
λ=0.001
λ=0.0001
0.9
γ=0.99
γ=0.95
γ=0.90
0
100
t=700
t=500
t=300
0.3
probability
γ0exp(-λtep)
reinforcement probability
1
0
0.2
0.1
20 40 60 80 100
tep
200
300
400
0
500
0
2
t
4
6
memory length
8
10
FIG. 1: (Color online) Properties of R-SGD. (a) Typical behavior of the reinforcement probability (Γ(t) = 1 − γ t ) as a
function of iterations. The inset shows an exponential decay of γ (γ0 = 0.99). Note that the learning time is measured in units
of epochs in the inset. (b) Probability distribution of memory length for R-SGD at different time-steps (t = 300, 500, and 700).
(γ0 , λ) = (0.9995, 0.0001).
(b)
(a)
0.4
1.2
1
error
error
0.8
test error R-SGD
train error R-SGD
test error Adam
train error Adam
0.3
test error R-SGD
train error R-SGD
test error BackProp
train error BackProp
0.6
0.4
0.2
49.1%
0.1
77.5%
0.2
0
0
20
40
60
epoch
80
100
0
0
20
40
60
epoch
80
100
FIG. 2: (Color online) Test performance of R-SGD compared with those obtained by using the vanilla BackProp (a) and Adam
(b), based on M = 1000 training examples. Note that the test performance is improved by 77.5%, compared with BackProp,
and by 49.1%, compared with Adam.
properties. This process is indicated by the temporary peak in both training and test errors for R-SGD. Remarkably,
even before or after this peak, there are a few less significant fluctuations in both training and test errors. These
fluctuations play an important role in the exploration of the parameter space.
Compared to state-of-the-art Adam, R-SGD still improves the final test performance by a significant amount (about
49.1%, see Fig. 2 (b)). Note that Adam is able to decrease both training and test errors very quickly, but the decrease
becomes slow after about 40 epochs. In contrast, R-SGD keeps decreasing both errors by a more significant amount
than Adam, despite the presence of slightly significant fluctuations. Another key feature of Fig. 2 (b) is that, a region
in the parameter space with low training error does not generally have low test error. The training error reached by
R-SGD is clearly higher than that of Adam, but the network architecture learned by R-SGD has nicer generalization
property. This observation is consistent with a recent study of maximizing local entropy in deep networks [19].
We then study the effects of reinforcement parameters (γ0 , λ) on the learning performance, as shown in Fig. 3. If
the exponential decay rate λ is large, R-SGD over-fits the data rapidly at around 17 epochs. This is because, γ decays
rapidly from γ0 , and thus a stochastic fluctuation at earlier stages of learning is strongly suppressed, which limits
severely the exploration ability of R-SGD in the high-dimensional parameter space. In this case, R-SGD is prone to
get stuck by bad regions with poor generalization performances. However, maintaining the identical small decay rate,
6
1.2
test error (0.9995, 0.001)
train error (0.9995, 0.001)
test error (0.99995, 0.0001)
train error (0.99995, 0.0001)
test error (0.9995, 0.0001)
train error (0.9995, 0.0001)
1
error
0.8
0.035
0.03
0.6
0.025
0.02
80 85 90 95 100
0.4
0.2
0
0
20
40
60
epoch
80
100
FIG. 3: (Color online) Effects of reinforcement parameters (γ0 ,λ) on the learning performance for R-SGD based on M = 1000
training examples. The inset is an enlarged view for the later stage (the worst performance with (0.9995,0.001) is omitted for
comparison).
0.2
Adam
R-SGD
BackProp
test error
0.15
0.1
0.05
0
500
1,000
1,500
2,000
M
FIG. 4: (Color online) Data size dependence of test performance for R-SGD, BackProp and Adam. The result is averaged
over ten independent runs.
we find that a larger value of γ0 leads to a smaller test error (inset of Fig. 3). For relatively large values of γ0 , the
learning performance is not radically different.
We also study the effects of training data size (M ) on the learning performances. Clearly, we see from Fig. 4, the
test error decreases with the training data size as expected. R-SGD outperforms the vanilla BackProp, and even
Adam.
For the simple toy model, SGDM with fixed momentum parameter could outperform Adam with a careful optimization of momentum parameter (e.g., ρt = 0.9 ∀t), but R-SGD still outperforms the classical SGDM (fixed momentum
parameter) by about 14.3% when M = 1000. By adaptively changing the momentum parameter whose value is
the same as the reinforcement probability of R-SGD (i.e., a smooth averaged version of R-SGD), SGDM can reach
a similar performance to that of R-SGD. Because the synthetic data in the toy model is relatively simple and the
reinforcement probability used in the synthetic data is rapidly saturated to one (Fig. 1 (a)), it is difficult to show the
performance difference between R-SGD and SGDM with the adaptive momentum parameter.√Therefore, for MNIST
classification task in the next section, we use the reinforcement probability of Γ(t) = 1 − 1/ t + 1, which does not
rapidly approach one, and compare R-SGD with different variants of momentum-based SGD.
7
0.06
0.04
1
0.02
reinf. prob
probability
t=700
t=500
t=300
0.5
0
0
0
0
500
2
1,000
4
t
6
8
10
memory length
FIG. 5: (Color online) Probability distribution of√memory length (Pt (L)) for R-SGD at different time-steps (t = 300, 500, and
700) for a reinforcement probability Γ(t) = 1 − 1/ t + 1 (inset) we used in learning MNIST dataset.
TABLE I: Test error of different methods on MNIST dataset with 2000 testing images and different training data sizes (10000
and 15000). The test error is averaged over five independent runs with 150 epochs.
Test error
10K
15K
BackProp 0.1047 ± 0.0020 0.0948 ± 0.0030
R-SGD 0.0535 ± 0.0021 0.0465 ± 0.0018
Adam 0.0535 ± 0.0016 0.0471 ± 0.0022
Methods
B.
Learning performance in MNIST dataset
Finally, we evaluate the test performance of R-SGD on MNIST dataset. The MNIST handwritten digits dataset
contains 60000 training images and an extra 10000 images for testing. Each image is one of ten handwritten digits (0
to 9) with 28 × 28 pixels. Therefore the input dimension is n1 = 784. For simplicity, we choose the network structure
as 784-100-200-10.
Although the reinforcement probability specified by (γ0 , λ) we used in the synthetic dataset works
√ on the MNIST
dataset (see an example in Fig. 7 (c)), we found that the reinforcement probability Γ(t) = 1 − 1/ t + 1 works better
(shortening the training time to reach a lower generalization error) for MNIST dataset, and furthermore, this choice
does not saturate the reinforcement probability to one within the explored range of learning time (Fig. 5), which
offers a nice candidate to demonstrate the performance difference between R-SGD and SGDM. Fig. 6 shows that
R-SGD improves significantly over BackProp, reaching a similar test performance to that of Adam with moderate
training data size (M = 10K). R-SGD achieves a test error of 0.0535 ± 0.0021, compared with BackProp reaching
0.1047 ± 0.0020, and Adam reaching 0.0535 ± 0.0016 (as in Table I). The test error is averaged over five independent
runs (different sets of training and test examples). Note that, as training size increases (e.g., M = 15K), the test
performance of R-SGD becomes slightly better than that of Adam (Table I). Adaptive methods such as Adam often
show faster initial progress, but their performances get quickly trapped by a test error plateau [20]. In contrast, as
shown in the inset of Fig. 6, R-SGD as a non-adaptive method is able to reach a lower test error by taking a few more
epochs (note that Adam needs more computer memory to store the uncentered variance of the gradients).
As shown in Fig. 7 (a), SGDM with adaptive momentum parameter reaches a higher test error than R-SGD, which
confirms that the switch-like event plays an important role in guiding R-SGD to a good region. In R-SGD, the
reinforcement along some weight components is turned off with a finite probability, then along these directions, only
the current gradients are used (the same as those used in BackProp). But the reinforcement may be turned on along
these directions once again during training. In contrast, for SGDM, the momentum term with adaptive momentum
parameter is always applied to update all weight components during training. This mechanism difference leads to
different test performances observed in Fig. 7 (a). The training dataset may change the error surface in practice. It
seems that the performance of R-SGD is robust to this change (Table I). NAG with adaptive momentum parameter
ρt = Γ(t) learns quickly but gets trapped by a slightly higher test error. In addition, a single epoch in NAG takes an
8
0.5
error
0.4
0.3
test Adam
train Adam
test R-SGD
train R-SGD
test BackProp
train BackProp
0.07
0.065
0.06
0.055
0.05
0.045
0.2
20
40
60
80
100
0.1
0
0
20
40
60
epoch
80
100
FIG. 6: (Color online) Learning performance of R-SGD on MNIST dataset based on M = 10000 training examples with 2000
testing images, compared with BackProp and Adam. Mini-batch size B = 250. The inset shows a training trajectory (test
error) when M = 15000, suggesting that Adam can be trapped by a higher plateau.
extra computational cost of M
B |W| due to the partial update.
For fixed-momentum-parameter SGDM (ρt does not change over time, unlike the reinforcement probability in RSGD), the learning performance gets worse (Fig. 7 (b)). When applied to a deep network with ReLU activation and
cross-entropy as the objective function, R-SGD still shows competitive performance (Fig. 7 (c)).
We also use bilinear interpolation method to qualitatively analyze the test error surface of the three algorithms
(BackProp, R-SGD and Adam). Using the bilinear interpolation method [21], one can visualize the error surface
in 3D subspace spanning four high-dimensional weight configurations. These four weight configurations defined as
{Wi }4i=1 are chosen either from one learning trajectory or from solutions obtained starting from four different random
initializations. Based on these four configurations, the error function is varied as a function of a new constructed
weight matrix specified by W = β(αW1 + (1 − α)W2 )) + (1 − β)(αW3 + (1 − α)W4 ), where α ∈ [0, 1] and β ∈ [0, 1]
are two control parameters. Results are shown in Fig. 8 for one trajectory interpolation. BackProp decreases the test
error slowly, and finally reaches a plateau and get stuck there. Adam decreases the error very quickly, and reaches a
more apparent and lower plateau than that of BackProp. Remarkably, R-SGD first decreases the error quickly, which
is followed by a plateau. The plateau is then passed and finally R-SGD reaches a region with less apparent flatness.
Interpolation results of four different solutions are shown in Fig. 9. Clearly, BackProp can get stuck by different
solutions of different qualities, depending on initializations. Both Adam and R-SGD can reach solutions of nearly the
same quality, despite different initializations. However, the subspace looks sharper for R-SGD than for Adam. This
rough analysis by using bilinear interpolation method implies that the high dimensional topology of weight space seen
by the three algorithms might be intrinsically different. However, to characterize necessary properties of a good region
with nice generalization capabilities requires a theoretical understanding of the entire high dimensional weight space
in terms of analyzing some non-local quantity. Establishing a theoretical relationship between learning performances
of various SGD algorithms and the intrinsic structure of the error surface is still an extremely challenging task in
future studies.
VI.
CONCLUDING REMARKS
In this paper, we propose a new type of effective strategy to guide SGD to overcome the plateau problem typical
in deep neural network learning. This strategy takes into account previous (accumulated) gradient information
in a probabilistic way when updating current weight matrix. It introduces a stochastic reinforcement to current
gradients, and thus enhances the exploration ability of the original SGD. This strategy is essentially different from an
independently random noise added to the gradient during the learning process [22, 23]. In fact, we add time-dependent
9
0.5
0.5
(a)
0.45
0.45
R-SGD
SGDM
NAG
0.35
0.3
0.4
test error
test error
0.4
0.25
0.35
0.25
0.2
0.2
0.15
0.15
0.1
0
10
20
30
40
50
60
0.1
70
(b)
0.3
R-SGD
SGDM ρ=0.1
SGDM ρ=0.5
SGDM ρ=0.9
0
10
20
epoch
30
40
50
60
70
epoch
classification error
0.7
0.6
(c)
Adam
R-SGD
SGDM
0.5
0.4
0.3
0.2
0.1
0
20
40
60
80
epoch
FIG. 7: (Color online) Comparison of R-SGD with variants of momentum-based SGD on MNIST dataset (M = 1000 training
examples and 1000 testing examples). Mini-batch size B = 100. (a) R-SGD compared with SGDM and NAG with adaptive
momentum parameter ρt = Γ(t) (a smooth averaged version of R-SGD). (b) R-SGD compared with prefixed-momentumparameter SGDM. (c) R-SGD compared with Adam and adaptive-momentum-parameter SGDM (ρt = Γ(t)). We use the same
reinforcement schedule as that used in Sec. V A for R-SGD. Deep learning is performed with ReLU activation (f (x) = max(0, x))
and cross-entropy cost function. The classification error is measured as the fraction of the testing examples which the trained
neural network mis-classifies.
Gaussian noise to gradients during training in our simulations using default hyper-parameters [22], whose performance
could not be comparable to that of R-SGD within 100-epochs training (or it requires longer convergence time).
A similar reinforcement strategy has been used in a single layer neural network with discrete weights [24], where
local fields in a belief propagation equation are reinforced. In our work, we study deep neural networks with continuous
weights, and thus the gradients are reinforced. The reinforced belief propagation is conjectured to be related to local
entropy maximization [6, 25] in discrete neural networks. Whether R-SGD reshapes the original non-convex error
surface in some way, such that searching for a parameter region of good-quality is facilitated, remains an interesting
open question. We leave this for future work.
In the current setting, the learning performance of R-SGD is comparable to (in MNIST) or even better than that of
Adam (in the synthetic dataset), which requires one-fold more computer memory to store the uncentered variance of
the gradients. The learning step-size of Adam is adaptively changed, which means that the step-size is automatically
tuned to be closer to zero when there is greater uncertainty about the direction of the true gradient [15], while RSGD uses the stochastic reinforcement of the gradient information to deal with this kind of uncertainty, and shows
comparable and even better performance.
A recent study argued that adaptive learning methods such as Adam generalize worse than SGD or SGDM on
CIFAR-10 image classification tasks [20]. It is thus very interesting to evaluate the performance of R-SGD on more
complicated deep network model and complex datasets, which are left for future systematic studies.
R-SGD may be able to avoid vanishing or exploding gradient problem typical in training a very deep network [4],
probably thanks to accumulation of gradients used stochastically. In addition, it may take effect in recurrent neural
10
(b )
1 .2
0 .1 6 5 0
1 .0
0 .3 0 0 6
0 .8
1 .2
0 .2 3 5 0
1 .0
0 .5 2 5 0
0 .8 4 3 1
0 .4
0 .9 7 8 8
1 .1 1 4
0 .2
0 .9 6 0 0
0 .6
1 .1 0 5
1 .2 5 0
0 .4
1
. 0
1 .2 5 0
. 0
1
0 .8 1 5 0
0 .8
0 .7 0 7 5
0 .6
0 .3 8 0 0
0 .6 7 0 0
0 .5 7 1 9
t e s t e r r o r
te s t e rro r
0 .4 3 6 3
0 .0 9 0 0 0
0
0 .2
. 8
0
. 6
0
0 .0
. 4
. 8
. 6
0
α
0
. 4
0
0
β
(a )
0.
0.
. 2
(c )
0
. 0
1 .0
0 .8
β
0 .6
0 .4
α
0 .6
0 .8
. 0
1 .0
0
0
0 .2
0
0 .2
0 .4
. 2
0
0 .0 8 5 0 0
1 .2
0 .2 3 0 6
0 .3 7 6 3
1 .0
0 .5 2 1 9
0 .6 6 7 5
0 .8 1 3 1
0 .9 5 8 8
0 .6
1 .1 0 4
1 .2 5 0
0 .4
1
0
0 .2
. 6
. 4
0.
0
0 .0
0
0
. 0
. 8
β
t e s t e r r o r
0 .8
. 2
0 .6
0 .4
0 .2
0
0
. 0
1 .0
0 .8
α
FIG. 8: (Color online) Bilinear interpolation of weight configurations selected from one learning trajectory with corresponding
epochs tep = 0, 30, 60, 80. The network is trained on M = 1000 examples with the other 1000 examples as test data. B = 100.
(a) BackProp. (b) R-SGD. (c) Adam.
network training [26]. In fact, previous gradient information at each step can be weighted before accumulation
according to its importance in guiding SGD. This is a very interesting direction for future studies on fundamental
properties of R-SGD.
Acknowledgments
We are grateful to the anonymous referee for many constructive comments. H.H. thanks Dr. Alireza Goudarzi
for a lunch discussion which later triggered the idea of this work. This work was supported by the program for
Brain Mapping by Integrated Neurotechnologies for Disease Studies (Brain/MINDS) from Japan Agency for Medical
Research and development, AMED, and by RIKEN Brain Science Institute.
[1] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton. Imagenet classification with deep convolutional neural networks.
In P. Bartlett, F.c.n. Pereira, C.j.c. Burges, L. Bottou, and K.q. Weinberger, editors, Advances in Neural Information
Processing Systems 25, pages 1106–1114, 2012.
[2] Ian Goodfellow, Yoshua Bengio, and Aaron Courville. Deep Learning. MIT Press, Cambridge, MA, 2016.
[3] Herbert Robbins and Sutton Monro. A stochastic approximation method. Ann. Math. Statist., 22:400–407, 1951.
11
(b )
0 .0 9 6 0 0
(a )
0 .5
0 .1 4 6 5
0 .4 5
0 .4 0
t e s t e r r o r
0 .2 2 5 3
te s t e r r o r
0 .2 5
0 .2 4 7 5
0 .4
0 .1 5 2 0
0 .2 9 8 0
0 .1 8 8 6
0 .3 5
0 .3 0
0 .1 9 7 0
0 .3 4 8 5
0 .3
0 .2 6 1 9
0 .2 9 8 5
0 .3 3 5 1
0 .3 9 9 0
0 .4 4 9 5
0 .5 0 0 0
0 .2
1
0 .3 7 1 8
0 .2 0
. 0
0
. 8
. 6
0 .1
5
0
0 .1
0 .4 4 5 0
0
1.
0
0
0 .4 0 8 4
. 8
. 6
β
1
. 0
. 4
0
0.
0 .2
8
0.
0
0
0.
β
0 . 4
(c )
0
. 0
1 .0
0 .8
0 .2
α
0 .6
4
α
0 .0
0 . 0
0 .4
6
0.
0 . 2
. 2
0 .5
0 .0 9 0 0 0
te s t e rro r
0 .4
0 .1 3 8 7
0 .1 8 7 5
0 .3
0 .2 3 6 3
0 .2 8 5 0
0 .2
0 .3 3 3 8
0 .3 8 2 5
0 .1
0 .4 3 1 3
. 0
0
0
0 .4 8 0 0
. 2
0
. 4
α
. 6
1.
0
0
0 .8
0 .6
. 8
0 .2
. 0
0 .0
1
0 .4
0
β
FIG. 9: (Color online) Bilinear interpolation of four different solutions obtained starting from four different random initializations. The network is trained on M = 1000 examples with the other 1000 examples as test data. B = 100. (a) BackProp. (b)
R-SGD. (c) Adam.
[4] Xavier Glorot and Yoshua Bengio. Understanding the difficulty of training deep feedforward neural networks. In Yee W.
Teh and D. M. Titterington, editors, Proceedings of the Thirteenth International Conference on Artificial Intelligence and
Statistics (AISTATS-10), volume 9, pages 249–256, 2010.
[5] Haiping Huang and Yoshiyuki Kabashima. Origin of the computational hardness for learning with binary synapses. Phys.
Rev. E, 90:052813, 2014.
[6] Carlo Baldassi, Christian Borgs, Jennifer T. Chayes, Alessandro Ingrosso, Carlo Lucibello, Luca Saglietti, and Riccardo
Zecchina. Unreasonable effectiveness of learning neural networks: From accessible states and robust ensembles to basic
algorithmic schemes. Proceedings of the National Academy of Sciences, 113(48):E7655–E7662, 2016.
[7] Alan J. Bray and David S. Dean. Statistics of critical points of gaussian fields on large-dimensional spaces. Phys. Rev.
Lett., 98:150201, 2007.
[8] Yan V. Fyodorov and Ian Williams. Replica symmetry breaking condition exposed by random matrix calculation of
landscape complexity. Journal of Statistical Physics, 129(5):1081–1116, 2007.
[9] A. Choromanska, M. Henaff, M. Mathieu, G. Ben Arous, and Y. LeCun. The loss surfaces of multilayer networks. ArXiv
e-prints 1412.0233, 2014.
[10] Y. Dauphin, R. Pascanu, C. Gulcehre, K. Cho, S. Ganguli, and Y. Bengio. Identifying and attacking the saddle point
problem in high-dimensional non-convex optimization. ArXiv e-prints 1406.2572, 2014.
[11] R. Ge, F. Huang, C. Jin, and Y. Yuan. Escaping From Saddle Points — Online Stochastic Gradient for Tensor Decomposition. ArXiv e-prints 1503.02101, 2015.
[12] Jimmy Ba, Roger Grosse, and James Martens. Distributed Second-Order Optimization using Kronecker-Factored Approximations. ICLR-2017, 2017.
[13] Y. Lecun, L. Bottou, Y. Bengio, and P. Haffner. Gradient-based learning applied to document recognition. Proceedings of
the IEEE, 86:2278–2324, 1998.
[14] Rumelhart David E., Hinton Geoffrey E., and Williams Ronald J. Learning representations by back-propagating errors.
12
Nature, 323:533–536, 1986.
[15] D. Kingma and J. Ba. Adam: A method for stochastic optimization. ArXiv e-prints 1412.6980, 2014.
[16] Ilya Sutskever, James Martens, George E. Dahl, and Geoffrey E. Hinton. On the importance of initialization and momentum
in deep learning. In Proceedings of the 30th International Conference on International Conference on Machine Learning Volume 28, ICML’13, pages 1139–1147, 2013.
[17] Y. Nesterov. Introductory lectures on convex optimization: A basic course, volume 87. Springer US, 2004.
[18] Hugo Larochelle, Yoshua Bengio, Jérome Louradour, and Pascal Lamblin. Exploring strategies for training deep neural
networks. J. Mach. Learn. Res., 10:1–40, 2009.
[19] P. Chaudhari, A. Choromanska, S. Soatto, Y. LeCun, C. Baldassi, C. Borgs, J. Chayes, L. Sagun, and R. Zecchina.
Entropy-SGD: Biasing Gradient Descent Into Wide Valleys. ArXiv e-prints 1611.01838, 2016.
[20] A. C. Wilson, R. Roelofs, M. Stern, N. Srebro, and B. Recht. The Marginal Value of Adaptive Gradient Methods in
Machine Learning. ArXiv e-prints 1705.08292, 2017.
[21] D. Jiwoong Im, M. Tao, and K. Branson. An Empirical Analysis of Deep Network Loss Surfaces. ArXiv e-prints 1612.04010,
2016.
[22] A. Neelakantan, L. Vilnis, Q. V. Le, I. Sutskever, L. Kaiser, K. Kurach, and J. Martens. Adding gradient noise improves
learning for very deep networks. ArXiv e-prints 1511.06807, 2015.
[23] P. Chaudhari and S. Soatto. The effect of gradient noise on the energy landscape of deep networks. ArXiv e-prints
1511.06485, 2015.
[24] A. Braunstein and R. Zecchina. Learning by message passing in networks of discrete synapses. Phys. Rev. Lett, 96:030201,
2006.
[25] Carlo Baldassi, Alessandro Ingrosso, Carlo Lucibello, Luca Saglietti, and Riccardo Zecchina. Subdominant dense clusters
allow for simple learning and high computational performance in neural networks with discrete synapses. Phys. Rev. Lett.,
115:128101, 2015.
[26] R. Pascanu, T. Mikolov, and Y. Bengio. On the difficulty of training Recurrent Neural Networks. ArXiv e-prints 1211.5063,
2012.
| 9cs.NE
|
arXiv:1801.00212v1 [math.PR] 31 Dec 2017
BENFORD’S LAW BEYOND INDEPENDENCE: TRACKING BENFORD BEHAVIOR
IN COPULA MODELS
REBECCA F. DURST AND STEVEN J. MILLER
A BSTRACT. Benford’s law describes a common phenomenon among many naturally occurring data
sets and distributions in which the leading digits of the data are distributed with the probability of a
first digit of d base B being logB d+1
d . As it often successfully detects fraud in medical trials, voting, science and finance, significant effort has been made to understand when and how distributions
exhibit Benford behavior. Most of the previous work has been restricted to cases of independent variables, and little is known about situations involving dependence. We use copulas to investigate the
Benford behavior of the product of n dependent random variables. We develop a method for approximating the Benford behavior of a product of n dependent random variables modeled by a copula
distribution C and quantify and bound a copula distribution’s distance from Benford behavior. We
then investigate the Benford behavior of various copulas under varying dependence parameters and
number of marginals. Our investigations show that the convergence to Benford behavior seen with
independent random variables as the number of variables in the product increases is not necessarily
preserved when the variables are dependent and modeled by a copula. Furthermore, there is strong
indication that the preservation of Benford behavior of the product of dependent random variables
may be linked more to the structure of the copula than to the Benford behavior of the marginal
distributions.
C ONTENTS
1. Introduction
2. Terms and Definitions
2.1. General Mathematics and Benford’s Law
2.2. Copulas
3. Testing for Benford Behavior of a Product
4. Testing For Benford Behavior: Examples
4.1. 2-Copulas with varying dependence parameter
4.2. n-Copulas
5. Benford Distance
5.1. Cases that are ǫ away from Benford.
5.2. Consequences of an L1 bound in R2 .
6. Applications, Future Work, and Conclusion
6.1. Fitting Copulas
6.2. Conclusion
Appendix A. Proofs for supporting Lemmas and Theorems
Appendix B. Computationally Testing for Benford Behavior: Examples
References
1
2
4
4
5
6
8
8
13
14
14
16
16
16
17
17
20
23
2
REBECCA F. DURST AND STEVEN J. MILLER
1. I NTRODUCTION
Benford’s law of digit bias applies to many commonly encountered data sets and distributions.
A set of data {xi }i∈I is said to be Benford base B if the probability of observing a value xi in the
set with the first digit d (where d is any integer from 1 to B − 1) is given by the equation
d+1
.
(1.1)
Prob (first digit of{xi }i∈I is d) base B = logB
d
These probabilities monotonically decrease, from base 10 about 30.103% of the time having a
leading digit of 1 to about 4.576% of the time starting with a 9.
Benford’s law was discovered in 1881 by the astronomer-mathematician Simon Newcomb who,
looking at his logarithm table, observed earlier pages were more heavily worn than later pages. As
logarithm tables are organized by leading digit, this led him to conclude that values with leading
digit 1 occurred more commonly than values with higher leading digits. These observations were
mostly forgotten for fifty years, when Benford [Benf] published his work detailing similar biases in
a variety of settings. Since then, the number of fields where Benford behavior is seen has rapidly
grown, including accounting, biology, computer science, economics, mathematics, physics and
psychology to name a few; see [BerH1, BerH2, Mil, Nig1, Rai] for a development of the general
theory and many applications. This prevalence of Benford’s law, particularly in naturally occurring
data sets and common distributions, has allowed it to become a useful tool in detecting fraud. One
notable example of this was its in 2009 to find evidence suggesting the presence of fraud in the
Iranian elections [Batt]; while Benford’s law cannot prove that fraud happened, it is a useful tool
for determining which sets of data are suspicious enough to merit further investigation (which is
of great importance given finite resources); see for example [Nig2, Sing].
To date, most of the work on the subject has involved independent random variables or deterministic processes (see though [B–, IMS] for work on dependencies in partition problems). Our
goal below is to explore dependent random variables through copulas, quantifying the connections
between various relations and Benford behavior.
Copulas are multivariate probability distributions restricted to the unit hypercube by transforming the marginals into uniform random variables via the probability integral transform (see Section
2 for precise statements). The term copulas was first defined by Abe Sklar in 1959, when he
published what is now known as Sklar’s Theorem (see Theorem 2.7), though similar objects were
present in the work of Wassily Hoeffding as early as 1940. Sklar described their purpose as linking
n-dimensional distributions with their one-dimensional margins. See [Nels] for a detailed account
of the presence and evolution of copulas.
Quoting the Encyclopedia of Statistical Sciences, Nelsen [Nels] writes: “Copulas [are] of interest to statisticians for two main reasons: Firstly, as a way of studying scale-free measures of
dependence; and secondly, as a starting point for constructing families of bivariate distributions,
sometimes with a view to simulation.” More specifically, copulas are widely used in application in
fields such as economics and actuarial studies; for example, Kpanzou [Kp] describes applications
in survival analysis and extreme value theory, and Wu [WVS] details the use of Archimedean copulas in economic modeling and risk management. Thus, as copulas are a convenient and useful
way to model dependent random variables, they are often employed in fields relating to finance and
economics. Since many of these areas are also highly susceptible to fraud, it is worth exploring
connections between copulas and Benford’s law, with the goal to develop data integrity tests.
BENFORD’S LAW BEYOND INDEPENDENCE: TRACKING BENFORD BEHAVIOR IN COPULA MODELS
3
Essentially, since so many dependencies may be modeled through copulas, it is natural to ask
when and how often these structures will display Benford behavior. In this paper, we investigate
when data modeled by a copula is close to Benford’s law by developing a method for approximating
Benford behavior. In Section 3, we develop this method for the product of n random variables
whose joint distribution are modeled by the copula C. We then apply this method in Section 4 to
directly investigate Benford behavior for various copulas and dependence parameters. We conclude
that Benford behavior depends heavily on the structure of the copula. Figure 1 shows a small
subset of the results covered in Section 4 that display the changing χ2 values of three different
copulas compared to a Benford distribution as their dependence parameters vary. These three
plots indicate how three different copulas modeling the same marginals may display drastically
different behavior. Furthermore, Figure 2 shows the χ2 values of a copula compared to a Benford
distribution as the number of marginals increases. As we will show, the behavior seen in this plot
indicates that the product of many random variables with dependence modeled by a copula will
not necessarily level-off like products of independent random variables, the log of which we may
expect to become more uniform as the number of variables increases.
F IGURE 1. The y-axes of these three plots represent the approximate values of the
copula PDF of log10 XY mod 1 at various values of x ∈ [0, 1], where X and Y are
the marginal distributions. In each case, the marginal distributions are N(0, 1), and
Pareto(1). A Benford variable is equivalent to log10 XY mod 1 = 1 at all points,
thus the red line at y = 1 represents the PDF of a Benford variable. We clearly see
that the first two cases are close to Benford at all points. The third case, however, is
highly variable and therefore does not display Benford behavior.
4
REBECCA F. DURST AND STEVEN J. MILLER
F IGURE 2. The χ2 values comparing the behavior of the product to a Benford PDF
as the number of marginals increases. We have 8 degrees of freedom and a significance level of 0.005, so we reject the hypothesis if the value exceeds 1.3.
The results of this paper extend current techniques for testing Benford’s law to situations where
independence is not guaranteed, allowing analyses like that carried out by Cuff et. al. [CLM] on
the Weibull and Durst et. al. [D–] on the Inverse Gamma distributions to be conducted in the
case of n dependent random variables. In Section 5, we restrict ourselves to n-tuples of random
variables in which at least one is a Benford distribution and develop a concept of distance between
our joint distribution and a Benford distribution, thus developing a concept of distance from a
Benford distribution in order to understand how much deviation from Benford one might expect of
a particular distribution. We then provide an upper bound for this distance using the L1 norm of the
n C(u ,u ,...,u )
n
1 2
function N(u1 , u2, . . . , un ) = 1 − ∂ ∂u
. In doing so, we draw an interesting connection
1 ∂u2 ...∂un
between the distance from a Benford distribution and a copula’s distance from the space of copulas
for which Cuv (u, v) = 1 for all u, v in [0, 1].
2. T ERMS
AND
D EFINITIONS
We abbreviate probability density functions by PDFs and cumulative distribution functions as
CDFs, and assume all CDFs are uniformly or absolutely continuous. All results below are standard;
see the references for proofs.
2.1. General Mathematics and Benford’s Law.
Lemma 2.1 (Barbalat’s Lemma (Lemma 2.1 in [FM])). Let t → F (t) be a differentiable function
with a finite limit as t → ∞. If F ′ is uniformly continuous, then F ′ (t) → 0 as t → ∞.
Definition 2.2 (Scientific Notation). Any real number, x, can be written in the form
x = SB (x) · B n ,
(2.1)
where n is an integer. We call B the base and SB (x) the significand.
We define strong Benford’s law base B (see, for example, [BerH2, Mil]). This is the definition
we primarily use in Section 3; strong indicates that we are studying the entire significand of the
number and not just its first digit. In Section 5, we will provide insight into how one may define
a weaker version of Benford’s law that permits the probabilities to be within ǫ of the theoretical
Benford probabilities.
BENFORD’S LAW BEYOND INDEPENDENCE: TRACKING BENFORD BEHAVIOR IN COPULA MODELS
5
Definition 2.3 (Strong Benford’s Law (see Definition 1.6.1 of [Mil])). A data set satisfies the
Strong Benford’s Law Base B if the probability of observing a leading digit of at most s in base B
is logB s.
Theorem 2.4 (Absorptive Property of Benford’s Law (see page 56 of [Tao]).). Let X and Y be
independent random variables. If X obeys Benford’s law, then the product W = XY obeys
Benford’s law regardless of whether or not Y obeys Benford’s law.
2.2. Copulas. All theorems and definitions in this section are from Nelsen [Nels] unless otherwise
stated.
Remark 2.5. In [Nels], functions are defined on the extended real line, [−∞, ∞]; thus f (t) is
defined when t = ±∞. We use this notation in order to maintain consistency with [Nels], as this
is one of the central texts in copula theory.
Definition 2.6 (n-Dimensional Copula). An n-dimensional copula, C, is a function satisfying the
following properties:
(1) the domain of C is [0, 1]n ,
(2) (n-increasing) the nth -order difference of C is greater than or equal to zero,
(3) (grounded) C(u1 , u2, . . . , un ) = 0 if uk = 0 for at least one k in {1, 2, . . . , n}, and
(4) C(1, 1, . . . , 1, uk , 1, , . . . , 1) = uk for some k in {1, 2, . . . , n}.
Theorem 2.7 (Sklar’s Theorm (Theorem 2.10.9 in [Nels])). Let H be a n-dimensional distribution
function with marginal CDFs F1 , F2 , . . . , Fn . Then there exists a n-copula C such that for all
(x1 , x2 , . . . , xn ) in [−∞, ∞]n ,
H(x1 , x2 , . . . , xn ) = C(F1 (x1 ), f2 (x2 ), . . . , Fn (xn )).
(2.2)
If all Fi are continuous, then C is unique; otherwise, C is uniquely determined on Range(F1 ) ×
Range(F2 ) × · · · × Range(Fn ). Conversely, if C is a copula and F1 , F2 , . . . , Fn are cumulative
distribution functions, then the function H defined by (2.2) is a distribution function with marginal
cumulative distribution functions F1 , F2 , . . . , Fn .
Theorem 2.8 (Extension of Theorem 2.4.2 in [Nels]). Let X1 , X2 , . . . , Xn be continuous random variables. Then they are independent if and only if their copula, CX1 ,X2 ,...,Xn , is given by
CX1 ,X2 ,...,Xn (x1 , x2 , . . . , xn ) = Π(x1 , x2 , . . . , xn ) = x1 x2 · · · xn ), where Π is called the product
copula.
Theorem 2.9 (Extension of Theorem 2.4.3 in [Nels]). Let X1 , X2 , . . . , Xn be continuous random variables with copula CX1 ,X2 ...Xn . If a1 , a2 , . . . , an are strictly increasing on Range(X1 ),
Range(X2 ), . . . , Range(Xn ), respectively, then Ca1 (X1 ),a2 (X2 )...an (Xn ) = CX1 ,X2 ...Xn . Thus CX1 ,X2 ...Xn
is invariant under strictly increasing transformations of X1 , X2 , . . . , Xn .
Remark 2.10. For the following three definitions, see [Nels] page 116 for the 2-copula formulas
and page 151 for the n-copula extension.
6
REBECCA F. DURST AND STEVEN J. MILLER
Definition 2.11 (Clayton Family of Copulas). A (n-dimensional) copula in the Clayton family is
given by the equation
−α
−α
C(u1 , u2 , . . . , un ) = max {u−α
1 + u2 + · · · + un + n − 1, 0},
(2.3)
where α ∈ [−1, ∞) \ {0} is a parameter related to dependence with α = 0 as the independence
case.
Definition 2.12 (Ali-Mikhail-Haq Family of Copulas). A (n-dimensional) copula in the Ali-MikhailHaq family is given by the equation
C(u1 , u2, . . . , un ) =
(1 − α)
(
Qn
i=1
1−α(1−ui )
)
ui
−α
,
(2.4)
where α ∈ [−1, 1) is a parameter related to dependence with α = 0 as the independence case.
Definition 2.13 (Gumbel-Barnett Family of Copulas). A (n-dimensional) copula in the GumbelBarnett family is given by the equation
1 + (α log u1 − 1)(α log u2 − 1) · · · (α log un − 1)
,
(2.5)
α
where α ∈ (0, 1] is a parameter related to dependence with α = 0 as the independence case.
C(u1, u2 , . . . , un ) = exp
3. T ESTING
FOR
B ENFORD B EHAVIOR
OF A
P RODUCT
We state the results below in arbitrary dimensions but for notational convenience give the proofs
for just two dimensions (as the generalization is straightforward).
Let X1 , X2 , . . . , Xn be continuous random variables with CDFs FX1 (x1 ), . . . , FXn (xn ). Let
their joint PDF be HX1 ,X2 ,...,Xn (X1 , X2 , . . . , Xn ). By Theorem (2.7), we know there exists a copula
C such that
(3.1)
HX1 ,X2 ,...,Xn (X1 , X2 , . . . , Xn ) = C(FX1 (X1 ), . . . , FXn (Xn )).
Assume X1 , . . . , Xn are such that their copula C is absolutely continuous. This allows us to define
∂nC
the joint probability density function (see [Nels], page 27) by ∂x1 ∂x
. Furthermore, we restrict
2 ···∂xn
ourselves to Xi such that all FXi are uniformly continuous, as this allows us to use Lemma (2.1)
to later ensure that the PDFs approach zero in their right and left end limits.
From here we have the following lemma.
Lemma 3.1. Given X1 , X2 , . . . , Xn continuous random variables with joint distribution modeled
by the absolutely continuous copula C, let Ui = logB Xi for all i ≤ n and for some base B,
and let the CDFs of each Ui be Fi (ui ). Also, let fi (ui ) be the PDF of Ui for all i. Finally, let
u0 = (u1 , u2 , . . . , un−1 , s + k − (u1 + u2 + · · · + un−1 )). Then
!
n
X
Ui ) mod 1 ≤ s =
Prob (
i=1
Z
s
∞
X
0 k=−∞
Z
∞
u1 =−∞
···
Z
∞
un−1 =−∞
∂ n C(F1 (u1 ), . . . , Fn−1 (un−1), Fn (un ))
∂u1 ∂u2 · · · ∂un
u0
du1 · · · dun−1. (3.2)
BENFORD’S LAW BEYOND INDEPENDENCE: TRACKING BENFORD BEHAVIOR IN COPULA MODELS
Therefore, the PDF of (U + V ) mod 1 is given by
Z ∞
∞ Z ∞
X
∂ n C(F1 (u1 ), . . . , Fn−1 (un−1 ), Fn (un ))
···
∂u1 ∂u2 · · · ∂un
u1 =−∞
un−1 =−∞
k=−∞
u0
du1 · · · dun−1.
7
(3.3)
See Appendix A for the proof.
If (3.3) equals 1 for all s, then our product is Benford. If it is not identically equal to 1 for
all s, then at each point we may assign a value ǫs that represents our distance from a Benford
distribution. Thus we have
Z ∞
∞ Z ∞
X
∂ n C(F1 (u1), . . . , Fn−1 (un−1 ), Fn (un ))
···
ǫs = 1 −
du1 · · · dun−1 .
∂u1 ∂u2 · · · ∂un
u0
un−1 =−∞
k=−∞ u1 =−∞
(3.4)
This formulation will form the basis of Section 5.
Unfortunately, the infinite sum and improper integral in (3.3) makes it highly impractical to use
in application unless we can determine a method to closely approximate it by a finite sum and finite
∂nC
, so we have the following properties
integral. We note that (3.3) is a PDF, and so is ∂x1 ∂x
2 ···∂xn
(for notational convenience we state them in the two-dimensional case; similar results hold for
n-dimensions).
R 1 P∞
R∞
(1) 0
k=−∞ −∞ Cu1 u2 (F1 (u1 ), F2 (s + k − u1 ))f1 (u1 )f2 (s + k − u1 )du1 ds = 1.
(2)
(3)
P∞
k=−∞
R∞
−∞
R∞
−∞
Cu1 u2 (F1 (u1), F2 (s + k − u1))f1 (u1 )f2 (s + k − u1 )du1 ≥ 0 for all s.
Cu1 u2 (F1 (u1 ), F2 (s + k − u1 ))f1 (u1 )f2 (s + k − u1 )du1 → 0 as k → ±∞.
(4) Cu1 u2 (F1 (u1 ), F2 (s + k − u1 ))f1 (u1 )f2 (s + k − u1 ) → 0 as u1 → ±∞.
Property (1) is simply the definition of a PDF, and Property (2) is a direct result of the fact that a
PDF is always positive. Properties 3 and 4 are required, under Lemma 2.1, by the convergence of
the integral in Property (1) and by the convergence of the sum.
From Properties (3) and (4) and the definition of convergence we obtain the following.
Lemma 3.2 (Approximating the PDF). Given U1 , . . . , Un continuous random variables modeled
by the copula C with marginal CDFs F1 , . . . , Fn and PDFs f1 , . . . , fn , then there exist a1 , . . . , an−1 ,
b1 , . . . , bn−1 , and c1 and c2 completely dependent on the Fi such that ai < bi for all i and c1 < c2
and
Z ∞
∞ Z ∞
X
∂ n C(F1 (u1 ), . . . , Fn−1 (un−1 ), Fn (un ))
···
du1 · · · dun−1
∂u1 ∂u2 · · · ∂un
u0
un−1 =−∞
k=−∞ u1 =−∞
Z bn−1
c2 Z b1
X
∂ n C(F1 (u1 ), . . . , Fn−1 (un−1), Fn (un ))
···
=
du1 · · · dun−1 + Ea,b,c (s)
∂u1 ∂u2 · · · ∂un
u0
un−1 =an−1
k=c1 u1 =a1
(3.5)
where Ea,b,c (s) → 0 as each ai and c1 go to −∞ and each bi and c2 go to ∞. Thus, for any ǫ > 0,
there exists (for each i) |ai |, |bi |, |c1 |, and |c2 | large enough such that |Ea,b,c (s)| ≤ ǫ.
8
REBECCA F. DURST AND STEVEN J. MILLER
The proof of this claim can be found in Appendix A.
Because s only ranges from 0 to 1, we can always find a value of s that maximizes Ea,b,c for any
given set of a, b, and c and set this to be the maximum error. Furthermore, since all fi should have
similar tail-end behavior, we do not have to worry about the divergence of one canceling out the
divergence of the other. Thus, for this analysis to work, it is sufficient to understand the tail-end
behavior of only one of the marginals.
In Appendix B, we provide several examples of this method for testing for Benford behavior
computationally with two variables.
4. T ESTING F OR B ENFORD B EHAVIOR : E XAMPLES
Now that an effective method for testing the Benford behavior of copulas has been established,
we investigate how this behavior varies for specific copulas and marginals. In all χ2 tests, we have
11 degrees of freedom and a significance level of 0.005, so we reject the hypothesis if the value
exceeds 2.6. Our main interest, however, is to observe the how and if these values trend towards
this critical value.
4.1. 2-Copulas with varying dependence parameter. The following figures display the nonerror values of (3.5) at various values of s for three different copulas. The red line in each plot
indicates the constant function y = 1 which will be achieved if the product XY is exactly Benford.
For each copula, we test three different pairings of marginals: (A) X 10N(0,1) and Y 10Exp(1) ,
(B) X 10Pareto(1) and Y 10N(0,1) , and (C) X 10Pareto(1) and Y 10Exp(1) . In each case, we vary
the dependence parameter, α and compare the results to the case of independence. Our Pareto
distribution has scale parameter xm = 1 and shape parameter αp = 2. We note that in some cases
the axes must be adjusted to be able to show any change in the Benford behavior.
4.1.1. Ali-Mikhail-Haq Copula: Considering the independence case, α = 0, in Figure 3 we note
that marginal pairings (A) and (B) have an approximately Benford product when independent.
Pairing (C), however, does not. From these plots, it is evident that the Ali-Mikhail-Haq copula
displays notably consistent Benford behavior, as each plot remains very close to the independence
case as α moves over its full range. This is reinforced by the corresponding plots in Figure 4,
which display the χ2 values of each marginal pairing for each value of alpha. We point out that
although each plot indicates a general trend away from Benford behavior (the constant function
1), the values for pairing (A) are all smaller than 10−7 , making them effectively 0. Similarly, the
values for pairing (B) appear to increase linearly, but they are all of order of 10−6. The values for
pairing B vary from order 10−2 to order 10−1 , suggesting that the behavior is both significantly
less Benford and more variable than the other two pairings.
BENFORD’S LAW BEYOND INDEPENDENCE: TRACKING BENFORD BEHAVIOR IN COPULA MODELS
9
F IGURE 3. The Ali-Mikhail-Haq 2-copula (see Definition 2.12) modeled on three
different sets of marginals with varying dependence parameter α ∈ [−1, 1). The
y-axes of these plots represent the approximate values of the copula PDF of
log10 XY mod 1 at various values of x ∈ [0, 1], where X and Y are the marginal
distributions. The red line represents the Benford distribution.
F IGURE 4. The χ2 values associated to the the preceding sets of plots for the AliMihkail-Haq copula. Each shows the comparison to Benford behavior as α increases. We have 11 degrees of freedom and a significance level of 0.005, so we
reject the hypothesis if the value exceeds 2.6. Clearly, only case (C) comes close
to rejecting the hypothesis.
4.1.2. Gumbel-Barnett Copula: These plots suggest that the Gumbel-Barnett copula undergoes
even less change over α than the Ali-Mikhail-Haq copula. For pairings (A) and (B), the range for
the plots must be restricted to [0.9999, 1.0001] and [0.995, 1.010], respectively, in order to show
any change at all. Pairing (C) is not nearly Benford, so its range is expected to vary (recall that the
function described by each plot should integrate to 1 in the continuous case). We note, however,
that the value at s = 0 in pairing (C) appears to vary over a range of 0.1 as α increases. The χ2
plots in Figure 6 reinforce this interpretation, as in each case the values vary over a significantly
small range.
10
REBECCA F. DURST AND STEVEN J. MILLER
This lack of variation is likely due to the actual formula of the copula,
C(x, y) = xye−αxy
(4.1)
In this case, we have the independence copula, C(x, y) = xy multiplied by a monotonic transformation of the independence copula, e−axy . Thus, it is possible that one or both of these elements
serves to preserve the Benford properties of the marginals.
F IGURE 5. The Gumbel-Barnett 2-copula (see Definition 2.13) modeled on three
different sets of marginals with varying dependence parameter α ∈ (0, 1]. The
y-axes of these plots represent the approximate values of the copula PDF of
log10 XY mod 1 at various values of x ∈ [0, 1], where X and Y are the marginal
distributions. The red line represents the Benford distribution.
BENFORD’S LAW BEYOND INDEPENDENCE: TRACKING BENFORD BEHAVIOR IN COPULA MODELS
11
F IGURE 6. The χ2 values associated to the the preceding sets of plots for the
Gumbel-Barnett copula. Each shows the comparison to Benford behavior as α increases. We have 11 degrees of freedom and a significance level of 0.005, so we
reject the hypothesis if the value exceeds 2.6. Despite the apparent variation, none
of these cases approach the critical value.
4.1.3. Clayton Copula: Unlike the previous two examples, the Clayton copula shows notable variance over α. Although it is not shown here, the independence case for Clayton copulas is α = 0.
For pairings (A) and (B), it appears that the plots diverge farther and farther away from y = 1 as α
moves away from 0. For pairing (C), the plots appear to get more random as α grows, and there is
no suggestion that Benford behavior may develop as we depart from independence. Furthermore,
the plots in Figure 8 show χ2 values that are significantly higher than those seen for the previous
two copulas, suggesting that the dependence imposed by Clayton copula tends to heavily alter any
Benford behavior of the marginals.
12
REBECCA F. DURST AND STEVEN J. MILLER
F IGURE 7. The Gumbel-Barnett 2-copula (see Definition 2.13) modeled on three
different sets of marginals with varying dependence parameter α ∈ (0, 1]. The
y-axes of these plots represent the approximate values of the copula PDF of
log10 XY mod 1 at various values of x ∈ [0, 1], where X and Y are the marginal
distributions. The red line represents the Benford distribution.
F IGURE 8. The χ2 values associated to the the preceding sets of plots for the Clayton copula. Each shows the comparison to Benford behavior as α increases. We
have 11 degrees of freedom and a significance level of 0.005, so we reject the hypothesis if the value exceeds 2.6. Unlike the previous two copulas, only case (B)
stays below the critical value. However, the behavior of the plot suggests it will
quickly surpass the critical value as α continues to increase.
The results from these three copulas suggest that the preservation of Benford behavior relies more heavily on the underlying structure of the copula than on the Benford behavior of the
marginals. Both the Ali-Mikhail-Haq copula and the Gumbel-Barnett copula formulas contain the
independence copula, C(x, y) = xy. The Clayton copula, however, does not contain the independence copula and is also the only copula of the three to show noticeable variation as the dependence
parameter changes.
BENFORD’S LAW BEYOND INDEPENDENCE: TRACKING BENFORD BEHAVIOR IN COPULA MODELS
13
4.2. n-Copulas. The previous results suggest that the underlying copula structure has a strong influence on the Benford behavior of 2-copulas. Thus the logical next step is to investigate whether
this holds true as we increase the number of Marginals. For all χ2 tests, we have 8 degrees of freedom and again take a significance level of 0.005. In practice, this means we reject the hypothesis
if the value exceeds 1.3.
We consider the most stable of the three previous copulas, the Gumbel-Barnett copula. We
fix α = 0.1 and set the log, base 10, of all marginals to be identically distributed according to the
Normal distribution with mean 0 and variance 1, our most Benford-like marginal. We then consider
cases where the copula has 2 to 7 marginals. We can see from Figure (9) that the Benford behavior
of the Gumbel-Barnett copula begins to fall apart as marginals are added. This is in direct contrast
to what would be expected from a central-limit type property, which should become increasingly
more uniform as variables are added. This is further reinforced by the χ2 values in Figure 10
and suggests that the dependence structure imposed by the copula prevents any leveling-off from
happening.
F IGURE 9. Gumbel-Barnett copula with two to seven marginals
14
REBECCA F. DURST AND STEVEN J. MILLER
F IGURE 10. The χ2 values comparing the behavior of the product to a Benford
PDF as the number of marginals increases. We have 8 degrees of freedom and a
significance level of 0.005, so we reject the hypothesis if the value exceeds 1.3.
5. B ENFORD D ISTANCE
Now that we know that we can test for Benford behavior of a product, regardless of dependence,
it would be prudent to know how often this behavior is expected to show up. In order to do this,
we investigate if the absorptive property of Benford products is common in dependent random
variables, or if its presence relies on some sort of proximity to independence.
To get an idea of this, let W be the space of all n-tuples of continuous random variables
(X1 , X2 , . . . , Xn ) for which at least one is Benford. Now let us assume that our set of marginals,
(X1 , X2 , . . . , Xn ), form an element in W. Then we know that their product, assuming independence, will always be Benford.
From this, we can restrict our Benford distance, (3.4), to W and define it as
ǫs,W =
∞ Z
X
k=−∞
∂ n C(F1 (u1 ), . . . , Fn−1 (un−1 ), Fn (un ))
···
1 −
∂u1 ∂u2 · · · ∂un
u1 =−∞
un−1 =−∞
∞
Z
∞
u0
du1du2 · · · dun−1
where u0 is defined as in Lemma 3.1. Therefore, our problem becomes to minimize the value of
ǫs,W = 0, as a proximity to 0 should indicate proximity to a Benford distribution.
5.1. Cases that are ǫ away from Benford. Rather than directly calculating the value of ǫs,W , it
may often be more convenient to provide a bound that depends only on the copula C. Note that if
n
),...,Fn−1 (un−1 ),Fn (un ))
the value of ∂ C(F1 (u1∂u
is identically 1 for all values of (u1 , u2 , . . . , un ), then the
1 tialu2 ···∂un
value of ǫs,W will be identically 0 and our product will be Benford. Even though this case does not
cover all situations in which our product will be Benford, it suggests that a product’s distance from
n
n−1 (un−1 ),Fn (un ))
and the
Benford may be related to the distance between the function ∂ C(F1 (u1 ),...,F
∂u1 ∂u2 ···∂un
constant function, 1. This brings us to the main result of this section.
Theorem 5.1. Let X1 , X2 , . . . , Xn be continuous random variables where (X1 , X2 , . . . , Xn ) ∈ W.
Assume also that they are jointly described by a copula C, where the function N(u1 , u2 , . . . , un ) =
n
n−1 (un−1 ),Fn (un ))
1 − ∂ C(F1 (u1 ),...,F
is in L1 (Rn ). Let Ui = logB Xi for each i and some base, B, and
∂u1 ∂u2 ···∂un
,
(5.1)
BENFORD’S LAW BEYOND INDEPENDENCE: TRACKING BENFORD BEHAVIOR IN COPULA MODELS
15
let Fi be the CDFs of Ui for each i. Then the L1 distance from Benford, defined by
Z ∞
Z 1 X
∞ Z ∞
∂ n C(F1 (u1 ), . . . , Fn−1 (un−1), Fn (un ))
du1 · · · dun−1 ds
···
1 −
u0
∂u1 tialu2 · · · ∂un
un−1 =−∞
0 k=−∞ u1 =−∞
(5.2)
is bounded above by the L1 norm of N. In other words
Z ∞
Z 1 X
∞ Z ∞
∂ n C(F1 (u1 ), . . . , Fn−1 (un−1), Fn (un ))
···
du1 · · · dun−1 ds
1 −
∂u1 tialu2 · · · ∂un
u0
un−1 =−∞
0 k=−∞ u1 =−∞
≤ k1 −
∂ n C(F1 (u1 ), . . . , Fn−1 (un−1), Fn (un ))
k L1 .
∂u1 ∂u2 · · · ∂un
(5.3)
We prove this for the two-dimensional case, as the results in n-dimensions proceed similarly.
We need the following result (see Appendix A for a proof).
Lemma 5.2. Given Cuv , F (u), and G(v) as defined before, we have
Z ∞Z ∞
k1 − Cuv (u, v)kL1 =
f (u)g(v)|1 − Cuv (F (u), G(v))|dudv.
−∞
(5.4)
−∞
Proof of Theorem 5.1. From the positivity of f and g we have
Z 1 X
∞ Z ∞
f (u)g(s + k − u)(1 − Cuv (F (u), G(s + k − u)))du ds
0
≤
k=−∞ −∞
∞ Z ∞
1 X
Z
0 k=−∞
f (u)g(s + k − u)|1 − Cuv (F (u), G(s + k − u))|du ds.
(5.5)
−∞
We investigate exactly what region (5.5) covers. The lines shown in Figure 11 are the sets
Ak = {(u, v) : v = s + k − u}. We integrate f (u)g(s + k − u)(1 − Cuv (F (u), G(s + k − u)))
along each of these lines and sum the results over k. The shaded region shows the area covered
when A2 is integrated over s from 0 to 1.
As all of our sums and integrals converge absolutely, by Fubini’s theorem we may switch our
sum and integral in (5.5) and get
Z 1 X
∞ Z ∞
f (u)g(s + k − u)|1 − Cuv (F (u), G(s + k − u))|du ds
0 k=−∞
∞
X
=
k=−∞
−∞
Z
1
0
Z
∞
f (u)g(s + k − u)|1 − Cuv (F (u), G(s + k − u))|du ds.
From this, we can quickly see that for any k,
Z 1Z ∞
f (u)g(s + k − u)|1 − Cuv (F (u), G(s + k − u))|du ds
0
(5.6)
−∞
(5.7)
−∞
is the integral of f (u)g(s + k − u)|1 − Cuv (F (u), G(s + k − u))| over a region in between and
including Ak and Ak+1, just like the shaded region in Figure 11. Therefore, (5.6) is the sum of the
16
REBECCA F. DURST AND STEVEN J. MILLER
F IGURE 11. The plane broken up into a few of the sections Ak .
integrals of f (u)g(s + k − u)|1 − Cuv (F (u), G(s + k − u))| over all of these (disjoint) regions
(over all k), which is equivalent to integrating over all of R2 , giving us
Z ∞Z ∞
f (u)g(v)|1 − Cuv (F (u), G(v))|dudv.
(5.8)
−∞
−∞
Finally, we know from Lemma 5.2, we know that this is equal to k1 − Cuv (u, v)kL1 .
5.2. Consequences of an L1 bound in R2 . What Theorem 5.1 provides is a way to understand the
behavior of our probabilities. To see this, let S ⊂ [0, 1] be the region over which ǫs,W > ǫN . If ǫs,W
is large on S, then the measure of RS must be small in order to conform to (5.3), which requires that
1
if k1 − Cuv (u, v)kL1 ≤ ǫN , then 0 ǫs,W ds ≤ ǫN as well. In fact, the following corollary proves
that Theorem 5.1 provides useful information regarding how large |S| can be.
Corollary 5.3. Let S ⊂ [0, 1] be the set {s : ǫs,W ≥ ǫ}. Then
k1 − Cuv (u, v)kL1
.
ǫ
Proof. This result comes directly from Markov’s Inequality:
Z
1 1
k1 − Cuv (u, v)kL1
|{s : ǫs,W ≥ ǫ}| ≤
ǫs,W ≤
.
ǫ 0
ǫ
|S| ≤
6. A PPLICATIONS , F UTURE WORK ,
AND
(5.9)
(5.10)
C ONCLUSION
6.1. Fitting Copulas. The results of Section 3 allow us to determine the Benford behavior of the
product n distributions jointly modeled by a specific copula. However, we may wish to go in the
BENFORD’S LAW BEYOND INDEPENDENCE: TRACKING BENFORD BEHAVIOR IN COPULA MODELS
17
other direction and, instead, find a copula that best fits n correlated data sets. Statisticians have
several methods for testing the goodness-of-fit to find the best choice of copula in these situations
(see [GQR] for some examples and an analysis of several forms of goodness-of-fit tests), but it is
not known whether or not these goodness-of-fit tests take Benford behavior into account. That is
to say, will the prescribed copula mimic the Benford behavior observed in the data?
The results Section 4 have shown us that the product of the same set of marginals will not
display the same Benford behavior when modeled by different copulas. Thus, Benford behavior
is not guaranteed. A natural next step is to investigate how the goodness-of-fit of a copula may or
may not be correlated with how well it preserves the expected Benford behavior of the product of
two or more marginals. A comparison between the L1 norm and well-known goodness of fit tests
would enable us to see whether or not a strong Benford fit corresponds to a well-fit distribution as
a whole. Furthermore, if a stronger Benford fit may be shown to correspond to a smaller L1 bound,
then we may be able to define this bound as a new goodness of fit test for distributions with one or
more Benford marginals.
6.2. Conclusion. In fields such as actuarial sciences and statistics Benford’s law is useful for
fraud detection. Furthermore, copulas are a highly effective tool for modelling systems with dependencies. In Section 3 we demonstrated that Benford behavior for dependent variables modeled
by a copula may be detected and therefore analyzed to investigate the product of the variables.
Thus these results indicate that the Benford’s law methods used by professionals on single-variate,
and/or independent data sets are now at the disposal of individuals who wish to model dependent
data via a copula. We then applied these results in Section 4 where we observed that the preservation of Benford behavior appears to rely more heavily on the structure of the copula than on the
marginals.
Essentially, the results of Section 3 permit analyses like those carried out in [CLM] and [D–] in
which a known distribution, in these cases the Weibull distribution and the inverse-gamma distribution, is analyzed to determine the conditions under which Benford behavior should arise. Once
these conditions are established, any non-Benford data set which is expected to come from such a
distribution may be considered suspicious enough to warrant a fraud investigation. In the case of
copulas, the results of Section 3 allow one to conduct this exact method of analysis on the product
of n random variables jointly modeled by a copula C.
Finally, in Section 5 we encountered a useful consequence of of considering a distribution’s L1
distance from a Benford distribution to determine a useful bound for this Benford distance. We
determined that the Benford distance of a product of n random variables will always be bounded
above by the distance between the copula PDF and the class of copulas whose PDFs are identically
1.
A PPENDIX A. P ROOFS
FOR SUPPORTING
L EMMAS
AND
T HEOREMS
Proof of Lemma 3.1.
Given X and Y continuous random variables with joint distribution modeled by the absolutely
continuous copula C, Let U = logB X and V = logB Y , for some base, B, and let the (marginal)
CDFs of U and V be F (u) and G(v), respectively. Also, let f (u) and g(v) be the PDFs of U and
18
REBECCA F. DURST AND STEVEN J. MILLER
V , respectively. Then
Prob ((U + V ) mod 1 ≤ s)
!
Z s X
∞ Z ∞
Cuv (F (u), G(s + k − u))f (u)g(s + k − u)du .
=
0
k=−∞
Therefore, the PDF of (U + V ) mod 1 is given by
∞ Z ∞
X
Cuv (F (u), G(s + k − u))f (u)g(s + k − u)du.
k=−∞
(A.1)
−∞
(A.2)
−∞
Proof. By the invariance of copulas under monotonically increasing functions (Theorem 2.9), we
know that the joint CDF of U and V is given by the same copula as X and Y . Thus, the joint CDF
of U and V is given by
C(F (U), G(V )).
(A.3)
Then, by definition, the joint PDF of U and V is given by the mixed partial derivative.
∂ ∂
∂
C(F (u), G(v)) = Cuv (F (u), G(v))f (u)g(v) + Cu (F (u), G(v)) f (u)
∂v ∂u
∂v
= Cuv (F (u), G(v))f (u)g(v).
(A.4)
= 0 since all dependence between U and V is modeled by C.
Note that we assume that du
dv
Note, also, that Prob (XY ≤ 10s ) = Prob ((U + V ) ≤ s). Thus we have
Prob ((U + V ) mod 1 ≤ s)
Z s+k−u
∞ Z ∞
X
Cuv (F (u), G(v))f (u)g(v)dvdu.
=
k=−∞
u=−∞
(A.5)
v=k−u
If XY is Benford, then (A.5) will equal s for all s. It is, however, easier to test the PDF then the
CDF. So we differentiate with respect to s. Let C1 (u, v) be the antiderivative of Cuv (F (u), G(v))f (u)g(v)
with respect to v. Then
Z s+k−u
∞ Z ∞
∂ X
Cuv (F (u), G(v))f (u)g(v)dvdu
∂s k=−∞ u=−∞ v=k−u
Z ∞
∞
∂ X
=
(
(C1 (u, s + k − u) − C1 (u, k − u))du
∂s k=−∞ u=−∞
∞ Z ∞
X
Cuv (F (u), G(s + k − u))f (u)g(s + k − u)du.
(A.6)
=
k=−∞
−∞
Proof of Lemma 3.2.
Given U and V , continuous random variables modeled by the copula C with marginals F and
G, respectively, there exist a1 , a2 , b1 , and b2 completely dependent on F or G such that a1 < a2
BENFORD’S LAW BEYOND INDEPENDENCE: TRACKING BENFORD BEHAVIOR IN COPULA MODELS
19
and b1 < b2 , and
∞ Z
X
k=−∞
=
∞
Cuv (F (u), G(s + k − u))f (u)g(s + k − u)du
−∞
b2 Z
X
k=b1
a2
Cuv (F (u), G(s + k − u))f (u)g(s + k − u)du + Ea,b (s)
(A.7)
a1
where Ea,b (s) → 0 as a1 , b1 → −∞ and a2 , b2 → ∞. Thus, for any ǫ > 0, there exists |a1 |, |a2 |,
|b1 |, and |b2 | large enough such that |Ea,b (s)| ≤ ǫ.
Proof. Since both the sum and the integral are convergent, the proof for a1 , a2 and b1 , b2 are nearly
identical, so we only provide the work here for a1 and a2 . The same steps may be used in the proof
for b1 and b2 . We also know that Cuv (F (u), G(s + k − u))f (u)g(s + k − u) must go to 0 as u goes
to ±∞ because of this convergence. Thus we choose to prove the case where f and/or g converge
faster than Cuv . If Cuv were to converge faster, the results derived here would still suffice. We
prove that for any ǫ > 0 we can find a1 and a2 such that, for all u ≤ a1 and all u ≥ a2 , we have
|Cuv (F (u), G(s + k − u))f (u)g(s + k − u)| ≤ ǫ.
Let ǫ > 0, set s and k to be constant, and assume Cuv is nonzero everywhere. If Cuv is zero
at any point, then we have a trivial case. Because F and G are CDFs, we know that f → 0 as
u → ±∞ and g → 0 as −u → ±∞, thus, we may choose af 1 , af 2 , ag1 , and ag2 such that, for all
u ≤ af 1 and all u ≥ af 2 , we have
r
ǫ
f (u) ≤
.
(A.8)
Cuv (F (u)G(s + k − u))
The same can be done for g such that, for all u ≥ ag1 and all u ≤ ag2 , we have
r
ǫ
.
g(s + k − u) ≤
Cuv (F (u)G(s + k − u))
(A.9)
Thus, we let a1 = min{af 1 , ag1 } and a2 = max{af 2 , ag2 }. then we have, for all u ≤ a1 and all
u ≥ a2 , we have
|Cuv (F (u), G(s + k − u))f (u)g(s + k − u)| ≤ ǫ.
Proof of Lemma 5.2.
Given Cuv , F (u), and G(v) as defined in Theorem 5.1, we have
Z ∞Z ∞
k1 − Cuv (u, v)kL1 =
f (u)g(v)|1 − Cuv (F (u), G(v))|dudv.
−∞
(A.10)
−∞
Proof. We know that u and v are defined on [0, 1]. Thus,
Z 1Z 1
k1 − Cuv (u, v)kL1 =
|1 − Cuv (u, v)|dudv.
0
0
(A.11)
20
REBECCA F. DURST AND STEVEN J. MILLER
However, by a simple change of variables u → F (u), v → G(v) (defended as CDFs, just like
before, so their derivatives are f (u) and g(v), both of which are greater than or equal to 0), we get
Z ∞Z ∞
k1 − Cuv (u, v)kL1 =
f (u)g(v)|1 − Cuv (F (u), G(v))|dudv.
(A.12)
−∞
−∞
A PPENDIX B. C OMPUTATIONALLY T ESTING
FOR
B ENFORD B EHAVIOR : E XAMPLES
In this section, we use Clayton copulas (see Definition 2.11) to determine the Benford behavior
of different combinations of marginals. We specifically look at marginals of the form X = 10U
and Y = 10V , where U and V are N[0, 1] or Exp[1]. In all analyses, we let α = 2 and B = 10. We
also provide the independence case for each set of marginals to allow for comparison.
Case 1: U and V ∼ N[0, 1].
Given our definition of X and Y , (3.3) we first determine acceptable values for a1 , b1 , a2 . and
b2 by using an error analysis to test whether or not −10 and 10 should be acceptable values for a1
and a2 .
We generated a list of the error caused by truncating the integral at these values for various
values of s. The first value of each triple in the list is s. The second is the lower error and the
third is the upper error. To determine the error caused by truncating the integral, we used the
approximation method detailed in Section 3. As the list shows, the error is on the order of 10−22
or smaller, indicating that our selections for a1 and a2 are good bounds. We took the sum from
k = −20 to k = 20 because we know this will be sufficient, as indicated by the convergence in
Figure (12) below.
In[262]:= errorsb =
Table[{N[Log[10, s]], ea[Log[10, s]], eb[Log[10, s]]}, {s, 1, 9}]
Out[262]= {{0., 6.86784*10^-22, 1.28213*10^-22}, {0.30103,
9.38169*10^-24, 1.28257*10^-22}, {0.477121, 3.03058*10^-25,
1.28274*10^-22}, {0.60206, 2.74232*10^-26,
1.28266*10^-22}, {0.69897, 4.3443*10^-27,
1.28249*10^-22}, {0.778151, 9.77379*10^-28,
1.28234*10^-22}, {0.845098, 2.79567*10^-28,
1.28223*10^-22}, {0.90309, 9.52164*10^-29,
1.28216*10^-22}, {0.954243, 3.70245*10^-29, 1.28213*10^-22}}
We now plot in Figure (12) the value of our truncated form of our PDF for different values of s.
The line y = 1 is included to demonstrate how close to 1 our PDF is for all values of s, suggesting
that the product of X and Y , with joint PDF modeled by a Clayton copula with α = 2 should
display Benford behavior.
BENFORD’S LAW BEYOND INDEPENDENCE: TRACKING BENFORD BEHAVIOR IN COPULA MODELS
21
F IGURE 12. U ∼ N[0, 1], V ∼ N[0, 1]
Case 2: U ∼ N[0, 1] and V ∼ Exp[1].
A similar analysis as before was conducted on this new set of variables. Through an identical
analysis, we defined the bounds for our integral to be a = −5 and b = 10, and provide the
accumulated errors in the code below where the first term in each pair is s and the second and third
are the lower and upper errors, respectively. As we can see, the errors are still very very small.
In[419]:= Table[{N[Log[10, s]], ea2[Log[10, s]], eb2[Log[10, s]]},
{s, 1, 9}]
Out[419]= {{0., 3.30411*10^-21, 1.23628*10^-22}, {0.30103,
2.43577*10^-21, 1.27151*10^-22}, {0.477121, 2.03887*10^-21,
1.31758*10^-22}, {0.60206, 1.79746*10^-21,
1.32924*10^-22}, {0.69897, 1.63021*10^-21,
1.32387*10^-22}, {0.778151, 1.50526*10^-21,
1.31045*10^-22}, {0.845098, 1.40717*10^-21,
1.2933*10^-22}, {0.90309, 1.32741*10^-21,
1.27456*10^-22}, {0.954243, 1.26084*10^-21, 1.25536*10^-22}}
We now plot in Figure (13) the value of our truncated form of our PDF for various s. We again
note how close the PDF remains to 1 for all values of s, suggesting that the product of X and Y ,
with joint PDF modeled by a Clayton copula with α = 2 should display Benford behavior.
22
REBECCA F. DURST AND STEVEN J. MILLER
F IGURE 13. U ∼ N[0, 1], V ∼ Exp[1]
As before, this is backed up by the following simulation.
Simulation 2
Case 3: U ∼ Exp[, 1] and V ∼ Exp[1].
Finally, we conduct our analysis on the case of two exponentials. Our error terms for a = 25 are
generated in the code below (By inspection, we can tell that Cuv (F (u), G(s + k − u))f (u)g(s +
k − u) will be zero for negative values of u). Again we choose k from 0 to 50, and the first term in
each pair is s.
In[363]:= Table[{N[Log[10, s]], N[eb1[Log[10, s]]]}, {s, 1, 9}]
Out[363]= {{0., 5.57839*10^-11}, {0.30103, 5.73736*10^-11}, {0.477121,
5.94524*10^-11}, {0.60206, 5.99786*10^-11}, {0.69897,
5.97362*10^-11}, {0.778151, 5.91306*10^-11}, {0.845098,
5.83566*10^-11}, {0.90309, 5.75112*10^-11}, {0.954243,
5.66447*10^-11}}
Now that we know a = 25 provides a small enough error, we plot, once again, the PDF for
various values of s, as shown in Figure (14). We quickly see that the PDF does not converge to 1
and actually changes for each value of s. Even though we only take our sum out to k = ±50, this
is enough to suggest that Benford behavior is unlikely.
BENFORD’S LAW BEYOND INDEPENDENCE: TRACKING BENFORD BEHAVIOR IN COPULA MODELS
23
F IGURE 14. U and V ∼ Exp[1]
Checking the Marginals
To understand why this might be the case, we took a look at the marginal distributions. We note
that X = 10U where U ∼ N[0, 1] is a closely Benford distribution with χ2 ≈ 0.9918, but Y = 10V
where V ∼ Exp[1] is not, with χ2 ≈ 0.7084. Thus, in the independent case we would expect that
two variables modeled like X, or any product with X, should yield a Benford distribution. The
product of two variables modeled like Y , however, should not be Benford.
R EFERENCES
[Batt]
[B–]
[Benf]
[BerH1]
[BerH2]
[CLM]
[D–]
[FM]
[GM1]
[GM2]
W. Battersby, Statistics hint at fraud in Iranian election New Scientist (June 24, 2009).
T. Becker, D. Burt, T. C. Corcoran, A. Greaves-Tunnell, J. R. Iafrate, J. Jing, S. J. Miller, J. D. Porfilio,
R. Ronan, J. Samranvedhya, F. Strauch, and B. Talbut, Benford’s Law and Continuous Dependent
Random Variables, preprint; arXiv version: http://arxiv.org/pdf/1309.5603.
F. Benford, The Law of Anomalous Numbers, Proceedings of the American Philosophical Society 78
(1938), no. 4, 551-572. Retrieved from http://www.jstor.org/stable/984802.
A. Berger and T. P. Hill, Benford Online Bibliography, http://www.benfordonline.net.
A. Berger and T. P. Hill, An Introduction to Benford’s Law, Princeton University Press, Princeton, 2015.
V. Cuff, A. Lewis, and S. J. Miller, The Weibull Distribution and Benford’s law, Involve 8 (2015), no.
5, 859-874. https://doi.org/10.2140/involve.2015.8.859.
R. F. Durst, C. Huynh, A. Lott, S. J. Miller, E. A. Palsson, W. Touw and G. Vriend, The Inverse Gamma
Distribution and Benford’s Law, preprint 2017. https://arxiv.org/pdf/1609.04106.
F. A. C. C. Fontes and L. Magni, A Generalization of Barbalat’s Lemma with Applications to Robust Model Predictive Control, Paper presented at Sixteenth International Symposium on Mathematical
Theory of Networks and Systems, Leuven, Belgium (July 2004).
C. Genest and J. MacKay, Copules archimadiennes et familles de lois bi- dimensionnelles dont les
marges sont donnaes, Canad J Statist 14 (1986), 145-159.
C. Genest and J. MacKay, The joy of copulas: Bivariate distributions with uni- form marginals, Amer
Statist 40 (1986), 280-285.
24
REBECCA F. DURST AND STEVEN J. MILLER
[GQR]
[IMS]
[Kp]
[Mil]
[MusSch]
[Nels]
[Nig1]
[Nig2]
[Rai]
[SF]
[Sing]
[Tao]
[WVS]
C. Genest, J.-F. Quessy and B. Remillard, Goodness-of-fit Procedures for Copula Models Based on the
Probability Integral Transformation, Scandinavian Journal of Statistics (2006) https://doi.org/
10.1111/j.1467-9469.2006.00470.
J. Iafrate, S. J. Miller and F. Strauch, Equipartitions and a distribution for numbers: A statistical model
for Benford’s law, Physical Review E 91 (2015), no. 6, 062138 (6 pages).
T. A. Kpanzou, Copulas in Statistics, African Institute of Mathematical Sciences (AIMS) (May 2007).
S. J. Miller (editor), Theory and Applications of Benford’s Law, 2015, Princeton University Press,
Princeton, NJ.
C. Muscalu and W. Schlag, The Hardy-Littlewood maximal function. In Cambridge studies in advanced
mathematics: Vol. 137. Classical and Multilinear Harmoic Analysis (Vol. 1), Cambridge University
Press, Cambridge, UK, 2013.
R. B. Nelsen, Springer Series in Statistics: An introduction to copulas (2nd ed.), Springer Science+Business Media, New York, NY, 2006
M. Nigrini, I’ve Got Your Number: How a Mathematical Phenomenon Can Help CPAs Uncover Fraud
and Other Irregularities, Journal of Accountancy (May 1999), 79–83.
M. Nigrini, The Use of Benford’s Law as an Aid in Analytical Procedures, Auditing: A Journal of
Practice & Theory 16 (1997), no. 2, 52–67.
R. A. Raimi, The First Digit Problem, Amer. Math. Monthly 83 (1976), no. 7, 521–538.
P. D. Scott and M. Fasli, (2001). Benford’s Law: An empirical investigation and a novel explanation,
CSM Technical Report No. 349, Department of Computer Science, Univ. Essex, http://dces.
essex.ac.uk/technical-reports/2001/CSM-349.pdf.
T. W. Singleton, Understanding and Applying Benford’s Law, ISACA Journal 3 (2011), 1-4.
T. Tao, An Epsilon of Room, II: pages from year three of a mathematical blog, American Mathematical
Society, 2011
F. Wu, E. Valdez and M. Sherris, Simulating Exchangeable Multivariate Copulas and its Applications,
Communications in Statistics - Simulation and Computation 36, (2006).
E-mail address: [email protected]
D EPARTMENT OF M ATHEMATICS AND S TATISTICS , W ILLIAMS C OLLEGE , W ILLIAMSTOWN , MA 01267
Current address: Department of Mathematics, Brown University, Providence, RI 02912
E-mail address: [email protected], [email protected]
D EPARTMENT OF M ATHEMATICS AND S TATISTICS , W ILLIAMS C OLLEGE , W ILLIAMSTOWN , MA 01267
| 10math.ST
|
Further study on the maximum number of bent
components of vectorial functions
arXiv:1801.06542v1 [cs.IT] 21 Jan 2018
Sihem Mesnager1 , Fengrong Zhang2 , Chunming Tang3 , Yong Zhou2
1. LAGA, Department of Mathematics, University of Paris VIII
(and Paris XIII and CNRS), Saint–Denis cedex 02, France.
E-mail: [email protected]
2. School of Computer Science and Technology, China University
of Mining and Technology, Xuzhou, Jiangsu 221116, China.
E-mail: {zhfl203,yzhou}@cumt.edu.cn
3. School of Mathematics and Information, China West Normal University, Nanchong, Sichuan
637002, China.
E-mail: [email protected]
Abstract. In 2018, Pott, at al. have studied in [IEEE Transactions on Information
Theory. Volume: 64, Issue: 1, 2018] the maximum number of bent components of
vectorial function. They have presented serval nice results and suggested several open
problems in this context. This paper is in the continuation of their study in which
we solve two open problems raised by Pott et al. and partially solve an open problem
raised by the same authors. Firstly, we prove that for a vectorial function, the property
of having the maximum number of bent components is invariant under the so-called
CCZ equivalence. Secondly, we prove the non-existence of APN plateaued having the
maximum number of bent components. In particular, quadratic APN functions cannot
have the maximum number of bent components. Finally, we present some sufficient
conditions that the vectorial function defined from F22k to F22k by its univariate
representation:
αx
2i
x+x
2k
+
ρ
X
γ
t
(j) 2 j
x
+
j=1
ρ
X
γ
t +k
(j) 2 j
x
j=1
!
has the maximum number of components bent functions, where ρ ≤ k. Further, we
i
k
t1
t1 +k
show that the differential spectrum of the function x2 (x + x2 + x2 + x2
+
t2 +k
t2
x2 + x2
) (where i, t1 , t2 satisfy some conditions) is different from the binomial
i
k
function F i (x) = x2 (x + x2 ) presented in the article of Pott et al.
Finally, we provide sufficient and necessary conditions so that the functions
T r12k
αx
2i
T re2k (x)
+
ρ
X
j=1
γ
(j)
j
(T re2k (x))2
!!
are bent.
Keywords: Vectorial functions, Boolean functions, Bent functions, Nonlinearity,
APN functions, Plateaued functions, CCZ equivalence.
1
Introduction
Vectorial (multi-output) Boolean functions, that is, functions from the vector space
Fn2 (of all binary vectors of length n) to the vector space Fm
2 , for given positive
integers n and m. These functions are called (n, m)-functions and include the (singleoutput) Boolean functions (which correspond to the case m = 1). In symmetric
cryptography, multi-output functions are called S-boxes. They are fundamental parts
of block ciphers. Being the only source of nonlinearity in these ciphers, S-boxes play
a central role in their robustness, by providing confusion (a requirement already
mentioned by C. Shannon), which is necessary to withstand known (and hopefully
future) attacks. When they are used as S-boxes in block ciphers, their number m of
output bits equals or approximately equals the number n of input bits. They can
also be used in stream ciphers, with m significantly smaller than n, in the place of
Boolean functions to speed up the ciphers.
We shall identify Fn2 with the Galois field F2n of order 2n but we shall always use
n
F2 when the field structure will not really be used. The component functions of F are
the Boolean functions v · F , that is, x ∈ F2n 7→ T r1m (vF (x)), where “·” stands for an
inner product in F2m (for instance: u·v := T r1m (uv), ∀u ∈ F2m , v ∈ F2m where ”T r1m ”
denotes the absolute trace over F2m ). In order to classify vectorial Boolean functions
that satisfy desirable nonlinearity conditions, or to determine whether, once found,
they are essentially new (that is, inequivalent in some sense to any of the functions
already found) we use some concepts of equivalence. For vectorial Boolean functions,
there exist essentially two kinds concepts of equivalence: the extended affine EAequivalence and the CCZ-equivalence (Carlet-Charpin-Zinoviev equivalence). Two
(n, r)-functions F and F ′ are said to be EA-equivalent if there exist affine automorphisms L from F2n to F2n and L′ from F2r to F2r and an affine function L′′ from
F2n to F2r such that F ′ = L′ ◦ F ◦ L + L′′ . EA-equivalence is a particular case of
CCZ-equivalence [4]. Two (n, r)-functions F and F ′ are said to be CCZ-equivalent
if their graphs GF := {(x, F (x)), x ∈ F2n } and G′F := {(x, F ′ (x)), x ∈ F2n } are
affine equivalent, that is, if there exists an affine permutation L of F2n × F2m such
that L(GF ) = G′F .
A standard notion of nonlinearity of an (n, m)-function F is defined as
N (F ) = min
nl(v · F ),
⋆
v∈F2m
(1)
where v·F denotes the usual inner product on F2m and nl(·) denotes the nonlinearity
of Boolean functions (see definition in Section 2). From the covering radius bound, it
is known that N (F ) 6 2n−1 − 2n/2−1 . The functions achieving this bound are called
(n, m)-bent functions. Equivalently, a vectorial Boolean function F : F2n → F2m is
said to be a vectorial bent function if all nonzero component functions of F are bent
(Boolean) functions. Bent Boolean functions have maximum Hamming distance to
the set of affine Boolean functions. The notion of bent function was introduced by
Rothaus [9] and attracted a lot of research of more than four decades. Such functions
are extremal combinatorial objects with several areas of application, such as coding
2
theory, maximum length sequences, cryptography. A survey on bent function can
be found in [6] as well as the book [15].
In [16], it is shown that (n, m)-bent functions exist only if n is even and m 6 n/2.
The notion of nonlinearity in (1) (denoted by N ), was first introduced by Nyberg
in [16], which is closely related to Matsui’s linear attack [13] on block ciphers. It has
been further studied by Chabaud and Vaudenay [7]. The nonlinearity is invariant under CCZ equivalence (and hence under extended affine equivalence). Budaghyan and
Carlet have proved in [1] that for bent vectorial Boolean functions, CCZ-equivalence
coincides with EA-equivalence.
The problem of construction vectorial bent functions has been considered in the
literature. Nyberg [16] investigated the constructions of vectorial bent functions; she
presented two constructions based on Maiorana-McFarland bent functions and PS
bent functions, respectively. In [17], Satoh, Iwata, and Kurosawa have improved the
first method of construction given in [16] so that the resulting functions achieve the
largest degree. Further, serval constructions of bent vectorial functions have been
investigated in some papers [5,10,11,14,18]. A complete state of the art can be found
in [15] (Chapter 12).
Very recently, Pott et al.[8] considered functions F2n → F2n of the form F i (x) =
k
i
2
x (x + x2 ), where n = 2k, i = 0, 1, · · · , n − 1. They showed that the upper bound
of number of bent component functions of a vectorial function F : F2n → F2n is
k
i
2n − 2n/2 (n even). In addition, they showed that the binomials F i (x) = x2 (x + x2 )
have such a large number of bent components, and these binomials are inequivalent
k
to the monomials x2 +1 if 0 < i < k. Further, the properties (such as differential
properties and complete Walsh spectrum) of the functions F i were investigated.
In this paper, we will consider three open problems raised by Pott et al [8].
In the first part, we prove that CCZ equivalence is preserved for vectorial functions
having the maximum number of bent components. Next, we consider APN plateaued
functions and investigate if they can have the maximum number of bent components.
We shall give a negative answer to this question. Finally, we consider the bentness
property of functions F22k → F22k of the form
ρ
ρ
X
X
tj +k
tj
i
k
,
γ (j) x2
(2)
γ (j) x2 +
G(x) = αx2 x + x2 +
j=1
j=1
where m ≤ k, γ (j) ∈ F2kand 0 ≤ tj ≤ k be a nonnegative integer.
In particular, we
t2 +k
t2
t1 +k
t1
k
t2
2
2
2
2
2
2
are inequivalent to
+x +x
x+x +x +x
show the functions x
t
k
x2 2 (x + x2 ), where t1 = 1 and gcd(t2 , k) 6= 1 . Here we use the concept of CCZequivalence when we speak about the equivalence of functions. The rest of the paper
is organized as follows. Some preliminaries are given in Section 2. In Section 3, we
prove our result on the stability under CCZ equivalence of a function having the
maximum number of bent components which solve Problem 4 in [8]. Next, in Section
4, we prove that APN plateaued functions cannot have the maximum number of
bent components, which partially solves Problem 8 in [8]. Finally, in Section 5 we
3
investigate Problem 2 in [8]. To this end, we provide several functions defined as
G(x) = xL(x) on F22k (where L(x) is a linear function on F22k ) such that the number
of bent components T r12k (αF (x)) is maximal.
2
Preliminaries and notation
Throughout this article, kEk denotes the cardinality of a finite set E, the binary field
is denoted by F2 and the finite field of order 2n is denoted by F2n . The multiplicative
group F∗2n is a cyclic group consisting of 2n − 1 elements. The set of all Boolean
functions mapping from F2n (or Fn2 ) to F2 is denoted by Bn .
Recall that for any positive integers k, and r dividing k, the trace function from
F2k to F2r , denoted by T rrk , is the mapping defined as:
k
T rrk (x)
:=
−1
r
X
r
ir
2r
k−r
x2 = x + x2 + x2 + · · · + x2
.
i=0
In particular, the absolute trace over F2 of an element x ∈ F2n equals T r1n (x) =
Pn−1 2i
i=0 x .
There exist several kinds of possible univariate representations (also called trace,
or polynomial, representations) of Boolean functions which are not all unique and
use the identification between the vector-space Fn2 and the field F2n . Any Boolean
function over Fn2 can be represented
P2n −1 in aj unique way as a polynomial in one variable
x ∈ F2n of the form f (x) = j=0 aj x , where a0 , a2n −1 ∈ F2 , aj ’s are elements of
F2n for 1 ≤ j < 2n − 1 such that aj 2 = a2i mod (2n −1) . The binary expansion of j is
j = j0 +j1 2+· · · jn−1 2n−1 and we denote j̄ = (j0 , j1 , · · · , jn−1 ). The algebraic degree
of f equals max{wt(ī) | aj 6= 0, 0 ≤ j < 2n } where wt(ī) = j0 + j1 + · · · + jn−1 .
Affine functions (whose set is denoted by An ) are those of algebraic degree at most
1. The Walsh transform of f ∈ Bn at λ ∈ F2n is defined as
P
n
(−1)f (x)+T r1 (λx) .
Wf (λ) =
x∈F2n
The nonlinearity of f ∈ Bn is defined as the minimum Hamming distance to the
set of all n-variable affine functions, i.e.,
nl(f ) = ming∈An d(f, g).
where d(f, g) is the Hamming distance between f and g. Following is the relationship
between nonlinearity and Walsh spectrum of f ∈ Bn
nl(f ) = 2n−1 −
1
max |Wf (λ)|.
2 λ∈F2n
P
By Parseval’s identity λ∈F2n Wf (λ)2 = 22n , it can be shown that max{|Wf (λ)| :
n
n
λ ∈ F2n } ≥ 2 2 which implies that nl(f ) ≤ 2n−1 − 2 2 −1 . If n is an even integer a
4
n
n
function f ∈ Bn is said to be bent if Wf (λ) ∈ {2 2 , −2 2 }, for all λ ∈ F2n . Moreover,
n+t
a function f ∈ Bn is said to be t-plateaued if Wf (λ) ∈ {0, ±2 2 }, for all λ ∈ F2n .
The integer t (0 ≤ t ≤ n) is called the amplitude of f . Note that a bent function is
a 0-plateaued function. In the following, “ <, > ” denotes the standard inner (dot)
product of two vectors, that is, < λ, x >= λ1 x1 + . . . + λn xn , where λ, x ∈ F2n . If we
identify the vector space F2n with the finite field F2n , we use the trace bilinear form
T r1n (λx) instead of the dot product, that is, < λ, x >= T r1n (λx), where λ, x ∈ F2n .
For vectorial functions F : Fn2 → Fm
2 , the extended Walsh-Hadamard transform
defined as,
P
(−1)<v,F (x)>+<u,x> ,
WF (u, v) =
x∈Fn
2
where F (x) = (f1 (x), f2 (x), · · · , fm (x)), u ∈ Fn2 , v ∈ Fm
2 .
Let F be a vectorial function from F2n into F2m . The linear combinaison of the
coordinates of F are the Boolean functions fλ : x 7→ T r1m (λF (x)), λ ∈ F2m , where
f0 is the null function. The functions fλ (λ 6= 0) are called the components of F .
A vectorial function is said to be bent (resp. t-plateaued) if all its components are
bent (resp. t-plateaued). A vectorial function F is called vectorial plateaued if all its
components are plateaued with possibly different amplitudes.
Let F : Fn2 → Fn2 be an (n, n)-function. For any a ∈ Fn2 , b ∈ Fn2 , we denote
∆F (a, b) = {x|x ∈ Fn2 , F (x ⊕ a) ⊕ F (x) = b},
δF (a, b) = k∆F (a, b)k,
Then, we have δ(F ) := maxa6=0,b∈Fn2 δF (a, b) ≥ 2 and the functions for which equality
holds are said to be almost perfect nonlinear (APN).
A nice survey on Boolean and vectorial Boolean functions for cryptography can
be found in [2] and [3], respectively.
3
The stability of a function having the maximum number of bent
components under CCZ equivalence
In [8], Pott et al. have shown that the maximum number of bent components of a
vectorial (n, n)-function F is 2n − 2k where k := n2 (n even). They left open the
problem whether the property of a function having the maximum number of bent
components is invariant under CCZ equivalence or not. In this section we solve this
problem by giving a positive answer in the following theorem.
Theorem 1. Let n = 2k and F, F ′′ : Fn2 → Fn2 be CCZ-equivalent functions. Then
F has 2n − 2k bent components if and only if F ′′ has 2n − 2k bent components.
Proof. Let F be a function with 2n − 2k bent components. Define
S = {v ∈ Fn2 : x →< v, F (x) > is not bent}.
5
By Theorem 3 of [8], S is a linear subspace of dimension k. Then, let U be any
k-dimensional subspace of Fn2 such that U ∩ S = {0}. Let v1 , · · · , vk be a basis of S
and u1 , · · · , uk be a basis of U . Define a new function F ′ : Fn2 → Fn2 as
F ′ (x) = (H(x), I(x))
where H(x) = (< v1 , F (x) >, · · · , < vk , F (x) >) and I(x) = (< u1 , F (x) >, · · · , <
uk , F (x) >). Then, F ′ is EA-equivalent to F . Recall that the property of a function
having the maximum number of bent components is invariant under EA equivalence.
Thus, F ′ has 2n −2k bent components. Since F and F ′′ are CCZ-equivalent functions,
F ′′ is CCZ-equivalent to F ′ , which has 2n − 2k bent components. Let L(x, y, z) =
(L1 (x, y, z), L2 (x, y, z), L3 (x, y, z)), (with L1 : Fn2 × F2k × F2k → Fn2 , L2 : Fn2 × F2k ×
F2k → Fk2 and L3 : Fn2 ×F2k ×F2k → Fk2 ) be an affine permutation of Fn2 ×F2k ×F2k which
maps the graph of F ′ to the graph of F ′′ . Then, the graph GF ′′ = {L(x, H(x), I(x)) :
x ∈ Fn2 }. Thus L1 (x, H(x), I(x)) is a permutation and for some affine function
L′1 : Fn2 × F2k → Fn2 and linear function L′′1 : Fk2 → Fn2 we can write L1 (x, y, z) =
L′1 (x, y) + L′′1 (z). For any element v of Fn2 we have
< v, L1 (x, H(x), I(x)) >= < v, L′1 (x, H(x)) > + < v, L′′1 (I(x)) >
= < v, L′1 (x, H(x)) > + < L′′∗
1 (v), I(x) >
′
′′
= < L′′∗
1 (v), I(x) > + < v , H(x) > + < v , x > +a, (3)
′′
where a ∈ F2 , v ′ ∈ Fk2 , v ′′ ∈ Fn2 and L′′∗
1 is the adjoint operator of L1 , in fact,
′′
L′′∗
1 is the linear permutation whose matrix is transposed of that of L1 . Since
L1 (x, H(x), I(x)) is a permutation, then any function < v, L1 (x, H(x), I(x)) > is
balanced (recall that this property is a necessary and sufficient condition) and, hence,
′
cannot be bent. From the construction of F ′ , < L′′∗
1 (v), I(x) > + < v , H(x) > + <
′′∗
′′
′′∗
v , x > +a is not bent if and only if L1 (v) = 0. Therefore, L1 (v) = 0 for any v ∈ Fn2 .
This means that L′′1 is null, that is, L1 (x, H(x), I(x)) = L′1 (x, H(x)). We can also
write Li (x, y, z) = L′i (x, y) + L′′i (z) for i ∈ {2, 3} where L′i : Fn2 × F2k → Fk2 are affine
functions and L′′i : F2k → Fk2 are linear functions. Set F1′′ (x) = L1 (x, H(x), I(x)) =
L′1 (x, H(x)) and F2′′ (x) = (L′2 (x, H(x)) + L′′2 (I(x)), L′3 (x, H(x)) + L′′3 (I(x))). Then,
F ′′ (x) = F2′′ ◦ F1′′−1 (x). For any v ∈ Fn2 and u = (u′ , u′′ ) ∈ Fk2 × Fk2 ,
X
′′
WF ′′ (v, u) =
(−1)<u,F (x)>+<v,x>
x∈Fn
2
=
X
(−1)<u,F
′′ ◦F ′′ (x)>+<v,F ′′ (x)>
1
1
x∈Fn
2
=
X
′′
′′
(−1)<u,F2 (x)>+<v,F1 (x)>
x∈Fn
2
=
X
′′
′′
′
′
′
(−1)<u,(L2 (I(x)),L3 (I(x)))>+<u,(L2 (x,H(x)),L3 (x,H(x)))>+<v,L3 (x,H(x))>
x∈Fn
2
=
X
′′∗ (u′ )+L′′∗ (u′′ ),I(x)>+<v ′ ,H(x)>+<v ′′ ,x>+a
3
(−1)<L2
x∈Fn
2
6
.
′′∗ ′′
′′∗ ′
′
′′∗ ′′
By the construction of I(x), if L′′∗
2 (u ) + L3 (u ) 6= 0, < L2 (u ) + L3 (u ), I(x) >
′
′′∗ ′′
+ < v ′ , H(x) > is bent. Thus, < u, F ′′ (x) > is bent when L′′∗
2 (u ) + L3 (u ) 6= 0,
k
′
′′
k
where u = (u , u ) ∈ F2 × F2 . For i = 2, 3, let Ai be the matrices of size k × k defined
as
L′′i (z) = zAi ,
where z = (z1 , · · · , zk ) ∈ Fk2 . Then,
′
′′∗ ′′
′ T
′′ T
L′′∗
2 (u ) + L3 (u ) =u A2 + u A3
T
A
=(u′ , u′′ ) 2T .
A3
(4)
Recall that L is a affine permutation. Hence, the rank of the linear function (L′′1 (z), L′′2 (z), L′′3 (z))
′′
′′
′′
k
k
n
k
′′
= (0, L′′2 (z),
L3 (z)) from F2 to F2 × F2 × F2 is k. By (L1 (z), L2 (z), L3 (z)) =
z 0|A2 |A3 , the rank of the matrix A2 |A3 is k. Thus, the rank of the matrix
T
T
A2
= A2 |A3 is also k. Set
T
A3
′
′′∗ ′′
S ′ ={(u′ , u′′ ) ∈ Fk2 × Fk2 : L′′∗
2 (u ) + L3 (u ) = 0}
T
A
={(u′ , u′′ ) ∈ Fk2 × Fk2 : (u′ , u′′ ) 2T = 0}.
A3
Then, S ′ is a linear subspace of dimension k. By the previous discussion, if u =
(u′ , u′′ ) ∈ Fn2 \ S ′ , the component function < u, F ′′ (x) > is bent. Thus, F ′′ (x) has at
least 2n − 2k bent components. From Theorem 3 in [8], F ′′ (x) has exactly 2n − 2k
bent components, which completes the proof.
⊔
⊓
4
The non-existence of APN plateaued functions having the
maximum number of bent components
In [8], the authors asked if APN functions could have the maximum number of bent
components or not. In this section we investigate the case of all APN plateaued
functions. The result is given the following theorem.
Theorem 2. Let F be a plateaued APN function defined on Fn2 (where n ≥ 4 is an
even positive integer). Then F cannot have the maximum number of bent components.
Proof. Let F be a plateaued APN function on Fn2 . Denote
Nt = {v ∈ Fn2 : WF (u, v) = ±2
7
n+t
2
},
where WF (u, v) =
have
(−1)v·F (x)+u·x
x∈Fn
2
P
X
WF4 (u, v) =
and t is a positive integer (0 ≤ t ≤ n). We
X
(2
n+tv
2
)2
2tv
X
v∈F2n
u,v∈Fn
2
=2n
WF2 (u, v)
u∈Fn
2
X
v∈Fn
2
=23n
X
X
WF2 (u, v)
u∈Fn
2
2tv
v∈Fn
2
=23n (N0 + N2 22 + · · · + Nn 2n ).
(5)
Since F is APN, we have
X
WF4 (u, v) = 23n (3 · 2n − 2).
(6)
u,v∈Fn
2
From Equations (5) and (6), we have
N0 + N2 22 + · · · + Nn 2n = 3 · 2n − 2.
n
Therefore, we have N0 ≡ 2 mod 4. Since n ≥ 4, 2n − 2 2 ≡ 0 mod 4. Hence,
n
N0 6= 2n − 2 2 .
Thus, F does not have the maximum number of bent components. In particular,
quadratic APN functions cannot have the maximum number of bent components.
⊔
⊓
5
New constructions of bent component functions of vectorial
functions
In this section we provide several functions defined as G(x) = xL(x) on F22k such
that the number of bent components T r12k (αF (x)) equals 22k − 2k , where L(x) is
a linear function on F22k . We first recall two lemmas which will be useful in our
context.
Lemma 1. [8] Let V = F22k and let <, > be a nondegenerate symmetric bilinear
form on V . If L : V → V is linear, we denote the adjoint operator by L∗ , i.e.,
< x, L(y) >=< L∗ (x), y > for all x, y ∈ V . The function f : V → F2 , defined by
x 7→< x, L(x) >, is bent if and only if L + L∗ is invertible.
Lemma 2. [8] Let V = F22k and < x, y >= T r1n (xy) be the trace bilinear form. If
i
L : V → V is defined by L(x) = αx2 , α ∈ V and for any i = 0, 1, · · · , n − 1, then
n−i
n−i
L∗ (x) = α2 x2 .
8
In [8], the authors presented a construction of bent functions through adjoint operators. We start by providing a simplified proof of [8, Theorem 4] (which is the main
result of their article).
Theorem 3. [8, Theorem 4] Let V = F22k and i be a nonnegative integer. Then,
the mapping Fiα defined by
k
i
Fiα (x) = T r12k αx2 (x + x2 )
is bent if and only if α ∈
/ F2k .
Proof. From the proof of [8, Theorem 4], we know
i
2k−i
L(x) + L∗ (x) = T rk2k (αx2 ) + α2
2k−i
.
T rk2k (x)
From Lemma 1, we need to show that L(x) + L∗ (x) = 0 if and only if x = 0.
Let ∇a = {x|T rk2k (x) = a, a ∈ F2k }. We all know T rk2k (x) is a surjection from
F22k to F2k and k∇a k = 22k−k for any a ∈ F2k . We also know ∇0 = F2k .
If α ∈
/ F2k , then L(x) + L∗ (x) = 0 if and only if
2k
i
T rk (αx2 ) = 0,
(7)
T rk2k (x) = 0,
i.e., x = 0. If for any x 6= 0, we always have L(x) + L∗ (x) 6= 0, then α ∈
/ F2k . In fact,
if α ∈ F2k , then
2k−i
2k−i
i
.
= αT rk2k (x) + αT rk2k (x)
L(x) + L∗ (x) = αT rk2k (x2 ) + αT rk2k (x)
Further, L(x) + L∗ (x) = 0 for any x ∈ F2k . Thus, we have Fiα (x) is bent if and only
⊔
⊓
if α ∈
/ F2k .
Now, we are going to present a first new family of bent functions through adjoint
operators.
Theorem 4. Let V = F22k and i be a nonnegative integer. Let t1 , t2 be two positive
t
k−t
k−t
integers such that 0 ≤ t1 , t2 ≤ k and both z 2 1 −1 + z 2 2 −1 + 1 = 0 and z 2 1 −1 +
t
z 2 2 −1 + 1 = 0 have no solutions on F2k . Then, the function Fiα defined on V by
t2 +k
t2
t1 +k
t1
k
i
)
(8)
+ x2 + x2
Fiα (x) = T r12k αx2 (x + x2 + x2 + x2
is bent if and only if α ∈
/ F2k .
Proof. We have
t +k
t
t +k
t
k
i
Fiα (x) = T r12k αx2 (x + x2 + x2 1 + x2 1 + x2 2 + x2 2 )
i
k
2k−t
i+k
i+2k−t
1
)
= T r12k (xαx2 ) + T r12k (xα2 x2 ) + T r12k (xα2 1 x2
i+k−t2
k−t
i+2k−t2
2k−t
i+k−t1
k−t
)
) + T r12k (xα2 2 x2
) + T r12k (xα2 2 x2
+T r12k (xα2 1 x2
2k
= T r1 (xL(x)),
(9)
9
where
i
k
2k−t
i+k
k−t
i+2k−t
i+k−t
1
1
+ α2 1 x2
L(x) = αx2 + α2 x2 + α2 1 x2
i+k−t2
k−t
i+2k−t2
2k−t
+ α2 2 x2
+α2 2 x2
i+k−t1 2k
k−t
i+k−t
k−t
k
i
i
1
)
+ (α2 1 x2
= αx2 + (αx2 )2 + α2 1 x2
k−t2 2i+k−t2 2k
k−t2 2i+k−t2
2
2
)
x
+ (α
x
+α
k−t2
k−t1
k 2
i
i
k 2
i
i
k
i
i
+ αx2 + (αx2 )2
= αx2 + (αx2 )2 + αx2 + (αx2 )2
According to Lemma 2, the adjoint operator L∗ (x) is
2k−i
2k−i
2k−i
k−i
2k−i
t −i
x2 1
x2 + α2
+ α2
x2
L∗ (x) = α2
2k−i 2k+t2 −i
t
−i
2k−i
k+t
−i
2k−i
x
x2 2 + α2
+α2 x2 1 + α2
k+t −i
t −i
k+t −i
t −i
k−i
2k−i
+ x2 + x2 1 + x2 1 + x2 2 + x2 2
x2
k−i
k
t −i
t −i
k
t −i
t −i
k
k−i
2k−i
(10)
x2 + (x2 )2 + x2 1 + (x2 1 )2 + x2 2 + (x2 2 )2
= α2
t −i
k
t −i
k
k−i
k
2k−i
(x + x2 )2 + (x + x2 )2 1 + (x + x2 )2 2
= α2
t +k−i
k
t +k−i
k
k−i
k
2k−i
+ (x + x2 )2 2
(x + x2 )2 + (x + x2 )2 1
= α2
2k−i
= α2
Thus, we have
k−t2
k−t1
k 2
i
i
k 2
i
i
k
i
i
+ αx2 + (αx2 )2
L(x) + L∗ (x) = αx2 + (αx2 )2 + αx2 + (αx2 )2
t +k−i
k
t +k−i
k
k−i
k
2k−i
.
+ (x + x2 )2 2
(x + x2 )2 + (x + x2 )2 1
+α2
k
k
k−i
t +k−i
k
t +k−i
∈
+ (x+ x2 )2 2
Note that we have L(x) ∈ F2k and (x+ x2 )2 + (x+ x2 )2 1
∗
F2k for any x ∈ F22k . From Lemma 1, it is sufficient to show that L(x) + L (x) is
invertible. That is, we need to show that L(x) + L∗ (x) = 0 if and only if x = 0.
t
t
k−t
k−t
Since both z 2 1 −1 + z 2 2 −1 + 1 = 0 and z 2 1 −1 + z 2 2 −1 + 1 = 0 have
k−t1
k 2
i
i
k
i
i
+
no solution in F2k , we have both αx2 + (αx2 )2 + αx2 + (αx2 )2
k−t
2
t +k−i
k
t +k−i
k
k−i
k
k 2
i
i
=0
+ (x + x2 )2 2
= 0 and (x + x2 )2 + (x + x2 )2 1
αx2 + (αx2 )2
if and only if
(
k
i
i
αx2 + (αx2 )2 = 0,
(11)
k
x + x2 = 0.
k−t
z
k−t
t
From the proof of Theorem 3, when both z 2 1 −1 +z 2 2 −1 +1 = 0 and z 2 1 −1 +
/ F2k .
+ 1 = 0 have no solution in F2k , we have Fiα (x) is bent if and only if α ∈
⊔
⊓
2t2 −1
We immediately have the following statement by setting t2 = k − t1 in the previous
theorem.
10
Corollary 1. Let V = F22k and i be a nonnegative integer. Let t1 , t2 be two positive
t
t
integers such that t1 + t2 = k and z 2 1 −1 + z 2 2 −1 + 1 = 0 has no solution in F2k .
Then, the mapping Fiα defined on V by
Fiα (x) = T r12k (αG(x))
t +k
t
t +k
t
k
i
.
is bent if and only if α ∈
/ F2k , where G(x) = x2 x + x2 + x2 1 + x2 1 + x2 2 + x2 2
The previous construction given by Theorem 4 can be generalized as follows.
Theorem 5. Let V = F22k and i be a nonnegative integer. Let t1 , t2 be two positive
k−t
k−t
k−t
k−t
integers such that 0 ≤ t1 , t2 ≤ k and both (γ (1) )2 1 z 2 1 −1 + (γ (2) )2 2 z 2 2 −1 +
t
t
k−i
k−i
1 = 0 and (γ (1) )2 z 2 1 −1 + (γ (2) )2 z 2 2 −1 + 1 = 0 have no solution in F2k , where
γ (1) , γ (2) ∈ F2k . Then, the mapping Fiα defined by
t2 +k
t2
t1 +k
t1
k
i
)
(12)
) + γ (2) (x2 + x2
Fiα (x) = T r12k αx2 x + x2 + γ (1) (x2 + x2
is bent if and only if α ∈
/ F2k .
Proof. We have
Fiα (x) = T r12k (xL(x)),
(13)
where
2k−t i+2k−t
i+k−t1
k−t
k−t
i+k
k
i
1
+ α2 1 x2
L(x) = αx2 + α2 x2 + (γ (1) )2 1 α2 1 x2
2k−t i+2k−t
i+k−t2
k−t
k−t
2
+ α2 2 x2
+ (γ (2) )2 2 α2 2 x2
k−t1
k 2
i
i
k−t
k
i
i
= αx2 + (αx2 )2 + (γ (1) )2 1 αx2 + (αx2 )2
k−t2
k 2
i
i
k−t
+ (γ (2) )2 2 αx2 + (αx2 )2
The adjoint operator L∗ (x) is
2k−i
L∗ (x) = α2
2k−i
+α2
2k−i
= α2
2k−i
+ α2
k+t −i
x2
2k−i t −i
k−i
x2 1
α2
+ (γ (1) )2
2k−i 2k+t2 −i
2k−i 2t2 −i
k−i
x
+ α2
x
α2
+ (γ (2) )2
2k−i
2k−i
x2
1
k
k
t −i
k
t +k−i
k−i
(x + x2 )2 1
k−i
(x + x2 )2 1
k−i
+ (γ (1) )2
k−i
+ (γ (1) )2
(x + x2 )2
k
k−i
x2
k−i
+ (γ (2) )2
k
t −i
(x + x2 )2 2
k−i
k
t +k−i
(x + x2 )2 2
(14)
k 2t1 +k−i
k−i
k 2k−i
2
(1)
2
2
+
(x + x )
+ (γ )
Note that we have L(x) ∈ F2k and (x + x )
t +k−i
k
k−i
∈ F2k for any x ∈ F22k . In order to show that L(x) + L∗ (x)
(γ (2) )2 (x + x2 )2 2
is invertible, we need to show that L(x) + L∗ (x) = 0 if and only if x = 0.
t
k−i
k−t
k−t
k−t
k−t
Since both (γ (1) )2 1 z 2 1 −1 + (γ (2) )2 2 z 2 2 −1 + 1 = 0 and (γ (1) )2 z 2 1 −1 +
t
k−i
(γ (2) )2 z 2 2 −1 + 1 = 0 have no solutions on F2k , we have both L(x) = 0 and
= α2
(x + x2 )2
11
+ (γ (2) )2
k
k−i
(x + x2 )2
if
k−i
+ (γ (1) )2
k
t +k−i
k
k−i
t +k−i
+ (γ (2) )2 (x + x2 )2 2
(x + x2 )2 1
(
k
i
i
αx2 + (αx2 )2 = 0,
k
x + x2 = 0.
= 0 if and only
(15)
/ F2k .
By using the proof of Theorem 3, we have Fiα (x) is bent if and only if α ∈
⊔
⊓
By the same process used to prove Theorem 5, one can get the following result.
Theorem 6. Let V = F22k and i, ρ be two nonnegative integers such that ρ ≤ k. Let
γ (j) ∈ F2k and 0 ≤ tj ≤ k be a nonnegative integer, where j = 1, 2, · · · , ρ. Assume
ρ
ρ
P
P
k−tj
k−tj
tj
k−i
−1 + 1 = 0 and
that both equations
z2
(γ (j) )2
(γ (j) )2 z 2 −1 + 1 = 0
j=1
j=1
have no solution in F2k . Then, the mapping Fiα defined on V by
Fiα (x) = T r12k (αG(x))
i
is bent if and only if α ∈
/ F2k , where G(x) = x2
T rk2k (x) +
ρ
P
j=1
(16)
!
tj
γ (j) (T rk2k (x))2
.
Lemma 3. [8] Let Fα (x) = T r12k (αG(x)), be a Boolean bent function for any α ∈
2k
k
2k
2k
2k
k
F2k
2 \ F2 , where G : F2 → F2 . Then, F : F2 → F2 , defined as F (x) = T rk (αG(x))
k
2k
is a vectorial bent function for any α ∈ F2 \ F2 .
According to Theorem 6 and Lemma 3, we immediately get the following theorem.
Theorem 7. Let G(x) be defined as in Theorem 6. Then, the mapping Fα defined
by
Fα (x) = T rk2k (αG(x))
(17)
is a vectorial bent function for any α ∈ F22k \ F2k .
In [8], the authors presented the differential spectrum of the functions G : F22k →
k
i
F22k defined by G(x) = x2 (x + x2 ). Their result is given below.
Lemma 4. [8] Let i be a nonnegative integer such that i < k. The differential
k
i
spectrum of the functions G(x) = x2 (x + x2 ), G : F22k → F22k , is given by,
{0, 2k }
if a ∈ F∗2k ,
(18)
δG (a, b) ∈
{0, 2gcd(i,k) } if a ∈ F22k \ F2k .
In particular, δG (a, b) = 2k only for a ∈ F∗2k and b ∈ F2k .
i
k
Now we are going to show that the differential spectrum of the functions x2 (x+x2 +
k
i
t +k
t
t +k
t
x2 1 +x2 1 +x2 2 +x2 2 ) is different from the one of the functions x 7→ x2 (x+x2 ).
12
Theorem 8. Let Fiα (x) = T r12k (αG(x)) be defined as Theorem 4, where G(x) =
t +k
t
t +k
t
k
i
x2 (x + x2 + x2 1 + x2 1 + x2 2 + x2 2 ). If there exists t1 = 1 and gcd(t2 , k) 6= 1
k−t
k−t
t
t
such that both z 2 1 −1 + z 2 2 −1 + 1 = 0 and z 2 1 −1 + z 2 2 −1 + 1 = 0 have no
solutions on F2k , then for i = t2 , there exist elements a ∈ F22k \ F2k such that the
number δG (a, b) is equal to 2 for any b ∈ F22k , which is neither 2gcd(i,k) nor 2k .
k
t
Proof. Let a ∈ F22k \ F2k such that τ 2 1 = τ , where τ = a + a2 6= 0 (since t1 |k). We
have
t
k
t
k
k
i
G(x) + G(x + a) = a2 x + x2 + (x + x2 )2 1 + (x + x2 )2 2
(19)
t
k
t
k
k
i
+(x + a)2 a + a2 + (a + a2 )2 1 + (a + a2 )2 2 .
Thus, for any x′ ∈ F22k , there must be one element b ∈ F22k such that G(x′ )+ G(x′ +
a) = b.
Let x′ , x′′ are the solutions of G(x) + G(x + a) = b. Hence
G(x′ ) + G(x′ + a) + G(x′′ ) + G(x′′ + a)
k
t
k
t
k
k
i
= a2 x′ + x′2 + (x′ + x′2 )2 1 + (x′ + x′2 )2 2 + x′′ + x′′2
t
k
t
k
+(x′′ + x′′2 )2 1 + (x′′ + x′′2 )2 2
t
k
t
k
k
i
+(x′ + x′′ )2 a + a2 + (a + a2 )2 1 + (a + a2 )2 2 = 0.
(20)
t
k
k
i
k
Since x+x2 ∈ F2k for any x ∈ F22k , (20) implies that (x′ +x′′ )2 a + a2 + (a + a2 )2 1
i
t
k
+(a + a2 )2 2 belongs to the multiplicative coset a2 F∗2k . Thus, we necessarily have
k
k
x′ + x′′ = aν, where ν ∈ F∗2k . Further, x′ + x′′ + (x′ + x′′ )2 = aν + a2 ν. Since t1 |t2 ,
from (20), we have
i
t
t
t
t
t
t
τ + τ 2 1 + τ 2 2 ν2 + τ ν + τ 2 1 ν2 1 + τ 2 2 ν2 2
i
(21)
t
t
= τ ν 2 + ν + ν 2 1 + ν 2 2 = 0.
If we set i = t2 , then from (21) we have δG (a, b) = 2 6= 2gcd(i,k) since gcd(i, k) =
gcd(t2 , k) 6= 1.
For one element b ∈ F22k , if for any x′ ∈ F22k , we always have G(x′ )+ G(x′ + a) 6=
b, then δG (a, b) = 0.
⊔
⊓
Theorem 9. Let i, ρ be two nonnegative integers such that ρ ≤ k. Let 0 ≤ tj ≤ k be
ρ
P
k−tj
−1 + 1 = 0
a nonnegative integer, where j = 1, 2, · · · , ρ. Assume that both
z2
j=1
and
ρ
P
j=1
tj
z2
−1
+ 1 = 0 have no solution in F2k . Then, the mapping Fiα defined by
Fiα (x) = T r12k (αG(x))
13
(22)
where G(x) =
i
x2
ther, if the number of the
!
ρ
P
tj
+
(T rk2k (x))2
j=1
ρ
P
tj
solutions of
z2
T rk2k (x)
j=1
2gcd(i,k) ,
is bent if and only if α ∈
/ F2k . Furi
+ z + z 2 = 0 on F2k is not equal to
then there exist elements a ∈ F22k \ F2k such that the number δG (a, b) does
not equal 2gcd(i,k) for any b ∈ F22k .
Proof. From Theorem 6, we know Fiα (x) is bent if and only if α ∈
/ F2k . We have
!
ρ
P
tj
i
(T rk2k (x))2
G(x) + G(x + a) = a2 T rk2k (x) +
j=1
!
(23)
ρ
P
tj
i
2k
2
2k
2
T rk (a) +
= b.
+(x + a)
(T rk (a))
j=1
k
Let a ∈ F22k \F2k such that a+a2 = 1. We need to show the number of solutions
of G(x) + G(x + a) = b is not equal to 2gcd(i,k) for any b ∈ F22k . Let ρ = gcd(i, k).
We suppose δG (a, b) = 2ρ and let x′ , x′′ are the solutions of (23) for some b. Hence
G(x′ ) + G(x′ + a) + G(x′′ ) + G(x′′ + a) !
ρ
P
tj
i
i
+ (x′ + x′′ )2 = 0
(T rk2k (x′ + x′′ ))2
= a2 T rk2k (x′ + x′′ ) +
(24)
j=1
since
ρ
P
j=1
tj
z2
−1 +1
= 0 have no solution in F2k , that is,
T rk2k (a) +
i
ρ
P
tj
(T rk2k (a))2
j=1
!
=
1. For any x ∈ F22k , (24) implies that (x′ + x′′ )2 belongs to the multiplicative
i
coset a2 F∗2k . Thus, we necessarily have x′ + x′′ = aν, where ν ∈ F∗2k . Further,
k
k
T rk2k (x′ + x′′ ) = x′ + x′′ + (x′ + x′′ )2 = aν + a2 ν. From (24), we have
ν
2i
+ν +
ρ
X
tj
ν 2 = 0.
(25)
j=1
We also know that the number of the solutions of
ρ
P
j=1
2gcd(i,k) ,
{x|T rk2k (x)
not equal to
thus, if a ∈
not equals 2gcd(i,k) for any b ∈ F22k
tj
i
z 2 + z + z 2 = 0 on F2k is
= 1, x ∈ F22k }, the number δG (a, b) is
⊔
⊓
Theorem 10. Let n = 2k, e be two positive integers. Let V = F22k and i be a nonnegative integer. Let E = {x|x ∈ F22k , T rk2k (x) ∈ F2e } and O = {x ∈ F22k , T rk2k (x) ∈
M }, where M = {y + T rek (y)|y ∈ F2k }. Let Fiα be the function defined on V by
i
(26)
Fiα (x) = T r12k αx2 T re2k (x) .
/ E. If ke is odd, then Fiα is bent if and
If ke is even, then Fiα is bent if and only if α ∈
/ O.
only if α ∈
/ O. Further, if k is odd and e = 2, then Fiα is bent if and only if α ∈
14
Proof. We have
2k−e
2e
e
i
)
Fiα (x) = T r12k αx2 (x + x2 + x2 + · · · + x2
2k−e
i
= T r12k (xαx2 ) + T r12k (xα2
i+e
e
+ · · · + T r12k (xα2 x2 )
= T r12k (xL(x)),
i+2k−e
x2
2k−2e
) + T r12k (xα2
i+2k−2e
x2
)
(27)
where
2k−e
i
2k−2e
i+2k−e
i+e
e
i+2k−2e
+ · · · + α2 x2
x2
+ α2
x2
L(x) = αx2 + α2
e
i
2k−2e
i
2k−e
i
i
+ · · · + (αx2 )2
+ (αx2 )2
= αx2 + (αx2 )2
i
= T re2k (αx2 ).
According to Lemma 2, the adjoint operator L∗ (x) is
2k−i
2k−i
2k−i
2e−i
2k−i
e−i
2k−i
+ ···
+ α2
x2
x2 + α2
+ α2
x2
2k−e−i
2e−i
e−i
2k−i
2k−i
+ · · · + x2
+ x2 + x2
x2
= α2
2k−i
2k−e 2
2e
e
2k−i
x + x2 + x2 + · · · + x2
= α2
L∗ (x) = α2
2k−i
= α2
Thus, we have
2k−e−i
x2
(28)
22k−i
.
T re2k (x)
i
2k−i
L(x) + L∗ (x) = T re2k (αx2 ) + α2
22k−i
.
T re2k (x)
From Lemma 1, it is sufficient to show that L(x) + L∗ (x) is invertible. That is, we
need to show that L(x) + L∗ (x) = 0 if and only if x = 0.
For ke being even, we have T re2k (x) = 0 if and only if x ∈ E. If α ∈
/ E, then
L(x) + L∗ (x) = 0 is only if
(
i
i
T re2k (αx2 ) = T rek T rk2k (αx2 ) = 0,
(29)
T re2k (x) = 0,
i.e., x = 0. If for any x 6= 0, we have L(x) + L∗ (x) 6= 0, then α ∈
/ E. In fact, if
suppose α ∈ E, then
22k−i
2k−i
i
T re2k (x)
L(x) + L∗ (x) = T re2k (αx2 ) + α2
22k−i
2k−i
i
T re2k (x)
= T rk2k (α)T rek T rk2k (x2 ) + α2
=0
for any x ∈ E. Hence, Fiα (x) is bent if and only if α ∈
/ E.
Similarly, for ke odd, we have T re2k (x) = 0 if and only if x ∈ O. We can prove
Fiα (x) is bent if and only if α ∈
/ O.
Similarly, for k odd and e = 2, we have T r22k (x) = 0 if and only if x ∈ O. We
/ O.
⊔
⊓
can prove Fiα (x) is bent if and only if α ∈
15
Remark 1. Note that Theorem 3 is special case of Theorem 10. It corresponds to
the case where e = k.
Similary to Theorem 6, we have the following statement.
Theorem 11. Let i, ρ be two nonnegative integers such that ρ ≤ k. Let γ (j) ∈
F2k and 0 ≤ tj ≤ k be a nonnegative integer, where j = 1, 2, · · · , ρ. Let both
ρ
ρ
P
P
k−tj
k−tj
tj
k−i
−1 + 1 = 0 and
z2
(γ (j) )2
(γ (j) )2 z 2 −1 + 1 = 0 have no solution
j=1
j=1
in F2k . Let E and O be defined as Theorem 10. Let the function Fiα be defined by
ρ
X
tj
i
Fiα (x) = T r12k αx2 T re2k (x) +
γ (j) (T re2k (x))2 .
(30)
j=1
If ke is even, then Fiα is bent if and only if α ∈
/ E. If ke is odd, then Fiα is bent if and
only if α ∈
/ O. Further, if k is odd and e = 2, then Fiα is bent if and only if α ∈
/ O.
6
Conclusions
This paper is in the line of a very recent paper published in the IEEE-transactions
Information Theory by Pott et al [8] in which several open problems have been raised.
In the present paper, we have established that the property of a function having the
maximal number of bent components is invariant under CCZ-equivalence which
gives an answer to an open problem in [8]. Next, we have proved the non-existence
of APN plateaued functions having the maximal number of bent components which
gives a partial answer to an open problem in [8]. Furthermore, we have exhibited
several bent functions Fαi for any α ∈ F22k \ F2k provided that some conditions hold.
In other words, the set of those α for which Fαi is bent is of maximal cardinality
22k − 2k . This provide an answer to another open problem in [8]. In addition, we
have studied the differential spectrum of certain functions and showed that it is not
equal to those studied in [8].
References
1. L. Budaghyan and C. Carlet.: On CCZ-equivalence and its use in secondary constructions of
bent functions, Proceedings of the International Workshop on Coding and Cryptography WCC
2009.
2. C. Carlet.: ‘Boolean Functions for Cryptography and Error Correcting Codes, Chapter of the
monography Boolean Models and Methods in Mathematics, Computer Science, and Engineering,
Y. Crama and P. Hammer eds, Cambridge University Press, pp. 257–397, 2010.
3. C. Carlet.: Vectorial Boolean Functions for Cryptography, Chapter of the monography Boolean
Models and Methods in Mathematics, Computer Science, and Engineering, Y. Crama and
P. Hammer eds, Cambridge University Press, pp. 398–469, 2010.
4. C. Carlet, P. Charpin and V. Zinoviev.: “Codes, bent functions and permutations suitable for
DES-like cryptosystems”, Designs, Codes and Cryptography, 15(2), pp. 125-156,1998.
16
5. C. Carlet and S. Mesnager.: On the construction of bent vectorial functions. Journal of Information and Coding Theory: Algebraic and Combinatorial Coding Theory Vol. 1, No. 2, pp.
133-148 (2010).
6. C. Carlet, S. Mesnager.: Four decades of research on bent functions”, Designs, Codes and
Cryptography, vol. 78, no. 1, pp. 5–50, 2016.
7. F. Chabaud and S. Vaudenay.: Links between differential and linear cryptanalysis. Proceedings
of EUROCRYPT’94, Lecture Notes in Computer Science 950, pp. 356-365 (1995).
8. A. Pott, E. Pasalic, A. Muratović-Ribić and S. Bajrić.: On the maximum number of bent
components of vectorial functions, IEEE Transactions on Information Theory. Volume: 64,
Issue: 1, pp. 403-411, 2018.
9. O. S. Rothaus.: On bent functions, J. Combin. Theory, Ser. A, vol. 20, pp. 300–305, May 1976.
10. Pasalic, E. and Zhang, W. G.: On Multiple Output Bent Functions, Inf. Process. Lett. vol.
112, no. 21, pp. 811–815, nov, 2012.
11. K. Feng and J. Yang.: Vectorial boolean functions with good cryptographic properties. Int. J.
Found. Comput. Sci., vol. 22(6):1271-1282, 2011
12. C. Tang, Y. Qi, M. Xu.: New Quadratic Bent Functions in Polynomial Forms with Coefficients
in Extension Fields, https://eprint.iacr.org/2013/405.pdf
13. M. Matsui.: Linear cryptanalysis method for DES cipher, In: Advances in Cryptology—
EUROCRYPT’93, Lecture Notes in Computer Science, vol.765, Berlin: Springer-Verlag,
pp.386–397, 1993.
14. S. Mesnager.: Bent vectorial functions and linear codes from o-polynomials. Journal Designs,
Codes and Cryptography. 77(1), pages 99-116 (2015).
15. S. Mesnager.: Bent functions: fundamentals and results, pp. 1–544, Springer, Switzerland 2016.
16. K. Nyberg, “Perfect non-linear S-boxes”, Proceedings of EUROCRYPT’ 91, Lecture Notes in
Computer Science, vol. 547, pp. 378–386, 1992.
17. T. Satoh,T. Iwata and K. Kurosawa.: On cryptographically secure vectorial Boolean functions.
Proceeding of Asiacrypt’99 Lecture Notes in Computer Science, Springer-Verlag, Berlin, pp.2028 (1999).
18. J. Wu, Y. Wei, and X. Wang.: An optimized method for multiple output bent functions. Acta
Electronica Sinica, vol. 33(3):521-523, 2005.
17
| 7cs.IT
|
Near Capacity Signaling over Fading Channels using Coherent Turbo Coded OFDM
and Massive MIMO
arXiv:1711.10104v1 [cs.IT] 28 Nov 2017
K. Vasudevan
Dept. of EE
IIT Kanpur
India
Email: [email protected]
Abstract—The minimum average signal-to-noise ratio (SNR) per
bit required for error-free transmission over a fading channel
is derived, and is shown to be equal to that of the additive
white Gaussian noise (AWGN) channel, which is −1.6 dB.
Discrete-time algorithms are presented for timing and carrier
synchronization, as well as channel estimation, for turbo coded
multiple input multiple output (MIMO) orthogonal frequency
division multiplexed (OFDM) systems. Simulation results show
that it is possible to achieve a bit error rate of 10−5 at an
average SNR per bit of 5.5 dB, using two transmit and two receive
antennas. We then propose a near-capacity signaling method in
which each transmit antenna uses a different carrier frequency.
Using the near-capacity approach, we show that it is possible to
achieve a BER of 2 × 10−5 at an average SNR per bit of just 2.5
dB, with one receive antenna for each transmit antenna. When
the number of receive antennas for each transmit antenna is
increased to 128, then a BER of 2×10−5 is attained at an average
SNR per bit of 1.25 dB. In all cases, the number of transmit
antennas is two and the spectral efficiency is 1 bit/transmission
or 1 bit/sec/Hz. In other words, each transmit antenna sends 0.5
bit/transmission. It is possible to obtain higher spectral efficiency
by increasing the number of transmit antennas, with no loss
in BER performance, as long as each transmit antenna uses a
different carrier frequency. The transmitted signal spectrum for
the near-capacity approach can be restricted by pulse-shaping.
In all the simulations, a four-state turbo code is used. The
corresponding turbo decoder uses eight iterations. The algorithms
can be implemented on programmable hardware and there is a
large scope for parallel processing.
Keywords–Channel capacity; coherent detection; frequency selective Rayleigh fading channel; massive multiple input multiple output (MIMO); orthogonal frequency division multiplexing
(OFDM); spectral efficiency; turbo codes.
I.
I NTRODUCTION
We begin this article with an open question: what is
the operating signal-to-noise (SNR) per bit or Eb /N0 of
the present day mobile phones [1]–[3]? The mobile phones
indicate a typical received signal strength of −100 dBm (10−10
mW), however this is not the SNR per bit.
The above question assumes significance since future wireless communications, also called the 5th generation or 5G [4]–
[7], is supposed to involve not only billions of people, but
also smart machines and devices, e.g., driverless cars, remotely
controlled washing machines, refrigerators, microwave ovens,
robotic surgeries in health care and so on. Thus, we have to
deal with an internet of things (IoT), which involves deviceto-human, human-to-device and device-to-device communications. Due to the large number of devices involved, it becomes
imperative that each device operates at the minimum possible
average SNR per bit required for error-free communication.
Depending on the application, there are different requirements on the communication system. Critical applications like
driverless cars and robotic surgeries require low to medium
bit rates, e.g., 0.1 – 10 Mbps and low latency (the time taken
to process the received information and send a response back
to the transmitter) of the order of a fraction of a millisecond.
Some applications like watching movies on a mobile phone
require high bit rates, e.g., 10 – 1000 Mbps for high density
and ultra high density (4k) video and can tolerate high latency,
of the order of a fraction of a second. Whatever the application,
the 5G wireless communication systems are expected to share
some common features like having a large number of transmit
and receive antennas also called massive multiple input multiple output (MIMO) [8]–[11] and the use of millimeter wave
carrier frequencies (> 100 GHz) [12]–[16], to accommodate
large bit-rates (> 1 Gbps) and large number of users. In this
paper we deal with the physical layer of wireless systems
that are also applicable to 5G. The main topics addressed
in this work are timing and carrier synchronization, channel
estimation, turbo codes and orthogonal frequency division
multiplexing (OFDM). Recall that OFDM converts a frequency
selective channel into a flat channel [17] [18].
Channel characteristics in the THz frequency range and
at 17 GHz for 5G indoor wireless systems is studied in [19]
[20]. Channel estimation for massive MIMO assuming spatial
correlation between the receive antennas is considered in [21]
[22]. In [23], a MIMO channel estimator and beamformer
is described. Uplink channel estimation using compressive
sensing for millimeter wave, multiuser MIMO systems is
considered in [24] [25].
Waveform design for spectral containment of the transmitted signal, is an important aspect of wireless telecommunications, especially in the uplink, where many users access a
base station. We require that the signal from one user does not
interfere with the other user. This issue is addressed in [26]–
[37]. Error control coding for 5G is discussed in [38] [39].
References to carrier and timing synchronization in OFDM
can be found in [2] [3] [40].
The capacity of single-user MIMO systems under different
assumptions about the channel impulse response (also called
the channel state information or CSI) and the statistics of
the channel impulse response (also called channel distribution
information or CDI) is discussed in [41]. The capacity of
MIMO Rayleigh fading channels in the presence of interference and receive correlation is discussed in [42]. The low SNR
capacity of MIMO fading channels with imperfect channel
state information is presented in [43].
The main contribution of this paper is to develop discretetime algorithms for coherently detecting multiple input, multiple output (MIMO), orthogonal frequency division multiplexed
(OFDM) signals, transmitted over frequency selective Rayleigh
fading channels. Carrier frequency offset and additive white
Gaussian noise (AWGN) are the other impairments considered
in this work. The minimum SNR per bit required for error-free
transmission over frequency selective MIMO fading channels
is derived. Finally we demonstrate how we can approach close
to the channel capacity.
To the best of our knowledge, other than the work in
[40], which deals with turbo coded single input single output
(SISO) OFDM, and [2] [3], which deal with turbo coded single
input multiple output (SIMO) OFDM, discrete-time algorithms
for the coherent detection of turbo coded MIMO OFDM
systems have not been discussed earlier in the literature.
Coherent detectors for AWGN channels is discussed in [44]
[45]. Simulations results for a 2×2 turbo coded MIMO OFDM
system indicate that a BER of 10−5 , is obtained at an average
SNR per bit of just 5.5 dB, which is a 2.5 dB improvement
over the performance given in [2]. If each transmit antenna
transmits at a different carrier frequency, then we show that
it is possible to achieve a BER of 2 × 10−5 at an average
SNR per bit of just 2.5 dB, with one receive antenna for each
transmit antenna. When the number of receive antennas for
each transmit antenna is increased to 128, then a BER of
2×10−5 is obtained at an average SNR per bit of 1.25 dB. In all
cases, the number of transmit antennas is two and the spectral
efficiency is 1 bit/transmission or 1 bit/sec/Hz. In other words,
each transmit antenna sends 0.5 bit/transmission. It is possible
to obtain higher spectral efficiency by increasing the number of
transmit antennas, with no loss in BER performance, as long
as each transmit antenna uses a different carrier frequency. It
is possible to band limit the transmitted signal using pulse
shaping. In all the simulations, a four-state turbo code is used.
The corresponding turbo decoder uses eight iterations.
This paper is organized as follows. Section II presents the
system model. The discrete-time algorithms and simulation
results for the coherent receiver are given in Section III.
Near-capacity signaling is presented in Section IV. Finally,
Section V concludes the paper.
II.
S YSTEM M ODEL
We assume a MIMO-OFDM system with Nt transmit and
Nr receive antennas, with QPSK modulation. The data from
each transmit antenna is organized into frames, as shown in
Figure 1(a), similar to [2] [3] [40]. Note the presence of
the cyclic suffix, whose purpose will be explained later. In
Figure 1(b), we observe that only the data and postamble
QPSK symbols are interleaved. The buffer QPSK symbols (B)
are sent to the IFFT without interleaving. In Figure 1, the
s̃1, n, nt
Lp
Preamble
s̃4, n, nt s̃k, 2, n, nt
Lcs
Lcp
Cyclic
Cyclic
suffix
prefix
s̃k, 3, n, nt
Ld
B
Data
Postamble
s̃k, n, nt
s̃5, n, nt
(a)
S̃k, 3, 0, nt
Buffer
B
Ld
B
s̃k, 3, 0, nt
Ld −
Data
Ld2
Data
Interleaver
Postamble
Lo
point
π(·)
IFFT
Buffer
B
S̃k, 3, Ld −1, nt
s̃k, 3, Ld −1, nt
(b)
Figure 1.
The frame structure in the time domain.
subscript k refers to the k th frame, n denotes the time index
in a frame and 1 ≤ nt ≤ Nt is the index to the transmit
antenna. The total length of the frame is
L = Lp + Lcs + Lcp + Ld .
(1)
Let us assume a channel span equal to Lh . The channel span
assumed by the receiver is [3] [40]
Lhr = 2Lh − 1
(2)
Note that Lh depends on the delay spread of the channel, and
is measured in terms of the number of symbols. Recall that,
the delay spread is a measure of the time difference between
the arrival of the first and the last multipath signal, as seen by
the receiver. Typically
Lh = d0 /(cTs )
(3)
where d0 is the distance between the longest and shortest
multipath, c is the velocity of light and Ts is the symbol
duration which is equal to the sample spacing of s̃k, n, nt in
Figure 1(a). We have assumed a situation where the mobile is
close to the base station and the longest path is reflected from
the cell edge, which is approximately equal to the cell diameter
d0 , as shown in Figure 2. The base stations, depicted by green
dots, are interconnected by a high data-rate backhaul, shown
by the blue lines. The cell edge is given by the red circles.
Note that d1 < d0 . In order to obtain symmetry, the backhaul
forms an equilateral triangle of length d1 . The base station is
at the center of each cell, whose diameter is d0 . For Lh = 10,
1/Ts = 107 bauds and c = 3 × 108 meters per sec, we get
d0 = 300 meters. Similarly with Lh = 10 and 1/Ts = 108
bauds we obtain d0 = 30 meters. In other words, as the baud
rate increases, the cell size needs to decrease, and consequently
the transmit power decreases, for the same channel span Lh .
The length of the cyclic prefix and suffix is [17]:
Lcp = Lcs = Lhr − 1.
(4)
d1
d0 /2
cell
Cell edge
Figure 2.
Arrangement of cells and base stations.
Throughout the manuscript, we use tilde to denote complex
quantities. However, complex QPSK symbols will be denoted
without a tilde, e.g., S1, n, nt . Boldface letters denote vectors or
matrices. The channel coefficients h̃k, n, nr , nt associated with
the receive antenna nr (1 ≤ nr ≤ Nr ) and transmit antenna
nt (1 ≤ nt ≤ Nt ) for the k th frame are C N (0, 2σf2 ) and
independent over time n, that is:
i
1 h
(5)
E h̃k, n, nr , nt h̃∗k, n−m, nr , nt = σf2 δK (m)
2
where “*” denotes complex conjugate and δK (·) is the Kronecker delta function. This implies a uniform power delay profile. Note that even though an exponential power delay profile
is more realistic, we have used a uniform power delay profile,
since it is expected to give the worst-case BER performance,
as all the multipath components have the same power. The
channel is assumed to be quasi-static, that is h̃k, n, nr , nt is
time-invariant over one frame and varies independently from
frame-to-frame, as given by:
i
1 h
(6)
E h̃k, n, nr , nt h̃∗j, n, nr , nt = σf2 δK (k − j)
2
over [−0.04, 0.04] radian [46]. We assume that ωk is fixed for
a frame and varies randomly from frame-to-frame.
Frequency domain
Time domain
S1, 0, nt
Preamble QPSK
s̃1, 0, nt
IFFT
symbols
s̃1, Lp −1, nt
S1, Lp −1, nt
Sk, 3, 0, nt
Data QPSK
s̃k, 3, 0, nt
IFFT
symbols
s̃k, 3, Ld −1, nt
Sk, 3, Ld −1, nt
r̃k, n, nr
Parallel
to
serial
and
add
cyclic
prefix
and
suffix
ỹk, n, nr
P
nt
Channel
h̃k, n, nr , nt
s̃k, n, nt
e j ωk n
w̃k, n, nr (AWGN)
where k and j denote the frame indices.
The AWGN noise samples w̃k, n, nr for the k th frame
2
at time n and receive antenna nr are C N (0, 2σw
). The
th
frequency offset ωk for the k frame is uniformly distributed
Figure 3.
Block diagram of the transmitter.
The block diagram of the transmitter is given in Figure 3.
can be approximated by a weighted Kronecker delta function
(this condition is usually satisfied by random sequences having
zero-mean; the approximation gets better as Lp increases).
With reference to Figs. 1(a) and 3, note that:
Lp −1
=
s̃1, n, nt
s̃k, 3, n, nt
=
s̃k, 2, n, nt
=
=
s̃4, n, nt
=
s̃5, n, nt
1 X
S1, i, nt e j 2πni/Lp
Lp i=0
We assume Sk, 3, i, nt ∈ {±1 ± j}. Since we require:
i
i
h
h
∆
= E |s̃k, 3, n, nt |2 = 2/Ld = σs2 (13)
E |s̃1, n, nt |2
p
we must have S1, i, nt ∈ Lp Nt /Ld (±1 ± j). In other words,
the average power of the preamble part must be equal to the
average power of the data part, in the time domain.
for 0 ≤ n ≤ Lp − 1
Ld −1
1 X
Sk, 3, i, nt e j 2πni/Ld
Ld i=0
for 0 ≤ n ≤ Ld − 1
s̃k, 3, Ld −Lcp +n, nt
for 0 ≤ n ≤ Lcp − 1
s̃1, n, nt
for 0 ≤ n ≤ Lcs − 1
s̃1, n, nt + s̃4, n−Lp , nt .
Due to the presence of the cyclic suffix in Figure 1 and
(7), and due to (12), we have
(7)
From (7), it is clear that the preamble is independent of the
frame k. However, each transmit antenna has its own preamble,
for the purpose of synchronization and channel estimation at
the receiver.
The preamble in the frequency domain, for each transmit
antenna is generated as follows. Let πp (i), for 0 ≤ i ≤ Lp − 1,
denote the interleaver map for the preamble. Let
T
Sr = Sr, 0 . . . Sr, Lp −1 L ×1
(8)
s̃5, n, nt ⋆ s̃∗1, Lp −1−n, mt
0 for Lp − 1 ≤ n ≤ Lp + Lhr − 2,
nt 6= mt
=
(2Lp /Ld )δK (n − Lp + 1)
for nt = mt
where “⋆” denotes linear convolution.
The signal for the k th frame and receive antenna nr can
be written as (for 0 ≤ n ≤ L + Lh − 2):
r̃k, n, nr
=
p
denote a random vector of QPSK symbols. The preamble
vector for the transmit antenna nt is first initialized by
T
S1, 0, nt . . . S1, Lp −1, nt L ×1
S1, nt =
Nt
X
(9)
ỹk, n, nr e j ωk n + w̃k, n, nr
e j ωk n + w̃k, n, nr
(15)
where s̃k, n, nt is depicted in Figure 1(a) and
ỹk, n, nr =
Nt
X
s̃k, n, nt ⋆ h̃k, n, nr , nt .
(16)
nt =1
Next, we substitute
S1, πp (i4 :i5 ), nt = Sr (i4 : i5 ).
(10)
where i4 : i5 denotes the range of indices from i4 to i5 , both
inclusive, and
i4
i5
s̃k, n, nt ⋆ h̃k, n, nr , nt
nt =1
=
p
= 0Lp ×1 .
(14)
= (nt − 1)Lp /Nt
= i4 + Lp /Nt − 1.
(11)
Note that the preamble in the frequency domain for each
transmit antenna has only Lp /Nt non-zero elements, the rest
of the elements are zero. Moreover, due to πp (·), the Lp /Nt
non-zero elements are randomly interspersed over the Lp
subcarriers in the frequency domain, for each transmit antenna.
By virtue of the preamble construction in (9), (10) and
(11), the preambles in the frequency and time domains corresponding to transmit antennas nt and mt satisfy the relation
(using Parseval’s energy theorem):
∗
S1, i, nt S1,
i, mt
=
⇒ s̃1, n, nt ⊙Lp s̃∗1, −n, mt
=
(2Nt Lp /Ld )δK (nt − mt )
for 0 ≤ i ≤ Lp − 1
0 for nt 6= mt ,
0 ≤ n ≤ Lp − 1
(12)
(2L
p /Ld )δK (n)
for nt = mt
where “⊙Lp ” denotes the Lp -point circular convolution. In
other words, the preambles corresponding to distinct transmit
antennas are orthogonal over Lp samples. Moreover, the autocorrelation of the preambles in frequency and time domain,
Note that any random carrier phase can be absorbed in the
channel impulse response.
The uplink and downlink transmissions between the mobiles and base station could be carried out using time division
duplex (TDD) or frequency division duplex (FDD). Time
division (TDMA), frequency division (FDMA), code division
(CDMA), orthogonal frequency division (OFDMA), for downlink transmissions and filterbank multicarrier (FBMC), for
uplink transmissions [26], are the possible choices for multiple
access (MA) techniques.
III.
R ECEIVER
In this section, we discuss the discrete-time receiver algorithms.
A. Start of Frame (SoF) and Coarse Frequency Offset Estimate
The start of frame (SoF) detection and coarse frequency
offset estimation is performed for each receive antenna 1 ≤
nr ≤ Nr and transmit antenna 1 ≤ nt ≤ Nt , as given by the
following rule (similar to (22) in [40] and (24) in [3]): choose
that value of m and νk which maximizes
(17)
r̃k, m, nr e−j νk m ⋆ s̃∗1, Lp −1−m, nt .
Let m̂k (·) denote the time instant and ν̂k (·) denote the coarse
estimate of the frequency offset (both of which are functions
of nr and nt ), at which the maximum in (17) is obtained. Note
that (17) is a two-dimensional search over m and νk , which
can be efficiently implemented in hardware, and there is a large
scope for parallel processing. In particular, the search over νk
involves dividing the range of ωk ([−0.04, 0.04]] radians) into
B1 frequency bins, and deciding in favour of that bin which
maximizes (17). In our simulations, B1 = 64 [3] [40].
Note that in the absence of noise and due to the properties
given in (14)
m̂k (nr , nt ) = Lp − 1 + argmaxm h̃k, m, nr , nt
(18)
where argmaxm corresponds to the value of m for which
h̃k, m, nr , nt is maximum. We also have
Lp − 1 ≤ m̂k (nr , nt ) ≤ Lp + Lh − 2.
(19)
If m̂k (·) lies outside the range in (19), the frame is declared
as erased (not detected). This implies that the peak in (17) is
due to noise, and not due to the channel. The results for SoF
detection at 0 dB SNR per bit for Lp = 512, 1024, 4096 are
given in Figs. 4, 5, and 6, respectively, for Nt = Nr = 2.
The parameter Z in the three figures denotes the correlation
magnitude given by (17).
The average value of the coarse frequency offset estimate
is given by
PNr PNt
ν̂k (nr , nt )
ω̂k = nr =1 nt =1
.
(20)
Nr Nt
B. Channel Estimation
We assume that the SoF has been estimated using (17)
with outcome m0, k given by (assuming the condition in (19)
is satisfied for all nr and nt ):
m0, k = m̂k (1, 1) − Lp + 1 0 ≤ m0, k ≤ Lh − 1
(21)
and the frequency offset has been perfectly canceled [3] [40].
Observe that any value of nr and nt can be used in the
computation of (21). We have taken nr = nt = 1. Define
m1, k = m0, k + Lh − 1.
(22)
For the sake of notational simplicity, we drop the subscript k
in m1, k , and refer to it as m1 . The steady-state, preamble part
of the received signal for the k th frame and receive antenna
nr can be written as:
r̃k, m1 , nr =
Nt
X
s̃5, nt h̃k, nr , nt + w̃k, m1 , nr
(23)
nt =1
where
r̃k, m1 , nr
=
w̃k, m1 , nr
=
h̃k, nr , nt
=
s̃5, nt
=
T
r̃k, m1 , nr . . . r̃k, m1 +Lp −1, nr
[Lp × 1] vector
T
w̃k, m1 , nr . . . w̃k, m1 +Lp −1, nr
[Lp × 1] vector
T
h̃k, 0, nr , nt . . . h̃k, Lhr −1, nr , nt
[L × 1] vector
hr
...
s̃5, 0, nt
s̃5, Lhr −1, nt
..
..
.
...
.
s̃5, Lp +Lhr −2, nt . . . s̃5, Lp −1, nt
[Lp × Lhr ] matrix
(24)
where Lhr is the channel length assumed by the receiver (see
(2)), s̃5, nt is the channel estimation matrix and r̃k, m1 , nr is
the received signal vector after cancellation of the frequency
offset. Observe that s̃5, nt is independent of m1 and due to the
relations in (12) and (14), we have
0Lhr ×Lhr
for nt 6= mt
H
s̃5, mt s̃5, nt =
(25)
(2Lp /Ld )ILhr for nt = mt
where ILhr is an Lhr × Lhr identity matrix and 0Lhr ×Lhr
is an Lhr × Lhr null matrix. The statement of the ML
channel estimation is as follows. Find ĥk, nr , mt (the estimate
of h̃k, nr , mt ) such that:
!H
Nt
X
s̃5, mt ĥk, nr , mt
r̃k, m1 , nr −
mt =1
Nt
X
r̃k, m1 , nr −
s̃5, mt ĥk, nr , mt
mt =1
!
(26)
is minimized. Differentiating with respect to ĥ∗k, nr , mt and
setting the result to zero yields [17] [47]:
−1 H
ĥk, nr , mt = s̃H
(27)
s̃5, mt r̃k, m1 , nr .
5, mt s̃5, mt
Observe that when m0, k = Lh − 1 in (21), and noise is absent
(see (29) in [40] and (35) in [3]), we obtain:
ĥk, nr , mt
= h̃k, 0, nr , mt
. . . h̃k, Lh −1, nr , mt
0 ... 0
T
. (28)
T
. (29)
Similarly, when m0, k = 0 and in the absence of noise:
ĥk, nr , mt
= 0 . . . 0 h̃k, 0, nr , mt
. . . h̃k, Lh −1, nr , mt
To see the effect of noise on the channel estimate in (27),
consider
−1 H
ũ = s̃H
(30)
s̃5, mt w̃k, m1 , nr .
5, mt s̃5, mt
It can be shown that
σ 2 Ld
∆
E ũũH = w ILhr = 2σu2 ILhr .
Lp
(31)
Therefore, the variance of the ML channel estimate (σu2 ) tends
to zero as Lp → ∞ and Ld is kept fixed. Conversely, when
Ld is increased keeping Lp fixed, there is noise enhancement
[2] [3]. The magnitude spectrum of the actual and estimated
channel for various preamble lengths are shown in Figs. 7, 8
and 9 for Nt = Nr = 2 and 0 dB average SNR per bit. Note
that H̃k, i, nr , nt denotes the Ld -point discrete Fourier transform
(DFT) of h̃k, n, nr , nt in (5).
C. Fine Frequency Offset Estimation
The fine frequency offset estimate is obtained using the
following rule: choose that value of time instant m and
frequency offset νk, f which maximizes:
∗
(32)
r̃k, m, nr e−j (ω̂k +νk, f )m ⋆ ỹ1,
k, L2 −1−m, nr , nt
1
0.005
1
0.005
30
0
30
0
20
−0.005
20
10
10
−0.005
1
1
0.005
30
0
0
30
20
20
−0.005
10
10
−0.005
Figure 4.
SoF detection at 0 dB SNR per bit, Lp = 512.
where
L2
ŷ1, k, m, nr , nt
= Lhr + Lp − 1
= s̃1, m, nt ⋆ ĥk, m, nr , nt
(33)
where ĥk, m, nr , nt is obtained from (27). The fine frequency
offset estimate (ν̂k, f (nr , nt )) is obtained by dividing the
interval [ω̂k − 0.005, ω̂k + 0.005] radian (ω̂k is given in (20))
into B2 = 64 frequency bins [44]. The reason for choosing
0.005 radian can be traced to Figure 5 of [3]. We find that
the maximum error in the coarse estimate of the frequency
offset is approximately 0.004 radian over 104 frames. Thus
the probability that the maximum error exceeds 0.005 radian is
less than 10−4 . However, from Table IV in this paper, we note
that the maximum error in the frequency offset is 2.4 × 10−2
radians for Lp = 512, and 1.1 × 10−2 for Lp = 1024, both of
which are larger than 0.005 radian. By observing this trend,
we expect that for larger values of Lp , say Lp = 4096, the
maximum error in the coarse frequency offset estimate would
be less than 0.005 radians. Increasing Lp would also imply an
increase in Ld , for the same throughput (see (54)). The average
value of the fine frequency offset estimate is given by:
PNr PNt
ν̂k, f (nr , nt )
ω̂k, f = nr =1 nt =1
.
(34)
Nr Nt
D. Super Fine Frequency Offset Estimation
The fine frequency offset estimate in (34) is still inadequate
for turbo decoding and data detection when Ld ≫ Lp [40].
Note that the residual frequency offset is equal to:
ωk − ω̂k − ω̂k, f .
(35)
This residual frequency offset is estimated by interpolating the
FFT output and performing postamble matched filtering at the
receiver [2] [3]. If the interpolation factor is I, then the FFT
size is ILd (interpolation in the frequency domain is achieved
by zero-padding the FFT input in the time domain, and then
taking the ILd -point FFT). Let
m2, k = m1, k + Lp + Lcs
(36)
where m1, k is defined in (22). Once again, we drop the
subscript k from m2, k and refer to it as m2 . Define the FFT
input in the time domain as:
r̃k, m2 , nr = [ r̃k, m2 , nr
. . . r̃k, m2 +Ld −1, nr ]
T
(37)
which is the data part of the received signal in (15) for the k th
frame and receive antenna nr , assumed to have the residual
frequency offset given by (35). The output of the ILd -point
FFT of r̃k, m2 , nr in (37) is denoted by
R̃k, i, nr =
LX
d −1
r̃k, m2 +n, nr e−j 2πin/(ILd )
(38)
n=0
for 0 ≤ i ≤ ILd − 1.
The coefficients of the postamble matched filter is obtained
as follows [2] [3]. Define
G̃′′k, i, nr =
Nt
X
nt =1
Ĥk, i3 , nr , nt Sk, 3, i, nt
for i0 ≤ i ≤ i1 (39)
where Ĥk, i, nr , nt is the Ld -point FFT of the channel estimate
1
1
0.005
0.005
30
0
30
0
20
−0.005
20
10
10
−0.005
1
1
0.5
0.005
0.005
30
0
30
0
20
−0.005
Figure 5.
20
10
10
−0.005
SoF detection at 0 dB SNR per bit, Lp = 1024.
residual frequency offset is given by:
in (27), and
i0
i1
i3
= B + Ld2
= i0 + Lo − 1
= B + π(i − B)
ω̂k, sf = 2π/(ILd Nr )
nr =1
(40)
where π(·) is the data interleaver map, B, Ld2 and Lo are
the lengths of the buffer, data, and postamble, respectively, as
shown in Figure 1(b). Let
′′
G̃k, i, nr for i0 ≤ i ≤ i1
G̃′k, i3 , nr =
(41)
0
otherwise
where 0 ≤ i3 ≤ Ld − 1, the relation between i3 and i is given
in (40). Next, we perform interpolation:
G̃′k, i, nr for 0 ≤ i ≤ Ld − 1
G̃k, i4 , nr =
(42)
0
otherwise
where 0 ≤ i4 ≤ ILd − 1 and i4 = iI. Finally, the postamble
matched filter is G̃∗k, ILd −1−i, nr , which is convolved with
R̃k, i, nr in (38). Note that due to the presence of the cyclic
prefix, any residual frequency offset in the time domain,
manifests as a circular shift in the frequency domain. The
purpose of the postamble matched filter is to capture this shift.
The role of the buffer symbols is explained in [2] [3]. Assume
that the peak of the postamble matched filter output occurs
at m3, k (nr ). Ideally, in the absence of noise and frequency
offset
m3, k (nr ) = ILd − 1.
Nr
X
(43)
In the presence of the frequency offset, the peak occurs to the
left or right of ILd − 1. The average superfine estimate of the
[m3, k (nr ) − ILd + 1] .
(44)
E. Noise Variance Estimation
The noise variance is estimated as follows, for the purpose
of turbo decoding:
!H
Nt
Nr
X
X
1
2
σ̂w =
s̃5, nt ĥk, nr , nt
r̃k, m1 , nr −
2Lp Nr n =1
nt =1
r
!
Nt
X
s̃5, nt ĥk, nr , nt . (45)
r̃k, m1 , nr −
nt =1
F. Turbo Decoding
In this section, we assume that the frequency offset has
been perfectly canceled, that is, r̃k, m2 , nr in (37) contains no
frequency offset. The output of the Ld -point FFT of r̃k, m2 , nr
for the k th frame is given by:
R̃k, i, nr =
Nt
X
H̃k, i, nr , nt Sk, 3, i, nt + W̃k, i, nr
(46)
nt =1
for 0 ≤ i ≤ Ld − 1, where H̃k, i, nr , nt is the Ld -point FFT
of h̃k, n, nr , nt and W̃k, i, nr is the Ld -point FFT of w̃k, n, nr . It
can be shown that [3] [40]
2
1
2
= Ld σw
E W̃k, i, nr
2
2
1
= Lh σf2 .
(47)
E H̃k, i, nr , nt
2
2
2
1
1
0.005
0.005
30
0
30
0
20
−0.005
20
10
−0.005
10
1
1
0.005
0.005
30
0
30
0
20
−0.005
Figure 6.
20
10
−0.005
SoF detection at 0 dB SNR per bit, Lp = 4096.
The received signal in (46), for the k th frame and ith subcarrier, can be written in matrix form as follows:
R̃k, i = H̃k, i Sk, 3, i + W̃k, i
(48)
where R̃k, i is the Nr × 1 received signal vector, H̃k, i is the
Nr × Nt channel matrix, Sk, 3, i is the Nt × 1 symbol vector
and W̃k, i is the Nr × 1 noise vector.
The generating matrix of each of the constituent encoders
is given by (41) in [3], and is repeated here for convenience:
2
1
+
D
G(D) = 1
.
(49)
1 + D + D2
For the purpose of turbo decoding, we consider the case where
Nr = Nt = 2. The details of turbo decoding can be found
in [3], and will not be discussed here. Suffices to say that
corresponding to the transition from state m to state n, at
decoder 1, for the k th frame, at time i, we define (for 0 ≤ i ≤
Ld2 − 1):
2
(50)
γ1, k, i, m, n = exp −Z1, k, i, m, n / 2Ldσ̂w
where Z1, k, i, m, n is given by
min
10
all Sm, n, 2
2
X
nr =1
R̃k, i, nr −
2
X
where Z2, k, i, m, n is given by
min
all Sm, n, 1
where Sm, n, nt denotes the QPSK symbol corresponding to
the transition from state m to state n in the trellis, at transmit
2
2
antenna nt . Observe that σ̂w
is the estimate of σw
obtained
from (45). Observe that the minimization in (51) is over all
possible QPSK symbols, at nt = 2 and index i. Similarly, for
nr =1
R̃k, i, nr −
2
X
2
Ĥk, i, nr , nt Sm, n, nt
.
(53)
nt =1
G. Summary of the Receiver Algorithms
The receiver algorithms are summarized as follows:
1)
2)
3)
(51)
nt =1
2
X
Now, (50) and (52) are used in the forward and backward
recursions of the BCJR algorithm [3].
4)
2
Ĥk, i, nr , nt Sm, n, nt
the transition from state m to state n, at decoder 2, for the k th
frame, at time i, we define (for 0 ≤ i ≤ Ld2 − 1):
2
(52)
γ2, k, i, m, n = exp −Z2, k, i, m, n / 2Ld σ̂w
5)
6)
7)
Estimate the start-of-frame and the frequency offset
(coarse) using (17), for each receive antenna. Obtain
the average value of the frequency offset (ω̂k ) using
(20).
Cancel the frequency offset by multiplying r̃k, n, nr in
(15) by e−j ω̂k n , and estimate the channel using (27),
for each nr and nt .
Obtain ỹ1, k, m, nr , nt from (33) and the fine frequency
offset using (34).
Cancel the frequency offset by multiplying r̃k, n, nr
in (15) by e−j(ω̂k +ω̂k, f )n , and estimate the channel
again using (27), for each nr and nt .
Obtain the average superfine frequency offset estimate using (44). Cancel the offset by multiplying
r̃k, n, nr in (15) by e−j(ω̂k +ω̂k, f +ω̂k, sf )n .
Obtain the noise variance estimate from (45).
Take the Ld -point FFT of r̃k, m2, nr and perform turbo
decoding.
8
Actual
Actual
8
Estimated
Estimated
6
6
4
4
2
2
0
0
0
1 000
2 000
3 000
4 000
0
1 000
2 000
3 000
4 000
8
Actual
Actual
6
Estimated
Estimated
6
4
4
2
2
0
0
0
Figure 7.
1 000
2 000
3 000
4 000
0
1 000
2 000
3 000
4 000
Magnitude spectrum of estimated and actual channel, Lp = 512.
TABLE II.
H. Simulation Results
In this section, we present the simulation results for the
proposed turbo coded MIMO OFDM system with Nt = Nr =
2. The SNR per bit is defined in (92). Note that one data
bit (two coded QPSK symbols) is sent simultaneously from
two transmit antennas. Hence, the number of data bits sent
from each transmit antenna is κ = 0.5, as given in (92).
We have also assumed that σf2 = 0.5. The frame parameters
are summarized in Table I. The maximum number of frames
simulated is 2.2 × 104 , at an average SNR per bit of 6.5 dB.
Each simulation run is over 103 frames. Hence, the maximum
number of independently seeded simulation runs is 22.
TABLE I.
Parameter
F RAME PARAMETERS .
Value
(QPSK symbols)
Lp
512, 1024
Ld
4096
B
4
Lo
256, 512
Ld2
3832, 3576
Lh
10
Lcp = Lcs
18
T =
Ld2
.
Ld + Lp + Lcp + Lcs
Lp
Lo
Ld2
T
512
256
3832
82.515%
1024
512
3576
69.356%
Table II. The BER simulation results for the turbo coded
MIMO OFDM system with Nt = Nr = 2 is shown in
Figure 10. Here “Id” denotes the ideal receiver. For the
practical receivers (“Pr”), the interpolation factor for superfine
frequency offset estimation is I = 16. The practical receiver
with Lp = 1024, Lo = 512 attains a BER of 10−5 at an
SNR per bit of 5.5 dB, which is 1 dB better than the receiver
with Lp = 512, Lo = 256. This is due to the fact that
the variance of the channel estimation error with Lp = 512
is twice that of Lp = 1024 (see (31)). This difference in
the variance of the channel estimation error affects the turbo
decoding process. Moreover, the practical receiver in Figure 10
with Lp = 1024, Lo = 512 is 2.5 dB better than the
practical receiver with one transmit and two receive antennas
in Figure 10 of [2].
TABLE III.
The throughput is defined as [2] [3]:
T HROUGHPUT.
P ROBABILITY OF FRAME ERASURE .
Frame configuration
Probability of erasure
Lp = 512, Lo = 256
2.98 × 10−2
Lp = 1024, Lo = 512
7 × 10−4
(54)
The throughput of various frame configurations is given in
The probability of frame erasure (this happens when (19)
8
8
6
Actual
Actual
Estimated
Estimated
6
4
4
2
2
0
0
0
1 000
2 000
3 000
4 000
0
1 000
2 000
3 000
4 000
6
Actual
5
Actual
6
Estimated
Estimated
4
4
3
2
2
1
0
0
0
Figure 8.
1 000
2 000
3 000
4 000
0
1 000
2 000
3 000
4 000
Magnitude spectrum of estimated and actual channel, Lp = 1024.
is not satisfied) at 0 dB SNR per bit is shown in Table III.
Clearly, as Lp increases, the probability of erasure decreases.
Finally, the root mean square (RMS) and maximum frequency
offset estimation error in radians, at 0 dB SNR per bit, is given
in Table IV.
Note that an infinite impulse response (IIR) filter could also
be used. However, it may have stability problems when the
cut-off frequency of the LPF is close to zero radians. If the
physical channel has 10 taps as given by (3), then the length
of the equivalent channel as seen by the receiver would be:
Lh
IV.
N EAR C APACITY S IGNALING
In Sections II and III, we had presented the discrete-time
algorithms for MIMO-OFDM. The inherent assumption in
these two sections was that all transmit antennas use the same
carrier frequency. The consequence of this assumption is that
the signal at each receive antenna is a linear combination of
the symbols from all the transmit antennas, as given in (46)
and (48). This makes the estimation of symbols, Sk, 3, i in (48),
complicated for large values of Nt and Nr (massive MIMO).
In this section, we assume that distinct transmit antennas use
different carrier frequencies. Thus, the signals from distinct
transmit antennas are orthogonal. To each transmit antenna,
we associate Nr receive antennas, that are capable of receiving
signals from one particular transmit antenna. The total number
of receive antennas is now Nr Nt .
In order to restrict the transmitted signal spectrum, it is
desirable to have a lowpass filter (LPF) at the output of
the parallel-to-serial converter in Figure 3, for each transmit
antenna. If we assume that the cut-off frequency of the LPF is
π/10 radians and its transition bandwidth is π/20 radians, then
the required length of the linear-phase, finite impulse response
(FIR) LPF with Hamming window would be [48]
8π/LLPF
⇒ LLPF
=
=
π/20
160.
= LLPF + 10 − 1
= 160 + 10 − 1
= 169.
The values of Lp and Ld in Figure 1(a) have to be suitably
increased to obtain a good estimate of the channel (see (31))
and maintain a high throughput (see 54). Let us denote the
impulse response of the LPF by pn . We assume that pn is
obtained by sampling the continuous-time impulse response
p(t) at a rate of 1/Ts , where Ts is defined in (3). Note that pn is
real-valued [48]. The discrete-time Fourier transform (DTFT)
of pn is [17] [18]:
P̃P (F )
=
LLPF
X−1
pn e−j 2πF nTs
n=0
=
∞
1 X
P̃ (F − m/Ts )
Ts m=−∞
(57)
where the subscript P denotes a periodic function, F denotes
the frequency in Hz and P̃ (F ) is the continuous-time Fourier
transform of p(t). Observe that:
1)
the digital frequency ω in radians is given by
ω = 2πF Ts
(55)
(56)
2)
P̃P (F ) is periodic with period 1/Ts
(58)
5
6
Actual
Actual
5
Estimated
Estimated
4
4
3
3
2
2
1
1
0
0
0
1 000
2 000
3 000
4 000
0
1 000
2 000
3 000
4 000
5
Actual
Actual
8
Estimated
4
Estimated
6
3
4
2
2
1
0
0
0
Figure 9.
1 000
2 000
3 000
4 000
0
1 000
2 000
3 000
4 000
Magnitude spectrum of estimated and actual channel, Lp = 4096.
TABLE IV.
Frame
configuration
F REQUENCY OFFSET ESTIMATION ERROR .
Est.
Error
Coarse
Fine
Superfine
RMS
1.71 × 10−3
3.38 × 10−4
5.85 × 10−5
Max.
2.4 × 10−2
1.6 × 10−2
2.6 × 10−4
RMS
3.3 × 10−4
9.2 × 10−5
4.3 × 10−5
Max.
1.2 × 10−2
3.9 × 10−4
1.82 × 10−4
Lp = 512
Lo = 256
Lp = 1024
Lo = 512
3)
there is no aliasing in the second equation of (57),
that is
P̃ (F )
for − 2T1 s < F < 2T1 s . (59)
P̃P (F ) =
Ts
Now, s̃k, n, nt in Figure 1(a) is passed through the LPF. Let
us denote the LPF output by ṽk, n, nt . After digital-to-analog
(D/A) conversion, the continuous-time signal is denoted by
ṽk, nt (t). The power spectral density of ṽk, nt (t) [17] [18]
2
1 σs2
·
· P̃ (F )
(60)
Ts 2
where we have assumed that the samples of s̃k, n, nt are
uncorrelated with variance σs2 given in (13). Thus the onesided bandwidth of the complex baseband signal ṽk, nt (t) is
1/(20Ts) Hz, for an LPF with cut-off frequency π/10 radians,
since 1/Ts corresponds to 2π radians. Thus, the passband
signal spectrum from a single transmit antenna would have
a two-sided bandwidth of 1/(10Ts ) Hz.
Sṽ (F ) =
The frame structure is given by Figure 1. The average
power of the preamble in the time domain must be equal to that
of the data, as given by (13). Due to the use of different carrier
frequencies for distinct transmit antennas, the same preamble
pattern can be used for all the transmit antennas. Therefore, the
subscript nt can be dropped from the preamble signal, both in
the time and frequency domain, in Figure 1(a) and (7). There
are also no zero-valued preamble symbols in the frequency
domain, that is [40]
q
S1, i ∈ Lp /Ld (±1 ± j)
(61)
for 0 ≤ i ≤ Lp − 1. The block diagram of the system for
near capacity signaling is shown in Figure 11. The received
signal vector at the output of the FFT for the Nr antennas
associated with the transmit antenna nt , for the k th frame and
ith (0 ≤ i ≤ Ld − 1) subcarrier is:
R̃k, i, nt = H̃k, i, nt Sk, 3, i, nt + W̃k, i, nt
(62)
(49). For decoder 1 and 0 ≤ i ≤ Ld2 − 1, we define [2]:
1.0e+00
1.0e-01
γ1, k, i, m, n =
BER
1.0e-02
γ1, k, i, m, n, l
(65)
l=1
where
1.0e-03
1.0e-04
γ1, k, i, m, n, l
1.0e-05
1.0e-06
Lp=512, Lo=256, Pr
Id
Lp=1024, Lo=512, Pr
1.0e-07
0
1
2
3
4
5
6
7
BER simulation results.
Carrier
frequency
f1
2
R̃
−
Ĥ
S
k,
i,
1,
l
k,
i,
1,
l
m,
n
= exp −
2
2Ld σ̂w
(66)
2
where σ̂w
is the average estimate of the noise variance over
all the Nr diversity arms, as given by (45) with Nt = 1, and
Sm, n is the QPSK symbol corresponding to the transition from
state m to n in the encoder trellis. Similarly at decoder 2, for
0 ≤ i ≤ Ld2 − 1, we have:
SNR per bit (dB)
Figure 10.
Nr
Y
γ2, k, i, m, n =
Rx 1
Nr
Y
γ2, k, i, m, n, l
(67)
l=1
Carrier
frequency
f1
where
γ2, k, i, m, n, l
Rx Nr
Carrier
frequency
fNt
2
R̃
−
Ĥ
S
k,
i,
2,
l
k,
i,
2,
l
m,
n
= exp −
. (68)
2
2Ld σ̂w
1.0e+00
Rx 1
Nr=1
Nr=2
Nr=128
1.0e-01
Bit error rate
Carrier
frequency
fNt
Rx Nr
1.0e-02
1.0e-03
1.0e-04
Figure 11.
Near capacity signaling.
1.0e-05
0
where Sk, 3, i, nt is given in (7), R̃k, i, nt , H̃k, i, nt and W̃k, i, nt
are Nr × 1 vectors given by:
R̃k, i, nt
=
H̃k, i, nt
=
W̃k, i, nt
=
R̃k, i, nt , 1
. . . R̃k, i, nt , Nr
H̃k, i, nt , 1
. . . H̃k, i, nt , Nr
W̃k, i, nt , 1
. . . W̃k, i, nt , Nr
T
T
T
.
(63)
Similar to (47), it can be shown that for 1 ≤ l ≤ Nr
2
1
E W̃k, i, nt , l
2
2
1
E H̃k, i, nt , l
2
=
2
Ld σw
=
Lh σf2 .
The synchronization and channel estimation algorithms are
identical to that given in Section III with Nt = 1.
In the turbo decoding operation we assume that Nt = 2.
The generating matrix of the constituent encoders is given by
1
1.5
2
2.5
SNR per bit (dB)
Figure 12.
BER results for near capacity signaling with Nt = 2.
The simulation results assuming an ideal coherent receiver
is given in Figure 12, for Ld = 4096, Lp = Lo = B = Lcp =
Lcs = 0 (the preamble, postamble, buffer, cyclic prefix or
suffix is not required, since this is an ideal coherent receiver),
Nt = 2 and different values of Nr .
TABLE V.
(64)
0.5
N UMBER OF SIMULATION RUNS FOR VARIOUS Nr .
Nr
SNR
per
bit (dB)
Max.
no. of
of runs
Time for
1 run
(minutes)
1
2.5
51
75
2
1.75
25
76
128
1.25
30
88
The maximum number of independently seeded simulation
runs for various Nr and SNR per bit is given in Table V. For
where we have used (64). It can be shown that
lower values of the SNR per bit, the number of runs is less
than the maximum. Each run is over 103 frames. The time
taken for one run using Scilab on an i5 processor is also given
in Table V. The total time taken to obtain Figure 12 is approximately three months. The channel coefficients H̃k, i, nt , l
in (63) are assumed to be complex Gaussian and independent
over i and l, that is
i
1 h
∗
= Lh σf2 δK (i1 − i2 )
E H̃k, i1 , nt , l H̃k,
i2 , nt , l
2 h
i
1
∗
E H̃k, i, nt , l1 H̃k,
= Lh σf2 δK (l1 − l2 ).
i, nt , l2
2
(69)
σ12 =
where for 0 ≤ i ≤ Ld − 1
2
E H̃k, i, nt , 1
Hk, i, nt , 1, I + j Hk, i, nt , 1, Q
2
E Hk,
i, nt , 1, I
The average SNR per bit is given by the second equation in
(92) with
Pav
Lh σf2
κ
= 2
= 0.5
= 0.5.
It is interesting to compare the average SNR per bit
definitions given by (70) and (92) for Nr = 1 in this paper,
with (38) in [40]. Observe that both definitions are identical.
However, in Figure 12, we obtain a BER of 2 × 10−5 at an
average SNR per bit of 2.5 dB, whereas in Figure 7 of [40]
we obtain a similar BER at 8 dB average SNR per bit, for the
ideal receiver. What could be the reason for this difference?
Simply stated, in this section we have assumed a 4096-tap
channel (see the first equation in (69) and equation (37) in
[40] with Lh = Ld ). However, in [40] we have considered a
10-tap channel. This is further explained below.
1)
In this section, the SNR per bit for the k th frame and
Nr = 1 is proportional to (see also (22) in [2])
SNRk, bit, 1
Ld −1
1 X
H̃k, i, nt , 1
∝
Ld
2
E Hk,
i, nt , 1, Q
2
(71)
i=0
where the subscript 1 in SNRk, bit, 1 denotes case (1)
and H̃k, i, nt , 1 is defined in (63). Note that H̃k, i, nt , 1
is a zero-mean Gaussian random variable which is
independent over i and variance given by (64). Moreover, the right hand side of (71) gives the estimate of
the variance of H̃k, i, nt , 1 . Let us now compute the
variance of the estimate of the variance in (71), that
is
!2
LX
d −1
2
1
H̃k, i, nt , 1 − 2Lh σf2 (72)
σ12 = E
Ld i=0
4
E Hk,
4 i, nt , 1, I
E Hk, i, nt , 1, Q
2
2
E Hk, i, nt , 1, I Hk,
j, nt , 1, I
2
2
E Hk,
i, nt , 1, Q Hk, j, nt , 1, Q
2
2
E Hk,
i, nt , 1, I Hk, j, nt , 1, Q
(70)
The turbo decoder uses eight iterations. Observe that in Figure 12, we obtain a BER of 2×10−5 at an average SNR per bit
of just 2.5 dB, for Nr = 1. It is also clear from Figure 12 that
increasing Nr follows the law of diminishing returns. In fact
there is only 1.25 dB difference in the average SNR per bit,
between Nr = 1 and Nr = 128, at a BER of 2 × 10−5 . The
slight change in slope at a BER of 2×10−5 is probably because
the BER needs to averaged over more number of simulation
runs. We do not expect an ideal coherent detector to exhibit
an error floor.
4
4L2h σf4
σH
=
Ld
Ld
2)
(73)
= 2Lh σf2
∆
2
= σH
= H̃k, i, nt , 1
2
= σH
/2
∆
2
= σH,
1
2
= σH
/2
∆
2
= σH,
1
4
= 3σH,
1
4
= 3σH,
1
4
= σH,
1
=
=
4
σH,
1
4
σH,
1
i 6= j
i 6= j
(74)
where we have used the first equation in (69) and
the assumption that Hk, i, nt , 1, I and Hk, j, nt , 1, Q are
independent for all i, j.
Let us now compute the SNR per bit for each frame
in [40]. Using the notation given in [40], we have
SNRk, bit, 2 ∝
Ld −1
1 X
H̃k, i
Ld i=0
2
(75)
where the subscript 2 in SNRk, bit, 2 denotes case (2).
Again, the variance of the estimate of the variance in
the right hand side of (75) is
!2
LX
d −1
2
1
H̃k, i − 2Lh σf2 . (76)
σ22 = E
Ld i=0
Observe that H̃k, i in (76) is obtained by taking
the Ld -point FFT of an Lh -tap channel, and the
autocorrelation of H̃k, i is given by (37) in [40]. Using
Parseval’s theorem we have
Ld −1
1 X
H̃k, i
Ld i=0
2
=
LX
h −1
h̃k, n
2
(77)
n=0
where h̃k, n denotes a sample of zero-mean Gaussian
random variable with variance per dimension equal
to σf2 . Note that h̃k, n is independent over n (see also
(1) in [40]). Substituting (77) and the first equation
of (74) in (76) we get
!2
LX
h −1
2
2
.
σ22 = E
(78)
h̃k, n − σH
n=0
It can be shown that
σ22 = 4Lh σf4
(79)
A PPENDIX
where we have used the following relations:
2
E h̃k, n
= 2σf2
hk, n, I + j hk, n, Q
E h2k, n, I
2
E hk, n, Q
E h4k, n, I
E h4k, n, Q
E h2k, n, I h2k, m, I
2
E hk, n, Q h2k, m, Q
E h2k, n, I h2k, m, Q
A. The Minimum Average SNR per bit for Error-free Transmission over Fading Channels
= h̃k, n
= σf2
In this appendix, we derive the minimum average SNR per
bit for error-free transmission over MIMO fading channels.
Consider the signal
= σf2
r̃n = x̃n + w̃n
= 3σf4
= 3σf4
=
=
=
σf4
σf4
σf4
n 6= m
n 6= m
(80)
where we have assumed that hk, n, I and hk, m, Q are
independent for all n, m.
Substituting
Lh
Ld
= 10
= 4096
(81)
in (74) and (80) we obtain
σ12
= 0.1σf4
σ22
= 40σf4 .
(83)
where x̃n is the transmitted signal (message) and w̃n denotes
samples of zero-mean noise, not necessarily Gaussian. All
the terms in (83) are complex-valued or two-dimensional
and are transmitted over one complex dimension. Here the
term dimension refers to a communication link between the
transmitter and the receiver carrying only real-valued signals.
We also assume that x̃n and w̃n are ergodic random processes,
that is, the time average statistics is equal to the ensemble
average. The time-averaged signal power over two-dimensions
is given by, for large values of N :
N −1
1 X
′
|x̃n |2 = Pav
.
N n=0
(84)
The time-averaged noise power per dimension is
(82)
Thus we find that the variation in the SNR per bit for each
frame is 400 times larger in case (2) than in case (1). Therefore,
in case (2) there are many frames whose SNR per bit is much
smaller than the average value given by (92), resulting in a
large number of bit errors. Conversely, the average SNR per
bit in case (2) needs to be much higher than in case (1) for
the same BER.
V.
for 0 ≤ n < N
N −1
N −1
1 X
1 X
2
|w̃n |2 = σ ′ w =
|r̃n − x̃n |2 .
2N n=0
2N n=0
The received signal power over two-dimensions is
N −1
1 X
|r̃n |2
N n=0
Finally, an ideal near capacity signaling is proposed, where
each transmit antenna uses a different carrier frequency. Simulation results for the ideal coherent receiver show that it is
possible to achieve a BER of 2 × 10−5 at an average SNR per
bit equal to 2.5 dB, with two transmit and two receive antennas.
When the number of receive antennas for each transmit antenna
is increased to 128, the average SNR per bit required to attain
a BER of 2 × 10−5 is 1.25 dB. The spectral efficiency of the
proposed near capacity system is 1 bit/sec/Hz. Higher spectral
efficiency can be obtained by increasing the number of transmit
antennas with no loss in BER performance. A pulse shaping
technique is also proposed to reduce the bandwidth of the
transmitted signal.
Future work could address the issues of peak-to-average
power ratio (PAPR).
=
=
C ONCLUSIONS
Discrete-time algorithms for the coherent detection of turbo
coded MIMO OFDM system are presented. Simulations results
for a 2 × 2 turbo coded MIMO OFDM system indicate that a
BER of 10−5 , is obtained at an SNR per bit of just 5.5 dB,
which is a 2.5 dB improvement over the performance given in
the literature. The minimum average SNR per bit for error-free
transmission over fading channels is derived and shown to be
equal to −1.6 dB, which is the same as that for the AWGN
channel.
(85)
N −1
1 X
|x̃n + w̃n |2
N n=0
N −1
1 X
2
2
|x̃n | + |w̃n |
N n=0
2
′
= Pav
+ 2σ ′ w
h
i
2
= E |x̃n + w̃n |
(86)
where we have assumed independence between x̃n and w̃n
and the fact that w̃n has zero-mean. Note that in (86) it is
necessary that either x̃n or w̃n or both, have zero-mean.
Next, we observe that (85) is the expression
for a 2N √
′
dimensional noise hypersphere with radius σw
2N . Similarly,
(86) is the expression for
qa 2N -dimensional received signal
hypersphere with radius
′ + 2σ ′ 2 ).
N (Pav
w
Now, the problem statement is: how many noise hyperspheres (messages) can fit into the received signal hypersphere,
such that the noise hyperspheres do not overlap (reliable
2
′
decoding), for a given N , Pav
and σ ′ w ? The solution lies in the
volume of the two hyperspheres. Note that a 2N -dimensional
hypersphere of radius R has a volume proportional to R2N .
Therefore, the number of possible messages is
N
!N
2
′
2
′
N Pav
+ 2σ ′ w
Pav
+ 2σ ′ w
=
(87)
M=
N
2σ ′ 2w
2N σ ′ 2w
over N samples (transmissions). The number of bits required
to represent each message is log2 (M ), over N transmissions.
Therefore, the number of bits per transmission, defined as the
channel capacity, is given by [49]
C
=
1
log (M )
N 2
P′
log2 1 + av
2σ ′ 2w
The average SNR per bit is
SNRav, b
=
bits per transmission (88)
=
over two dimensions or one complex dimension (here again the
term “dimension” implies a communication link between the
transmitter and receiver, carrying only real-valued signals. This
is not to be confused with the 2N -dimensional hypersphere
mentioned earlier or the M -dimensional orthogonal constellations in [18]).
=
=
Proposition A.1: Clearly, the channel capacity is additive
over the number of dimensions. In other words, channel
capacity over D dimensions, is equal to the sum of the
capacities over each dimension, provided the information is
independent across dimensions [2]. Independence of information also implies that, the bits transmitted over one dimension
is not the interleaved version of the bits transmitted over any
other dimension.
Proposition A.2: Conversely, if C bits per transmission are
sent over 2Nr dimensions, (Nr complex dimensions), it seems
reasonable to assume that each complex dimension receives
C/Nr bits per transmission [2].
The reasoning for Proposition A.2 is as follows. We assume
that a “bit” denotes “information”. Now, if each of the Nr
antennas (complex dimensions) receive the “same” C bits of
information, then we might as well have only one antenna,
since the other antennas are not yielding any additional information. On the other hand, if each of the Nr antennas receive
“different” C bits of information, then we end up receiving
more information (CNr bits) than what we transmit (C bits),
which is not possible. Therefore, we assume that each complex
dimension receives C/Nr bits of “different” information.
Note that, when
x̃n
=
Nt
X
H̃k, n, nr , nt Sk, 3, n, nt
nt =1
w̃n
= W̃k, n, nr
(89)
as given in (46), the channel capacity remains the same as in
(88). We now define the average SNR per bit for MIMO systems having Nt transmit and Nr receive antennas. We assume
that κ information bits are transmitted simultaneously from
each transmit antenna. The amount of information received
by each receive antenna is κNt /Nr bits per transmission,
over two dimensions (due to Proposition A.2). Assuming
independent channel frequency response and symbols across
different transmit antennas, the average SNR of R̃k, i, nr in
(46) can be computed from (47) as:
SNRav =
2Lh σf2 Pav Nt
P′
= av
2
2Ld σw
2σ ′ 2w
(90)
for κNt /Nr bits, where
i
h
Pav = E |Sk, 3, i, nt |2 .
(91)
2Lh σf2 Pav Nt Nr
·
2
2Ldσw
κNt
Lh σf2 Pav Nr
2κ
Ld σw
′
Pav
Nr
2 · κN .
′
2σ w
t
(92)
Moreover, for each receive antenna we have
C = κNt /Nr
bits per transmission
(93)
over two dimensions. Substituting (92) and (93) in (88) we get
C
=
⇒ SNRav, b
=
log2 (1 + C · SNRav, b )
2C − 1
.
C
(94)
Clearly as C → 0, SNRav, b → ln(2), which is the minimum
SNR required for error-free transmission over MIMO fading
channels.
R EFERENCES
[1] K. Vasudevan, “Coherent turbo coded mimo ofdm,” in ICWMC 2016,
The 12th International Conference on Wireless and Mobile Communications, Nov. 2016, pp. 91–99, [Online].
[2] ——, “Coherent detection of turbo-coded ofdm signals transmitted
through frequency selective rayleigh fading channels with
receiver diversity and increased throughput,” Wireless Personal
Communications, vol. 82, no. 3, 2015, pp. 1623–1642. [Online].
Available: http://dx.doi.org/10.1007/s11277-015-2303-8
[3] ——, “Coherent detection of turbo-coded OFDM signals transmitted
through frequency selective rayleigh fading channels with receiver
diversity and increased throughput,” CoRR, vol. abs/1511.00776, 2015.
[Online]. Available: http://arxiv.org/abs/1511.00776
[4] J. G. Andrews et al., “What will 5g be?” IEEE Journal on Selected
Areas in Communications, vol. 32, no. 6, June 2014, pp. 1065–1082.
[5] O. Galinina, A. Pyattaev, S. Andreev, M. Dohler, and Y. Koucheryavy,
“5g multi-rat lte-wifi ultra-dense small cells: Performance dynamics,
architecture, and trends,” IEEE Journal on Selected Areas in Communications, vol. 33, no. 6, June 2015, pp. 1224–1240.
[6] C. L. I et al., “New paradigm of 5g wireless internet,” IEEE Journal
on Selected Areas in Communications, vol. 34, no. 3, Mar. 2016, pp.
474–482.
[7] M. Agiwal, A. Roy, and N. Saxena, “Next generation 5g wireless
networks: A comprehensive survey,” IEEE Communications Surveys
Tutorials, vol. 18, no. 3, thirdquarter 2016, pp. 1617–1655.
[8] F. Rusek et al., “Scaling up mimo: Opportunities and challenges with
very large arrays,” IEEE Signal Processing Magazine, vol. 30, no. 1,
Jan. 2013, pp. 40–60.
[9] J. Hoydis, S. ten Brink, and M. Debbah, “Massive mimo in the ul/dl
of cellular networks: How many antennas do we need?” IEEE Journal
on Selected Areas in Communications, vol. 31, no. 2, Feb. 2013, pp.
160–171.
[10] E. Bjrnson, E. G. Larsson, and M. Debbah, “Massive mimo for maximal
spectral efficiency: How many users and pilots should be allocated?”
IEEE Transactions on Wireless Communications, vol. 15, no. 2, Feb.
2016, pp. 1293–1308.
[11] K. L. Wong, C. Y. Tsai, J. Y. Lu, D. M. Chian, and W. Y. Li, “Compact
eight mimo antennas for 5g smartphones and their mimo capacity
verification,” in 2016 URSI Asia-Pacific Radio Science Conference
(URSI AP-RASC), Aug. 2016, pp. 1054–1056.
[12] Z. Pi and F. Khan, “An introduction to millimeter-wave mobile broadband systems,” IEEE Communications Magazine, vol. 49, no. 6, June
2011, pp. 101–107.
[35]
[13]
T. S. Rappaport et al., “Broadband millimeter-wave propagation measurements and models using adaptive-beam antennas for outdoor urban
cellular communications,” IEEE Transactions on Antennas and Propagation, vol. 61, no. 4, Apr. 2013, pp. 1850–1859.
[14]
——, “Millimeter wave mobile communications for 5g cellular: It will
work!” IEEE Access, vol. 1, 2013, pp. 335–349.
[15]
T. S. Rappaport, W. Roh, and K. Cheun, “Mobile’s millimeter-wave
makeover,” IEEE Spectrum, vol. 51, no. 9, Sept. 2014, pp. 34–58.
[37]
[16]
S. Niknam, A. A. Nasir, H. Mehrpouyan, and B. Natarajan, “A multiband ofdma heterogeneous network for millimeter wave 5g wireless
applications,” IEEE Access, vol. 4, 2016, pp. 5640–5648.
[38]
[17]
K. Vasudevan, Digital Communications and Signal Processing, Second
edition (CDROM included). Universities Press (India), Hyderabad,
www.universitiespress.com, 2010.
[18]
——, “Digital Communications and Signal Processing, Third edition,”
2016, URL: http://home.iitk.ac.in/∼ vasu/book0.pdf [accessed: 2017-0201].
[19]
N. Khalid and O. B. Akan, “Wideband thz communication channel measurements for 5g indoor wireless networks,” in 2016 IEEE International
Conference on Communications (ICC), May 2016, pp. 1–6.
[20]
N. R. Zulkefly et al., “Channel characterization for indoor environment
at 17 ghz for 5g communications,” in 2015 IEEE 12th Malaysia
International Conference on Communications (MICC), Nov. 2015, pp.
241–245.
[21]
Q. Guo, G. Gui, and F. Li, “Block-partition sparse channel estimation
for spatially correlated massive mimo systems,” in 2016 8th International Conference on Wireless Communications Signal Processing
(WCSP), Oct. 2016, pp. 1–4.
[22]
[23]
M. Kuerbis, N. M. Balasubramanya, L. Lampe, and A. Lampe, “On the
use of channel models and channel estimation techniques for massive
mimo systems,” in 2016 24th European Signal Processing Conference
(EUSIPCO), Aug. 2016, pp. 823–827.
V. Sridhar, T. Gabillard, and A. Manikas, “Spatiotemporal-mimo channel estimator and beamformer for 5g,” IEEE Transactions on Wireless
Communications, vol. 15, no. 12, Dec. 2016, pp. 8025–8038.
[36]
[39]
[40]
[41]
[42]
[43]
[44]
[45]
[24]
H. Xie, F. Gao, and S. Jin, “An overview of low-rank channel estimation
for massive mimo systems,” IEEE Access, vol. 4, 2016, pp. 7313–7321.
[46]
[25]
Q. Wang, Z. Zhou, J. Fang, and Z. Chen, “Compressive channel
estimation for millimeter wave multiuser mimo systems via pilot reuse,”
in 2016 8th International Conference on Wireless Communications
Signal Processing (WCSP), Oct. 2016, pp. 1–6.
[47]
[48]
[26]
B. Farhang-Boroujeny, “Ofdm versus filter bank multicarrier,” IEEE
Signal Processing Magazine, vol. 28, no. 3, May 2011, pp. 92–112.
[27]
Y. Chen, F. Schaich, and T. Wild, “Multiple access and waveforms
for 5g: Idma and universal filtered multi-carrier,” in 2014 IEEE 79th
Vehicular Technology Conference (VTC Spring), May 2014, pp. 1–5.
[28]
B. Farhang-Boroujeny and H. Moradi, “Ofdm inspired waveforms
for 5g,” IEEE Communications Surveys Tutorials, vol. 18, no. 4,
Fourthquarter 2016, pp. 2474–2492.
[29]
E. Basar, “Index modulation techniques for 5g wireless networks,” IEEE
Communications Magazine, vol. 54, no. 7, July 2016, pp. 168–175.
[30]
S. Venkatesan and R. A. Valenzuela, “Ofdm for 5g: Cyclic prefix
versus zero postfix, and filtering versus windowing,” in 2016 IEEE
International Conference on Communications (ICC), May 2016, pp.
1–5.
[31]
P. Weitkemper et al., “Hardware experiments on multi-carrier waveforms for 5g,” in 2016 IEEE Wireless Communications and Networking
Conference, April 2016, pp. 1–6.
[32]
Z. Hraiech, F. Abdelkefi, and M. Siala, “Pops-ofdm with different
tx/rx pulse shape durations for 5g systems,” in 2015 5th International
Conference on Communications and Networking (COMNET), Nov.
2015, pp. 1–6.
[33]
X. Zhang, L. Chen, J. Qiu, and J. Abdoli, “On the waveform for 5g,”
IEEE Communications Magazine, vol. 54, no. 11, Nov. 2016, pp. 74–
80.
[34]
A. A. Zaidi et al., “Waveform and numerology to support 5g services
and requirements,” IEEE Communications Magazine, vol. 54, no. 11,
Nov. 2016, pp. 90–98.
[49]
G. Berardinelli, K. I. Pedersen, T. B. Sorensen, and P. Mogensen,
“Generalized dft-spread-ofdm as 5g waveform,” IEEE Communications
Magazine, vol. 54, no. 11, Nov. 2016, pp. 99–105.
C. An, B. Kim, and H. G. Ryu, “Design of w-ofdm and nonlinear
performance comparison for 5g waveform,” in 2016 International Conference on Information and Communication Technology Convergence
(ICTC), Oct. 2016, pp. 1006–1009.
Y. Qi and M. Al-Imari, “An enabling waveform for 5g – qam-fbmc:
Initial analysis,” in 2016 IEEE Conference on Standards for Communications and Networking (CSCN), Oct. 2016, pp. 1–6.
R. Garzn-Bohrquez, C. A. Nour, and C. Douillard, “Improving turbo
codes for 5g with parity puncture-constrained interleavers,” in 2016
9th International Symposium on Turbo Codes and Iterative Information
Processing (ISTC), Sept. 2016, pp. 151–155.
L. Guo, Z. Ning, Q. Song, Y. Cui, and Z. Chen, “Toward efficient 5g
transmission: Ser performance analysis for asynchronous physical-layer
network coding,” IEEE Access, vol. 4, 2016, pp. 5083–5097.
K. Vasudevan, “Coherent detection of turbo coded ofdm signals transmitted through frequency selective rayleigh fading channels,” in Signal
Processing, Computing and Control (ISPCC), 2013 IEEE International
Conference on, Sept. 2013, pp. 1–6.
A. Goldsmith, S. A. Jafar, N. Jindal, and S. Vishwanath, “Capacity
limits of mimo channels,” IEEE Journal on Selected Areas in Communications, vol. 21, no. 5, June 2003, pp. 684–702.
Y. Wang and D. W. Yue, “Capacity of mimo rayleigh fading channels in
the presence of interference and receive correlation,” IEEE Transactions
on Vehicular Technology, vol. 58, no. 8, Oct. 2009, pp. 4398–4405.
F. Benkhelifa, A. Tall, Z. Rezki, and M. S. Alouini, “On the low
snr capacity of mimo fading channels with imperfect channel state
information,” IEEE Transactions on Communications, vol. 62, no. 6,
June 2014, pp. 1921–1930.
K. Vasudevan, “Iterative Detection of Turbo Coded Offset QPSK in the
Presence of Frequency and Clock Offsets and AWGN,” Signal, Image
and Video Processing, Springer, vol. 6, no. 4, Nov. 2012, pp. 557–567.
——, “Design and development of a burst acquisition system for
geosynchronous satcom channels,” CoRR, vol. abs/1510.07106, 2015.
[Online]. Available: http://arxiv.org/abs/1510.07106
H. Minn, V. K. Bhargava, and K. B. Letaief, “A Robust Timing
and Frequency Synchronization for OFDM Systems,” IEEE Trans. on
Wireless Commun., vol. 2, no. 4, July 2003, pp. 822–839.
S. Haykin, Adaptive Filter Theory, 3rd ed. Prentice Hall, 1996.
J. G. Proakis and D. G. Manolakis, Digital Signal Processing: Principles, Algorithms and Applications, 2nd ed. Maxwell MacMillan,
1992.
J. G. Proakis and M. Salehi, Fundamentals of Communication Systems.
Pearson Education Inc., 2005.
| 7cs.IT
|
On Inductive Abilities of Latent Factor Models
for Relational Learning
Théo Trouillon
[email protected]
Univ. Grenoble Alpes, 700 avenue Centrale, 38401 Saint Martin d’Hères, France
arXiv:1709.05666v1 [cs.LG] 17 Sep 2017
Éric Gaussier
[email protected]
Univ. Grenoble Alpes, 700 avenue Centrale, 38401 Saint Martin d’Hères, France
Christopher R. Dance
[email protected]
NAVER LABS Europe, 6 chemin de Maupertuis, 38240 Meylan, France
Guillaume Bouchard
[email protected]
Bloomsbury AI, 115 Hampstead Road, London NW1 3EE, United Kingdom
University College London, Gower St, London WC1E 6BT, United Kingdom
Abstract
Latent factor models are increasingly popular for modeling multi-relational knowledge
graphs. By their vectorial nature, it is not only hard to interpret why this class of models
works so well, but also to understand where they fail and how they might be improved. We
conduct an experimental survey of state-of-the-art models, not towards a purely comparative
end, but as a means to get insight about their inductive abilities. To assess the strengths
and weaknesses of each model, we create simple tasks that exhibit first, atomic properties of
binary relations, and then, common inter-relational inference through synthetic genealogies.
Based on these experimental results, we propose new research directions to improve on
existing models.
1. Introduction
In many machine learning fields, research is drifting away from first-order logic methods.
Most of the time, this drift is justified by better predictive performances and scalability of
the new methods. It is especially true with link prediction, a core problem of statistical
relational learning (Getoor and Taskar, 2007), where latent factor models became more
popular than logic-based models (Nickel et al., 2011; Bordes et al., 2013b; Trouillon et al.,
2016).
Link prediction in knowledge graphs—also known as knowledge graph completion—
operates on predicates of pairs of entities: the objects of knowledge graphs. Each different
predicate symbol is called a relation, and a grounded relation is called a fact. For example, given the entities Alice, Eve and Bob and the relations mother and grandmother, if
mother(Alice,Eve) and mother(Eve,Bob) are true facts, then grandmother(Alice,Bob)
is also true. Inferring this last fact from the first two however, requires knowing that the
mother of one’s mother is one’s grandmother, which can be expressed by the first-order
formula: ∀x∀y∀z mother(x, y) ∧ mother(y, z) ⇒ grandmother(x, z).
Logic-based link prediction consists in using both observed facts and logical rules to
infer the truth of unobserved facts. It can be achieved deterministically by logical deduction,
or probabilistically to cope with uncertainty of the data (Richardson and Domingos, 2006;
Trouillon, Gaussier, Dance & Bouchard
Kersting and De Raedt, 2001). Beyond known problems such as complexity or brittleness,
an obvious limitation arises in this setup: logical rules over the knowledge graph relations
are required for inference, and many knowledge graphs only provide observed facts (Dong
et al., 2014; Auer et al., 2007). In this case one must either handcraft rules, or learn them,
generally through inductive logic programming (ILP) methods (Muggleton and De Raedt,
1994; Dzeroski and Lavrac, 1994).
Latent factor models do not suffer this limitation, as the learned model is never represented
explicitly in a symbolic way, but rather as vectorial embeddings of the entities and relations.
Such representations can make the model difficult to interpret, and although they show
better predictive abilities, it has not yet been explored how well those models are able to
overcome this absence of logical rules, and how their inference abilities differ from logic-based
models.
To do so, we evaluate state-of-the-art latent factor models for relational learning on
synthetic tasks, each designed to target a specific inference ability, and see how well they
discover structure in the data. As we are only interested in evaluating inductive abilities of
these models, and not their ability to cope with uncertainty, we design synthetic experiments
with noise-free deterministic data. The choice of this very favorable setup for deterministic
logical inference clarifies the approach followed in this paper and its very purpose: we do
not evaluate latent factor models as an end, but as a means to point out their weaknesses
and stimulate research towards models that do not suffer from combinatorial complexity—
as advocated by Bottou (2014). Computational complexity, and namely polynomiality,
could turn out to be the very criterion for machine intelligence (Aaronson, 2013). Beyond
complexity, one could also argue against explicitly learning logical expressions to tackle
knowledge graph completion that, “when solving a given problem, try to avoid solving a
more general problem as an intermediate step” (Vapnik, 1995).
We first evaluate the models on the three main properties of binary relations: reflexivity,
symmetry and transitivity, and their combinations. We do so by experimentally testing their
ability to learn these patterns from facts, and their robustness to missing data. Then we
set up tasks that represent real reasoning over family genealogies. On this data, we explore
different types of training/testing splits that map to different types of inference.
The remainder of the paper is organized as follows. We first review the literature in
Section 2, before presenting formally the link-prediction task, the evaluated latent factor
models and the optimization procedure in Section 3. Experiments on learning properties of
relations are presented along with their results in Section 4, and experiments description and
results for family genealogies are reported in Section 5. Finally, we propose new research
directions from these results in Section 6.
2. Related Work
Artificial intelligence is becoming more driven by its empirical successes than by the quest
for a principled formalisation of reasoning, making it more of an empirical science than a
theoretical one. Experimental design is a key skill of empirical scientists, and a well-designed
experiment should expose model limitations to enable improving on them. Indeed, seeking
falsification is up to now the best definition of science (Popper, 1934). In machine learning,
it is extremely simple to come up with an experiment that will fail. However it is less easy
2
On Inductive Abilities of Latent Factor Models for Relational Learning
to think of one that brings an informative failure—when one thinks of a failing experiment
at all. The bAbI data set (Weston et al., 2015), proposing a set of 20 prerequisite tasks
for reasoning over natural language, is an example of an informative experiment, by the
specific reasoning type that each task targets. Inspired by the idea of this work, we designed
simple tasks for relational learning that assess basic properties of relations, as well as simple
reasonings such as kinship relations.
Learning from knowledge graphs and more generally relational data is an old problem
of artificial intelligence (Davis et al., 1993). Many contributions have been made using
inductive logic programming for relational data during the last decades (Muggleton, 1995;
Lisi, 2010; Galárraga et al., 2015). Handling inference probabilistically gave birth to the
statistical relational learning field (Getoor and Taskar, 2007), and link prediction has always
been one of the main problems in that field. Different probabilistic logic-based inference
models have been proposed (Ngo and Haddawy, 1997; Wellman et al., 1992; Kersting and
De Raedt, 2001). The main contribution along this line of research is probably Markov
Logic Networks (MLNs) (Richardson and Domingos, 2006). MLNs take as input a set of
first-order rules and facts, build a Markov random field between facts co-occuring in possible
groundings of the formulae, from which they learn a weight over each of these rules that
represents their likeliness of being applied at inference time. Some other proposals followed
a purely probabilistic approach (Friedman et al., 1999; Heckerman et al., 2007).
The link-prediction problem has recently drawn attention from a wider community. Driven
by the W3C standard data representation for the semantic web, the resource description
framework (Cyganiak et al., 2014), various knowledge graphs—also called knowledge bases—
have been collaboratively or automatically created in recent years such as DBpedia (Auer
et al., 2007), Freebase (Bollacker et al., 2008) or the Google Knowledge Vault (Dong et al.,
2014). Since the Netflix challenge (Koren et al., 2009), latent factor models have taken the
advantage over probabilistic and symbolic approaches in the link-prediction task. In terms of
prediction performances first, but also in scalability. This rise of predictive performances and
speed enabled many applications including automated personal assistants and recommender
systems (Ma et al., 2015; Koren, 2008).
Statistical models for learning in knowledge graphs are summarized in a recent review
(Nickel et al., 2016a), and among them latent factor models. We discuss these models in
detail in the following section. One notable latent factor model that is not tested in this
paper is the holographic embeddings model (Nickel et al., 2016b), as it has been shown to be
equivalent to the ComplEx model (Trouillon and Nickel, 2017; Hayashi and Shimbo, 2017).
The ComplEx model (Trouillon et al., 2016) is detailed in the next section. Also, the latent
factor model proposed by Jenatton et al. (2012) is not included as it is a combination of
uni-, bi- and trigram terms that will be evaluated in separate models to understand the
contribution of each modeling choice in different situations.
Not all latent models are actually factorization models. Among these are a variety
of neural-network models, including the neural tensor networks (Socher et al., 2013), or
the multi-layer perceptron used in Dong et al. (2014). We did not survey these models
in this work and focus on latent factor models, that is models that can be expressed as a
factorization of the knowledge graph represented as a tensor.
Similarly to our approach, Bowman et al. (2015) learned some natural logic operations
directly from data with recurrent neural tensor networks, to tackle natural language pro3
Trouillon, Gaussier, Dance & Bouchard
cessing tasks such as entailment or equivalence. Natural logic is a theoretical framework
for natural language inference that uses natural language strings as the logical symbols.
Bouchard et al. (2015) compared the squared and logistic losses when learning transitive
and sequential synthetic relations, and Singh et al. (2015) also investigated a few synthetic
examples for relational learning on different latent factor models.
Following a different goal, other approaches formalised the encoding of logical operations
as tensor operations. Smolensky et al. (2016) applied it to the bAbI data set reasoning tasks,
and Grefenstette (2013) to general Boolean operations.
Advances in bringing both worlds together include the work of Rocktäschel et al. (2015,
2014) and Demeester et al. (2016), where a latent factor model is used, as well as a set of
logical rules. An error-term over the rules is added to the classical latent factor objective
function. In Rocktäschel and Riedel (2016), a fully differentiable neural theorem prover is
used to handle both facts and rules, whereas Minervini et al. (2017) use adversarial training
to do so. Wang and Cohen (2016) learned first-order logic embeddings from formulae learned
by ILP. Similar proposals for integrating logical knowledge in distributional representations
of words include the work of Lewis and Steedman (2013). Conversely, Yang et al. (2015)
learn a latent factor model over the facts only, and then try to extract rules from the learned
embeddings. (Yoon et al., 2016) proposed to use projections of the subject and object entity
embeddings that conserve transitivity and symmetry.
3. The Link-Prediction Task and Models
This section formally defines the link-prediction problem in knowledge graphs, as well as the
notations that will be used throughout this paper. We then introduce the state-of-the-art
models that will be tested in the experimental sections.
3.1 Link-Prediction in Knowledge Graphs
A knowledge graph stores facts about a set of entities E, and a set of relations R, in the form
of facts r(s, o), that we also write as triples (r, s, o), where the relation r ∈ R and the subject
and object entities s, o ∈ E. Each fact is associated with its truth value yrso ∈ {−1, 1}. For
example, the fact first used(Cardano, imaginary numbers) is true (Cardano, 1545), thus
it has a corresponding truth value yrso = 1. To false facts we attribute the value −1. We
denote the set of all possible triples for a given entity set and relation set with T = R × E × E.
In the link-prediction task we observe the truth values of a given set of training triples
TΩ ⊆ T , that together form the observed facts set Ω = {((r, s, o), yrso ) | (r, s, o) ∈ TΩ }. The
task then consists in predicting the truth values of a disjoint set of unobserved triples
(r0 , s0 , o0 ) ∈ T \ TΩ .
Each model is defined by its scoring function φ(r, s, o; Θ), where Θ are the latent
parameters of this model—the entity and relation embeddings—and φ(r, s, o; Θ) : C|Θ| → R
assigns a real-valued score to the fact r(s, o). As some models are real-valued and some
other models are complex-valued, we define the space of the parameters C|Θ| directly over
the complex space.
We define the probability of a given fact r(s, o) to be true as
P (yrso = 1) = σ(φ(r, s, o; Θ))
4
(1)
On Inductive Abilities of Latent Factor Models for Relational Learning
Model
Scoring Function φ
Parameters Θ
CP (Hitchcock, 1927)
hwr , us , vo i
wr , us , vo ∈ RK
RESCAL (Nickel et al., 2011)
eTs Wr eo
Wr ∈ RK , es , eo ∈ RK
TransE (Bordes et al., 2013b)
−||(es + wr ) − eo ||q
wr , es , eo ∈ RK
F model (Riedel et al., 2013)
u>
d wr
wr , ud ∈ RK
DistMult (Yang et al., 2015)
hwr , es , eo i
wr , es , eo ∈ RK
ComplEx (Trouillon et al., 2016)
Re(hwr , es , ēo i)
wr , es , eo ∈ CK
2
Table 1: Scoring functions of the evaluated latent factor models for a given fact r(s, o), along
with the representation of their parameters. In the F model, d indexes all possible pairs of
entities: d = (s, o) ∈ E × E.
where σ is a sigmoid function. We here use the classical logistic function σ(x) = 1+e1−x .
In the remainder of the paper, the following notations will also be used: the number of
entities is denoted by Ne = |E|, and the number of relations by Nr = |R|. The ith row of a
complex matrix X ∈ Cn×m is written xi ∈ Cm . By a slight abuse of notation, for entities
i ∈ E and relations r ∈ R, we will write their corresponding rows in the embedding matrices
as xi or xr , where xi , xr ∈ Cm .
Let us also define the trilinear product of three vectors over the complex space:
ha, b, ci =
K
X
aj bj cj
j=1
= a> (b
c)
(2)
where a, b, c ∈ CK , and
is the Hadamard product, that is the element-wise product
between two vectors of same length.
3.2 The Models
In the following we present in detail the model scoring functions and parameters that we
compare in this experimental survey. We chose to compare only the most popular and
best-performing link-prediction models. The models’ scoring functions and parameters are
summarized in Table 1.
Each of the following models uses latent representations of variable length, that is
controlled by the hyper-parameter K ∈ Z++ . We start by introducing the most natural
model, a general decomposition for tensors: the Canonical-Polyadic (CP) decomposition
(Hitchcock, 1927), also know as CANDECOMP (Caroll and Chang, 1970), and PARAFAC
(Harshman, 1970).
Canonical-Polyadic Decomposition (CP) A natural way to represent a knowledge
graph mathematically is to frame it as a 3rd -order, partially observed, binary tensor Y ∈
{−1, 1}Nr ×Ne ×Ne , where the value at index (r, s, o) is the truth value of the corresponding
triple: yrso .
5
Trouillon, Gaussier, Dance & Bouchard
The Canonical-Polyadic decomposition involves one latent matrix for each dimension
of the decomposed tensor, so in our case we have three latent matrices as Y is a 3rd order
tensor. The dimension of the learned representations K is also often called the rank of the
decomposition. The scoring function is
φ(r, s, o; Θ) = hwr , us , vo i
(3)
where U, V ∈ RNe ×K are the embedding matrices of entities depending on whether they
appear as subject (U ) of the triple or as object (V ), and W ∈ RNr ×K is the embedding
matrix of the relations.
This model is a very general tensor decomposition, though it is not really tailored to
our problem, since our tensor is a stack of Nr square matrices where rows and columns
represent the same underlying objects: the entities. Indeed, its completely decorrelated
representations ui and vi of the same entity i ∈ E make it harder for this model to generalize,
as we will see in the experiments.
RESCAL RESCAL (Nickel et al., 2011) differs from the CP decomposition in two points:
there is only one embedding per entity instead of having one embedding for entities as
subject and another one for entities as objects; and each relation is represented by a matrix
embedding instead of a vector. Its scoring function is
φ(r, s, o; Θ) = e>
s Wr e o
(4)
where E ∈ RNe ×K is the embedding matrix of the entities, and W ∈ RNr ×K×K the embedding
tensor of the relations. Thus Wr ∈ RK×K is the embedding matrix of the relation r.
RESCAL was the first model to propose unique embeddings for entities—simultaneously
with Bordes et al. (2011)—which yielded significant performance improvement, and since then
unique entity embeddings have been adopted by most of the subsequent models. However,
its matrix representations of relations makes its scoring function time and space complexity
quadratic in the rank K of the decomposition. This also leads to potential overfitting.
F model This model proposed by Riedel et al. (2013) maps all possible subject and object
entity pairs d = (s, o) ∈ E × E to a single dimension. Each row in the entity embedding
matrix corresponds to one pair of entities. The scoring function computes the dot product
of the embedding of the pair d with the embedding of the relation r:
φ(r, s, o; Θ) = e>
d wr
2
(5)
where E ∈ RNe ×K is the embedding matrix of the pairs of entities, and W ∈ RNr ×K the
embedding matrix of the relations. It is actually a decomposition of the matrix that results
from a specific unfolding of the Y tensor.
Its pairwise nature gives this model an advantage over non-compositional pairs of entities.
However, its memory complexity is quadratic in the number of entities Ne . In practice,
unobserved pairs of entities are not stored in memory as they are useless. Though this is
also the weak point of this model: it cannot predict scores for unobserved pairs of entities
since it only learns pairwise representations.
6
On Inductive Abilities of Latent Factor Models for Relational Learning
TransE The TransE model (Bordes et al., 2013b) imposes a geometrical structural bias
on the model: the subject entity vector should be close to the object entity vector once
translated by the relation vector. For a given q-norm (generally q = 1 or q = 2) over the
embedding space:
φ(r, s, o; Θ) = −||(es + wr ) − eo ||q
(6)
where E ∈ RNe ×K is the embedding matrix of the entities, and W ∈ RNr ×K is the embedding
matrix of the relations. Deriving the norm in the scoring function exposes another perspective
on the model and unravels its factorial nature, as it gives a sum of bilinear terms as explored
by Garcı́a-Durán et al. (2014):
>
>
φ(r, s, o; Θ) ≈ e>
s e o + e o wr − e s wr
(7)
where constant multipliers and norms of the embeddings have been ignored here. These
bigram terms will help in some specific situations as shown in Section 5.
It is difficult to capture symmetric relations with this model in a multi-relational setting.
Indeed, having φ(r, s, o; Θ) = φ(r, o, s; Θ) implies either es = eo , or wr> (eo − es ) = 0. Since
es 6= eo in general for s 6= o, and wr is in general not the zero vector—in order to share
latent dimensions’ information with the other relation embeddings—modeling symmetric
relations such as similar, cousin, or related implies a strong geometrical constraint on
entity embeddings: their difference must be orthogonal to the relation embedding wr . The
model thus has to make a trade-off between (i) correctly modelling the symmetry of the
relation r, (ii) not zeroing its relation embedding wr , and (iii) not altering too much the
entity embeddings to meet the orthogonality requirement between wr and (eo − es ) for all
e, o ∈ E.
DistMult The DistMult model (Yang et al., 2015) can be seen as a simplification of the
RESCAL model, where the unique representation of entities is kept, while the representation
of the relations is brought back to vectors instead of matrices:
φ(r, s, o; Θ) = hes , wr , eo i
(8)
where E ∈ RNe ×K is the embedding matrix of the entities, and W ∈ RNr ×K the embedding
matrix of the relations.
The major drawback of this model is its symmetry over the subject and object entity
roles. Indeed we have φ(r, s, o; Θ) = φ(r, o, s; Θ), for all s, o ∈ E. But many antisymmetric
relations appear in knowledge graphs such as older, partOf, hypernym. One does not want
to assign the same score to older(a,b) as to older(b,a)!
ComplEx The ComplEx model (Trouillon et al., 2016, 2017) can be seen as a complexvalued version of the DistMult model. The latent matrices of the entities and relations are
in the complex domain instead of the real domain. The scoring function is the real part of
the trilinear product of the entities and relation embeddings:
φ(r, s, o; Θ) = Re(hes , wr , eo i)
7
(9)
Trouillon, Gaussier, Dance & Bouchard
where E ∈ CNe ×K is the embedding matrix of the entities, and W ∈ CNr ×K the embedding
matrix of the relations. We write Re(a) the real part of the complex
number a ∈ C, and
√
Im(a) its imaginary part: a = Re(a) + iIm(a), where i = −1 and Re(a), Im(a) ∈ R.
Note that the conjugate of the object entity embedding is used, and introduces asymmetry
between the subject and object entities. We use the same notations for complex vectors:
eo = Re(eo ) − iIm(eo ), where Re(eo ), Im(eo ) ∈ RK .
This model solves the symmetry problem of DistMult by having slightly different
representations of entities as subject or object through the use of the complex conjugate.
These representations are still tightly linked which yields good generalization properties,
unlike CP. Yet this slight difference allows the model to retain a vectorial representation
of the relations, and thus a linear time and space complexity, unlike RESCAL, and to do
so without any loss of expressiveness—ComplEx is able to decompose exactly all possible
knowledge graphs (Trouillon et al., 2017).
3.3 Training the Models
All models have been reimplemented within the same framework for experimental fairness.
We describe their common optimization scheme in this section.
As previously mentioned, we used the logistic function for σ, as it provides better results
compared to squared error (Nickel and Tresp, 2013; Bouchard et al., 2015). We minimized
the negative log-likelihood with L2 regularization applied entity-wise and relation-wise over
their vector embeddings of the considered model:
X
L(Ω; Θ) =
log(1 + exp(−yrso φ(r, s, o; Θ))) + λ||Θ{r,s,o} ||22 ,
(10)
((r,s,o),yrso )∈Ω
as we found it to perform better than the ranking loss used in previous studies (Bordes
et al., 2013b; Nickel et al., 2016b). For RESCAL’s relation embeddings Wr ∈ RK×K , the
Frobenius norm is used for regularization ||Wr ||F .
The loss is optimized through stochastic gradient descent with mini-batches (10 batches
for the relation properties experiment, and 100 for the families experiment), AdaGrad (Duchi
et al., 2011) with an initial learning rate of 0.1, and early stopping when average precision
decreased on the validation set calculated every 50 epochs. The λ regularization parameter
was validated over the values {0.3, 0.1, 0.03, 0.01, 0.003, 0.001, 0.0003, 0.0} for each model for
each factorization rank K. Parameters are initialized from a centered unit-variance Gaussian
distribution. The complete algorithm is detailed in Appendix A.
Models are evaluated using average precision as it is classically used in the field (Richardson and Domingos, 2006; Nickel et al., 2011). We also computed the average precision of a
deterministic logic inference engine, by applying the corresponding rules that were used to
generate each data set. For each fact r(s, o) in the test set, its probability P (yrso = 1) is set
to 1 if the fact can be logically deduced true from the facts of the training and validation
sets, 0 if it can be deduced to be false, and 0.5 otherwise.
For the TransE model, we trained it with L1 and L2 norms, results are reported for
the best performing one in each experiment. As in the original paper, we did not use
regularization over the parameters but instead we enforced entity embeddings to have unit
norm ||ei ||2 = 1 for all i ∈ E (Bordes et al., 2013b). With the F model, prediction of
8
On Inductive Abilities of Latent Factor Models for Relational Learning
unobserved entity pairs in the training set is handled through random Gaussian embeddings,
similarly to their initialization.
Results are evaluated with average precision, where for each factorization rank the models
with best validated λ are kept. Average precisions are macro-averaged over 10 runs, and
error bars show the standard deviation over these 10 runs.
To assess whether latent factor models are able to generalize from data without any
first-order logic rules, we generate synthetic data from such rules, and evaluate the model in
a classical training, validation and test splitting of the data. The proportion of positives
and negatives is respected across the sets.
We first consider rules corresponding to relation properties, then rules corresponding to
inter-relations reasonings about genealogical data. We also explore robustness to missing
data, as well as different training/testing splits of the data. Keeping the data deterministic
and simple also allows us to write the corresponding logical rules of each experiment, and
simulate test metrics of what perfect induction would yield to get an upper-bound on the
performance of any method. All data sets are made available1 .
4. Learning Relation Properties
In this section we define the three main properties of binary relations, and devise different
experimental setups for learning them individually or jointly, and with more or less observed
data.
4.1 Experimental Design
Relations in knowledge graphs have different names in the different areas of mathematics.
Logicians call them binary predicates, as they are Boolean-valued functions of two variables.
For set theorists, they are binary endorelations, as they operate on two elements of a single
set, in our case the set of entities E. In set theory, relations are characterized by three
main properties: reflexivity/irreflexivity, symmetry/antisymmetry and transitivity. The
definitions of these properties are given in first-order logic in Table 2.
Different combinations of these properties define basic building blocks of set theory such
as equivalence relations that are reflexive, symmetric and transitive relations, or partial
orders that are reflexive, antisymmetric and transitive relations (Halmos, 1998). Examples
are given in Table 3.
Property
Reflexivity
Irreflexivity
Symmetry
Antisymmetry
Transitivity
Definition
∀a r(a, a)
∀a ¬r(a, a)
∀a∀b r(a, b) ⇒ r(b, a)
∀a∀b r(a, b) ∧ r(b, a) ⇒ a = b
∀a∀b∀c r(a, b) ∧ r(b, c) ⇒ r(a, c)
Table 2: Definitions of the main properties of binary relations.
1. https://github.com/ttrouill/induction_experiments
9
Trouillon, Gaussier, Dance & Bouchard
Table 3: Different types of binary relations in set theory. From Wikipedia page on binary
relations (Wikipedia, 2004).
There are many such common examples of these combinations in knowledge graphs, as
there are many hierarchical and similarity relations. For example, the relations older and
father are both strict hierarchies, thus antisymmetric and irreflexive. But one is transitive
(older) whereas the other is not, and that makes all the difference at inference time. Similarly
for symmetric relations, such as has-the-same-parents-as and friend, your sibling’s
parents are also yours which makes the first relation transitive, whereas your friend’s friends
are not necessarily yours. Note that this makes the has-the-same-parents-as relation
reflexive—it is thus an equivalence relation.
Relational learning models must be able to handle relations that exhibit each of the
possible combinations of these properties, since they are all very common, but imply different
types of reasoning, as already acknowledged by Bordes et al. (2013a). Given that a relation
can be reflexive, irreflexive, or neither; symmetric, antisymmetric, or neither; and transitive
or not, we end up with 18 possible combinations. However we will not address the cases of
little interest where (i) none of these properties are true, (ii) only reflexivity or irreflexivity
is true, (iii) the irreflexive, symmetric and transitive case as the only consistent possibility is
that all facts are false, and (iv) the irreflexive transitive case that again must be either all
false, or antisymmetric—and thus corresponds to an already existing case—to be consistent.
Indeed, if one observes two true facts r(s, o) and r(o, s), by application of the transitivity
rule, r(s, s) and r(o, o) must be true, which explains the inconsistency of cases (iii) and (iv),
as they are irreflexive. This leaves us with 13 cases of interest. To evaluate the ability of
models to learn these properties, we generate random 50 × 50 matrices that exhibit each
combination.
To do so, we sample random square sign matrices Y ∈ {−1, 1}Ne ×Ne . First we fill the
diagonal with 1, −1 or missing depending on reflexivity/irreflexivity or none. Then we
make successive passes over the data to make it [anti-]symmetric and/or transitive, until
all of the properties are true over the whole matrix. A pass to make a matrix symmetric
consists in assigning yji ← yij for all i, j ∈ 1, . . . , Ne where i < j, and yji ← −yij to make
it antisymmetric. A pass to make a matrix transitive consists in assigning yij ← 1 if there
10
On Inductive Abilities of Latent Factor Models for Relational Learning
exists a k ∈ 1, . . . , Ne such that yik = ykj = 1, for all i, j ∈ 1, . . . , Ne . When no more
assignment is made during the passes it means the desired properties are true, and the
relation generation is finished.
We also sample each matrix under the constraint of having a balanced number of positives
and negatives up to ±1%. Though there are many more negatives than positives in real
knowledge graphs, in practice negatives are generally subsampled or generated to match the
number of positive facts (Bordes et al., 2013b; Nickel et al., 2016b).
We first learn each relation individually as in a single relation knowledge graph, and
then jointly. In the joint case, note that since each relation is generated independently, there
is no signal shared across the relations that would help predicting facts of one relation from
facts of another relation, thus only the ability to learn each relation patterns is tested. The
proportion of observed facts is generally very small in real knowledge graphs. To assess
models robustness to missing data, we also reduce the proportion of the training set when
learning the different relations jointly.
Results are averaged over a 10-fold cross-validation, with 80% training, 10% validation
and 10% test in the individual learning case. In the joint learning case, the training
percentage varies between 80% and 10%, the validation set size is kept constant at 10%, and
the test set contains the remaining samples—between 10% and 80%.
4.2 Results
Results are first reported on each relation, then jointly and with decreasing proportion of
training data.
4.2.1 Individual Relation Learning
First of all, results were identical for all models whether the relations were reflexive, irreflexive,
or neither (unobserved). This tells us that reflexivity and irreflexivity are not so important
in practice as they do not add or remove any quality in the prediction of latent factor models.
We report only results for different combinations of symmetry/antisymmetry and transitivity
in the main text. Results of combinations including reflexivity and irreflexivity are reported
in Appendix B.
Figure 1 shows the average precision for each model over the generated symmetric and
antisymmetric relations. Surprisingly, on such simple relations with 80% of observed data,
only ComplEx and RESCAL manage to learn from the symmetric and antisymmetric
patterns, with a non-negligible advantage for the ComplEx model. Moreover, we can
see that with higher ranks, RESCAL starts overfitting as its average precision decreases
presumably due to its quadratic number of parameters with respect to the rank, since each
relation r is represented by a matrix Wr ∈ RK×K , as showed in Section 3.2.
The CP model probably fails due to its uncorrelated representations of entities as subject
and as objects, which makes it unable to model symmetry and antisymmetry. DistMult
unsurprisingly fails in the non-symmetric cases, due to the symmetric nature of its scoring
function, and thus succeeds in the symmetric case. More unexpectedly, the TransE model
has a hard time on antisymmetry, but performs well on the symmetric relation, by zeroing
its relation embedding, as explained in Section 3.2. The F model, cannot actually generalize
in a single relation case, as it has one single embedding for each (ordered) entity pair. For
11
Trouillon, Gaussier, Dance & Bouchard
Figure 1: Generated symmetric (left) and antisymmetric (right) relations with 50 entities.
Average precision for each rank ranging from 5 to 50 for each model.
Figure 2: Generated transitive relation with 50 entities. Average precision for each rank
ranging from 5 to 50 for each model.
any fact r(s, o) in the test set, the entity pair (s, o) has never been seen in the training set,
and thus has a random embedding vector.
Figure 3 shows results for the symmetric transitive and antisymmetric transitive relations,
and Figure 2 for the transitive only relations. Almost all models, except the F model and
DistMult in the non-symmetric cases, perfectly generalize with very low-rank. This tells
us that transitivity actually brings so much structure in the data that it makes the problem
very easy for latent factor models when most of the data is observed.
Most state-of-the-art latent factor models are surprisingly unable to model all the basic
properties of binary relations. Though most of the time those relations are learnt together,
but also with much less observed facts. We next assess the models ability to learn these five
relations together, and their robustness to sparse observations by gradually decreasing the
size of the training set.
12
On Inductive Abilities of Latent Factor Models for Relational Learning
Figure 3: Generated symmetric and transitive (left) and antisymmetric and transitive (right)
relations with 50 entities. Average precision for each rank ranging from 5 to 50 for each
model.
4.2.2 Joint Learning
Figure 4 shows the results when all five above relations are jointly learned, for different
proportions of the training set: 80%, 40%, 20%, 10%. As expected the scores drop, and the
gap between the—deterministic logic—upper-bound and latent factor models widen with
the decrease of training data. ComplEx proves to be the most robust to missing data down
to 20%, but match logical inference only with 80% of training data.
RESCAL again overfits with the rank increasing, but is the best performing model
with 10% of the training set, up to rank K = 30. This suggests that having richer relation
representations than entity representations, that is with more parameters, can be profitable
for learning relation properties from little data. However the reason why the variance of
RESCAL’s average precision decreases again for K ≥ 40 remains mysterious.
The CP and TransE models seem to be more sensitive to missing data as their curves
progressively get away from RESCAL’s one with the percentage of observed data decreasing.
DistMult, being a symmetric model, is below the other models in the four settings as some
of the relations are not symmetric.
Since each relation is generated independently, having observed the entity pair (s, o) in
the other relations does not help the F model, and it thus fails here too. At 10%, we see
that the latent factor models cannot match logical inference, suggesting that the number of
examples is not sufficient to learn these properties.
Finally, in the last setting with 10% of the training set, the best models are still 10
points below the best achievable average precision, showing that they need a large amount
of training data to correctly learn these basic properties of binary relations.
These results should be taken cautiously as this experiment does not state that in general
at least 80% of the facts should be observed in order to learn these properties correctly.
Indeed, here the 5 relations are completely uncorrelated, while in real knowledge graphs they
generally are correlated and thus share information. Also, as often in machine learning, the
ratio between the number of parameters and the number of data points is more informative
13
Trouillon, Gaussier, Dance & Bouchard
Figure 4: Joint learning of the 5 relations with 50 entities: one symmetric, one antisymmetric,
one transitive, one symmetric and transitive, and one antisymmetric and transitive. Average
Precision for each factorization rank ranging from 5 to 50 of each model. Top-Left: 80%
train set, Top-Right: 40% train set, Bottom-Left: 20% train set, Bottom-Right: 10% train
set.
about generalization than the number of data points alone.
Relation Properties Experiments Summary:
• Only ComplEx manages to learn each combination near perfectly.
• RESCAL is the most robust to missing data with low ranks.
• There is room for improvement on all models when a lot of data is missing.
5. Learning Inter-Relational Patterns: Family Relationships
We generated family trees and their corresponding kinship relations and facts, and designed
three different splits of the data. The three splits try to assess different inductive properties
of the latent models, by giving more or less supporting facts in the training set.
14
On Inductive Abilities of Latent Factor Models for Relational Learning
Figure 5: Example of a generated family tree.
5.1 Experimental Design
Predicting family relationships is an old task in AI, popularised by Hinton’s kinship data
set (Hinton, 1986). Generated synthetic families for testing link-prediction models have
also been recently proposed (Garcı́a-Durán et al., 2014). In this public dataset, generated
families are all intertwined with each other in it. We here want each family to be disjoint
from the other ones, such that there is no true fact between entities of two different families,
and we will see why below.
We propose here to generate family relations from synthetic family trees, namely:
mother, father, husband, wife, son, daughter, brother, sister, uncle, aunt, nephew,
niece, cousin, grandfather, grandson, grandmother and granddaughter.
We sample five families independently that span over three generations from a unique
couple, with three children per couple of random sex, where husbands, wives and their
parents were added to wed the middle generation. Figure 5 shows an example of such a
family tree. The whole data set totals 115 entities—23 persons per family—and the 17
relations mentioned above. Each family thus consists in 8993 true and false facts.
Within these traditional families that feature only married heterosexual couples that do
not divorce and have children, one can figure out that the relations mother, father, son
and daughter are sufficient to deduce the 13 remaining ones. Examples of rules that allow
deducing these 13 relations from the 4 main ones are shown in Table 4. From this fact, we
devise three different splits of the data.
Let us first introduce notations for each subset of the observed facts set Ω. As each
family is independent from the four others, we will denote each entity set of each family from
1 to 5: E 1 , . . . , E 5 , where E i ∩ E j = ∅ with i 6= j. Accordingly, we will write the observed
facts of each family Ω1 , . . . , Ω5 , where for all ((r, s, o), yrso ) ∈ Ωi we have s, o ∈ E i . Observed
fact sets that contain only the 4 main relations mother, father, son and daughter are
denoted by Ω4main , and the facts involving the 13 other relations by Ω13other . We thus have
for each family i: Ωi = Ωi4main ∪ Ωi13other . Figure 6 draws the corresponding tensor with each
subset of the observed facts. Finally, let the sampling function Sp (Ω) be a uniformly random
15
Trouillon, Gaussier, Dance & Bouchard
∀a∀b∀c father(a, c) ∧ mother(b, c) ⇒ husband(a, b)
∀a∀b∀c father(a, c) ∧ mother(b, c) ⇒ wife(b, a)
∀a∀b∀c daughter(a, c) ∧ son(b, c) ⇒ sister(a, b)
∀a∀b∀c daughter(a, c) ∧ son(b, c) ⇒ brother(b, a)
∀a∀b∀c father(a, b) ∧ father(b, c) ⇒ grandfather(a, c)
∀a∀b∀c son(a, b) ∧ son(b, c) ⇒ grandson(a, c)
∀a∀b∀c mother(a, b) ∧ mother(b, c) ⇒ grandmother(a, c)
∀a∀b∀c daughter(a, b) ∧ daughter(b, c) ⇒ granddaughter(a, c)
∀a∀b∀c∀d son(a, b) ∧ daughter(b, c) ∧ son(d, c) ⇒ uncle(d, a)
∀a∀b∀c∀d daughter(a, b) ∧ son(b, c) ∧ daughter(d, c) ⇒ aunt(d, a)
∀a∀b∀c∀d son(a, b) ∧ daughter(b, c) ∧ son(d, c) ⇒ nephew(a, d)
∀a∀b∀c∀d daughter(a, b) ∧ son(b, c) ∧ daughter(d, c) ⇒ niece(a, d)
∀a∀b∀c∀d∀e son(a, b) ∧ daughter(b, c) ∧ son(d, c) ∧ daughter(e, d) ⇒ cousin(a, e)
Table 4: Examples of rules to deduce all relations from the four relations: mother, father,
son and daughter.
Figure 6: Tensor representation of the observed subsets for the family experiments. The
part in dark orange represents the sets containing the four relations mother, father, son
and daughter, while the part in light orange represents the 13 other relations.
subset of Ω of size |Sp (Ω)| = dp|Ω|e, with 0 ≤ p ≤ 1, p being the proportion of the set that
is randomly sampled.
We propose to split the data in three different ways to explore inductive abilities of the
models. The first split is the classical random split between training, validation and test
16
On Inductive Abilities of Latent Factor Models for Relational Learning
(a) Random split
(b) Evidence split
(c) Family split
Figure 7: Tensor representation of the three different splits. Green sets are always contained
in the training set Ωtrain , whereas blue sets are split among training, validation and test sets.
sets, it will mainly serve as a control experiment for the other splits. In the second split, we
aim at evaluating whether latent factor models are able to leverage this information. To do
so, we ensure that all the relations mother, father, son and daughter of the five families
are in the training set, and we split the 13 remaining ones between training, validation and
test set. Formally: Ωtrain = Ω4main ∪ Sp (Ω13other ). We will call this splitting scheme the
evidence split, as the training set always contains the sufficient evidence to deduce the 13
other relations—that is the four main ones.
Thirdly, we assess the ability of latent factor models to transfer knowledge learnt from
a family to another, that is between disjoint set of entities. In this split, the training set
always contains all the relations for four out of the five families plus all the mother, father,
son and daughter of the fifth family, while the 13 other relations of this fifth family are split
between training, validation and test set. Formally: Ωtrain = Ω1−4 ∪ Ω54main ∪ Sp (Ω513other ).
We will call it the family split. Figure 7 shows tensor drawings of the three splits.
For each split we explore different values of p ∈ {0.8, 0.4, 0.2, 0.1}. We also run with
p = 0 in the last (family) split, which corresponds to Ωtrain = Ω1−4 ∪ Ω54main , that is 4 entirely
observed family, plus the 4 main relations of the fifth one. Observe that it only makes sense
to have p = 0 in this last split. If latent factor models have expected inductive abilities,
they would be able to understand genealogical reasoning from the four first families, and use
this learned information to correctly predict the 13 other relations of the fifth family from
its four main ones. Note that in the last two splits, a deterministic logic inference system
makes perfect predictions—given rules such as the ones shown in Table 4—for any value of
p. The number of facts in the training, validation and test sets of each split are summarized
in Table 5.
Similar splits of data have already been proposed to evaluate rule-based inference models
(for example the UW-CSE dataset (Richardson and Domingos, 2006)), which are able of
such transfer of reasoning between disjoint sets of entities. Interestingly, such data sets have
17
Trouillon, Gaussier, Dance & Bouchard
Split
Random
Evidence
Family
Set
Ωtrain = Sp (Ω)
Ωvalid = S0.1 (Ω)
Ωtest = S(0.9−p) (Ω)
Ωtrain = Ω4main ∪ Sp (Ω13other )
Ωvalid = S0.1 (Ω13other )
Ωtest = S(0.9−p) (Ω13other )
Ωtrain = Ω1−4 ∪ Ω54main ∪ Sp (Ω513other )
Ωvalid = S0.1 (Ω513other )
Ωtest = S(0.9−p) (Ω513other )
0.8
35973
4496
4496
38089
3438
3438
43589
688
688
Size with p =
0.4
0.2
0.1
17987 8994
4496
4496
4496
4496
22482 31475 35973
24334 17457 14019
3438
3438
3438
17193 24070 27508
40839 39463 38776
688
688
688
3438
4814
5501
0
38088
688
6189
Table 5: Training, validation and test set numbers for each split for each value of p.
rarely been reused in the subsequent latent factor model literature. Results reported next
might give us a hint why this is the case.
5.2 Results
Results are reported for each split separately. In each of them we again decrease progressively
the amount of training data.
5.2.1 Random Split
In the first random split, we try to evaluate the quantity of training data needed to learn to
reason in genealogies. Figure 8 shows the average precision of each model for ranks ranging
from 5 to 50, for each value of p. Only ComplEx and RESCAL are able to generalize
almost perfectly with 80% of observed data, which first tells us that these models are indeed
capable to learn such genealogical reasonings. As many relations are antisymmetric, it is
no surprise that DistMult and TransE cannot reach perfect predictions, as they already
failed in the antisymmetric synthetic relation.
The ComplEx model generalizes quickly with small ranks, but is outrun by RESCAL—
with small ranks—and TransE when the percentage of observed data decreases below
p = 0.2. We conjecture that TransE’s robustness is due to its bilinear terms, and especially
the one that involves the subject and the object embeddings—e>
s eo —as shown in Section 3.2,
that can give high scores to pairs of entities belonging to the same family. For RESCAL,
its richer representations of relations by matrices probably help here, as long as the rank is
not too big which clearly causes overfitting.
The CP decomposition scores drop quickly with the proportion of observed data, because
of its unrelated subject and object representations. The F model here fails again, for a
simple reason: these relations are exclusive between themselves for a given pair of entities
(s, o). Indeed, if father(s, o) is true for example, then none of the other relations between
s and o will be true—at least not in these families. Hence if the F model has to predict the
score of the fact r(s, o), it has no other true triple involving (s, o) to support its decision. It
will also have troubles on the two next splits for the same reason. Note that in this split,
18
On Inductive Abilities of Latent Factor Models for Relational Learning
Figure 8: Average Precision for each factorization rank ranging from 5 to 50 of each model
on the families experiment with the random split. Top-left: p = 0.8, top-right: p = 0.4,
bottom-left: p = 0.2, bottom-right: p = 0.1.
the logic upper-bound is not given as one would need to know all possible rules to deduce
the 17 relations from each of them—and not only from the four main ones—to compute this
upper-bound.
5.2.2 Evidence Split
In this split, we recall that all the mother, father, son and daughter relations are always
in the train set for the 5 families. The value of p ranging from 0.8 to 0.1 corresponds here to
the proportion of the 13 other relations that are also put into the training set. The test and
validation sets are only composed of these 13 relations.
Compared to the random split setting, we can see in Figure 9 that the performances
of the models decrease more slowly with the percentage of observed data. This shows that
latent factor models are able to use the information provided by these four relations from
which all of the others can be deduced.
RESCAL is here clearly the best model for all values of p, as long as K is not too
big. It exhibits again a behavior that seems to have two equilibria distributed around a
pivotal K at which average precision suddenly drops, with high variance of the predictions
around that K. ComplEx also seems to show a lighter overfitting with high values of K
19
Trouillon, Gaussier, Dance & Bouchard
Figure 9: Average Precision for each factorization rank ranging from 5 to 50 of each model
on the families experiment with the evidence split. Top-left: p = 0.8, top-right: p = 0.4,
bottom-left: p = 0.2, bottom-right: p = 0.1.
when p ≤ 0.2. TransE confirms an advantage with p = 0.1 with a notable rise of average
precision compared to the random split. CP, DistMult and the F model fail again for the
same reasons as in the random split.
However, given the rules to deduce the 13 other relations from the four main ones,
recall that a logical inference engine is able to reach an average precision of one. Though
improvement compared to the random split setup is large, the gap with logical inference is
still wide with p = 0.1 and p = 0.2, showing that latent factor models have troubles making
the link between the four main relations and the 13 other ones when limited training data is
available. This could be due to the imbalance in the number of each relation in the training
set that this split introduces, biasing the entity embeddings towards a better reconstruction
of the 4 main relations, to the detriment of the generalization over the 13 remaining ones.
Weighting the facts in accordance with the preponderance of each relation in the dataset
could improve performances here.
5.2.3 Family Split
In this last split, all the mother, father, son and daughter are in the train set for all
families, but also all the 13 other relations of four out of the five families. The value of p
20
On Inductive Abilities of Latent Factor Models for Relational Learning
Figure 10: Average Precision for each factorization rank ranging from 5 to 50 of each model
on the families experiment with the family split. Top-left: p = 0.8, top-right: p = 0.4,
middle-left: p = 0.2, middle-right: p = 0.1, bottom: p = 0.0.
corresponds here to the amount of the 13 other relations of the fifth family only that are in
the training set too.
The curves in Figure 10 show a clear improvement over the previous ones in Figure 9.
RESCAL is again the best model as it reaches average precisions ≥ 0.9 even down to
p = 0.1—with small ranks again. ComplEx is in these cases the best with high ranks,
though much below RESCAL’s best scores when p = 0.1.
21
Trouillon, Gaussier, Dance & Bouchard
Does that mean these models were able to exploit the additional information? Yes and
no. We conjecture that the better results for p ranging from 0.8 to 0.1 are partly due to the
relation imbalance problem—explained in the previous split—being much smaller here, as
all the relations of four families are given in in the training set.
To ensure that models indeed did not generalized from the four perfectly informed
families, we reduced the proportion p of the 13 other relations of the fifth family that are
in the training set to zero—which thus constitute the whole validation and test sets. And
though the models are provided with four perfectly informed families, and all the needed
facts to predict the missing ones in the fifth family, they fail in this last setting as shown in
the bottom plot of Figure 9. RESCAL and TransE resist better than the other models
again in this last setting with p = 0.
This is easily explained, as disconnected sets of entities, here families, correspond to
different blocks in the tensor Y, as shown in Figure 6. Entities that are in different
families s, o ∈ Ωi , s0 , o0 ∈ Ωj , i 6= j, are never involved together in an observed fact:
((s, r, o0 ), ysro0 ), ((s0 , r, o), ys0 ro ) ∈
/ Ω, for any relation r ∈ R. Thus when learning their
embeddings es , eo and es0 , eo0 , the only link they share is the embedding of the relation r
that is involved in the scoring functions φ(r, s, o) and φ(r, s0 , o0 ). This interpretation is also
supported by RESCAL scores, which benefits from its higher number of parameters of its
relation representations Wr ∈ RK×K , which increases the amount of information shared
across the families.
To formally bring this problem to light, let us re-frame the tensor approximation problem
as a system of inequalities. As P (yrso = 1) = σ(φ(r, s, o; Θ)), true triples should have a
score φ(r, s, o; Θ) > 0, and false triples φ(r, s, o; Θ) < 0. For the sake of the example, we
will consider factorizing the two relations sister and grandfather, with two families of 2
persons each, using the DistMult model. Observing the true fact sister(a,b) or the true
fact sister(b,a) where a, b ∈ E, allows us to deduce that a and b are not the grandfather
of each other:
• sister(a, b) ⇒ ¬grandfather(a, b)
• sister(a, b) ⇒ ¬grandfather(b, a)
• sister(b, a) ⇒ ¬grandfather(a, b)
• sister(b, a) ⇒ ¬grandfather(b, a)
Similarly to the family split with p = 0, let us have both relations fully observed for
a first family that contains entities a, b ∈ E 1 , and only the facts of the relation sister
observed for entities of a second family c, d ∈ E 2 . The resulting 2 × 4 × 4 partially observed
binary tensor is:
a
a −1
1
sister : b
c
d
b
1
−1
c
−1
1
d
a
a −1
,
1
−1
−1
grandfather : b
c
d
22
b
−1
−1
c
d
·
·
·
·
(11)
On Inductive Abilities of Latent Factor Models for Relational Learning
where · and empty spaces are unobserved facts. From the first, fully observed family we
wish to learn the above rules and the irreflexivity of the grandfather relation, to correctly
complete the grandfather facts between entities c and d.
As the observed blocks—and the block we wish to recover—are symmetric here, there is
no expressiveness issue with using DistMult. Decomposing this tensor with the DistMult
model with K = 2 such that true facts have probability P (yrso = 1) > 0.5 and false facts
have probability P (yrso = 1) < 0.5, amounts to solving the following system of inequalities:
ws1 e2a1 + ws2 e2a2 < 0
ws1 e2b1 + ws2 e2b2 < 0
ws1 ea1 eb1 + ws2 ea2 eb2 > 0
wg1 e2a1 + wg2 e2a2 < 0
w e2 + w e2 < 0
g1 b1
g2 b2
(12)
wg1 ea1 eb1 + wg2 ea2 eb2 < 0
2
2
ws1 ec1 + ws2 ec2 < 0
ws1 e2d1 + ws2 e2d2 < 0
w e e + w e e > 0
s1 c1 d1
s2 c2 d2
where ei ∈ R2 is the embedding of entity i ∈ E, ws ∈ R2 is the embedding of the relation
sister, and wg ∈ R2 is the embedding of the relation grandfather. The six first inequalities
involve the entities a and b, and the three lower ones involve the entities c and d.
Correctly reconstructing the grandfather facts between c and d would thus require their
embeddings to satisfy the same three additional inequalities:
2
2
wg1 ec1 + wg2 ec2 < 0
.
(13)
wg1 e2d1 + wg2 e2d2 < 0
wg1 ec1 ed1 + wg2 ec2 ed2 < 0
However, it is easy to check that arbitrary solutions to the system (12) for ec and ed does not
necessarily satisfy the system (13), and hence does not necessarily predict the grandfather
facts between c and d correctly. Also, this would be true even if we added more families like
a and b with both relations fully observed, as this would not add more constraints on ec and
ed .
This explains why all models fail in the family split with p = 0: nothing encourages less
constrained entities to have embeddings that resemble the ones of similar, more constrained
entities; and adding more examples of more constrained entities does not help.
Family Experiments Summary:
• RESCAL is the best model in all different splits, but overfits with a too big K.
• RESCAL and TransE are the most robust to missing data.
• ComplEx behaves well with more data and hardly overfits.
23
Trouillon, Gaussier, Dance & Bouchard
• Relation imbalance in the training set can be a problem when the test set is distributed
differently, and could be easily fixed by weighting the facts accordingly.
• The absence of explicit parameter sharing between entity representations prevents
knowledge transfer between disjoint sets of entities.
6. Future Research Directions
Overall, the ComplEx model proved to have the more stable generalization abilities
across all the synthetic experiments. Most models showed a good ability to learn basic
relation properties, except on antisymmetry where only ComplEx succeeded. This said,
when decreasing the size of the training set down to 10% on joint learning of the relation
properties, the best models were 10 points of average precision behind the best possible
score. Improving models towards learning basic binary relation properties from less data
thus seems a promising direction.
Some models showed their advantages in some specific settings. RESCAL and TransE
showed a good robustness when a lot of data is missing in the family experiments, thanks to
the bilinear terms for TransE, and the rich matrix relation representations of RESCAL.
The F model was not fit for these experiments, but its pairwise terms are known to give it
an advantage for non-compositional pairs of entities (Welbl et al., 2016).
Different possible combinations seem promising. The behaviour of RESCAL and
ComplEx on symmetric and antisymmetric experiments suggests that encoding these
patterns through complex conjugation is more stable than using the non-commutative matrix
product. But RESCAL’s matrix representations of relations helped a lot in the family
experiments, as long as the rank was not too high, suggesting that there might be a middle
ground between K and K 2 to be found for the parametric representation of the relations.
Using tridiagonal or pentadiagonal (or more) symmetric matrices for relation representations
within the ComplEx model could be an answer to these problems.
Combining the scoring functions of the TransE and F models with ComplEx could also
lead to a more robust model. The combination of bilinear and trilinear terms has already
been explored within real-valued models (Garcı́a-Durán et al., 2014), also with vectorial
weights over each term (Jenatton et al., 2012), as well as combining different pairwise terms
(Welbl et al., 2016; Singh et al., 2015), which yielded better performance in all cases.
The main defect of latent factor models that this experimental survey points to is their
low ability to transfer knowledge between disjoint set of entities, as shown in the last family
split with p = 0. Real knowledge graphs might not have fully disjoint subsets, but rather
some less-connected sub-graphs, between which this effect is likely to appear too. We believe
improving this ability of latent factor models is key.
One already-pursued way to harness this problem is to enable latent factor models to
make use of logic rules (Rocktäschel et al., 2015; Demeester et al., 2016). As already said,
those rules are not always available, and thus latent factor models should be improved in
order to have this ability to learn from disjoint subsets, while still operating without rules.
Intuitively, sharing parameters across all entity representations could also solve this
issue, as used in Bayesian clustered factorization models (Sutskever et al., 2009). Though
those models have known scalability issues. A possible, more scalable way to implement
a shared parametrization between the entity embeddings E ∈ CNe ×K is through a nested
24
On Inductive Abilities of Latent Factor Models for Relational Learning
factorization, where the matrix E is itself expressed as a low-rank factorization, as it has
already been proposed for the relation embeddings (Jenatton et al., 2012). Another one
could be a suited regularization over the whole matrix E: in most proposals E is regularized
row-wise with ||ei ||22 for all i ∈ E—as shown in Equation (10).
Another linked limitation of latent factor models—that does not require experiments
to be shown—is their inability to generalize to new entities without retraining. Indeed for
new facts involving a new entity i, its embedding ei ∈ CK is unknown. But in a logic-based
setting, only the new facts involving the new entity are necessary to infer other facts from
known rules. Some recent works started tackling this problem: Verga et al. (2017) proposed
a solution for the F model, by expressing entity pair embeddings as combinations of the
relation embeddings in which they appear. Hamaguchi et al. (2017) used graph neural
networks to handle unseen entities at test time.
The evidence split in the family experiments also pointed out a potential problem of
imbalance in the distribution of the relations across the facts when the train and test sets are
distributed differently. Correcting this imbalance via down-weighting the facts involving the
most frequent relations could be a solution, as well as sharing the parametrization between
the relations.
A non-mentioned aspect of the problem in this paper is the theoretical learnability
of such logic formulas, a field that has been extensively covered (Valiant, 1984; Kearns
and Valiant, 1994; Muggleton and De Raedt, 1994; Dzeroski and Lavrac, 1994). However
logic learnability by latent factor models has not yet been specifically studied. Recently
established links between sign-matrices complexity—specifically the sign-rank (Linial et al.,
2007)—and VC-dimension open the door to such theoretical study (Alon et al., 2016), and
possible extensions to the tensor case. This being said, theoretical guarantees generally come
under the condition that the training and test sets are drawn from the same distribution,
which is not the case in the last two splits of the family experiments: a theoretical analysis
of the learnability of such cases might require a new theoretical framework for statistical
learning.
7. Conclusion
We experimentally surveyed state-of-the-art latent factor models for link prediction in
knowledge graphs, in order to assess their ability to learn (i) binary relation properties,
and (ii) genealogical relations, directly from observed facts, as well as their robustness to
missing data. Latent factor models yield good performances in the first case, while having
more difficulties in the second one. Specifically, we show that such models do not reason
as it is generally meant for logical inference engines, as they are unable to transfer their
predictive abilities between disjoint subsets of entities. The different behaviors of the models
in each experimental setup suggest possible enhancements and research directions, including
combining them, as well as it exposes each model’s advantages and limitations.
Acknowledgments
This work was supported in part by the Association Nationale de la Recherche et de la
Technologie through the CIFRE grant 2014/0121.
25
Trouillon, Gaussier, Dance & Bouchard
References
Aaronson, S. (2013). Why philosophers should care about computational complexity. In
B. J. Copeland, C. P. and Shagrir, O., editors, Computability: Turing, Gödel, Church,
and Beyond, pages 261–328. MIT Press.
Alon, N., Moran, S., and Yehudayoff, A. (2016). Sign rank versus vc dimension. In Conference
on Learning Theory, pages 47–80.
Auer, S., Bizer, C., Kobilarov, G., Lehmann, J., and Ives, Z. (2007). DBpedia: A nucleus
for a web of open data. In International Semantic Web Conference, Busan, Korea, pages
11–15. Springer.
Bollacker, K., Evans, C., Paritosh, P., Sturge, T., and Taylor, J. (2008). Freebase: a collaboratively created graph database for structuring human knowledge. In ACM SIGMOD
International Conference on Management of Data, pages 1247–1250.
Bordes, A., Usunier, N., Garcia-Duran, A., Weston, J., and Yakhnenko, O. (2013a). Irreflexive
and hierarchical relations as translations. arXiv preprint arXiv:1304.7158.
Bordes, A., Usunier, N., Garcia-Duran, A., Weston, J., and Yakhnenko, O. (2013b). Translating embeddings for modeling multi-relational data. In Advances in Neural Information
Processing Systems, pages 2787–2795.
Bordes, A., Weston, J., Collobert, R., and Bengio, Y. (2011). Learning structured embeddings
of knowledge bases. In AAAI Conference on Artificial Intelligence.
Bottou, L. (2014). From machine learning to machine reasoning. Machine Learning,
94(2):133–149.
Bouchard, G., Singh, S., and Trouillon, T. (2015). On approximate reasoning capabilities
of low-rank vector spaces. AAAI Spring Symposium on Knowledge Representation and
Reasoning: Integrating Symbolic and Neural Approaches.
Bowman, S. R., Potts, C., and Manning, C. D. (2015). Recursive neural networks can
learn logical semantics. In ACL Workshop on Continuous Vector Space Models and their
Compositionality.
Cardano, G. (1545). Artis Magnæ, Sive de Regulis Algebraicis Liber Unus.
Caroll, J. D. and Chang, J.-J. (1970). Analysis of individual differences in multidimensional
scaling via n-way generalization of Eckart–Young decomposition. Psychometrika, 35:283–
319.
Cyganiak, R., Wood, D., and Lanthaler, M. (2014). RDF 1.1 concepts and abstract syntax.
W3C Recommendation.
Davis, R., Shrobe, H., and Szolovits, P. (1993). What is a knowledge representation? AI
magazine, 14(1):17–33.
26
On Inductive Abilities of Latent Factor Models for Relational Learning
Demeester, T., Rocktäschel, T., and Riedel, S. (2016). Lifted rule injection for relation
embeddings. In Empirical Methods in Natural Language Processing, pages 1389–1399.
Dong, X., Gabrilovich, E., Heitz, G., Horn, W., Lao, N., Murphy, K., Strohmann, T., Sun, S.,
and Zhang, W. (2014). Knowledge vault: A web-scale approach to probabilistic knowledge
fusion. In ACM SIGKDD International Conference on Knowledge Discovery and Data
Mining, pages 601–610.
Duchi, J., Hazan, E., and Singer, Y. (2011). Adaptive subgradient methods for online learning
and stochastic optimization. Journal of Machine Learning Research, 12:2121–2159.
Dzeroski, S. and Lavrac, N. (1994). Inductive logic programming: techniques and applications.
Ellis Horwood, New York.
Friedman, N., Getoor, L., Koller, D., and Pfeffer, A. (1999). Learning probabilistic relational
models. In International Joint Conference on Artificial Intelligence, volume 99, pages
1300–1309.
Galárraga, L., Teflioudi, C., Hose, K., and Suchanek, F. M. (2015). Fast rule mining in
ontological knowledge bases with amie+. The VLDB Journal, 24(6):707–730.
Garcı́a-Durán, A., Bordes, A., and Usunier, N. (2014). Effective blending of two and
three-way interactions for modeling multi-relational data. In Joint European Conference
on Machine Learning and Knowledge Discovery in Databases, pages 434–449. Springer.
Getoor, L. and Taskar, B. (2007). Introduction to Statistical Relational Learning. MIT Press.
Grefenstette, E. (2013). Towards a formal distributional semantics: Simulating logical calculi
with tensors. In Joint Conference on Lexical and Computational Semantics.
Halmos, P. R. (1998). Naive set theory. Springer Science & Business Media.
Hamaguchi, T., Oiwa, H., Shimbo, M., and Matsumoto, Y. (2017). Knowledge transfer
for out-of-knowledge-base entities: A graph neural network approach. arXiv preprint
arXiv:1706.05674.
Harshman, R. A. (1970). Foundations of the PARAFAC procedure: models and conditions
for an explanatory multimodal factor analysis. UCLA Working Papers in Phonetics,
16:1–84.
Hayashi, K. and Shimbo, M. (2017). On the equivalence of holographic and complex
embeddings for link prediction. arXiv preprint arXiv:1702.05563.
Heckerman, D., Meek, C., and Koller, D. (2007). Probabilistic entity-relationship models,
prms, and plate models. Introduction to statistical relational learning, pages 201–238.
Hinton, G. E. (1986). Learning distributed representation of concepts. In Conference of the
Cognitive Science Society.
Hitchcock, F. L. (1927). The expression of a tensor or a polyadic as a sum of products.
Journal of Mathematical Physics, 6(1):164–189.
27
Trouillon, Gaussier, Dance & Bouchard
Jenatton, R., Bordes, A., Le Roux, N., and Obozinski, G. (2012). A latent factor model
for highly multi-relational data. In Advances in Neural Information Processing Systems,
pages 3167–3175.
Kearns, M. and Valiant, L. (1994). Cryptographic limitations on learning boolean formulae
and finite automata. Journal of the ACM, 41(1):67–95.
Kersting, K. and De Raedt, L. (2001). Towards combining inductive logic programming
with bayesian networks. In International Conference on Inductive Logic Programming,
pages 118–131.
Koren, Y. (2008). Factorization meets the neighborhood: A multifaceted collaborative
filtering model. In ACM SIGKDD International Conference on Knowledge Discovery and
Data Mining, pages 426–434.
Koren, Y., Bell, R., and Volinsky, C. (2009). Matrix factorization techniques for recommender
systems. Computer, 42(8):30–37.
Lewis, M. and Steedman, M. (2013). Combining distributional and logical semantics.
Transactions of the Association for Computational Linguistics, 1:179–192.
Linial, N., Mendelson, S., Schechtman, G., and Shraibman, A. (2007). Complexity measures
of sign matrices. Combinatorica, 27(4):439–463.
Lisi, F. A. (2010). Inductive logic programming in databases: From datalog to. Theory and
Practice of Logic Programming, 10(3):331–359.
Ma, Y., Crook, P. A., Sarikaya, R., and Fosler-Lussier, E. (2015). Knowledge graph inference
for spoken dialog systems. In IEEE International Conference on Acoustics, Speech and
Signal Processing, pages 5346–5350.
Minervini, P., Demeester, T., Rocktäschel, T., and Riedel, S. (2017). Adversarial sets for
regularising neural link predictors. In Conference on Uncertainty in Artificial Intelligence.
Muggleton, S. (1995). Inverse entailment and progol. New generation computing, 13(3-4):245–
286.
Muggleton, S. and De Raedt, L. (1994). Inductive logic programming: Theory and methods.
The Journal of Logic Programming, 19:629–679.
Ngo, L. and Haddawy, P. (1997). Answering queries from context-sensitive probabilistic
knowledge bases. Theoretical Computer Science, 171(1):147–177.
Nickel, M., Murphy, K., Tresp, V., and Gabrilovich, E. (2016a). A review of relational
machine learning for knowledge graphs. Proceedings of the IEEE, 104(1):11–33.
Nickel, M., Rosasco, L., and Poggio, T. A. (2016b). Holographic embeddings of knowledge
graphs. In AAAI Conference on Artificial Intelligence, pages 1955–1961.
Nickel, M. and Tresp, V. (2013). Logistic tensor factorization for multi-relational data. arXiv
preprint arXiv:1306.2084.
28
On Inductive Abilities of Latent Factor Models for Relational Learning
Nickel, M., Tresp, V., and Kriegel, H.-P. (2011). A three-way model for collective learning on
multi-relational data. In International Conference on Machine Learning, pages 809–816.
Popper, K. (1934). Logik der Forschung. Mohr Siebeck.
Richardson, M. and Domingos, P. (2006). Markov logic networks. Machine Learning,
62(1-2):107–136.
Riedel, S., Yao, L., McCallum, A., and Marlin, B. M. (2013). Relation extraction with
matrix factorization and universal schemas. In North American Chapter of the Association
of Computational Linguistics: Human Language Technologies, pages 74–84.
Rocktäschel, T., Bosnjak, M., Singh, S., and Riedel, S. (2014). Low-dimensional embeddings
of logic. In Workshop on semantic parsing at ACL.
Rocktäschel, T. and Riedel, S. (2016). Learning knowledge base inference with neural
theorem provers. Workshop on Automated Knowledge Base Construction at NAACL-HLT,
pages 45–50.
Rocktäschel, T., Singh, S., and Riedel, S. (2015). Injecting logical background knowledge
into embeddings for relation extraction. In North American Chapter of the Association
for Computational Linguistics: Human Language Technologies, pages 1119–1129.
Singh, S., Rocktäschel, T., and Riedel, S. (2015). Towards combined matrix and tensor
factorization for universal schema relation extraction. In Workshop on Vector Space
Modeling for Natural Language Processing at NAACL-HLT, pages 135–142.
Smolensky, P., Lee, M., He, X., Yih, W.-t., Gao, J., and Deng, L. (2016). Basic reasoning
with tensor product representations. arXiv preprint arXiv:1601.02745.
Socher, R., Chen, D., Manning, C. D., and Ng, A. (2013). Reasoning with neural tensor
networks for knowledge base completion. In Advances in Neural Information Processing
Systems, pages 926–934.
Sutskever, I., Tenenbaum, J. B., and Salakhutdinov, R. R. (2009). Modelling relational
data using bayesian clustered tensor factorization. In Advances in neural information
processing systems, pages 1821–1828.
Trouillon, T., Dance, C. R., Welbl, J., Riedel, S., Gaussier, É., and Bouchard, G.
(2017). Knowledge graph completion via complex tensor factorization. arXiv preprint
arXiv:1702.06879. To appear in the Journal of Machine Learning Research.
Trouillon, T. and Nickel, M. (2017). Complex and holographic embeddings of knowledge
graphs: a comparison. International Workshop on Statistical Relational AI.
Trouillon, T., Welbl, J., Riedel, S., Gaussier, E., and Bouchard, G. (2016). Complex
embeddings for simple link prediction. In International Conference on Machine Learning,
volume 48, pages 2071–2080.
Valiant, L. G. (1984). A theory of the learnable. Communications of the ACM, 27(11):1134–
1142.
29
Trouillon, Gaussier, Dance & Bouchard
Vapnik, V. N. (1995). The Nature of Statistical Learning Theory. Springer-Verlag New York,
Inc.
Verga, P., Neelakantan, A., and McCallum, A. (2017). Generalizing to unseen entities and
entity pairs with row-less universal schema. In European Chapter of the Association of
Computational Linguistics.
Wang, W. Y. and Cohen, W. W. (2016). Learning first-order logic embeddings via matrix
factorization. In International Joint Conference on Artificial Intelligence, pages 2132–2138.
Welbl, J., Bouchard, G., and Riedel, S. (2016). A factorization machine framework for
testing bigram embeddings in knowledge base completion. In Workshop on Automated
Knowledge Base Construction at NAACL-HLT, pages 103–107.
Wellman, M. P., Breese, J. S., and Goldman, R. P. (1992). From knowledge bases to decision
models. The Knowledge Engineering Review, 7(01):35–53.
Weston, J., Bordes, A., Chopra, S., Rush, A. M., van Merriënboer, B., Joulin, A., and
Mikolov, T. (2015). Towards AI-complete question answering: A set of prerequisite toy
tasks. arXiv preprint arXiv:1502.05698.
Wikipedia (2004). Binary relation — Wikipedia, the free encyclopedia. https://en.
wikipedia.org/wiki/Binary_relation.
Yang, B., Yih, W.-t., He, X., Gao, J., and Deng, L. (2015). Embedding entities and relations
for learning and inference in knowledge bases. In International Conference on Learning
Representations.
Yoon, H.-G., Song, H.-J., Park, S.-B., and Park, S.-Y. (2016). A translation-based knowledge
graph embedding preserving logical property of relations. In North American Chapter
of the Association of Computational Linguistics: Human Language Technologies, pages
907–916.
30
On Inductive Abilities of Latent Factor Models for Relational Learning
Appendix A. Learning Algorithm
Algorithm 1 describes the stochastic gradient descent algorithm used to learn the evaluated
models, with the AdaGrad learning-rate updates (Duchi et al., 2011). The parameters are
initialized from a zero-mean normal distribution with unit variance. Squared gradients
are accumulated to compute AdaGrad learning rates, then gradients are updated. Every
s iterations, the parameters Θ are evaluated over the evaluation set Ωv , through the
evaluate AP(Ωv ; Θ) function. The optimization process is stopped when average precision
decreases compared to the last evaluation (early stopping). The sample batch of size b(Ω, b)
function sample uniformly b true and false triples uniformly at random from the training set
Ω.
Algorithm 1 Stochastic gradient descent with AdaGrad
Input Training set Ω, validation set Ωv , learning rate α ∈ R++ , rank K ∈ Z++ , L2
regularization factor λ ∈ R+ , batch size b ∈ Z++ , maximum iteration m ∈ Z++ , validate
every s ∈ Z++ iterations, AdaGrad regularizer = 10−8 .
Output Trained embeddings Θ.
Θi ∼ N (0k , I k×k ) for each i ∈ E
Θr ∼ N (0k , I k×k ) for each r ∈ R
gΘi ← 0k for each i ∈ E
gΘr ← 0k for each r ∈ R
previous score ← 0
for i = 1, . . . , m do
for j = 1, . . . , |Ω|/b do
Ωb ← sample batch of size b(Ω, b)
for ((r, s, o), yrso ) in Ωb do
for v in Θ do
// AdaGrad updates:
gv ← gv + (∇v L({((r, s, o), yrso )}; Θ))2
// Gradient updates:
v ← v − gvα+ ∇v L({((r, s, o), yrso )}; Θ)
end for
end for
end for
// Early stopping
if i mod s = 0 then
current score ← evaluate AP(Ωv ; Θ)
if current score ≤ previous score then
break
end if
previous score ← current score
end if
end for
return Θ
31
Trouillon, Gaussier, Dance & Bouchard
Appendix B. Results with Reflexivity and Irreflexivity
In this appendix we report results of the individual learning of combinations of relation properties including reflexivity and irreflexivity. Those results are included for completeness as
they are similar to the cases that are neither reflexive nor irreflexive, reported in Section 4.2.1.
Figure 12 shows results for the 5 combinations with reflexivity, and Figure 11 for the 3
combinations with irreflexivity. The irreflexive transitive case, and the irreflexive symmetric
transitive case are not reported as they are not consistent, as explained in Section 4.1. The
single noticeable difference is in the symmetric irreflexive case, where all models perform
slightly worse compared to the symmetric and symmetric reflexive cases, especially TransE.
Figure 11: Generated irreflexive relations with 50 entities, combined with symmetry (top-left),
antisymmetry (top-right) and antisymmetry and transitivity (bottom). Average precision
for each rank ranging from 5 to 50 for each model.
32
On Inductive Abilities of Latent Factor Models for Relational Learning
Figure 12: Generated reflexive relations with 50 entities, combined with symmetry (top-left),
antisymmetry (top-right), transitivity (middle), symmetry and transitivity (bottom-left)
and antisymmetry and transitivity (bottom-right). Average precision for each rank ranging
from 5 to 50 for each model.
33
| 2cs.AI
|
arXiv:1704.07548v1 [cs.AI] 25 Apr 2017
Semi-supervised Bayesian Deep Multi-modal Emotion Recognition
Changde Du1 , Changying Du2 , Jinpeng Li1 , Wei-long Zheng3 , Bao-liang Lu 3 , Huiguang He1
1
Research Center for Brain-Inspired Intelligence,
Institute of Automation, Chinese Academy of Sciences (CAS), Beijing, China
2
Laboratory of Parallel Software and Computational Science, Institute of Software, CAS, Beijing, China
3
Department of Computer Science and Engineering, Shanghai Jiao Tong University, Shanghai, China
{duchangde2016, huiguang.he}@ia.ac.cn
Abstract
In emotion recognition, it is difficult to recognize
human’s emotional states using just a single modality. Besides, the annotation of physiological emotional data is particularly expensive. These two aspects make the building of effective emotion recognition model challenging. In this paper, we first
build a multi-view deep generative model to simulate the generative process of multi-modality emotional data. By imposing a mixture of Gaussians
assumption on the posterior approximation of the
latent variables, our model can learn the shared
deep representation from multiple modalities. To
solve the labeled-data-scarcity problem, we further
extend our multi-view model to semi-supervised
learning scenario by casting the semi-supervised
classification problem as a specialized missing data
imputation task. Our semi-supervised multi-view
deep generative framework can leverage both labeled and unlabeled data from multiple modalities, where the weight factor for each modality can
be learned automatically. Compared with previous emotion recognition methods, our method is
more robust and flexible. The experiments conducted on two real multi-modal emotion datasets
have demonstrated the superiority of our framework over a number of competitors.
1 Introduction
With the development of human-computer interaction, emotion recognition has become increasingly important. Since
human’s emotion contains many nonverbal cues, various
modalities ranging from facial expressions, voice, Electroencephalogram (EEG), eye movements to other physiological signals can be used as the indicators of emotional
states [Calvo and D’Mello, 2010]. In real-world applications,
it is difficult to recognize the emotional states using just a single modality, because signals from different modalities represent different aspects of emotion and provide complementary information. Recent studies show that integrating multiple modalities can significantly boost the emotion recognition accuracy [Verma and Tiwary, 2014; Pang et al., 2015;
Lu et al., 2015; Liu et al., 2016; Soleymani et al., 2016;
Zhang et al., 2016].
The most successful approach to fuse the information from
multiple modalities is based on deep multi-view representation learning [Ngiam et al., 2011; Andrew et al., 2013;
Srivastava and Salakhutdinov, 2014; Wang et al., 2015;
Chandar et al., 2016]. For example, [Pang et al., 2015]
proposed to learn a joint density model for emotion analysis
with a multi-modal Deep Boltzmann Machine (DBM) [Srivastava and Salakhutdinov, 2014]. This multi-modal DBM
is exploited to model the joint distribution over visual, auditory, and textual features. [Liu et al., 2016] proposed
a multi-modal emotion recognition method by using multimodal Deep Autoencoders (DAE) [Ngiam et al., 2011], in
which the joint representations of EEG and eye movement
signals were extracted. Nevertheless, there are still limitations with these deep multi-modal emotion recognition methods, e.g., their performances depend on the amount of labeled
data.
By using the modern sensor equipments, we can easily
collect massive physiological signals, which are closely related to people’s emotional states. Despite the convenience
of data acquisition, the data labeling procedure requires lots
of manual efforts. Therefore, in most cases only a small
set of labeled samples is available, while the majority of
whole dataset is left unlabeled. Traditional emotion recognition approaches only utilized the limited amount of labeled
data, which may result in severe overfitting. The most attractive way to deal with this issue is based on Semi-supervised
Learning (SSL), which builds more robust model by exploiting both labeled and unlabeled data [Schels et al., 2014;
Jia et al., 2014; Zhang et al., 2016]. Though multi-modal approaches have been widely implemented for emotion recognition, very few of them explored SSL simultaneously. To
the best of our knowledge, only [Zhang et al., 2016] proposed an enhanced multi-modal co-training algorithm for
semi-supervised emotion recognition, but its shallow structure is hard to capture the high-level correlation between different modalities.
Amongst existing SSL approaches, the most competitive
one is based on deep generative models, which employs the
Deep Neural Networks (DNNs) to learn discriminative features and casts the semi-supervised classification problem as
a specialized missing data imputation task. [Kingma et al.,
2014] and [Maaløe et al., 2016] have shown that deep generative models and approximate Bayesian inference exploiting
recent advances in scalable variational methods [Kingma and
Welling, 2014; Rezende et al., 2014] can provide state-of-theart performance for semi-supervised classification. Though
the Variational Autoencoder (VAE) framework [Kingma and
Welling, 2014] has shown great advantages in SSL, its potential merits remain under-explored. For example, until recently, there was no successful multi-view extension for it.
The main difficulty lies in its inherent assumption that the
posterior approximation should be conditioned on the data
point, which is natural to single-view data but becomes problematic for multi-view case.
In this paper, we propose a novel semi-supervised multiview deep generative framework for multi-modal emotion
recognition. Our framework combines the advantages of deep
multi-view representation learning and Bayesian modeling,
thus it has sufficient flexibility and robustness in learning joint
features and classifier. Our main contributions can be summarized as follows.
• We propose a multi-view extension for VAE by imposing a mixture of Gaussians assumption on the posterior
approximation of the latent variables. For multi-view
learning, this is critical for fully exploiting the information from multiple views.
• We introduce a semi-supervised multi-modal emotion
recognition framework based on multi-view VAE. Our
framework can leverage both labeled and unlabeled samples from multiple modalities and the weight factor for
each modality can be learned automatically, which is
critical for building a robust emotion recognition system.
• We demonstrate the superiority of our framework and
provide insightful observations on two real multi-modal
emotion datasets.
2
Multi-view Variational Autoencoder for
Semi-supervised Emotion Recognition
The VAE framework has recently been introduced as a robust
model for latent feature learning [Kingma and Welling, 2014;
Rezende et al., 2014]. However, the single-view architecture
in VAE can’t effectively deal with multi-view data. In this
section, we first build a multi-view VAE, which can learn the
shared deep representation from multi-view data. And then,
we extend it to the semi-supervised scenario. Assume we
are faced with multi-view data that appears as pairs (X, y) =
({x(v) }Vv=1 , y), with observation x(v) from the v-th view and
the corresponding class label y.
2.1
Gaussian Prior and Mixture of Gaussians Posterior
Typically, both the prior p(z) and the approximate posterior
qφ (z|X) are assumed to be Gaussian distributions [Kingma
and Welling, 2014; Rezende et al., 2014] in order to maintain
mathematical and computational tractability. Although this
assumption has leaded to favorable results on several tasks,
it is clearly a restrictive and often unrealistic assumption.
Specifically, the choice of a Gaussian distribution for p(z)
and qφ (z|X) imposes a strong uni-modal structure assumption on the latent space. However, for data distributions that
are strongly multi-modal, the uni-modal Gaussian assumption inhibits the model’s ability to extract and represent important structure in the data. To improve the flexibility of
the model, one way is to impose a mixture of Gaussians assumption on p(z). However, it has the risk of creating separate “islands” of discontinuous manifolds that may break the
meaningfulness of the representation in the latent space.
To learn more powerful and expressive models – in particular, models with multi-modal latent variable structures for
multi-modal emotion recognition applications – we seek a
mixture of Gaussians for qφ (z|X), while preserving p(z) as
a standard Gaussian. Thus (cf. Fig. 1b),
p(z) = N (z|0, I) ,
V
X
qφ (z|X) =
λ(v) N z|µφ(v) (x(v) ), Σφ(v) (x(v) ) , (2)
v=1
Multi-view Variational Autoencoder
DNN-parameterized Likelihoods
We assume the latent variable z can generate multi-view features {x(v) }Vv=1 . Specifically, we assume z generates x(v) for
any v ∈ {1, ..., V }, with the following generative model (cf.
Fig. 1a):
pθ(v) (x(v) |z) = f (x(v) ; z, θ(v) ),
where f (x(v) ; z, θ(v) ) is a suitable likelihood function (e.g. a
Gaussian for continuous observation or Bernoulli for binary
observation), which is formed by a non-linear transformation
of the latent variable z. This non-linear transformation is essential to allow for higher moments of the data to be captured
by the density model, and we choose these non-linear functions to be DNNs, referred to as the generative networks, with
parameters {θ(v) }Vv=1 . Note that, the likelihoods for different
data views are assumed to be independent of each other, with
different nonlinear transformations.
The Bayesian Canonical Correlation Analysis (CCA)
model [Klami et al., 2013] can be seen as a special case of
our model, where linear shallow transformations were used
to generate each data view and only two different views were
considered. [Wang et al., 2016] used a similar deep nonlinear generative process as ours to construct deep Bayesian
CCA model, but during inference they construct the variational posterior approximation from just one view and ignore
the rest one. Such a choice is convenient for inference and
computation, but only seeks suboptimal solutions as it doesn’t
fully exploit the data. As shown in the following, we assume
the variational approximation to the posterior of latent variables to be a mixture of Gaussians, utilizing information from
multiple views.
(1)
where the mean µφ(v) and the covariance Σφ(v) are nonlinear
functions of the observation x(v) , with variational parameter
φ(v) . As in our generative model, we choose these nonlinear
functions to be DNNs, referred to as the inference networks.
λ(v) is the non-negative normalized weight factor for the vPV
th view, i.e., λ(v) > 0 and v=1 λ(v) = 1. By conditioning the posterior approximation on the data point, we avoid
p(z)
qφ (z|X)
p(y)
p(z)
qϕ (y|X)
qφ (z|X, y)
z
z
y
z
y
z
x(1)
x(V )
x(1)
x(V )
x(1)
x(V )
pθ(1) (x(1) |z)
pθ(V ) (x(V ) |z)
(a) Generative model P
x(1)
x(V )
pθ(1) (x(1) |y, z) p (V ) (x(V ) |y, z)
θ
(b) Inference model Q
(a) Generative model P
(b) Inference model Q
Figure 1: Graphical model of the multi-view VAE, where
X = {x(v) }Vv=1 .
Figure 2: Graphical model of the semiMVAE for semisupervised multi-view learning, where X = {x(v) }Vv=1 .
variational parameters per data point, instead only requiring
to fit global variational parameters. Note that, our mixed
Gaussian assumption on the variational approximation distinguishes our method from all existing ones using the autoencoding variational framework [Kingma and Welling, 2014;
Wang et al., 2016; Burda et al., 2016; Kingma et al., 2016;
Serban et al., 2016; Maaløe et al., 2016]. For multi-view
learning, this is critical for fully exploiting the information
from multiple views.
2.3
2.2
Semi-supervised Emotion Recognition
In semi-supervised classification, only a subset of the samples have corresponding class labels, and we focus on using
the multi-view VAE to build a model (semiMVAE) that learns
classifier from both labeled and unlabeled multi-view data.
Since the emotional data is continuous, we choose the Gaussian likelihoods. Then the generative model P is defined as
QV
p(y)p(z) v=1 pθ(v) (x(v) |y, z) (cf. Fig. 2a):
Variational Lower Bound
The variational lower bound on the marginal likelihood for a
single labeled data point is
Z
log pθ (X, y) = log pθ (X, y, z) dz
z
pθ (X, y, z)
≥ Eqφ (z|X,y) log
qφ (z|X, y)
X
V
= Eqφ (z|X,y)
log pθ(v) (x(v) |y, z) + log p(y)
v=1
+ log p(z) − log qφ (z|X, y)
≥ Eqφ (z|X,y)
pθ(v) (x
|y, z) = N µθ(v) (y, z),
log pθ(v) (x(v) |y, z) + log p(y)
v=1
X
X
V
V
(v)
(l)
+ log p(z) −
λ · log
λ · ωv,l
p(y) = Cat (y|π) ,
p(z) = N (z|0, I) ,
(v)
X
V
v=1
diag(σθ2(v) (y, z))
, (3)
where Cat(·) denotes the categorical distribution, y is treated
as a latent variable for the unlabeled data points, and the mean
µθ(v) and variance σθ2(v) are nonlinear functions of y and z,
with parameter θ(v) . The inference model Q is defined as
qϕ (y|X)qφ (z|X, y) (cf. Fig. 2b):
qϕ (y|X) = Cat (y|πϕ (X)) ,
(4)
V
X
qφ (z|X, y) =
λ(v) N z|µφ(v) (x(v) , y), Σφ(v) (x(v) , y) ,
(5)
where ωv,l = N µφ(v) |µφ(l) , Σφ(v) + Σφ(l) . Note that, the
Shannon entropy Eqφ (z|X,y) [− log qφ (z|X, y)] is hard to compute analytically, and we have used the Jensen’s inequality to
derive a lower bound of it:
Eqφ (z|X,y) [− log qφ (z|X, y)]
Z
V
X
(v)
=−
λ
N z|µφ(v) , Σφ(v) log qφ (z|X, y) dz
v=1
v=1
where qφ (z|X, y) is assumed to be a mixture of Gaussians
to combine the information from multiple data views. Intuitively, qφ (z|X, y), pθ(v) (x(v) |y, z) and qϕ (y|X) correspond
to the encoder, the decoder and the classifier, respectively.
For brevity, we omit the explicit dependencies on x(v) , y
and z for the moment variables mentioned above hereafter.
In principle, µθ(v) , σθ2(v) , πϕ , µφ(v) and Σφ(v) can be implemented by various DNN models, e.g., Multiple Layer Perceptrons (MLP) and Convolutional Neural Networks (CNN).
l=1
≡ −L(X, y),
≥−
V
X
λ(v) log
v=1
X
V
l=1
· N z|µφ(l) , Σφ(l)
=−
V
X
v=1
λ(l)
λ(v) log
X
V
l=1
Z
N z|µφ(v) , Σφ(v)
dz
λ(l) N µφ(v) |µφ(l) , Σφ(v) + Σφ(l)
.
For unlabeled data, we further introduce the variational dis-
tribution qϕ (y|X) for y:
Z Z
pθ (X, y, z) dy dz
log pθ (X) = log
z y
pθ (X, y, z)
≥ Eqϕ,φ (y,z|X) log
qϕ,φ (y, z|X)
= Eqϕ (y|X) − L(X, y) − log qϕ (y|X)
≡ −U(X),
φ(l) and λ(l) can be efficiently estimated using the following
Monte-Carlo estimators,
∂
Eq (z|X,y) [log pθ(v) (x(v) |y, z)]
∂θ(v) φ
V
X
∂
(v)
(l)
=
λ(l) EN (ǫ(l) |0,I)
log
p
x
|y,
z
(v)
θ
∂θ(v)
l=1
(6)
with qϕ,φ (y, z|X) = qϕ (y|X)qφ (z|X, y). The objective function for the entire dataset is now:
X
X
U(X) ,
J =
L(X, y) +
(7)
(X,y)∈Sl
X∈Su
where Sl and Su are labeled and unlabeled dataset, respectively. The classification accuracy can be improved by introducing an explicit classification loss for labeled data. The
extended objective function is:
X
[− log qϕ (y|X)] ,
F =J +α·
(8)
(X,y)∈Sl
where the hyper-parameter α is a weight between generative
and discriminative learning. We set α = β · (Nl + Nu ),
where β is a scaling constant, and Nl and Nu are the numbers of labeled and unlabeled data points in one minibatch,
respectively. Note that, the classifier qϕ (y|X) is also used at
test phase for the prediction of unseen data.
2.4
Eqφ (z|X,y) [log pθ(v) (x(v) |y, z)]
V
X
l=1
(10)
∂
Eq (z|X,y) [log pθ(v) (x(v) |y, z)]
∂φ(l) φ
h ∂
∂
= λ(l) (l) EN (ǫ(l) |0,I)
log pθ(v) (x(v) |y, z(l) )
∂φ
∂z(l)
∂µ (l)
∂Rφ(l) (l) i
φ
·
+
ǫ
∂φ(l)
∂φ(l)
T
λ(l) X ∂
≈
log pθ(v) (x(v) |y, z(l,t) )
T t=1 ∂z(l,t)
∂µ (l)
∂Rφ(l) (l,t)
φ
,
(11)
+
ǫ
·
∂φ(l)
∂φ(l)
∂
Eq (z|X,y) [log pθ(v) (x(v) |y, z)]
∂λ(l) φ
= EN (ǫ(l) |0,I) log pθ(v) (x(v) |y, z(l) )
≈
Reparameterization Trick
The reparameterization trick is a vital component of the algorithm, because it allows us to easily take the derivative of
Eqφ (z|X,y) [log pθ(v) (x(v) |y, z)] with respect to the variational
parameters φ. However, the use of a mixture of Gaussians
for the variational distribution qφ (z|X, y) makes the application of reparameterization trick challenging. It can be shown
that, for any v ∈ {1, ..., V }, Eqφ (z|X,y) [log pθ(v) (x(v) |y, z)]
can be rewritten, using the location-scale transformation for
the Gaussian distribution, as:
T
V
λ(l) X X ∂
log pθ(v) x(v) |y, z(l,t) ,
(v)
T t=1
∂θ
l=1
Optimization
Eq. (8) provides a unified objective function for optimizing the parameters of encoder, decoder and classifier networks. This optimization can be done jointly, without resort to the variational EM algorithm, using the stochastic backpropagation technique [Kingma and Welling, 2014;
Rezende et al., 2014].
=
≈
T
1 X
log pθ(v) x(v) |y, z(l,t) ,
T t=1
(12)
where z(l) is evaluated at z(l) = µφ(l) + Rφ(l) ǫ(l) and
z(l,t) = µφ(l) + Rφ(l) ǫ(l,t) with ǫ(l,t) ∼ N (0, I). In practice,
it suffices to use a small T (e.g. T = 1) and then estimate the
gradient using minibatches of data points. We use the same
random numbers ǫ(l,t) for all estimators to have lower variance. The gradient w.r.t. ϕ is omitted here, since it can be
derived straightforwardly by using traditional reparameterization trick [Kingma et al., 2014].
The gradients of the loss for semiMVAE (Eq. (8)) can then
be computed by a direct application of the chain rule and estimators presented above. During optimization we can use
the estimated gradients in conjunction with standard stochastic gradient based optimization methods such as SGD, RMSprop or Adam [Kingma and Ba, 2014]. Overall, the model
can be trained with reparameterization trick for backpropagation through the mixed Gaussian latent variables.
3 Experiments
λ(l) EN (ǫ(l) |0,I) log pθ(v) (x(v) |y, µφ(l) + Rφ(l) ǫ(l) ) ,
(9)
In this section, we present extensive experimental results
to demonstrate the effectiveness of the proposed semisupervised multi-view framework for emotion recognition.
where Rφ(l) R⊤
= Σφ(l) and l ∈ {1, ..., V }.
φ(l)
3.1
Gradients of the Objective
While the expectations on the right hand side of Eq. (9)
still cannot be solved analytically, their gradients w.r.t. θ(v) ,
Data description Two multi-modal emotion datasets,
the SEED dataset1 [Lu et al., 2015] and the DEAP
1
Experimental Testbed and Setup
http://bcmi.sjtu.edu.cn/%7Eseed/index.html
dataset2 [Koelstra et al., 2012], were used in our experiments.
The SEED dataset contains EEG and eye movement signals from 15 subjects during watching 15 movie clips, where
each movie clip lasts about 4 minutes long. The EEG signals
were recorded from 62 channels and the eye movement signals contained information about blink, saccade fixation and
so on. In our experiment, we used the data from 9 subjects
across 3 sessions, totally 27 data files. For each data file, data
from watching the 1-9 movie clips were used as training set,
while data from watching the 10-12 movie clips were used as
validation set and the rest (13-15) were used as testing set.
The DEAP dataset contains EEG and peripheral physiological signals of 32 participants. Signals were recorded when
they were watching 40 one-minute duration music videos.
The EEG signals were recorded from 32 channels, whereas
the peripheral physiological signals were recorded from 8
channels. The participants, using values from 1 to 9, rated
each music video in terms of the levels of valence, arousal
and so on. In our experiment, the valence-arousal space was
divided into four quadrants according to the ratings. The
threshold we used was 5, leading to four classes of data. Considering the fuzzy boundary of emotions and the variations of
participants’ ratings possibly associated with individual difference in rating scale, we discarded the samples whose ratings of arousal and valence are between 3 and 6. The dataset
was randomly divided into 10-folds, where 8 folds for training, one fold for validation and the last fold for testing. The
size of testing set is relative small, because some graph-based
semi-supervised baselines are hard to deal with large dataset.
Feature selection For SEED dataset, Lu et al. have extracted the Differential Entropy (DE) features and 33 eye
movement features from EEG and eye movement signals [Lu
et al., 2015]. We also used these features in our experiments.
For DEAP dataset, we extracted the DE features from EEG
and peripheral physiological signals. The DE features can be
calculated in four frequency bands: theta (4-8Hz), alpha (814Hz), beta (14-31Hz), and gamma (31-45Hz), and we used
all band’s features. The details of the data used in our experiments were summarized in Table 1.
Table 1: The details of the datasets used in our experiments.
Datasets #Instances
SEED
DEAP
22734
21042
#Features
310(EEG), 33(Eye)
128(EEG), 32(Phy.)
#Training #Validation #Testing #Classes
13473
16834
4725
2104
4536
2104
3
4
Compared methods We compared our semiMVAE with
a broad range of solutions, including supervised learning,
transductive and inductive semi-supervised learning. We
briefly summarize the various baselines in the following.
• MAE: the multi-view extension of deep autoencoders,
which can be used to extract the joint representations
from multiple modalities [Ngiam et al., 2011].
• DCCA: the full deep neural network extension of
Canonical Correlation Analysis (CCA). DCCA can learn
deep nonlinear mappings of two views, which are maximally correlated [Andrew et al., 2013].
• DCCAE: a deep multi-view representation learning
model which combines the advantages of the DCCA
2
http://www.eecs.qmul.ac.uk/mmv/datasets/deap/download.html
and deep autoencoders. In particular, DCCAE consists of two autoencoders and optimizes the combination
of canonical correlation between the learned bottleneck
representations and the reconstruction errors of the autoencoders [Wang et al., 2015].
• AMMSS: a graph-based multi-view semi-supervised
classification algorithm, which can integrate heterogeneous features from both labeled and unlabeled data [Cai
et al., 2013].
• AMGL: a latest auto-weighted multiple graph learning
framework, which can be applied to multi-view semisupervised classification task [Nie et al., 2016].
• semiVAE: a single-view semi-supervised deep generative model proposed in [Kingma et al., 2014]. We evaluate semiVAE’s performance for each modality and the
concatenation of all modalities, respectively.
For MAE, DCCA and DCCAE, we used the Support Vector
Machines3 (SVM) and transductive SVM4 (TSVM) for supervised learning and transductive semi-supervised learning,
respectively.
Parameter setting For semiMVAE, we considered multiple layer perceptrons as the type of inference and generative
networks. On both datasets, we set the structures of the inference and generative networks for each view as ‘100-5030’ and ‘30-50-100’, respectively. We used the Adam optimizer [Kingma and Ba, 2014] with a learning rate η =
3 × 10−4 in training. The scaling constant β was selected
from {0.1, 0.5, 1} throughout the experiments. The weight
factor for each view was initialized with λ(v) = 1/V , where
V is the number of views. For MAE, DCCA and DCCAE,
we considered the same setups (network structure, learning
rate, etc.) as our semiMVAE. For AMMSS, we tuned the parameters as suggested in [Cai et al., 2013]. For AMGL and
semiVAE, we used their default settings.
3.2
Performance Evaluation
To simulate semi-supervised learning scenario, on both
datasets, we randomly labeled different proportions of samples in the training set, and remained the rest samples in
the training set unlabeled. For transductive semi-supervised
learning, we trained models on the dataset consisting of the
testing data and labeled data belonging to training set. For inductive semi-supervised learning, we trained models on the
entire training set consisting of the labeled and unlabeled
data. For supervised learning, we trained models on the labeled data belonging to training set, and test their performance on the testing set. Table 2 presents the classification
accuracies of all methods on SEED and DEAP datasets. The
proportions of labeled samples in the training set vary from
1% to 3%. Several observations can be drawn as follows.
First, the average accuracy of semiMVAE significantly surpasses the baselines in all cases. Second, by examining
semiMVAE against supervised learning approaches trained
on very limited labeled data, we can find that semiMVAE always outperforms them. This encouraging result shows that
3
4
http://www.csie.ntu.edu.tw/%7Ecjlin/liblinear/.
http://svmlight.joachims.org/.
2% labeled
3% labeled
Inductive
semisupervised
learning
MAE+SVM
DCCA+SVM
DCCAE+SVM
AMMSS
AMGL
MAE+TSVM
DCCA+TSVM
DCCAE+TSVM
semiMVAE
semiVAE (Eye)
semiVAE (EEG)
semiVAE (Concat.)
semiMVAE
.814±.031
.809±.035
.819±.036
.731±.055
.711±.047
.818±.035
.811±.031
.823±.040
.861±.037
.753±.024
.768±.041
.803±.035
.880±.033
.896±.024
.891±.035
.893±.034
.839±.036
.817±.023
.910±.025
.903±.024
.907±.027
.931±.020
.849±.055
.861±.040
.876±.043
.955±.020
.925±.024
.923±.028
.923±.027
.912±.018
.886±.028
.931±.026
.928±.021
.929±.023
.960±.021
.899±.049
.919±.026
.926±.044
.968±.015
DEAP data
Algorithms
1% labeled
2% labeled
3% labeled
MAE+SVM
DCCA+SVM
DCCAE+SVM
AMMSS
AMGL
MAE+TSVM
DCCA+TSVM
DCCAE+TSVM
semiMVAE
semiVAE (Phy.)
semiVAE (EEG)
semiVAE (Concat.)
semiMVAE
.353±.027
.359±.016
.361±.023
.303±.029
.291±.027
.376±.025
.379±.021
.384±.022
.424±.020
.366±.024
.374±.019
.383±.019
.421±.019
.387±.014
.400±.014
.403±.017
.353±.024
.341±.021
.403±.031
.408±.024
.412±.027
.441±.013
.389±.048
.397±.013
.404±.016
.439±.025
.411±.016
.416±.018
.419±.013
.386±.014
.367±.019
.417±.026
.421±.017
.425±.021
.456±.013
.402±.034
.407±.016
.416±.012
.451±.022
Supervised
learning
Transductive
semisupervised
learning
Supervised
learning
Transductive
semisupervised
learning
Inductive
semisupervised
learning
1
0.5
0.95
0.48
0.9
0.46
0.85
0.8
0.75
inductive
transductive
0.7
0.65
Average accuracy
Average accuracy
semiMVAE can effectively leverage the useful information
from unlabeled data. Third, multi-view semi-supervised algorithms AMMSS and AMGL perform worst in all cases.
We attribute this to the fact that graph-based shallow models AMMSS and AMGL can’t extract the deep features from
the original data. Fourth, the performances of three TSVM
based semi-supervised methods are moderate. Although
MAE+TSVM, DCCA+TSVM and DCCAE+TSVM can also
integrate multi-modality information from unlabeled samples, their two-stage learning can’t obtain the global optimal
model parameters. Finally, compared with the single-view
semi-supervised method semiVAE, our multi-view method is
more effective in integrating multiple modalities.
0.42
0.4
0.36
inductive
transductive
0.34
0.5% 0.8% 1%
2%
3%
4%
5%
0.46
1% labeled
2% labeled
3% labeled
0.44
0.42
0.4
0.38
1% labeled
2% labeled
3% labeled
0.36
0.34
0.32
10%
30%
50%
70%
90%
10%
30%
50%
70%
Unlabeled proportion
Unlabeled proportion
(a) SEED dataset
(b) DEAP dataset
90%
Figure 4: Inductive semiMVAE’s performance with different
proportions of unlabeled samples in the training set.
models. Figs. 3 and 4 show the changes of semiMVAE’s average accuracy on both datasets with different proportions of
labeled and unlabeled samples in the training set. We can observe that both labeled and unlabeled samples can effectively
boost the classification accuracy of semiMVAE.
Instead of treating each modality equally, our semiMVAE
can weight each modality and perform classification simultaneously. Fig. 5a shows the learned weight factors by inductive semiMVAE on SEED and DEAP datasets (1% labeled).
From it, we can observe that EEG modality has the highest weight on both datasets, which is consistent with single
modality’s performance of semiVAE shown in Table 2 and
the results in previous work [Lu et al., 2015].
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
1
0.9
EEG
EEG
Phy.
Eye
SEED
DEAP
0.8
0.7
SEED
DEAP
0.6
0.5
0.4
0.3
0.2
0.01 0.05 0.1
Dataset
(a)
0.5
1
5
10
β
(b)
Figure 5: (a) Learned weight factors by inductive semiMVAE. (b) The impact of scaling constant β.
The scaling constant β controls the weight of discriminative learning in semiMVAE. Fig. 5b shows the performance
of inductive semiMVAE with different β values (1% labeled).
From it, we can find that the scaling constant β can be chosen
from {0.1, 0.5, 1}, where semiMVAE achieves good results.
0.44
0.38
0.48
Average accuracy
1% labeled
0.98
0.96
0.94
0.92
0.9
0.88
0.86
0.84
0.82
0.8
0.78
0.76
Average accuracy
Algorithms
Average accuracy
SEED data
Weight factor
Table 2: Comparison with several supervised and semisupervised methods on SEED and DEAP with few labels.
Results (mean±std) were averaged over 20 independent runs.
0.5% 0.8% 1% 2% 3% 4% 5%
Labeled proportion
Labeled proportion
(a) SEED dataset
(b) DEAP dataset
Figure 3: semiMVAE’s performance with different proportions of labeled samples in the training set.
The proportion of labeled and unlabeled samples in the
training set will affect the performance of semi-supervised
4 Conclusion
This paper proposes a semi-supervised multi-view deep generative framework for emotion recognition, which can leverage both labeled and unlabeled data from multiple modalities. The key to the framework are two parts: 1) multiview VAE can fully integrate the information from multiple
modalities and 2) semi-supervised learning can overcome the
labeled-data-scarcity problem. Experimental results on two
real multi-modal emotion datasets demonstrate the effectiveness of our approach.
References
[Andrew et al., 2013] Galen Andrew, Raman Arora, Jeff A
Bilmes, and Karen Livescu. Deep canonical correlation
analysis. In ICML, pages 1247–1255, 2013.
[Burda et al., 2016] Yuri Burda, Roger Grosse, and Ruslan
Salakhutdinov. Importance weighted autoencoders. In
ICLR, 2016.
[Cai et al., 2013] Xiao Cai, Feiping Nie, Weidong Cai, and
Heng Huang. Heterogeneous image features integration via multi-modal semi-supervised learning model. In
ICCV, pages 1737–1744, 2013.
[Calvo and D’Mello, 2010] Rafael A Calvo and Sidney
D’Mello. Affect detection: An interdisciplinary review
of models, methods, and their applications. IEEE Transactions on Affective Computing, 1(1):18–37, 2010.
[Chandar et al., 2016] Sarath Chandar, Mitesh M Khapra,
Hugo Larochelle, and Balaraman Ravindran. Correlational neural networks. Neural computation, 28(2):257–
285, 2016.
[Jia et al., 2014] Xiaowei Jia, Kang Li, Xiaoyi Li, and
Aidong Zhang. A novel semi-supervised deep learning
framework for affective state recognition on EEG signals.
In International Conference on Bioinformatics and Bioengineering (BIBE), pages 30–37. IEEE, 2014.
[Kingma and Ba, 2014] Diederik Kingma and Jimmy Ba.
Adam: A method for stochastic optimization. arXiv
preprint arXiv:1412.6980, 2014.
[Kingma and Welling, 2014] Diederik P Kingma and Max
Welling. Auto-encoding variational bayes. In ICLR, 2014.
[Kingma et al., 2014] Diederik P Kingma, Shakir Mohamed,
Danilo Jimenez Rezende, and Max Welling. Semisupervised learning with deep generative models. In NIPS,
pages 3581–3589, 2014.
[Kingma et al., 2016] Diederik P Kingma, Tim Salimans,
and Max Welling. Improving variational inference with
inverse autoregressive flow. In NIPS, 2016.
[Klami et al., 2013] Arto Klami, Seppo Virtanen, and
Samuel Kaski. Bayesian canonical correlation analysis.
Journal of Machine Learning Research, 14(1):965–1003,
2013.
[Koelstra et al., 2012] Sander Koelstra, Christian Muhl, Mohammad Soleymani, Jong-Seok Lee, Ashkan Yazdani,
Touradj Ebrahimi, Thierry Pun, Anton Nijholt, and Ioannis Patras. Deap: A database for emotion analysis; using physiological signals. IEEE Transactions on Affective
Computing, 3(1):18–31, 2012.
[Liu et al., 2016] Wei Liu, Wei-Long Zheng, and Bao-Liang
Lu. Multimodal emotion recognition using multimodal
deep learning. arXiv preprint arXiv:1602.08225, 2016.
[Lu et al., 2015] Yifei Lu, Wei-Long Zheng, Binbin Li, and
Bao-Liang Lu. Combining eye movements and EEG to
enhance emotion recognition. In IJCAI, pages 1170–1176,
2015.
[Maaløe et al., 2016] Lars Maaløe, Casper Kaae Sønderby,
Søren Kaae Sønderby, and Ole Winther. Auxiliary deep
generative models. In ICML, pages 1445–1453, 2016.
[Ngiam et al., 2011] Jiquan Ngiam, Aditya Khosla, Mingyu
Kim, Juhan Nam, Honglak Lee, and Andrew Y Ng. Multimodal deep learning. In ICML, pages 689–696, 2011.
[Nie et al., 2016] Feiping Nie, Jing Li, Xuelong Li, et al.
Parameter-free auto-weighted multiple graph learning: A
framework for multiview clustering and semi-supervised
classification. In IJCAI, pages 1881–1887, 2016.
[Pang et al., 2015] Lei Pang, Shiai Zhu, and Chong-Wah
Ngo. Deep multimodal learning for affective analysis and
retrieval. IEEE Transactions on Multimedia, 17(11):2008–
2020, 2015.
[Rezende et al., 2014] Danilo Jimenez Rezende, Shakir Mohamed, and Daan Wierstra. Stochastic backpropagation
and approximate inference in deep generative models. In
NIPS, pages 1278–1286, 2014.
[Schels et al., 2014] Martin Schels, Markus Kächele,
Michael Glodek, David Hrabal, Steffen Walter, and
Friedhelm Schwenker. Using unlabeled data to improve
classification of emotional states in human computer
interaction. Journal on Multimodal User Interfaces,
8(1):5–16, 2014.
[Serban et al., 2016] Iulian V Serban, II Ororbia, G Alexander, Joelle Pineau, and Aaron Courville.
Multimodal variational encoder-decoders.
arXiv preprint
arXiv:1612.00377, 2016.
[Soleymani et al., 2016] Mohammad Soleymani, Sadjad
Asghari-Esfeden, Yun Fu, and Maja Pantic. Analysis
of EEG signals and facial expressions for continuous
emotion detection.
IEEE Transactions on Affective
Computing, 7(1):17–28, 2016.
[Srivastava and Salakhutdinov, 2014] Nitish Srivastava and
Ruslan Salakhutdinov. Multimodal learning with deep
boltzmann machines. Journal of Machine Learning Research, 15:2949–2980, 2014.
[Verma and Tiwary, 2014] Gyanendra K Verma and
Uma Shanker Tiwary. Multimodal fusion framework: A
multiresolution approach for emotion classification and
recognition from physiological signals. NeuroImage,
102:162–172, 2014.
[Wang et al., 2015] Weiran Wang, Raman Arora, Karen
Livescu, and Jeff A Bilmes. On deep multi-view representation learning. In ICML, pages 1083–1092, 2015.
[Wang et al., 2016] Weiran Wang, Xinchen Yan, Honglak
Lee, and Karen Livescu. Deep variational canonical correlation analysis. arXiv: 1610.03454, 2016.
[Zhang et al., 2016] Zixing Zhang, Fabien Ringeval, Bin
Dong, Eduardo Coutinho, Erik Marchi, and Björn
Schüller. Enhanced semi-supervised learning for multimodal emotion recognition. In ICASSP, pages 5185–5189.
IEEE, 2016.
| 2cs.AI
|
Unsupervised Grammar Induction with Depth-bounded PCFG
arXiv:1802.08545v2 [cs.CL] 26 Feb 2018
Lifeng Jin
Department of Linguistics
The Ohio State University
[email protected]
Timothy Miller
Finale Doshi-Velez
Boston Children’s Hospital &
Harvard University
Harvard Medical School
[email protected]
[email protected]
William Schuler
Department of Linguistics
The Ohio State University
[email protected]
Abstract
There has been recent interest in applying cognitively or empirically motivated
bounds on recursion depth to limit the
search space of grammar induction models
(Ponvert et al., 2011; Noji and Johnson, 2016;
Shain et al., 2016). This work extends this
depth-bounding approach to probabilistic
context-free grammar induction (DB-PCFG),
which has a smaller parameter space than
hierarchical sequence models, and therefore
more fully exploits the space reductions of
depth-bounding. Results for this model on
grammar acquisition from transcribed childdirected speech and newswire text exceed or
are competitive with those of other models
when evaluated on parse accuracy. Moreover,
grammars acquired from this model demonstrate a consistent use of category labels,
something which has not been demonstrated
by other acquisition models.
1 Introduction
Grammar acquisition or grammar induction
(Carroll and Charniak, 1992) has been of interest to linguists and cognitive scientists for
decades.
This task is interesting because a
well-performing acquisition model can serve
as a good baseline for examining factors of
grounding
(Zettlemoyer and Collins, 2005;
Kwiatkowski et al., 2010), or as a piece of evidence (Clark, 2001; Zuidema, 2003) about the
Distributional Hypothesis (Harris, 1954) against the
poverty of the stimulus (Chomsky, 1965). Unfortunately, previous attempts at inducing unbounded
Lane Schwartz
Department of Linguistics
University of Illinois at Urbana-Champaign
[email protected]
context-free
grammars
(Johnson et al., 2007;
Liang et al., 2009) converged to weak modes of a
very multimodal distribution of grammars. There
has been recent interest in applying cognitively or
empirically motivated bounds on recursion depth to
limit the search space of grammar induction models
(Ponvert et al., 2011;
Noji and Johnson, 2016;
Shain et al., 2016).
Ponvert et al. (2011) and
Shain et al. (2016) in particular report benefits
for depth bounds on grammar acquisition using
hierarchical sequence models, but either without the
capacity to learn full grammar rules (e.g. that a noun
phrase may consist of a noun phrase followed by a
prepositional phrase), or with a very large parameter
space that may offset the gains of depth-bounding.
This work extends the depth-bounding approach
to directly induce probabilistic context-free grammars,1 which have a smaller parameter space
than hierarchical sequence models, and therefore
arguably make better use of the space reductions
of depth-bounding.
This approach employs a
procedure for deriving a sequence model from a
PCFG (van Schijndel et al., 2013), developed in the
context of a supervised learning model, and adapts
it to an unsupervised setting.
Results for this model on grammar acquisition from transcribed child-directed speech and
newswire text exceed or are competitive with those
of other models when evaluated on parse accuracy. Moreover, grammars acquired from this model
demonstrate a consistent use of category labels, as
shown in a noun phrase discovery task, something
which has not been demonstrated by other acquisi1
https://github.com/lifengjin/db-pcfg
tion models.
2 Related work
This paper describes a Bayesian Dirichlet model
of depth-bounded probabilistic context-free grammar (PCFG) induction. Bayesian Dirichlet models have been applied to the related area of latent variable PCFG induction (Johnson et al., 2007;
Liang et al., 2009), in which subtypes of categories
like noun phrases and verb phrases are induced on
a given tree structure. The model described in this
paper is given only words and not only induces categories for constituents but also tree structures.
There are a wide variety of approaches to
grammar induction outside the Bayesian modeling
paradigm.
The CCL system (Seginer, 2007a)
uses deterministic scoring systems to generate bracketed output of raw text.
UPPARSE
(Ponvert et al., 2011) uses a cascade of HMM chunkers to produce syntactic structures. BMMM+DMV
(Christodoulopoulos et al., 2012)
combines
an unsupervised part-of-speech (POS) tagger
BMMM and an unsupervised dependency grammar inducer DMV (Klein and Manning, 2004).
The BMMM+DMV system alternates between phases of inducing POS tags and
inducing dependency structures.
A large
amount
work
(Klein and Manning, 2002;
Klein and Manning, 2004;
Bod, 2006;
Berg-Kirkpatrick et al., 2010;
Gillenwater et al., 2011;
Headden et al., 2009;
Bisk and Hockenmaier, 2013;
Scicluna and de la Higuera, 2014;
Jiang et al., 2016;
Han et al., 2017) has been
on grammar induction with input annotated
with POS tags, mostly for dependency grammar
induction. Although POS tags can also be induced, this separate induction has been criticized
(Pate and Johnson, 2016) for missing an opportunity to leverage information learned in grammar
induction to estimate POS tags. Moreover, most
of these models explore a search space that includes syntactic analyses that may be extensively
center embedded and therefore are unlikely to
be produced by human speakers. Unlike most of
these approaches, the model described in this paper
uses cognitively motivated bounds on the depth of
human recursive processing to constrain its search
of possible trees for input sentences.
Some previous work uses depth bounds in the
form of sequence models (Ponvert et al., 2011;
Shain et al., 2016), but these either do not
produce complete phrase structure grammars
(Ponvert et al., 2011) or do so at the expense of large
parameter sets (Shain et al., 2016). Other work implements depth bounds on left-corner configurations
of dependency grammars (Noji and Johnson, 2016),
but the use of a dependency grammar makes the
system impractical for addressing questions of
how category types such as noun phrases may be
learned. Unlike these, the model described in this
paper induces a PCFG directly and then bounds it
with a model-to-model transform, which yields a
smaller space of learnable parameters and directly
models the acquisition of category types as labels.
Some induction models learn semantic
grammars from text annotated with semantic predicates (Zettlemoyer and Collins, 2005;
Kwiatkowski et al., 2012). There is evidence humans use semantic bootstrapping during grammar
acquisition (Naigles, 1990), but these models typically rely on a set of pre-defined universals, such as
combinators (Steedman, 2000), which simplify the
induction task. In order to help address the question
of whether such universals are indeed necessary
for grammar induction, the model described in
this paper does not assume any strong universals
except independently motivated limits on working
memory.
3 Background
Like
Noji and Johnson (2016)
and
Shain et al. (2016), the model described in this
paper defines bounding depth in terms of memory elements required in a left-corner parse. A
left-corner parser (Rosenkrantz and Lewis, 1970;
Johnson-Laird, 1983;
Abney and Johnson, 1991;
Resnik, 1992) uses a stack of memory elements
to store derivation fragments during incremental
processing. Each derivation fragment represents
a disjoint connected component of phrase structure a/b consisting of a top sign a lacking a bottom
sign b yet to come. For example, Figure 1 shows
the derivation fragments in a traversal of a phrase
S
NP
VP
NP
N
NP
VP
Probabilities from these distributions are then multiplied together to define a transition model M over
hidden states:
D
N
broke
VP
man
NP
the
N
horse
the
cart
D
RC
pulled
NP
the
2. boolean variables for decisions to ‘fork out’ ft
and ‘join in’ jt derivation fragments (in
Johnson-Laird (1983) terms, to shift with or
without match and to predict with or without
match).
RC
bought
D
tom signs, adt and bdt , of derivation fragments
at each depth level d (which correspond to left
and right children in tree structure), and
M[qt−1 ,qt ] = P(qt | qt−1 )
= P( ft pt jt a1..D
b1..D
| qt−1 )
t
t
1. preterminal labels pt and labels of top and bot-
(1b)
= P( ft | qt−1 )
Figure 1: Derivation fragments before the word man in
a left-corner traversal of the sentence The cart the horse
the man bought pulled broke.
structure tree for the sentence The cart the horse the
man bought pulled broke. Immediately before processing the word man, the traversal has recognized
three fragments of tree structure: two from category
NP to category RC (covering the cart and the horse)
and one from category NP to category N (covering
the). Derivation fragments at every time step are
numbered top-down by depth d to a maximum depth
of D. A left-corner parser requires more derivation
fragments — and thus more memory — to process
center-embedded constructions than to process leftor right-embedded constructions, consistent with
observations that center embedding is more difficult
for humans to process (Chomsky and Miller, 1963;
Miller and Isard, 1964).
Grammar acquisition
models (Noji and Johnson, 2016; Shain et al., 2016)
then restrict this memory to some low bound: e.g.
two derivation fragments.
For sequences of observed word tokens wt for
time steps t ∈ {1..T }, sequence models like
Ponvert et al. (2011) and Shain et al. (2016) hypothesize sequences of hidden states qt . Models like
Shain et al. (2016) implement bounded grammar
rules as depth bounds on a hierarchical sequence
model implementation of a left-corner parser, using
random variables within each hidden state qt for:
(1a)
def
· P(pt | qt−1 ft )
· P( jt | qt−1 ft pt )
· P(a1..D
| qt−1 ft pt jt )
t
· P(b1..D
| qt−1 ft pt jt a1..D
)
t
t
(1c)
For example, just after the word horse is recognized in Figure 1, the parser store contains two
derivation fragments yielding the cart and the horse,
both with top category NP and bottom category RC.
The parser then decides to fork out the next word the
based on the bottom category RC of the last derivation fragment on the store. Then the parser generates a preterminal category D for this word based on
this fork decision and the bottom category of the last
derivation fragment on the store. Then the parser decides not to join the resulting D directly to the RC
above it, based on these fork and preterminal decisions and the bottom category of the store. Finally
the parser generates NP and N as the top and bottom
categories of a new derivation fragment yielding just
the new word the based on all these previous decisions, resulting in the store state shown in the figure.
The model over the fork decision (shift with
or without match) is defined in terms of a depthspecific sub-model θF,d̄ , where ⊥ is an empty derivation fragment and d̄ is the depth of the deepest nonempty derivation fragment at time step t − 1:
def
P( ft | qt−1 ) = PθF,d̄ ( ft | bd̄t−1 ); d̄ = max{bdt−1 ,⊥} (2)
d
The model over the preterminal category label is
then conditioned on this fork decision. When there
is no fork, the preterminal category label is deterministically linked to the category label of the bottom sign of the deepest derivation fragment at the
previous time step (using ~φ as a deterministic indicator function, equal to one when φ is true and zero
otherwise). When there is a fork, the preterminal
category label is defined in terms of a depth-specific
sub-model θP,d̄ :2
P(pt | qt−1
if ft = 0
def
~pt =bd̄t−1
ft ) =
d̄
Pθ (pt | b ) if ft = 1
P,d̄
t−1
(3)
The model over the join decision (predict with or
without match) is also defined in terms of a depthspecific sub-model θJ,d̄ with parameters depending
on the outcome of the fork decision:3
d̄
def
PθJ,d¯ ( jt | bd̄−1
t−1 at−1 ) if ft =0 (4)
P( jt | qt−1 ft pt ) =
Pθ ¯ ( jt | bd̄ pt ) if ft =1
J,d+1
t−1
Decisions about the top categories of derivation
fragments a1..D
(which correspond to left siblings in
t
tree structures) are decomposed into fork- and joinspecific cases. When there is a join, the top category
of the deepest derivation fragment deterministically
depends on the corresponding value at the previous
time step. When there is no join, the top category is
defined in terms of a depth-specific sub-model:4
def
| qt−1 ft pt jt ) =
PθA (a1..D
t
φd̄−2 · ~atd̄−1 = ad̄−1
t−1
d̄
d̄
φd̄−1 · PθA,d¯ (at | bd̄−1
t−1 at−1 )
φd̄−1 · ~ad̄t = ad̄t−1
d̄+1 | bd̄ p )
φ · Pθ
¯ (at
d̄−0
A,d+1
t−1 t
· ψd̄+0
· ψd̄+1
· ψd̄+1
· ψd̄+2
if
if
if
if
ft ,
ft ,
ft ,
ft ,
jt = 0, 1
jt = 0, 0
jt = 1, 1
jt = 1, 0
(5)
Decisions about the bottom categories b1..D
t
(which correspond to right children in tree structures) also depend on the outcome of the fork and
join variables, but are defined in terms of a side- and
2
Here, again, d̄ = maxd {bdt−1 , ⊥}.
Again, d̄ = maxd {bdt−1 , ⊥}.
1..d¯
d¯
4
, ψd̄ = ~atd̄+1..D = ⊥, and again,
= at−1
Here φd̄ = ~a1..
t
d
,
⊥}.
d̄ = maxd {bt−1
3
depth-specific sub-model in every case:5
def
PθB (b1..D
| qt−1 ft pt jt a1..D
)=
t
t
d̄−1 | bd̄−1 ad̄ ) · ψ
φd̄−2 · PθB,R,d−1
¯ (bt
d̄+0
t−1 t−1
· ψd̄+1
φd̄−1 · PθB,L,d¯ (bd̄t | ad̄t ad̄t−1 )
· ψd̄+1
φd̄−1 · PθB,R,d¯ (bd̄t | bd̄t−1 pt )
d̄+1
d̄+1
φ · Pθ
| at pt ) · ψd̄+2
¯ (bt
d̄−0
B,L,d+1
if
if
if
if
ft , jt = 0, 1
ft , jt = 0, 0
ft , jt = 1, 1
ft , jt = 1, 0
(6)
In
a
sequence
model
inducer
like
Shain et al. (2016), these depth-specific models
are assumed to be independent of each other and
fit with a Gibbs sampler, backward sampling
hidden variable sequences from forward distributions using this compiled transition model M
(Carter and Kohn, 1996), then counting individual
sub-model outcomes from sampled hidden variable
sequences, then resampling each sub-model using
these counts with Dirichlet priors over a, b, and p
models and Beta priors over f and j models, then
re-compiling these resampled models into a new M.
However, note that with K category labels this
model contains DK 2 + 3DK 3 separate parameters
for preterminal categories and top and bottom categories of derivation fragments at every depth level,
each of which can be independently learned by the
Gibbs sampler. Although this allows the hierarchical sequence model to learn grammars that are more
expressive than PCFGs, the search space is several
times larger than the K 3 space of PCFG nonterminal
expansions. The model described in this paper instead induces a PCFG and derives sequence model
distributions from the PCFG, which has fewer parameters, and thus strictly reduces the search space
of the model.
4 The DB-PCFG Model
Unlike Shain et al. (2016), the depth-bounded probabilistic context-free grammar (DB-PCFG) model
described in this paper directly induces a PCFG
and then deterministically derives the parameters
of a probabilistic left-corner parser from this single source. This derivation is based on an existing
derivation of probabilistic left-corner parser models from PCFGs (van Schijndel et al., 2013), which
¯
¯
1..d
d
5
, ψd̄ = ~btd̄+1..D = ⊥, and again,
= bt−1
Here φd̄ = ~b1..
t
d̄ = maxd {bdt−1 , ⊥}.
was developed in a supervised parsing model, here
adapted to run more efficiently within a larger unsupervised grammar induction model.6
A PCFG can be defined in Chomsky normal form
as a matrix G of binary rule probabilities with one
row for each of K parent symbols c and one column for each of K 2 +W combinations of left and
right child symbols a and b, which can be pairs of
nonterminals or observed words from vocabulary W
followed by null symbols ⊥:7
G=
X
P(c → a b | c) δc (δa ⊗ δb )⊤
(7)
a,b,c
A depth-bounded grammar is a set of side- and
depth-specific distributions:
GD = {G s,d | s ∈ {L, R}, d ∈ {1..D}}
(8)
The posterior probability of a depth-bounded
model GD given a corpus (sequence) of words w1..T
is proportional to the product of a likelihood and a
prior:
P(GD | w1..T ) ∝ P(w1..T | GD ) · P(GD )
(9)
The likelihood is defined as a marginal over
bounded PCFG trees τ of the probability of that tree
given the grammar times the product of the probability of the word at each time step or token index t
6
More specifically, the derivation differs from that of
van Schijndel et al. (2013) in that it removes terminal symbols
from conditional dependencies of models over fork and join decisions and top and bottom category labels, substantially reducing the size of the derived model that must be run during induction.
7
This definition assumes a Kronecker delta function δi , defined as a vector with value one at index i and zeros everywhere
else, and a Kronecker product M ⊗ N over matrices M and N,
which tiles copies of N weighted by values in M as follows:
M[1,1] N
M ⊗ N = M[2,1] N
..
.
M[1,2] N
M[2,2] N
..
.
· · ·
· · ·
. .
.
(1’)
The Kronecker product specializes to vectors as single-column
matrices, generating vectors that contain the products of all
combinations of elements in the operand vectors.
given this tree:8
P(w1..T | GD ) =
X
τ
P(τ | GD ) ·
Y
P(wt | τ) (10)
t
The probability of each tree is defined to be the product of the probabilities of each of its branches:9
Y
P(τ | GD ) =
PGD (τη → τη0 τη1 | τη )
(11)
τη ∈τ
The probability P(GD ) is itself an integral
over the product of a deterministic transform φ
from an unbounded grammar to a bounded grammar P(GD | G) = ~GD = φ(G) and a prior over unbounded grammars P(G):
Z
P(GD ) =
P(GD | G) · P(G) · dG
(12)
Distributions P(G) for each nonterminal symbol
(rows) within this unbounded grammar can then be
sampled from a Dirichlet distribution with a symmetric parameter β:
G ∼ Dirichlet(β)
(13)
which then yields a corresponding transformed
sample in P(GD ) for corresponding nonterminals.
Note that this model is different than that of
Shain et al. (2016), who induce a hierarchical HMM
directly.
A depth-specific grammar GD is (deterministically) derived from G via transform φ with probabilities for expansions constrained to and renormalized
over only those outcomes that yield terminals within
a particular depth bound D. This depth-bounded
grammar is then used to derive left-corner expectations (anticipated counts of categories appearing
as left descendants of other categories), and ultimately the parameters of the depth-bounded leftcorner parser defined in Section 3. Counts for G are
then obtained from sampled hidden state sequences,
and rows of G are then directly sampled from the
posterior updated by these counts.
8
This notation assumes the observed data w1..T is a single
long sequence of words, and the hidden variable τ is a single
large but depth-bounded tree structure (e.g. a right-branching
discourse structure). Since the implementation is incremental,
segmentation decisions may indeed be treated as hidden variables in τ, but the experiments described in Section 5 are run on
sentence-segmented input.
9
Here, η is a node address, with left child η0 and right child
η1, or with right child equal to ⊥ if unary.
4.1
Depth-bounded grammar
In order to ensure the bounded version of G is a consistent probability model, it must be renormalized in
transform φ to assign a probability of zero to any
derivation that exceeds its depth bound D. For example, if D = 2, then it is not possible to expand a
left sibling at depth 2 to anything other than a lexical
item, so the probability of any non-lexical expansion
must be removed from the depth-bounded model,
and the probabilities of all remaining outcomes must
be renormalized to a new total without this probability. Following van Schijndel et al. (2013), this can
be done by iteratively defining a side- and depthspecific containment likelihood h(i)
s,d for left- or rightside siblings s ∈ {L, R} at depth d ∈ {1..D} at each iteration i ∈ {1..I},10 as a vector with one row for each
nonterminal or terminal symbol (or null symbol ⊥)
in G, containing the probability of each symbol generating a complete yield within depth d as an s-side
sibling:
=0
(14a)
h(0)
s,d
(i−1)
(i−1)
G (1 ⊗ δ⊥ + hL,d ⊗ hR,d ) if d ≤ D + 1
=
h(i)
L,d
0
if d > D + 1
(14b)
δT
if d = 0
(i)
(i−1)
(i−1)
hR,d =
G (1 ⊗ δ⊥ + hL,d+1 ⊗ hR,d ) if 0 < d ≤ D
0
if d > D
(14c)
where ‘T’ is a top-level category label at depth zero.
A depth-bounded grammar G s,d can then be defined to be the original grammar G reweighted and
renormalized by this containment likelihood:11
GL,d =
GR,d =
Experiments described in this article use I = 20 following
observations of convergence at this point in supervised parsing.
h(I)
L,d
(I)
G diag(1 ⊗ δ⊥ + h(I)
L,d+1 ⊗ hR,d )
h(I)
R,d
(15a)
(15b)
This renormalization ensures the depth-bounded
model is consistent. Moreover, this distinction between a learned unbounded grammar G and a derived bounded grammar G s,d which is used to derive a parsing model may be regarded as an instance
of Chomsky’s (1965) distinction between linguistic
competence and performance.
The side- and depth-specific grammar can then be
used to define expected counts of categories occurring as left descendants (or ‘left corners’) of rightsibling ancestors:
E(1)
d = GR,d (diag(1) ⊗ 1)
(16a)
GL,d
E(i−1)
d
PI
(i)
i=1 Ed
(16b)
E(i)
d
E+d
=
=
(diag(1) ⊗ 1)
(16c)
This left-corner expectation will be used to estimate the marginalized probability over all grammar rule expansions between derivation fragments,
which must traverse an unknown number of left children of some right-sibling ancestor.
4.2
Depth-bounded parsing
Again following van Schijndel et al. (2013), the fork
and join decision, and the preterminal, top and bottom category label sub-models described in Section 3 can now be defined in terms of these sideand depth-specific grammars G s,d and depth-specific
left-corner expectations E+d .
First, probabilities for no-fork and yes-fork outcomes below some bottom sign of category b at
depth d are defined as the normalized probabilities,
respectively, of any lexical expansion of a right sibling b at depth d, and of any lexical expansion following any number of left child expansions from b
11
10
(I)
G diag(1 ⊗ δ⊥ + h(I)
L,d ⊗ hR,d )
where diag(v) is a diagonalization of a vector v:
0
· · ·
v[1]
0
v[2]
diag(v) =
..
. .
.
.
(2’)
4.3
at depth d:
PθF,d (0 | b) =
PθF,d (1 | b) =
δb ⊤ GR,d (1 ⊗ δ⊥ )
(17a)
δb ⊤ (GR,d + E+d GL,d ) (1 ⊗ δ⊥ )
δb ⊤ E+d GL,d (1 ⊗ δ⊥ )
δb ⊤ (GR,d + E+d GL,d ) (1 ⊗ δ⊥ )
(17b)
Gibbs sampling
Grammar induction in this model then follows a forward-filtering backward-sampling algorithm (Carter and Kohn, 1996). This algorithm first
computes a forward distribution vt over hidden
states at each time step t from an initial value ⊥:
The probability of a preterminal p given a bottom category b is simply a normalized left-corner
expected count of p under b:
def δb
PθP,d (p | b) =
⊤
E+d δ p
δb ⊤ E+d 1
(18)
Yes-join and no-join probabilities below bottom
sign b and above top sign a at depth d are then defined similarly to fork probabilities, as the normalized probabilities, respectively, of an expansion to
left child a of a right sibling b at depth d, and of an
expansion to left child a following any number of
left child expansions from b at depth d:
PθJ,d (1 | b a) =
PθJ,d (0 | b a) =
δb ⊤ GR,d (δa ⊗ 1)
(19a)
δb ⊤ (GR,d + E+d GL,d ) (δa ⊗ 1)
δb ⊤ E+d GL,d (δa ⊗ 1)
δb ⊤ (GR,d + E+d GL,d ) (δa ⊗ 1)
(19b)
The distribution over category labels for top
signs a above some top sign of category c and below a bottom sign of category b at depth d is defined
as the normalized distribution over category labels
following a chain of left children expanding from b
which then expands to have a left child of category c:
PθA,d (a | b c) =
δb ⊤ E+d diag(δa ) GL,d (δc ⊗ 1)
δb ⊤ E+d diag(1) GL,d (δc ⊗ 1)
(23a)
⊤
vt = vt−1 M diag(L δwt )
(23b)
The algorithm then samples hidden states backward
from a multinomial distribution given the previously
sampled state qt+1 at time step t+1 (assuming input
parameters to the multinomial function are normalized):
qt ∼ Multinom( diag(vt ) M diag(L δwt+1 ) δqt+1 )
(24)
Grammar rule applications C are then counted
from these sampled sequences:12
⊤ if f , j = 0, 1
¯
¯ )
δbd−1
(δad¯ ⊗ δbd−1
t t
t
t−1
t−1
X
if ft , jt = 0, 0
δad¯ (δad¯ ⊗ δbd̄ )⊤
t
t
t−1
C=
⊤
δbd¯ (δ pt ⊗ δbd¯ )
if ft , jt = 1, 1
t
t
t−1
δad+1
¯
(δ pt ⊗ δbd̄+1 )⊤ if ft , jt = 1, 0
t
t
X
⊤
+
δ pt (δwt ⊗ δ⊥ )
(25)
t
and a new grammar G is sampled from a Dirichlet
distribution with counts C and a symmetric hyperparameter β as parameters:
G ∼ Dirichlet( C + β )
⊤
δa Gs,d (δc ⊗ δb )
δa ⊤ Gs,d (δc ⊗ 1)
⊤
(20)
The distribution over category labels for bottom
signs b below some sign a and sibling of top sign c
is then defined as the normalized distribution over
right children of grammar rules expanding from a
to c followed by b:
PθB,s,d (b | a c) =
v0 ⊤ = δ⊥ ⊤
(21)
Finally, a lexical observation model L is defined
as a matrix of unary rule probabilities with one row
for each combination of store state and preterminal
symbol and one column for each observation symbol:
L = 1 ⊗ G (diag(1) ⊗ δ⊥ )
(22)
(26)
This grammar is then used to define transition and
lexical models M and L as defined in Sections 3
through 4.2 to complete the cycle.
4.4
Model hyper-parameters and priors
There are three hyper-parameters in the model. K is
the number of non-terminal categories in the grammar G, D is the maximum depth, and β is the parameter for the symmetric Dirichlet prior over multinomial distributions in the grammar G.
As seen from the previous subsection, the prior
is over all possible rules in an unbounded PCFG
12
Again, d̄ = maxd {adt−1 , ⊥}.
grammar. Because the number of non-terminal categories of the unbounded PCFG grammar is given as
a hyper-parameter, the number of rules in the grammar is always known. It is possible to use nonparametric priors over the number of non-terminal
categories, however due to the need to dynamically
mitigate the computational complexity of filtering
and sampling using arbitrarily large category sets,
this is left for future work.
5 Evaluation
The DB-PCFG model described in Section 4
is evaluated first on synthetic data to determine
whether it can reliably learn a recursive grammar
from data with a known optimum solution, and
to determine the hyper-parameter value for β
for doing so. Two experiments on natural data
are then carried out. First, the model is run on
natural data from the Adam and Eve parts of the
CHILDES corpus (MacWhinney, 1992) to compare with other grammar induction systems on a
human-like acquisition task. Then data from the
Wall Street Journal section of the Penn Treebank
(Marcus et al., 1993) is used for further comparison
in a domain for which competing systems are optimized. The competing systems include UPPARSE
(Ponvert et al., 2011)13 , CCL (Seginer, 2007a)14 ,
BMMM+DMV with undirected dependency
features (Christodoulopoulos et al., 2012)15 and
UHHMM (Shain et al., 2016).16
For the natural language datasets, the variously
parametrized DB-PCFG systems17 are first validated
on a development set, and the optimal system is then
run until convergence with the chosen hyperparameters on the test set. In development experiments,
the log-likelihood of the dataset plateaus usually after 500 iterations. The system is therefore run at
least 500 iterations in all test set experiments, with
one iteration being a full cycle of Gibbs sampling.
a) X1
X1
b)
b
X2
d)
X1
X2
X1 X2
X1
X2
X1 X2
b
a
b
a
X1 X2
X1 X2
a
c) X2
a
b
a
b
Figure 2:
Synthetic left-branching (a,b) and rightbranching (c,d) datasets.
The system is then checked to see whether the loglikelihood has plateaued, and halted if it has.
The DB-PCFG model assigns trees sampled from
conditional posteriors to all sentences in a dataset
in every iteration as part of the inference. The system is further allowed to run at least 250 iterations
after convergence and proposed parses are chosen
from the iteration with the greatest log-likelihood after convergence. However, once the system reaches
convergence, the evaluation scores of parses from
different iterations post-convergence appear to differ
very little.
5.1
Synthetic data
Following
Liang et al. (2009)
and
Scicluna and de la Higuera (2014),
an
initial
set of experiments on synthetic data are used
to investigate basic properties of the model—in
particular:
1. whether the model is balanced or biased in favor of left- or right-branching solutions,
2. whether the model is able to posit recursive
structure in appropriate places, and
3. what hyper-parameters enable the model to find
optimal modes more quickly.
The risk of bias in branching structure is important because it might unfairly inflate induction results on languages like English, which are heavily
right branching. In order to assess its bias, the model
13
https://github.com/eponvert/upparse
is evaluated on two synthetic datasets, each consist14
https://github.com/DrDub/cclparser
15
ing of 200 sentences. The first dataset is a leftBMMM:https://github.com/christos-c/bmmm
DMV:https://code.google.com/archive/p/pr-toolkit/ branching corpus, which consists of 100 sentences
16
https://github.com/tmills/uhhmm/tree/coling16
of the form a b and 100 sentences of the form a b b
17
The most complex configuration that would run on
, with optimal tree structures as shown in Figure 2
available GPUs was D=2, K=15. Analysis of full WSJ
(a) and (b). The second dataset is a right-branching
(Schuler et al., 2010) shows 47.38% of sentences require depth
2, 38.32% require depth 3 and 6.26% require depth 4.
corpus, which consists of 100 sentences of the form
a)
X1
X1
X2
X1 X2
b
a
c)
X3
X3
X1
c
X1 X2
a
b
d)
X3
X1 X2
c
b
X1
X3
X1 X2
c
a
X1
System
CCL
UPPARSE
UHHMM
BMMM+DMV
DB-PCFG
X3
b
X3
a
b)
X3
b
X3
X1
X2
X1 X2
b
a
b
X1
X2
X1 X2
b
c
b
Figure 3: Synthetic center-embedding structure. Note
that tree structures (b) and (d) have depth 2 because they
have complex sub-trees spanning a b and a b b, respectively, embedded in the center of the yield of their roots.
a b and 100 sentences of the form a a b , with optimal tree structures as shown in Figure 2 (c) and (d).
Results show both structures (and both corresponding grammars) are learnable by the model, and result in approximately the same log likelihood. These
synthetic datasets are also used to tune the β hyperparameter of the model (as defined in Section 4) to
enable it to find optimal modes more quickly. The
resulting β setting of 0.2 is then used in induction on
the CHILDES and Penn Treebank corpora.
After validating that the model is not biased,
the model is also evaluated on a synthetic centerembedding corpus consisting of 50 sentences each
of the form a b c; a b b c; a b a b c; and a b b a b b c,
which has optimal tree structures as shown in Figure 3.18 Note that the (b) and (d) trees have depth 2
because they each have a complex sub-tree spanning
a b and a b b embedded in the center of the yield of
the root. Results show the model is capable of learning depth 2 (recursive) grammars.
Finally, as a gauge of the complexity of this task,
results of the model described in this paper are compared with those of other grammar induction mod18
Hyperparameters
D1K15
D1K30
D1K45
D2K15
X3
X1
a
Recall
71.1
80.7
37.7
83.2
100.0
F1
76.7
85.7
37.7
90.5
100.0
Table 1: The performance scores of unlabeled parse evaluation of different systems on synthetic data.
X3
X1
Precision
83.2
91.4
37.7
99.2
100.0
Here, in order to more closely resemble natural language
input, tokens a, b, and c are randomly chosen uniformly from
{a1 , . . . , a50 }, {b1 , . . . , b50 } and {c1 , . . . , c50 }, respectively.
Precision
57.1
52.8
44.4
44.0
Recall
70.7
65.4
54.9
54.5
F1
63.2
58.5
49.1
48.7
Table 2: PARSEVAL results of different hyperparameter
settings for the DB-PCFG system on the Adam dataset.
Hyperparameter D is the number of possible depths, and
K is the number of non-terminals.
els on the center-embedding dataset. In this experiment, all models are assigned hyper-parameters
matching the optimal solution. The DB-PCFG is
run with K=5 and D=2 and β=0.2 for all priors, the
BMMM+DMV (Christodoulopoulos et al., 2012) is
run with 3 preterminal categories, and the UHHMM
model is run with 2 active states, 4 awaited states
and 3 parts of speech.19 Table 1 shows the PARSEVAL scores for parsed trees using the learned grammar from each unsupervised system. Only the DBPCFG model is able to recognize the correct tree
structures and the correct category labels on this
dataset, showing the task is indeed a robust challenge. This suggests that hyper-parameters optimized on this dataset may be portable to natural data.
5.2
Child-directed speech corpus
After setting the β hyperparameter on synthetic
datasets, the DB-PCFG model is evaluated on
14,251 sentences of transcribed child-directed
speech from the Eve section of the Brown corpus of CHILDES (MacWhinney, 1992). Hyperparameters D and K are set to optimize performance on the Adam section of the Brown Corpus of
CHILDES, which is about twice as long as Eve. Fol19
It is not possible to use just 2 awaited states, which is the
gold setting, since the UHHMM system errors out when the
number of categories is small.
System
CCL
UPPARSE
UHHMM
BMMM+DMV
UHHMM-F
DB-PCFG
Right-branching
Precision
50.5
60.5
55.5
63.5
62.9
64.5
68.7
Recall
53.5
51.9
69.3
63.3
68.4
80.5
85.8
F1
51.9
55.9
61.7
63.4
65.6
71.6∗∗
76.3
Table 3: PARSEVAL scores on Eve dataset for all competing systems. These are unlabeled precision, recall and
F1 scores on constituent trees without punctuation. Both
the right-branching baseline and the best performing system are in bold. (**: p < 0.0001, permutation test)
lowing Seginer (2007a), Ponvert et al. (2011) and
Shain et al. (2016), these experiments leave all
punctuation in the input for learning, then remove
it in all evaluations on development and test data.
Model performance is evaluated against Penn
Treebank style annotations of both Adam and Eve
corpora (Pearl and Sprouse, 2013). Table 2 shows
the PARSEVAL scores of the DB-PCFG system
with different hyperparameters on the Adam corpus
for development.The simplest configuration, D1K15
(depth 1 only with 15 non-terminal categories), obtains the best score, so this setting is applied to
the test corpus, Eve. Results of the D=1, K=15
DB-PCFG model on Eve are then compared against
those of other grammar induction systems which use
only raw text as input on the same corpus. Following Shain et al. (2016) the BMMM+DMV system is
run for 10 iterations with 45 categories and its output is converted from dependency graphs to constituent trees (Collins et al., 1999). The UHHMM
system is run on the Eve corpus using settings
in Shain et al. (2016), which also includes a postprocess option to flatten trees (reported here as
UHHMM-F).
Table 3 shows the PARSEVAL scores for all the
competing systems on the Eve dataset. The rightbranching baseline is still the most accurate in terms
of PARSEVAL scores, presumably because of the
highly right-branching structure of child-directed
speech in English. The DB-PCFG system with only
one memory depth and 15 non-terminal categories
achieves the best performance in terms of F1 score
and recall among all the competing systems, signif-
icantly outperforming other systems (p < 0.0001,
permutation test).20
The Eve corpus has about 5,000 sentences with
more than one depth level, therefore one might expect a depth-two model to perform better than a
depth-one model, but this is not true if only PARSEVAL scores are considered. This issue will be revisited in the following section with the noun phrase
discovery task.
5.3
NP discovery on child-directed speech
When humans acquire grammar, they do not only
learn tree structures, they also learn category types:
noun phrases, verb phrases, prepositional phrases,
and where each type can and cannot occur.
Some of these category types — in particular,
noun phrases — are fairly universal across languages, and may be useful in downstream tasks such
as (unsupervised) named entity recognition. The
DB-PCFG and other models that can be made to
produce category types are therefore evaluated on a
noun phrase discovery task.
Two metrics are used for this evaluation. First,
the evaluation counts all constituents proposed by
the candidate systems, and calculates recall against
the gold annotation of noun phrases. This metric is
not affected by which branching paradigm the system is using and reveals more about the systems’
performances. This metric differs from that used by
Ponvert et al. (2011) in that this metric takes NPs at
all levels in gold annotation into account, not just
base NPs.21
The second metric, for systems that produce category labels, calculates F1 scores of induced categories that can be mapped to noun phrases. The
first 4,000 sentences are used as the development
set for learning mappings from induced category labels to phrase types. The evaluation calculates precision, recall and F1 of all spans of proposed categories against the gold annotations of noun phrases
20
Resulting
scores
are
better
when
applying
Shain et al. (2016) flattening to output binary-branching
trees. For the D=1, K=15 model, precision and F1 can be
raised to 70.31% and 74.33%. However, since the flattening is
a heuristic which may not apply in all cases, these scores are
not considered to be comparable results.
21
Ponvert et al. (2011) define base NPs as NPs with no NP
descendants, a restriction motivated by their particular task
(chunking).
System
CCL
UPPARSE
UHHMM
BMMM+DMV
DB-PCFG (D1K15)
DB-PCFG (D1K30)
DB-PCFG (D1K45)
DB-PCFG (D2K15)
Right-branching
NP Recall
35.5
69.1
61.4
71.3
75.7
78.6
76.9
85.1
64.2
NP agg F1
27.4
61.2
28.7
60.7
64.0
65.9
-
Table 4: Performances of different systems for noun
phrase recall and aggregated F1 scores on the Eve dataset.
in the development set, and aggregates the categories
ranked by their precision scores so that the F1 score
of the aggregated category is the highest on the development set. The evaluation then calculates the F1
score of this aggregated category on the remainder
of the dataset, excluding this development set.
The UHHMM system is the only competing system that is natively able to produce labels for proposed constituents. BMMM+DMV does not produce constituents with labels by default, but can
be evaluated using this metric by converting dependency graphs into constituent trees, then labeling
each constituent with the part-of-speech tag of the
head. For CCL and UPPARSE, the NP agg F1 scores
are not reported because they do not produce labeled
constituents.
Table 4 shows the scores for all systems on the
Eve dataset and four runs of the DB-PCFG system
on these two evaluation metrics. Surprisingly the
D=2, K=15 model which has the lowest PARSEVAL scores is most accurate at discovering noun
phrases. It has the highest scores on both evaluation metrics. The best model in terms of PARSEVAL scores, the D=1, K=15 DB-PCFG model, performs poorly among the DB-PCFG models, despite
the fact that its NP Recall is higher than the competing systems. The low score of NP agg F1 of
DB-PCFG at D1K15 shows a diffusion of induced
syntactic categories when the model is trying to find
a balance among labeling and branching decisions.
The UPPARSE system, which is proposed as a base
NP chunker, is relatively poor at NP recall by this
definition.
The right-branching baseline does not perform
well in terms of NP recall. This is mainly because
noun phrases are often left children of some other
constituent and the right branching model is unable to incorporate them into the syntactic structures
of whole sentences. Therefore although the rightbranching model is the best model in terms of PARSEVAL scores, it is not helpful in terms of finding
noun phrases.
5.4
Penn Treebank
To further facilitate direct comparison to previous
work, we run experiments on sentences from the
Penn Treebank (Marcus et al., 1993). The first experiment uses the sentences from Wall Street Journal part of the Penn Treebank with at most 20 words
(WSJ20). The first half of the WSJ20 dataset is used
as a development set (WSJ20dev) and the second
half is used as a test set (WSJ20test). We also extract
sentences in WSJ20test with at most 10 words from
the proposed parses from all systems and report results on them (WSJ10test). WSJ20dev is used for
finding the optimal hyperparameters for both DBPCFG and BMMM-DMV systems.22
Table 5 shows the PARSEVAL scores of all systems. The right-branching baseline is relatively
weak on these two datasets, mainly because formal
writing is more complex and uses more non-rightbranching structures (e.g., subjects with modifiers
or parentheticals) than child-directed speech. For
WSJ10test, both the DB-PCFG system and CCL are
able to outperform the right branching baseline. The
F1 difference between the best-performing previouswork system, CCL, and DB-PCFG is highly significant. For WSJ20test, again both CCL and DBPCFG are above the right-branching baseline. The
difference between the F scores of CCL and DBPCFG is very small compared to WSJ10, however
it is also significant.
It is possible that the DB-PCFG is being penalized for inducing fully binarized parse trees. The
22
Although UHHMM also needs tuning, in practice we find
that this system is too inefficient to be tuned on a development
set, and it requires too many resources when the hyperparameters become larger than used in previous work. We believe that
further increasing the hyperparameters of UHHMM may lead to
performance increase, but the released version is not scalable to
larger values of these settings. We also do not report UHHMM
on WSJ20test for the same scalabilty reason. The results of
WSJ10test of UHHMM is induced with all WSJ10 sentences.
System
CCL
UPPARSE
UHHMM
BMMM+DMV(K10)
UHHMM-F
DB-PCFG (D2K15)
Right-branching
WSJ10test
Precision Recall
F1
63.4
71.9
67.4
54.7
48.3
51.3
49.1
63.4
55.3
36.2
40.6
38.2
57.1
54.4
55.7
64.5
82.6
72.4∗∗
55.1
70.5
61.8
WSJ20test
Precision Recall
F1
60.1
61.7
60.9∗∗
47.8
40.5
43.9
25.3
29.0
27.0
53.0
70.5
60.5
41.5
55.3
47.4
Table 5: PARSEVAL scores for all competing systems on WSJ10 and WSJ20 test sets. These are unlabeled precision,
recall and F1 scores on constituent trees without punctuation (**: p <0.0001, permutation test).
System
CCL
UPPARSE
DB-PCFG (D2K15)
Right-branching
WSJ10
Precision Recall
75.3
76.1
74.6
66.7
65.5
83.6
55.2
70.0
F1
75.7
70.5
73.4
61.7
WSJ40
Precision Recall
58.7
55.9
60.0
49.4
47.0
63.6
35.4
47.4
F1
57.2
54.2
54.1
40.5
Table 6: Published PARSEVAL results for competing systems. Please see text for details as the systems are trained
and evaluated differently.
accuracy of the DB-PCFG model is dominated by
recall rather than precision, whereas CCL and other
systems are more balanced. This is an important distinction if it is assumed that phrase structure is binary (Kayne, 1981; Larson, 1988), in which
case precision merely scores non-linguistic decisions about whether to suppress annotation of nonmaximal projections. However, since other systems
are not optimized for recall, it would not be fair to
use only recall as a comparison metric in this study.
Finally, Table 6 shows the published results of different systems on WSJ. The CCL results come from
Seginer (2007b), where the CCL system is trained
with all sentences from WSJ, and evaluated on sentences with 40 words or fewer from WSJ (WSJ40)
and WSJ10. The UPPARSE results come from
Ponvert et al. (2011), where the UPPARSE system
is trained using 00-21 sections of WSJ, and evaluated on section 23 and the WSJ10 subset of section
23. The DB-PCFG system uses hyperparameters optimized on the WSJ20dev set, and is evaluated on
WSJ40 and WSJ10, both excluding WSJ20dev. The
results are not directly comparable, but the results
from the DB-PCFG system is competitive with the
other systems, and numerically have the best recall
scores.
6 Conclusion
This paper describes a Bayesian Dirichlet model
of depth-bounded PCFG induction. Unlike earlier
work this model implements depth bounds directly
on PCFGs by derivation, reducing the search space
of possible trees for input words without exploding
the search space of parameters with multiple sideand depth-specific copies of each rule. Results for
this model on grammar acquisition from transcribed
child-directed speech and newswire text exceed or
are competitive with those of other models when
evaluated on parse accuracy. Moreover, grammars
acquired from this model demonstrate a consistent
use of category labels, something which has not
been demonstrated by other acquisition models.
In addition to its practical merits, this model may
offer some theoretical insight for linguists and other
cognitive scientists. First, the model does not assume any universals except independently motivated
limits on working memory, which may help address
the question of whether universals are indeed necessary for grammar induction. Second, the distinction
this model draws between its learned unbounded
grammar G and its derived bounded grammar GD
seems to align with Chomsky’s (1965) distinction
between competence and performance, and has the
potential to offer some formal guidance to linguistic
inquiry about both kinds of models.
Acknowledgments
The authors would like to thank Cory Shain and
William Bryce and the anonymous reviewers for
their valuable input. Computations for this project
were partly run on the Ohio Supercomputer Center (1987). This research was funded by the Defense Advanced Research Projects Agency award
HR0011-15-2-0022. The content of the information
does not necessarily reflect the position or the policy of the Government, and no official endorsement
should be inferred.
References
Steven P. Abney and Mark Johnson. 1991. Memory requirements and local ambiguities of parsing strategies.
J. Psycholinguistic Research, 20(3):233–250.
Taylor Berg-Kirkpatrick, Alexandre Bouchard-Côté,
John DeNero, and Dan Klein. 2010. Painless unsupervised learning with features. In Human Language
Technologies: The 2010 Annual Conference of the
North American Chapter of the Association for Computational Linguistics, number June, pages 582–590.
Yonatan Bisk and Julia Hockenmaier. 2013. An HDP
Model for Inducing Combinatory Categorial Grammars. Transactions Of The Association For Computational Linguistics.
Rens Bod. 2006. Unsupervised parsing with U-DOP.
Proceedings of the Conference on Computational Natural Language Learning, pages 85–92.
Glenn Carroll and Eugene Charniak. 1992. Two Experiments on Learning Probabilistic Dependency Grammars from Corpora. Working Notes of the Workshop
on Statistically-Based {NLP} Techniques, (March):1–
13.
C K Carter and R Kohn. 1996. Markov Chain Monte
Carlo in Conditionally Gaussian State Space Models.
Biometrika, 83(3):589–601.
Ohio Supercomputer Center. 1987. Ohio supercomputer
center. http://osc.edu/ark:/19495/f5s1ph73.
Noam Chomsky and George A. Miller. 1963. Introduction to the formal analysis of natural languages. In
Handbook of Mathematical Psychology, pages 269–
321. Wiley, New York, NY.
Noam Chomsky. 1965. Aspects of the Theory of Syntax.
MIT Press, Cambridge, MA.
Christos Christodoulopoulos, Sharon Goldwater, and
Mark Steedman. 2012. Turning the pipeline into a
loop: iterated unsupervised dependency parsing and
PoS induction. Proceedings of the Annual Conference of the North American Chapter of the Association for Computational Linguistics: Human Language
Technologies-HLT; Workshop on the Induction of Linguistic Structure, pages 96–99.
Alexander Clark. 2001. Unsupervised induction of
stochastic context-free grammars using distributional
clustering. Proceedings of the 2001 workshop on
Computational Natural Language Learning, 7:1–8.
Michael Collins, Lance Ramshaw, Jan Hajič, and
Christoph Tillmann. 1999. A Statistical Parser for
Czech. Proceedings of the 37th Annual Meeting of the
Association for Computational Linguistics (ACL ’99),
(405):505–512.
Jennifer Gillenwater, Kuzman Ganchev, João Graça, Fernando Pereira, and Ben Taskar. 2011. Posterior Sparsity in Unsupervised Dependency Parsing. Journal of
Machine Learning Research, 12:455–490.
Wenjuan Han, Yong Jiang, and Kewei Tu. 2017. Dependency Grammar Induction with Neural Lexicalization
and Big Training Data. EMNLP, pages 1684–1689.
Zellig Harris. 1954. Distributional structure. In Jerry A
Fodor and Jerrold J Katz, editors, The Structure of
Language: readings in the philosophy of language,
volume 10, pages 33–49. Prentice-Hall.
William P. Headden, III, Mark Johnson, and David McClosky. 2009. Improving unsupervised dependency
parsing with richer contexts and smoothing. Proceedings of Human Language Technologies: The 2009 Annual Conference of the North American Chapter of the
Association for Computational Linguistics (NAACL
’09), (June):101–109.
Yong Jiang, Wenjuan Han, and Kewei Tu. 2016. Unsupervised Neural Dependency Parsing . Proceedings of
the 2016 Conference on Empirical Methods in Natural
Language Processing (EMNLP-16), (61503248):763–
771.
Mark Johnson, Thomas L Griffiths, and Sharon Goldwater. 2007. Adaptor grammars: A framework for specifying compositional nonparametric Bayesian models.
In NIPS, volume 19, page 641.
Philip N. Johnson-Laird. 1983. Mental models: Towards a cognitive science of language, inference, and
consciousness. Harvard University Press, Cambridge,
MA, USA.
Richard Kayne. 1981. Unambiguous Paths. In R May
and J Koster, editors, Levels of Syntactic Representation, pages 143–183. Foris Publishers.
Dan Klein and Christopher D Manning. 2002. A generative constituent-context model fo rimproved grammar
induction. Proceedings of the Annual Meeting of the
Association for Computational Linguistics.
Dan Klein and Christopher Manning. 2004. Corpusbased induction of syntactic structure: models of dependency and constituency. Proceedings of the Annual Meeting on Association for Computational Linguistics, 1:478–485.
Tom Kwiatkowski, Luke S Zettlemoyer, Sharon Goldwater, and Mark Steedman. 2010. Inducing Probabilistic CCG Grammars from Logical Form with HigherOrder Unification. In EMNLP, pages 1223–1233.
Tom Kwiatkowski, Sharon Goldwater, Luke S. Zettlemoyer, and Mark Steedman. 2012. A probabilistic model of syntactic and semantic acquisition from
child-directed utterances and their meanings. In Proceedings of EACL 2012.
Richard K. Larson. 1988. On the double object construction.
Percy Liang, Michael I Jordan, and Dan Klein. 2009.
Probabilistic Grammars and Hierarchical Dirichlet
Processes. The Handbook of Applied Bayesian Analysis.
B. MacWhinney. 1992. The CHILDES project: tools for
analyzing talk, volume 8. Lawrence Elrbaum Associates, Mahwah, NJ, third edition.
Mitchell P Marcus, Beatrice Santorini, and Mary Ann
Marcinkiewicz. 1993. Building a large annotated corpus of English: the Penn Treebank. Computational
Linguistics, 19(2):313–330.
George A. Miller and Stephen Isard. 1964. Free recall
of self-embedded english sentences. Information and
Control, 7:292–303.
Letitia R. Naigles. 1990. Children use syntax to learn
verb meanings. The Journal of Child Language,
17:357–374.
Hiroshi Noji and Mark Johnson. 2016. Using Leftcorner Parsing to Encode Universal Structural Constraints in Grammar Induction. Proceedings of the
Conference on Empirical Methods in Natural Language Processing, pages 33–43.
John K Pate and Mark Johnson. 2016. Grammar induction from ( lots of ) words alone. In Proceedings
of nternational Conference on Computational Linguistics, pages 23–32.
Lisa Pearl and Jon Sprouse. 2013. Syntactic Islands
and Learning Biases: Combining Experimental Syntax
and Computational Modeling to Investigate the Language Acquisition Problem. Language Acquisition,
20(1):23–68, 1.
Elias Ponvert, Jason Baldridge, and Katrin Erk. 2011.
Simple Unsupervised Grammar Induction from Raw
Text with Cascaded Finite State Models. Proceedings
of the Annual Meeting of the Association for Computational Linguistics, (1999):1077–1086.
Philip Resnik. 1992. Left-corner parsing and psychological plausibility. In Proceedings of COLING, pages
191–197, Nantes, France.
Stanley J. Rosenkrantz and Philip M. Lewis, II. 1970.
Deterministic left corner parser. In IEEE Conference
Record of the 11th Annual Symposium on Switching
and Automata, pages 139–152.
William Schuler, Samir AbdelRahman, Tim Miller, and
Lane Schwartz. 2010. Broad-Coverage Parsing Using Human-Like Memory Constraints. Computational
Linguistics, 36(1):1–30.
James Scicluna and Colin de la Higuera. 2014. PCFG Induction for Unsupervised Parsing and Language Modelling. In Proceedings of the Conference on Empirical Methods in Natural Language Processing, pages
1353–1362.
Yoav Seginer. 2007a. Fast Unsupervised Incremental
Parsing. In Proceedings of the Annual Meeting of
the Association of Computational Linguistics, number
June, pages 384–391.
Yoav Seginer. 2007b. Learning Syntactic Structure.
Ph.D. thesis.
Cory Shain, William Bryce, Lifeng Jin, Victoria Krakovna, Finale Doshi-Velez, Timothy Miller,
William Schuler, and Lane Schwartz. 2016. MemoryBounded Left-Corner Unsupervised Grammar Induction on Child-Directed Input. Proceedings of the International Conference on Computational Linguistics:
Technical Papers, pages 964–975.
Mark Steedman. 2000. The syntactic process. MIT
Press/Bradford Books, Cambridge, MA.
Marten van Schijndel, Andy Exley, and William Schuler.
2013. A Model of Language Processing as Hierarchic
Sequential Prediction. Topics in Cognitive Science,
5(3):522–540.
Luke Zettlemoyer and Michael Collins. 2005. Learning
to map sentences to logical form: Structured classification with probabilistic categorial grammars. In Proceedings of the Proceedings of the Twenty-First Conference Annual Conference on Uncertainty in Artificial
Intelligence (UAI-05), pages 658–666, Arlington, Virginia. AUAI Press.
Willem Zuidema. 2003. How the Poverty of the Stimulus Solves the Poverty of the Stimulus. Advances in
Neural Information Processing Systems 15, 15:51.
| 2cs.AI
|
Symmetric Randomized Dependent Rounding ∗
David G. Harris†
Thomas Pensyl‡
Aravind Srinivasan§
Khoa Trinh¶
arXiv:1709.06995v1 [cs.DS] 20 Sep 2017
Abstract
Dependent rounding is a popular technique in designing approximation algorithms. It allows us
to randomly round a fractional solution x to an integral vector X such
P that E[X]
P = x, all Xi ’s are
(“cylindrically”) negatively correlated, and the cardinality constraint i Xi = i xi holds with probability
one. P
One can also essentially replace the cardinality constraint by a general linear constraint
P
i ai Xi =
i ai xi if the ai ’s are non-negative. However, in certain problems, the coefficients ai may
have mixed signs; furthermore, some applications require correlation inequalities that are much more
general than negative correlation. In this paper, we propose a generalized dependent rounding technique,
called symmetric randomized dependent rounding, which can round x to an almost-integer vector so that
any given linear constraint is satisfied with probability one and such that all the variables are nearly independent. We give two illustrative examples of this technique: a randomized algorithm for the knapsack
center problem with new average approximation guarantees and an improved bi-criteria approximation
ratio for the knapsack median problem.
∗
Research supported in part by NSF Awards CNS-1010789 and CCF-1422569, and by research awards from Adobe, Inc.
Department of Computer Science, University of Maryland, College Park, MD 20742.
Email:
[email protected]
‡
Department of Computer Science, University of Maryland, College Park, MD 20742. Email: [email protected]
§
Department of Computer Science and Institute for Advanced Computer Studies, University of Maryland, College Park, MD
20742. Email: [email protected]
¶
Department of Computer Science, University of Maryland, College Park, MD 20742. Email: [email protected]
†
1
Contents
1
Introduction
2
Our results
2.1 The main rounding algorithm . . . . .
2.2 Comparisons to related work . . . . .
2.3 Applications . . . . . . . . . . . . . .
2.3.1 The knapsack center problem
2.3.2 The knapsack median problem
3
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
5
5
6
7
7
8
3
Symmetric Randomized Dependent Rounding (SRDR)
8
3.1 Upper bound on near-independence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
3.2 Lower bound on near-independence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
4
The knapsack center problem
11
4.1 An LP relaxation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
4.2 Proof of Theorem 2.3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
4.3 Proof of Theorem 2.4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
5
An improved bi-factor approximation algorithm for the knapsack median problem
20
5.1 Pruning “big” facilities and Computing a bi-point solution . . . . . . . . . . . . . . . . . . 21
5.2 Corner cases: a ≤ 1/5 or a ≥ 4/5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
5.3 Main case: a ∈ [1/5, 4/5] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
6
Acknowledgment
26
A Negative correlation in symmetric randomized dependent rounding
27
A.1 Upper bound on near-independence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
A.2 Lower bound on near-independence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
B Omitted proofs for Section 4
33
2
1
Introduction
Dependent rounding is a popular technique in designing approximation algorithms. It allows us to randomly
round a fractional solution x to an integral vector X
Psuch thatPE[X] = x, all Xi ’s are (“cylindrically”)
negatively correlated, and the cardinality constraint i Xi = i xi holds with probability one. Over the
last two decades, increasingly-sophisticated dependent-rounding techniques have been used to obtain good
solutions to optimization problems on matroid- and matroid-intersection- polytopes, subject in some cases
to additional constraints; see, e.g., [1, 4, 7, 9, 16]. For example, in the context of facility-location problems,
Charikar and Li [5] applied dependent rounding in a clever way to obtain a 3.25-approximation for the
classic k-median problem. More recently, Byrka et. al. [3] showed the “near-independence” property for
small subsets of variables when running dependent rounding on a random permutation of the input vector.
Dependent rounding is often used to round a fractional vector subject to (almost) satisfying a single linear
constraint with non-negative coefficients [1, 16].
P
P
It is not difficult to replace the cardinality constraint by a general linear constraint i ai Xi = i ai xi if
the ai ’s are non-negative. Unfortunately, the current approaches do not work when the coefficients ai ’s have
mixed signs. Furthermore, some applications require correlation inequalities that are much more general
than negative correlation. In fact, our problems considered in this paper require a new dependent-rounding
method which can preserve a set of hard “clustering contraints” and a set of soft knapsack constraints
while guaranteeing near-negative-correlation among any subset of the variables. We guarantee this by our
Symmetric Randomized Dependent Rounding (SRDR) technique, which utilizes three major modifications
to [1, 16]: (i) choosing a pair of indices randomly to co-round, in order to mitigate the inevitable positive
correlation that arises since the ai ’s have mixed signs, (ii) updating the variables “symmetrically” and in a
slowed-down manner, and (iii) stopping the process “early” when there are only O(1/) fractional values
left. (Note that in many applications, it suffices to obtain an “almost” integral solution. For example, in one
of our results, the left-over fractional variables belong to disjoint clusters and all such fractional clusters can
be rounded independently. Another notable example is the approach used in [3] for the k-median problem:
the remaining fractional variables result in some O(1) extra open facilities and a special “postprocessing”
step by Li and Svensson [14] can be used to correct the solution.)
In a different context, Chekuri et. al. [6, 7] and Peres et. al. [15] gave powerful rounding schemes
which can round a point inside a matroid polytope (with the rounded coordinates finally being negatively
correlated) or a point inside a matroid-intersection polytope (with good concentration bounds for any linear
function of the rounded coordinates); however, their techniques are not applicable here as we also have to
deal with clustering-type constraints arising naturally in facility-location problems.
In this paper, we present two applications of SRDR: a randomized algorithm for the Multi Knapsack
Center (MKC) problem with new average approximation guarantees and an improved bi-criteria approximation ratio for the Knapsack Median (KM) problem. Furthermore, as illustrated by a general example at
the end of this section, SRDR can be a valuable tool when combined with other rounding techniques.
The (Multi) Knapsack Center problem: An instance I of the MKC problem consists of a set V of n
vertices, a symmetric distance metric d on V , and an m × n weight matrix M , which corresponds to m
non-negative weight functions (scaled so that the corresponding budgets are all equal to one). Our goal is to
choose a set S ⊆ V of facilities (“centers”) so that (1) all m knapsack constraints are satisfied — that is,
X
Mki ≤ 1
i∈S
for all k = 1, . . . , m and (2) the objective function R := maxi∈V Ti , where Ti := minj∈S d(i, j), is
minimized. We shall refer to Ti as the connection cost of vertex i.
3
The Knapsack Center (KC) problem (in the case m = 1) was first studied by Hochbaum & Shmoys
in [10], under the name “weighted k-center”. They gave a 3-approximation algorithm and proved that this
is best possible unless P = NP; see also [12]. More recently, Chen et. al. [8] considered the case m > 1.
They showed that this problem is not approximable to within any constant factor, and gave a pseudo 3approximation algorithm which may violate all but one knapsack constraint by a factor of (1 + ).
Our result here is, for any fixed > 0, a randomized polynomial-time algorithm with small budget
violation that guarantees for each client i that Ti ≤ 3R with probability one, and E[Ti ] ≤ (1 + 2/e + )R.
We do so by applying a type of dependent rounding in the context of a partition matroid M. More formally,
suppose we are given a fractional solution y which is in the matroid polytope of M and satisfies m additional
packing constraints, we round y to a random Y such that: (i) Y is a base of M, (ii) Y satisfies
the packing
P
constraints with at most (1 + ) multiplicative violation, and (iii) for any set S with i∈S yi ≥ 1, we
have Pr[∧i∈S (Yi = 0)] ≤ 1/e + . This dependent rounding scheme has two major advantages over
an independent rounding. First, it achieves a run-time exponent of roughly 1/, as compared to 1/2 for
independent rounding. Secondly, we achieve results for the important case m = 1 which satisfy the packing
constraint exactly (this is discussed in Section 4.2); this is not possible by independent rounding. (The
powerful framework of [7] cannot be applied as it could output an “almost-base” of M.) We are thus able
to obtain a finer version of the optimality result of [10]: we ensure that every client gets service that is at
most 3-approximate with probability 1, but which is also much better on average. This can be quite helpful
in flexible facility location, where the problem is solved repeatedly (such as a streaming-service provider
periodically reshuffling its service locations using an algorithm such as ours), thus guaranteeing improved
average service for every client even after relatively-few repeats.
The Knapsack Median problem: The Knapsack Median (KM) problem is quite similar to the Knapsack
Center problem, except that we aim to minimize the total connection cost. More formally, an instance I of
this problem consists of a set F of facilities, a set C of clients, a symmetric distance metric d on F ∪ C, and
a weight function w on F. Again, we P
assume that all the weights are scaled so that our P
budget B = 1. We
want to find a set S ⊆ F such that (a) j∈C min{d(i, j) : i ∈ S} is minimized and (b) i∈S wi ≤ 1. This
problem was first studied by Krishnaswamy et. al. [13]. The current best-known approximation for KM is
17.46 due to Byrka et. al. [2]. The special case of KM when all facilities in F have unit weight, known as
the classic k-median problem, can be approximated to within a factor of (2.675 + ) [3]. In [2], the authors
show that one can achieve a bi-criteria approximation ratio of 3.05 if allowed to slightly violate the knapsack
constraint by a factor of (1+). Here we show that SRDR can be applied in √
one crucial step of Li-Svensson’s
algorithm for the k-median problem to yield an improved bi-criteria (1 + 3 + )-approximation ratio for
KM.
More-general applicability: The above are two specific applications of SRDR; however, we believe SRDR
is of much more general applicability when combined with other (rounding) ideas in problem-specific settings. As a stand-alone example, let us consider the scenario of Section 4.2 which can be reduced to the
following rounding problem in clustering. (The notation here is slightly different.) Given y1 , . . . yn ∈ [0, 1],
non-negative weights w1 , w2 , . . . , wn , and disjoint sets F1 , F2 , . . . , Fm ⊆ [n], satisfying1
• y(Fj ) = 1 for j = 1, . . . , m,
P
•
i wi yi = 1
That is, we have a fractional solution to a partition-matroid constraint that also satisfies a weight (knapsack)
constraint. It is easy to round this such that, say, the final weight is at most 1. However, suppose we want
a significant extension of this, where, informally, any set S with y(S) = O(1) has the probability that all
1
As is standard, we let y(S) denote
P
i∈S
yi .
4
its entries get rounded down bounded essentially by the case where the yi ’s are rounded independently (see
(1)). Given constants ∈ (0, 1) and a > 0, we aim for a constant t = t(, a) such that the following is true.
We want to (efficient) round y to an almost-{0, 1} vector Y – one with at most a constant t many entries still
lying in (0, 1) – such that
• Y (Fj ) = 1 for j = 1, . . . , m with probability one,
P
•
i wi Yi = 1 with probability one,
• For all S ⊆ [n] such that y(S) ≤ a, we would like to upper-bound Pr[∧i∈S (Yi = 0)] by what we
would get if the Yi ’s were independent; however, we cannot reason directly about Pr[∧
Qi∈S (Yi = 0)]
since a few (at most t many) entries in S could be yet-unrounded. Thus we bound E[ i∈S (1 − Yi )]
(which is a surrogate for the value “Pr[∧i∈S (Yi = 0)]” we are aiming for – note that these two are
equal in case Y was a fully-rounded {0, 1} vector):
!1−/a
E[
Y
(1 − Yi )] ≤
i∈S
Y
(1 − yi )
≤ e−y(S)+ .
(1)
i∈S
Note that this is a fairly general problem; we are not aware of any other technique that can achieve these
results. It is thus our belief that in combination with other (rounding) techniques, SRDR can be a useful tool
in general.
Organization: We formally describe our results in Section 2. In Section 3, we develop SRDR and prove
its guarantees. The new results for the Knapsack Center problem and the Knapsack Median problem are
discussed in Sections 4 and 5, respectively.
2
Our results
Our main contribution is the Symmetric Randomized Dependent Rounding (SRDR) method. We first discuss
its advantages over previous dependent rounding techniques in detail. Then we apply SRDR to obtain new
results for several facility-location problems.
2.1
The main rounding algorithm
The main results of our rounding algorithm are summarized in Theorem 2.1. We shall apply the algorithm
in Theorem 2.1 as a black-box in the applications below.
Theorem 2.1. Given vectors x ∈ [0, 1]n , a = (a1 , . . . , an ) ∈ Rn , and t ∈ N, there exists a randomized
algorithm which can round x in expected O(n2 ) time and return a vector X ∈ [0, 1]n with
Pat most t
fractional
values.
Both
the
weighted
sum
and
all
the
marginal
probabilities
are
preserved:
i ai Xi =
P
a
x
with
probability
one,
and
E[X
]
=
x
for
all
i
∈
[n].
Let
S,
T
be
disjoint
subsets
of
[n].
Then:
i
i
i i i
1. We have the upper correlation bound:
p
Y
p
Y Y
Y
Y
Y
E Xi
(1 − Xj ) ≤ E
Xi
(1 − Xj ) ≤ xi
(1 − xj ) ,
i∈S
j∈T
i∈S
j∈T
i∈S
(2)
j∈T
where p = 1 − 1/(t + 1). Furthermore, if ai ≥ 0 and aj ≤ 0 for all i ∈ S, j ∈ T , or if ai ≤ 0 and
aj ≥ 0 for all i ∈ S, j ∈ T , then inequality (2) holds for p = 1.
5
2. Let S 0 = {i ∈ S : ai > 0} ∪ {i ∈ T : ai < 0} and T 0 = {i ∈ S : ai < 0} ∪ {i ∈ T : ai > 0}. If
|S 0 |, |T 0 | ≤ t, then we have the lower correlation bound:
p
Y
p
Y
Y
Y
Y Y
E Xi
(1 − Xj ) ≥ E
Xi
(1 − Xj ) ≥ xi
(1 − xj ) ,
(3)
i∈S
j∈T
where p = 1 + max
p=1+
i∈S
n
|T 0 |−1
|S 0 |−1
t+1−|S 0 | , t+1−|T 0 |
j∈T
i∈S
j∈T
o
. In particular, if |S ∪ T | ≤ t, inequality (3) holds for
|S∪T |−1
t+1−|S∪T | .
3. Note from (2) that cylinder negative correlation holds in particular if all the ai have the same sign:
"
#
"
#
Y
Y
Y
Y
E
xi , and E
Xi ≤
(1 − Xi ) ≤
(1 − xi ).
(4)
i∈S
i∈S
2.2
i∈S
i∈S
Comparisons to related work
The near-independence property of dependent rounding was studied by Byrka et. al. [3]. Their idea is to
randomly permute the vector x before applying the dependent rounding of [16]. Then they argue that any
two variables are “far” from each other and unlikely to be rounded together in a single round, which implies
that any small groups of the variables are nearly independent. We summarize their results in Theorem 2.2
for comparison.
Theorem 2.2 ( [3]). Given vectors x ∈ [0, 1]n and a = (a1 , . . . , an ) ∈ Rn+ such that mini {xi , 1 − xi } ≥ α
i ai
and max
mini ai ≤ 1 + α for some constant α ≥ 0, there exists a randomized algorithm which can round x in
O(n) time and return a vector X ∈ [0, 1]nP
with at mostPone fractional value. Both the weighted sum and all
the marginal probabilities are preserved: i ai Xi = i ai xi with probability one, and E[Xi ] = xi for all
i ∈ [n]. Let S, T be disjoint subsets of [n]. Then:
"
#
t−1
Y
Y
8t(t − 1)
8t
1−
λ
≤
E
λ,
X
(1
−
X
)
≤
1
+
i
i
3nα2
3nα2
i∈S
where t = |S ∪ T | and λ =
Q
i∈S
xi
Q
j∈T (1
i∈T
− xj ).
Our results here are improvements in several ways. First, our upper-bound (2) and lower-bound (3)
in Theorem 2.1 only depend on the remaining number of fractional variables (i.e., are independent of the
number of terms n and of the ratio amax /amin ). Second, the “target” probabilities xi need not be bounded
away from 0 or 1. Third, Theorem 2.1 does not require the weights to be non-negative as in [3] and the
gap between amin and amax can be arbitrarily large. Finally, we show in Section 4.3 that our technique can
be generalized to work with multiple linear constraints while slightly violating these soft constraints and
still guaranteeing the “near-negative-correlation” property. We thus expect our rounding theorem to be of
independent interest.
Recall that each iteration of the standard dependent-rounding technique of [16] co-rounds two variables
x1 , x2 such that the sum a1 x1 + a2 x2 is preserved and the expected values of x1 , x2 do not change. If
a1 a2 > 0, then an increase in x1 will lead to a decrease in x2 for the sum to remain the same, and vice
versa. This explains why we obtain negative correlation for all techniques in [3, 6, 9, 16]. If the weights are
arbitrary, then it may be that a1 a2 < 0 and so we may have positive correlation between the variables x1 , x2 .
Our SRDR technique employs the following ideas to reduce this positive correlation. First, we randomly
6
pick a pair (x1 , x2 ) to co-round in each iteration so that the probability that a1 a2 < 0 is only about the 1/2 in
worst case. Next, instead of enforcing either x1 or x2 to be integral after a single step, we allow both x1 and
x2 to remain fractional in some cases. For example, suppose a1 = +1, a2 = −1, x1 = 0.1, and x2 = 0.2.
The normal approach will round (x1 , x2 ) to (0, 0.1) with probability 8/9 and (0.9, 1) with probability 1/9.
Our idea is to round this pair “symmetrically”, getting (0, 0.1) with probability 1/2 and (0.2, 0.3) with
probability 1/2; importantly, this symmetric amount of change (a parameter called δ in SRDR) is chosen
globally, instead of on the particular ai , xi chosen (at random). All of these make our analyses of SRDR
possible.
Note that the rounding step of the (2.675 + )-approximation algorithm for k-median in [3] relies on the
method in Theorem 2.2. SRDR is versatile enough so that we can replace this method by SRDR. Although
this simplifies the main algorithm in [3], it will require opening some O(1/) additional facilities in the
pseudo-solution. By exploiting the fact that all facilities have uniform weight in k-median, the authors
in [3] applied a geometric grouping step which allows only O(log(1/)) additional facilities. However, this
method cannot be applied when facilities have different weights as in the knapsack median problem below.
2.3
Applications
A typical application of dependent rounding is to convert a fractional solution, for example the solution to
some LP relaxation, into an integral solution. For SRDR, this is typically a two-part process. First, the
SRDR algorithm
converts the fractional solution into a nearly integral solution, which preserves the weight
P
function
ai xi , while approximately preserving the moments. Next, one must apply a problem-specific
“end-game” to convert the partially integral solution to a fully-integral solution.
There are a number of possible strategies for this second step.
P One attractive option is to apply independent rounding; this can cause a small change in the weight ai xi . However, since only the t fractional
values of xi are affected by this second rounding step, this will cause a much smaller change than if one
applied independent rounding to the original fractional values.
Another possible option,Pwhich we will use in Section 4, is to force all fractional values of x to be zero
or one in order to minimize i ai xi . This ensures that the final solution has smaller weight than the original
fractional solution; again, since there are only t fractional values, which are the only entries of x affected by
this discretization step, this process does not affect the large-scale behavior of the solution.
2.3.1
The knapsack center problem
We let R denote the optimal radius of the given instance. For the standard Knapsack Center problem with one
constraint, we give a polynomial-time algorithm which returns a feasible solution such that (1) all vertices
are within distance 3R from some chosen center and (2) almost all vertices have expected connection cost
at most about (1 + 2/e)R ≈ 1.74R. For the Multi Knapsack Center problem, we show that it is possible to
obtain a similar result (where all vertices have small expected connection cost) while slightly violating the
knapsack constraints. (Again, the violation is likely unavoidable because it is NP-hard to approximate this
problem to within any constant factor.)
O
1
δγ 2
Theorem 2.3. For any δ, γ > 0, there exists an algorithm which runs in n
time and returns a
feasible solution the Knapsack Center problem (i.e., with m = 1 knapsack constraint) such that Ti ≤ 3R
for all i ∈ V . Moreover, there is a set U ⊆ V (which is deterministic, not a random variable), such that:
1. |U | ≥ (1 − δ)n,
2. ∀i ∈ U, E[Ti ] ≤ (1 + 2/e + γ)R,
3. ∀i ∈ V, Ti ≤ 3R with probability one.
7
Theorem 2.4. For any γ, ∈ (0, 1/2), there is an algorithm for the
Multi Knapsack
Center problem (i.e.,
with m ≥ 1 knapsack constraints) which runs in expected time n
S such that
O −1 m2
q
log(m/γ)
γ
and returns a solution
1. M Y ≤ (1 + )~1, where Y is the indicator vector of S,
2. ∀i ∈ V, E[Ti ] ≤ (1 + 2/e + γ)R,
3. ∀i ∈ V, Ti ≤ 3R with probability one.
2.3.2
The knapsack median problem
We show that SRDR can be applied in a crucial step of Li-Svensson’s algorithm [14] to make it work for the
KM problem, yielding the following result.
Theorem 2.5. For any > 0 and an instance I = (F, C, d, w) of the Knapsack Median problem, there
exists a polynomial-time algorithm
which returns a solution S ⊆ F such that w(S) ≤ 1 + and such that
√
the cost of S is at most (1 + 3 + ) times the optimal cost.
3
Symmetric Randomized Dependent Rounding (SRDR)
In this section, we introduce the SRDR scheme and prove that it obeys certain correlation properties. Suppose we are given vector x ∈ [0, 1]n and a weight vector a ∈ Rn . Let frac(x) = {i : 0 < xi < 1} denote the
set of indices of fractional values of x. Also, for any set I = {i1 , i2 , . . . , ik }, let xI := {xi1 , xi2 , . . . , xik }
denote the subvector of x indexed by I. Let ei denote the unit vector in the ith dimension in the standard
basis. The algorithm to round x is SRDR below, which crucially uses Algorithm ROUND.
Algorithm 1 ROUND(x, a)
1: δ ← mini |ai | min{xi , 1 − xi }
2: Randomly choose a pair i∗ , j ∗ ∈ [n] where i∗ < j ∗
3: With probability 1/2,
x ← x + (δ/ai∗ )ei∗ − (δ/aj ∗ )ej ∗ ,
else,
x ← x − (δ/ai∗ )ei∗ + (δ/aj ∗ )ej ∗ .
4:
return x
Algorithm 2 SRDR(x, a, t)
For each i such that ai = 0, draw xi as an independent Bernoulli-xi variable.
while |frac(x)| > t do
3:
xfrac(x) ← ROUND(xfrac(x) , afrac(x) )
# Apply ROUND to fractional elements of x
4: return x
1:
2:
8
Proposition 3.1. Given vectors x ∈ [0, 1]n , a ∈ Rn , and integer t ≥ 1, the algorithm SRDR(x, a, t)
2
will return a vector X ∈ [0, 1]n with at most t fractional values
P in expected
P O(n ) time. Moreover, the
weighted sum and marginal probabilities are both preserved: i ai Xi = i ai xi with probability one, and
E[Xi ] = xi for all i ∈ [n].
Proof. The runtime of SRDR is determined by the number of calls to ROUND. Consider a single such call.
We know δ > 0, since all elements of xfrac(x) are in (0, 1), and all elements of afrac(x) are non zero (due to
step 1 of SRDR). Let i0 be a minimizer in step 1 of ROUND. If i0 is chosen in step 2, then at least one of the
2
two possible outcomes of step 3 will produce xi0 ∈ {0, 1}. Thus, with probability at least frac(x)
· 12 ≥ n1 ,
each call to ROUND will reduce the number of fractional elements of x by at least 1. In expectation, the
process will terminate after O(n(n − t)) = O(n2 ) calls.
Step 1 of SRDR preserves both the marginal probability (by definition of a Bournoulli-p variable) and
the weighted sum (because ai = 0). Step 3 of ROUND preserves both the weighted sum and the marginal
probabilities of xi∗ , xj ∗ by construction. By induction, these properties are preserved in the final output.
We now proceed to analyze SRDR. Many of the proofs, which are somewhat technical, are deferred to
Appendix A.
3.1
Upper bound on near-independence
Proposition 3.2. Let X = ROUND(x, a) for vectors x ∈ (0, 1)n and a ∈ (R \ {0})n , n ≥ 2. For any set
S ⊆ [n],
"
#
Y p
Y p
E
Xi
xi
≤
i∈S
holds for p = 1 −
1
n.
i∈S
Furthermore, if all weights in aS have the same sign, the inequality holds for p = 1.
Proof. See Appendix A.
Proposition 3.3. Let X = ROUND(x, a) for vectors x ∈ (0, 1)n and a ∈ (R \ {0})n , n ≥ 2. For any
disjoint sets S, T ⊆ [n],
Y
p
Y Y
p
Y
E
Xi
(1 − Xj ) ≤
xi
(1 − xj ) ,
(5)
i∈S
j∈T
i∈S
j∈T
holds for p = 1 − n1 . If ai > 0 and aj < 0 for all i ∈ S, j ∈ T , OR if ai < 0 and aj > 0 for all i ∈ S,
j ∈ T , then 5 holds for p = 1.
Proof. Define vectors x0 , a0 and random vector X 00 as
(
(
1 − xi i ∈ T
−ai
0
0
xi :=
,
ai :=
xi
i 6∈ T
ai
i∈T
.
i 6∈ T
Xi00
(
1 − Xi
:=
Xi
i∈T
.
i 6∈ T
(6)
Let X 0 = ROUND(x0 , a0 ). It is straightforward to verify that X 0 and X 00 have the same joint probability
distribution. Then by Proposition 3.2,
"
#
Y
p
Y
p
Y
p Y Y
p
Y
00
≤
x0i =
xi
(1 − xj )
E
Xi
(1 − Xj ) = E
Xi
i∈S
j∈T
i∈S∪T
i∈S∪T
i∈S
j∈T
for p = 1 − n1 . If ai > 0 for all i ∈ S and aj < 0 for all j ∈ T , then all elements of a0S∪T are positive. If
ai < 0 for all i ∈ S and aj > 0 for all j ∈ T , then all elements of a0S∪T are negative. In either case, the
above steps hold for p = 1.
9
We can now prove the upper bound of Theorem 2.1:
Proof of Theorem 2.1 (Part 1). Define S 0 := {i ∈ S | ai 6= 0} and T 0 := {i ∈ T | ai 6= 0}. Observe that
for each i ∈ (S \ S 0 ) ∪ (T \ T 0 ), Xi is an independent Bournoulli-xi , as determined by step 1 of SRDR. In
this case, E[Xip ] = E[Xi ] = xi ≤ xpi (for p ≤ 1), and similarly E[(1 − Xi )p ] ≤ (1 − xi )p . Thus,
p i
h Y
p i
Y
Y
Y
h Y
Y
E (1 − Xi )p E
Xi
(1 − Xj )
E Xip
E
Xi
(1 − Xj )
=
i∈S
≤
Y
xpi
i∈S\S 0
i∈S 0
i∈T \T 0
i∈S\S 0
j∈T
Y
(1 − xi )p E
h Y
i∈S 0
i∈T \T 0
Xi
Y
(1 −
j∈T 0
p i
Xj )
.
(7)
j∈T 0
Now, at each call to ROUND, xfrac(x) and afrac(x) satisfy the conditions for Proposition 3.3, with n =
|frac(x)| ≥ t + 1 ≥ 2. Furthermore, by Jensen’s inequality, (5) additionally holds for any p ≤ 1 −
1/|frac(x)|, namely for p = 1 − 1/(t + 1). Then if X = SRDR(x, a, t), we can show by induction that
h Y
Y
Y
Y
p i
p
E
Xi
(1 − Xj )
xi
(1 − xj ) .
(8)
≤
i∈S 0
j∈T 0
i∈S 0
j∈T 0
Combining (7) and (8) yields (2) for p = 1 − 1/(t + 1).
If ai > 0 and aj < 0 for all i ∈ S 0 , j ∈ T 0 , OR if ai < 0 and aj > 0 for all i ∈ S 0 , j ∈ T 0 , then (5)
holds for p = 1, as do (8) and (2) by the same reasoning as above. Finally, we relax the strict inequalities as
S and T may contain i with ai = 0.
3.2
Lower bound on near-independence
In this section, we show the lower bound in Theorem 2.1.
Proposition 3.4. Let X = ROUND(x, a) for vectors x ∈ (0, 1)n and a ∈ (R \ {0})n , n ≥ 2. For set
S ⊆ [n], define S1 := {i ∈ S | ai > 0} and S2 := {i ∈ S | ai < 0}. If |S1 |, |S2 | ≤ n − 1, then
h Y p i Y p
Xi
xi
E
≥
i∈S
i∈S
n
o
|S1 |−1 |S2 |−1
holds for p = 1 + max n−|S
,
.
1 | n−|S2 |
Proof. See Appendix A.
Proposition 3.5. Let X = ROUND(x, a) for vectors x ∈ (0, 1)n and a ∈ (R \ {0})n , n ≥ 2. Let S, T ⊆ [n]
be disjoint sets. Define S 0 := {i ∈ S | ai > 0} ∪ {i ∈ T | ai < 0} and T 0 := {i ∈ S | ai < 0} ∪ {i ∈ T |
ai > 0}. If |S 0 |, |T 0 | ≤ n − 1, then
Y
p
Y Y
p
Y
(9)
E
Xi
(1 − Xj ) ≥
xi
(1 − xj ) ,
i∈S
holds for p = 1 + max
n
i∈S
j∈T
|S 0 |−1 |T 0 |−1
n−|S 0 | , n−|T 0 |
o
.
10
j∈T
Proof. This proof is very similar to that of Proposition 3.3. Reusing the definitions for x0 ,a0 , and X 00 from
(6), observe that S 0 = {i ∈ S ∪ T | a0i > 0} and T 0 = {i ∈ S ∪ T | a0i < 0}. Then if |S 0 |, |T 0 | ≤ n − 1, we
may apply Proposition 3.4 to the show the following
"
#
Y
p
Y
p
Y
p Y Y
p
Y
E
Xi
(1 − Xj ) = E
Xi00
≥
x0i =
xi
(1 − xj ) ,
i∈S
j∈T
where p = 1 + max
i∈S∪T
|S 0 |−1
|T 0 |−1
n−|S 0 | , n−|T 0 |
i∈S∪T
i∈S
j∈T
.
Proof of Theorem 2.1 (Part 2). This proof is very similar to that of Theorem 2.1(1). Define S 00 := {i ∈ S |
ai 6= 0} and T 00 := {i ∈ T | ai 6= 0}. For each i ∈ (S \ S 00 ) ∪ (T \ T 00 ), Xi is an independent Bournoulli-xi ,
so we have E[Xip ] = E[Xi ] = xi ≥ xpi (for p ≥ 1), and similarly E[(1 − Xi )p ] ≥ (1 − xi )p . Thus,
Y
p
p
Y
Y
Y p Y
Y
(1 − xi )p E
Xi
(1 − Xj ) .
(10)
xi
(1 − Xj ) ≥
E
Xi
i∈S
i∈S\S 00
j∈T
i∈S 00
i∈T \T 00
j∈T 00
Now, consider a single call ROUND(xfrac(x) , afrac(x) ). Assuming |S 0 |, |T 0 | ≤ t as in the theorem statement, we have |S 0 ∩ frac(x)| ≤ |S 0 | ≤ t ≤ frac(x) − 1 and |T 0 ∩ frac(x)| ≤ |T 0 | ≤ t ≤ frac(x) − 1, so the
conditions for Proposition 3.5 are met. Furthermore, by Jensen’s inequality, (9) additionally holds for any
|S 0 ∩frac(x)|−1
|S 0 |−1
|T 0 ∩frac(x)|−1
|T 0 |−1
p ≥ 1 + max frac(x)−|S
0 ∩frac(x)| , frac(x)−|T 0 ∩frac(x)| , namely for p = 1 + max
t+1−|S 0 | , t+1−|T 0 | . Then
if X = SRDR(x, a, t), we can show by induction that
Y
p
Y
p
Y
Y
E
Xi
(1 − Xj ) ≥
xi
(1 − xj ) .
(11)
i∈S 00
j∈T 00
i∈S 00
j∈T 00
Combining (10) and (11) yields (3).
Lastly, observe that max{|S 0 |, |T 0 |} ≤ |S 0 ∪ T 0 | = |S ∪ T |. This implies that |S ∪ T | ≤ t is a stronger
|S 0 |−1
|T 0 |−1
|S∪T |−1
assumption than |S 0 |, |T 0 | ≤ t, and also that p = 1 + t+1−|S∪T
| is larger than 1 + max t+1−|S 0 | , t+1−|T 0 |
(and so is valid by Jensen’s inequality).
4
The knapsack center problem
In this section, we shall prove Theorem 2.3 and Theorem 2.4.
4.1
An LP relaxation
Suppose R is the optimal radius. Note that there are only n2 possible values for the optimal radius R.
Thus, we can guess this value in O(n2 ) time. We define the polytope P(I, R) containing points (x, y)
which satisfy the following constraints (C1)—(C7). Given a solution to this LP, our goal is to (randomly)
round y to an integral solution.
P
(C1)
i∈V :d(i,j)≤R xij = 1 for all j ∈ V (all clients should get connected to some center),
(C2) xij ≤ yi for all i, j ∈ V (vertex j can only connect to center i if it is open),
(C3) M y ≤ ~1 (the m knapsack constraints),
(C4) 0 ≤ xij , yi ≤ 1 for all i, j ∈ V .
11
By splitting vertices we may ensure that we satisfy the additional constraints
(C5) For all i, j ∈ V , we have xij ∈ {0, yi },
(C6) For all i ∈ V , we have xii = yi .
It will be useful to enforce an additional constraint on the maximum size of the entries of M :
Proposition 4.1. For any ρ > 0, we can find a solution to the LP satisfying (C1) – (C6), and which satisfies
the additional property
(C7) For any i ∈ V , if Mki ≥ ρ for any k = 1, . . . , m, we have yi ∈ {0, 1}.
The running time for this process in nO(m/ρ) .
Proof. We say that an item i is big if Mki ≥ ρ for any k = 1, . . . , m. Suppose we fix any optimal solution
S. Observe that there can be at most m/ρ big facilities in S. We can guess the set of such facilities in time
nO(m/ρ) . For any big facility i that we guess is in S, we set yi = 1; for any big facility that we guess is
outside S, we set yi = 0. This procedure will only check at most nO(m/ρ) possible cases. If our guess is
correct, then P(I, R) 6= ∅.
For any j ∈ V , let Fj := {i ∈ V : xij > 0}. We refer to these sets as clusters. It is easy to verify that
y(Fj ) = 1 due to (C5). We now form a subset V 0 ⊆ V such that all the clusters Fj , Fj 0 for j, j 0 ∈ V 0 are
pairwise disjoint, and such that V 0 is maximal with this property.
We also define
[
F0 = i ∈ V | yi > 0, i ∈
/
Fj .
0
j∈V
By property (C6), we have j ∈ Fj if yj > 0, which implies that F0 and V 0 are disjoint.
We now partition V into a set of groups. There are two types of groups. First, for each j ∈ V 0 , we define
the group Gj to be simply Fj . Next, for each j ∈ F0 , we create a group Gj which consists of two items,
namely j and a new dummy item Dummy(j), which has y(Dummy(j)) = 1 − yj , which has distance ∞
from all i ∈ V , and which has cost zero according to all m knapsack constraints. (If we select this dummy
item, it simply means that we do not choose to select item j ∈ F0 .) We let W denote the set of groups.
The following notation will be useful throughout: for any set X ⊆ [n], we define frac(X) to be the set
of items i ∈ X such that yi ∈
/ {0, 1}; here y will always denote the current value of the vector y ∈ [0, 1]n .
In the first step of both algorithms, we simplify y to reduce the number of fractional items in each Gj
to at most m + 1 (this is automatically the case for j ∈ F0 ). We do so using the following algorithm
K NAPSACK I NTRAGROUP R EDUCE:
Algorithm 3 K NAPSACK I NTRAGROUP R EDUCE (y, X)
1:
2:
3:
4:
5:
while |frac(X)| > m + 1 do
Let δ ∈ Rn , δ 6= 0 be such that M δ = 0, δ(X) = 0, and δi = 0 ∀i ∈
/ frac(X)
Choose scaling factors a, b > 0 such that
• y + aδ ∈ [0, 1]n and y − bδ ∈ [0, 1]n
• there is at least one entry of y + aδ which is equal to zero or one
• there is at least one entry of y − bδ which is equal to zero or one
b
With probability a+b
, update y ← y + aδ; else, update y ← y − bδ.
return y
12
Proposition 4.2. One can find a vector δ ∈ Rn as claimed in line 2 of K NAPSACK I NTRAGROUP R EDUCE.
Proof. This line is only executed when we have at least m + 2 variables in X. On the other hand, we
have only m knapsack constraints (M δ = 0) and the additional linear constraint δ(X) = 0. This system is
underdetermined and the claim follows.
Proposition 4.3. Suppose y 0 = K NAPSACK I NTRAGROUP R EDUCE(y, X) for any X ⊆ [n]. Then y 0 (X) =
y(X) and M y = M y 0 and E[yi0 ] = yi for all i ∈ V .
Proof. In each round of K NAPSACK I NTRAGROUP R EDUCE, we update y by either y 0 := y + aδ or y 0 :=
y − bδ. Since δ is chosen so that δ(X) = 0, we have y 0 (X) = y(X) + aδ(X) = y(X) or y 0 (X) =
y(X) − bδ(X) = y(X). Similarly, as M δ = 0 we have M y = M y 0 .
Also, for any i ∈ V we have
E[yi0 ] = yi +
b
a
(aδi ) −
(bδi ) = yi .
a+b
a+b
The claim follows by induction on all iterations.
4.2
Proof of Theorem 2.3
Now we show that, when m = 1 (the standard Knapsack Center problem), one can satisfy the knapsack
constraint with no violation while guaranteeing that the expected approximation ratio of at least (1 − δ)n
vertices is at most 1 + 2/e + γ for any γ > 0.
High-level ideas: First, by a preprocessing step (the P RUNE algorithm), we find a fractional solution
in which each vertex i with yi > 0 will only serve (fractionally) at most n other vertices. (We do not
really need to filter out “large” vertices having weight more than ρ in this case.) Next, we use the procedure
K NAPSACK I NTRAG ROUP R EDUCE to reduce the number of fractional variables inside Gj down to 2 for all
j ∈ W . Note that the opening mass y(Gj ) = 1 remains unchanged in the process. Define Xj to be the
indicator for choosing the fractional vertex of higher weight in Gj . We then use SRDR to round vector X
into an “almost” integral solution. That is, there will be at most some O(1) groups Gj containing exactly
two fractional vertices. Finally, we open the vertex with smaller weight in each fractional group.
We let M denote the weight function; since we are considering here the case that m = 1, we view M as
an n-long vector where Mi is the weight of node i ∈ V .
(We note that it is formally possible for this process to return S which contains some dummy items.
As these dummy items have infinite distance and zero weight, they contribute nothing and can simply be
discarded.)
Proposition 4.4. The procedure P RUNE runs in nO(1/) time and will return EITHER an optimal solution
OR a set S of open centers along with a fractional solution (x, y) for the residual instance I 0 = (V, d, m),
in which each center i only serves ≤ n other vertices fractionally (i.e., |{j ∈ V : xij > 0}| ≤ n for all
i ∈ V ).
Proof. Fix any optimal solution to the given instance. In Algorithm 4, whenever we find a vertex i which
serves ≥ n other vertices, we simply guess two cases: whether or not i is in the optimal solution. Note that
for the guess “i is in the optimal solution”, we open i and remove at least n other vertices from the instance.
M
(Because we assume that the budget constraint has RHS value of 1, this requires rescaling M to 1−M
.
i
0
In the other case, we remove i from V . Observe that if our guess is correct, P(I , R) will not be empty
in the next step. The algorithm stops when the current fractional solution satisfies all the properties in the
claim.
13
Algorithm 4 P RUNE (, M, V, S, R)
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
if V = ∅ then
return S // we obtain an optimal solution in this case
I 0 ← (M, V )
if P(I 0 , R) 6= ∅ then
Compute a solution (x, y) ∈ P(I 0 , R)
if there exists some center i such that Mi ≤ 1 and |{j ∈ V : xij > 0}| ≥ n then
Let X ← {j ∈ V : xij > 0}
M
return P RUNE (, 1−M
, V \ X, S ∪ {i}, R) if it is not FALSE
i
return P RUNE (, M, V \ {i}, S, R))
else
return (the solution (x, y), the set of remaining vertices V , the scaled weight matrix M , and the
set of open centers S)
else
return FALSE
Algorithm 5 S TANDARD K NAPSACK SRDR (M, V, t)
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
(x, y, V, M, S0 ) ← P RUNE(, M, V, ∅, R) and return S if it already covers all nodes in V
for each j ∈ V 0 do update y ←K NAPSACK I NTRAGROUP R EDUCE (y, Gj ).
Let W 0 denote the set of groups j ∈ W with frac(Gj ) = 2.
for each j ∈ W 0 do
Suppose Gj = {i1 (j), i2 (j)} and Mi1 (j) ≤ Mi2 (j)
Set xj ← yi2 (j) and aj ← Mi2 (j) − Mi1 (j)
X ← SRDR(x, a, t)
for each j ∈ W 0 do
if Xj = 1 then
Set yi1 (j) ← 0, yi2 (j) ← 1
else
Set yi1 (j) ← 1, yi2 (j) ← 0
return S = S0 ∪ {i ∈ V | yi = 1}
To bound the running time of this algorithm, we can visualize its execution by a binary tree in which
each node is either a vertex chosen in line 6 or a leaf. Each non-leaf node of the tree has two children
corresponding to two decisions whether or not it is in the optimal solution. Indeed, the running time of the
algorithm is bounded by the number of paths from the root to any leaf of this tree. Because (1) the length
of any path is at most n and (2) the number of vertices chosen to be in the optimal solution in this path is at
most 1/, the number of such paths is at most nO(1/) .
For the rest of this section, we will be working on the residual instance (M, V ) returned by P RUNE. In
the first part of the analysis, we show an upper bound on the probability that a given vertex k ∈ V has no
open facility in Fk . This is done by defining a natural potential function, which is an estimation for this
probability. We will analyze its change after applying SRDR.
For each j ∈ W we let Cj = Fk ∩ Gj be the set of vertices that vertex k is “interested in” from group j.
Let y denote the current fractional solution at the beginning of line 7 of S TANDARD K NAPSACK SRDR (just
14
before the execution of SRDR). We define a potential function:
Y
S :=
(1 − y(Cj )) .
j∈W
Let x0 be the value of the vector after executing SRDR. Let us next define a vector y 0 as follows: for
0
0
each j ∈ W , we set yi1(j)
= x0j , yi2(j)
= 1 − x0j . For all other indices i we set yi0 = yi . We define
S 0 :=
Y
1 − y 0 (Cj ) .
j∈W
Proposition 4.5. Conditioned on any value of y and S, we have
E[(S 0 )1−1/(t+1) ] ≤ S 1−1/(t+1) .
Proof. Let p = 1 − 1/(t + 1). First, note that if j ∈ W − W 0 , then the execution of SRDR does not change
the value of y(Cj ). Similarly, we have y 0 (Cj ) = y(Cj ) if Cj has size zero, or if Cj contains both the items
in Gj . So let us define the sets A, B as
A = {j ∈ W : Cj = {i1 (j)}}
B = {j ∈ W : Cj = {i2 (j)}}
Thus, it suffices to show that
E[
Y
Y
(1 − y 0 (Cj ))p ] ≤
j∈A∪B
(1 − y(Cj ))p
j∈A∪B
Note that y(Cj ) = xj for j ∈ B and y(Cj ) = 1 − xj for j ∈ A, and similarly for y 0 . Thus, it suffices to
show that
Y
Y
Y
Y
p
p
E[
x0j
(1 − x0j ) ] ≤
xj
(1 − xj )
j∈A
j∈B
j∈A
j∈B
This is precisely the upper-correlation property of the SRDR (Theorem 2.1(1)).
Proposition 4.6. We have that
E[S 0 ] ≤ (1/e)1−1/(t+1) .
Q
Proof. Let y 0 be the original fractional solution y at line 1 and S0 := j∈W (1 − y0 (Cj )). We have
P
Q
0
− i∈C yi0
j
S0 ≤ j∈W e
= e−y (Fk ) = 1/e. Next, since the clusters Gj are processed independently in line
1, and marginals are preserved by Proposition 4.3, we have
Y
Y
E[S] =
(1 − E[yi1 ]) =
(1 − yi0 ) = S0 ≤ 1/e.
j∈W
j∈W
Proposition 4.5 and Jensen’s inequality give
E[S 0 ] ≤ E[(S 0 )1−1/(t+1) ] ≤ E[S 1−1/(t+1) ] ≤ E[S]1−1/(t+1) ≤ (1/e)1−1/(t+1) .
15
Proof of Theorem 2.3. Feasibility of S: Recall that after calling the procedure P RUNE at line 1, the fractional solution (x, y) is feasible for the residual instance (V, M ). It suffices to show that we do not violate
the scaled knapsack constraint of the residual instance when rounding y. Proposition 4.3 ensures that the
knapsack constraint is preserved (fractionally) by K NAPSACK I NTRAG ROUP R EDUCE.
We can interpret lines 8 – 12 of S TANDARD K NAPSACK SRDR as a two-part process. First, we set
yi1 (j) = 1 − Xj , yi2 (j) = Xj . Next, for each j ∈ W 0 with Xj ∈ (0, 1), we set yi1 (j) = 1, yi2 (j) =
0. We refer to the latter stepP
as rounding down j. Every aspect of the S TANDARD K NAPSACK SRDR
preserves the marginal weight i Mi yi , except this rounding-down
step. Since Mi1 (j) ≤ Mi2 (j) for every
P
j ∈ W 0 , the rounding-down step can only
decrease
M
y
.
As
a consequence
P
P i i0 i
P
P of this, atPthe end0 of
S TANDARD K NAPSACK SRDR, we have i Mi yi ≤ i Mi yi . Thus i∈S Mi ≤ i∈S0 Mi + i Mi yi ≤
1.
Cost analysis: For any k ∈ V , we have Tk ≤ R if there is a facility opened in Fk , and Tk ≤ 3R
otherwise. (By construction, Fk ∩ Fj 6= ∅ for some j ∈ V 0 and we always open one center inside Gj . Then
the distance from k to this center is at most dkj + R ≤ 3R by triangle inequality.) Note that there are at
most 2t fractional vertices in V . For each vertex j, let qj denote the probability that j is adjacent to such an
unrounded vertex.
X
XX
qj ≤
Pr[k is fractional]
(by the union-bound)
j∈V
j∈V k∈Fj
=
X
Pr[k is fractional]
1
j∈Fk
k∈V
≤ (n)
X
X
Pr[k is fractional]
(by the pre-processing step)
k∈V
≤ (2t)n
We say that a vertex is good if qj ≤ 1/t and bad otherwise. Then the number of bad vertices is at most
2t2 n. We let U be the set of good vertices. Let y 00 be the vector y after finishing the for-loop at lines 9–10
and let
Y
S 00 :=
(1 − y 00 (Cj )).
j∈W
Now fix any k ∈ U . We have that Fk ∩ S = ∅ iff S 00 = 1, and S 00 = 0 otherwise. Let E denote the event
that there are no fractional vertices in Fk before line 9. Note that Pr[E] = 1 − qk ≥ 1 − 1/t. Then, for t
large enough,
Pr[S ∩ Fk = ∅] = Pr[S ∩ Fk = ∅ ∧ E] + Pr[S ∩ Fk = ∅ ∧ ¬E]
≤ Pr[S 00 = 1 ∧ E] + Pr[¬E]
≤ Pr[S 0 = 1 ∧ E] + 1/t
(if E then S 0 = S 00 )
≤ E[S 0 ] + 1/t
≤ (1/e)1−1/(t+1) + 1/t ≤ 1/e + 2/t
Thus,
E[Tk ] ≤ R + (2R) Pr[S ∩ Fk = ∅] ≤ (1 + 2/e + 2/t)R
Now for any δ, γ > 0, by setting t = 2/γ and =
δ
,
2t2
the number of badvertices
is ≤ δn and, for any
k ∈ U we have E[Tk ] ≤ (1 + 2/e + γ)R. The running time is nO(1/) = n
16
O
1
δγ 2
.
4.3
Proof of Theorem 2.4
Here we describe a generalization of SRDR for rounding a fractional solution of the Multi Knapsack Center
problem. Again, we will round y to an “almost” integral vector. During this process, whenever a fractional
variable becomes integral (i.e., yi = 0 or yi = 1), then it stays at that values for the remainder of the
algorithm. Furthermore, if yi = 1, then i will be put into S. Thus the fractional clusters frac(Gj ) will
continually shrink; so frac(Gj ) always denotes the vertices in Gj which are fractional for the current value
of y.
Finally, we use dependent rounding (independently on each group) to completely transform y to an
integral vector, which will be our solution S.
We summarize these three steps via the following algorithm K NAPSACK D EPENDENT ROUND (which
we abbreviate as KDR).
Algorithm 6 KDR (y, M, t)
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
0
for each
P j ∈ V do update y ←K NAPSACK I NTRAGROUP R EDUCE (y, Gj );
while j∈W (|frac(Gj )| − 1) > t do
J := ∅
for each j ∈ W do
|frac(Gj )|−1
With probability pj = 3m × P
, update J ← J ∪ {j}
u∈W (|frac(Gu )|−1)
Select two distinct vertices ij1 , ij2 ∈ frac(Gj ) uniformly at random
if |J| ≥ m + 1 then
Choose δ ∈ Rn , δ 6= 0 such that
•
M δ = 0, y + δ ∈ [0, 1]n , and y − δ ∈ [0, 1]n
•
There is at least one entry of y + δ or y − δ which is equal to zero or one
•
For each j ∈ J, δij1 = −δij2
•
δi = 0 if i is not a selected vertex from some Gj where j ∈ J
With probability 1/2, update y ← y + δ; else, update y ← y − δ
For each j ∈ W select a vertex Xj ∈ Gj , wherein we select Xj = i with probability yi
return S = {Xj | j ∈ W }
We first show that these procedures are well-defined and terminate in expected polynomial time.
Proposition 4.7. One can find a vector δ ∈ Rn as claimed in line 8 of KDR.
Proof. For each j ∈ J, let xij1 = Uj and xij2 = −Uj for some variable Uj to be determined. Set all
other entries of x to be zero. Again, the system of M x = 0 has |J| ≥ m + 1 variables Uj and only m
constraints, hence is underdetermine. So let x be any non-zero solution vector. Let a ∈ R be maximal such
that y + ax ∈ [0, 1]n and y − ax ∈ [0, 1]n . Then one may verify that a < ∞ and setting δ = ax achieves
the claimed result.
Proposition 4.8. Suppose y 0 = KDR (y, M, t). Then y 0 (Gj ) = y(Gj ) for all j ∈ W and E[yi0 ] = yi for all
i ∈ V , and M y = M y 0 .
Proof. In each round of Algorithm 6 that y is changed, we update it by either y 0 := y + δ or y 0 := y − δ.
Note that δ is chosen so that, for any j ∈ W , δ(Gj ) = δij1 + δij2 = 0. Thus, y 0 (Gj ) = y(Gj ).
Similarly, δ is chosen so that M δ = 0, so that M y = M y 0 .
Also, for any i ∈ V , we have
E[yi0 ] = yi + (1/2)δi − (1/2)δi = yi .
The claim follows by induction on all iterations.
17
Proposition 4.9. The probability distribution in line 10 of KDR is valid, i.e. y(Gj ) = 1.
Proof. In the LP solution, we have y(Gj ) = 1 by construction (C5) and by construction of Gj . After line 1,
we have y(Gj ) = 1 by Proposition 4.3. Before line 9, we have y(Gj ) = 1 by Proposition 4.8.
Proposition 4.10. When we run KDR(y, M, t) with t ≥ 3m2 , then for all iterations of Algorithm 6 the
probability vector p is valid, i.e. p ∈ [0, 1]n .
Proof. We clearly have pj ≥ 0.
P
Line 5 of that algorithm is only executed as long as u∈W (|frac(Gu )| − 1) > t. Also, after applying
K NAPSACK I NTRAGROUP R EDUCE to all clusters, we have |frac(Gj )| ≤ m + 1 for each j ∈ W . The
fractional clusters can only shrink during the execution of Algorithm 6, and so |frac(Gj )| ≤ m + 1 for all
iterations of Algorithm 6. Thus we have
|frac(Gj )| − 1
m
≤1
≤ 3m ×
t
u∈W (|frac(Gu )| − 1)
pj = 3m × P
as desired.
Proposition 4.11. In any iteration of KDR, conditional on all prior state, there is a probability of at
least 1/10 that a variable becomes rounded. Thus, KDR terminates with probability one and has expected
poly(n) running time.
Proof. In any round with |J| ≥ m + 1, there is at least a 1/2 probability of producing P
at least one quantized
variable. Also, observe that |J| is a sum of independent random variables with mean j∈W pj = 3m, and
2
hence by the Chernoff-Hoeffding bound the probability that it is ≤ m is at most e−m(2/3) ≤ 0.65. Thus,
there is a probability of at least (1 − 0.65) × (1/2) ≥ 1/10 that there is a rounded variable.
This implies that the expected number of iterations of that algorithm is O(n). It is clear that individual
steps of the algorithm can be executed in polynomial time.
Proposition 4.12. Let y denote the fractional vector immediately before line 10 of KDR. Then M y ≤ ~1.
Proof. In the original fractional solution we have M y ≤ ~1. By Propositions 4.3,4.8, M y does not change
through any of the intermediate steps of the algorithm.
Next, we will analyze E[Tk ] for a given vertex k ∈ V . To do so, we we show an upper bound on
the probability that there is no open facility in Fk . Recall that, for each j ∈ W , Cj = Fk ∩ Gj . Let y
denoteP
the current fractional solution at the beginning of some round q of KDR. For each j ∈ W we define
cj := i∈Cj yi and a potential function:
S=
Y
(1 − cj )
j∈W
In Appendix B, we show that S does not increase too much in expectation over the execution of KDR).
Proposition 4.13. Let S denote the value of the potential function at the beginning of the dependent rounding
process (immediately after line 1 of KDR), and let S 0 denote
the value of the potential function at the end
P
of all rounds (i.e. at line 10 of KDR, after terminating j∈W (|frac(Gj )| − 1) ≤ t. Then
E[S 0 ] ≤ S +
Proof. See Appendix B.
18
180m2
.
t
Proposition 4.14. For any vertex k ∈ V , the probability that S ∩ Fk = ∅ when KDR terminates is at most
2
1/e + 180m
t .
Proof. In this proof, let y 0 be the original fractional vector y which is the input of Algorithm KDR. Let y 1
be the modified vector y after executing lines 1. Let y 2 be the modified vector immediately before line 10.
Finally, let y 3 = Y , where Yi is the indicator variable for S.
Now fix any vertex k ∈ V . If k ∈ V 0 then there is an open facility in Fk with probability 1. Suppose
k∈
/ V 0 . We define
Y
X
S0 :=
(1 −
yi0 )
j∈W
S1 :=
Y
i∈Cj
j∈W
S2 :=
Y
S3 :=
yi1 )
i∈Cj
X
(1 −
j∈W
Y
X
(1 −
yi2 )
i∈Cj
X
(1 −
j∈W
yi3 )
i∈Cj
Observe that Xj ∈ Gj and Xj ∈ S. As the sets Gj are all pairwise disjoint, then S ∩ Gj = {Xj }. This
implies that |S ∩ Cj | is equal to zero or one, for each j ∈ W . Also, the entries of y 3 are all zero or one.
2
Hence, S3 = 1 if S ∩ Fk = ∅, and S3 = 0 otherwise.
So, it suffices to show that E[S3 ] ≤ 1/e + 180m
t .
P
Q
− i∈C yi0
j
Now, note that we have S0 ≤ j∈W e
= e−y(Fk ) = 1/e.
Next, since the clusters Gj are processed independently in lines 1, and marginals are preserved by
Proposition 4.3, we have
Y
Y
E[S1 ] =
(1 − E[yi1 ]) =
(1 − yi0 ) = S0 .
j∈W
j∈W
Next Proposition 4.13 shows that
E[S2 ] ≤ E[S1 ] +
180m2
.
t
Finally, the clusters Gj are processed independently in line 10, and again marginals are preserved by
Property (P1), so E[S3 ] = E[S2 ].
2
In total, E[S3 ] ≤ 1/e + 180m
t .
Next, we use concentration bounds to show that there is a good probability of not violating the knapsack
constraints by too much.
2
Proposition 4.15. With probability at least 1 − m exp − 2
, we have that M Y ≤ (1 + )~1.
2
tρ
P
Proof. Immediately before line 10 of KDR, we have j∈W (|frac(Gj )| − 1) ≤ t; also by Proposition 4.12,
we have M y ≤ 1. This means that there are at most t groups Gj with frac(Gj ) > 0 (that is, the value of y
for the vertices in that cluster are not completely integral).
Now suppose we condition on the state before line 10. Mk Y can be viewed as a sum of independent
random variables (for each j ∈ W , the value of Mk,Xk ). However, at most t of these variables are random
— the remaining are deterministic (if frac(Gj ) = 0, then there is exactly one entry k ∈ Gj with yk = 1,
and Xj = k).
19
Thus we can view Mk Y is a weighted sum of ` ≤ t independent random variables, in which each
variable is bounded in the range [0, ρ]. We may apply Hoeffding’s bound: the probability that such a sum
2
exceeds its mean by is at most exp(− 2
).
tρ2
This is the probability of violating a single knapsack constraint. By the union bound, the total probability
2
of violating any constraint is at most m exp(− 2
).
tρ2
Proof of Theorem 2.4. By rescaling, it suffices to show that E[Ti ] ≤ (1 + 2/e + O(γ))R. We will choose
parameters ρ, t and obtain a fractional solution which satisfies (C1) – (C7). We will independently repeat
KDR until M Y ≤ (1 + )~1; thus it suffices to show that after a single application of KDR we have
1. M Y ≤ (1 + )~1 with constant probability
2. E[Ti | M Y ≤ (1 + )~1] ≤ (1 + 2/e + O(γ))R.
We have Ti ≤ R if there is a facility opened in Fi , and Ti ≤ 3R otherwise. So, by Proposition 4.14,
the expected distance is at most (1 + 2(1/e + 180m2 /t))R. By Proposition 4.15, the probability of satis2
fying (approximately) the knapsack constraints is at least 1 − m exp(− 2
). Hence, the expected distance
tρ2
conditional on satisfying knapsack constraints is at most
1 + 2/e + 320m2 /t
R.
E[Ti | M Y ≤ (1 + )~1] ≤
2
1 − m exp − 2
2
tρ
Now set t = m2 /γ and set ρ =
√
m
√
γ
ln(m/γ)
. Then
22
m exp − 2 = γ 2 /m ≤ 1/4
tρ
Thus
E[Ti | M Y ≤ (1 + )~1] ≤
1 + 2/e + 320γ
R ≤ R 1 + 2/e + O(γ)
2
1 − γ /m
The work is
solving the LP to ensure (C1) — (C7); by Proposition 4.1, this takes time
dominated by
nO(m/ρ)
5
=n
O −1 m2
q
log(m/γ)
γ
.
An improved bi-factor approximation algorithm for the knapsack median
problem
√
In this section, we show that one can obtain a bi-factor (1 + 3 + )-approximation algorithm for the KM
problem if allowed to slightly violate the budget constraint by a factor
√ of (1 + ). This algorithm is inspired
by the work of Li-Svenssson [14], in which they gave a (1 + 3 + )-approximation algorithm for the
k-median problem. The idea is to compute a “bi-point” solution which is a convex combination of a feasible
solution and another√
pseudo solution. Then rounding this solution will result in a solution whose cost can
be bounded by (1 + 3 + ) times the optimal cost but would slightly violate the knapsack constraint. The
rounding step can be done using SRDR and some O(1) left-over fractional variables will have to be rounded
up to 1, resulting in a small violation in the total weight. In [14], the authors use a postprocessing step to
correct the solution, making the cardinality constraint to be preserved exactly. However, this step does not
seem to work for the knapsack median problem.
20
5.1
Pruning “big” facilities and Computing a bi-point solution
Let I0 = (F0 , C, d, w) be any KM instance and > 0 be some small parameter. Let us also fix an optimal
integral solution S0 of I0 . Note that S0 may contain at most 1/ facilities which have weight greater than or
equal to . Let F = {i ∈ F0 : wi ≥ } and S = S0 ∩ F . Then |S | ≤ 1/. Hence one may “guess” this
set S in nO(1/) time.
From now on, suppose that we already have the “correct” set S . Define B := 1 − w(S ). Let I =
(F, C, d, w) denote the residual instance in which (i) we eliminate all “big” facilities: F = (F0 \ F ) ∪ S
and (ii) the weights of facilities in S are set to zero: wi := 0 ∀i ∈ S .
Lemma 5.1. Suppose there is an algorithm A which returns a solution S = S 0 ∪S 00 for the residual instance
I where S 00 ∩ S = ∅ and c = |S 00 | is a constant, w(S 0 ) ≤ B, and cost I (S) ≤ αOPTI for some constant
α > 0. Then we have that
(i) w(S) ≤ 1 + c,
(ii) cost I0 (S) ≤ αOPTI0 .
Proof. Since all facilities in S 00 have weight at most , w(S 00 ) ≤ c. Also, w(S 0 ) ≤ B ≤ 1. Thus,
w(S) = w(S 0 ) + w(S 00 ) ≤ 1 + c.
Now observe that OPTI ≤ OPTI0 because, by construction of I, the chosen optimal integral of I0 is
also a feasible solution of I. Therefore,
cost I0 (S) = cost I (S) ≤ αOPTI ≤ αOPTI0 .
In the rest of this section, we aim to design such an algorithm A for the residual instance I. We now
compute the so-called bi-point solution as in [14].
Theorem 5.2. There is a polynomial-time algorithm to compute two sets F1 , F2 ⊆ F and constants a, b ≥ 0
and a+b = 1 such that w(F1 ) ≤ B ≤ w(F2 ), a·w(F1 )+b·w(F2 ) ≤ B and a· cost I (F1 )+b· cost I (F2 ) ≤
2OPTI . (The pair (F1 , F2 ) is called a bi-point solution of I.)
Proof. One can extend the algorithm to construct a bi-point solution for the k-median problem in [11, 17]
so that it also works for the weighted case.
WLOG, we may assume that d1 ≥ d2 . Now for each client j ∈ C, we let i1 (j), i2 (j) denote the
closest facilities to j in F1 and F2 , respectively. Next, we adopt the definition of stars as in [14]. Each
facility i2 ∈ F2 will be associated with the closest facility i1 ∈ F1 , breaking ties arbitrarily. Then, for
each facility i ∈ F1 , we define a star centered at i as a set of {i} and its associated facilities, which will be
referred to as leaves. Let Si denote the setPof leaves of the star with center i. Let
P d1 (j) = d(j, i1 (j)) and
d2 (j) = d(j, i2 (j)). Finally, define d1 := j∈C d1 (j) = cost I (F1 ) and d2 := j∈C d2 (j) = cost I (F2 ).
21
5.2
Corner cases: a ≤ 1/5 or a ≥ 4/5
Recall that F1 is a feasible solution as w(F1 ) ≤ B. Suppose we take F1 as our solution. Consider the
following strategy: we connect each client j to the center i of the star Si where i2 (j) ∈ Si . Fix any client j
and let Si be the star containing i2 (j). By triangle inequality and definition of stars, we have that
d(j, i) ≤ d(j, i2 (j)) + d(i2 (j), i) ≤ d2 (j) + d(i2 (j), i1 (j)) ≤ d2 (j) + d1 (j) + d2 (j) = 2d2 (j) + d1 (j).
P
Therefore, the connection cost of this assignment is bounded by j∈C (2d2 (j) + d1 (j)) = 2d2 + d1 . Now
assuming we want to close the center i and open all the facilities Si , how does the connection cost change?
Observe that the client j can now be connected to i2 (j) in this case so that we can “save” the cost by
(2d2 (j) + d1 (j)) − d2 (j) = d1 (j) + d2 (j). Our algorithm will aim to find a subset of stars such that when
closing all the centers and opening all the leaves of these stars, the amount of “saving” is maximized.
For each i ∈ F1 , let zi be the indicator for the event that i is closed and all facilities in Si is opened.
Also, define δ(i) := {j ∈ C : i2 (j) ∈ Si }. The LP relaxation for our problem is as follows.
LPstar-rounding : maximize
X X
(d1 (j) + d2 (j))zi
i∈F1 j∈δ(i)
subject to
X
(w(Si ) − wi ) zi ≤ B − w(F1 )
i∈F1
0 ≤ zi ≤ 1 ∀i ∈ F1 .
We are now ready to describe the algorithm. We first compute a basic optimal solution z ∗ of LPstar-rounding .
It is not difficult to see that z ∗ has at most one fractional value. Suppose zi∗∗ is fractional for some i∗ ∈ F1 .
Consider the following LP.
X
X
LPSi∗ : maximize
(d1 (j) + d2 (j))ti
i∈Si∗ j∈C:i2 (j)=i
subject to
X
wi ti ≤ w(Si∗ )zi∗
i∈Si∗
0 ≤ ti ≤ 1 ∀i ∈ Si∗ .
Next, we compute a basic optimal solution t∗ for LPSi∗ . Again, t∗ contains at most one fractional value.
Suppose t∗i∗∗ is fractional for some i∗∗ ∈ Si∗ . Let
[
S := {i∗ , i∗∗ } ∪
Si ∪ {i ∈ Si∗ : t∗i = 1}.
∗
i∈F1 :zi =1
Our algorithm will return the solution A := arg minZ∈{S,F1 } cost I (Z).
Theorem 5.3. If a ≤ 1/5 or a ≥ 4/5, then there is a bi-factor (2.73 + )-approximation algorithm for KM.
Proof. Let us analyze the above-mentioned algorithm. Let S 00 = {i∗ , i∗∗ }. If wi∗ = 0 or wi∗∗ = 0, then we
remove the corresponding facility from S 00 . Let S 0 = S \ S 00 . By construction and the fact that z ∗ and t∗ are
22
feasible solutions of LPstar-rounding and LPSi∗ respectively, we have that
w(S 0 ) =
X
w(Si ) +
X
wi
i∈Si∗ :t∗i =1
i∈F1 :zi∗ =1
≤
X
w(Si ) + w(Si∗ )zi∗∗
i∈F1 :zi∗ =1
≤ w(F1 ) +
X
(w(Si ) − wi )zi∗ ≤ B.
i∈F1
Next, we claim that the amount of saving by using S will be at least b(d1 + d2 ); and hence, cost I (S) ≤
2d2 + d1 − b(d1 + d2 ) = ad1 + (1 + a)d2 . Observe that setting ti := zi∗∗ for all i ∈ Si∗ gives a feasible
solution of LPSi∗ . Thus, the amount of saving by Si∗ is at least
X
X
(d1 (j) + d2 (j))zi∗∗ =
X
(d1 (j) + d2 (j))zi∗∗ .
j∈δ(i∗ )
i∈Si∗ j∈C:i2 (j)=i
Then the amount of saving by S is at least
X X
X X
(d1 (j) + d2 (j))zi∗ ≥
(d1 (j) + d2 (j))b = b(d1 + d2 ),
i∈F1 j∈δ(i)
i∈F1 j∈δ(i)
because setting zi := b for all i ∈ F1 also yields a feasible solution of LPstar-rounding .
Since the algorithm will choose the better solution between S and F1 , we have that
cost I (A) ≤ min{d1 , ad1 + (1 + a)d2 }.
Then the approximation ratio with respect to the cost of the bi-point solution is
d1
ad1 + (1 + a)d2
1
a + (1 + a)r
min
,
≤
max
min
,
ad1 + bd2
ad1 + bd2
a + (1 − a)r a + (1 − a)r
r≥0,a∈[4/5,1]∨a∈[0,1/5]
15
≤ .
11
Therefore, we have
cost I (A) ≤
15
30
(ad1 + bd2 ) ≤ OPTI < 2.73OPTI .
11
11
Finally, the theorem follows from Lemma 5.1.
5.3
Main case: a ∈ [1/5, 4/5]
In this section, we will apply dependent rounding in such a way that (i) each facility in F1 is open with
probability ≈ a, (ii) each facility in F2 is open with probability ≈ b, and (iii) the budget is not violated by
too much. To this end, for each i ∈ F1 , let Zi ∈ {0, 1} be a random indicator for the event “i is closed and
all facilities in Si are open.” (If Zi = 0 then i is open and all facilities in Si are closed.) Let t ≥ 1 be a
constant to be determined. The main algorithm is as follows.
We will now analyze the algorithm ROUND S TARS.
Claim 5.4. We have that |F 0 ∪ L0 | ≤ 2t and w(S \ (F 0 ∪ L0 )) ≤ B.
23
Algorithm 7 ROUND S TARS (t, F1 , F2 , a, b, B)
1:
2:
3:
4:
5:
6:
7:
8:
Initialize zi ← b and ci ← w(Si ) − wi for all i ∈ F1
Z ← SRDR(z, c, t)
Let F 0 :=S{i ∈ F1 : 0 < Zi < 1} be the set of ≤ t left-over fractional values of Z
Let L := i∈F 0 Si
for each i ∈ L: set ri ← Zk where k ∈ F 0 is the center of the star containing i.
R ← SRDR(r, w, t)
Let L0 := {i ∈ L : 0 <
RS
i < 1} be the set of ≤ t left-over fractional values of R
return S = F 0 ∪ L0 ∪
i∈F1 :Zi =1 Si ∪ {i ∈ L : Ri = 1}
Proof. The first claim is trivial due to construction of F 0 and L0 . By the property of SRDR, we have
X
X
X
wi ≤
wi Ri =
w(Sk )Zk .
i∈L:Ri =1
k∈F 0
i∈L
Therefore,
w(S \ (F 0 ∪ L0 )) =
X
w(Si ) +
i∈F1 :Zi =1
≤
X
X
wi
i∈L:Ri =1
w(Si ) +
X
w(Sk )Zk
k∈F 0
i∈F1 :Zi =1
=
X
w(Si )Zi
i∈F1
≤ w(F1 ) +
X
(w(Si ) − wi )Zi
i∈F1
= w(F1 ) +
X
(w(Si ) − wi )b
i∈F1
= aw(F1 ) + bw(F2 ) ≤ B,
where the penultimate equality follows from the fact that SRDR preserves the weighted sum of Zi ’s.
Now let us fix a client j and analyze the expected connection cost of j. Note that ROUND S TARS
guarantees that whenever a center i of some star is closed, all of the leaves in Si will be open. Thus, we will
use the following strategy to get j connected: if i2 (j) is open, then assign j to i2 (j). Else, if i1 (j) is open,
assign j to i1 (j). Finally, if both i1 (j) and i2 (j) are closed, the center of the star containing i2 (j), say i,
must be open and we connect j to i. For simplicity, let i1 , i2 denote the event that i1 (j) and i2 (j) are open
respectively.
Claim 5.5. The following bounds hold:
• Pr[ī2 ] ≤ 1 − b,
• Pr[ī1 ī2 ] ≤ (1 + 2/t)b(1 − b).
Proof. Conditioned on any vector Z returned by SRDR, if Zi ∈ {0, 1} then Pr[ī2 ] = 1 − Zi . If i ∈ F 0 ,
conditioned on any vector R returned by SRDR, we have Pr[i2 ] ≥ Ri2 (j) because i2 (j) will be open even
if Ri2 (j) is fractional. This implies that Pr[i2 ] ≥ E[Ri2 (j) ] = Zi or Pr[ī2 ] ≤ 1 − Zi . Summing over all
possible Z’s, we get
Pr[ī2 ] ≤ E[1 − Zi ] = 1 − zi = 1 − b.
24
Again, conditioned on any vector Z, we claim that
Pr[ī1 ī2 ] ≤ Zi1 (j) (1 − Zi ).
Indeed, if Zi1 (j) = 0 or Zi1 (j) ∈ (0, 1), then i1 (j) will be opened by the algorithm and Pr[ī1 ī2 ] = 0. If
Zi1 (j) = 1, then i1 (j) will be closed and Pr[ī1 ī2 ] = Pr[ī2 ] = 1 − Zi = Zi1 (j) (1 − Zi ) as in the above case.
Thus, summing over all possible Z’s, we obtain
Pr[ī1 ī2 ] ≤ E[Zi1 (j) (1 − Zi )]
≤ (zi1 (j) (1 − zi ))1−1/(t+1)
= (b(1 − b))1−1/(t+1) ,
where the second inequality is due to the near independence property of SRDR. Since a ∈ [1/5, 4/5], we
have b ∈ [1/5, 4/5] and b(1−b) ≥ 4/25. Hence, (b(1−b))−1/(t+1) ≤ (4/25)−1/(t+1) ≤ 1+2/t. Therefore,
we get
Pr[ī1 ī2 ] ≤ (1 + 2/t)b(1 − b).
Now the expected connection cost j can be bounded as follows.
E[ cost I (j)] ≤ Pr[i2 ]d2 (j) + Pr[i1 ī2 ]d1 (j) + Pr[ī1 ī2 ](2d2 (j) + d1 (j))
= Pr[i2 ]d2 (j) + Pr[ī2 ]d1 (j) + 2 Pr[ī1 ī2 ]d2 (j)
= d2 (j) + (d1 (j) − d2 (j)) Pr[ī2 ] + 2 Pr[ī1 ī2 ]d2 (j)
≤ d2 (j) + (d1 (j) − d2 (j))(1 − b) + (1 + 2/t)b(1 − b)(2d2 (j)).
Summing over all clients j, we get
E[ cost I (S)] ≤ d2 + (d1 − d2 )(1 − b) + (1 + 2/t)b(1 − b)(2d2 ).
√
Theorem 5.6. If a ∈ [1/5, 4/5], then there is a bi-factor (1 + 3 + )-approximation algorithm for KM.
Proof. First of all, we remove all facilities i for which wi = 0 from F 0 and L0 . Observe that Claim 5.4 still
holds. It means that we can write S = S 0 ∪ S 00 where S 00 = F 0 ∪ L0 , S 0 = S \ S 0 , |S 00 | ≤ 2t, w(S 0 ) ≤ B
and S 00 does not contain “big facilities”: S 00 ∪ S = ∅. Then, by Lemma 5.1, we have
w(S) ≤ 1 + 2t.
Now suppose we will take the better solution between S and F1 (which is feasible). In other words, let
us consider the set A := arg minZ∈{S,F1 } cost I (Z). We have
E[ cost I (A)] ≤ min{d1 , d2 + (d1 − d2 )(1 − b) + (1 + 2/t)b(1 − b)(2d2 )}
≤ (1 + 2/t) min{d1 , d2 + (d1 − d2 )(1 − b) + b(1 − b)(2d2 )}
d1
d2 + (d1 − d2 )(1 − b) + b(1 − b)(2d2 )
,
= (1 + 2/t)(ad1 + bd2 ) min
ad1 + bd2
ad1 + bd2
≤ (1 + 2/t) · (2OPTI ) · C0 ,
25
where
d2 + (d1 − d2 )(1 − b) + b(1 − b)(2d2 )
d1
,
C0 = min
ad1 + bd2
ad1 + bd2
d1
d2 + (d1 − d2 )(1 − b) + b(1 − b)(2d2 )
= min
,
(1 − b)d1 + bd2
(1 − b)d1 + bd2
r + (1 − r)(1 − b) + b(1 − b)(2r)
1
,
≤
max
min
1 − b + br
1 − b + br
r∈[0,1], b∈[1/5,4/5]
√
1+ 3
≤
.
2
Therefore, we get
E[ cost I (A)] ≤ (1 + 2/t) · (1 +
√
3) · OPTI
≤ (1 + 6/t) · OPTI .
In conclusion, for any γ > 0, setting t = 6/γ and = γ/(2t) = γ 2 /12 gives
(i) w(A) ≤ 1 + γ,
(ii) E[ cost I (A)] ≤ (1 + γ)OPTI0 .
6
Acknowledgment
We thank Jason Shuster at NASA for the proof of Proposition A.3.
References
[1] A. Ageev and M. Sviridenko. Pipage rounding: a new method of constructing algorithms with proven
performance guarantee. Journal of Combinatorial Optimization, 8(3):307–328, 2004.
[2] J. Byrka, T. Pensyl, B. Rybicki, J. Spoerhase, A. Srinivasan, and K. Trinh. An improved approximation
algorithm for knapsack median using sparsification. In Algorithms - ESA 2015 - 23rd Annual European
Symposium, Patras, Greece, September 14-16, 2015, Proceedings, pages 275–287, 2015.
[3] J. Byrka, T. Pensyl, B. Rybicki, A. Srinivasan, and K. Trinh. An improved approximation for kmedian, and positive correlation in budgeted optimization. In Proceedings of the Annual ACM-SIAM
Symposium on Discrete Algorithms, (SODA 2015), pages 737–756, 2015.
[4] G. Călinescu, C. Chekuri, M. Pál, and J. Vondrák. Maximizing a monotone submodular function
subject to a matroid constraint. SIAM J. Comput., 40(6):1740–1766, 2011.
[5] M. Charikar and S. Li. A dependent LP-rounding approach for the k-median problem. Automata,
Languages, and Programming (ICALP), pages 194–205, 2012.
[6] C. Chekuri, J. Vondrák, and R. Zenklusen. Dependent randomized rounding via exchange properties
of combinatorial structures. In 51th Annual IEEE Symposium on Foundations of Computer Science,
pages 575–584, 2010.
26
[7] C. Chekuri, J. Vondrák, and R. Zenklusen. Multi-budgeted matchings and matroid intersection via
dependent rounding. In Proceedings of the Twenty-Second Annual ACM-SIAM Symposium on Discrete
Algorithms, SODA 2011, San Francisco, California, USA, January 23-25, 2011, pages 1080–1097,
2011.
[8] D. Z. Chen, J. Li, H. Liang, and H. Wang. Matroid and Knapsack Center Problems. In Integer Programming and Combinatorial Optimization: 16th International Conference, IPCO 2013, Valparaı́so,
Chile, March 18-20, 2013. Proceedings, pages 110–122. Springer: Berlin, Heidelberg, 2013.
[9] R. Gandhi, S. Khuller, S. Parthasarathy, and A. Srinivasan. Dependent rounding and its applications to
approximation algorithms. Journal of the ACM, 53:324–360, 2006.
[10] D. S. Hochbaum and D. B. Shmoys. A unified approach to approximation algorithms for bottleneck
problems. J. ACM, 33(3):533–550, 1986.
[11] K. Jain, M. Mahdian, and A. Saberi. A new greedy approach for facility location problems. In STOC,
pages 731–740, 2002.
[12] S. Khuller, R. Pless, and Y. J. Sussmann. Fault tolerant k-center problems. Theor. Comput. Sci.,
242(1-2):237–245, 2000.
[13] R. Krishnaswamy, A. Kumar, V. Nagarajan, Y. Sabharwal, and B. Saha. The matroid median problem.
In Proceedings of the annual ACM-SIAM Symposium on Discrete Algorithms (SODA), pages 1117–
1130. SIAM, 2011.
[14] S. Li and O. Svensson. Approximating k-median via pseudo-approximation. In Symposium on Theory
of Computing Conference, STOC’13, Palo Alto, CA, USA, June 1-4, 2013, pages 901–910, 2013.
[15] Y. Peres, M. Singh, and N. K. Vishnoi. Random walks in polytopes and negative dependence. In
Innovations in Theoretical Computer Science (ITCS), 2017.
[16] A. Srinivasan. Distributions on level-sets with applications to approximation algorithms. In IEEE
Symposium on Foundations of Computer Science, pages 588–597, 2001.
[17] D. P. Williamson and D. B. Shmoys. The Design of Approximation Algorithms. Cambridge University
Press, 2011.
A
Negative correlation in symmetric randomized dependent rounding
The following proposition introduces notation and preliminary analysis used for proving the lower and upper
bounds on near-independence.
Proposition A.1. Let X = ROUND(x, a) for vectors x ∈ (0, 1)n and a ∈ (R \ {0})n . For S ⊆ [n] and
c ≥ 0, define
(
δ
Y p
Y p
i∈S
Λ :=
Xi ,
λ :=
xi ,
yi := ai xi
,
Ai := (1 + yi )p ,
Bi := (1 − yi )p .
0
i
∈
6
S
i∈S
i∈S
Then
1
E[Λ|(i, j)] := E[Λ|{i∗ , j ∗ } = {i, j}] = λ(Ai Bj + Bi Aj ).
2
(Note: The variables defined above are functions of S and/or p, but we omit this to avoid notational clutter.)
27
Proof. First suppose i, j ∈ S. Because X and x differ only in elements i∗ and j ∗ , we may express
h Xp Xp
i
j
E[Λ|{i∗ , j ∗ } = {i, j}] = E λ · pi · p (i, j)
xi xj
1
(xi + δ/ai )p (xj − δ/aj )p 1
(xi − δ/ai )p (xj + δ/aj )p
= λ·
·
+
λ
·
·
2
2
xpi
xpj
xpi
xpj
1
= λ (1 + yi )p (1 − yj )p + (1 − yi )p (1 + yj )p
(12)
2
1
= λ(Ai Bj + Bi Aj )
2
Lastly, observe equation (12) is still true even when i or j are not in S.
A.1
Upper bound on near-independence
Here we will prove the upper bound of Theorem 2.1. We start with a technical lemma. Note the binomial
coefficient kp is defined for real p ≥ 0 and integer k ≥ 0 to be:
Qk−1
(p − i)
p
:= i=0
.
k
k!
Proposition A.2. For all x ∈ [−1, 1], n ≥ 2, and p = 1 − n1 ,
n(1 + x)2p + n(1 − x)2p + 2(n − 2)(1 − x2 )p ≤ 4(n − 1).
Proof. First expand the LHS using the generalized binomial theorem:
n(1 + x)2p + n(1 − x)2p + 2(n − 2)(1 − x2 )p
X 2p
X 2p
X p
k
k k
=n
x +n
(−1) x + 2(n − 2)
(−1)k x2k
k
k
k
k≥0
k≥0
k≥0
X 2p
X p
=2n
x2` + 2(n − 2)
(−1)k x2k
2`
k
`≥0
k≥0
X 2p
p
k
=2
n
+ (n − 2)
(−1) x2k .
2k
k
(13)
k≥0
Now define fk (n) := n
2p
2k
+ (n − 2)
p
k
(−1)k for k = 0, 1, 2 . . . where p = 1 − 1/n. Then
f0 (n) = n · 1 + (n − 2) · 1 · 1 = 2(n − 1),
1
2p(2p − 1)
f1 (n) = n
− (n − 2)p = 2p (p − 1)n + 1 = 2p − n + 1 = 0.
2
n
n
n
We claim the following recurrence holds, which can be shown algebraically using n−k+1
k
k−1 = k :
p + 1
2p
2p
.
fk (n) = 1 −
fk−1 (n, p) − n
k
2k − 1 2(k − 1)
2p
2p
When k ≥ 2, n ≥ 2, and p = (1 − n1 ), we have (1 − p+1
k ) ∈ (0, 1), 2(k−1) ≥ 0, and n 2k−1 > 0. Thus the
recurrence implies fk (n) < 1 · (fk−1 (n) − 0 · 0) = fk−1 (n) < . . . < f1 (n) = 0. Continuing from (13),
X
n(1 + x)2p + n(1 − x)2p + 2(n − 2)(1 − x2 )p = 2
fk (n)x2k ≤ 2f0 (n) = 4(n − 1).
k≥0
28
Proposition 3.2. Let X = ROUND(x, a) for vectors x ∈ (0, 1)n and a ∈ (R \ {0})n , n ≥ 2. For any set
S ⊆ [n],
"
#
Y p
Y p
(14)
E
≤
xi
Xi
i∈S
i∈S
holds for p = 1 −
1
n.
Furthermore, if all weights in aS have the same sign, the inequality holds for p = 1.
Proof. Fix set S ∈ [n]. First consider the case that all weights in aS have the same sign. Then all values in
yS also have the same sign or are zero. Applying Proposition A.1 with p = 1, we see
1
1
E[Λ|(i, j)] = λ(Ai Bj + Bi Aj ) = λ (1 + yi )(1 − yj ) + (1 − yi )(1 + yj ) = λ(1 − yi yj ) ≤ λ.
2
2
Since this is true for all pairs (i, j), it holds that E[Λ] ≤ λ, which is equivalent to (14).
We now consider the general case. In the following summations, i, j ∈ [n].
X
E[Λ] =
Pr[{i∗ , j ∗ } = {i, j}] E[Λ|(i, j)]
i<j
X 1 1
=
n · λ(Ai Bj + Bi Aj )
2
2
i<j
1X
(Ai + Bi )(Aj + Bj ) − (Ai − Bi )(Aj − Bj )
2
2
i<j
X
2 X
X
2 X
λ
1
= n ·
(Ai + Bi ) −
(Ai + Bi )2 −
(Ai − Bi ) +
(Ai − Bi )2
4 2 2
i
i
i
i
X
X
X
λ
(Ai + Bi )2 −
(Ai + Bi )2 +
≤ n n
(Ai − Bi )2
(15)
8 2
i
i
i
λ X
n(A2i + Bi2 ) + 2(n − 2)Ai Bi .
E[Λ] = n
8 2 i
=
λ
·
n
2
To get (15) we applied the Cauchy-Schwarz inequality and the nonnegativity of squares. Now expand Ai
and Bi . By observing that yi ∈ [0, 1] and fixing p = 1 − n1 , we may apply Proposition A.2:
E[Λ] =
≤
A.2
8
λ X
n(1 + yi )2p + n(1 − yi )2p + 2(n − 2)n(1 − yi2 )2p
n
8
λ X
4(n − 1) =
n
2
2
i
i
λ
8
4n(n − 1) = λ.
n
2
Lower bound on near-independence
Here we will prove the lower bound of Theorem 2.1. We start with a technical lemma.
Proposition A.3. For all x ∈ [−1, 1], n ≥ 2, 1 ≤ s < n, and p ≥ 1 +
s−1
n−s
≥ 1,
(n − s) (1 + x)p + (1 − x)p + (s − 1)(1 − x2 )p ≥ 2n − s − 1.
29
(16)
Proof. Define
f (x) := (n − s) (1 + x)p + (1 − x)p + (s − 1)(1 − x2 )p − (2n − s − 1).
We wish to show f (x) ≥ 0. Observe f (0) = 2(n − s) + (s − 1) − 2n + s + 1 = 0. Since f is continuous,
it thus suffices to show f 0 (x) ≥ 0 for x ∈ [0, 1).
f 0 (x) = p(n − s)(1 + x)p−1 − p(n − s)(1 − x)p−1 − 2xp(s − 1)(1 − x2 )p−1 .
When x ∈ [0, 1), the quantity p(1 + x)p−1 (1 − x)p−1 > 0. Thus it is equivalent to show g(x) ≥ 0, where
g(x) :=
f 0 (x)
= (n − s)(1 − x)1−p − (n − s)(1 + x)1−p − 2x(s − 1).
p(1 − x)p−1 (1 + x)p−1
Again, since g(0) = 0 and g is continuous (for x ∈ [0, 1)), it suffices to show that g 0 (x) ≥ 0 for x ∈ [0, 1).
We do this using Jensen’s inequality, since x−p is convex when p ≥ 0.
g 0 (x) = −(1 − p)(n − s)(1 − x)−p − (1 − p)(n − s)(1 + x)−p − 2(s − 1)
= (p − 1)(n − s) (1 − x)−p + (1 + x)−p − 2(s − 1)
≥ (p − 1)(n − s) · 2 − 2(s − 1)
s − 1
(n − s) · 2 − 2(s − 1) = 0.
≥
n−s
Thanks to Jason Shuster at NASA for the above proof.
Proposition 3.4. Let X = ROUND(x, a) for vectors x ∈ (0, 1)n and a ∈ (R \ {0})n , n ≥ 2. For set
S ⊆ [n], define S1 := {i ∈ S | ai > 0} and S2 := {i ∈ S | ai < 0}. If |S1 |, |S2 | ≤ n − 1, then
h Y p i Y p
E
Xi
≥
xi
i∈S
i∈S
o
|S1 |−1 |S2 |−1
.
,
holds for p = 1 + max n−|S
1 | n−|S2 |
n
Proof. Using the notation of Proposition A.1, we need to show E[Λ] ≥ E[λ], or equivalently (since λ is
constant), E[ Λλ ] ≥ 1. We decompose E[ Λλ ] into a sum of expectations conditional on (i, j).
hΛi X h
i hΛ
i
i
1 X hΛ
=
Pr {i∗ , j ∗ } = {i, j} E
(i, j) = n
(i, j)
(17)
E
E
λ
λ
λ
2
i<j
i<j
We will partition the conditional expectations according to the signs of yi and yj .
Case yi yj < 0. Notice Ai > Bi when yi > 0 and Ai < Bi when yi < 0. Since yi and yj are opposite
signs: −(Ai −Bi )(Aj −Bj ) > 0. Also observe for x, y ≥ 2, it is true that xy = 2x+2y−4+(x−2)(y−2) ≥
2x + 2y − 4. By Bernoulli’s inequality, Ai + Bi = (1 + yi )p + (1 − yi )p ≥ 1 + pyi + 1 − pyi = 2. Therefore
(Ai + Bi )(Aj + Bj ) ≥ 2(Ai + Bi ) + 2(Aj + Bj ) − 4. Altogether we get
hΛ
i 1 1
E
(i, j) ∧ (yi yj < 0) = · λ(Ai Bj + Bi Aj )
λ
λ 2
1
= (Ai + Bi )(Aj + Bj ) − (Ai − Bi )(Aj − Bj )
4
1
1
> (Ai + Bi ) + (Aj + Bj ) − 1.
(18)
2
2
30
P
Now sum (18) over all pairs (i, j) ∈ S1 × S2 (i.e. those pairs such that yi yj < 0.) Define ȳ1 := s11 i∈S1 yi
P
and ȳ2 := s12 i∈S2 yi . Also define s1 := |S1 |, s2 := |S2 |. We observe the function xp is convex (for p ≥ 1)
and apply Jensen’s inequality to get a quantity in terms of ȳ1 and ȳ2 :
X
E
hΛ
i<j
yi yj <0
i X X 1
1
(i, j) >
(Ai + Bi ) + (Aj + Bj ) − 1
λ
2
2
i∈S1 j∈S2
s1 X
s2 X
(Ai + Bi ) +
(Ai + Bi ) − s1 s2
2
2
i∈S1
i∈S2
s2 X
s1 X
=
((1 + yi )p + (1 − yi )p ) +
((1 + yi )p + (1 − yi )p ) − s1 s2
2
2
=
i∈S1
i∈S2
s2
s1
≥ s1 ((1 + ȳ1 )p + (1 − ȳ1 )p ) + s2 ((1 + ȳ2 )p + (1 − ȳ2 )p ) − s1 s2
2
2
s1 s2
p
p
(1 + ȳ1 ) + (1 − ȳ1 ) + (1 + ȳ2 )p + (1 − ȳ2 )p − 2
=
2
(19)
Case yi yj > 0. Applying Jensen’s inequality for xp , we get
E
hΛ
i 1 1
(i, j) ∧ (yi yj > 0) = · λ(Ai Bj + Bi Aj )
λ
λ 2
1
1
= (1 + yi )p (1 − yj )p + (1 − yi )p (1 + yj )p
2
2
1
1
= (1 + yi − yj − yi yj )p + (1 − yi + yj − yi yj )p
2
2
≥ (1 − yi yj )p
(20)
Now sum (20) over all pairs
that is, where either i, j ∈ S1 or i, j ∈ S2 . Maclaurin’s
Pm(i, j) where yi yj > 0,P
m
2
inequality states that m1
y
y
≤
(1/n
i=1 yi ) . This together with Jensen’s gives:
( 2 ) 1≤i<j≤m i j
X
i<j
yi yj >0
E
hΛ
i
i
i
X hΛ
X hΛ
(i, j) =
(i, j) +
(i, j)
E
E
λ
λ
λ
i<j
i,j∈S1
≥
X
i<j
i,j∈S1
i<j
i,j∈S2
(1 − yi yj )p +
X
(1 − yi yj )p
i<j
i,j∈S2
p
p
s2
1 X
s1
1 X
≥
1 − s1
yi yj +
1 − s2
yi yj
2
2
2
2
i<j
i,j∈S1
s1
s2
2 p
≥
(1 − ȳ1 ) +
(1 − ȳ22 )p
2
2
i<j
i,j∈S2
(21)
Case yi yj = 0. Define S3 := [n] \ S. Then yi = 0 ⇐⇒ i ∈ S3 . When yi is 0, then Ai = Bi = 1, so
E[Λ|(i, j) ∧ (yi = 0)] = 12 λ(Aj + Bj ). If yi = yj = 0, then E[Λ|(i, j) ∧ (yi = yj = 0)] = λ. Again we
31
use Jensen’s:
hΛ
i X X hΛ
i X X hΛ
i
i
X
X hΛ
E
(i, j) =
E
(i, j) +
E
(i, j) +
E
(i, j)
λ
λ
λ
λ
i<j
yi yj =0
i∈S3 j∈S1
=
i<j
i,j∈S3
i∈S3 j∈S2
X X 1
X X 1
X
((1 + yj )p + (1 − yj )p ) +
((1 + yj )p + (1 − yj )p ) +
1
2
2
i∈S3 j∈S1
i∈S3 j∈S2
i<j
i,j∈S3
X s1
X s2
n − s1 − s2
p
p
p
p
≥
((1 + ȳ1 ) + (1 − ȳ1 ) ) +
((1 + ȳ2 ) + (1 − ȳ2 ) ) +
2
2
2
i∈S3
i∈S3
n−s1 −s2
=
s1 (1+ ȳ1 )p + (1− ȳ1 )p + s2 (1+ ȳ2 )p + (1− ȳ2 )p ) + n−s1 −s2 −1 .
2
(22)
Combining (17), (19), (21), and (22), we can bound E[ Λλ ]:
h i
i
i
i
hΛ
hΛ
hΛ
X
X
X
n
Λ
E
=
(i, j) +
(i, j) +
(i, j)
E
E
E
2
λ
λ
λ
λ
i<j
yi yj =0
i<j
yi yj =0
i<j
yi yj =0
s1 s2
s1
s2
(1+ ȳ1 )p + (1− ȳ1 )p + (1+ ȳ2 )p + (1− ȳ2 )p − 2 +
(1− ȳ12 )p +
(1− ȳ22 )p
2
2
2
n−s1 −s2
s1 (1 + ȳ1 )p + (1 − ȳ1 )p + s2 (1 + ȳ2 )p + (1 − ȳ2 )p ) + n−s1 −s2 −1
+
2
s
s1 (n − s1 )
1
p
p
(1 − ȳ12 )p
=
(1 + ȳ1 ) + (1 − ȳ1 ) +
2
2
s
s2 (n − s2 )
n−s1 −s2
2
2 p
p
p
+
(1 − ȳ2 ) − s1 s2 +
. (23)
(1 + ȳ2 ) + (1 − ȳ2 ) +
2
2
2
n
o
s1 −1 s2 −1
Finally, we’d like to apply Proposition A.3 separately for ȳ1 and ȳ2 . Fix p = 1 + max n−s
,
.
1 n−s2
Observe that the means ȳ1 , ȳ2 ∈ (0, 1), and recall that s1 , s2 ≤ n − 1 and n ≥ 2 are given. If we further
assume that s1 , s2 ≥ 1, then all the necessary conditions for Proposition A.3 hold and we may apply it twice
to (23) to get:
h i
s1
s2
n
Λ
n − s1 − s2
E
≥ (2n − s1 − 1) + (2n − s2 − 1) − s1 s2 +
2
λ
2
2
2
1
1
1
= (2n − 1)(s1 + s2 ) − (s1 + s2 )2 + (n − (s1 + s2 ))(n − 1 − (s1 + s2 ))
2
2
2
n(n − 1)
n
=
=
.
2
2
Else at least one of s1 and s2 is 0. If both are zero, the LHS of (23) becomes n2 . If just one is zero (s1 for
example), then we apply Proposition A.3 only once:
h i
s
n
Λ
s2 (n − s2 )
n − s2
2
p
p
2 p
E
≥
(1 + ȳ2 ) + (1 − ȳ2 ) +
(1 − ȳ2 ) +
.
2
2
2
λ
2
s2
1
n
≥ (2n − s2 − 1) + (n − s2 )(n − s2 − 1) =
.
2
2
2
≥
32
B
Omitted proofs for Section 4
P
For each j ∈ W , let us define the random variable µj := i∈Cj δi . We also define the random variable Uj
as follows:
1. We set Uj = 0 if j ∈
/ J.
2. Otherwise, if ij1 , ij2 ∈ Cj or ij1 , ij2 ∈
/ Cj , we set Uj = 0.
3. Otherwise, we set Uj = min(yij1 , yij2 ).
Observe that the random variables Uj are independent (as each only depends on the choices made within
Gj ). The random variables µj are highly interdependent.
Proposition B.1. For any j ∈ W , we have |µj | ≤ Uj .
Proof. If j ∈
/ J, then necessarily µj = 0 = Uj and we are done. So suppose j ∈ J. By construction and
our choice of δ, if ij1 , ij2 ∈ Cj or ij1 , ij2 ∈
/ Cj then µj = 0 = Uj .
Finally, observe that we must have |δij1 | ≤ yij1 as y ± δ ∈ [0, 1]n . Similarly, |δij2 | ≤ yij2 . As |δij1 | =
|δij2 | and either µj = δij1 or µj = δij2 , it follows that |µj | ≤ min(yij1 , yij2 ) as we have claimed.
Proposition B.2. For any j ∈ W , we have
E[Uj ] ≤ pj
2cj (1 − cj )
.
|Fj | − 1
Proof. First, observe that j ∈ J with probability pj , and this is independent of ij1 , ij2 .
Suppose we enumerate Cj = {1, . . . , l} and Gj − Cj = {l + 1, . . . , r} where r = |Gj |. As ij1 , ij2 are
chosen uniformly at random, we have
Pl
E[Uj ] = pj
u=1
Pr
v=l+1 min(yu , yv )
.
r
2
(24)
Now, we claim that for a fixed value of the sum cj = y1 + . . . + yl and yl+1 + . . . + yr = 1 − cj , the
RHS of (24) is maximized when y1 = y2 = . . . = yl = cj /l and yl+1 = . . . = yr = (1 − cj )/r. For,
suppose (WLOG) that l > 1 and y1 > y2 and we decrement y1 by some small x > 0 and increment y2 by
x. We choose x sufficiently small so that we have
y1 > yv > y2
y1 − x > yv > y2 + x
iff
P
for v = l + 1, . . . , r. Then E[Uj ] changes by v∈[l+1,r]:y1 >ye >y2 x > 0.
Thus,
Pl Pr
min(cj /l, (1 − cj )/(r − l))
l(r − l) min(cj /l, (1 − cj )/(r − l))
E[Uj ] ≤ pj u=1 v=l+1
= pj
r
r
2
2
Now, suppose that we ignore the constraint that l is an integer in [r] and relax this to the weaker condition
that l ∈ [0, r]. We then have
E[Uj ] ≤ pj
maxl∈[0,r] T (l)
r
where
2
T (l) = l(r − l) min(
cj 1 − cj
,
)
l r−l
We claim that the maximum value of T (l) can only occur when cj /l = (1 − cj )/(r − l). For, suppose
that cj /l < (1 − cj )/(r − l). Then, within an open ball of l, we have T (l) = (r − l)cj , and hence
33
T 0 (l) = −cj 6= 0. Similarly, if cj /l > (1 − cj )/(r − l), we have T (l) = l(1 − cj ) and T 0 (l) = −cj 6= 0.
Thus, we have that T (l) is maximized at l = cj r yielding T (l) ≤ rcj (1 − cj ) and hence
E[Uj ] ≤ pj
2cj (1 − cj )
.
r−1
Proposition B.3. Suppose we condition on all the state at the beginning of the dependent rounding round q.
Then
X E[Uj ]
.
E[S 0 ] ≤ S cosh
1 − cj
j∈W
Proof. Let us first condition on the random variable δ (which includes all the random choices up to but not
including line 9 in Algorithm 6). With probability 1/2 we change y to y + δ and with probability 1/2 we
change y to y − δ. Thus,
Y
Y
E[S 0 | δ] = 1/2
(1 − cj − µj ) + 1/2
(1 − cj + µj )
j∈W
j∈W
X Y
= 1/2
(−µj )
=
µj
X⊆W j∈X
|X| even
=S
X Y
X⊆W j∈X
|X| even
≤S
Y
X Y
(1 − cj ) + 1/2
j∈W −X
X⊆W j∈X
X Y
Y
X⊆W j∈X
Y
(µj )
(1 − cj )
j∈W −X
(1 − cj )
j∈W −X
µj
1 − cj
Y
(as S =
(1 − cj ))
j∈W
X Y |µj |
1 − cj
X⊆W j∈X
|X| even
≤S
X Y
X⊆W j∈X
|X| even
Uj
.
1 − cj
(by Proposition B.1)
Integrating over the randomness involved in J, ij1 , ij2 , δ gives:
X
Y Uj
E[S 0 ] ≤ S
E
1 − cj
X⊆W
|X| even
=S
j∈X
X Y E[Uj ]
1 − cj
(as Uk are independent)
X⊆W j∈X
|X| even
∞ X Y
∞
X
X
E[Uj ]
≤S
=S
1 − cj
v=0 X⊆W j∈X
|X|=2v
E[Uj ] 2v
j∈W 1−cj
P
(2v)!
v=0
34
X E[Uj ]
.
= S cosh
1 − cj
j∈W
Proposition B.4. Suppose that t ≥ 6m2 . Then
18m2
E[S 0 ] ≤ S + P
|frac(Gj )| − 1
j∈W
2
Proof. We have
X E[Uj ]
E[S 0 ] ≤ S cosh
(Proposition B.3)
1 − cj
j∈W
X
|frac(Gj )| − 1
2cj (1 − cj )
(Proposition B.2)
≤ S cosh
(3m) × P
×
(|frac(G
)|
−
1)(1
−
c
)
|frac(G
)|
−
1
j
j
j
j∈W
j∈W
X
6m
cj
= S cosh P
j∈W (|frac(Gj )| − 1) j∈W
"
!
#
X
−6m
−6m
≤ S cosh P
ln(1 − cj ) = S + S cosh P
ln S − S .
j∈W |frac(Gj )| − 1
j∈W |frac(Gj )| − 1
j∈W
Now let z =
defined by
6m
P
j∈W (|frac(Gj )|−1)
6m
t
. Observe that z ≤
≤ 1. Now, consider the function f : [0, 1] → R
f (s) = s cosh(−z ln s) − s.
Simple analysis shows that for z < 1, this achieves a maximum value at s =
2z 2
f (s) ≤
2
z+1
1
z
−1
1 − z2
Thus,
E[S 0 ] ≤ S + f (S) ≤ S + P
1/z
so that
≤ z 2 /2.
18m2
j∈W
1−z
1+z
|frac(Gj )| − 1
2 .
We are now ready to analyze multiple rounds of the dependent rounding process.
Proposition 4.13. Let S denote the value of the potential function at the beginning of the dependent rounding
process (immediately after line 1 of KDR), and let S 0 denote
the value of the potential function at the end
P
of all rounds (i.e. at line 10 of KDR, after terminating j∈W (|frac(Gj )| − 1) ≤ t. Then
E[S 0 ] ≤ S +
180m2
.
t
2
Proof. First, suppose that t ≤ 6m2 . Then RHS is at least equal to 180m
= 30. As S 0 ∈ [0, 1] with certainty,
6m2
this result holds vacuously. So we suppose for the remainder of the proof that t > 6m2 .
35
P
Let Sq denote the potential function after q rounds (so S0 = S) and let γq = j∈W (frac(|Gj |) − 1)
after q rounds (recall that the vector y, and consequently the fractional clusters frac(Gj ), change after each
round.) By Proposition B.4, conditional on the state at the beginning of round q, we have
E[Sq+1 | state at round q] ≤ Sq +
18m2
γq2
We also have that Sq+1 = Sq with certainty if γq ≤ t, so can write:
E[Sq+1 | state at round q] ≤ Sq +
18m2 [γq > t]
,
γq2
and thus, by iterated expectations, we have that, for all q ≥ 0,
q−1
X
18 E[γz > t]m2
E[Sq ] ≤ S +
.
γz2
z=0
We can write this equivalently as:
" q
#
∞
∞
X
X
X
1
1
2
E
[γz = u] ≤ S + 18m
E [#z : γz = u] .
u2
u2
2
E[Sq ] ≤ S + 18m
u=t+1
z=0
u=t+1
By Proposition 4.11, there is a probability of at least 1/10 that there is a rounded variable in any
given round, in which case γz decreases by at least one. We now claim that for any integer u, we have
E [#z : γz = u] ≤ 10. For, suppose we condition on that γz = u for the first time at z = z0 . Then, in each
subsequent round, there is a probability of at least 1/10 that γz is no longer equal to u. So the expected
number of rounds for which γz = u is dominated by a geometric random variable with probability 1/10,
and hence has mean ≤ 10.
Thus, for any q ≥ 0 we have:
2
E[Sq ] ≤ S + 18m
∞
∞
X
X
1
1
180m2
2
E
[#z
:
γ
=
u]
≤
S
+
180m
≤
S
+
z
u2
u2
t
u=t+1
u=t+1
With probability one, the dependent rounding process terminates at some finite (random) number of
rounds Q, that is, S 0 = SQ . Observe that SQ ∈ [0, 1] with certainty, and if q > Q we have that Sq = SQ .
Thus, for any q ≥ 0, we have SQ ≤ Sq + [Q > q]. Taking expectation, this implies that
E[SQ ] ≤ E[Sq ] + Pr[Q > q] ≤ S +
180m2
+ Pr[Q > q].
t
Now take the limit of both sides as q → ∞. Observe that Q is finite with probability one so that
Pr[Q > q] → 0, and hence
180m2
.
E[S 0 ] = E[SQ ] ≤ S +
t
36
| 8cs.DS
|
On the Underapproximation of Reach Sets of Abstract
Continuous-Time Systems
Ievgen Ivanov
Taras Shevchenko National University of Kyiv
[email protected]
We consider the problem of proving that each point in a given set of states (“target set”) can indeed be
reached by a given nondeterministic continuous-time dynamical system from some initial state. We
consider this problem for abstract continuous-time models that can be concretized as various kinds
of continuous and hybrid dynamical systems.
The approach to this problem proposed in this paper is based on finding a suitable superset S of
the target set which has the property that each partial trajectory of the system which lies entirely in S
either is defined as the initial time moment, or can be locally extended backward in time, or can be
locally modified in such a way that the resulting trajectory can be locally extended back in time.
This reformulation of the problem has a relatively simple logical expression and is convenient
for applying various local existence theorems and local dynamics analysis methods to proving reachability which makes it suitable for reasoning about the behavior of continuous and hybrid dynamical
systems in proof assistants such as Mizar, Isabelle, etc.
1
Introduction
Real-time embedded and cyber-physical systems (CPS) [4, 13] are important types of artificial systems
that combine software and hardware and interact closely with external devices and the physical environment. One of the important aspect of such systems is their dynamical behavior which depends on
the behavior of the physical environment with which they interact. Normally, the latter behavior of the
environment can be represented using the modeling notion of a global continuous (real) time, so on a certain level of abstraction it makes sense to model the dynamical behavior of an embedded/cyber-physical
system also in terms of a global continuous time. On other levels, more close to the software, one may
wish to focus on the discrete-continuous nature of this behavior and consider alternative time models that
are convenient in this case, like hybrid time (used in hybrid automata) [6], superdense time [13], nonglobal time models, etc. Still continuous-time models play an important role in the definitions of various
classes of discrete-continuous (hybrid) models [6] and special kinds of hybrid models admit reformulation in the form of certain kinds of well-known continuous-time models [6], one of the most general
among which are differential inclusions [2]. To illustrate this generality, ordinary differential equations
(x′ (t) = f (x(t),t)), differential equations with discontinuous right-hand side (x′ (t) = f (x(t),t), where
f is discontinuous, under various notions of solution like Carathéodory solution [5], Filippov solution
[5], etc.), implicit differential equations ( f (t, x(t), x′ (t)) = 0), differential inequalities (x′ (t) ≥ f (x(t),t)),
certain hybrid dynamical systems [6], switched systems [14], etc. can be reformulated [15, p. 312] in
the form of a differential inclusion [2] x′ (t) ∈ F(x(t),t), where F is a function from a subset of Rn+1
n
to 2R . (Note, however, that a differential inclusion admits different notions of solution, e.g. ordinary
solution [3, 5], classical solution [3], g-solution [3], etc., and in the case of the mentioned reformulation,
different notions of solutions for an initial model correspond to different kinds of solutions of the resulting inclusion). Despite their generality, differential inclusions cannot be a common mathematical ground
E. Ábrahám and S. Bogomolov (Eds.): 3rd International
Workshop on Symbolic and Numerical Methods for Reachability
Analysis (SNR 2017)
EPTCS 247, 2017, pp. 46–51, doi:10.4204/EPTCS.247.4
c I. Ivanov
This work is licensed under the
Creative Commons Attribution License.
I. Ivanov
47
suitable for describing all dynamical aspects of embedded systems/CPS. One of the reasons is that they
are concrete models that cannot encompass abstractions of various elements of a system in a convenient
way. E.g., they require a specific structure of the state space in which their solutions/trajectories take
values (Rn or more general topological vector spaces). If one uses a differential inclusion to model the
behavior of a whole CPS, one is forced to assume that the state space of the CPS is e.g. Rn , encode the
states of various (software, hardware) layers from which a CPS consists in real vectors, and encode the
description of the behavior of software in derivative constraints. Such an encoding may be used, if it is
necessary for automatic analysis/verification of a system, but it is arguably unsuitable for the purpose of
specifying system behavior and reasoning about or refining system behavior specification.
For such purposes abstract continuous-time dynamical models are desirable (that do not impose
restrictions on the state space), that admit refinement to concrete dynamical models (e.g. those that can
be described by state transition systems, differential equations, switched systems, etc.) and have certain
general analysis principles and property proof principles that can be concretized after refinement. Such
abstract dynamical models can be found, e.g. in various variants of the mathematical systems theory (a
list of references can be found in [11]).
An abstract dynamical system model usually describes a set of abstract trajectories which are mappings from a time scale to an abstract state space. The behavior of a system is defined by the set of its
possible trajectories. In some cases, a class of abstract models can be described by stating axiomatically
the assumed properties of trajectories of its systems.
Following this approach, in [9, 11] we defined a class of abstract dynamical system models which
we called Nondeterministic Complete Markovian Systems (NCMS) which can be seen as a subclass of
O. Hájek solution systems [7, 8].
A NCMS is a triple (T, Q, Tr), where T is a time scale (which is assumed to be the real interval
[0, +∞)), Q is a nonempty set (state space), and Tr ⊆ (T →Q)
˜
is a set of partial functions from T to
Q which are called trajectories. To be a NCMS, by the definition the set Tr must satisfy 3 rather nonrestrictive properties (be closed under proper restrictions, be Markovian, and be complete in the sense
that each nonempty chain of trajectories in the sense of the subtrajectory relation must have a supremum
in Tr, see [12] for the details). The mentioned Markovian property in this case is not formally related to
the Markov processes in the probability theory, but has a rather similar idea: informally, it states that at
any time moment, the possible future evolution of the state of the system depends on its current state, but
not on the way by which the system reached it.
One can also view NCMS as a generalization of state transition systems to continuous time, since the
properties of Tr stated in the definition of NCMS are satisfied for transition systems, when interpreted
in terms of discrete time and transition system runs. By themselves NCMS describe evolutions of the
state of a system (trajectories). If necessary, it is also possible to associate “observable traces” with
trajectories (which are quantities that evolve with trajectories or are simply pointwise projections of
trajectories). After this association one gets an analog of a labeled transition system (LTS) in continuous
time. We call this analog a labeled NCMS. This notion is defined in [12].
Many concrete dynamical models are instances of NCMS [11, 9] (which are also examples of
O. Hájek’s solution systems), including models the behavior of which is described by classical and
Carátheodory solutions of ordinary differential equations, ordinary solutions of differential inequalities
and inclusions, switched systems, various hybrid dynamical systems.
The main feature of NCMS (not available for general O. Hájek’s solution systems) is that they admit
reduction of certain forms of analysis of their global-in-time behavior to analysis of their local-in-time
behavior (which can be done using methods depending on the means used to specify this behavior –
differential equations, inclusions, etc.; in all such cases the set of well-known local-in-time analysis
48
On the Underapproximation of Reach Sets of Abstract Continuous-Time Systems
methods is quite large, including solution approximation methods with guaranteed error bounds, etc.).
One example of this reduction is the method of proving the existence of global (i.e. defined on T )
trajectories which is based on proving the existence of some trajectories defined in a neighborhood of
each time moment described in [9].
Like in the case of other types of dynamical systems, one can study the behavior of a NCMS by
analyzing its reach set. A t-reach set is defined as the set of states which can be reached by a NCMS at
the time moments in [0,t] by following trajectories that start at the time moment 0. Reach sets can have
complex structure, so for system behavior analysis it is usually more useful to consider their under- and
over-approximations by well-behaved sets. One can consider this as a computational problem (compute
under- and over-approximations of the t-reach set), or as a decision problem (check if a given candidate
set is an under- or over-approximation of the t-reach set).
In this paper we focus on the latter case, and, moreover, focus on methods of proving that its answer
is affirmative. This view is useful when one considers applying interactive proof assistants such as Mizar
system, Isabelle, Coq, etc. to the verification of cyber-physical systems.
In this context, an obvious approach to proving that a given set A is an overapproximation of a t-reach
set is to show that it contains an invariant set (i.e. a set of states I such that each trajectory that starts in I
remains forever in I) which includes the set of initial states of a NCMS.
Consider the problem of proving that a given set A is an underapproximation of a t-reach set of
a NCMS. Below we propose an approach to this problem based on finding a suitable superset S of A
which has the property which, informally, means that that each partial trajectory of the system which
lies entirely in S either is defined as the initial time moment, or can be locally extended backward in
time, or can be locally modified in such a way that the resulting trajectory can be locally extended back
in time. This condition has a relatively simple logical expression and is convenient for applying a wide
range of local existence theorems and local dynamics analysis methods such as linearization of system’s
dynamics, various series expansions, approximations, singularity analysis to proving reachability. This
is important in the case of using proof assistants for verifying bounds of t-reach sets of continuous- or
discrete-continuous models. For an example of formalization of NCMS and related results in the proof
assistant Isabelle, we refer to [1].
2
Preliminaries
We will use the following notation: N = {1, 2, 3, ...} is the set of natural numbers; R is the set of real
numbers, R+ is the set of nonnegative real numbers, f : A → B is a total function from a set A to B;
f : A→B
˜ is a partial function from A to B, 2A is the power set of a set A, f |A is the restriction of a function
f to a set A. For a function f : A→B
˜ we will use the notation f (x) ↓ ( f (x) ↑) to denote that f (x) is
defined, or, respectively, is undefined on the argument x.
We will not distinguish the notions of a function and a functional binary relation. When we write
that a function f : A→B
˜ is total or surjective, we mean that f is total on the set A specifically ( f (x) is
defined for all x ∈ A), or, respectively, is onto B (for each y ∈ B there exists x ∈ A such that y = f (x)).
For any f : A→B
˜ denote dom( f ) = {x | f (x) ↓}, range( f ) = {y |∃x ∈ dom( f ) y = f (x)}. For any
partial functions f , g the notation f (x) ∼
= g(x) will mean the strong equality: f (x) ↓ if and only if g(x) ↓,
and f (x) ↓ implies f (x) = g(x). We will denote by f ◦ g the functional composition: ( f ◦ g)(x) ∼
= f (g(x))
and by f |A the restriction of a function f to a set A, i.e. a function defined on dom( f ) ∩ A such that the
graph of f |A is a subset of the graph of f .
Denote by T the non-negative real time scale [0, +∞). We will assume that T is equipped with a
I. Ivanov
49
topology induced by the standard topology on R.
The symbols ¬, ∨, ∧, ⇒, ⇔ will denote the logical operations of negation, disjunction, conjunction,
implication, and equivalence respectively.
Let us denote by T the set of all intervals in T (connected subsets) which have the cardinality greater
than one. Let Q be a set (a state space) and Tr be some set of functions of the form s : A → Q, where
A ∈ T. We will call the elements of Tr (partial) trajectories.
Definition 1 ([9, 10]) A set of trajectories Tr is closed under proper restrictions (CPR), if s|A ∈ Tr for
each s ∈ Tr and A ∈ T such that A ⊆ dom(s).
Let us introduce the following notation: if f , g are partial functions, f ⊑ g means that the graph of f
is a subset of the graph of g, and f ⊏ g means that the graph of f is a proper subset of g.
Definition 2 Let s1 , s2 ∈ Tr be trajectories. Then s1 is called a subtrajectory of s2 , if s1 ⊑ s2 .
The pair (Tr, ⊑) is a possibly empty partially ordered set.
Definition 3 ([9, 10]) A CPR set of trajectories Tr is
(1) Markovian, if for each s1 , s2 ∈ Tr and t0 ∈ T such that t0 = sup dom(s1 ) = inf dom(s2 ), s1 (t0 ) ↓,
s2 (t0 ) ↓, and s1 (t0 ) = s2 (t0 ), the following function s belongs to Tr: s(t) = s1 (t), if t ∈ dom(s1 ) and
s(t) = s2 (t), if t ∈ dom(s2 ).
(2) complete, if each non-empty chain in (Tr, ⊑) has a supremum.
Definition 4 ([9, 10]) A nondeterministic complete Markovian system (NCMS) is a triple (T, Q, Tr),
where Q is a set (called the state space) and Tr (called the set of trajectories) is a set of functions
s : T →Q
˜ such that dom(s) ∈ T, which is CPR, complete, and Markovian.
Examples of representation of sets of trajectories of well-known continuous and discrete-continuous
dynamical models such as ordinary differential equations, differential inclusions, switched systems, etc.
as can be found in [12, 9, 7, 8].
3
Main Result
Definition 5 Let I1 , I2 ∈ T and s1 : I1 → Q, s2 : I2 → Q. Then s2 is called a backward extension of s1 , if
s1 ⊏ s2 and t2 ≤ t1 for each (t1 ,t2 ) ∈ dom(s1 ) × dom(s2 ).
Definition 6 Let s1 : I1 → Q where I1 ∈ T.
Then a τ -backward escape from s1 at time d, where d is a non-maximal element of dom(s1 ), is a
function s2 : [c, d] → Q for some c ∈ T , c < d such that τ = d − c and s2 (d) = s1 (d).
Definition 7 A trajectory s ∈ Tr of a NCMS (T, Q, Tr) is called
• initial, if s(0) ↓ (i.e. s is defined at the initial time moment 0);
• S-valued, where S is a set, if range(s) ⊆ S.
Definition 8 Let Σ = (T, Q, Tr) be a NCMS and t0 ∈ (0, +∞).
• The t0 -reach set of Σ is the set
{q ∈ Q | ∃s ∈ Tr0 ∃t ∈ dom(s) ∩ [0,t0 ] q = s(t)},
where Tr0 is the set of all initial trajectories of Σ (i.e. the set of states that are attained at some
time in [0,t0 ] by some initial trajectory).
50
On the Underapproximation of Reach Sets of Abstract Continuous-Time Systems
• An underapproximation of a t0 -reach set of Σ is a subset of the t0 -reach set of Σ.
• The t0 -right range set of Σ is the set
{q ∈ Q | ∃s ∈ Tr : dom(s) ⊆ [0,t0 ] ∧ max dom(s) ↓ ∧q = s(max dom(s))}
(i.e. the set of states that are attainted at the right end of some trajectory defined within [0,t0 ], but
not necessarily an initial trajectory).
• Σ is f -backward extensible, where f : R+ → R+ is a function of class K (i.e. a continuous, strictly
increasing function such that f (0) = 0), if for each s ∈ Tr at least one of the following holds:
a) s is an initial trajectory;
b) there exists s′ ∈ Tr such that s′ is a backward extension of s;
c) dom(s) has no minimum element and there exists s′ ∈ Tr, t ∈ T – a non-maximal element of
dom(s), and τ ≥ f (t − inf dom(s)) such that s′ is a τ -backward escape from s at time t.
• A sub-NCMS of Σ is a NCMS Σ = (T, Q′ , Tr′ ) such that Q ⊆ Q′ , Tr ⊆ Tr′ .
Theorem 1 Let f : R+ → R+ be a function of class K, Σ be a NCMS, t0 ∈ (0, +∞), A ⊆ Q.
Then A is an underapproximation of the t0 -reach set of Σ
if and only if A is a subset of the t0 -right range set of some f -backward extensible sub-NCMS of Σ.
Proof. We will use the terminology of [9, 12, 11].
“If” The idea is to apply the method of proving the existence of infinite-in-time trajectories of NCMS
described in [9], but in reversed time and to show that locally defined trajectories that end in A
can be continued backward all the way until the initial time moment using left (time revered)
extensibility measures [12]. Let A be a subset of the t0 -right range set of Σ′ = (T, Q′ , Tr′ ), where
Σ′ is a f -backward extensible sub-NCMS of Σ. Let f − (a, b) = a − f (a − b). From the conditions
of the f -backward extensible NCMS it follows that Σ′ satisfies the local backward extensibility
property and each left dead-end path in Σ′ is f − -escapable. Then by the theorem about the leftdead end path (a time reversal of the theorem about the right-dead end path as formulated in [12]),
each left dead-end path in Σ′ is strongly escapable.
Let q ∈ A. Then q = s(max dom(s)) for some s ∈ Tr′ such that dom(s) ⊆ [0,t0 ]. Denote b =
max dom(s). Then s ∈ Tr. If s is an initial trajectory or is a subtrajectory of an initial trajectory of
Σ′ , then q is attained by an initial trajectory of Σ at a time in [0,t0 ], so q is in the t0 -reach set of Σ.
Assume the contrary, i.e. s is not a subtrajectory of an initial trajectory of Σ′ . Suppose that there
exists a trajectory of Σ′ which is a backward extension of s. Then the completeness property of
NCMS implies that Tr′ contains a maximal (in the sense of the subtrajectory relation ⊑) backward
extension of s (among backward extensions of s in Tr′ ). Denote it as s′ . Then s′ cannot be an initial
trajectory, since otherwise s would be a subtrajectory of an initial trajectory of Σ′ . Also, s′ has no
backward extension in Tr′ . Since Σ′ is f -backward extensible, dom(s′ ) does not have a minimal
element. Then dom(s′ ) = (a, b] for some a ∈ T such that a < b. Then s′ is a left dead-end path
and it is strongly escapable. Moreover, q = s′ (b) and b ≤ t0 . This implies that q is attained by an
initial trajectory of Σ at some time in [0,t0 ], so q belongs to the t0 -reach set of Σ.
Suppose that there is no trajectory of Σ′ which is a backward extension of s. Also s is not an initial
trajectory. Since Σ′ is f -backward extensible, dom(s) does not have a minimal element. Then s is
a left dead-end path in Σ′ and it is strongly escapable. Moreover, q = s(b) and b ≤ t0 . This implies
that q is attained by an initial trajectory of Σ at some time in [0,t0 ], so q is in the t0 -reach set of Σ.
I. Ivanov
51
“Only if” Let A be a subset of the t0 -reach set of Σ. Let Tr0 be the set of initial trajectories of Σ,
Tr′ = {s′ : I → Q | I ∈ T ∧ ∃s ∈ Tr0 s′ = s|I }, and Σ′ = (T, Q, Tr′ ). It is easy to see that Tr′
is CPR, Markovian, and complete (in the sense of Definition 3), so Σ′ is a NCMS. Moreover, it
is a sub-NCMS of Σ, and is f -backward extensible, since each non-initial trajectory of Σ′ has a
backward extension in Tr′ . Moreover, for each initial trajectory s of Σ, s|[0,t0 ] is a trajectory with
the domain in [0,t0 ] and by the CPR property of the NCMS, its range is attainted at the right ends
of trajectories of Σ′ defined within [0,t0 ], so A is a subset of the t0 -right range set of Σ′ . Thus A is
a subset of the t0 -right range set of Σ′ , a f -backward extensible sub-NCMS of Σ.
An obvious way of obtaining a sub-NCMS of a NCMS Σ is restricting the state space of Σ. More
specifically, it is easy to check that if S ⊆ Q and TrS is the set of all S-valued trajectories of Σ, then
ΣS = (T, S, TrS ) is a sub-NCMS of Σ. This implies the above-mentioned approach to proving that a given
set A is an underapproximation of a t0 -reach set of a NCMS: to prove this it is sufficient to choose a
function f of class K and find/guess a superset S of A which has the property that A is a subset of the
t0 -right range set of ΣS and ΣS is f -backward extensible.
References
[1] https://code.google.com/archive/p/isabelle-behavioral-approach/downloads.
[2] Jean-Pierre Aubin (2009): Viability Theory. Springer Nature, doi:10.1007/978-0-8176-4910-4.
[3] A. Bacciotti (2005): Generalized solutions of differential inclusions and stability. Ital. J. Pure Appl. Math,
pp. 183–192.
[4] R. Baheti & H. Gill (2011): Cyber-physical systems. The Impact of Control Technology, pp. 161–166.
[5] A. F. Filippov (1988): Differential Equations with Discontinuous Righthand Sides. Springer Netherlands,
doi:10.1007/978-94-015-7793-9.
[6] R. Goebel, R. G. Sanfelice & A. Teel (2009): Hybrid dynamical systems. Control Systems, IEEE 29(2), pp.
28–93, doi:10.1109/MCS.2008.931718.
[7] O. Hájek (1967): Theory of processes, I. Czechoslovak Mathematical Journal 17, pp. 159–199.
[8] O. Hájek (1967): Theory of processes, II. Czechoslovak Mathematical Journal 17(3), pp. 372–398.
[9] Ievgen Ivanov (2013): A Criterion for Existence of Global-in-Time Trajectories of Non-deterministic Markovian Systems. In: ICT in Education, Research, and Industrial Applications, Springer, pp. 111–130,
doi:10.1007/978-3-642-35737-4 7.
[10] Ievgen Ivanov (2013): On Existence of Total Input-Output Pairs of Abstract Time Systems. In: Information
and Communication Technologies in Education, Research, and Industrial Applications, Springer, pp. 308–
331, doi:10.1007/978-3-319-03998-5 16.
[11] Ievgen Ivanov (2014): On Representations of Abstract Systems with Partial Inputs and Outputs. In: Lecture
Notes in Computer Science, Springer, pp. 104–123, doi:10.1007/978-3-319-06089-7 8.
[12] Ievgen Ivanov (2016): On Local Characterization of Global Timed Bisimulation for Abstract
Continuous-Time Systems. In: Coalgebraic Methods in Computer Science, Springer, pp. 216–234,
doi:10.1007/978-3-319-40370-0 13.
[13] E.A. Lee & S.A. Seshia (2017): Introduction to embedded systems: A cyber-physical systems approach,
Second Edition. MIT Press.
[14] Daniel Liberzon (2003):
Switching
doi:10.1007/978-1-4612-0017-8.
in
Systems
and
Control.
Birkhäuser
Boston,
[15] Dušan Repovš & Pavel V. Semenov (2013): Continuous Selections of Multivalued Mappings. In: Recent
Progress in General Topology III, Atlantis Press, pp. 711–749, doi:10.2991/978-94-6239-024-9 17.
| 3cs.SY
|
The final version of this paper has been published in “International Conference on Communication and Computing (ICC-2014)”
http://www.elsevierst.com/conference_book_download_chapter.php?cbid=86#chapter41
A Fuzzy MLP Approach for Non-linear Pattern Classification
Tirtharaj Dash, H.S. Behera
Department of Computer Science and Engineering
Veer Surendra Sai University of Technology (VSSUT), Burla, Odisha 768018, India
Email: [email protected], [email protected]
Abstract: In case of decision making problems, classification of pattern is a complex and crucial task. Pattern classification using
multilayer perceptron (MLP) trained with back propagation learning becomes much complex with increase in number of layers,
number of nodes and number of epochs and ultimate increases computational time [31]. In this paper, an attempt has been made
to use fuzzy MLP and its learning algorithm for pattern classification. The time and space complexities of the algorithm have
been analyzed. A training performance comparison has been carried out between MLP and the proposed fuzzy-MLP model by
considering six cases. Results are noted against different learning rates (α) ranging from 0 to 1. A new performance evaluation
factor ‗convergence gain‘ has been introduced. It is observed that the number of epochs drastically reduced and performance
increased compared to MLP. The average and minimum gain has been found to be 93% and 75% respectively. The best gain is
found to be 95% and is obtained by setting the learning rate to 0.55.
Keywords: classification, Fuzzy MLP, MLP, pattern, perceptron, UCI
1. Introduction
Pattern Classification has been one of the most frequently needed tasks in different fields of science and engineering
and in many other fields [1]. In real world, every object has certain set of attributes. These attributes help in keeping
that object to a certain group, called class. A particular group follows a certain pattern. A pattern classification
problem can be defined as a problem where an object is assigned into a predefined group or class based on a set of
observed attributes related to that object [1]. Let us take a set of examples to support this statement. We classify the
programming languages as high-level or low-level, object-oriented or object-based depending on a set of properties
possessed by these languages. Several problems in business, science, industry, and medicine can be treated as
classification problems. Examples include bankruptcy prediction [2,3], stock market forecasting [2,3], credit
scoring, medical diagnosis [4,5], medical data classification [4,5], power quality control and classification [6],
handwritten character recognition [7], signature verification [8,9], fingerprint recognition [10,11], speech
recognition [12] etc.
Finding solutions to such classification problems has been a crucial research area in the field of technology. Recent
works have been done using statistical techniques or using data mining techniques, such as neural computing.
Statistical methods use probability theory, decision theory [13] to classify a set of data; whereas, neural computing is
a technique which uses a neural network (NN) model for classifying the input data [14]. With increasing demands of
the problems, the NN approach for classifying pattern is becoming popular and new NN models have been
developed in the process. The reason behind this fact is that, the statistical methods use certain set of assumptions
and conditions to satisfy the solution. And these developed models cannot be applied directly to any such related
problems [1]. This limitation of statistical methods has been reduced by the use of NN techniques. The first and
foremost advantage of using NN techniques is that, NNs are nonlinear which let them adjust to the real world input
data without any explicit specification or external conditions. Secondly, they are universal functional approximation
models which can approximate any function with arbitrary accuracy [1,15,16]. NN techniques perform statistical
analysis of the processed data to support the second advantage [1].
Significant progress has been made in recent years in the area of pattern recognition and neuro-computing. Feedforward Multilayer Perceptrons (MLPs) are the most widely studied NN classifiers. A technique hybrid of fuzzy
rules and NN, called neurofuzzy has been applied to model some real world problems such as medical diagnosis
[17]. However, this technique has not been applied to pattern classification problem. A novel model was developed
by Kang and Brown [18] for classifying data. This is an unsupervised NN, called as adaptive function neural
Cite as: Dash, T., & Behera, H.S. (2014). A Fuzzy MLP Approach for Non-linear Pattern Classification. In
Proc: K.R. Venugopal, S.C. Lingareddy (eds.) International Conference on Communication and Computing
(ICC- 2014), Bangalore, India (June 12-14, 2014), Computer Networks and Security, 314-323.
The final version of this paper has been published in “International Conference on Communication and Computing (ICC-2014)”
http://www.elsevierst.com/conference_book_download_chapter.php?cbid=86#chapter41
network and has no hidden layer in its architecture. A few examples of data classification applied in various fields
have been listed below in the Table-1 with citation to corresponding works.
Table 1 Classification in real world problems
Application area
Bankruptcy prediction
Medical diagnosis
Handwritten signature recognition
Fingerprint recognition
System fault identification
Data classification
References
[19,20,21,22,23]
[4,5,17,24,25]
[8,9,26,27]
[10,28,29]
[30]
[31,32,33]
Although the concept of Fuzzy MLP has been proposed by Mitra et al. [34] for classification and rule generation
purposes, the model has not been applied successfully to pattern recognition or classification problems. This paper
focuses on development of a novel Fuzzy MLP model for pattern recognition and classification. Performance of this
model has been compared with that of the classical MLP model developed for the same problems. It should be noted
that both fuzzy-MLP and classical MLP nets uses the concept of back propagation for learning patterns. The issues
discussed in this paper can also apply to other neural network models.
The overall organization of this paper is as follows. A summary of preliminaries has been given in section 2 for
refreshing the concept of MLP. Section 3 elaborates the proposed fuzzy-MLP model. Section 4 gives the
experimental details and results followed by the conclusion section in section 5. A list of papers referred for this
work has been given in reference section.
2.
Preliminaries
2.1. The MLP architecture
The MLP is n-layer architecture, where n≥3. The layers are (i) input layer, (ii) output layers and (iii) hidden layer(s).
The number of hidden layers may be increased or decreased based on the problem for which the model is developed.
It is a feed-forward neural network, where non-linear elements, called neurons are arranged in layers. The data flow
from the input layer to the output layer via hidden layer(s) [35]. The MLP net learns through supervised learning
algorithms. A generalized delta rule (∆-rule) is used for this process which has been given in pseudo-code section in
the following section. The architecture of MLP is shown in Fig. 1. It should be noted that if there are n number of
neurons in input layer, there can be a maximum of 2n+1 number of neurons in hidden layers, i.e. m ≤ 2n+1.
The symbols used in the algorithms are given in the Table 2. It should be noted that, in this work, we have employed
binary sigmoid as the activation function which is defined as f(x) = 1/(1+exp(-x)). Following equations are used for
various updations mentioned in the above given piece of pseudo-code.
n
Z in
b1
X iVini
(1)
i 1
Where, Vini represents the incoming input-hidden layer weight to the considered node for which net input Zin is
being calculated. It should be noted that all the input, processing and output are carried out with decimal values
only.
m
yin
b2
Z iWi
(2)
i 1
W
{
b2
V
{
E
f ' ( y in ) Z i } (
E
f ' ( y in )
E
f ' ( y in ) W
Wold )
(3)
(4)
f ' ( Z in ) X } (
Vold )
(5)
Cite as: Dash, T., & Behera, H.S. (2014). A Fuzzy MLP Approach for Non-linear Pattern Classification. In
Proc: K.R. Venugopal, S.C. Lingareddy (eds.) International Conference on Communication and Computing
(ICC- 2014), Bangalore, India (June 12-14, 2014), Computer Networks and Security, 314-323.
The final version of this paper has been published in “International Conference on Communication and Computing (ICC-2014)”
http://www.elsevierst.com/conference_book_download_chapter.php?cbid=86#chapter41
b1
E
f ' ( y in ) W
f ' ( Z in )
(6)
Table 2 Symbols used in algorithms
Symbol
W
∆W
V
∆V
b1
∆b1
b2
∆b2
α
µ
f
Name and Purpose
Hidden-output layer weight (refer Fig. 1)
change in weight, W
Input-hidden layer weight (refer Fig. 1)
change in weight, V
bias value used in hidden layer
change in b1
bias value used in output layer
change in b2
learning rate
momentum factor
activation function
Fig. 1 MLP architecture
Pseudo-code for training MLP:
Initialize W, V, b1, b2, α, µ
Set mean-squared-error, MSE = 0;
while (termination criteria is not satisfied)
do
for each training pattern
for each hidden neuron
Calculate input to the hidden neuron (Zin) using Equation-1
Calculate activation of hidden neuron, Z = f(Zin) (refer Fig. 1)
end for
Calculate net input to the output neuron (yin) using Equation-2
Calculate activation of output neuron, Y = f(yin) (refer Fig. 1)
Compute the error, E = T–Y; where T is the corresponding target
/*Back propagate the error to update weights*/
Calculate ∆W using Equation-3
Update b2 using Equation-4
for each hidden neuron
Calculate ∆V using Equation-5
Update b1 using Equation-6
end for
end for
end while
3.
Fuzzy MLP for pattern classification
In this work, an attempt has been made to introduce the concept of fuzzy membership with the MLP learning
algorithm. The corresponding learning algorithm has been elaborated in the next section. We call the combined
network formed thereof as Fuzzy-MLP. The membership function used in this model is the S-shaped (S-shaped)
membership function or the spline-based membership function (MF) [36,37]. It should be noted that this type of MF
is considered in this work because of its simplicity and it normalizes the input value to a certain range. However,
Cite as: Dash, T., & Behera, H.S. (2014). A Fuzzy MLP Approach for Non-linear Pattern Classification. In
Proc: K.R. Venugopal, S.C. Lingareddy (eds.) International Conference on Communication and Computing
(ICC- 2014), Bangalore, India (June 12-14, 2014), Computer Networks and Security, 314-323.
The final version of this paper has been published in “International Conference on Communication and Computing (ICC-2014)”
http://www.elsevierst.com/conference_book_download_chapter.php?cbid=86#chapter41
other kind of MF can also be implemented for future researches in the area of pattern recognition and classification.
This section is divided into three subsections as follows. Section-3.1 gives a brief description of S-shaped MF used
in this work. Section-3.2 and 3.3 describes the proposed fuzzy-MLP architecture and corresponding learning
algorithm respectively. Theoretical analysis of the proposed algorithm has been done in section-3.4.
3.1. Spline-based membership function (S-shaped MF)
A Spline-based MF is defined in equation-7 below. Where a, b locate the extremes of sloped portion of the curve as
shown in Fig. 2.
0, x
f ( x, a , b )
a
x a
2
b a
1 2
1, x
2
,a
x b
b a
b
x
2
,
a b
2
a b
2
x
(7)
b
Fig. 2. S-shaped MF
3.2. Fuzzy-MLP network architecture
The proposed architecture is similar to that of the classical MLP architecture and has been shown in Fig.-3 below.
Fig. 3. Fuzzy-MLP architecture
3.3. Fuzzy-MLP learning pseudo-code
initialize W, V, b1, b2, α, µ, MSE=0
fuzzify the training input pattern using S-shaped MF
input the fuzzified values to the input layer neurons
while (termination criteria is not satisfied)
do
for each training pattern
for each hidden neuron
Calculate input to the hidden neuron (Zin)
Calculate activation of hidden neuron, Z = f(Zin) (refer Fig. 3)
end for
Calculate net input to the output neuron (yin)
Cite as: Dash, T., & Behera, H.S. (2014). A Fuzzy MLP Approach for Non-linear Pattern Classification. In
Proc: K.R. Venugopal, S.C. Lingareddy (eds.) International Conference on Communication and Computing
(ICC- 2014), Bangalore, India (June 12-14, 2014), Computer Networks and Security, 314-323.
The final version of this paper has been published in “International Conference on Communication and Computing (ICC-2014)”
http://www.elsevierst.com/conference_book_download_chapter.php?cbid=86#chapter41
Calculate activation of output neuron, Y = f(yin) (refer Fig. 3)
Y=Defuzzify the fuzzy output Y (µY)
Compute the error, E = T–Y; where T is the corresponding target
/*Back propagate the error to update weights
Calculate ∆W
Update b2
for each hidden neuron
Calculate ∆V
Update b1
end for
end for
update MSE = (MSE + E2)/n
end while
3.4. Theoretical analysis of the proposed algorithm
In this section, time and space complexity analysis of the proposed algorithm has been carried out and represented
briefly. In the proposed fuzzy-MLP model, number of input neurons is ‗n‘; number of hidden units is ‗m‘ and
number of output neuron is 1. Due to three layered architecture, the computation will be carried out in 4 major
phases: forward (2-phases) and back-propagation (2-phases). The first forward computation is from input to hidden
layer computation and it will occur in O(n×m) times. The next forward computation is from hidden to output layer
computation and will take O(m) times. Similarly, in the back propagation process, the steps output to hidden layer
and hidden to input layer computations will occur in O(m) and O(m×n) times respectively. So, the time complexity
of the algorithm can be calculated as follows.
Or,
T(m,n) = 2×O(n×m) + 2×O(m)
T(m,n) = O(n×m); as n >> 1
In the proposed algorithm, one extra data matrix will be created which will store the fuzzified patterns for processing
by the NN in the hidden layer. So, the overall space complexity is O(n×m). However, in the current era of
computers, space complexity is not a problem anymore, as memory is abundantly available for computation. So, the
result section is restricted to the time of processing only.
4.
Experimental Details
The proposed algorithm has been evaluated with six public domain datasets from the University of California at
Irvine (UCI) machine learning repository (http://www.ics.uci.edu/mlearn/MLRepository.html). The results have been
compared with that obtained by using MLP. This section is divided into following subsections (i) description of
dataset, (ii) experimental setup, and (iii) classification result.
4.1. Description of the datasets
IRIS: This is most popular dataset for pattern classification. It is based on multivariate characteristics of flower
plant species. Those are length and thickness of petals and sepals. The dataset contains three classes such as Iris
Setosa, Iris Versicolor and Iris Virginica of 50 instances each. The dataset contains 150 instances and 5 attributes (4
predicting and 1 target). All the attribute values are real.
ABALONE: The dataset contains physical characteristics of abalone such as sex, shell length, shell diameter,
height, whole weight, shucked weight, viscera weight, shell weight and number of rings in the shell. The first seven
characteristics determine number of rings in the shell and are treated as target class attribute in this work.
BREAST-CANCER-WISCONSIN (BCW): This dataset is obtained from University of Wisconsin Hospitals, USA
and contains attributes related to two major types of breast cancers; benign and malignant. The predicting attributes
Cite as: Dash, T., & Behera, H.S. (2014). A Fuzzy MLP Approach for Non-linear Pattern Classification. In
Proc: K.R. Venugopal, S.C. Lingareddy (eds.) International Conference on Communication and Computing
(ICC- 2014), Bangalore, India (June 12-14, 2014), Computer Networks and Security, 314-323.
The final version of this paper has been published in “International Conference on Communication and Computing (ICC-2014)”
http://www.elsevierst.com/conference_book_download_chapter.php?cbid=86#chapter41
are clump thickness, uniformity of cell size, shape, marginal adhesion, single epithelial cell size, bare nuclei, bland
chromatin, normal nucleoli, and mitoses.
GLASS: The predicting attributes in this dataset are refractive index (RI), measurements of Na, Mg, Al, Si, K,
Ca, Ba, Fe, which determine the quality of glass and for what purpose the glass is fit to be used. The types can be
building-windows-float-processed,
building-windows-non-float-processed,
vehicle-windows-float-processed,
vehicle-windows-non-float-processed, containers, tableware, and headlamps.
SOYBEAN: This is a small data set containing only 47 patterns but is having maximum number of attributes that
is 36. This makes the NN to expand in the input as well as the hidden layers.
WINE: This dataset was generated as the results of a chemical analysis of wines grown in the same region in
Italy but derived from three different cultivars. The analysis determined the quantities of 13 constituents found in
each of the three types of wines. A few attributes are alcohol, malic acid, ash, magnesium etc.
Table 3. Dataset properties
Dataset
Iris
Abalone
Breast-cancer-Wisconsin (BCW)
Glass
Soybean
Wine
Number of
patterns
150
4177
699
214
47
178
Number of attributes
(including target attribute)
5
9
10
10
36
14
Number of
target classes
3
29
2
7
4
3
4.2. Experimental setup
Both MLP and fuzzy-MLP algorithms are implemented using MATLAB R2010a environment installed in a
personal computer having Windows 32-bit XP professional operating system and 2GB of main memory. The
processor is Intel dual processor system (Intel Core2Duo) and each processor has equal computation speed of 2
GHz.
4.3. Experimental results
As mentioned earlier, the results obtained by employing both fuzzy-MLP and MLP algorithms are compared for
each datasets. In this paper, convergence of MSE is considered for performance evaluation of pattern classification
using both the algorithms. However, the learning parameters are varied from their minimum to maximum value and
the results are noted. Number of nodes in the hidden node has set to 3/2 of the number of nodes in input layer; i.e.
3×n/2. In this work, the learning rate (α) is the most crucial parameter which affects the learning process of the NN.
So, it is varied from 0 to 1 in the learning process and a result for each α is noted for each dataset as shown in next
subsections. The typical values of α those were chosen are 0.05, 0.10, 0.25, 0.40, 0.55, 0.70, 0.85 and 0.99. The
momentum factor (µ) is set to 0.50 for all the experiments. As the fuzzy-MLP is converging very much faster than
the MLP net, the number of epoch is set to 100. A table has been maintained for each datasets to show the minimum
MSE obtained by MLP and fuzzy-MLP. The table also shows the time consumed by both the algorithms for the
process of pattern classification. To check the efficiency of fuzzy-MLP algorithm over classical MLP model, a
factor called convergence gain (Cg) has been introduced and is defined in equation-8.
Cg
Min _ MSEMLP
Min _ MSEFuzzy
Min _ MSEMLP
MLP
(8)
4.3.1. Case-1 (IRIS dataset)
The training performance of the fuzzy-MLP and MLP algorithms are evaluated with Iris dataset. The result
summary has been given in Table 5. It should be noted that the simulation time presented in the table is an average
Cite as: Dash, T., & Behera, H.S. (2014). A Fuzzy MLP Approach for Non-linear Pattern Classification. In
Proc: K.R. Venugopal, S.C. Lingareddy (eds.) International Conference on Communication and Computing
(ICC- 2014), Bangalore, India (June 12-14, 2014), Computer Networks and Security, 314-323.
The final version of this paper has been published in “International Conference on Communication and Computing (ICC-2014)”
http://www.elsevierst.com/conference_book_download_chapter.php?cbid=86#chapter41
of 5 executions. As number of input attributes for Iris dataset is 4, the number of neurons in hidden layer is set to 6.
Results are shown for 100 epochs and µ is set to 0.50. It is clear from the above table that when α is set to 0.99, the
propose fuzzy-MLP algorithm is converging to a minimum error of 0.016 (approx.) within 100 epochs only where as
the MLP algorithm is converging only to 1.67 (approx.) with this number of epochs. The convergence gain is
approximately 99% when α is set to 0.99. However, the overall performance of the proposed algorithm is better than
the classical MLP algorithm when the convergence gain is considered. The simulation time of the proposed
algorithm is also lesser than that of MLP algorithm in almost all α settings. A plot has been given in Fig. 4 by taking
number of epoch in X-axis and MSE in Y-axis to illustrate the convergence of error in fuzzy-MLP algorithm for
each α. The Fig. also shows that α = 0.99 is the best value for pattern classification in Iris dataset. However, α can
also be set to 0.85 for which the result of convergence is approximately equal to that in α = 0.99.
4.3.2. Case-2 (ABALONE dataset)
The result summary for Abalone dataset has been given in Table 6. Here number of hidden units is equal to
12.Unlike results in Iris dataset, the minimum MSE obtained by MLP and fuzzy-MLP algorithms are having small
difference of 0.39. The gain also shows that the later one is not showing better efficiency as compared to that in the
Iris dataset. The plot given in Fig. 5 shows the convergence of error when α is set to the different values. It can be
seen from table-5 and the above given plot that the convergence of error is hardly depending on the learning rate for
the Abalone dataset. However, with α=0.05 and 0.10, the error is converging to its minimum after 15 epochs, where
as for other α values it‘s getting to its minimum within 10 epochs.
4.3.3. Case-3 (BCW dataset)
The result summary for BCW dataset has been given in Table 7. Here number of hidden units is equal to 14. The
fuzzy-MLP algorithm outperforms the classical MLP algorithm with a convergence gain factor of approximately
98%. As the BCW dataset is large, it can be proposed that the proposed algorithm can be applied to larger datasets in
real life scenarios. A plot has been given in Fig. 6 shows the convergence of error when α is set to the different
values.
4.3.4. Case-4 (GLASS dataset)
The result summary for Glass dataset has been given in Table 8. Here number of hidden units is set to 15, as the
number of input attributes is 10. It can be seen from the table that the maximum gain obtained is 99% when α is set
within a range of 0.25-0.99. The plot given in Fig. 7 shows the convergence of error when α is set to the different
values between 0 and 1.
4.3.5. Case-5 (SOYBEAN dataset)
The result summary for Soybean dataset has been given in Table 9. Here number of hidden units is equal to 51. It
can be noted that the Soybean dataset is the dataset is having maximum input characteristic, i.e. 34. And the average
gain obtained is approximately 95%. Therefore, the proposed algorithm can be applied to problems with larger
predicting attributes. The plot given in Fig. 8 shows the convergence of error when α is set to the different values. It
can be seen from Fig. 8 that, the MSE obtained by the employing proposed fuzzy-MLP algorithm are converging to
their corresponding minimum with different α after 25 epochs. The plots with α=0.05 and α=0.25 are not
converging within 100 epochs. However, observing the plots, the converging is inversely proportional to α. That
means the result with maximum α is converging with minimum epoch and result with minimum α is converging
with more number of epochs.
4.3.6. Case-6 (WINE dataset)
The result summary for Wine dataset has been given in Table 10. Here number of hidden units is equal to 18.
Observing the MSEs obtained by fuzzy-MLP algorithm with α set to 0.05 and 0.10 there is a difference of 0.27. The
plot given in Fig. 9 shows the convergence of error when α is set to the different values. The convergence obtained
with the Wine dataset is similar to that of Glass dataset where the MSE plot is gradually increasing after 40 epochs.
Table 5 Performance of fuzzy-MLP for Iris dataset
α
Minimum MSE
MLP
Fuzzy-
Convergence
Gain (Cg)
Simulation Time
(sec)
MLP
Fuzzy-
0.05
0.10
0.25
1.66686
1.66676
1.66670
MLP
0.07183
0.04323
0.03665
0.9569
0.9741
0.9780
62.82
62.77
56.06
MLP
55.82
55.79
50.55
Cite as: Dash, T., & Behera, H.S. (2014). A Fuzzy MLP Approach for Non-linear Pattern Classification. In
Proc: K.R. Venugopal, S.C. Lingareddy (eds.) International Conference on Communication and Computing
(ICC- 2014), Bangalore, India (June 12-14, 2014), Computer Networks and Security, 314-323.
The final version of this paper has been published in “International Conference on Communication and Computing (ICC-2014)”
http://www.elsevierst.com/conference_book_download_chapter.php?cbid=86#chapter41
0.40
0.55
0.70
0.85
0.99
1.66669
1.66668
1.66668
1.66667
1.66667
0.02718
0.02175
0.01873
0.01687
0.01566
0.9837
0.9870
0.9888
0.9899
0.9906
57.05
58.96
55.18
60.62
60.81
56.70
47.94
56.11
49.46
46.78
0.05
0.10
0.25
0.40
0.55
0.70
0.85
0.99
Table 6 Performance of fuzzy-MLP for Abalone dataset
Minimum MSE
α
MLP
0.05
0.10
0.25
0.40
0.55
0.70
0.85
0.99
0.5105
0.5024
0.4952
0.4894
0.4843
0.4815
0.4794
0.4780
FuzzyMLP
0.1045
0.1034
0.1035
0.1035
0.1035
0.1036
0.1037
0.10394
Convergence
Gain (Cg)
0.7952
0.7942
0.7909
0.7885
0.7862
0.7848
0.7836
0.7826
Simulation Time
(sec)
FuzzyMLP
MLP
1832.82 1844.7
838.68
885.39
866.88
863.60
846.67
841.34
842.05
838.90
841.62
840.23
834.07
847.16
824.41
843.72
Minimum MSE
FuzzyMLP
0.04232
0.04122
0.04128
0.04150
0.04169
0.04183
0.04195
0.04207
MLP
0.05
0.10
0.25
0.40
0.55
0.70
0.85
0.99
Convergence
Gain (Cg)
3.75824
3.75823
3.75823
3.75822
3.75822
3.75822
3.75822
3.75822
0.9887
0.9890
0.9890
0.9890
0.9889
0.9889
0.9888
0.9888
Minimum MSE
α
Minimum MSE
MLP
Convergence
Gain (Cg)
Fuzzy-
MLP
53.37
48.71
41.70
41.64
42.81
45.66
43.78
41.59
44.92
44.53
43.56
43.81
42.39
42.30
44.64
42.69
0.05
0.10
0.25
0.40
0.55
0.70
0.85
0.99
Simulation Time
(sec)
FuzzyMLP
MLP
152.17
156.42
154.13
158.57
156.24
149.07
148.39
151.29
156.81
150.67
147.51
147.36
149.75
148.83
153.26
153.62
Convergence
Gain (Cg)
FuzzyMLP
0.35186
0.35183
0.35131
0.10119
0.03035
0.20958
0.02967
0.02975
MLP
4.31924
4.31922
4.31902
4.31918
4.31918
4.31917
4.31917
4.31916
0.9185
0.9185
0.9187
0.9766
0.9930
0.9515
0.9931
0.9931
Simulation Time
(sec)
FuzzyMLP
MLP
10.68
11.65
12.06
12.08
11.09
11.55
10.99
9.43
9.61
9.20
9.38
9.83
10.17
10.19
10.38
10.16
Table 10 Performance of fuzzy-MLP for Wine dataset
Minimum MSE
α
0.05
0.10
0.25
0.40
0.55
0.70
0.85
0.99
Simulation Time
(sec)
MLP
Fuzzy-
Convergence
Gain (Cg)
FuzzyMLP
0.36559
0.08701
0.06913
0.06440
0.06271
0.06352
0.06679
0.06226
MLP
Table 8 Performance of fuzzy-MLP for Glass dataset
α
0.9840
0.9874
0.9902
0.9908
0.9913
0.9916
0.9918
0.9920
Table 9 Performance of fuzzy-MLP for Soybean dataset
Table 7 Performance of fuzzy-MLP for BCW dataset
α
MLP
0.12105
0.09537
0.07441
0.06987
0.06626
0.06396
0.06217
0.06067
7.57481
7.57479
7.57478
7.57476
7.57474
7.57472
7.57471
7.57471
1.47754
1.47754
1.47753
1.47753
1.47753
1.47753
1.47753
1.47753
0.7526
0.9411
0.9532
0.9564
0.9576
0.9570
0.9548
0.9579
Simulation Time
(sec)
FuzzyMLP
MLP
37.71
36.65
37.73
42.44
37.40
35.42
35.91
40.93
35.34
36.01
36.99
36.90
39.09
39.30
39.81
40.36
Convergence plot for Abalone Dataset
Convergence plot for Iris Dataset
0.45
0.4
Mean-squared-error
0.35
0.3
0.25
0.2
0.15
alpha=0.05
alpha=0.10
alpha=0.25
alpha=0.40
alpha=0.55
alpha=0.70
alpha=0.85
alpha=0.99
0.4
0.35
Mean-squared-error
alpha=0.05
alpha=0.10
alpha=0.25
alpha=0.40
alpha=0.55
alpha=0.70
alpha=0.85
alpha=0.99
0.3
0.25
0.2
0.1
0.15
0.05
0
0
10
20
30
40
50
60
70
80
90
Epoch
Fig.4. Convergence plot obtained by employing Fuzzy
MLP algorithm for Iris dataset
100
0.1
0
10
20
30
40
50
60
70
80
90
Epoch
Fig. 5. Convergence plot obtained by employing fuzzyMLP algorithm for Abalone dataset
Cite as: Dash, T., & Behera, H.S. (2014). A Fuzzy MLP Approach for Non-linear Pattern Classification. In
Proc: K.R. Venugopal, S.C. Lingareddy (eds.) International Conference on Communication and Computing
(ICC- 2014), Bangalore, India (June 12-14, 2014), Computer Networks and Security, 314-323.
100
The final version of this paper has been published in “International Conference on Communication and Computing (ICC-2014)”
http://www.elsevierst.com/conference_book_download_chapter.php?cbid=86#chapter41
Convergence plot for BCW Dataset
Convergence plot for Soybean Dataset
0.7
alpha=0.05
alpha=0.10
alpha=0.25
alpha=0.40
alpha=0.55
alpha=0.70
alpha=0.85
alpha=0.99
Mean-squared-error
0.5
0.25
0.4
0.3
0.2
0.15
0.2
0.1
0.1
0.05
0
0
10
20
30
40
50
60
70
80
90
alpha=0.05
alpha=0.10
alpha=0.25
alpha=0.40
alpha=0.55
alpha=0.70
alpha=0.85
alpha=0.99
0.3
Mean-squared-error
0.6
0.35
0
0
100
10
20
30
40
50
60
70
80
90
100
Epoch
Epoch
Fig. 8. Convergence plot obtained by employing fuzzyMLP algorithm for Soybean dataset
Fig.6. Convergence plot obtained by employing fuzzyMLP algorithm for BCW dataset
Convergence plot for Wine Dataset
Convergence plot for Glass Dataset
0.7
0.6
Mean-squared-error
0.5
0.4
0.3
0.4
0.35
Mean-squared-error
alpha=0.05
alpha=0.10
alpha=0.25
alpha=0.40
alpha=0.55
alpha=0.70
alpha=0.85
alpha=0.99
alpha=0.05
alpha=0.10
alpha=0.25
alpha=0.40
alpha=0.55
alpha=0.70
alpha=0.85
alpha=0.99
0.3
0.25
0.2
0.15
0.2
0.1
0.1
0.05
0
0
10
20
30
40
50
60
70
80
90
Epoch
Fig. 7. Convergence plot obtained by employing fuzzyMLP algorithm for Glass dataset
100
0
0
10
20
30
40
50
60
70
80
90
Epoch
Fig. 9. Convergence plot obtained by employing fuzzyMLP algorithm for Wine dataset
For all the datasets, the proposed algorithm results a higher gain when compared with the MLP model. However, as
the MSE varies with α value, it will be interesting to analyze a plot for the gain against each α. A plot has been given
in Fig. 10 by α-value in X-axis and average gain (average Cg) in Y-axis. It should be noted that average gain against
an α is obtained by calculating the average of all the gains obtained with the six datasets used in this work. Fig. 10
clearly reveals that best gain, which is close to 95%, is obtained when the parameter α is tuned to 0.55, 0.85 and
0.99. So, it should be noted that for any datasets including the six UCI datasets used in this work, the proposed NN
model will achieve a higher convergence gain. However, it will also be important to check if the simulation time is
affected by α. Therefore, a plot has been given in Fig. 11 to show a comparative characteristic of average simulation
time for different α values. As mentioned in previous section, the simulation time is obtained by calculating mean of
simulation time obtained with five consecutive executions to make the result error free. In this plot, α has been taken
in X-axis and average simulation time is shown in Y-axis.
Cite as: Dash, T., & Behera, H.S. (2014). A Fuzzy MLP Approach for Non-linear Pattern Classification. In
Proc: K.R. Venugopal, S.C. Lingareddy (eds.) International Conference on Communication and Computing
(ICC- 2014), Bangalore, India (June 12-14, 2014), Computer Networks and Security, 314-323.
100
The final version of this paper has been published in “International Conference on Communication and Computing (ICC-2014)”
http://www.elsevierst.com/conference_book_download_chapter.php?cbid=86#chapter41
0.96
220
0.95
215
210
0.94
Avg. simulation time (seconds)
Average Convergence Gain (Cg)
MLP
Fuzzy-MLP
0.93
0.92
0.91
205
200
195
190
0.9
185
0.89
0
0.1
0.2
0.3
0.4
0.5
Learning rate ( )
0.6
0.7
0.8
0.9
1
Fig. 10. Average gain obtained against different
learning rate (α)
180
0.1
0.2
0.3
0.4
0.5
0.6
Learning rate ( )
0.7
0.8
0.9
Fig. 11. Average simulation time obtained against
different learning rate (α)
In the previous discussion, it is seen that when α is tuned to 0.55, 0.85 or 0.99, the proposed model obtains a
maximum of 95% convergence gain. So, it will be wise to check the simulation time against these three α values
only. Fig. 11 shows, for α=0.55, the proposed model is trained within 190 seconds whereas the MLP model is taking
more than 190 seconds for the purpose. For rest two α values, the simulation is approximately equal to that obtained
by using the MLP model and it is more than that obtained by setting α to 0.55. Therefore, it could be concluded that
the best α is 0.55 for all the future tests.
4.3.7. Statistical analysis of result (ANOVA)
ANOVA test on the results has been conducted carefully to see if there is any difference between the groups on the
resulted data. The t-test result using ANOVA has been given in Table-11 below. It should be noted that the test has
been carried out using the SPSS software package [38].
Table 11 t-test on the result data
t-value
Mean
Difference
95% Confidence Interval
of the Difference
Lower
Upper
MSEMLP
9.285
3.24734596
2.5433497
3.9513422
MSEFuzzy-MLP
6.970
.09052234
.0643786
.1166661
Cg
81.684
.9387638
.915630
.961897
TMSE
3.947
214.68277
105.2108
324.1548
TFuzzy-MSE
3.925
215.65191
105.0514
326.2524
Variables
5.
Conclusion
The fuzzy-MLP for pattern classification has been developed. The input patterns are fuzzified by using spline (Sshaped) membership function and then input to the MLP model. The results obtained shows that, the proposed
model converges to its minimum MSE within 100 epochs and achieves a convergence gain of 93%. The proposed
algorithm outperforms MLP for all the six UCI datasets used in this work. As future work, it will be advantageous to
optimize the network with an optimization algorithm.
References
1.
2.
G. P. Zhang, ―Neural Networks for Classification: A Survey‖, IEEE Transactions on Systems, Man, and Cybernetics—Part
C: Applications and Reviews, Vol. 30, No. 4, November 2000, pp. 451-462.
L. Liu, X. Peng, ―An Improved Diagonal Recurrent Neural Network and Its Application in Nonlinear Time Series
Forecasting‖, Information Engineering Letters, ISSN: 2160-4114 Volume 2, Number 1, March, 2012, pp. 18-25.
Cite as: Dash, T., & Behera, H.S. (2014). A Fuzzy MLP Approach for Non-linear Pattern Classification. In
Proc: K.R. Venugopal, S.C. Lingareddy (eds.) International Conference on Communication and Computing
(ICC- 2014), Bangalore, India (June 12-14, 2014), Computer Networks and Security, 314-323.
1
The final version of this paper has been published in “International Conference on Communication and Computing (ICC-2014)”
http://www.elsevierst.com/conference_book_download_chapter.php?cbid=86#chapter41
3.
4.
5.
6.
7.
8.
9.
10.
11.
12.
13.
14.
15.
16.
17.
18.
19.
20.
21.
22.
23.
24.
25.
26.
27.
28.
29.
30.
31.
32.
33.
T. Quah, B. Srinivasan, ―Improving returns on stock investment through neural network selection‖, Expert Systems with
Applications 17 (1999) 295–301.
M. Karabataka, M. C. Ince, ―An expert system for detection of breast cancer based on association rules and neural network‖,
Expert Systems with Applications 36 (2009) 3465–3469.
W. Sibanda, P. Pretorious, ―Novel Application of Multi-Layer Perceptrons (MLP) Neural Networks to Model HIV in South
Africa using Seroprevalence Data from Antenatal Clinics‖, International Journal of Computer Applications (0975 – 8887)
Volume 35– No.5, December 2011, pp. 26-31.
M. Uyara, S. Y., M. T. Gencoglu, ―An expert system based on S-transform and neural network for automatic classification
of power quality disturbances‖, Expert Systems with Applications 36 (2009) 5962–5975.
C. L. Liu, F. Yin, D. H. Wang, Q. F. Wang, ―Online and offline handwritten Chinese character recognition: benchmarking
on new databases‖, Pattern Recognition, 46(1), 155-162, 2013.
M. Quraishi, A. Das, S. Roy, ―A novel signature verification and authentication system using image transformations and
Artificial Neural Network‖, In Computer and Information Technology (WCCIT), 2013 World Congress on (pp. 1-6). IEEE.
R. Kumar, J. D. Sharma, B. Chanda, ―Writer-independent off-line signature verification using surroundedness feature‖,
Pattern Recognition Letters, 33(3), 301-308, 2012.
T. Dash, T. Nayak, S. Chattopadhyay, F. A. Rabhi, ―A Fingerprint Verification Tool using Adaptive Resonance Theory
Nets‖, American Journal of Biomedical Engineering 2013, 3(2): 31-40; doi: 10.5923/j.ajbe.20130302.01.
J.C. Yang, D.S. Park, ―A fingerprint verification algorithm based on tessellated invariant moment features‖,
Neurocomputing, 71(2008), 1939-1946.
C. Kurian, K. Balakrishnan, ―Continuous speech recognition system for Malayalam language using PLP cepstral
coefficient,‖ Journal of Computing and Business Research, Vol.3, Issue.1, 2012.
P. O. Duda, P. E. Hart, ―Pattern Classification and Scene Analysis‖, New York: Wiley, 1973.
R. P. Lippmann, ―Pattern classification using neural networks,‖ IEEE Commun. Mag., pp. 47–64, Nov. 1989.
K. Hornik, ―Approximation capabilities of multilayer feedforward networks‖, Neural Networks, vol. 4, pp. 251–257, 1991.
G. Cybenko, ―Approximation by superpositions of a sigmoidal function‖, Math. Contr. Signals Syst., vol. 2, pp. 303–314,
1989.
S. Chattopadhyay, ―Neurofuzzy models to automate the grading of old - age depression‖, Expert Systems; doi:
10.1111/exsy.12000, 2012.
M. Kang, D.P. Brown, ―A modern learning adaptive functional neural network applied to handwritten digit
recognition‖, Information Sciences 178(2008), 3802–3812.
G. Zhang, M. Y. Hu, E. B. Patuwo, D. Indro, ―Artificial neural net-works in bankruptcy prediction: General framework and
cross-validation analysis‖, Eur. J. Oper. Res., vol. 116, pp. 16–32, 1999.
R. L. Wilson, R. Sharda, ―Bankruptcy prediction using neural networks‖, Decision Support Syst., vol. 11, pp. 545–557,
1994.
M. Leshno, Y. Spector, ―Neural network prediction analysis: The bankruptcy case‖, Neurocomputing, vol. 10, pp. 125–147,
1996.
R. C. Lacher, P. K. Coats, S. C. Sharma, L. F. Fant, ―A neural net-work for classifying the financial health of a firm‖, Eur. J.
Oper. Res., vol. 85, pp. 53–65, 1995.
O. Akbilgic, H. Bozdogan, M.E. Balaban, ―A novel Hybrid RBF Neural Networks model as a forecaster‖, Statistics and
Computing (2013). DOI 10.1007/s11222-013-9375-7
A. Suhasini, S. Palanivel, V. Ramalingam, ―Multimodel decision support system for psychiatry problem‖, Expert Systems
with Applications 38(2011), 4990–4997.
S-C YU, U.H. Lin, ―Applications of fuzzy theory on health care: an example of depression disorder classification based on
FCM‖, WSES Transactions of Information Science and Applications 5(2008), 31–36.
R. Bajaj, S. Chaudhury, ―Signature verification using multiple neural classifiers‖, Pattern Recognition 3 (1), 1–7.
J. P. Drouhard, R. Sabourin, M. Godbout, ―A neural network approach to offline signature verification using direction pdf‖,
Pattern Recognition 29 (3), 415–424.
K. Mali, S. Bhattacharya, ―Fingerprint Recognition system based on Global and local structures‖, International Journal of
Computer Science and Engineering (2011), 3(1), 161-172.
S.K. Sahu, S. Sahani, P.K. Jena, S. Chattopadhyay, ―Fingerprint Identification System using Tree based Matching,‖
International Journal of Computer Applications, 53(2012), 11-16.
J. C. Hoskins, K. M. Kaliyur, and D. M. Himmelblau, ―Incipient fault detection and diagnosis using artificial neural
networks,‖ in Proc. Int. Joint Conf. Neural Networks, 1990, pp. 81–86.
S. Dehuria, R. Roy, S-B Cho, A. Ghosh, ―An improved swarm optimized functional link artificial neural network (ISOFLANN) for classification‖, The Journal of Systems and Software 85 (2012) 1333– 1345.
P.G. Espejo, S. Ventura, F. Herrera, ―A Survey on the Application of Genetic Programming to Classification‖, IEEE
Transactions on Systems, Man, And Cybernetics—Part C: Applications And Reviews, Vol. 40, No. 2, March 2010, pp. 121144.
S. Marinai, M. Gori, G. Soda, ―Artificial Neural Networks for Document Analysis and Recognition‖, IEEE Transactions on
Pattern Analysis And Machine Intelligence, Vol. 27, No. 1, January 2005, pp.23-35.
Cite as: Dash, T., & Behera, H.S. (2014). A Fuzzy MLP Approach for Non-linear Pattern Classification. In
Proc: K.R. Venugopal, S.C. Lingareddy (eds.) International Conference on Communication and Computing
(ICC- 2014), Bangalore, India (June 12-14, 2014), Computer Networks and Security, 314-323.
The final version of this paper has been published in “International Conference on Communication and Computing (ICC-2014)”
http://www.elsevierst.com/conference_book_download_chapter.php?cbid=86#chapter41
34. S. Mitra, R. K. De, S. K. Pal, ―Knowledge-Based Fuzzy MLP for Classification and Rule Generation‖, IEEE Transactions
on Neural Networks, Vol. 8, No. 6, November 1997, pp. 1338-1350.
35. S.N. Sivanandan, S.N. Deepa, ―Principle of Soft Computing (2nd Edition)‖, WILEY-INDIA, ISBN - 978- 81- 265 - 2741- 0,
2011.
36. J. Bezdek, ―Pattern recognition with fuzzy objective function algorithms‖, Plenum Press, New York, 1981.
37. T.J. Ross, ―Fuzzy Logic with Engineering Applications‖, 2nd Ed., John Wiley and Sons, Ltd, 2004, ISSN. 0-470-86075-8.
38. S.J. Coakes, S. Lyndall, ―SPSS: Analysis without anguish using SPSS version 14.0 for Windows‖, John Wiley & Sons, Inc.,
2009.
Cite as: Dash, T., & Behera, H.S. (2014). A Fuzzy MLP Approach for Non-linear Pattern Classification. In
Proc: K.R. Venugopal, S.C. Lingareddy (eds.) International Conference on Communication and Computing
(ICC- 2014), Bangalore, India (June 12-14, 2014), Computer Networks and Security, 314-323.
| 9cs.NE
|
The average size of the kernel of a matrix
and orbits of linear groups
Tobias Rossmann
arXiv:1704.02668v1 [math.NT] 9 Apr 2017
Department of Mathematics, University of Auckland, Auckland,
New Zealand
Let O be a compact discrete valuation ring of characteristic zero. Given a
module M of matrices over O, we study the generating function encoding the average
sizes of the kernels of the elements of M over finite quotients of O. We prove
rationality and establish fundamental properties of these generating functions and
determine them explicitly for various natural families of modules M . Using p-adic
Lie theory, we then show that special cases of these generating functions enumerate
orbits and conjugacy classes of suitable linear p-groups.
Contents
1
2
3
4
5
6
7
8
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . .
Elementary properties of average sizes of kernels . . . . . . .
Basic algebraic and analytic properties of ZM (T ) and ζM (s)
Rationality of ZM (T ) and p-adic integration . . . . . . . . .
Full matrix algebras, classical Lie algebras, and relatives . .
Constant rank spaces . . . . . . . . . . . . . . . . . . . . . .
Orbits and conjugacy classes of linear groups . . . . . . . . .
Further examples . . . . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
1
6
8
11
18
23
26
31
1 Introduction
This article is devoted to certain generating functions ZM (T ) (“ask zeta functions”)
attached to modules M of matrices over compact discrete valuation rings. The coefficients
of ZM (T ) encode the average sizes of the kernels of the elements of M over finite quotients
2010 Mathematics Subject Classification. 15B33, 15B52, 05A15, 11M41, 11S80, 20D60, 20D15, 20E45
Keywords. Random matrices, average size of a kernel, p-adic integration, orbits of linear groups,
conjugacy classes, finite p-groups, pro-p groups, unipotent groups, matrix Lie algebras
The author gratefully acknowledges the support of the Alexander von Humboldt-Foundation in the
form of a Feodor Lynen Research Fellowship.
of the base ring. Prior to formally defining these functions and stating our main results,
we briefly indicate how our study of the functions ZM (T ) is motivated by questions from
(both finite and infinite) group theory and probabilistic linear algebra.
Conjugacy classes of finite groups. Given a finite group G, let k(G) denote the number
of its conjugacy classes. It is well-known that k(G) coincides with the number of the
(ordinary) irreducible characters of G. Let Ud 6 GLd be the group scheme of upper
unitriangular d × d matrices. Raised as a question in [36], “Higman’s conjecture” asserts
that k(Ud (Fq )) is given by a polynomial in q for each fixed d > 1.
Numerous people have contributed to confirming Higman’s conjecture for small d. In
particular, building on a long series of papers, Vera-López and Arregi [65] established
Higman’s conjecture for d 6 13. Using a different approach, Pak and Soffer [54] recently
provided a confirmation for d 6 16. While Higman’s conjecture remains open in general
and despite some evidence suggesting that it may fail to hold for large d (see [54]),
it nonetheless influenced and inspired numerous results on related questions; see, in
particular, work of Isaacs [40, 41] on character degrees of so-called algebra groups and
work of Goodwin and Röhrle [32–35] on conjugacy classes of unipotent elements in groups
of Lie type.
Orbits of linear groups. All rings in this article are assumed to be commutative and
unital. Let R be a ring and let V be an R-module with |V | < ∞. Given a linear group
G 6 GL(V ), it is a classical problem (for R = Fq ) to relate arithmetic properties of
the number of orbits of G on its natural module V to geometric and group-theoretic
properties of G; see e.g. [29] and the references therein. This problem is closely related to
the enumeration of irreducible characters (and hence of conjugacy classes). In particular,
if G is a finite p-group of nilpotency class less than p, then the Kirillov orbit method
establishes a bijection between the ordinary irreducible characters of G and the coadjoint
orbits of G on the dual of its associated Lie ring; cf. [31] and see [53] for applications of
such techniques to enumerative problems.
Rank distributions and the average size of a kernel. In addition to group-theoretic
problems such as those indicated above, this article is also inspired by questions and
results from probabilistic linear algebra. For an elementary example, to the author’s
knowledge, the number
r−1
Y
q d−i − 1
.
(q e − q i ) i+1
q
−1
i=0
of d × e matrices of rank r with entries in a finite field Fq was first recorded by Landsberg [47]. More recently, probabilistic questions surrounding the distribution of ranks in
sets of matrices over finite fields have been studied, see e.g. [3, 6, 13, 25] and [46, Ch. 3];
for applications, see [49, 63].
Let R be a ring, let V and W be R-modules with |V |, |W | < ∞, and let M ⊂ Hom(V, W )
be a submodule. In the following, we are primarily interested in the case that R is finite
2
and M ⊂ Md×e (R) acts by right-multiplication on V = Rd . The average size of the
kernel of the elements of M is
ask(M ) := ask(M | V ) := |M |−1
X
|Ker(a)|.
a∈M
Linial and Weitz [50] gave the following formula for ask(Md×e (Fq )); the same result
appeared (with a different proof) in a recent paper by Fulman and Goldstein [28, Lem. 3.2]
which also contains further examples of ask(M ).
Proposition 1.1. ask(Md×e (Fq )) = 1 + q d−e − q −e .
As we will see later, for a linear p-group G 6 GL(V ) with a sufficiently strong Lie
theory, |V /G| and k(G) are both instances of ask(g) for suitable linear Lie algebras g.
Orbit-counting and conjugacy class zeta functions. In the literature, numbers of the
form |V /G|, k(G), and ask(M | V ) for R-modules V and W , a linear group G 6 GL(V ),
and M ⊂ Hom(V, W ) were primarily studied in the case that R = Fq is a finite field.
Instead of individual numbers, we consider families of such numbers obtained by replacing
Fq by the finite quotients of suitable rings.
We will use the following notation throughout this article. Let K be a non-Archimedean
local field and let O be its valuation ring—equivalently, O is a compact discrete valuation
ring with field of fractions K; we occasionally write OK instead of O and similarly below.
For example, K could be the field Qp of p-adic numbers (in which case O = Zp is the
ring of p-adic integers) or the field Fq ((t)) of formal Laurent series over Fq (in which
case O = Fq [[t]]). Let P denote the maximal ideal of O. Let K := O/P be the residue
field of K and let q and p denote the size and characteristic of K, respectively. We write
Pn = P · · · P for the nth ideal power of P; on the other hand, On = O × · · · × O denotes
the nth Cartesian power of O. Let On = O/Pn and O∞ = O.
Definition 1.2. Let G 6 GLd (O) be a subgroup.
(i) Let Gn 6 GLd (On ) denote the image of G under the natural map GLd (O)
GLd (On ). The conjugacy class zeta function of G is Zcc
G (T ) =
(ii) The orbit-counting zeta function of G is Zoc
G (T ) =
∞
P
∞
P
k(Gn )T n .
n=0
|Odn /G|T n .
n=0
−s
The study of these generating functions usually involves the Dirichlet series Zcc
G (q )
−s
and Zoc
G (q ) which justifies calling them “zeta functions”. Conjugacy class zeta functions were introduced by du Sautoy [22] who established their rationality for O =
Zp . Berman et al. [5] investigated Zcc
G(O) (T ) for Chevalley groups G. Special cases
of the functions Zoc
(T
)
have
previously
appeared in the literature. In particular,
G
Avni et al. [2, Thm E] determined orbit-counting zeta functions associated with the
coadjoint representations of GL3 and group schemes of the form GU3 .
3
Conjugacy class and orbit-counting zeta functions are natural analogues of the numbers
of conjugacy classes and orbits of finite groups from above. For example, it is a natural
generalisation of Higman’s conjecture to ask, for each fixed d, whether Zcc
Ud (OK ) (T ) is
given by a rational function in qK and T as a function of K.
The definition of ZM (T ). We now introduce the protagonist of this article. Let V and
W be finitely generated O-modules. We frequently write Vn = V ⊗On and Wn = W ⊗On ,
where, in the absence of subscripts, tensor products are always taken over O. Given
a submodule M ⊂ Hom(V, W ), we let Mn denote the image of M under the natural
map Hom(V, W ) → Hom(Vn , Wn ), a 7→ a ⊗ idOn ; we note that the natural surjection
M ⊗ On Mn need not be an isomorphism. Hence, for a submodule M ⊂ Md×e (O), we
obtain Mn ⊂ Md×e (On ) by reducing the entries of all matrices in M modulo Pn . This
article is devoted to generating functions of the following form.
Definition 1.3. Let M ⊂ Md×e (O) be a submodule and V = Od . Define the ask zeta
function of M to be
ZM (T ) := ZM yV (T ) :=
Zask
M (T )
=
∞
X
ask(Mn | Vn )T n ∈ Q[[T ]].
n=0
In contrast to the probabilistic flavour of the work on the numbers ask(M | V ) cited
above, our investigations of the functions ZM (T ) draw upon results and techniques that
have been previously applied in asymptotic group theory and, specifically, the theory of
zeta functions (representation zeta functions, in particular) of groups and other algebraic
structures; for recent surveys of this area, see [45, 68, 69].
Results I: fundamental properties and examples of ask zeta functions. Our central
structural result on the functions ZM (T ) is the following.
Theorem 1.4. Let O be the valuation ring of a non-Archimedean local field of characteristic zero. Let M ⊂ Md×e (O) be a submodule. Then ZM (T ) is rational, i.e. ZM (T ) ∈ Q(T ).
For example, Z{0d×e } (T ) = 1/(1 − q d T ). At the other extreme, we will obtain the
following generalisation of Proposition 1.1.
Proposition 1.5. Let O be the valuation ring of a non-Archimedean local field of
arbitrary characteristic. Let q be the residue field size of O. Then
ZMd×e (O) (T ) =
1 − q −e T
.
(1 − T )(1 − q d−e T )
(1.1)
Note that since ZMd×e (O) (T ) = 1 + (1 + q d−e − q −e )T + O(T 2 ), Proposition 1.5 indeed
generalises Proposition 1.1. Apart from proving Proposition 1.5, in §5, we will also determine ZM (T ) for traceless (Corollary 5.8), symmetric (Proposition 5.11), anti-symmetric
(Proposition 5.9), upper triangular (Proposition 5.13), and diagonal (Corollary 5.15)
matrices. We will also explain why many of our formulae are of the same shape as (1.1).
4
Our proofs of Theorem 1.4, Proposition 1.5, and various other results in this article rest
upon expressing the functions ZM (T ) in terms of p-adic integrals (Theorem 4.5). These
integrals can then be studied using powerful techniques developed over the past decades,
primarily in the context of Igusa’s local zeta function (see [17, 38] for introductions). Our
use of these techniques is similar to and inspired by their applications in the theory of zeta
functions of groups, primarily in the study of representation growth (see [1, 42, 61, 67]).
In particular, Theorem 1.4 follows from rationality results going back to Igusa and Denef.
Using a theorem of Voll [67] we will furthermore see that the identity
= (−q d T ) · ZMd×e (O) (T )
ZMd×e (O) (T )
(q,T )→(q −1 ,T −1 )
is no coincidence (Theorem 4.15). Our p-adic formalism is also compatible with our
previous computational work (summarised in [57]) which allows us to explicitly compute
numerous further examples of ZM (T ); see §8 for some of these.
While “random matrices” over local fields have been studied before (see e.g. [27]), the
author is not aware of previous applications of the particular techniques employed (and
the point of view taken) here.
Results II: ask zeta functions and asymptotic group theory. We say that a formal
power series F (T ) ∈ Q[[T ]] has bounded denominators if there exists a non-zero a ∈ Z
such that aF (T ) ∈ Z[[T ]]. As usual, for a ring R and R-module V , let gl(V ) denote the
Lie algebra associated with the associative algebra End(V ); that is, gl(V ) = End(V ) as
R-modules and the Lie bracket of gl(V ) is defined in terms of multiplication in End(V )
via [a, b] = ab − ba.
Theorem 1.6. Let O be the valuation ring of a non-Archimedean local field of characteristic zero. Let g ⊂ gld (O) be a Lie subalgebra. Then Zg (T ) has bounded denominators.
Theorem 1.6 is based on a connection between ZM (T ) and orbit-counting zeta functions.
For a sketch, let G 6 GLd (O) act on V = Od . As before, we write On = O/Pn and
Vn = V ⊗ On . Then G acts on each of the finite sets Vn and, extending our previous
definition of the orbit-counting zeta function Zoc
G (T ) (Definition 1.2(ii)), we let
Zoc,m
(T ) =
G
∞
X
|Vn /G| · T n−m ∈ Z[[T ]];
n=m
oc,0
hence, Zoc
G (T ) := ZG (T ). In the setting of Theorem 1.6, by linearising the orbit-counting
lemma using p-adic Lie theory, we will see that for sufficiently large m, there exists
Gm 6 GLd (O) with q dm Zg (T ) = Zoc,m
Gm (T ). Theorem 1.6 then follows immediately.
In addition to using group theory to deduce properties of ask zeta functions such as
Theorem 1.6, we will see that, conversely, our methods for studying ask zeta functions
allow us to deduce results on both orbit-counting and conjugacy class zeta functions.
As we will now sketch, this direction is particularly fruitful for unipotent groups. For a
Lie algebra g over a ring R, let ad : g → gl(g) denote its adjoint representation given by
5
ad(a) : b 7→ [b, a] for a ∈ g. Let nd (R) ⊂ gld (R) denote the Lie algebra of strictly upper
triangular d × d matrices.
Theorem 1.7. Let O be the valuation ring of a finite extension of the field Qp of p-adic
numbers. Let g ⊂ nd (O) be a Lie subalgebra and let G := exp(g) 6 Ud (O). Suppose that
p > d and that g is an isolated submodule of nd (O) (i.e. the O-module quotient nd (O)/g
ask
cc
ask
is torsion-free). Then Zoc
G (T ) = Zg (T ) and ZG (T ) = Zad(g) (T ).
For example, we may use Theorem 1.7 and the methods for computing ZM (T ) developed
below in order to compute “generic” conjugacy class zeta functions arising from all
unipotent algebraic groups of dimension at most 5 over a number field (see §8.3).
Outline. In §§2–§3, we collect elementary facts on ask(M | V ) and ZM (T ). We then
derive expressions for ZM (T ) in terms of p-adic integrals in §4. In §5, we use these to
compute explicit formulae for ZM (T ) for various modules M . Next, in §6, we discuss a
relationship between the functions ZM (T ) and “constant rank spaces” studied extensively
in the literature. In §7, we explore the aforementioned connection between ask, conjugacy
class, and orbit-counting zeta functions. In particular, we prove Theorems 1.6–1.7.
Finally, given that most of the explicit formulae for ZM (T ) obtained in §§5–6 are quite
tame, §8 contains a number of “wild” examples of ZM (T ) and Zcc
G (T ).
Acknowledgement
The author is grateful to Angela Carnevale for discussions leading to Corollary 5.15 and
to Eamonn O’Brien and Christopher Voll for helpful comments and suggestions.
2 Elementary properties of average sizes of kernels
We collect some elementary observations on average sizes of kernels. Throughout, let R
be a ring, let V and W be R-modules with |V |, |W | < ∞, and let M ⊂ Hom(V, W ) be a
submodule.
2.1 Rank varieties
In the case of a finite field R = Fq , ask(M | V ) admits a natural geometric interpretation.
Namely, by choosing a basis of M , we may identify M = A`Fq (Fq ), where ` = dimFq (M )
and A`Fq = Spec(Fq [X1 , . . . , X` ]). We may then decompose A`Fq =
r̀
Vi , where Vi is
i=0
) · q d−i−` ; in fact, by
the subvariety of maps of rank i. Then ask(M | V ) = di=0 #Vi (Fq
replacing q by q f on the right-hand side, we express ask(M ⊗Fq Fqf | V ⊗Fq Fqf ) using
a formula which is valid for any f > 1. However, even for M = Md×e (Fq ), this approach
yields a fairly complicated interpretation of Proposition 1.1.
P
6
2.2 Kernels and orbits
One simple yet crucial observation contained in the proof of [50, Thm 1.1] is the following
connection between the sizes of the kernels Ker(a) (a ∈ M ) and those of the “orbits”
xM := {xa : a ∈ M } ≈ M/ cM (x), where x ∈ V and cM (x) := {a ∈ M : xa = 0}; note
that in contrast to orbits under group actions, the sets xM always overlap.
Lemma 2.1 (Cf. [50, Thm 1.1]). ask(M | V ) =
P
|xM |−1 .
x∈V
We give two proofs of this lemma. The first is a combinatorial version of a probabilistic
argument in the proof of [50, Thm 1.1]. We include it here since our terminology is
different from theirs; similar arguments appear in [51].
First proof of Lemma 2.1. By computing #{(x, a) ∈ V × M : xa = 0} in two ways, we
P
P
obtain
|Ker(a)| =
|cM (x)|. As xM ≈R M/ cM (x), we have |cM (x)| = |M |/|xM |
a∈M
x∈V
and thus ask(M | V ) = |M |−1
P
|cM (x)| =
x∈M
|xM |−1 .
P
x∈M
Our second proof of Lemma 2.1 already hints at the connection between average sizes
of kernels and orbits of linear groups, a subject further explored in §7. Recall that
for a finite group G acting on a finite set X, the orbit-counting lemma asserts that
P
|X/G| = |G|−1
|Fix(g)|, where Fix(g) = {x ∈ X : xg = x}.
g∈G
Second proof of Lemma 2.1. The rule a 7→ a∗ := 10 a1 yields an isomorphism of (M, +)
`
onto a subgroup M ∗ of GL(V ⊕ W ). We claim that the natural bijection V ⊕ W → W
V
induces a bijection (V ⊕ W )/M ∗ →
`
W/xM . Indeed, (x, y)a∗ = (x, y + xa) for
x∈V
(x, y) ∈ V ⊕ W . As Fix(a∗ ) = Ker(a) ⊕ W , the orbit-counting lemma yields
|W |
X
|xM |−1 = |(V ⊕ W )/M ∗ | = |M ∗ |−1
x∈V
X
|Fix(a∗ )| = |W | · ask(M | V ).
a∈M
In order to deduce Proposition 1.1 from Lemma 2.1, note that x Md×e (Fq ) = Feq for
each non-zero x ∈ Fdq whence ask(Md×e (Fq )) = 1 + (q d − 1)q −e .
2.3 Direct sums
Lemma 2.2. Let V 0 and W 0 be R-modules with |V 0 |, |W 0 | < ∞. Let M 0 ⊂ Hom(V 0 , W 0 )
be a submodule. We regard M ⊕ M 0 as a submodule of Hom(V ⊕ V 0 , W ⊕ W 0 ) in the
natural way. Then ask(M ⊕ M 0 | V ⊕ V 0 ) = ask(M | V ) · ask(M 0 | V 0 ).
Proof.
ask(M ⊕ M 0 | V ⊕ V 0 ) = |M ⊕ M 0 |−1 ·
X
|Ker(a ⊕ a0 )|
(a,a0 )∈M ⊕M 0
−1
= |M |
0 −1
|M |
·
X
(a,a0 )∈M ⊕M 0
0
0
|Ker(a)| · |Ker(a0 )|
= ask(M | V ) · ask(M | V ).
7
Corollary 2.3. Let R be finite, M ⊂ Md×e (R) be a submodule, and M̃ ⊂ M(d+1)×e (R)
be obtained from M by adding a zero row to the elements of M in some fixed position.
Then ask(M̃ ) = ask(M ) · |R|.
2.4 Matrix transposes
Following Kaplansky [43], we call R an elementary divisor ring if for each a ∈ Md×e (R)
(and all d, e > 1), there exist u ∈ GLd (R) and v ∈ GLe (R) such that uav is a diagonal
matrix (padded with zeros according to the shape of a). For example, any quotient of a
principal ideal domain is an elementary divisor ring (regardless of whether the quotient
is an integral domain or not). Write a> for the transpose of a.
Lemma 2.4. Let R be a finite elementary divisor ring and let M ⊂ Md×e (R) be a
submodule. Then ask(M > | V ) = ask(M | V ) · |R|e−d .
Proof. Let a =
h
diag(a1 ,...,ar ) 0
0
0
i
∈ Md×e (R). Then Ker(a) consists of those x ∈ Rd
with ai xi = 0 for 1 6 i 6 r and Ker(a> ) consists of those y ∈ Re with ai yi = 0 for
1 6 i 6 r.
2.5 Reduction modulo a and base change R → R/a
Let V and W be finitely generated R-modules, the underlying sets of which need not
be finite. As before, let M ⊂ Hom(V, W ) be a submodule. Let a / R with |R/a| < ∞.
Define Va := V ⊗R R/a, Wa := W ⊗R R/a, and let Ma be the image of the natural map
M ,→ Hom(V, W ) → Hom(Va , Wa ). In general, the natural surjection M ⊗R R/a Ma
need not be injective. However, if M is finitely generated, then M ⊗R R/a is finite and
we obtain the following expression for ask(Ma | Va ).
Lemma 2.5. Suppose that M is finitely generated. Then
ask(Ma | Va ) = |M ⊗R R/a|−1
X
|Ker(ā | Va )|.
ā∈M ⊗R R/a
3 Basic algebraic and analytic properties of ZM (T ) and ζM (s)
3.1 Average sizes of kernels and Dirichlet series: ζM (s)
While our main focus is on the generating functions ZM (T ) from the introduction, it is
natural to also consider a global analogue. First suppose that R is a ring which contains
only finitely many ideals a of a given finite norm |R/a|. Given a submodule M ⊂ Md×e (R)
acting on V = Rd and an ideal a / R, let Va = (R/a)d , Wa = (R/a)e , and let Ma denote
the image of the natural map M ,→ Md×e (R) Md×e (R/a) (cf. §2.5).
8
Definition 3.1.
(i) Define a formal Dirichlet series
ζM (s) =
X
ask(Ma | Va ) · |R/a|−s ,
a
where the sum extends over the ideals of finite norm of R.
(ii) Let αM ∈ [−∞, ∞] denote the abscissa of convergence of ζM (s).
3.2 Abscissae of convergence: local case
Let K be a local field of arbitrary characteristic with valuation ring O and residue field
size q. Let M ⊂ Md×e (O) be a submodule acting on V = Od . Then ζM (s) = ZM (q −s ).
Moreover, if O has characteristic zero, then Theorem 1.4 (proved in §4.3.3) implies that
αM is precisely the largest real pole of (the meromorphic continuation of) ζM (s).
Recall that, unless otherwise indicated, tensors products are taken over O.
Definition 3.2. The generic orbit rank of M is gor(M ) := max dimK (xM ⊗ K).
x∈V
Our choice of terminology will be justified by Proposition 4.12.
Proposition 3.3. max d − gor(M ), 0 6 αM 6 d.
n(d−s)
Proof. The upper bound follows since ask(Mn | Vn ) 6 |Vn | = q nd and ∞
n=0 q
converges for Re(s) > d. Similarly, the lower bound follows from Lemma 2.1 and
ask(Mn | Vn ) > max(|Vn |/q n gor(M ) , 1).
P
Let 0 6 r 6 e. Let M ⊂ Md×e (O) be obtained from Md×r (O) by inserting e − r
zero columns in some fixed positions. Then gor(M ) = r, ζM (s) = ζMd×r (O) (s), and it
will follow from Proposition 1.5 that αM = max(d − r, 0). In particular, the bounds
in Proposition 3.3 are optimal. We note that Example 8.1 below illustrates that the
meromorphic continuation of ζM (s) (cf. Theorem 1.4) may have real poles less than d − e.
3.3 Abscissae of convergence in the global case and Euler products
Let k be a number field with ring of integers o. Let Vk denote the set of non-Archimedean
places of k. For v ∈ Vk , let kv be the v-adic completion of k and let ov be its valuation
ring. We let qv denote the size of the residue field Kv of kv . For an o-module U and
v ∈ Vk , we write Uv := U ⊗o ov (regarded as an ov -module).
Let M ⊂ Md×e (o) be a submodule. For v ∈ Vk , we may identify Mv with the
ov -submodule of Md×e (ov ) generated by M .
Proposition 3.4. Let M ⊂ Md×e (o) be a submodule. Then:
(i) αM 6 d + 1.
(ii) ζM (s) =
Q
v∈Vk
ζMv (s).
9
Proof. Let V = od . The proof of (i) is similar to that of Proposition 3.3. Namely, for
P
each a / o, ask(Ma | Va ) 6 |o/a|d and a |o/a|d−s = ζk (s − d) converges for Re(s) > d + 1,
where ζk (s) is the Dedekind zeta function of k. For (ii), it suffices to show that for
non-zero coprime ideals a, b / o, ask(Mab | Vab ) = ask(Ma | Va ) · ask(Mb | Vb ).
To that end, the natural isomorphism o/ab → o/a × o/b yields an (o-module) isomorphism M ⊗o o/ab → (M ⊗o o/a) × (M ⊗o o/b) which is compatible with the corresponding isomorphism Vab → Va × Vb in the evident way. Hence, for ā ∈ M ⊗o o/ab
corresponding to (āa , āb ) ∈ (M ⊗o o/a) × (M ⊗o o/b), we obtain an isomorphism
Ker(ā | Vab ) → Ker(āa | Va ) × Ker(āb | Vb ). Part (ii) thus follows from Lemma 2.5.
Example 3.5. Let ζk (s) denote the Dedekind zeta function of k. Then Proposition 1.5
and Proposition 3.4(ii) imply that ζMd×e (o) (s) = ζk (s)ζk (s − d + e)/ζk (s + e).
Further analytic properties of ζM (s) in a global setting will be derived in §4.4.
3.4 Hadamard products
Recall that the Hadamard product F (T ) ? G(T ) of formal power series F (T ) =
P∞
P∞
n
n
n=0 an T and G(T ) =
n=0 bn T with coefficients in some common ring is
F (T ) ? G(T ) =
∞
X
an bn T n .
n=0
The following is an immediate consequence of Lemma 2.2.
Corollary 3.6. Let K be a local field of arbitrary characteristic with valuation ring O.
Let M = A ⊕ B ⊂ Md+e (O) for submodules A ⊂ Md (O) and B ⊂ Me (O). Then
ZM (T ) = ZA (T ) ? ZB (T ).
We note that Hadamard products of rational generating functions are well-known to
be rational (see [60, Prop. 4.2.5]).
Corollary 3.7. Let M ⊂ Md×e (O) be a submodule. Define f = max(d, e). Let M̃ ⊂
Mf (O) be obtained from M by adding f − d zero rows and f − e zero columns to the
elements of M in some fixed positions. Then ZM̃ (T ) = ZM (q f −d T ).
Thus, various questions on the series ZM (T ) are reduced to the case of square matrices.
3.5 Rescaling
Let O be the valuation ring of a non-Archimedean local field K of arbitrary characteristic.
Let M ⊂ Md×e (O) be a submodule, V = Od , and W = Oe .
Definition 3.8. For m > 0, let Zm
M (T ) :=
Note that ZM (T ) = Z0M (T ). For an
common value of (U m )n and (Un )m .
Vn ) = |Vn | = q nd .
∞
P
n=m
ask(Mn | Vn ) · T n−m ∈ Q[[T ]].
O-module U , let U m = Pm U and write Unm for the
Clearly, if n 6 m, then ask(Mnm | Vn ) = ask({0} |
10
dm · Z (T ).
Proposition 3.9. Zm
M
M m (T ) = q
Proof. It suffices to show that ask(Mnm | Vn ) = q dm · ask(Mn−m | Vn−m ) for n > m.
Choose π ∈ P\P2 . Observe that multiplication by π m induces an O-module isomorphism
Mn−m → Mnm and a monomorphism Vn−m → Vn with image Vnm . For a ∈ M ,
Ker(π m a | Vn ) = x ∈ Vn : x(π m a) ≡ 0
(mod W n )
(mod W n−m )
= x ∈ Vn : xa ≡ 0
= x ∈ Vn : x + V n−m ∈ Ker(a | Vn−m )
has size |Ker(a | Vn−m )| · q dm . We conclude that
ask(Mnm | Vn ) = |Mnm |−1
X
|Ker(a | Vn )|
a∈Mnm
= |Mn−m |−1
X
|Ker(a | Vn−m )| · q dm
a∈Mn−m
=q
dm
· ask(Mn−m | Vn−m ).
4 Rationality of ZM (T ) and p-adic integration
Unless otherwise stated, in this section, K is a non-Archimedean local field of arbitrary
characteristic with valuation ring O. Given a submodule M ⊂ Md×e (O), we use the
original definition of ask(M | V ) as well as the alternative formula in Lemma 2.1 to
derive two types of expressions for ZM (T ) in terms of p-adic integrals.
In §4.1, we describe a general setting for rewriting certain generating functions as
integrals. By specialising to the case at hand, in §4.2, we obtain two expressions for
ZM (T ) (Theorem 4.5) in terms of functions KM and OM that we introduce. In §4.3, we
derive explicit formulae (in terms of the absolute value of K and polynomials over O) for
these functions. These formulae serve two purposes. First, using established rationality
results from p-adic integration, they allow us to deduce Theorem 1.4. Secondly, these
formulae, in particular the one based on OM (and hence on Lemma 2.1), lie at the heart
of explicit formulae such as Proposition 1.5 in §5.
4.1 Generating functions and p-adic integrals
Let Z be a free O-module of finite rank d and let U ⊂ Z be a submodule. By the
elementary divisor theorem, there exists a unique sequence (λ1 , . . . , λd ) with 0 6 λ1 6
L
· · · 6 λd 6 ∞ such that Z/U ≈ di=1 Oλi as O-modules; recall that O∞ = O. We
call (λ1 , . . . , λd ) the submodule type of U within Z. Recall that the isolator isoZ (U )
of U in Z is the preimage of the torsion submodule of Z/U under the natural map
Z Z/U —equivalently, isoZ (U ) is the smallest direct summand of Z which contains U .
If U is isolated in Z (i.e. U = isoZ (U )), then we may naturally identify U ⊗ On with the
image Un of U ,→ Z Zn := Z ⊗ On . In that case, if U has rank `, say, then |Un | = q `n .
The general case is only slightly more complicated.
11
We let ν denote the valuation on K with value group Z. Let | · | be the absolute value
on K with |π| = q −1 for π ∈ P \ P2 ; we write kAk = sup(|a| : a ∈ A).
Lemma 4.1. Let U ⊂ Z be a submodule. Suppose that U has submodule type (λ1 , . . . , λ` )
within isoZ (U ). Let Un denote the image of the natural map U ,→ Z Zn := Z ⊗ On .
P̀
(n−min(λi ,n))
Then |Un | = q i=1
Q̀ kπλi ,yk
. In particular, for y ∈ O with ν(y) = n, |Un | =
i=1
|y|
.
P`
Proof. Clearly, |Un | = q i=1 min(n−λi ,0) and the first identity follows from min(0, n − a) =
n − min(n, a); the second claim then follows immediately.
The following result is concerned with generating functions associated with a given
family of “weight functions” fn : Un → R>0 . Given a free O-module W of finite rank, let
µW denote the Haar measure on W with µW (W ) = 1.
Lemma 4.2. Let U ⊂ Z be a submodule. Suppose that U has submodule type (λ1 , . . . , λ` )
within isoZ (U ). Let Un denote the image of U ,→ Z Zn := Z ⊗ On and let πn : U → Un
denote the natural map. Let N > 0 and, for n > N , let fn : Un → R>0 be given. Define
F : U × PN \ {0} → R>0 ,
(x, y) 7→ fν(y) (πν(y) (x))
and extend F to a map U × PN → R>0 via F (x, 0) ≡ 0. Let
g : PN × C → R>0 ,
(y, s) 7→ |y|s−`−1 ·
`
Y
kπ λi , yk,
i=1
where we set g(0, s) ≡ 0. (Note that g(y, s) = |y|s−1 · |Uν(y) | for y 6= 0 by Lemma 4.1.)
P
fn (x̄) = O(q δn ). Then, writing t = q −s , for all s ∈ C
Suppose that δ > 0 satisfies
x̄∈Un
with Re(s) > δ,
∞ X
X
n=N x̄∈Un
fn (x̄)tn = (1 − q −1 )−1
Z
F (x, y) · g(y, s) dµU ×O (x, y).
(4.1)
U ×PN
Proof. First note that the left-hand side of (4.1) converges for Re(s) > δ. Further note
that we may ignore the case y = 0 on the right-hand side as it occurs on a set of measure
zero.
Let U n = Ker(πn : U Un ). Given (x, y) ∈ U × PN \ {0} with n := ν(y), the map F
is constant on the open set (x + U n ) × (y + Pn+1 ); in particular, F is measurable.
Let Rn ⊂ U be a complete and irredundant set of representatives for the cosets of U n
and let Wn = Pn · O× = Pn \ Pn+1 . By Lemma 4.1, µU (U n ) = |Un |−1 =
Q̀
i=1
12
|y|
kπ λi ,yk
for
any y ∈ Wn ; moreover, µO (Wn ) = (1 − q −1 )q −n . The claim now follows via
Z
F (x, y) · g(y, s) dµU ×O (x, y) =
U ×PN
∞ X
X
Z
fn (πn (x)) · g(y, s)
| {z }
n=N x∈Rn(x+U n )×W
n
=|Un | ·(qt)n
= (1 − q −1 )
= (1 − q −1 )
∞ X
X
dµU ×O (x, y)
fn (x̄) · µ(U n × Wn ) · |Un | · (qt)n .
n=N x̄∈Un
∞ X
X
fn (x̄)tn .
n=N x̄∈Un
Remark 4.3. The introduction of the additional variable y to express a generating
function as an integral in Lemma 4.2 mimics similar formulae of Jaikin-Zapirain [42, §4]
and Voll [67, §2.2].
4.2 ZM (T ) and the functions KM and OM
Let M ⊂ Md×e (O) be a submodule, V = Od , and W = Oe .
Definition 4.4. Define
!
a⊗O/(y)
KM : M × O → N0 ∪ {∞},
(a, y) 7→ #Ker V ⊗ O/(y) −−−−−→ W ⊗ O/(y)
OM : V × O → N0 ∪ {∞},
(x, y) 7→ # Im xM ,→ W W ⊗ O/(y) .
and
Note that for y 6= 0, 1 6 KM (a, y) 6 |y|−d and 1 6 OM (x, y) 6 |y|− gor(M ) (see
Definition 3.2); these are the same estimates as in Proposition 3.3.
Using Lemmas 2.1 and 4.2, we obtain the following formulae for ZM (t) (where t = q −s ).
Theorem 4.5. For s ∈ C with Re(s) > d,
(1 − q −1 ) · ZM (q −s ) =
Z
|y|s−1 · KM (a, y) dµM ×O (a, y) =
Z
|y|s−d−1
dµV ×O (x, y).
OM (x, y)
V ×O
M ×O
Proof. The given formulae for (1 − q −1 ) · ZM (q −s ) are based on ask(Mn | Vn ) = |Mn |−1 ·
P
−1 (Lemma 2.1), respectively. In detail,
a∈Mn |Ker(a)| and ask(Mn | Vn ) =
x̄∈Vn |x̄Mn |
the first equality follows from Lemma 4.2 with U = M , Z = Md×e (O), F (a, y) =
|Mν(y) |−1 KM (a, y) and g(y, s) = |y|s−1 · |Mν(y) | (for y =
6 0). For the second equality,
use Lemma 4.2 with U = Z = V , F (x, y) = |(x + yOd )Mν(y) |−1 = OM (x, y)−1 and
g(y, s) = |y|s−d−1 (for y 6= 0).
P
4.3 Explicit formulae for KM and OM
As before, let M ⊂ Md×e (O) be a submodule. In order to use Theorem 4.5 for theoretical
investigations or explicit computations of ZM (T ), we need to produce sufficiently explicit
formulae for KM (a, y) or OM (x, y).
13
4.3.1 The sizes of kernels and images
Let a ∈ Md×e (O) have rank r over K. Fix π ∈ P \ P2 . By the elementary divisor
theorem, there are 0 6 λ1 6 · · · 6 λr , u ∈ GLd (O), and v ∈ GLe (O) such that
"
#
diag(π λ1 , . . . , π λr ) 0
uav =
.
0
0
(4.2)
We call (λ1 , . . . , λr ) the equivalence type of a.
For a set of polynomials f (Y ), we write f (y) = {f (y) : f ∈ f }.
Lemma 4.6. Let a ∈ Md×e (O) have rank r over K and equivalence type (λ1 , . . . , λr ).
Let fi (Y ) ⊂ Z[Y ] := Z[Yij : 1 6 i 6 d, 1 6 j 6 e] be the set of non-zero i × i minors of
the generic d × e matrix [Yij ] ∈ Md×e (Z[Y ]). (Hence, f0 (Y ) = {1}.) Let an ∈ Md×e (On )
be the image of the matrix a under the natural map Md×e (O) Md×e (On ). Then:
(i) kfi (a)k = q −λ1 −···−λi for 0 6 i 6 r.
(ii) (Cf. [67, §2.2])
kfi−1 (a)k
kfi (a)∪π n fi−1 (a)k
= q min(λi ,n) for 1 6 i 6 r.
(iii) |Ker(an )| = q min(λ1 ,n)+···+min(λr ,n)+(d−r)n .
(iv) |Im(an )| = q rn−(min(λ1 ,n)+···+min(λr ,n)) .
Proof. The first part is elementary linear algebra. Part (ii) then follows from
q −λ1 −···−λi−1 +min(λ1 +···+λi ,λ1 +···+λi−1 +n) = q min(λi ,n) .
For 0 6 i 6 n, the map On → On given by multiplication by π i has kernel and image
size q i and q n−i , respectively. Parts (iii)–(iv) thus follow from equation (4.2).
4.3.2 A formula for KM
We use Lemma 4.6 in order to derive a formula for KM (a, y).
Definition 4.7. The generic element rank of M is grk(M ) := max rkK (a).
a∈M
By the following, grk(M ) is the rank of a “generic” matrix in M in any meaningful
sense.
Proposition 4.8. Let (a1 , . . . , a` ) be an O-basis of M and let λ1 , . . . , λ` be algebraically
independent over K. Then:
(i) grk(M ) = rkK(λ1 ,...,λ` ) (λ1 a1 + · · · + λ` a` ).
(ii) µM ({a ∈ M : rkK (a) < grk(M )}) = 0.
14
Proof. Let r denote the right-hand side in (i). Then r is the largest number such that some
r × r minor, m(λ1 , . . . , λ` ) say, of λ1 a1 + · · · + λ` a` is non-zero. In particular, grk(M ) 6 r.
Conversely, since m(λ1 , . . . , λ` ) 6= 0, we find c1 , . . . , c` ∈ O with m(c1 , . . . , c` ) 6= 0 whence
grk(M ) > r. Finally, the well-known fact (provable using induction and Fubini’s theorem)
that the zero locus of a non-zero polynomial over K has measure zero implies (ii).
Excluding a set of measure zero, we thus obtain the following formula for KM (a, y).
Corollary 4.9. Let N = {a ∈ M : rkK (a) < grk(M )} and let fi (Y ) be the set of nonzero i × i minors of the generic d × e matrix. Then for all a ∈ M \ N and y ∈ O \ {0},
grk(M )
KM (a, y) = |y|grk(M )−d ·
Y
i=1
kfi−1 (a)k
.
kfi (a) ∪ yfi−1 (a)k
Proof. Immediate from Lemma 4.6(ii)–(iii).
Hence, using Theorem 4.5, we conclude that
grk(M )
Z
ZM (q −s ) = (1 − q −1 )−1 |y|s+grk(M )−d−1 ·
Y
i=1
M ×O
kfi−1 (a)k
dµM ×O (a, y). (4.3)
kfi (a) ∪ yfi−1 (a)k
4.3.3 Rationality and variation of the place
As in the proof of Proposition 4.8, we may replace the fi (Y ) in (4.3) by polynomials in a
chosen system of coordinates of M . We may thus interpret the integral in (4.3) as being
defined in terms of valuations of polynomial expressions in dimK (M ⊗ K) + 1 variables.
Integrals of this form have been studied extensively. In particular, using well-known
results going back to work of Igusa and Denef (see [17]), initially for a single polynomial
and later extended to families of polynomials (see, in particular, work of du Sautoy and
Grunewald [23], Veys and Zúñiga-Galindo [66], and Voll [67]), we obtain the following
two results, the first of which implies and refines Theorem 1.4.
Corollary 4.10. Let O be the valuation ring of a local field of characteristic zero and
let M ⊂ Md×e (O) be a submodule. Then ZM (T ) ∈ Q(T ). More precisely, there exist
f (T ) ∈ Z[T ], non-zero (a1 , b1 ), . . . , (ar , br ) with (ai , bi ) ∈ Z × N0 , and m ∈ N0 such that
ZM (T ) =
f (T )
.
q m (1 − q a1 T b1 ) · · · (1 − q ar T br )
Moreover, in a global setting, the dependence of Euler factors on the place is as follows.
Corollary 4.11. Let k be a number field with ring of integers o. Recall the notation
from §3.3. Let M ⊂ Md×e (o) be a submodule. There are separated o-schemes V1 , . . . , Vr
of finite type and rational functions W1 (X, T ), . . . , Wr (X, T ) ∈ Q(X, T ) (which can be
written over denominators of the same shape as those in Corollary 4.10) such that the
following holds: for almost all v ∈ Vk ,
ZMv (T ) =
r
X
#Vi (Kv ) · Wi (qv , T ).
i=1
15
(4.4)
4.3.4 A formula for OM
As in the case of KM , we can produce a formula for OM .
Let X = (X1 , . . . , Xd ) be algebraically independent over K. Let C(X) ∈ M`×e (O[X])
with O[X]d · C(X) = X · (M ⊗ O[X]). For example, we may choose generators a1 , . . . , a`
of M as an O-module and take
Xa1
..
C(X) = . ∈ M`×e (O[X]).
Xa`
Let gi (X) be the set of non-zero i × i minors of C(X). Note that if x ∈ V , then xM is
the row span of C(x) over O so that, in particular, dimK (xM ⊗ K) = rkK (C(x)).
Recall the definition of gor(M ) in Definition 3.2. The following is proved in the same
way as Proposition 4.8.
Proposition 4.12. Let X = (X1 , . . . , Xd ) be algebraically independent over K. Then:
(i) gor(M ) = dimK(X) X · (M ⊗ K(X)) = rkK(X) (C(X)).
(ii) µV ({x ∈ V : dimK (xM ⊗ K) < gor(M )}) = 0.
The following analogue of Corollary 4.13 is obtained using Lemma 4.6(ii),(iv).
Corollary 4.13. Let Z = {x ∈ V : dimK (xM ⊗ K) < gor(M )}. Then for all x ∈ V \ Z
and y ∈ O \ {0},
gor(M )
OM (x, y) = |y|
− gor(M )
Y
i=1
kgi (x) ∪ ygi−1 (x)k
.
kgi−1 (x)k
(4.5)
Theorem 4.5 thus provides us with the following counterpart of (4.3):
Z
gor(M )
ZM (q −s ) = (1 − q −1 )−1 |y|s+gor(M )−d−1 ·
V ×O
Y
i=1
kgi−1 (x)k
dµV ×O (x, y). (4.6)
kgi (x) ∪ ygi−1 (x)k
Despite the essentially identical shapes of the integrals in (4.3) and (4.6), either type
might be vastly more useful for explicit computations of specific examples. In particular, §5
is concerned with examples of ZM (T ) that can be easily computed using (4.6) and §6
considers the analogous situations for (4.3). A very similar phenomenon was exploited by
O’Brien and Voll [53, §5] in their enumeration of conjugacy classes of certain relatively
free p-groups.
Remark 4.14. We note that the integrals in (4.3)–(4.6) are almost of the same shape as
those in [1, Eqn (1.4)]. Moreover, our matrix C(X) plays the same role as the commutator
matrix R(Y ) in [1, §3.1] and [61, §2.2.2]. These similarities can be clarified further by
rewriting our integrals slightly; see the proof of Theorem 4.15 below.
16
4.4 Local functional equations and global analytic properties
Functional equations under “inversion of the prime” are a common (but not universal)
phenomenon in the theory of local zeta functions. Denef and Meuser [18] showed that
for a homogeneous polynomial over a number field, almost all of its associated local
Igusa zeta functions satisfy such a functional equation. Vastly generalising their result,
Voll [67] established functional equations for numerous types of zeta functions arising
in asymptotic algebra and expressible in terms of p-adic integrals. Using the formalism
developed above, we may deduce the following; recall the notation from §3.3.
Theorem 4.15. Let k be a number field with ring of integers o. Let M ⊂ Md×e (o) be a
submodule. Recall that ζMv (s) = ZMv (qv−s ) for v ∈ Vk . Then for almost all v ∈ Vk ,
ζMv (s)
qv →qv−1
= (−q d−s ) · ζMv (s).
Remark 4.16. The operation “qv → qv−1 ” can be unambiguously defined in terms of an
arbitrary formula of the form (4.4); see [18,67] and cf. [56, Cor. 5.3]. If ZMv (T ) = W (qv , T )
for almost all v ∈ Vk and some W (X, T ) ∈ Q(X, T ), then Theorem 4.15 asserts that
W (X −1 , T −1 ) = (−X d T ) · W (X, T ); see [56, §5].
Proof of Theorem 4.15. We rewrite the second integral in Theorem 4.5 in a form that
allows us to invoke Voll’s results from [67, §2.1]. Let K = kv for v ∈ Vk and let O, | · |,
etc. be as before. Fix π ∈ P \ P2 . Clearly, OMv (πx, πy) = OMv (x, y) for x ∈ V := Od
and y ∈ O \ {0}. Writing t = qv−s , Theorem 4.5 and a change of variables thus show that
(1 − qv−1 )−1 · tZMv (t) =
R
PV ×P
|y|s−d−1
OMv (x,y)
Hv (s) := 1 + (1 − q
dµV ×O (x, y). Let
−1 −1
Z
)
(V \PV )×P
|y|s−d−1
dµV ×O (x, y).
OMv (x, y)
(4.7)
As OMv (x, y) = 1 for all x ∈ V and y ∈ O× , ZMv (t) = Hv (s)/(1 − t). We may then use
the surjection GLd (O) → V \ PV which sends a matrix to its first row to rewrite (4.7)
in terms of an integral over GLd (O) × P; cf. [1, §4]. In the setting of the explicit formula
for OMv (x, y) derived in §4.3.4, we may assume that C(X) is a matrix of linear forms
whence each gi (X) consists of homogeneous polynomials of degree i. This allows us to
use [67, Cor. 2.4] which shows that Hv (s) qv →q−1 = q d Hv (s) for almost all v ∈ Vk .
v
Based on work of du Sautoy and Grunewald [23], Duong and Voll [26] studied analytic
properties of Euler products of functions of the same form as the Hv (s) in (4.7). In
particular, their findings allow us to deduce the following.
Theorem 4.17 (Cf. [26, Thm A]). Let k be a number field with ring of integers o. Let
M ⊂ Md×e (o) be a submodule and V = od .
(i) The abscissa of convergence αM of ζM (s) is a rational number.
17
(ii) There exists δ > 0 such that ζM (s) admits meromorphic continuation to {s ∈ C :
Re(s) > αM − δ}. This continued function has a pole of order βM , say, at s = αM
but no other poles on the line Re(s) = αM .
(iii) αM and βM are (and δ can be chosen to be) invariant under base change of M
from o to the ring of integers of an arbitrary finite extension of k.
5 Full matrix algebras, classical Lie algebras, and relatives
In this section, let O be the valuation ring of a non-Archimedean local field of arbitrary
characteristic. Apart from proving Proposition 1.5, we compute examples of Zg (T ),
where g ranges over various infinite families of matrix Lie algebras. At the heart of these
computations lies the notion of “O-maximality” introduced in §5.1.
5.1 O-maximality
Let M ⊂ Md×e (O) be a submodule and V = Od . As we will now see, if OM (x, y) is
(generically) as large as possible, then ZM (T ) coincides with ZMd×gor(M ) (T ).
Lemma 5.1. Let x ∈ V \ {0} and y ∈ O \ {0}. Then OM (x, y) 6 |y|− gor(M ) kx, ykgor(M ) .
Proof. Let C(X) ∈ M`×e (O[X]) with O[X]d · C(X) = X · (M ⊗ O[X]). As 0Od · M =
{0Oe }, the constant terms of all non-zero polynomials in C(X) vanish whence kC(x)k 6
kxk. Thus, if C(x) has equivalence type (λ1 , . . . , λr ), then |π λi | 6 |π λ1 | = kC(x)k 6 kxk
for 1 6 i 6 r. Define m and n via q −m = kxk and n = ν(y). Then, by Lemma 4.6(iv),
Pr
OM (x, y) = q rn−
i=1
min(λi ,n)
6 q gor(M )(n−min(m,n)) = |y|− gor(M ) kx, ykgor(M ) .
The above inequality is sharp (cf. the comments after Proposition 3.3):
Lemma 5.2. Let x ∈ V \ {0} and y ∈ O \ {0}. Then OMd×e (O) (x, y) = |y|−e kx, yke .
Proof. Let (e1 , . . . , ed ) ⊂ Od be the standard basis. As x Md×e (O) is generated by
{gcd(x1 , . . . , xd )ei : i = 1, . . . , e}, the claim follows from Lemma 4.6(iv).
Definition 5.3. We say that M is O-maximal if OM (x, y) = |y|− gor(M ) kx, ykgor(M ) for
all (x, y) ∈ V × O outside a set of measure zero.
Proposition 1.5 serves as a blueprint for ZM (T ) whenever M is O-maximal:
Corollary 5.4. If M is O-maximal, then ZM (T ) = ZMd×gor(M ) (O) (T ).
Proof. Combine (4.6) and Lemma 5.2.
We will repeatedly use the following lemma to prove O-maximality.
Lemma 5.5. Let X = (X1 , . . . , Xd ). Let C(X) ∈ M`×e (O[X]) with O[X]d C(X) =
X · (M ⊗ O[X]). Suppose that for i = 1, . . . , gor(M ), the ideal of O[X] generated by the
i × i minors of C(X) contains X1i , . . . , Xdi . Then M is O-maximal.
18
Proof. As in §4.3.4, let gi (X) be the set of non-zero i × i minors of C(X). Let hi (X) =
{X α : α ∈ Nd0 , α1 + · · · + αd = i}. Let x ∈ Od with x1 · · · xd 6= 0 and y ∈ O \ {0}.
Since the constant term of each f (X) ∈ gi (X) vanishes for 1 6 i 6 gor(M ) (see
the proof of Lemma 5.1), each monomial in f (X) has total degree at least i whence
|f (x)| 6 khi (x)k and therefore kgi (x)k 6 khi (x)k = kxi1 , . . . , xid k. By our assumption,
kgi (x)k > kxi1 , . . . , xid k = kxki ; in particular, dimK (xM ⊗ K) = gor(M ). Thus, by
Corollary 4.13, M is OM -maximal since
gor(M )
− gor(M )
Y
OM (x, y) = |y|
max kxki , |y| · kxki−1
i−1
kxk
i=1
= |y|− gor(M ) kx, ykgor(M ) .
For a geometric interpretation of Lemma 5.5 in a global setting, see Proposition 6.10.
5.2 Proof of Proposition 1.5
Our proof of Proposition 1.5 and other computations in §5.3 rely on the following.
Lemma 5.6. Let a0 , . . . , ar ∈ C and write σj = a0 + · · · + aj . Suppose that the integral
Z
Fr (a0 , . . . , ar ) :=
|y|a0 kx1 , yka1 · · · kx1 , . . . , xr , ykar dµOr ×O (x, y),
Or ×O
is absolutely convergent. Then
Fr (a0 , . . . , ar ) =
r−1
Y 1 − q −σj −j−2
1 − q −1
·
.
1 − q −σr −r−1 j=0 1 − q −σj −j−1
In particular, in the special case a1 = · · · = ar−1 = 0, we obtain
Fr (a0 , 0, . . . , 0, ar ) =
(1 − q −1 )(1 − q −a0 −r−1 )
.
(1 − q −a0 −ar −r−1 )(1 − q −a0 −1 )
Proof. Both claims follow by induction from the identities (a) F0 (a0 ) =
1−q −1
1−q −1−a0
and
1−q −a0 −2
1−q −a0 −1
(b) Fr (a0 , . . . , ar ) = Fr−1 (a0 + a1 + 1, a2 , . . . , ar ) ·
for r > 1. The formula for
F0 (a0 ) in (a) is well-known and easily proved. By performing a change of variables
according to whether
|x1 | 6 |y| or|x1 | > |y|, we find that Fr (a0 , . . . , ar ) = Fr−1 (a0 + a1 +
R
1, a2 , . . . , ar ) · 1 + P |y|a0 dµO (y) whence (b) follows readily.
Proof of Proposition 1.5. Using (4.6) and Lemma 5.2, we obtain
ZMd×e (O) (q
−s
) = (1 − q
−1 −1
Z
|y|s+e−d−1 kx, yk−e dµOd ×O (x, y).
)
(5.1)
Od ×O
whence the claim follows from Lemma 5.6.
19
Remark 5.7.
(i) Let M ⊂ Md×e (O) be any submodule. By combining Lemma 5.1 and (5.1), we
thus obtain another interpretation of the lower
bound in Proposition 3.3 in the
form αM > αMd×gor(O) (O) = max d − gor(O), 0 .
(ii) We note that (1.1) could also be derived in an elementary fashion (without using
p-adic integration) using Lemma 2.1 and ad hoc computations with generating
functions. Such an approach quickly becomes cumbersome for more complicated
examples such as most of those in §8. The author is, moreover, unaware of
elementary proofs of general results such as Theorems 1.4 and 4.15.
5.3 Classical Lie algebras and relatives
Reminder. Let R be a ring. Recall the definitions of the special linear, orthogonal,
and symplectic Lie algebras
sld (R) = {a ∈ gld (R) : trace(a) = 0},
sod (R) = {a ∈ gld (R) : a + a> = 0}
("
sp2d (R) =
(assuming char(R) 6= 2), and
#
)
a
b
: a, b, c ∈ Md (R), b = b> , c = c> .
c −a>
These are Lie subalgebras of gld (R) and gl2d (R), respectively. Finally, we let trd (R)
and nd (R) denote the Lie subalgebras of gld (R) consisting of upper triangular matrices
and strictly upper triangular matrices, respectively.
We now determine Zg (T ), where g is one of the Lie algebras from above. Of course,
the case g = gld (O) is covered by Proposition 1.5. Next, clearly, Zsl1 (O) (T ) = 1/(1 − qT ).
The general case of sld (O) offers nothing new.
Corollary 5.8. Let d > 1. Then Zsld (O) (T ) = Zgld (O) (T ) =
1−q −d T
.
(1−T )2
Proof. It suffices to show that x · sld (O) ⊃ x · gld (O) for all x ∈ Od \ {0}. Let x` have
minimal valuation among the entries of x. Let ej ∈ Od be the jth unit vector. By our
proof of Lemma 5.2, x · gld (O) is generated by {x` ej : 1 6 j 6 d}. Note that x · sld (O)
is spanned by all xi ej and xi ei − xj ej for 1 6 i, j 6 d with i 6= j. It thus only remains
to show that x` e` ∈ x · sld (O). Since d > 1, we may choose j 6= `. Since |xj | 6 |x` |,
x
x` e` = (x` e` − xj ej ) + xj` x` ej ∈ x · sld (O) whence the claim follows.
Proposition 5.9. Let char(K) 6= 2. Then Zsod (O) (T ) = ZMd×(d−1) (O) (T ) =
1−q 1−d T
(1−T )(1−qT ) .
Remark 5.10.
(i) It is instructive to first determine ask(sod (Fq )) for odd q. If F is any field with
char(F ) 6= 2, then it is easy to see that x · sod (F ) = x⊥ for all x ∈ F d \ {0}, where
20
the orthogonal complement is taken with respect to the standard inner product. In
particular, if x 6= 0, then dimF (x · sod (F )) = d − 1. Using Lemma 2.1, we conclude
d −1
that ask(sod (Fq )) = 1 + qqd−1
= 1 − q 1−d + q for odd q; this identity was first
proved probabilistically by Fulman and Goldstein [28, Lem. 5.3]. We note that for
char(K) 6= 2 and x ∈ Od , while we still have an inclusion x · sod (O) ⊂ x⊥ , equality
does not, in general, hold; indeed, x⊥ is always an isolated submodule of Od .
(ii) While we assumed that char(K) 6= 2 in Proposition 5.9, we do allow char(O/P) = 2.
Note, however, that in this case, sod (O)n (= sod (O) ⊗ On ) is properly contained in
the set of all skew-symmetric d × d-matrices over On .
Proof of Proposition 5.9. Part (i) of the preceding remark implies that gor(sod (O)) =
d − 1. Given elements z1 , . . . , z` of some ring, we recursively define a matrix
−z2 z1
z1
−z3
..
.
m(z1 , . . . , z` ) := −z`
0
..
.
..
.
z1 .
m(z2 , . . . , z` )
0
By induction, we may assume that ±X2i , . . . , ±Xdi are i × i minors of m(X2 , . . . , Xd )
for all 1 6 i 6 d − 2 so that ±X1j , . . . , ±Xdj are j × j minors of m(X1 , . . . , Xd ) for
1 6 j 6 d − 1. Let eij ∈ Md (O) be the elementary matrix with 1 in position (i, j) and
zeros elsewhere. Then the eij − eji for 1 6 i < j 6 d generate sod (O) as an O-module
whence the rows of m(X1 , . . . , Xd ) span X · sod (O[X]). Thus, sod (O) is O-maximal by
Lemma 5.5 and the claim follows from Corollary 5.4.
For a ring R, let Symd (R) = {a ∈ Md (R) : a> = a}.
Proposition 5.11. ZSymd (O) (T ) = Zgld (O) (T ) =
1−q −d T
.
(1−T )2
Proof. This proof is similar to that of Proposition 5.9 and we use the same notation. By
considering the images of the first unit vector in Od under the matrices e11 and e1j + ej1
(2 6 j 6 d), we find that gor(Symd (O)) = d. Given z1 , . . . , z` , recursively define
0
m (z1 , . . . , z` ) :=
z1
z2
..
.
z`
0
..
.
0
21
z1
..
.
z1 .
m0 (z2 , . . . , z` )
An induction similar to the one in the proof of Proposition 5.9 shows that X1i , . . . , Xdi
are i × i minors of m0 (X1 , . . . , Xd ) for 1 6 i 6 d. As X · Symd (O[X]) is the row span of
m0 (X1 , . . . , Xd ) over O[X], the claim follows from Lemma 5.5 and Corollary 5.4.
Proposition 5.12. Zsp2d (O) (T ) = Zgl2d (O) (T ) =
1−q −2d T
.
(1−T )2
Proof. We proceed along the same lines as the preceding two proofs. Let eij denote the
usual elementary matrix, now of size 2d × 2d. Using these matrices, it is easy to see
that (1, 0, . . . , 0) · sp2d (O) = O2d whence gor(sp2d (O)) = 2d. As an O-module, sp2d (O)
is generated by the following matrices: (i) eij − ed+j,d+i (1 6 i, j 6 d), (ii) ei,d+i , ed+i,i
(1 6 i 6 d), and (iii) ei,d+j +ej,d+i , ed+i,j +ed+j,i (1 6 i < j 6 d). Write X = (X1 , . . . , Xd )
and X 0 = (X10 , . . . , Xd0 ). Define m0 (z1 , . . . , z` ) as in the proof of Proposition 5.11. Then
(X, X 0 ) · sp2d (O[X, X 0 ]) is generated by the rows of
−X10
..
.
X1
0
m̃(X, X ) :=
Xd
..
.
X1 −Xd0
..
.
..
..
.
−X10
..
.
.
−Xd0
m0 (X)
Xd
m0 (X 0 )
.
Using what we have shown about the minors of m0 (X) in the proof of Proposition 5.11,
we conclude that Xij and (Xi0 )j (i = 1, . . . , d; j = 1, . . . , 2d) arise as j × j minors of
m̃(X, X 0 ). Again, the claim thus follows from Lemma 5.5 and Corollary 5.4.
In contrast to the above examples, neither nd (O) nor trd (O) (for d > 1) is O-maximal.
Proposition 5.13.
(i) Znd (O) (T ) =
(1−T )d−1
.
(1−qT )d
(ii) Ztrd (O) (T ) = Znd+1 (O) (q −1 T ) =
(1−q −1 T )d
.
(1−T )d+1
Proof. Since nd+1 (O) is obtained from trd (O) by adding a zero row and a zero column,
by Corollary 3.6, it suffices to prove (i). Let ei denote the ith unit vector in Od and let
x ∈ Od \ {0}. Then x nd (O) is generated by
n
0, x1 , 0, . . . , 0 , 0, 0, gcd(x1 , x2 ), 0, . . . , 0 , . . . , 0, . . . , 0, gcd(x1 , . . . , xd−1 )
o
;
in particular, gor(nd (O)) = d − 1. Moreover, by (4.6) and Lemma 5.2,
Ond (O) (x, y) = |y|1−d · kx1 , yk · kx1 , x2 , yk · · · kx1 , . . . , xd−1 , yk.
Hence, using (4.6) and Lemma 5.6, (1 − q −1 )Znd (O) (q −s ) = Fd−1 (s − 2, −1, . . . , −1).
22
5.4 Diagonal matrices
Let dd (O) ⊂ gld (O) be the subalgebra of diagonal matrices. Clearly d1 (O) = gl1 (O) so
−1 T
that Zd1 (O) (T ) = 1−q
. It turns out that the functions Zdd (O) (T ) have essentially been
(1−T )2
computed by Brenti [9, Thm 3.4] in a different context; the author is grateful to Angela
Carnevale for pointing this out. We first recall some terminology. Let Bn = {±1}oSn be the
group of signed permutations on n letters, regarded
as a subgroup of the symmetric group
on {±1, . . . , ±n}. For σ ∈ Bn , let dB (σ) := # i ∈ {1, . . . , n} : (i − 1)σ > iσ (where we
P
wrote 0σ = 0), N(σ) := # i ∈ {1, . . . , n} : iσ < 0 , and Bn (X, Y ) :=
X N(σ) Y dB (σ) .
σ∈Bn
Theorem 5.14 ([9, Thm 3.4(ii)]).
∞
P
(i(X + 1) + 1)n Y i =
i=0
Bn (X,Y )
(1−Y )n+1
for n > 1.
The following marks a departure from the simplicity of previous examples of ZM (T ).
Corollary 5.15. Zdd (O) (T ) =
Bd (−q −1 ,T )
.
(1−T )d+1
Proof. By Corollary 3.6, Zdd (O) (T ) is the dth Hadamard power of Zd1 (O) (T ). Since
Zd1 (O) (T ) =
1−q −1 T
(1−T )2
=
P∞
i=0 (1
+ i − iq −1 )T i , the claim follows from Theorem 5.14.
Example 5.16.
(i) Zd2 (O) (T ) =
1 + T − 4q −1 T + q −2 T 2 + q −2 T
.
(1 − T )3
(ii) Zd3 (O) (T ) =
1 + 4T − 12q −1 T + 6q −2 T − q −3 T + T 2 − 6q −1 T 2 + 12q −2 T 2 − q −3 T 3 − 4q −3 T 2
.
(1 − T )4
We note that permutation statistics have previously featured in explicit formulae for
representation zeta functions [10, 61]; see also [11].
6 Constant rank spaces
By a constant rank space over a field F , we mean a subspace M ⊂ Md×e (F ) such
that all non-zero elements of M have the same rank, say r; we then say that M has
constant rank r. Such spaces have been studied extensively in the literature (see e.g.
[4, 7, 39, 64, 70]), often in the context of vector bundles on projective space. A problem
of particular interest is to find, for given d and r, the largest possible dimension of a
subspace of Md (C) of constant rank r. Apart from trivial examples such as band matrices
(see Example 6.6 below), the construction of constant rank spaces (in particular those
of large dimension) seems to be challenging. Note that if M ⊂ Md×e (Fq ) has constant
rank r and dimension `, then
ask(M ) = q −` q d + (q ` − 1)q d−r = q d−` + q d−r − q d−`−r .
(6.1)
In §6.1, we consider a natural analogue, K-minimality, of the concept of O-maximality
introduced and studied in §5.1. We then derive conceptual interpretations of both of
these notions in a global setting in §6.2—in particular, we will see that K-minimality is
closely related to constant rank spaces.
23
6.1 K-minimality
Let O be the valuation ring of a non-Archimedean local field K of arbitrary characteristic.
Let M ⊂ Md×e (O) be a submodule. Recall the definition of KM from Definition 4.4.
Lemma 6.1. Let F : O` → M be an O-module isomorphism, w ∈ O` \ {0}, and
y ∈ O \ {0}. Then KM (F (w), y) > |y|grk(M )−d kw, yk− grk(M ) .
Proof. Let a := F (w) have equivalence type (λ1 , . . . , λr ) (see §4.3.1); of course, r 6
grk(M ). Let m := min(ν(w1 ), . . . , ν(w` )), n = ν(y), and a := F (w). Then m 6
min(ν(aij ) : 1 6 i, j 6 d) = λ1 6 λ2 6 · · · 6 λr and Lemma 4.6(iii) shows that
KM (F (w), y) > q r min(m,n)+(d−r)n > q grk(M ) min(m,n)+(d−grk(M ))n
= |y|grk(M )−d kw, yk− grk(M ) .
Definition 6.2. We say that M is K-minimal if there exists an O-module isomorphism
F : O` → M such that KM (F (w), y) = |y|grk(M )−d kw, yk− grk(M ) for all (w, y) ∈ O` × O
outside a set of measure zero.
Clearly, if x ∈ On and a ∈ GLn (O), then kxak = kxk. We conclude that if the
condition in the preceding definition is satisfied for some isomorphism F : O` → M , then
it holds for all of them. By the elementary divisor theorem we may thus characterise
K-minimality as follows.
Lemma 6.3. M is K-minimal if and only if M is an isolated submodule of Md×e (O)
and KM (a, y) = |y|grk(M )−d ka, yk− grk(M ) for all (a, y) ∈ M × O outside a set of measure
zero.
Lemma 5.6 now implies the following.
Proposition 6.4. Let M be K-minimal. Let r = grk(M ) and ` = dimK (M ⊗ K). Then
ZM (T ) =
1 − q d−`−r T
.
(1 − q d−` T )(1 − q d−r T )
The following sufficient condition for K-minimality is proved similarly to Lemma 5.5.
Lemma 6.5. Let (a1 , . . . , a` ) be an O-basis of M . Suppose that for 1 6 i 6 grk(M ), the
ideal generated by the i × i minors of X1 a1 + · · · + X` a` ∈ Md×e (O[X1 , . . . , X` ]) contains
X1i , . . . , X`i . Then M is K-minimal.
Example 6.6 (Band matrices). Let r > 1 and define
x1
x
2
.
..
Br =
x
r
..
..
.
.
..
.
..
.
x1
: x1 , . . . , xr ∈ O ⊂ M(2r−1)×r (O).
x2
..
.
xr
By Lemma 6.5 and Proposition 6.4 (with d = 2r − 1 and ` = r), ZBr (T ) =
24
1−q −1 T
.
(1−q r−1 T )2
6.2 A global interpretation
Henceforth, let k be a number field with ring of integers o; recall the notation from §3.3.
Let k̄ be an algebraic closure of k. Let M ⊂ Md×e (o) be a submodule. The following can
be proved similarly to Proposition 4.8 (and Proposition 4.12).
Lemma 6.7.
(i) max rkk (a) = max rkk̄ (ā) = grk(Mv ) for all v ∈ Vk .
a∈M
ā∈M ⊗o k̄
(ii) max dimk (x · (M ⊗o k)) = max dimk̄ (x̄ · (M ⊗o k̄)) = gor(Mv ) for all v ∈ Vk .
x∈od
x̄∈k̄d
Extending Definitions 3.2 and 4.7, we let grk(M ) and gor(M ) be the common number
in (i) and (ii), respectively. K-minimality is closely related to constant rank spaces:
Proposition 6.8.
(i) Let (a1 , . . . , a` ) ⊂ M be a k-basis of M ⊗o k. Let Ii be the ideal of k[X1 , . . . , X` ]
generated by the i × i minors of X1 a1 + · · · + X` a` . Then M ⊗o k̄ is a constant
rank space if and only if X1i , . . . , Xdi ∈ Ii for i = 1, . . . , grk(M ).
(ii) If M ⊗o k̄ is a constant rank space, then Mv is K-minimal for almost all v ∈ Vk .
Our proof of Proposition 6.8 relies on the following observation.
Lemma 6.9. Let F be a field. Let I = hf1 , . .√. , fr i / F [X] = F [X1 , . . . , Xd ], where each
fi is homogeneous of degree n. Suppose that I = hX1 , . . . , Xd i. Then X1n , . . . , Xdn ∈ I.
Proof. It clearly suffices to show that Xdn ∈ I. We freely use basic facts from the theory
of Gröbner bases; see e.g. [14]. Throughout, we consider the lexicographic monomial
order with Xd ≺ · · · ≺ X1 . Let lt(f ) denote the leading term of a non-zero f ∈ F [X].
Let G be the reduced Gröbner basis of I. Since each fi is homogeneous of degree n,
Buchberger’s algorithm and the uniqueness of G show that each element of G is also
homogeneous of degree n. As I is 0-dimensional, we may write G = {g1 , . . . , g` } such
that gi ∈ F [Xi , . . . , Xd ] for 1 6 i 6 d and such that each √
lt(gi ) is a power of Xi ; since
gi is homogeneous of degree n, lt(gi ) = Xin . Since Xd ∈ I, there exists m > 1 with
Xdm ∈ I. The multivariate division algorithm modulo G thus reduces Xdm to 0. As G
is reduced, the only element of G whose leading term is a power of Xd is gd . Since the
latter belongs to F [Xd ], the division algorithm actually takes places within F [Xd ] whence
Xdm is divisible by gd . This only leaves the possibility gd = Xdn ∈ I.
Proof of Proposition 6.8.
(i) Let V(I) ⊂ k̄ ` be the algebraic set corresponding to I / k[X] := k[X1 , . . . , X` ] and
let r = grk(M ). Then M ⊗o k̄ is a constant rank space if and only
√ if V(Ii ) ⊂ {0} for
i = 1, . . . , r which, by Hilbert’s Nullstellensatz, is√equivalent to Ii ⊃ hX1 , . . . , Xr i
for i = 1, . . . , r. By the preceding lemma, Xj ∈ Ii if and only if Xji ∈ Ii .
25
(ii) Lemma 5.5 shows that (i) implies (ii); note that if f ∈ o[X] belongs to the k[X]-ideal
generated by g1 , . . . , gr ∈ o[X], then f also belongs to the ov [X]-ideal generated
by the gi , at least for almost all v ∈ Vk .
In view of Lemma 2.1, we say that a subspace M 0 ⊂ Md×e (F ) (where F is a field)
has constant orbit dimension if all F -spaces xM 0 for x ∈ F d \ {0} have the same
dimension. The following counterpart of Proposition 6.8 is then proved in the same way.
Proposition 6.10. Let (a1 , . . . , a` ) ⊂ M be a k-basis of M ⊗o k. Let Ji be the ideal of
k[X] := k[X1 , . . . , X` ] generated by the i × i minors of
Xa1
..
. .
Xa`
Then M ⊗o k̄ has constant orbit dimension if and only if X1i , . . . , Xdi ∈ Ji for i =
1, . . . , gor(M ). In that case, Mv is O-maximal for almost all v ∈ Vk .
7 Orbits and conjugacy classes of linear groups
In this section, we use p-adic Lie theory to relate ask, orbit-counting, and conjugacy class
zeta functions. In §7.1, we recall properties of saturable pro-p groups and Lie algebras.
In §7.2, we prove that orbit-counting zeta functions over Zp are rational. In §7.3 we
compare group stabilisers and Lie centralisers under suitable hypotheses and this allows
us to deduce Theorem 1.6 in §7.4. Finally, in §7.5, we prove Theorem 1.7.
7.1 Reminder: saturable pro-p groups and Lie algebras
We briefly recall Lazard’s [48] notion of (p-)saturability of groups and Lie algebras using
González-Sánchez’s [30] equivalent formulation.
Let g be a Lie Zp -algebra whose underlying Zp -module is free of finite rank. A potent
filtration of g is a central series g = g1 ⊃ g2 ⊃ · · · of ideals (i.e. [gi , g] ⊂ gi+1 for all i)
with
∞
T
gi = 0 and such that [gi ,p−1 , g] := [gi , g, . . . , g] := [. . . [[gi , g], g], . . . , g] ⊂ pgi+1
i=1
| {z }
p−1
for all i > 1. We say that g is saturable if it admits a potent filtration.
Proposition 7.1. ( [1, Prop. 2.3]) Let O be the valuation ring of a non-Archimedean local
field K of characteristic zero and residue characteristic p. Let e denote the ramification
index of K/Qp . Let g be a Lie O-algebra whose underlying O-module is free of finite
e
rank. Let m > p−1
. Then gm (= Pm g) is saturable as a Zp -algebra.
We note that, as before, subalgebras of an O-algebra are understood to be Osubalgebras; whenever we consider Zp -subalgebras, we will explicitly state as much.
26
Similarly to the case of Lie algebras, a torsion-free finitely generated pro-p group G
which admits a central series G = G1 > G2 > · · · of closed subgroups with
Gpi+1
∞
T
Gi = 1
i=1
and [Gi ,p−1 G] 6
is saturable.
If g is a saturable Lie Zp -algebra, then the underlying topological space of g can
be endowed with the structure of a saturable pro-p group using the Hausdorff series.
Conversely, every saturable pro-p group gives rise to a saturable Lie Zp -algebra and
these two functorial operations furnish mutually quasi-inverse equivalences between
the categories of saturable Lie Zp -algebras and saturable pro-p groups (defined as full
subcategories of all Lie Zp -algebras and pro-p groups, respectively); see [30, §4] for an
overview and [48, Ch. 4] for details. While the general interplay between subalgebras
and subgroups is subtle, we note the following fact.
Lemma 7.2. Let g be a saturable Lie Zp -algebra and let h be a saturable subalgebra of
finite additive index. Let G and H be the saturable pro-p groups associated with g and h
via the Hausdorff series (so that H 6 G). Then |G : H| = |g : h|.
Proof. This can be proved in the same way as [31, Lem. 3.2(4)].
7.2 Orbits of p-adic linear groups
Let O be the valuation ring of a non-Archimedean local field K. Recall that P denotes the
maximal ideal of O and that q and p denote the size and characteristic of the residue field
of K, respectively. Further recall the definition of Zoc
G (T ) (Definition 1.2(ii)). Although
we will not need it in the sequel, since it might be of independent interest, we note the
following rationality statement for Zoc
G (T ), a consequence of a powerful model-theoretic
result due to Cluckers [37, App. A].
Theorem 7.3. Let G 6 GLd (Zp ). Then ZG (T ) ∈ Q(T ). More precisely, there are
Q
a1 , . . . , ar ∈ Z and b1 , . . . , br ∈ N such that ri=1 (1 − pai T bi )Zoc
G (T ) ∈ Q[T ].
Lemma. Let Ḡ be the closure of G 6 GLd (O) in GLd (O). Then Zoc
(T ) = Zoc
G (T ).
Ḡ
Proof. The GLid (O) form a fundamental system of neighbourhoods of the identity whence
the claim follows from G · GLid (O) = Ḡ · GLid (O) for all i > 0.
Proof of Theorem 7.3. Define an equivalence relation ∼n on Zdp via
x ∼n y
:⇔
∃g ∈ G. x ≡ yg
(mod pn ).
Our theorem will follow immediately from [37, Thm A.2] once we have established that
∼n is definable (uniformly in n) in the subanalytic language used in [37, App. A]. By
the preceding lemma, we may assume that G = Ḡ. It then follows from the well-known
structure theory of p-adic analytic groups (see [19, 48]) that there exists an open saturable
(or, more restrictively, uniform) subgroup H 6 G of the form H = exp(p2 h), where
27
h ⊂ gld (Zp ) is a suitable saturable (or uniform) Zp -subalgebra. Let T be a transversal
for the right cosets of H in G. Then, for x, y ∈ Zdp , x ∼n y if and only if
_
∃a ∈ h. x exp(p2 a) ≡ yt−1
(mod pn ).
t∈T
The claim thus follows since the d × d-matrix exponential exp(p2 X) is given by d2 power
series in d2 variables which all converge on gld (Zp ).
Remark 7.4. Let K/Qp be a finite extension. Then we may regard G 6 GLd (O) as
a Zp -linear group of degree d|K : Qp | via the regular representation of O over Zp . In
particular, Theorem 7.3 implies that Zoc
G (T ) is rational provided that K/Qp is unramified.
We conclude this subsection with some elementary examples of Zoc
G (T ).
Example 7.5.
(i) It is easy to see that the rule x 7→ min(ν(x1 ), . . . , ν(xd ), n) induces a bijection
Odn / GLd (O) → {0, . . . , n} whence
∞
X
Zoc
GLd (O) (T ) =
(n + 1)T n =
n=0
1
.
(1 − T )2
(ii) Let p =
6 2. The number of orbits of h−1i 6 GL1 (O) on On is 1 + (q n − 1)/2 so that
Zoc
h−1i (T ) =
1
1 1
1
2 − qT − T
+
−
=
.
1−T
2 1 − qT
1−T
2(1 − qT )(1 − T )
(iii) Let G = h[ 01 10 ]i 6 GL2 (O). It is easy to see that
Zoc
G (T ) =
∞
X
q n (q n + 1) n
T =
n=0
2
2 − q 2 T − qT
.
2(1 − q 2 T )(1 − qT )
7.3 Lie centralisers and group stabilisers
Let O be the valuation ring of a non-Archimedean local field K ⊃ Qp . Let e denote the
ramification index of K/Qp . As expected, for suitable matrix algebras and groups, the
equivalence between saturable pro-p groups and Lie Zp -algebras recalled in §7.1 can be
made explicit using exponentials and logarithms.
In line with our previous notation (see §3.5), we write glm
d (O) := {a ∈ gld (O) : a ≡ 0
(mod Pm )}. Moreover, we define GLm
(O)
:=
{a
∈
GL
(O)
: a ≡ 1 (mod Pm )}.
d
d
e
Proposition 7.6 ([44, Lem. B.1]). Let m > p−1
. The formal exponential and logam
m
rithm series converge on gld (O) and GLd (O), respectively, and define mutually inverse
m
m
m
bijections exp : glm
d (O) → GLd (O) and log : GLd (O) → gld (O).
e
Hence, if m > p−1
and g ⊂ glm
d (O) is a saturable subalgebra, then we may identify
the saturable pro-p group associated with g in the sense of §7.1 with exp(g) 6 GLm
d (O).
For x ∈ Od and n > 0, we write x mod Pn for the image of x in Odn .
28
Lemma 7.7. Let g ⊂ gled (O) (= p gld (O)) be a saturable subalgebra. Then
cg (x mod Pn ) = a ∈ g : xa ≡ 0
(mod Pn )
is a saturable subalgebra of g for all x ∈ Od .
Proof. Write cn := cg (x mod Pn ); obviously, cn is a subalgebra of g. Let g = γ1 (g) ⊃
γ2 (g) ⊃ · · · be the lower central series of g. Then γp (cn ) ⊂ cn+e . Indeed, each element,
a say, of γp (cn ) is a sum of matrix products c(ph) for suitable c ∈ cn and h ∈ gld (O) and
clearly, xc(ph) ≡ 0 (mod Pn+e ). Let g = g1 ⊃ g2 ⊃ · · · be a potent filtration of g. Then
cn = cn ∩g1 ⊃ cn ∩g2 ⊃ · · · is a central series of cn with
∞
T
(cn ∩gi ) = 0. It is in fact a potent
i=1
filtration for pg∩cn+e = pcn whence [(cn ∩gi ),p−1 cn ] ⊂ pgi+1 ∩γp (cn ) ⊂ pg∩cn+e = pcn .
Lemma 7.8. Let m >
exp(a) = 1 + au.
e
p−1
1
and a ∈ glm
d (O). Then there exists u ∈ GLd (O) with
m
1
i
Proof. Let g(X) := ∞
i=0 (i+1)! X so that exp(X) = 1 + Xg(x) in Q[[X]]. Let a ∈ gld (O)
be non-zero. Let b be an entry of a of minimal valuation. Then νK (bi /(i + 1)!) is a
1
ai for i > 0. It is well-known that
lower bound for the valuation of each entry of (i+1)!
νp ((i + 1)!) 6 i/(p − 1) (see e.g. [12, Lem. 4.2.8(1)]) so that νK ((i + 1)!) 6 ei/(p − 1).
Therefore, νK (bi /(i + 1)!) > i(νK (b) − e/(p − 1)) =: f (i). Clearly, f (i) → ∞ as i → ∞
and f (i) > 0 for i > 0. Hence, g(a) converges in Md (O) to an element of GL1d (O).
P
By combining the preceding two lemmas, we obtain the following.
e
), let g ⊂ glm
Corollary 7.9. Let m > max(e − 1, p−1
d (O) be a saturable subalgebra, and
d
n
let x ∈ O . Then exp(cg (x mod P )) = Stexp(g) (x mod Pn ).
Proof. Let a ∈ g and write exp(a) = 1 + au for u ∈ GL1d (O). Then
a ∈ cg (x mod Pn ) ⇔ xa ≡ 0
⇔ xau ≡ 0
(mod Pn )
(mod Pn )
⇔ x exp(a) ≡ x(1 + au) ≡ x (mod Pn )
⇔ exp(a) ∈ Stexp(g) (x mod Pn ).
7.4 Proof of Theorem 1.6
Let O be the valuation ring of a non-Archimedean local field K ⊃ Qp and let e be the
ramification index of K/Qp .
e
Proposition 7.10. Let m > max(e − 1, p−1
), let g ⊂ glm
d (O) be a saturable subalgebra,
and let G = exp(g). Then Zg (T ) = Zoc
(T
).
G
29
Proof. Write V = Od and Vn = V ⊗ On . Recall that gn denotes the image of g under
the natural map gld (O) → gld (On ). For n > 0, by combining Lemma 2.1, Lemma 7.2,
and Corollary 7.9, we obtain
|Vn /G| =
X
|xG|−1 =
x∈Vn
=
X
X
x∈Vn
|g : cg (x)|
−1
x∈Vn
=
X
|G : StG (x)|−1
=
X
|gn : cgn (x)|−1
x∈Vn
−1
|x gn |
= ask(gn | Vn ).
x∈Vn
e
Proof of Theorem 1.6. Let m > max(e − 1, p−1
). Propositions 3.9 and 7.10 show that
oc,m
dm
m
q Zg (T ) = Zgm (T ) = Zexp(gm ) (T ) ∈ Z[[T ]].
7.5 Orbits and conjugacy classes of unipotent groups
Let O be the valuation ring of a finite extension K/Qp . Recall that nd (O) ⊂ gld (O) is
the Lie algebra of all strictly upper triangular matrices and that Ud denotes the group
scheme of upper unitriangular d × d matrices. The following is well-known.
Proposition 7.11. Let p > d.
(i) All Zp -subalgebras of nd (O) and closed subgroups of Ud (O) are saturable.
(ii) exp and log define polynomial bijections between nd (O) and Ud (O).
Proof. Noting that subalgebras of nd (O) and closed subgroups of Ud (O) are nilpotent
of class at most d − 1 < p, their lower central series constitute potent filtrations. This
proves (i). Part (ii) follows since ad = 0 for a ∈ nd (O).
A simple variation of Proposition 7.11 yields the following.
Corollary 7.12. Let g ⊂ gld (O) be a subalgebra. Suppose that g is nilpotent of class at
most c and that ac+1 = 0 (in Md (O)) for all a ∈ g. Further suppose that p > c.
(i) g is saturable.
(ii) G := exp(g) is a saturable subgroup of GLd (O).
(iii) exp and log define mutually inverse polynomial bijections between g and G.
By going through the proof of Theorem 1.6, we now easily obtain the following.
oc
Corollary 7.13. Let the notation be as in Corollary 7.12. Then Zask
g (T ) = Zexp(g) (T )
and thus, in particular, Zask
g (T ) ∈ Z[[T ]].
This proves the first part of Theorem 1.7. We note that Corollary 5.15 shows that we
may not, in general, relax the assumptions in Corollary 7.13 and merely assume that
g ⊂ gld (O) is a nilpotent subalgebra. The following completes the proof of Theorem 1.7.
30
Proposition 7.14. Let g ⊂ gld (O) be an isolated subalgebra consisting of nilpotent
ask
matrices. Suppose that p > d. Then Zcc
exp(g) (T ) = Zad(g) (T ).
Proof. By Engel’s theorem, g is GLd (K)-conjugate to a subalgebra of nd (K) and hence
nilpotent of class at most d − 1; in particular, ad(a)d = 0 for all a ∈ g. By Corollary 7.12,
G := exp(g) is a saturable subgroup of GLd (O).
Let Ad : G → GL(g) denote the adjoint representation of G; hence, Ad(g) : g → g, a 7→
log(exp(a)g ) = ag for g ∈ G. Recall that Gn denotes the image of G in GLd (On ) and gn
that of g in gld (On ). Clearly, a ≡ 0 (mod Pn ) if and only if exp(a) ≡ 1 (mod Pn ) for
a ∈ g. We may thus identify conjugacy classes of Gn with Ad(G)-orbits on gn . As g is
oc
isolated within gld (O), we may identify gn = g ⊗ On and obtain Zcc
G (T ) = ZAd(G)yg (T ).
By Corollary 7.12, ad(g) is a saturable subalgebra of gl(g). The Hausdorff series
P
1
shows that log(exp(b)exp(a) ) = ∞
i=0 i! [b,i a] for a, b ∈ g (see [31, Eqn (3)]) whence
Ad(exp(a)) = exp(ad(a)) for all a ∈ g. Thus, Ad(G) = exp(ad(g)) and Corollary 7.13
ask
ask
shows that Zoc
Ad(G)yg (T ) = Zlog(Ad(G))yg (T ) = Zad(g)yg (T ).
Using the well-known equivalence between unipotent algebraic groups and nilpotent
finite-dimensional Lie algebras over a field of characteristic zero (see [16, Ch. IV]),
Corollary 7.13 and Proposition 7.14 now imply the following global result.
Corollary 7.15. Let k be a number field with ring of integers o. Let G 6 Ud ⊗Z k be
a unipotent algebraic group over k and let G 6 Ud ⊗Z o be the associated o-form of G
(i.e. the schematic closure of G). Let g ⊂ nd (k) be the Lie algebra of G and g = g ∩ nd (o).
ask
cc
ask
Then for almost all v ∈ Vk , Zoc
G(ov ) (T ) = Zgv (T ) and ZG(ov ) (T ) = Zad(gv ) (T ).
Note, in particular, that Theorem 4.15 therefore establishes functional equations for
orbit-counting and conjugacy class zeta functions arising from unipotent algebraic groups.
Corollary 7.16. Let k be a number field with ring of integers o. Let G 6 Ud ⊗Z k
and H 6 Ue ⊗Z k be unipotent algebraic groups over k with o-forms G and H as above.
cc
Suppose that for almost all v ∈ Vk , Zcc
G(ov ) (T ) = ZH(ov ) (T ). Then dimk (G) = dimk (H).
Proof. Theorem
4.15, Corollary 7.15, and [56, §5] allow us to recover dimk (G) from
Zcc
(T
)
for any finite set S ⊂ Vk .
G(ov )
v∈V \S
k
We note that there are examples of non-isomorphic groups G and H (of the same
cc
dimension) which satisfy Zcc
G(ov ) (T ) = ZH(ov ) (T ) for almost all v ∈ Vk ; see Table 2 in §8.3.
8 Further examples
8.1 Computer calculations: Zeta
The author’s software package Zeta [59] for Sage [62] can compute numerous types of
“generic local” zeta functions in fortunate (“non-degenerate”) cases. The techniques
used by Zeta were developed over the course of several papers; see [57] for an overview
and references to other pieces of software that Zeta relies upon. When performing
31
computations, Zeta proceeds by attempting to explicitly compute certain types of p-adic
integrals. Fortunately, the integrals in (4.3) and (4.6) can both be encoded in terms
of the “representation data” introduced in [58, §5] whence the author’s computational
techniques apply verbatim to the functions ZM (T ) for O = Z. In detail, given a
submodule M ⊂ Md×e (Z), Zeta can be used to attempt to construct a rational function
W (X, T ) ∈ Q(X, T ) with the following property: for almost all primes p and all finite
extensions K/Qp , ZM ⊗Z OK (T ) = W (qK , T ); we note that for various reasons, Zeta may
well fail to construct W (X, T ) even if it exists. Given M ⊂ Md×e (Z), Zeta can also be
used to attempt to construct a formula as in Corollary 4.11.
The remainder of this section is devoted to a number of examples of functions ZM (T )
and Zcc
G (T ) (via Theorem 1.7) computed with the help of Zeta. Throughout, K is a
non-Archimedean local field of characteristic zero with valuation ring O, residue field
size q, and residue characteristic p.
8.2 Examples of ask zeta functions
Example 8.1 (Small poles and unbounded denominators). Let
a b
M = b c
a d
a
d : a, b, c, d ∈ O .
c
Then for sufficiently large p,
ZM (T ) =
1 + 5q −1 T − 12q −2 T + 5q −3 T + q −4 T 2
.
(1 − q −1 T )(1 − T )2
Hence, the real poles of ζM (s) are −1 and 0; it is easy to see that gor(M ) = 3 (see
Definition 3.2). This example illustrates that, in contrast to the case of Md×e (O),
d − gor(M ) is generally not a lower bound for the real poles of ζM (s). Note that ZM (T )
has unbounded denominators—the author has found comparatively few modules of square
matrices with this property (and initially suspected they did not exist).
Example 8.2 (Y 2 = X 3 − X). Let c(q) denote the number of points in P2 (Fq ) on the
elliptic curve Y 2 = X 3 − X. In [20, 21], du Sautoy constructed a nilpotent group of Hirsch
length 9 whose local (normal) subgroup zeta function at a prime p depends on c(p). Due
to the “wild” behaviour of c(p) as a function of p, he thus disproved earlier predictions
on the growth of (normal) subgroups of nilpotent groups. His construction has since
been used to demonstrate that other group-theoretic counting problems can be “wild”
(e.g. the enumeration of representations [68, Ex. 2.4] or of “descendants” [24])—we will
now see that the present setting is no exception. Namely, let
z
E = x
y
x y
z 0 : x, y, z ∈ O .
0 x
For sufficiently large p,
32
1 + c(q) · q −1 − 1 − 2c(q) · q −2 + 2q −1 + c(q) · q −3 − 4q −2 T + q −3 T 2 + q −3 T
ZE (T ) =
.
(1 − T )3
In particular, despite all the examples of functions ZM (T ) that we encountered so far,
Corollary 4.11 accurately reflects the general dependence of ZMv (T ) on a place v ∈ Vk
for a module of matrices M over the ring of integers o of a number field k. However, just
as in the study of zeta functions of groups, it remains unclear if anything meaningful can
be said about the varieties Vi ⊗o k “required” to produce formulae (4.4) as M varies over
all modules of matrices over o.
Note that we may derive a group-theoretic counting problem (namely,
the enumeration
1 E
∗
of orbits) from E at the cost of doubling the dimension: for E := 0 1 6 GL6 (O), the
second proof of Lemma 2.1 shows that ZE ∗ (T ) = ZE (q 3 T ). This construction of a group
attached to an elliptic curve previously featured in work of Boston and Isaacs [8] on
numbers of conjugacy classes of p-groups of a given order.
Example 8.3. Suppose that p 6= 2 and let
0 x2
0 0
0 0
M=
0 0
0 0
0
0
−x3 0
0
x1 x22 − x23
0
x1
0
0
0
x1
0
0
0
0
0
0
x1
x5
x4
: x1 , . . . , x5 ∈ O .
x3
x2
0
For sufficiently large p,
ZM (T ) = (+q 36 T 19 − 4q 35 T 19 − q 34 T 20 + q 35 T 18 + 8q 34 T 19 − 2q 34 T 18 − 2q 33 T 19 − q 34 T 17
− 6q 33 T 18 − q 32 T 19 + 3q 33 T 17 + 5q 32 T 18 + 3q 32 T 17 + 6q 31 T 18 − q 32 T 16 − 12q 31 T 17
− 2q 30 T 18 − 9q 30 T 17 − q 31 T 15 + 14q 30 T 16 + 14q 29 T 17 + 4q 30 T 15 + 5q 29 T 16
− 3q 28 T 17 − 14q 29 T 15 − 41q 28 T 16 + q 29 T 14 + 12q 28 T 15 + 26q 27 T 16 − 2q 28 T 14
+ 46q 27 T 15 − 4q 26 T 16 − 7q 27 T 14 − 73q 26 T 15 + 2q 27 T 13 − 24q 26 T 14 + 32q 25 T 15
− 2q 26 T 13 + 103q 25 T 14 − 3q 24 T 15 + 6q 25 T 13 − 98q 2 4T 14 − q 2 5T 1 2 − 89q 2 4T 13
+ 29q 23 T 14 + 8q 24 T 12 + 176q 23 T 1 3 − 2q 22 T 1 4 + 35q 23 T 12 − 115q 22 T 13 + q 23 T 11
− 178q 22 T 12 + 25q 21 T 13 − 15q 22 T 11 + 223q 21 T 12 − 2q 20 T 13 + 119q 21 T 11
− 100q 20 T 12 + q 21 T 10 − 262q 20 T 11 + 16q 19 T 12 − 39q 20 T 10 + 214q 19 T 11 − q 18 T 12
+ 176q 19 T 10 − 61q 18 T 11 + 3q 19 T 9 − 280q 18 T 10 + 3q 17 T 11 − 61q 18 T 9 + 176q 17 T 10
− q 18 T 8 + 214q 17 T 9 − 39q 16 T 10 + 16q 17 T 8 − 262q 16 T 9 + q 15 T 10 − 100q 16 T 8
+ 119q 15 T 9 − 2q 16 T 7 + 223q 15 T 8 − 15q 14 T 9 + 25q 15 T 7 − 178q 14 T 8 + q 13 T 9
− 115q 14 T 7 + 35q 13 T 8 − 2q 14 T 6 + 176q 13 T 7 + 8q 12 T 8 + 29q 13 T 6 − 89q 12 T 7
− q 11 T 8 − 98q 12 T 6 + 6q 11 T 7 − 3q 12 T 5 + 103q 11 T 6 − 2q 10 T 7 + 32q 11 T 5 − 24q 10 T 6
33
+ 2q 9 T 7 − 73q 10 T 5 − 7q 9 T 6 − 4q 10 T 4 + 46q 9 T 5 − 2q 8 T 6 + 26q 9 T 4 + 12q 8 T 5 + q 7 T 6
− 41q 8 T 4 − 14q 7 T 5 − 3q 8 T 3 + 5q 7 T 4 + 4q 6 T 5 + 14q 7 T 3 + 14q 6 T 4 − q 5 T 5
− 9q 6 T 3 − 2q 6 T 2 − 12q 5 T 3 − q 4 T 4 + 6q 5 T 2 + 3q 4 T 3 + 5q 4 T 2 + 3q 3 T 3 − q 4 T
− 6q 3 T 2 − q 2 T 3 − 2q 3 T − 2q 2 T 2 + 8q 2 T + qT 2 − q 2 − 4qT + T
/ q 2 (1 − q 10 T 5 )(1 − q 8 T 4 )(1 − q 5 T 3 )(1 − q 4 T 2 )2 (1 − q 3 T 2 )(1 − q 2 T )(1 − qT )2 .
Since ZM (T ) = 1 + (2q 2 + 4q + 4q −1 − q −2 − 8)T + O(T 2 ), we see that, in contrast to
O-maximal or K-minimal cases (see §5–6), the complexity of ask(M1 | V1 ) is in general a
poor indicator of that of ZM (T ).
We note that by Corollary 7.13 and since ZM (T ) 6∈ Z[[T ]], the module M cannot be a
Lie subalgebra of n6 (O). Indeed, this is readily verified directly even though M is listed
as a Lie algebras in [55, Table 5].
Rojas’s article [55] provides numerous examples of Lie subalgebras g ⊂ nd (Z), say, for
d 6 6. For many of these, we may use Zeta to compute Zg⊗Z Zp (T ). Here, we only include
one example.
Example 8.4. Let p 6= 2 and let
0 x1
0 0
g = 0 0
0 0
0
0
x2
2
x1
0
0
0
− x23
0
x1
0
0
x5
x4
x3 : x1 , . . . , x5 ∈ O .
x2
0
Then g is a Lie subalgebra of n5 (O) of nilpotency class 4, listed as L5,6 (de Graaf’s [15]
notation) in [55, Table 3]. For sufficiently large p,
Zg (T ) =
+ q 8 T 7 − 3q 8 T 6 + q 8 T 5 + q 7 T 6 + 2q 7 T 5 − 2q 6 T 5 − 2q 6 T 4 − q 5 T 5 + 6q 5 T 4
− 3q 4 T 4 − 3q 4 T 3 + 6q 3 T 3 − q 3 T 2 − 2q 2 T 3 − 2q 2 T 2 + 2qT 2 + qT + T 2
− 3T + 1 / (1 − q 5 T 3 )(1 − q 4 T 2 )(1 − q 2 T )(1 − qT )2 .
Numerous examples (including the case of Md×e (O)) show that ζM (s) may have a pole
at zero and Example 8.1 shows that negative poles can arise even for modules of square
matrices. In contrast, all of the author’s computations are consistent with the following
question having a positive answer.
Question 8.5. Let k be a number field with ring of integers o. Let g ⊂ nd (o) be a Lie
subalgebra. Is it the case that for almost all v ∈ Vk , every real pole of ζgv (s) is positive?
Supposing that Question 8.5 indeed has a positive answer, if G 6 Ud ⊗Z k is an
algebraic group over k with associated o-form G 6 Ud ⊗Z o (see Corollary 7.15), then we
−s
may evaluate the meromorphic continuation of Zoc
G(ov ) (q ) at s = 0 for almost all v ∈ Vk .
It would then be interesting to see if one can interpret the resulting rational numbers,
say in terms of properties of the orbit space odv /G(ov ).
34
8.3 Examples of conjugacy class zeta functions
Let k be a number field with ring of integers o. Morozov [52] classified nilpotent Lie
algebras of dimension at most 6 over an arbitrary field of characteristic zero—equivalently,
he classified unipotent algebraic groups of dimension at most 6 over these fields. A recent
computer-assisted version of this classification (valid for fields of characteristic 6= 2) is
due to de Graaf [15]. We use his notation and let Ld,i (or Ld,i (a)) denote the ith Lie
k-algebra (with parameter a) given in [15, §4].
Table 1 provides a complete list of “generic conjugacy class zeta function” associated
with nilpotent Lie k-algebras of dimension at most 5 in the following sense: for each such
algebra g, let G be its associated unipotent algebraic group over k. After choosing an
embedding G 6 Ud ⊗Z k, we obtain an o-form G of G as in Corollary 7.15. Then for
almost all v ∈ Vk and all finite extensions K/kv , Zcc
G(O) (T ) is given in Table 1.
In contrast to dimension at most 5, Zeta is unable to compute generic conjugacy class
zeta functions associated with every nilpotent Lie k-algebra of dimension 6. Nevertheless,
Table 2 contains numerous examples of such functions; we only included examples
corresponding to ⊕-indecomposable algebras. Clearly, generic conjugacy class zeta
functions of direct products of algebraic groups are Hadamard products of the zeta
functions corresponding to the factors.
We note that L3,2 ≈ n3 (K) and L6,19 (−1) ≈ n4 (K); an incorrect formula for Zcc
U3 (O) (T )
was given in [5, §8.2].
g
L1,1
L2,1
L3,1
L3,2
L4,1
L4,2
L4,3
L5,1
L5,2
L5,3
L5,4
L5,5
L5,6
L5,7
L5,8
L5,9
Zcc
G(O)) (T )
1/(1 − qT )
1/(1 − q 2 T )
1/(1 − q 3 T )
(1 − T )/((1 − q 2 T )(1 − qT ))
1/(1 − q 4 T )
(1 − qT )/((1 − q 3 T )(1 − q 2 T ))
(1 − T )/(1 − q 2 T )2
1/(1 − q 5 T )
2
(1 − q T )/((1 − q 4 T )(1 − q 3 T ))
(1 − qT )/(1 − q 3 T )2
(1 − T )/((1 − q 4 T )(1 − qT ))
1−T −qT +q 2 T +q 2 T 2 −q 3 T 2 −q 4 T 2 +q 4 T 3
(1−q 5 T 2 )(1−q 3 T )(1−qT )
1−2T +qT 2 +q 2 T −2q 3 T 2 +q 3 T 3
(1−q 5 T 2 )(1−q 2 T )(1−qT )
(1 − T )/((1 − q 3 T )(1 − q 2 T ))
(1 − qT )/(1 − q 3 T )2
(1 − T )/((1 − q 3 T )(1 − q 2 T ))
Table 1: Complete list of generic conjugacy class zeta functions of unipotent algebraic
groups of dimension at most 5
35
g
L6,10 , L6,25 , L6,26
L6,11 , L6,12 , L6,20
L6,16
L6,17
L6,18
L6,19 (0)
L6,19 (−1), L6,21 (0)
L6,21 (1)
L6,22 (0)
L6,23 , L6,24 (0)
Zcc
G(O) (T )
(1 − qT )/((1 − q 4 T )(1 − q 3 T ))
(1 −
1−2qT +q 2 T +q 4 T 2 −2q 5 T 2 +q 6 T 3
(1−q 6 T 2 )(1−q 3 T )2
qT )(1 − T )/((1 − q 2 T )2 (1 − q 3 T ))
1−T −qT +q 2 T +q 3 T 2 −q 4 T 2 −q 5 T 2 +q 5 T 3
(1−q 6 T 2 )(1−q 3 T )(1−q 2 T )
(1 − T )/((1 − q 2 T )(1 − q 4 T ))
1+T −3qT −q 2 T +q 3 T 2 +3q 4 T 2 −q 5 T 2 −q 5 T 3
(1−q 3 T )3 (1−q 2 T )
(1 − qT )2 /((1 − q 3 T )2 (1 − q 2 T ))
1−T −qT +q 2 T +q 2 T 2 −q 3 T 2 −q 4 T 2 +q 4 T 3
(1−q 5 T 2 )(1−q 3 T )(1−q 2 T )
1−qT −q 2 T +q 3 T +q 4 T 2 −q 5 T 2 −q 6 T 2 +q 7 T 3
(1−q 7 T 2 )(1−q 4 T )(1−q 2 T )
1−2qT +q 3 T +q 3 T 2 −2q 5 T 2 +q 6 T 3
(1−q 7 T 2 )(1−q 3 T )(1−q 2 T )
Table 2: Examples of generic conjugacy class zeta functions in dimension 6
References
[1] N. Avni, B. Klopsch, U. Onn, and C. Voll, Representation zeta functions of compact p-adic analytic
groups and arithmetic groups, Duke Math. J. 162 (2013), no. 1, 111–197.
[2]
, Similarity classes of integral p-adic matrices and representation zeta functions of groups of
type A2 , Proc. Lond. Math. Soc. (3) 112 (2016), no. 2, 267.
[3] G. V. Balakin, The distribution of the rank of random matrices over a finite field, Teor. Verojatnost.
i Primenen. 13 (1968), 631–641.
[4] L. B. Beasley, Spaces of matrices of equal rank, Linear Algebra Appl. 38 (1981), 227–237.
[5] M. N. Berman, J. Derakhshan, U. Onn, and P. Paajanen, Uniform cell decomposition with applications
to Chevalley groups, J. Lond. Math. Soc. (2) 87 (2013), no. 2, 586–606.
[6] J. Blömer, R. Karp, and E. Welzl, The rank of sparse random matrices over finite fields, Random
Structures Algorithms 10 (1997), no. 4, 407–419.
[7] A. Boralevi, D. Faenzi, and E. Mezzetti, Linear spaces of matrices of constant rank and instanton
bundles, Adv. Math. 248 (2013), 895–920.
[8] N. Boston and I. M. Isaacs, Class numbers of p-groups of a given order, J. Algebra 279 (2004), no. 2,
810–819.
[9] F. Brenti, q-Eulerian polynomials arising from Coxeter groups, European J. Combin. 15 (1994),
no. 5, 417–441.
[10] F. Brenti and A. Carnevale, Proof of a conjecture of Klopsch-Voll on Weyl groups of type A (2017).
To appear in Trans. Amer. Math. Soc.
[11] A. Carnevale and C. Voll, Orbit Dirichlet series and multiset permutations (preprint) (2016).
arXiv:1607.01292.
[12] H. Cohen, Number theory. Vol. I. Tools and Diophantine equations, Graduate Texts in Mathematics,
vol. 239, Springer, New York, 2007.
[13] C. Cooper, On the distribution of rank of a random matrix over a finite field, Proceedings of the Ninth
International Conference “Random Structures and Algorithms” (Poznan, 1999), 2000, pp. 197–212.
[14] D. A. Cox, J. Little, and D. O’Shea, Ideals, varieties, and algorithms, Fourth, Undergraduate Texts
in Mathematics, Springer, Cham, 2015.
36
[15] W. A. de Graaf, Classification of 6-dimensional nilpotent Lie algebras over fields of characteristic
not 2, J. Algebra 309 (2007), no. 2, 640–653.
[16] M. Demazure and P. Gabriel, Groupes algébriques. Tome I: Géométrie algébrique, généralités, groupes
commutatifs, Masson & Cie, Éditeur, Paris; North-Holland Publishing Co., Amsterdam, 1970. Avec
un appendice ıt Corps de classes local par Michiel Hazewinkel.
[17] J. Denef, Report on Igusa’s local zeta function, Astérisque 201-203 (1991), Exp. No. 741, 359–386
(1992). Séminaire Bourbaki, Vol. 1990/91.
[18] J. Denef and D. Meuser, A functional equation of Igusa’s local zeta function, Amer. J. Math. 113
(1991), no. 6, 1135–1152.
[19] J. D. Dixon, M. P. F. du Sautoy, A. Mann, and D. Segal, Analytic pro-p groups, Second, Cambridge
Studies in Advanced Mathematics, vol. 61, Cambridge University Press, Cambridge, 1999.
[20] M. P. F. du Sautoy, A nilpotent group and its elliptic curve: non-uniformity of local zeta functions
of groups, Israel J. Math. 126 (2001), 269–288.
[21]
[22]
, Counting subgroups in nilpotent groups and points on elliptic curves, J. Reine Angew. Math.
549 (2002), 1–21.
, Counting conjugacy classes, Bull. London Math. Soc. 37 (2005), no. 1, 37–44.
[23] M. P. F. du Sautoy and F. J. Grunewald, Analytic properties of zeta functions and subgroup growth,
Ann. of Math. (2) 152 (2000), no. 3, 793–833.
[24] M. P. F. du Sautoy and M. Vaughan-Lee, Non-PORC behaviour of a class of descendant p-groups, J.
Algebra 361 (2012), 287 –312.
[25] J.-G. Dumas, R. Gow, G. McGuire, and J. Sheekey, Subspaces of matrices with special rank properties,
Linear Algebra Appl. 433 (2010), no. 1, 191–202.
[26] D. H. Dung and C. Voll, Uniform analytic properties of representation zeta functions of finitely
generated nilpotent groups (2015). To appear in Trans. Amer. Math. Soc. arXiv:1503.06947.
[27] S. N. Evans, Elementary divisors and determinants of random matrices over a local field, Stochastic
Process. Appl. 102 (2002), no. 1, 89–102.
[28] J. Fulman and L. Goldstein, Stein’s method and the rank distribution of random matrices over finite
fields, Ann. Probab. 43 (2015), no. 3, 1274–1314.
[29] M. Giudici, M. W. Liebeck, C. E. Praeger, J. Saxl, and P. H. Tiep, Arithmetic results on orbits of
linear groups, Trans. Amer. Math. Soc. 368 (2016), no. 4, 2415–2467.
[30] J. González-Sánchez, On p-saturable groups, J. Algebra 315 (2007), no. 2, 809–823.
[31] J. González-Sánchez, Kirillov’s orbit method for p-groups and pro-p groups, Comm. Algebra 37
(2009), no. 12, 4476–4488.
[32] S. M. Goodwin and G. Röhrle, Calculating conjugacy classes in Sylow p-subgroups of finite Chevalley
groups, J. Algebra 321 (2009), no. 11, 3321–3334.
[33]
, On conjugacy of unipotent elements in finite groups of Lie type, J. Group Theory 12 (2009),
no. 2, 235–245.
[34]
, Rational points on generalized flag varieties and unipotent conjugacy in finite groups of Lie
type, Trans. Amer. Math. Soc. 361 (2009), no. 1, 177–206.
[35]
, Counting conjugacy classes in the unipotent radical of parabolic subgroups of GLn (q), Pacific
J. Math. 245 (2010), no. 1, 47–56.
[36] G. Higman, Enumerating p-groups. I. Inequalities, Proc. London Math. Soc. (3) 10 (1960), 24–30.
[37] E. Hrushovski, B. Martin, S. Rideau, and R. Cluckers, Definable equivalence relations and zeta
functions of groups (preprint) (2015). arXiv:0701011.
[38] J. ichi Igusa, An introduction to the theory of local zeta functions, AMS/IP Studies in Advanced
Mathematics, vol. 14, Providence, RI: American Mathematical Society, 2000.
37
[39] I. Ilani, Zeta functions related to the group SL2 (Zp ), Israel J. Math. 109 (1999), 157–172.
[40] I. M. Isaacs, Characters of groups associated with finite algebras, J. Algebra 177 (1995), no. 3,
708–730.
[41]
, Counting characters of upper triangular groups, J. Algebra 315 (2007), no. 2, 698–719.
[42] A. Jaikin-Zapirain, Zeta function of representations of compact p-adic analytic groups, J. Amer.
Math. Soc. 19 (2006), no. 1, 91–118 (electronic).
[43] I. Kaplansky, Elementary divisors and modules, Trans. Amer. Math. Soc. 66 (1949), 464–491.
[44] B. Klopsch, On the Lie theory of p-adic analytic groups, Math. Z. 249 (2005), no. 4, 713–730.
[45]
, Representation growth and representation zeta functions of groups, Note Mat. 33 (2013),
no. 1, 107–120.
[46] V. F. Kolchin, Random graphs, Encyclopedia of Mathematics and its Applications, vol. 53, Cambridge
University Press, Cambridge, 1999.
[47] G. Landsberg, Ueber eine Anzahlbestimmung und eine damit zusammenhängende Reihe, J. Reine
Angew. Math. 111 (1893), 87–88.
[48] M. Lazard, Groupes analytiques p-adiques, Inst. Hautes Études Sci. Publ. Math. 26 (1965), 389–603.
[49] X. Li, W. H. Mow, and F.-L. Tsang, Rank distribution analysis for sparse random linear network
coding, International Symposium on Network Coding (NetCod), 2011, pp. 1–6.
[50] N. Linial and D. Weitz, Random vectors of bounded weight and their linear dependencies (unpublished
manuscript) (2010). http://dimacs.rutgers.edu/~dror/pubs/rand_mat.pdf.
[51] S. R. Lockard, Random vectors over finite fields, Ph.D. Thesis, 2007. http://tigerprints.clemson.
edu/all_dissertations/54.
[52] V. V. Morozov, Classification of nilpotent Lie algebras of sixth order, Izv. Vysš. Učebn. Zaved.
Matematika 1958 (1958), no. 4 (5), 161–171.
[53] E. A. O’Brien and C. Voll, Enumerating classes and characters of p-groups, Trans. Amer. Math. Soc.
367 (2015), no. 11, 7775–7796.
[54] I. Pak and A. Soffer, On Higman’s k(Un (Fq )) conjecture (preprint) (2015). arXiv:1507.00411.
[55] N. Rojas, Minimal representations for 6-dimensional nilpotent Lie algebra, Linear Algebra Appl. 15
(2016), no. 10, 19 pages.
[56] T. Rossmann, Stability results for local zeta functions of groups and related structures (preprint)
(2015). arXiv:1504.04164.
[57]
, A framework for computing zeta functions of groups, algebras, and modules (preprint) (2016).
http://www.math.uni-bielefeld.de/~rossmann/files/1489.pdf.
[58]
, Topological representation zeta functions of unipotent groups, J. Algebra (2016).
doi:10.1016/j.jalgebra.2015.09.050.
[59]
, Zeta, version 0.3.2, 2016. See http://www.math.uni-bielefeld.de/~rossmann/Zeta/.
[60] R. P. Stanley, Enumerative combinatorics. Volume 1, Second, Cambridge Studies in Advanced
Mathematics, vol. 49, Cambridge University Press, Cambridge, 2012.
[61] A. Stasinski and C. Voll, Representation zeta functions of nilpotent groups and generating functions
for Weyl groups of type B, Amer. J. Math. 136 (2014), no. 2, 501–550.
[62] W. A. Stein et al., Sage Mathematics Software (Version 7.4), The Sage Development Team, 2016.
Available from http://www.sagemath.org/.
[63] C. Studholme and I. F. Blake, Random matrices and codes for the erasure channel, Algorithmica 56
(2010), no. 4, 605–620.
[64] J. Sylvester, On the dimension of spaces of linear transformations satisfying rank conditions, Linear
Algebra Appl. 78 (1986), 1–10.
38
[65] A. Vera-López and J. M. Arregi, Conjugacy classes in unitriangular matrices, Linear Algebra Appl.
370 (2003), 85–124.
[66] W. Veys and W. A. Zúñiga-Galindo, Zeta functions for analytic mappings, log-principalization of
ideals, and Newton polyhedra, Trans. Amer. Math. Soc. 360 (2008), no. 4, 2205–2227.
[67] C. Voll, Functional equations for zeta functions of groups and rings, Ann. of Math. (2) 172 (2010),
no. 2, 1181–1218.
[68]
, A newcomer’s guide to zeta functions of groups and rings, Lectures on profinite topics in
group theory, 2011, pp. 99–144.
[69]
, Zeta functions of groups and rings—recent developments, Groups St Andrews 2013, 2015,
pp. 469–492.
[70] R. Westwick, Spaces of matrices of fixed rank, Linear and Multilinear Algebra 20 (1987), no. 2,
171–174.
39
| 4math.GR
|
Better Tradeoffs for Exact Distance Oracles in Planar Graphs∗
Paweł Gawrychowski, Shay Mozes, Oren Weimann, and Christian Wulff-Nilsen
arXiv:1708.01386v1 [cs.DS] 4 Aug 2017
Abstract
1.5
We present an O(n )-space distance oracle for directed planar graphs that answers distance
queries in O(log n) time. Our oracle both significantly simplifies and significantly improves the
recent oracle of Cohen-Addad, Dahlgaard and Wulff-Nilsen [FOCS 2017], which uses O(n5/3 )space and answers queries in O(log n) time. We achieve this by designing an elegant and efficient
point location data structure for Voronoi diagrams on planar graphs.
We further show a smooth tradeoff between space and query-time. For any S ∈ [n, n2 ], we
show an oracle of size S that answers queries in Õ(max{1, n1.5 /S}) time. This new tradeoff
is currently the best (up to polylogarithmic factors) for the entire range of S and improves by
polynomial factors over all the previously known tradeoffs for the range S ∈ [n, n5/3 ].
∗
This research was supported in part by Israel Science Foundation grant 794/13.
1
Introduction
Computing shortest paths is a classical and fundamental algorithmic problem that has received
considerable attention from the research community for decades. A natural data structure problem
in this context is to compactly store information about distances in a graph in such a way that
the distance between any pair of query vertices can be computed efficiently. A data structure that
support such queries is called a distance oracle. Naturally, there is a tradeoff between the amount of
space consumed by a distance oracle, and the time required by distance queries. Another quantity
of interest is the preprocessing time required for constructing the oracle.
Distance oracles in planar graphs. It is natural to consider the class of planar graphs in this
setting since planar graphs arise in many important applications involving distances, most notably in
navigation applications on road maps. Moreover, planar graphs exhibit many structural properties
that facilitate the design of very efficient algorithms. Indeed, distance oracles for planar graphs
have been extensively studied. These oracles can be divided into two groups: exact distance oracles
which always output the correct distance, and approximate distance oracles which allow a small
stretch in the distance output. For approximate distance oracles, one can obtain near-linear space
and near-constant query-time at the cost of a (1 + ) stretch (for any fixed ) [14–16, 26, 29]. In this
paper we focus on the tradeoff between space and query-time of exact distance oracles for planar
graphs.
Exact distance oracles. The following results as well as ours all hold for directed planar graphs
with real arc-lengths (but no negative length cycles). Djidjev [8] and Arikati et al. [1] obtained
distance oracles with the following tradeoff between space and query-time. For any S ∈ [n, n2 ],
they show an oracle with space S and√query-time of O(n2 /S 2 ). For S ∈ [n4/3 , n1.5 ], Djidjev’s oracle
achieves an improved bound of O(n/ S). This bound (up to polylogarithmic factors) was extended
to the entire range S ∈ [n, n2 ] in a number of papers [3, 6, 10, 22, 24]. Wulff-Nilsen [28] showed how
to achieve constant query-time with O(n2 (log log n)4 / log n) space, improving the above tradeoff for
close to quadratic space. Very recently, Cohen-Addad, Dahlgaard, and Wulff-Nilsen [7], inspired
by the ideas of Cabello [4] made significant progress by presenting an oracle with O(n5/3 ) space
and O(log n) query-time. This is the first oracle for planar graphs that achieves truly subquadratic
space and subpolynomial query-time. They also showed that with S ≥ n1.5 space, a query-time of
O(n2.5 /S 1.5 log n) is possible. To
√ summarize, prior to the results described in the current paper the
best known tradeoff was Õ(n/ S) query-time for space S ∈ [n, n1.5 ], Õ(n2.5 /S 1.5 ) query-time for
space S ∈ [n1.5 , n5/3 ], and O(log n) query-time for S ∈ [n5/3 , n2 ].
Our results and techniques. In this paper we show a distance oracle with O(n1.5 ) space and
√
O(log n) query-time. More generally, for any r ≤ n we construct a distance oracle with O(n1.5 / r +
√
n log r log(n/r)) space and O( r log n log r) query-time. This improves the currently best known
tradeoffs for essentially the entire range of S: for space S ∈ [n, n1.5 ] we obtain an oracle with
Õ(n1.5 /S) query-time, while for space S ∈ [n1.5 , n2 ], our oracle has query-time of O(log n).
To explain our techniques we need the notion of an additively weighted Voronoi diagram on a
planar graph. Let P = (V, E) be a directed planar graph, and let S ⊆ V be a subset of the vertices,
which are called the sites of the Voronoi diagram. Each site u ∈ S has a weight ω(u) ≥ 0 associated
with it. The distance between a site u ∈ S and a vertex v ∈ V , denoted by d(u, v), is defined as
1
ω(u) plus the length of the u-to-v shortest path in P . The additively weighted Voronoi diagram of
(S, ω) within P , denoted VD(S, ω), is a partition of V into pairwise disjoint sets, one set Vor(u) for
each site u ∈ S. The set Vor(u), which is called the Voronoi cell of u, contains all vertices in V that
are closer (w.r.t. d(·, ·)) to u than to any other site in S (assuming that the distances are unique).
There is a dual representation VD∗ (S, ω) of a Voronoi diagram VD(S, ω) as a planar graph with
O(|S|) vertices and edges. See Section 2.
We obtain our results using point location in additively weighted Voronoi diagrams. This approach is also the one taken in [7]. However, our construction is arguably simpler and more elegant
than that of [7]. Our main technical contribution is a novel point location data structure for Voronoi
diagrams (see below). Given this data structure, the description of the preprocessing and query
algorithms of our O(n1.5 )-space oracle are extremely simple and require a few lines each. In a
nutshell, the construction is recursive, using simple cycle separators. We store a Voronoi diagram
for each node u of the graph. The sites of this diagram are the vertices of the separator and the
weights are the distances from u to each site. To get the distance from u to v it suffices to locate
the node v in the Voronoi diagram stored for u using the point location data structure. Since the
√
cycle separator has O( n) vertices, this yields an oracle requiring O(n1.5 ) space.
The oracles for the tradeoff are built upon this simple oracle by storing Voronoi diagrams for
just a subset of the nodes in a graph (the so called boundary vertices of an r-division). This requires
less space, but the query-time increases. This is because a node u now typically does not have a
dedicated Voronoi diagram. Therefore, to find the distance from u to v, we now we need to locate
v in multiple Voronoi diagrams stored for nodes in the vicinity of u.
As we mentioned above, our main technical tool is a data structure that supports point location
queries in Voronoi diagrams in O(log n) time. This is summarized in the following theorem.
Theorem 1. Let P be a directed planar graph with real arc-lengths, r vertices, and no negative
length cycles. Let S be a set of b sites that lie on a single face (hole) of P . We can preprocess P
in Õ(b · r) time and O(b · r) space so that, given the dual representation of any additively weighted
Voronoi diagram, V D∗ (S, ω), we can extend it in O(b) time and space to support the following
queries. Given a vertex v of P , report in O(log b) time the site u such that v belongs to Vor(u).
A data structure for the same task was described in [7]. Our data structure is both significantly
simpler and more efficient. Roughly speaking, the idea is as follows. We prove that VD∗ (S, ω) is
a ternary tree. This allows us to use a straightforward centroid decomposition of depth O(log b)
for point location. To locate the voronoi cell Vor(u) containing a node v we traverse the centroid
decomposition. At any given level of the decomposition we only need to known which of the three
subtrees in the next level contains Vor(u). To this end we associate with each centroid node three
shortest paths. These paths partition the plane into three parts, each containing exactly one of
the three subtrees. Identifying the desired subtree then boils down to determining the position of
v relative to these three shortest paths. We show that this can be easily done by examining the
preorder number of v in the shortest path trees rooted at three sites.
Roadmap. Theorem 1 is proved in Section 4 under a simplifying assumption that every site
belongs to its Voronoi cell. This suffices to design our distance oracle with space O(n1.5 ) and querytime O(log n), which is done in Section 3, assuming that Theorem 1 holds. Section 5 describes
the improved space to query-time tradeoff. Finally, in Section 6 we describe how to remove the
simplifying assumption. Additional details and some omitted proofs appear in the appendix.
2
2
Preliminaries
We assume that shortest paths are unique. This can be ensured in linear time by a random perturbation of the edge lengths [21, 23] or deterministically in near-linear time using lexicographic
comparisons [5, 13]. It will also be convenient to assume that graphs are strongly connected; if not,
we can always triangulate them with bidirected edges of infinite length.
Separators in planar graphs. Given a planar embedded graph G, a Jordan curve separator is
a simple closed curve in the plane that intersects the embedding of G only at vertices. Miller [20]
√
showed that any n-vertex planar embedded graph has a Jordan curve separator of size O( n) such
that the number of vertices on each side of the curve is at most 2n/3. In fact, the balance of 2/3
can be achieved with respect to any weight function on the vertices, not necessarily the uniform
one. Miller also showed that the vertices of the separator ordered along the curve can be computed
in O(n) time. An r-division [11] of a planar graph G, for some r ∈ (1, n), is a decomposition of G
√
into O(n/r) pieces, where each piece has at most r vertices and O( r) boundary vertices (vertices
shared with other pieces). There is an O(n) time algorithm that computes an r-division of a planar
graph with the additional property that, in every piece, the number of faces of the piece that are
not faces of the original graph G is constant [18, 27] (such faces are called holes).
Voronoi diagrams on planar graphs. Recall the definition of additively weighted Voronoi
diagrams VD(S, ω) from the introduction. We write just VD when the particular S and ω are not
important, or when they are clear from the context.
We restrict our discussion to the case where the sites S lie on a single face, denoted by h. We
work with a dual representation of VD(S, ω), denoted VD∗ (S, ω) or simply VD∗ . Let P ∗ be the
planar dual of P . Let VD∗0 be the subgraph of P ∗ consisting of the duals of edges uv of P such that u
and v are in different Voronoi cells. Let VD∗1 be the graph obtained from VD∗0 by contracting edges
incident to a degree-2 vertex one after the other until no degree 2 vertices remain. The vertices
of VD∗1 are called Voronoi vertices. A Voronoi vertex f ∗ is dual to a face f such that the nodes
incident to f belong to at least three different Voronoi cells. In particular, h∗ (i.e., the dual vertex
corresponding to the face h to which all the sites are incident) is a Voronoi vertex. Each face of
VD∗1 corresponds to a cell Vor(vi ), hence there are at most |S| Voronoi vertices, and, by sparsity of
planar graphs, the complexity (i.e., the number of nodes, edges and faces) of VD∗1 is O(|S|). Finally,
we define VD∗ to be the graph obtained from VD∗1 after replacing the node h∗ by multiple copies,
one for each incident edge. The original Voronoi vertices are called real. See Figure 1.
Given a planar graph P with r nodes and a set S of b sites on a single face h, one can compute
any additively weighted Voronoi diagram VD(S, ω) naively in Õ(r) time by adding an artificial
source node, connecting it to every site s with an edge of length ω(s), and computing the shortest
path tree. The dual representation VD∗ (S, ω) can then be obtained in additional O(r) time by
following the constructive description above. There are more efficient algorithms [4, 12] when one
wants to construct many different additively weighted Voronoi diagrams for the same set of sites S.
The basic approach is to invest superlinear time in preprocessing P , but then construct VD(S, ω)
for multiple choices of ω in Õ(|S|) each instead of Õ(r). Since the focus of this paper is on the
tradeoff between space and query-time, and not on the preprocessing time, the particular algorithm
used for constructing the Voronoi diagrams is less important.
3
Figure 1: A planar graph (black edges) with four sites on the infinite face together with the dual
Voronoi diagram VD∗ (in blue). The sites are shown together with their corresponding shortest
path trees (in turquoise, red, yellow, and green). Two of the Voronoi vertices (in blue) are real.
3
The Oracle
In this section we describe our distance oracle assuming Theorem 1. Let G be a directed planar graph
with non-negative arc-lengths. At a high level, our oracle is based on a recursive decomposition
of G into pieces using Jordan curve separators. Each piece P = (V, E) is a subgraph of G. The
boundary vertices of P are vertices of P that are incident (in G) to edges not in P . The holes of P
are faces of P that are not faces of G. Note that every boundary vertex of P is incident to some
hole of P .
A piece R = (V, E) is decomposed into two smaller pieces on the next level of the decomposition
p
as follows. We choose a Jordan curve separator C = (v1 , v2 , . . . , vk ), where k = O( |V |). This
separates the plane into two parts and defines two smaller pieces P and Q corresponding to, respectively, the subgraphs of R inside and the outside of C. Every edge of R is assigned to either P or
Q. Thus, on every level of the recursive decomposition into pieces, an edge of G appears in exactly
one piece. The separators in levels congruent to 0 modulo 3 are chosen to balance the total number
of nodes. The separators in levels congruent to 1 modulo 3 are chosen to balance the number of
boundary nodes. The separators in levels congruent to 2 modulo 3 are chosen to balance the number
of holes. This guarantees that the number of holes in each piece is constant, and that the number of
vertices and boundary vertices decrease exponentially along the recursion. In particular, the depth
of the decomposition is logarithmic in |V |. These properties are summarized in the following lemma
whose proof is in the appendix.
Lemma 2. Choosing the separators as described above guarantees that (i) each piece has O(1) holes,
`/3
(ii) the number of nodes in a piece on the `-th level in the decomposition is O(n/c1 ), for some
constant c1 > 1, (iii) the number of boundary nodes in a piece on the `-th level in the decomposition
4
√
`/3
is O( n/c2 ), for some constant c2 > 1.
Preprocessing. We compute a recursive decomposition of G using Jordan separators as described
above. For each piece R = (VR , ER ) in the recursive decomposition we perform the following
preprocessing. We compute and store, for each boundary node v of R, the shortest path tree TvR in
R rooted at v. Additionally, we store for every node u of R the distance from v to u and the distance
from u to v in the whole G. For a non-terminal piece R, let P = (VP , EP ) and Q = (VQ , EQ ) be
the two pieces into which R is separated. For every node u ∈ VQ and for every hole h of P we store
an additively weighted Voronoi diagram VD(Sh , ω) for P , where the set of sites Sh is the set of
boundary nodes of P incident to the hole h, and the additive weights ω correspond to the distances
in G from u to each site in Sh . We enhance each Voronoi diagram with the point location data
structure of Theorem 1. We also store the same information with the roles of Q and P exchanged.
Query. To compute the distance from u to v, we traverse the recursive decomposition starting
from the piece that corresponds to the whole initial graph G. Suppose that the current piece is
R = (V, E), which is partitioned into P and Q with a Jordan curve separator C. If v ∈ C then,
because the nodes of C are boundary nodes in both P and Q, we return the additive weight ω(v)
in the Voronoi diagram stored for u, which is equal to the distance from u to v in G. Similarly, if
u ∈ C then we retrieve and return the distance from u to v in the whole G. The remaining case is
that both u and v belong to a unique piece P or Q. If both u and v belong to the same piece on
the lower level of the decomposition, we continue to that piece. Otherwise, assume without loss of
generality that u ∈ Q and v ∈ P . Then, the shortest path from u to v must go through a boundary
node vi of P . We therefore perform a point location query for v in each of the Voronoi diagrams
stored for u and for some hole h of P . Let s1 , . . . , sg be the sites returned by these queries, where
g = O(1) is the number of holes of P . The distance in G from u to si is ω(si ), and the distance in
P from si to v is stored in TsPi . We compute the sum of these two terms for each si , and return the
minimum sum computed.
Analysis. First note that the query-time is O(log n) since, at each step of the traversal, we either
descend to a smaller piece in O(1) time or terminate after having found the desired distance in
O(log n) time by O(1) queries to a point location structure.
Next, we analyze the space. Consider a piece R with O(1) holes. Let n(R) and b(R) denote the
number of nodes and boundary nodes of R, respectively. The trees TuR and the stored distances in
G require a total of O(b(R) · n(R)) space. Let R be further decomposed into pieces P and Q. We
bound the space used by all Voronoi diagrams created for R. Recall that every Voronoi diagram
and point location structure corresponds to a node u of P and a hole of Q, or vice versa. The size
of each additively weighted Voronoi diagram stored for a node of P is O(b(Q)), so O(n(P ) · b(Q))
for all nodes of P . The additional space required by Theorem 1 is also O(n(P ) · b(Q)). Finally, for
every node of R we record if it belongs to the Jordan curve separator used to further divide R, and,
if not, to which of the resulting two pieces it belongs. This takes only O(n(R)) space. The total
space for each piece R is thus O(n(R) · b(R)) plus O(n(P ) · b(Q) + n(Q) · b(P )) if R is decomposed
into P and Q.
We need to bound the sum of O(n(R) · b(R)) over all the pieces R. Consider all pieces
R
,
Because these pieces
P1 R2 , . . . , Rs on the same level ` in the decomposition.
P are edge-disjoint,
√
`
i n(Ri ) = O(n). Additionally, b(Ri ) = O( n/c ) for any i, where c > 1, so
i O(n(Ri ) · b(Ri )) =
5
O(n1.5 /c` ). Summing over all levels `, this is O(n1.5 ). The sum of O(n(P ) · b(Q) + n(Q) · b(P )) over
all pieces R that are decomposed into P and Q can be analysed with the same reasoning to obtain
that the total size of the oracle is O(n1.5 ).
Finally, we analyze the preprocessing time. For each piece R, the preprocessing of Theorem 1
takes Õ(n(R) · b(R)). Then, we compute O(n(R)) different additively weighted Voronoi diagrams
for R. Each diagram is built in Õ(n(R)) time, and its representation is extended in O(b(R)) time
to support point location with Theorem 1. The total preprocessing time for R is hence Õ((n(R))2 ),
which sums up to Õ(n2 ) overall by Lemma 2. We also need to compute the distances between pairs
of vertices in G. This can be also done in Õ(n2 ) total time by computing the shortest path tree
rooted at each vertex in Õ(n) [19].
4
Point Location in Voronoi Diagrams
In this section we prove Theorem 1. Let P be a piece (i.e., a planar graph), and S be a set of sites
that lie on a single face (hole) h of P .
Our goal is to preprocess P once in O(|P ||S|) time and space, and then, given any additively
weighted Voronoi diagram VD∗ (S, ω) (denoted VD∗ for short), preprocess it in O(|S|) time and
space so as to answer point location queries in O(log |S|) time.
We assume that the hole h incident to all nodes in S is the external face. We assume that all
nodes of P ∗ , except for h∗ , have degree 3. This can be achieved by triangulating P with infinite
length edges. We also assume that all nodes incident to the external face belong to S and denote
them s1 , s2 , . . . , sb , according to their clockwise order on h. We assume b ≥ 3 (for any constant b
point location is trivial).
Recall that for a site u and a vertex v we define d(u, v) as ω(u) plus the length of the u-to-v
shortest path in P . We further assume that no Voronoi cell is empty. That is, we assume that, for
every pair of distinct sites u, u0 ∈ S, ω(u) < d(u0 , u). If this assumption does not hold, let S 0 be
the subset of the sites whose Voronoi cells are non empty. We can embed inside the hole h infinite
length edges between every pair of consecutive sites in S 0 , and then again triangulate with infinite
length edges. This results in a new face h0 whose vertices are the sites in S 0 . Replacing S with S 0
and h with h0 enforces the assumption. Note that since this transformation changes P , it is not
suitable when working with Voronoi diagrams constructed by algorithms that preprocess P , such
as the ones in [4, 12] In Section 6 we prove Theorem 1 without this assumption.
4.1
Preprocessing for P
The preprocessing for P consists of computing shortest path trees Tv for every boundary node v ∈ S,
decorated with some additional information which we describe next. We stress that the additional
information does not depend on any weights ω (which are not available at preprocessing time).
Let Ti be the shortest path tree in P rooted at si . For a technical reason that will become clear
soon, we add some artificial vertices to Ti . For each face f of P other than h, we add an artificial
vertex vf whose embedding coincides with the embedding of the dual vertex f ∗ . Let yf be closest
vertex to si in P that is incident to f . We add a zero length arc yf vf to Ti . Note that vf is a leaf
of Ti . Let pi,f be the shortest si -to-vf path in Ti . We say that a vertex v of Ti is to the right (left)
of pi,f if the shortest si -to-v path emanates right (left) of pi,f . Note that, since vf is a leaf of Ti , v
is either right of pi,f , left of pi,f , or a vertex of pi,f ; the goal of adding the artificial vertices vf is to
6
guarantee that these are the only options. The following proposition can be easily obtained using
preorder numbers and a lowest common ancestor (LCA) data structure [2] for Ti .
Proposition 3. There is a data structure with O(|P |) preprocessing time that can decide in O(1)
time if for a given query vertex v and query face f , v is right of pi,f , left of pi,f , or a vertex of pi,f .
We compute and store the shortest path trees Ti rooted at each site si , along with preorder
numbers and LCA data structures required by Proposition 3. This requires preprocessing time
Õ(|P ||S|) by computing each Ti in Õ(|P |) time [19], and can be stored in O(|P ||S|) space.
4.2
Handling a Voronoi diagram VD∗ (S, ω)
We now describe how to handle an additively weighted Voronoi diagram VD∗ = VD∗ (S, ω). This
consists of a preprocessing stage and a query algorithm. Handling VD∗ crucially relies on the fact
that, under assumption that each site is in its own Voronoi cell, VD∗ is a tree.
Lemma 4. VD∗ is a tree.
Proof. Suppose that VD∗ contains a cycle C ∗ . Since the degree of each copy of h∗ is one, the cycle
does not contain h∗ . Therefore, since all the sites are on the boundary of the hole h, the vertices of
P enclosed by C ∗ are in a Voronoi cell that contains no site, a contradiction.
To prove that VD∗ is connected, observe that in VD∗1 , every Voronoi cell is a face (cycle) going
through h∗ . Let C ∗ denote this cycle. If C ∗ is disconnected in VD∗ then, in VD∗1 , C ∗ must visit h∗
at least twice. But this implies that the cell corresponding to C ∗ contains more than a single site,
contradiction our assumption. Thus, the boundary of every Voronoi cell is a connected subgraph of
VD∗ . Since the boundaries of the cell of si and the cell of si+1 both contain the dual of the edge
si si+1 , it follows that the entire modified VD∗ is connected.
We briefly describe the intuition behind the design of the point location data structure. To find
the Voronoi cell Vor(s) to which a query vertex v belongs, it suffices to identify an edge e∗ of VD∗
that is adjacent to Vor(s). Given e∗ we can simply check which of its two adjacent cells contains v
by comparing the distances from the corresponding two sites to v. Our point location structure is
based on a centroid decomposition of VD∗ into connected subtrees, and on the ability to determine,
in constant time, which of the subtrees is the one that contains the desired edge e∗ .
Preprocessing. The preprocessing consists of just computing a centroid decomposition of VD∗ .
A centroid of an n-node tree T is a a node u ∈ T such that removing u and replacing it with copies,
one for each edge incident to u, results in a set of trees, each with at most n+1
2 edges. A centroid
always exists in a tree with more than one edge. In every step of the centroid decomposition of VD∗ ,
we work with a connected subtree T ∗ of VD∗ . Recall that there are no nodes of degree 2 in VD∗ .
If there are no nodes of degree 3, then T ∗ consists of a single edge of VD∗ , and the decomposition
terminates. Otherwise, we choose a centroid f ∗ , and partition T ∗ into the three subtrees T0∗ , T1∗ , T2∗
obtained by splitting f ∗ into three copies, one for each edge incident to f ∗ . Clearly, the depth of
the recursive decomposition is O(log |S|). The decomposition can computed in O(|S|) time and be
represented as a ternary tree, which we call the decomposition tree, in O(|S|) space.
7
Point location query. We first describe the structure that gives rise to the efficient query, and
only then describe the query algorithm. Consider a centroid f ∗ used at some step of the decomposition. Let si0 , si1 , si2 denote the three sites adjacent to f ∗ , listed in clockwise order along h. Let
f be the face of P whose dual is f ∗ . Let y0 , y1 , y2 be the three vertices of f , such that yj is the
vertex of f in Vor(sij ). Let e∗j be the edge of V D∗ incident to f ∗ that is on the boundary of the
Voronoi cells of sij and sij−1 (indices are modulo 3). Let Tj∗ be the subtree of T that contains e∗j .
Let pj denote the shortest sj -to-vf path. Note that the vertex preceding vf in pj is yj . See Figure 2
(right).
si 0
si0
p0
si 2
y2
f
si2
y0
∗
p2
y1
y0
e0
y2
e1
f⇤
e2
y1
p1
si1
si 1
Figure 2: Illustration of the setting and proof of Lemma 5. Left: A decomposition of VD∗ (shown
in blue) by a centroid f ∗ into three subtrees, and a corresponding partition of P into three regions
delimited by the paths pi (shown in red, yellow, and turquoise). Right: a schematic illustration of
the same scenario.
Lemma 5. Let s be the site such that v ∈ Vor(s). If T ∗ contains all the edges of VD∗ incident to
Vor(s), and if v is closer to site sij than to sites sij−1 , sij+1 (indices are modulo 3), then one of the
following is true:
• s = sij ,
• v is to the right of pj and all the boundary edges of Vor(s) are contained in Tj∗ ,
∗ .
• v is to the left of pj and all the boundary edges of Vor(s) are contained in Tj+1
Proof. In the following, let rev(q) denote the reverse of a path q. See Figure 2 for an illustration of
the proof.
Let p be the shortest path from sij to v. If p is a subpath of pj , then s = sij . Assume that p
emanates right of pj (the other case is symmetric). First observe that the path consisting of the
concatenation pj ◦ rev(pj−1 ) intersects VD∗ only at f ∗ . This is because, apart from the artificial
arc yj vf , each shortest path pj is entirely contained in the Voronoi cell of sj . Therefore, none of the
subtrees Tj∗0 contains an edge dual to pj ◦ rev(pj−1 ). Since the path pj ◦ rev(pj−1 ) starts on h, ends
on h and contains no other vertices of h, it partitions the embedding into two subgraphs, one to the
8
right of pj ◦ rev(pj−1 ), and the other to its left. Since e∗j is the only edge of T ∗ that emanates right
of pj ◦ rev(pj−1 ), the only edges of T ∗ in the right subgraph are those of Tj∗ .
Next observe that p does not cross pj (since shortest paths from the same source do not cross),
and does not cross pj−1 (since v is closer to sij than to sij−1 ). Since we assumed p emanates right
of pj , the only edges of T ∗ whose duals belong to p are edges of Tj∗ . Consider the last edge e∗ of
p that is not strictly in Vor(s). If e∗ does not exist then p consists only of edges of Vor(sij ), so
s = sij . If e∗ does exist then it is an incident to Vor(s). By the statement of the lemma all edges of
VD∗ incident to Vor(s) are in T ∗ . Therefore, by the discussion above, e∗ ∈ Tj∗ . We have established
that some edge of VD∗ incident to Vor(s) is in Tj∗ . It remains to show that all such edges are in
Tj∗ . The only two Voronoi cells that are partitioned by the path pj ◦ rev(pj−1 ) are Vor(sij ) and
Vor(sij−1 ). Since v is closer to sij than to sij−1 , s 6= sij−1 . Hence either s = sij , or all the edges of
VD∗ incident to Vor(s) are in Tj∗ .
We can finally state and analyze the query algorithm. We have already argued that, to locate
the Voronoi cell Vor(s) to which v belongs, it suffices to show how to find an edge e∗ incident to
Vor(s). We start with the tree T ∗ = VD∗ which trivially contains all edges of VD∗ incident to
Vor(s). We use the notation from Lemma 5. Note that we can determine in constant time which of
the three sites sij is closest to v by explicitly comparing the distances stored in the shortest path
trees Tj∗ . We use Proposition 3 to determine, in constant time, whether v is right of pj , left of pj ,
or a node on pj . In the latter case, by Lemma 5, we can immediately infer that v is in the Voronoi
cell of sij . In the former two cases we recurse on the appropriate subtree containing all the edges of
VD∗ incident to Vor(s). The total time is dominated by the depth of the centroid decomposition,
which is O(log |S|).
5
The Tradeoff
In this section we generalize the construction presented in Section 3 to yield a smooth tradeoff
between space and query-time. In the following, an MSSP data structure refers to Klein’s multiplesource shortest paths data structure [17]. From now on we assume that all arc-lengths are nonnegative. This can be ensured with a standard transformation that computes shortest paths from a
designated source node in O(n log2 n) time [19] and then appropriately modifies all lengths to make
them non-negative while keeping the same shortest paths.
5.1
Preprocessing
The data structure achieving the tradeoff is recursive using Jordan curve separators as described in
Section 3; at each recursive level we have a piece R = (VR , ER ), which is decomposed by a Jordan
curve separator C into P = (VP , EP ) and Q = (VQ , EQ ), where C is chosen to balance the number of
nodes, the number of boundary nodes, or the number of holes, depending on the remainder modulo
3 of the recursive level. The main difference compared to the oracle of Section 3 is that we do not
store an additively weighted Voronoi diagram of P for each node u in Q (and similarly we do not
store a diagram of Q for each node of P ). Instead, we use an r-division to decrease the number
√
of stored Voronoi diagrams by a factor of r. Additionally, we stop the decomposition when the
number of vertices drops below r. More specifically, for every non-terminal piece R in the recursive
decomposition such that n(R) > r that is decomposed into P and Q with a Jordan curve separator
C, we store the following:
9
1. For each hole h of P , an MSSP data structure capturing the distances in P from all the
boundary nodes of P incident to h to all nodes of P . The MSSP data structure is augmented
with predecessor and preorder information (see below).
2. An r-division for Q, denoted DQ , with O(1) MSSP data structures for each piece of DQ , one
for each hole of the piece. All the boundary nodes of Q (in particular, all nodes of C) are
considered as boundary nodes of DQ (see below).
3. For each boundary node u of DQ , and for each boundary node v of P , the distance dG (u, v)
from u to v in G, and also the distance dG (v, u) from v to u in G.
4. For each boundary node u of DQ , and for each hole h of P , an additively weighted Voronoi
diagram VD(Sh , ω) for P , where the set of sites Sh is the set of boundary nodes of P incident
to the hole h, and the additive weights ω correspond to the distances in G from u to each site
in Sh . We enhance each Voronoi diagram with the point location data structure of Theorem 1.
We also store the same information with the roles of Q and P exchanged. For a terminal piece R,
i.e. when n(R) ≤ r, instead of further subdividing R we revert to the oracle
p of Fakcharoenphol and
Rao [10], which needs O(n(R) log n(R)) space, answers a query in O( n(R) log2 n(R)) time, and
can be constructed in O(n(R) log2 n(R)) time. We also construct an r-division DR for R together
with the MSSP data structures. The boundary nodes of R are considered as boundary nodes of
DR . For each boundary node u ∈ ∂DR , and for each boundary node v ∈ R, we store the distance
dG (u, v) from u to v in G and, for each hole of R, an enhanced additively weighted Voronoi diagram
VD(Sh , ω) for R, where the set of sites Sh is the set of boundary nodes of R incident to the hole h,
and the additive weights ω correspond to the distances in G from u to each site in Sh .
The MSSP data structure in item 1 is a modification of the standard MSSP of Klein [17], where
we change the interface of the persistent dynamic tree representing the shortest path tree rooted
at the boundary nodes of R incident to h, as stated by the following lemma whose proof is in the
appendix.
Lemma 6. Consider a directed planar embedded graph on n nodes with non-negative arc-lengths,
and let v1 , v2 , . . . , vs be the nodes on the boundary of its infinite face, in clockwise order. Then, in
O(n log n) time and space, we can construct a representation of all shortest path trees Ti rooted at
vi , that allow answering the following queries in O(log n) time:
• for a vertex vi and a vertex v ∈ V , return the length of the vi -to-v path in Ti .
• for a vertex vi and vertices u, v ∈ V , return whether u is an ancestor of v in Ti .
• for a vertex vi and vertices u, v ∈ V , return whether u occurs before v in the preorder traversal
of Ti .
In the r-division in item 2 we extend the set of boundary nodes ∂DQ of DQ to also include all the
boundary nodes of Q. In more detail, DQ is obtained from Q by the same recursive decomposition
process as the one used to partition G; on every level of the recursive decomposition we choose a
Jordan curve separator as to balance the total number of nodes, boundary nodes, or holes, depending
on the remainder of the level modulo 3, and terminate the recursion when the number of nodes in
√
a piece is O(r) and the number of boundary nodes is O( r). Every piece of DQ consists of O(r)
√
nodes and, because the boundary nodes of Q are incident to O(1) holes, its O( r) boundary nodes
10
are incident to O(1) holes. Because of the boundary nodes inherited from Q, the number of pieces in
√
DQ is not O(n(Q)/r), and |∂DQ | is not O(n(Q)/ r). We will analyze |∂DQ | later. The MSSP data
structures in item 2 stored for every piece of DQ are the standard structures of Klein. The distances
in item 3 are stored explicitly. The point location mechanism used for the Voronoi diagrams in
item 4 is the one described in Section 4 with the following important modification. Instead of
storing the shortest path trees rooted at every site of the Voronoi diagram explicitly to report
distances, preorder numbers, and ancestry relations in O(1) time, use the MSSP data structure
stored in item 1. Clearly, with such queries one can implement Proposition 3 in O(log(|P |)) time
instead of O(1).
5.2
Query
To compute the distance from u to v, we traverse the recursive decomposition starting from the
piece that corresponds to the whole initial graph G as in Section 3. Eventually, we reach a piece
R = (VR , ER ) such that u, v ∈ VR and either n(R) ≤ r, or n(R) > r and R is decomposed into P
and Q with a Jordan curve separator C such that either u ∈ C, or v ∈ C, or u and v are separated
by C.
We first consider the case when n(R) > r. If u ∈ C or v ∈ C then, because the nodes of C
are boundary nodes the r-divisions DP and DQ , the distance from u to v in G can be extracted
from item 3. Otherwise, u and v are separated by C, and we assume without loss of generality that
u ∈ Q and v ∈ P . Let Q0 be the piece of the r-division DQ that contains u. Any path from u to v
must visit a boundary node of Q0 . Thus, we can iterate over the boundary nodes u0 of Q0 , retrieve
dQ0 (u, u0 ) (from the MSSP data structure in item 2), and then, for each hole h of P , use the Voronoi
diagram VD(Sh , ω) for P (item 4) to find the node v 0 ∈ Sh that minimizes dG (u0 , v 0 ) + dP (v 0 , v)
(computed from item 3 and item 1). The minimum value of dQ0 (u, u0 ) + dG (u0 , v 0 ) + dP (v 0 , v) found
during this computation corresponds to the shortest path from u to v.
The remaining possibility is that n(R) ≤ r. Then the shortest path from u to v either visits
some boundary node of R or not. To check the former case, we proceed similarly as above: we find
the piece R0 of the r-division DR that contains u, iterate over the boundary nodes u0 of Q0 , retrieve
dR0 (u, u0 ), and use the Voronoi diagram VD(Sh , ω) for R to find the node v 0 ∈ Sh that minimizes
dG (u0 , v 0 ) + dP (v 0 , v). To check the latter case, we query the oracle of Fakcharoenphol and Rao [10]
stored for R, and return the minimum of these two distances.
5.3
Analysis
For a piece R, we denote by n(R) and b(R) the number of nodes and boundary nodes of R, respectively. We first analyze the query-time. In O(log n) time we reach the appropriate piece R. Then,
√
we iterate over O( r) boundary nodes. For each of them, we first spend O(log r) time to retrieve the
distance from u to u0 . Then, we need O(log(b(P )) log(n(P ))) time to queryp
the Voronoi diagram.
If n(R) ≤ r, this changes into O(log(b(R)) log(n(R))) and additional O( n(R) log2 (n(R))) =
√
O( r log2 r) time for the oracle of Fakcharoenphol and Rao [10]. Thus, the total query-time is
√
O( r log2 n).
We bound the space required by the data structure for a piece R which is divided into pieces P
and Q. Each MSSP data structure in item 1 requires O(n(P ) log(n(P ))) space, and there are O(1) of
them. Representing the r-division DQ and the MSSP data structures for all the pieces in item 2 can
be done within O(n(Q) log r) space. Then, for every boundary node of DQ the distances in item 3
11
and the O(1) Voronoi diagrams in item 4 can be stored in O(b(P )) space. Thus, we need to analyze
the total number of boundary nodes of D. As we explained above, |∂DQ | would be simply O(n(Q)/r)
√
if not for the additional boundary nodes of Q. We claim that |∂DQ | = O(n(Q)/ r + b(Q)).
To prove the claim we slightly modify the reasoning used by Klein, Mozes, and Sommer [18]
to bound the total number of boundary nodes in an r-division without additional boundary vertices. They analyzed the same recursive decomposition process of a planar graph Q on n nodes by
separating to balance the number of nodes, boundary nodes, or holes, depending on the remainder
modulo 3 of the current level.1 Let T be a tree representing this process, and x̂ be the root of T .
Every node x of T corresponds to a piece. For example, the piece corresponding to the root x̂ is all
of Q. We denote by n(x) and b(x) the number of nodes and boundary nodes, respectively, of the
piece corresponding to x. Define Sr to be the set of rootmost nodes y of T such that n(y) ≤ r.
P
√
Lemma 7.
x∈Sr b(x) = O(n(Q)/ r + b(Q))
Proof. For a node x of T and a set S ofPdescendants of x such that no node of S is an ancestor
of any other, define L(x, S) := −n(x) + y∈S n(y). Essentially, L(x, S) counts the number of new
boundary nodes with multiplicities created when replacing x by all pieces in S. Lemma 8 in [18]
√
states that L(x̂, Sr ) = O(n/ r). I.e., the number of new boundary nodes (with multiplicities)
√
created when replacing the single piece Q by the pieces in Sr is O(n(Q)/ r). We assume that each
node of Q has constant degree (this can be guaranteed with a standard transformation). Thus, each
boundary vertex of Q appears in a constant number of pieces in Sr . Since the number of boundary
vertices in Q is b(Q), the lemma follows.
√
Let Sr0 (x) be the set of rootmost descendants y of x such that b(y) ≤ c0 r, whereSc0 is a fixed
known constant. The r-division found by the recursive decomposition process is Sr0 = x∈Sr Sr0 (x).
√
Indeed, each piece x in Sr0 has n(x) ≤ r, b(x) ≤ c0 r, and O(1) holes. This is true by definition of
Sr0 , even though, instead of starting with a graph with no boundary nodes, we start with a graph
containing b(Q) boundary nodes incident to O(1) holes.
The following claim is proved in Lemma 9 of [18].
√ }
Lemma 8 (Lemma 9 of [18]). |Sr0 (x)| ≤ max{1, 40b(x)
c0 r
√
Corollary 9. |∂DQ | = O(n(Q)/ r + b(Q))
Proof.
|∂DQ | ≤
X
x∈Sr0
!
X
X
√
√
√
40
b(x) ≤ c0 r
|Sr0 (x)| ≤ c0 r |Sr | + 0 √
b(x) = O(n(Q)/ r + b(Q)).
c r
x∈Sr
s∈Sr
Here, the first inequality follows by definition of DQ and Sr0 . The second inequality follows by
√
definition of Sr0 (x) and by the fact that for any x ∈ Sr0 , b(x) ≤ c0 r. The third inequality follows by
Lemma 8. The last inequality follows from the fact that |Sr | = O(n(Q)/r), and from Lemma 7.
We have shown that, starting the recursive decomposition of Q with b(Q) boundary nodes inci√
dent to O(1) holes, we obtain an r-division D consisting of pieces containing O(r) nodes and O( r)
1
In [18] simple cycle separators are used (rather than Jordan curve separators), and thus every piece along the
recursion needs to be re-triangulated. The analysis of the number of boundary nodes, however, is the same.
12
√
boundary nodes incident to O(1) holes, and O(n(Q)/ r + b(Q)) boundary nodes overall. Consequently, the space required by the data structure for a piece R with n(R) > r that is separated into
√
pieces P and Q is O(n(P ) log(n(P )+n(Q) log r+n(Q)/ r+b(Q))b(P )), plus a symmetric term with
√
the roles of P and Q exchanged. If n(R) ≤ r the space is O(n(R) log(n(R))+(n(R)/ r+b(R))b(R)).
Overall, O(n(R) log(n(R)) sums up O(n log n log(n/r)). On the `-th level of the decomposition,
√
√
√
√
O(n(P )/ r · b(Q) + n(Q)/ r · b(P )) sums up to O(n1.5 /( r · c` )), so O(n1.5 / r) over all levels.
√
O(b(P ) · b(Q)) can be bounded by O(b(Q) · n), so we only need to bound the total number of
boundary nodes in all pieces of the recursive decomposition of the whole graph. For the terminal
pieces R, it directly follows from Lemma 7 (with Q being the entire graph G) that the total number
√
of boundary nodes is O(n/ r), but we also need to analyse the non-terminal pieces. Because the
size of a piece decreases by a constant factor after at most three steps of the recursive decomposition
process, it suffices to bound only the total number of boundary nodes for pieces in the sets Sri , for
ri = r · 2i , i = 0, 1, . . . , log(n/r). By applying Lemma 7, with r = ri and Q = G we get that the
√
√
total number of boundary nodes for pieces in Sri is O(n/ ri ), which sums up to O(n/ r) over all i.
√
Thus, the sum of O(b(P ) · b(Q)) over all non-terminal pieces O(n1.5 / r). For all terminal pieces R,
√
√
√
O(n(R)/ r·b(R)) adds up to O(n1.5 / r). O(b(R)·b(R)) can be bounded by O(b(R)· n), which we
√
√
have already shown to be O(n1.5 / r) overall. Thus, the total space is O(n1.5 / r +n log n log(n/r)).
The preprocessing time can be analyzed similarly as in Section 3, except that now we need to
√
compute only O(n(P )/ r + b(P )) Voronoi diagrams for P , each in Õ(n(P )) time. As shown above,
√
√
the overall number of boundary nodes is O(n/ r), so this is Õ(n2 / r) total time. Additionally, we
need to compute the distance between pairs of vertices of P in G (item 3). One of these vertices is
√
always a boundary node of the r-division, so overall we need Õ(n/ r) single-source shortest paths
√
computations in G, which takes Õ(n2 / r) total time. Additionally, we need to construct the oracles
√
when n(R) ≤ r in O(n log2 r) total time. Thus, the total construction time is Õ(n2 / r) overall.
5.4
Improved query-time
The final step in this section is to replace log2 n with log n · log r in the query-time. This is done by
observing that the augmented MSSP data structure takes linear space, but for smaller values of r
we can actually afford to store more data. In the appendix we show the following lemma.
Lemma 10. For any r ∈ [1, n], the representation from Lemma 6 can be modified to allow answer√
√
ing queries in O(log r) time in O(s · n/ r + n log r) space after O(s · n/ r log r + n log n) time
preprocessing.
√
This decreases the query-time to O( r log n log r) at the expense of increasing the space taken
√
by the MSSP data structures in item 1 to O(n(P )/ r · b(P ) + n(P ) log r). Summing over all levels
√
` and including the space used by all other ingredients, this is O(n1.5 / r + n log r log(n/r)).
6
Removing the Assumption on Sites
We now remove the assumption that all the vertices on the hole h are sites of the Voronoi diagram
whose Voronoi cells are non-empty. Recall that VD∗ (S, ω) is obtained from V D1∗ by replacing h∗
with multiple copies, one copy h∗e for each edge e of VD∗0 (S, ω) incident to h∗ . Consider the proof of
Lemma 4. The argument showing that VD∗ (S, ω) contains no cycles still holds. However, because
now vertices incident to the hole are not necessarily sites, the argument showing connectivity fails,
13
d ∗ by identifying
and indeed, VD∗ (S, ω) might be a forest. We turn the forest VD∗ into a tree VD
certain pairs of copies of h∗ as follows. Consider the sequence Eh of edges of VD∗ incident to h∗ ,
ordered according to their clockwise order on the face h. Each pair of consecutive edges e, e0 in Eh
delimits a subpath Q of the boundary walk of h. Note that Q belongs to a single Voronoi cell of
some site s ∈ S . If Q does not contain s we connect the two copies h∗e and h∗e0 with an artificial
d ∗ . See Figure 3.
edge. We denote the resulting graph by VD
d ∗ obtained by
Figure 3: Left: a Voronoi diagram VD∗ (blue) forms a forest. Right: the tree VD
adding three artificial edges (thicker blue lines).
d ∗ is a tree.
Lemma 11. VD
d ∗ is connected and has no cycles. Consider the Voronoi cell of a site s ∈ S.
Proof. We show that VD
In VD∗1 (i.e., before splitting h∗ ) the boundary of this cell is a non-self-crossing cycle C ∗ . Consider
the restriction of C to edges incident to h∗ . Consider two consecutive edges e, e0 in the restriction.
If e and e0 are not consecutive in C ∗ (i.e., if e and e0 do not meet at h∗ ), then they remain connected
in VD∗ (i.e., after splitting h∗ ). If e and e0 are consecutive on C ∗ , then they are also consecutive in
d ∗ . This is true unless the
Eh , so they become disconnected in VD∗ , but get connected again in VD
subpath Q of the face h delimited by e and e0 contains s, but this only happens for one pair of edges
d ∗ . Now, since
in C ∗ . Therefore, since C ∗ was 2-connected (a cycle) in VD∗1 , it is 1-connected in VD
adjacent Voronoi cells share edges, the boundaries of any two adjacent Voronoi cells are connected.
d ∗ is connected that the boundaries of all cells are
It follows from the fact that the dual graph of VD
connected after the identification step.
d ∗ contains a cycle C ∗ . Then C ∗ must also be a cycle in VD∗1 . Since every
Assume that VD
cycle in VD∗1 contains h∗ and encloses at least one site, C ∗ contains a copy of h∗ and encloses
at least one site. Consider a decomposition of C ∗ into maximal segments between copies of h∗ .
d ∗ , whenever two segments of C ∗ are connected with an artificial edge, the
By construction of VD
segment of the boundary of h delimited by these two segments and enclosed by C ∗ does not contain
a site. Since all the sites are on the boundary of h, it follows that C ∗ does not enclose any sites, a
contradiction.
d∗
We next describe how to extend the point location data structure. First observe that since VD
d ∗ each copy
is a tree, it has a centroid decomposition. In VD∗ , the copies of h∗ are all leaves. In VD
d ∗ is still 3. If
of h∗ is incident to at most two artificial edges. Hence the maximum degree in VD
14
d ∗ is not a copy of h∗ then Lemma 5 holds. We need a version of Lemma 5 for
the centroid of VD
the case when the centroid is a copy of h∗ with degree greater than 1 (i.e., incident to one or two
artificial edges). This is in fact a simpler version of Lemma 5. The difference between a copy of h∗
and a Voronoi vertex f ∗ is that f ∗ is a triangular face incident to three specific vertices y0 , y1 , y2 ,
whereas h∗ is incident to all vertices of the hole h. Recall that we connect two copies h∗e and h∗e0 if
the segment Q of the boundary of h delimited by the edges e and e0 belongs to the Voronoi cell of
a site s but does not contain s. When we add this artificial edge, we associate with h∗e and with h∗e0
an arbitrary primal vertex y on Q. Thus, each copy f of h∗ is associated with at most two primal
vertices.
We describe the case where the centroid of T ∗ is a copy ĥ∗ of h∗ with degree 3. The case of
degree 2 and one associated vertex is similar. In the case of degree 3, ĥ∗ is incident to one edge
e1 ∈ Eh , and to two artificial edges which we denote e0 , and e2 , so that the counterclockwise order
of edges around ĥ∗ is e0 , e1 , e2 . Removing ĥ∗ breaks T ∗ into three subtrees. Let Tj∗ be the subtree
of T ∗ rooted at the endpoint of ej that is not ĥ∗ . Recall that, since the degree of ĥ∗ is 3, it has
two associated vertices, y0 , y1 , where yj belongs to the subpath of the boundary of h delimited by
ej and ej+1 . Let sij be the site such that yj belongs to Vor(sij ). Let pj be the shortest path from
sij to yj . See Figure 4.
e0
y0
ĥ⇤
e2
e1
y1
Figure 4: Illustration of the case when the centroid is a copy ĥ∗ of h∗ of degree three. ĥ∗ has two
incident artificial edges (e0 , e2 ), and one Voronoi edge (e1 ). The two vertices y0 , y1 associated with
ĥ∗ are shown, as well as the shortest paths p0 and p1 .
d ∗ incident
Lemma 12. Let s be the site such that v ∈ Vor(s). If T ∗ contains all the edges of VD
to Vor(s), and if v is closer to site sij than to site sij+1 (indices are modulo 2), then one of the
following is true:
• s = sij ,
d ∗ incident to Vor(s) are contained in T ∗ ,
• v is to the left of pj and all the edges of VD
j
15
d ∗ incident to Vor(s) are contained in T ∗ .
• v is to the right of pj and all the edges of VD
j+1
Proof. Observe that all the edges of pj belong to Vor(sij ), while for every i ∈ {0, 1, 2}, the duals of
edges of Ti∗ have endpoints in two different Voronoi cells. Therefore, the paths pj do not cross the
trees Ti∗ . Since p0 and p1 are paths that start and end on the boundary of h and do not cross each
other, they partition G into three subgraphs {Gi }2i=0 . Let G0 be the subgraph to the left of p0 , G1
the subgraph to the right of p0 and to the left pf p1 , and G2 the graph to the right of p1 . It follows
from the above that each subtree Ti∗ belongs to the subgraph Gi .
The remainder of the proof is almost identical to that of Lemma 5. Let p be the shortest path
from sij to v. If p is a subpath of pj then s = sij . Otherwise, assume p emanates left of p0 (the
other cases are similar). Consider the last edge e∗ of p that is not strictly in Vor(s). If e∗ does
not exist then s = si0 . If it does exist, then it must be en edge of T0∗ . Since the only Voronoi cell
d ∗ incident to Vors belong to T ∗ .
partitioned by p0 is that of si0 , either s = si0 , or all edges of VD
0
References
[1] S. Arikati, D. Z. Chen, L. P. Chew, G. Das, M. Smid, and C. D. Zaroliagis. Planar spanners
and approximate shortest path queries among obstacles in the plane. In ESA, pages 514–528,
1996.
[2] M. A. Bender and M. Farach-Colton. The LCA problem revisited. In LATIN, pages 88–94,
2000.
[3] S. Cabello. Many distances in planar graphs. Algorithmica, 62(1-2):361–381, 2012.
[4] S. Cabello. Subquadratic algorithms for the diameter and the sum of pairwise distances in
planar graphs. In SODA, pages 2143–2152, 2017.
[5] S. Cabello, E. Chambers, and J. Erickson. Multiple-source shortest paths in embedded graphs.
SIAM Journal on Computing, 42(4):1542–1571, 2013.
[6] D. Z. Chen and J. Xu. Shortest path queries in planar graphs. In STOC, pages 467–478, 2000.
[7] V. Cohen-Addad, S. Dahlgaard, and C. Wulff-Nilsen. Fast and compact exact distance oracle
for planar graphs. In FOCS, 2017. To appear.
[8] H. Djidjev. On-line algorithms for shortest path problems on planar digraphs. In WG, pages
151–165, 1996.
[9] J. R. Driscoll, N. Sarnak, D. D. Sleator, and R. E. Tarjan. Making data structures persistent.
J. Comput. Syst. Sci., 38(1):86–124, 1989.
[10] J. Fakcharoenphol and S. Rao. Planar graphs, negative weight edges, shortest paths, and near
linear time. J. Comput. Syst. Sci., 72(5):868–889, 2006.
[11] G. N. Frederickson. Fast algorithms for shortest paths in planar graphs, with applications.
SIAM Journal on Computing, 16(6):1004–1022, 1987.
16
[12] P. Gawrychowski, H. Kaplan, S. Mozes, M. Sharir, and O. Weimann. Voronoi diagrams on
planar graphs, and computing the diameter in deterministic Õ(n5/3 ) time. Arxiv 1704.02793,
2017. Submitted to SODA’18.
[13] D. Hartvigsen and R. Mardon. The all-pairs min cut problem and the minimum cycle basis
problem on planar graphs. Journal of Discrete Mathematics, 7(3):403–418, 1994.
[14] K. Kawarabayashi, P. N. Klein, and C. Sommer. Linear-space approximate distance oracles for
planar, bounded-genus and minor-free graphs. In ICALP, pages 135–146, 2011.
[15] K. Kawarabayashi, C. Sommer, and M. Thorup. More compact oracles for approximate distances in undirected planar graphs. In SODA, pages 550–563, 2013.
[16] P. Klein. Preprocessing an undirected planar network to enable fast approximate distance
queries. In SODA, pages 820–827, 2002.
[17] P. N. Klein. Multiple-source shortest paths in planar graphs. In SODA, pages 146–155, 2005.
[18] P. N. Klein, S. Mozes, and C. Sommer. Structured recursive separator decompositions for
planar graphs in linear time. In STOC, pages 505–514, 2013. Full version at https://arxiv.
org/abs/1208.2223.
[19] P. N. Klein, S. Mozes, and O. Weimann. Shortest paths in directed planar graphs with negative
lengths: A linear-space o(n log2 n)-time algorithm. ACM Trans. Algorithms, 6(2):30:1–30:18,
2010.
[20] G. L. Miller. Finding small simple cycle separators for 2-connected planar graphs. J. Comput.
Syst. Sci., 32(3):265–279, 1986. A preliminary version in 16th STOC, 1984.
[21] R. Motwani and P. Raghavan. Randomized algorithms. Press Syndicate of the University of
Cambridge, 1995.
[22] S. Mozes and C. Sommer. Exact distance oracles for planar graphs. In SODA, pages 209–222,
2012.
[23] K. Mulmuley, U. Vazirani, and V. Vazirani. Matching is as easy as matrix inversion. Combinatorica, 7(1):105–113, 1987.
[24] Y. Nussbaum. Improved distance queries in planar graphs. In WADS, pages 642–653, 2011.
[25] D. D. Sleator and R. E. Tarjan. A data structure for dynamic trees. J. Comput. Syst. Sci.,
26(3):362–391, 1983.
[26] M. Thorup. Compact oracles for reachability and approximate distances in planar digraphs.
Journal of the ACM, 51(6):993–1024, 2004.
[27] F. van Walderveen, N. Zeh, and L. Arge. Multiway simple cycle separators and i/o-efficient
algorithms for planar graphs. In SODA, pages 901–918, 2013.
[28] C. Wulff-Nilsen. Algorithms for planar graphs and graphs in metric spaces, 2010. PhD thesis,
University of Copenhagen.
17
[29] C. Wulff-Nilsen. Approximate distance oracles for planar graphs with improved query timespace tradeoff. In SODA, pages 351–362, 2016.
18
A
Missing Proofs
Lemma 2. Choosing the separators as described above guarantees that (i) each piece has O(1) holes,
`/3
(ii) the number of nodes in a piece on the `-th level in the decomposition is O(n/c1 ), for some
constant c1 > 1, (iii) the number of boundary nodes in a piece on the `-th level in the decomposition
√
`/3
is O( n/c2 ), for some constant c2 > 1.
Proof. The number of holes increases by at most one in every recursive call, but decreases by a
constant multiplicative factor every 3 recursive calls, and is initially equal to 0, so part (i) easily
follows. The number of nodes never increases, and decreases by a constant multiplicative factor every
3 recursive calls, and is initially equal to n, so part (ii) follows. The situation with the number of
√
boundary nodes is slightly more complex, because it increases by O( n) in every recursive call, and
decreases by a constant multiplicative factor every 3 recursive calls, where n is the number of nodes
in the current piece. For simplicity, we analyze a different process, in which the number of boundary
√
nodes decreases by a constant multiplicative factor and then increases by O( n) in every recursive
call. The asymptotic behavior of these two processes is identical. Thus, we want to analyze the
following recurrence:
√
n
b(` + 1) = b(`)/c + 0 ` .
(c )
for some constants c, c0 > 1. Then
b(` + 1) =
`
X
i=0
√ X
`
n
n
c0 i
=
).
(
c
ci (c0 )`−i
(c0 )`
√
i=0
We consider two cases:
1. c0 ≤ c, then b(` + 1) ≤
√
n (c`0 )` ≤
2. c0 > c, then b(` + 1) = O(
√
√n
( c0 )`
for ` large enough.
√
n
).
c`
√
In both cases, b(`) = O(
n
)
c`2
for some constant c2 > 1 as claimed.
Lemma 6. Consider a directed planar embedded graph on n nodes with non-negative arc-lengths,
and let v1 , v2 , . . . , vs be the nodes on the boundary of its infinite face, in clockwise order. Then, in
O(n log n) time and space, we can construct a representation of all shortest path trees Ti rooted at
vi , that allow answering the following queries in O(log n) time:
• for a vertex vi and a vertex v ∈ V , return the length of the vi -to-v path in Ti .
• for a vertex vi and vertices u, v ∈ V , return whether u is an ancestor of v in Ti .
• for a vertex vi and vertices u, v ∈ V , return whether u occurs before v in the preorder traversal
of Ti .
Proof. We proceed as in the original implementation of MSSP, that is, we represent every Ti with a
persistent link-cut tree. We present some of the details, understanding them is required to explain
how to implement the queries.
19
In MSSP, we start with constructing T1 with Dijkstra’s algorithm in O(n log n) time. Then, we
iterate over i = 2, 3, . . . , s. The current Ti is maintained with a persistent link-cut tree of Sleator
and Tarjan [25]. The gist of MSSP is that every edge of the graph goes in and out of the shortest
path tree at most once, and that we can efficiently retrieve the edges that should be removed from
and added to Ti−1 to obtain Ti (in O(log n) time per edge). Thus, if we are able to remove or insert
an edge from Ti−1 in O(log n) time, the total update time is O(n log n). With a link-cut tree, we can
indeed remove or insert edge in such time (note that we prefer the worst-case version instead of the
simpler implementation based on splay trees). We make our link-cut tree partially persistent with
a straightforward application of the general technique of Driscoll et al. [9]. This requires that the
in-degree of the underlying structure is O(1), which is indeed the case if the degrees of the nodes in
the graph (and hence in every Ti ) are O(1). This can be guaranteed by replacing a node of degree
d > 4 by a cycle on d nodes, where every node has degree 3. We now verify that the in-degree is
O(1) for such a structure by presenting a high-level overview of link-cut trees.
The edges of a rooted tree are partitioned into solid and dashed. There is at most one solid
edge incoming into any node, so we obtain a partition of the tree into node-disjoint solid paths. For
every solid path, we maintain a balanced search tree on a set of leaves corresponding to the nodes
of the path in the natural top-bottom order when read from left to right. To obtain a worst-case
time bound, Sleator and Tarjan use biased binary trees. Every node stores a pointer to the leaf in
the corresponding biased binary tree, and additionally the topmost node of a heavy path stores a
pointer to its parent in the represented tree (together with the cost of the corresponding edge). The
nodes of every biased binary tree store standard data (a pointer to the left child, the right child,
and the parent) and, additionally, every inner node (that corresponds to a fragment of a solid path)
stores the total cost of the corresponding fragment. An additional component of the link-cut tree is
a complete binary tree on n leaves corresponding to the nodes of the tree (called 1, 2, . . . , n). This
is required, so that we can access a node of the tree on demand in O(log n) time, as random access
is not allowed in this setting. The access pointer points to the root of the complete binary tree.
Now we can indeed verify that the in-degree of the structure is O(1).
Assuming that a representation of every Ti with a partially persistent link-cut tree is available,
we can answer the queries as follows.
First, consider calculating the distance from vi to some v ∈ V . We retrieve the access pointer of
Ti and navigate the complete binary tree to reach the node v. Then, we navigate up in the link-cut
representation of Ti starting from v. In every step, we traverse a biased binary tree starting from
a leaf corresponding to some ancestor u of v. Conceptually, this allows us to jump to the topmost
node of the solid path containing u. While doing so, we accumulate the total cost of the prefix of
that solid path ending at u. Then, we follow the pointer from the topmost node of the current solid
path to reach its parent in Ti , add its cost to the answer, and continue in the next biased binary
tree. It is easy to see that in the end we obtain the total cost of the path from v to the root of Ti ,
and by the properties of biased binary trees the total number of steps is O(log n).
Second, consider checking if u is an ancestor of v in Ti . We navigate in the link-cut representation
of Ti starting from u and marking the visited solid paths (in more detail, every solid path stores a
timestamp of the most recent visit; the timestamps are not considered a part of the original partially
persistent structure and the current time is increased after each query). Then, we navigate starting
from v, but stop as soon as we reach a solid path already visited in the previous step. For u to be
an ancestor of v, this must be the path containing u, and furthermore u must be on the left of v in
the corresponding biased binary tree. This can be all checked in O(log n) time.
20
Third, consider checking if u occurs before v in the preorder traversal of Ti . By proceeding as
in the previous paragraph we can identify the LCA of u and v, denoted w. Assuming that w 6= u
and w 6= v, we can also retrieve the edge outgoing from w leading to the subtree containing u, and
similarly for v, in O(log n) total time. We can also retrieve the edge incoming to w from its parent
in O(1) additional time. Then, we check the cyclic order on the edges incident to w in the graph
to determine if u comes before v in the preorder traversal of Ti (this is so that we do not need to
think about an embedding of Ti while maintaining the link-cut representation).
Lemma 6. Consider a directed planar embedded graph on n nodes with non-negative arc-lengths,
and let v1 , v2 , . . . , vs be the nodes on the boundary of its infinite face, in clockwise order. Then, in
O(n log n) time and space, we can construct a representation of all shortest path trees Ti rooted at
vi , that allow answering the following queries in O(log n) time:
• for a vertex vi and a vertex v ∈ V , return the length of the vi -to-v path in Ti .
• for a vertex vi and vertices u, v ∈ V , return whether u is an ancestor of v in Ti .
• for a vertex vi and vertices u, v ∈ V , return whether u occurs before v in the preorder traversal
of Ti .
Proof. We construct an r-division R of the graph. The structure consists of parts: micro components
and macro component.
Consider a shortest path tree Ti . We construct a new (smaller) tree Ti0 as follows. First, mark
in Ti vi and all boundary nodes of R. Then, Ti0 is the tree induced by the marked nodes in Ti
(in other words: for any two marked nodes we also mark their LCA in Ti , and then construct Ti0
by connecting every marked node to its first marked ancestor with an edge of length equal to the
√
total length of the path connecting them in Ti . Then, |Ti0 | = O(n/ r). Intuitively, Ti0 gives us a
high-level overview of the whole Ti . We augment it with the usual preorder numbers and LCA data
structure. We call this the macro component.
For every piece Rj of R, consider the subgraph of Ti consisting of all edges belonging to Rj .
This subgraph is a collection of trees rooted at some of the boundary nodes of Rj . We represent
this forest Ti,j with a persistent link-cut forest. While sweeping through the nodes v1 , v2 , . . . , vs ,
every edge of the graph goes in and out of the shortest path tree at most once. Hence, every edge
of Rj goes in and out of Ti,j at most once. Thus, the persistent link-cut representation of Ti,j
takes O(|R| log |R|) time and space. To answer a query concerning Ti,j , we first need to retrieve the
corresponding version of the link-cut forest. This can be done with a predecessor search, if we store
a sorted list of the values of i together with a pointer to the corresponding version, in O(log r) time,
as there are at most O(|R|) versions. Then, a query concerning Ti,j can be answered in O(log r)
time.
We claim that combining the micro and macro components allows us to answer any query in
O(log r) time. Consider calculating the distance from vi to some v ∈ V . We retrieve the piece Rj
containing v and, by using the micro component, find the root r of the tree containing v in the
forest Ti,j together with the distance from r to v in O(log r) time. Then, r is a boundary node,
so the macro component allows us to find the distance from vi to r in O(1) time. Other queries
can be processed similarly by first look at the pieces containing u and v, then replacing them by
appropriate boundary nodes, and finally looking at the macro component.
√
The total space is clearly O(n log r) to represent all the micro components, and O(s · n/ r)
for the macro component. To bound the preprocessing time, observe that constructing the macro
21
√
component requires extracting O(n/ r) nodes from the persistent link-cut representation of Ti . If,
instead of accessing them one by one, we work with all of them at the same time, can be seen to
√
take O(n/ r log r) time by the convexity of log.
22
| 8cs.DS
|
A
Liquid Cloud Storage
arXiv:1705.07983v1 [cs.DC] 22 May 2017
Michael G. Luby, Qualcomm Technologies, Inc. ([email protected])
Roberto Padovani, Qualcomm Technologies, Inc. ([email protected])
Thomas J. Richardson, Qualcomm Technologies, Inc. ([email protected])
Lorenz Minder, Qualcomm Technologies, Inc. ([email protected])
Pooja Aggarwal, Qualcomm Technologies, Inc. ([email protected])
A liquid system provides durable object storage based on spreading redundantly generated data across a
network of hundreds to thousands of potentially unreliable storage nodes. A liquid system uses a combination
of a large code, lazy repair, and a flow storage organization. We show that a liquid system can be operated to
enable flexible and essentially optimal combinations of storage durability, storage overhead, repair bandwidth
usage, and access performance.
CCS Concepts: •Information systems → Distributed storage; Information storage technologies; Storage recovery strategies; •Computer systems organization → Maintainability and maintenance; Reliability; Availability; Redundancy;
Additional Key Words and Phrases: distributed information systems, data storage systems, data warehouses, information
science, information theory, information entropy, error compensation, time-varying channels, error correction codes, ReedSolomon codes, network coding, signal to noise ratio, throughput, distributed algorithms, algorithm design and analysis,
reliability, reliability engineering, reliability theory, fault tolerance, redundancy, robustness, failure analysis, equipment failure.
ACM Reference Format:
ACM V, N, Article A (January YYYY), 44 pages.
DOI: http://dx.doi.org/10.1145/0000000.0000000
1. INTRODUCTION
Distributed storage systems generically consists of a large number of interconnected storage nodes,
with each node capable of storing a large quantity of data. The key goals of a distributed storage
systems are to store as much source data as possible, to assure a high level of durability of the source
data, and to minimize the access time to source data by users or applications.
Many distributed storage systems are built using commodity hardware and managed by complex
software, both of which are subject to failure. For example, storage nodes can become unresponsive, sometimes due to software issues where the node can recover after a period of time (transient
node failures), and sometimes due to hardware failure in which case the node never recovers (node
failures). The paper [8] provides a more complete description of different types of failures.
A sine qua non of a distributed storage system is to ensure that source data is durably stored,
where durability is often characterized in terms of the mean time to the loss of any source data
(MTTDL) often quoted in millions of years. To achieve this in the face of component failures a
fraction of the raw storage capacity is used to store redundant data. Source data is partitioned into
objects, and, for each object, redundant data is generated and stored on the nodes. An often-used
trivial redundancy is triplication in which three copies of each object are stored on three different
nodes. Triplication has several advantages for access and ease of repair but its overhead is quite high:
two-thirds of the raw storage capacity is used to store redundant data. Recently, Reed-Solomon (RS)
codes have been introduced in production systems to both reduce storage overhead and improve
MTTDL compared to triplication, see e.g., [10], [8], and [19]. As explained in [19] the reduction
in storage overhead and improved MTTDL achieved by RS codes comes at the cost of a repair
bottleneck [19] which arises from the significantly increased network traffic needed to repair data
lost due to failed nodes.
The repair bottleneck has inspired the design of new types of erasure codes, e.g., local reconstruction codes (LR codes) [9], [19], and regenerating codes (RG codes) [6], [7], that can reduce
repair traffic compared to RS codes. The RS codes, LR codes and RG codes used in practice typically use small values of (n, k, r), which we call small codes. Systems based on small codes, which
ACM Journal Name, Vol. V, No. N, Article A, Publication date: January YYYY.
2
OVERVIEW
we call small code systems, therefore spread data for each object across a relatively small number of
nodes. This situation necessitates using a reactive repair strategy where, in order to achieve a large
MTTDL, data lost due to node failures must be quickly recovered. The quick recovery can demand
a large amount network bandwidth and the repair bottleneck therefore remains.
These issues and proposed solutions have motivated the search for an understanding of tradeoffs.
Tradeoffs between storage overhead and repair traffic to protect individual objects of small code
systems are provided in [6], [7]. Tradeoffs between storage overhead and repair traffic that apply to
all systems are provided in [14].
Initially motivated to solve the repair bottleneck, we have designed a new class of distributed
storage systems that we call liquid systems: Liquid systems use large codes to spread the data stored
for each object across a large number of nodes, and use a lazy repair strategy to slowly repair data
lost from failed nodes.1 In this paper we present the design and its most important properties. Some
highlights of the results and content of paper include:
— Liquid systems solve the repair bottleneck, with less storage overhead and larger MTTDL than
small code systems.
— Liquid systems can be flexibly deployed at different storage overhead and repair traffic operating
points.
— Liquid systems avoid transient node failure repair, which is unavoidable for small code systems.
— Liquid systems obviate the need for sector failure data scrubbing, which is necessary for small
code systems.
— Liquid system regulated repair provides an extremely large MTTDL even when node failures are
essentially adversarial.
— Simulation results demonstrate the benefits of liquid systems compared to small code systems.
— Liquid system prototypes demonstrate improved access speeds compared to small code systems.
— Detailed analyses of practical differences between liquid systems and small code systems are
provided.
— Liquid systems performance approach the theoretical limits proved in [14].
2. OVERVIEW
2.1. Objects
We distinguish two quite different types of data objects. User objects are organized and managed
according to semantics that are meaningful to applications that access and store data in the storage
system. Storage objects refers to the organization of the source data that is stored, managed and
accessed by the storage system. This separation into two types of objects is common in the industry,
e.g., see [3]. When user objects (source data) are to be added to the storage system, the user objects
are mapped to (often embedded in) storage objects and a record of the mapping between user objects
and storage objects is maintained. For example, multiple user objects may be mapped to consecutive
portions of a single storage object, in which case each user object constitutes a byte range of the
storage object. The management and maintenance of the mapping between user objects and storage
objects is outside the scope of this paper.
The primary focus of this work is on storing, managing and accessing storage objects, which
hereafter we refer to simply as objects. Objects are immutable, i.e., once the data in an object is
determined then the data within the object does not change. Appends to objects and deletions of
entire objects can be supported. The preferred size of objects can be determined based on system
tradeoffs between the number of objects under management and the granularity at which object
repair and other system operations are performed. Generally, objects can have about the same size,
1 The
term liquid system is inspired by how the repair policy operates: instead of reacting and separately repairing chunks of
data lost from each failed node, the lost chunks of data from many failed nodes are lumped together to form a “liquid” and
repaired as a flow.
2
2
OVERVIEW
2.2
Storage and repair
or at least a minimum size, e.g., 1 GB. The data size units we use are KB = 1024 bytes, MB = 1024
KB, GB = 1024 MB, TB = 1024 GB, and PB = 1024 TB.
2.2. Storage and repair
The number of nodes in the system at any point in time is denoted by M , and the storage capacity of
each node is denoted by S, and thus the overall raw capacity is Draw = M · S. The storage overhead
Dsrc
expressed as a fraction is β = 1 − D
, and expressed as a per cent is β · 100%, where Dsrc is the
raw
aggregate size of objects (source data) stored in the system.
When using an (n, k, r) erasure code, each object is segmented into k source fragments and an
encoder generates r = n − k repair fragments from the k source fragments. An Encoding Fragment
ID, or EFI, is used to uniquely identify each fragment of an object, where EFIs 0, 1, . . . , k − 1
identify the source fragments of an object, and EFIs k, . . . , n − 1 identify the repair fragments
generated from source fragments. Each of these n = k + r fragments is stored at a different node.
An erasure code is MDS (maximum distance separable) if the object can be recovered from any k
of the n fragments.
A placement group maps a set of n fragments with EFIs {0, 1, . . . , n − 1} to a set of n of the M
nodes. Each object is assigned to one of the placement groups, and the selected placement group
determines where the fragments for that object are to be stored. To ensure that an equal amount of
data is stored on each node, each placement group should be assigned an equal amount of object
data and each node should accommodate the same number of placement groups.
Dsrc
For the systems we describe, D
= nk and thus β = nr , i.e., the amount of source data stored
raw
in the system is maximized subject to allotting enough space to store n fragments for each object.
Object loss occurs when the number of fragments available for an object is less than k, and thus
the object is no longer recoverable. A repairer is an agent that maintains recoverability of objects
by reading, regenerating and writing fragments lost due to node failures, to ensure that each object
has at least k available fragments. In principle the repairer could regenerate r missing fragments
for an object when only k fragments remain available. However, this repair policy results in a small
MTTDL, since one additional node failure before regenerating the r missing fragments for the
object causes object loss.
A repairer determines the rate at which repair occurs, measured in terms of a read repair rate,
since data read by a repairer is a fair measure of repair traffic that moves through the network. (The
amount of data written by a repairer is essentially the same for all systems, and is generally a small
fraction of data read by a repairer.)
Of primary interest is the amount of network bandwidth Rpeak that needs to be dedicated to repair
to achieve a given MTTDL for a particular system, as this value of Rpeak determines the network
infrastructure needed to support repair for the system. Generally, we set a global upper bound Rpeak
on the allowable read repair rate used by a repairer, and determine the MTTDL achieved for this
setting of Rpeak by the system. The average read repair rate Ravg is also of some interest, since this
determines the average amount of repair traffic moved across the network over time.
The fixed-rate repair policy works in similar ways for small code systems and liquid systems in
our simulations: The repairer tries to use a read repair rate up to Rpeak whenever there are any objects
to be repaired.
2.3. Small code systems
Replication, where each fragment is a copy of the original object is an example of a trivial MDS
erasure code. For example, triplication is a simple (3, 1, 2) MDS erasure code, wherein the object
can be recovered from any one of the three copies. Many distributed storage systems use replication.
A Reed-Solomon code (RS code) [2], [17], [12] is an MDS code that is used in a variety of applications and is a popular choice for storage systems. For example, the small code systems described
in [8], [10] use a (9, 6, 3) RS code, and those in [19] use a (14, 10, 4) RS code.
3
2.4
Liquid systems
2
OVERVIEW
When using a (n, k, r) small code for a small code system with M nodes, an important consideration is the number of placement groups P to use. Since n M, a large number P of placement
groups are typically used to smoothly spread the fragments for the objects across the nodes, e.g.,
Ceph [22] recommends P = 100·M
n . A placement group should avoid mapping fragments to nodes
with correlated failures, e.g., to the same rack and more generally to the same failure domain. Pairs
of placement groups should avoid mapping fragments to the same pair of nodes. Placement groups
are updated as nodes fail and are added. These and other issues make it challenging to design placement groups management for small code systems.
Reactive repair is used for small code systems, i.e., the repairer quickly regenerates fragments
as soon as they are lost from a failed node (reading k fragments for each object to regenerate the
one missing fragment for that object). This is because, once a fragment is lost for an object due to
a node failure, the probability of r additional node failures over a short interval of time when r is
small is significant enough that repair needs to be completed as quickly as practical. Thus, the peak
read repair rate Rpeak is higher than the average read repair rate Ravg , and Ravg is k times the node
failure erasure rate. In general, reactive repair uses large bursts of repair traffic for short periods of
time to repair objects as soon as practical after a node storing data for the objects is declared to have
permanently failed.
The peak read repair rate Rpeak and average read repair rate Ravg needed to achieve a large MTTDL
for small code systems can be substantial. Modifications of standard erasure codes have been designed for storage systems to reduce these rates, e.g., local reconstruction codes (LR codes) [9],
[19], and regenerating codes (RG codes) [6], [7]. Some versions of LR codes have been used in
deployments, e.g., by Microsoft Azure [10]. The encoding and decoding complexity of RS, LR, and
RG codes grows non-linearly (typically quadratric or worse) as the values of (n, k, r) grow, which
makes the use of such codes with large values of (n, k, r) less practical.
2.4. Liquid systems
We introduce liquid systems for reliably storing objects in a distributed set of storage nodes. Liquid
systems use a combination of a large code with large values of (n, k, r), lazy repair, and a flow
storage organization. Because of the large code size the placement group concept is not of much
importance: For the purposes of this paper we assume that one placement group is used for all
objects, i.e., n = M and a fragment is assigned to each node for each object.
The RaptorQ code [20], [13] is an example of an erasure code that is suitable for a liquid system.
RaptorQ codes are fundamentally designed to support large values of (n, k, r) with very low encoding and decoding complexity and to have exceptional recovery properties. Furthermore, RaptorQ
codes are fountain codes, which means that as many encoded symbols n as desired can be generated on the fly for a given value of k. The fountain code property provides flexibility when RaptorQ
codes are used in applications, including liquid systems. The monograph [20] provides a detailed
exposition of the design and analysis of RaptorQ codes.
The value of r for a liquid system is large, and thus lazy repair can be used, where lost fragments
are repaired at a steady background rate using a reduced amount of bandwidth. The basic idea is
that the read repair rate is controlled so that objects are typically repaired when a large fraction of
r fragments are missing (ensuring high repair efficiency), but long before r fragments are missing
(ensuring a large MTTDL), and thus the read repair rate is not immediately reactive to individual
node failures. For a fixed node failure rate the read repair rate can in principle be fixed as described
in Appendix A. When the node failure rate is unknown a priori, the algorithms described in Section 9
and Appendix A should be used to adapt the read repair rate to changes in conditions (such as node
failure rate changes) to ensure high repair efficiency and a large MTTDL.
We show that a liquid system can be operated to enable flexible and essentially optimal combinations of storage durability, storage overhead, repair bandwidth usage, and access performance,
exceeding the performance of small code systems.
Information theoretic limits on tradeoffs between storage overhead and the read repair rate are introduced and proved in [14]. The liquid systems for which prototypes and simulations are described
4
3
SYSTEM FAILURES
Clients
Object
Object
Object
HTTP REST based
Object interface
(S3 or SWIFT API)
Liquid Access Tier
Access Proxy
Access Proxy
RaptorQ
Codec
RaptorQ
Codec
Flow Data Organization
Fragment
Store
Fragment
Store
Disk
Disk
Storage Server
Background Lazy Repair Process
Liquid Storage Tier
Fragment
Store
Fragment
Store
Disk
Disk
Storage Server
Fig. 1: A liquid system architecture.
in this paper are similar to the liquid systems described in Section 9.A of [14] that are within a factor
of two of optimal. More advanced liquid systems similar to those described in Section 9.B of [14]
that are asymptotically optimal would perform even better.
Fig. 1 shows a possible liquid system architecture. In this architecture, a standard interface, such
as the S3 or SWIFT API, is used by clients to make requests to store and access objects. Object
storage and access requests are distributed across access proxy servers within the liquid access tier.
Access proxy servers use a RaptorQ codec to encode objects into fragments and to decode fragments
into objects, based on the flow storage organization. Access proxy servers read and write fragments
from and to the storage servers that store the fragments on disks within the liquid storage tier. Lazy
repair operates as a background process within the liquid storage tier. Further details of this example
liquid system architecture are discussed in Section 8.
3. SYSTEM FAILURES
Since distributed storage systems consist of a large number of components, failures are common and
unavoidable. The use of commodity hardware can additionally increase the frequency of failures and
outages, as can necessary maintenance and service operations such as software upgrades or system
reconfiguration on the hardware or software level.
Different failure types can affect the system in various ways. For example, a hard drive can fail in
its entirety, or individual sectors of it can become unreadable. Node failures can be permanent, e.g.
if the failure results in the entire node needing to be replaced, or it can be transient, if it is caused
issues such as network outages, reboots, software changes. Network failures can cause subsets of
many nodes to be unreachable, e.g. a full rack or even a complete site.
5
3.1
Node failures
3
SYSTEM FAILURES
Moreover, distributed storage systems are heterogeneous in nature, with components that are
commonly being upgraded and replaced. Device models do not remain the same; for example,
a replacement hard drive will often be of a different kind, with different and possibly unknown
expected lifetime. Even within a device model, different manufacturing runs can result in variation
in expected lifetimes. Further, upgrades may require a large set of machines to be replaced in a short
time frame, leading to bursts of increased churn. Changes to the underlying technology, such as a
transition from hard drives to SSDs, are also likely to profoundly impact the statistics and nature of
storage failures.
For these reasons, the failure behavior of a distributed storage system is difficult to model. Results
obtained under complicated realistic error models are difficult to interpret, and hard to generalize.
The results in this paper are obtained under a fairly straightforward failure model which focuses on
two types of failures: Node failures and sector failures.
3.1. Node failures
Although node failure processes can be difficult to accurately model, analysis based on a random
node failure models can provide insight into the strengths and weaknesses of a practical system, and
can provide a first order approximation to how a practical system will operate. The time till failure
of an individual node is often modeled by a Poisson random variable with rate λ, in which case
the average time till an individual node fails is 1/λ. Nodes are assumed to fail independently. Since
failed nodes are immediately replaced, the aggregate node failure process is Poisson with rate λ · M .
Our simulations use this model with 1/λ = 3 years, and also use variants of this model where the
node failure rate bursts to 3 times this rate (1/λ = 1 year) or 10 times this rate (1/λ = 0.3 years)
for periods of time.
Nodes can become unresponsive for a period of time and then become responsive again (transient
node failure), in which case the data they store is temporarily unavailable, or nodes can become and
remain unresponsive (node failure), in which case the data they store is permanently unavailable.
Transient node failures are an order of magnitude more common than node failures in practice, e.g.,
see [8].
Unresponsive node are generally detected by the system within a few seconds. However, it is
usually unknown whether the unresponsiveness is due to a transient node failure or a node failure.
Furthermore, for transient node failures, their duration is generally unknown, and can vary in an
extremely large range: Some transient node failures have durations of just seconds when others can
have durations of months, e.g., see [8].
Section 9 discusses more general and realistic node failure models and introduces repair strategies
that automatically adjust to such models.
3.2. Sector failures
Sector failures are another type of data loss that occurs in practice and can cause operational issues.
A sector failure occurs when a sector of data stored at a node degrades and is unreadable. Such data
loss can only be detected when an attempt is made to read the sector from the node. (The data is
typically stored with strong checksums, so that the corruption or data loss becomes evident when
an attempt to read the data is made.) Although the sector failure data loss rate is fairly small, the
overhead of detecting these types of failures can have a negative impact on the read repair rate and
on the MTTDL.
Data scrubbing is often employed in practice to detect sector failures: An attempt is made to read
through all data stored in the system at a regular frequency. As an example, [4] reports that Dropbox
scrubs data at the frequency of once each two weeks, and reports that read traffic due to scrubbing
can be greater than all other read data traffic combined. Most (if not all) providers of data storage
systems scrub data, but do not report the frequency of scrubs. Obviously, scrubbing can have an
impact on system performance.
The motivation for data scrubbing is that otherwise sector failures could remain undetected for
long periods of time, and thus have a large negative impact on the MTTDL. This impact can be
6
4
REPAIR
substantial even if the data loss rate due to sector failures is relatively much smaller than the data
loss rate due to node failures. The reason for this is that time scale for detecting and repairing sector
failures is so much higher than for detecting and repairing node failures.
We do not employ a separate scrubbing process in our simulations, but instead rely on the repair
process to scrub data: Since each node fails on average in 1/λ years, data on nodes is effectively
scrubbed on average each 1/λ years. Sector failures do not significantly affect the MTTDL of liquid
systems, but can have a significant affect on the MTTDL for small code systems.
We model the time to sector failure of an individual 4 KB sector of data on a node as a Poisson
random variable with rate λSF , and thus the average time till an individual 4 KB sector of data on a
node experiences a sector failure is 1/λSF , where sector failures of different sectors are independent.
Based on sector failure statistics derived from the data of [25], our simulations use this model with
1/λSF = 5 · 108 years. Thus, the data loss rate due to sector failures is more than 108 times smaller
than the data loss rate due to node failures in our simulations.
4. REPAIR
Although using erasure coding is attractive in terms of reducing storage overheads and improving
durability compared to replication, it can potentially require large amounts of bandwidth to repair
for fragment loss due to node failures. Since the amount of required repair bandwidth can be high
and can cause system-wide bottlenecks, this is sometimes referred to as a repair bottleneck. For
example, [19] states the goal of their work as follows:
Our goal is to design more efficient coding schemes that would allow a large fraction of
the data to be coded without facing this repair bottleneck. This would save petabytes of
storage overheads and significantly reduce cluster costs.
The repair bottleneck sometimes has proven to be an obstacle in the transition from simple replication to more storage efficient and durable erasure coding in storage clusters. It simply refers to
the fact that, when using a standard erasure code such as a RS code, the loss of a fragment due
to a node failure requires the transfer of k fragments for repair, thus increasing k-fold the required
bandwidth and I/O operations in comparison to replication. For example, [19] states the following
when describing a deployed system with 3000 nodes:
. . . the repair traffic with the current configuration is estimated around 10-20% of the
total average of 2 PB/day cluster network traffic. As discussed, (14,10,4) RS encoded
blocks require approximately 10x more network repair overhead per bit compared to
replicated blocks. We estimate that if 50% of the cluster was RS encoded, the repair
network traffic would completely saturate the cluster network links.
The recognition of a repair bottleneck presented by the use of standard MDS codes, such as RS
codes, has led to the search for codes that provide locality, namely where the repair requires the
transfer of ` < k fragments for repair. The papers [9], [10], [19] introduce LR codes and provide
tradeoffs between locality and other code parameters. We discuss some properties of small code
systems based on LR codes in Section 5.2.
The following sub-sections provide the motivation and details for the repair strategies employed
in our simulations.
4.1. Repair initiation timer
In practice, since the vast majority of unresponsive nodes are transient node failures, it is typical to
wait for a repair initiation time TRIT after a node becomes unresponsive before initiating repair. For
example, TRIT = 30 minutes for the small code system described in [10], and thus the policy reaction
is not immediate but nearly immediate.
In our simulations repairers operate as follows. If a node is unresponsive for a period at most
TRIT then the event is classified as a transient node failure and no repair is triggered. After being
unresponsive for a period of TRIT time, a node failure is declared, the node is decommissioned, all
7
4.2
Repair strategies
4
REPAIR
fragments stored on the node are declared to be missing (and will be eventually regenerated by the
repairer). The decommissioned node is immediately replaced by a new node that initially stores no
data (to recover the lost raw capacity).
Transient node failures may sometimes take longer than 30 minutes to resolve. Hence, to avoid
unnecessary repair of transient node failures, it is desirable to set TRIT longer, such as 24 hours.
A concern with setting TRIT large is that this can increase the risk of source data loss, i.e., it can
significantly decrease the MTTDL.
Since a large value for TRIT has virtually no impact on the MTTDL for liquid systems, we can
set TRIT to a large value and essentially eliminate the impact that transient node failures have on the
read repair rate. Since a large value for TRIT has a large negative impact on the MTTDL for small
code systems, TRIT must be set to a small value to have a reasonable MTTDL, in which case the
read repair rate can be adversely affected if the unresponsive time period for transient node failure
extends beyond TRIT and triggers unnecessary repair. We provide some simulation results including
transient node failures in Section 5.4.
4.2. Repair strategies
In our simulations, a repairer maintains a repair queue of objects to repair. An object is added to the
repair queue when at least one of the fragments of the object have been determined to be missing.
Whenever there is at least one object in the repair queue the repair policy works to repair objects.
When an object is repaired, at least k fragments are read to recover the object and then additional
fragments are generated from the object and stored at nodes which currently do not have fragments
for the object.
Objects are assigned to placement groups to determine which nodes store the fragments for the
objects. Objects within a placement group are prioritized by the repair policy: objects with the
least number of available fragments within a placement group are the next to be repaired within
that placement group. Thus, objects assigned to the same placement group have a nested object
structure: The set of available fragments for an object within a placement group is a subset of the
set of available fragments for any other object within the placement group repaired more recently.
Thus, objects within the same placement group are repaired in a round-robin order. Consecutively
repaired objects can be from different placement groups if there are multiple placement groups.
4.3. Repair for small code system
In our simulations, small code systems (using either replication or small codes) use reactive repair,
i.e., Rpeak is set to a significantly higher value than the required average repair bandwidth, and thus
repair occurs in a burst when a node failure occurs and the lost fragments from the node are repaired
as quickly as possible. Typically only a small fraction of objects are in the repair queue at any point
in time for a small code system using reactive repair.
We set the number of placement groups to P = 100·M
to the recommended Ceph [22] value,
n
and thus there are 100 placement groups assigned to each of the M nodes, with each node storing
fragments for placement groups assigned to it. The repair policy works as follows:
— If there are at most 100 placement groups that have objects to repair then each such placement
group repairs at a read repair rate of Rpeak /100.
— If there are more than 100 placement groups that have objects to repair then the 100 placement
groups with objects with the least number of available fragments are repaired at a read repair rate
of Rpeak /100.
This policy ensures that, in our simulations, the global peak read repair rate is at most Rpeak , that
the global bandwidth Rpeak is fully utilized in the typical case when one node fails and needs repair (which triggers 100 placement groups to have objects to repair), and that the maximum traffic
from and to each node is not exorbitantly high. (Unlike the case for liquid systems, there are still
significant concentrations of repair traffic in the network with this policy.)
8
4
REPAIR
4.3
Repair for small code system
1E+00
1E-01
Object Loss Probability
1E-02
1E-03
1E-04
(a)
1E-05
(b)
1E-06
1E-07
~2 days
1E-08
~10 months
1E-09
1E-10
0.05
0.5
5
Time t (in months)
50
Fig. 2: Cumulative probability of object loss versus time for: (a) (14, 10, 4) small code ; (b)
(3010, 2150, 860) large code.
The need for small code systems to employ reactive repair, i.e., use a high Rpeak value, is a consequence of the need to achieve a large MTTDL. Small codes are quite vulnerable, as the loss of
a small number of fragments can lead to object loss. For instance, with the (14, 10, 4) RS code the
loss of five or more fragments leads to object loss. Fig. 2 illustrates the drastically different time
scales at which repair needs to operate for a small code system compared to a liquid system. In this
example, nodes fail on average every three years and Fig. 2 shows the probability that N ≤ k − 1
of n nodes are functioning at time t when starting with N = n nodes functioning at time zero. As
shown, a failed node needs to be repaired within two days to avoid a 10−10 probability object loss
event for a (14, 10, 4) RS code, whereas a failed node needs to be repaired within 10 months for a
(3010, 2150, 860) large code. Note that the storage overhead β = r/n = 0.286 is the same in these
two systems. A similar observation was made in [21].
Exacerbating this for a small code system is that there are many placement groups and hence
many combinations of small number of node failures that can lead to object loss, thus requiring
a high Rpeak value to ensure a target MTTDL. For example, P = 100·M
= 21500 for a (14, 10, 4)
n
small code in a M = 3010 node system, and thus an object from the system is lost if any object from
any of the 21500 placement groups is not repaired before more than four nodes storing fragments
for the object fail. Furthermore, a small code system using many placement groups is difficult to
analyze, making it difficult to determine the value of Rpeak that ensures a target MTTDL.
Although small codes outperform replication in both storage overhead and reliability metrics,
they require a high Rpeak value to achieve a reasonable MTTDL, which leads to the repair bottleneck
described earlier. For example, Rpeak used by small code system can be controlled to be a fraction
of the total network bandwidth available in the cluster, e.g., 10% of the total available bandwidth is
mentioned in [10] and 10% − 20% was estimated in [19] to protect a small fraction of the object
data with a small code. However, using these fractions of available bandwidth for repair may or may
not achieve a target MTTDL. As we show in Section 5 via simulation, small code systems require
a large amount of available bandwidth for bursts of time in order to achieve a reasonable target
9
4.4
Repair for fixed rate liquid system
5
REPAIR SIMULATION RESULTS
MTTDL, and in many cases cannot achieve a reasonable MTTDL due to bandwidth limitations, or
due to operating with a large TRIT value.
4.4. Repair for fixed rate liquid system
The simplest repair policy for a liquid system is to use a fixed read repair rate: the value of the
peak read repair rate Rpeak is set and the read repair rate is Rpeak whenever there are objects in the
repair queue to be processed. A liquid system employs a repairer that can be described as lazy:
The value of Rpeak is set low enough that the repair queue contains almost all objects all the time.
Overall, a liquid system using lazy repair uses a substantially lower average repair bandwidth and a
dramatically lower peak repair bandwidth than a small code system using reactive repair.
There are two primary reasons that a liquid system can use lazy repair: (1) the usage of a large
code, i.e., as shown in Fig. 2, the (3010, 2150, 860) large code has substantially more time to repair
than the (14, 10, 4) small code to achieve the same MTTDL; (2) the nested object structure.
A liquid system can store a fragment of each object at each node, and thus uses a single placement
group for all objects. For example, using a (3010, 2150, 860) large code, an object is lost if the
system loses more than 860 nodes storing fragments for the object before the object is repaired.
Since there is only one placement group the nested object structure implies that if the object with
the fewest number of available of fragments can be recovered then all objects can be recovered.
This makes it simpler to determine an Rpeak value that achieves a target MTTDL for a fixed node
failure rate. The basic intuition is that the repair policy should cycle through and repair the objects
fast enough to ensure that the number of nodes that fail between repairs of the same object is at
most r. Note that Dsrc /Rpeak is the time between repairs of the same object when the aggregate size
of all objects is Dsrc and the peak global read repair bandwidth is set to Rpeak . Thus, the value of
Rpeak should be set so that the probability that there are more than r node failures in time Dsrc /Rpeak
is extremely tiny. Eq. (3) from Appendix A provides methodology for determining a value of Rpeak
that achieves a target MTTDL.
Unlike small code systems, wherein TRIT must be relatively small, a liquid systems can use large
TRIT values, which has the benefit of practically eliminating unnecessary repair due to transient
failures.
5. REPAIR SIMULATION RESULTS
The storage capacity of each node is S = 1 PB in all simulations. The system is fully loaded with
source data, i.e., Dsrc = S · M · (1 − β), where β is the storage overhead expressed as a fraction.
There are node failures in all simulations and node lifetimes are modeled as independent exponentially distributed random variables with parameter λ. Unless otherwise specified, we set 1/λ to
3 years. When sector failures are also included, sector lifetimes are also exponentially distributed
with 1/λSF fixed to 5 · 108 years.
At the beginning of each simulation the system is in a perfect state of repair, i.e., all n fragments
are available for all objects. For each simulation, the peak global read repair bandwidth is limited
to a specified Rpeak value. The MTTDL is calculated as the number of years simulated divided by
one more than the observed number of times there was an object loss event during the simulation.
The simulation is reinitialized to a perfect repair state and the simulation is continued when there
is an object loss event. To achieve accurate estimates of the MTTDL, each simulation runs for a
maximum number of years, or until 200 object loss events have occurred.
Fig. 3 presents a comparison of repair traffic for a liquid system and a small code system. In this
simulation, lazy repair for the liquid system uses a steady read repair rate of 704 Gbps to achieve
an MTTDL of 108 years, whereas the reactive repair for the small code system uses read repair rate
bursts of 6.4 Tbps to achieve an MTTDL slightly smaller than 107 years. For the small code system,
a read repair rate burst starts each time a node fails and lasts for approximately 4 hours. One of the
3010 nodes fails on average approximately every 8 hours.
In Fig. 4, we plot Rpeak against achieved MTTDL for a cluster of M = 402 nodes. The liquid
systems use fixed read repair rate equal to the indicated Rpeak and two cases of β = 33.3% and
10
5
REPAIR SIMULATION RESULTS
Fixed node failure rate with 402 nodes
5.1
7000
Peak Read Repair Rate [Gbps]
6000
5000
4000
3000
2000
1000
0
0
4
8
12
16
20
24
Time [Hours]
Fig. 3: Read repair rate comparison over a 24 hour period of reactive repair using a (14, 10, 4) small
code (solid line) and lazy repair using a (3010, 2150, 860) large code (dash line). Arrows indicate
times of node failures.
β = 16.7% are shown. The curves depict MTTDL bounds as calculated according to Eq. (3) of
Appendix A. As can be seen, the bounds and the simulation results agree quite closely. Fig. 4 also
shows simulation results for a small code system with storage overhead 33.3%, which illustrates a
striking difference in performance between small code systems and liquid systems.
5.1. Fixed node failure rate with 402 nodes
Fig. 5 plots the simulation results for Rpeak against MTTDL for a wide variety of M = 402 node
systems. Each simulation was run for 109 years or until 200 object loss events occurred. The number
of nodes, the node failure rate, and the small codes used in these simulations are similar to [10].
The square icons correspond to 16.7% storage overhead and the circle icons correspond to 33.3%
storage overhead. For the small code systems, a (18, 15, 3) small code is used for 16.7% storage
overhead, and a (9, 6, 3) small code is used for 33.3% storage overhead. For the liquid systems, a
(402, 335, 67) large code is used for 16.7% storage overhead, and a (402, 268, 134) large code is
used for 33.3% storage overhead.
The unshaded icons correspond to systems where there are only node failures, whereas the shaded
icons correspond to systems where there are both node failures and sector failures.
The small code systems labeled as “SC, 30min” and “SC, 24hr” use the read repair rate strategy
described in Section 4.3 based on the indicated Rpeak value, with TRIT set to 30 minutes and 24 hours
respectively. The triplication systems labeled as “Trip, 30min” and “Trip, 24hr” use the read repair
rate strategy described in Section 4.3 based on the indicated Rpeak value with TRIT is set to 30 minutes
and 24 hours respectively.
The liquid systems labeled as “LiqF, 24hr” use a fixed read repair rate set to the shown Rpeak ,
which was calculated according to Eq. (3) from Appendix A with a target MTTDL of 107 years.
The value of TRIT is set to 24 hours As would be expected since Eq. (3) is a lower bound, the observed
MTTDL in the simulations is somewhat above 107 years.
11
5.1
Fixed node failure rate with 402 nodes
Peak Global Repair Bandwidth [Gbps]
10000
5
REPAIR SIMULATION RESULTS
(c)
1000
(b)
(a)
100
10
1.0E+10
1.0E+9
1.0E+8
1.0E+7
1.0E+6
1.0E+5
1.0E+4
1.0E+3
1.0E+2
1.0E+1
1.0E+0
MTTDL [Years]
Fig. 4: Plot of MTTDL bound vs. Rpeak for liquid systems with: (a) (402, 268, 134) code (33.3%
storage overhead); (b) (402, 335, 67) code (16.7% storage overhead). Round markers on curve are
simulation results. Square markers (c) are simulation results for small code system with (9, 6, 3)
code (33.3% storage overhead).
The liquid systems labeled as “LiqR, 24hr” have TRIT set to 24 hours and use a version of the
regulator algorithm described in Section 9 to continually and dynamically recalculate the read repair rate according to current conditions. The regulator used (per object) node failure arrival rate
estimates based on the average of the last 76 · r − F node failure inter-arrival times where F denotes
the number of missing fragments for an object at the time it forms the estimate. The maximum read
repair rate was limited to the shown Rpeak value, which is about three times the average rate. In these
runs there were no object loss events in the 109 years of simulation when there are both node failures and sector failures. This is not unexpected, since the estimate (using the MTTDL lower bound
techniques described in Section 9 but with a slightly different form of node failure rate estimation)
for the actual MTTDL is above 1025 years for the 33.3% storage overhead case and above 2.5 · 1012
years for the 16.7% storage overhead case. These estimates are also known to be conservative when
compared to simulation data. The same estimate shows that with this regulator an MTTDL of 109
years would be attained with the smaller overhead of r = 115 instead of r = 134. As shown in
Table II, the average read repair rate Ravg is around 1/3 of Rpeak , and the 99% read repair rate R99% is
around 1/2 of Rpeak . The large MTTDL achieved by the regulator arises from its ability to raise the
repair rate when needed. Other versions of the regulator in which the peak read repair rate is not so
limited can achieve much larger MTTDL. For example, a version with a read repair rate limit set to
be about five times the average rate for the given node failure rate estimate achieves an MTTDL of
around 1039 years for the 33.3% storage overhead case.
As can be seen, the value of Rpeak required for liquid systems is significantly smaller than that
required for small code systems and for triplication. Furthermore, although regulated rate liquid
system can provide an MTTDL of greater than 1025 years even when TRIT is set to 24 hours and
there are sector failures in addition to node failures, the small code systems and triplication do not
provide as good an MTTDL even when TRIT is set to 30 minutes and there are no sector failures, and
provide a poor MTTDL when TRIT is set to 24 hours or when there are sector failures. The MTTDL
12
5
REPAIR SIMULATION RESULTS
5.2
Fixed node failure rate with 3010 nodes
100,000
16.7% Overhead
16.7% Overhead + SF
33.3% Overhead
SC, 24hr
SC, 30min
SC, 30min
33.3% Overhead + SF
66.7% Overhead
Peak Repair Bandwidth (in Gbps)
10,000
Trip, 30min
SC, 30min
1,000
Trip, 24hr
SC, 24hr
SC, 30min
LiqR, 24hr
LiqR, 24hr
LiqF, 24hr
100
LiqF, 24hr
10
1.0E+10
1.0E+9
1.0E+8
1.0E+7
1.0E+6
1.0E+5
1.0E+4
1.0E+3
1.0E+2
1.0E+1
MTTDL (in years)
Fig. 5: Peak read repair rate versus MTTDL for a M = 402 node system for fixed 1/λ.
would be further degraded for small code systems if both TRIT were set to 24 hours and there were
sector failures.
5.2. Fixed node failure rate with 3010 nodes
Fig. 6 shows detailed simulation results, plotting the value of Rpeak and the resulting value of MTTDL
for various M = 3010 node systems. Each simulation was run for 108 years or until 200 object loss
events occurred. The number of nodes, the node failure rate, and the small code systems used in
these simulations are similar to [19]. The square icons correspond to 14.3% storage overhead and
the circle icons correspond to 28.6% storage overhead. For the small code systems, a (24, 20, 4)
small code is used for 14.3% storage overhead, and a (14, 10, 4) small code is used for 28.6%
storage overhead. For the liquid systems, a (3010, 2580, 430) large code is used for 14.3% storage
overhead, and a (3010, 2150, 860) large code is used for 28.6% storage overhead. The remaining
parameters and terminology used in Fig. 6 are the same as in Fig. 5. There were no object loss events
in the 108 years of simulation for any of the liquid systems shown in Fig. 6.
The systems “LiqR, 24hr” liquid systems used a regulated repair rate similar to that described
for the 402 node system described above. The target repair efficiency was set at 32 r. In this case the
average repair rate is higher than for the fixed rate case because the fixed rate case was set using the
target MTTDL which, due the larger scale, admits more efficient repair than the 402 node system
for the same target MTTDL. Estimates of MTTDL for the regulated case using techniques from
Section 9 indicate that the MTTDL is greater than an amazing 10140 years for the 28.6% storage
overhead case and 1066 years for the 14.3% storage overhead case even without the benefit of node
failure rate estimation. As before, the regulated repair system could be run at substantially lower
13
5.2
Fixed node failure rate with 3010 nodes
5
REPAIR SIMULATION RESULTS
100,000
14.3% Overhead
14.3% Overhead + SF
28.6% Overhead
28.6% Overhead + SF
SC, 30min
SC, 30min
SC, 24hr
Peak Repair Bandwidth (in Gbps)
66.7% Overhead
Trip, 30min
10,000
Trip, 24hr
SC, 30min
LiqR, 24hr
SC, 30min
SC, 24hr
LiqR, 24hr
LiqF, 24hr
1,000
LiqF, 24hr
100
1.0E+9
1.0E+8
1.0E+7
1.0E+6
1.0E+5
1.0E+4
1.0E+3
1.0E+2
1.0E+1
MTTDL (in years)
Fig. 6: Peak read repair rate versus MTTDL for a M = 3010 node system for fixed 1/λ.
overhead while still maintaining a large MTTDL. For the “LiqR, 24hr” liquid systems, the average
read repair rate Ravg is around 1/3 of Rpeak , and the 99% read repair rate R99% is around 1/2 of Rpeak .
The conclusions drawn from comparing small code systems to liquid systems shown in Fig. 5
are also valid for Fig. 6 but to an even greater degree. The benefits of the liquid system approach
increase as the system size increases.
The paper [19] describes a repair bottleneck for a small code system, hereafter referred to as the
Facebook system, which has around 3000 nodes and storage capacity per node of approximately
S = 15 TB. The Facebook system uses a (14, 10, 4) RS code to protect 8% of the source data and
triplication to protect the remaining 92% of the source data. The amount of repair traffic estimated
in [19] for the Facebook system is around 10% to 20% of the total average of 2 PB/day of cluster
network traffic. The paper [19] projects that the network would be completely saturated with repair
traffic if even 50% of the source data in the Facebook system were protected by the RS code.
Taking into account the mix of replicated and RS protected source data and the mixed repair
traffic cost for replicated and RS protected source data, there are around 6.5 to 13 node failures per
day in the Facebook system. An average of 6.5 to 13 node failures per day for a 3000 node system
implies that each individual node on average fails in around 115 to 230 days, which is around 2.5 to
5 times the node failure rate of 1/λ = 3 years we use in most of our simulations. If all source data
were protected by a (14, 10, 4) RS code in the Facebook system then the repair traffic per day would
average around 1 to 2 PB. However, the node failure per day statistics in [19] have wide variations,
which supports the conclusions in [19] that protecting most of the source data with the RS code in
the Facebook system will saturate network capacity.
Note that 1 to 2 PB/day of repair traffic for the Facebook system implies an average read repair
rate of around 90 to 180 Gbps. If the storage capacity per node were S = 1 PB instead of S = 15
14
5
REPAIR SIMULATION RESULTS
5.3
Varying node failure rate
TB then the Facebook system average read repair rate Ravg would be approximately 6 to 12 Tbps,
which is around 2.5 to 5 times the average read repair rate Ravg for the very similar “SC, 30min”
small code system shown in Fig. 6. This relative difference in the value of Ravg makes sense, since
the Facebook system node failure rate is around 2.5 to 5 times larger than for the“SC, 30min” small
code system.
The “SC, 30min” small code system shown in Fig. 6 achieves an MTTDL of just under 107 years
using a peak read repair rate Rpeak = 6.4 Tbps and using TRIT = 30 minutes. Assume the node
failure rate for the Facebook system is 5 times the node failure rate for the “SC, 30min” small code
system shown in Fig. 6, and suppose we set TRIT = 30/5 = 6 minutes for the Facebook system.
The scaling observations in Section 5.5 show that this Facebook system achieves a MTTDL of just
under 107 /5 = 2 · 106 years when Rpeak = 6.4 · 5 = 32 Tbps.
The average node failure rates reported in [19] for the Facebook system are considerably higher
than the node failure rates we use in our simulations. One possible reason is that the Facebook
system uses a small TRIT value, e.g., TRIT = 10 minutes, after which unresponsive nodes trigger repair.
This can cause unnecessary repair of fragments on unresponsive nodes that would have recovered if
TRIT were larger.
It is desirable to set TRIT as large as possible to avoid unnecessary repair traffic, but only if a
reasonable MTTDL can be achieved. The results in Fig. 6 indicate that a small code system cannot
offer a reasonable MTTDL with TRIT = 24 hours. On the other hand, the results from Section 5.4
indicate that the average read repair rate increases significantly (indicating a significant amount of
unnecessary repair) if a smaller value of TRIT is used for a small code system when there are transient
node failures. In contrast, a regulated liquid system can provide a very large MTTDL operating with
TRIT = 24 hours and with sector failures, thus avoiding misclassification of transient node failures
and unnecessary repair traffic.
Results for small code systems using LR codes [9], [10], [19] can be deduced from results for
small code systems using RS codes. A typical (14, 10, 2, 2) LR code (similar to the one described
in [10] which defines a (16, 12, 2, 2) LR code) partitions 10 source fragments into two groups of
five, generates one local repair fragment per group and two global repair fragments for a total of
14 fragments, and thus the storage overhead is the same as for a (14, 10, 4) RS code. At least five
fragments are read to repair a lost fragment using the LR code, whereas 10 fragments are read to
repair a lost fragment using the RS code. However, there are fragment loss patterns the LR code
does not protect against that the RS code does. Thus, the LR code operating at half the Rpeak value
used for the RS code achieves a MTTDL that is lower than the MTTDL for the RS code.
Similarly, results for [19] which use a (16, 10, 4, 2) LR code can be compared against the
(14, 10, 4) RS code that we simulate. The LR code operating at half the Rpeak value used for the
RS code achieves an MTTDL that may be as large as the MTTDL for the RS code. However, the
LR code storage overhead is β = 0.375, which is higher than the RS code storage overhead of
β = 0.286.
5.3. Varying node failure rate
In this subsection we demonstrate the ability of the regulated repair system to respond to bursty node
failure processes. The node failures in all simulations described in this subsection are generated by
a time varying periodic Poisson process repeating the following pattern over each ten year period
of time: the node failure rate is set to 1/λ = 3 years for the first nine years of each ten year period,
and then 1/λ = 1 year for the last year of each ten year period. Thus, in each ten year period, the
node failure rate is the same as it was in the previous subsections for the first nine years, followed
by a node failure rate that is three times higher for the last year.
Except for using variable λ for node failures, Fig. 7 is similar to Fig. 5. The small code systems
labeled as “SC, 30min” in Fig. 7 use the read repair rate strategy described in Section 4.3 based on
the shown Rpeak value, and TRIT is set to 30 minutes. Thus, even when the node failure rate is lower
during the first nine years of each period, the small code system still sets read repair rate to Rpeak
whenever repair is needed.
15
5.4
Transient node failures
5
REPAIR SIMULATION RESULTS
100,000
16.7% Overhead
33.3% Overhead
Peak Repair Bandwidth (in Gbps)
SC, 30min
10,000
SC, 30min
LiqR, 24hr
1,000
LiqR, 24hr
100
1.0E+10
1.0E+9
1.0E+8
1.0E+7
1.0E+6
1.0E+5
1.0E+4
MTTDL (in years)
Fig. 7: Peak read repair rate versus MTTDL for a M = 402 node system for variable 1/λ.
The liquid systems labeled as “LiqR, 24hr” in Fig. 7 use the regulator algorithm described in
Section 9 to continually and dynamically recalculate the read repair rate according to current conditions, allowing a maximum read repair rate up to the shown Rpeak value, and TRIT is set to 24 hours.
In these runs there were no object loss events in the 109 years of simulation with both node failures
and sector failures. [[We may want to compute something here TBD?]]
Fig. 18 is an example trace of the read repair rate as function of time for the liquid system labeled
“LiqR, 24hr” in Fig. 7, which shows how the the read repair rate automatically adjusts as the node
failure rate varies over time. In these simulations, the average read repair rate Ravg is around 1/9 of
Rpeak during the first nine years of each period and around 1/3 of Rpeak during the last year of each
period, and the 99% read repair rate R99% is around 1/6 of Rpeak during the first nine years of each
period and around 1/2 of Rpeak during the last year of each period. Thus, the regulator algorithm
does a good job of matching the read repair rate to what is needed according to the current node
failure rate.
Except for using variable λ for node failures, Fig. 8 is similar to Fig. 6. The conclusions comparing small code systems to liquid systems shown in Fig. 7 are also valid for Fig. 8.
5.4. Transient node failures
In this subsection we demonstrate the ability of the liquid systems to efficiently handle transient
node failure processes.
Fig. 10 plots the simulation results for Ravg against MTTDL for a number of M = 402 node systems. Like in previous subsections, the node lifetimes for node failures are modeled as independent
exponentially distributed random variables with parameter λ and is set to 1/λ = 3 years. The occurrence times of transient node failures are modeled as independent exponentially distributed random
variables with parameter λTF and is set to 1/λTF = 0.33 years. Thus, transient node failures occur
at 9 times the rate at which node failures occur, consistent with [8]. The durations of transient node
failures are modeled with log-logistic random variables, having a median of 60 seconds and shape
parameter 1.1. These choices were made so as to mimic the distribution provided in [8]. Figure 9
16
5
REPAIR SIMULATION RESULTS
5.4
Transient node failures
100,000
14.3% Overhead
28.6% Overhead
Peak Repair Bandwidth (in Gbps)
SC, 30min
SC, 30min
LiqR, 24hr
10,000
LiqR, 24hr
1,000
1.0E+9
1.0E+8
1.0E+7
1.0E+6
MTTDL (in years)
Fig. 8: Peak read repair rate versus MTTDL for a M = 3010 node system for variable 1/λ.
Fig. 9: Log-logistic transient failure duration model versus measured data
shows the graph from [8] with the log-logistic fitted curve overlaid. With this model, less than 10%
of the transient node failures last for more than 15 minutes.
The unshaded markers mark simulations with just node failures and the shaded markers mark
simulations with both node failures and transient node failures.
The Rpeak values in all simulations is the same as Rpeak value used correspondingly in Fig. 5.
The small code systems labeled as “SC, 30min” and “SC, 24hr“ in Fig. 10 use the read repair rate
strategy described in Section 4.3, with Rpeak value of 6400 Gbps and TRIT is set to 30 minutes and 24
hours respectively. The liquid system labeled as “LiqR, 24hr” in Fig. 10 uses the regulator algorithm
described in Section 9, with Rpeak value of 311 Gbps and TRIT is set to 24 hours.
17
5.5
Scaling parameters
5
REPAIR SIMULATION RESULTS
300
Permanent + Transient Node Failures
Permanent Node Failures
Average Repair Bandwidth (in Gbps)
250
SC, 30min
SC, 24hr
SC, 30min
200
150
100
LiqR, 24hr
50
0
1.0E+10
1.0E+9
1.0E+8
1.0E+7
1.0E+6
1.0E+5
1.0E+4
1.0E+3
1.0E+2
1.0E+1
MTTDL (in years)
Fig. 10: Average read repair rate versus MTTDL for a M = 402 node system for transient 1/λ.
As evident, there is no difference in the Ravg for the liquid system between simulations with
transient node failures and those without. No object loss events were observed for the liquid system
in the 109 years of simulation with both node failures and transient node failures. The Ravg for small
code system labeled as “SC, 30min“ is however higher for the simulation with transient node failures
and achieves an MTTDL that is less than half of what is achieved when there are no transient node
failures. The Ravg for small code system labeled as “SC, 24hr“ is the same between simulations with
transient node failures and those without, but they fail to achieve a high enough MTTDL and do not
provide adequate protection against object loss.
Thus, the liquid systems do a better job of handling transient node failures without requiring more
effective average bandwidth than small code systems.
5.5. Scaling parameters
The simulation results shown in Section 5 are based on sets of parameters from published literature,
and from current trends in deployment practices. As an example of a trend, a few years ago the
storage capacity of a node was around S = 16 TB, whereas the storage capacity of a node in some
systems is S = 1 PB or more, and thus the storage capacity per node has grown substantially.
Repair bandwidths scale linearly as a function of the storage capacity S per node (keeping other
input parameters the same). For example, the values of (Rpeak , Ravg ) for a system with S = 512 TB
can be obtained by scaling down by a factor of two the values of (Rpeak , Ravg ) from simulation results
for S = 1 PB, whereas the MTTDL and other values remain unchanged.
Repair bandwidths and the MTTDL scale linearly as a function of concurrently scaling the node
failure rate λ and the repair initiation timer TRIT . For example, the values of (Rpeak , Ravg ) for a system
with 1/λ = 9 years and TRIT = 72 hours can be obtained by scaling down by a factor of three the
18
6
ENCODING AND DECODING OBJECTS
EFIs
0
1
2
3
4
5
ENC
Object
6
7
8
Repair fragments
Fig. 11: Example of block storage organization. A data object is segmented into k = 6 source
fragments and encoded by a small code to generate r = 3 repair fragments.
values of (Rpeak , Ravg ) from simulation results for 1/λ = 3 years and TRIT = 24 hours, whereas the
MTTDL can be obtained by scaling up by a factor of three, and other values remain unchanged.
6. ENCODING AND DECODING OBJECTS
In a block storage organization commonly used by small code systems, objects are partitioned into
k successive source fragments and encoded to generate r = n − k repair fragments (we assume the
small code is MDS). In effect a fragment is a symbol of the small code, although the small code
may be intrinsically defined over smaller symbols, e.g., bytes. A high level representation of a block
storage organization for a simple example using a (9, 6, 3) Reed-Solomon code is shown in Fig. 11,
where the object runs from left to right across the source fragments as indicated by the blue line.
A relatively small chunk of contiguous data from the object that resides within a single source
fragment can be accessed directly from the associated node if it is available. If, however, that node is
unavailable, perhaps due to a transient node failure or node failure, then k corresponding chunks of
equal size must be read from other nodes in the placement group and decoded to generate the desired
chunk. This involves reading k times the size of the missing chunk and performing a decoding
operation, which is referred to as a degraded read [19], [18]. Reading k chunks can incur further
delays if any of the nodes storing the chunks are busy and thus non-responsive. This latter situation
can be ameliorated if a number of chunks slightly larger than k are requested and decoding initiated
as soon as the first k chunks have arrived, as described in [11].
A flow storage organization, used by liquid systems, operates in a stream fashion. An (n, k, r)
large code is used with small symbols of size Ssize, resulting in a small source block of size Bsize =
k · Ssize. For example, with k = 1024 and symbols of size Ssize = 64 Bytes, a source block is of
size Bsize = 64 KB. An object of size Osize is segmented into
N=
Osize
Bsize
source blocks, and the k source symbols of each such source block is independently erasure encoded
into n symbols. For each i = 0, . . . , n − 1, the fragment with EFI i consists of the concatenation of
the ith symbol from each of the N consecutive source blocks of the object. An example of a flow
storage organization showing a (1536, 1024, 512) large code with N = 24, 576 is shown in Fig. 12,
where each row corresponds to a source block and runs from left to right across the source symbols
as indicated by the dashed blue lines, and the object is the concatenation of the rows from top to
bottom.
A chunk of data consisting of a consecutive set of source blocks can be accessed as follows. For
a fragment with EFI i, the consecutive portion of the fragment that corresponds to the ith symbol
from each of consecutive set of source blocks is read. When such portions from at least k fragments
are received (each portion size a 1/k-fraction of the chunk size), the chunk of data can be recovered
by decoding the consecutive set of source blocks in sequence. Thus, the amount of data that is read
to access a chunk of data is equal to the size of the chunk of data. Portions from slightly more than
k fragments can be read to reduce the latency due to nodes that are busy and thus non-responsive.
Note that each access requires a decoding operation.
Let us contrast the two organizations with a specific example using an object of size Osize = 1.5
GB. Consider an access request for the 32 MB chunk of the object in positions 576 MB through
19
6
ENCODING AND DECODING OBJECTS
EFIs
1023
…
…
…
Source block 24575
1024
ENC
Object
…
…
1535
…
2
Source block 0
…
1
…
0
Repair fragments
Fig. 12: Example of a flow storage organization. A data object is segmented into 24, 576 source
blocks, and each source block is segmented into k = 1024 source symbols and encoded by a
large code to generate r = 512 repair symbols. The concatenation of the ith symbol from each of
the source blocks forms the fragment with EFI i. Thus the vertical columns labeled by the EFIs
correspond to fragments.
0
1
2
3
4
5
6
7
8
Block Storage Access
…
1024
Source block 0
Source block 9216
Source block 9727
…
…
1535
… … …
…
1023
… … …
2
… … …
1
… … …
0
Source block 24575
Flow Storage Access
Fig. 13: Comparison of accessing a 32 MB chunk of an object for block storage organization and
flow storage organization. The shaded region indicates the chunk of the object to be accessed.
608 MB, which is shown as the shaded region in Fig. 13 with respect to both a block storage
organization and a flow storage organization.
With the block storage organization shown in Fig. 11, the 32 MB chunk is within the fragment
with EFI = 2 as shown in Fig. 13. If the node that stores this fragment is available then the chunk
can be read directly from that node. If that node is unavailable, however, then in order to recover
the 32 MB chunk the corresponding 32 MB chunks must be retrieved from any six available nodes
of the placement group and decoded. In this case a total of 192 MB of data is read and transferred
through the network to recover the 32 MB chunk of data.
With the flow storage organization shown in Fig. 12, the chunk of size 32 MB corresponds to 512
consecutive source blocks from source block 9216 to block 9727 as shown in Fig. 13, which can be
retrieved by reading the corresponding portion of the fragment from each of at least 1024 fragments.
From the received portions, the chunk can be recovered by decoding the 512 source blocks. Thus,
in a flow storage organization a total of 32 MB of data is read and transferred through the network
to recover the original 32 MB chunk of data.
Naturally, the symbol size, source block size, and the granularity of the data to be accessed should
be chosen to optimize the overall design, taking into account constraints such as whether the physical storage is hard disk or solid state.
20
7
PROTOTYPE IMPLEMENTATION
7. PROTOTYPE IMPLEMENTATION
Our team implemented a prototype of the liquid system. We used the prototype to understand and
improve the operational characteristics of a liquid system implementation in a real world setting,
and some of these learnings are described in Section 8. We used the prototype to compare access
performance of liquid systems and small code systems. We used the prototype to cross-verify the
lazy repair simulator that uses a fixed read repair rate as described in Section 4.4 to ensure that they
both produce the same MTTDL under the same conditions. We used the prototype to validate the
basic behavior of a regulated read repair rate as described in Section 9.
The hardware we used for our prototype consists of a rack of 14 servers. The servers are connected
via 10 Gbit full duplex ethernet links to a switch, and are equipped with Intel Xeon CPUs running
at 2.8 GHz, with each CPU having 20 cores. Each server is equipped with an SSD drive of 600 GB
capacity that is used for storage.
The liquid prototype system software is custom written in the Go programming language. It
consists of four main modules:
— Storage Node (SN) software. The storage node software is a HTTP server. It is used to store
fragments on the SSD drive. Since a liquid system would generally use many more than 14 storage
nodes, we ran many instances of the storage node software on a small number of physical servers,
thereby emulating systems with hundreds of nodes.
— An Access Generator creates random user requests for user data and dispatches them to accessors.
It also collects the resulting access times from the accessors.
— Accessors take user data requests from the access generator and create corresponding requests
for (full or partial) fragments from the storage nodes. They collect the fragment responses and
measure the amount of time it takes until enough fragments have been received to recreate the
user data. There are in general multiple accessors running in the system.
— A Repair Process is used to manage the repair queue and regenerate missing fragments of stored
objects. The repair process can be configured to use a fixed read repair rate, or to use a regulated
read repair rate as described in Section 9.
In addition, the team developed a number of utilities to test and exercise the system, e.g., utilities to
cause node failures, and tools to collect data about the test runs.
7.1. Access performance setup
The users of the system are modeled by the access generator module. We set a target user access load
ρ (0 < ρ < 1) on the system as follows. Let C be the aggregate capacity of the network links to the
storage nodes, and let s be the size of the user data requests that will be issued. Then C/s requests
per unit of time would use all available capacity. We set the mean time between successive user
s
requests to t = ρ·C
, so that on average a fraction ρ of the capacity is used. The access generator
module uses an exponential random variable to generate the timing of each successive user data
request, where the mean of the random variable is t.
As an example, there is a 10 Gbps link to each of the six storage nodes in the setup of Fig. 14,
and thus C = 60 Gbps. If user data requests are of size s = 10 MB = 80 · 220 bits, and the target
load is ρ = 0.8 (80% of capacity), then the mean time between requests is set to
80 · 220
≈ 1.75 milliseconds.
0.8 · 60 · 109
The access generator module round-robins the generated user data requests across the accessor
modules. When an accessor module receives a generated user data request from the access generator
module, the accessor module is responsible for making fragment requests to the storage nodes and
collecting response payloads; it thus needs to be fairly high performance. Our implementation of
an accessor module uses a custom HTTP stack, which pipelines requests and keeps connections
statically alive over extended periods of time. Fragment requests are timed to avoid swamping the
t=
21
7.2
Access performance tests
7
PROTOTYPE IMPLEMENTATION
switch with response data and this reduces the risk of packet loss significantly. The server network
was also tuned, e.g., the TCP retransmission timer was adjusted for the use case of a local network.
See [5] for a more detailed discussion of network issues in a cluster environment. These changes
in aggregate resulted in a configuration that has low response latency and is able to saturate the
network links fully.
7.2. Access performance tests
The prototype was used to compare access performance of liquid systems and small code systems.
Our goal was to evaluate access performance using liquid system implementations, i.e., understand
the performance impact of requesting small fragments (or portions of fragments) from a large number of storage servers. For simplicity we evaluated the network impact on access performance.
Fragment data was generally in cache and thus disk access times were not part of the evaluation.
We also did not evaluate the computational cost of decoding. Some of these excluded aspects are
addressed separately in Section 8.
Figure 14 displays the setup that we use for testing access speed. Each physical server has a
full duplex 10 GbE network link which is connected over a switch to all the other server blades.
Testing shows that we are able to saturate all of the 10 Gbps links of the switch simultaneously.
Eight of our 14 physical servers are running accessors, and one of the eight is running the access
generator as well. The remaining six servers are running instances of the storage node software.
This configuration allows us to test the system under a load that saturates the network capacity to
the storage nodes: The aggregate network capacity between the switch and the accessors is 80 Gbps,
which is more than the aggregate network capacity of 60 Gbps between the switch and the storage
nodes, so the bottleneck is the 60 Gbps between the switch and the storage nodes.
During access performance tests there are no node failures, all n fragments for each object are
available, and the repair process is disabled. For all tests, 67 storage node instances run on each
storage server, which emulates a system of 402 nodes. We operate with a storage overhead of 33.3%,
a (402, 268, 134) large code is used for the liquid system, and a (9, 6, 3) small code is used for the
small code system. We tested with 10 MB and 100 MB user data request sizes.
We run the access prototype in three different configurations. The “Liq” configuration models
liquid system access of user data. For a user data request of size s, the accessor requests k + E
fragment portions, each of size s/k, from a random subset of distinct storage nodes, and measures
the time until the first k complete responses are received (any remaining up to E responses are
discarded silently by the accessor). Each fragment portion is of size around 25.5 KB when s = 10
MB, and around 255 KB when s = 100 MB. We use E = 30, and thus the total size of requested
fragment portions is around 10.75 MB when s = 10 MB, and around 107.5 MB when s = 100 MB.
See [11] for an example of this access strategy.
The “SC” configuration models small code system normal access of user data, i.e., the requested
user data is stored on a storage node that is currently available. For a user data request of size s,
the accessor requests one fragment portion of size s from a randomly selected storage node, and
measures the time until the complete response is received. The total size of the requested fragment
portion is 10 MB when s = 10 MB, and 100 MB when s = 100 MB.
The “SC-Deg” configuration models small code system degraded access of user data, i.e., the
requested user data is stored at a storage node that that is currently unavailable, e.g., see [19],
[18]. For a user data request of size s, the accessor requests k fragment portions of size s from a
random subset of distinct storage nodes and measures the time until the first k complete responses
are received. The total size of requested fragment portions is 60 MB when s = 10 MB, and 600 MB
when s = 100 MB. Most user data is stored at available storage nodes, and thus most accesses are
normal, not degraded, in operation of a small code system. Thus, we generate the desired load on
the system with normal accesses to user data, and run degraded accesses at a rate that adds only a
nominal aggregate load to the system, and only the times for the degraded accesses are measured by
the accessors. Similar to the “Liq” configuration, we could have requested k + 1 fragment portions
22
7
PROTOTYPE IMPLEMENTATION
Access Blade
7.2
Access Blade
Access Blade
Accessor
Accessor
Accessor
Access performance tests
Access Blade
Accessor
Access Blade
Access Blade
Accessor
Accessor
Access Blade
Access Blade
Accessor
Accessor
Access
Generator
Switch 10 GbE
80 Gbps
aggregate
link capacity
60 Gbps aggregate
link capacity
Storage Blade
Storage Blade
SN
SN
SN
SN
SN
SN
SN
SN
Storage Blade
...
SN
SN
SN
SN
SN
SN
SN
SN
SN
SN
SN
SN
SN
SN
SN
SN
SN
SN
SN
SN
SN
SN
SN
...
SN
Storage Blade
Storage Blade
SN
SN
...
...
SN
SN
SN
SN
SN
SN
SN
SN
SN
SN
SN
...
Storage Blade
SN
SN
SN
SN
SN
...
SN
SN
SN
SN
Fig. 14: Setup used for access performance testing
for the “SC-Deg” configuration to decrease the variation in access times, but at the expense of even
more data transfer over the network.
Figures 15 and 16 depict access time results obtained with the above settings under different
system loads. The average time for liquid system accesses is less than for small code system normal
accesses in most cases, except under very light load when they are similar on average. Even more
striking, the variation in time is substantially smaller for liquid systems than for small code systems
under all loads. Liquid system accesses are far faster (and consume less network resources) than
small code system degraded accesses.
Our interpretation of these results is that liquid systems spread load equally, whereas small code
systems tend to stress individual storage nodes unevenly, leading to hotspots. With small code systems, requests to heavily loaded nodes result in slower response times. Hotspots occur in small code
systems when appreciably loaded, even when the fragment requests are distributed uniformly at random. It should be expected that if the requests are not uniform, but some data is significantly more
popular than other data, then response times for small code systems would be even more variable.
23
7.3
Repair tests and verification of the simulator 8
IMPLEMENTATION CONSIDERATIONS
400
99%
350
50%
1%
Response Time (in ms)
300
250
200
150
100
50
0
Liq
SC
SC-Deg
20% load
Liq
SC
SC-Deg
50% load
Liq
SC
SC-Deg
66% load
Liq
SC
SC-Deg
80% load
Fig. 15: Access Results for 10 MB object requests.
7.3. Repair tests and verification of the simulator
We used the prototype repair process to verify our liquid system repair simulator. We started from
realistic sets of parameters, and then sped them up by a large factor (e.g., 1 million), in such a way
that we could run the prototype repair process and observe object loss in a realistic amount of time.
We ran the prototype repair process in such a configuration, where node failures were generated
artificially by software. The resulting measured statistics, in particular the MTTDL, were compared
to an equivalent run of the liquid system repair simulator. This allowed us to verify that the liquid
system repair simulator statistics matched those of the prototype repair process and also matched
our analytical predictions.
8. IMPLEMENTATION CONSIDERATIONS
Designing a practical liquid system poses some challenges that are different from those for small
code systems. In the following sections we describe some of the issues, and how they can be addressed.
8.1. Example Architecture
As discussed in Section 2.4, Fig. 1 shows an example of a liquid system architecture. This simplified
figure omits components not directly related to source data storage, such as system management
components or access control units.
Concurrent capacity to access and store objects scales with the number of access proxy servers,
while source data storage capacity can be increased by adding more storage nodes to the system, and
24
8
IMPLEMENTATION CONSIDERATIONS
8.2
Metadata
4000
99%
3500
50%
1%
Response Time (in ms)
3000
2500
2000
1500
1000
500
0
Liq
SC
20% load
SC-Deg
Liq
SC
SC-Deg
50% load
Liq
SC
SC-Deg
66% load
Liq
SC
SC-Deg
80% load
Fig. 16: Access Results for 100 MB object requests.
thus these capacities scale independently with this architecture. Storage nodes (storage servers) can
be extremely simple, as their only function is to store and provide access to fragments, which allows
them to be simple and inexpensive. The lazy repair process is independent of other components; thus
reliability is handled by a dedicated system.
8.2. Metadata
The symbol size used by a liquid system with a flow storage organization is substantially smaller
than the size used by a small code system with a block storage organization. For a small code system
each symbol is a fragment, whereas for a liquid system there are many symbols per fragment and,
as described in Section 6, the symbol to fragment mapping can be tracked implicitly. For both types
of systems, each fragment can be stored at a storage node as a file.
The names and locations of each fragment are explicitly tracked for small code systems. There
are many more fragments to track for a liquid system because of the use of a large code, and explicit
tracking is less appealing. Instead, a mapping from the placement group to available nodes can be
tracked, and this together with a unique name for each object can be used to implicitly derive the
names and locations of fragments for all objects stored at the nodes for liquid systems.
8.3. Network usage
The network usage of a liquid system is fairly different from that of a small code system. As described in Section 7.2, liquid systems use network resources more smoothly than small code systems, but liquid systems must efficiently handle many more requests for smaller data sizes. For
example, a small code system using TCP might open a new TCP connection to the storage node for
25
8.4
Storage medium
8
IMPLEMENTATION CONSIDERATIONS
each user data request. For a liquid system, user data requests are split into many smaller fragment
requests, and opening a new TCP connection for each fragment request is inefficient. Instead, each
access proxy can maintain permanent TCP connections to each of the storage nodes, something that
modern operating systems can do easily. It is also important to use protocols that allow pipeline
data transfers over the connections, so that network optimizations such as jumbo packets can be
effectively used. For example the HTTP/2 protocol [1] satisfies these properties.
Technologies such as RDMA [16] can be used to eliminate processing overhead for small packets
of data. For example, received portions of fragments can be placed autonomously via RDMA in the
memory of the access proxy, and the access proxy can decode or encode one or more source blocks
of data concurrently.
8.4. Storage medium
A key strength of liquid systems is that user data can be accessed as a stream, i.e., there is very little
startup delay until the first part of a user data request is available, and the entire user data request
is available with minimal delay, e.g., see Section 7.2. Nevertheless, random requests for very small
amounts of user data is more challenging for a liquid system.
The storage medium used at the storage nodes is an important consideration when deciding the
large code parameters (n, k, r). Hard disk drives (HDDs) efficiently support random reads of data
blocks of size 500 KB or larger, whereas solid state drives (SSDs) efficiently support random reads
of data blocks of size 4 KB or larger. We refer to these data block sizes as the basic read size of the
storage medium. The parameters (n, k, r) should be chosen so that if s is the typical requested user
data size then s/k is at least the basic read size, which ensures that a requested fragment portion
size is at least the basic read size. In many cases this condition is satisfied, e.g., when the requested
user data size is generally large, or when the storage medium is SSD.
In other cases, more advanced methods are required to ensure that a liquid system is efficient.
For example, user data request performance can be improved by assigning different size user data
to different types of objects: Smaller user data is assigned to objects that are stored using smaller
values of k, and larger user data is assigned to objects that are stored using larger values of k, where
in each case s/k is at least the basic read size if s is the requested user data size. In order to achieve
a large MTTDL for all objects, the storage overhead nr is set larger for objects stored using smaller
values of k than for objects stored using larger values of k. Overall, since the bulk of the source data
is likely to be large user data, the overall storage overhead remains reasonable with this approach.
Only a small proportion of user data is frequently accessed in many use cases, e.g., in the use case
described in [15], only about 10% of the user data is accessed frequently, and this user data is easy
to identify based on the amount of time it has been stored in the storage system. Thus, a caching
layer can be used to store and make such user data available without having to request the user data
from the storage system. For a liquid system, it makes most sense to cache user data that is both
small and frequently accessed.
Another strategy to minimize requests for small user data is to store related user data together in
objects and coalesce accesses. Suppose for example, the storage system stores web content. To load
a web page, dozens of small files of user data are typically loaded, such as HTML content and small
images. Since the same content is accessed whenever such a web page is displayed, the system can
store and access this user data together.
8.5. RaptorQ software performance
With small code systems, user data can usually be accessed directly without requiring erasure decoding. With liquid systems, each access of user data typically requires erasure decoding.
RaptorQ codes are the large codes of choice for liquid systems. The encoding and decoding
complexity of RaptorQ codes is inherently linear by design [20], i.e., for a (n, k, r) RaptorQ code,
the encoding complexity is linear in n, and the decoding complexity is linear in k, which makes the
use of a RaptorQ code with large values of (n, k, r) practical.
26
8
IMPLEMENTATION CONSIDERATIONS
Number of EFIs used to decode
Fraction of failed decodings
8.6
1000
4.9 · 10−3
1001
2.4 · 10−5
RaptorQ hardware performance
1002
1.3 · 10−7
1003
1.1 · 10−9
1004
0
Table I: Decode failures in 1011 decodings of a (1200, 1000, 200) RaptorQ code
Software implementations of RaptorQ codes have been deployed in a number of applications,
including real-time applications that require low encoding and decoding complexity. Encode and
decode speeds of over 10 Gbps are achieved using a software implementation of RaptorQ running
on one core of a 20 core Intel Xeon CPU ES-2680 server running at 2.8 GHz, with code parameters (n, k, r) = (1500, 1000, 500) and 256 byte symbols. Aggregate encode and decode speeds of
around 150 Gbps are achieved using all 20 cores of the same server, with the same code parameters
and symbol size.
These numbers suggest that RaptorQ decoding would add a fractional extra cost to the data processing pipeline: It is expected that other transformations are performed when accessing user data,
such as for example checksumming, decompression and decryption. Of those MD5 checksums can
be computed on the same machine at a rate of at most about 4 Gbps. As for decompression, Google’s
run-time performance oriented Snappy codec achieves reportedly about 4 Gbps on a Core i7 processor [23], suggesting that RaptorQ coding is not a bottleneck.
8.6. RaptorQ hardware performance
Hardware based implementations have the potential to further cut down on power usage by the
RaptorQ codes. Based on a hardware RaptorQ code chip design by Qualcomm for a different application, our estimate is that the power usage in an ASIC implementation would be about 3 Watts per
100 Gbps encoding or decoding speed.
To put this into perspective, we compare this to the power use of an SSD to access data. For
example, the Samsung SM863a 2 TB enterprise SSD drive [24], achieves a data transfer rate of up
to about 4.1 Gbps and uses about 2.5 W of power when reading. This translates to a power use of
about 61 W to achieve 100 Gbps. Thus, the power needed for the RaptorQ code is about 5% of the
power needed to read the data off a drive. If other necessary aspects such as network transfer, etc,
were incorporated as well, the RaptorQ code share drops well below 5%.
8.7. RaptorQ decode reliability
Extremely high durability is required for storage systems. MDS codes have the property that an
object can be recovered from any k fragments for the object. RaptorQ codes are essentially MDS,
but not literally MDS: it is not necessarily the case that an object can be recovered from any k
fragments. We present simulation results showing the probability of object loss due to RaptorQ
codes not being MDS can be essentially ignored.
Consider a liquid system using a (1200, 1000, 200) RaptorQ code, where the target w = 50, i.e.,
the number of available fragments for any object is always at least k + w = 1050. If 1/λ = 3
years, then on average there are 400 node failures per year. We group node failures together into
consecutive batches of w − 4 = 46, and consider testing for decoding using the set of k + w EFIs
for available fragments at the beginning of the batch minus the w − 4 = 46 EFIs for fragments lost
within the batch, and thus we decode from k + w − (w − 4) = k + 4 = 1004 EFIs. Note that
if decoding is possible from this set of 1004 EFIs then decoding is possible for all objects at each
point during the batch.
As shown in Table I, there were no RaptorQ decoding failures in an experiment where 1011
random sets of 1004 EFIs were tested for decodability. Since there are on average 400 node failures
per year and the batch size is 46, there will be on average 400
46 < 10 such batches per year, and
thus these 1011 decoding tests cover 10 billion years of simulation time. Thus there would be no
decoding failures for RaptorQ within 10 billion years, which is a thousand times a target MTTDL
of 10 million years.
27
9
READ REPAIR RATE REGULATION
9. READ REPAIR RATE REGULATION
The simple Poisson process model of node failures discussed in Section 3.1 is useful to benchmark the reliability of a distributed storage system. However, node failures in real systems are not
necessarily faithful to a simple statistical model.
Detailed statistical assumptions of the node failure process do not significantly affect reactive
repair decisions for a small code system, i.e., the actions taken by the reactive repair process are
largely event driven and immediate. For example, repairs are executed as soon as fragments are lost
due to node failures, with no dependency on assumptions on future node failures. The anticipated
node failure rate is only germane to the initial provisioning of repair resources (e.g., total repair
bandwidth).
This is in sharp contrast to lazy repair for a liquid system, where repairs are postponed to achieve
greater repair efficiency. The degree to which repairs can be postponed is clearly dependent on
some expectation of future node failures. The tradeoff between repair efficiency and MTTDL rests
on assumptions about the long term node failure rate. In the liquid system with fixed read repair rate
this dependence appears prominently in the choice of the read repair rate. Lower read repair rate
results in greater repair efficiency but higher probability of object loss. This is clearly evident in the
(tight) MTTDL estimate for the fixed rate system derived in Appendix A.
For example, consider a liquid system using a (402, 268, 134) large code. If the node failure rate
λ is 1/3 per year and the read repair rate is set so that λ · T = 0.21 then the achieved MTTDL
(estimate) is 3.6 · 109 years. If instead λ is 10% higher then the MTTDL is 1.0 · 107 years, and if
λ is 20% higher then the MTTDL is 8.7 · 104 years. Thus, the MTTDL of a fixed read repair rate
design is sensitive to error in knowledge of λ. If there is uncertainty in the node failure rate then
either some repair efficiency must be sacrificed or a reduction in MTTDL must be tolerated for a
fixed read repair rate design.
Fortunately, liquid systems can be fitted with an adjustment algorithm that smoothly and automatically adjusts the read repair rate in response to node failure conditions so that accurate a priori
knowledge of node failure rates is not required. Suppose for example that the node failure rate increases at some point in time, possibly due to component quality issues. This will result in a relative
increase of the number of missing fragments among objects in the repair queue. Moreover, over
time the increase in node failure rate can be detected by a node failure rate estimator. Both of these
occurrences can signal to the lazy repair process a need to increase the read repair rate. Properly
designed, a regulated read repair rate process can achieve significantly larger MTTDLthan a fixed
read repair rate process running at the same average read repair rate, at the cost of some variation in
read repair rate.
Even in the case where the node failure process is a Poisson process of known rate, regulation of
the read repair rate can bring significant benefit. To develop a good regulation method it is helpful to
consider in more detail the behavior of a fixed read repair rate process. We assume a liquid system
with n nodes and a node failure rate λ. We assume a fixed read repair rate that we characterize using
T, the time required to repair all objects in the system. We work in the large number of objects limit
where the size of a single object is a negligible fraction of all source data. Objects are repaired in
a fixed periodic sequence. We view all objects as residing in a repair queue and we use x to denote
the relative position in the queue. Thus, an object at position x = 1 is under repair and the object in
position x = 0 has just been repaired. An object in position x will be repaired after a time (1−x)·T
and was repaired a time x · T in the past.
For an object currently in position x let f (x) · n denote the number of its fragments that have
been erased due to node failures.2 Assume 0 ≤ x ≤ y ≤ 1 and 0 ≤ s ≤ t ≤ 1, and that the read
repair rate is characterized by T. The number of erased fragments f (y) · n the object will have when
2 Note
here that while f (x) · n is integer valued we generally consider f (x) as arbitrarily valued on [0, 1]. We will make
frequent use of this slight abuse of notation.
28
9
READ REPAIR RATE REGULATION
it reaches position y is given by the transition probability
Pr [f (y) · n = t · n | f (x) · n = s · n] = B (1 − t) · n, (t − s) · n, e−λ·T ·(y−x)
(1)
where B(n, m, q) is the probability mass at m of a binomially distributed random variable with
parameters (n, q). That is,
n
B(n, m, q) =
· q n−m · q̄ m
m
where we have introduced the notation q̄ = 1 − q. Applying this to the particular case with initial
condition f (0) · n = 0 we have Pr [f (x) · n = m] = B(n, m, e−λ·T ·x ) . Using Stirling’s approximation ln(n!) = n · ln(n) + O(ln(n)) we can write ln( Pr [f (x) · n = f ]) = −n · E(f, λ · T · x) +
O(ln(n)) where
E(f, λ · T · x) = −H(f ) + f¯ · λ · T · x − f · ln(1 − e−λ·T ·x )
where H(f ) = −f · log2 (f ) − f¯ · log2 (f¯) is the entropy function. As a function of f, the exponent
E(f, λ · T · x) is minimized at f = 1 − e−λ·T ·x , which is also the expected value of f (x), being the
solution to the differential equation
d(1 − f (x))
= −λ · T · (1 − f (x))
dx
which governs the expected value. Thus, in the large system limit the fraction of erased fragments
as function of repair queue position concentrates around the function 1 − e−λ·T ·x . In particular we
identify n · (1 − e−λ·T ) as the number of fragments that are typically repaired. We can interpret
the quantity ftar := 1 − e−λ·T as a target fraction of repair fragments. Note that ftar and λ · T are
in a one-to-one relationship so, given the node failure rate λ, we can view ftar as a specification of
the read repair rate. This perspective on specification of system behavior is a convenient one for
consideration of regulated read repair rate systems.
To gain a further understanding of the fixed read repair rate system behavior it is helpful to consider typical behavior under object loss. We assume that object loss occurs when an object has more
than r = β · n fragments erased and consider the number of erased fragments as a function of queue
position for such an object. Thus, we consider the distribution Pr [f (x) · n | f (1) · n > β · n] ,
which, using Bayes rule, can be written as
Pr [f (x) · n = f · n | f (1) · n > β · n]
Pr [f (1) · n > β · n | f (x) · n = f · n] · Pr [f (x) · n = f · n)]
=
Pr [f (1) · n > β · n]
Using (1) and Stirling’s approximation as above, we find that the in the large n limit the solution
concentrates around f (1) · n = β · n, and, more generally, f (x) = fβtar · (1 − e−λ·T ·x ) which is
simply a scaled version of the nominal queue function 1 − e−λ·T ·x . Note that this solution satisfies
(x))
the equation d(1−f
= λ · fβtar · (1 − fβtar · f (x)) . Thus, for small x, where f (x) is small, the
dx
solution corresponds to a typical behavior for the system with node failure rate λ · fβtar . Another
likely contribution to typical object loss is a skewing of node failures to those nodes which, for the
given object in the given repair cycle, have not previously failed. This effect can explain the second
factor 1 − fβtar · f (x) which replaces 1 − f (x) in the equation for the expected value. Thus, we
observe that object loss typically involves a sustained increase in the observed node failure rate over
an entire repair cycle and possible skewing of the node failure selection. The fluctuation in node
failure rate can be detected by a node failure rate estimator. Both the increased node failure rate and
the node selection skewing can be detected from the atypical behavior of the repair queue. We shall
29
9
READ REPAIR RATE REGULATION
first consider a regulator that assumes a known fixed node failure rate λ and responds only to the
repair queue, and later extend the design to incorporate estimators of node failure arrival rates.
Under regulated repair we would expect that if the repair queue state is typical, i.e., the fraction
of erased fragments for an object in position x is near its nominal value of 1 − e−λ·T ·x then the read
repair rate should be kept near its nominal value. If, however, at some points in the repair queue
the number of missing fragments is larger than the nominal amount then the read repair rate should
be increased. As a general philosophy we adopt the view that each object in the repair queue takes
responsibility for itself to ensure that the probability of object loss for that object is kept small. Each
object, knowing its position in the queue and knowing its number of erased fragments, desires a
certain system read repair rate such that, if adopted, would adequately protect that object against
object loss. The repair system will apply the highest of the desired read repair rates thus satisfying
all read repair rate requests for all the objects.
The regulator design defines a function φ(f, x) such that an object in position x with a fraction f
missing fragments requests a read repair rate corresponding to a system repair time of λ−1 φ(f, x).
Recall that a read repair rate corresponds to time required to repair all object in the system T and
that λ · T is the expected number of times a single node fails during the time it takes to repair all
objects. In this form, φ(f, x) expresses a desired read repair rate in terms of the desired value for
λ · T. In general φ(f, x) will be monotonically decreasing in f and increasing in x. Under these
assumptions, as well as some assumptions on node failure rate estimators, it follows that the read
repair rate will always be determined by certain critical objects in the repair queue. The critical
objects are those that were at the tail of the queue when a node failure occurred (see Appendix A
for more detail). Note that the fixed read repair rate case corresponds to setting φ constant.
As a basis for designing the function φ(f, x) we adopt the notion that objects desire to recover
certain properties of the nominal trajectory. One quantity of particular interest is the probability of
object loss for an object. Given a position x in the queue and a fraction f (x) < β erased fragments
the probability that the object will experience object loss can be expressed using (1). For an object
on the nominal trajectory, f (x) = 1 − eλ·T ·x = 1 − (1 − ftar )x , this probability is easily seen to
be decreasing in x. Thus, if the object enjoys typical behavior early in the repair queue then its
probability of object loss decreases. As a design principle, therefore, we can stipulate that an object
in position x with a fraction f of erased fragments will request a system repair parameter λ · T, as
represented by the function φ(f, x), such that the probability of object loss for that object would be
the same as an object at that position on the nominal trajectory. Under this stipulation, we obtain
φ(f, x) implicitly as the solution to
X
X
B(f¯tarx · n, (f¯tarx − s̄) · n, f¯tar1−x ) =
B(f¯ · n, (f¯ − s̄) · n, e−φ(f,x)·(1−x) ) .
s·n>β·n
s·n>β·n
Note that in this context the quantity β ·n represents a particular threshold that the regulator attempts
avoid reaching and need not necessarily correspond with the erasure correcting limit of the used
erasure code. For that reason we introduce a different notation fT for the target hard threshold. In
practice one will typically have fT ' β.
Simpler expressions of the above principle for deriving appropriate φ functions can be obtained
in a number of ways. One possibility is note that the above sums are dominated by the term with
s · n = β · n and then, using the Stirling approximation, to equate the rate functions on both sides.
Another approach is to use a Gaussian approximation of the above binomial distributions and equate
the integrals that correspond to the above sums. Under the Gaussian approximation we obtain the
equation
¯ ¯ 2 ¯
1 − e−(1−x)·φ(f,x)
ftar
fe − fT
· ¯
(2)
=
1 − e−(1−x)·φnom
f¯tar − f¯T
fe
where φnom = − ln(1 − ftar ) and fe denotes the expected fraction of erased fragments that will
be repaired for an object in position x with f erased fragments assuming that the read repair rate
30
9
READ REPAIR RATE REGULATION
9.1
Regulated read repair rate simulation
is subsequently fixed according to φ(f, x), i.e., we have f¯e = f¯e−φ(f,x)·(1−x) . Substituting this
into the above we obtain a quadratic equation for fe whose solution then determines φ. A good
approximation is
1 − e−(1−x)·φ(f,x)
φ(f, x)
'
−(1−x)·φ
nom
φnom
1−e
under which the level curves of φ(f, x) correspond exactly to asymptotic trajectories for the regulated system that can be parameterized by their implied target fragment repair fraction fe . This
implies that an object that is controlling the repair rate expects on average to hold the repair constant until it is repaired.
For the simulation results given below we used the Gaussian approximation approach to obtain
φ. It is not hard to show that φ is bounded above (read repair rate never goes to 0). A lower bound
is 0 (infinite read repair rate) which is requested if the f (x) is sufficiently close to fT . In practice
the read repair rate will obviously be limited by available resources, and such a limit can be brought
into the definition of φ by limiting its minimum value and for simulations we typically saturate φ at
some postive value, often specified as fraction of φnom .
Recall that λ−1 φ represents desired time to repair all objects precisely once. This indicates how to
incorporate estimates for the node failure rate into the operation of the regulator: one simply replaces
λ in the expression λ−1 φ with a suitable estimate of λ. In general, an estimator of the node failure
rate will be based on past node failures . One pertinent quantity in this context is the appropriate
time scale to use to form the estimate. We may choose to vary the time scale as a function of the
number of erased fragments the object experiences, or other aspects of the node failure process.
Let us assume a system with a total of O objects of equal size. Then each object has a position in
the queue x with x · O ∈ {0, ..., O − 1}. When an object is repaired the position x of each object
is incremented by 1/O. According to the regulated read repair rate process the repair time for the
next object is then given by
minx φ(f (x), x)
λ·O
where f (x) denotes the fraction of erased fragments for the object in position x. Here we tacitly
assume f (x) < fT for all x but for analytical purposes we may extend the definition of φ to all f
and x in [0, 1]. In practice φ is effectively bounded from below by the repair speed limitations of
the system. In our simulations we typically assume a floor for φ of γ · φnom with γ set to a small
fraction, e.g. γ = 13 . Note further that we always have φ(0, 0) = φnom = − ln(1 − ftar ) and so this
gives a lower bound on the read repair rate and we can impose a ceiling on the function φ of value
φnom without altering system behavior.
The choice of ftar involves a tradeoff: lower values of ftar lead to higher read repair rates, but also
a more reliable system, and one that is potentially less prone to large read repair rate variations,
whereas a larger value of ftar gives higher repair efficiency at the price of larger read repair rate
fluctuations and a reduced object loss safety margin of the system. For simplicity, all the results
that we present in this section with the repair adjustment algorithm were obtained by running the
algorithm with ftar = 2·r
3 , which is a reasonable choice in all but some contrived circumstances.
For certain forms of failure rate estimators it is feasible to compute lower bounds on MTTDL.
The details of the models and the calculations can be found in Appendix A. The bounds can be used
to designing system parameters such as fT and more generally φ so as to maximize repair efficiency
while ensuring a desired MTTDL.
9.1. Regulated read repair rate simulation
Fig. 17 illustrates how the regulated read repair rate automatically varies over time when the node
failure rate is 1/3 years for a liquid system using a (402, 268, 134) large code with storage overhead
33.3%, and the first row of Table II shows a comparison between the regulated repair rate with
31
9.1
Regulated read repair rate simulation
9
READ REPAIR RATE REGULATION
250
Regulated Repair Rate (Gbps)
200
150
100
50
0
0
5
10
15
20
25
30
35
40
45
Time (years)
Fig. 17: Repair regulator for a (402, 268, 134) liquid system with fixed 1/λ = 3 years.
an average of Ravg = 107 Gbps and a fixed read repair rate of Rpeak = 104 Gbps (set so that the
MTTDL is ≈ 107 years). In this table, the MTTDL of 109 for the regulated read repair rate
indicates there were no object losses in 109 years of simulation (based on estimates using the results
from Appendix A, the MTTDL is orders of magnitude larger than 109 years), and Rpeak is a hard
upper limit on the used read repair rate. The variability of the regulated read repair rate is moderate,
e.g., 99.99% of the time the read repair rate is less than 231 Gbps. The second row of Table II shows
a comparison using a (402, 335, 67) large code with storage overhead 16.7%.
n
402
402
System parameters
r
β
S
134 33.3% 1 PB
67 16.7% 1 PB
Ravg
106 Gbps
298 Gbps
Regulated read repair rate
R99%
R99.99%
Rpeak
154 Gbps 226 Gbps 311 Gbps
513 Gbps 975 Gbps 1183 Gbps
MTTDL
109 years
109 years
Fixed read repair rate
Rpeak
MTTDL
104 Gbps 3.2 · 107 years
394 Gbps 9.0 · 107 years
Table II: Comparison of regulated read repair rate versus fixed read repair rate.
Fig. 18 shows the behavior of regulated read repair rate when the node failure rate is varied as
described in Section 5.3. The upper portion of Fig. 18 displays the average node failure rate varying
over time, and the lower portion of Fig. 18 displays the corresponding read repair rate chosen by the
regulator. This is a stress test of the regulator, which appropriately adjusts the read repair rate to the
quickly changing node failure rate to avoid object losses while at the same time using an appropriate
amount of bandwidth for the current node failure rate. There were no object losses in 109 years of
simulation using the regulator running in the conditions depicted in Fig. 18, and based on estimates
using the results from Appendix A the MTTDL is orders of magnitude larger than 109 years. The
results from Appendix A show that the regulator is effective at achieving a large MTTDL when the
node failure rate varies even more drastically.
A repair adjustment algorithm could additionally incorporate more advanced aspects. For example, since the bandwidth consumed by the read repair rate is a shared resource, the read repair rate
32
10
CONCLUSIONS
Node Failure Rate (1/year)
1.2
1.0
0.8
0.6
0.4
0.2
0.0
Regulated Repair rate (Gbps)
800
700
600
500
400
300
200
100
0
0
5
10
15
20
25
30
35
40
45
50
55
Time (years)
Fig. 18: Regulated repair for a (402, 268, 134) liquid system with node failure rate varying by a
factor of three
may also be adjusted according to other bandwidth usage, e.g., the read repair rate can automatically
adjust to higher values when object access or storage rates are low, and conversely adjust to lower
values when object access or storage rates are high. Furthermore, the read repair rate could be regulated to take into account known future events, such as scheduled upgrades or node replacements.
10. CONCLUSIONS
We introduce a new comprehensive approach to distributed storage, liquid systems, which enable
flexible and essentially optimal combinations of storage reliability, storage overhead, repair bandwidth usage, and access performance. The key ingredients of a liquid system are a low complexity large code, a flow storage organization and a lazy repair strategy. The repair regulator design
we present provides further robustness to liquid systems against varying and/or unexpected node
failure. Our repair and access simulations establish that a liquid system significantly exceeds performance of small code systems in all dimensions, and allows superior operating and performance
trade-offs based on specific storage deployment requirements.
We address the practical aspects to take into consideration while implementing a liquid system
and provide an example architecture. A liquid system eliminates network and compute hot-spots
and the need for urgent repair of failed infrastructure.
While not detailed in the paper, we believe a liquid system provides superior geo-distribution
flexibility and is applicable to all kinds of Object, File, Scale-out, Hyper-converged, and HDD/SSD
architectures/use-cases. There are also optimizations that the team has worked on that involve pregeneration of EFIs and result in better storage efficiencies at the storage nodes.
Acknowledgment
Rick Dynarski, Jim Kelleman, and Todd Mizenko developed the prototype software for the liquid system described in Section 7, which provides a basic validation of the overall liquid system
33
A
MTTDL ANALYSIS OF LIQUID SYSTEM
properties presented herein. This prototype software is used to cross validate the simulation results
presented in Section 5.
Christian Foisy helped to produce the RaptorQ performance results presented in Section 8.5, and
also helped cross verify the prototype software for the liquid system. Menucher Menuchehry and
Valerie Wentworth provided high level comments that greatly improved the presentation.
We would like to thank Rick, Jim, Todd, Christian, Menucher, and Valerie for their valuable
contributions.
A. MTTDL ANALYSIS OF LIQUID SYSTEM
In this section we analyze regulated repair of the liquid system and derive some bounds on MTTDL.
In the case of constant repair rate the bound is particularly tight. Although we typically assume that
the node failure process is a time invariant Poisson process the analysis can be applied to cases
where the rate is not known and time varying. We assume throughout that the number of nodes M
is the same as the code length n, and we use n to denote this number.
The framework for the analysis ignores the possibility that an object may be unrepairable. In effect
we suppose that as each object leaves the repair queue (because it reaches the head of the queue
and, normally, is repaired) a new object enters at the tail. This notion is an analytical convenience
that takes advantage of the relative ease of determining or estimating the asymptotic steady state
distributions of this system. We assume a large number of objects O, and for some of the analysis
we ignore effects of a finite number of objects and consider the infinite object limit. This has no
significant effect on the results but simplifies the analysis considerably.
For the analysis we generally assume that the repair rate is updated upon object repair. In the
limit of infinitely many objects this amounts to continuous update. With a finite number of objects
the queue state, as sampled at object repair times, the system forms a Markov chain with a unique
invariant distribution. In the case where the node failure rate is known it is a finite state chain. Of
particular interest is the marginal distribution of the number of erased fragments for the object at
the head of the queue. In the case of fixed repair rate we can express this distribution in closed form,
and for the regulated case we will derive a bound on the cumulative distribution.
A.1. MTTDL Estimates
We define the MTTDL as the expected time given a perfect initial queue state (all objects have
zero erased fragments) until an object achieves r + 1 missing fragments prior to completing repair.
For the system with node failure rate estimation, we also need to specify the initialization of the
estimator. The system is typically designed to achieve a very large MTTDL. We will show in the
constant repair rate case that we can obtain a sharp closed-form lower bound on the MTTDL. The
bound can be applied even in the case of a mismatch between an assumed node failure rate and an
actual node failure rate. In the regulated repair case the analysis is more difficult, but we obtain an
efficiently computable generalization of the lower bound for the fixed node failure arrival rate case.
We further extend the analysis to cases where the node failure rate is estimated based on the node
failure realization.
Between node failure events the system evolves in a purely deterministic fashion; objects simply
advance in the repair queue according to the repair rate policy. Under a time invariant Poisson model
the node failure interarrival times are independent exponentially distributed random variables and
the selection of the failing node is uniformly random over all of the nodes. Thus, the repair can be
viewed as a random process consisting of a sequence of samples of the queue state taken upon node
failure. The process is generally Markov where the state space consists of a pair of queue states,
that just prior to the node failure and that after, and, when there is failure rate estimation, a state
representation of the failure rate estimator. The transition to the next state has two components.
First, there is the time until the next node failure, which is an exponential random variable. Given
that time the queue state advances deterministically from the state after the current node failure to
the state just prior to the next. The second component is the selection of the failing node. That choice
determines which objects in the queue experience additional fragment loss and which do not. The
34
A
MTTDL ANALYSIS OF LIQUID SYSTEM
A.1
MTTDL Estimates
update of the state of the failure rate estimators can also occur at this point. Because of the nested
nature of the process there will be a point x in the queue such that objects in positions less than
x experience fragment loss and those more than x do not, because they were already missing the
fragment associated to the failing node. Note that the queue remains monotonic: the set of erased
fragments for an object in position x is a subset of the erased fragments for an object in position y
whenever x ≤ y. In the case of node failure rate estimation, those estimates will be updated upon
node failure.
A.1.1. The Constant Repair Rate Case. When the repair rate is held constant the deterministic
update portion of the process is particularly simple. Let q denote the distribution of the number of
erased fragments for the object at the head of the repair queue immediately prior to node failure.
Because node failure is a time invariant Poisson process the distribution q is also the distribution
of the number of erased fragments for the object at the head of the repair queue at an arbitrary
time. In other words, q is the the distribution of the number of erased fragments upon repair under
continuous (steady state) operation. We therefore have
q(s) = B(n, s, e−λT ).
Let F (t) denote the number of erased fragments for the object at the head of the repair queue at
time t. If t is a node failure time let F (t−) and F (t+) denote value of F (t) immediately prior
to and after the node failure. The probability of simultaneous node failures is 0 so we assume
F (t−) ≤ F (t+) ≤ F (t−) + 1. Let t be a node failure time, then we have
n−r
Pr [F (t+) = r + 1 ∧ F (t−) = r] =
q(r) .
n
Note that this event F (t+) = r + 1 ∧ F (t−) = r corresponds to a data loss event and that the
first time this occurs, starting from the perfect queue state, is the first data loss event. We will
refer to this event as a data loss transition. Thus, the MTTDL is by definition the expected time
to the first data loss transition starting from a perfect queue state. Applying Kac’s theorem to the
−1
is the expected number of node failure events between data
Markov chain we have ( n−r
n · q(r))
loss transitions under steady state operation. The expected time between such events is given by
(λ·(n−r)·q(r))−1 since λ·n is the rate of node failure. To obtain a rigorous bound on MTTDL some
further analysis is needed because of differing assumptions on the initial condition. The rigorous
correction is negligible in regimes of interest so we use the estimate
1
. MTTDL
λ · (n − r) · q(r)
(3)
as a basis for choosing repair rates (T ) as a function of MTTDL.
Starting at time 0 let TS denote the expected value of the first data loss transition time where
the initial queue state is distributed according to the the
P steady state distribution conditioned on
F (0) ≤ r. Let us also introduce the notation q(> r) = s>r q(s).
L EMMA A.1. Assume a constant repair rate with system repair time T. Then we have the
following bound
1
T
−
≤ MTTDL
λ · (n − r) · q(r) 1 − q(> r)
P ROOF. We prove the Lemma by establishing the following two inequalities
1
T + TS
1 − q(> r)
1
1
≤
T + TS .
λ · (n − r) · q(r)
1 − q(> r)
TS ≤ MTTDL ≤
35
A.2
Regulated Repair Rate
A
MTTDL ANALYSIS OF LIQUID SYSTEM
Consider a fixed node failure process realization. Among all initial queue conditions with F (0) ≤
r the perfect queue will have a maximal time to data loss. Thus, TS ≤ MTTDL.
Assume the system is initialized at time 0 with an arbitrary initial queue state and let us consider
the expected time until the next data loss transition. Note that k ·T is time when the object initially at
the tail of the queue is repaired for the k-th time. Let k ∗ denote the smallest k > 0 where upon repair
that object has at most r erased fragments. Then the queue state at time t = k ∗ · T is the steady state
distribution conditioned on the object at the head of the queue having at most r erased fragments,
i.e., conditioned on F (t) ≤ r. Hence, the expected time to the next data loss transition can be upper
bounded by P
E(k ∗ · T ) + TS . Now, the probability that k ∗ = k is (1 − q(> r)) · q(> r)k−1 hence
T
T
∗
E(k ·T ) ≤ k>1 (1−q(> r))·q(> r)k−1 ·k·T = 1−q(>r)
. It follows that MTTDL ≤ 1−q(>r)
+TS
and that the expected time between data loss transitions, which is (λ · (n − r) · q(r))−1 , is also upper
T
+ TS .
bounded by 1−q(>r)
A.2. Regulated Repair Rate
For the regulated repair rate case the memory introduced into the system by the selection of the
maximum requested repair rate makes it quite difficult to analyze. In particular the rate of advancement of the repair queue depends on the entire state of the queue, so there is no convenient way to
reduce the size of the Markov chain state space, as was effectively done in the fixed repair rate case.
The addition of node failure rate estimation further complicates the problem. It turns out, however,
that by approximating queue advancement with a greedy form of operation we can obtain a related
lower bound on the MTTDL. When applied to the constant repair rate case the bound is smaller than
the one above by a factor of (1 − ftar ).
Although we shall consider the case of known node failure arrival rate, much of the analysis is
common with the case including node failure estimation. Thus, we will first develop some general
results that admit node failure rate estimation.
We introduce the notation FDis (s, t) to denote the number of distinct node failures in the interval
(s, t). As before, if s is a node failure time we use s− and s+ to denote times infinitisemally before
or after the node failure respectively. For convenience we assume that φ(f, x) is bounded below by
a positive constant, is defined for all f and x and is non-increasing in f and non-decreasing in x.
For computation we may also assume that φ(f, x) is bounded above by the nominal value λ · T so
that, in the known node failure rate case, T is an upper bound on the time between repairs for a
given object. Recall that since λ · T = φ(0, 0) we always have inf x∈[0,1] φ(f (x), x) ≤ λ · T so this
restriction does not affect the operation of the system. In the case with node failure rate estimation
we will typically assume some minimum postive value for the estimated node failure rate so that
there will be some finite time T during which a complete system repair is guaranteed.
When we admit failure rate estimation we suppose that each object has it’s own failure rate
estimate. That estimate λ̂ figures in the requested repair rate through the implied requested system
repair time of λ̂−1 · φ(f, x). For an object that entered the queue at time s we write its node failure
rate estimate at time t as λ̂(s, t). The estimate is allowed to depend on past node failures. It can
depend on s only through the partition of node failures into ‘past’ and ‘future’ implied by time s.
Consequently, all objects that enter the queue between the same pair of successive node failures use
the same value for λ̂. In other words, as a function of s, the estimate λ̂(s, t) is piecewise constant in
s with discontinuity points only at node failure times. We remark that this can be relaxed to having
λ̂(s, t) non-decreasing in s between node failure points. In practice the node failure rate estimate
used will generally be a function of relatively recent node failure history which can be locally stored.
Note that the estimator for an object may depend on its number of erased fragments.
We call an object in the repair queue a critical object if it was at the tail of the queue at a node
failure point. Letting s(x, t) denote the queue entry time of the object in position x of the queue at
time t, an object is critical if s(x, t) is a node failure time. To be more precise, a critical object is
36
A
MTTDL ANALYSIS OF LIQUID SYSTEM
A.2
Regulated Repair Rate
an object that was the last object completely repaired prior to a node failure. Such an object has an
erased fragment while it is still at the tail of the queue.
Let χ(t) denote the set of critical object queue locations x ∈ [0, 1] at time t. Let f (x, t) =
FDis (s(x, t), t)/n and let λ̂x (t) denote λ̂(s(x, t), t). Under the current assumptions the repair regulator need only track the critical objects in the system.
L EMMA A.2. The repair rate is always determined by a critical object, i.e.,
−1
inf λ̂−1
x (t) · φ(f (x, t), x) = min λ̂x (t) · φ(f (x, t), x) .
x∈[0,1]
x∈χ(t)
P ROOF. Of all objects that entered the queue between the same pair of successive node failures,
the critical object has the minimal value of x and all such objects have the same value of λ̂ and f.
Since φ(f, x) is non-decreasing in x the result follows.
To obtain a bound on the MTTDL we will adopt a greedy approximation in which we effectively
assume that objects control the repair rate during their time in the repair queue. To this end, for any
time s set y(s, s) = 0 and for t ≥ s let us define y(s, t) as the solution to
d
y(s, t) = λ̂(s, t) · (φ(FDis (s, t), y(s, t)))−1
dt
which is well defined as long as y(s, t) ≤ 1. Further, let τ (s) be the minimal t such that y(s, t) = 1.
An interpretation of y is that it is the position in the queue of an object that was at the tail of the
queue at time s and is designated to be the object that determines the system repair rate as long as it
remains in the queue. Let x(s, t) denote the actual position of the object at time t given that it is at
the tail of the queue at time s.
L EMMA A.3. For any s we have x(s, t) ≥ y(s, t).
P ROOF. For each fixed f the function φ(f, z) is non-increasing in z ∈ [0, 1]. Thus if 0 ≤ x ≤
y ≤ 1 then (φ(f, x))−1 − (φ(f, y))−1 ≥ 0 . Let us fix s and assume 0 ≤ y(s, t), x(s, t) ≤ 1. Then
we have
−1
−1
d
(x(s, t) − y(s, t)) = inf λ̂−1
− λ̂−1 (s, t) · φ(FDis (s, t), y(s, t))
z (t) · φ(f (z, t), z)
dt
z∈[0,1)
−1
−1
− φ(FDis (s, t), y(s, t))
≥ λ̂−1 (s, t) φ(FDis (s, t), x(s, t))
d
Thus, x(s, t) − y(s, t) < 0 implies dt
(x(s, t) − y(s, t)) ≥ 0 and since x(s, s) = y(s, s) = 0 we see
that x(t, s) − y(t, s) < 0 cannot occur by the mean value theorem.
For data loss to occur there must be node failure times s ≤ t such that both FDis (s−, t+) ≥ r + 1
and x(s−, t) < 1.3 In this event data loss has occured to the critical object that was at the tail of the
queue immediately prior to time s. By Lemma A.3 this implies that FDis (s−, τ (s)) ≥ r + 1. The
condition FDis (s−, τ (s)) ≥ r + 1 denotes a greedy data loss where the object arriving at the tail of
the queue at time s greedily controls the repair rate and suffers data loss. We refer to the node failure
at time s as a greedy data loss node failure. Note that the greedy data loss actually occurs after time
s.
A.2.1. The Regulated Repair Case with Fixed Failure Rate. We now assume that the estimated
repair rate λ̂ is constant and may or may not be equal to the actual node failure rate λ. In this case
there is no ambiguity in the intialization of λ̂ and the MTTDL can be lower bounded by the expected
value of s where s is the time of the first greedy data loss node failure.
3 Here
we may interpret x(s−, t) as the limit of x(s0 , t) as s0 approaches s from below.
37
A.2
Regulated Repair Rate
A
MTTDL ANALYSIS OF LIQUID SYSTEM
Assume s is a node failure time and let q now denote the distribution of FDis (s−, τ (s−)). By
Kac’s theorem, 1/q(> r) is the expected number of node failures between greedy data loss node
failures. Hence 1/(λ · n · q(> r)) is the expected time between greedy data loss node failures.
Let TG denote the expected time time until the first greedy data loss node failure assuming an
unconditioned future node failure process. Note that the time to a greedy data loss is independent of
the initial queue state. It follows that 1/(λ · n · q(> r)) is upper bounded by T + TG . This is because
if we condition on 0 being a greedy data loss node failure time, then the node failure process after
time T is independent of this condition. Recalling that TG ≤ MTTDL we obtain
1
− T ≤ MTTDL
λ · n · q(> r)
In regimes of interest we will have q(> r) ' q(r + 1) and assuming a fixed repair rate we have
n · q(r + 1) = n · B(n − 1, r, e−λ·T ) = eλ·T · (n − r) · B(n, r, e−λ·T ) . If λ̂ = λ then we see that the
resulting MTTDL bound is approximately a factor e−λ·T = 1 − ftar worse than the previous one.
Given φ it is not difficult to quantize the system to compute the distribution q. To perform the
computation we consider quantizing x with spacing δx. 4 This is essentially equivalent to assuming a finite number, O = δ −1 , of objects and we describe the quantization from that perspective.
Assume that repair rates are updated upon object repair. Under regulated repair the repair system
then selects the next object repair time, which is now given by the minimum of δ · φ(f (jδ), jδ)
for j = 0, . . . , O − 1. Assume that the next repair time ∆ is given and that we are given that the
object in position k · δ has a random number of erased fragments F that is distributed according
to pk (F ). Then the distribution of the number of erased fragments for the object when it reaches
position (k + 1) · δ is given by
X
pk+1 (F ) =
pk (F 0 ) · B(n − F 0 , F − F 0 , e−λ·∆ ) .
0≤F 0 ≤F
To compute the distribution q we proceed similarly but, in effect, track the greedy controlling object
as it moves through the repair queue, assuming that that object is the one determining the repair
rate. Letting qk (F ) denote the distribution of the number of missing fragments when the object is
(greedily) in position kδ. We have
X
0
qk+1 (F ) =
qk (F 0 ) · B(n − F 0 , F − F 0 , e−δ·φ(F /n,kδ) ) .
0≤F 0 ≤F
Note that the distribution q is simply qO (·) assuming we initialize with q0 (F ) = 1F −1 (the distribution with unit mass at F = 1.) Using the above expression we see that the distribution q can be
computed with complexity that scales as n · O.
Note that if the above computation we instead use the initial condition 1F then the resulting
distribution q provides an upper bound on the steady state distribution of the number of erased
fragments upon repair. More precisely, the resulting q(> s) upper bounds the probability that more
than
Pn s fragments of an object are erased when that object is repaired. Using this value of q we obtain
s=0 q(> s) as an upper bound on the expected number of repaired fragments, which represents
the repair efficiency of the system.
In Fig. 19 we plot the computed cumulative distribution of the number of missing fragments
upon repair and the computed upper bound. The five shown bounds are different only in the limit
placed on the maximum repair rate which is indicated as a factor above the nominal rate. The
“1x” case corresponds to the fixed repair rate policy. Already by allowing a “2x” limit a dramatic
improvement is observed. It is observed that the greedy bound tracks the true distribution, as found
through simulation for the “3x” limited case, with some shift in the number of missing fragments.
4 An
analog of Lemma A.3 can be obtained for the quantized system with the addition of a small correction term, but the
impact on the results is negligible for reasonable size systems.
38
A
MTTDL ANALYSIS OF LIQUID SYSTEM
A.2
Regulated Repair Rate
1.0E+00
1.0E-01
Computed bound 1x
1.0E-02
1.0E-03
Computed bound 2x
Cumulative probability p (missing >= f)
1.0E-04
Computed bound 3x
1.0E-05
1.0E-06
Computed bound 5x
1.0E-07
Computed bound 100x
1.0E-08
1.0E-09
Simulation 3x
1.0E-10
1.0E-11
1.0E-12
1.0E-13
1.0E-14
1.0E-15
1.0E-16
1.0E-17
1.0E-18
1.0E-19
1.0E-20
80
90
100
110
120
130
140
Number of missing fragments f
Fig. 19: Comparison of simulated and computed results for regulation with known node failure
arrival rate. Different curves place different limits on the peak repair rate, indicated as factors above
the nominal rate
Although somewhat conservative, it is evident that the computed bound can provide an effective
design tool and that rate regulation can provide dramatic improvements in MTTDL.
Fig. 20 shows how the cumulative distribution of the number of missing fragments changes with
node failure rate mismatch. We set the repair rate limit to 10 times the nominal rate in increased the
node failure rate by factors of 1.25, 1.5, and 2.0 respectively. The regulator responds by increasing
the average repair rate and reaching a new asymptotic queue state equilibrium with respectively
larger repair targets. This is indicated in the shifts of the cumulative distribution. For very large shifts
in failure rate this level of adaptation would be inadequate to ensure data durability. Fortunately node
failure rate estimation is easily incorporated into the regulator design, leading to a significantly more
robust system.
A.2.2. The Regulated Repair Case with Failure Rate Estimation. The rate regulator as described
above is designed to concentrate the repair queue near its nominal behavior when the actual node
failure rate matches the design’s assumed failure rate. As demonstrated, the regulator can compensate for some variation in the failure rate. If the actual node failure rate is persistently different from
the design assumed rate then the regulator typical behavior is shifted relative to the nominal one. By
using an estimator of the node failure rate the system can bring the typical behavior back toward the
nominal one. Moreover, by using an estimate of the node failure rate the system can adapt to much
larger variations in the actual node failure rate.
39
A.2
Regulated Repair Rate
A
MTTDL ANALYSIS OF LIQUID SYSTEM
1.0E+00
1.0E-01
Bound 10x, Rate factor 2.0
Cumulative probability p (missing >= f)
1.0E-02
Bound 10x, Rate factor 1.5
1.0E-03
Bound 10x, Rate factor 1.25
1.0E-04
Bound 10x, Rate factor 1.0
1.0E-05
1.0E-06
1.0E-07
1.0E-08
1.0E-09
1.0E-10
1.0E-11
1.0E-12
1.0E-13
1.0E-14
80
90
100
110
120
130
140
Number of missing fragments f
Fig. 20: Cumulative distribution bounds with mismatch between assumed design node failure rate
and actual node failure rate.
Many forms of estimators may be considered. We shall consider some relatively simple forms
that are amenable to computation analysis similar to the fixed arrival rate case. These estimates are
based on first order filtering of some function of the node failure interarrival times. A typical update
of the node failure interarrival time will take the form
ξ(T̂ ) ← α · ξ(T̂ ) + ᾱ · ξ(T ) .
where T is a most recent node failure interarrival time and the functional ξ is an invertible function.
A convenient case to consider is ξ(T ) = T but other potentially interesting forms are ξ(T ) = log T
or ξ(T ) = 1/T, or regularized versions of these functions such as ξ(T ) = log(1+T ). For simplicity
we will focus on the case ξ(T ) = T. The node failure rate estimate λ̂ is simply given as T̂ −1 .
Corresponding to our assumptions on failure rate estimators, we assume that the value of α used
by an object is a function only of the number of erased fragments for that object and the total number
of node failures that have occurred since that object entered the repair queue, for which we introduce
the notation FAll (s, t). Let us now consider how to compute the distribution of the greedy behavior.
In this case we take a slightly different approach then previously. We aim to recursively compute
the joint greedy queue position and failure rate estimate distribution at node failure times. Since a
node failure may or may not result in a fragment erasure for the greedy object, we must also jointly
consider the distribution of the number of erased fragments.
40
A
MTTDL ANALYSIS OF LIQUID SYSTEM
A.2
Regulated Repair Rate
Let is fix s and consider F (t) = FDis (s, t), F + (t) = FAll (s, t), λ̂(s, t), y(s, t), for t ≤ τ (s) and
extend these definitions for t ≥ τ (s). Specifically, for t ≥ τ (s) set y(s, t) = 1, set F (t) = F (τ (s)),
set λ̂(s, t) = λ̂(s, τ (s)). We will leave F + (t) = FAll (s, t) unchanged, i.e. F + (t) continues to count
node failures. In short we track F (t), y(s, t), and λ̂(s, t) until y(s, t) = 1 which then marks an
absorbing state.
We will now briefly discuss how, give F + , we can computationally update a joint distribution
of F, y, λ̂ to the next node failure point. Assume that at time t0 the greedy object has an estimate
λ̂(s, t0 ) and greedy queue position y = y(s, t0 ) and a certain value of F. Then the joint distribution of y and λ̂(s, t) at the next node failure point lies on a curve parameterized by t1 , the next
node failure time. Note that since we assume that no additional node failures occur, we have that
λ̂(s, t), F (t), and F + (t) are all constant and we can drop the dependence on s and t. Note also that
α is detemined and is independent of the next node failure time. For t ≥ t0 such that y(s, t) < 1 we
have
Z t
y(s, t) = y(s, t0 ) + λ̂
φ−1 (F/n, u)du .
t0
Assuming t = t1 is the next node failure the update of λ̂ is given by
T̂ (s, t) = α · T̂ (s, t0 ) + ᾱ(t − t0 )
and we obtain the curve y(s, t1 ), T̂ (s, t1 ) for y(s, t1 ) < 1. Note, however, that at y(s, t1 ) = 1 we
actually have the discontinuity T̂ (s, t1 ) = T̂ (s, t0 ) since in this case the object reached the head of
the queue prior to additional node failures. To obtain the probability density on this curve observe
that until the time of the next node failure t1 we have
Z y(s,t)
t − t0 = λ̂−1
φ(F/n, u)du .
y(s,t0 )
The distribution of t1 − t0 is exponential with rate λ · n so the probability that t − t0 is greater than
0
a quantity t0 > 0 is given by e−λ·n·t . Since y(s, t) is monotonic in t, the probability that y(s, t1 ) is
greater than z where y(s, t0 ) ≤ z < 1 is given by
Pr [y(s, t1 ) > z] = e
−λ·n·λ̂−1
Rz
y(s,t0 )
φ(F/n,u)du
and this gives the distribution along the curve y(s, t1 ), T̂ (s, t1 ).
Let us introduce the notation QF + (F, y, λ̂) to denote the conditional density of the extended
definitions of F, y, λ̂ conditioned on F + . Using the above described calculations we can update this
distribution to the distribution at the next node failure point. We will use the notation Q̃F + (y, λ̂)
to denote this updated distribution. Assuming y < 1, the next node failure is incrementing with
probability (1 − F/n) thus for y < 1 we have
QF + (F, y, λ̂) = (F/n)Q̃F + −1 (F, y, λ̂) + (1 − (F − 1)/n)Q̃F + −1 (F − 1, y, λ̂) .
and for y = 1 we have
QF + (F, y, λ̂) = Q̃F + −1 (F, y, λ̂) .
With suitable quantization of y and λ̂ the above procedure can be used to compute the greedy
distribution.
To obtain the greedy bound for an arbitrary initial time we initialize the distrbution Q0 so that
the marginal distribution of y is 1y (and, necessarily, F = 0 with probability 1.) Given a Poisson
node failure model and a particular estimator the distribution λ̂(s, s) may be determined. In general
41
REFERENCES
REFERENCES
the distribution of λ̂ can be set according to the desired assumptions of the computation. The limit
as F + tends to +∞ of QF + (F, 1, λ̂) is the greedy approximation to the distribution upon repair of
F and λ̂. The probability of large F + with y < 1 decays exponentially in F + , so this computation
converges quickly in practice.
To obtain a bound on the MTTDL we need to perform the corresponding computation for critical
objects. This requires only one additional initialization step in which the distribution Q1 (F, y, λ̂) is
modified to project out y and put all probability mass on y = 0. Note that when F + = 1 we have
F = 1 with probability 1. With such a modification of the initialization process we can obtain a
lower bound on the MTTDL as before. We remark, however, that in this case we take the definition
of MTTDL to include a random initialization of λ̂ according to the distribution of λ̂(s, s) and the
above computation should also use this initial condition.
Fig. 21 shows the comulative distribution of the number of missing fragments upon repair q as
obtained from simulation and the computed bound. The simulation objects used for an estimate
of node failure rate the average of the node failure interrarrival times for the last 155 − F node
failures, where F denotes the number of erased fragments in a object. Note that the value of 155
was chosen slightly larger than r +1 = 135. The computed bound used a slightly different estimator
consistent with the above analysis. The estimator used is a first order filter but, to best model the
corresponding simulation, the time constant of the filter was adjusted as a function of F to match
155 − F. In both cases the peak repair rate was limited to 3 times the nominal rate as determined by
the actual node failure rate. It can be observed from these the plots that even compared with know
node failure arrival rate, the estimated case can give larger MTTDL. This is intuitive since data loss
is generally associated to increases in node failure rate and the increase in the estimated node failure
rate case gives rise to faster repair. In addition, as shown previously, a regulator with rate estimation
can accomodate large swings in node failure rates.
REFERENCES
[1] M. Belshe, R. Peon, M. Thomson. IETF RFC7540. Hypertext Transfer Protocol Version 2 (HTTP/2). Internet Engineering Task Force, May 2015.
[2] J. Bloemer, M. Kalfane, M. Karpinski, R. Karp, M. Luby, D. Zuckerman. An XOR-Based Erasure-Resilient Coding
Scheme. ICSI Technical Report, No. TR-95-048, August 1995.
[3] B. Calder, J. Wang, A. Ogus, N. Nilakantan, A. Skjolsvold, S. McKelvie, Y. Xu, S. Srivastav, J. Wu, H. Simitci, J. Haridas,
C. Uddaraju, H. Khatri, A. Edwards, V. Bedekar, S. Mainali, R. Abbasi, A. Agarwal, M. Fahim ul Haq, M. Ikram
ul Haq, D. Bhardwaj, S. Dayanand, A. Adusumilli, M. McNett, S. Sankaran, K. Manivannan, L. Rigas. Windows
Azure Storage: A Highly Available Cloud Storage Service with Strong Consistency. SOSP ’11. Oct. 23-26, 2011,
Cascais, Portugal.
[4] J. Cowling. https://code.facebook.com/posts/253562281667886/data-scale-june-2016-recap/
[5] Y. Chen, R. Griffith, D. Zats, A. D. Joseph, R. Katz. Understanding TCP Incast and Its Implications for Big Data
Workloads University of California at Berkeley, Technical Report, 2012
[6] A. Dimakis, P. Godfrey, Y. Wu, M. Wainwright, K. Ramchandran. Network coding for distributed storage systems. IEEE
Infocom, May 2007
[7] A. Dimakis, P. Godfrey, Y. Wu, M. Wainwright, K. Ramchandran. Network coding for distributed storage systems. IEEE
Transactions on Information Theory, Vol. 56, No. 9, pp. 4539-4551, September 2010
[8] D. Ford, F. Labelle, F. Popovici, M. Stokely, V. Truong, L. Barroso, C. Grimes, and S. Quinlan. Availability in globally distributed storage systems. Proceedings of the 9th USENIX Symposium on Operating Systems Designs and
Implementation, pp. 1-7 Oct. 2010.
[9] P. Gopalan, C. Huang, H. Simitci, S. Yekhanin. On the Locality of Codeword Symbols. IEEE Trans. Inf. Theory, vol. 58,
No. 11, pp. 6925-6934, Nov. 2012.
[10] C. Huang, H. Simitci, Y. Xu, A. Ogus, B. Calder, P. Gopalan, J. Li, and S. Yekhanin. Erasure Coding in Windows Azure
System. USENIX Annual Tech. Conference, Boston, MA, 2012.
[11] G. Joshi, Y. Liu, E. Soljanin. Coding for fast content download. Communication, Control, and Computing (Allerton),
50th Annual Allerton Conference, IEEE, pp. 326-333, Oct. 2012.
[12] J. Lacan, V. Roca, J. Peltotalo, S. Peltotalo. IETF RFC5510. Reed-Solomon Forward Error Correction (FEC) Schemes.
Internet Engineering Task Force, April 2009.
42
REFERENCES
REFERENCES
1.0E+00
Simulated Data 3x
1.0E-01
Computed bound 3x
Cumulative probability p (missing >= f)
1.0E-02
1.0E-03
1.0E-04
1.0E-05
1.0E-06
1.0E-07
1.0E-08
1.0E-09
1.0E-10
1.0E-11
1.0E-12
60
70
80
90
100
110
120
130
Number of missing fragments f
Fig. 21: Comparison of simulated and computed bound for repair regulation with failure rate estimation and a peak repair rate limit of 3 times the nomimal rate.
[13] M. Luby, A. Shokrollahi, M. Watson, T. Stockhammer, L. Minder. IETF RFC6330. RaptorQ Forward Error Correction
Scheme for Object Delivery. Internet Engineering Task Force, August 2011.
[14] M. Luby. Capacity Bounds for Distribute Storage. submitted to IEEE Transactions on Information Theory, October 2016.
[15] S. Muralidhar, W. Lloyd, S. Roy, C. Hill, E. Lin, W. Liu, S. Pan, S. Shankar, V. Sivakumar, L. Tang, S. Kumar. f4: Facebook’s warm BLOB storage system. 11th USENIX conference on Operating Systems Design and Implementation,
pp. 383-398, 2014.
[16] R. Recio, B. Metzler, P. Culley, J. Hilland, D. Garcia IETF RFC5040 A Remote Direct Memory Access Protocol Specification Internet Engineering Task Force, October 2007.
[17] L. Rizzo. Effective erasure codes for reliable computer communication protocols. ACM SIGCOMM computer communication review, Vol. 27, No. 2, pp. 24-36, April 1997.
[18] R. Rodrigues, B. Liskov. High Availability in DHTs: Erasure Coding vs. Replication. Peer-to-Peer Systems IV, pp. 226239, Springer Berlin Heidelberg, 2005
[19] M. Sathiamoorty, M. Asteris, D. Papailiopoulos, A. Dimakis, R. Vadali, S. Chen, and D. Borthakur. XORing Elephants:
Novel Erasure Codes for Big Data. Proceedings of the VLDB Endowment, pp. 325-336 Vol. 6, No. 5, 2013.
[20] A. Shokrollahi, M. Luby. Raptor Codes. Foundations and Trends in Communications and Information Theory, 2011,
Vol. 6: No 3-4, pp 213-322.
[21] H. Weatherspoon, Hakim, J. Kubiatowicz. Erasure coding vs. replication: A quantitative comparison. Peer-to-Peer Systems, pp. 328-337, Springer-Verlag, 2002.
[22] http://docs.ceph.com/docs/master/rados/configuration/pool-pg-config-ref/
[23] https://github.com/google/snappy/blob/master/README
[24] SM863a Specification Sheet http://www.samsung.com/semiconductor/minisite/ssd/product/enterprise/sm863a.html
43
REFERENCES
REFERENCES
[25] Hard Drive Data and Stats https://www.backblaze.com/b2/hard-drive-test-data.html
44
| 3cs.SY
|
Optimal Subhourly Electricity Resource Dispatch Under Multiple Price Signals
With High Renewable Generation Availability
David P. Chassin1,2 , Sahand Behboodi2 and Ned Djilali2
SLAC National Accelerator Laboratory, Menlo Park CA (USA)
2
Inst. for Integrated Energy Systems, and Mechanical Engineering Dept, University of Victoria, Victoria BC (Canada)
1
arXiv:1711.04175v1 [cs.SY] 11 Nov 2017
Abstract
This paper proposes a system-wide optimal resource dispatch strategy that enables a shift from a primarily energy
cost-based approach, to a strategy using simultaneous price signals for energy, power and ramping behavior. A formal
method to compute the optimal sub-hourly power trajectory is derived for a system when the price of energy and ramping
are both significant. Optimal control functions are obtained in both time and frequency domains, and a discrete-time
solution suitable for periodic feedback control systems is presented. The method is applied to North America Western
Interconnection for the planning year 2024, and it is shown that an optimal dispatch strategy that simultaneously
considers both the cost of energy and the cost of ramping leads to significant cost savings in systems with high levels of
renewable generation: the savings exceed 25% of the total system operating cost for a 50% renewables scenario.
Keywords: Electricity pricing, bulk electric system, optimal energy dispatch, optimal ramping, renewable integration,
resource allocation
Highlights
• A method to minimize the cost of subhourly dispatch
of bulk electric power systems.
• Dispatch based on simultaneous use of energy and
ramping costs yields significant savings
• Savings from optimal dispatch increase as transmission constraints increase.
• Savings from optimal dispatch increase as variable
generation increases.
1. Introduction
The growth of renewable electricity generation resources
is driven in part by climate-change mitigation policies that
seek to reduce the long-term societal costs of continued dependence on fossil-based electricity generation and meet
growing electric system load using lower cost resources.
However, each class of renewable generation comes with
one or more disadvantages that can limit the degree to
which they may be effectively integrated into bulk system
operations.
Hydro-electric generation has long been employed as a
significant renewable electric energy and ramping resource.
But climate change may jeopardize the magnitude and certainty with which the existing assets can meet demand
[1, 2]. Concerns about population displacement, habitat loss and fishery sustainability often limit the growth
Preprint submitted to Applied Energy
of new hydro-electric generation assets, placing additional
constraints on new ramping response resources, such as
requiring the use of additional reserves and ramping resources. Shifts in both load and hydro-electric generation
potentially increase uncertainty in long term planning and
further enhance the need for technological configurations
that support operational flexibility [3].
Wind power has seen rapid growth, but concern about
system reliability has limited the amount of wind generation that can be supported without additional planning and operational measures, such as committing more
carbon-intensive firming resources [4]. Solar resources are
also becoming increasingly available but the intermittency
challenges are similar to those of wind. In addition, residential rooftop solar resources are challenging the classical
utility revenue model [5], can cause voltage control issues
in distribution systems [6], and in extreme cases can result
in overgeneration [7]. Taken together these considerations
have given rise to questions about the reliable, robust control and optimal operation of an increasingly complex bulk
electricity system [8].
The conventional utility approach to addressing renewable generation variability is to allocate additional firm
generation resources to replace all potentially non-firm
renewables resources. These firm resources are typically
fast-responding thermal fossil resources or hydro resources
when and where available. For new renewable resources
the impact of this approach is quantified as an intermittency factor, which discounts for instance the contribution
of wind in addition to its capacity factor and limits the deNovember 15, 2017
gree to which renewables can contribute to meeting peak
demand [9]. However, the intermittency factor does not
account for the ramping requirements created by potentially fast-changing renewable resources [10]. The need for
fast-ramping resources discourages the dispatch of highefficiency fossil and nuclear generation assets and can encourage reliance on low-efficiency fossil-fuel resources for
regulation services and reserves [11].
One solution to overcoming the renewable generation
variability at the bulk electric level is to tie together a number of electric control areas into a super-grid so that they
can share generation and reserve units through optimal
scheduling of system interties [12]. In an interconnected
system, the combined power fluctuations are smaller than
the sum of the variations in individual control areas. Furthermore, fast-acting energy storage systems and demand
response programs can provide required ancillary services
such as real-time power balancing [13] and frequency regulation [14] if they are equipped with suitable control mechanisms. A competitive market framework in which energy
resources participate to sell and buy ancillary service products can accelerate the transition to a high-renewable scenario by supporting the long-term economic sustainability
of flexible resources.
Concerns about the financial sustainability of utilities
under high level of renewables are also beginning to arise.
The question is particularly challenging when one seeks solutions that explicitly maximize social welfare rather than
simply minimizing production cost [15]. The growth of
low-marginal cost renewable resources can lead one to expect utility revenues to decline to the point where they can
no longer recover their long term average costs. But this
conclusion may be erroneous if one fails to consider both
the impact of demand own-price elasticity, as well as the
impact of load control automation on substitution elasticity. The latter type of demand response can significantly
increase the total ramping resource on peak and decrease
ramping resource scarcity. One option for replacing energy resource scarcity rent is increasing fixed payments.
But this may lead to economic inefficiencies as well as an
unraveling of the market-based mechanisms built so far.
Another option is to enable payments based on ramping resource scarcity rent through existing markets for ancillary
services. At the present time, the majority of resources
continue to be dispatched based on the energy marginal
cost merit order. But it is not unreasonable to consider
how one might operate a system in which the energy price
is near zero and resources are dispatched instead according
the ramping cost merit order.
In the presence of high levels of variable generation,
the scheduling problem is a co-optimization for allocating
energy and ramping resources [16]. Under existing energy
deregulation policies, there is usually a market in which
energy producers compete to sell energy, and a separate
market in which they compete to sell power ramping resources for flexibility. Producers get paid for their energy
deliveries in the energy market and for power ramping flex-
ibility in the flexibility market. But today’s dual-pricing
mechanism is dominated by the energy markets, which
drives generation resources to secure revenue primarily in
the energy market, and only deliver residual ramping resources in the flexibility market. Meanwhile poor access to
energy markets leads loads and storage to seek participation primarily in the flexibility market while only revealing their elasticities to the energy market. This relegates
loads and storage to only a marginal role in the overall
operation of the system, which is the motivation for seeking policy solutions to improving their access to wholesale
energy markets, such as FERC Orders 745 and 755.
This paper proposes a system-wide optimal resource
dispatch strategy that enables a shift from primarily energy cost-based approach to primarily ramping cost-based
one. This optimal dispatch answers the question of what
power schedule to follow during each hour as a function of
the marginal prices of energy, power and ramping over the
hour1 . The main contributions of this paper are (1) the
derivation of the formal method to compute the optimal
sub-hourly power trajectory for a system when the cost
of energy and ramping are both of the same order, (2)
the development of an optimal resource allocation strategy based on this optimal trajectory, and (3) a simulation
method to evaluate the cost savings of choosing the optimal trajectory over the conventional sub-hourly dispatch
used in today’s system operation.
In Section 2 we develop the optimal control function
in both time and frequency domains. In the case of the
frequency domain optimal control function the solution is
presented as a continuous function. A discrete-time solution suitable for periodic feedback control systems is presented in Section 3. In Section 4 we examine the performance of this optimal dispatch solution in terms of varying prices for a given “typical” hour and in Section 5, we
analyze the cost savings in an interconnection that models the Western Electric Coordinating Council (WECC)
system for the year 2024 under both low (13%) and high
(50%) renewable generation scenarios. Finally, in Section 6
we discuss some of the consequences that appear to arise
from this new paradigm and our perspectives on possible
future research on this topic.
2. Methodology
Consider a utility’s cost minimization problem over a
time interval T . The utility’s customers purchase their net
energy use E(T ) at a pre-determined retail price. So in today’s systems, profit maximization and cost minimization
are essentially the same problem. For each hour the utility pays for energy delivered at a real-time locationallydependent wholesale price that is also dependent on demand under typical deregulated nodal pricing markets.
1 We define the marginal price of a product or service as the change
in its price when the quantity produced or delivered is increased by
one unit.
2
Q [MW]
R(Q, Q̇) [$/MW]
R(Q, Q̇) [$/MW]
P (Q) [$/MWh]
a
b
Q [MW]
From this we form a second-order ordinary differential
equation describing the critical load trajectory
c
a
µ
Q̈ − Q = .
c
2c
where µ = λ − aQZ . Using the Laplace transform we find
the critical system response in s-domain is
Q̇ [MW/h]
Fig. 1: Power (left) and ramp (center and right) price functions.
Q̂(s) =
,
(2)
Q̇0
µ
µ
cosh ωt +
Q(t) = Q0 +
sinh ωt − ,
2a
ω
2a
(3)
where Q0 and Q̇0 are initial power and ramp values.
We can determine whether this solution is an extremum
by computing the second variation
Z T
∂2C
(T ) =
[α(v)2 + 2β(vv 0 ) + γ(v 0 )2 ]dt
∂Q2
0
Z T
=
H(t)dt,
0
with H(t) > 0 for all v 6= 0 subject to v(0) = 0 = v(T ).
We then have
α=
∂2G
= 2a,
∂Q2
β=
∂2G
= b,
∂Q∂ Q̇
γ=
∂2G
= 2c.
∂ Q̇2
Thus for all a, b, c > 0, H(t) > 0 and Q(t) is a minimizer.
Since the only physical meaningful non-zero values of a
and c are positive, this is satisfactory. We will examine
cases when a and c are zero separately. Note that when
Q̇ < 0, we have b < 0, so that the sign of b does not affect
the general solution.
RT
Given the constraints 0 Q(t)dt = ET and Q(T ) =
QT , which come from the hour-ahead schedule, we obtain
the solution for µ and Q̇0 for the case where a, c > 0:
−1
µ
A B
E∆
=
,
(4)
C D
Q∆
Q̇0
0
Given the dispatch from Q(0) to Q(T ) and the scheduled
RT
energy use E(T ) = 0 Q(t)dt we augment the cost function with the Lagrange multiplier λ so that we have
Z T
a(Q − Qz )Q + b(Q − Qz )|Q̇| + cQ̇2 + λQ dt
where
sinh ωT − ωT
2aω
cosh ωT − 1
C=
2a
sinh ωT
E∆ = ET −
Q0
ω
A=
0
=
µ
2c
where ω 2 = ac . The general time-domain solution for the
critical function over the interval 0 ≤ t < T is
The utility’s scheduled energy use is forecast for each hour
based on their customers’ expected net energy use, which
is then used to compute the utility’s net load over that
hour. We assume that over any interval T the utility may
incur additional costs for any deviation in actual net load
from the scheduled load.
The price function at the operating point is split up
∂P
(measured in
into the marginal price of energy a = ∂Q
∂R
2
$/MW ·h), the marginal price of power b = ∂Q
(mea2
sured in $/MW ), and the marginal price of ramping c =
∂R
(measured in $·h/MW2 ). In order to reflect resource
∂ Q̇
scarcity all cost functions are assumed to be quadratic so
that the price function for each is linear as shown in Fig. 1.
The marginal prices a and b determine prices as a function
of the power demand Q, and the marginal price c determines prices based on the ramp rates Q̇. The cost parameters arise from the schedule and may vary from hour to
hour, but do not change within any given hour. Any of
the marginal prices may be zero or positive depending on
the market design and prevailing conditions in the system.
For the purposes of this paper, we will assume that they
cannot be negative.
Over the time interval T the total cost of both the
power trajectory Q(t) and the ramping trajectory Q̇(t)
given the power price P (t) = aQ(t) and ramp price R(t) =
bQ(t) + cQ̇(t), respectively, is given by
Z T
C(T ) =
P [Q(t)]Q(t) + R[Q(t), Q̇(t)]Q̇(t)dt.
(1)
Z
Q0 s2 + Q̇0 s +
s(s2 − ω 2 )
T
G(t, Q, Q̇)dt,
0
where the |Q̇| represents the magnitude of the ramp rate
Q̇, and QZ is the amount of must-take generation having
zero or effectively zero marginal energy cost. Then the
optimal dispatch trajectory Q(t) is the critical function
obtained by solving the Euler-Lagrange equation
cosh ωT − 1
ω2
sinh ωT
D=
ω
B=
Q∆ = QT − Q0 cosh ωT .
When a = 0, the cost of energy is zero and only the
ramping cost is considered. Then the time-domain solution is
µ
Q(t) = t2 + Q̇0 t + Q0 ,
(5)
4c
∂G
d ∂G
−
= 0.
∂Q dt ∂ Q̇
3
/
µ
/
Q0
Q̇0
/
/ 2
1
4a
/+
O
1
2
/
−
/+
/ 2
1
2ω
/
ωeωts z
z−1
−
/+
O
/Q
ωe−ωts z
z−1
Fig. 2: Optimal dispatch controller diagram with discrete update time ts .
with
T3
12c
T2
C=
4c
E∆ = ET − Q0 T
A=
B=
T2
2
D=T
Fig. 3: Optimal discrete time control for various values of ω
at ts = 5 minutes.
Q∆ = QT − Q0 ,
which gives the critical response in s-domain
Q̂(s) =
which handles the transition from the initial system load
µ
Q0 to the scheduled load QE = − 2as
. The central response is dominated by the scheduled load solution
h µ i
(s).
K2 = L−1 −
2as
µ
Q̇0
Q0
+ 2 +
.
3
4cs
s
s
When c = 0, there is no scarcity for ramping so that the
ramping price is based only on the marginal energy cost
of additional units that are dispatched. Then we have the
time-domain solution
µ
Q(t) = − ,
(6)
2a
Finally, the terminal response is dominated by the reversetime solution
"
#
Q0
Q̇0
µ
ωt
−1
2 + 2ω + 4a
K3 e = L
(s),
s−ω
with
µ=−
2aET
.
T
which handles the transition from the scheduled load to
the terminal load QT . A discrete-time controller that implements the solution of Eq. 7 is shown in Figure 2. The
controller implements the three main components to the
optimal solution with step inputs µ, Q0 , and Q̇0 . Note
that the marginal prices a, b and c for the entire hour are
constants in the controller blocks, which makes the controller design linear time-invariant within each hour, but
time-variant over multiple hours. The discrete-time solution is then
K1 τ k + K2 + K3 τ −k : a > 0, c > 0
∗
µ 2 2
Q (k) =
t k + Q̇0 ts k + Q0 : a = 0, c > 0
4c µs
− 2a
: a > 0, c = 0
This gives the critical response is s-domain
Q̂(s) = −
µ
,
2as
µ
and the initial and final ramps from Q(0) to − 2a
and from
µ
− 2a to Q(T ) are limited by the ramping limits of the responding units.
3. Optimal Dispatch Controller
The partial fraction expansion of Eq. 2 is
K2
K3
K1
+
+
,
s+ω
s
s−ω
where K1 =
Q̇0
2ω
Q0
2
−
Q̇0
2ω
+
µ
4a ,
µ
K2 = − 2a
, and K3 =
(7)
Q0
2
where τ = eωts .
The discrete-time dispatch p
control is illustrated in Figure 3 for various values of ω = a/c. When the value of ω
is large, the optimal dispatch is dominated by the energy
cost and the cost of high ramp rates is negligible compared
to the energy cost. The result is a dispatch that moves as
quickly as possible to scheduled load QE . In the limit of
zero ramping cost, the optimal response is a step function2 .
+
µ
4a ,
+
with the values of the parameters are computed
from Eq. 4.
The initial response of the optimal controller is dominated by the forward-time solution
"
#
Q0
Q̇0
µ
−
+
2ω
4a
K1 e−ωt = L−1 2
(s),
s+ω
2 Step responses are only possible by generation or load tripping,
which is not considered as part of the conventional control strategy.
4
As the cost of ramping increases relative to the energy cost,
the optimal dispatch begins to reduce the ramp rate while
still following a trajectory that satisfies the hourly energy
delivery requirement. In the limit of zero energy cost, the
optimal dispatch trajectory is parabolic.
Power (MW)
Q(t)
4. Performance Evaluation
In this section we develop the cost performance metric of the optimal dispatch control design. The optimal
dispatch cost function is found by evaluating Equation 1
using Equations 3, 5 and 6. Thus when a, b, c > 0 we have3
QE
QT
Q0
0
10
20
30
40
50
60
10
20
30
40
50
60
10
20
30
40
50
60
t
Q(t)
Power (MW)
sinh 2ωT
a(A2 + B 2 ) + bABω
C(T ) =
2ω
sinh2 ωT
+
b(A2 + B 2 )ω + 4aAB
2ω
cosh ωT − 1
+
[(bAω + 2aB)C − (aB + bAω)Qz ]
ω
sinh ωT
+
[(bBω + 2aA)C − (aA + bBω)Qz ]
ω2
+ aC − aCQz T.
QT
QE
Q0
0
where A = Q0 + µ/2a, B = Q̇0 /ω and C = −µ/2a. For
the case when a = 0 we have
t
Q(t)
C(T ) = 21 bA2 T 4 + [bB + 43 cA]AT 3
Power (MW)
+ [ 12 b(B 2 + 2AC) + 2cAB − bAQz ]T 2
+ [(bC + cB)B − bBQz ]T,
where A = µ/4c, B = Q̇0 , and C = Q0 . When c = 0 we
have
ET
C(T ) = aET
− Qz .
T
QE
t
Time (min)
Fig. 4: Conventional power dispatch for base case: (top) significant negative schedule error requiring over-production, (center) small negative, zero and positive schedule error requiring
over (red), normal (black) and under (blue) production, and
(bottom) significant positive schedule error requiring underproduction.
The cost of the base case is then
with the initial and terminal ramp rates
and
Q0
0
We use as the base case a conventional unit dispatch
strategy that requires generators ramp to their new operating point during the 20 minutes spanning the top of the
hour. Accordingly the generators begin ramping 10 minutes before the hour and end ramping 10 minutes after
the hour. In the aggregate for a given hour this strategy
is illustrated in Figure 4 where
6
Q0 + QT
QE =
ET −
,
5
12
Q̇0 = 6(QE − Q0 )
QT
Q̇T = 6(QT − QE ).
Cbase (T ) =
−
Three cases are shown: overproduction to compensate for
a lack of generation in previous hours (top), scheduled
production (center), and underproduction to compensate
for extra generation in previous hours (bottom).
+
2
2
aT
18 (QT + QT QE + 14QE + QE Q0
aT
12 (QT + 10QE + Q0 )Qz
| 2b (QE − Q0 )|(QE + Q0 − 2Qz )
+ Q20 )
+ | 2b (QT − QE )|(QT + QE − 2Qz )
+
3 Note that if the ramp rate Q̇ changes sign at the time t =
c
tanh−1 (− B
) and 0 < tc < T , then we must divide the cost integral
A
into two parts to account for the absolute value of Q̇ on b terms.
2
6c
T (QT
− 2QT QE + 2Q2E − 2Q0 QE + Q20 ).
The zero-order hold ramp discrete form of Equation 1
gives us the cost of operating with a discrete control time-
1
ω
5
step ts , i.e.,
T /ts
C ∗ (T ) =
X
P ∗ [Q∗ (k)]Q∗ (k) + R∗ [Q∗ (k), Q̇∗ (k)]Q̇∗ (k) ts
k=0
T /ts
=
X
ats
4
h
Q∗ (k)2 + 2Q∗ (k)Q̇∗ (k)
k=0
i
+ Q̇∗ (k)2 − 2Qz [Q∗ (k) + Q̇∗ (k)]
+
1
2
+
c
ts
h
b(Q̇∗ (k) − Q∗ (k)) (Q̇∗ (k) + Q∗ (k) − 2Qz
h
i
Q∗ (k)2 − 2Q∗ (k)Q̇∗ (k) + Q̇∗ (k)2
i
where Q∗ (k) = Q(kts ) and Q̇∗ (k) = Q∗ (k + 1). We evaluate the performance of the control strategy for different
control update rates ts using two future scenarios, one for
low renewables where ω > 1, and one for high renewables
where ω < 1 for both unconstrained and constrained transmission operating conditions.
5. Case Study: WECC 2024
In this section we examine the cost savings associated
with using the optimal control solution on the WECC
2024 base case model introduced in [12]. The WECC 2024
model is a 20-area base case used by WECC for planning
studies. The 20-area model combines a number of smaller
control areas based on the anticipated intertie transfer limits reported in the WECC 2024 common case [17]. In this
model constraints within control areas are ignored, while
internal losses are approximated. The peak load, annual
energy production and demand consumption are forecast,
including intermittent wind, solar, and run-of-river hydro
for the entire year.
The model also includes a hypothetical market for each
consolidated control area, with a flat zero-cost supply curve
for all renewable and must-take generation resources and
a constant positive supply curve slope for all dispatchable
units. The hourly generation of intermittent resources is
provided by the base case model and incorporated into the
supply curve so that there is effectively no marginal cost of
production for renewable energy and must take generation.
All generating units are paid the hourly clearing price, and
when the marginal energy price in a control area is zero
then renewable generation may be curtailed. As a result,
under the high renewable scenario, zero energy prices are
commonplace and renewable generation is curtailed more
frequently. Demand response is similarly considered for
each control area and the output of this scheduling model
provides the hourly nodal prices required to satisfy the
transmission constraints, if any.
The low renewables case is the WECC forecast for the
year 2024, which correspond to 29.5 GW (16.1%) of renewable capacity and 140.8 TWh (13.4%) of annual renewable generation. The high renewables case is given as
400% of capacity of the WECC forecast for the year 2024,
6
which corresponds to 117.8 GW (63.5%) and 523.9 TWh
(49.6%), respectively. The blended energy price of operations is $130.6/MWh and $50.2/MWh for the low and
high renewables cases, respectively.
The ramping price was not considered in the WECC
2024 base case model. For this study we have assumed
that the ramping energy cost is based on the marginal energy cost for the dispatchable generation and the demand
response, as well as the cost of changing the dispatchable
generation output, as shown in Table 1. In both cases, the
marginal price of power b is the average marginal price of
energy a over the hour. In the low renewables case the
marginal price of ramping c is the marginal price of power
b multiplied by 12 seconds. In the high renewables case,
c is the marginal price of power b multiplied by 49 hours.
The value of ω is approximately 121 times greater in the
low renewable case than it is in the high renewable case.
Note that a is zero when renewables are curtailed while
b is assumed to also be zero because curtailed renewables
and demand response are presumed to be dispatchable.
The values of the ramping response constant c were also
selected such that the overall cost of operating the system
remains more or less constant when going from the low to
high renewables scenarios under the base case. This allows
us to evaluate the impact of the optimal control strategy
without involving the question of revenue adequacy under
the high renewables scenario. Given that there are few
markets from which to determine these values, we must
be satisfied with this assumption for now.
The statistical nature of the intermittency and load
forecast errors and their connection to load following and
regulation was studied at length in [18]. The authors
showed that consolidated control of WECC could yield
both cost savings and performance improvements. In particular, the study showed that with high accuracy control
1% standard deviation in load forecast was expected, with
0% real-time mean error at 0.15% standard deviation at
peak load. However, for the purposes of a preliminary
study like the one presented in this paper, we will consider
the scheduling error to be Gaussian with a mean error of 0
MW and a standard deviation of 100 MW. We believe that
energy and flexibility markets should be efficient enough
to remove all systematic error from the price signals leaving only the random noise that is satisfactorily modeled
by Gaussian noise.
The comparison of the conventional and optimal dispatch for a typical case is shown in Figure 5. The conventional control strategy is shown in dotted lines, with
the 10 minute optimal-dispatch trajectory shown as solid
lines. Note that the ramp rate is constant between discrete
control updates. The evaluation is completed with the
marginal prices and marginal costs at 100 GW, as shown
in Table 1. The energy schedule changes according to a
varying energy error remaining at the end of the previous dispatch interval. A −5% error represents an energy
deficits of 5 GWh for a 105 GWh schedule, while a +5%
error represents an energy surplus of 5 GWh.
Table 3: WECC 2024 cost savings from optimal dispatch under different transmission constraint and renewable scenarios
Total
Scenario Energy
Model
(TWh)
Unconstrained:
Low
1054.6
High
1067.2
Constrained:
Low
1054.5
1055.7
High
Base case
($/MWh)
Price
Optimal
($/MWh)
119.5
101.8
119.35 (-0.1%)
67.29 (-51.2%)
174.8
367.8
174.55 (-0.2%)
87.96 (-318.2%)
Table 4: Summary of energy and price impacts of optimal
dispatch control for the WECC 2024 base case
Total
Scenario Energy
(TWh)
Model
Unconstrained:
Low
1054.6
High
1067.2
Constrained:
Low
1054.5
High
1055.7
Fig. 5: Single hour optimal dispatch for low and high renewables with a ramp from 100 GW to 110 GW using a 10-minute
discrete-time dispatch control rate, with hourly energy schedule
correction errors varying from −5% to +5%.
Table 1: Marginal prices and marginal costs for 105 GWh
schedule at 100 GW initial power and 10 GW/h ramp for cases
in Figure 5
Variable
Marginal
a
b
c
Marginal
P
R
ω
Base case
prices:
1.27 × 10−3
1.27 × 10−3
4.23 × 10−6
costs:
133.09
133.13
17.3
Study case
Units
6.34 × 10−4
6.34 × 10−4
3.09 × 10−2
$/MW2 ·h
$/MW2
$·h/MW2
66.55
375.19
0.1433
$/MW·h
$/MW
h−1
Cost
Optimal
($B/y)
Savings
($B/y)
125.9
77.8
0.16 (0.1%)
30.85 (28.4%)
184.1
231.2
0.26 (0.1%)
157.12 (40.5%)
Price
Optimal
($/MWh)
119.5
101.8
119.35 (-0.1%)
67.29 (-51.2%)
174.8
367.8
174.55 (-0.2%)
87.96 (-318.2%)
2. For the low renewables case, the energy cost is roughly
10 times the ramping cost, while for the high renewables case the ramping cost is roughly 10 times the
energy cost for the nominal schedule. This was necessary to ensure that costs were the same for both
cases.
3. The marginal power price b for both cases is equal to
the marginal energy price a of the respective case.
We considered the performance degradation resulting
from longer dispatch intervals by evaluating the performance using 5 minute updates, 1 minute updates, and 4
second discrete control timesteps but found no appreciable
difference in the economic performance. The results shown
in Table 2 are shown for the 5 minute dispatch interval.
The output of the presented discrete control method is a
load profile that does not necessarily lead to the scheduled
hourly energy, because the load trajectory over each time
intervals (which is linear) is slightly different from the optimal load trajectory (that often has a curvature). One
approach to deal with this energy deficiency is to use a
higher time resolution, so that the trajectories lay on each
other more precisely. Another approach is to adjust the
targeted load such that it delivers the scheduled energy
over each time interval. In this case, the discrete control
load is not necessarily equal to the optimal load.
Generally at low levels of renewables savings are not
possible using the optimal control strategy. The cost savings observed in the extreme low renewables dispatch cases
in Table 2 are due to the fact that discrete dispatch control
Table 2: Single hour cost savings under low and high renewable for a ramp from 100 GW to 110 GW at 5 minute discrete
dispatch control update rate, with varying energy error redispatch
Scenario Base case
Model
($B/y)
Unconstrained:
Low
126.0
High
108.6
Constrained:
Low
184.4
High
388.3
Base case
($/MWh)
The marginal prices in Table 1 are chosen to satisfy the
following conditions:
1. The system operating cost is roughly $100/MWh at
a system load of 100 GW.
7
Fig. 6: WECC 2024 load duration (top) and optimal dispatch savings duration (bottom) using discrete optimal control at 5minute dispatch rate for the unconstrained (left) and constrained (right) high renewables scenario. The scatter plots are the
corresponding cost (top) and load (bottom) values for the durations curves shown.
follows the optimal trajectory sampling every ts seconds.
This dispatch error can result in small over or underproduction depending on the degree of asymmetry in the optimal trajectory.
At higher levels of renewables the savings are potentially more significant. In addition, the savings are maximum when dispatch tracks the original schedule, which
suggests that there may be a strong economic incentive to
avoid carrying over energy tracking error from one schedule interval to the next.
The interconnetion-wide scheduling solution in [12] includes a 20-area constrained solution. The hourly energy prices for each area are computed considering both
supply and demand energy price elasticities. The energy
prices are computed for the interconnection-wide surplusmaximizing schedule over the entire year. The marginal
power price is the price of energy for the schedule hour.
The marginal price of ramping is 1/300 marginal price of
power in the low renewable case, and 49 times the marginal
price of power in the high renewable case. The costs, savings and price impact of using this scheduling solution
compared to the base case are presented in Tables 3 and
4. The unconstrained solution is evidently less costly because the combined system-wide fluctuations are smaller
than the sum of the individual of the variations in each
balancing authority.
The WECC 2024 system-wide load and savings duration curves4 are shown in Figure 6. The potential savings
are very significant for all scenarios, with the highest savings being found when high levels of renewable resources
are available. The savings when more transmission constraints are active are augmented considerably with respect to unconstrained system conditions.
6. Discussion
The significance of the results shown in Figure 3 cannot
be understated. First we observe that when a >> c, the
optimal response is very similar to the conventional dispatch strategy, giving us some assurance that today’s operations are very nearly optimal. However, when a << c
today’s hourly dispatch strategy is not optimal. As the
fraction of cost attributed to energy decreases relative to
the cost attributed to ramping, we see that ω decreases
and the value of changing the dispatch strategy increases
dramatically. In the limit of a very high renewable scenario
the savings achievable using the optimal dispatch strategy
can be extremely significant. Failure to adopt an optimal
dispatch such as the one proposed could result in major
and likely unnecessary costs. Utilities will inevitably find
it necessary to mitigating these costs, either by reducing
the amount of renewables, by increasing the revenues from
their customers, or by developing some kind of optimal resource allocation strategy such as the one proposed.
A sensitivity analysis of the savings as a function of
the marginal price of ramping c shows that the savings are
not overly sensitive to changes in our assumption of the
cost of ramping scarcity. Figure 7 shows that for a 50%
decrease in c, we observe a 10.3% decrease in savings, while
a 50% increase in c results in a 3.9% increase in savings.
This suggests that the savings from employing the optimal
dispatch strategy is quite robust to our uncertainty about
the marginal price of ramping resources.
In any financially sustainable future scenario, we must
consider how the long-term average costs and fixed costs
are recovered under the pricing mechanism. We have assumed in this study that renewable generation and utilities cannot sustainably continue employing complex power
purchasing agreements and subsidies to hedge against energy price volatility. Instead all parties should come to
rely on separate real-time pricing mechanisms for energy,
power and ramping response of the resources they control.
4 A duration curve shows the number of hours per year that a
time-series quantity is above a particular value. It is obtained by
sorting the time-series data in descending order of magnitude and
plotting the resulting monotonically descending curve.
8
marginal cost of additional resource capacity is never considered for optimal subhourly dispatch control. This is
consistent with the expectation that sunk costs should not
be a factor in the selection of which units to dispatch at
what level, at least to the extent that these costs are not
entering into the energy or ramping costs.
In the presence of significant renewables, the energy
marginal cost does not entirely reflect the grid condition
without considering the cost of ramping up and down services. Therefore, the energy price cannot be solely used as
a control signal to the generation and load units to achieve
the optimal utilization of resources. In order to quantify
the value of ramping product we suggest using a market
framework in which flexible generation and load resources
compete to sell their ancillary service products at the bulk
electric system level. As renewable level rises the energy
marginal cost decrease (smaller a value) because renewables are zero-generation cost resources, but the ramping
marginal cost increases (larger c value) because the system
requires more flexibility to handle the generation variation.
In long run, inflexible units get retired and more flexible
units are built to support the renewable integration since
flexibility will be a revenue source rather than energy.
The availability of high renewables can lead to situations where low cost energy is being supplied to areas with
high cost flexibility resources through constrained interties. The optimal strategy avoids dispatching these high
cost flexilibity resources to the extent possible by reducing the ramping schedule. The more transmission capacity
is available, the lower the overall cost, but we note that
even when the system is constrained, the cost of optimally
dispatching flexibility resources can be significantly lower
under the high renewables case than under a low renewables scenario.
It seems that the use of energy-only market designs run
counter to the results of this study. Flexilibity resource
markets may become increasingly important, even in regions that are not dominated by local renewable generation. This is especially true in cases where adequate transmission capacity is available for renewables in remote regions to displace local dispatchable generation. This may
give rise to a new set of challenges for utility and system
operators as they seek a revenue model that not only provides for operating costs, but also maintains the coupling
between retail demand response and wholesale supply and
retail delivery constraints. If the cost of the wholesale system becomes increasingly dominated by ramping resource
constraints, while retail continues to use energy prices to
encourage consumer efficiency, then retail behavior will be
not affected as much by short-term wholesale price fluctuations. This trend runs against the desire for more engaged
consumers who can respond to system conditions in real
time. Clearly a new utility revenue model is needed if the
transformation to a high renewable modus operandi is to
occur successfully in the coming decades.
Fig. 7: Sensitivity of savings to marginal price of ramping
resources.
Shifting revenue from resource allocation mechanisms
based primarily on energy resource scarcity to ones based
primarily on flexibility resource scarcity can be expected to
have a significant impact on the cost of subhourly resource
dispatch. The optimal strategy for low renewable conditions very closely matches the strategy employed today
when moving hour-to-hour from one scheduled operating
point to another. Indeed, the optimal dispatch strategy
does not offer any significant cost savings when overall
pricing is dominated by energy resource scarcity.
However, as increasing amounts of renewables are introduced, the scarcity rents may shift from energy to flexibility resources. The optimal subhourly dispatch strategy
may be expected to change with increasing emphasis on
avoiding high ramp rates over sustained periods at the expense of maintaining a constant power level over the hour.
The relationship between existing price signals for various grid services and the three principle price components
needed to implement this optimal strategy requires further investigation. It is evident that the marginal price a
represents a linearization of the energy price itself at the
current operating point. But it is not clear yet whether
and to what degree the marginal prices b and c can be
connected to any existing price signals, such as the capacity price or the price of ancillary services like frequency
regulation resources, generation reserves, and demand response. The links do suggest themselves based on both the
resource behaviors and physical dimensions of the parameters. However, it is not certain yet whether this will be
simply a matter of obtaining a linearization of the services’
cost functions at the appropriate operating point.
Additionally, it is instructive to note that the marginal
price of redispatched power b is not important to the optimal dispatch strategy, insofar as the parameter does not
appear in Eq. 2. This leads one to conclude that to the
extent capacity limits do not affect either energy or ramping scarcity rents (or are already captured in them), the
9
7. Conclusions
[8] I. Dobson, B. A. Carreras, V. E. Lynch, and D. E. Newman.
Complex systems analysis of series of blackouts: Cascading failure, critical points, and self-organization. Chaos: An Interdisciplinary Journal of Nonlinear Science, 17(2):026103, 2007.
[9] G. Boyle. Renewable electricity and the grid: the challenge of
variability. Earthscan, 2012.
[10] Y.V. Makarov, C. Loutan, J. Ma, and P. de Mello. Operational
impacts of wind generation on California power systems. IEEE
Transactions on Power Systems, 24(2):1039–1050, 2009.
[11] M. Nyberg. Thermal efficiency of gas–fired generation in California: 2012 update. California Energy Commission, 2013.
[12] S. Behboodi, D. P. Chassin, N. Djilali, and C. Crawford.
Interconnection-wide hour-ahead scheduling in the presence of
intermittent renewables and demand response: A surplus maximizing approach. Applied Energy, 189:336–351, 2017.
[13] S. Behboodi, D. P. Chassin, N. Djilali, and C. Crawford. Transactive control of fast-acting demand response based on thermostatic loads in real-time retail electricity markets. Applied
Energy, in press, 2017.
[14] D. P. Chassin, S. Behboodi, Y. Shi, and N. Djilali. H2-optimal
transactive control of electric power regulation from fast-acting
demand response in the presence of high renewables. Applied
Energy, in press, 2017.
[15] G. A. Stern, J. H. Yan, P. B. Luh, and W. E. Blankson. What
objective function should be used for optimal auctions in the
iso/rto electricity market? In Power Engineering Society General Meeting, pages 10–pp. IEEE, 2006.
[16] Y. T. Tan and D. S. Kirschen. Co-optimization of energy and
reserve in electricity markets with demand-side participation in
reserve services. In Power Systems Conference and Exposition,
pages 1182–1189. IEEE, 2006.
[17] WECC’s
System
Adequacy
Planning
Department.
WECC 2024 common case.
https://www.wecc.biz/
SystemAdequacyPlanning/Pages/Datasets.aspx.
[Online;
accessed 21-July-2017].
[18] R. Diao, N. Samaan, Y. Makarov, R. Hafen, and J. Ma. Planning for variable generation integration through balancing authorities consolidation. IEEE Power and Energy Society General Meeting, pages 1–8, 2012.
The principal finding of this paper is that the use of an
optimal dispatch strategy that considers both the cost of
energy and the cost of ramping resources simultaneously
leads to significant cost savings in systems with high levels
of renewable generation. For the WECC 2024 common
case the savings can exceed 25% of total operating costs
in a 50% renewables scenario.
As the bulk power interconnection resource mix shifts
from primarily dispatchable non-zero marginal cost resources (e.g., natural gas) to primarily non-dispatchable
zero marginal cost renewable resources (e.g., run-of-river
hydro, wind, solar) we expect a steady shift in bulk system costs from energy resource scarcity rent to ramping
resource scarcity rent. While the total revenue must remain largely the same for financial sustainability, operating strategies must adapt to reflect changes in resource
scarcity costs.
Acknowledgments
This work was funded in part by the Pacific Institute
for Climate Solutions at the University of Victoria and
the California Energy Commission under Project EPC-15047. SLAC National Accelerator Laboratory is operated
for the US Department of Energy by Stanford University
under Contract No. DE-AC02-76SF00515. The authors
also thank Francesco Carducci for his very helpful suggestions during the preparation of this paper.
References
References
[1] M. S. Markoff and A. C. Cullen. Impact of climate change on
Pacific Northwest hydropower. Climatic Change, 87(3-4):451–
469, 2008.
[2] M. K. Chandel, L. F. Pratson, and R. B. Jackson. The potential
impacts of climate-change policy on freshwater use in thermoelectric power generation. Energy Policy, 39(10):6234 – 6242,
2011. Sustainability of biofuels.
[3] S. C. Parkinson and N. Djilali. Robust response to hydroclimatic change in electricity generation planning. Climatic
Change, 130(4):475–489, 2015.
[4] M. A. Ortega-Vazquez and D. S. Kirschen. Estimating the
spinning reserve requirements in systems with significant wind
power generation penetration. IEEE Transactions on Power
Systems, 24(1):114–124, Feb 2009.
[5] G. Blackburn, C. Magee, and V. Rai. Solar valuation and the
modern utility’s expansion into distributed generation. The
Electricity Journal, 27(1):18–32, 2014.
[6] M. J. E. Alam, K. M. Muttaqi, and D. Sutanto. Mitigation of
rooftop solar PV impacts and evening peak support by managing available capacity of distributed energy storage systems.
IEEE Transactions on Power Systems, 28(4):3874–3884, Nov
2013.
[7] P. Denholm, M. O’Connell, G. Brinkman, and J. Jorgenson.
Overgeneration from solar energy in California: A field guide to
the duck chart. National Renewable Energy Laboratory, 2015.
10
| 3cs.SY
|
THE USE OF BAD PRIMES IN RATIONAL
RECONSTRUCTION
arXiv:1207.1651v2 [math.AC] 7 Oct 2013
JANKO BÖHM, WOLFRAM DECKER, CLAUS FIEKER, AND GERHARD PFISTER
Abstract. A standard method for finding a rational number from its
values modulo a collection of primes is to determine its value modulo
the product of the primes via Chinese remaindering, and then use Farey
sequences for rational reconstruction. Successively enlarging the set
of primes if needed, this method is guaranteed to work if we restrict
ourselves to “good” primes. Depending on the particular application,
however, there may be no efficient way of identifying good primes.
In the algebraic and geometric applications we have in mind, the
final result consists of an a priori unknown ideal (or module) which is
found via a construction yielding the (reduced) Gröbner basis of the
ideal. In this context, we discuss a general setup for modular and, thus,
potentially parallel algorithms which can handle “bad” primes. A key
new ingredient is an error tolerant algorithm for rational reconstruction
via Gaussian reduction.
1. Introduction
Rational reconstruction, in conjunction with Chinese remaindering, provides a standard way of obtaining results over the rational numbers from
results in characteristic p > 0. This is of particular use in the design of
parallel algorithms and in situations where the growth of intermediate results matters. Classical applications are the computation of polynomial
greatest common divisors (see [Wang 1981, Encarnación 1995]) and Gröbner
bases (see [Arnold 2003]). The goal of the modular Gröbner basis algorithm presented in [Arnold 2003] is to compute the Gröbner basis of an
ideal already known. That is, the ideal is given by a finite set of generators. In contrast, there are constructions which yield a priori unknown
ideals by finding their (reduced) Gröbner bases. Prominent examples are
the computation of normalization and the computation of adjoint curves (see
[Böhm et al. 2013a, Böhm et al. 2013b]). Here, for the purpose of modularization, we expect that the respective construction applies to a given set
of input data in characteristic zero as well as to the modular values of the
input data. In such a situation, problems may arise through the presence of
“bad” primes of different types.
Usually, a first step to resolve these problems is to show that the bad
primes are “rare”. Then the different types of bad primes are addressed. For
example, a prime for which it is a priori clear that the modular construction
does not work will simply be rejected. Depending on the application, however, there may be bad primes which can only be detected a posteriori, that
Date: October 9, 2013.
Key words and phrases. rational reconstruction, Farey map.
1
THE USE OF BAD PRIMES IN RATIONAL RECONSTRUCTION
2
is, after the true characteristic zero result has been found. For such an application, we must ensure that the reconstruction algorithm used will return
the correct rational number even in the presence of bad primes. In this note,
we derive such an algorithm. Based on this algorithm, we describe a general
scheme for computing ideals (or modules) in the algebraic and geometric
applications we have in mind, addressing in particular how various types of
bad primes are handled. This scheme has already been successfully used in
the aforementioned papers [Böhm et al. 2013a, Böhm et al. 2013b].
To begin, in Section 2, we recall the classical approach to rational reconstruction which is based on the lifting of modular to rational results by
computing Farey preimages via Euclidean division with remainder. Section
3 contains a short discussion of the different types of bad primes. In Section
4, we present the new lifting algorithm which is based on Gaussian reduction,
and discuss the resulting error tolerant reconstruction algorithm. Finally,
in Section 5, we present our general scheme for applications in commutative
algebra and algebraic geometry. We finish by giving an explicit example
which involves a bad prime that can only be detected a posteriori.
To summarize, we focus on the presentation of a general setup for modular
computations based on error tolerant rational reconstruction. We do not
discuss implementation details or performance questions. In fact, for the
applications we have in mind, the time used for rational reconstruction can
be neglected in view of the more involved parts of the respective algorithms.
2. Reconstruction of a single rational number
We describe the reconstruction of a single unknown number x ∈ Q. In
practical applications, this number will be a coefficient of an object to be
computed in characteristic 0 (for example, a vector, polynomial, or Gröbner
basis). Here, we suppose that we are able to verify the correctness of the
computed object (in some applications by a comparably easy calculation, in
others by a bound on the size of the coefficients).
We use the following notation: Given an integer N ≥ 2 and a number
x = ab ∈ Q with gcd(a, b) = 1 and gcd(b, N ) = 1, the value of x modulo N is
a
:= (a + N Z)(b + N Z)−1 ∈ Z/N Z.
xN :=
b N
We also write x ≡ r mod N if r ∈ Z represents xN .
In what follows, we suppose that in the context of some application, we
are given an algorithm which computes the value of the unknown number
x ∈ Q modulo any prime p, possibly rejecting the prime. For reference
purposes, we formulate this in the black box type Algorithm 1.
Algorithm 1 Black Box Algorithm x mod p
Input: A prime number p.
Output: false or an integer 0 ≤ s ≤ p − 1 such that x ≡ s mod p.
Assumption: There are only finitely many primes p where the return value
is false.
Once the values of x modulo each prime in a sufficiently large set of primes
P have been computed, we may find x via a lifting procedure consisting of
THE USE OF BAD PRIMES IN RATIONAL RECONSTRUCTION
3
two steps: First, useQChinese remaindering to obtain the value of x modulo
the product N := p∈P p. Second, compute the preimage of this value
under the Farey rational map which is defined as follows.
For an integer B > 0, set
o
na
∈ Q | gcd(a, b) = 1, 0 ≤ a ≤ B, 0 < |b| ≤ B ,
FB =
b
and for m ∈ Z/N Z, let
a
o
na
∈ Q | gcd(a, b) = 1, gcd(b, N ) = 1,
=m
QN,m =
b
b N
be the set of rational numbers whose value modulo N is m. Then QN =
S
N −1
pm=0 QN,m is a subring of Q with identity. If B is an integer with B ≤
(N − 1)/2, then the Farey map
a
a
7→
,
ϕB,N : FB ∩ QN → Z/N Z,
b
b N
is well–defined and injective (but typically not surjective). To obtain the
injective map with the largest possible image for a given N , we tacitly
suppose in what follows that B is chosen as large as possible for N .
Algorithm 2 below will return ϕ−1
B,N (r) if r is in the image of the Farey
map, and false otherwise (see, for example, [Kornerup and Gregory 1983,
Wang 1981, Wang et al. 1982, Collins et al. 1994]).
Algorithm 2 Farey Preimage
Input: Integers N ≥ 2 and 0 ≤ r ≤ N − 1.
Output: false or a rational number ab with gcd(a, b) = 1, gcd(b, N ) = 1,
p
p
a
(N − 1)/2, 0 < |b| ≤ (N − 1)/2.
b ≡ r mod N , 0 ≤ a ≤
1: (a0 , b0 ) := (N, 0), (a1 , b1 ) := (r, 1), i := −1
2: while 2a2i+2 > N − 1 do
3:
i := i + 1
4:
divide ai by ai+1 to find qi , ai+2 , bi+2 such that
(ai , bi ) = qi (ai+1 , bi+1 ) + (ai+2 , bi+2 )
and 0 ≤ ai+2 < ai+1
if 2b2i+2 ≤ N − 1 and gcd(ai+2 , bi+2 ) = 1 then
i+2
6:
return abi+2
7: return false
5:
Combining Algorithm 2 with Chinese remaindering as indicated above,
we get the classical reconstruction Algorithm 3.
Note that Algorithm 3 works correctly since we suppose that our Black
Box Algorithm 1 either returns false or a correct answer. For most applications, however, there exist primes p which are bad in the sense that
the algorithm under consideration returns a wrong answer modulo p which
can only be detected a posteriori. In this note, we show that if there are
only finitely many such primes, they can just be ignored. More precisely,
we show that in Algorithm 3, we may call the black box type Algorithm 4
below instead of Algorithm 1, provided we then call the lifting Algorithm 5
from Section 4 instead of Algorithm 2.
THE USE OF BAD PRIMES IN RATIONAL RECONSTRUCTION
4
Algorithm 3 Reconstruction of a Rational Number
Input: Algorithm 1 and a way to verify that a computed number equals x.
Output: x.
1: N := 1, r := 0
2: p := 2
3: loop
4:
let s be the return value of Algorithm 1 applied to p
5:
if s = false then
6:
continue with Step 13
7:
find 1 = eN + f p and set r := rf p + seN , N := N p
8:
let y be the return value of Algorithm 2 applied to N and r
9:
if y = false then
10:
continue with step 13
11:
if y = x then
12:
return y
13:
p :=NextPrime(p)
Algorithm 4 Black Box Algorithm x mod p With Errors
Input: A prime number p.
Output: false or an integer 0 ≤ s ≤ p − 1.
Assumption: There are only finitely many primes p where either the return
value is false or x 6≡ s mod p.
3. Types of bad primes for modular algorithms
In this section, we suppose that we are given an algorithm implementing
a construction which applies in any characteristic, together with a set of
input data over the rationals.
We call a prime p good for the given algorithm and input data if the
algorithm applied to the modulo p values of the input returns the reduction
of the characteristic zero result. Otherwise, the prime is called bad. In
what follows, to make the discussion in the subsequent sections more clear,
we specify various types of bad primes and describe their influence on the
design of algorithms.
A prime p is bad of type-1 if the modulo p reduction of the characteristic
zero input is either not defined or for obvious reasons not a valid input for
the algorithm. For example, if the input is a polynomial, type-1 primes
arise from prime factors of a denominator of a coefficient. Type-1 primes
are harmless with regard to rational reconstruction as they can be detected
and, thus, rejected from the beginning, at no additional cost.
For a bad prime p of type-2, it turns out only in the course of the construction that the computation in characteristic p cannot be carried through.
For example, an algorithm for inverting a matrix will not work for a prime
dividing the determinant. Since, typically, the determinant is not known,
the failure will only turn out eventually. Type-2 primes waste computation
time, but with regard to rational reconstruction they are detected before
the Chinese remainder step and do, thus, not influence the final lifting.
THE USE OF BAD PRIMES IN RATIONAL RECONSTRUCTION
5
Consider an invariant whose characteristic zero value coincides with the
characteristic p value for all good primes p, and suppose that this value is
known a priori. Moreover assume that the algorithm computes this invariant
at some point of the construction. Then a prime p is bad of type-3 if the
value in characteristic p differs from the expected one. Like type-2 primes,
bad primes of type-3 waste computation time for computing modular results
which will then be discarded, but do not influence the final lifting. Examples
of possible invariants are the dimension or the degree or of a variety. Note
that the computation of an invariant for detecting a type-3 prime may be
expensive. Dropping the computation of the invariant in the design of the
algorithm, if possible, will turn a type-3 prime into a prime of different type.
This includes primes of type-5 below.
Now suppose that some invariant associated to the modular output is
computed by the algorithm, and that the a priori unknown characteristic
zero value of this invariant coincides with the characteristic p value for all
good primes p. Then a prime is bad of type-4 if this invariant does not
have the correct value. Such a prime cannot be detected a priori. However,
if there are only finitely many such primes, they can be eliminated with
arbitrarily high probability by a majority vote over several primes. Type-4
bad primes may occur, for example, in modular Gröbner basis computations,
where we use the leading ideal as an invariant for voting. Handling type4 primes is expensive not only since we waste computation time, but also
since we have to store the modular results for the majority vote. Again,
these primes are eventually detected and, hence do not enter the Chinese
remainder step.
Bad primes other than those discussed so far are called bad primes of
type-5. This includes primes which cannot be discovered by any known
means without knowledge of the characteristic zero result. Example 5.12
below shows that type-5 bad primes indeed occur in algebraic geometry.
Type-5 bad primes enter the Chinese remainder step and are, thus, present
during the final lifting process. Considering Algorithm 3, calling black box
Algorithm 4 instead of Algorithm 1, we will be in a situation where always
either Algorithm 2 or the comparison x = y will return false. As a result,
the algorithm will not terminate.
Due to their nature, bad primes create hardly ever a practical problem.
Typically, there are only very few bad primes for a given instance, and these
will not be encountered if the primes used are chosen at random. On the
other hand, for some of the modern algorithms in commutative algebra, we
have no theoretical argument eliminating type-5 bad primes. Hence, we need
error tolerant reconstruction, which ensures termination provided there are
only finitely many bad primes.
4. Reconstruction with bad primes
To design our error tolerant reconstruction algorithm, we turn rational
reconstruction into a lattice problem.
To begin with, given an integer N ≥ 2, we define the subset CN ⊆ Z/N Z
of elements applied to which Algorithm 5 below will return a rational number
(and not false). Let CN be the set of all r ∈ Z/N Z such that there are
THE USE OF BAD PRIMES IN RATIONAL RECONSTRUCTION
6
integers u, v ∈ Z with u ≥ 0, v 6= 0, and gcd(u, v) = 1 which satisfy the
following condition:
there is an integer q ≥ 1 with q|N and such that
u2 + v 2 <
N
q2
and u ≡ vr mod
(1)
N
q.
In Lemma 4.2 below, we will prove that the rational number uv = uq
vq ∈ Q
is uniquely determined by condition (1). Hence, we have a well–defined map
ψN : CN → Q.
Note that the image of the Farey map ϕB,N , with B =
jp
k
(N − 1)/2 , is
contained in CN : If r ∈ im(ϕB,N ), then ϕ−1
B,N (r) satisfies Condition (1) with
−1
q = 1. Furthermore, ϕB,N (r) = ψN (r).
Typically, the inclusion im(ϕB,N ) ⊆ CN is strict:
Example 4.1. For N = 2 · 13, we have B = 3, hence
im(ϕB,N ) = 0, 1, 2, 3, 8, 9, 17, 18, 23, 24, 25 ,
and the rational numbers which can be reconstructed by Algorithm 2 are
the elements of
1 2
FB ∩ QN = 0, ±1, ±2, ±3, ± , ±
.
3 3
On the other hand,
CN = {r | 0 ≤ r ≤ 25, r 6= 5, 21} ,
and Algorithm 5 will reconstruct the rational numbers in
1 1 2 4
.
ψN (CN ) = 0, ±1, ±2, ±3, ±4, ± , ± , ± , ±
2 3 3 3
Note that the denominator of 21 = ψN (7) = ψN (20) is not coprime to N . In
both cases, q = 2: We have 1 ≡ 2 · 7 mod 13 and 1 ≡ 2 · 20 mod 13.
Now, fix 0 ≤ r ≤ N − 1 such that r ∈ CN , and consider the lattice
Λ = ΛN,r := h(N, 0), (r, 1)i of discriminant N . Let u, v, q correspond to r as
in Condition (1). Then (uq, vq) ∈ ΛN,r .
Lemma 4.2. With notation as above, all (x, y) ∈ Λ with x2 + y 2 < N are
collinear. That is, they define the same rational number xy .
Proof. Let λ = (x, y), µ = (c, d) ∈ Λ be vectors with x2 + y 2 , c2 + d2 < N .
Then yµ − dλ = (yc − xd, 0) ∈ Λ, so N |(yc − xd). Since |yc − xd| < N by
Cauchy–Schwarz, we get yc = xd, as claimed.
Next, consider integers N ′ , M ≥ 2, with gcd(M, N ′ ) = 1, and such that
N = N ′ M . Let a ≥ 0, b 6= 0 be integers such that gcd(b, N ′ ) = 1, and let
a ≡ bs mod N ′ , with 0 ≤ s ≤ N ′ − 1. Let 0 ≤ t ≤ M − 1 be another integer,
and let 0 ≤ r ≤ N − 1 be the Chinese remainder lift satisfying r ≡ s mod N ′
and r ≡ t mod M . In practical applications, we think of N ′ and M as
the product of good and bad primes, respectively. By the following lemma,
Algorithm 5 below applied to N and r will return a/b independently of the
possibly “wrong result” t, provided that M ≪ N ′ .
THE USE OF BAD PRIMES IN RATIONAL RECONSTRUCTION
7
Lemma 4.3. With notation as above, suppose that (a2 + b2 )M < N ′ . Then,
for all (x, y) ∈ Λ = h(N, 0), (r, 1)i with (x2 + y 2 ) < N , we have xy = ab .
Furthermore, if gcd(a, b) = 1 and (x, y) is a shortest nonzero vector in Λ,
we also have gcd(x, y)|M .
Proof. From a ≡ bs mod N ′ , we get a − bs = k1 N ′ for some k1 . Moreover,
s ≡ r mod N ′ gives r = s+k2 N ′ . Now (aM, bM )−bM (r, 1) = (aM −brM, 0)
and aM − brM = M (a − br) = M (a − b(s + k2 N ′ )) = M (a − bs) − k2 bN =
k1 N − k2 bN , thus (aM, bM ) ∈ Λ. Since (a2 + b2 )M < N ′ , Lemma 4.2 gives
a
aM
x
2
2
b = bM = y for all (x, y) ∈ Λ such that (x + y ) < N .
For the second statement, write A := (aM, bM ) and X := (x, y). By
Lemma 4.2, there is a λ = st ∈ Q, with gcd(s, t) = 1, and such that λX = A.
The Euclidean Algorithm gives integers e, f with et + sf = 1, hence
X
X
= (et + sf ) = eX + f A ∈ Λ.
t
t
Since X is a shortest vector in the lattice, it follows that t = ±1, hence
A = ±sX. Since gcd(a, b) = 1, we conclude that gcd(x, y)|M .
The use of Lemma 4.3 is twofold. From a theoretical point of view, it
allows us to ignore type-5 bad primes in the design of modular algorithms
– as long as there are only finitely many of them. This makes the design of
modular algorithms much simpler. From a practical point of view, it allows
us to avoid expensive computations of invariants to eliminate bad primes
of any type. Moreover, factorizing the gcd of the components of a shortest
lattice element can help us to identify bad primes (see Example 4.5 below).
Lemma 4.3 yields the correctness of both the new lifting Algorithm 5
and the resulting reconstruction Algorithm 3, calling black box Algorithm
4 instead of Algorithm 1, and lifting Algorithm 5 instead of Algorithm 2.
In applications, the termination can be based either on the knowledge of
a priori bounds on the height of xy or on an a posteriori verification of the
result. It should be mentioned that both methods are used: some problems
allow for easy verification, while others yield good bounds.
Remark 4.4. Algorithm 5, which is just a special case of Gaussian reduction,
will always find a shortest vector in the lattice generated by (N, 0) and (r, 1).
Moreover, bi 6= 0 for all i > 0 since in every step the vector (ai , bi ) gets
shorter and, hence, cannot be equal to (N, 0).
Even though Algorithm 5 looks more complicated than Algorithm 2, the
bit–complexity of both algorithms is the same: O(log2 N ). See the discussion
in [Nguyen el al. 2009, Section 3.3].
Example 4.5. We reconstruct the rational number
13
12
using the modulus
N = 38885 = 5 · 7 · 11 · 101.
With notation as above, a = 13, b = 12, r = 22684, and the Farey bound is
k
jp
(N − 1)/2 = 139.
B=
THE USE OF BAD PRIMES IN RATIONAL RECONSTRUCTION
8
Algorithm 5 Error Tolerant Lifting
Input: Integers N ≥ 2 and 0 ≤ r ≤ N − 1.
Output: ψN (r) if r ∈ CN and false otherwise.
1: (a0 , b0 ) := (N, 0), (a1 , b1 ) := (r, 1), i := −1
2: repeat
3:
i =i+1
4:
set
h(ai , bi ), (ai+1 , bi+1 )i
qi =
k(ai+1 , bi+1 )k2
5:
6:
7:
8:
9:
10:
set
(ai+2 , bi+2 ) = (ai , bi ) − qi (ai+1 , bi+1 )
until
+
≥ a2i+1 + b2i+1
2
2
if ai+1 + bi+1 < N then
i+1
return abi+1
else
return false
a2i+2
b2i+2
Algorithm 2 applied to this data will correctly return
Algorithm 5 which generates the sequence
13
12 .
Similarly for
(38885, 0) = 2 · (22684, 1) + (−6483, −2),
(22684, 1) = −3 · (−6483, −2) + (3235, −5),
(−6483, −2) = 2 · (3235, −5) + (−13, −12),
(3235, −5) = −134 · (−13, −12) + (1493, −1613).
Now, bad primes will enter the picture. Consider the Chinese remainder
isomorphism
χ : Z/5Z × Z/7Z × Z/11Z × Z/101Z → Z/38885Z.
The preimage of r =
13
12 N
is
χ−1 (r) = (4, 4, 2, 60).
That is, r is the solution to the simultaneous congruences
x ≡ 4 mod 5
x ≡ 4 mod 7
x ≡ 2 mod 11
x ≡ 60 mod 101.
If we make 101 a bad prime by changing the congruence x ≡ 60 mod 101 to
x ≡ 61 mod 101, we obtain
χ(4, 4, 2, 61) = 16524.
THE USE OF BAD PRIMES IN RATIONAL RECONSTRUCTION
9
Algorithm 5 then computes
(38885, 0) = 2 · (16524, 1) + (5837, −2),
(16524, 1) = 3 · (5837, −2) + (−987, 7),
(5837, −2) = 6 · (−987, 7) + (−85, 40),
(−987, 7) = 10 · (−85, 40) + (−137, 393).
85
13
Hence the output −40
= 17
8 6= 12 is not the desired lift. The reason for this
is that 101 is not small enough compared to its cofactor in N . Algorithm
2, on the other hand, returns false since the reduction process will also
terminate with (85, −40) and these numbers are not coprime. Note that
Algorithm 6 in Section 5 below will detect an incorrect lift either by the
procedure pTest (with a very high probability) or the subsequent verification step over the rationals (carried through only if pTest returns true).
As a consequence, in both cases, the set of primes will be enlarged (without
discarding previous results). Eventually, the good primes will outweigh the
bad ones and Algorithm 5, when called from Algorithm 6, will return the
correct lift.
For example, replace the congruence x ≡ 4 mod 7 by x ≡ 2 mod 7, so that
χ(4, 2, 2, 60) = 464.
Then Algorithm 5 yields
(38885, 0) = 84 · (464, 1) + (−91, −84),
(464, 1) = −3 · (−91, −84) + (191, −251),
and terminates with the correct lift
13
91
= .
84
12
Algorithm 2, on the other hand, will again return false since the reduction
also terminates with the numbers (91, 84) which are not coprime.
Since
(132 + 122 ) · 7 < 5 · 11 · 101,
Lemma 4.3 shows that 7 is small enough compared to its cofactor in N .
Hence, the wrong result 2 modulo the bad prime 7 does not influence the
result of the lift. In fact, all other possible congruences modulo 7 will
lead to the same output. Note that the bad prime can be detected as
gcd(91, 84, N ) = 7. Furthermore, note that in the example the lifting process involving the bad prime requires fewer steps than the process relying
on good primes only.
5. A setup for applications in algebra and geometry
In this section, we discuss a general computational setup for applications in commutative algebra and algebraic geometry which requires error tolerance. A setup of this type occurs, for example, when computing
normalization or when computing adjoint curves. See [Böhm et al. 2013a,
Böhm et al. 2013b] and Example 5.11 below.
To begin, fix a global monomial ordering > on the monoid of monomials in
the variables X = {X1 , . . . , Xn }. Consider the polynomial rings W = Q[X]
THE USE OF BAD PRIMES IN RATIONAL RECONSTRUCTION
10
and, given an integer N ≥ 2, WN = (Z/N Z)[X]. If T ⊆ W or T ⊆ WN
is a set of polynomials, then denote by LM(T ) := {LM(f ) | f ∈ T } its set
of leading monomials. If f ∈ W is a polynomial such that N is coprime to
any denominator of a coefficient of f , then its reduction modulo N is the
polynomial fN ∈ WN obtained by mapping each coefficient x of f to xN as
described in Section 2. If H = {h1 , . . . , hs } ⊂ W is a Gröbner basis such that
N is coprime to any denominator in any hi , set HN = {(h1 )N , . . . , (hs )N }.
If J ⊆ W is any ideal, its reduction modulo N is the ideal
JN = hfN | f ∈ J ∩ Z[X]i ⊆ WN .
Notation: From now on, let I ⊂ W be a fixed ideal.
Remark 5.1. For practical purposes, I is given by a set of generators. Fix
one such set f1 , . . . , fr . Then we realize the reduction of I modulo a prime
p via the following equality which holds for all but finitely many primes p:
Ip = h(f1 )p , ..., (fr )p i ⊆ Wp .
More precisely, when running the modular Algorithm 6 described below, we
incorporate the following: if one of the (fi )p is not defined (that is, p is
bad of type-1 for the given set of generators), we reject the prime1. If all
(fi )p are defined, we work with the ideal on the right hand side instead of
Ip . Note that is possible to detect primes with Ip 6= h(f1 )p , ..., (fr )p i (which
are hence of type-3). Indeed, Ip can be found using Gröbner bases (see
[Adams and Loustaunau 1994, Cor. 4.4.5] and [Arnold 2003, Lem. 6.1]).
However, we suggest to skip this computation: finitely many bad primes
will not influence the result if we use error tolerant rational reconstruction
as in Algorithm 5.
To simplify our presentation in what follows, we will systematically ignore the primes discussed in Remark 5.1. We suppose that we are given a
construction which associates to I a uniquely determined ideal U (0) ⊆ W ,
and to each reduction Ip , with p a prime number, a uniquely determined
ideal U (p) ⊆ Wp , where we make the following assumption:
Assumption: We ask that U (0)p = U (p) for all but finitely many p.
We write G(0) for the uniquely determined reduced Gröbner basis of U (0),
and G(p) for that of U (p). In the applications we have in mind, we wish
to construct the unknown ideal U (0) from a collection of its characteristic
p counterparts U (p). Technically, given a finite set of primes P, we wish to
construct G(0) by computing the G(p), p ∈ P, and lifting the G(p) coefficientwise to characteristic zero. Here, to identify Gröbner basis elements
corresponding to each other, we require that LM(G(p)) = LM(G(q)) for all
p, q ∈ P. This leads to condition (1b) below:
Definition 5.2. With notation as above, we define:
(1) A prime number p is called lucky if the following hold:
(a) U (0)p = U (p) and
(b) LM(G(0)) = LM(G(p)).
1Note that rescaling to integer coefficients is not helpful: reducing the rescaled generators may yield the wrong leading ideal. See Remark 5.3.
THE USE OF BAD PRIMES IN RATIONAL RECONSTRUCTION
Otherwise p is called unlucky.
(2) If P is a finite set of primes, set
Y
N′ =
p and M =
p∈P lucky
Y
11
p.
p∈P unlucky
Then P is called sufficiently large if
N ′ > (a2 + b2 ) · M
for all coefficients
a
b
of polynomials in G(0) (assume gcd(a, b) = 1).
Note that a prime p violating condition (1a) is of type-5, while (1b) is a
type-4 condition.
Remark 5.3. A modular algorithm for the fundamental task of computing
Gröbner bases is presented in [Arnold 2003] and [Idrees et al. 2011]. In
contrast to our situation here, where we wish to find the ideal U (0) by
computing its reduced Gröbner basis G(0), Arnold’s algorithm starts from an
ideal which is already given. If p is a prime number, J ⊆ W is an ideal, H(0)
is the reduced Gröbner basis of J, and H(p) is the reduced Gröbner basis of
Jp , then p is lucky for J in the sense of Arnold if LM(H(0)) = LM(H(p)).
It is shown in [Arnold 2003, Thm. 5.12 and 6.2] that if J is homogeneous
and p is lucky for J in this sense, then H(0)p is well–defined and equal
to H(p). Furthermore, by [Arnold 2003, Cor. 5.4 and Thm. 5.13], all but
finitely many primes are Arnold–lucky for a homogeneous J. Using weighted
homogenization as in the proof of [Idrees et al. 2011, Thm. 2.4], one shows
that these results also hold true in the non–homogeneous setup.
∂f
5
11
9
Example 5.4. Consider the ideal J = h ∂f
∂x , ∂y i, where f = x + y + xy +
3
9
x y . The leading terms of the lexicographical Gröbner basis with integral
coprime coefficients are as follows:
264627y 39 + . . . ,
12103947791971846719838321886393392913750065060875xy 8 − . . . ,
40754032969602177507873137664624218564815033875x4 + . . . .
Hence, the Arnold unlucky primes for J are
3, 5, 11, 809, 65179, 531264751, 431051934846786628615463393.
With respect to our notion of lucky as introduced in Definition 5.2.(1),
we first observe:
Lemma 5.5. The set of unlucky primes is finite.
Proof. By our general assumption, U (0)p = U (p) for all but finitely many
primes p. Given a prime p such that U (0)p = U (p), we have LM(G(0)) =
LM(G(p)) if p does not divide the denominator of any coefficient of any
polynomial occurring when testing whether G(0) is a Gröbner basis using
Buchberger’s criterion. The result follows.
The type-5 condition (1a) cannot be checked a priori: We compute G(p)
and, thus, U (p) on our way, but U (0)p is only known to us after G(0) and,
thus, U (0) have been found. However, this is not a problem if we apply
THE USE OF BAD PRIMES IN RATIONAL RECONSTRUCTION
12
error tolerant rational reconstruction: the finitely many bad primes leading
to an ideal U (p) different from U (0)p will not influence the final result:
Lemma 5.6. If P is a sufficiently large set of primes satisfying condition
(1b), then the reduced Gröbner bases G(p), p ∈ P, lift via Algorithm 5 to
the reduced Gröbner basis G(0).
Proof. If a prime p satisfies (1b), then p is Arnold–lucky for U (0). Hence,
as remarked above, G(0)p = G(p). Since P is sufficiently large,
Q by Lemma
4.3, the coefficients of the Chinese remainder lift G(N ), N = p∈P , have a
lift to characteristic zero. By the uniqueness statement of Lemma 4.3, this
lift coincides with G(0).
Lemma 5.5 guarantees, in particular, that a sufficiently large set P of
primes satisfying condition (1b) exists. So from a theoretical point of view,
the idea of finding G(0) is now as follows: Consider such a set P, compute
the reduced Gröbner bases G(p), p ∈ P, and lift the results to G(0) as
described above.
From a practical point of view, we face the problem that we cannot test a
priori whether P is sufficiently large and satisfies condition (1b). However,
to remedy the situation, we can proceed in the following randomized way:
First, fix an integer t ≥ 1 and choose a set of t primes P at random.
Second, compute GP = {G(p) | p ∈ P}, and use a majority vote with
respect to the type-4 condition (1b):
deleteByMajorityVote: Define an equivalence relation on P by setting
p ∼ q :⇐⇒ LM(G(p)) = LM(G(q)). Then replace P by an equivalence class
of largest cardinality2 , and change GP accordingly.
Now all Gröbner bases in GP have the same leading ideal. Hence, we can
lift the Gröbner bases in GP to a set of polynomials G ⊆ W . Furthermore,
if P is suffciently large, all primes in P satisfy (1b). However, since we do
not know whether P is sufficiently large, a final verification in characteristic
zero is needed. As this may be expensive, especially if G 6= G(0), we first
perform a test in positive characteristic:
pTest: Randomly choose a prime number p ∈
/ P such that p does not divide
the numerator or denominator of any coefficient occurring in a polynomial
in G or {f1 , . . . , fr }. Return true if Gp = G(p), and false otherwise.
If pTest returns false, then P is not sufficiently large or the extra prime
chosen in pTest is bad. In this case, we enlarge the set P by t primes not
used so far, and repeat the whole process. If pTest returns true, however,
then most likely G = G(0). Hence, it makes sense to verify the result over
the rationals. If the verification fails, we enlarge P and repeat the process.
We summarize this approach in Algorithm 6 (recall that we ignore the
primes from Remark 5.1 in our presentation).
Remark 5.7. If Algorithm 6 requires more than one round of the loop, we
have to use a weighted cardinality count in deleteByMajorityVote:
2When computing the cardinality, take Remark 5.7 into account.
THE USE OF BAD PRIMES IN RATIONAL RECONSTRUCTION
13
Algorithm 6 Reconstruction of an Ideal
Input: An algorithm to compute G(p) from Ip , for each prime p, and a way
of verifying that a computed Gröbner basis equals G(0).
Output: The Gröbner basis G(0).
1: choose a list P of random primes
2: GP = ∅
3: loop
4:
for p ∈ P do
5:
compute G(p) ⊆ Wp
6:
GP = GP ∪ {G(p)}
7:
(P, GP ) = deleteByMajorityVote(P, GP )
8:
lift GP to G ⊆ W via Chinese remaindering and Algorithm 5
9:
if the lifting succeeds and pTest(I, G, P) then
10:
if G = G(0) then
11:
return G
12:
enlarge P with primes not used so far
when enlarging P, the total weight of the elements already present must
be strictly smaller than the total weight of the new elements. Otherwise,
though highly unlikely in practical terms, it may happen that only unlucky
primes are accumulated.
Remark 5.8. Our lifting process works since reduced Gröbner bases are
uniquely determined. In practical terms, however, there is often no need
to reduce the Gröbner bases involved: it is only required that the construction associating the Gröbner bases to I and its reductions yields uniquely
determined results.
Remark 5.9. We may allow that the given construction does not work for
finitely many primes p (which are then bad of type-2). In this case, the
respective primes will be rejected.
Remark 5.10. Depending on the particular implementation of the construction, type-3 primes (in addition to those considered in Remark 5.1) may
occur. In such a situation, it is often cheaper to rely on error tolerance
rather than spending computation time to detect these primes.
Example 5.11. If K is any field, and I ⊆ K[X] is a prime ideal, the normalization A of the domain A = K[X]/I is the integral closure of A in its
field of fractions Q(A). If K is perfect, the normalization algorithm given
in [Greuel et al. 2010] will find a “valid denominator” d ∈ A and an ideal
U ⊆ A such that 1d U = A ⊆ Q(A). In fact, U is uniquely determined
if we fix d. In practical terms, d and U are a polynomial and an ideal in
K[X], respectively. If K = Q, we can apply the modular version of the algorithm (see [Böhm et al. 2013a]). This version relies on choosing a “universal
denominator” d which is used over the rationals as well as in finite characteristic. Given a prime number p, it may then happen that Ip is not a prime
ideal (a type-2 condition), that the leading ideal of Ip does not coincide with
that of I (a type-3 condition), that dp is not defined (a type-1 condition),
or that dp is not a valid denominator (a type-2 condition). In accordance
THE USE OF BAD PRIMES IN RATIONAL RECONSTRUCTION
14
with the general setup, the numerator ideal U (p) is obtained by computing
the Gröbner basis G(p), and LM(G(0)) = LM(G(p)) and U (0)p = U (p) are
type-4 and type-5 conditions, respectively.
The normalization algorithm mentioned in the previous Example 5.11
finds A by successively enlarging A in form of an endomorphism ring of a so
called test ideal J ⊆ A. For practical purposes, the radical of the Jacobian
ideal is used for J. The following example shows that, for the algorithm
computing the radical of the Jacobian, bad primes p exist which satisfy the
type-4 condition (1b) but violate the type-5 condition (1a) in Definition 5.2.
In particular, these primes cannot be eliminated by a majority vote on the
leading ideal.
Example 5.12. We construct a sextic curve C = V (I) ⊂ P2C , given by an
ideal I = hf i ⊂ Q[x, y, z], such that 5 is a bad prime of type-5 for computing
the radical of the singular locus of C. The basic idea is to construct a
curve which has two double points in characteristic 0, which coincide when
reducing modulo 5, while one additional double point appears.
For
L = hy, x − 4zi2 ∩ hy, x + 6zi2 ⊆ Q[x, y, z]
the reduced Gröbner basis with respect to the degree reverse lexicographical
ordering is
G = y 2 , (x − 4z)(x + 6z)y, (x − 4z)2 (x + 6z)2 .
Note, that both L and
L5 = y 2 , (x + z)2 y, (x + z)4 ⊆ F5 [x, y, z]
have the same leading monoid y 2 , x2 y, x4 .
Writing generators of
M = L5 ∩ hy, x − zi2 ⊂ L5
in terms of the Gröbner basis of L5 yields the representation
M = y 2 , (x − z) · (x + z)2 y, (x2 + 3xz + z 2 ) · (x + z)4 .
Now consider a generic homogeneous element of degree 6 in
G1 , (x − z) · G2 , (x2 + 3xz + z 2 ) · G3 ⊂ Q[x, y, z].
For practical puposes, a random element will do, for example,
f = x6 + y 6 + 7x5 z + x3 y 2 z − 31x4 z 2 − 224x3 z 3 + 244x2 z 4 + 1632xz 5 + 576z 6 .
For the ideal I = hf i the prime 5 is bad of type-5 with respect to the
algorithm
p
I 7→ I + Jac(I),
where Jac(I) denotes the Jacobian ideal of I: First note, that no coefficient
of f is divisible by 5. In particular, LM(I) = x6 = LM(I5 ), so 5 is
Arnold–lucky for I. We compute
p
U (0) = I + Jac(I) = hy, x − 4zi ∩ hy, x + 6zi ,
p
U (5) = I5 + Jac(I5 ) = y, x2 − z 2 = hy, x − zi ∩ hy, x + zi ,
and
U (0)5 = y, (x + z)2 .
THE USE OF BAD PRIMES IN RATIONAL RECONSTRUCTION
Hence
15
LM(U (0)) = y, x2 = LM(U (5)),
but
U (0)5 6= U (5).
Acknowledgements. We would like to thank the referees who made valuable
suggestions to improve the presentation of this paper.
References
[Adams and Loustaunau 1994] Adams, W. W.; Loustaunau, P.: An introduction to
Gröbner bases, Graduate Studies in Mathematics, 3, AMS (1994).
[Arnold 2003] Arnold, E. A.: Modular algorithms for computing Gröbner bases, J. Symb.
Comput. 35, 403–419 (2003).
[Böhm et al. 2013a] Böhm, J.; Decker, W.; Laplagne, S.; Pfister, G.; Steenpaß, A.; Steidel,
S.: Parallel Algorithms for Normalization. J. Symb. Comp. 51, 99–114 (2013).
[Böhm et al. 2013b] Böhm, J.; Decker, W.; Laplagne, S.; Seelisch, F.: Local to global
algorithms for the Gorenstein adjoint ideal of a curve. In preparation.
[Collins et al. 1994] Collins, George E., Encarnación, Mark J.: Efficient Rational Number
Reconstruction. J. Symb. Comput. 20, 287–297 (1995).
[Encarnación 1995] Encarnación, Mark J.: Computing GCDs of polynomials over algebraic number fields. J. Symb. Comput. 20, 299–313 (1995).
[Greuel et al. 2010] Greuel, G.–M.; Laplagne, S.; Seelisch. F.: Normalization of rings. J.
Symb. Comput. 45, 887–901 (2010).
[Idrees et al. 2011] Idrees, N.; Pfister, G.; Steidel, S.: Parallelization of Modular Algorithms. J. Symb. Comput. 46, 672–684 (2011).
[Nguyen el al. 2009] Nguyen, P.Q.; Stehlé, D.: Low–dimensional lattice basis reduction
revisited. ACM Transactions on Algorithms, Paper 46 (2009).
[Kornerup and Gregory 1983] Kornerup, P.; Gregory, R. T.: Mapping Integers and Hensel
Codes onto Farey Fractions. BIT Numerical Mathematics 23(1), 9–20 (1983).
[Wang 1981] Wang, P. S.: A p–adic algorithm for univariate partial fractions. Proceedings
SYMSAC ’81, 212–217 (1981).
[Wang et al. 1982] Wang, P. S.; Guy, M. J. T.; Davenport, J. H.: P–adic reconstruction
of rational numbers. SIGSAM Bull, 2–3 (1982).
Fachbereich Mathematik, Technical University Kaiserslautern, Postfach
3049, 67653 Kaiserslautern, Germany
E-mail address: [email protected]
Fachbereich Mathematik, Technical University Kaiserslautern, Postfach
3049, 67653 Kaiserslautern, Germany
E-mail address: [email protected]
Fachbereich Mathematik, Technical University Kaiserslautern, Postfach
3049, 67653 Kaiserslautern, Germany
E-mail address: [email protected]
Fachbereich Mathematik, Technical University Kaiserslautern, Postfach
3049, 67653 Kaiserslautern, Germany
E-mail address: [email protected]
| 0math.AC
|
Information Bottleneck in Control Tasks with
Recurrent Spiking Neural Networks
Madhavun Candadai Vasu and Eduardo J. Izquierdo
arXiv:1706.01831v1 [cs.NE] 6 Jun 2017
Cognitive Science, School of Informatics and Computing
Indiana University, Bloomington, U.S.A.
{madcanda,edizquie}@indiana.edu
Abstract. The nervous system encodes continuous information from
the environment in the form of discrete spikes, and then decodes these
to produce smooth motor actions. Understanding how spikes integrate,
represent, and process information to produce behavior is one of the
greatest challenges in neuroscience. Information theory has the potential to help us address this challenge. Informational analyses of deep and
feed-forward artificial neural networks solving static input-output tasks,
have led to the proposal of the Information Bottleneck principle, which
states that deeper layers encode more relevant yet minimal information
about the inputs. Such an analyses on networks that are recurrent, spiking, and perform control tasks is relatively unexplored. Here, we present
results from a Mutual Information analysis of a recurrent spiking neural
network that was evolved to perform the classic pole-balancing task. Our
results show that these networks deviate from the Information Bottleneck
principle prescribed for feed-forward networks.
Keywords: Spiking neurons, Evolutionary neural networks, Recurrent
networks, Information theory, Information Bottleneck
1
Introduction
Deep Learning systems have surpassed other algorithms and even humans at
several tasks [1,2,3,4]. While their applications continue to grow, deep learning
systems are still considered black-box optimization methods. One of the most
vital features behind their success is their ability to extract relevant yet minimal
information as it progresses into deeper and deeper layers [5]. This is an extension
of Rate Coding Theory [6] presented as the Information Bottleneck principle [5,7].
The information bottleneck principle has been primarily focused on systems that
are (a) feedforward, and (b) in an open-loop, decoupled from their environment.
Neuroscientists, on the other hand, have long been studying the principles
behind encoding and representation of environmental information in neural activity using principles of information theory [9] and rate distortion theory [8].
Continuous variables from the environment are encoded as discrete spikes in the
brain, which are then decoded to produce smooth continuous movement. Due to
experimental limitations, an informational analysis of a closed-loop brain-bodyenvironment behaviour system is not yet feasible.
2
Information Bottleneck in Spiking RNNs: Candadai Vasu et. al.
We take a radically different approach to understanding information flow in a
behaviorally-functional agent. We artificially evolve embodied agents controlled
by recurrent spiking neural networks to perform a task. For this paper, we focus
on a non-Markovian version of the classical pole balancing task. Pole balancing
has been explored quite extensively as a benchmark for control using neural
networks [10,11]. With continuous states and actions, this task serves as an ideal
setting to study the transformation of the continuous input information into
spikes and then back to continuous action. While the typical task is Markovian,
and thus too trivial for an informational analysis, it can easily be transformed
into a non-Markovian task by making the available information to the agent
limited. Our approach to pole balancing incorporates an agent-centric “visual”
interface to the pole. Therefore, information that is typically available, like the
pole’s angle and angular velocity, the agent’s position and velocity, are not fed
directly to the network. Ultimately, the minimal nature of the task makes it
tractable for an investigation of a recurrent network in a closed-loop task.
The parameters of the recurrent spiking neural network that balances the
pole were optimized using an evolutionary algorithm. Evolving neural networks
as opposed to hand-designing them allows maximum flexibility for exploring
the parameter space. While evolutionary algorithms have been very commonly
used in several fields [16], recently, they have been proven to be efficient for
optimizing deep neural networks as well [14,15]. Moreover, due to the stochastic
nature of the optimization, running the algorithm several times provides us with
an ensemble of solutions that solve the problem. This allows the analysis of not
just one solution but several to evaluate consistency of results.
The paper is organized as follows. In the first section we report on the agent,
task, optimization technique, and analysis method. The section that follows
presents an informational analysis for the best and top performing agents. In
the last section we summarize the results.
2
Methods
Agent design. The agent lives in a 1-dimensional world with the pole attached
to its center. Seven equidistant rays of “vision” with overlapping receptive fields
spanning 36◦ provide it with sensory stimuli (Fig. 1A,B). The control network
of the agent has three primary components: sensory units, spiking interneurons,
and motor units. There is one sensory units per ray, which merely pass on the
signal received from the rays. Sensory units are fully connected to N interneurons
(here N = 2), modeled by Izhikevich spiking neuron model [13]. The model has
4 parameters per neuron and is governed by a two-dimensional system of ordinary differential equations [13]. Interneurons are recurrently connected (Fig. 1C).
Therefore, each interneuron receives weighted input from each sensory unit, Si ,
and from other spiking interneurons, Ii , as follows:
Si + Ii =
7
X
j=1
s
wji
sj +
N
X
j=1
i
wji
oi
(1)
Information Bottleneck in Spiking RNNs
B
Sensory response
A
3
C
1
0.5
0
-15
0
15
Fig. 1. Task set up and agent design. [A] The agent (gray circle) moves left/right along
the horizontal axis (gray arrows) sensing the pole (black rod) through the seven vision
rays (color dashed lines), with a range of 36 degrees. [B] Sensory rays have a linearly
diffused receptive field from their centers and overlap at the edges with the receptive
fields of adjacent rays. [C] The agent has 7 vision sensors (black) connected to two
recurrent spiking interneurons (red) connected to two motor units (gray).
s
where sj is the input at the j th sensory unit, wji
is the strength of the connection
th
th
i
from the j sensory unit to the i spiking interneuron, wji
is the strength of
th
th
the recurrent connections from the j to the i spiking neuron, and oi is the
output of the neuron. The sign of all outgoing connections from an interneuron
depends on its excitatory or inhibitory nature, as identified by a binary parameter. Finally, the layer of interneurons feeds into the two motor neurons, that has
the following state equation:
τm ṁi = −mi +
N
X
m
wji
ōj
i = 1, 2
(2)
j=1
ōj (t) =
hj
1 X
oj (t − k)
hj
(3)
k=0
m
where mi represents the motor neuron state, wji
is the strength of the connection
th
th
from the j spiking interneuron to the i motor neuron, ōj represents the firing
rate code, the moving average over a window of length hj for the output of spiking
interneuron j. Finally, the difference in output between the motor neurons results
in a net force that acts upon the agent, moving it along the track. The network
was simulated using Euler integration with step size 0.01.
Pole Balancing Task Design. The agent can move left/right to balance a
pole for as long as possible. The pole-balancing task was implemented based on
published descriptions [12]. The force from the agent, computed as the difference
between motor unit outputs, affects the angular acceleration of the pole and
acceleration of the agent. The physical parameters such as mass, length and
coefficient of friction were all set as per the same reference. While typically polebalancers receive as input the angle of the pole (θ), its angular velocity (ω), the
agent’s position (x) and velocity (v), our implementation was modified to only
sense the pole through the sensory rays.
4
Information Bottleneck in Spiking RNNs: Candadai Vasu et. al.
Evolutionary Algorithm. The network was optimized using a real valued evolutionary algorithm. A solution was encoded by 38 parameters, including the
intrinsic parameters of the Izhikevich neurons, the neuron-specific size of the
window for estimating rate code, all connection weights (sensory-interneuron,
interneuron-interneuron, interneuron-motor), and the time constant, bias and
gain for the motor units. Parameters were constrained to certain ranges: connection strengths ∈ [−50, 50]; motor unit biases ∈ [−4, 4]; time-constants ∈ [1, 2].
The range of intrinsic parameters and the polarity of the outgoing weights from
the inter-neuron depended on a binary inhibitory/excitatory neuron flag parameter in the genotype [13]. The population consisted of 100 individuals.
Fitness Function. Performance was estimated by averaging over 16 trials,
starting at pole angles θ0 between ±12◦ , in increments of 3◦ , and two initial
angular
velocities, ω0 = ±0.001. The fitness function to be maximized was f =
PT
( t=1 cos(θt ))/T , where T = 500s is the maximum duration of the run. The pole
was considered balanced if it was within the sensory range of the agent. Also,
the track length was fixed at 45 units, beyond which the trial was terminated.
Mutual Information. The amount of information contained in one variable
about another was estimated using Mutual Information (MI). We quantified the
information that neurons contain about pole angle (θ), angular velocity (ω),
agent position (x) and agent velocity (v) by calculating their probability distributions (using a binning method with each bin of width 0.01):
M I(N, X) =
X X
n∈N x∈X
3
p(x, n)log
p(x, n)
p(x)p(n)
(4)
Results
Performance of evolutionary optimization. While pole balancing is a well-known
benchmark, it was also a relatively easy task to optimize. The evolutionary
algorithm found fairly good solutions (around 75% performance) at the very
first random draw of the population. Fig. 2A shows the performance of the
best agent in every generation over time. Out of the 100 evolutionary runs, 99
converged to over 99% fitness with only two spiking interneurons.
Network structure and performance of the best agent. The network structure
of the best agent from the 100 runs is shown in Fig. 2B. The behavioral traces
of this agent on the 16 trials specified in the previous section are shown in
Fig. 3A. This agent achieved a fitness of 99.4%. To test for generalization, its
performance was evaluated on a finer and broader range of conditions, postoptimization (Fig. 2C): initial pole angle θ0 in the range [−45, 45] and initial
pole angular velocity ω0 in the range [−0.01, 0.01]. As can be seen, the agent
generalizes well within and outside the region it was evolved on. Note that θ0
that are beyond 18◦ on either side are beyond the range of the sensory rays.
Encoding of environmental variables. The different network elements that
manipulate the input are shown in Fig. 3B. Sensory signals first act on neurons’
potential. The neuron fires based on dynamics in its potential, which is then
Information Bottleneck in Spiking RNNs
A
S1 S2 S3 S4 S5
B
1
S6 S7
C 0.01
5
1
40
Fitness
5
0.9
0
15
0.8
0
0.5
50
0.6
N1
N2
-20
20
0.8
0.005
20
10
0
0.4
-0.005
0.2
-40
1
100
Generation #
1.5
150
-0.01
M1
M2
-20
0
20
0
Fig. 2. Optimization performance. [A] Fitness of the best individual in the population
vs. generations for 100 evolutionary runs. Best run in blue, top 10 in light blue and the
rest in gray. [B] Network structure of the best agent. The width of the edges indicate
the magnitude of the weights and are also color coded for polarity. The sensory units
in black are identified as S1-S7, spiking interneurons N1,N2 in red and motor units
M1,M2 in grey. [C] Generalization performance over a broader and finely grained set
of initial conditions that were tested during evolution (marked by cross marks).
interpreted by its rate. The motor units then convert this discrete spike rate
to smooth continuous movement. Note that a single neuron has three levels of
informational content - continuous valued potential, binary spikes, and discrete
spike rate codes. Fig. 3C shows traces for the highlighted trial in Fig. 3A of
the best agent. Although θ is the most directly available information, unlike the
standard practice of directly providing it, the sensory rays provide an agentcentric perspective of θ. MI between each of the network elements with θ, see
Fig. 3D, revealed that internal potential of neurons have relatively more information about θ than the spike rate and so does the motor units. Albeit only for
one trial in one agent, this shows that the bottleneck does not always become
narrow in control tasks. MI also reveals that indirect encoding of ω, x and v all
happen in the very first stage of the network, neuron potential (black bars in
Fig. 4). This can be attributed to the recurrent nature of connections between
the interneurons and also their rather complex non-linear internal dynamics.
Analysis of the Information Bottleneck. All available information, as shown
by MI in V , is not necessarily used in controlling movement, as shown by relatively lower information in R. To further study the bottleneck, we compared
the amount of information contained in neuron potentials, V , versus the rate
coded outputs of the neurons, R. Note that the spikes themselves do not have
any information about the environment but, in fact, encode them in its rate.
For each of the environmental variables a paired samples t-test was conducted
with a significance threshold of p < 0.05. This revealed that there is a significant
drop in the amount of information between V and R (Fig. 4) robustly across the
top ten agents. This can be attributed to the loss due to the discretization of
information available in V as spikes. However, the information in R is sufficient
to perform the behavior with great accuracy and so this is in fact an efficient
encoding of information. The minimal yet relevant information that is encoded
in R is interpreted by the motor units. They integrate R from the interneurons
6
Information Bottleneck in Spiking RNNs: Candadai Vasu et. al.
A
E
1
I(V2, )
0.8
0.6
0.4
0.2
0
Trial-by-trial
Across trials
0
0.5
1
I(V1, )
B
Sensory Input
C
D
1
0.5
Mutual Information
Information about
Mutual
with 𝜃
0
0
50
100
150
200
250
0
0.5
1
50
Neuron Potential
(V1,V2)
Neuron Spikes
(S1,S2)
Mutual Information
0
0
-50
0
50
100
150
200
250
0
50
100
150
200
250
0
50
100
150
200
250
0
50
100
150
200
250
0.5
1
0.8
0.6
0.4
0.2
0.04
Neuron Rate
Codes (R1,R2)
0.02
0
0.15
Motor Unit States
(M1,M2)
0.1
0.05
0
Fig. 3. Behavior of the best agent and one of its trials in detail. [A] Angle of the pole
over time on 16 trials. One of the trials (θ0 = −6◦ and ω0 = 0.001) is highlighted
and explored further. [B] Information flowchart: from sensory input rays through the
spiking inter-neuron layer, composed of 3 levels of processing (neuron potential, spiking
activity and rate code), and then the motor units. [C] A sample trace for each of the
components corresponding to each box in B from the trial highlighted in A. Each color
in the first figure is matched with the sensory rays in Fig. 1. The blue and orange colors
indicate interneurons 1 and 2 respectively. The green and brown traces corresponds to
the left and right motor neurons respectively. [D] Mutual Information about the pole
angle (θ) for the highlighted trial for each of the components. [E] Mutual information
about θ in the neuron potential (V ). A comparison between the trial-by-trial MI vs.
the total MI across trials.
Neuron Potential (V)
Spike Rate Code (R)
Motor Unit States (M)
Mutual Information
Information Bottleneck in Spiking RNNs
7
1.5
1
0.5
0
x
v
Fig. 4. MI about the four environmental variables: pole angle (θ), pole angular
velocity(ω), agent position (x) and agent velocity (v) in the network elements: neuron potentials (V ), rate coded outputs of neurons (R), and motor unit states (M ).
Paired samples t-test yielded highly significant (p 0.05) differences between the
information contained about θ, ω and v in V versus R and R versus M .
and their outputs directly impact θ, ω, x and v and so this is where a deviation
from the IB principle is expected. Statistical analyses of the MI between R and
the motor units state, M , using the paired samples t-test yielded highly significant (p 0.05) increase in information about all environmental variables in
M (Fig. 4). This shows that the IB for control tasks is not always a filtering of
information but is rather filtering followed by an expansion at the control layer.
Context sensitive information encoding. From previous analysis, we know
that components of the network encode information about the environment. But
what information do they encode? Typically, when a neuron is said to encode
information about a feature of the environment it is thought to be a consistent,
context-independent code. To explore this idea further, we compared the MI the
neuron potential has about the pole’s angle I(V, θ) on a trial by trial basis to
the information that same neuron has across all trials about the same feature
of the environment on the best agent (Fig. 3E) and the top 10 agents. A onesample t-test of the combined MI with the distribution of trial-by-trial MI values
yielded a highly significant difference (p 0.05). This means that the combined
information is significantly lower than trial-by-trial information, and therefore
that encoding is highly context-dependent across all evolved pole-balancers.
4
Discussion
In this paper, we have presented results from an information theoretic analysis
of recurrent spiking neural networks that were evolved to perform a continuous
control task: agent-centric non-Markovian pole balancing. Our results can be
summarized as follows. First, networks with as few as two spiking neurons could
be evolved to perform this task. Second, through the use of MI, we show that the
spiking network encoded environmental variables of interest that were directly
and indirectly available. Third, we show that the information bottleneck from
the neuron potential to its firing rate is an efficient filtering/compression, which
was followed by an increase in information at the control layer on account of
8
Information Bottleneck in Spiking RNNs: Candadai Vasu et. al.
their causal effect on the environment. This is a phenomenon that we expect to
arise in control tasks in general, and plan to explore further with different tasks
and types of networks. Perhaps, this can develop into an optimization method
for neural network control. Finally, we show that the information encoded in the
spiking neurons vary across trials, causing the across-trial combined information
to be significantly lower. This can mean either that the same stimuli are encoded
in different ways (redundancy) or that different stimuli are mapped on to the
same encoding (generalization) or both. This warrants further analysis to understand the encoding in more detail and more interestingly, to understand how
the context helps disambiguate the generalized representations during a trial.
References
1. Bengio, Y., Courville, A., & Vincent, P. (2013). Representation learning: A review
and new perspectives. IEEE transactions on pattern analysis and machine intelligence, 35(8), 1798-1828.
2. Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2012). Imagenet classification with
deep convolutional neural networks. In Advances in neural information processing
systems (pp. 1097-1105).
3. Mnih, V., Kavukcuoglu, K., Silver, D., Graves, A., Antonoglou, I., Wierstra, D.,
& Riedmiller, M. (2013). Playing atari with deep reinforcement learning. arXiv
preprint arXiv:1312.5602.
4. Silver, D., Huang, A., Maddison, C. J., Guez, A., Sifre, L., Van Den Driessche, G.,
... & Dieleman, S. (2016). Mastering the game of Go with deep neural networks and
tree search. Nature, 529(7587), 484-489.
5. Tishby, N., Pereira, F. C., & Bialek, W. (2000). The information bottleneck method.
arXiv preprint physics/0004057.
6. Berger, T. (1971). RateDistortion Theory. Encyclopedia of Telecommunications.
7. Tishby, N., & Zaslavsky, N. (2015, April). Deep learning and the information bottleneck principle. In Information Theory Workshop (ITW), 2015 IEEE (pp. 1-5).
8. Simoncelli, E. P., & Olshausen, B. A. (2001). Natural image statistics and neural
representation. Annual review of neuroscience, 24(1), 1193-1216.
9. Borst, A., & Theunissen, F. E. (1999). Information theory and neural coding. Nature
neuroscience, 2(11), 947-957.
10. Pasemann, F., & Dieckmann, U. (1997). Evolved neurocontrollers for polebalancing. Biological and Artificial Computation: From Neuroscience to Technology,
1279-1287.
11. Onat, A., Nishikawa, Y., & Kita, H. (1998). Q-learning with recurrent neural networks as a controller for the inverted pendulum problem.
12. Barto, A. G., Sutton, R. S., & Anderson, C. W. (1984).Neuron-like adaptive elements that can solve difficult learning control problems. Behavioural Processes.9,89.
13. Izhikevich, E. M. (2003). Simple model of spiking neurons. IEEE Transactions on
neural networks, 14(6), 1569-1572.
14. Miikkulainen, R., Liang, J., Meyerson, E., Rawal, A., Fink, D., Francon,
O., ... & Hodjat, B. (2017). Evolving Deep Neural Networks. arXiv preprint
arXiv:1703.00548.
15. Salimans, T., Ho, J., Chen, X., & Sutskever, I. (2017). Evolution Strategies as a
Scalable Alternative to Reinforcement Learning. arXiv preprint arXiv:1703.03864.
16. De Jong, K. A. (2006). Evolutionary computation: a unified approach. MIT press.
| 7cs.IT
|
Statistical Watermarking for Networked Control Systems
arXiv:1709.08617v1 [math.OC] 25 Sep 2017
Pedro Hespanhol, Matthew Porter, Ram Vasudevan, and Anil Aswani
Abstract— Watermarking can detect sensor attacks in control
systems by injecting a private signal into the control, whereby
attacks are identified by checking the statistics of the sensor
measurements and private signal. However, past approaches assume full state measurements or a centralized controller, which
is not found in networked LTI systems with subcontrollers.
Since generally the entire system is neither controllable nor
observable by a single subcontroller, communication of sensor
measurements is required to ensure closed-loop stability. The
possibility of attacking the communication channel has not been
explicitly considered by previous watermarking schemes, and
requires a new design. In this paper, we derive a statistical watermarking test that can detect both sensor and communication
attacks. A unique (compared to the non-networked case) aspect
of the implementing this test is the state-feedback controller
must be designed so that the closed-loop system is controllable
by each sub-controller, and we provide two approaches to design
such a controller using Heymann’s lemma and a multi-input
generalization of Heymann’s lemma. The usefulness of our
approach is demonstrated with a simulation of detecting attacks
in a platoon of autonomous vehicles. Our test allows each vehicle
to independently detect attacks on both the communication
channel between vehicles and on the sensor measurements.
I. I NTRODUCTION
A major challenge in designing networked control systems
is ensuring resilience of subcontrollers to malicious attacks
[1], [1]–[3]. Unlike systems with a centralized controller
where sensor attacks occur only on the output measurements,
networked control systems are also susceptible to attacks on
the communication channel used to transfer measurements
between subcontrollers. Though cryptography and cybersecurity [4]–[7] can secure communication channels, many
networked control system applications feature rapid reconfiguration of the network and cannot operate in real-time with
the overhead required to establish secure communication
channels. For instance, consider a platoon of autonomous
vehicles where vehicles rapidly enter or exit the platoon.
This paper develops a statistical watermarking approach
for detecting malicious sensor and communication attacks
on networked LTI systems. Our first contribution is to design
a watermarking test using null hypothesis testing [8]–[12],
and this requires characterizing the statistics of states and
private watermarking signals under the dynamics of multiple subcontrollers within the networked system. A unique
This work was supported by the UC Berkeley Center for Long-Term
Cybersecurity, and by Ford Motor Company.
Pedro Hespanhol and Anil Aswani are with the Department of Industrial Engineering and Operations Research, University of California,
Berkeley, CA 94720, USA [email protected],
[email protected]
Matthew Porter and Ram Vasudevan are with the Department of Mechanical Engineering, University of Michigan, Ann Arbor, MI 48109, USA
[email protected], [email protected]
feature (as compared to the non-networked setting) of the
watermarking scheme is it requires the state-feedback control
to be such that the closed-loop system is controllable by each
subcontroller, because otherwise a subcontroller could not
independently verify the lack of an attack. Our second contribution is to provide two approaches to constructing such
a state-feedback control, and this partly involves deriving a
multi-input generalization of Heymann’s lemma [13], [14].
A. Watermarking for LTI Systems
Watermarking been proposed for detecting sensor attacks
in LTI systems [8]–[12], [15]–[17]. The idea is that a random
signal is added to the control, and the statistics of the sensor
measurements and random signal are compared to check for
sensor attacks. Some schemes use null hypothesis testing
[8]–[12], while dynamic watermarking [15]–[17] ensures
only attacks adding a zero-average-power signal to the sensor
measurements can remain undetected. Most watermarking
[8]–[12], [17] is for LTI systems with a centralized controller,
and only the approaches of [15], [16] can be used with
networked LTI systems; however, the approaches [15], [16]
require full state observation, which is not the case for many
systems. Our first contribution is to develop watermarking
for networked LTI systems with partial state observation.
B. Security for Intelligent Transportation Systems
Intelligent transportation systems (ITS) [16], [18]–[23]
may benefit from watermarking. For instance, [16] considered the use of dynamic watermarking to detect sensor
attacks in a network of autonomous vehicles coordinated
by a supervisory controller; the watermarking approach was
successfully able to detect attacks. However, large-scale
deployments of ITS must be resilient in the face of partial
state observations and partially distributed control structures.
For example, vehicle platoons are susceptible to malicious
interference of GPS and the communication channel between
vehicles [24]–[26]. Our third contribution is to conduct
a simulation that shows the efficacy of our watermarking
scheme in detecting attacks on a vehicle platoon.
C. Outline
Sect. II provides a model of the networked LTI system
we consider, and specifies a model for communication and
sensor attacks. Next, Sect. III presents an example to give
intuition about the new challenges with designing watermarking for networked systems. We construct a statistical watermarking test in Sect. IV, which allows each subcontroller to
independently check for the presence of communication or
sensor attacks. Our tests require a state-feedback controller
such that the closed-loop system is controllable by each subcontroller, and Sect. V provides two methods for constructing
such a controller. Last, Sect. VI conducts simulations of an
autonomous vehicle platoon. We show that our approach is
able to detect the presence or absence of sensor attacks and
attacks on the communication channel between vehicles.
II. N ETWORKED LTI S YSTEM AND ATTACK M ODALITIES
A. Dynamics and Measurement Model
We study a setting with κ subcontrollers. The subscripts
i or j denote the i-th or j-th subcontroller, and the subscript
n indicates time. Consider the LTI system with dynamics
Pκ
xn+1 = Axn + i=1 Bi ui,n + wn ,
(1)
where x ∈ Rp is the state, ui ∈ Rqi is the input of the ith subcontroller, and w ∈ Rp is a zero mean i.i.d. process
noise with a jointly Gaussian distribution and covariance
ΣW . Each subcontroller steers a subset of the actuators, and
each subcontroller makes the partial state observations
yi,n = Ci xn + zi,n + vi,n ,
(2)
mi
where yi ∈ R is the observation of the i-th subcontroller,
zi ∈ Rmi is zero mean i.i.d. measurement noise with a jointly
Gaussian distribution and covariance ΣZ,I , and vi ∈ Rmi
should be interpreted as an additive measurement disturbance
that is added by an attacker.
B. Network Communication Model
The LTI system here is networked in the following sense:
The dynamics and partial observations are such that for
B = B1 · · · Bκ
(3)
T
T
T
C = C1 · · · Cκ
(4)
we have that (A, B) is stabilizable and (A, C) is detectable.
In general, (A, Bi ) is not stabilizable for some (or all) i,
and similarly (A, Ci ) is not detectable for some (or all)
i. Thus coordination is required between subcontrollers to
ensure closed-loop stability, and networking arises because
we assume each subcontroller communicates its own partial
state observations to all other subcontrollers. (Our setting
assumes communication has zero cost.) Consider the values
si,j,n = yj,n + νi,j,n ,
(5)
where si,j ∈ Rmj is the value communicated to subcontroller
i of the measurement made by subcontroller j, and νi,j,n ∈
Rmj should be interpreted as an additive communication
disturbance added by an attacker. Clearly νi,i,n ≡ 0 for all i,
since the i-th subcontroller already has its own measurement.
C. Controller and Observer Structure
The idea of statistical watermarking in this context will be
to superimpose a private (and random) excitation signal ei,n
known in value to only the i-th subcontroller but unknown in
value to the attacker or to the other subcontrollers. We will
apply the control input ui,n = Ki x̂i,n +ei,n , where x̂i,n is the
observer-estimated state (the subscript i here indicates that
each subcontroller operates its own observer, and that x̂i,n is
the state estimated by the observer of the i-th subcontroller)
and ei,n are i.i.d. Gaussian with zero mean and constant
variance ΣE,I fixed by the subcontrollers.
Now letPKi be constant state-feedback gain matrices such
κ
that A + i=1 Bi Ki is Schur stable, and let Li be constant
observer gain matrices. It will be useful to define
K T = K1T · · · KκT
(6)
L = L1 · · · Lκ
(7)
Then the closed-loop system with private excitation is
Pκ
xn+1 = Axn + j=1 Bj (Kj x̂j,n + ej,n ) + wn
Pκ
Pκ
x̂i,n+1 = (A + j=1 Bj Kj + j=1 Lj Cj )x̂i,n +
Pκ
(8)
− j=1 Lj Cj xn + Bi ei,n +
Pκ
− j=1 Lj (zj,n + vj,n + νi,j,n )
These equations represent the fact that each subcontroller has
its own observer using the measurements that it has received.
It is not clear a priori that this closed-loop system is stable
since each observer may start at a different initial condition.
This concern is resolved by the following result:
Proposition 1: Let Ki and Li be constant state-feedback
and observer gains such that A + BK, A + LC, and A +
BK + LC are Schur stable. The closed-loop system (8) is
Schur stable with no private excitation ej,n ≡ 0, process
noise wn ≡ 0, measurement noise zj,n ≡ 0, measurement
attack vj,n ≡ 0, and communication attack νi,j,n ≡ 0.
Proof: Consider the change of variables from the states
x, x̂i to the states x, δ1 , di where δ1 = x̂1 −x and di = x̂i −x̂1
for i = 2, . . . , κ. Then inserting this change of variables into
(8) gives
Pκ
Pκ
xn+1 = (A + j=1 Bj Kj )xn + ( j=1 Bj Kj )δ1,n +
Pκ
j=2 Bj Kj dj,n
Pκ
Pκ
δ1,n+1 = (A + j=1 Lj Cj )δ1,n − j=2 Bj Kj dj,n (9)
Pκ
Pκ
di,n+1 = (A + j=1 Bj Kj + j=1 Lj Cj )di,n ,
for i = 2, . . . , κ
If we put x, δi into a single vector x̌, then the dynamics
x̌n+1 = Ǎx̌n are such that Ǎ is a block upper-triangular
matrix with A + BK, A + LC, and A + BK + LC on the
diagonal. This means Ǎ is Schur stable since we assumed
A + BK, A + LC, and A + BK + LC are Schur stable.
Remark 1: This result implies that the separation principle
does not hold. Fortunately, this is not a substantial impediment from the standpoint of design. Given a K such that
A + BK is stable, we can solve an LMI formulation [27],
[28] to choose (when feasible) a C such that both A + LC
and A + BK + LC are Schur stable. In particular, suppose
there exists a positive definite matrix Q 0 and general
matrix R such that the following two LMI’s
Q
AT Q + C T R
0
QT A + RT C
Q
Q
(A + BK)T Q + C T R
0
QT (A + BK) + RT C
Q
(10)
are satisfied. Then choosing L = Q−1 RT ensures that A +
LC and A+BK +LC are Schur stable. Convex optimization
can be used to determine if these LMI’s have a solution, and
compute a solution if possible.
For the purpose of designing our test, it will be useful to
define another change of variables on the states. Consider
the change of variables from the states x, x̂i to the states
x, δi where δi = x̂i − x for i = 2, . . . , κ. If there is no
measurement attack vj,n ≡ 0 and no communication attack
νi,j,n ≡ 0, then a straightforward calculation gives
Pκ
xn+1 = (A + BK)xn + j=1 Bj (Kj δj,n + ej,n ) + wn
Pκ
δi,n+1 = (A + BK + LC)δi,n + Bi ei,n − j=1 Lj zj,n +
Pκ
− j=1 Bj (Kj δj,n + ej,n ) − wn
(11)
If we define ∆T = δ1T · · · δκT and E T = eT1 · · · eTκ ,
then the above dynamics for the δi can be written as
∆n+1 = A∆n + blkdiag(B1 , . . . , Bκ )En +
T
Pκ
− 1 · · · 1 ⊗ −wn + j=1 Lj zj,n + Bj ej,n ,
(12)
where ⊗ is the Kronecker product, blkdiag(B1 , . . . , Bκ ) is
the block diagonal matrix with B1 , . . . , Bκ on the diagonals,
and A is the corresponding matrix defined to make the above
equivalent to (11). This will be used to define our test.
III. I NTUITION FOR WATERMARKING D ESIGN
Watermarking for networked systems faces new challenges
not encountered in the non-networked setting. To illustrate
the new difficulty, consider the networked LTI system with
1 1
1
0 C1 = 1 0
(13)
A=
B1 =
B2 =
0 1
0
1 C2 = 0 1
In this example, (A, B1 ) is not stabilizable and (A, C2 ) is
not detectable. And so coordination is required between the
subcontrollers to stabilize the system.
For instance, the choice K = − 12 I makes A + BK Schur
stable, and implementing the corresponding output-feedback
controller requires communication of partial observations between the two subcontrollers. In this case, the design is such
that the first subcontroller cannot inject any watermarking
signal into the second state, while the second subcontroller
cannot inject any watermarking signal into the first state. This
is problematic because this means each subcontroller cannot
verify the accuracy of the communicated state information.
However, suppose we instead choose
1 1 1
(14)
K=−
2 1 1
Then A + BK is Schur stable. More importantly, (A +
BK, B1 ) and (A + BK, B2 ) are controllable with this K.
Thus each subcontroller can inject a private watermarking
signal known only to the subcontroller, and such that this
signal can be used to verify the accuracy of the communicated state information and of the partial observations.
This example shows that designing watermarking differs in
the networked and non-networked cases. The networked case
requires designing both the state-feedback controller and the
corresponding tests to detect attacks; whereas watermarking
in the non-networked case only requires designing the the
corresponding tests to detect attacks [8]–[12], [17].
IV. S PECIFICATION OF S TATISTICAL T EST
Though watermarking for networked systems requires designing both the state-feedback controller and watermarking
tests, we first focus on the latter. We construct a statistical test
using the framework of null hypothesis testing, after assuming the existence of a state-feedback controller satisfying:
0
Condition 1: Let ki,j
= min{k ≥ 0 | Cj (A + BK)k Bi 6=
0
≤ p − 1.
0}. For each i and j, there exists a ki,j
This condition is itself nontrivial because it may be that
Cj (A+BK)k Bi ≡ 0 for all k ≥ 0. Approaches to synthesize
a state-feedback controller K to ensure the above condition
holds will be shown in the next section. This property is
important because it means the watermarking signal of the
i-th subcontroller is seen in the j-th output when the system
is controlled by perfect-information state-feedback.
A. Variable Definitions
Now before specifying the test, it is useful to define some
variables. Suppose we have K, L such that A+BK, A+LC,
and A + BK + LC are Schur stable. Let Σ∆ be the positive
semidefinite matrix that solves the Lyapunov equation
Σ∆ = AΣ∆ AT +
blkdiag(B1 , . . . , Bκ )ΣE blkdiag(B1 , . . . , Bκ )T +
T
1 ··· 1
..
.. ⊗ Σ + Pκ L Σ (L )T +
− ...
W
j
j=1 j Z,J
.
.
1 ··· 1
Bj ΣE,J (Bj )T , (15)
where ΣE = blkdiag(ΣE,1 , . . . , ΣE,κ ). A solution exists because the above is a Lyapunov equation and since
Proposition P
1 ensured stability. Note by construction Σ∆ =
N −1
as-limN N1 n=0 ∆n ∆Tn when there is no attack (i.e.,
vi,n ≡ 0 and νi,j,n ≡ 0 for all i, j, n). If we divide
Σ∆ ∈ Rκp×κp into sub-matrices with dimension p × p, then
define DI ∈ Rp×p to be the i × i-th sub-matrix of Σ∆ .
Lastly, we consider the matrix dynamics
E(∆n+1 eTi,t ) = AE(∆n eTi,t ))+
fi ⊗ Bi ΣE,I · 1(t = n)+
T
− 1 · · · 1 ⊗ (Bi ΣE,I ) · 1(t = n), (16)
where 1(·) is an indictor function, and the vector fi has a
one in the i-th position and is zero otherwise. This means
Σ∆,I,k := E(∆n eTi,n−k−1 ) = Ak fi ⊗ Bi ΣE,I +
T
− Ak 1 · · · 1 ⊗ (Bi ΣE,I ) . (17)
If we divide Σ∆,I,k ∈ Rκp×qi in sub-matrices of size p × qi ,
then let QI,J,k ∈ Rp×qi be the j-th sub-matrix of Σ∆,I,k .
Algorithm 1 Compute State-Feedback K for Proposition 2
b1
kb1 k
x1 :=
for all k ∈ {1, . . . , p − 1} do
xk+1 := λk kbbk+1
k+1 k
uk := B −1 xk+1 − Axk
end for
xp+1 := λp kbb11 k
up := B −1 xp+1 − Ax
p
X := x1 · · · xp
U := u1 . . . up
K := U X −1
B. Definition of Test
Our statistical watermarking test will involve the (secondorder) statistical characterization of the vectors defined as
0 −1
ei,n−ki,j
ψn,i,j =
,
(18)
Cj x̂i,n − si,j,n
and this characterization will be used to specify the distribution corresponding to the null hypothesis of no attack.
Theorem 1: If we have that vi,n ≡ 0 and νi,j,n ≡ 0, then
PN −1
as-limN N1 n=0 ψn ψnT = RI,J , where
"
#
ΣE,I
QTI,J,k0 CjT
i,j
RI,J =
.
(19)
0
Cj QI,J,ki,j
Cj Di CjT + ΣZ,J
Moreover, we have that as-limn E(ψn ) = 0.
Proof: First note that we have as-limn E(δi,n ) = 0
by the stability from Proposition 1. But Cj x̂i,n − si,j,n =
Cj δi,n − zj,n , and so E(Cj x̂i,n − si,j,n ) = Cj E(δi,n ).
This implies as-limn E(Cj x̂i,n − si,j,n ) = 0, which proves
as-limn E(ψn ) = 0 since the ei have zero mean.
Next observe the upper block triangle of (19) is correct by
0
construction of QI,J,ki,j
and by definition of the ei , and so
we only have to prove that the lower-right block is correct.
In particular, note that E((Cj δi,n − zn )(Cj δi,n − zn )T ) =
E((Cj δi,n )(Cj δi,n )T )+ΣZ,I since zi,n is independent of δi,n
by (11). This implies that we have that as-limn E((Cj δi,n −
zn )(Cj δi,n − zn )T ) = Cj Di CjT + ΣZ,I .
This result
Pn+` means Tthat asymptotically the summation
with ` ≥ (mi +qi ) has a Wishart
Si,j = 1` n+1 ψn,i,j ψn,i,j
distribution with ` degrees of freedom and a scale matrix that
matches (19), and we use this to define a statistical test. In
particular, we check if the negative log-likelihood
Pκ
L = j=1 (1 − ` + mi + qi ) · log det Sn,i,j +
Pκ
−1
trace
R
·
S
(20)
n,i,j
I,J
j=1
Algorithm 2 Compute State-Feedback K for Proposition 3
choose any v such that Bv ∈ ∩κi=1 range(Bi )
compute K 0 satisfying Heymann’s lemma for Bv
compute G such that A + BK 0 + BvG is Schur stable
K := K 0 + vG
while an acceptance corresponds to the lack of detection of
an attack. This notation emphasizes the fact that achieving
a specified false error rate α (a false error in our context
corresponds to detecting an attack when there is no attack
occurring) requires changing the threshold τ (α).
V. D ESIGNING THE S TATE -F EEDBACK
We provide two approaches for designing a state-feedback
controller that satisfies
Condition 1. The first applies when B
Pk
is square (i.e., i=1 qi = p); though it generalizes to skinny
Pk
B (i.e., i=1 qi < p) in some cases, we do not prove this.
The first approach relies upon a multi-input generalization
(which we construct and prove) of Heymann’s lemma [13],
[14]. The second approach applies to B of arbitrary size
where the range spaces of Bi have a nonempty intersection.
A. Multiple Input Heymann’s Lemma
Heymann’s lemma [13], [14] is used to prove arbitrary
pole placement of controllable, multiple input LTI systems by
allowing a reduction to the case of arbitrary pole placement
of a controllable, single input LTI system. Formally, it says
Lemma 1 (Heymann’s Lemma): If (A, B) is controllable,
then for any b = Bv 6= 0 there exists K (that depends on b)
such that (A + BK, b) is controllable.
We need a multiple input generalization of Heymann’s
Lemma. Let bi denote the i-th column of the matrix B. Then
Proposition 2: If B is full rank and square-shaped (i.e.,
B ∈ Rp×p ); then there exists a single K such that A + BK
is Schur stable and (A + BK, bi ) is controllable for all i.
Proof: We prove this result stepwise. Since B is full
rank and square-shaped, its columns are linearly independent.
Consider any λ with 0 < |λ| < 1, and define x1 = kbb11 k and
xn+1 = Axn +Bun . Now suppose there exists u1 , . . . , uk−1
such that xi = λi−1 kbbii k for i = 1, . . . , k. If k < p, then there
exists a uk satisfying λk kbbk+1
− λk−1 A kbbkk k = Buk since
k+1 k
B is full rank. Hence by definition of the dynamics on x
there exists uk such that xk+1 = λk kbbk+1
. If k = p, then
k+1 k
b
corresponding to this Wishart distribution and the summations of Sn,i,j is large by conducting the hypothesis test
(
reject,
if L(Sn ) > τ (α)
(21)
accept, if L(Sn ) ≤ τ (α)
there exists a up satisfying λp kbb11 k − λp−1 A kbpp k =
since B is full rank. So by definition of the dynamics
there exists up such that xp+1 = λp kbb11 k .
Next define the matrices
U = u1 · · · up
h
i
b
R = kbb11 k · · · kbpp k
Λ = diag 1, λ, . . . , λp−1
where τ (α) is a threshold that controls the false error rate
α. A rejection corresponds to the detection of an attack,
and K = U R−1 Λ−1 . The matrices Λ and R are invertible
by construction since 0 < |λ| < 1 and B is invertible. Note
Bup
on x
(22)
(23)
(24)
that by definition U = KΛB. Finally, note for any i we have
h
i
b
b
kbi k · kbbii k · · · kbpp k kbb11 k · · · kbi−1
·Λ=
i−1 k
p−1
bi (A + BK)bi · · · (A + BK) bi . (25)
has rank(C0 ) = p, and so the controllability matrix
C = Bi (A + BK)Bi . . . (A + BK)p−1 Bi
1000
800
0
also has rank(C) = p since the columns of C are a superset
of the columns of C0 . Thus by Sylvester’s rank inequality,
we have rank(Cj C) ≥ rank(Cj ) + rank(C) − p = rank(C).
But rank(Cj ) ≥ 1 since Cj 6= 0. Combining this with the
earlier inequality gives rank(Cj C) ≥ 1, and so Cj C 6= 0.
0
This means ki,j
≤ p − 1 exists since Cj C is a block matrix
consisting of the blocks Cj (A + BK)k Bi .
B. Nonempty Intersection of Inputs
We next consider B with arbitrary shape, such that the
range spaces of Bi have a nonempty intersection. Our
Algorithm 2 designs a K for this case, and it uses Heymann’s
lemma [13], [14]. The next result proves its correctness.
Proposition 3: Suppose Cj 6= 0 for all j and that we
have ∩κi=1 range(Bi ) 6= ∅. If (A, B) is controllable, then
Algorithm 2 computes a K such that A + BK is Schur
stable and that Condition 1 is satisfied.
Proof: First note that v exists by assumption, and
so we can compute K 0 by Heymann’s lemma. This means
(A + BK 0 , Bv) is controllable, which implies that we can
compute G such that A + BK 0 + BvG is Schur stable. This
10
20
30
40
50
40
50
40
50
Test by Car 2
1400
1200
1000
800
0
10
20
30
Test by Car 3
1400
1200
1000
800
0
10
20
30
Time (s)
Fig. 1. Value of (20) for Simulation of Vehicle Platoon, with a Negative
Log-Likelihood Threshold for α = 0.05 False Detection Error Rate
proves that A + BK is Schur stable since K = K 0 + vG.
Next consider any i ∈ {1, . . . , κ}. Since (A + BK 0 , Bv)
is controllable, this means that (A + BK 0 + BvG, Bv)
is controllable. (This uses the well known fact that statefeedback does not affect controllability.) As a result, we have
C0 = Bv
(27)
unattacked
attacked
(0.05)
1200
Negative Log Liklihood
The left side has full rank by the assumptions on B, and the
right side is the observability matrix for the (A + BK, bi ).
This proves (A + BK, bi ) is observable for all i since we
have shown that the observability matrix has full rank.
We conclude by proving that the above designed K makes
A + BK Schur stable. Consider any x ∈ Rp , and observe
that by the assumptions on B there exists z ∈ Rp such that
x = Rz. But, as in (25), by construction (A + BK)p kbbii k =
λp kbbii k for all i. Hence (A + BK)p x = (A + BK)p Rz =
λp Rz = λp x for all x. This means all the eigenvalues of
(A + BK)p are λp , and using the spectral mapping theorem
implies the eigenvalues of A + BK are roots of λp . Thus
the magnitude of the eigenvalues of A + BK are |λ|, which
means that A + BK is Schur stable since 0 < |λ| < 1.
Though the above is an existence result, a state-feedback
matrix K satisfying Proposition 2 can be computed using
Algorithm 1. The correctness of this algorithm follows from
the construction used in the proof of Proposition 2. Also, the
next result proves that this K satisfies Condition 1.
Corollary 1: Suppose Cj 6= 0 for all j. If B is full rank
and square-shaped (i.e., B ∈ Rp×p ); then there exists a K
such that A+BK is Schur stable and that Condition 1 holds.
Proof: Consider any i ∈ {1, . . . , κ}, and choose s to
be any index such that the s-th column in B belongs to Bi .
Proposition 2 says (A + BK, bs ) is controllable. This means
the controllability matrix
C0 = bs (A + BK)bs . . . (A + BK)p−1 bs
(26)
Test by Car 1
1400
(A + BK)Bv
...
(A + BK)p−1 Bv
(28)
has rank(C0 ) = p. But by assumption, there exists vi such
that Bi vi = Bv. So if we define
C = Bi
(A + BK)Bi
...
(A + BK)p−1 Bi , (29)
then we have that C0 = C blkdiag(vi , . . . , vi ). Thus p ≥
rank(C) ≥ rank(C0 ) = p. Thus by Sylvester’s rank inequality, we have rank(Cj C) ≥ rank(Cj ) + rank(C) − p =
rank(C). But rank(Cj ) ≥ 1 since Cj 6= 0. Combining this
with the earlier inequality gives rank(Cj C) ≥ 1, and so
0
Cj C 6= 0. This means ki,j
≤ p − 1 exists since Cj C is a
block matrix consisting of the blocks Cj (A + BK)k Bi .
VI. S IMULATION : AUTONOMOUS V EHICLE P LATOONING
We apply a standard model [29] for error kinematics of
speed control of vehicles to generate a model for a three
car platoon.
The state vector
for the three car platoon is
xT = e1 d1 e2 d2 e3 d3 , where ei is deviation from the
desired velocity for car i, and di is deviation from the desired
following distance between car i and car i+1. The discretized
dynamics for a timestep of 0.05 seconds
1
0
0
0
1
− 1 1
0
20
20
0
0
1
0
A=
1
0
1
0 − 20
0
0
0
0
are then:
0
0
0
1
20
1
R EFERENCES
(30)
and
1
20
− 1
800
B1 =
0
0
0
1
800
1
B2 =
201
−
800
0
B3 =
0
0
0
0
1
800
1
20
.
(31)
Assuming each car measures its own velocity and the distance to the car in front of it, we have
C1 = 1 0 0 0 0
0 1 0 0 0
C2 =
0 0 1 0 0
(32)
0 0 0 1 0
C2 =
0 0 0 0 1
We assume that the process and measurement noise had
variance ΣW = 5×10−5 ·I and ΣZ,I = 10−3 ·I, respectively.
We applied our statistical test (20) with each car using a
watermarking signal with variance ΣE,I = 0.2, where
K1 = −1 0.1 0 0 0
K2 = 1 −1 −2 0.1 0
(33)
K3 = 0.5 −0.5 0.5 −1 −2
and
0
0
0
0
−0.5
(34)
We conducted two simulations, where the platoon was unattacked and attacked. In the attack simulations, the attacker
chose v1,n and ν2,3,n to be zero mean i.i.d. jointly Gaussian
random variables with variance 0.5 and 0.2 · I, respectively.
Fig. 1 shows the results of applying our statistical test (20),
and the plots show that our statistical watermarking test can
detect the presence or absence of an attack.
0.05
−0.5
−0.5
0
L1 =
0 L2 = 0
0
0
0
0
0
0
0
0
−0.5
L3 = 0.05
−0.5
0
0
0
VII. C ONCLUSION
This paper developed statistical watermarking to detect
sensor and communication attacks on networked LTI systems. Unlike the non-networked case, watermarking in the
networked case requires a K so that A + BK is controllable
with respect to each Bi . This ensures the private watermarking signal of each subcontroller is seen in the output of
all subcontrollers. We provided two algorithms to compute
such a controller. The efficacy of our watermarking approach
was demonstrated by a simulation of a three car platoon.
One possible direction for future work is to explore the
performance of our watermarking test under specific types
of attacks, such as replay attacks or network disturbances.
[1] A. A. Cárdenas, S. Amin, and S. Sastry, “Research challenges for the
security of control systems.” in HotSec, 2008.
[2] M. Abrams and J. Weiss, “Malicious control system cyber security
attack case study–Maroochy water services, australia,” MITRE, 2008.
[3] R. Langner, “Stuxnet: Dissecting a cyberwarfare weapon,” IEEE
Security & Privacy, vol. 9, no. 3, pp. 49–51, 2011.
[4] B. Parno, M. Luk, E. Gaustad, and A. Perrig, “Secure sensor network
routing: A clean-slate approach,” in ACM CoNEXT, 2006.
[5] V. Kumar, J. Srivastava, and A. Lazarevic, Managing cyber threats:
issues, approaches, and challenges. Springer, 2006, vol. 5.
[6] W. Wang and Z. Lu, “Cyber security in the smart grid: Survey and
challenges,” Computer Networks, vol. 57, no. 5, pp. 1344–1371, 2013.
[7] K.-D. Kim and P. R. Kumar, “Cyber–physical systems: A perspective
at the centennial,” Proc. of IEEE, vol. 100, pp. 1287–1308, 2012.
[8] Y. Mo and B. Sinopoli, “Secure control against replay attacks,” in
Allerton Conference. IEEE, 2009, pp. 911–918.
[9] Y. Mo, E. Garone, A. Casavola, and B. Sinopoli, “False data injection
attacks against state estimation in wireless sensor networks,” in Proc.
of IEEE CDC, 2010, pp. 5967–5972.
[10] Y. Mo, R. Chabukswar, and B. Sinopoli, “Detecting integrity attacks
on scada systems,” IEEE CST, vol. 22, no. 4, pp. 1396–1407, 2014.
[11] S. Weerakkody, Y. Mo, and B. Sinopoli, “Detecting integrity attacks
on control systems using robust physical watermarking,” in Proc. of
IEEE CDC, 2014, pp. 3757–3764.
[12] Y. Mo, S. Weerakkody, and B. Sinopoli, “Physical authentication
of control systems: Designing watermarked control inputs to detect
counterfeit sensor outputs,” IEEE Control Systems, vol. 35, no. 1, pp.
93–109, 2015.
[13] M. Heymann, “Comments on “on pole assignment in multi-input controllable linear systems”,” IEEE Transactions on Automatic Control,
vol. 13, no. 6, pp. 748–749, 1968.
[14] M. L. Hautus, “A simple proof of heymann’s lemma,” IEEE Transactions on Automatic Control, vol. 22, no. 5, pp. 885–886, 1977.
[15] B. Satchidanandan and P. Kumar, “Dynamic watermarking: Active
defense of networked cyber-physical systems,” Proc. of IEEE, 2016.
[16] W.-H. Ko, B. Satchidanandan, and P. Kumar, “Theory and implementation of dynamic watermarking for cybersecurity of advanced
transportation systems,” in Proc. of IEEE CNS, 2016, pp. 416–420.
[17] P. Hespanhol, M. Porter, R. Vasudevan, and A. Aswani, “Dynamic
watermarking for general LTI systems,” in IEEE Conference on
Decision and Control, 2017, to appear.
[18] H. Gonzalez and E. Polak, “On the perpetual collision-free rhc of
fleets of vehicles,” Journal of optimization theory and applications,
vol. 145, no. 1, pp. 76–92, 2010.
[19] A. Aswani and C. Tomlin, “Game-theoretic routing of GPS-assisted
vehicles for energy efficiency,” in ACC, 2011, pp. 3375–3380.
[20] W. Zhang, M. Kamgarpour, D. Sun, and C. J. Tomlin, “A hierarchical
flight planning framework for air traffic management,” Proceedings of
the IEEE, vol. 100, no. 1, pp. 179–194, 2012.
[21] R. Vasudevan, V. Shia, Y. Gao, R. Cervera-Navarro, R. Bajcsy, and
F. Borrelli, “Safe semi-autonomous control with enhanced driver
modeling,” in ACC, 2012, pp. 2896–2903.
[22] S. Mohan and R. Vasudevan, “Convex computation of the reachable
set for hybrid systems with parametric uncertainty,” in Proc. of ACC,
2016, pp. 5141–5147.
[23] G. Como, E. Lovisari, and K. Savla, “Convexity and robustness of
dynamic network traffic assignment for control of freeway networks,”
IFAC-PapersOnLine, vol. 49, no. 3, pp. 335–340, 2016.
[24] D. P. Shepard, T. E. Humphreys, and A. A. Fansler, “Evaluation
of the vulnerability of phasor measurement units to GPS spoofing
attacks,” International Journal of Critical Infrastructure Protection,
vol. 5, no. 3, pp. 146–153, 2012.
[25] T. Nighswander, B. Ledvina, J. Diamond, R. Brumley, and D. Brumley,
“GPS software attacks,” in ACM CCS, 2012, pp. 450–461.
[26] C. Bonebrake and L. R. O’Neil, “Attacks on GPS time reliability,”
IEEE Security Privacy, vol. 12, no. 3, pp. 82–84, May 2014.
[27] S. Boyd, L. El Ghaoui, E. Feron, and V. Balakrishnan, Linear matrix
inequalities in system and control theory. SIAM, 1994.
[28] M. C. de Oliveira, J. Bernussou, and J. C. Geromel, “A new discretetime robust stability condition,” Systems & control letters, vol. 37,
no. 4, pp. 261–265, 1999.
[29] A. G. Ulsoy, H. Peng, and M. Çakmakci, Automotive control systems.
Cambridge University Press, 2012.
| 3cs.SY
|
&+(3/D-ROOD&DOLIRUQLD0DUFK
1
Design, implementation and deployment of the Saclay muon
reconstruction algorithms (Muonbox/y) in the Athena software
framework of the ATLAS experiment
A. Formica
DAPNIA/SEDI, CEA/Saclay, 91191 Gif-sur-Yvette CEDEX, France
This paper gives an overview of a reconstruction algorithm for muon events in ATLAS experiment at CERN. After a short
introduction on ATLAS Muon Spectrometer, we will describe the procedure performed by the algorithms Muonbox and
Muonboy (last version) in order to achieve correctly the reconstruction task. These algorithms have been developed in Fortran
language and are working in the official C++ framework Athena, as well as in stand alone mode. A description of the
interaction between Muonboy and Athena will be given, together with the reconstruction performances (efficiency and
momentum resolution) obtained with MonteCarlo data.
1. INTRODUCTION
This paper summarizes the results collected in ATLAS
Technical Design Reports (TDR) [1] [2] and in the
recent analysis of Data Challenge 1 (DC1).
2. MUON RECONSTRUCTION IN ATLAS
A robust muon identification and high precision
momentum measurement is crucial to fully exploit the
physics potential of the LHC. The muon energy of
physics interest ranges over a large interval from a few
GeV, HJ for %-physics studies, up to a few TeV, where
one might expect the presence of new physics. From the
point of view of muon detection, the ATLAS apparatus
is characterised by two high precision tracking systems,
namely the Inner Detector and the Muon System, and a
thick calorimeter that ensures safe hadron filtering with
high purity for muons of momentum above 3 GeV. The
Muon System is designed to reach high reconstruction
efficiency and a momentum resolution at a few % level
RYHU D ODUJH IUDFWLRQ RI WKH
VSDFH covered by the
detector, in a wide range of muon energies (S from 10
GeV up to a few TeV). The Muon System is composed
by: 1) the Muon Drift Tubes (MDT), that are arranged in
chambers in three (four) layers around the beam axis: the
inner, the middle, and the outer layer. Each MDT
chamber contains two superlayers which each consist of
3 or 4 tube layers. An MDT chamber is able to measure a
’local straight track segment’ or track vector. The muon
trigger chambers are Resistive Plate Chambers (RPC) in
the barrel and Thin Gap Chambers (TGC) in the endcaps and are physically connected to the MDT chambers.
A package of MDT and RPC/TGP chambers is called a
station. The whole of the spectrometer is in a toroidal
magnetic field (see Figure 1).
Muon reconstruction in ATLAS experiment is a very
challenging task, due to several effects: the large muon
system size (44m length and 22m Ø) implies higher
extrapolation uncertainties; the open air-core toroid that
leads to a rather inhomogeneous magnetic field affects
the measurement for events with low S momentum, and
finally the precision chambers that can measure only the
coordinate giving the bending direction with an accuracy
of 80µm, while the second coordinate is given by trigger
chambers with an accuracy of 1cm. High level of physics
background is also expected.
Open air-core Toroid
EndCap MDT
TGC
Barrel MDT
Figure 1: ATLAS detection of HÆ4µ simulated event. The MDT chambers traversed by the muons are shown
7+/7
2
&+(3/D-ROOD&DOLIRUQLD0DUFK
3.
3. MUONBOX (MUONBOY) ALGORITHMS
These algorithms have been developed in F77
(Muonbox) and F90 (Muonboy). The usage of the latter
language allows many improvements in the code, since
F90 contains advanced features like dynamic memory
management, recursion and polymorphism
3.1. Strategy for pattern recognition
The strategy of the pattern recognition algorithm,
described below, can be summarized in four main points:
1) identifications of ’regions of activity’ (ROA) in the
muon system, through the RPC/TGC systems; 2)
reconstruction of ’local straight track segments’ in each
muon station of these regions of activity; 3) combination
of tracks segments of different muon stations to form
muon track candidates; 4) global track fit of the muon
track candidates through the full system.
4.
5.
an accurate JHRPHWU\ GHVFULSWLRQ of the whole
detector (AMDB database), needed to correctly take
into account the traversed material during the
tracking procedure;
DOLJQPHQW
FRUUHFWLRQV
to nominal chamber
positions are also important for tracking the muons;
the PDJQHWLF ILHOG (figure 3) has to be known with a
precision of 20-30 gauss (10-3) that is required in the
fit procedure. The magnetic field will be measured
by sensors installed in the detector. A service has
been developed in order to provide magnetic field
values for every given space point:
I
& &
( [, \ , ] ) → ( % , ∇% )
Using these parameters and the previously described
strategy for the pattern recognition, we are able to
reconstruct tracks inside the spectrometer
6HJPHQWFRPELQDWLRQ
52$GHILQLWLRQ
/RFDO7UDFNVHJPHQWV
3.3. Building Tracks
Figure 2: pattern recognition strategy
In the first step, ROAs LQWKH VSDFHDUHLGHQWLILHG
using information from the trigger chambers. The size of
WKHVHUHJLRQVLVURXJKO\G × d × 0.4 and they are
centered where there exists at least one RPC/TGC hit in
both coordinates (among up to seven layers of
RPC/TGC). Then, straight track segments are first
reconstructed individually, in the bending plane, in each
muon station intersecting with these ROAs. We will
enter in the detail of segment combination later in this
paper (paragraph 3.3).
3.2. Input Parameters for Reconstruction
The needed input parameters for the algorithm are:
GULIW WLPH (dt) measurement in MDT tubes and
ROA hits coming from trigger chambers. The drift
time is corrected accounting for Lorentz angle
(magnetic field effect), tube length (signal
propagation delay), and time of flight;
2. the 57 relations (time vs distance), that are
measured using real tracks;
1.
7+/7
Starting from straight track segments from each muon
station we can combine them and perform track
reconstruction in the whole spectrometer. During this
procedure, the DOJRULWKP WDNHV LQWR DFFRXQW IRU UD\V
2
DQG WXEH LQHIILFLHQFLHV WKDW FDQ GHJUDGH WKH
of the
track. The position and direction of track segments found
in the outer and middle stations of the muon
spectrometer allow a first rough estimate of the
momentum of corresponding candidate muons. Each of
these strict track segments is then extrapolated to the first
middle or outer (or even inner when middle or outer is
missing) station found using tracking in the magnetic
field. A momentum scan around the first estimate is used
searching for best matching between tracks. Fit on the
candidate track is made performing full tracking at each
step of the minimisation procedure. In this way the
algorithm can fine tune position, direction, and
momentum of the track, by taking into account for
inefficient tubes (due to dead time in the detector), and
for missing hits along the track trajectory.
Finally, a last global re-fit is performed, starting from
the best result of the previous fits, but using this time
&+(3/D-ROOD&DOLIRUQLD0DUFK
3
directly the raw information available information (i.e.
the drift time values and hit strips instead of the prereconstructed straight track segments). Eventually, the
selection of reconstructed muons is made according to
the value of the 2 of this last global fit. Those single hits
that are spoiled by UD\V DQG QHXWURQ EDFNJURXQG are
cleaned up during this final fit. Multiple scattering and
energy loss in the spectrometer matter are taken into
account as well, using the geometry description.
In order to combine the tracks reconstructed in the
Inner Detector and the Muon System, Muonbox/y apply
a strategy based on the statistical combination of the two
independent measurements using the parameters of the
reconstructed tracks and their covariance matrices. We
will not enter here in a detailed description of this
method, which can be found in [2].
3.4. Tracking Back to the Beam
In figures 5 and 6 we show the reconstruction
efficiency and the resolution for the Muon spectrometer
alone as well as for the combined reconstruction with the
Inner Detector. The plots reproduce the TDR result, and
are obtained running Muonbox algorithm inside Athena
framework, on a data sample coming from DC1 (Data
Challenge 1).
Another relevant task of the reconstruction is to
perform backtracking from Muon System down to beam
region. This procedure requires an accurate knowledge
of the amount and the nature of the material traversed by
muon trajectory in such a way to correctly account for
energy losses of the muons along the track. Muon
momentum is corrected using an energy loss
parameterization. A second method that will be
implemented in future can use the calorimeter energy
measured in the traversed cells. This second method is
foreseen however only for high S muons, having higher
probability of a catastrophic energy loss. The
reconstruction program is able to predict with good
accuracy (see figure 4) the energy loss over a wide range
of muon momenta (5-500 GeV). A complete set of
output parameters is provided by Muonbox and
Muonboy algorithms: the track parameters and the
covariance matrix of the fit. Track parameters are given
in different positions of the detector: the entrance of the
spectrometer, the entrance of the calorimeters, and at the
perigee level (i.e. the closest distance of approach with
the beam axis). The covariance matrix represents the
errors on tracking procedure, and takes into account for
multiple scattering and fluctuations in energy losses.
3.5. Combination with Inner Detector
Muon tracks are also reconstructed in the Inner
Detector. The combination of measurements made in
the Muon System with the ones from the Inner
Detector improves the momentum resolution in the
momentum range 6 GeV < S < 100 GeV. The
4. ALGORITHM PERFORMANCES
Figure 5: Muonbox efficiency on DC1 data
The reconstruction efficiency is affected by the energy
deposit of the muon in the detector before the Muon
System for low S muons.
The main contribution to reconstruction resolution are
coming from the Muon System alignment (30 µm for
Barrel, 50µm for End-Cap) and from the space resolution
of the precision chambers at high momentum (S > 200
GeV), from energy loss fluctuations at S < 200 GeV,
and from multiple scattering fluctuations that are almost
constant in S .
Figure 6: Muonbox: S resolution on DC1 data
The time performance (time per event) of the
reconstruction is summarized in figure 7, for three
different versions of the algorithm: Muonbox in blue,
Muonboy in red and an accelerated version in violet. The
measurement were performed on a DEC alpha station
(700 MHz), running the algorithm in a stand alone mode.
matching of the muon track reconstructed independently
in the Inner Detector and in the Muon System allows the
rejection of muons from secondary interactions as well
as the ones from π/.decays in flight.
7+/7
&+(3/D-ROOD&DOLIRUQLD0DUFK
4
Figure 7: Muonbox/y time performances
Figure 6: Muonbox: S resolution on DC1 data
The time performance (time per event) of the
reconstruction is summarized in figure 7, for three
different versions of the algorithm: Muonbox in blue,
Muonboy in red and an accelerated version in violet. The
measurement were performed on a DEC alpha station
(700 MHz), running the algorithm in a stand alone mode.
The accelerated version of the algorithm is foreseen to
improve the time per event performances in order to
have the possibility of using this reconstruction for high
level trigger purposes.
Figure 8: Diagrams of the interface between Muonboy and Athena framework
5. USING MUONBOX/Y
5.1. Athena framework
Athena is the official C++ framework that provides
services and algorithms for physics analysis in ATLAS.
7+/7
Algorithms inside Athena can access the hits information
needed for the reconstruction from the Transient Data
Store, and put the reconstruction output (track parameters
and covariance matrix) in the same place, from where
other users can access it for further analysis. We recall
here that the output parameters are given in different
position of the detector, so to allow combination of muon
3ODFH&RQIHUHQFH7LWOH/RFDWLRQDQG'DWHV+HUH
spectrometer tracks with the Inner Detector tracks (see
figure 8). An interface has been developed in order to run
Muonbox/y algorithms in Athena. This interface takes
care of the information transmission from the C++ world
to the reconstruction algorithms.
5.2. Stand Alone Mode
The algorithms can run in stand alone mode by reading
all needed input from external files (geometry and
magnetic field map as well). This feature has been very
useful in the debugging of the algorithm itself, and for a
fast feedback in the analysis of TestBeam data in H8 at
CERN, during 2002 and 2003 years. For its modularity,
this reconstruction is used also in an event display (see
7+/7
5
figure 1) that gives a visual representation of the
geometry of the detector, and can simulate and
reconstruct muon events at once.
Acknowledgments
The authors wish to thank Saclay Muon Software
Group for the help and guidance in preparing this
document.
References
[1]
[2]
ATLAS TDR 10, CERN/LHCC 97-22 , May 1997.
ATLAS TDR 14, CERN/LHCC 99-14 , May 1999.
| 5cs.CE
|
Evolutionary Trees and the Ising Model on the Bethe Lattice:
A Proof of Steel’s Conjecture∗
arXiv:math/0509575v3 [math.PR] 28 Jul 2009
Constantinos Daskalakis†
Elchanan Mossel‡
Sébastien Roch§
April 16, 2018
Abstract
A major task of evolutionary biology is the reconstruction of phylogenetic trees from molecular data.
The evolutionary model is given by a Markov chain on a tree. Given samples from the leaves of the
Markov chain, the goal is to reconstruct the leaf-labelled tree.
It is well known that in order to reconstruct a tree on n leaves, sample sequences of length Ω(log n)
are needed. It was conjectured by M. Steel that for the
√ CFN/Ising evolutionary model, if the mutation
probability on all edges of the tree is less than p∗ = ( 2 − 1)/23/2 , then the tree can be recovered from
sequences of length O(log n). The value p∗ is given by the transition point for the extremality of the
free Gibbs measure for the Ising model on the binary tree. Steel’s conjecture was proven by the second
author in the special case where the tree is “balanced.” The second author also proved that if all edges
have mutation probability larger than p∗ then the length needed is nΩ(1) .
Here we show that Steel’s conjecture holds true for general trees by giving a reconstruction algorithm
that recovers the tree from O(log n)-length sequences when the mutation probabilities are discretized
and less than p∗ . Our proof and results demonstrate that extremality of the free Gibbs measure on
the infinite binary tree, which has been studied before in probability, statistical physics and computer
science, determines how distinguishable are Gibbs measures on finite binary trees.
Keywords: Phylogenetics, CFN model, Ising model, phase transitions, reconstruction problem, Jukes
Cantor.
1
Introduction
In this paper we prove a central conjecture in mathematical phylogenetics [Ste01]: We show that every
phylogenetic tree with short, discretized edges on n leaves can be reconstructed from sequences of length
O(log n), where by short we mean that the mutation probability on every edge is bounded above by the
critical transition probability for the extremality of the Ising model on the infinite binary tree.
∗
2000 Mathematics Subject Classification. Primary 60K35, 92D15; Secondary 60J85, 82B26. Key words and phrases. Phylogeny, phase transition, Ising model.
†
CSAIL, MIT. Work done at U.C Berkeley and Microsoft Research. Partially supported by CIPRES (NSF ITR grant # NSF EF
03-31494). Email: [email protected].
‡
Statistics and Computer Science, U.C. Berkeley, and Weizmann Institute of Science, Rehovot, Israel. Supported
by a Miller fellowship in Statistics and Computer Science, by a Sloan fellowship in Mathematics and by NSF grants
DMS-0504245, DMS-0528488 DMS-0548249 (CAREER), ONR grant N0014-07-1-05-06 and ISF grant 1300/08. Email:
[email protected].
§
Department of Mathematics, UCLA. Work at U.C. Berkeley and Microsoft Research. Partially supported by CIPRES (NSF
ITR grant # NSF EF 03-31494). Email: [email protected].
This result establishes that the extremality of the free Gibbs measure for the Ising model on the infinite binary tree, studied in probability, statistical physics and computer science, determines the sampling
complexity of the phylogenetic problem, a central problem in evolutionary biology. We proceed with background on the phylogenetic problem, on the reconstruction problem and a statement of our results.
Phylogenetic Background. Phylogenies are used in evolutionary biology to model the stochastic evolution
of genetic data on the ancestral tree relating a group of species. The leaves of the tree correspond to (known)
extant species. Internal nodes represent extinct species. In particular the root of the tree represents the most
recent ancestor of all species in the tree. Following paths from the root to the leaves, each bifurcation indicates a speciation event whereby two new species are created from a parent. We refer the reader to [SS03] for
an excellent introduction to phylogenetics. The underlying assumption is that genetic information evolves
from the root to the leaves according to a Markov model on the tree. It is further assumed that this process
is repeated independently a number of times denoted k. Thus each node of the tree is associated with a
sequence of length k. The vector of the i’th letter of all sequences at the leaves is called the i’th character.
One of the major tasks in molecular biology, the reconstruction of phylogenetic trees, is to infer the topology
of the tree from the characters at the leaves.
In this paper we will be mostly interested in two evolutionary models, the so-called Cavender-FarrisNeyman (CFN) [Cav78, Far73, Ney71] and Jukes-Cantor (JC) [JC69] models. In the CFN model the states
at the nodes of the tree are 0 and 1 and their a priori probability at the root is uniform. To each edge e
corresponds a mutation probability p(e) which is the probability that the state changes along the edge e.
Note that this model is identical to the free Gibbs measure of the Ising model on the tree. See [Lyo89]. In
the JC model the states are A, C, G and T with a priori probability 1/4 each. To each edge e corresponds a
mutation probability p(e) and it is assumed that every state transitions with probability p(e) to each of the
other states. This model is equivalent to the ferromagnetic Potts model on the tree.
Extremality and the Reconstruction Problem. A problem that is closely related to the phylogenetic problem is that of inferring the ancestral state, that is, the state at the root of the tree, given the states at the
leaves. This problem was studied earlier in statistical physics, probability and computer science under the
name of reconstruction problem, or extremality of the free Gibbs measure. See [Spi75, Hig77, Geo88]. The
reconstruction problem for the CFN model was analyzed in [BRZ95, EKPS00, Iof96, BKMP05, MSW04b,
BCMR06]. In particular, the role of the reconstruction problem in the analysis of the mixing time of Glauber
dynamics on trees was established in [BKMP05, MSW04b].
Roughly speaking, the reconstruction problem is solvable when the correlation between the root and
the leaves persists no matter how large the tree is. When it is unsolvable, the correlation decays to 0 for
large trees. The results of [BRZ95, EKPS00, Iof96, BKMP05, MSW04b, BCMR06] show that for the CFN
model, if for all e it holds that p(e) ≤ pmax < p∗ , then the reconstruction problem is solvable, where
√
2−1
∗
p = √
≈ 15%.
8
If, on the other hand, for all e it holds that p(e) ≥ pmin > p∗ and the tree is balanced in the sense that all
leaves are at the same distance from the root, then the reconstruction problem is unsolvable. Moreover in
this case, the correlation between the root state and any function of the character states at the leaves decays
as n−Ω(1) .
Our Results. M. Steel [Ste01] conjectured that when 0 < pmin ≤ p(e) ≤ pmax < p∗ for all edges e,
one can reconstruct with high probability the phylogenetic tree from O(log n) characters. Steel’s insightful
conjecture suggests that there are deep connections between the reconstruction problem and phylogenetic
reconstruction.
2
This conjecture has been proven to hold for trees where all the leaves are at the same graph distance from
the root—the so-called “balanced” case—in [Mos04]. It is also shown there that the number of characters
needed when p(e) ≥ pmin > p∗ , for all e, is nΩ(1) . The second result intuitively follows from the fact that
the topology of the part of the tree that is close to the root is essentially independent of the characters at the
leaves if the number of characters is not at least nΩ(1) .
The basic intuition behind Steel’s conjecture is that: since in the regime where p(e) ≤ pmax < p∗ there
is no decay of the quality of reconstructed sequences, it should be as easy to reconstruct deep trees as it
is to reconstruct shallow trees. In [ESSW99] (see also [Mos07]) it is shown that “shallow” trees can be
reconstructed from O(log n) characters if all mutation probabilities are bounded away from 0 and 1/2 (the
results of [ESSW99] also show that in this regime, any tree can be recovered from sequences of polynomial
length). The same high-level reasoning has also yielded a complete proof that O(log n) characters suffice
for a percolation-type mutation model when all edges are short [MS04]. See [MS04] for details.
Here we prove Steel’s conjecture for general trees under the assumption that the mutation probabilities
are discretized. We show that, if 0 < pmin ≤ p(e) ≤ pmax < p∗ for all edges e of the tree, then the tree
can be reconstructed from c(pmin , pmax )(log n + log 1/δ) characters with error probability at most δ. The
discretization assumption amounts to assuming that all edge lengths are a multiple of a small constant ∆.
(See below for a formal definition of “edge length.”) This result further implies that sequences of logarithmic
length suffice to reconstruct phylogenetic trees in the Jukes-Cantor model, when all the edges are sufficiently
short.
Compared to [Mos04], our main technical contribution is the design and analysis of a tree-building procedure that uses only “local information” to build parts of the tree, while also maintaining “disjointness”
between the different reconstructed subtrees. The disjointness property is crucial in order to maintain the
conditional independence properties of the Gibbs distribution on the original tree—for the purpose of performing estimation of the states at internal nodes of the tree. Note that in the balanced case of [Mos04] this
property can be achieved in a straightforward manner by building the tree “level-by-level.”
1.1
Formal Definitions and Main Results
Trees and Path Metrics. Let T be a tree. Write V(T ) for the nodes of T , E(T ) for the edges of T and
L(T ) for the leaves of T . If the tree is rooted, then we denote by ρ(T ) the root of T . Unless stated otherwise,
all trees are assumed to be binary (all internal degrees are 3) and it is further assumed that L(T ) is labelled.
Let T be a tree equipped with a length function onP
its edges, d : E(T ) → R+ . The function d will also
denote the induced path metric on V(T ): d(v, w) = {d(e) : e ∈ pathT (v, w)}, for all v, w ∈ V(T ),
where pathT (x, y) is the path (sequence of edges) connecting x to y in T .
We will further assume below that the length of all edges is bounded between f and g for all e ∈ E(T ).
In other words, for all e ∈ E(T ), f ≤ d(e) ≤ g.
Markov Model of Evolution. The evolutionary process is determined by a rooted tree T = (V, E)
equipped with a path metric d and a rate matrix Q. We will be mostly interested in the case where
1
Q = −1
1 −1 corresponding to the CFN model and in the case where Q is a 4 × 4 matrix given by
Qi,j = 1 − 4 · 1{i = j} corresponding to the Jukes-Cantor model. To edge e of length d(e) we associate
the transition matrix M e = exp(d(e)Q).
In the evolutionary model on the tree T rooted at ρ each vertex iteratively chooses its state from the state
at its parent by an application of the Markov transition rule M e , where e is the edge connecting it to its
parent. We assume that all edges in E are directed away from the root. Thus the probability distribution on
3
the tree is the probability distribution on {0, 1}V ({A, C, G, T }V ) given by
Y
(x→y)
Mσ(x),σ(y) ,
µ[σ] = π(σ(ρ))
(x→y)∈E
where π is given by the uniform distribution at the root, so that π(0) = π(1) = 1/2 for the CFN model and
π(A) = π(C) = π(G) = π(T ) = 1/4 for the JC model. We let the measure µ denote the marginal of µ on
the set of leaves which we identify with [n] = {1, . . . , n}. Thus
X
µ(σ) =
{µ(τ ) : ∀i ∈ [n], τ (i) = σ(i)}.
The measure µ defines the probability distribution at the leaves of the tree.
We note that both for the CFN model and for the JC model, the transition matrices M e admit a simple
alternative representation. For the CFN model, with probability p(e) = (1 − exp(−2d(e)))/2, there is
a transition and, otherwise, there is no transition. Similarly for the JC model with probability p(e) =
(1 − exp(−4d(e)))/4 each of the three possible transitions occur. In particular, defining
g∗ =
ln 2
,
4
(1)
we may formulate the result on the reconstruction problem for the phase transition of the CFN model as
follows: “If d(e) ≤ g < g ∗ for all e then the reconstruction problem is solvable.”
Phylogenetic Reconstruction Problem. We will be interested in reconstructing phylogenies in the regime
where the reconstruction problem is solvable. The objective is to reconstruct the underlying tree whose
internal nodes are unknown from the collection of sequences at the leaves. Since for both the CFN model
and the JC model, the distribution µ[σ] described above is independent of the location of the root we can
only aim to reconstruct the underlying unrooted topology. Let T represent the set of all binary topologies
(that is, unrooted undirected binary trees) and MCFN
f,g the family of CFN transition matrices, as described
above, which correspond to distances d satisfying:
0 < f ≤ d ≤ g < g∗,
where g ∗ is given by (1) and f is an arbitrary positive constant. Let T ⊗ MCFN
f,g denote the set of all
unrooted phylogenies, where the underlying topology is in T and all transition matrices on the edges are
CFN
in MCFN
f,g . Rooting T ∈ T ⊗ Mf,g at an arbitrary node, let µT be the measure at the leaves of T as
described above. It is well known, e.g. [ESSW99, Cha96], that different elements in T ⊗ MCFN
f,g correspond
to different measures; therefore we will identify measures with their corresponding elements of T ⊗ MCFN
f,g .
1
k
We are interested in finding a (efficiently computable) map Ψ such that Ψ(σ∂ , . . . , σ∂ ) ∈ T , where σ∂ =
k
σ∂i i=1 are k characters at the leaves of the tree. Moreover, we require that for every distribution µT ∈
1
k
T ⊗ MCFN
f,g , if the characters σ∂ , . . . , σ∂ are generated independently from µT , then with high probability
Ψ(σ∂1 , . . . , σ∂k ) = T . The problem of finding an efficiently computable map Ψ is called the phylogenetic
reconstruction problem for the CFN model. The phylogenetic reconstruction problem for the JC model is
defined similarly. In [ESSW99], it is shown that there exists a polynomial-time algorithm that reconstructs
the topology from k = poly(n, log 1/δ) characters, with probability of error δ. Our results are the following.
CFN
We first define a subset of T ⊗ MCFN
f,g . In words, the ∆-Branch Model (∆-BM) is a subset of T ⊗ Mf,g
where the edge lengths d(e), e ∈ E, are discretized. This extra assumption is made for technical reasons.
See Section 7.3.
4
Definition 1.1 (∆-Branch Model) Let ∆ > 0. We denote by Y[∆] the subset of T ⊗ MCFN
f,g where all
d(e)’s are multiples of ∆. We call Y[∆] the ∆-Branch Model (∆-BM).
Theorem 1 (Main Result) Consider the ∆-Branch Model above with: 0 < f ≤ g < g ∗ and ∆ >
0. Then there exists a polynomial-time algorithm that reconstructs the topology of the tree from k =
c(f, g, ∆)(log n + log 1/δ) characters with error probability at most δ. In particular,
c(f, g, ∆) =
c(g)
.
min{∆2 , f 2 }
Moreover, the value g ∗ given by (1) is tight.
Corollary 2 (Jukes-Cantor Model) Consider the JC model on binary trees where all edges satisfy
∗
∗
0 < f ≤ g < gJC
, where gJC
:= g ∗ /2.
Under the ∆-BM, there exists a polynomial-time algorithm that reconstructs the topology of the tree from
c0 (f, g, ∆)(log n + log 1/δ) characters with error probability at most δ and
c0 (f, g, ∆) =
c0 (g)
.
min{∆2 , f 2 }
∗ corresponds to the so-called Kesten-Stigum bound [KS66] on the reconstruction threshold,
The value gJC
which has been conjectured to be tight for the JC model [MM06, Sly08].
Theorem 1 and Corollary 2 extend also to cases where the data at the leaves is given with an arbitrary level
∗ are tight. See [JM04].
of noise. For this robust phylogenetic reconstruction problem both values g ∗ and gJC
The results stated here were first reported without proof in [DMR06]. Note that in [DMR06] the result
were stated without the discretization assumption which is in fact needed for the final step of the proof. This
is further explained in subsection 7.3.
1.2
Organization of the Paper
Roughly speaking, our reconstruction algorithm has two main components. First, the statistical component
consists in
1. [Ancestral Reconstruction] the reconstruction of sequences at internal nodes;
2. [Distance Estimation] the estimation of distances between the nodes.
The former—detailed in Section 2—is borrowed from [Mos04] where Steel’s conjecture is proved for the
special case of balanced trees. In general trees, however, distance estimation is complicated by nontrivial
correlations between reconstruction biases. We deal with these issues in Section 3.
Second, the combinatorial component of the algorithm—which uses quartet-based ideas from phylogenetics and is significantly more involved than [Mos04]—is detailed in Sections 4 and 5. A full description of
the algorithm as well as an example of its execution can also be found in Section 5. Proof of the correctness
of the algorithm is provided in Section 6 and 7.
5
1.3
Notation
Throughout we fix 0 < f ≤ g < g 0 < g ∗ , ∆ > 0, and γ > 3. By definition of g ∗ , we have 2e−4g > 1 and
0
0
2e−4g > 1. We let θ = e−2g and θ0 = e−2g .
2
Reconstruction of Ancestral Sequences
Ancestral Reconstruction. In this section we state the results of [Mos04] on ancestral reconstruction
using recursive majority and we briefly explain how these results are used in the reconstruction algorithm.
Following [Mos04], we use state values ±1 instead of 0/1. Furthermore, we use the parameter θ(e) =
1 − 2p(e) = e−2d(e) . Note that θ(e) measures the correlation between the states at the endpoints of e.
Because the CFN model is ferromagnetic, we have 0 ≤ θ(e) ≤ 1. In terms of θ we have reconstruction
solvability whenever θ(e) ≥ θ > θ∗ , for all edges e, where the value θ∗ satisfies 2θ∗2 = 1.
For the CFN model both the majority algorithm [Hig77] and the recursive majority algorithm [Mos98]1
are effective in reconstructing the root value. Note that for other models, in general, most simple reconstruction algorithms are not effective all the way to the reconstruction threshold [Mos01, MP03, JM04].
However, as noted in [Mos04], there is an important difference between the two reconstruction algorithms
when different edges e have different values of θ(e). Suppose that θ(e) > θ0 > θ∗ for all edges e. Then, the
recursive majority function is effective in reconstructing the root value with probability bounded away from
1/2 (as a function of θ0 ). On the other hand, it is easy to construct examples where the majority function
reconstructs the root with probability tending to 1/2 as the tree size increases.
The difference between the two algorithms can be roughly stated as follows. When different edges have
different θ-values, different “parts” of the tree should have different weights in calculating the reconstructed
value at the root. Indeed, when all θ-values are known, an appropriate weighted majority function can
estimate the root value correctly with probability bounded away from 1/2 [MP03]. However, when the
θ-values are unknown, using uniform weights may result in an arbitrarily inaccurate estimator.
Recursive majority, on the other hand, does not require knowledge of the θ-values to be applied successfully. This essentially follows from the fact that the majority function is “noise-reducing” in the following
sense. Suppose θ0 > θ∗ . Then, as we shall see shortly, there exists an integer ` and noise level q < 1/2 such
that majority on the `-level binary tree has the following property: if all leaf values are given with stochastic
noise at most q, then the majority of these values differs from the actual root state with probability at most q.
Therefore, recursive application of the majority function—` levels at a time—results in an estimator whose
error is at most q for any number of levels.
d :
Properties of Recursive Majority. We proceed with a formal definition of recursive majority. Let Maj
d
{−1, 1} → {−1, 1} be the function defined as follows
!
d
X
d 1 , . . . , xd ) = sign
Maj(x
xi + 0.5ω ,
i=1
d outputs the majority
where ω is ±1 with probability 1/2 independently of the xi ’s. In other words, Maj
value of its input arguments, unless there is a tie in which case it outputs ±1 with probability 1/2 each.
For consistency, we denote all statistical estimators with a hat. Note in particular that our notation differs
from [Mos04].
1
See below for a definition of these estimators.
6
Next we define the “noise-reduction” property of majority. The function η below is meant to measure
the noise level at the leaves.
Definition 2.1 (Correlation at the root) Let T = (V, E) be a tree rooted at ρ with leaf set ∂T . For functions θ̃ : E → [0, 1] and η̃ : ∂T → [0, 1], let CFN(θ̃, η̃) be the CFN model on T where θ(e) = θ̃(e) for all
e which are not adjacent to ∂T , and θ(e) = θ̃(e)η̃(v) for all e = (u, v), with v ∈ ∂T . Let
h
i
h
i
d ∂T ) σρ = +1 = E −Maj(σ
d ∂T ) σρ = −1 ,
MajCorr(θ̃, η̃) = E +Maj(σ
where σ is one sample drawn from CFN(θ̃, η̃).
2
Proposition 2.2 (Noise reduction of majority [Mos04]) Let b and θmin be such that bθmin
> h2 > 1.
`
Then there exist ` = `(b, θmin ), α = α(b, θmin ) > h > 1 and β = β(b, θmin ) > 0, such that any
CFN(θ, η) model on the `-level b-ary tree satisfying mine∈E θ(e) ≥ θmin and minv∈∂T η(v) ≥ ηmin > 0
also satisfies:
MajCorr(θ, η) ≥ min{αηmin , β}.
(2)
In particular if ηmin ≥ β then:
MajCorr(θ, η) ≥ β.
(3)
General Trees. Recursive application of Proposition 2.2 allows the reconstruction of the root value on any
balanced binary tree with correlation at least β. However, below we consider general trees. In particular,
when estimating the sequence at an internal node u of the phylogenetic tree, we wish to apply Proposition 2.2
to a subtree rooted at u and this subtree need not be balanced. This can be addressed by “completing” the
subtree into a balanced tree (with number of levels a multiple of `) and taking all added edges to have length
d(e) = 0, that is, θ(e) = 1. Fix ` and β so as to satisfy Proposition 2.2 with b = 2 and θmin = θ. (Recall
that for the proof of Theorem 1 we assume θ(e) ≥ θ > θ∗ , for all edges e.) Consider the following recursive
function of x = (x1 , x2 , . . .):
d 0` (x1 ) = x1 ,
Maj
d j` (x1 , . . . , x2j` ) = Maj
d Maj
d j−1
d j−1
Maj
` (x1 , . . . , x2(j−1)` ), . . . , Maj` (x2j` −2(j−1)` +1 , . . . , x2j` ), ,
d ρ,T (σ∂T ) that
for all j ≥ 1. Now, if σ∂T is a character at the leaves of T , let us define the function Anc
estimates the ancestral state at the root ρ of T using recursive majority as follows:
1. Let Te be the tree T (minimally) completed with edges of θ-value 1 so that Te is a complete binary tree
with a number of levels a multiple of `, say J`;
2. Assign to the leaves of Te the value of their ancestor leaf in T under σ∂T ;
3. Let σ̃ be the leaf states of Te arranged in pre-order;
4. Compute
d ρ,T (σ∂T ) := Maj
d J` (σ̃).
Anc
From Proposition 2.2, we get:
7
Proposition 2.3 (Recursive Majority) Let T = (V, E) be a tree rooted at ρ with leaf set ∂T . Let σ∂T be
one sample drawn from CFN(θ, η) with θ(e) ≥ θ > θ∗ for all edges e and η(v) = 1 for all v ∈ ∂T . Then,
we can choose ` and β > 0 so that
1+β
d
d
P Ancρ,T (σ∂T ) = σρ σρ = +1 = P Ancρ,T (σ∂T ) = σρ σρ = −1 ≥
.
(4)
2
k
i
is
In the remainder of the paper, we will use Proposition 2.3 with θ(e) = e−2d(e) . Also, if σ∂T = σ∂T
i=1
d
a collection of k characters at the leaves of T , we extend the function Ancρ,T to collections of characters in
the natural way:
k
d ρ,T (σ∂T ) := Anc
d ρ,T (σ i )
.
(5)
Anc
∂T
3
i=1
Distance Estimation
Throughout this section, we fix a tree T on n leaves. Recall the definition of our path metric d from
k
t
Section 1.1. We assume that σ∂T = σ∂T
are k i.i.d. samples (or characters) at the leaves of T
t=1
generated by the CFN model with parameters d(e) ≤ g for all e ∈ E(T ) (a lower bound on d is not required
k
in this section, but it will be in the next one). We think of σ∂T = (σut )u∈∂T t=1 as n sequences of length k
and we sometimes refer to σu = (σut )kt=1 as the “sequence at u”.
Distance between leaves. As explained in Section 1.1, a basic concept in phylogenetic reconstruction
is the notion of a metric on the leaves of the tree. The distance between two leaves is a measure of the
correlation between their respective sequences. We begin with the definition of a natural distance estimator.
Definition 3.1 (Correlation between sequences) Let u, v be leaves of T . The following quantity
"
! #
k
X
1
1
d (σu , σv ) = − ln
Dist
σut σvt
,
2
k
t=1
(6)
+
d (σu , σv ) =
is an estimate of d(u, v). In cases where the sum inside the log is non-positive, we define Dist
+∞.
d The proof follows from standard
The next proposition provides a guarantee on the performance of Dist.
concentration inequalities. See, e.g., [ESSW99]. An important point to note is that, if we use short sequences
of length O(log n), the guarantee applies only to short distances, that is, distances of order O(1).
d For all ε > 0, M > 0, there exists c = c(ε, M ; γ) > 0, such that if
Proposition 3.2 (Accuracy of Dist)
the following conditions hold:
• [Short Distance] d(u, v) < M ,
• [Sequence Length] k = c0 log n, for some c0 > c,
then
d u , σv ) < ε,
d(u, v) − Dist(σ
with probability at least 1 − O(n−γ ).
8
Proof: By Azuma’s inequality we get
" k
#
h
i
X
1
t
t
−2d(u,v)−2ε
d u , σv ) > d(u, v) + ε = P
P Dist(σ
σu σv < e
k
t=1
" k
#
X
1
σut σvt < e−2d(u,v) − (1 − e−2ε )e−2d(u,v)
= P
k
t=1
#
" k
1X t t
1 1
−2ε −2M
σu σv < E[σu σv ] − (1 − e )e
≤ P
k
t=1
2 !
(1 − e−2ε )e−2M
≤ exp −
2k(2/k)2
0
= n−c K ,
with K depending on M and ε. Above, we used that
1
d(u, v) = − ln E[σu1 σv1 ].
2
A similar inequality holds for the other direction.
Distance Between Internal Nodes. Our reconstruction algorithm actually requires us to compute distances between internal nodes. Note that, in general, we do not know the true sequences at internal nodes of
the tree. Therefore, we need to estimate distances by applying (6) to reconstructed sequences. An obvious
issue with this idea is that reconstructed sequences are subject to a systematic bias.
d v,Tv (σ∂T ). Then, the
Definition 3.3 (Bias) Suppose v is the root of a subtree Tv of T . Let σ
bv = Anc
v
quantity
1
λ(Tv , v) = − ln(E[σv1 · σ
bv1 ]+ ),
2
is called the reconstruction bias at v on Tv . (Note the similarity to (6).) We denote
B(g̃) = sup λ(Tv , v),
where the supremum is taken over all rooted binary trees with edge lengths at most g̃. Note that for all g̃,
0 < g̃ < g ∗ , we have by Proposition 2.3
0 < B(g̃) < +∞.
We also denote
β(g̃) = e−2B(g̃) .
Since B(g) > 0, we cannot estimate exactly the internal sequences. Hence Proposition 3.2 cannot be
used directly to estimate distances between internal nodes. We deal with this issue as follows. Consider the
configuration in Figure 3.1. More precisely:
Definition 3.4 (Basic Disjoint Setup (Preliminary Version)) Root T at an arbitrary vertex. Note that, by
reversibility, the CFN model on a tree T can be rerooted arbitrarily without changing the distribution at
the leaves. Denote by Tx the subtree of T rooted at x. We consider two internal nodes u1 , u2 that are
not descendants of each other in T . Denote by vϕ , wϕ the children of uϕ in Tuϕ , ϕ = 1, 2. We call this
configuration the Basic Disjoint Setup (Preliminary Version).
9
u1
v1
Tv1
u2
w1
v2
Tw1
Tv2
w2
Tw2
c and routines I S S PLIT in Figure 4.1 and I S S HORT in
Figure 3.1: Basic disjoint setup for estimator Int
Figure 3.6. The subtrees Tv1 , Tw1 , Tv2 , Tw2 are edge-disjoint (see Definition 3.9 below).
We need a few basic combinatorial definitions.
Definition 3.5 (Restricted Subtree) Let V 0 ⊆ V be a subset of the vertices of T . The subtree of T restricted
to V 0 is the tree T 0 obtained by 1) keeping only nodes and edges on paths between vertices in V 0 and 2) by
then contracting all paths composed of vertices of degree 2, except the nodes in V 0 . We sometimes use the
notation T |V 0 := T 0 . See Figure 3.2 for an example.
Definition 3.6 (Quartet) A quartet is a set of four leaves. More generally, for any set Q = {v1 , w1 , v2 , w2 }
of four nodes of T , we think of Q as a quartet on T |Q . We say that a quartet Q = {v1 , w1 , v2 , w2 } is
nondegenerate if none of the nodes in Q are on a path between two other nodes in Q. There are three
possible (leaf-labeled) topologies on a nondegenerate quartet, called quartet splits, one for each partition
of the four leaves into two pairs. In Figure 3.1, the correct quartet split on Q is {{v1 , w1 }{v2 , w2 }} which
we denote by v1 w1 |v2 w2 .
To compute the distance between u1 and u2 , we think of the path between u1 and u2 as the internal edge
of the quartet Q = {v1 , w1 , v2 , w2 }. The reconstructed sequences at x ∈ Q also suffer from a systematic
bias. However, we prove in Proposition 3.8 that the bias does not affect the computation of the length of
the internal edge of the quartet. Indeed, as depicted in Figure 3.3, we treat the systematic errors introduced
by the reconstruction as “extra edges” attached to the nodes in Q = {v1 , w1 , v2 , w2 }, with corresponding
b = {v̂1 , ŵ1 , v̂2 , ŵ2 }. Our estimator Int
c is then obtained from a classical distance-based
“endpoints” Q
b
computation—known in phylogenetics as the Four-Point Method [Bun71]—applied to the “extra nodes” Q
(see (7) below). For this idea to work, it is crucial that the systematic error in the reconstructed sequences
at x ∈ Q be independent given the true sequences at x ∈ Q. This is the case when u1 and u2 are not
descendant of each other because of the Markov property.
c We let σ
d x,Tx (σ∂T ) be the reconstructed sequence at
We now define our distance estimator Int.
bx = Anc
x
x.
10
Figure 3.2: Restricting the top tree to its white nodes.
u1
v1
v̂1
u2
w1
v2
ŵ1
v̂2
w2
ŵ2
c where reconstruction bias is represented by dashed edges.
Figure 3.3: Basic setup for estimator Int
11
c Consider the Basic Disjoint Setup (Preliminary Version). Then,
Definition 3.7 (Distance Estimator Int)
we let
1 d
d σw , σ
d σv , σ
d σv , σ
c (b
Dist(b
σ v1 , σ
bv2 ) + Dist(b
bw2 ) − Dist(b
bw1 ) − Dist(b
bw2 ) .
Int
σ v1 , σ
bw1 ; σ
bv2 , σ
bw2 ) =
1
1
2
2
(7)
d quantities is +∞, the function Int
c is set to +∞.
If one of the Dist
c As in Proposition 3.2, the guaranThe next proposition provides a guarantee on the performance of Int.
tee applies only to O(1) distances, if we use sequences of length O(log n).
c Consider the Basic Disjoint Setup (Preliminary Version). Let Q =
Proposition 3.8 (Accuracy of Int)
{v1 , w1 , v2 , w2 }. For all ε > 0, M > 0, there exists c = c(ε, M ; γ, g) > 0 such that, if the following
hold:
• [Edge Lengths] For all e ∈ E(T ), d(e) ≤ g < g ∗ ;
• [Short Distances] We have d(x, y) < M , ∀x, y ∈ Q;
• [Sequence Length] The sequences used have length k = c0 log n, for some c0 > c;
then, with probability at least 1 − O(n−γ ),
c (b
d(u1 , u2 ) − Int
σv1 , σ
bw1 ; σ
bv2 , σ
bw2 ) < ε.
Proof: First note that, for all t ∈ {1, . . . , k}, conditioned on {σxt }x∈Q , the values {b
σxt }x∈Q are independent
by the Markov property. Therefore, as we observed before, the systematic errors introduced by the reconb = {v̂1 , ŵ1 , v̂2 , ŵ2 } (see Figure 3.3).
struction process can be treated as “extra edges”, with endpoints Q
Moreover, from (4), it follows that, ∀x ∈ Q, ∀t ∈ {1, . . . , k},
t
1 + β(g)
P σ
bx = σxt |σxt ≥
,
2
(8)
where β(g) > 0 is specified by Definition 3.3. From (8), the lengths of the “extra edges”—equal to the
reconstruction biases—are at most B(g).
Appealing to Proposition 3.2 with constants 2ε and M + 2B(g) and the union bound, we argue that with
probability at least 1 − O(n−γ ) for all x, y ∈ Q
ε
d σx , σ
Dist(b
by ) − d(x̂, ŷ) < ,
2
where we extend d to the “extra nodes” x̂, ŷ by setting
1
d(x, x̂) = − ln(E[σx σ
bx ]+ ),
2
and likewise for ŷ (see Figure 3.3). Observing that d(u1 , u2 ) = 21 (d(v̂1 , v̂2 ) + d(ŵ1 , ŵ2 ) − d(v̂1 , ŵ1 ) − d(v̂2 , ŵ2 ))
and using the above we get
c (b
d(u1 , u2 ) − Int
σv1 , σ
bw1 ; σ
bv2 , σ
bw2 ) < ε.
12
Algorithm D ISTANCE E STIMATE
bacc ;
Input: Two nodes u1 , u2 ; a rooted forest F; accuracy radius R
Output: Distance estimate ν;
• [Children] Let wϕ , vϕ be the children of uϕ in F for ϕ = 1, 2 (if uϕ is a leaf, set wϕ = vϕ = uϕ );
d x,T F (σ∂T F );
• [Sequence Reconstruction] For x ∈ {v1 , w1 , v2 , w2 }, set σ
bx = Anc
x
x
• [Accuracy Cutoff] If there is {x, y} ⊆ {v1 , w1 , v2 , w2 } such that
d (b
bacc ,
Dist
σx , σ
by ) > R
return +∞;
• [Distance Computation] Otherwise return
c (b
ν = Int
σ v1 , σ
bw1 ; σ
bv2 , σ
bw2 ) .
Figure 3.4: Routine D ISTANCE E STIMATE.
u3
u2
u4
u1
u5
u7
u8
u6
Figure 3.5: The subtrees T |{u1 ,u2 ,u3 ,u8 } and T |{u4 ,u5 ,u6 ,u7 } are edge-disjoint. The subtrees T |{u1 ,u5 ,u6 ,u8 }
and T |{u2 ,u3 ,u4 ,u7 } are edge-sharing.
Distances Between Restricted Subtrees. In fact, we need to apply Proposition 3.8 to restricted subtrees
of the true tree. Indeed, in the reconstruction algorithm, we maintain a “partially reconstructed subforest” of
the true tree, that is, a collection of restricted subtrees of T . For this more general setup, we use the routine
D ISTANCE E STIMATE detailed in Figure 3.4.
To generalize Proposition 3.8, we need a few definitions. First, the notion of edge disjointness is borrowed from [Mos07].
Definition 3.9 (Edge Disjointness) Denote by pathT (x, y) the path (sequence of edges) connecting x to y
in T . We say that two restricted subtrees T1 , T2 of T are edge disjoint if
pathT (x1 , y1 ) ∩ pathT (x2 , y2 ) = ∅,
for all x1 , y1 ∈ L(T1 ) and x2 , y2 ∈ L(T2 ). We say that T1 , T2 are edge sharing if they are not edge disjoint.
See Figure 3.5 for an example. (If T1 and T2 are directed, we take this definition to refer to their underlying
undirected version.)
13
Definition 3.10 (Legal Subforest) We say that a tree is a rooted full binary tree if all its internal nodes have
degree 3 except for the root which has degree 2. A restricted subtree T1 of T is a legal subtree of T if:
1. It is a rooted full binary tree, that is, if it has a unique node of degree 2—its root;
2. And if all its leaves are leaves of T .
Moreover, we say that a forest
F = {T1 , . . . , Tα },
is legal subforest of T if the Tϕ ’s are edge-disjoint legal subtrees of T . We denote by ρ(F) the set of roots
of F.
Definition 3.11 (Dangling Subtrees) We say that two edge-disjoint legal subtrees T1 , T2 of T are dangling
if there is a choice of root ρ∗ for T not in T1 or T2 that is consistent with the rooting of both T1 and T2 , that
is, pointing the edges of T away from ρ∗ is such that the edges of T1 and T2 are directed away from their
respective roots.
See Figure 3.7 below for an example where two legal, edge-disjoint subtrees are not dangling. We generalize
the basic configuration of Figure 3.1 as follows:
Definition 3.12 (Basic Disjoint Setup (Dangling)) Let T1 = Tu1 and T2 = Tu2 be two legal subtrees of
T rooted at u1 and u2 respectively. Assume further that T1 and T2 are edge-disjoint and dangling. Denote
by vϕ , wϕ the children of uϕ in Tϕ , ϕ = 1, 2. If uϕ is a leaf, we let instead vϕ = wϕ = uϕ . We call this
configuration the Basic Disjoint Setup (Dangling).
We make another important remark about the more general setup considered here. Note that even though
the true tree satisfies the “short edge length” condition (that is, d(e) ≤ g) by assumption, partially reconstructed subtrees may not—because their edges are actually paths in the true tree. Therefore, an important
step of the reconstruction algorithm is to make sure that all restricted edges of the partially reconstructed
subforest are short enough for recursive majority to be accurate. (This explains why we need to use two
constants g < g 0 below g ∗ .) Refer to routine L OCAL C HERRY in Figure 5.4 in Section 5.
We then obtain the following generalization of Proposition 3.8. Note that the routine D ISTANCE E S bacc . As we show in the proof below, this ensures that
in Figure 3.4 uses an “accuracy cutoff” R
distances that are too long are rejected, in which case the value +∞ is returned.
TIMATE
Proposition 3.13 (Accuracy of D ISTANCE E STIMATE) Consider the Basic Disjoint Setup (Dangling). Let
bacc > M +2B(g 0 ), there exists c = c(ε, M, R
bacc ; γ, g 0 ) >
Q = {v1 , w1 , v2 , w2 }. For all ε > 0, M > 0, and R
0 such that, if the following hold:
• [Edge Length] It holds that d(e) ≤ g 0 < g ∗ , ∀e ∈ E(Tx ), x ∈ Q;
• [Sequence Length] The sequences used have length k = c0 log n, for some c0 > c;
then, with probability at least 1 − O(n−γ ), the following holds: letting ν be the output of D ISTANCE E STI MATE in Figure 3.4, we have that if one of the following hold
1. [Short Distances] We have d(x, y) < M , ∀x, y ∈ Q;
2. [Finite Estimate] ν < +∞;
14
then
|d(u1 , u2 ) − ν| < ε.
Proof: The first half of the proposition follows immediately from Propositions 3.2 and 3.8. Refer to Figure
3.3.
The second part follows from a “double window” argument as in [ESSW99, Theorem 9]. Let 0 < τ <
such that
bacc = M + 2B(g 0 ) − 1 ln(2τ ).
R
2
By assumption, ν < +∞ and therefore for all pairs {x, y} ⊆ {v1 , w1 , v2 , w2 } we have
1
d (b
Dist
σx , σ
by ) ≤ M + 2B(g 0 ) − ln(2τ ).
2
It follows directly from Azuma’s inequality that a pair {x, y} such that
d(x̂, ŷ) ≥ M + 2B(g 0 ) −
1
2
(9)
1
ln τ,
2
satisfies (9) with probability at most
" k
#
X
1
1
0
d (b
P Dist
σx , σ
by ) ≤ M + 2B(g 0 ) − ln 2τ
= P
σ
bxt σ
byt ≥ 2τ e−2[M +2B(g )]
2
k
t=1
#
" k
1X t t
−2d(x̂,ŷ)
−2[M +2B(g 0 )]
σ
bx σ
by ≥ e
+ τe
≤ P
k
t=1
2
−2[M +2B(g 0 )]
τ
e
≤ exp −
2k(2/k)2
0
= n−c K ,
for a constant K > 0 depending on M, g 0 , τ . The result follows by applying the first part of the proposition
with M replaced by M + 2B(g 0 ) − 12 ln τ . (Note that d(x, y) ≤ d(x̂, ŷ), for all x, y ∈ Q.)
We also need a simpler variant of D ISTANCE E STIMATE (Figure 3.4) whose purpose is to test whether
the internal path of a quartet is longer than g. We record this variant in Figure 3.6.
Detecting Long Distances When T1 And T2 Are Not Dangling. Roughly speaking, our reconstruction
algorithm works by progressively merging subtrees that are close in the true tree. (See Section 5 for further
details.) Hence, the algorithm needs to tell whether or not two subtrees are sufficiently close to be considered
for this merging operation. However, as we explain in the next sections, we cannot guarantee that the Basic
Disjoint Setup (Dangling) in Proposition 3.13 applies to all situations encountered during the execution of
the algorithm. Instead we use a special “multiple test” to detect long distances. This test is performed by
the routine D ISTORTED M ETRIC detailed in Figure 3.8.
The routine has a further important property. During the course of the algorithm, since we only have
partial information about the structure of the tree, we may not always know whether or not two subtrees
are dangling—and therefore whether or not D ISTANCE E STIMATE in Figure 3.4 returns accurate distance
estimates. The “multiple test” in D ISTORTED M ETRIC is such that, if the routine returns a finite estimate,
that estimate is accurate. We proceed with an explanation of these properties.
We further generalize the basic configuration of Figure 3.1 as follows:
15
Algorithm I S S HORT
bacc ; tolerance ε;
Input: Two pairs of nodes (v1 , w1 ), (v2 , w2 ); rooted forest F; accuracy radius R
Output: Boolean value and length estimate;
d x,T F (σ∂T F );
• [Sequence Reconstruction] For x ∈ {v1 , w1 , v2 , w2 }, set σ
bx = Anc
x
x
• [Accuracy Cutoff] If there is {x, y} ⊆ {v1 , w1 , v2 , w2 } such that
d (b
bacc ,
Dist
σx , σ
by ) > R
return +∞;
• [Internal Edge Length] Set
c (b
ν = Int
σ v1 , σ
bw1 ; σ
bv2 , σ
bw2 ) ;
• [Test] If ν < g + ε return (TRUE, ν), o.w. return (FALSE, 0);
Figure 3.6: Routine I S S HORT.
Definition 3.14 (Basic Disjoint Setup (General)) Let T1 = Tx1 and T2 = Tx2 be two restricted subtrees
of T rooted at x1 and x2 respectively. Assume further that T1 and T2 are edge-disjoint, but not necessarily
dangling. Denote by yϕ , zϕ the children of xϕ in Tϕ , ϕ = 1, 2. Let wϕ be the node in T where the path
between T1 and T2 meets Tϕ , ϕ = 1, 2. Note that wϕ may not be in Tϕ since Tϕ is restricted, ϕ = 1, 2.
If wϕ 6= xϕ , assume without loss of generality that wϕ is in the subtree of T rooted at zϕ , or on the edge
(xϕ , zϕ ), ϕ = 1, 2. We call this configuration the Basic Disjoint Setup (General). See Figure 3.7. Let
d(T1 , T2 ) be the length of the path between w1 and w2 (in the path metric d).
The key point to note is that when computing the distance between y1 and y2 rather than the distance
between x1 and x2 , then the assumptions of Proposition 3.13 are satisfied (as the claim holds in this case
by rooting the tree at any node along the path connecting w1 and w2 ). Hence, if T1 and T2 are far apart,
the distance between y1 and y2 is correctly estimated as being large. On the other hand, if T1 and T2 are
dangling and close, then distances between all pairs in {y1 , z1 } × {y2 , z2 } are accurately estimated as being
short.
Proposition 3.15 (Accuracy of D ISTORTED M ETRIC) Consider the Basic Disjoint Setup (General) with
bacc > M + 2B(g 0 ) + 4g 0 there exists
F = {T1 , T2 } and Q = {y1 , z1 , y2 , z2 }. For all ε > 0, M > 0, and R
0
bacc ; γ, g ) > 0 such that, if the following hold:
c = c(ε, M, R
• [Edge Length] It holds that d(e) ≤ g 0 < g ∗ , ∀e ∈ E(Tx ), x ∈ Q; also, d(xϕ , yϕ ), d(xϕ , zϕ ) ≤ g 0 ,
ϕ = 1, 2;
• [Weight Estimates] We are given weight estimates
h : E(F) → R+ ,
such that |h(e) − d(e)| < ε/16, ∀e ∈ E(F);
• [Sequence Length] The sequences used have length k = c0 log n, for some c0 > c;
then, with probability at least 1 − O(n−γ ), the following holds: letting ν be the output of D ISTORTED M ETRIC in Figure 3.8, we have that if one of the following hold
1. [Dangling Case] T1 and T2 are dangling and d(T1 , T2 ) < M ;
16
x2
z2
y2
x1
y1
z1
u2
w2
v2
u1
w1
v1
T2
T1
Figure 3.7: Basic Disjoint Setup (General). The rooted subtrees T1 , T2 are edge-disjoint but are not assumed
to be dangling. The white nodes may not be in the restricted subtrees T1 , T2 . The case w1 = x1 and/or
w2 = x2 is possible. Note that if we root the tree at any node along the dashed path, the subtrees rooted at
y1 and y2 are edge-disjoint and dangling (unlike T1 and T2 ).
2. [Finite Estimate] ν < +∞;
then
|ν − d(x1 , x2 )| < ε.
Proof: Choose c as in Proposition 3.13 with parameters ε/16 and M + 4g 0 . As in Figure 3.8, let
bacc ) − h(x1 , r1 ) − h(x2 , r2 ),
D0 (r1 , r2 ) := D ISTANCE E STIMATE(r1 , r2 ; F; R
for all pairs (r1 , r2 ) ∈ {y1 , z1 } × {y2 , z2 } (where D ISTANCE E STIMATE is defined in Figure 3.4). In Case
1, the result follows directly from Proposition 3.13 and the remarks above Proposition 3.15. In particular,
note that, with probability 1 − O(n−γ ),
D0 (r1 , r2 ) − d(x1 , x2 ) <
3ε
,
16
for all (r1 , r2 ) ∈ {y1 , z1 } × {y2 , z2 } and therefore
max
n
D
0
(1) (1)
(r1 , r2 )
−D
0
(2) (2)
(r1 , r2 )
:
(ϕ) (ϕ)
(r1 , r2 )
o
3ε
∈ {y1 , z1 } × {y2 , z2 }, ϕ = 1, 2 < 2
< ε/2,
16
and D ISTORTED M ETRIC returns a finite value (namely ν := D0 (z1 , z2 ); see Figure 3.8) that is accurate
within 3ε/16 < ε.
In Case 2, the condition ν < +∞ implies in particular that all four distance estimates are equal up
to ε/2. From the remark above the statement of the proposition, at least one such distance, say D0 (y1 , y2 )
w.l.o.g., is computed within the Basic Disjoint Setup (Dangling) and we can therefore apply Proposition 3.13
again. In particular, we have
3ε
D0 (y1 , y2 ) − d(x1 , x2 ) < ,
16
17
and therefore
|ν − d(x1 , x2 )| <
ε 3ε
+
< ε,
2 16
(where again ν := D0 (z1 , z2 )).
Remark. Note that D ISTORTED M ETRIC outputs +∞ in two very distinct situations. In particular, if
D ISTORTED M ETRIC outputs +∞, then either the subtrees are too far to obtain an accurate estimate or T1
and T2 are not dangling (or both). This convention will turn out to be convenient in the statement of the full
algorithm.
Algorithm D ISTORTED M ETRIC
bacc ; tolerance ε;
Input: Two nodes x1 , x2 ; a rooted forest F; edge lengths {h(e)}e∈F ; accuracy radius R
Output: Distance ν;
• [Children] Let yϕ , zϕ be the children of xϕ in F for ϕ = 1, 2 (if xϕ is a leaf, set zϕ = yϕ = xϕ );
• [Distance Computations] For all pairs (r1 , r2 ) ∈ {y1 , z1 } × {y2 , z2 }, compute
bacc ) − h(x1 , r1 ) − h(x2 , r2 );
D0 (r1 , r2 ) := D ISTANCE E STIMATE(r1 , r2 ; F; R
• [Multiple Test] If
n
o
(1) (1)
(2) (2)
(ϕ) (ϕ)
max D0 (r1 , r2 ) − D0 (r1 , r2 ) : (r1 , r2 ) ∈ {y1 , z1 } × {y2 , z2 }, ϕ = 1, 2 < ε/2,
return ν := D0 (z1 , z2 ), otherwise return ν := +∞ (return ν := +∞ if any of the distances above is +∞).
Figure 3.8: Routine D ISTORTED M ETRIC.
4
Quartet Tests
Before giving a complete description of the reconstruction algorithm, we introduce some important combinatorial tools. As we discussed previously, we make use of a key concept from phylogenetics—the notion of
a quartet. In the previous section, we showed how to estimate accurately distances between internal nodes
of the tree. In this section, we explain how to use such estimates to perform topological tests on quartets.
Those tests form the basic building blocks of the combinatorial component of the algorithm. As before, we
fix a tree T on n leaves. Moreover, we assume that our path metric d on T satisfies d(e) ≥ f for all e ∈ T .
Note that an upper bound on d is not explicitly required in this section since we assume that appropriate
distance estimates are handed to us.
Splits. The routine I S S PLIT in Figure 4.1 performs a classic test to decide the correct split of a quartet. It is
based on the so-called Four-Point Method [Bun71]. Proposition 4.1 guarantees the correctness of I S S PLIT.
Its proof is omitted. We consider once again the Basic Disjoint Setup (Dangling) of Section 3.
Proposition 4.1 (Correctness of I S S PLIT) Consider the Basic Disjoint Setup (Dangling). Let
Q = {v1 , w1 , v2 , w2 },
18
Algorithm I S S PLIT
Input: Two pairs of nodes (v1 , w1 ), (v2 , w2 ); a distance matrix D on these four nodes;
Output: TRUE or FALSE;
• [Internal Edge Length] Set
ν=
1
(D(w1 , w2 ) + D(v1 , v2 ) − D(w1 , v1 ) − D(w2 , v2 )) ;
2
(set ν = +∞ if any of the distances is +∞)
• [Test] If ν < f /2 return FALSE, o.w. return TRUE.
Figure 4.1: Routine I S S PLIT.
and let D be the distance matrix on the four nodes of Q passed to I S S PLIT in Figure 4.1. Assume that
d(e) ≥ f for all edges on the subtree restricted to Q. If
|d(x, y) − D(x, y)| <
f
,
4
∀x, y ∈ Q,
then the call I S S PLIT((v1 , w1 ), (v2 , w2 ); D) returns TRUE, whereas the calls I S S PLIT((v1 , w2 ), (v2 , w1 ); D)
and I S S PLIT((v1 , v2 ), (w1 , w2 ); D) return FALSE.
Collisions. As we discussed before, for routines D ISTANCE E STIMATE (Figure 3.4) and I S S PLIT (Figure 4.1) to work, we need a configuration as in Figure 3.1 where two edge-disjoint subtrees are connected
by a path that lies “above” them. However, for reasons that will be explained in later sections (see also the
discussion in Section 3), in the course of the reconstruction algorithm we have to deal with configurations
where two subtrees are not dangling, as in Figure 3.7. We use the following definitions. Recall the Basic
Disjoint Setup (General) in Figure 3.7.
Definition 4.2 (Collisions) Suppose that T1 and T2 are legal subtrees of T and suppose they are not dangling. We say that T1 collides into T2 at edge e2 = (u2 , v2 ) (u2 is the parent of v2 in T2 ), if the path
pathT (ρ(T1 ), ρ(T2 ))
has non-empty intersection with edge e2 (i.e., with pathT (u2 , v2 )) but with no other edge in the subtree of
T2 rooted at v2 . See Figure 3.7. We sometimes say that the trees T1 , T2 collide. We say that the collision is
within distance M if d(T1 , T2 ) ≤ M . (See Definition 3.14.)
An important step of the reconstruction algorithm is to detect collisions, at least when they are within a
short distance. Routine I S C OLLISION defined in Figure 4.3 and analyzed in Proposition 4.4 below performs
this task. We consider the following configuration, which we call the “basic collision setup.”
Definition 4.3 (Basic Collision Setup) We have two legal subtrees Tx0 and Tu rooted at x0 and u in T . We
let v, w be the children of u. We assume that we are in either of the configurations depicted in Figure 4.2,
that is:
a. Either the path between Tx0 and Tu attaches in the middle of the edge (u, v)—in other words, Tx0
collides into Tu at edge (u, v);
b. Or it goes through u—in other words, Tx0 and Tu are dangling.
19
x0
x0
u
v
u
v
w
v1
v2
v1
a) with collision
w
v2
b) without collision
Figure 4.2: Basic setup of the I S C OLLISION routine in Figure 4.3. The subtrees Tx0 and Tu are edgedisjoint. The configuration on the left shows a collision, whereas the one on the right does not. Note that in
the collision case, the internal path of the quartet {x0 , w, v1 , v2 } is shorter than the edge (u, v). This is the
basic observation behind the routine I S C OLLISION. We call x0 the “reference point”.
Algorithm I S C OLLISION
Input: Four nodes x0 , v, w, u; an edge length h; a rooted forest and a distance matrix (F, D);
Output: TRUE or FALSE;
• [Children] Let v1 , v2 be the children of v in F (or v1 = v2 = v if v is a leaf);
• [Internal Edge Length] Set
ν=
1
(D(v1 , x0 ) + D(v2 , w) − D(v1 , v2 ) − D(x0 , w)) ;
2
(set ν = +∞ if any of the distances is +∞)
• [Test] If (h − ν) > f /2 return TRUE, else return FALSE;
Figure 4.3: Routine I S C OLLISION.
We call x0 the “reference point”.
The purpose of I S C OLLISION is to distinguish between the two configurations above.
Proposition 4.4 (Correctness of I S C OLLISION) Consider the Basic Collision Setup. In particular, assume that one of the two configurations in Figure 4.2 holds. Let Q = {x0 , v1 , v2 , w} and let D and h be the
parameters passed to I S C OLLISION in Figure 4.3. Assume that all edges in the tree satisfy d(e) ≥ f . If
|d(x, y) − D(x, y)| <
f
,
8
∀x, y ∈ Q,
and
f
,
4
then the routine I S C OLLISION returns TRUE if and only if Tx0 collides into Tu at edge (u, v).
|h − d(u, v)| <
Proof: Whether or not there is a collision, the quantity ν computed in the routine is the length of the internal
path of the quartet split x0 w|v1 v2 . If there is not a collision, then this path is actually equal to the path
20
corresponding to the edge (u, v). Otherwise, the length of the path is shorter than d(u, v) by at least f by
assumption. (See Figure 4.2.) The proof follows.
5
Reconstruction Algorithm
We proceed with a formal description of the reconstruction algorithm. A detailed example can be found in
Section 5.2. The reader may want to take a look at the example before reading the details of the algorithm.
5.1
Description of the Algorithm
Cherry Picking. Recall that in a 3-regular tree a cherry is a pair of leaves at graph distance 2. Roughly
speaking, our reconstruction algorithm proceeds from a simple idea: it builds the tree one layer of cherries
at a time. To see how this would work, imagine that we had access to a “cherry oracle,” that is, a function
C(u, v, T ) that returns the parent of the pair of leaves {u, v} if the latter forms a cherry in the tree T (and
say 0 otherwise). Then, we could perform the following “cherry picking” algorithm:
• Currently undiscovered tree:
T 0 := T ;
• Repeat until T 0 is empty,
– For all (u, v) ∈ L(T 0 ) × L(T 0 ), if
w := C(u, v, T 0 ) 6= 0, set Parent(u) := Parent(v) := w;
– Remove from T 0 all cherries discovered at
this step.
Unfortunately, the cherry oracle cannot be simulated from short sequences at the leaves. Indeed, as we
discussed in Section 3, short sequences provide only “local” metric information on the structure of the tree.
See the example in Section 5.2 for an illustration of the problems that may arise. Nevertheless, the above
scheme can be roughly followed by making a number of modifications which we now describe.
The high-level idea of the algorithm, which we call B LINDFOLDED C HERRY P ICKING (BCP), is to
apply the cherry picking scheme above with two important differences:
• [Reconstructed Sequences] Leaf sequences provide only local metric information “around the leaves.”
To gain information about higher, internal nodes of the tree, we reconstruct sequences at the internal
nodes of our partially reconstructed subforest and compute local metric information “around these
nodes.” By repeating this process, we gain information about higher and higher parts of the tree.
• [Fake Cherries] Moreover, because of the local nature of our information, some of the cherries we
pick may in fact turn out not to be cherries—that is, they correspond to a path made of more than
two edges in the true tree. (See Section 5.2 for an example.) As it turns out, this only becomes
apparent once a larger fraction of the tree is reconstructed, at which point a subroutine detects the
“fake” cherries and removes them.
The full algorithm is detailed in Figures 5.3, 5.4, 5.5, and 5.6. We now explain its main components. The
bacc , and ε will be set in Section 6.
parameters k, R
21
z
x
y
Tx
Ty
Figure 5.1: Adding a cherry (x, z, y) to F ⊇ {Tx , Ty }.
Adding a Cherry. The algorithm BCP maintains a partially reconstructed subforest of the true tree, or
more precisely, a legal subforest F of T . The main operation we employ to “grow” our partially reconstructed subforest is the merging of two subtrees of F at their roots—an operation we call “adding a cherry,”
in reference to the cherry picking algorithm above. Suppose the current forest F contains two edge-disjoint
legal subtrees Tx and Ty . We merge them by creating a new node z and adding the edges (z, x) and (z, y)
as in Figure 5.1. We sometimes denote the pair of edges {(z, x), (z, y)} by (x, z, y). We call this operation
adding cherry (x, z, y) to F.
Identifying “Local” Cherries. As we explained above, we cannot hope to identify with certainty the
cherries of the unexplored part of the tree from short sequences at the leaves. Instead, we settle for detecting
what we refer to as “local” cherries, roughly, cherries in a “local” neighbourhood around the roots of the
current reconstructed subforest. More precisely, a “local” cherry is a pair of roots of the current subforest
that passes a series of tests as detailed below.
To determine whether two roots v1 , w1 of the current forest form a “local” cherry, our routine L O CAL C HERRY 2 in Figure 5.4 performs three tests:
1. [Short Distance] The nodes v1 , w1 are at a short distance (roughly 2g);
2. [Local Cherry] For all pairs of roots v2 , w2 at short distance (roughly 5g), the quartet
Q = {v1 , w1 , v2 , w2 },
admits the split v1 w1 |v2 w2 ;
3. [Edge Lengths] The edges connecting v1 , w1 to their (hypothetical) parent are short (roughly g).
The routine L OCAL C HERRY has three key properties, proved in Section 6:
1. [Edge Disjointness] It preserves the edge-disjointness of the current forest;
2. [Restricted Subforest] It builds a forest that is always a legal restriction of the true tree;
2
In [DMR06], the routine was called C HERRY ID.
22
3. [Short Edges] It guarantees that all edges of the restricted forest are short (smaller than g 0 ).
These properties are crucial for the proper operation of the algorithm (in particular for the correctness of
routines such as D ISTORTED M ETRIC (Figure 3.8) and I S S HORT (Figure 3.6) as seen in Section 3).
Finally, another important property of L OCAL C HERRY is that it is guaranteed to detect true cherries—at
least those that are “locally witnessed.” We now define this notion more precisely. For a distance matrix D
and a set of nodes N , we denote
D(N ) = max {D(x, y) : {x, y} ⊆ N } .
Definition 5.1 (Witnessed Cherry) Let F be a forest with path metric D. We say that a pair of leaves
{u, v} is an M -witnessed cherry in F if {u, v} is a cherry in F and there are at least two other leaves u0 , v 0
s.t.
D(Q) ≤ M,
where Q = {u, v, u0 , v 0 } (the leaves u0 , v 0 will act as “witnesses” of the cherry {u, v}).
Detecting Collisions. The merging of subtrees through “local” cherries that are not actual cherries eventually results in collisions between subtrees of the current forest such as in Figure 3.7. Such configurations
are undesirable since they do not allow to complete the reconstruction by simple merging of the subtrees at
their roots. Therefore, we seek to detect these collisions using the routine 3 in Figure 5.5.
After adding a new layer of “local” cherries, D ETECT C OLLISION checks whether new collisions can be
found. For this, we use routine I S C OLLISION in Figure 4.3 from Section 4. We actually perform two such
I S C OLLISION tests for each candidate edge in the target subtree Tu1 (see Figure 5.5). This is done precisely
for the same reasons that we did a “multiple test” in routine D ISTORTED M ETRIC (Figure 3.8) in Section 3:
It guarantees that at least one test is performed under the Basic Disjoint Setup (Dangling), which is needed
for I S C OLLISION to be correct. See Section 6 for details. Also, to make sure that we are always in either of
the configurations in Figure 4.2, we scan the nodes of the target subtree Tu1 in “reverse breath-first search
(BFS) order”:
1. Order the vertices in subtree Tu1 according to breath-first search v1 , . . . , vs ;
2. Scan the list in reverse order vs , . . . , v1 .
This way, we never find ourselves “above” a collision in Tu1 (because otherwise we would have encountered
the collision first).
Removing Collisions. Once a collision is identified we remove it using the routine in Figure 5.6.4 As
seen in Figure 5.2, the routine essentially removes the path from the collision all the way to the root of the
corresponding subtree. For a rooted forest F, we use the notation ParentF (x) to indicate the unique parent
of node x.
We note that we actually go through the “Collision Identification/Removal” step twice to make sure that
we detect collisions occurring on edges adjacent to the roots. See the proof of Lemma 6.5.
3
4
In [DMR06], the routine was called FAKE C HERRY.
In [DMR06], the routine was called B UBBLE.
23
z0
v
Figure 5.2: The R EMOVE C OLLISION in Figure 5.6 routine removes all white nodes and dashed edges. The
dotted curvy line indicates the location of a collision to be removed.
We prove in Proposition 7.4 below that, at each iteration of BCP, at least one cherry is added that will not
be removed at a later stage. Hence, the algorithm makes progress at every iteration and eventually recovers
the full tree.
Implementation. We argue that the running time of the algorithm is O(n5 k) and that with the use of
appropriate data structures it can be reduced further to O(n3 k). Let us start with the naive analysis. The
distance estimations in the initialization step of the algorithm take overall time O(n2 k), since there are
O(n2 ) pairs of leaves and each distance estimation between leaves takes linear time in the sequence length k.
Now, in every iteration of the algorithm:
• The Local Cherry Identification step takes overall time O(n4 k), since we consider O(n2 ) pairs of
roots in the for-loop of this step, and each call of L OCAL C HERRY requires O(n2 + nk) time—O(n2 )
time for the [Local Cherry] step and O(nk) time for the [Edge Lengths] step, in which the I S S HORT
call involves O(1) ancestral sequence reconstructions on trees of size O(n).
• The Collision Removal step requires O(n3 k) time in each iteration. Indeed, it performs O(n2 ) distance computations and each of these takes O(nk) time, since it requires O(1) ancestral sequence
reconstructions on trees of size O(n). It also performs O(n2 ) calls to D ETECT C OLLISION and R E MOVE C OLLISION , each of which is a linear time operation on trees of size O(n).
Since each iteration requires O(n4 k) time and there are O(n) iterations (see proof in Section 7), the overall
running time is O(n5 k). The above analysis is wasteful in 1) not reusing the already reconstructed ancestral
sequences and in 2) performing various tests on pairs of nodes that are far apart in the tree. With the use of
appropriate data structures, we could keep track of the “local” neighborhood of each node and restrict many
computations to these neighborhoods. This should bring down the running time to O(n3 k) with a constant
that would depend explicitly on f, g. The details are left to the reader.
24
Algorithm B LINDFOLDED C HERRY P ICKING
Input: Samples at the leaves {σu }u∈[n] ;
Output: Estimated topology;
• 0) Initialization:
– [Iteration Counter] i := 0;
– [Rooted Subforest] F0 := [n];
bacc );
– [Local Metric] For all u, v ∈ F0 , set D0 (u, v) = D ISTANCE E STIMATE(u, v; F0 ; R
• 1) Local Cherry Identification:
– Iteration: i;
– Set Fi+1 := Fi ;
– For all (v1 , w1 ) ∈
ρ(Fi )
2
,
∗ [Main Step] Compute (IsCherry, lv , lw ) := L OCAL C HERRY ((v1 , w1 ); (Fi , Di ));
∗ If IsCherry = TRUE,
· [Update Forest] Create new node u1 and add cherry (v1 , u1 , w1 ) to Fi+1 ;
· [Edge Lengths] Set h(u1 , v1 ) := lv and h(u1 , w1 ) := lw ;
• 2) Collision Removal:
F
– [Update Metric] For all x1 , x2 ∈ ρ(Fi+1 ), for all uϕ ∈ Txϕi+1 , ϕ = 1, 2, set
bacc ; ε);
Di+1 (u1 , u2 ) = D ISTORTED M ETRIC(u1 , u2 ; Fi+1 ; {h(e)}e∈Fi+1 ; R
– Set F := Fi+1 and D := Di+1 ;
– For all (u0 , u1 ) ∈ ρ(F) × ρ(F),
∗ Set HasCollision := FALSE;
∗ If u1 is not a leaf,
· [Main Step] Compute (HasCollision, z) := D ETECT C OLLISION ((u0 , u1 ); (F, D));
∗ If HasCollision = TRUE,
· [Update Forest] Compute Fi+1 := R EMOVE C OLLISION(z; Fi+1 );
– [Second Pass] Set F := Fi+1 and repeat the previous step;
• 3) Termination:
– If |ρ(Fi+1 )| ≤ 3,
∗ Join nodes in ρ(Fi+1 ) (star if 3, single edge if 2);
∗ Return (tree) Fi+1 ;
– Else, set i := i + 1, and go to Step 1.
Figure 5.3: Algorithm B LINDFOLDED C HERRY P ICKING.
5.2
Example
We give a detailed example of the execution of the algorithm. Consider the tree depicted in Figure 5.7a. It
is made of a large complete binary tree (on the right) with a small 3-level complete binary tree attached to
its root (on the left). All edges have length g, except (v, x), (x, v 0 ) and the edges adjacent to the root of the
small tree which have length g/2. In the figure, the subtree currently discovered by BCP is made of solid
arrows and full circles. The remaining (undiscovered) tree is in dotted lines and empty circles. Assume
25
Algorithm L OCAL C HERRY
Input: Two nodes (v1 , w1 ); current forest and distance matrix (F, D);
Output: Boolean value and length estimates;
• Set IsCherry := TRUE and lv = lw = 0;
• [Short Distance]
– If D(v1 , w1 ) > 2g + ε, then IsCherry := FALSE;
• [Local Cherry]
n
– Set N = (v2 , w2 ) ∈
ρ(F )
2
o
: D({v1 , w1 , v2 , w2 }) ≤ 5g + ε ;
– If N is empty, then IsCherry := FALSE; Else, for all (v2 , w2 ) ∈ N ,
∗ If I S S PLIT ((v1 , w1 ), (v2 , w2 ); D) = FALSE then set IsCherry := FALSE and break;
• [Edge Lengths]
– If IsCherry = TRUE,
∗ Let x1 , x2 be the children of v1 in F (or let x1 = x2 = v1 if v1 is a leaf);
∗ Let z0 be the closest node to v1 in ρ(F) − {v1 , w1 } under D;
bacc ; ε/16 ;
∗ Set (bv , lv ) := I S S HORT (x1 , x2 ), (w1 , z0 ); F; R
∗ Repeat previous steps switching the roles of v1 and w1 ;
∗ Set IsCherry := bv ∧ bw ;
• Return (IsCherry, lv , lw );
Figure 5.4: Routine L OCAL C HERRY.
Algorithm D ETECT C OLLISION
Input: Two roots u0 , u1 ; directed forest and distance matrix (F, D);
Output: Boolean value and node;
• Set HasCollision := FALSE and z := 0;
• Let x0 , y0 be the children of u0 in F;
• Scan through all nodes v in Tu1 (except u1 ) in reverse BFS manner,
– Let w := SisterF (v) and u := ParentF (v);
– [Collision Test] Compute
bx := I S C OLLISION(x0 , v, w, u; h(u, v); (F, D)),
and
by := I S C OLLISION(y0 , v, w, u; h(u, v); (F, D));
– If HasCollision := bx ∧ by = TRUE then set z := v and break;
• Return (HasCollision, z);
Figure 5.5: Routine D ETECT C OLLISION.
that the length of the sequences at the leaves allows us to estimate accurately distances up to 5g (the actual
constants used by the algorithm can be found later).
26
Algorithm R EMOVE C OLLISION
Input: Node v; rooted forest F;
Output: Rooted forest;
• If v is not in F or v is a root in F, return F;
• Let z0 be the root of the subtree of F in which v lies;
• Set x := v;
• While x 6= z0 ,
– Set x := ParentF (x);
– Remove node x and its adjacent edges below it;
• Return the updated forest F.
Figure 5.6: Routine R EMOVE C OLLISION.
First Level. We first join into cherries all pairs of leaves that “look” like g-cherries in this “local” metric.
We are guaranteed to find all true g-cherries. However, consider pairs of leaves such as u, v for which there
is no “local evidence” that it does not form a cherry. Even though u, v is not a cherry, it is joined into a cherry
by BCP. Figure 5.7a depicts the current forest after the first iteration of BCP. Before proceeding further, we
d from Section 2 to obtain reconstructed sequences at the roots of the current forest
apply our estimator Anc
and recompute the “local” metric.
Removing a Fake Cherry. We subsequently proceed to join “local” g-cherries one layer at a time, reconstructing internal sequences as we do so. After many iterations, we find ourselves in the situation of Figure 5.7b where most of the large complete tree has been reconstructed (assume for now that edges (u0 , v 00 ),
(v 0 , v 00 ), (u, v 0 ), (v, v 0 ) represented in dashed lines are present). Now, the new information coming from
sequences at y1 , . . . , y4 provides evidence that (u, v 0 , v) is not a cherry and that there is in fact a node x
on edge (v, v 0 ). For example, the quartet {y1 , y2 , u, v} suggests that u, v forms a cherry with a 3g/2-edge,
which cannot hold in a true cherry. At this point, we remove the “fake” cherry (u, v 0 , v) as well as all
cherries built upon it, here only (u0 , v 00 , v 0 ). Note that we have removed parts of the tree that were in fact
reconstructed correctly (e.g., the path between u and u0 ).
Rediscovering Removed Parts. Subsequently, BCP continues to join “local” cherries and “rediscovers”
the parts of the tree that were removed. For instance, in Figure 5.7c, the edge (u0 , v 00 ) is reconstructed again
but this time it forms a cherry with (u00 , v 00 ) rather than (v 0 , v 00 ).
Final Step. Eventually, the full tree is correctly reconstructed except maybe for a few (at most 3) remaining
edges. Those can be added separately. For example in Figure 5.7d only the three edges around x remain to
be uncovered. Note that the reconstructed tree has a root which is different from that of the original tree.
6
Analysis I: Induction Step
In this section and the next, we establish that BCP reconstructs the phylogeny correctly. In this section, we
establish a number of combinatorial properties of the current forest Fi grown by BCP. Then, in the next
section, we prove that the “correctly reconstructed subforest” of Fi increases in size at every iteration.
27
v 00
v0
u
u0
x
v
v0
x
v
u
y1 y2 y3 y4
b) BCP removes a fake cherry
a) first level of local cherries
u00
v 00
u
0
x
c) with the extra information,
BCP rediscovers part of the tree
z
d) only three extra edges need to be added
Figure 5.7: Illustration of BCP’s unraveling.
28
Parameters. Let δ > 0 be the error tolerance. Each application of Proposition 3.15 has an error of
O(n−γ ). We will do O(n3 ) distance estimations so that by the union bound we require O(n3−γ ) ≤ δ. Let
bcol > 6g. Set M > R
bcol + 4g 0 ; take R
bacc > M + 2B(g 0 ) + 4g 0 ; and choose k to
ε < 18 min{f, g 0 − g} and R
bacc ,
be equal to the maximum sequence length requirement for Proposition 3.15 with parameters , M and R
bacc .
for Proposition 3.2 with parameters and M , and for Proposition 3.13 with parameters ε/16, M and R
Induction Step. The following proposition establishes a number of properties of the forest grown by BCP.
We assume here that the conclusion of Proposition 3.15 holds at every iteration of the algorithm, until the
tree is fully recovered. In the next section, we will prove that the latter is indeed true with high probability.
Proposition 6.1 (Properties of Fi ) Denote by Fi = {Tu : u ∈ ρ(Fi )} the current forest at the beginning
of BCP’s i-th iteration. Also, assume that the conclusion of Proposition 3.15 holds for each call of the
routine D ISTORTED M ETRIC throughout the execution of the algorithm. Then, ∀i ≥ 0 (until the algorithm
stops),
1. [Legal Subforest] Fi is a legal subforest of T ;
2. [Edge Lengths] ∀u ∈ ρ(Fi ), Tu has edge lengths at most g 0 ;
3. [Weight Estimation] The estimated lengths {h(e)}e∈E(Fi ) of the edges in Fi are within ε/16 of their
right values;
bcol . (See Definition 4.2.)
4. [Collisions] There is no collision within distance R
Proof of Proposition 6.1: The proof is by induction on i.
i = 0: The set ρ(F0 ) consists of the leaves of T . The claims are therefore trivially true.
i > 0: Assume the claims are true at the beginning of the j-th iteration for all j ≤ i. By doing a step-by-step
analysis of the i-th iteration, we show that the claims are still true at the beginning of the (i + 1)-st iteration.
We first analyze the routine L OCAL C HERRY (Figure 5.4). For a legal subforest F of T , we denote the
“remaining” forest by
Fe = T − F.
More precisely, if F = {T1 , . . . , Tα } then Fe is the forest obtained from T as follows:
1. Remove all edges of T in the union of the trees T1 , . . . , Tα . In particular, for those edges of a Ti
representing a path in T , we remove all corresponding edges of T .
2. The nodes of Fe are all the endpoints of the remaining edges of T . All other nodes of T are discarded.
Note that the set Fe is in fact a subforest of T .
Lemma 6.2 (Local Cherry Identification) Let Fi be the current forest at the beginning of the i-th iteration. Then we have the following:
• If {v1 , w1 } is a 5g-witnessed cherry in Fei , then it passes all tests in L OCAL C HERRY;
29
• If {v1 , w1 } passes all the tests in L OCAL C HERRY, then
d(u1 , v1 ) ≤ g + 2ε,
and
d(u1 , w1 ) ≤ g + 2ε,
where u1 is the parent of {v1 , w1 } (as defined by L OCAL C HERRY and the [Update Forest] step in the
main loop). Moreover the length estimates satisfy
|d(u1 , v1 ) − lv | <
ε
,
16
|d(u1 , w1 ) − lw | <
ε
.
16
and
• If two pairs {v1 , w1 } and {v2 , w2 } both pass all tests in L OCAL C HERRY, then it must be that the
paths pathT (v1 , w1 ) and pathT (v2 , w2 ) are non-intersecting.
Proof: Suppose first that {v1 , w1 } is a 5g-witnessed cherry in Fei with witness {v2 , w2 }. Since there is no
bcol > 6g by assumption, we are in the dangling case. Moreover, all edge weights below
collision within R
{v1 , w1 , v2 , w2 } have been estimated within ε/16 and all corresponding edge weights (possibly corresponding to paths) are ≤ g 0 . Therefore, by Propositions 3.15 and 4.1, D ISTORTED M ETRIC (Figure 3.8) is accurate
within ε, I S S PLIT (Figure 4.1) returns the correct answer and I S S HORT (Figure 3.6) is accurate within ε/16.
Hence, {v1 , w1 } passes the three tests in L OCAL C HERRY.
Conversely, suppose {v1 , w1 } passes all tests in L OCAL C HERRY. By our assumptions, when D ISTORTreturns a finite value, it is accurate within ε. Let z0 be as in Figure 5.4 (that is, z0 is the closest
node to v1 in ρ(F) − {v1 , w1 } under the distorted metric). The fact that {v1 , w1 } previously passed the
[Local Cherry] test implies in particular that the diameter of {v1 , w1 , z0 } must be less than distance 5g + 2ε.
bcol > 6g. Let u1 be
In particular, there is no collision between the subtrees rooted at v1 , w1 , and z0 since R
the intersection of {v1 , w1 , z0 }. Then, I S S HORT returns an estimate within ε/16 < ε which in turn implies
that d(u1 , v1 ) and d(u1 , w1 ) are ≤ g + 2ε.
ED M ETRIC
For the third part, assume by contradiction that the paths pathT (v1 , w1 ) and pathT (v2 , w2 ) intersect.
Then by the triangle inequality, the diameter of {v1 , w1 , v2 , w2 } is at most 5g. In particular, when L O CAL C HERRY is applied to {v1 , w1 }, the pair {v2 , w2 } is considered in the [Local Cherry] test and, since
bcol > 6g, I S S PLIT correctly returns FALSE. That is a contradiction.
there is no collision within R
We can now prove Claims 1, 2, and 3 of Proposition 6.1.
Lemma 6.3 (Claims 1, 2, and 3) Let Fi+1 be the current forest at the beginning of the (i + 1)-st iteration.
The Claims 1, 2, and 3 of the induction hypothesis hold for Fi+1 .
Proof: Since R EMOVE C OLLISION (Figure 5.6) only removes edges from the current forest, it is enough
to prove that after the completion of the Local Cherry Identification step the resulting forest satisfies
Claims 1, 2, and 3.
Claim 1. By the induction hypothesis, Fi is legal. We only need to check that Fi+1 is edge-disjoint.
Suppose on the contrary that the forest is not edge-disjoint. Also, suppose that, along the execution of the
Local Cherry Identification step, the forest stops being edge-disjoint when cherry (v1 , u1 , w1 ) is added to
Fi . Then one of the following must be true:
30
1. There is an “old” root z ∈ ρ(Fi ) such that pathT (v1 , w1 ) is edge-sharing with TzFi . But then there is
a collision in Fi within distance 2g + 2ε which contradicts the induction hypothesis (Claim 4).
2. There is a “new” root z ∈ ρ(Fi+1 )\ρ(Fi ) with corresponding cherry (x, z, y) such that pathT (v1 , w1 )
is edge-sharing with pathT (x, y). In that case v1 w1 |xy is not the correct split and
¯ 1 , w1 , x, y}) ≤ 5g,
d({v
by the triangle inequality and Proposition 3.15. But then, by Lemma 6.2, L OCAL C HERRY (Figure 5.4)
rejects {v1 , w1 } when performing the [Local Cherry] test with witness {x, y}—a contradiction.
Claim 2 and 3. Follows from the induction hypothesis and Lemma 6.2.
It remains to prove Claim 4 of Proposition 6.1. This follows from the following analysis of D ETECTC OLLISION (Figure 5.5). Note that, since Claim 4 holds for all iterations j ≤ i, it must be the case that
any new collision between two trees involves at least one of the new edges of these trees added in the Local
Cherry Identification step. We call an edge deep if it is not adjacent to a root in the current forest. Otherwise
we call the edge a top edge. We show that the first pass of the Collision Removal step removes all collisions
into deep edges. At the beginning of the second pass, all collisions (if any) must be into top edges. We show
that the second pass cleans those up. We first prove that there are no false positives in the Collision Removal
step.
Lemma 6.4 (Collision Removal: No False Positive) Let Fi+1 be the current forest at the beginning of the
first or second pass of the Collision Removal step of the i-th iteration, and let u0 , u1 ∈ ρ(Fi+1,1 ) be the roots
F
F
of the trees T0 = Tu0i+1,1 and T1 = Tu1i+1,1 . Let v be a node in T1 , and suppose that T0 does not collide
into T1 below v or on the edge immediately above it. Then no collision is detected in the corresponding step
of D ETECT C OLLISION.
Proof: Let x0 , y0 be the children of u0 . It suffices to show that either bx is FALSE or by is FALSE (see
Figure 5.5). Without loss of generality we can assume that the path connecting u0 to u1 does not pass
through x0 . In particular, we are in the case b) of Figure 4.2. If any of the distances passed to I S C OLLISION
is +∞, I S C OLLISION returns FALSE. Otherwise, by our assumption on the output of D ISTORTED M ETRIC,
the assumptions of Proposition 4.4 are satisfied. Hence, I S C OLLISION returns FALSE in that case as well.
Lemma 6.5 (Collision Removal) The first and second passes of the Collision Removal step satisfy:
1. Let Fi+1,1 be the current forest at the beginning of the first pass of the Collision Removal step of the
F
F
i-th iteration, and let u0 , u1 ∈ ρ(Fi+1,1 ). Suppose T0 = Tu0i+1,1 collides into T1 = Tu1i+1,1 within
bcol on a deep edge e = (u, v) of T1 . Then D ETECT C OLLISION in Figure 5.5 correctly
distance R
detects the collision.
2. Let Fi+1,2 be the current forest at the beginning of the second pass of the Collision Removal step of
F
F
the i-th iteration, and let u0 , u1 ∈ ρ(Fi+1,2 ). Suppose T0 = Tu0i+1,2 collides into T1 = Tu1i+1,2 within
bcol . Then D ETECT C OLLISION correctly detects the collision.
distance R
Proof: 1. Denote by xϕ , yϕ the children of uϕ , ϕ = 0, 1. Consider the Basic Collision Setup of Section 4.
By the remark above the statement of the lemma, the path coming from e enters T0 through a top edge, or at
u0 . (See Figure 6.1.) Let
31
u1
y1
x1
u0
B0→1
x0
y0
u
w
v
v1
v2
T1
T0
Figure 6.1: Configuration in the proof of Lemma 6.5.
A0→1 = {z ∈ V(T1 ) : e is not in the subtree of T1 rooted at z} ,
and
B0→1 = V(T1 ) \ A0→1 .
Denote by z the current-node variable used by D ETECT C OLLISION as it scans the tree T1 in a reverse BFS
manner. Observe that, for all z ∈ A0→1 − {v}, the Basic Collision Setup of Proposition 4.4 holds for both
x0 and y0 . Hence, by Lemma 6.4, I S C OLLISION in Figure 4.3 returns the correct answer. Furthermore, in
bcol and
the case of v, I S C OLLISION returns TRUE for both x0 and y0 since the collision is within R
¯ 0 , y0 , v1 , v2 , w}) ≤ R
bcol + 4g 0 ≤ M,
d({x
by Claim 2. For w, I S C OLLISION returns FALSE because in that case h(u, w) − ν ≤ −f + 3ε ≤ f /2, where
ν is the estimated length of the internal edge of {x0 , w1 , w2 , v} or {y0 , w1 , w2 , v} with w1 , w2 the children
of w. Finally, by scanning the nodes in reverse BFS order, we guarantee that v and w are encountered before
any node in B0→1 . Hence, D ETECT C OLLISION identifies correctly the collision on edge e.
2. Observe that, after a call to R EMOVE C OLLISION, the set of edges in the remaining forest is a subset of
what it used to be. In particular, the subset of these edges involved in a collision decreases in size. Moreover,
from the first part of the lemma, at the end of the first pass there is no collision remaining into deep edges.
bcol will be found and
Given this, the argument above implies that any remaining collision within distance R
removed in the second pass.
7
Analysis II: Tying It All Together
In the previous section, we showed that provided Proposition 3.15 holds at every iteration the forest built is
“well-behaved.” Below, we finish the proof of our main theorem by showing that Proposition 3.15 indeed
holds until termination and that the algorithm eventually converges. We also discuss the issues involved in
extending O(log n)-reconstruction beyond the ∆-Branch Model in Section 7.3.
32
7.1
Quantifying Progress
Our main tool to assess the progress of the algorithm is the notion of a fixed subforest—in words, a subforest
of the current forest that will not be modified again under the normal operation of the algorithm.
Definition 7.1 (Fixed Subforest) Let F be a legal subforest of T . Let u ∈ V(F). We say that u is fixed if
TuF is fully reconstructed, or in other words, TuF can be obtained from T by removing (at most) one edge
adjacent to u and keeping the subtree containing u. Note that descendants of a fixed node are also fixed. We
denote by F ∗ the rooted subforest of F made of all fixed nodes of F. We say that F ∗ is the maximal fixed
subforest of F.
Let Fi = {T1 , . . . , Tα } be the current forest at iteration i with remaining forest Fei = {T10 , . . . , Tβ0 }. Assuming the conclusion of Proposition 6.1 holds, each leaf v in Fei satisfies exactly one of the following:
• Fixed Root: v is a root of a fully reconstructed tree Ta ∈ Fi (that is, Ta is in Fi∗ );
• Colliding Root: v is a root of a tree Ta ∈ Fi that contains a collision (that is, Ta is not in Fi∗ );
• Collision Node: v belongs to a path connecting two vertices in Ta ∈ Fi but is not the root of Ta (that
is, it lies in the “middle” of an edge of Ta ).
Note in particular that the fixed roots of Fei are roots in Fi∗ (although not all roots of the maximal fixed
subforest Fi∗ are fixed roots in Fei as they may lie below a colliding root). We also need the notion of a fixed
bundle—in words, a cherry (along with witnesses) that will be picked by the algorithm at the next iteration
and remain until termination.
Definition 7.2 (Fixed Bundle) A bundle in Fei is a group of four leaves in Fei such that:
• Any two leaves in the bundle are at topological distance at most 5 in T ;
• It includes at least one cherry of Fei .
A fixed bundle is a bundle in Fei whose leaves are fixed roots.
The following lemma is the key to our convergence argument. It ensures that a fixed bundle always exists
and hence that progress is made at every iteration.
Proposition 7.3 (Existence of a Fixed Bundle) Assume Fei has at least 4 leaves and satisfies the conclusion of Proposition 6.1. Then, Fei contains at least one fixed bundle.
Proof: To avoid confusion between the forests Fi and Fei , we will refer to the forest Fei as the anti-forest, to
its trees as anti-trees and to its leaves as anti-leaves. We first make a few observations:
1. The anti-forest Fei is binary, that is, all its nodes have degrees in {1, 3}. Indeed, note first that since
T is binary, one cannot obtain nodes of degree higher than 3 by removing edges from T . Assume by
contradiction that there is a node, u, of degree 2 in Fei . Let w1 , w2 , w3 be the neighbors of u in T
and assume that w3 ∈
/ Fei . By construction of the anti-forest Fei the edge (u, w3 ) is in the forest Fi
(possibly included in an edge of Fi corresponding to a path of T ). Moreover, u is a node of degree 1
in Fi . But this is a contradiction because the only nodes of degree 1 in a legal subforest of T are the
leaves of T and u cannot be a leaf of T .
33
2. A binary tree (or anti-tree) T0 with 4 or more leaves (or anti-leaves) contains at least one bundle.
Indeed, let L0 be the leaves of T0 . We call R0 = L0 the level-0 leaves of T0 . Now remove all cherries
from T0 and let R1 be the roots of the removed cherries—the level-1 leaves. Denote by T1 the tree so
obtained and note that its leaves L1 contain R1 as well as some remaining level-0 leaves. Consider
the cherries of T1 (there is at least one). If any such cherry is made of two level-1 leaves, then the
corresponding (descendant) level-0 leaves of T0 form a bundle in T0 and we are done. (In that case,
the diameter of the bundle is 4.) Suppose there is no such cherry. Note that there is no cherry in
T1 formed from two level-0 leaves as those were removed in constructing T1 . Hence, all remaining
cherries of T1 must contain at least one level-1 leaf. Now, merge all such cherries to obtain T2 . Denote
by R2 the roots of the removed cherries—the level-2 leaves. Once again, by the argument above, any
cherry of T2 contains at least one level-2 leaf. Any such cherry (there is at least one) provides a
bundle by considering its descendant level-0 leaves in T0 . (If the second leaf in the cherry is level-0,
the diameter of the bundle is 4. If it is level-1 or level-2, the diameter is 5.) This proves the claim.
bcol > 6g (see Definition 4.2). In particular,
3. By Claim 4 in Proposition 6.1, there is no collision within R
b
collision nodes are at distance at least Rcol > 6g from any other anti-leaf (collision nodes, fixed roots,
colliding roots) in Fei . Therefore, if an anti-tree in Fei contains a collision node, then it has ≥ 4
anti-leaves and, from the previous observation, it contains at least one bundle. Moreover, this bundle
cannot contain a collision node since in a bundle all anti-leaves are at distance at most 5g and collision
bcol > 6g from all other anti-leaves in Fei .
nodes are at distance at least R
4. From the previous observations, we get the following: if a tree in Fei contains a collision, then either
it has a fixed bundle or it has at least one colliding root.
We now proceed with the proof. Assume first that there is no collision node in Fei . Then, there cannot
be any colliding root either because by definition a colliding root is the root of a tree containing a collision.
In particular, Fei is composed of a single anti-tree whose anti-leaves are all fixed roots. Then, since by
assumption Fei has at least 4 anti-leaves, there is a fixed bundle by Observation 2 above.
Assume instead that there is a collision node. Let Te0 be an anti-tree in Fei with such a collision node,
say c0 . Then by Observation 4, either 1) Te0 has a fixed bundle in which case we are done, or 2) one of Te0 ’s
anti-leaves is a colliding root, say r0 . In the latter case, let T 0 be the tree in Fi whose root is r0 . The tree T 0
contains at least one collision node (included in an edge corresponding to a path of T ). This collision node,
say c1 , is also contained in an anti-tree in Fei , say Te1 . Repeat the argument above on Te1 , and so on.
We claim that this process generates a simple path P in T starting from the node c0 above, passing
through an alternating sequence of colliding roots and collision nodes r0 , c1 , r1 , c2 , . . ., and eventually
reaching a fixed bundle. Indeed, since there is no cycle in T , the only way for P not to be simple is for
it to “reverse on itself.” But note that by definition, for all j we have cj 6= rj and rj 6= cj+1 . Moreover the
simple paths cj → rj and rj → cj+1 belong respectively to the anti-forest Fei and the forest Fi (possibly
with subpaths collapsed into edges). In particular, their edges (in T ) cannot intersect. Hence, P is simple.
Finally, since T is finite, this path cannot be infinite, and we eventually find a fixed bundle.
7.2
Proof of the Main Theorem
Consider now the ∆-Branch Model. The proof of convergence works by considering first the hypothetical
case where all distance estimates computed by the algorithm are perfectly accurate, that is the case where
we have “perfect” local information. We denote by {(F0 )i }i≥0 the sequence of forests built under this
assumption. Note that, up to arbitrary choices (tie breakings, orderings, etc.), this sequence is deterministic.
We now show that it terminates in a polynomial number of steps with the correct tree.
34
Proposition 7.4 (Progress Under Perfect Local Information) Assume (F0 )i = {Tu : u ∈ ρ((F0 )i )} is
the current forest at the beginning of BCP’s i-th iteration under perfect local information with corresponding maximal fixed subforest (F0 )∗i . Then for all i ≥ 0 (before the termination step), (F0 )∗i ⊆ (F0 )∗i+1 and
|V((F0 )∗i+1 )| > |V((F0 )∗i )|.
Proof of Proposition 7.4: We first argue that (F0 )∗i ⊆ (F0 )∗i+1 . Note that the only routine that removes
edges is R EMOVE C OLLISION in Figure 5.6. Since R EMOVE C OLLISION only removes edges above identified collisions and (F0 )∗i is fully reconstructed, it suffices to show that collisions identified by D ETECTC OLLISION in Figure 5.5 are actual collisions. This follows from Lemma 6.4.
To prove |V((F0 )∗i+1 )| > |V((F0 )∗i )|, assume (F0 )i = {T1 , . . . , Tα } with remaining forest (Fe0 )i =
{T10 , . . . , Tβ0 }. From Proposition 7.3, it follows that (Fe0 )i contains at least one fixed bundle. This immediately implies the second claim. Indeed, by Lemma 6.2 note that the cherry in the fixed bundle is found by
L OCAL C HERRY in Figure 5.4 during the (i + 1)-st iteration and is not removed by the Collision Removal
step from the argument above.
Proof of Theorem 1: To prove our main theorem under the ∆-Branch Model, we modify our reconstruction
algorithm slightly by rounding the estimates in Proposition 3.15 to the closest multiple of ∆. Also, we
choose a number of samples large enough so that the distance estimation error is smaller than ∆/3. In that
case, we simply mimic the algorithm under perfect local information. Note that by Proposition 7.4 there are
at most O(n) iterations until termination under perfect local information. By a union bound, it follows that
Proposition 3.15 holds true for all pairs of subtrees in {(F0 )i }i≥0 with high probability.
We can now conclude the proof. By Proposition 6.1, the current forest at each iteration is correctly
reconstructed. By Proposition 7.4 after O(n) iterations there remain at most three nodes in ρ(Fi ) and at
that point, from Proposition 6.1 Claim 4, we have that Fi ≡ Fi∗ . Therefore the remaining task is to join the
remaining roots and there is only one possible topology. So when the BCP algorithm terminates, it outputs
the tree T (as an undirected tree) with high probability.
The tightness of the value g ∗ = ln42 is justified by the polynomial lower bound [Mos04] on the number
of required characters if the mutation probability p on all edges of the tree satisfies 2(1 − 2p)2 < 1.
7.3
Beyond the ∆-Branch Model?
Extending Theorem 1 to continuous edge lengths appears far from trivial. The issue arises in the final
union bound over all applications of Proposition 3.15 (more specifically, the ancestral state reconstruction
step) which is valid except with inverse polynomial probability over the generated sequences (for a fixed
subtree of T ). Indeed, note that in general there are super-polynomially many restricted subtrees of the
true tree T where all edges (paths in T ) are shorter than g ∗ . Therefore, using only a simple union bound,
we cannot hope to guarantee that ancestral state reconstruction is successful simultaneously on all possible
partial reconstructed subtrees.
In the previous subsection, we avoided this problem by showing that under the ∆-BM assumption the
algorithm follows a deterministic realization path of polynomial length. Moving beyond this proof would
require proving that the ancestral state reconstruction can be performed on the random forests generated
by the algorithm—but this is not straightforward as the partially reconstructed forest is generated by the
same data that is used to perform the ancestral estimation. We conjecture that the correlation created by
this “bootstrap” is mild enough to allow our algorithm to work in general, but we cannot provide a rigorous
proof at this point.
We remark that Mossel’s earlier result [Mos04] on the balanced case with continuous edge lengths is
not affected by the issue above because, there, the reconstruction of the tree occurs one level at a time (there
35
is no collision). Hence, ancestral state reconstruction is performed only on fully reconstructed subtrees—of
which there are only polynomially many.
Note finally that, even under the discretization assumption made in this paper, achieving log(n)-reconstruction is nontrivial and does not follow from previous techniques. In particular, it can be shown that
all previous rigorous reconstruction algorithms for general trees require polynomial sequence lengths even
when all edge lengths are identical and below g ∗ . See [Roc08] for a formal argument of this type.
8
Conclusion
The proof of Steel’s Conjecture [Ste01] provides tight results for the phylogenetic reconstruction problem.
However, many theoretical and practical questions remain:
• Can the discretization assumption be removed? We conjecture that the answer is yes.
• Can the results be extended to other mutation models? In subsequent work, Roch [Roc09] showed
that our results hold for so-called General Time-Reversible (GTR) mutation models below the KestenStigum bound. Can the results be extended to deal with “rates across sites” (see e.g. [Fel04])?
• What is the optimal g-value for the Jukes-Cantor model? Is it identical to the critical value gq=4 of
the reconstruction problem for the so-called “Potts model” with q = 4? We note that it is a long
standing open problem to find gq=4 . The best bounds known are given in [MP03, MSW04a]. See
also [MM06, Sly08].
Acknowledgments
S.R. thanks Martin Nowak and the Program for Evolutionary Dynamics at Harvard University where part
of this work was done. C.D. and E.M. thank Satish Rao for interesting discussions. E.M. thanks M. Steel
for his enthusiastic encouragement for studying the connections between the reconstruction problem and
phylogenetics. We thank Allan Sly for helpful discussions.
References
[BCMR06] Christian Borgs, Jennifer T. Chayes, Elchanan Mossel, and Sébastien Roch. The Kesten-Stigum
reconstruction bound is tight for roughly symmetric binary channels. In FOCS, pages 518–530,
2006.
[BKMP05] N. Berger, C. Kenyon, E. Mossel, and Y. Peres. Glauber dynamics on trees and hyperbolic
graphs. Probab. Theory Rel., 131(3):311–340, 2005. Extended abstract by Kenyon, Mossel and
Peres appeared in proceedings of 42nd IEEE Symposium on Foundations of Computer Science
(FOCS) 2001, 568–578.
[BRZ95]
P. M. Bleher, J. Ruiz, and V. A. Zagrebnov. On the purity of the limiting Gibbs state for the
Ising model on the Bethe lattice. J. Statist. Phys., 79(1-2):473–482, 1995.
[Bun71]
P. Buneman. The recovery of trees from measures of dissimilarity. In Mathematics in the
Archaelogical and Historical Sciences, pages 187–395. Edinburgh University Press, Edinburgh,
1971.
36
[Cav78]
J. A. Cavender. Taxonomy with confidence. Math. Biosci., 40(3-4), 1978.
[Cha96]
Joseph T. Chang. Full reconstruction of Markov models on evolutionary trees: identifiability
and consistency. Math. Biosci., 137(1):51–73, 1996.
[DMR06]
Constantinos Daskalakis, Elchanan Mossel, and Sébastien Roch. Optimal phylogenetic reconstruction. In STOC’06: Proceedings of the 38th Annual ACM Symposium on Theory of
Computing, pages 159–168, New York, 2006. ACM.
[EKPS00] W. S. Evans, C. Kenyon, Y. Peres, and L. J. Schulman. Broadcasting on trees and the Ising
model. Ann. Appl. Probab., 10(2):410–433, 2000.
[ESSW99] P. L. Erdös, M. A. Steel, L. A. Székely, and T. A. Warnow. A few logs suffice to build (almost)
all trees (part 1). Random Struct. Algor., 14(2):153–184, 1999.
[Far73]
J. S. Farris. A probability model for inferring evolutionary trees. Syst. Zool., 22(4):250–256,
1973.
[Fel04]
J. Felsenstein. Inferring Phylogenies. Sinauer, New York, New York, 2004.
[Geo88]
H. O. Georgii. Gibbs measures and phase transitions, volume 9 of de Gruyter Studies in Mathematics. Walter de Gruyter & Co., Berlin, 1988.
[Hig77]
Y. Higuchi. Remarks on the limiting Gibbs states on a (d + 1)-tree. Publ. Res. Inst. Math. Sci.,
13(2):335–348, 1977.
[Iof96]
D. Ioffe. On the extremality of the disordered state for the Ising model on the Bethe lattice.
Lett. Math. Phys., 37(2):137–143, 1996.
[JC69]
T. H. Jukes and C. Cantor. Mammalian protein metabolism. In H. N. Munro, editor, Evolution
of protein molecules, pages 21–132. Academic Press, 1969.
[JM04]
S. Janson and E. Mossel. Robust reconstruction on trees is determined by the second eigenvalue.
Ann. Probab., 32:2630–2649, 2004.
[KS66]
H. Kesten and B. P. Stigum. Additional limit theorems for indecomposable multidimensional
Galton-Watson processes. Ann. Math. Statist., 37:1463–1481, 1966.
[Lyo89]
R. Lyons. The Ising model and percolation on trees and tree-like graphs. Comm. Math. Phys.,
125(2):337–353, 1989.
[MM06]
Marc Mézard and Andrea Montanari. Reconstruction on trees and spin glass transition. J. Stat.
Phys., 124(6):1317–1350, 2006.
[Mos98]
E. Mossel. Recursive reconstruction on periodic trees. Random Struct. Algor., 13(1):81–97,
1998.
[Mos01]
E. Mossel. Reconstruction on trees: beating the second eigenvalue. Ann. Appl. Probab.,
11(1):285–300, 2001.
[Mos04]
E. Mossel. Phase transitions in phylogeny. Trans. Amer. Math. Soc., 356(6):2379–2404, 2004.
[Mos07]
E. Mossel. Distorted metrics on trees and phylogenetic forests. IEEE/ACM Trans. Comput. Bio.
Bioinform., 4(1):108–116, 2007.
37
[MP03]
E. Mossel and Y. Peres. Information flow on trees. Ann. Appl. Probab., 13(3):817–844, 2003.
[MS04]
E. Mossel and M. Steel. A phase transition for a random cluster model on phylogenetic trees.
Math. Biosci., 187(2):189–203, 2004.
[MSW04a] F. Martinelli, A. Sinclair, and D. Weitz. Fast mixing for independent sets, colorings, and other
models on trees. In Proceedings of the 15th ACM-SIAM Symposium on Discrete Algorithms,
pages 449–458, 2004.
[MSW04b] F. Martinelli, A. Sinclair, and D. Weitz. Glauber dynamics on trees: boundary conditions and
mixing time. Comm. Math. Phys., 250(2):301–334, 2004.
[Ney71]
J. Neyman. Molecular studies of evolution: a source of novel statistical problems. In S. S. Gupta
and J. Yackel, editors, Statistical desicion theory and related topics, pages 1–27. Academic
Press, New York, 1971.
[Roc08]
Sébastien Roch. Sequence-length requirement for distance-based phylogeny reconstruction:
Breaking the polynomial barrier. In FOCS, pages 729–738, 2008.
[Roc09]
Sébastien Roch. Phase transition in distance-based phylogeny reconstruction. Preprint, 2009.
[Sly08]
A. Sly.
Reconstruction of symmetric Potts models.
http://arxiv.org/abs/0811.1208, 2008.
[Spi75]
F. Spitzer. Markov random fields on an infinite tree. Ann. Probability, 3(3):387–398, 1975.
[SS03]
C. Semple and M. Steel. Phylogenetics, volume 22 of Mathematics and its Applications series.
Oxford University Press, 2003.
[Ste01]
M. Steel. My Favourite Conjecture. Preprint, 2001.
38
Preprint available at
| 5cs.CE
|
QUICKEST DETECTION OF A MINIMUM OF
TWO POISSON DISORDER TIMES
arXiv:cs/0509029v4 [cs.CE] 25 Apr 2007
ERHAN BAYRAKTAR
∗ AND
H. VINCENT POOR†
Abstract. A multi-source quickest detection problem is considered. Assume there are two independent Poisson
processes X 1 and X 2 with disorder times θ1 and θ2 , respectively; that is, the intensities of X 1 and X 2 change at random
unobservable times θ1 and θ2 , respectively. θ1 and θ2 are independent of each other and are exponentially distributed.
Define θ , θ1 ∧ θ2 = min{θ1 , θ2 } . For any stopping time τ that is measurable with respect to the filtration generated by
the observations define a penalty function of the form
ˆ
˜
Rτ = P(τ < θ) + cE (τ − θ)+ ,
where c > 0 and (τ − θ)+ is the positive part of τ − θ. It is of interest to find a stopping time τ that minimizes the above
performance index. This performance criterion can be useful for example in the following scenario: There are two assembly
lines that produce products A and B, respectively. Assume that the malfunctioning (disorder) of the machines producing A
and B are independent events. Later, the products A and B are to be put together to obtain another product C. A product
manager who is worried about the quality of C will want to detect the minimum of the disorder times (as accurately as
possible) in the assembly lines producing A and B. Another problem to which we can apply our framework is the internet
surveillance problem: A router receives data from, say, n channels. The channels are independent and the disorder times
of channels are θ1 , · · · , θn . The router is said to be under attack at θ = θ1 ∧ · · · ∧ θn . The administrator of the router is
interested in detecting θ as quickly as possible.
Since both observations X 1 and X 2 reveal information about the disorder time θ, even this simple problem is more
involved than solving the disorder problems for X 1 and X 2 separately. This problem is formulated in terms of a three
dimensional sufficient statistic, and the corresponding optimal stopping problem is examined. The solution is characterized
by iterating a suitable functional operator.
AMS subject classifications. 62L10, 62L15, 62C10, 60G40.
Key words. Change detection, Poisson processes, optimal stopping.
1. Introduction. Consider two independent Poisson processes X i = {Xti : t ≥ 0} i ∈ {1, 2} with
the same arrival rate β. At some random unobservable times θ1 and θ2 , with distributions
P(θi > t) = (1 − πi )e−λt for t ≥ 0,
P(θi = 0) = πi
(1.1)
the arrival rates of the Poisson processes X 1 and X 2 change from β to α, respectively, i.e.,
Xti
−
Z
t
hi (s)ds,
0
t ≥ 0, i = 1, 2,
(1.2)
are martingales, in which
hi (t) = [β1{s<θi } + α1{s≥θi } ],
t ≥ 0, i = 1, 2.
(1.3)
Here α and β are known positive constants. We seek a stopping rule τ that detects the instant θ = θ1 ∧ θ2
of the first regime change as accurately as possible given the past and the present observations of the
processes X 1 and X 2 . More precisely, we wish to choose a stopping time τ of the history of the processes
X 1 and X 2 that minimizes the following penalty function
(1.4)
Rτ = P(τ < θ) + cE (τ − θ)+ .
∗ E. Bayraktar is with the Department of Mathematics, University of Michigan, Ann Arbor, MI 48109, USA, email:
[email protected]
† H. V. Poor is with the School of Engineering and Applied Science, Princeton University, Princeton, NJ 08544, USA,
email: [email protected]
1
2
E. BAYRAKTAR AND H. V. POOR
The first term in (1.4) penalizes the frequency of false alarms, and the second term penalizes the detection
delay. The disorder time demarcates two regimes, and in each of these regimes the decision maker uses
distinctly different strategies. Therefore, it is in the decision maker’s interest to detect the disorder time
as accurately as possible from its observations. Here, we are solving the case when a decision maker has
two identical and independent sources to process. In the Section 9 we discuss how our analysis can be
extended to non-identical sources.
Quickest detection problems arise in a variety of applications such as seismology, machine monitoring,
finance, health, and surveillance, among others (see e.g. [1], [10], [7], [11] and [14]). Because Poisson
processes are often used to model abrupt changes, Poisson disorder problems have potential applications
e.g. to the effective control and prevention of infectious diseases, quickest detection of quality and
reliability problems in industrial processes, and surveillance of Internet traffic to protect network servers
from the attacks of malicious users. This is because the number of patients infected, number of defected
items produced and number of packets arriving at a network node are usually modeled by Poisson
processes. In these examples the disorder time corresponds to the time when an outbreak occurs, when
a machine in an assembly line breaks down or when a router is under attack, respectively. The multisource quickest detection problem considered here can be applied to tackle these problems when there are
multiple sources of information. For example in the monitoring of industrial processes the minimum of
disorder times represents the first time when one of many assembly lines in a plant breaks down during
the production of a certain type of item. Let us be more specific: Assume that there are two assembly
lines that produce products A and B, respectively. Assume also that the malfunctioning (disorder) of the
machines producing A and B are independent events. Later, the products A and B are to be put together
to obtain another product C. A product manager who is worried about the quality of C will want to
detect the minimum of the disorder times (as accurately as possible) in the assembly lines producing A
and B. The performance function (1.4) is an appropriate choice because the product manager will worry
about the quality of the end product C, not of the individual pieces seperately. Another problem to
which we can apply our framework is the internet surveillance problem: A router receives data from, say,
n channels. The channels are independent and the disorder times of channels are θ1 , · · · , θn . The router
is said to be under attack at θ = θ1 ∧ · · · ∧ θn . The administrator of the router is interested in detecting
θ as quickly as possible.
The one dimensional Poisson disorder problem, i.e., the problem of detecting θ1 as accurately as possible
given the observations from the Poisson process X 1 has recently been solved (see [2], [3] and the references
therein). The two-dimensional disorder problem we have introduced cannot be reduced to solving the
corresponding one-dimensional disorder problems since both X 1 and X 2 reveal some information about θ
whenever these processes jump. That is, if we take the minimum of the optimal stopping times that solve
the one dimensional Poisson disorder problems, then we obtain a stopping time that is a sub-optimal
solution to (1.4) (see Remark 4.1).
We will show that the quickest detection problem of (1.4) can be reduced to an optimal stopping problem for a three-dimensional piece-wise deterministic Markov process. Continuous-time Markov optimal
stopping problems are typically solved by formulating them as free boundary problems associated with
the infinitesimal generator of the Markov process. In this case, however the infinitesimal generator contains differential delay operators. Solving free boundary problems involving differential delay operators
is a challenge even in the one dimensional case and the smooth fit principle is expected to fail (see [2],
[3] and the references therein). Instead as in [4] and [6] we work with an integral operator, iteration of
which generates a monotonically increasing sequence of functions converging exponentially to the value
function of the optimal stopping problem. That is, using the integral operator we reduce the problem
to a sequence of deterministic optimization problems. This approach provides a new numerical method
for calculating and characterizing the value function and the continuation region in addition to providing
information about the shape and the location of the optimal continuation region. Using the structure
of the paths of the piece-wise deterministic Markov process we also provide a non-trivial bound on the
optimal stopping time which can be used to obtain approximate stopping strategies.
The remainder of this paper is organized as follows. In Sections 2 and 3, we restate the problem
of interest under a suitable reference measure P0 that is equivalent to P. Working under the reference
Quickest Detection of a Minimum of Two Poisson Disorder Times
3
measure P0 reduces the computations considerably, since under this measure the observations X 1 and X 2
are simple Poisson processes that are independent of the disorder times. Here we show that the quickest
detection problem reduces to solving an optimal stopping problem for a three-dimensional statistic.
In Section 4, we analyze the path behavior of this sufficient statistic. In Section 5, we provide a tight
upper bound on the continuation region of the optimal stopping problem, which can be used to determine
approximate detection rules besides helping us to determine the location and the shape of the continuation
region. Here, we also show that the smallest optimal stopping time of the problem under consideration has
finite expectation. In Section 6, we convert the optimal stopping problem into sequences of deterministic
optimal stopping problems using a suitably defined integral operator. In Section 7, we construct optimal
stopping times from sequences of stopping (alarm) times that sound before the processes X 1 and X 2
jump a certain number of times. In Section 8 we discuss the structure of the optimal stopping regions.
And finally, we discuss how to extend our approach to the case with more than two sources, and to the
case when the jump sizes are random and the jump size distribution changes at the time of disorder.
2. Problem Description. Let us start with a probability space (Ω, F , P0 ) that hosts two independent Poisson processes X 1 and X 2 , both of which have rate β, as well as two independent random
variables θ1 and θ2 independent of the Poisson processes with distributions
and P0 (θi > t) = (1 − πi )e−λi t ,
P0 (θi = 0) = πi
(2.1)
for 0 ≤ t < ∞, i ∈ {1, 2} and for some known constants πi ∈ [0, 1) and λ > 0 for i ∈ {1, 2}. We denote
by F = {Ft }0≤t<∞ the filtration generated by X 1 and X 2 , i.e., Ft = σ(Xs1 , Xs2 , 0 ≤ s ≤ t), and denote
by G = {Gt }0≤t<∞ the initial enlargement of F by θ1 and θ2 , i.e., Gt , σ(θ1 , θ2 , Xs2 , Xs2 : 0 ≤ s ≤ t). The
processes X 1 and X 2 satisfy (1.2) under a new probability measure P, which is characterized by
dP
dP0
, Zt , Zt1 Zt2 ,
(2.2)
Gt
where
Zti , exp
Z
t
log
0
hi (s−)
β
dXsi −
Z
0
t
[hi (s) − β]ds ,
(2.3)
for t ≥ 0 and i ∈ {1, 2} are exponential martingales (see e.g. [5]). Under this new probability measure P
of (2.2), θ1 and θ2 have the same distribution as they have under the measure P0 , i.e., their distribution is
given by (1.1). This holds because θ1 and θ2 are G0 -measurable and dP/dP0 G0 = 1, i.e., P and P0 coincide
on G0 . Under the new probability measure P the processes X 1 and X 2 have measurable intensities h1 and
h2 respectively. That is to say that (1.2) holds. In other words, the probability space (Ω, F , P) describes
the model posited in (1.1) and (1.2). Now, our problem is to find a quickest detection rule for the disorder
times θ1 ∧ θ2 , which is adapted to the history F generated by the observed processes X 1 and X 2 because
the complete information (concerning θ1 and θ2 ) embodied in G is not available. We will achieve our goal
by finding an F stopping time that minimizes (1.4).
In terms of the exponential likelihood processes
Lit ,
Xti
α
exp(−(α − β)t), t ≥ 0, i ∈ {1, 2},
β
(2.4)
we can write
Zti = 1{θi >t} + 1{θi ≤t}
Lit
Liθi
(2.5)
Let us introduce the posterior probability process
E0 Zt 1{θ≤t} Ft
,
Πt , P θ ≤ t Ft =
E0 Zt Ft
(2.6)
4
E. BAYRAKTAR AND H. V. POOR
where the second equality follows from the Bayes formula (see e.g. [9]). Then it follows that from (2.5)
and (2.6) that
1 − Πt =
(1 − π)e−2λt
,
E0 Zt Ft
where
(2.7)
π , 1 − (1 − π1 )(1 − π2 ).
(2.8)
Let us now introduce the odds-ratio process
Φt ,
Πt
, 0 ≤ t < ∞.
1 − Πt
(2.9)
Then observe from (2.6) and (2.7) that
E0 Zt 1{θ≤t} |Ft = (1 − π)e−λt Φt ,
(2.10)
t ≥ 0. Now, we will write the penalty function of (1.4) in terms of the odds-ratio process.
Z ∞
Z ∞
E (τ − θ)+ = E
E0 1{τ >t} E0 Zt 1{θ≤t} Ft dt
1{τ >t} 1{θ≤t} dt =
0
0
Z τ
= (1 − π)E0
e−2λt Φt dt.
(2.11)
0
Since {τ < θ} ∈ Gθ we can write
P(τ < θ) = E0 Zθ 1{τ <θ} = P0 (τ < θ) = (1 − π) 1 − λE0
Z
τ
0
e
−2λt
dt
,
(2.12)
where the second equality follows since Zθ = 1 almost surely under P0 . Using (2.11) and (2.12) we can
write the penalty function as
Z τ
λ
dt .
(2.13)
Rτ (π1 , π2 ) = 1 − π + c(1 − π)E0
e−2λt Φt −
c
0
On the other hand the following lemma obtains a representation for the odds ratio process Φ.
Lemma 2.1. Let us denote
Φit
#
"
P(θi ≤ t|Ft )
eλt
Lit
i
,
E0 1{θi ≤t} i Ft =
,
1 − πi
1
−
P(θi ≤ t|Ft )
L θi
(2.14)
for t ≥ 0 and i ∈ {1, 2}. Then we can write the odds-ratio process Φ as
Φt = Φ1t + Φ2t + Φ1t Φ2t ,
t≥0
(2.15)
Proof. From (2.10)
e2λt
E0 Zt 1{θ≤t} |Ft ,
(1 − π)
"
(
"
#
#
2
L1t
L
e2λt
P0 (θ1 > t)E0 1{θ1 ≤t} 1 Ft1 + P(θ2 > t)E0 1{θ2 ≤t} 2t Fti
=
(1 − π)
L θ1
L θi
"
# "
#)
L2
L1
+ E0 1{θ1 ≤t} 1t Ft1 E0 1{θ2 ≤t} 2t Ft2
L θ1
L θ2
Φt =
(2.16)
Quickest Detection of a Minimum of Two Poisson Disorder Times
5
The second equality follows from (2.2), (2.5) and the independence of the sigma algebras Ft1 and Ft2 .
Now the claim follows from (2.1), (2.8) and (2.14).
Using the fact that the likelihood ratio process Li is the unique solution of the equation
dLit = [(α/β) − 1]Lit− (dX i − α dt),
Li0 = 1,
(2.17)
(see e.g. [13]) and by means of the chain-rule we obtain
dΦit = (λ + (λ − α + β)Φit )dt + [(α/β) − 1]Φit dXti ,
Φi0 =
πi
,
1 − πi
(2.18)
for t ≥ 0 and i ∈ {1, 2} (see [3]). If we let
×
1
2
1 2
Φ+
t , Φt + Φt , Φt , Φt Φt , t ≥ 0,
(2.19)
then using a change of variable formula for jump processes gives
+
×
×
1
2
dΦ×
t = [λΦt + 2aΦt ]dt + (α/β) − 1)Φt d(Xt + Xt ),
+
1
1
2
2
dΦ+
t = [2λ + aΦt ]dt + ((α/β) − 1)[Φt dXt + Φt dXt ]
(2.20)
+
with Φ×
0 = π1 π2 /[(1 − π1 )(1 − π2 )], and Φ0 = π1 /(1 − π1 ) + π2 /(1 − π2 ), where a , λ − α + β. Note that
1
2
Xt , Xt + Xt , t ≥ 0, is a Poisson process with rate 2β under P0 .
It is clear from (2.18) and (2.20) that
Υ , (Φ× , Φ+ , Φ1 ),
(2.21)
is a piece-wise deterministic Markov process; therefore the original change detection problem with penalty
function (1.4) has been reformulated as (2.13) and (2.18)-(2.21), which is an optimal stopping problem for
a two dimensional Markov process driven by three dimensional piecewise-deterministic Markov process.
We will denote by A the infinitesimal generator of Υ. Its action on a smooth test function f : B3+ → R
is given by
[Af ](φ× , φ+ , φ1 ) = Dφ× f (φ× , φ+ , φ1 )[λφ+ + 2aφ+ ] + Dφ+ f (φ× , φ+ , φ1 )[2λ + aφ+ ]
α
α × +
α
+ Dφ1 f (φ× , φ+ , φ1 )[λ + aφ1 ] + β f
φ ,φ +
− 1 φ1 , φ1 − f (φ× , φ+ , φ1 )
β
β
β
α × α +
α
+β f
φ , φ −
− 1 φ1 , φ1 − f (φ× , φ+ , φ1 ) .
β
β
β
(2.22)
Let us denote
√
B2+ , {(x, y) ∈ R2+ : y ≥ 2 x} and
(2.23)
√
B3+ , {(x, y, z) ∈ R3+ : y ≥ 2 x, y ≥ z}.
(2.24)
Now, for every (φ× , φ+ , φ1 ) ∈ B3+ , let us denote denote by x(t, φ× ), y(t, φ+ ) and z(t, φ1 ), t ∈ R, the
solutions of
d
x(t, φ× ) = [λy(t, φ+ ) + 2ax(t, φ× )]dt, x(0, φ× ) = φ× ,
dt
d
y(t, φ+ ) = [2λ + ay(t, φ+ )]dt, y(0, φ+ ) = φ+ ,
dt
d
z(t, φ1 ) = [λ + az(t, φ1 )]dt, z(0, φ1 ) = φ1 .
dt
(2.25)
6
E. BAYRAKTAR AND H. V. POOR
The solutions of (2.25), when a 6= 0, are explicitly given by
2λ
λ2
λ
λ2
φ+ +
,
x(t, φ× ) = 2 + e2at φ× − 2 + e2at (1 − e−at )
a
a
a
a
2λ
2λ
y(t, φ+ ) = −
,
+ eat φ+ +
a
a
λ
λ
.
z(t, φ1 ) = − + eat φ+ +
a
a
(2.26)
Otherwise, x(t, φ× ) = φ× + λtφ× + λ2 t2 , y(t, φ+ ) = φ+ + 2λt, and z(t, φ1 ) = φ1 + λt. Note that the
solution (x, y, z) of the system of equations in (2.25) satisfy the semi-group property, i.e., for every s, t ∈ R
x(t + s, φ0 ) = x(s, x(t, φ0 )),
y(t + s, φ1 ) = y(s, y(t, φ1 )),
and z(t + s, φ1 ) = z(s, z(t, φ1 )).
(2.27)
Note from (2.18), (2.20) and (2.26) that
+
1
1
+
×
Φ×
t = x(t − σn , Φσn ), Φt = y(t − σn , Φσn ), Φt = z(t − σn , Φσn ) σn ≤ t < σn+1 , n ∈ N,
(2.28)
and
α ×
α
Φ
, Φ1σn+1 = Φ1σn+1 1{Xσ1
and,
1
6=Xσ
},
n+1
n+1−
β σn+1−
β
–
»
–
«
„
«
»
„
α +
α
α
− 1 Φ1σn+1− 1{Xσ1
Φσn+1 − −
− 1 Φ1σn+1− 1{Xσ2
= Φ+
1
2
6=Xσ
} +
6=Xσ
}
σn+1 − +
n+1
n+1−
n+1
n+1−
β
β
β
(2.29)
Φ×
σn+1 =
Φ+
σn+1
Here, for any function h, h(t−) , lims↑t h(t). Note that an observer watching Υ is able to tell whenever
the processes X 1 and X 2 jump (see (2.18), (2.20)), i.e., the filtration generated by Υ is the same as F.
3. An Optimal Stopping Problem. Let us denote the set of F stopping times by S. The value
function of the quickest detection problem
U (π1 , π2 ) , inf Rτ (π1 , π2 )
(3.1)
π1
π1 π2 π1 + π2 − 2π1 π2
,
,
,
U (π1 , π2 ) = (1 − π) 1 + cV
1−π
1−π
1 − π1
(3.2)
τ ∈S
can be written as
where V is the value function of the optimal stopping problem
Z τ
(φ× , φ+ , φ1 )
×
+
−λt
×
+
1
e h Φt , Φt dt ,
V (φ , φ , φ ) , inf E0
τ ∈S
(3.3)
0
(φ× , φ+ , φ1 )
in which (φ× , φ+ , φ1 ) ∈ D3+ , and h(x, y) , x + y − λ/c. Here E0
+
×
+
1
1
given that Φ×
0 = φ , Φ0 = φ , Φ0 = φ .
is the expectation under P0
+
It is clear from (3.3) that for both optimal stopping problems it is not optimal to stop before (Φ×
t , Φt ),
t ≥ 0, leaves the advantageous region defined by
C0 , {(φ× , φ+ ) ∈ B2+ : φ× + φ+ ≤ λ/c}.
(3.4)
C , {(φ× , φ+ , φ1 ) ∈ B3+ : φ× + φ+ ≤ λ/c}.
(3.5)
Let us also denote
Also note that the only reason not to stop at the time of the first exit from the region C0 is the prospect
+
of (Φ×
t , Φt ), t ≥ 0 returning to C0 at a future time.
Remark 3.1. It is reasonable to question our choice of statisitc, since it is clear that (Φ1t , Φ2t )t≥0
contains all the information X has to offer. Our choice (Υt )t≥0 , which is defined in (2.21), is motivated
Quickest Detection of a Minimum of Two Poisson Disorder Times
7
by the mere desire of having a concave value function U and a convex optimal stopping region. The
concavity is due to the linearity of the function h (see Lemma 6.2 and its proof along with Lemmas 6.1
and (6.1), and Theorem 6.1).
If we had chosen to work with (Φ1t , Φ2t )t≥0 , then the relevant optimal stopping problem becomes
Z τ
(a,b)
W (a, b) , inf E0
e−λt h̃ Φ1t , Φ2t dt ,
τ ∈S
(3.6)
0
in which
h̃(x, y) , x + y + xy,
(x, y) ∈ R2+ .
(3.7)
Since h(·, ·) is non-linear the concavity of the value function, i.e. W (·, ·) is not concave. In fact, W (x, y) =
V (xy, x + y, x). The function V is concave but W is not. So there is a trade off between concavity and
the dimension of the statistic to be used.
In what follows, for the sake of the simplicity of notation, when the meaning is clear, we will drop the
(φ× , φ+ , φ1 )
superscripts of the expectation operators E0
.
4. Sample Paths of Ψ = (Φ× , Φ+ ). It is illustrative to look at the sample paths of the sufficient
statistic Ψ, to understand the nature of the problem. Indeed, this way, for a certain parameter range, we
will be able to identify the optimal stopping time without any further analysis.
From (2.26), we see that, if a > 0, then the paths of the processes Φ× and Φ+ increase between the
jumps, and otherwise the paths of the processes Φ× and Φ+ mean-revert to the levels 2λ2 /a2 and −2λ/a
respectively. Also observe that Φ× , Φ+ increase (decrease) with a jump if α ≥ β (β > α). See (2.20).
Case I: α ≥ β, a > 0. The following theorem follows from the description of the behavior of the paths
above.
Theorem 4.1. If α > β and a > 0, then the stopping rule
+
τ0 , inf{t ≥ 0 : Φ×
t + Φt ≥ λ/c}
(4.1)
is optimal for (3.3).
Proof. Under the hypothesis of the theorem and whenever a path of (Φ× , Φ+ ) leaves C0 it never returns.
In Section 5, we will identify another case (another range of parameters) in which the advantageous
region C0 is the optimal continuation region and the stopping time τ0 is optimal (see Cases II-b-i-1 and
II-b-ii-1).
Remark 4.1. Let κi , inf{t ≥ 0 : Φit ≥ λ/c}. If α ≥ β and a > 0, then κi is the optimal stopping
time for the one dimensional disorder problem with disorder time θi ([3]). Let us define κ , κ1 ∧ κ2 .
+
Since with this choice of parameters Φ×
κ + Φκ > λ/c, it follows that τ0 < κ almost surely. Therefore, if
we follow the rule dictated by the stopping time κ, then we pay an extra penalty for detection delay. This
example illustrates that solving the two one dimensional quickest detection problems separately in order
to minimize the penalty function of (1.4) is suboptimal.
In what follows, we will consider the remaining cases: α ≥ β and a < 0; α < β.
5. Construction of a Bound on the Continuation Region. In this section the purpose is to
show that the continuation region of (3.3) is bounded. The construction of upper bounds is carried out
in the next two theorems. These upper bounds are tight as the next theorem shows and might be used to
construct useful approximations to the two optimal stopping times solving the problems defined in (3.3).
We will carry out the analysis for a = λ − α + β 6= 0. A similar analysis for this case can be similarly
performed. As a result of Theorem 5.3 we are also able to conclude that the (smallest) optimal stopping
time has a finite expectation.
8
E. BAYRAKTAR AND H. V. POOR
The first two theorems in this section assume that an optimal stopping time of (3.3) exists and in
particular the stopping time
τ ∗ (a, b, c) , inf{t ≥ 0 : V (Υt ) = 0, Υ0 = (a, b, c)},
(5.1)
is optimal. In Section 7, we will verify that this assumption holds. We will denote by
Γ , {(a, b, c) ∈ B3+ : v(a, b, c) = 0},
C , B3+ − Γ,
(5.2)
the optimal stopping region and optimal continuation region of (3.3) respectively.
Theorem 5.1. In this theorem the standing assumption is that α ≥ β and that a < 0 (Case II).
(Case II-a) Let us further assume that λ/a2 − 2/a ≤ 1/c and denote
1
3λ − 2a
D0 , (x, y) ∈ B2+ : x ·
+y·
+k >0 ,
2(λ − a)
2(λ − a)(2λ − a)
(5.3)
in which
λ
1
1
λ2
1
k, 2− −
+ 2
+
+2
2a
a 2c 2a (λ − a) 2λ − a
λ λ2
− 2
a
a
.
(5.4)
Let (φ0 , φ1 ) ∈ D0 ∩ (B2+ − C0 ). Then for any φ2 ≤ φ1 , (φ0 , φ1 , φ2 ) is in the stopping region of (3.3).
(Case II-b) Assume that λ/a2 − 2/a ≥ 1/c (standing assumption in the rest of the theorem). Consider
the four different possible ranges of parameters:
Case II-b-i: If λ + a ≤ 0
• and if −a/c − 1 ≤ 0 (Case II-b-i-1), then C in (3.5) is the optimal continuation region for (3.3).
• Else if −a/c − 1 > 0 (Case II-b-i-2), then a superset of the continuation region can be constructed
as follows. Let us introduce the line segment
C , {(x, y) ∈ B2+ : x + y = λ/c},
(5.5)
and define
C1 , (x, y) ∈ B2+ : x = x(t, x∗ ), y = y(t, 0) for t ∈ [0, t∗ ]
[
(5.6)
a
2λ
a
2λ
λ
1+
,y>−
1+
.
(x, y) ∈ C : x < +
c
λ−a
c
λ−a
c
2λ
2λ
1 + ac ) = 0 and x∗ = x(−t∗ , λc + λ−a
1 + ac ) The curve
Here, t∗ is the solution of y(−t∗ , − λ−a
C1 separates R2+ into two connected regions. Let us denote the region that lies above the curve
C1 by
D1 , {(x, y) ∈ B2+ : there exists a positive number ỹ(x) < y such that (x, ỹ(x)) ∈ C1 }.
(5.7)
Then [(B2+ − D1 ) × R+ ] ∩ B3+ is an upper bound on the continuation region of (3.3).
Case II-b-ii: If λ + a > 0
• and if −a/c − 1 < 0 (Case II-b-ii-1), then C in (3.5) is the optimal continuation region for (3.3).
• Else if −a/c − 1 > 0, then [(B2+ − D1 ) × R+ ] ∩ B3+ is an upper bound on the continuation region
of (3.3).
Note that all the supersets of the continuation we constructed are bounded subsets of R3+ .
Proof.
Note that
Φ+
t ≥ y(t, φ1 ),
Φ×
t ≥ x(t, φ0 ), t ≥ 0,
(5.8)
9
Quickest Detection of a Minimum of Two Poisson Disorder Times
×
×
+
almost surely if Φ+
0 = φ1 and Φ0 = φ0 . This is because Φ and Φ increase with jumps.
From this observation we obtain the following inequality
Z τ
Z
+
inf E0
e−2λs h(Φ×
,
Φ
)ds
≥
inf
E
0
s
s
τ ∈S
0
=
Z
inf
t∈[0,∞]
τ ∈S
τ
e−2λs h(x(s, φ0 ), y(s, φ1 ))ds
0
t
e−2λs h(x(s, φ0 ), y(s, φ1 ))ds.
(5.9)
(5.10)
0
Note that if for a given (φ0 , φ1 ) the expression in (5.10) is equal to zero, then the infimum on the left
hand side of (5.9) is attained by setting τ = 0. In what follows we will find a subset of the stopping
region of the optimal stopping problem using this argument.
Case II-a: λ/a2 − 2/a ≤ 1/c. In this case the mean reversion level of the path (x(·, φ0 ), y(·, φ1 )),
(φ0 , φ1 ) ∈ B2+ , namely (λ2 /a2 , −2λ/a), is inside the region C0 which is defined in (3.4). In this case, for
any (φ0 , φ1 ) ∈ B2+ − C0 the minimizer topt (φ0 , φ1 ) of the expression in (5.10) is either 0 or ∞ by the
following argument. For any (φ0 , φ1 ) ∈ B2+ − C0 the path (x(·, φ0 ), y(·, φ1 )) is in the advantageous region
C0 all the time except for possibly a finite duration. Therefore if
Z ∞
e−2λs h(x(s, φ0 ), y(s, φ1 ))ds < 0,
(5.11)
0
then in order to minimize (5.10) it is never optimal to stop. On the other hand if (5.11) is positive, then
it is not worth taking the journey into the advantageous region and it is optimal to stop immediately in
order to minimize (5.10).
We shall find the pairs (φ0 , φ1 ) for which topt = 0. Using (2.26) we can write
Z ∞
a
1
−2λs
+ φ1
+ k,
e
h(x(s, φ0 ), y(s, φ1 ))ds = φ0 −
α−β
(α − β)2
0
(5.12)
where k is given by (5.4). Note that if (φ0 , φ1 ) ∈ D0 ∩ (B2+ − C0 ), then by (5.9) and (5.10) we can see
that the infimum in (5.10) is equal to 0. Therefore [((B2+ − D0 ) ∪ C0 ) × R+ ] ∩ B3+ is a superset of the
optimal continuation region of (3.3).
Case II-b: λ/a2 − 2/a ≥ 1/c. In this case the mean reversion level of t → (x(t, φ0 ), y(t, φ1 )) is outside
C0 . Therefore, the minimizer of (5.10) is topt (φ0 , φ1 ) ∈ {0, tc(φ0 , φ1 ), ∞} where tc (φ0 , φ1 ) is the exit time
of the path (x(t, φ0 ), y(t, φ1 )) from C0 . The derivative
d
[x(t, φ0 ) + y(t, φ1 )] = (λ + a)y(t, φ1 ) + 2ax(t, φ1 ) + 2λ
dt
(5.13)
vanishes if (x(t, φ0 ), y(t, φ1 )) meets the line segment
L = {(x, y) ∈ B2+ : (λ + a)y + 2ax + 2λ = 0}.
Note that the mean reversion level belongs to L, i.e.,
2
λ
2λ
∈ L.
,−
a2
a
(5.14)
(5.15)
Case II-b-i: λ + a < 0. (In addition to α > β, a < 0 and λ/a2 − 2/a ≥ 1/c.) In this case the line L is
decreasing (as a function of x).
Case II-b-i-1: −a/c − 2 < 0. (In addition to α > β, a < 0, λ/a2 − 2/a ≥ 1/c and λ + a < 0.) In this
case the line segment C in (5.5) lies entirely below L. Assume that a path (x(·, φ0 ), y(·, φ1 )) originating
at (φ0 , φ1 ) ∈ B2+ − C0 enters C0 at time t0 > 0. This path must leave C0 at time t1 < ∞ since the mean
reversion level (λ2 /a2 , −λ/a) ∈
/ C0 . This implies that for any t ∈ (t0 , t1 ) x(t, φ0 ) + y(t, φ1 ) < λ/c and
10
E. BAYRAKTAR AND H. V. POOR
x(t0 , φ0 ) + y(t0 , φ1 ) = λ/c. This yields a contradiction, because λ + a < 0 together with (5.13) implies
that t → x(t, φ0 ) + y(t, φ1 ) is increasing below the line segment L. Therefore the minimizer topt (φ0 , φ1 )
of (5.10) is equal to 0 if (φ0 , φ1 ) ∈
/ C0 , and it is equal to tc (φ0 , φ1 ) if (φ0 , φ1 ) ∈ C0 . From (5.9) we can
conclude that C is equal to the optimal continuation region of (3.3).
( λc
I
I
Case II-b-i-2 : −a/c − 1 > 0.
In this case the line segments C and L intersect at I∗ = (x , y ) ,
2λ
a
2λ
a
+ λ−a 1 + c , − λ−a 1 + c ). By running the paths backward in time, we can find x such that
(x∗ , 0) = x −t∗ , xI , y −t∗ , y I .
(5.16)
By the semi-group property (2.27), we have
x(t∗ , x∗ ) = x t∗ , x −t∗ , xI
= x 0, xI = xI .
= x t∗ + (−t∗ ), xI
(5.17)
Similarly, y(t∗ , 0) = y I . The function t → x(t, x∗ ) + y(t, 0) is decreasing on (0, t∗ ) and increasing on
(t∗ , ∞). It follows that the path t → (x(t, x∗ ), y(t, 0)) is tangential to C at I and lies above the region
C0 .
We will now show that if a path (x(·, φ0 ), y(·, φ1 )) originates in D1 , then it stays in D1 . Let us first
consider a pair (φ0 , φ1 ) ∈ D1 such that φ1 < −2λ/a. Consider the curve
P , (x, y) ∈ B2+ : x = x(t, x∗ ), y = y(t, 0) for t ∈ [0, ∞) .
(5.18)
The following remark will be useful in completing the proof.
Remark 5.1. The semi-group property in (2.27) implies that two distinct curves (x(·, φa0 ) , y(·, φa1 ))
and (x(·, φb0 ), y(·, φb1 )) do not intersect. If
(x(ta , φa0 ), y(ta , φa1 )) = (x(tb , φb0 ), y(tb , φb1 )) = (φ0 , φ1 )
(5.19)
for some ta , tb ∈ R then (2.27) implies that
(x(t, φa0 ), y(t, φa1 )) = (x(ta + (t − ta ), φa0 ), y(ta + (t − ta ), φa1 ))
= (x(t − ta , φ0 ), y(t − ta , φ1 )) = (x(tb + (t − ta ), φb0 ), y(tb + (t − ta ), φb1 ))
b
a
= (x(t − t +
t, φb0 ), y(tb
a
−t +
t, φb1 )),
for all
(5.20)
t ∈ R,
i.e., the two curves are identical after a reparametrization.
If the point (φ0 , φ1 ) lies above P , and we recall that P lies above C0 , then by Remark 5.1 the
path (x(·, φ0 ), y(·, φ1 )) will lie above C0 . If the point (φ0 , φ1 ) lies between P and C0 , then the path
(x(·, φ0 ), y(·, φ1 )) will lie below the line segment L. This observation together with the fact that λ + a <
0, implies (using (5.13)) that the function t → x(t, φ0 ) + y(t, φ1 ) is increasing. Therefore the path
(x(·, φ0 ), y(·, φ1 )) cannot intersect C0 .
Now let us consider a pair (φ0 , φ1 ) ∈ D1 such that φ1 > −2λ/a. If (φ0 , φ1 ) lies above L, then the
function t → x(t, φ0 ) + y(t, φ1 ) is decreasing and its range is [φ0 + φ1 , 2λ/a(λ/a − 1)), which is always
above λ/c, and therefore the path (x(·, φ0 ), y(·, φ1 )) does not enter C0 . If the point (φ0 , φ1 ) lies below L
then t → x(t, φ0 )+y(t, φ1 ) is increasing. This monotonicity implies that the path (x(·, φ0 ), y(·, φ1 )) cannot
visit C0 . If φ1 = −2λ/a, then y(t, φ1 ) = −2λ/a for all t ≥ 0. x(t, φ0 ) increases or decreases depending
on whether (φ0 , −2λ/a) is below or above L. Therefore if (φ0 , −2λ/a) ∈
/ C0 , then (x(·, φ0 ), y(·, φ1 )) never
visits C0 . These arguments show that if a path (x(·, φ0 ), y(·, φ1 )) originates in D1 , then it stays in D1 .
Therefore if (φ0 , φ1 ) ∈ D1 , then the infimum in (5.10) is equal to 0 (by (5.9) and (5.10)). Therefore
[(B2+ − D1 ) × R+ ] ∩ B3+ is a superset of the optimal continuation region of (3.3).
Case II-b-ii: λ + a > 0. In this case L is increasing (as a function of x). The function t → x(t, φ0 ) +
y(t, φ1 ) is increasing if (φ0 , φ1 ) lies above L, and it is decreasing otherwise.
11
Quickest Detection of a Minimum of Two Poisson Disorder Times
Case II-b-ii-1: −a/c − 1 ≤ 0. In this case the line segments L and C do not intersect. Let us first
consider a pair (φ0 , φ1 ) ∈ B2+ such that φ1 < −2λ/a. If (φ0 , φ1 ) ∈
/ C0 lies above the line segment L, then
t → x(t, φ0 ) + y(t, φ1 ) is increasing and the path (x(·, φ0 ), y(·, φ1 )) cannot enter C0 . Consider the curve
λ
2
, y = y(t, 0) for t ∈ [0, ∞) ,
(5.21)
P̃ , (x, y) ∈ B+ : x = x t, −
a
which starts at the intersection of L with the x-axis. The semi-group property Remark 5.1 implies that
no path starting to the right of P̃ intersects P̃ and therefore lies to the right of the region C0 . Therefore,
if (φ0 , φ1 ) is below the line segment L, then the path (x(·, φ0 ), y(·, φ1 )) never visits the advantageous
region C0 . (Note that if the path (x(·, φ0 ), y(·, φ1 )) meets the line L at time tL (φ0 , φ1 ), then t →
(x(t, φ0 ) + y(t, φ1 )) is decreasing (increasing) on [0, tL ] ([tL , ∞).)
Now let us consider a point (φ0 , φ1 ) ∈ B2+ − C0 such that φ1 > −2λ/a. Then t → (x(t, φ0 ) + y(t, φ1 )) is
increasing on [0, tL (φ0 , φ1 )] and is decreasing on (tL (φ0 , φ1 ), ∞) (it decreases to −2λ/a + λ2 /a2 > λ/c).
And the monotonicity of t → x(t, φ0 )+y(t, φ1 ) on [0, tL (φ0 , φ1 )] implies that x(t, φ0 )+y(t, φ1 ) > λ/c for t ∈
[0, tL (φ0 , φ1 )]. If φ1 = −2λ/a, then y(t, φ1 ) = −2λ/a for all t ≥ 0. x(t, φ0 ) increases (decreases) depending
on whether (φ0 , −2λ/a) is above or below L. These arguments show that if a path (x(·, φ0 ), y(·, φ1 ))
originates in B2+ − C0 , then it stays in B2+ − C0 . Therefore the minimizer topt (φ0 , φ1 ) of (5.10) for any
φ0 , φ1 ∈ B2+ − C0 is equal to zero. Now using (5.9) and (5.10) the optimal continuation region of (3.3) is
equal to C.
Case II-b-ii-2: −a/c−1 > 0. In this case the line segments C and L intersect at I = (xI , y I ). Arguments
similar to those of Case II-b-i-2 show that [(B2+ − D1 ) × R+ ] ∩ B3+ , in which D1 is defined in (5.7), is a
superset of the optimal continuation region of (3.3).
Theorem 5.2. Let us assume that α < β ( Case III: α < β) and define
λ + 2β
.
D2 , (x, y) ∈ B2+ : x + y ≥
c
(5.22)
Then [(B2+ − D1 ) × R+ ] ∩ B3+ , which is a bounded region in R3+ , is an upper bound on the continuation
region of (3.3).
Proof. Note that in this case a > 0. The paths of the processes Φ× , Φ+ increase between the jumps
and decrease with a jump. If τ ∈ S then there is a constant t ≥ 0 such that τ ∧ σ1 = t ∧ σ1 almost surely.
Hence we can write
Z τ
Z τ ∧σ1
Z τ
e−λs h(Ψs )ds
E0
e−λs h(Ψs )ds + E0 1{τ ≥σ1 }
e−λs h(Ψs )ds = E0
0
0
= E0
Z
t∧σ1
0
Z
t∧σ1
e−λs h(Ψs )ds + E0 1{t≥σ1 }
Z
σ1
τ
e−λs h(Ψs )ds
σ1
1
e−λs h(Ψs )ds − E0 1{t≥σ1 } e−λσ1
≥ E0
c
0
Z t
2β
e−(λ+2β)s h(x(s, φ0 ), y(s, φ1 )) −
ds,
=
c
0
(5.23)
using also the fact that σ1 has exponentialRdistribution with
rate 2β. From (5.23) it follows that if
τ
x(s, φ0 ) + y(s, φ1 ) − (λ + 2β)/c > 0, then E0 0 e−λs h(Ψs )ds > 0 for every stopping time τ 6= 0, τ ∈ S.
Since the paths x(t, φ0 ), y(t, φ1 ) are increasing we can conclude that stopping immediately is optimal for
(3.3). That is τ = 0 is optimal for (3.3) if (φ0 , φ1 ) ∈ D2 and φ2 ≤ φ1 , in which D2 is as in (5.22).
Theorems 5.1 and 5.2 can be used to determine approximate detection rules besides helping us to
determine the location and the shape of the continuation region. As we have seen in Cases II-b-i-1 and
II-b-ii-1, these approximate rules turn out to be tight. The next theorem is essential in proving the fact
that the smallest optimal stopping time of (3.3) has a finite expectation.
Theorem 5.3. Let τD be the exit time of the process Υ from a bounded region D ⊂ B3+ . Then
E0φ
×
,φ+ ,φ1
[τD ] < ∞ for every (φ× , φ+ , φ1 ) ∈ B3+ . Hence τ ∗ defined in (5.1) has a finite expectation.
12
E. BAYRAKTAR AND H. V. POOR
Proof. Let f (φ× , φ+ , φ1 ) , φ× + φ+ . Then it follows from (2.22) that
α
α +
φ + φ× +
− 1 φ1 − φ× − φ+
[Af ](φ× , φ+ , φ1 ) = λφ+ + 2aφ× + 2λ + aφ+ + β
β
β
α
α + α ×
φ + φ −
− 1 φ1 − φ× − φ+ = 2λ(φ× + φ+ + 1) ≥ 2λ,
β
β
β
β
(5.24)
for every (φ× , φ+ , φ1 ) ∈ B3+ . Since f is bounded on D, and τD ∧ t is a bounded F-stopping time, we have
E0 [f (ΥτD ∧t )] = f (Υ0 ) + E0
Z
τD ∧t
0
[Af ](Υs )ds ≥ 2λE0 [τD ∧ t].
(5.25)
On the other hand
E0 [f (ΥτD ∧t )] ≤
α
ξ,
β
in which ξ = min{a ∈ R+ : for any (x, y, z) ∈ D, max(x, y, z) ≤ a} < ∞. An application of the monotone
convergence theorem implies that E0 [τD ] < ∞.
The results of this section can be used to determine approximate detection rules besides helping us to
determine the location and the shape of the continuation region. As we have seen in Cases II-b-i-1 and
II-b-ii-1, these approximate rules turn out to be tight.
6. Optimal Stopping with Time Horizon σn . In this section, we will first approximate the
optimal stopping problem (3.3) by a sequence of optimal stopping problems. Let us denote
Z τ ∧σn
a,b,c
×
+
−λt
(6.1)
e h Φt , Φt dt ,
Vn (a, b, c) , inf E0
τ ∈S
0
for all (a, b, c) ∈ B3+ and n ∈ N. Here, σn is the nth jump time of the process X.
Observe that (Vn )n∈N is a decreasing sequence and they each of its members satisfy −1/c < Vn < 0.
Therefore the pointwise limit limn Vn exist. It can be shown that more is true using the fact that the
function h is bounded from below and σn is a sum of independent exponential random variables.
Lemma 6.1. For any (a, b, c) ∈ B3+
1
0 ≤ Vn (a, b, c) − V (a, b, c) ≤
c
Proof. For any τ ∈ S,
Z
Z τ
+
dt
=
E
E0
e−λt h Φ×
,
Φ
0
t
t
0
0
τ ∧σn
2β
2β + λ
n
(6.2)
.
Z
+
dt
+
E
e−λt h Φ×
,
Φ
1
0
{τ ≥σn }
t
t
τ
σn
+
e−λt h(Φ×
t , Φt )dt
(6.3)
The first term on the right-hand-side of (6.3) is greater than Vn . Since h(·, ·) > −λ/c we can show that
the second term is greater than
n
Z τ
1 φ0 ,φ1 −λσn
2β
1
λ φ0 ,φ1
−λs
e
ds ≥ − E0
1{τ ≥σn }
.
(6.4)
≥−
e
− E0
c
c
c λ + 2β
σn
To show the last inequality we have used the fact that σn is a sum of n independent and identically
distributed exponential random variables with rate 2β. Now, the proof of the lemma follows immediately.
Quickest Detection of a Minimum of Two Poisson Disorder Times
13
As in [4] and [6] to calculate the value functions Vn iteratively we introduce the functional operators
J, Jt . These operators are defined through their actions on bounded functions g : B3+ → R as follows:
Z t∧σ1
1
+
×
−λs
×
+
−λσ1
)
and,
,
Φ
,
Φ
g(Φ
e
h(Φ
,
Φ
)ds
+
1
e
[Jg](t, a, b, c) , Ea,b,c
{t≥σ
}
σ1 ,
σ1
σ1
s
s
0
1
0
(6.5)
[Jt g](a, b, c) , inf [Jg](s, a, b, c), t ∈ [0, ∞].
s∈[t,∞]
Observe that
α 1
α
α ×
+
1
1
+
1
Φ
,
Φ
+
−
1
Φ
,
Φ
1 1
g
)
=
E
,
Φ
,
Φ
E0 1{t≥σ1 } e−λσ1 g(Φ×
0
σ1 −
σ1
σ1
σ1
β σ1 − σ1 −
β
β σ1 − {Xσ1 6=Xσ1 − }
α × α +
α
+g
Φσ1 − , Φσ1 − −
− 1 Φ1σ1 − , Φ1σ1 − 1{Xσ2 6=Xσ2 − } 1{t≥σ1 } e−λσ1
1
1
β
β
β
Z t
α
α
α
1
x(s, a), y(s, b) +
− 1 z(s, c), z(s, c) ds
2βe−(λ+2β)s g
=
2 0
β
β
β
Z t
α
1
α
α
−(λ+2β)s
x(s, a), y(s, b) −
− 1 z(s, c), z(s, c) ds.
+
2βe
g
2 0
β
β
β
(6.6)
To derive (6.6) we used the fact that σ1 has exponential distribution with rate 2β, the dynamics in (2.20),
and the fact that conditioned on the event that there is a jump it has 1/2 probability of coming from X 1
and
Using (6.6) and Fubini’s theorem we can write
Z t
[Jg](t, a, b, c) =
e−(λ+2β)s (h + β · g ◦ (F1 + F2 ))(x(s, a), y(s, b), z(s, c))ds,
(6.7)
0
where
Fi (a, b, c) =
α
a,
β
i−1
2−i !
α
α
α
b + (−1)i
c
− 1 c,
β
β
β
i ∈ {1, 2}.
(6.8)
Using (2.26) it can be shown that
lim [Jg](t, a, b, c) = [Jg](∞, a, b, c) < ∞.
t→∞
(6.9)
Lemma 6.2. For every bounded function f , the mapping J0 f is bounded. If f is a concave function,
then J0 f is also a concave function. If f1 ≤ f2 , then J0 f1 ≤ J0 f2 .
Proof. The third assertion of the lemma directly follows from the representation (6.7). The first
assertion holds since h is bounded from below and J0 f (a, b, ) ≤ Jf (0, a, b, c) = 0. The second assertion
follows from the linearity of the functions x(t, ·), y(t, ·), h(·, ·), F1 (·, ·, ·) and F2 (·, ·, ·).
Using Lemma 6.2 we can prove the following corollary:
Corollary 6.1. Let us define a sequences of function (vn )n∈N by
v0 , 0,
vn , J0 vn−1 .
(6.10)
Then every n ∈ N, vn is bounded and concave; and vn+1 ≤ vn . Therefore, the limit v = limn vn exist, and
is bounded and concave. Moreover, vn , for all n ∈ N and v are increasing in each of their arguments.
Proof. The proof of the first part directly follows from Lemma 6.2. That vn for all n ∈ N and v are
increasing in each of their arguments follows from the fact that these functions are bounded from above
and below and that they are concave.
14
E. BAYRAKTAR AND H. V. POOR
We will need the following lemma to give a characterization of the stopping times of the filtration F
(see [5]).
Lemma 6.3. For every τ ∈ S, there are Fσn measurable random variables ξn : Ω → ∞ such that
τ ∧ σn+1 = (σn + ξn ) ∧ σn+1 P0 almost surely on {τ ≥ σn }. The main theorem of this section can be
proven by induction using Lemma 6.3 and the strong Markov property.
Theorem 6.1. For every n ∈ N, vn defined in Corollary 6.1 is equal to Vn . For ε ≥ 0, let us denote
rnε (a, b, c) , inf{t ∈ (0, ∞] : [Jvn ](t, (a, b, c)) ≤ [J0 vn ](a, b, c) + ε}.
And let us define a sequence of stopping times by S1ε , r0ε (Υ0 ) ∧ σ1 and
(
ε/2
ε/2
rn (Υ0 )
if σ1 ≥ rn (Υ0 )
ε
Sn+1 ,
ε/2
σ1 + Sn ◦ θσ1 otherwise.
(6.11)
(6.12)
Here θs is the shift operator on Ω, i.e., Xt ◦ θs = Xs+t . Then Snε is an ε-optimal stopping time of
(6.1), i.e.,
"Z ε
#
Ea,b,c
0
Sn
0
e−λt h(Ψt )dt ≤ vn (a, b, c) + ε,
(6.13)
+
in which Ψt = (Φ×
t , Φt ), t ≥ 0.
Proof. See Appendix.
Theorem 6.1 shows that the value function Vn of the optimal stopping problem defined in (6.1) and
the function vn introduced in Corollary 6.1 by an iterative application of the operator J0 are equal.
This implies that the value function of the optimal stopping problem of (6.1) can be found by solving a
sequence of deterministic minimization problems.
7. Optimal Stopping Time. Theorem 7.1. τ ∗ defined in (5.1) is the smallest optimal stopping
time for (3.3).
This theorem shows that Γ defined in (5.2) is indeed an optimal stopping region. We will divide the
proof of this theorem into several lemmas.
The following dynamic programming principle can be proven by the special representation of the
stopping times of a jump process (Lemma 6.3) and the strong Markov property.
Lemma 7.1. For any bounded function g : B3+ → R we have
[Jt g](a, b, c) = [Jg](t, a, b, c) + e−(λ+2β)t [J0 g](x(t, a), y(t, b), z(t, c)).
(7.1)
rn (a, b, c) , rn0 (a, b, c),
(7.2)
Let us denote
which is well defined because of (6.9) and the continuity of the function t → [Jf ](t, a, b, c), t ≥ 0. (See
(6.11) for the definition of rn0 .) Let us also denote
r(a, b, c) , inf{t ≥ 0 : [JV ](t, a, b, c) = J0 V (a, b, c)}.
(7.3)
Corollary 7.1. The functions rn and r defined by (7.2) and (7.3) respectively, satisfy
rn (a, b, c) = {t ≥ 0 : vn+1 (x(t, a), y(t, b), z(t, c)) = 0}
r(a, b, c) = {t ≥ 0 : v(x(t, a), y(t, b), z(t, c)) = 0},
(7.4)
Quickest Detection of a Minimum of Two Poisson Disorder Times
15
with the convention that inf ∅ = 0. Together with (7.4), Corollary 6.1 implies that rn (a, b, c) ↑ r(a, b, c)
as n ↑ ∞.
Proof.
Suppose that rn (a, b, c) < ∞. Then from (6.11) it follows that
[Jvn ](rn (a, b, c), a, b, c) = [J0 vn ](a, b, c) = [Jrn (a,b,c) vn ](a, b, c)
= [Jvn ](rn (a, b, c), a, b, c) + e−(λ+2β)rn (a,b,c) vn+1 (x(rn (a, b, c), a), y(rn (a, b, c), b), z(rn (a, b, c), c)).
(7.5)
Here the second equality follows from the definition of the operator Jt in (6.5) and the third equation
follows from Lemma 7.1 and the fact that J0 vn = vn+1 . From (7.5) it follows that vn+1 (x(rn (a, b, c), a),
y(rn (a, b, c), b), z(rn (a, b, c), c)) = 0. For t ∈ (0, rn (a, b, c)) we have [Jvn ](t, a, b, c) > [J0 vn ](a, b, c)
= [Jrn (a,b,c) vn ](a, b, c) = Jt vn (a, b, c), since the function s → [Js vn ](a, b, c) is non-decreasing. Using
Lemma 7.1 we can write
[J0 vn ](a, b, c) = [Jt vn ](a, b, c) = [Jvn ](t, a, b, c) + e−(λ+2β) vn+1 (x(t, a), y(t, b), z(t, c)),
(7.6)
which implies that vn+1 (x(t, a), y(t, b), z(t, c)) < 0 for all t < rn (a, b, c).
Now, suppose that rn (a, b, c) = ∞. Then vn+1 (x(t, a), y(t, b), z(t, c)) < 0 for every t ∈ (0, ∞) which
can be shown using same arguments as above. Therefore {t > 0 : vn+1 (x(t, a), y(t, b), z(t, c)) = 0} = ∅
and (7.4) holds.
The proof for the representation of r can be proven using the same line of argument and the fact that
J0 V = V . The fact that J0 V = V can be proven by the dominated convergence theorem, since the
sequences (vn (a, b, c))n≥0 and ([Jvn ](t, a, b, c))n≥0 are decreasing, and since vn is bounded function for
all n ∈ N.
In the next lemma we construct optimal stopping times for the family of problems introduced in (6.1).
Lemma 7.2. Let us denote Sn , Sn0 , where Snε is defined in Theorem 6.1 for ε ≥ 0. Then the sequence
(Sn )n∈N is an almost surely increasing sequence. Moreover Sn < τ ∗ almost surely for all n.
Proof. Since r1 > 0, using Corollary 7.1 we can write
r1 − r0 ,
S2 − S1 = σ1 − r0 + S1 ◦ θσ1 ,
S1 ◦ θσ1
if σ1 > r1
if r0 < σ1 ≤ r1 > 0.
if σ1 ≤ r0
(7.7)
Now, let us assume that Sn − Sn−1 > 0 almost surely. From Lemma 7.1 we have that rn > rn−1 . Using
this fact and the induction hypothesis we can write
rn − rn−1 ,
Sn+1 − Sn = σ1 − rn−1 + Sn ◦ θσ1 ,
(Sn − Sn−1 ) ◦ θσ1
which proves the first assertion of the lemma.
if σ1 > rn
if rn−1 < σ1 ≤ rn > 0,
if σ1 ≤ rn−1
(7.8)
From Corollary 7.1 and the definition of τ ∗ it follows that τ ∗ ∧ σ1 = r ∧ σ1 . Therefore τ ∗ ∧ σ1 >
r0 ∧ σ1 = S1 , since r0 < r. Now, we will assume that Sn < τ ∗ and show that Sn+1 < τ ∗ . On {σ1 ≤ rn }
we have that
Sn+1 = σ1 + Sn ◦ θσ1 < σ1 + τ ∗ ◦ θσ1 .
(7.9)
Since τ ∗ ∧ σ1 = r ∧ σ1 and r > rn , if σ1 ≤ rn , then τ ∗ ∧ σ1 = σ1 . Because τ ∗ is a hitting time, on the set
{σ1 ≤ rn } ⊂ {σ1 ≤ τ ∗ } the following holds
Sn+1 ≤ σ1 + τ ∗ ◦ θσ1 = τ ∗ .
16
E. BAYRAKTAR AND H. V. POOR
On the other hand if σ1 > rn , then τ ∗ ∧ σ1 = r ∧ σ1 > rn . Therefore on {σ1 > rn }, Sn+1 = rn < τ ∗ .
This concludes the proof of the second assertion.
+
∗
∗
∗
Lemma 7.3. Let us denote Ψt = (Φ×
t , Φt ), t ≥ 0. If S , limn Sn , then S = τ almost surely.
∗
Moreover, τ is an optimal stopping time, i.e.,
"Z ∗
#
V (a, b, c) = Ea,b,c
0
S
e−λs h(Ψs )ds .
0
Proof. The limit S ∗ , limn Sn exists since (Sn )n∈N is increasing and Sn ≤ τ ∗ < ∞ (as a corollary of
Theorem 5.3) for all n. Let us show that S ∗ is optimal.
"Z
#
#
"
Z Sn
Sn
e−λt h(Ψt )dt
e−λt h(Ψt )dt ≤ lim inf E0
E0 lim
n
n
0
0
(7.10)
= lim Vn (a, b, c) = V (a, b, c).
n
The first inequality follows from Fatou’s Lemma, which we can apply since
√
Z ∞
Z Sn
2
, a.s.
e−λt h(Ψt )dt ≥
e−λt h(Ψt )dt ≥ −
c
0
0
The first equality in (7.10) follows from Theorem 6.1. Now it can be seen from (7.10) that S ∗ is an
optimal stopping time. Taking the limit of (6.12) as n → ∞ and using Corollary 7.1 we conclude that
τ ∗ = S ∗.
Proof of Theorem 7.1 The proof of the optimality of τ ∗ follows directly from Lemma 7.3. We will show
that τ ∗ is the smallest optimal stopping time.
Given any F-stopping time τ < τ ∗ , let us define
τ̃ , τ + τ ∗ ◦ θτ
Then the stopping time τ̃ satisfies
"Z
#
Ea,b,c
0
τ̃
Ea,b,c
0
"Z
τ
Ea,b,c
0
"Z
τ
= Ea,b,c
0
Z
e−λs h(Ψs )ds =
0
=
< Ea,b,c
0
e−λs h(Ψs )ds +
0
Z
τ̃
e−λs h(Ψs )ds
τ
e
−λs
h(Ψs )ds + e
τ
Z0 τ
−λτ
Z
τ ∗ ◦θτ
e
#
−λs
h(Ψs+τ )ds
0
0
0
(7.11)
#
(7.12)
−λs
−λτ
e
h(Ψs )ds + e
V (Υτ )
e−λs h(Ψs )ds .
Here the third equality follows from the strong Markov property of the process Υ and the inequality
follows since V (Υτ ) < 0. Equation (7.12) shows that any optimal stopping time τ < τ ∗ cannot be
optimal.
8. Structure of the Continuation and Stopping Regions Regions. Let us recall (5.2) and
denote
Γn , {(a, b, c) ∈ B3+ : vn (a, b, c) = 0},
Cn , B3+ − Γn .
(8.1)
Quickest Detection of a Minimum of Two Poisson Disorder Times
17
We have shown in Theorem 7.1 that that Γ of (5.2) is the optimal stopping region for (3.3) and the
first hitting time τ ∗ of Υ to this set is optimal. On the other hand although Γn is an optimal stopping
region for (6.1), the description of the optimal stopping times Sn0 (see (6.12)) is more involved. These
optimal stopping times are not hitting times of the sets Γn . Sn0 prescribes to stop if Υ hits Γn before it
jumps. Otherwise if there is a jump before Υ reaches Γn , then Sn0 prescribes to stop when the process
hits Γn−1 before the next jump, and so on.
Theorem 6.1 shows that Vn of (6.1) and the functions vn introduced in Corollary 6.1 are equal. Therefore, their respective limits V and v are also equal. Recall that V n converges to V uniformly and the
convergence rate is exponential (see Lemma 6.1). Since (vn )n∈N is a decreasing sequence with limit v the
stopping regions in (8.1) are nested and satisfy Γ ⊂ · · · ⊂ Γn ⊂ Γn−1 ⊂ · · · Γ1 and Γ = ∩∞
n=1 Γn .
By Corollary 6.1 we know that each vn is concave and bounded, which also implies that the limit v is
concave and bounded. This in turn implies that the stopping regions Γn and Γ are convex and closed.
Since we show in Section 5 that the continuation region is bounded, it can readily be shown that the
stopping regions Γn and Γ are the epigraphs of some mappings γn and γ which are convex and strictly
decreasing and the numbers xn , inf{y ∈ R+ : γn (y) = 0} and x , inf{y ∈ R+ : γ(x) = 0} are finite.
9. Extensions.
9.1. Non-identical Sources. Consider two independent Poisson processes X 1 and X 2 with arrival
rates β1 and β2 respectively. At some random unobservable times θ1 and θ2 , with distributions
P(θi > t) = (1 − πi )e−λi t for t ≥ 0,
P(θi = 0) = πi
the arrival rates of the Poisson processes X 1 and X 2 change from βi to αi , respectively, i.e.,
Z t
Xti −
hi (s)ds, t ≥ 0, i = 1, 2,
(9.1)
(9.2)
0
are martingales, in which
hi (t) = [βi 1{s<θi } + αi 1{s≥θi } ],
t ≥ 0, i = 1, 2.
(9.3)
Here α1 , α2 , β1 and β2 are known positive constants. Then Equation the dynamics of Φ× defined in
(2.19) becomes
+
×
1
2
1
2
dΦ×
(9.4)
t = [λ2 Φt + λ1 Φt + (a1 + a2 )Φt ]dt + Φt ((α1 /β1 ) − 1)dXt + ((α2 /β2 ) − 1)dXt ,
in which ai = λi − αi + βi , i ∈ {1, 2}. Let us introduce
Z t
(a1 +a2 )t
x(t, φ0 ) = e
φ0 +
e(a1 +a2 )(t−u) (λ2 y(u, φ1 ) + λ1 z(u, φ2 ))du, in which
0
λ2
λ
λ2
λ1
a1 t
a2 t
, z(t, φ2 ) = − + e
φ1 +
φ2 +
.
y(t, φ1 ) = − + e
a1
a
a2
a2
(9.5)
1
2
Then Φ×
t , Φt and Φt , t ≥ 0 can be written as
2
1
1
+
×
Φ×
t = x(t − σn , Φσn ), Φt = y(t − σn , Φσn ), Φt = z(t − σn , Φσn ) σn ≤ t < σn+1 , n ∈ N,
(9.6)
and
„
«
α1
α2
=
1 1
+
1 2
Φ×
1
2
σn+1− ,
β1 {Xσn+1 6=Xσn+1− } β2 {Xσn+1 6=Xσn+1− }
(9.7)
α2
α1
1
1
2
1{Xσ1
Φ
1
Φ
,
Φ
=
.
Φ1σn+1 =
1
2
2
σ
σ
σ
6=Xσ
}
n+1
(n+1)−
n+1
n+1−
β1
β2 {Xσn+1 6=Xσn+1− } (n+1)−
` × 1 2´
Choosing Υt = Φt , Φt , Φt , t ≥ 0 as the Markovian statistic to work with, we can extend our analysis to deal
with non-identical sources.
Φ×
σn+1
18
E. BAYRAKTAR AND H. V. POOR
9.2. When there are more than two sources. We have solved a two-source quickest detection
problem in which the aim is to detect the minimum of two disorder times. Our approach can easily be generalized
to problems including several dimensions. To clarify how this generalization works, let us show what the sufficient
statistics are when there are three independent sources. Assume that the observations come from the independent
sources X 1 , X 2 and X 3 . Let Φt be the odds ratio defined in (2.9). Then
Φt = Φ1t + Φ2t + Φ3t + Φ1t Φ2t + Φ1t Φ3t + Φ2t Φ3t + Φ1t Φ2t Φ3t ,
(i,j)
(9.8)
(x)
in which Φi , i ∈ {1, 2, 3} is defined as in (2.14). Let us denote Φt
, Φit Φjt , i, j ∈ {1, 2, 3} and Φt , Φ1t Φ2t Φ3t ,
t ≥ 0. The dynamics of these processes can be written as
«
„
α
(i,j)
(i,j)
(i,j)
− 1 Φt d(Xti + Xtj ),
= [λ(Φit + Φjt ) + 2(λ − α + β)Φt ]dt +
dΦt
β
(9.9)
„
«
i
”
h “
α
(x)
(x)
(1,2)
(1,3)
(2,3)
(x)
dt +
+ 3(λ − α + β)Φt
dΦt = λ Φt
+ Φt
+ Φt
− 1 Φt d(Xt1 + Xt2 + Xt3 ).
β
(1,3)
(1,2)
(2,3)
We can see from (2.13) and (9.9) that Υ , (Φ1 , Φ2 , Φ3 , Φt , Φt , Φt , Φ(x) ) is a 7 dimensional Markovian
statistic whose natural filtration is equal to the filtration generated by X 1 , X 2 and X 3 . From this one can see
that the results of Sections 6 and 7 can be extended to the three dimensional case since these results rely only on
the fact that the sufficient statistic Υ is a strong Markov process. The boundedness of the continuation region can
also be shown as in Section 5 since these results can be derived from the sample path properties of the sufficient
statistic.
As a result, our results are applicable for decision making with large-scale distributed networks of information
sources. In the future, using the techniques developed here, we would like to solve a multi-source detection
problem where the observations come from correlated sources. We also would like to extend our results and
develop change detection algorithms that can be applied effectively to multiple source data that involves both
continuous and discrete event phenomena.
9.3. When the jump size of the observations are random. Consider two independent compound
Poisson processes X i = {Xti : t ≥ 0}, i ∈ {1, 2}, where
i
Xti
=
X0i
+
Nt
X
Yji ,
(9.10)
j=1
in which N i , i ∈ {1, 2} are two independent Poisson processes whose common rate β > 0 changes to α at some
random unobservable times θi , i ∈ {1, 2}, respectively. The random variables Yji ∈ Rd , i ∈ {1, 2}, which are also
termed as ‘marks’, are independent and identically distributed with a common distribution, ν, which is called
as the ‘mark distribution’. At the change time θi the mark distribution of the process X i changes from ν to µ.
We will assume that µ is absolutely continuous with respect to ν and denote the Radon-Nikodym derivative by
r(y) , dµ
(y), y ∈ Rd . In this case Lit in (2.4) becomes
dν
i
Lit = e−(α−β)t
Nt „
Y
α
k=1
β
r(Yki )
«
(9.11)
The likelihood ratio process Li is the unique solution of the stochastic differential equation (see e.g. [8])
«
«
„
„
Z
α
r(y) − 1 p(dtdy) , Li0 = 1,
dLit = Lit −(α − β)dt +
β
y∈Rd
(9.12)
where p is a random measure that is defined as
pi ((0, t] × A) ,
∞
X
1{σ i ≤t} 1{Y i ∈A} ,
k
k
t ≥ 0,
(9.13)
k=1
and for any A that is a Borel measurable subset of Rd . Here σki is the kth jump time of the process X i . Now
using the change of variable formula for semi-martingales (see e.g. [12]), we can write
„
«
Z
α
πi
dΦit = (λ + (λ − α + β)Φit )dt + Φit−
,
(9.14)
r(y) − 1 pi (dtdy), Φi0 =
β
1
−
πi
d
y∈R
19
Quickest Detection of a Minimum of Two Poisson Disorder Times
r(Yn )Φiσn − at the nth jump time of the process X i . Using a change
for t ≥ 0, and i ∈ {1, 2}. Note that Φiσn = α
β
of variable formula for semi-martingales, the dynamics of Φ× and Φ+ in (2.19) can be written as
«
„
Z
α
+
×
×
r(y)
−
1
(p1 + p2 )(dtdy),
dΦ×
=
[λΦ
+
aΦ
]dt
+
Φ
t
t
t
t−
β
y∈Rd
«
„
«
(9.15)
„
Z
Z
α
α
+
1
r(y) − 1 p1 (dtdy) + Φ2t−
r(y) − 1 p2 (dtdy)
dΦ+
t = [2λ + aΦt ]dt + Φt−
β
β
y∈Rd
y∈Rd
+
with initial conditions Φ×
0 = π1 π2 /[(1 − π1 )(1 − π2 )], and Φ0 = π1 /(1 − π1 ) + π2 /(1 − π2 ).
The bounds on the continuation region constructed for the simple Poisson disorder problem in Section 5 can
also be shown to bound the continuation region of the compound Poisson disorder problem. On the other hand
the results in Sections 6 and 7 can be shown to hold. The only change will be the form of the operator J in (6.7).
But this new operator can be shown to share the same properties as its counterpart for the un-marked case.
Acknowledgment. This work was supported in part by the U.S. Army Pantheon Project and National
Science Foundation under Grant DMS-0604491. We are grateful to the two referees for their detailed comments
that helped us improve the manuscript. We also would like to thank Semih Sezer for insightful comments.
10. Appendix. Proof of Theorem 6.1. We will prove only that Vn = vn and Snε is an ε-optimal stopping
time of (6.1).
The proof will be carried out in three steps.
(i) First we will show that Vn ≥ vn . To establish this fact, it is enough to show that for any stopping time τ ∈ S
–
»Z τ ∧σn
−λt
e
h(Ψ
)dt
≥ vn (a, b, c).
(10.1)
Ea,b,c
t
0
0
In order to prove (10.1) we will show that
»Z τ ∧σn−k+1
–
–
»Z τ ∧σn
e−λt h(Ψt )dt + 1{τ ≥σn−k+1 } e−λσn−k+1 vk−1 (Υσn−k+1 ) ,
e−λt h(Ψt )dt ≥ E0
E0
(10.2)
0
0
for k ∈ {1, 2, ..., n + 1}. Note that (10.1) follows from (10.2) if we set k = n + 1. In what follows we will to show
(10.2) by induction.
When k = 1, (10.2) is satisfied since v0 = 0. Assume that (10.2) holds for 1 ≤ k ≤ n + 1. Let us denote the
right-hand-side of (10.2) by ρk−1 . We can write ρk−1 = ρ1k−1 + ρ2k−1 , where
»Z τ ∧σn−k
–
e−λt h(Ψt )dt
and
ρ1k−1 , E0
0
"
!#
(10.3)
Z τ ∧σn−k+1
.
e−λt h(Ψt )dt + 1{τ ≥σn−k+1 } e−λσn−k+1 vk−1 (Υσn−k+1 )
ρ2k−1 , E0 1{τ ≥σn−k }
σn−k
Now by Lemma 6.3, there exists an Fσn−k -measurable random variable ξn−k such that
τ ∧ σn−k+1 = (σn−k + ξn−k ) ∧ σn−k+1
almost surely on {τ ≥ σn−k }.
(10.4)
Equation (10.4) together with the strong Markov property of Υ (with respect to the filtration F) implies that
i
h
ρ2k−1 = E0 1{τ ≥σn−k } e−λσn−k fk−1 (ξn−k , Υσn−k ) ,
(10.5)
in which
fk−1 (r, (a, b, c)) , Ea,b,c
0
»Z
r∧σ1
0
e−λt h(Ψt )dt + 1{r≥σ1 } e−λσ1 vk−1 (Υσ1 )
= Jvk−1 (r, (a, b, c)) ≥ J0 vk−1 (a, b, c) = vk (a, b, c),
–
(10.6)
in which the second equality and the first inequality follow from (6.5) and the last equality follows from (6.10).
Therefore
i
h
ρ2k−1 ≥ E0 1{τ ≥σn−k } e−λσn−k vk (Υσn−k ) .
(10.7)
20
E. BAYRAKTAR AND H. V. POOR
Now using (10.2), (10.3) and (10.7) we obtain that (10.2) holds when k is replaced by k + 1. At this point we
have proved by induction that (10.2) holds for k = 1, 2, ..., n + 1.
(ii) The converse of (i), Vn ≤ vn , follows from (6.13), since Snε ≤ σn by construction (see (6.12)).
(iii) What is left to prove is (6.13). If n = 1, then the left-hand-side of (6.13) becomes
#
"Z ε
r0 (a,b,c)∧σ1
Ea,b,c
0
e−λt h(Ψt )dt = Jv0 (r0ε (a, b, c), a, b, c)
(10.8)
0
≤ J0 v0 (a, b, c) + ε = v1 (a, b, c) + ε.
ε/2
ε
Now, suppose that (6.13) holds for all ε > 0 for some n. Using the fact that Sn+1
∧ σ1 = rn ∧ σ1 almost surely
and the strong Markov property of Υ, we can write
#
"Z ε
#
"Z ε
ε
Z Sn+1
Sn+1
Sn+1 ∧σ1
−λt
−λt
−λt
ε
E0
e
h(Ψt )dt = E0
e
h(Ψt )dt
e
h(Ψt )dt + 1{Sn+1 ≥σ1 }
0
= E0
0
"Z
ε/2
rn
0
(a,b,c)∧σ1
σ1
#
(10.9)
i
h
e−λt h(Ψt )dt + E0 1{rε/2 (a,b,c)≥σ } e−λσ1 gn (Υσ1 ) ,
in which
gn (a, b, c) ,
Ea,b,c
0
"Z
n
1
ε/2
Sn
e
−λt
#
h(Ψt )dt ≤ vn (a, b, c) + ε/2.
0
(10.10)
The inequality in (10.10) follows from the induction hypothesis. Using (10.10) we can write (10.9) as
#
"Z ε
» Z rnε/2 (a,b,c)∧σ1
–
Sn+1
a,b,c
−λt
−λσ1
−λt
a,b,c
E0
vn (Υσ1 ) + ε/2
e
h(Ψt )dt + 1{rε/2 (a,b,c)≥σ } e
e
h(Ψt )dt ≤ E0
n
0
0
=
Jvn (rnε/2 (a, b, c), a, b, c)
1
+ ε/2 ≤ vn+1 (a, b, c) + ε.
(10.11)
This proves (6.13) when n is replaced by n + 1.
REFERENCES
[1] M. Basseville and I. V. Nikiforov, Detection of Abrupt Changes: Theory and Application, Prentice Hall, Englewood
Cliffs, NJ., 1993.
[2] E. Bayraktar and S. Dayanik, Poisson disorder problem with exponential penalty for delay, Mathematics of Operations Research, 31 (2) (2006), pp. 217–233.
[3] E. Bayraktar, S. Dayanik, and I. Karatzas, The standard Poisson disorder problem revisited, Stochastic Processes
and Their Applications, 115 (9) (2005), pp. 1437–1450.
, Adaptive poisson disorder problem, Annals of Applied Probability, 16 (3) (2006), pp. 1190–1261.
[4]
[5] P. Brémaud, Point Processes and Queues, Springer-Verlag, Berlin, 1981.
[6] M. H. A. Davis, Markov Models and Optimization, Chapman & Hall, London, 1993.
[7] P. Dube and R. Mazumdar, A framework for quickest detection of traffic anomalies in networks, preprint, Purdue
University, (2001).
[8] J. Jacod and A. Shiryaev, Limit Theorem for Stochastic Processes, Springer-Verlag, Berlin, 1987.
[9] R. S. Lipster and A. N. Shiryaev, Statistics of Random Processes, Springer-Verlag, Berlin, 2001.
[10] S. N. Neftci, Optimal prediction of cyclical downturns, Journal of Economic Dynamics and Control, 4 (1982), pp. 225–
241.
[11] H. V. Poor, Quickest detection with exponential penalty for delay, Annals of Statistics, 26 (1998), pp. 2179–2205.
[12] P. Protter, Stochastic Integration and Differential Equations, Springer-Verlag, Berlin, 1990.
[13] D. Revuz and M. Yor, Continuous Martingales and Brownian Motion, Springer-Verlag, Berlin, 1999.
[14] C. Sonesson and D. Bock, A review and discussion of prospective statistical surveillance in public health, Journal
of Royal Statistical Society, A., 166 (2003), pp. 5–21.
| 5cs.CE
|
arXiv:1606.08036v2 [math.GR] 29 Dec 2017
THE BOUNDED AND PRECISE WORD PROBLEMS FOR
PRESENTATIONS OF GROUPS
S. V. IVANOV
Abstract. We introduce and study the bounded word problem and the precise word problem for groups given by means of generators and defining relations. For example, for every finitely presented group, the bounded word
problem is in NP, i.e., it can be solved in nondeterministic polynomial time,
and the precise word problem is in PSPACE, i.e., it can be solved in polynomial space. The main technical result of the paper states that, for certain
finite presentations of groups, which include the Baumslag-Solitar one-relator
groups and free products of cyclic groups, the bounded word problem and
the precise word problem can be solved in polylogarithmic space. As consequences of developed techniques that can be described as calculus of brackets,
we obtain polylogarithmic space bounds for the computational complexity of
the diagram problem for free groups, for the width problem for elements of
free groups, and for computation of the area defined by polygonal singular
closed curves in the plane. We also obtain polynomial time bounds for these
problems.
Contents
1.
2.
3.
4.
5.
6.
7.
8.
Introduction
Preliminaries
Proof of Proposition 1.1
Calculus of Brackets for Group Presentation (1.2)
Proofs of Theorem 1.2 and Corollary 1.3
Calculus of Brackets for Group Presentation (1.4)
Proof of Theorem 1.4
Minimizing Diagrams over (1.2) and Proofs of Theorem 1.5 and
Corollary 1.6
9. Construction of Minimal Diagrams over (1.4) and Proof of Theorem 1.7
10. Polygonal Curves in the Plane and Proofs of Theorems 1.8, 1.9 and
Corollary 1.10
References
2
8
12
13
34
38
58
62
72
81
93
2010 Mathematics Subject Classification. Primary 20F05, 20F06, 20F10, 68Q25, 68U05; Secondary 52B05, 20F65, 68W30.
Key words and phrases. Presentations of groups, diagrams, word problems, polylogarithmic
space, polygonal curves.
The work on this article was supported in smaller part by NSF grant DMS 09-01782.
1
2
S. V. IVANOV
1. Introduction
Suppose that a finitely generated group G is defined by a presentation by means
of generators and defining relations
G = h A k R = 1, R ∈ R i ,
(1.1)
where A = {a1 , . . . , am } is a finite alphabet and R is a set of defining relators
which are nonempty cyclically reduced words over the alphabet A±1 := A ∪ A−1 .
Let F(A) denote the free group over A, let |W | be the length of a word W over the
alphabet A±1 , and let N(R) denote the normal closure of R in F(A). The notation
(1.1) implies that G is the quotient group F(A)/N(R). We will say that a word W
over A±1 is equal to 1 in the group G, given by (1.1), if W ∈ N(R) in which case we
G
also write W = 1. Recall that the presentation (1.1) is called finite if both the sets
A and R are finite, in which case G is called finitely presented. The presentation
(1.1) is called decidable if there is an algorithm that decides whether a given word
over A±1 belongs to R.
The classical word problem for a finite group presentation (1.1), put forward
by Dehn [10] in 1911, asks whether, for a given word W over A±1 , it is true that
G
W = 1. The word problem is said to be solvable (or decidable) for a decidable
presentation (1.1) if there exists an algorithm which, given a word W over A±1 ,
G
decides whether or not W = 1.
Analogously to Anisimov [1], one might consider the set of all words W over
G
G
A±1 , not necessarily reduced, such that W = 1 as a language L(G) := {W | W =
1} over A±1 and inquire about a computational complexity class K which would
contain this language. If L(G) is in K, we say that the word problem for G is
in K. For example, it is well known that the word problem is in P, i.e., solvable
in deterministic polynomial time, for surface groups, for finitely presented groups
with small cancellation condition C ′ (λ), where λ ≤ 16 , for word hyperbolic groups,
for finitely presented groups given by Dehn presentations etc., see [13], [18]. On
the other hand, according to results of Novikov [25], [26] and Boone [6], [7], based
on earlier semigroup constructions of Turing [39] and Post [34], see also Markov’s
papers [22], [23], there exists a finitely presented group G for which the word problem
G
is unsolvable, i.e., there is no algorithm that decides whether W = 1. The proof
of this remarkable Novikov–Boone theorem was later significantly simplified by
Borisov [8], see also [37].
In this paper, we introduce and study two related problems which we call the
bounded word problem and the precise word problem for a decidable group presentation (1.1).
The bounded word problem for a decidable presentation (1.1) inquires whether,
for given a word W over A±1 and an integer n ≥ 0 written in unary, denoted
1n , one can represent the word W as a product in F(A) of at most n conjugates
of some words in R±1 := R ∪ R−1 , i.e., whether there are R1 , . . . , Rk ∈ R±1 and
S1 , . . . , Sk ∈ F(A) such that W = S1 R1 S1−1 . . . Sk Rk Sk−1 in F(A) and k ≤ n.
Equivalently, for given an input (W, 1n ), the bounded word problem asks whether
there exists a disk diagram over (1.1), also called a van Kampen diagram, whose
boundary label is W and whose number of faces is at most n.
As above for the word problem, we say that the bounded word problem for a decidable presentation (1.1) is solvable if there exists an algorithm that, given an input
BOUNDED AND PRECISE WORD PROBLEMS FOR GROUP PRESENTATIONS
3
(W, 1n ), decides whether W is a product in F(A) of at most n conjugates of words
in R±1 . Analogously, if the language of those pairs (W, 1n ), for which the bounded
word problem has a positive answer, belongs to a computational complexity class
K, then we say that the bounded word problem for G is in K.
The precise word problem for a decidable presentation (1.1) asks whether, for
given a word W over A±1 and a nonnegative integer 1n , one can represent the word
W as a product in F(A) of n conjugates of words in R±1 and n is minimal with this
property. Equivalently, for given an input (W, 1n ), the precise word problem asks
whether there exists a disk diagram over (1.1) whose boundary label is W , whose
number of faces is n, and there are no such diagrams with fewer number of faces.
The definitions for the precise word problem for (1.1) being solvable and being in
a complexity class K are similar to the corresponding definitions for the (bounded)
word problem.
In the following proposition we list basic comparative properties of the standard
word problem and its bounded and precise versions.
Proposition 1.1. (a) There exists a decidable group presentation (1.1) for which
the word problem is solvable while the bounded and precise word problems are not
solvable.
(b) If the bounded word problem is solvable for (1.1), then the precise word
problem is also solvable.
(c) For every finite group presentation (1.1), the bounded word problem is in
NP, i.e., it can be solved in nondeterministic polynomial time, and the precise word
problem is in PSPACE, i.e., it can be solved in polynomial space.
(d) There exists a finite group presentation (1.1) for which the bounded and
precise word problems are solvable while the word problem is not solvable.
(e) There exists a finitely presented group (1.1) for which the bounded word problem is NP-complete and the precise word problem is NP-hard.
Note that a number of interesting results on solvability of the word problem,
solvability of the bounded word problem and computability of the Dehn function
for decidable group presentations can be found in a preprint of Cummins [9] based
on his PhD thesis written under the author’s supervision.
It is of interest to look at the bounded and precise word problems for finitely
presented groups for which the word problem could be solved very easily. Curiously,
even for very simple presentations such as h a, b k a = 1 i and h a, b k ab = ba i, for
which the word problem is obviously in L, i.e., solvable in deterministic logarithmic
space, it does not seem to be possible to solve the bounded and precise word
problem in logarithmic space. In this article, we will show that the bounded and
precise word problems for these “simple” presentations and their generalizations
can be solved in polynomial time. With much more effort, we will also prove that
the bounded and precise word problems for such presentations can be solved in
polylogarithmic space. Similarly to [2], we adopt NC-style notation and denote
Lα := DSPACE((log s)α ), i.e., Lα is the class of decision problems that can be
solved in deterministic space O((log s)α ), where s is the size of input.
Theorem 1.2. Let the group G2 be defined by a presentation of the form
G2 := h a1 , . . . , am k aki i = 1, ki ∈ Ei , i = 1, . . . , m i,
(1.2)
where for every i, one of the following holds: Ei = {0} or, for some integer ni > 0,
Ei = {ni } or Ei = ni N = {ni , 2ni , 3ni , . . . }. Then both the bounded and precise
4
S. V. IVANOV
word problems for (1.2) are in L3 and in P. Specifically, the problems can be solved
in deterministic space O((log |W |)3 ) or in deterministic time O(|W |4 log |W |).
It is worth mentioning that, to prove L3 part of Theorem 1.2, we will not devise
a concrete algorithm that solves the bounded and precise word problems for presentation (1.2) in deterministic space O((log |W |)3 ). Instead, we will develop a certain
nondeterministic procedure that solves the bounded word problem for presentation
(1.2) nondeterministically in space O((log |W |)2 ) and time O(|W |) and then use
Savitch’s theorem [38] on conversion of nondeterministic computations in space S
and time T into deterministic computations in space O(S log T ).
The proof of P part of Theorem 1.2 is much easier. Here our arguments are
analogous to folklore arguments [35], [36] that solve the precise word problem for
presentation h a, b k a = 1, b = 1 i in polynomial time and that utilize the method of
dynamic programming. Interestingly, these folklore arguments are strikingly similar
to the arguments that have been used in computational biology to efficiently solve
the problem of planar folding of long chains such as RNA and DNA biomolecules,
see [27], [28], [29], [40].
The techniques to prove L3 part of Theorem 1.2 and their generalizations, that
occupy a significant part of this article and that could be described as calculus of
bracket systems, have applications to other problems. For example, Grigorchuk
and Kurchanov [11] defined the width of an element W of the free group F(A) over
A as the minimal number h = h(W ) so that
W = S1 akj11 S1−1 . . . Sh akjhh Sh−1
(1.3)
in F(A), where aj1 , . . . , ajh ∈ A, S1 , . . . , Sh ∈ F(A) and k1 , . . . , kh are some integers.
Alternatively, the width h(W ) of W can be defined as an integer such that the
precise word problem holds for the pair (W, h(W )) for the presentation (1.2) in
which Ei = N for every i. Grigorchuk and Kurchanov [11] found an algorithm that
computes the width h(W ) for given W ∈ F(A) and inquired whether computation
of the width h(W ) can be done in deterministic polynomial time. Ol’shanskii [31]
gave a different geometric proof to this result of Grigorchuk and Kurchanov and
suggested some generalizations.
Majumdar, Robbins, and Zyskin [20], [21] introduced and investigated the spelling
length h1 (W ) of a word W ∈ F(A) defined by a similar to (1.3) formula in which
kj = ±1 for every j. Alternatively, the spelling length h1 (W ) is an integer such
that the precise word problem holds for the pair (W, h1 (W )) for the presentation
(1.2) in which Ei = {1} for every i.
As a corollary of Theorem 1.2, we obtain a positive solution to the problem of
Grigorchuk and Kurchanov and also compute both the width and the spelling length
of W in cubic logarithmic space. We remark that Riley [36] gave an independent
solution to the Grigorchuk–Kurchanov problem.
Corollary 1.3. Let W be a word over A±1 and n ≥ 0 be an integer. Then the
decision problems that inquire whether the width h(W ) or the spelling length h1 (W )
of W is equal to n belong to L3 and P. Specifically, the problems can be solved in
deterministic space O((log |W |)3 ) or in deterministic time O(|W |4 log |W |).
Making many technical modifications but keeping the general strategy of arguments unchanged, we will obtain similar to Theorem 1.2 results for Baumslag–
Solitar one-relator groups.
BOUNDED AND PRECISE WORD PROBLEMS FOR GROUP PRESENTATIONS
5
Theorem 1.4. Let the group G3 be defined by a presentation of the form
n2
G3 := h a1 , . . . , am k a2 an1 1 a−1
2 = a1 i,
(1.4)
where n1 , n2 are some nonzero integers. Then both the bounded and precise word
problems for (1.4) are in L3 and in P. Specifically, the problems can be solved
in deterministic space O((max(log |W |, log n)(log |W |)2 ) or in deterministic time
O(|W |4 ).
As another application of our techniques, we will obtain a solution in polylogarithmic space for the (minimal) diagram problem for presentation (1.2) which
includes the case of the free group F(A) = hA k ∅i over A without relations. Recall that the diagram problem for a decidable presentation (1.1) is a search problem
G
that, given a word W over A±1 with W = 1, asks to algorithmically construct a disk
diagram ∆ over (1.1) whose boundary ∂∆ is labeled by W , denoted ϕ(∂∆) ≡ W ,
for the definitions see Sect. 2. Analogously, the minimal diagram problem for a
decidable presentation (1.1) is a search problem that, given a word W over A±1
G
with W = 1, asks to algorithmically construct a disk diagram ∆ over (1.1) such
that ϕ(∂∆) ≡ W and ∆ contains a minimal number of faces.
Recall that, according to Lipton and Zalcstein [16], the word problem for the free
group F(A), given by presentation (1.1) with R = ∅, is in L. However, construction
of an actual diagram ∆ over F(A) for a word W over A±1 such that ϕ(∂∆) ≡ W ,
is a different matter and it is not known whether this construction could be done
in polylogarithmic space (note that it is easy to construct such a diagram, which
is a tree, in polynomial time). In fact, many results of this article grew out of
the attempts to solve the diagram problem for free groups with no relations in
subpolynomial space.
Theorem 1.5. Both the diagram problem and the minimal diagram problem for
group presentation (1.2) can be solved in deterministic space O((log |W |)3 ) or in
deterministic time O(|W |4 log |W |).
G
Furthermore, let W be a word such that W =2 1 and let
τ (∆) = (τ1 (∆), . . . , τsτ (∆))
be a tuple of integers, where the absolute value |τi (∆)| of each τi (∆) represents
the number of certain vertices or faces in a disk diagram ∆ over (1.2) such that
ϕ(∂∆) ≡ W . Then, in deterministic space O((log |W |)3 ), one can algorithmically
construct such a minimal diagram ∆ which is also smallest relative to the tuple
τ (∆) (the tuples are ordered lexicographically).
We point out that the case of the free group F(A) = hA k ∅i with no relations
is covered by Theorem 1.5 and, since there are no relations, every diagram ∆ over
F(A) = hA k ∅i is minimal. Hence, for the free group F(A), Theorem 1.5 implies
the following.
Corollary 1.6. There is a deterministic algorithm that, for given word W over the
F(A)
alphabet A±1 such that W = 1, where F(A) = hA k ∅i is the free group over A,
constructs a pattern of cancellations of letters in W that result in the empty word
and the algorithm operates in space O((log |W |)3 ).
6
S. V. IVANOV
Furthermore, let ∆ be a disk diagram over F(A) that corresponds to a pattern of
cancellations of letters in W , i.e., ϕ(∂∆) ≡ W , and let
τ (∆) = (τ1 (∆), . . . , τsτ (∆))
be a tuple of integers, where the absolute value |τi (∆)| of each τi (∆) represents
the number of vertices in ∆ of certain degree. Then, also in deterministic space
O((log |W |)3 ), one can algorithmically construct such a diagram ∆ which is smallest
relative to the tuple τ (∆).
Here is the analogue of Theorem 1.5 for presentations (1.4) of one-relator Baumslag–
Solitar groups.
Theorem 1.7. Suppose that W is a word over the alphabet A±1 such that the
bounded word problem for presentation (1.4) holds for the pair (W, n). Then a
minimal diagram ∆ over (1.4) such that ϕ(∂∆) ≡ W can be algorithmically constructed in deterministic space O(max(log |W |, log n)(log |W |)2 ) or in deterministic
time O(|W |4 ).
In addition, if |n1 | = |n2 | in (1.4), then the minimal diagram problem for presentation (1.4) can be solved in deterministic space O((log |W |)3 ) or in deterministic
time O(|W |3 log |W |).
As further applications of the techniques of the proof of Theorems 1.4, 1.7, we
obtain computational results on discrete homotopy of polygonal closed curves in
the plane.
Let T denote a tessellation of the plane R2 into unit squares whose vertices are
points with integer coordinates. Let c be a finite closed path in T so that edges of
c are edges of T. Consider the following two types of elementary operations over c.
If e is an oriented edge of c, e−1 is the edge with an opposite to e orientation, and
ee−1 is a subpath of c so that c = c1 ee−1 c2 , where c1 , c2 are subpaths of c, then
the operation c → c1 c2 over c is called an elementary homotopy of type 1. Suppose
that c = c1 uc2 , where c1 , u, c2 are subpaths of c, and a boundary path ∂s of a unit
square s of T is ∂s = uv, where u, v are subpaths of ∂s and either of u, v could be
of zero length, i.e., either of u, v could be a single vertex of ∂s. Then the operation
c → c1 v −1 c2 over c is called an elementary homotopy of type 2.
Theorem 1.8. Let c be a finite closed path in a tessellation T of the plane R2
into unit squares so that edges of c are edges of T. Then a minimal number m2 (c)
such that there is a finite sequence of elementary homotopies of type 1–2, which
turns c into a single point and which contains m2 (c) elementary homotopies of type
2, can be computed in deterministic space O((log |c|)3 ) or in deterministic time
O(|c|3 log |c|), where |c| is the length of c.
Furthermore, such a sequence of elementary homotopies of type 1–2, which turns
c into a single point and which contains m2 (c) elementary homotopies of type 2,
can also be computed in deterministic space O((log |c|)3 ) or in deterministic time
O(|c|3 log |c|).
We remark that this number m2 (c) defined in Theorem 1.8 can be regarded as
the area “bounded” by a closed path c in T. Clearly, if c is simple, i.e., c has no
self-intersections, then m2 (c) is the area of the compact region bounded by c. If c is
simple then the area bounded by c can be computed in logarithmic space O(log |c|)
as follows from the “shoelace” formula for the area of a simple polygon.
BOUNDED AND PRECISE WORD PROBLEMS FOR GROUP PRESENTATIONS
7
More generally, assume that c is a continuous closed curve in R2 , i.e., c is the
image of a continuous map S1 → R2 , where S1 is a circle. Consider a homotopy
H : S1 × [0, 1] → R2 that turns the curve c = H(S1 × {0}) into a point H(S1 × {1})
so that for every t, 0 ≤ t < 1, H(S1 × {t}) is a closed curve. Let A(H) denote
the area swept by the curves H(S1 × {t}), 0 ≤ t < 1, and let A(c) denote the
infimum inf H A(H) over all such homotopies H. As above, we remark that this
number A(c) can be regarded as the area defined (or “bounded”) by c. Note that
this number A(c) is different from the signed area of c defined by applying the
“shoelace” formula to singular polygons.
Other applications that we discuss here involve polygonal (or piecewise linear)
closed curves in the plane and computation and approximation of the area defined
by these curves in polylogarithmic space or in polynomial time.
We say that c is a polygonal closed curve in the plane R2 with given tessellation T
into unit squares if c consists of finitely many line segments c1 , . . . , ck , k > 0, whose
endpoints are vertices of T, c = c1 . . . ck , and c is closed. If ci ⊂ T then the T-length
|ci |T of ci is the number of edges of T in ci . If ci 6⊂ T then the T-length |ci |T of ci is
the number of connected components in ci \ T. We assume that |ci |T > 0 for every
P
i and set |c|T := ki=1 |ci |T .
Theorem 1.9. Suppose that n ≥ 1 is a fixed integer and c is a polygonal closed
curve in the plane R2 with given tessellation T into unit squares. Then, in deterministic space O((log |c|T )3 ) or in deterministic time O(|c|n+3
log |c|T ), one can
T
1
compute a rational number rn such that |A(c) − rn | < |c|n .
T
In particular, if the area A(c) defined by c is known to be an integer multiple of
1
n
L , where L > 0 is a given integer and L < |c|T /2, then A(c) can be computed in
deterministic space O((log |c|T )3 ) or in deterministic time O(|c|n+3
log |c|T ).
T
Corollary 1.10. Let K ≥ 1 be a fixed integer and let c be a polygonal closed curve
in the plane R2 with given tessellation T into unit squares such that c has one of
the following two properties (a)–(b).
(a) If ci , cj are two nonparallel line segments of c then their intersection point,
1
if it exists, has coordinates that are integer multiples of K
.
(b) If ci is a line segment of c and ai,x , ai,y are coprime integers such that the
line given by an equation ai,x x + ai,y y = bi , where bi is an integer, contains ci , then
max(|ai,x |, |ai,y |) ≤ K.
Then the area A(c) defined by c can be computed in deterministic space O((log |c|T )3 )
log |c|T ), where n depends on K.
or in deterministic time O(|c|n+3
T
In particular, if T∗ is a tessellation of the plane R2 into equilateral triangles of
unit area, or into regular hexagons of unit area, and q is a finite closed path in T∗
whose edges are edges of T∗ , then the area A(q) defined by q can be computed in
deterministic space O((log |q|)3 ) or in deterministic time O(|q|5 log |q|).
It is tempting to try to lift the restrictions of Corollary 1.10 to be able to compute,
in polylogarithmic space, the area A(c) defined by an arbitrary polygonal closed
curve c in the plane equipped with a tessellation T into unit squares. However, in
the general situation, this idea would not work because the rational number A(c)
might have an exponentially large denominator, hence, A(c) could take polynomial
space just to store (let alone the computations), see an example in the end of
Sect. 10.
8
S. V. IVANOV
We remark in passing that there are decision problems in NP that are not known
to be NP-complete or in P, called NP-intermediate problems, that are solvable
in polylogarithmic space. For example, a restricted version of the NP-complete
clique problem asks whether a graph on n vertices contains a clique with at most
[log n] vertices, where [k] is the integer part of k, and this restriction is obviously
a problem solvable in nondeterministic space O([log n]2 ). More natural examples
of such NP-intermediate problems would be a decision version of the problem on
finding a minimum dominating set in a tournament, [24], [33], and the problem on
isomorphism of two finite groups given by their multiplication tables, [4], [17].
2. Preliminaries
0
If U, V are words over an alphabet A±1 := A ∪ A−1 , then U = V denotes the
equality of U, V as elements of the free group F(A) whose set of free generators is A.
The equality of natural images of words U, V in the group G, given by presentation
G
(1.1), is denoted U = V .
The letter-by-letter equality of words U, V is denoted U ≡ V . If U ≡ aεi11 . . . aεiℓℓ ,
where ai1 , . . . , aiℓ ∈ A and ε1 , . . . , εℓ ∈ {±1}, then the length of U is |U | = ℓ. A
nonempty word U over A±1 is called reduced if U contains no subwords of the form
aa−1 , a−1 a, where a ∈ A.
Let ∆ be a 2-complex and let ∆(i) denote the set of nonoriented i-cells of ∆,
~
~
i = 0, 1, 2. We also consider the set ∆(1)
of oriented 1-cells of ∆. If e ∈ ∆(1)
−1
−1
then e denotes e with the opposite orientation, note that e 6= e . For every
~
e ∈ ∆(1),
let e− , e+ denote the initial, terminal, resp., vertices of e. In particular,
−1
(e )− = e+ and (e−1 )+ = e− . The closures of i-cells of ∆ are called vertices,
edges, faces when i = 0, 1, 2, resp.
A path p = e1 . . . eℓ in ∆ is a sequence of oriented edges e1 , . . . , eℓ of ∆ such that
(ei )+ = (ei+1 )− , i = 1, . . . , ℓ − 1. The length of a path p = e1 . . . eℓ is |p| = ℓ. The
initial vertex of p is p− := (e1 )− and the terminal vertex of p is p+ := (eℓ )+ . We
allow the possibility that |p| = 0 and p = p− .
A path p is called reduced if |p| > 0 and p contains no subpath of the form ee−1 ,
where e is an edge. A path p is called closed if p− = p+ .
A cyclic path is a closed path with no distinguished initial vertex. A path
p = e1 . . . eℓ is called simple if the vertices (e1 )− , . . . , (eℓ )− , (eℓ )+ are all distinct. A
closed path p = e1 . . . eℓ is simple if the vertices (e1 )− , . . . , (eℓ )− are all distinct.
A disk diagram, also called a van Kampen diagram, ∆ over a presentation (1.1)
is a planar connected and simply connected finite 2-complex which is equipped with
a labeling function
~
ϕ : ∆(1)
→ A ∪ A−1 = A±1
~
such that, for every e ∈ ∆(1),
ϕ(e−1 ) = ϕ(e)−1 and, for every face Π of ∆, if
~
∂Π = e1 . . . eℓ is a boundary path of Π, where e1 , . . . , eℓ ∈ ∆(1),
then
ϕ(∂Π) := ϕ(e1 ) . . . ϕ(eℓ )
is a cyclic permutation of one of the words in R±1 = R ∪ R−1 .
A disk diagram ∆ over presentation (1.1) is always considered with an embedding
∆ → R2 into the plane R2 . This embedding makes it possible to define positive
(=counterclockwise) and negative (=clockwise) orientations for boundaries of faces
of ∆, for the boundary path ∂∆ of ∆, and, more generally, for boundaries of disk
subdiagrams of ∆. It is convenient to assume, as in [15], [30], that the boundary
BOUNDED AND PRECISE WORD PROBLEMS FOR GROUP PRESENTATIONS
9
path ∂Π of every face Π of a disk diagram ∆ has the positive orientation while the
boundary path ∂∆ of ∆ has the negative orientation.
If o ∈ ∂∆ is a vertex, let ∂|o ∆ denote a boundary path of ∆ (negatively oriented)
starting (and ending) at o. Using this notation, we now state van Kampen lemma
on geometric interpretation of consequences of defining relations.
Lemma 2.1. Let W be a nonempty word over A±1 and let a group G be defined
G
by presentation (1.1). Then W = 1 if and only if there is a disk diagram over
presentation (1.1) such that ϕ(∂|o ∆) ≡ W for some vertex o ∈ ∂∆.
Proof. The proof is straightforward, for details the reader is referred to [15], [30],
see also [18].
A disk diagram ∆ over presentation (1.1) with ϕ(∂∆) ≡ W is called minimal
if ∆ contains a minimal number of faces among all disk diagrams ∆′ such that
ϕ(∂∆′ ) ≡ W .
A disk diagram ∆ over (1.1) is called reduced if ∆ contains no two faces Π1 , Π2
such that there is a vertex v ∈ ∂Π1 , v ∈ ∂Π2 and the boundary paths ∂|v Π1 , ∂|v Π2
of the faces, starting at v, satisfy ϕ(∂|v Π1 ) ≡ ϕ(∂|v Π2 )−1 . If ∆ is not reduced and
Π1 , Π2 is a pair of faces that violates the definition of being reduced for ∆, then
these faces Π1 , Π2 are called a reducible pair (cf. similar definitions in [15], [18],
[30]). A reducible pair of faces Π1 , Π2 can be removed from ∆ by a surgery that
cuts through the vertex v and identifies the boundary paths ∂|v Π1 and (∂|v Π2 )−1 ,
see Fig. 2.1, more details can be found in [18], [30]. As a result, one obtains a disk
diagram ∆′ such that ϕ(∂∆′ ) ≡ ϕ(∂∆) and |∆′ (2)| = |∆(2)| − 2, where |∆(2)| is
the number of faces in ∆. In particular, a minimal disk diagram is always reduced.
v′
Π1
Π2
v
v ′′
Fig. 2.1
Note that if o ∈ ∂∆ is a vertex and ∂|o ∆ = q1 q2 is a factorization of the boundary
path ∂|o ∆ of ∆, q1 is closed, 0 < |q1 |, |q2 | < |∂∆|, then the notation ∂|o ∆ is in fact
ambiguous, because the path q2 q1 also has the form ∂|o ∆. To avoid this (and other)
G
type of ambiguity, for a given pair (W, ∆), where W = 1 and ∆ is a disk diagram for
W as in Lemma 2.1, we consider a “model” simple path PW such that |PW | = |W |,
PW is equipped with a labeling function ϕ : P~W (1) → A±1 on the set P~W (1) of its
oriented edges so that ϕ(e−1 ) = ϕ(e)−1 and ϕ(PW ) ≡ W .
It will be convenient to identify vertices of PW with integers 0, 1, . . . , |W | so that
(PW )− = 0 and the numbers corresponding to vertices increase by one as one goes
along PW from (PW )− = 0 to (PW )+ = |W |. This makes it possible to compare
vertices v1 , v2 ∈ PW (0) by the standard order ≤ defined on the integers, consider
vertices v1 ± 1 etc.
G
For a given pair (W, ∆), where W = 1, let
α : PW → ∂∆
10
S. V. IVANOV
be a continuous cellular map that preserves dimension of cells, ϕ-labels of edges,
and has the property that α((PW )− ) = α(0) = o, where o is a fixed vertex of ∂∆
with ϕ(∂|o ∆) ≡ W .
If v is a vertex of PW , let
PW (fact, v) = p1 p2
denote the factorization of PW defined by v so that (p1 )+ = v. Analogously, if
v1 , v2 are vertices of PW with v1 ≤ v2 , we let
PW (fact, v1 , v2 ) = p1 p2 p3
denote the factorization of PW defined by v1 , v2 so that (p2 )− = v1 and (p2 )+ = v2 .
Note that if v1 = v2 , then p2 = {v1 } and |p2 | = 0. Clearly, |p2 | = v2 − v1 .
Making use of the introduced notation, consider a vertex v of PW and let
PW (fact, v) = p1 p2 . Define ∂|v ∆ := α(p2 )α(p1 ). This notation ∂|v ∆, in place of
∂|α(v) ∆, will help us to avoid the potential ambiguity when writing ∂|α(v) ∆. In particular, if W̄ is a cyclic permutation of W so that the first k, where 0 ≤ k ≤ |W |−1,
letters of W are put at the end of W , then ϕ(∂|k ∆) ≡ W̄ . It is clear that
ϕ(∂|0 ∆) ≡ W .
Consider the following property.
(A) Suppose that ∆ is a disk diagram over (1.2). If Π is a face of ∆ and e ∈ ∂Π
is an edge then e−1 ∈ ∂∆.
We now state a lemma in which we record some simple properties of disk diagrams over (1.2) related to property (A).
Lemma 2.2. Let ∆ be a disk diagram over presentation (1.2). Then the following
hold true.
(a) If ∆ has property (A), then the degree of every vertex of ∆ is at most 2|∂∆|,
the boundary path ∂Π of every face Π of ∆ is simple, and
X
|∂Π| ≤ |∂∆|.
|∆(2)| ≤ |∂∆|,
Π∈∆(2)
′
(b) There exists a disk diagram ∆ over (1.2) such that ϕ(∂∆′ ) ≡ ϕ(∂∆),
|∆ (2)| ≤ |∆(2)| and ∆′ has property (A).
′
Proof. (a) Let v be a vertex of ∆. By property (A), v ∈ ∂∆ and if e is an edge such
that e− = v, then either e ∈ ∂∆ or e ∈ ∂Π, where Π is a face of ∆, and e−1 ∈ ∂∆.
This implies that deg v ≤ 2|∂∆|.
If the boundary path ∂Π of a face Π of ∆ is not simple, then there is a factorization ∂Π = u1 u2 , where u1 , u2 are closed subpaths of ∂Π and 0 < |u1 |, |u2 | < |∂Π|,
see Fig. 2.2. Clearly, the edges of one of the paths u1 , u2 do not belong to the
boundary ∂∆ of ∆, contrary to property (A) of ∆.
u1
u2
Π
∆2
Fig. 2.2
∂Π = u1 u2
BOUNDED AND PRECISE WORD PROBLEMS FOR GROUP PRESENTATIONS
11
P
The inequality |∆(2)| ≤ |∂∆| and its stronger version Π∈∆(2) |∂Π| ≤ |∂∆| are
immediate from property (A).
(b) Suppose that a disk diagram ∆ over (1.2) does not have property (A).
First assume that ∆ contains a face Π whose boundary path ∂Π is not simple.
Then, as in the proof of part (a), there is a factorization ∂Π = u1 u2 , where u1 , u2 are
closed subpaths of ∂Π and 0 < |u1 |, |u2 | < |∂Π|. Renaming u1 and u2 if necessary,
we may assume that u2 bounds a disk subdiagram ∆2 of ∆ such that ∆2 does not
contain Π, see Fig. 2.2. If u2 is not simple, then we can replace u2 by its closed
subpath u′2 such that 0 < |u′2 | < |u2 | and u′2 bounds a disk subdiagram ∆′2 that
contains no Π. Hence, choosing a shortest path u2 as above, we may assume that
u2 is simple. By Lemma 2.1 applied to ∆2 with ∂∆2 = u2 , we have
G
ϕ(u2 ) ≡ ϕ(∂∆2 ) =2 1.
(2.1)
aεk
i ,
Denote ϕ(∂Π) ≡
where ε = ±1, k ∈ Ei .
Note that the group G2 is the free product of cyclic groups generated by the
images of generators a1 , . . . , am and the image of aj has order nj > 0 if Ej 6= {0}
G
or the image of aj has infinite order if Ej = {0}. Hence, an equality aℓj =2 1, where
ℓ 6= 0, implies that Ej 6= {0} and nj divides ℓ.
2
It follows from (2.1) and 0 < |u2 | < |∂Π| that ϕ(u2 ) ≡ aεk
i , where k2 ∈ Ei ,
ε(k−k2 )
, where k − k2 ∈ Ei .
k2 < k. Therefore, Ei = {ni , 2ni , . . . } and ϕ(u1 ) ≡ ai
ε(k−k2 )
′
′
. Now we
Hence, we can consider a face Π such that ϕ(∂Π ) ≡ ϕ(u1 ) ≡ ai
take the subdiagram ∆2 out of ∆ and replace the face Π with ϕ(∂Π) ≡ aεk
i by the
ε(k−k2 )
. Doing this results in a disk diagram ∆′ such that
face Π′ with ϕ(∂Π′ ) ≡ ai
ϕ(∂∆′ ) ≡ ϕ(∂∆′ ) and |∆′ (2)| < |∆(2)| as |∆2 (2)| > 0.
Assume that, for every face Π in ∆, the boundary path ∂Π is simple. Also,
assume that the property (A) fails for ∆. Then there are faces Π1 , Π2 , Π1 6= Π2 ,
and an edge e such that e ∈ ∂Π1 and e−1 ∈ ∂Π2 . Consider a disk subdiagram Γ of ∆
that contains Π1 , Π2 and Γ is minimal with this property relative to |Γ(2)| + |Γ(1)|.
Since ∂Π1 , ∂Π2 are simple paths, it follows that ∂Γ = r1 r2 , where r1−1 is a subpath
of ∂Π1 and r2−1 is a subpath of ∂Π2 . Denote ϕ(∂Π1 ) ≡ aiε1 k1 , where ε1 = ±1 and
k1 ∈ Ei . Clearly, ϕ(∂Π2 ) ≡ ai−ε1 k2 , where k2 ∈ Ei and
0
ϕ(∂Γ) ≡ ϕ(r1 )ϕ(r2 ) = aεk
i ,
G
2
where ε = ±1 and k ≥ 0. As above, we observe that k ∈ Ei following from aεk
i = 1.
′
′
′ ′
′
Hence, we may consider a disk diagram Γ such that ∂Γ = r1 r2 , where ϕ(r1 ) ≡
ϕ(r1 ), ϕ(r2′ ) ≡ ϕ(r2 ), and Γ′ contains a single face Π such that ϕ(∂Π) ≡ a−εk
if
i
k 6= 0 or Γ′ contains no faces if k = 0. We take the subdiagram Γ out of ∆ and
replace Γ with Γ′ , producing thereby a disk diagram ∆′ such that ϕ(∂∆′ ) ≡ ϕ(∂∆′ )
and |∆′ (2)| < |∆(2)|.
We now observe that if property (A) fails for ∆ then there is a face Π in ∆
such that ∂Π is not simple or there are distinct faces Π1 , Π2 and an edge e such
that e ∈ ∂Π1 and e−1 ∈ ∂Π2 . In either case, as was shown above, we can find a
disk diagram ∆′ such that ϕ(∂∆′ ) ≡ ϕ(∂∆′ ) and |∆′ (2)| < |∆(2)|. Now obvious
induction on |∆(2)| completes the proof of part (b).
In view of Lemma 2.2(b), we will be assuming in Sects. 4–5, 8 that if ∆ is a disk
diagram over presentation (1.2), then ∆ has property (A).
12
S. V. IVANOV
3. Proof of Proposition 1.1
Proposition 1.1. (a) There exists a decidable group presentation (1.1) for which
the word problem is solvable while the bounded and precise word problems are not
solvable.
(b) If the bounded word problem is solvable for (1.1), then the precise word
problem is also solvable.
(c) For every finite group presentation (1.1), the bounded word problem is in
NP, i.e., it can be solved in nondeterministic polynomial time, and the precise word
problem is in PSPACE, i.e., it can be solved in polynomial space.
(d) There exists a finite group presentation (1.1) for which the bounded and
precise word problems are solvable while the word problem is not solvable.
(e) There exists a finitely presented group (1.1) for which the bounded word problem is NP-complete and the precise word problem is NP-hard.
Proof. (a) We will use the construction of [12, Example 3] suggested by C. Jockush
and I. Kapovich. Consider the group presentation
h a, b k ai = 1, ai bki = 1, i ∈ N i,
(3.1)
where K = {k1 , k2 , . . . } is a recursively enumerable but not recursive subset of the
set of natural numbers N with the indicated enumeration and k1 = 1. It is clear
that the set of relations is decidable and this presentation defines the trivial group,
hence the word problem is solvable for (3.1). On the other hand, it is easy to see
that the bounded word problem for a pair (bk , 2), where k ∈ N, holds true if and
only if k ∈ K. Analogously, the precise word problem for a pair (bk , 2) holds true if
and only if k ∈ K. Since the set K is not recursive, it follows that both the bounded
word problem and the precise word problem for presentation (3.1) are unsolvable.
(b) Note that the precise word problem holds true for a pair (W, n) if and only
if the bounded word problem is true for (W, n) and the bounded word problem is
false for (W, n − 1). This remark means that the solvability of the bounded word
problem for (1.1) implies the solvability of the precise word problem for (1.1). On
the other hand, the bounded word problem holds for a pair (W, n) if and only if
the precise word problem holds for (W, k) with some k ≤ n. This remark means
that the solvability of the precise word problem for (1.1) implies the solvability of
the bounded word problem for (1.1), as required.
(c) Suppose that presentation (1.1) is finite, i.e., both A and R are finite, and we
are given a pair (W, 1n ). It follows from definitions and Lemma 2.1 that the bounded
word problem holds for the pair (W, 1n ) if and only if there is a disk diagram ∆
~
such that ϕ(∂∆) ≡ W and |∆(2)| ≤ n. Observe that |∆(1)|
≤ M n + |W |, where
M = max{|R| : R ∈ R} is a constant. Therefore, the size of a disk diagram ∆ with
ϕ(∂∆) ≡ W is bounded by a linear function in n + |W | and such a diagram ∆ can
be used as a certificate to verify in polynomial time that the bounded word problem
holds for the pair (W, 1n ). Thus the bounded word problem for finite presentation
(1.1) is in NP.
Recall that the precise word problem holds for a pair (W, 1n ) if and only if the
bounded word problem is true for (W, 1n ) and the bounded word problem is false
for (W, 1n−1 ). As we saw above, the bounded word problem for (1.1) is in NP,
hence, the complement of the bounded word problem for (1.1) is in coNP. Since
BOUNDED AND PRECISE WORD PROBLEMS FOR GROUP PRESENTATIONS
13
both coNP and NP are subsets of PSPACE, it follows that the precise word problem
for finite presentation (1.1) is in PSPACE.
(d) According to Boone [6], [7] and Novikov [25], [26], see also [18], there exists a
finite group presentation (1.1) such that the word problem for this presentation is
not solvable. In view of part (c) both the bounded word problem and precise word
problem for this presentation are solvable.
(e) According to Birget, Sapir, Ol’shanskii and Rips [5], there exists a finite
group presentation (1.1) whose isoperimetric function is bounded by a polynomial
p(x) and for which the word problem is NP-complete. It follows from definitions
G
that if W = 1 and ∆ is a minimal diagram over presentation (1.1) such that
ϕ(∂∆) ≡ W then |∆(2)| ≤ p(|W |). Therefore, the bounded word problem, whose
input is (W, 1n ), where n ≥ p(|W |), is equivalent to the word problem, whose input
is W . Since the latter problem is NP-complete, it follows that the bounded word
problem for (1.1) is NP-hard. By part (c), the bounded word problem for (1.1) is
in NP, whence the bounded word problem for (1.1) is NP-complete.
Note that the word problem for given word W is equivalent to the disjunction
of the claims that the precise word problem holds for the pairs (W, 11 ), (W, 12 ), . . . ,
(W, 1p(|W |) ). Since p(x) is a polynomial, it follows that the precise word problem
for this presentation (1.1) is NP-hard.
4. Calculus of Brackets for Group Presentation (1.2)
As in Theorem 1.2, consider a group presentation of the form
G2 = h a1 , . . . , am k aki i = 1, ki ∈ Ei , i = 1, . . . , m i,
(1.2)
where, for every i, one of the following holds: Ei = {0}, or, for some integer ni > 0,
Ei = {ni }, or Ei = ni N = {ni , 2ni , 3ni , . . . }.
G
Suppose that W is a nonempty word over A±1 , W =2 1 and ∆ is a disk diagram
over presentation (1.2) such that ϕ(∂|0 ∆) ≡ W and ∆ has property (A). Recall
that the existence of such a diagram ∆ follows from Lemmas 2.1–2.2(b).
Lemma 4.1. Suppose that ∆ is a disk diagram over presentation (1.2) and ∆
contains no faces, i.e., ∆ is a tree or, equivalently, ∆ is a disk diagram over presentation F (A) = h a1 , . . . , am k ∅ i of the free group F (A) with no relations, and
assume that ϕ(∂|0 ∆) ≡ W , where |W | > 2. Then there are vertices v1 , v2 ∈ PW
such that v1 < v2 , α(v1 ) = α(v2 ) and if PW (fact, v1 , v2 ) = p1 p2 p3 is the factorization of PW defined by v1 , v2 , then
min(|p2 |, |p1 | + |p3 |) ≥ 13 |∂|0 ∆| = 13 |W |.
Proof. It is easy to verify that if |W | ≤ 6, then Lemma 4.1 is true. Hence, we may
assume that |W | > 6.
For every pair v1′ , v2′ of vertices of PW such that v1′ < v2′ and α(v1′ ) = α(v2′ ),
consider the factorization PW (fact, v1′ , v2′ ) = p′1 p′2 p′3 and pick such a pair that
maximizes min(|p′2 |, |p′1 | + |p′3 |). Let v1 , v2 be such a maximal pair and denote
PW (fact, v1 , v2 ) = p1 p2 p3 . Arguing on the contrary, assume that
min(|p2 |, |p1 | + |p3 |) < 31 |W |.
(4.1)
14
S. V. IVANOV
Denote qi = α(pi ), i = 1, 2, 3. Let e1 , . . . , ek , f1 , . . . , fℓ , k, ℓ ≥ 1, be all edges that
start at the vertex α(v1 ) = α(v2 ) so that
−1
q2 = e1 s1 e−1
1 . . . e k sk e k ,
q3 q1 = f1 t1 f1−1 . . . fℓ tℓ fℓ−1 ,
where s1 , . . . , sk and t1 , . . . , tℓ are subpaths of q2 and q3 q1 , resp., see Fig. 4.1.
s2
...
e1 e2
s1
fℓ
tℓ
ek
f1
f2
sk
t1
...
t2
Fig. 4.1
First we assume that |p2 | ≥ |p1 | + |p3 |. Then, in view of inequality (4.1),
|p2 | > 23 |W |,
|p1 | + |p3 | < 31 |W |.
(4.2)
Suppose that for some i we have
1
|ei si e−1
i | ≥ 2 |W |.
(4.3)
Pick vertices v1′ , v2′ ∈ PW for which if PW (fact, v1′ , v2′ ) = p′1 p′2 p′3 then α(p′2 ) =
ei si ei−1 . If k > 1, then |ei si e−1
i | < |p2 | and we have a contradiction to the maximality of the pair v1 , v2 because α(v1′ ) = α(v2′ ). Hence, k = 1 and i = 1.
Now we pick vertices v1′ , v2′ ∈ PW for which if PW (fact, v1′ , v2′ ) = p′1 p′2 p′3 then
α(p′2 ) = s1 . Note |p′2 | = |s1 | = |p2 | − 2 > 32 |W | − 2 ≥ 31 |W | for |W | > 6 and
|p′1 | + |p′3 | = |p1 | + |p3 | + 2 < 13 |W | + 2 ≤ 23 |W | for |W | > 6. Hence, either
min(|p′2 |, |p′1 | + |p′3 |) ≥ 31 |W |
(4.4)
min(|p′2 |, |p′1 | + |p′3 |) > min(|p2 |, |p1 | + |p3 |)
(4.5)
if |p′2 | ≤ |p′1 | + |p′3 | or
|p′2 |
|p′1 |
|p′3 |.
if
>
+
In either case, we obtain a contradiction to the maximality of
the pair v1 , v2 because α(v1′ ) = α(v2′ ). Thus it is shown that the inequality (4.3) is
1
false, hence, for every i = 1, . . . , k, we have |ei si e−1
i | < 2 |W |.
1
′
′
Assume |ei si e−1
i | ≥ 3 |W | for some i. Pick vertices v1 , v2 ∈ PW for which if
−1
−1
′
′
′ ′ ′
′
PW (fact, v1 , v2 ) = p1 p2 p3 then α(p2 ) = ei si ei . Since |ei si ei | < 21 |W |, it follows
that min(|p′2 |, |p′1 | + |p′3 |) = |p′2 | ≥ 31 |W |. A contradiction to the maximality of the
1
pair v1 , v2 proves that |ei si e−1
i | < 3 |W | for every i = 1, . . . , k. According to (4.2),
2
|p2 | > 3 |W |, hence, k ≥ 3 and, for some i ≥ 2, we obtain
1
3 |W |
−1
2
≤ |e1 s1 e−1
1 . . . ei si ei | ≤ 3 |W |.
This means the existence of vertices v1′ , v2′ ∈ PW for which, if
PW (fact, v1′ , v2′ ) = p′1 p′2 p′3 ,
BOUNDED AND PRECISE WORD PROBLEMS FOR GROUP PRESENTATIONS
15
−1
then the paths p′1 , p′2 , p′3 have the properties that α(p′2 ) = e1 s1 e−1
and
1 . . . e i si e i
min(|p′2 |, |p′1 | + |p′3 |) ≥ 31 |W |.
This contradiction to the maximality of the pair v1 , v2 completes the first main case
|p2 | ≥ |p1 | + |p3 |.
Now assume that |p2 | < |p1 | + |p3 |. In this case, we repeat the above arguments
with necessary changes. By the inequality (4.1), |p2 | < 13 |W | and |p1 |+|p3 | > 23 |W |.
Suppose that for some j we have
|fj tj fj−1 | ≥ 21 |W |
(4.6)
Pick vertices v1′ , v2′ ∈ PW so that α(v1′ ) = α(v2′ ), v1′ < v2′ and, if
PW (fact, v1′ , v2′ ) = p′1 p′2 p′3 ,
then either α(p′2 ) = fj tj fj−1 in case when fj tj fj−1 is a subpath of one of q1 , q3 , or
α(p′3 )α(p′1 ) = fj tj fj−1 in case when fj tj fj−1 has common edges with both q1 and
q3 . In either case,
min(|p′2 |, |p′1 | + |p′3 |) > min(|p2 |, |p1 | + |p3 |)
whenever ℓ > 1. By the maximality of the pair v1 , v2 , we conclude that ℓ = 1 and
j = 1.
In the case ℓ = j = 1, we consider two subcases: min(|p1 |, |p3 |) > 0 and
min(|p1 |, |p3 |) = 0.
Assume that min(|p1 |, |p3 |) > 0. Then we can pick vertices v1′ , v2′ ∈ PW for which,
if PW (fact, v1′ , v2′ ) = p′1 p′2 p′3 , then the subpaths p′1 , p′2 , p′3 of PW have the properties
that α(p′2 ) = f1−1 q2 f1 and α(p′3 )α(p′1 ) = t1 . Similarly to the above arguments that
led to inequalities (4.4)–(4.5), it follows from the inequality |W | > 6 that either
min(|p′2 |, |p′1 | + |p′3 |) ≥ 31 |W |
if |p′1 | + |p′3 | < |p′2 | or
min(|p′2 |, |p′1 | + |p′3 |) > min(|p2 |, |p1 | + |p3 |)
if |p′1 | + |p′3 | ≥ |p′2 |. In either case, we obtain a contradiction to the maximality of
the pair v1 , v2 .
Now assume that min(|p1 |, |p3 |) = 0. For definiteness, let |pi | = 0, i ∈ {1, 3}.
Then we can pick vertices v1′ , v2′ ∈ PW for which α(v1′ ) = α(v2′ ), v1′ < v2′ and, if
PW (fact, v1′ , v2′ ) = p′1 p′2 p′3 , then the subpaths p′1 , p′2 , p′3 of PW have the properties
that p′2 = p4−i , p′i = p2 , |p′4−i | = |pi | = 0. Hence, |p′2 | > |p′1 | + |p′3 | and
min(|p′2 |, |p′1 | + |p′3 |) = min(|p2 |, |p1 | + |p3 |).
This means that the subcase min(|p1 |, |p3 |) = 0 is reduced to the case |p′2 | ≥ |p′1 | +
|p′3 | which was considered above.
The case ℓ = j = 1 is complete and it is shown that the inequality (4.6) is false,
hence, for every j = 1, . . . , ℓ, we have |fj tj fj−1 | < 21 |W |.
Suppose that |fj tj fj−1 | ≥ 31 |W | for some j. Pick vertices v1′ , v2′ ∈ PW so that if
PW (fact, v1′ , v2′ ) = p′1 p′2 p′3 , then either α(p′2 ) = fj tj fj−1 in case when fj tj fj−1 is a
subpath of one of q1 , q3 , or α(p′3 )α(p′1 ) = fj tj fj−1 in case when fj tj fj−1 has common
edges with both q1 and q3 . Since |fj tj fj−1 | < 21 |W |, it follows that
min(|p′2 |, |p′1 | + |p′3 |) ≥ 31 |W |.
16
S. V. IVANOV
A contradiction to the maximality of the pair v1 , v2 proves that |fj tj fj−1 | < 31 |W |
for every j = 1, . . . , ℓ. Since |p1 | + |p3 | > 32 |W |, we get ℓ ≥ 3 and, for some j ≥ 2,
we obtain
−1
−1
2
1
3 |W | ≤ |f1 t1 f1 . . . fj tj fj | ≤ 3 |W | .
This means the existence of vertices v1′ , v2′ ∈ PW for which α(v1′ ) = α(v2′ ) and if
PW (fact, v1′ , v2′ ) = p′1 p′2 p′3 then the subpaths p′1 , p′2 , p′3 have the following properties:
Either α(p′2 ) = f1 t1 f1−1 . . . fj tj fj−1 or α(p′3 )α(p′1 ) = f1 t1 f1−1 . . . fj tj fj−1 . The it is
clear that
min(|p′2 |, |p′1 | + |p′3 |) ≥ 31 |W |.
This contradiction to the choice of the pair v1 , v2 completes the second main case
when |p2 | < |p1 | + |p3 |.
Lemma 4.2. Suppose ∆ is a disk diagram with property (A) over presentation
(1.2) and ϕ(∂|0 ∆) ≡ W with |W | > 2. Then one of the following two claims holds.
(a) There are vertices v1 , v2 ∈ PW such that α(v1 ) = α(v2 ), v1 < v2 , and if
PW (fact, v1 , v2 ) = p1 p2 p3 is the factorization of PW defined by v1 , v2 , then
min(|p2 |, |p1 | + |p3 |) ≥ 61 |W |.
(b) There exists a face Π in ∆ with |∂Π| ≥ 2 and there are vertices v1 , v2 ∈ PW
such that α(v1 ), α(v2 ) ∈ ∂Π, v1 < v2 , and if PW (fact, v1 , v2 ) = p1 p2 p3 then
min(|p2 |, |p1 | + |p3 |) ≥ 61 |W |.
(4.7)
~
In addition, if (∂Π)−1 = e1 . . . e|∂Π| , where ei ∈ ∆(1),
and ∂∆ = e1 h1 . . . e|∂Π| h|∂Π| ,
where hi is a closed subpath of ∂∆, then, for every i, hi is a subpath of either α(p2 )
or α(p3 )α(p1 ) and |hi | ≤ 56 |W |, see Fig. 4.2.
h3
e3
h2
Π
e2
h1
Fig. 4.2
e1
e|∂Π|
h|∂Π|
h|∂Π|−1
Proof. Since ∆ has property (A), it follows that if e ∈ ∂Π, where Π is a face of ∆,
then e−1 ∈ ∂∆ and if e ∈ ∂∆, then e−1 ∈ ∂Π′ ∪ ∂∆, where Π′ is a face of ∆.
Consider a planar graph Γ∆ constructed from ∆ as follows. For every face Π of
∆, we pick a vertex oΠ in the interior of Π. The vertex set of Γ∆ is
V (Γ∆ ) := ∆(0) ∪ {oΠ | Π ∈ ∆(2)} ,
where ∆(i) is the set of i-cells of ∆, i = 0, 1, 2. For every face Π of ∆, we delete
nonoriented edges of ∂Π and draw |∂Π| nonoriented edges that connect oΠ to all
vertices of ∂Π, see Fig. 4.3.
BOUNDED AND PRECISE WORD PROBLEMS FOR GROUP PRESENTATIONS
Π
EΠ
17
oΠ
Fig. 4.3
We draw these edges so that their interiors are disjoint and are contained in the
interior of Π. Let EΠ denote the set of these |∂Π| edges. The set E(Γ∆ ) of nonoriented edges of Γ∆ is ∆(1), without edges of faces of ∆, combined with ∪Π∈∆(2) EΠ ,
hence,
E(Γ∆ ) := ∪Π∈∆(2) EΠ ∪ (∆(1) \ {e | e ∈ ∆(1), e ∈ ∂Π, Π ∈ ∆(2)}) .
It follows from definitions that |V (Γ∆ )| = |∆(0)| + |∆(2)|, |E(Γ∆ )| = |∆(1)| and
that Γ∆ is a tree. Assigning labels to oriented edges of Γ∆ , by using letters from
A±1 , we can turn Γ∆ into a disk diagram over presentation h a1 , . . . , am k ∅ i of
the free group F (A).
~
and
Denote ϕ(∂|o′ Γ∆ ) ≡ W ′ for some vertex o′ ∈ ∆(0), where |W ′ | = |∆(1)|,
′
′
let α (PW ′ ) = ∂|o′ Γ∆ = ∂|0 Γ∆ . Since |W | ≥ |W | > 2, Lemma 4.1 applies to Γ∆
and yields the existence of vertices u1 , u2 ∈ PW ′ such that α′ (u1 ) = α′ (u2 ) in ∂Γ∆ ,
u1 < u2 , and if PW ′ (fact, u1 , u2 ) = r1 r2 r3 , then
min(|r2 |, |r1 | + |r3 |) ≥ 31 |W ′ |.
(4.8)
′
First suppose that α (u1 ) is a vertex of ∆. It follows from the definition of the
tree Γ∆ that there is a factorization PW = p1 p2 p3 of the path PW such that the
vertex α((p2 )− ) = α((p2 )+ ) is α′ (u1 ) ∈ ∆(0) and |pi | ≤ |ri | ≤ 2|pi |, i = 1, 2, 3.
Indeed, to get from ∂∆ to ∂Γ∆ we replace every edge e ∈ ∂Π, Π ∈ ∆(2), by two
edges of EΠ , see Fig. 4.3. Hence, if r is a subpath of ∂Γ∆ and p is a corresponding
to r subpath of ∂∆ with r− = p− ∈ ∆(0), r+ = p+ ∈ ∆(0), then |p| ≤ |r| ≤ 2|p|.
Then it follows from (4.8) that
min(|p2 |, |p1 | + |p3 |) ≥
1
2
min(|r2 |, |r1 | + |r3 |) ≥ 16 |W ′ | ≥ 61 |W |,
as required.
Now assume that α′ (u1 ) = α′ (u2 ) = oΠ for some face Π ∈ ∆(2). Let e1 , . . . , ek ,
f1 , . . . , fℓ , k, ℓ ≥ 0, be all oriented edges of Γ∆ that start at the vertex α′ (u1 ) =
α′ (u2 ) = oΠ so that
−1
α′ (r2 ) = e1 s1 e−1
1 . . . e k sk e k ,
α′ (r3 )α′ (r1 ) = f1 t1 f1−1 . . . fℓ tℓ fℓ−1 ,
where s1 , . . . , sk and t1 , . . . , tℓ are closed subpaths of α′ (r2 ) and α′ (r3 )α′ (r1 ), resp.
Since α′ (u1 ) = α′ (u2 ) = oΠ , it follows that k+ℓ = |∂Π|. Since min(|r2 |, |r1 |+|r3 |) ≥
1
′
3 |W | > 0 is an integer, we also have that k, ℓ ≥ 1 and |∂Π| > 1. If |r3 | > 0, we
consider vertices u′1 := u1 + 1, u′2 := u2 + 1. On the other hand, if |r3 | = 0, then
|r1 | > 0 and we consider vertices u′1 := u1 − 1, u′2 := u2 − 1. In either case, denote
PW ′ (u′1 , u′2 ) = r1′ r2′ r3′ . Then |r2′ | = |r2 | and |r1′ | + |r3′ | = |r1 | + |r3 |, hence, by virtue
of (4.8),
min(|r2′ |, |r1′ | + |r3′ |) ≥ 31 |W ′ |.
Note that the vertices α′ ((r2′ )− ), α′ ((r2′ )+ ) belong to the boundary ∂Π. Hence,
as above, there is also a factorization PW (fact, v1 , v2 ) = p1 p2 p3 such that α(v1 ) =
α′ (u′1 ), α(v2 ) = α′ (u′2 ) and |pi | ≤ |ri′ | ≤ 2|pi |, i = 1, 2, 3. Therefore,
min(|p2 |, |p1 | + |p3 |) ≥
1
2
min(|r2′ |, |r1′ | + |r3′ |) ≥ 16 |W ′ | ≥ 61 |W |,
18
S. V. IVANOV
as required.
It remains to observe that it follows from the definition of vertices v1 , v2 that
every hi is a subpath of one of the paths α(p2 ), α(p3 )α(p1 ). In particular, |hi | ≤
5
6 |W |, as desired.
In the definitions below, we assume that ∆ is a disk diagram over presentation
(1.2) such that ∆ has property (A), ϕ(∂|0 ∆) ≡ W , |W | > 0, and that the pair
(W, ∆) is fixed.
A 6-tuple
b = (b(1), b(2), b(3), b(4), b(5), b(6))
of integers b(1), b(2), b(3), b(4), b(5), b(6) is called a bracket for the pair (W, ∆)
if b(1), b(2) satisfy the inequalities 0 ≤ b(1) ≤ b(2) ≤ |W | and, in the notation
PW (fact, b(1), b(2)) = p1 p2 p3 , one of the following two properties (B1)–(B2) holds
true.
(B1) b(3) = b(4) = b(5) = 0, α(b(1)) = α(b(2)), and the disk subdiagram ∆b of
∆, defined by ∂|b(1) ∆b = α(p2 ), contains b(6) faces, see Fig. 4.4(B1).
(B2) b(3) > 0 and α(b(1)), α(b(2)) ∈ ∂Π, where Π is a face of ∆ such that
εb(4)
ϕ(∂Π) ≡ ab(3) , b(4) > 0, ε = ±1, and if ∆b is the disk subdiagram
of ∆, defined by ∂|b(1) ∆b = α(p2 )u, where u is the subpath of ∂Π with
−b(5)
u− = α(b(2)) and u+ = α(b(1)), then ϕ(u) ≡ ab(3) and |∆b (2)| = b(6), see
Fig. 4.4(B2).
α(p2 ) = a(b)
α(p2 ) = a(b)
∆b
∆b
α(p1 )α(p2 )α(p3 ) = ∂∆
α(b(1)) = α(b(2))
u
Π
α(p1 )
∂Π
α(p3 )
α(0)
Fig. 4.4(B1)
α(p1 )
α(0)
α(p3 )
Fig. 4.4(B2)
A bracket b is said to have type B1 or type B2 if the property (B1) or (B2), resp.,
holds for b. Note that the equality b(4) = 0 in property (B1) and the inequality
b(4) > 0 in property (B2) imply that the type of a bracket is unique.
The boundary subpath α(p2 ) of the disk subdiagram ∆b associated with a bracket
b is denoted a(b) and is called the arc of b, see Figs. 4.4(B1)–4.4(B2).
For example, bv = (v, v, 0, 0, 0, 0) is a bracket of type B1 for every vertex v of
PW , called a starting bracket at v. Note that a(bv ) = α(v) = α(bv (1)).
BOUNDED AND PRECISE WORD PROBLEMS FOR GROUP PRESENTATIONS
19
The final bracket for (W, ∆) is bF = (0, |W |, 0, 0, 0, |∆(2)|), it has type B1 and
a(bF ) = ∂|0 ∆.
Let B be a set of brackets for the pair (W, ∆), perhaps, B is empty, B = ∅. We
say that B is a bracket system if, for every pair b, c ∈ B of distinct brackets, either
b(2) ≤ c(1) or c(2) ≤ b(1). In particular, the arcs of distinct brackets in B have no
edges in common. A bracket system consisting of a single final bracket is called a
final bracket system.
Now we describe four kinds of operations over brackets and over bracket systems:
additions, extensions, turns, and mergers. Let B be a bracket system.
Additions.
Suppose b is a starting bracket, b 6∈ B, and B ∪ {b} is a bracket system. Then
we may add b to B thus making an addition operation over B.
Extensions.
Suppose b ∈ B, b = (b(1), b(2), b(3), b(4), b(5), b(6)), and e1 a(b)e2 is a subpath of
∂|0 ∆, where a(b) is the arc of b and e1 , e2 are edges one of which could be missing.
Assume that b is of type B2, in particular, b(3), b(4) > 0. Using the notation of
ε1
the condition (B2), suppose e−1
1 is an edge of ∂Π, and ϕ(e1 ) = ab(3) , where ε1 = ±1.
If |b(5)| ≤ b(4) − 2 and ε1 b(5) ≥ 0, then we consider a bracket b′ such that
b′ (1) = b(1) − 1, b′ (2) = b(2), b′ (3) = b(3),
b′ (4) = b(4), b′ (5) = b(5) + ε1 , b′ (6) = b(6).
Note that a(b′ ) = e1 a(b). We say that b′ is obtained from b by an extension of type
1 (on the left). If (B \ {b}) ∪ {b′ } is a bracket system, then replacement of b ∈ B
with b′ in B is called an extension operation over B of type 1.
On the other hand, if |b(5)| = b(4)− 1 and ε1 b(5) ≥ 0, then we consider a bracket
b′ such that
b′ (1) = b(1) − 1, b′ (2) = b(2), b′ (3) = b′ (4) = b′ (5) = 0, b′ (6) = b(6) + 1.
In this case, we say that b′ is obtained from b by an extension of type 2 (on the
left). Note that a(b′ ) = e1 a(b) and b′ has type B1. If (B \ {b}) ∪ {b′ } is a bracket
system, then replacement of b ∈ B with b′ in B is called an extension operation
over B of type 2.
2
Analogously, assume that b has type B2, e2−1 is an edge of ∂Π, and ϕ(e2 ) = aεb(3)
,
where ε2 = ±1.
If |b(5)| ≤ b(4) − 2 and ε2 b(5) ≥ 0, then we consider a bracket b′ such that
b′ (1) = b(1), b′ (2) = b(2) + 1, b′ (3) = b(3),
b′ (4) = b(4), b′ (5) = b(5) + ε2 , b′ (6) = b(6).
Note that a(b′ ) = a(b)e2 and b′ has type B2. We say that b′ is obtained from b
by an extension of type 1 (on the right). If (B \ {b}) ∪ {b′ } is a bracket system,
then replacement of b ∈ B with b′ in B is called an extension operation over B of
type 1.
On the other hand, if |b(5)| = b(4) − 1, then we may consider a bracket b′ such
that
b′ (1) = b(1), b′ (2) = b(2) + 1, b′ (3) = b′ (4) = b′ (5) = 0, b′ (6) = b(6) + 1.
Note that a(b′ ) = a(b)e2 and b′ has type B1. We say that b′ is obtained from b
by an extension of type 2 (on the right). If (B \ {b}) ∪ {b′ } is a bracket system,
20
S. V. IVANOV
then replacement of b ∈ B with b′ in B is called an extension operation over B of
type 2.
Assume that b ∈ B is a bracket of type B1, e1 a(b)e2 is a subpath of ∂|0 ∆, both
′
e1 , e2 exist, and e1 = e−1
2 . Consider a bracket b of type B1 such that
b′ (1) = b(1) − 1, b′ (2) = b(2) + 1, b′ (3) = b′ (4) = b′ (5) = 0, b′ (6) = b(6).
Note that a(b′ ) = e1 a(b)e2 . We say that b′ is obtained from b by an extension of
type 3. If (B \ {b}) ∪ {b′ } is a bracket system, then replacement of b ∈ B with b′ in
B is called an extension operation over B of type 3.
Turns.
Let b ∈ B be a bracket of type B1. Then b(3) = b(4) = b(5) = 0. Suppose
Π is a face in ∆ such that Π is not in the disk subdiagram ∆b , associated with
Π
, where
b and bounded by the arc a(b) of b, and α(b(1)) ∈ ∂Π. If ϕ(∂Π) = aεn
j
′
′
ε = ±1, nΠ ∈ Ej , then we consider a bracket b with b (i) = b(i) for i = 1, 2, 5, 6,
and b′ (3) = j, b′ (4) = nΠ . Note that b′ has type B2. We say that b′ is obtained
from b by a turn operation. Replacement of b ∈ B with b′ in B is also called a
turn operation over B. Note that (B \ {b}) ∪ {b′ } is automatically a bracket system
(because so is B).
Mergers.
Now suppose that b, c ∈ B are distinct brackets such that b(2) = c(1) and one
of b(3), c(3) is 0. Then one of b, c is of type B1 and the other has type B1 or B2.
Consider a bracket b′ such that b′ (1) = b(1), b′ (2) = c(2), and b′ (i) = b(i) + c(i) for
i = 3, 4, 5, 6. Note that a(b′ ) = a(b1 )a(b2 ) and b′ is of type B1 if both bi , bj have
type B1 or b′ is of type B2 if one of b, c has type B2. We say that b′ is obtained
from b, c by a merger operation. Taking both b, c out of B and putting b′ in B is a
merger operation over B. Note that (B \ {b, c}) ∪ {b′ } is automatically a bracket
system.
We will say that additions, extensions, turns and mergers, as defined above, are
elementary operations over brackets and bracket systems for the pair (W, ∆).
Assume that one bracket system Bℓ is obtained from another bracket system
B0 by a finite sequence Ω of elementary operations and B0 , B1 , . . . , Bℓ is the corresponding to Ω sequence of bracket systems. Such a sequence B0 , B1 , . . . , Bℓ of
bracket systems will be called operational.
We will say that the sequence B0 , B1 , . . . , Bℓ has size bounded by (k1 , k2 ) if
ℓ ≤ k1 and, for every i, the number of brackets in Bi is at most k2 . Whenever it is
not ambiguous, we will also say that Ω has size bounded by (k1 , k2 ) if so does the
corresponding to Ω sequence B0 , B1 , . . . , Bℓ of bracket systems.
Lemma 4.3. There exists a sequence of elementary operations that converts the
empty bracket system for (W, ∆) into the final bracket system and has size bounded
by (8|W |, |W | + 1).
Proof. For every k with 0 ≤ k ≤ |W |, consider a starting bracket (k, k, 0, 0, 0, 0) for
(W, ∆). Making |W | + 1 additions, we get a bracket system
BW = {(k, k, 0, 0, 0, 0) | 0 ≤ k ≤ |W |}
of |W | + 1 starting brackets. Now, looking at the disk diagram ∆, we can easily
find a sequence of extensions, turns and mergers that converts BW into the final
bracket system, denoted Bℓ . Note that the inequality b(4) ≤ |W | for every bracket
BOUNDED AND PRECISE WORD PROBLEMS FOR GROUP PRESENTATIONS
21
b of intermediate systems follows from definitions and Lemma 2.2. To estimate the
total number of elementary operations, we note that the number of additions is
|W | + 1. The number of extensions is at most |W | because every extension applied
to a bracket system B increases the number
X
η(B) :=
(b(2) − b(1))
b∈B
by 1 or 2 and η(BW ) = 0, whereas η(Bℓ ) = |W |. The number of mergers is |W |
because the number of brackets |B| in a bracket system B decreases by 1 if B is
obtained by a merger and |BW | = |W | + 1, |Bℓ | = 1. The number of turns does
not exceed the total number of additions, extensions, and mergers, because a turn
operation is applied to a bracket of type B1 and results in a bracket of type B2 to
which a turn operation may not be applied, whence a turn operation succeeds an
addition, or an extension, or a merger. Therefore, the number of turns is at most
3|W | + 1 and so ℓ ≤ 6|W | + 2 ≤ 8|W |.
Lemma 4.4. Suppose there is a sequence Ω of elementary operations that converts
the empty bracket system E for (W, ∆) into the final bracket system F and has size
bounded by (k1 , k2 ). Then there is also a sequence of elementary operations that
transforms E into F and has size bounded by (11|W |, k2 ).
Proof. Assume that the sequence Ω has an addition operation which introduces
a starting bracket c = (k, k, 0, 0, 0, 0) with 0 ≤ k ≤ |W |. Since the final bracket
system contains no starting brackets, c must disappear and an elementary operation
is applied to c. If a merger is applied to c and another bracket b and the merger
yields b
c, then b
c = b. This means that the addition of c and the merger could be
skipped without affecting the sequence otherwise. Note that the size of the new
sequence Ω′ is bounded by (k1 − 2, k2 ). Therefore, we may assume that no merger
is applied to a starting bracket in Ω.
Now suppose that a turn is applied to c, yields c′ and then a merger is applied to
′
c , b and the merger produces b
c. Note that c′ has type B2 and b has type B1. Then
it is possible to apply a turn to b and get b′ with b′ = b
c. Hence, we can skip the
addition of c, the turn of c, the merger and use, in their place, a turn of b. Clearly,
the size of the new sequence Ω′ is bounded by (k1 − 2, k2 ).
Thus, by induction on k1 , we may assume that, for every starting bracket c which
is added by Ω, an extension is applied to c or an extension is applied to c′ and c′ is
obtained from c by a turn.
Now we will show that, for every starting bracket c which is added by Ω, there
are at most 2 operations of additions of c in Ω. Arguing on the contrary, assume
that c1 , c2 , c3 are the brackets equal to c whose additions are done in Ω. By the
above remark, for every i = 1, 2, 3, either an extension is applied to ci , resulting
in a bracket b
ci , or a turn is applied to ci , resulting in c′i , and then an extension is
applied to c′i , resulting in a bracket b
ci .
Let c1 , c2 , c3 be listed in the order in which the brackets b
c1 , b
c2 , b
c3 are created by
Ω. Note that if b
c1 is obtained from c1 by an extension (with no turn), then the
extension has type 3 and
c1 (1) = c1 (1) − 1, b
b
c1 (2) = c1 (2) + 1.
This means that brackets b
c2 , b
c3 could not be created by an extension after b
c1
appears, as d(2) ≤ d′ (1) or d(1) ≥ d′ (2) for distinct brackets d, d′ ∈ B of any
22
S. V. IVANOV
bracket system B. This contradiction proves that b
c1 is obtained from c′1 by an
′
extension. Then c1 has type B2, the extension has type 1 or 2 (and is either on
the left or on the right). Similarly to the forgoing argument, we can see that if
b
c1 is obtained by an extension on the left/right, then b
c2 must be obtained by an
extension on the right/left, resp., and that b
c3 cannot be obtained by any extension.
This contradiction proves that it is not possible to have in Ω more than two additions
of any starting bracket c. Thus, the number of additions in Ω is at most 2|W | + 2.
As in the proof of Lemma 4.3, the number of extensions is at most |W | and the
number of mergers is ≤ 2|W | + 1. Hence, the number of turns is ≤ 5|W | + 3 and the
total number of elementary operations is at most 10|W |+ 6 ≤ 11|W | as desired.
Lemma 4.5. Let there be a sequence Ω of elementary operations that transforms
the empty bracket system for the pair (W, ∆) into the final bracket system and has
size bounded by (k1 , k2 ) and let c be a starting bracket for (W, ∆). Then there is
also a sequence of elementary operations that converts the bracket system {c} into
the final bracket system and has size bounded by (k1 + 1, k2 + 1).
Proof. Let
B0 , B1 , . . . , Bℓ
(4.9)
be the corresponding to Ω sequence of bracket systems, where B0 is empty and Bℓ
is final.
First suppose that c(1) = 0 or c(1) = |W |.
Assume that no addition of c is used in Ω. Setting Bi′ := Bi ∪ {c}, i = 0, 1, . . . , ℓ,
we obtain an operational sequence of bracket systems B0′ , . . . , Bℓ′ that starts with
{c} and ends in the bracket system Bℓ′ = {c, dF }, where dF is the final bracket for
′
can be obtained from Bi′ by the same elementary operation
(W, ∆). Note that Bi+1
that was used to get Bi+1 from Bi . A merger operation applied to Bℓ′ yields the
′
′
final bracket system Bℓ+1
= {dF } and B0′ , . . . , Bℓ′ , Bℓ+1
is a desired sequence of
bracket systems of size bounded by (k1 + 1, k2 + 1).
Now suppose that an addition of c is used in Ω, Bi∗ +1 = Bi∗ ∪{c} is obtained from
Bi′ by addition of c, and i∗ is minimal with this property. Define Bj′ := Bj ∪ {c}
′
′
:= Bℓ . Then B0′ , . . . , Bℓ−1
is a
for j = 0, . . . , i∗ and Bi′∗ +1 := Bi∗ +2 , . . . , Bℓ−1
desired operational sequence of bracket systems that starts with {c}, ends in the
final bracket system, and has size bounded by (k1 − 1, k2 + 1).
We may now assume that 0 < c(1) < |W |. Let Bi∗ be the first bracket system
of the sequence (4.9) such that Bi∗ ∪ {c} is not a bracket system. The existence
of such Bi∗ follows from the facts that B0 ∪ {c} is a bracket system and Bℓ ∪ {c}
is not. Since B0 ∪ {c} is a bracket system, it follows that i∗ ≥ 1 and Bi∗ −1 ∪ {c}
is a bracket system. Since Bi∗ −1 ∪ {c} is a bracket system and Bi∗ ∪ {c} is not,
there is a bracket b ∈ Bi∗ such that b(1) < c(1) < b(2) and b is obtained from a
bracket d1 ∈ Bi∗ −1 by an extension or b is obtained from brackets d1 , d2 ∈ Bi∗ −1 by
a merger. In either case, it follows from definitions of elementary operations that
di (j) = c(1) for some i, j ∈ {1, 2}. Hence, we can use a merger applied to di (j)
and c which would result in di (j), i.e., in elimination of c from Bi∗ −1 ∪ {c} and in
getting thereby Bi∗ −1 from Bi∗ −1 ∪ {c}. Now we can see that the original sequence
of elementary operations, together with the merger Bi∗ −1 ∪ {c} → Bi∗ −1 can be
used to produce the following operational sequence of bracket systems
B0 ∪ {c}, . . . , Bi∗ −1 ∪ {c}, Bi∗ −1 , Bi∗ , . . . , Bℓ .
Clearly, the size of this new sequence is bounded by (k1 + 1, k2 + 1), as required.
BOUNDED AND PRECISE WORD PROBLEMS FOR GROUP PRESENTATIONS
23
Lemma 4.6. There exists a sequence of elementary operations that converts the
empty bracket system for the pair (W, ∆) into the final bracket system and has size
bounded by
(11|W |, C(log |W | + 1)),
where C = (log 65 )−1 .
Proof. First suppose that |W | ≤ 2. Then ∆ consists of a single edge, or of a single
face Π with |∂Π| ≤ 2, or of two faces Π1 , Π2 with |∂Π1 | = |∂Π2 | = 1. In each
of these three cases, we can easily find a sequence of elementary operations that
transforms the empty bracket system for (W, ∆) into the final bracket system by
using a single addition, at most two turns, and at most two extensions. Hence, the
size of the sequence is bounded by
(5, 1) ≤ (11|W |, C(log |W | + 1)
as C > 1.
Assuming |W | > 2, we proceed by induction on the length |W |. By Lemma 4.2
applied to (W, ∆), we obtain either the existence of vertices v1 , v2 ∈ PW such that
v1 < v2 , α(v1 ) = α(v2 ) and, if PW (fact, v1 , v2 ) = p1 p2 p3 , then
min(|p2 |, |p1 | + |p3 |) ≥ 61 |W |
(4.10)
or we get the existence of a face Π in ∆ and vertices v1 , v2 ∈ PW with the properties
stated in part (b) of Lemma 4.2.
First assume that Lemma 4.2(a) holds true and ∂|0 ∆ = q1 q2 q3 , where qi = α(pi ),
i = 1, 2, 3. Consider disk subdiagrams ∆1 , ∆2 of ∆ given by ∂|v2 ∆2 = q2 , ∂|0 ∆1 =
q1 q3 . Denote W2 ≡ ϕ(q2 ), W1 ≡ ϕ(q1 )ϕ(q3 ) and let PWi = PWi (Wi , ∆i ), i = 1, 2,
denote the corresponding paths such that α1 (PW1 ) = q1 q3 and α2 (PW2 ) = q2 .
Since |W1 |, |W2 | < |W |, it follows from the induction hypothesis that there is
a sequence Ω2 of elementary operations for (W2 , ∆2 ) that transforms the empty
bracket system into the final system and has size bounded by
(11|W2 |, C(log |W2 | + 1)).
(4.11)
Let B2,0 , . . . , B2,ℓ2 denote the corresponding to Ω2 sequence of bracket systems,
where B2,0 is empty and B2,ℓ2 is final.
We also consider a sequence Ω1 of elementary operations for (W1 , ∆1 ) that transforms the bracket system {c0 }, where
c0 := (|p1 |, |p1 |, 0, 0, 0, 0)
(4.12)
is a starting bracket, into the final bracket system. It follows from the induction
hypothesis and Lemma 4.5 that there is such a sequence Ω1 of size bounded by
(11|W1 | + 1, C(log |W1 | + 1) + 1).
(4.13)
Let B1,0 , . . . , B1,ℓ1 denote the corresponding to Ω1 sequence of bracket systems,
where B1,0 = {c0 } and B1,ℓ1 is final.
We will show below that these two sequences Ω2 , Ω1 of elementary operations,
the first one for (W2 , ∆2 ) and the second one for (W1 , ∆1 ), could be modified and
combined into a single sequence of elementary operations for (W, ∆) that transforms
the empty bracket system into the final system and has size with the desired upper
bound.
24
S. V. IVANOV
First we observe that every bracket b = (b(1), . . . , b(6)) for (W2 , ∆2 ) naturally
gives rise to a bracket bb = (bb(1), . . . , bb(6)) for (W, ∆). Specifically, define
bb := (b(1) + |p1 |, b(2) + |p1 |, b(3), b(4), b(5), b(6)).
b2,j denote the bracket system obtained from B2,j by replacing every bracket
Let B
b2,0 , . . . , B
b2,ℓ2 is a sequence of bracket systems for (W, ∆)
b ∈ B2,j with bb. Then B
that changes the empty bracket system into
b2,ℓ2 = {(|p1 |, |p1 | + |p2 |, 0, 0, |∆2 (2)|)}.
B
Define a relation on the set of pairs (b, i), where b ∈ B1,i , that is the reflexive
and transitive closure of the relation (c, i + 1) ≻ (b, i), where c ∈ B1,i+1 is obtained
from b, b′ ∈ B1,i by an elementary operation σ, where b′ could be missing. It follows
from the definitions that is a partial order on the set of such pairs (b, i) and that
if (b2 , i2 ) (b1 , i1 ) then i2 ≥ i1 and
b2 (1) ≤ b1 (1) ≤ b1 (2) ≤ b2 (2).
Note that the converse need not hold, e.g., if b1 is a starting bracket and i1 = i2 ,
b1 (1) = b1 (2) = b2 (1), b1 6= b2 , then the above inequalities hold but (b2 , i2 ) 6
(b1 , i1 ).
Now we observe that every bracket d = (d(1), . . . , d(6)), d ∈ B1,i , for (W1 , ∆1 )
naturally gives rise to a bracket
b
b
db = (d(1),
. . . , d(6))
for (W, ∆) in the following fashion.
If (d, i) is not comparable with (c0 , 0), where c0 is defined in (4.12), by the
relation and d(1) ≤ d(2) ≤ |p1 |, then
db := d.
If (d, i) is not comparable with (c0 , 0) by the relation and |p1 | ≤ d(1) ≤ d(2),
then
db := (d(1) + |p2 |, d(2) + |p2 |, d(3), d(4), d(5), d(6)).
If (d, i) (c0 , 0), then
db := (d(1), d(2) + |p2 |, d(3), d(4), d(5), d(6) + |∆2 (2)|).
Note that the above three cases cover all possible situations because if (d, i)
is not comparable with (c0 , 0) by the relation , then d(2) ≤ |p1 | = c0 (1) or
d(1) ≥ |p1 | = c0 (2).
b1,i := {db | d ∈ B1,i }. Then B
b1,0 , . . . , B
b1,ℓ1 is a sequence of bracket
As above, let B
systems for (W, ∆) that changes the bracket system
b1,0 = B
b2,ℓ2 = {(|p1 |, |p1 p2 |, 0, 0, 0, |∆2 (2)|)}
B
into the final bracket system
b1,ℓ1 = (0, |p1 | + |p2 | + |p3 |, 0, 0, 0, |∆1(2)| + |∆2 (2)|).
B
b1,0 , . . . , B
b1,ℓ1 is an operational
More specifically, it is straightforward to verify that B
b 1 of the
sequence of bracket systems for (W, ∆) which corresponds to an analogue Ω
sequence of elementary operations Ω1 for (W1 , ∆1 ) so that if a bracket b ∈ B1,i ,
i ≥ 1, is obtained from brackets d1 , d2 ∈ B1,i−1 , where one of d1 , d2 could be missing,
BOUNDED AND PRECISE WORD PROBLEMS FOR GROUP PRESENTATIONS
25
b1,i is obtained from db1 , db2 ∈ B
b1,i−1
by an elementary operation σ of Ω1 , then bb ∈ B
b 1 and σ
by an elementary operation σ
b of Ω
b has the same type as σ.
Thus, with the indicated changes, we can now combine the foregoing sequences
of bracket systems for (W2 , ∆2 ) and for (W1 , ∆1 ) into a single sequence of bracket
systems for (W, ∆) that transforms the empty bracket system into the bracket
system {(|p1 |, |p1 p2 |, 0, 0, 0, |∆2(2)|)} and then continues to turn the latter into the
final bracket system. It follows from definitions and bounds (4.11)–(4.13) that the
size of thus constructed sequence is bounded by
(11|W1 | + 11|W2 | + 1, max(C(log |W1 | + 1) + 1, C(log |W2 | + 1)))
Therefore, in view of Lemma 4.4, it remains to show that
max(C(log |W1 | + 1) + 1, C(log |W2 | + 1)) ≤ C(log |W | + 1).
In view of inequality (4.10),
max(C(log |W1 | + 1) + 1, C(log |W2 | + 1)) ≤ C(log( 56 |W |) + 1) + 1,
and C(log( 56 |W |) + 1) + 1 ≤ C(log |W | + 1) if C ≥ (log 56 )−1 . Thus the first main
case, when Lemma 4.2(a) holds for the pair (W, ∆), is complete.
Now assume that Lemma 4.2(b) holds true for the pair (W, ∆) and Π is the
face in ∆ with |∂Π| ≥ 2, v1 , v2 ∈ PW are the vertices of PW such that v1 < v2 ,
α(v1 ), α(v2 ) ∈ ∂Π, and if PW (fact, v1 , v2 ) = p1 p2 p3 then
min(|p2 |, |p1 | + |p3 |) ≥ 16 |W |.
~
Furthermore, if (∂Π)−1 = e1 . . . e|∂Π| , where ei ∈ ∆(1),
and the cyclic boundary
∂∆ of ∆ is ∂∆ = e1 h1 . . . e|∂Π| h|∂Π| , where hi is a closed subpath of ∂∆, then, for
every i, hi is a subpath of either α(p2 ) or α(p3 )α(p1 ) and |hi | ≤ 56 |W |.
For i = 2, . . . , |∂Π|, denote Wi := ϕ(hi ) and let ∆i be the disk subdiagram of
∆ with ∂|(hi )− ∆i = hi . We also consider a path PWi with ϕ(PWi ) ≡ Wi and let
αi : PWi → ∆i denote the map whose definition is analogous to that of α : PW → ∆
(note that |Wi | = 0 is now possible).
By the induction hypothesis on |W | applied to (Wi , ∆i ) (the case |Wi | = 0
is vacuous), there is a sequence Ωi of elementary operations for (Wi , ∆i ) that
transforms the empty bracket system into the final one and has size bounded by
(10|Wi |, C(log |Wi | + 1)).
Making a cyclic permutation of indices of ei , hi in ∂∆ = e1 h1 . . . e|∂Π| h|∂Π| if
necessary, we may assume that
PW = s2 f2 q2 f3 q3 . . . f|∂Π| q|∂Π| f1 s1 ,
where α(fi ) = ei , i = 1, . . . , |∂Π|, α(qj ) = hj , j = 2, . . . , |∂Π|, and α(s1 )α(s2 ) = h1 ,
see Fig. 4.2. Note that |s1 | = 0 or |s2 | = 0 is possible.
Let Bi,0 , . . . , Bi,ℓi denote the corresponding to Ωi sequence of bracket systems,
where Bi,0 is empty and Bi,ℓi is final. As in the above arguments, we can easily
convert every bracket b ∈ ∪j Bi,j , where i > 1, into a bracket bb for (W, ∆) by using
the rule
bb := (b(1) + |s2 f2 . . . fi |, b(2) + |s2 f2 . . . fi |, b(3), b(4), b(5), b(6)).
bi,j := {bb | b ∈ Bi,j }, becomes an operbi,0 , . . . , B
bi,ℓi , where B
Then the sequence B
ational sequence of bracket systems for (W, ∆) that transforms the empty bracket
26
S. V. IVANOV
bi,ℓi = {dbi }, where
system into B
dbi = (|s2 f2 . . . fi |, |s2 f2 . . . fi qi |, 0, 0, 0, |∆i (2)|)}, i > 1.
bi,0 , . . . , B
bi,ℓi corresponds to
We also remark that the sequence of bracket systems B
b i of the sequence of elementary operations Ωi for (Wi , ∆i ) so that
an analogue Ω
if a bracket b ∈ Bi,j , j ≥ 1, is obtained from brackets g1 , g2 ∈ Bi,j−1 , where one
bi,j is
of g1 , g2 could be missing, by an elementary operation σ of Ωi , then bb ∈ B
b
b
obtained from b
g1 , b
g2 ∈ Bi,j−1 by an elementary operation σ
b of Ωi and σ
b has the
same type as σ.
Π
b
b
Denote ϕ((∂Π)−1 ) = aεn
jΠ . Using the operational sequence B2,0 , . . . , B2,ℓ2 , we
b
convert the empty bracket system into {d2 }. Applying a turn operation to db2 ,
we change db2 (3) from 0 to jΠ and d2 (4) from 0 to nΠ . Note that nΠ ≤ |W | by
Lemma 2.2, hence this is correct to do so. Then we apply two extensions of type 1
on the left and on the right to increase db2 (2) by 1 and to decrease db2 (1) by 1, see
Fig. 4.2. Let
de2 = (|s2 |, |s2 f2 q2 f3 |, jΠ , nΠ , 2ε, |∆2 (2)|)
denote the bracket of type B2 obtained this way. Now, starting with the bracket
system {de2 }, we apply those elementary operations that are used to create the
b3,ℓ3 , and obtain the bracket system {de2 , db3 }. Applying a merger
b3,0 , . . . , B
sequence B
to de2 , db3 , we get
db′ = (|s2 |, |s2 f2 q2 f3 q3 |, jΠ , nΠ , 2ε, |∆2 (2)| + |∆3 (2)|).
3
Let de3 be obtained from db′3 by extension of type 1 on the right, so
de3 = (|s2 |, |s2 f2 q2 f3 q3 f4 |, jΠ , nΠ , 3ε, |∆2 (2)| + |∆3 (2)|).
Iterating in this manner, we will arrive at a bracket system consisting of the single
bracket
X
db′|∂Π| = (|s2 |, |s2 f2 . . . q|∂Π| |, jΠ , nΠ , ε(|∂Π| − 1),
|∆i (2)|).
i≥2
Applying to db′|∂Π| an extension of type 2 on the right along the edge f1 = α(e1 ),
see Fig. 4.2, we obtain the bracket
X
de|∂Π| = (|s2 |, |W | − |s1 |, 0, 0, 0, 1 +
|∆i (2)|)
i≥2
of type B1.
For i = 1, we let W1 := ϕ(s2 )ϕ(s1 ) and let ∆1 be the disk subdiagram of ∆ with
∂|(s2 )− ∆1 = s2 s1 . Referring to the induction hypothesis on |W | for (W1 , ∆1 ) and
to Lemma 4.5, we conclude that there is a sequence Ω′1 of elementary operations
that changes the starting bracket
c2 := (|s2 |, |s2 |, 0, 0, 0, 0)
(4.14)
into the final bracket system and has size bounded by
(11|W1 | + 1, C(log |W1 | + 1) + 1).
′
′
′
Let B1,0
, . . . , B1,ℓ
′ denote the corresponding to Ω1 sequence of bracket systems,
1
′
b1,i
where B1,0
= {c2 } and B ′ ′ is final. Similarly to the above construction of B
1,ℓ1
′
from B1,i , we will make changes over brackets b ∈ ∪j B1,j
so that every b becomes a
BOUNDED AND PRECISE WORD PROBLEMS FOR GROUP PRESENTATIONS
27
b ′ := {bb | b ∈ B ′ }, then the sequence B
b′ , . . . , B
b′ ′
bracket bb for (W, ∆) and if B
1,0
1,i
1,i
1,ℓ1
′
b1,0
transforms the bracket system B
= {b
c2 } = {de|∂Π| } into the final one for (W, ∆).
′
Specifically, define a relation ′ on the set of all pairs (b, i), where b ∈ B1,i
, that
′
is reflexive and transitive closure of the relation (c, i + 1) ≻ (b, i), where a bracket
c ∈ B1,i+1 is obtained from brackets b, b′ ∈ B1,i by an elementary operation σ,
where b′ could be missing. As before, it follows from the definitions that ′ is a
partial order on the set of all such pairs (b, i) and that if (b2 , i2 ) ′ (b1 , i1 ) then
i2 ≥ i1 and
b2 (1) ≤ b1 (1) ≤ b1 (2) ≤ b2 (2).
′
Furthermore, every bracket b = (b(1), . . . , b(6)), where b ∈ B1,i
, for (W1 , ∆1 ),
naturally gives rise to a bracket
bb = (bb(1), . . . , bb(6))
for (W, ∆) in the following fashion.
If the pair (b, i) is not comparable with (c2 , 0), where c2 is defined in (4.14), by
the relation ′ and b(2) ≤ |s2 |, then
bb := b.
If the (b, i) is not comparable with (c2 , 0) by the relation ′ and |s1 | ≤ b(1), then
bb := (b(1) + |W | − |h1 |, b(2) + |W | − |h1 |, b(3), b(4), b(5), b(6)).
If (d, i) ′ (c2 , 0), then
bb := (b(1), b(2) + |W | − |h1 |, b(3), b(4), b(5), b(6) + 1 +
X
|∆i (2)|).
i≥2
As before, we note that these three cases cover all possible situations because if
(b, i) is not comparable with (c2 , 0) by the relation ′ , then either
b(2) ≤ |W | − |h1 | = c2 (1)
or b(1) ≥ |W | − |h1 | = c2 (2).
b′ , . . . , B
b′
Similarly to the foregoing arguments, we check that B
1,0
1,ℓ1 is an operational sequence of bracket systems for (W, ∆) which corresponds to an analogue
b ′1 of the sequence of elementary operations Ω′1 for (W1 , ∆1 ) so that if a bracket
Ω
′
′
b ∈ B1,i
, i ≥ 1, is obtained from brackets g1 , g2 ∈ B1,i−1
, where one of g1 , g2 could
′
b
b ′ is obtained from
be missing, by an elementary operation σ of Ω1 , then b ∈ B
1,i
b′
b ′1 and σ
b
has
the
same type as σ.
g1 , gb2 ∈ B
b
by
an
elementary
operation
σ
b
of
Ω
1,i−1
Summarizing, we conclude that there exists a sequence of elementary operations
b 2, . . . , Ω
b |∂Π| , that transforms the empty bracket sysΩ, containing subsequences Ω
b′ ,
tem for (W, ∆) into the bracket system B = {de|∂Π| } and then, via subsequence Ω
1
continues to transform
B = {de|∂Π| } = {b
c2 }
into the final bracket system for (W, ∆). It follows from the induction hypothesis
for the pairs (Wi , ∆i ) and definitions that the size of thus constructed sequence Ω
is bounded by
X
|Wi |) + 2|∂Π| + 2, max {C(log |Wi | + 1) + 1}).
((11·
1≤i≤|∂Π|
1≤i≤|∂Π|
28
S. V. IVANOV
In view of Lemma 4.4, it suffices to show that
max {C(log |Wi | + 1) + 1} ≤ C(log |W | + 1).
1≤i≤|∂Π|
Since |Wi | ≤ 56 |W |, the latter inequality holds true, as in the above case, for C =
(log 65 )−1 .
Let W be an arbitrary nonempty word over the alphabet A±1 , not necessarily
representing the trivial element of the group given by presentation (1.2) (and W is
not necessarily reduced). As before, let PW be a labeled simple path with ϕ(PW ) ≡
W and let vertices of PW be identified along PW with integers 0, . . . , |W | so that
(PW )− = 0, . . . , (PW )+ = |W |.
A 6-tuple
b = (b(1), b(2), b(3), b(4), b(5), b(6))
of integers b(1), b(2), b(3), b(4), b(5), b(6) is called a pseudobracket for the word W
if b(1), b(2) satisfy the inequalities 0 ≤ b(1) ≤ b(2) ≤ |W | and one of the following
two properties (PB1), (PB2) holds for b.
(PB1) b(3) = b(4) = b(5) = 0 and 0 ≤ b(6) ≤ |W |.
(PB2) b(3) ∈ {1, . . . , m}, b(4) ∈ Eb(3) , where the set Eb(3) is defined in (1.2),
0 < b(4) ≤ |W |, −b(4) < b(5) < b(4), and 0 ≤ b(6) ≤ |W |.
We say that a pseudobracket b has type PB1 if the property (PB1) holds for b.
We also say that a pseudobracket b has type PB2 if the property (PB2) holds for
b. Clearly, these two types are mutually exclusive.
Let p denote the subpath of PW such that p− = b(1) and p+ = b(2), perhaps,
p− = p+ and |p| = 0. The subpath p of PW is denoted a(b) and is called the arc of
the pseudobracket b.
For example, bv = (v, v, 0, 0, 0, 0) is a pseudobracket of type PB1 for every vertex
v ∈ PW and such bv is called a starting pseudobracket. Note that a(b) = {v}. A
final pseudobracket for W is c = (0, |W |, 0, 0, 0, k), where k ≥ 0 is an integer. Note
that a(c) = PW .
Observe that if b is a bracket for the pair (W, ∆) of type B1 or B2, then b is also
a pseudobracket of type PB1 or PB2, resp., for the word W .
Let B be a finite set of pseudobrackets for W , perhaps, B is empty. We say
that B is a pseudobracket system for W if, for every pair b, c ∈ B of distinct
pseudobrackets, either b(2) ≤ c(1) or c(2) ≤ b(1). It follows from the definitions
that every bracket system for (W, ∆) is also a pseudobracket system for the word
W . B is called a final pseudobracket system for W if B consists of a single final
pseudobracket for W .
Now we describe four kinds of elementary operations over pseudobrackets and
over pseudobracket systems: additions, extensions, turns, and mergers, which are
analogous to those definitions for brackets and bracket systems, except there are
no any diagrams and no faces involved.
Let B be a pseudobracket system for a nonempty word W over A±1 .
Additions.
Suppose b is a starting pseudobracket, b 6∈ B, and B ∪ {b} is a pseudobracket
system. Then we may add b to B thus making an addition operation over B.
Extensions.
BOUNDED AND PRECISE WORD PROBLEMS FOR GROUP PRESENTATIONS
29
Suppose b ∈ B, b = (b(1), b(2), b(3), b(4), b(5), b(6)), and e1 a(b)e2 is a subpath of
PW , where a(b) is the arc of b and e1 , e2 are edges one of which could be missing.
1
Assume that b is of type PB2. Suppose ϕ(e1 ) = aεb(3)
, where ε1 = ±1.
If |b(5)| ≤ b(4) − 2 and ε1 b(5) ≥ 0, then we consider a pseudobracket b′ of type
PB2 such that
b′ (1) = b(1) − 1, b′ (2) = b(2), b′ (3) = b(3),
b′ (4) = b(4), b′ (5) = b(5) + ε1 , b′ (6) = b(6).
Note that a(b′ ) = e1 a(b). We say that b′ is obtained from b by an extension of type
1 (on the left). If (B \ {b}) ∪ {b′ } is a pseudobracket system, then replacement of
b ∈ B with b′ in B is called an extension operation over B of type 1.
On the other hand, if |b(5)| = b(4) − 1 and ε1 b(5) ≥ 0, then we consider a
pseudobracket b′ of type PB1 such that
b′ (1) = b(1) − 1, b′ (2) = b(2), b′ (3) = b′ (4) = b′ (5) = 0, b′ (6) = b(6) + 1.
In this case, we say that b′ is obtained from b by an extension of type 2 (on the left).
Note that a(b′ ) = e1 a(b) and b′ has type PB1. If (B \ {b}) ∪ {b′ } is a pseudobracket
system, then replacement of b ∈ B with b′ in B is called an extension operation
over B of type 2.
2
, where ε2 = ±1.
Analogously, assume that b has type PB2 and ϕ(e2 ) = aεb(3)
If |b(5)| ≤ b(4) − 2 and ε2 b(5) ≥ 0, then we consider a pseudobracket b′ such that
b′ (1) = b(1), b′ (2) = b(2) + 1, b′ (3) = b(3),
b′ (4) = b(4), b′ (5) = b(5) + ε2 , b′ (6) = b(6).
Note that a(b′ ) = a(b)e2 and b′ has type PB2. We say that b′ is obtained from b by
an extension of type 1 (on the right). If (B \ {b}) ∪ {b′ } is a pseudobracket system,
then replacement of b ∈ B with b′ in B is called an extension operation over B of
type 1.
On the other hand, if |b(5)| = b(4) − 1 and ε2 b(5) ≥ 0, then we consider a
pseudobracket b′ such that
b′ (1) = b(1), b′ (2) = b(2) + 1, b′ (3) = b′ (4) = b′ (5) = 0, b′ (6) = b(6) + 1.
Note that a(b′ ) = a(b)e2 and b′ has type PB1. We say that b′ is obtained from b by
an extension of type 2 (on the right). If (B \ {b}) ∪ {b′ } is a pseudobracket system,
then replacement of b ∈ B with b′ in B is called an extension operation over B of
type 2.
Assume that b ∈ B is a pseudobracket of type PB1, e1 a(b)e2 is a subpath of PW ,
where a(b) is the arc of b and e1 , e2 are edges with ϕ(e1 ) = ϕ(e2 )−1 . Consider a
pseudobracket b′ of type PB1 such that
b′ (1) = b(1) − 1, b′ (2) = b(2) + 1, b′ (3) = b′ (4) = b′ (5) = 0, b′ (6) = b(6).
Note that a(b′ ) = e1 a(b)e2 . We say that b′ is obtained from b by an extension of
type 3. If (B \ {b}) ∪ {b′ } is a pseudobracket system, then replacement of b ∈ B
with b′ in B is called an extension operation over B of type 3.
Turns.
Let b ∈ B be a pseudobracket of type PB1. Then, by the definition, b(3) = b(4) =
b(5) = 0. Pick j ∈ {1, . . . , m} such that Ej 6= {0}. Consider a pseudobracket b′
with b′ (i) = b(i) for i = 1, 2, 4, 5, 6, and b′ (3) = j. Note that b′ is of type PB2 and
30
S. V. IVANOV
a(b′ ) = a(b). We say that b′ is obtained from b by a turn operation. Replacement of
b ∈ B with b′ in B is also called a turn operation over B. Note that (B \ {b}) ∪ {b′ }
is a pseudobracket system because so is B.
Mergers.
Now suppose that b, c ∈ B are distinct pseudobrackets such that b(2) = c(1)
and one of b(3), c(3) is 0. Then one of b, c is of type PB1 and the other has type
PB1 or PB2. Consider a pseudobracket b′ such that b′ (1) = b(1), b′ (2) = c(2), and
b′ (i) = b(i) + c(i) for i = 3, 4, 5, 6. Note that a(b′ ) = a(b)a(c) and b′ is of type PB1
if both b, c have type PB1 or b′ is of type PB2 if one of b, c has type PB2. We say
that b′ is obtained from b, c by a merger operation. Taking both b, c out of B and
putting b′ in B is a merger operation over B.
As before, we will say that additions, extensions, turns and mergers, as defined
above, are elementary operations over pseudobrackets and pseudobracket systems
for W .
Assume that one pseudobracket system Bℓ is obtained from another pseudobracket system B0 by a finite sequence Ω of elementary operations and B0 , B1 , . . . , Bℓ
is the corresponding to Ω sequence of pseudobracket systems. Such a sequence
B0 , B1 , . . . , Bℓ of pseudobracket systems is called operational. We say that a sequence B0 , B1 , . . . , Bℓ of pseudobracket systems has size bounded by (k1 , k2 ) if ℓ ≤ k1
and, for every i, the number of pseudobrackets in Bi is at most k2 . Whenever it
is not ambiguous, we also say that Ω has size bounded by (k1 , k2 ) if so does the
corresponding to Ω sequence B0 , B1 , . . . , Bℓ of pseudobracket systems.
The significance of pseudobracket systems and elementary operations over pseudobracket systems introduced above is revealed in the following lemma.
Lemma 4.7. Suppose that the empty pseudobracket system B0 for W can be transformed by a finite sequence Ω of elementary operations into a final pseudobracket
G
system {bF }. Then W =2 1, i.e., W represents the trivial element of the group
defined by presentation (1.2), and there is a disk diagram ∆ over (1.2) such that
∆ has property (A), ϕ(∂|0 ∆) ≡ W and |∆(2)| = bF (6).
Proof. Let B0 , B1 , . . . , Bℓ be the corresponding to Ω sequence of pseudobracket
systems, where B0 is empty and Bℓ is final. Consider the following Claims (C1)–
(C2) for a pseudobracket b ∈ Bi , i = 1, . . . , ℓ, in which a(b) denotes the arc of b.
G
(C1) If a pseudobracket b ∈ Bi has type PB1 then ϕ(a(b)) =2 1 in the group G2
given by presentation (1.2) and there is a disk diagram ∆b over (1.2) such
that ∆b has property (A), ϕ(∂∆b ) ≡ ϕ(a(b)) and |∆b (2)| = b(6).
G
b(5)
(C2) If a pseudobracket b ∈ Bi has type PB2, then ϕ(a(b)) =2 ab(3) and there is
a disk diagram ∆b over (1.2) such that ∆b has property (A), ∂∆b = pq −1 ,
b(5)
where p, q −1 are subpaths of ∂∆b , ϕ(p) ≡ ϕ(a(b)), ϕ(q) ≡ ab(3) , |∆b (2)| =
b(6), and if e is an edge of q −1 then e−1 ∈ p.
By induction on i ≥ 1, we will prove that Claims (C1)–(C2) hold true for every
pseudobracket b′ ∈ Bi . Note that if b′ is a starting pseudobracket, then b′ is of type
PB1 and Claim (C1) is obviously true for b′ (Claim (C2) is vacuously true for b′ ).
Since B1 consists of a starting pseudobracket, the base of induction is established.
BOUNDED AND PRECISE WORD PROBLEMS FOR GROUP PRESENTATIONS
31
To make the induction step from i − 1 to i, i ≥ 2, we consider the cases corresponding to the type of the elementary operation that is used to get Bi from
Bi−1 .
Suppose that Bi is obtained from Bi−1 by an elementary operation σ and b′ ∈ Bi
is the pseudobracket obtained from b, c ∈ Bi−1 by application of σ, here one of b, c
or both, depending on type of σ, could be missing. By the induction hypothesis,
Claims (C1)–(C2) hold for every pseudobracket of Bi different from b′ and it suffices
to show that the suitable Claim (C1)–(C2) holds for b′ .
If Bi is obtained from Bi−1 by an addition, then it suffices to refer to the above
remark that Claims (C1)–(C2) hold for a starting pseudobracket.
Suppose that Bi is obtained from Bi−1 by an extension of type 1 and let b′ ∈ Bi
be the pseudobracket created from b ∈ Bi−1 by an extension of type 1 on the left
(the “right” subcase is symmetric).
Note that both b and b′ have type PB2. By the induction hypothesis, which is
Claim (C2) for b, there exists a disk diagram ∆b such that ∆b has property (A),
b(5)
∂∆b = pq −1 , ϕ(p) ≡ ϕ(a(b)), ϕ(q) ≡ ab(3) , |∆b (2)| = b(6),
and if e is an edge of q −1 then e−1 ∈ p. Here and below we use the notation of the
definition of an extension of type 1.
1
, ε1 = ±1.
Let e1 denote the edge of PW such that a(b′ ) = e1 a(b) and ϕ(e1 ) = aεb(3)
ε1
Consider a “loose” edge f with ϕ(f ) = ab(3) . We attach the vertex f+ to the vertex
p− = q− of ∆b to get a new disk diagram ∆′b such that ∂∆′b = f pq −1 f −1 , see
Fig. 4.5. Note that property (A) holds for ∆′b ,
1
ϕ(a(b)) ≡ ϕ(e1 a(b)) ≡ ϕ(a(b′ )),
ϕ(f p) ≡ aεb(3)
b(5)+ε1
ϕ(q −1 f −1 ) ≡ ab(3)
b′ (5)
≡ ab(3) ,
|∆′b (2)| = |∆b (2)| = b(6) = b′ (6), and if e is an edge of q −1 then e−1 ∈ p. Thus,
Claim (C2) holds for the pseudobracket b′ with ∆b′ := ∆′b .
p
∆′b
∆b
f
q
Fig. 4.5
Assume that Bi is obtained from Bi−1 by an extension of type 2 and let b′ ∈ Bi
be the pseudobracket obtained from b ∈ Bi−1 by an extension of type 2 on the left
(the “right” subcase is symmetric).
Note that b has type PB2, while b′ has type PB1. By the induction hypothesis,
which is Claim (C2) for b, there exists a disk diagram ∆b such that ∆b has property
(A),
b(5)
∂∆b = pq −1 , ϕ(p) ≡ ϕ(a(b)), ϕ(q) ≡ ab(3) , |∆b (2)| = b(6),
and if e is an edge of q −1 then e−1 ∈ p. Here and below we use the notation of the
definition of an extension of type 2.
1
, ε1 = ±1.
Let e1 denote the edge of PW such that a(b′ ) = e1 a(b) and ϕ(e1 ) = aεb(3)
ε1
Consider a “loose” edge f with ϕ(f ) = ab(3) . We attach the vertex f+ to p− = q−
32
S. V. IVANOV
and the vertex f− to p+ = q+ of ∆b . Since ε1 b(5) ≥ 0 and |b(5)| = b(4) − 1, it
ε1 b(4)
and we can attach a
follows that ε1 + b(5) = ε1 |b(4)|. Therefore, ϕ(qf ) ≡ ab(3)
−ε b(4)
so that the boundary path ∂Π is identified with
new face Π with ϕ(∂Π) ≡ ab(3)1
the path q −1 f −1 , see Fig. 4.6. This way we obtain a new disk diagram ∆′b such
that ϕ(∂|f− ∆′b ) ≡ ϕ(f )ϕ(p) and |∆′b (2)| = |∆b (2)| + 1, see Fig. 4.6. It follows from
construction of ∆′b that property (A) holds for ∆′b
G
b(5)+ε1
1
ϕ(a(b)) ≡ ϕ(e1 a(b)) ≡ ϕ(a(b′ )) =2 1, ϕ(q −1 f −1 ) ≡ ab(3)
ϕ(f p) ≡ aεb(3)
b′ (5)
≡ ab′ (3) ,
and |∆′b (2)| = |∆b (2)| + 1 = b(6) + 1 = b′ (6). Therefore, Claim (C1) holds for the
pseudobracket b′ if we set ∆b′ := ∆′b .
∆′b
p
q
∆b
Π
f
Fig. 4.6
Suppose that an extension of type 3 was applied to get Bi from Bi−1 and b′ ∈ Bi
is the pseudobracket created from b ∈ Bi−1 by the operation.
Note that both b and b′ have type PB1. By the induction hypothesis, which is
Claim (C1) for b, there exists a disk diagram ∆b such that ∆b has property (A)
and
−b(5)
ϕ(∂∆b ) ≡ ϕ(a(b))ab(3) , |∆b (2)| = b(6).
Here and below we use the notation of the definition of an extension of type 3.
Denote ∂∆b = p, where ϕ(p) ≡ ϕ(a(b)) and let e1 , e2 be the edges of PW such
that a(b′ ) = e1 a(b)e2 and ϕ(e1 ) = ϕ(e2 )−1 . Consider a “loose” edge f with ϕ(f ) =
ϕ(e1 ). We attach the vertex f+ to the vertex p− of ∆b to get a new disk diagram
∆′b such that ∂∆′b = f pf −1 , see Fig. 4.7. Since ∆b has property (A),
ϕ(∂∆′b ) ≡ ϕ(f pf −1 ) ≡ ϕ(e1 )ϕ(a(b))ϕ(e1 )−1 ≡ ϕ(a(b′ ))
and |∆′b (2)| = |∆b (2)| = b(6) = b′ (6), it follows that Claim (C1) holds for the
pseudobracket b′ with ∆b′ := ∆′b .
p
∆b
f
∆′b
Fig. 4.7
Suppose that Bi is obtained from Bi−1 by a turn operation and let b′ ∈ Bi be
the pseudobracket created from b ∈ Bi−1 by the turn operation. By the definition
of a turn, b has type PB1 and b′ has type PB2. By the induction hypothesis, which
is Claim (C1) for b, there exists a disk diagram ∆b such that ∆b has property (A)
ϕ(∂∆b ) ≡ ϕ(a(b)),
|∆b (2)| = b(6).
BOUNDED AND PRECISE WORD PROBLEMS FOR GROUP PRESENTATIONS
33
Here and below we use the notation of the definition of a turn operation.
Denote ∆′b := ∆b and let ∂∆′b := pq −1 , where p, q −1 are subpaths of ∂∆′b such
that ϕ(p) ≡ ϕ(a(b)) and |q| = 0. Clearly, ∆′b has property (A). Since a(b) = a(b′ ),
b′ (5) = b(5) = 0, and
|∆′b (2)| = |∆b (2)| = b(6) = b′ (6),
b′ (5)
it follows that ϕ(p) ≡ ϕ(a(b′ )), ϕ(q) ≡ ab′ (3) . Hence, Claim (C2) holds for the
pseudobracket b′ with ∆b′ := ∆′b .
Finally, assume that Bi results from Bi−1 by a merger operation and let b′ ∈ Bi
be the pseudobracket created from pseudobrackets b, c ∈ Bi−1 , where b(2) = c(1),
by the operation. By the definition of a merger operation, one of the pseudobrackets
b, c must have type PB1. By the induction hypothesis, there are disk diagrams
∆b , ∆c for b, c, resp., as stated in Claims (C1)–(C2). Denote ∂∆b = pb qb−1 , ∂∆c =
pc qc−1 , where ϕ(pb ) ≡ ϕ(a(b)), ϕ(pc ) ≡ ϕ(a(c)), and, for x ∈ {b, c}, |qx | = 0 if x has
x(5)
type PB1 or ϕ(qx ) ≡ ax(3) if x has type PB2. Note that |∆x (2)| = x(6).
Consider a disk diagram ∆′ obtained from ∆b , ∆c by identification of the vertices
(pb )+ and (pc )− , see Fig. 4.8. Then
|∆′ (2)| = |∆b (2)| + |∆c (2)|,
∂∆′ = pb pc qc−1 qb−1 .
Note that
ϕ(pb pc ) ≡ ϕ(a(b))ϕ(a(c)) ≡ ϕ(a(b′ )),
|∆′ (2)| = b(6) + c(6)
b(5)
and |qc−1 qb−1 | = 0 if both b, c have type PB1, or ϕ(qb qc ) ≡ ab(3) if b has type PB2
c(5)
and c has type PB1, or ϕ(qb qc ) ≡ ac(3) if b has type PB1 and c has type PB2.
Therefore, it follows that Claim (C1) holds true for b′ with ∆b′ = ∆′ if both b, c
have type PB1 or Claim (C2) holds for b′ with ∆b′ = ∆′ if one of b, c has type PB2.
pb
∆′
pc
∆b
∆c
qb
qc
Fig. 4.8
All possible cases are discussed and the induction step is complete. Claims
(C1)–(C2) are proven.
We can now finish the proof of Lemma 4.7. By Claim (C1) applied to the
pseudobracket bF = (0, |W |, 0, 0, 0, bF (6)) of the final pseudobracket system Bℓ =
{bF }, there is a disk diagram ∆bF over (1.2) such that ∆bF has property (A),
ϕ(∂|0 ∆bF ) ≡ W and |∆bF (2)| = bF (6). Thus, ∆bF is a desired disk diagram and
Lemma 4.7 is proven.
Lemma 4.8. Suppose W is a nonempty word over A±1 and n ≥ 0 is an integer.
Then W is a product of at most n conjugates of words R±1 , where R = 1 is a
relation of presentation (1.2), if and only if there is a sequence Ω of elementary
operations such that Ω transforms the empty pseudobracket system for W into a
34
S. V. IVANOV
final pseudobracket system {bF }, where bF (6) ≤ n, and Ω has size bounded by
(11|W |, C(log |W | + 1)), where C = (log 56 )−1 .
Proof. Assume that W is a product of at most n conjugates of words R±1 , where
R = 1 is a relation of presentation (1.2). By Lemmas 2.1–2.2(b), there is a
disk diagram ∆ over (1.2) such that ∆ has property (A), ϕ(∂|0 ∆) ≡ W and
|∆(2)| ≤ min(n, |W |). Then Lemma 4.6 applies and yields a sequence Ω of elementary operations over bracket systems for (W, ∆) that converts the empty bracket
system for (W, ∆) into the final one and has size bounded by (11|W |, C(log |W |+1)).
It follows from arguments of Lemma 4.6 and Lemma 2.2(a) that if b is a bracket of
one of the intermediate bracket systems, associated with Ω, then b(4) ≤ |W |. Observe that every bracket b ∈ B and every intermediate bracket system B for (W, ∆),
associated with Ω, could be considered as a pseudobracket and a pseudobracket
system for W , resp., we automatically have a desired sequence of pseudobracket
systems.
Conversely, the existence of a sequence Ω of elementary operations over pseudobracket systems, as specified in Lemma 4.8, implies, by Lemma 4.7, that there
is a disk diagram ∆ over (1.2) such that ∆ has property (A), ϕ(∂|0 ∆) ≡ W and
|∆(2)| = bF (6) ≤ n. The existence of such a disk diagram ∆ means that W is a
product of |∆(2)| = bF (6) ≤ n conjugates of words R±1 , where R = 1 is a relation
of presentation (1.2).
5. Proofs of Theorem 1.2 and Corollary 1.3
Theorem 1.2. Let the group G2 be defined by a presentation of the form
G2 := h a1 , . . . , am k aki i = 1, ki ∈ Ei , i = 1, . . . , m i,
(1.2)
where for every i, one of the following holds: Ei = {0} or, for some integer ni > 0,
Ei = {ni } or Ei = ni N = {ni , 2ni , 3ni , . . . }. Then both the bounded and precise
word problems for (1.2) are in L3 and in P. Specifically, the problems can be solved
in deterministic space O((log |W |)3 ) or in deterministic time O(|W |4 log |W |).
Proof. We start with L3 part of Theorem 1.2. First we discuss a nondeterministic
algorithm which solves the bounded word problem for presentation (1.2) and which
is based on Lemma 4.8.
Given an input (W, 1n ), where W is a nonempty word (not necessarily reduced)
over the alphabet A±1 and n ≥ 0 is an integer, written in unary notation as 1n ,
we begin with the empty pseudobracket system and nondeterministically apply
a sequence of elementary operations of size bounded by (11|W |, C(log |W | + 1)),
where C = (log 65 )−1 . If such a sequence of elementary operations results in a
final pseudobracket system {(0, |W |, 0, 0, 0, n′)}, where n′ ≤ n, then our algorithm
accepts and, in view of Lemma 4.8, we may conclude that W is a product of n′ ≤ n
conjugates of words R±1 , where R = 1 is a relation of (1.2). It follows from the
definitions and Lemma 4.8 that the number of elementary operations needed for
this algorithm to accept is at most 11|W |. Hence, it follows from the definition of
elementary operations over pseudobracket systems for W that the time needed to
run this nondeterministic algorithm is O(|W |). To estimate the space requirements
of this algorithm, we note that if b is a pseudobracket for W , then b(1), b(2) are
integers in the range from 0 to |W |, hence, when written in binary notation, will
take at most C ′ (log |W |+1) space, where C ′ is a constant. Since b(3), b(4), b(5), b(6)
BOUNDED AND PRECISE WORD PROBLEMS FOR GROUP PRESENTATIONS
35
are also integers that satisfy inequalities
0 ≤ b(3) ≤ m, 0 ≤ b(4) ≤ |W |, |b(5)| ≤ b(4), 0 ≤ b(6) ≤ |W |,
and m is a constant, it follows that the total space required to run this algorithm
is at most
5C ′ (log |W | + 1 + log m)C(log |W | + 1) = O((log |W |)2 ).
Note that this bound is independent of n because it follows from Lemma 2.2(b)
that if W is a product of n′ conjugates of words R±1 , where R = 1 is a relation of
(1.2), then it is possible to assume that n′ ≤ |W |.
Furthermore, according to Savitch’s theorem [38], see also [3], [32], the existence
of a nondeterministic algorithm that recognizes a language in space S and time
T implies the existence of a deterministic algorithm that recognizes the language
in space O(S log T ). Therefore, by Savitch’s theorem [38], there is a deterministic
algorithm that solves the bounded word problem for presentation (1.2) in space
O((log |W |)3 ).
To solve the precise word problem for presentation (1.2), suppose that we are
given a pair (W, 1n ) and wish to find out if W is a product of n conjugates of
words R±1 , where R = 1 is a relation of (1.2), and n is minimal with this property.
By Lemma 2.2, we may assume that n ≤ |W |. Using the foregoing deterministic
algorithm, we consecutively check whether the bounded word problem is solvable
for the two pairs (W, 1n−1 ) and (W, 1n ). It is clear that the precise word problem
for the pair (W, 1n ) has a positive solution if and only if the bounded word problem
has a negative solution for the pair (W, 1n−1 ) and has a positive solution for the pair
(W, 1n ) and that these two facts can be verified in deterministic space O((log |W |)3 ).
Now we describe an algorithm that solves the precise word problem for presentation (1.2) in polynomial time. Our arguments are analogous to folklore arguments
[35], [36] that solve the precise word problem for presentation h a, b k a = 1, b = 1 i
in polynomial time and that are based on the method of dynamic programming.
For a word U over A±1 consider the following property.
(E) U is nonempty and if b1 , b2 ∈ A±1 are the first, last, resp., letters of U then
b1 6= b−1
2 and b1 6= b2 .
Lemma 5.1. Let ∆ be a disk diagram over presentation (1.2) such that ∆ has
property (A) and the word W ≡ ϕ(∂|0 ∆) has property (E). Then there exists a
factorization ∂|0 ∆ = q1 q2 such that |q1 |, |q2 | > 0 and (q1 )+ = (q2 )− = α(0). In
G
G
particular, ϕ(q1 ) =2 1, ϕ(q2 ) =2 1.
Proof. Let ∂|0 ∆ = e1 qe2 , where e1 , e2 are edges, q is a subpath of ∂|0 ∆.
−1
Suppose e−1
1 is an edge of ∂|0 ∆. Since e1 6= e2 by property (E) of W , it follows
−1
−1
that e1 is an edge of q and q = r1 e1 r2 , where r1 , r2 are subpaths of q. Hence,
q1 = e1 r1 e−1
1 and q2 = r2 e2 are desired paths.
Hence, we may assume that e−1
1 is an edge of the boundary path ∂Π1 of a face Π1 .
Arguing in a similar fashion, we may assume that e−1
2 is an edge of the boundary
path ∂Π2 of a face Π2 . If Π1 = Π2 then, in view of relations of the presentation
(1.2), we have ϕ(e1 ) = ϕ(e2 ) contrary to property (E) of W . Hence, Π1 6= Π2 .
Since for every edge f of ∂Π1 the edge f −1 belongs to ∂∆ by property (A),
we obtain the existence of a desired path q1 of the form q1 = e1 if |∂Π1 | = 1 or
36
S. V. IVANOV
q1 = e1 rf −1 if |∂Π1 | > 1, where r is a subpath of ∂|0 ∆, f is the edge of ∂Π1 such
that e−1
1 f is a subpath of ∂Π1 , f− = α(0), and |q1 | < |∂∆|.
G
If U =2 1, let µ2 (U ) denote the integer such that the precise word problem for
G2
presentation (1.2) holds for the pair (U, µ2 (U )). If U 6= 1, we set µ2 (U ) := ∞.
G
Lemma 5.2. Let U be a word such that U =2 1 and U has property (E). Then there
exists a factorization U ≡ U1 U2 such that |U1 |, |U2 | > 0 and
µ2 (U ) = µ2 (U1 ) + µ2 (U2 ).
Proof. Consider a disk diagram ∆ over (1.2) such that ϕ(∂|0 ∆) ≡ U and |∆(2)| =
µ2 (U ). By Lemma 2.2(b), we may assume that ∆ has property (A). Hence, it
follows from Lemma 5.1 applied to ∆ that there is a factorization ∂|0 ∆ = q1 q2
such that |q1 |, |q2 | > 0 and (q1 )+ = (q2 )− = α(0). Denote Ui := ϕ(qi ) and let
∆i be the subdiagram of ∆ bounded by qi , i = 1, 2. Since ∆ is a minimal disk
diagram for U , it follows that ∆i is a minimal disk diagram for Ui , i = 1, 2. Hence,
|∆i (2)| = µ2 (Ui ) and
µ2 (U ) = |∆(2)| = |∆1 (2)| + |∆2 (2)| = µ2 (U1 ) + µ2 (U2 ),
as required.
Let U be a nonempty word over A±1 and let U (i, j), where 1 ≤ i ≤ |U | and
0 ≤ j ≤ |U |, denote the subword of U that starts with the ith letter of U and has
length j. For example, U = U (1, |U |) and U (i, 0) is the empty subword.
in
If aj ∈ A, let |U |aj denote the total number of occurrences of letters aj , a−1
j
G
U . Note that we can decide whether U =2 1 in time O(|U |2 ) by cancelling subwords
±n
−1
aj j , a−1
j aj , aj aj , and checking whether the word obtained by a process of such
cancellations is empty.
Let W be a nonempty word over A±1 . Define a parameterized word W [i, j, k, ℓ],
where 1 ≤ i ≤ |W |, 0 ≤ j ≤ |W |, 1 ≤ k ≤ m, and ℓ is an integer that satisfies
|ℓ| ≤ |W |ak − |W (i, j)|ak so that
W [i, j, k, ℓ] := W (i, j)aℓk
(5.1)
and the word W (i, j)aℓk is not empty, i.e., j + |ℓ| ≥ 1.
Note that the total number of such parameterized words W [i, j, k, ℓ] is bounded
by O(|W |3 ), here m = |A| is a constant as the presentation (1.2) is fixed. Let
S2 (W ) denote the set of all parameterized words W [i, j, k, ℓ]. Elements W [i, j, k, ℓ]
and W [i′ , j ′ , k ′ , ℓ′ ] of S2 (W ) are defined to be equal if and only if the quadruples
(i, j, k, ℓ), (i′ , j ′ , k ′ , ℓ′ ) are equal. Hence, we wish to distinguish between elements
of S2 (W ) and actual words represented by elements of S2 (W ). It is clear that we
′
can have W (i, j)aℓk ≡ W (i′ , j ′ )aℓk′ when W [i, j, k, ℓ] 6= W [i′ , j ′ , k ′ , ℓ′ ].
If U is the word represented by W [i, j, k, ℓ], i.e., U ≡ W (i, j)aℓk , then we denote
this by writing
⋆
U = W [i, j, k, ℓ].
We introduce a partial order on the set S2 (W ) by setting
W [i′ , j ′ , k ′ , ℓ′ ] ≺ W [i, j, k, ℓ]
BOUNDED AND PRECISE WORD PROBLEMS FOR GROUP PRESENTATIONS
37
if j ′ < j or j ′ = j and |ℓ′ | < |ℓ|. In other words, we partially order the set S2 (W ) by
using the lexicographical order on pairs (j, |ℓ|) associated with elements W [i, j, k, ℓ]
of S2 (W ).
Define
µ2 (W [i, j, k, ℓ]) := µ2 (W (i, j)aℓk ).
To compute the number µ2 (W ) = µ2 (W [1, |W |, 1, 0]) in polynomial time, we use
the method of dynamic programming in which the parameter is (j, |ℓ|). In other
words, we compute the number µ2 (W [i, j, k, ℓ]) by induction on parameter (j, |ℓ|).
The base of induction (or initialization) for j + |ℓ| = 1 is obvious as µ2 (W [i, j, k, ℓ])
is 0 or 1 depending on the presentation (1.2).
To make the induction step, we assume that the numbers µ2 (W [i′ , j ′ , k ′ , ℓ′ ]) are
already computed for all W [i′ , j ′ , k ′ , ℓ′ ] whenever (j ′ , |ℓ′ |) ≺ (j, |ℓ|) and we compute
the number µ2 (W [i, j, k, ℓ]).
G2
If W (i, j)aℓk 6= 1, we set µ2 (W [i′ , j ′ , k ′ , ℓ′ ]) := ∞.
G
Assume that W (i, j)aℓk =2 1.
Suppose that the word W (i, j)aℓk has property (E). Consider all possible factorizations for W (i, j)aℓk of the form
W (i, j)aℓk ≡ U1 U2 ,
(5.2)
where |U1 |, |U2 | ≥ 1. Let us show that either of the words U1 , U2 can be represented
in a form W [i′ , j ′ , k ′ , ℓ′ ] so that W [i′ , j ′ , k ′ , ℓ′ ] ≺ W [i, j, k, ℓ].
First suppose |U1 | ≤ |W (i, j)| = j. If i + |U1 | ≤ |W |, we have
⋆
U1 = W [i, |U1 |, 1, 0],
⋆
U2 = W [i + |U1 |, j − |U1 |, k, ℓ].
On the other hand, if i + |U1 | = |W | + 1, we have
⋆
U1 = W [i, |U1 |, 1, 0],
⋆
U2 = W [|W |, 0, k, ℓ].
Now suppose |U1 | > |W (i, j)| = j. Let ℓ1 , ℓ2 be integers such that ℓ1 , ℓ2 , ℓ have
the same sign, ℓ = ℓ1 + ℓ2 , and j + |ℓ| = |U1 | + ℓ2 . Then we have
⋆
U1 = W [i, |U1 |, k, ℓ1 ],
⋆
U2 = W [1, 0, k, ℓ2].
Note that the induction parameter (j ′ , |ℓ′ |) for indicated representations of U1 , U2
is smaller than the parameter (j, |ℓ|) for the original parameterized word W [i, j, k, ℓ].
It follows from Lemma 5.2 applied to the word W (i, j)aℓk that there is a factorization
W (i, j)aℓk ≡ U1′ U2′
such that |U1′ |, |U2′ | ≥ 1 and µ2 (W (i, j)aℓk ) = µ2 (U1′ ) + µ2 (U2′ ). Hence, taking the
minimum
min(µ2 (U1 ) + µ2 (U2 ))
(5.3)
over all factorizations (5.2), we obtain the number µ2 (W (i, j)aℓk ).
Assume that the word W (i, j)aℓk has no property (E), i.e., W (i, j)aℓk ≡ bU b−1 or
W (i, j)aℓk ≡ b or W (i, j)aℓk ≡ bU b, where b ∈ A±1 and U is a word.
First suppose that W (i, j)aℓk ≡ bU b−1 . Note that the word U can be represented
in a form W [i′ , j ′ , k ′ , ℓ′ ] so that W [i′ , j ′ , k ′ , ℓ′ ] ≺ W [i, j, k, ℓ]. Indeed, if |ℓ| = 0 then
⋆
U = W [i + 1, j − 2, k, 0].
38
S. V. IVANOV
On the other hand, if |ℓ| > 0 then
⋆
U = W [i + 1, j − 1, k, ℓ1 ],
where |ℓ1 | = |ℓ| − 1 and ℓ1 ℓ ≥ 0. Hence, the number µ2 (U ) is available by induction hypothesis. Since µ2 (W [i, j, k, ℓ]) = µ2 (U ), we obtain the required number
µ2 (W [i, j, k, ℓ]).
In case W (i, j)aℓk ≡ b, the number µ2 (W [i, j, k, ℓ]) is either one or ∞ depending
on the presentation (1.2).
Finally, consider the case when W (i, j)aℓk ≡ bU b. Denote b = aδk1 , where ak1 ∈ A
and δ = ±1.
If j = 0, i.e., W (i, j)aℓk ≡ aℓk , then k = k1 and the number µ2 (W [i, j, k, ℓ])
can be easily computed in time O(log |W |) as |ℓ| ≤ |W | and we can use binary
representation for ℓ.
⋆
Suppose j > 0. If ℓ = 0 then U b2 = W [i + 1, j − 1, k1 , δ]. If |ℓ| > 0 then k = k1 ,
⋆
ℓ and δ have the same sign, and U b2 = W [i + 1, j − 1, k, ℓ + δ].
In either subcase j = 0 or j > 0, we obtain
⋆
U b2 = W [i′ , j ′ , k ′ , ℓ′ ],
W [i′ , j ′ , k ′ , ℓ′ ] ≺ W [i, j, k, ℓ],
µ2 (W [i, j, k, ℓ]) = µ2 (W [i′ , j ′ , k ′ , ℓ′ ]).
This completes our inductive procedure of computation of numbers µ2 (W [i, j, k, ℓ])
for all W [i, j, k, ℓ] ∈ S2 (W ).
Since the length of every word W (i, j)aℓk is at most |W |, it follows that our
computation of the number µ2 (W [i, j, k, ℓ]) can be done in time O(|W | log |W |)
including additions of binary representations of numbers µ2 (U1 ), µ2 (U2 ) to compute
the minimum (5.3). Since the cardinality |S2 (W )| of the set S2 (W ) is bounded by
O(|W |3 ), we conclude that computation of numbers µ2 (W [i, j, k, ℓ]) for all words
W [i, j, k, ℓ] ∈ S2 (W ) can be done in deterministic time O(|W |4 log |W |). This means
that both the bounded word problem and the precise word problem for presentation
(1.2) can be solved in time O(|W |4 log |W |).
The proof of Theorem 1.2 is complete.
Corollary 1.3. Let W be a word over A±1 and n ≥ 0 be an integer. Then the
decision problems that inquire whether the width h(W ) or the spelling length h1 (W )
of W is equal to n belong to L3 and P. Specifically, the problems can be solved in
deterministic space O((log |W |)3 ) or in deterministic time O(|W |4 log |W |).
Proof. Observe that the decision problem asking whether the width (resp. the
spelling length) of a word W is n is equivalent to the precise word problem whose
input is (W, n) for presentation (1.2) in which Ei = N (resp. Ei = {1}) for every
i = 1, . . . , m. Therefore, a reference to proven Theorem 1.2 shows that the problem
asking whether the width (resp. the spelling length) of W is n belongs to both L3
and P and yields the required space and time bounds for this problem.
6. Calculus of Brackets for Group Presentation (1.4)
As in Theorem 1.4, consider a group presentation
G3 = h a1 , a2 , . . . , am k a2 an1 1 a2−1 = an1 2 i,
(1.4)
±1
where n1 , n2 are nonzero integers. Suppose W is a nonempty word over A such
G
that W =3 1, n ≥ 0 is an integer and W is a product of at most n conjugates
BOUNDED AND PRECISE WORD PROBLEMS FOR GROUP PRESENTATIONS
39
−n2 ±1
of the words (a2 an1 1 a−1
) . Then, as follows from arguments of the proof
2 a1
of Lemma 2.1, there exists a disk diagram ∆ over presentation (1.4) such that
ϕ(∂|0 ∆) ≡ W and |∆(2)| ≤ n. Without loss of generality, we may assume that ∆
is reduced. Consider the word ρa1 (W ) over A±1 \ {a1 , a−1
1 } which is obtained from
W by erasing all occurrences of a±1
letters.
1
An oriented edge e of ∆ is called an ai -edge if ϕ(e) = a±1
i . We also consider a
tree ρa1 (∆) obtained from ∆ by contraction of all a1 -edges of ∆ into points and
subsequent identification of all pairs of edges e, f such that e− = f− and e+ = f+ . If
an edge e′ of ρa1 (∆) is obtained from an edge e of ∆ this way, we set ϕ(e′ ) := ϕ(e).
It is easy to check that this definition is correct. This labeling function turns the
tree ρa1 (∆) into a disk diagram over presentation h a2 , . . . , am k ∅ i.
Two faces Π1 , Π2 in a disk diagram ∆ over (1.4) are termed related, denoted
Π1 ↔ Π2 , if there is an a2 -edge e such that e ∈ ∂Π1 and e−1 ∈ ∂Π2 . Consider the
minimal equivalence relation ∼2 on the set of faces of ∆ generated by the relation
↔. An ai -band, where i > 1, is a minimal subcomplex Γ of ∆ (Γ is not necessarily
simply connected) that contains an ai -edge f , and, if there is a face Π in ∆ with
f ∈ (∂Π)±1 , then Γ must contain all faces of the equivalence class [Π]∼2 of Π.
Hence, an ai -band Γ is either a subcomplex consisting of a single nonoriented edge,
denoted {f, f −1 }, where f is an ai -edge, i > 1, and f, f −1 ∈ ∂∆, or Γ consists of all
faces of an equivalence class [Π]∼2 when i = 2. Clearly, the latter case is possible
only if i = 2. In this latter case, if Γ contains faces but Γ has no face in [Π]∼2 whose
boundary contains an a2 -edge f with f −1 ∈ ∂∆, Γ is called a closed a2 -band.
Lemma 6.1. Suppose that ∆ is a reduced disk diagram over presentation (1.4).
Then there are no closed a2 -bands in ∆ and every ai -band Γ of ∆, i > 1, is a disk
subdiagram of ∆ such that
∂|(f1 )− Γ = f1 s1 f2 s2 ,
where f1 , f2 are edges of ∂∆ with ϕ(f1 ) = ϕ(f2 )−1 = ai , s1 , s2 are simple paths
−kn2
1
for some integer k ≥ 0, Γ contains k faces, and
with ϕ(s1 ) ≡ akn
1 , ϕ(s2 ) ≡ a1
∂Γ is a simple closed path when k > 0.
Proof. Since ∆ is reduced, it follows that if two faces Π1 , Π2 are related by the
relation ∼, then Π1 is not a mirror copy of Π2 , i.e., ϕ(Π1 ) 6≡ ϕ(Π2 )−1 .
Assume that there is a closed a2 -band Γ0 in ∆. Then it follows from the above
remark that there is a disk subdiagram ∆0 of ∆, surrounded by Γ0 , such that
ϕ(∆0 ) ≡ a1k0 ni , where i = 1 or i = 2 and k0 = |Γ0 (2)| > 0. This, however, is a
contradiction to the Magnus’s Freiheitssatz, see [18], [19]. Recall that the Magnus’s
Freiheitssatz for a one-relator group presentation
G0 = h A k R = 1 i
±1
claims that every letter a ∈ A , that appears in a cyclically reduced word R over
G
A±1 , will also appear in W ±1 whenever W is reduced and W =0 1.
Hence, no a2 -band Γ in ∆ is closed and, therefore, ∂Γ = f1 s1 f2 s2 as described
in Lemma’s statement. If k > 0 and ∂Γ is not a simple closed path, then a proper
subpath of ∂Γ bounds a disk subdiagram ∆1 such that ϕ(∆1 ) is a proper subword of
1 −1 −kn2
ϕ(∂Γ) ≡ a2 akn
, where k > 0. Then it follows from Lemma 2.1 that either
1 a2 a1
G
′
G
aℓ1 =3 1, where ℓ 6= 0, or aℓ1 aε2 =3 1, where ε = ±1. The first equality contradicts
Magnus’s Freiheitssatz and the second one is impossible in the abelianization of
G3 .
40
S. V. IVANOV
The factorization ∂|(f1 )− Γ = f1 s1 f2 s2 of Lemma 6.1 will be called the standard
boundary of an ai -band Γ, where i > 1. Note that |s1 | = |s2 | = 0 is the case when
k = 0, i.e., Γ contains no faces and ∂Γ = f1 f2 with f2−1 = f1 .
An alternative way of construction of the tree ρa1 (∆) from ∆ can now be described as contraction of the paths s1 , s2 of the standard boundary of every a2 -band
Γ of ∆ into points, identification of the edges f1 , f2−1 of (∂Γ)±1 , and contraction
into points of all a1 -edges e with e ∈ ∂∆.
Let s be a path in a disk diagram ∆0 over presentation (1.4). We say that s is
a special arc of ∆0 if either |s| = 0 or |s| > 0, s is a reduced simple path, and s
consists entirely of a1 -edges. In particular, if s is a special arc then every subpath
of s, s−1 is also a special arc, and s− 6= s+ whenever |s| > 0.
Lemma 6.2. Suppose that ∆0 is a reduced disk diagram over presentation (1.4)
and s, t are special arcs in ∆0 such that s+ = t− . Then there are factorizations
s = s1 s2 and t = t1 t2 such that s2 = t−1
1 and s1 t2 is a special arc, see Fig. 6.1.
s2 = t−1
1
s1
t2
Fig. 6.1
Proof. First we observe that if ∆1 is a disk subdiagram of ∆0 such that every edge
of ∂∆1 is an a1 -edge, then it follows from Lemma 6.1 that ∆1 contains no faces,
i.e., |∆1 (2)| = 0, and so ∆1 is a tree.
We now prove this Lemma by induction on the length |t| ≥ 0. If |t| = 0 then our
claim is true with s1 = s and t2 = t. Assume that |t| > 0 and t = re, where e is an
edge. Since r is also a special arc with |r| < |t|, the induction hypothesis applies to
the pair s, r and yields factorizations s = s′1 s′2 and r = r1 r2 such that s′2 = (r1 )−1
and s′1 r2 is a special arc. If s′1 r2 e is also a special arc, then the factorizations
s = s′1 s′2 and t = t1 t2 , where t1 = r1 and t2 = r2 e, have the desired property and
the induction step is complete.
Hence, we may assume that s′1 r2 e is not a special arc. Note that if |r2 | > 0, then
the path s′1 r2 e is reduced because s′1 r2 and r2 e are special arcs. Assume that s′1 r2 e
is not reduced. Then |r2 | = 0 and the last edge of s′1 is e−1 . Denote s′1 = s′11 e−1 .
Then, letting s1 := s′11 , s2 := t−1 and t = t1 t2 , where t1 := t and |t2 | = 0, we
have s2 = t−1 and s1 t2 is a desired special arc. Thus, we may assume that the
path s′1 r2 e is reduced but not simple. Note that the path s′1 r2 e consists entirely of
a1 -edges. Since s′1 r2 and r2 e are simple, it follows that e+ belongs to s′1 and defines
a factorization s′1 = s11 s12 , where |s12 | > 0. Hence, the path s12 r2 e is a simple
closed path which bounds a disk subdiagram ∆1 whose boundary ∂∆1 consists of
a1 -edges. By the observation made above, ∆1 is a tree, which, in view of the fact
that ∂∆1 is a simple closed path, implies that ∂∆1 = e−1 e. Therefore, |r2 | = 0
and s12 = e−1 . This means that the path s′1 r2 e is not reduced, contrary to our
assumption. This contradiction completes the proof.
Let U be a word over A±1 . If a ∈ A is a letter, then the a-length |U |a of U is
the number of occurrences of a and a−1 in U . We also define the complementary
a-length by |U |ā := |U | − |U |a .
BOUNDED AND PRECISE WORD PROBLEMS FOR GROUP PRESENTATIONS
41
From now on we assume, unless stated otherwise, that W is a nonempty word
G
over A±1 such that W =3 1 and ∆ is a reduced disk diagram over presentation (1.4)
such that ϕ(∂|0 ∆) ≡ W .
The definitions of the path PW , of the map α : PW → ∆ and those of related
notions, given in Sect. 2, are retained and are analogous to those used in Sects. 4–5.
Lemma 6.3. Suppose that |W |ā1 > 2. Then there exist vertices v1 , v2 ∈ PW such
that v1 < v2 and if PW (fact, v1 , v2 ) = p1 p2 p3 , then the following hold true. There is
a special arc r in ∆ such that r− = α(v1 ), r+ = α(v2 ), ϕ(r) ≡ aℓ1 for some ℓ, and
min(|ϕ(p2 )|ā1 , |ϕ(p1 )|ā1 + |ϕ(p3 )|ā1 ) ≥ 31 |ϕ(∂|0 ∆)|ā1 = 31 |W |ā1 .
~
In addition, if |∆(1)|
a1 := |W |a1 +(|n1 |+|n2 |)|∆(2)| denotes the number of a1 -edges
~
of ∆, then |r| ≤ 21 |∆(1)|
a1 .
Proof. Consider the tree ρa1 (∆) constructed from ∆ as above by collapsing a1 -edges
e into points and subsequent identification of edges f, g with f− = g− , f+ = g+ .
Then ρa1 (∆) is a disk diagram over (1.4) with no faces and there is a surjective
continuous cellular map
β1 : ∆ → ρa1 (∆)
which preserves labels of edges e with ϕ(e) 6= a±1
1 and sends a1 -edges into vertices.
It is also clear that if
α1 : Pρa1 (W ) → ρa1 (∆)
is the corresponding to the pair (ρa1 (W ), ρa1 (∆)) map, then there is a cellular
continuous surjective map
β : PW → Pρa1 (W )
which preserves labels of edges e with ϕ(e) 6= a±1
1 , sends a1 -edges into vertices, and
makes the following diagram commutative.
PW
yβ
Pρa1 (W )
α
−−−−−−−−→
α
∆
yβ1
−−−−−−1−−−→ ρa1 (W )
By Lemma 4.1 applied to ρa1 (∆), there are vertices u1 , u2 ∈ Pρa1 (W ) such that
α1 (u1 ) = α1 (u2 ) and if Pρa1 (W ) (fact, u1 , u2 ) = s1 s2 s3 , then
min(|s2 |, |s1 | + |s3 |) ≥ 31 |∂ρa1 (∆)| = 31 |W |ā1 .
(6.1)
It follows from definitions and Lemma 6.1 that there are some vertices v1 , v2 ∈ PW
such that β(vj ) = uj , j = 1, 2, the vertices α(v1 ), α(v2 ) belong to ∂Γ, where Γ is
an ai -band in ∆, i > 1, and the vertices α(v1 ), α(v2 ) can be connected along ∂Γ
by a simple reduced path r such that ϕ(r) ≡ aℓ1 for some integer ℓ, where ℓ = 0 if
i > 2. Clearly, r is a special arc of ∆. Denote PW (fact, v1 , v2 ) = p1 p2 p3 . Then it
follows from definitions that |ϕ(pi )|ā1 = |si |, i = 1, 2, 3, hence the inequality (6.1)
yields that
min(|ϕ(p2 )|ā1 , |ϕ(p1 )|ā1 + |ϕ(p3 )|ā1 ) ≥ 31 |∂ρa1 (∆)| = 13 |W |ā1 ,
as required.
42
S. V. IVANOV
Furthermore, since r is simple, reduced and r− 6= r+ unless |r| = 0, the path r
contains every a1 -edge e of ∆ at most once and if r contains e then r contains no
~
e−1 . Since the total number |∆(1)|
a1 of a1 -edges of ∆ is equal to
~
|∆(1)|
a1 = |ϕ(∂∆)|a1 + (|n1 | + |n2 |)|∆(2)|,
we obtain the desired inequality
1
1
~
|r| = |ϕ(r)|a1 ≤ 21 |∆(1)|
a1 = 2 |W |a1 + 2 (|n1 | + |n2 |)|∆(2)|.
A quadruple
b = (b(1), b(2), b(3), b(4)),
of integers b(1), b(2), b(3), b(4) is called a bracket for the pair (W, ∆) if b(1), b(2)
satisfy the inequalities 0 ≤ b(1) ≤ b(2) ≤ |W |, and, in the notation
PW (fact, b(1), b(2)) = p1 p2 p3 ,
the following holds true. There exists a special arc r(b) in ∆ such that r(b)− =
0
b(3)
α(b(1)), r(b)+ = α(b(2)), and ϕ(r(b)) = a1 . Furthermore, if ∆b is the disk
subdiagram of ∆ defined by ∂|b(1) ∆b = α(p2 )r(b)−1 (such ∆b is well defined as r(b)
is a special arc in ∆), then |∆b (2)| = b(4), see Fig. 6.2.
α(p2 )
∆b
α(b(1))
α(b(2))
r(b)
α(p1 )
α(0)
α(p3 )
Fig. 6.2
This disk subdiagram ∆b of ∆, defined by ∂|b(1) ∆b = α(p2 )r(b)−1 , is associated
with the bracket b. The boundary subpath α(p2 ) of ∆b is denoted a(b) and called
the arc of the bracket b. The path r(b) is termed the complementary arc of b.
For example, bv = (v, v, 0, 0) is a bracket for every vertex v of PW , called a
starting bracket at v = b(1). Note that a(b) = α(v), r(b) = α(v), and ∆b = {α(v)}.
The final bracket for (W, ∆) is c = (0, |W |, 0, |∆(2)|). Observe that a(c) = ∂|0 ∆,
r(c) = {α(0)} and ∆c = ∆.
Let B be a finite set of brackets for the pair (W, ∆), perhaps, B is empty. We say
that B is a bracket system for (W, ∆) if, for every pair b, c ∈ B of distinct brackets,
either b(2) ≤ c(1) or c(2) ≤ b(1).
Now we describe three kinds of elementary operations over a bracket system B
for the pair (W, ∆): additions, extensions, and mergers.
Additions.
Suppose b is a starting bracket, b 6∈ B, and B ∪ {b} is a bracket system. Then
we may add b to B thus making an addition operation over B.
Extensions.
BOUNDED AND PRECISE WORD PROBLEMS FOR GROUP PRESENTATIONS
43
Suppose B is a bracket system, b ∈ B is a bracket and e1 a(b)e2 is a subpath
of the boundary path ∂|0 ∆, where a(b) is the arc of b and e1 , e2 are edges one of
which could be missing.
Assume that ϕ(e1 ) = aε1 , where ε = ±1. Since e1 and r(b) are special arcs of ∆
and e1 r(b) is a path, Lemma 6.2 applies and yields that either e1 r(b) is a special
arc or r(b) = e−1
1 r1 , where r1 is a subpath of r(b) and r1 is a special arc. In the first
case, define r(b′ ) := e1 r(b). In the second case, we set r(b′ ) := r1 . Note that, in
0 b(3)+ε
either case, r(b′ ) is a special arc and ϕ(r(b′ )) = a1
. In either case, we consider
a bracket b′ such that
b′ (1) = b(1) − 1, b′ (2) = b(2), b′ (3) = b(3) + ε, b′ (4) = b(4).
In either case, we have that a(b′ ) = e1 a(b), r(b′ ) is defined as above, and ∆b′ is the
disk subdiagram whose boundary is ∂∆b′ = a(b′ )r(b′ )−1 . We say that b′ is obtained
from b by an extension of type 1 (on the left). If (B \ {b}) ∪ {b′} is a bracket system,
then replacement of b ∈ B with b′ in B is called an extension operation over B of
type 1.
Similarly, assume that ϕ(e2 ) = aε1 , where ε = ±1. Since r(b) is a simple path, it
follows as above from Lemma 6.2 that either r(b)e2 is a special arc or r(b) = r2 e−1
2 ,
where r2 is a subpath of r(b) and r2 is a special arc. In the first case, define
r(b′ ) := r(b)e2 . In the second case, we set r(b′ ) := r2 . Note that, in either case,
0 b(3)+ε
r(b′ ) is a special arc and ϕ(r(b′ )) = a1
. In either case, we consider a bracket b′
such that
b′ (1) = b(1), b′ (2) = b(2) + 1, b′ (3) = b(3) + ε, b′ (4) = b(4).
In either case, we have that a(b′ ) = a(b)e2 , r(b′ ) is defined as above, and ∆b′ is the
disk subdiagram whose boundary is ∂∆b′ = a(b′ )r(b′ )−1 . We say that b′ is obtained
from b by an extension of type 1 (on the right). If (B \ {b}) ∪ {b′ } is a bracket
system, then replacement of b ∈ B with b′ in B is called an extension operation
over B of type 1.
Now suppose that ϕ(e1 ) = ϕ(e2 )−1 = aε2 , ε = ±1, and there is an a2 -bond Γ
whose standard boundary is ∂|(ei )− Γ = ei si e3−i s3−i , where i = 1 if ε = 1 and i = 2
if ε = −1. Recall that the standard boundary of an a2 -bond starts with an edge f
with ϕ(f ) = a2 .
0
First we assume that |b(3)| 6= 0, i.e., ϕ(si ) 6= 1. By Lemma 6.1, the paths
s1 , s2 are special arcs. Moreover, Γ consists of |b(3)|/|ni | faces. Recall that ∆
is reduced. It follows from Lemma 6.2 applied to special arcs r(b) and s−1
that
i
r(b) = si . Consider a bracket b′ such that b′ (1) = b(1) − 1, b′ (2) = b(2) + 1,
ε
′
′
b′ (3) = (|n−1
1 n2 |) b(3), and b (4) = b(4) + |b(3)|/|ni |. Note that a(b ) = e1 a(b)e2
−1
′
and r(b) = si+1 . We say that b is obtained from b by an extension of type 2. If
(B \ {b}) ∪ {b′} is a bracket system, then replacement of b ∈ B with b′ in B is called
an extension operation over B of type 2.
±1
′
Assume that e1 = e−1
2 and ϕ(e1 ) 6= a1 . Then we may consider a bracket b such
′
′
′
′
that b (1) = b(1) − 1, b (2) = b(2) + 1, b (3) = b(3), and b (4) = b(4). Since the path
G
a(b) is closed, it follows from Lemma 2.1 that ϕ(a(b)) =3 1 and so b′ (3) = b(3) = 0
by Magnus’s Freiheitssatz. Note that a(b′ ) = e1 a(b)e2 and |r(b′ )| = 0. We say that
b′ is obtained from b by an extension of type 3. If (B \ {b}) ∪ {b′ } is a bracket
system, then replacement of b ∈ B with b′ in B is called an extension operation
over B of type 3.
44
S. V. IVANOV
Mergers.
Suppose that b1 , b2 are distinct brackets in B such that b1 (2) = b2 (1). Consider
the disk diagram ∆bi , associated with the bracket bi , i = 1, 2, and let ∂|bi (1) ∆bi =
a(bi )r(bi )−1 be the boundary of ∆bi , where a(bi ), r(bi ) are the arcs of bi , i = 1, 2.
Since r(b1 ), r(b2 ) are special arcs with r(b1 )− = r(b2 )+ , it follows from Lemma 6.2
that there are factorizations r(b1 ) = r1 r0 , r(b2 ) = r0−1 r2 such that the path r(b′ ) :=
r1 r2 is a special arc of ∆. Note that r(b′ )− = α(b1 (1)), r(b′ )+ = α(b2 (2)), and the
disk subdiagram ∆0 of ∆ defined by ∂∆0 = a(b1 )a(b2 )r(b′ )−1 contains b1 (4) + b2 (4)
faces. Therefore, we may consider a bracket b′ such that b′ (1) = b1 (1), b′ (2) = b2 (2),
b′ (3) = b1 (3) + b2 (3), and b′ (4) = b1 (4) + b2 (4). Note that a(b′ ) = a(b1 )a(b2 ),
r(b′ ) = r1 r2 , and ∆b′ = ∆0 . We say that b′ is obtained from b1 , b2 by a merger
operation. If (B \ {b1 , b2 }) ∪ {b′ } is a bracket system, then taking both b1 , b2 out
of B and putting b′ in B is a merger operation over B.
We will say that additions, extensions, and mergers, as defined above, are elementary operations over brackets and bracket systems for (W, ∆).
Assume that one bracket system Bℓ is obtained from another bracket system
B0 by a finite sequence Ω of elementary operations and B0 , B1 , . . . , Bℓ is the
corresponding to Ω sequence of bracket systems. As before, such a sequence
B0 , B1 , . . . , Bℓ of bracket systems is called operational. We say that the sequence
B0 , B1 , . . . , Bℓ has size bounded by (k1 , k2 ) if ℓ ≤ k1 and, for every i, the number of
brackets in Bi is at most k2 . Whenever it is not ambiguous, we also say that Ω has
size bounded by (k1 , k2 ) if so does the corresponding to Ω sequence B0 , B1 , . . . , Bℓ
of bracket systems.
We now study properties of brackets and bracket systems.
~
Lemma 6.4. If b is a bracket for the pair (W, ∆), then |b(3)| ≤ 21 |∆(1)|
a1 .
Proof. By the definition of a bracket, the complementary arc r(b) of b is a special
arc in ∆, hence, either |r(b)| = 0 or, otherwise, |r(b)| > 0 and r(b) is a simple,
reduced path consisting entirely of a1 -edges. Since the total number of a1 -edges in
1 ~
~
∆ is |∆(1)|
a1 , it follows that |b(3)| ≤ |r(b)| ≤ 2 |∆(1)|a1 .
Lemma 6.5. Suppose that b, c are two brackets for the pair (W, ∆) and b(1) = c(1),
b(2) = c(2). Then b = c.
Proof. Consider the complementary arcs r(b), r(c) of brackets b, c, resp. Since
r(b)− = r(c)− , r(b)+ = r(c)+ , and r(b), r(c)−1 are special arcs of ∆, it follows from
Lemma 6.2 applied to special arcs r(b), r(c)−1 that r(b) = r(c). This means that
b(3) = c(3), ∆b = ∆c and b(4) = c(4). Therefore, b = c, as desired.
Lemma 6.6. There exists a sequence of elementary operations that converts the
empty bracket system for (W, ∆) into the final bracket system and has size bounded
by (4|W |, |W |).
Proof. For every k with 0 ≤ k ≤ |W |, consider a starting bracket (k, k, 0, 0) for
(W, ∆). Making |W | + 1 additions, we get a bracket system
BW = {(k, k, 0, 0) | 0 ≤ k ≤ |W |}
of |W | + 1 starting brackets. Now, looking at ∆, we can easily find a sequence
of extensions and mergers that converts BW into the final bracket system Bℓ . To
estimate the total number of elementary operations, we note that the number of
BOUNDED AND PRECISE WORD PROBLEMS FOR GROUP PRESENTATIONS
45
additions is |W | + 1. The number of extensions is at most |W | because every
extension applied to a bracket system B increases the number
X
η(B) :=
(b(2) − b(1))
b∈B
by 1 or 2 and η(BW ) = 0, while η(Bℓ ) = |W |. The number of mergers is |W |
because the number of brackets |B| in a bracket system B decreases by 1 if B is
obtained by a merger and |BW | = |W | + 1, |Bℓ | = 1. Hence, ℓ ≤ 3|W | + 1 ≤ 4|W |,
as required.
Lemma 6.7. Suppose there is a sequence Ω of elementary operations that converts
the empty bracket system E for (W, ∆) into the final bracket system F and has size
bounded by (k1 , k2 ). Then there is also a sequence of elementary operations that
changes E into F and has size bounded by (7|W |, k2 ).
Proof. Assume that the sequence Ω has an addition operation which introduces a
starting bracket c = (k, k, 0, 0) with 0 ≤ k ≤ |W |. Since the final bracket system
contains no starting bracket, c must disappear and an elementary operation is
applied to c. If a merger is applied to c and another bracket b and the merger
yields b
c, then b
c = b. This means that the addition of c and the merger could be
skipped without affecting the sequence otherwise. Note that the size of the new
sequence Ω′ is bounded by (k1 − 2, k2 ). Therefore, we may assume that no merger
is applied to a starting bracket in Ω. We also observe that, by the definition, an
extension of type 2 is not applicable to a starting bracket (because of the condition
b(3) 6= 0).
Thus, by an obvious induction on k1 , we may assume that, for every starting
bracket c which is added by Ω, an extension of type 1 or 3 is applied to c.
Now we will show that, for every starting bracket c which is added by Ω, there
are at most 2 operations of addition of c in Ω. Arguing on the contrary, assume
that c1 , c2 , c3 are brackets equal to c whose additions are used in Ω. By the above
remark, for every i = 1, 2, 3, an extension of type 1 or 3 is applied to ci , resulting
in a bracket b
ci .
Let c1 , c2 , c3 be listed in the order in which the brackets b
c1 , b
c2 , b
c3 are created by
Ω. Note that if b
c1 is obtained from c1 by an extension of type 3, then
c1 (1) = c1 (1) − 1,
b
c1 (2) = c1 (2) + 1.
b
This means that brackets b
c2 , b
c3 could not be created by extensions after b
c1 appears,
as b(2) ≤ c(1) or b(1) ≥ c(2) for distinct brackets b, c ∈ B of any bracket system
B. This observation proves that b
c1 is obtained from c1 by an extension of type
1. Similarly to the forgoing argument, we can see that if b
c1 is obtained by an
extension of type 1 on the left/right, then b
c2 must be obtained by an extension
on the right/left, resp., and that b
c3 cannot be obtained by any extension. This
contradiction proves that it is not possible to have in Ω more than two additions
of any starting bracket c. Thus, the number of additions in Ω is at most 2|W | + 2.
As in the proof of Lemma 6.6, the number of extensions is ≤ |W | and the number
of mergers is at most 2|W | + 1. Hence, the total number of elementary operations
is at most 5|W | + 3 ≤ 7|W | as desired.
Lemma 6.8. Let there be a sequence Ω of elementary operations that transforms
the empty bracket system for (W, ∆) into the final bracket system and has size
bounded by (k1 , k2 ) and let c be a starting bracket for (W, ∆). Then there is also
46
S. V. IVANOV
a sequence of elementary operations that converts the bracket system {c} into the
final bracket system and has size bounded by (k1 + 1, k2 + 1).
Proof. The proof is analogous to the proof of Lemma 4.5. The arguments in the
cases when c(1) = 0 or c(1) = |W | are retained verbatim.
Assume that 0 < c(1) < |W |. As before, let B0 , B1 , . . . , Bℓ be the corresponding
to Ω operational sequence of bracket systems, where B0 is empty and Bℓ is final.
Let Bi∗ be the first bracket system of the sequence such that Bi∗ ∪ {c} is not a
bracket system. The existence of Bi∗ follows from the facts that B0 ∪ {c} is a
bracket system and Bℓ ∪ {c} is not. Since B0 ∪ {c} is a bracket system, it follows
that i∗ ≥ 1 and Bi∗ −1 ∪ {c} is a bracket system. Since Bi∗ −1 ∪ {c} is a bracket
system and Bi∗ ∪ {c} is not, there is a bracket b ∈ Bi∗ such that b(1) < c(1) < b(2)
and b is obtained from a bracket d1 ∈ Bi∗ −1 by an extension or b is obtained from
brackets d1 , d2 ∈ Bi∗ −1 by a merger. In either case, it follows from definitions of
elementary operations that dj (k) = c(1) for some j, k ∈ {1, 2}. Hence, we can
use a merger applied to dj and c which would result in dj , i.e., in elimination of
c from Bi∗ −1 ∪ {c} and in getting thereby Bi∗ −1 from Bi∗ −1 ∪ {c}. Now we can
see that the original sequence of elementary operations, together with the merger
Bi∗ −1 ∪ {c} → Bi∗ −1 can be used to produce the following operational sequence of
bracket systems
B0 ∪ {c}, . . . , Bi∗ −1 ∪ {c}, Bi∗ −1 , . . . , Bℓ .
Clearly, the size of this new sequence is bounded by (k1 + 1, k2 + 1), as desired.
Lemma 6.9. Let PW = p1 p2 , let the path α(p2 ) be a special arc of ∆ and let
there exist a sequence Ω of elementary operations that transforms the empty bracket
system for (W, ∆) into the final bracket system and has size bounded by (ℓ1 , ℓ2 ).
Then there is also a sequence Ωp1 of elementary operations such that Ωp1 transforms
the empty bracket system for (W, ∆) into the bracket system {(0, |p1 |, −k, |∆(2)|)},
0
where ϕ(p2 ) = ak1 , and Ωp1 has size bounded by (ℓ1 , ℓ2 ). In addition, for every
bracket b ∈ B, where B is an intermediate bracket system of the corresponding to
Ωp1 operational sequence of bracket systems, it is true that b(2) ≤ |p1 |.
Proof. Let B0 , B1 , . . . , Bℓ be the corresponding to Ω operational sequence of bracket
systems, where B0 is empty and Bℓ = {(0, |W |, 0, |∆(2)|)} is the final bracket system
for (W, ∆).
For every bracket system Bi , we will construct an associated bracket system Bi′
as follows. Let b ∈ Bi be a bracket. If b(1) > |p1 |, then we disregard b.
Suppose that b(1) ≤ |p1 | and b(2) > |p1 |. Let pb be a subpath of p2 such that
0 k
(pb )− = (p2 )− , (pb )+ = b(2) and let ϕ(pb ) = a1p , see Fig. 6.3. Since r(b) and α(pb )
are special arcs in ∆, it follows from Lemma 6.2 that the reduced path rp , obtained
from r(b)α(pb )−1 by canceling pairs of edges of the form ee−1 , is a special arc such
that
(rp )− = α(b(1)), (rp )+ = α((pb )+ ) = α(|p1 |).
Hence, we can define a bracket
b′ := (b(1), |p1 |, b(3) − kp , b(4))
whose complementary arc r(b′ ) is rp , r(b′ ) := rp . For every such b ∈ Bi , we include
b′ in Bi′ .
BOUNDED AND PRECISE WORD PROBLEMS FOR GROUP PRESENTATIONS
α(p1 )
α(|p1 |)
α(b(1))
47
α(pb )
α(b(2))
r(b)
α(p1 )
α(0)
Fig. 6.3
If b ∈ Bi satisfies b(2) ≤ |p1 |, then we add b to Bi′ .
Going over all brackets b ∈ Bi as described above, we obtain a new bracket
′
′
system Bi′ associated with Bi . Observe that Bi+1
is either identical to Bi′ or Bi+1
′
is obtained from Bi by a single elementary operation which is identical to the
elementary operation that is used to get Bi+1 from Bi .
′
Moreover, if Bi+1
6= Bi′ and Bi+1 is obtained from Bi by application of an
elementary operation σ to a bracket b1 ∈ Bi or to brackets b1 , b2 ∈ Bi (in the
case when σ is a merger) and this application results in c, written c = σ(b1 ) or
′
c = σ(b1 , b2 ), then Bi+1
is obtained from Bi′ by application of σ ′ , where σ ′ has the
same type as σ, to the bracket b′1 ∈ Bi′ or to the brackets b′1 , b′2 ∈ Bi′ (in the case
when σ is a merger) and c′ = σ ′ (b′1 ) or c′ = σ ′ (b′1 , b′2 ).
Indeed, this claim is immediate for additions, extensions of type 1 and mergers.
On the other hand, it follows from the definitions of extensions of type 2 and 3 that
an extension of type 2 or 3 can be applied only to a bracket b1 with b1 (2) ≤ |p1 |
and results in a bracket c = σ(b1 ) with c(2) ≤ |p1 |. Hence, in this case, b′1 = b1 ,
c′ = c, and our claim holds true again.
It follows from foregoing observations that a new sequence B0′ , B1′ , . . . , Bℓ′ , when
repetitions are disregarded, is operational, has size bounded by (ℓ1 , ℓ2 ), B0′ is empty
and Bℓ′ consists of the single bracket (0, |p1 |, −k, |∆(2)|), as desired. The last in
equality b(2) ≤ |p1 | in lemma’s statement is immediate from the definitions.
Lemma 6.10. Let PW = p1 p2 p3 be a factorization of the path PW such that α(p2 )
is a special arc of ∆ and let there exist a sequence of elementary operations of size
bounded by (ℓ1 , ℓ2 ) that transforms the empty bracket system for (W, ∆) into the
final bracket system. Then there is also a sequence of elementary operations that
transforms a bracket system consisting of the single bracket c0 := (|p1 |, |p1 p2 |, k2 , 0),
0
where ϕ(p2 ) = ak12 , into the final bracket system for (W, ∆) and has size bounded
by (ℓ1 + 1, ℓ2 + 1).
Proof. Consider a sequence Ω of elementary operations that transforms the empty
bracket system for (W, ∆) into the final bracket system and has size bounded by
(ℓ1 , ℓ2 ). Let B0 , B1 , . . . , Bℓ be the corresponding to Ω sequence of bracket systems,
where B0 is empty and Bℓ = {(0, |W |, 0, |∆(2)|)} is the final bracket system for
(W, ∆).
Define i∗ to be the minimal index so that Bi∗ contains a bracket b such that
b(1) ≤ |p1 | and b(2) ≥ |p1 p2 |, i.e., the arc a(b) of b contains the path α(p2 ). Since
Bℓ has this property and B0 does not, such an i∗ exists and 0 < i∗ ≤ ℓ.
First suppose that |p2 | = 0. Then c0 = (|p1 |, |p1 |, 0, 0) and we can define the
following operational sequence of bracket systems Bi′ := Bi ∪{c0 } for i = 0, . . . , i∗ −
48
S. V. IVANOV
1. By the minimality of i∗ , there is a unique bracket b∗ ∈ Bi∗ such that b∗ (1) = |p1 |
or b∗ (2) = |p1 |.
Suppose b∗ (1) = b∗ (2) = |p1 |. Then Bi∗ is obtained from Bi∗ −1 by addition of
c0 , Bi∗ = Bi∗ −1 ∪ {c0 }, and we can consider the following operational sequence of
bracket systems
B0′ = {c0 }, B1′ , . . . , Bi′∗ −2 , Bi′∗ −1 = Bi∗ , Bi∗ +1 , . . . , Bℓ
that transforms {c0 } into the final bracket system and has size bounded by (ℓ1 −
1, ℓ2 + 1).
Suppose b∗ (1) < b∗ (2). Then a merger of b∗ and c0 yields b∗ , hence this merger
turns Bi′∗ = Bi∗ ∪ {c0 } into Bi∗ . Now we can continue the original sequence Ω of
elementary operations to get Bℓ . Thus the following
B0′ = {c0 }, B1′ , . . . , Bi′∗ , Bi∗ , Bi∗ +1 , . . . , Bℓ
is an operational sequence of bracket systems that transforms {c0 } into the final
bracket system and has size bounded by (ℓ1 + 1, ℓ2 + 1).
Now assume that |p2 | > 0. For every bracket system Bi , where i < i∗ , we
construct an associated bracket system Bi′ in the following manner.
Let b ∈ Bi be a bracket. If b(1) ≥ |p1 | and b(2) ≤ |p1 p2 | and b(1) − |p1 | + |p1 p2 | −
b(2) > 0, i.e., the arc a(b) of b is a proper subpath of the path α(p2 ), then we
disregard b.
Suppose that b(1) ≤ |p1 | and |p1 | < b(2) ≤ |p1 p2 |, i.e., the arc a(b) overlaps with
a prefix subpath of the path α(p2 ) of positive length. Let p21 denote a subpath
of p2 given by (p21 )− = (p2 )− and (p21 )+ = b(2), see Fig. 6.4. Then it follows
from Lemma 6.2 applied to special arcs r(b), α(p21 )−1 of ∆ that the reduced path,
obtained from r(b)α(p21 )−1 by canceling pairs of edges of the form ee−1 , is a special
arcs of ∆. Therefore, we may consider a bracket
b′ = (b(1), |p1 |, b(3) − k21 , b(4)),
0
where ϕ(p21 ) = ak121 . For every such a bracket b ∈ Bi , we include b′ into Bi′ .
α(p1 )
α(|p1 |)
α(p21 )
∆b
α(b(1))
α(b(2))
r(b)
α(p1 )
α(p22 )
α(0)
α(|p1 p2 |)
α(p3 )
Fig. 6.4
Suppose |p1 | ≤ b(1) < |p1 p2 | and |p1 p2 | ≤ b(2), i.e., the arc a(b) overlaps with a
suffix of the path α(p2 ) of positive length. Let p22 denote a subpath of p2 given by
(p22 )− = b(1) and (p22 )+ = (p2 )+ , see Fig. 6.5. Then it follows from Lemma 6.2
applied to special arcs α(p22 )−1 , r(b) of ∆ that the reduced path, obtained from
α(p22 )−1 r(b) by canceling pairs of edges of the form ee−1 , is a special arcs of ∆.
Hence, we may consider a bracket
b′ = (|p1 p2 |, b(2), b(3) − k22 , b(4)),
0
where ϕ(p22 ) = ak122 . For every such a bracket b ∈ Bi , we include b′ into Bi′ .
BOUNDED AND PRECISE WORD PROBLEMS FOR GROUP PRESENTATIONS
α(p22 )
α(|p1 p2 |)
49
α(p3 )
∆b
α(b(1))
α(b(2))
r(b)
α(p21 )
α(0)
α(|p1 |)
α(p1 )
Fig. 6.5
α(p3 )
If b ∈ Bi satisfies either b(2) ≤ |p1 | or b(1) ≥ |p1 p2 |, then we just add b to Bi′
without alterations.
We also put the bracket c0 = (|p1 |, |p1 p2 |, k2 , 0) in every Bi′ , i = 0, . . . , i∗ − 1.
Going over all brackets b ∈ Bi , as described above, and adding c0 , we obtain a
new bracket system Bi′ associated with Bi .
′
Observe that for every i < i∗ − 1 the following holds true. Either Bi+1
is
′
′
′
identical to Bi or Bi+1 is obtained from Bi by a single elementary operation which
is identical to the elementary operation that is used to get Bi+1 from Bi . Moreover,
′
if Bi+1
6= Bi′ and Bi+1 is obtained from Bi by application of an elementary operation
σ to a bracket b1 ∈ Bi or to brackets b1 , b2 ∈ Bi (in the case when σ is a merger)
′
and this application of σ results in c, written c = σ(b1 ) or c = σ(b1 , b2 ), then Bi+1
′
′
′
is obtained from Bi by application of σ , where σ has the same type as σ, to the
bracket b′1 ∈ Bi′ or to the brackets b′1 , b′2 ∈ Bi′ (in the case when σ is a merger) and
c′ = σ ′ (b′1 ) or c′ = σ ′ (b′1 , b′2 ), resp. Indeed, this claim is immediate for additions,
extensions of type 1 and mergers. On the other hand, it follows from the definitions
of extensions of types 2–3 and from the definition of i∗ that if σ is an extension of
type 2 or 3 then σ can only be applied to a bracket b1 such that b1 (2) ≤ |p1 | or
|p1 p2 | ≤ b1 (1) and this application results in a bracket c = σ(b1 ) with c(2) ≤ |p1 |
or c(1) ≤ |p1 p2 |. Hence, in this case, b′1 = b1 , c′ = c and our claim holds true.
Since the bracket system Bi∗ contains a bracket d such that α(p2 ) is a subpath
of a(d) and i∗ is the minimal index with this property, it follows from the definition
of elementary operations and from the facts that α(p2 ) consists of a1 -edges and
|p2 | > 0 that either d is obtained from a bracket b1 ∈ Bi∗ −1 by an extension of type
1 or d is obtained from brackets b2 , b3 ∈ Bi∗ −1 by a merger.
First suppose that d is obtained from a bracket b1 ∈ Bi∗ −1 by an extension of
type 1. In this case, we pick the image b′1 ∈ Bi′∗ −1 of b1 and use a merger operation
over b′1 , c0 ∈ Bi′∗ −1 to get a bracket c3 . Let Bi′∗ denote the new bracket system
obtained from Bi′∗ −1 by the merger of b′1 , c0 ∈ Bi′∗ −1 , hence,
Bi′∗ := (Bi′∗ −1 \ {c0 , b′1 }) ∪ {c3 }.
Since d is obtained from b1 by an extension of type 1, it follows that, for every
b ∈ Bi∗ −1 , b 6= b1 , we have either b(1) ≥ d(2) ≥ |p1 p2 | or b(2) ≤ d(1) ≤ |p1 |.
Therefore, every bracket b ∈ Bi∗ −1 , b 6= b1 , has an image b′ ∈ Bi′∗ −1 and b′ = b.
This, together with equalities c3 (1) = d(1), c3 (2) = d(2) and Lemma 6.5, implies
that c3 = d and Bi′∗ = Bi∗ . Thus we can consider the following sequence of bracket
systems
B0′ , . . . , Bi′∗ −1 , Bi′∗ = Bi∗ , Bi∗ +1 , . . . , Bℓ
(6.2)
50
S. V. IVANOV
that starts at B0′ = {c0 } and ends in the final bracket system Bℓ for (W, ∆). It is
clear that the size of this sequence is bounded by (ℓ1 , ℓ2 + 1). Hence, after deletion
of possible repetitions in (6.3), we obtain a desired operational sequence.
Now assume that d is obtained from brackets b2 , b3 ∈ Bi∗ −1 by a merger and let
b2 (2) = b3 (1). To define B ′∗ 1 , we apply a merger operation to b′2 and c0 , which
i −2
results in a bracket c2 . To define Bi′∗ , we apply a merger operation to c2 and b′3
which results in a bracket c3 .
As above, we observe that since d is obtained from b2 , b3 by a merger, it follows
that, for every bracket b ∈ Bi∗ −1 , b 6∈ {b1 , b2 }, we have either b(1) ≥ d(2) ≥ |p1 p2 |
or b(2) ≤ d(1) ≤ |p1 |. Therefore, every bracket b ∈ Bi∗ −1 , b 6∈ {b1 , b2 }, has an image
b′ ∈ Bi′∗ −1 and b′ = b. This, together with equalities c3 (1) = d(1), c3 (2) = d(2) and
Lemma 6.5, implies that c3 = d and Bi′∗ = Bi∗ . Thus we can consider the following
sequence of bracket systems
′
1 , Bi∗
i −2
B0′ , . . . , Bi′∗ −1 , B ′∗
= Bi∗ , Bi∗ +1 , . . . , Bℓ
(6.3)
that starts at B0′ = {c0 } and ends in the final bracket system Bℓ . Clearly, the size
of the sequence (6.3) is bounded by (ℓ1 + 1, ℓ2 + 1). Hence, after deletion of possible
repetitions in (6.3), we obtain a desired operational sequence.
Lemma 6.11. There exists a sequence of elementary operations that converts the
empty bracket system for (W, ∆) into the final bracket system and has size bounded
by
(6.4)
(7|W |, C(log |W |ā1 + 1)),
where C = (log 56 )−1 and log |W |ā1 := 0 if |W |ā1 = 0.
Proof. The arguments of this proof are similar to those of the proof of Lemma 4.6.
Now we proceed by induction on |W |ā1 ≥ 0. To establish the basis for induction,
we first check our claim in the case when |W |ā1 ≤ 2.
Let |W |ā1 ≤ 2. Then, looking at the tree ρa1 (∆), we see that |W |ā1 = 0 or
|W |ā1 = 2. If |W |ā1 = 0, then it follows from Lemma 6.1 applied to ∆ that ∆ is a
tree consisting of a1 -edges and we can easily convert the empty bracket system for
(W, ∆) into the final bracket system by using a single addition and |W | extensions
of type 1. The size of the corresponding sequence is bounded by (|W | + 1, 1) and
the bound (6.4) holds as C ≥ 1.
If now |W |ā1 = 2, then it follows from Lemma 6.1 applied to ∆ that either ∆ is a
tree consisting of a1 -edges and two (oriented) ai -edges, i > 1, that form an ai -band
Γ0 with no faces or ∆ contains a single a2 -band Γ that has faces and all edges of
∆, that are not in Γ, are a1 -edges that form a forest whose trees are attached to Γ.
In either case, we can convert the empty bracket system for (W, ∆) into the final
bracket system by using a single addition and ≤ |W | extensions of type 1 and 3
if ∆ is a tree or of type 1 and 2 if ∆ is not a tree. The size of the corresponding
sequence of elementary operations is bounded by (|W | + 1, 1) and the bound (6.4)
holds as C ≥ 1. The base step is complete.
Making the induction step, assume |W |ā1 > 2. By Lemma 6.3 applied to (W, ∆),
we obtain vertices v1 , v2 ∈ PW such that if PW (fact, v1 , v2 ) = p1 p2 p3 then there is
a special arc r in ∆ such that r− = α(v1 ), r+ = α(v2 ) and
min(|ϕ(p2 )|ā1 , |ϕ(p1 )|ā1 + |ϕ(p3 )|ā1 ) ≥ 13 |W |ā1 ,
(6.5)
BOUNDED AND PRECISE WORD PROBLEMS FOR GROUP PRESENTATIONS
51
Let ∂|0 ∆ = q1 q2 q3 , where qi = α(pi ), i = 1, 2, 3. Consider disk subdiagrams
∆1 , ∆2 of ∆ given by ∂|v1 ∆2 = q2 r−1 and by ∂|0 ∆1 = q1 rq3 , see Fig. 6.6. Denote
W2 ≡ ϕ(q2 r−1 ),
W1 ≡ ϕ(q1 rq3 )
and let PWi = PWi (Wi , ∆i ), i = 1, 2, denote the corresponding paths such that
0
α1 (PW1 ) = q1 rq3 and α2 (PW2 ) = q2 r−1 . We also denote ϕ(r) = ak1r .
α(p2 ) = q2
∆2
α(v1 )
α(p1 ) = q1
r
∆1
α(v2 )
α(p3 ) = q3
α(0)
Fig. 6.6
Since |W1 |ā1 , |W2 |ā1 < |W |ā1 by (6.5), it follows from the induction hypothesis
b i of elementary operations for (Wi , ∆i ) that transforms
that there is a sequence Ω
the empty bracket system into the final system and has size bounded by
(7|Wi |, C(log |Wi |ā1 + 1)),
(6.6)
where i = 1, 2.
Furthermore, it follows from the bound (6.6) with i = 2 and from Lemma 6.9
applied to ∆2 whose boundary label has factorization ϕ(∂|0 ∆2 ) ≡ W2 ≡ W21 W22 ,
b 2 of elemenwhere W21 ≡ ϕ(p2 ) and W21 ≡ ϕ(r)−1 , that there exists a sequence Ω
tary operations for the pair (W2 , ∆2 ) that transforms the empty bracket system
B2,0 into the bracket system B2,ℓ2 = {(0, |p2 |, kr , |∆2 (2)|)} and has size bounded
by
(7|W2 |, C(log |W2 |ā1 + 1)) .
(6.7)
b 2 sequence of bracket systems.
Let B2,0 , B2,1 , . . . , B2,ℓ2 denote the associated with Ω
It follows from the bound (6.6) with i = 1 and from Lemma 6.10 applied to
∆1 whose boundary label has factorization ϕ(∂|0 ∆1 ) ≡ W1 ≡ W11 W12 W13 , where
W11 ≡ ϕ(p1 ), W12 ≡ ϕ(r), W13 ≡ ϕ(p3 ), that there exists a sequence Ω1 of elementary operations for (W1 , ∆1 ) that transforms the bracket system B1,0 = {cr },
where
cr := (|p1 |, |p1 | + |r|, kr , 0),
into the final bracket system B1,ℓ1 and has size bounded by
(7|W1 | + 1, C(log |W1 |ā1 + 1)) .
(6.8)
Let B1,0 , B1,1 , . . . , B1,ℓ1 be the associated with Ω1 sequence of bracket systems.
Now we will show that these two sequences Ω2 , Ω1 of elementary operations,
the first one for (W2 , ∆2 ), and the second one for (W1 , ∆1 ), could be modified and
combined into a single sequence of elementary operations for (W, ∆) that converts
the empty bracket system into the final bracket system and has size with a desired
bound.
52
S. V. IVANOV
Note that every bracket b = (b(1), b(2), b(3), b(4)), b ∈ ∪j B2,j , for (W2 , ∆2 ), by
Lemma 6.9, has the property that b(2) ≤ |p2 | and b naturally gives rise to the
bracket
bb := (b(1) + |p1 |, b(2) + |p1 |, b(3), b(4))
b2,j denote the bracket system for (W, ∆) obtained from B2,j by
for (W, ∆). Let B
b2,0 , . . . , B
b2,ℓ2
replacing every bracket b ∈ B2,j with bb. Then it is easy to verify that B
is an operational sequence of bracket systems for (W, ∆) that changes the empty
bracket system into
b2,ℓ2 = {(|p1 |, |p1 | + |p2 |, kr , |∆2 (2)|)}
B
and has size bounded by (6.7).
Consider a relation ≻1 on the set of all pairs (b, i), where b ∈ B1,i , i = 0, . . . , ℓ1 ,
defined so that (c, i + 1) ≻1 (b, i) if and only if c ∈ B1,i+1 is obtained from brackets
b, b′ ∈ B1,i by an elementary operation σ, here b′ is missing if σ is an extension.
Define a relation on the set of all pairs (b, i), where b ∈ B1,i , i = 0, . . . , ℓ1 ,
that is the reflexive and transitive closure of the relation ≻1 . Clearly, is a partial
order on the set of such pairs (b, i) and if (b2 , i2 ) (b1 , i1 ) then i2 ≥ i1 and
b2 (1) ≤ b1 (1) ≤ b1 (2) ≤ b2 (2).
Now we observe that every bracket d = (d(1), d(2), d(3), d(4)), d ∈ B1,i , for
b
b
b
b
(W1 , ∆1 ) naturally gives rise to a bracket db = (d(1),
d(2),
d(3),
d(4))
for (W, ∆) in
the following manner.
If (d, i) is not comparable with (cr , 0) by the relation and d(1) ≤ |p1 |, then
db := d.
If (d, i) is not comparable with (cr , 0) and |p1 | + |r| ≤ d(2), then
db := (d(1) + |p2 | − |r|, d(2) + |p2 | − |r|, d(3), d(4)).
If (d, i) (cr , 0), then
db := (d(1), d(2) + |p2 | − |r|, d(3), d(4) + |∆2 (2)|).
Note that the above three cases cover all possible situations.
b1,i := {db | d ∈ B1,i }. Then it is easy to verify that B
b1,0 , . . . , B
b1,ℓ1
As before, let B
is an operational sequence of bracket systems for (W, ∆) that changes the bracket
system
b1,0 = B
b2,ℓ2 = {(|p1 |, |p1 | + |p2 |, kr , |∆2 (2)|)}
B
b1,ℓ1 = {(0, |p1 | + |p2 | + |p3 |, 0, |∆1 (2)| + |∆2 (2)|)}.
into the final bracket system B
Thus, with the indicated changes, we can now combine the foregoing modified
sequences of bracket systems for (W2 , ∆2 ) and for (W1 , ∆1 ) into a single operational
sequence
b1,0 , . . . , B
b1,ℓ1
b2,0 , . . . , B
b2,ℓ2 = B
B
of bracket systems for (W, ∆) that transforms the empty bracket system into the
bracket system {(|p1 |, |p1 | + |p2 |, kr , |∆2 (2)|)} and then continues to turn the latter
into the final bracket system. It follows from definitions and bounds (6.7)–(6.8)
that the size of thus constructed sequence is bounded by
(7|W1 | + 7|W2 | + 1, max(C(log |W1 |ā1 + 1) + 1, C(log |W2 |ā1 + 1)))
BOUNDED AND PRECISE WORD PROBLEMS FOR GROUP PRESENTATIONS
53
Therefore, in view of Lemma 6.7, it remains to show that
max(C(log |W1 |ā1 + 1) + 1, C(log |W2 |ā1 + 1)) ≤ C(log |W |ā1 + 1).
In view of the inequality (6.5),
max(C(log |W1 |ā1 + 1) + 1, C(log |W2 |ā1 + 1)) ≤ C(log( 65 |W |ā1 ) + 1) + 1,
and C(log( 65 |W |ā1 ) + 1) + 1 ≤ C(log |W |ā1 + 1) for C = (log 56 )−1 .
Let W be an arbitrary nonempty word over the alphabet A±1 , not necessarily
representing the trivial element of the group given by presentation (1.4). As before,
let PW be a simple labeled path with ϕ(PW ) ≡ W and let vertices of PW be
identified along PW with integers 0, 1, . . . , |W | so that (PW )− = 0, . . . , (PW )+ =
|W |.
A quadruple
b = (b(1), b(2), b(3), b(4))
of integers b(1), b(2), b(3), b(4) is called a pseudobracket for the word W if b(1), b(2)
satisfy the inequalities 0 ≤ b(1) ≤ b(2) ≤ |W | and b(4) ≥ 0.
Let p denote the subpath of PW with p− = b(1), p+ = b(2), perhaps, p− = p+
and |p| = 0. The subpath p of PW is denoted a(b) and called the arc of the
pseudobracket b.
For example, bv = (v, v, 0, 0) is a pseudobracket for every vertex v ∈ PW and
such bv is called a starting pseudobracket. Note that a(b) = v = b(1). A final
pseudobracket for W is c = (0, |W |, 0, k), where k ≥ 0 is an integer. Note that
a(c) = PW .
Observe that if b is a bracket for the pair (W, ∆), then b is also a pseudobracket
for the word W .
Let B be a finite set of pseudobrackets for W , perhaps, B is empty. We say that
B is a pseudobracket system if, for every pair b, c ∈ B of distinct pseudobrackets,
either b(2) ≤ c(1) or c(2) ≤ b(1). As before, B is called a final pseudobracket
system if B contains a single pseudobracket which is final. Clearly, every bracket
system for (W, ∆) is also a pseudobracket system for the word W .
Now we describe three kinds of elementary operations over pseudobrackets and
over pseudobracket systems: additions, extensions, and mergers, which are analogous to those definitions for brackets and bracket systems, except there are no any
diagrams and faces involved.
Let B be a pseudobracket system for a word W .
Additions.
Suppose b is a starting pseudobracket, b 6∈ B, and B ∪ {b} is a pseudobracket
system. Then we can add b to B thus making an addition operation over B.
Extensions.
Suppose B is a pseudobracket system, b ∈ B is a pseudobracket and e1 a(b)e2 is
a subpath of PW , where a(b) is the arc of b and e1 , e2 are edges one of which could
be missing.
Assume that ϕ(e1 ) = aε11 , where ε1 = ±1. Then we consider a pseudobracket b′
such that
b′ (1) = b(1) − 1, b′ (2) = b(2), b′ (3) = b(3) + ε1 , b′ (4) = b(4).
54
S. V. IVANOV
Note that a(b′ ) = e1 a(b). We say that b′ is obtained from b by an extension of type
1 (on the left). If (B \ {b}) ∪ {b′ } is a pseudobracket system, then replacement of
b ∈ B with b′ in B is called an extension operation over B of type 1.
Similarly, assume that ϕ(e2 ) = aε12 , where ε2 = ±1. Then we consider a pseudobracket b′ such that
b′ (1) = b(1), b′ (2) = b(2) + 1, b′ (3) = b(3) + ε2 , b′ (4) = b(4).
Note that a(b′ ) = a(b)e2 . We say that b′ is obtained from b by an extension of type
1 (on the right). If (B \ {b}) ∪ {b′ } is a pseudobracket system, then replacement of
b ∈ B with b′ in B is called an extension operation over B of type 1.
Now suppose that both edges e1 , e2 do exist and ϕ(e1 ) = ϕ(e2 )−1 = aε2 , ε = ±1.
We also assume that b(3) 6= 0 and that b(3) is a multiple of n1 if ε = 1 or b(3) is a
multiple of n2 if ε = −1. Consider a pseudobracket b′ such that
b′ (1) = b(1)−1, b′ (2) = b(2)+1, b′ (3) = (|n1−1 n2 |)ε b(3), b′ (4) = b(4)+|b(3)|/|ni(ε) |,
where i(ε) = 1 if ε = 1 and i(ε) = 2 if ε = −1. Note that a(b′ ) = e1 a(b)e2 . We
say that b′ is obtained from b by an extension of type 2. If (B \ {b}) ∪ {b′ } is a
pseudobracket system, then replacement of b ∈ B with b′ in B is called an extension
operation over B of type 2.
Assume that both e1 , e2 do exist, ϕ(e1 ) = ϕ(e2 )−1 , ϕ(e1 ) 6= a±1
1 , and b(3) = 0.
Consider a pseudobracket b′ such that
b′ (1) = b(1) − 1, b′ (2) = b(2) + 1, b′ (3) = b(3), b′ (4) = b(4).
Note that a(b′ ) = e1 a(b)e2 . We say that b′ is obtained from b by an extension of
type 3. If (B \ {b}) ∪ {b′ } is a pseudobracket system, then replacement of b ∈ B
with b′ in B is called an extension operation over B of type 3.
Mergers.
Suppose that b1 , b2 are distinct pseudobrackets in B such that b1 (2) = b2 (1).
Consider a pseudobracket b′ such that
b′ (1) = b1 (1), b′ (2) = b2 (2), b′ (3) = b1 (3) + b2 (3), b′ (4) = b1 (4) + b2 (4).
Note that a(b′ ) = a(b1 )a(b2 ). We say that b′ is obtained from b1 , b2 by a merger
operation. Taking both b1 , b2 out of B and putting b′ in B is a merger operation
over B.
We will say that additions, extensions, and mergers, as defined above, are elementary operations over pseudobrackets and pseudobracket systems for W .
Assume that one pseudobracket system Bℓ is obtained from another pseudobracket system B0 for W by a finite sequence Ω of elementary operations and
B0 , B1 , . . . , Bℓ is the corresponding to Ω sequence of pseudobracket systems. As
above, we say that the sequence B0 , B1 , . . . , Bℓ is operational and that B0 , B1 , . . . , Bℓ
has size bounded by (k1 , k2 ) if ℓ ≤ k1 and, for every i, the number of pseudobrackets
in Bi is at most k2 . Whenever it is not ambiguous, we also say that Ω has size
bounded by (k1 , k2 ) if so does the corresponding to Ω sequence B0 , B1 , . . . , Bℓ of
pseudobracket systems.
Lemma 6.12. Suppose that the empty pseudobracket system B0 for W can be transformed by a finite sequence Ω of elementary operations into a final pseudobracket
G
system Bℓ = {(0, |W |, 0, k)}. Then W =3 1 and there is a reduced disk diagram ∆
over presentation (1.4) such that ϕ(∂|0 ∆) ≡ W and |∆(2)| ≤ k.
BOUNDED AND PRECISE WORD PROBLEMS FOR GROUP PRESENTATIONS
55
Proof. Let B0 , B1 , . . . , Bℓ be the corresponding to Ω sequence of pseudobracket
systems, where B0 is empty and Bℓ is final. Consider the following claim.
G
c(3)
(D1) If c is a pseudobracket of Bi , i = 1, . . . , ℓ, then ϕ(a(c)) =3 a1 , where a(c)
is the arc of c, and there is a disk diagram ∆c over presentation (1.4) such
c(3)
that ∂|0 ∆c = sr−1 , where ϕ(s) ≡ ϕ(a(c)), r is simple, ϕ(r) ≡ a1 and
|∆c (2)| = c(4).
Proof of Claim (D1). By induction on i ≥ 1, we will prove that Claim (D1)
holds for every pseudobracket c ∈ Bi , i = 1, . . . , ℓ.
The base step of this induction is obvious since B1 consists of a starting pseudobracket c for which we have a disk diagram ∆c consisting of a single vertex.
To make the induction step from i to i + 1, i ≥ 1, we consider the cases corresponding to the type of the elementary operation that is used to get Bi+1 from
Bi .
Suppose that Bi+1 is obtained from Bi by an elementary operation σ and c ∈
Bi+1 is the pseudobracket obtained from b1 , b2 ∈ Bi by application of σ, denoted
c = σ(b1 , b2 ). Here one of b1 , b2 or both, depending on type of σ, could be missing.
By the induction hypothesis, Claim (D1) holds for every pseudobracket of Bi+1
different from c and it suffices to show that Claim (D1) holds for c.
If c ∈ Bi+1 is obtained by an addition, then Claim (D1) holds for c because it
holds for any starting pseudobracket.
Let c ∈ Bi+1 be obtained from a pseudobracket b ∈ Bi by an extension of type
1 on the left and a(c) = e1 a(b), where ϕ(e1 ) = aε11 , ε1 = ±1, and a(c), a(b) are the
arcs of c, b, resp. By the induction hypothesis applied to b, there is a disk diagram
∆b over presentation (1.4) such that ∂|0 ∆b = sr−1 , where ϕ(s) ≡ ϕ(a(b)), r is
b(3)
simple, ϕ(r) ≡ a1 , and |∆b (2)| = b(4). First we assume that the integers ε1 , b(3)
satisfy ε1 b(3) ≥ 0. Then we consider a new “loose” edge f such that ϕ(f ) = aε11 . We
attach the vertex f+ to the vertex s− of ∆b and obtain thereby a disk diagram ∆′b
such that ∂∆′b = f sr−1 f −1 , see Fig. 6.7(a). Since ϕ(a(c)) ≡ ϕ(f s) ≡ aε11 ϕ(a(b)),
f r is simple and
ε +b(3)
ϕ(f r) ≡ a11
, |∆′b (2)| = |∆b (2)| = b(4),
it follows that we can use ∆′b as a required disk diagram ∆c for c.
s
∆′b
s
∆′b
∆b
f
r
Fig. 6.7(a)
∆b
g
r′
r = gr′
Fig. 6.7(b)
Now suppose that the integers ε1 , b(3) satisfy ε1 b(3) < 0, i.e., b(3) 6= 0 and ε1 ,
b(3) have different signs. Then we can write r = gr′ , where g is an edge of r and
ϕ(g) = a1−ε1 . Let ∆′b denote the diagram ∆′b whose boundary has factorization
∂|0 ∆′b = (g −1 s)(r′ )−1 , see Fig. 6.7(b). Note that ϕ(a(c)) ≡ ϕ(g −1 s), r′ is simple,
56
S. V. IVANOV
and
ε +b(3)
ϕ(r′ ) ≡ a11
, |∆′b (2)| = |∆b (2)| = b(4).
Hence, it follows that we can use ∆′b as a desired diagram ∆c for c. The case of an
extension of type 1 on the right is similar.
Let c ∈ Bi+1 be obtained from a pseudobracket b ∈ Bi by an extension of type 2
and assume that a(c) = e1 a(b)e2 , where ϕ(e1 ) = ϕ(e2 )−1 = a2 (the subcase when
ϕ(e1 ) = ϕ(e2 )−1 = a−1
2 is similar). Let a(c), a(b) denote the arcs of pseudobrackets
c, b, resp. According to the definition of an extension of type 2, b(3) is a multiple of
n1 , say, b(3) = n1 ℓ and ℓ 6= 0. By the induction hypothesis, there is a disk diagram
∆b over presentation (1.4) such that ∂|0 ∆b = sr−1 , where ϕ(s) ≡ ϕ(a(b)), r is
b(3)
simple, ϕ(r) ≡ a1 , and |∆b (2)| = b(4). Since b(3) = n1 ℓ, there is a disk diagram
Γ over (1.4) such that
b(3)
∂Γ = f1 s1 f2 s2 , ϕ(f1 ) = a2 , ϕ(f2 ) = a−1
2 , ϕ(s1 ) ≡ a1
b(3)n2 /n1
, ϕ(s−1
2 ) ≡ a1
,
and Γ consists of |ℓ| > 0 faces. Note that Γ itself is an a2 -band with |ℓ| faces.
Attaching Γ to ∆b by identification of the paths s1 and r that have identical labels,
we obtain a disk diagram ∆′b such that ∂|0 ∆′b = f1 sf2 s2 and |∆′b (2)| = |∆b (2)| + |ℓ|,
b(3)n2 /n1
see Fig. 6.8. Then ϕ(a(c)) ≡ ϕ(f1 sf2 ), s2−1 is simple, ϕ(s−1
and we
2 ) ≡ a1
′
see that ∆b satisfies all desired conditions for ∆c .
s
∆′b
∆b
r = s1
s1
f1
Γ
f2
s2
Fig. 6.8
Now let c ∈ Bi+1 be obtained from a pseudobracket b ∈ Bi by an extension of
type 3 and let a(c) = e1 a(b)e2 , where ϕ(e1 ) = ϕ(e2 )−1 , ϕ(e1 ) = aεj , ε = ±1 and
j > 1. By the definition of an extension of type 3, we have b(3) = 0. Hence, by the
induction hypothesis, there is a disk diagram ∆b over (1.4) such that ∂|0 ∆b = sr−1 ,
where ϕ(s) ≡ ϕ(a(b)), |r| = 0 and |∆b (2)| = b(4). Consider a new “loose” edge f
such that ϕ(f ) = aεj . We attach the vertex f+ to the vertex s− = s+ of ∆b and
obtain thereby a disk diagram ∆′b such that ∂∆′b = f sf −1 r′ , where |r′ | = 0, see
Fig. 6.9. Since ϕ(a(c)) ≡ ϕ(f sf −1 ) and |∆′b (2)| = |∆b (2)| = b(4), it follows that we
can use ∆′b as a desired diagram ∆c for c.
s
∆b
f
Fig. 6.9
∆′b
BOUNDED AND PRECISE WORD PROBLEMS FOR GROUP PRESENTATIONS
57
Let c ∈ Bi+1 be obtained from pseudobrackets b1 , b2 ∈ Bi by a merger and
a(c) = a(b1 )a(b2 ), where a(c), a(b1 ), a(b2 ) are the arcs of the pseudobrackets c, b1 , b2 ,
resp.
By the induction hypothesis, there are disk diagrams ∆b1 , ∆b2 over presentation
(1.4) such that
b (3)
∂|0 ∆bj = sj rj−1 , ϕ(sj ) ≡ ϕ(a(bj )), ϕ(rj ) ≡ a1j
, |∆bj (2)| = bj (4),
and rj is a simple path for j = 1, 2. Assume that the numbers b1 (3), b2 (3) satisfy the condition b1 (3)b2 (3) ≥ 0. We attach ∆b1 to ∆b2 by identification of
the vertices (s1 )+ and (s2 )− and obtain thereby a disk diagram ∆′b such that
∂∆′b = s1 s2 (r1 r2 )−1 and |∆′b (2)| = |∆b1 (2)| + |∆b2 (2)|, see Fig. 6.10(a). Note
b (3)+b2 (3)
. Hence,
that ϕ(a(c)) ≡ ϕ(s1 s2 ), r1 r2 is a simple path and ϕ(r1 r2 ) ≡ a11
′
∆b is a desired disk diagram ∆c for c.
s1
∆′b
s2
∆b1
∆b2
r1
r2
s1
∆′b
∆b1
r12 = r2−1
r11
∆b2
s2
Fig. 6.10(a)
Fig. 6.10(b)
Now suppose that b1 (3), b2 (3) satisfy the condition b1 (3)b2 (3) < 0. For definiteness, assume that |b1 (3)| ≥ |b2 (3)| (the case |b1 (3)| ≤ |b2 (3)| is analogous). Denote
r1 = r11 r12 , where |r12 | = |r2 | = |b2 (3)|. Then we attach ∆b2 to ∆b1 by identi−b (3)
fication of the paths r12 and r2−1 that have identical labels equal to a1 2 , see
−1
Fig. 6.10(b). Doing this produces a disk diagram ∆′b such that ∂∆′b = s1 s2 r11
and
′
|∆b (2)| = |∆b1 (2)|+ |∆b2 (2)|. Note that ϕ(a(c)) ≡ ϕ(s1 s2 ), r11 is a simple path and
b (3)+b2 (3)
ϕ(r11 ) ≡ a11
. Hence ∆′b is a desired disk diagram ∆c for c. The induction
step is complete and Claim (D1) is proven.
To finish the proof of Lemma 6.12, we note that it follows from Claim (D1)
applied to the pseudobracket bF = (0, |W |, 0, k) of the final system Bℓ = {bF } that
there is a disk diagram ∆bF such that ∂|0 ∆bF = sr−1 , where ϕ(s) ≡ W , |r| = 0 and
|∆bF (2)| ≤ k. It remains to pick a reduced disk diagram ∆ with these properties
of ∆bF .
Lemma 6.13. Suppose W is a nonempty word over A±1 and n ≥ 0 is an inte−n2 ±1
ger. Then W is a product of at most n conjugates of the words (a2 an1 1 a−1
)
2 a1
if and only if there is a sequence Ω of elementary operations such that Ω transforms the empty pseudobracket system for W into a final pseudobracket system
{(0, |W |, 0, n′)}, where n′ ≤ n, Ω has size bounded by
(7|W |, C(log |W |ā1 + 1)),
58
S. V. IVANOV
where C = (log 65 )−1 , and, if b is a pseudobracket of an intermediate pseudobracket
system of the corresponding to Ω sequence of pseudobracket systems, then
|b(3)| ≤ 21 (|W |a1 + (|n1 | + |n2 |)n).
(6.9)
Proof. Assume that W is a product of at most n conjugates of the words
−n2 ±1
(a2 an1 1 a−1
) .
2 a1
It follows from Lemma 2.1 that there is a reduced disk diagram ∆ over presentation
(1.4) such that ϕ(∂|0 ∆) ≡ W and |∆(2)| ≤ n. Applying Lemma 6.11 to the pair
(W, ∆), we obtain a sequence of elementary operations over bracket systems for
(W, ∆) that converts the empty bracket system for (W, ∆) into the final bracket
system {(0, |W |, 0, |∆(2)|)} and has size bounded by (7|W |, C(log |W |ā1 +1)). Since
every bracket b and every bracket system B for (W, ∆) could be considered as a
pseudobracket and a pseudobracket system for W , resp., we automatically have a
desired sequence of pseudobracket systems. The inequality (6.9) follows from the
definition of a bracket for (W, ∆) and the inequalities
1
~
|b(3)| ≤ 21 |∆(1)|
a1 = 2 (|W |a1 + (|n1 | + |n2 |)|∆(2)|),
|∆(2)| ≤ n.
Conversely, the existence of a sequence of elementary operations over pseudobracket systems for W , as specified in Lemma 6.13, implies, by virtue of Lemma 6.12,
G
that W =3 1 and that there exists a reduced disk diagram ∆ such that ϕ(∂|0 ∆) ≡ W
and |∆(2)| = n′ ≤ n. Hence, W is a product of n′ ≤ n conjugates of words
−n2 ±1
(a2 an1 1 a−1
) , as required.
2 a1
7. Proof of Theorem 1.4
Theorem 1.4. Let the group G3 be defined by a presentation of the form
n2
G3 := h a1 , . . . , am k a2 an1 1 a−1
2 = a1 i,
(1.4)
where n1 , n2 are some nonzero integers. Then both the bounded and precise word
problems for (1.4) are in L3 and in P. Specifically, the problems can be solved
in deterministic space O((max(log |W |, log n)(log |W |)2 ) or in deterministic time
O(|W |4 ).
Proof. As in the proof of Theorem 1.2, we start with L3 part of Theorem 1.4. First
we discuss a nondeterministic algorithm which solves the bounded word problem
for presentation (1.4) and which is now based on Lemma 6.13.
Given an input (W, 1n ), where W is a nonempty word (not necessarily reduced)
over the alphabet A±1 and n ≥ 0 is an integer, written in unary notation as
1n , we begin with the empty pseudobracket system and nondeterministically apply a sequence of elementary operations of size ≤ (7|W |, C(log |W |ā1 + 1)), where
C = (log 65 )−1 . If such a sequence of elementary operations results in a final pseudobracket system, consisting of a single pseudobracket {(0, |W |, 0, n′ )}, where n′ ≤ n,
then our algorithm accepts and, in view of Lemma 6.13, we may conclude that W
−n2 ±1
) . It follows from
is a product of n′ ≤ n conjugates of the words (a2 an1 1 a−1
2 a1
definitions and Lemma 6.13 that the number of elementary operations needed for
this algorithm to accept is at most 7|W |. Hence, it follows from the definition of
elementary operations over pseudobracket systems for W that the time needed to
run this nondeterministic algorithm is bounded by O(|W |). To estimate the space
requirements of this algorithm, we note that if b is a pseudobracket for W , then
BOUNDED AND PRECISE WORD PROBLEMS FOR GROUP PRESENTATIONS
59
b(1), b(2) are integers in the range from 0 to |W |, hence, when written in binary
notation, will take at most C ′ (log |W | + 1) space, where C ′ is a constant. Since
b(3), b(4) are also integers that satisfy inequalities
|b(3)| ≤ 12 (|W |a1 + (|n1 | + |n2 |)n) ≤ |W |(|n1 | + |n2 |)(n + 1),
(7.1)
0 ≤ b(4) ≤ n,
(7.2)
and |n1 | + |n2 | is a constant, it follows that the total space required to run this
algorithm is at most
(2C ′ (log |W | + 1) + C ′ log(|W |(|n1 | + |n2 |)(n + 1))) · C(log |W |ā1 + 1)+
+ (C ′ log(n + 1)) · C(log |W |ā1 + 1) = O((max(log |W |, log n) log |W |).
As in the proof of Theorem 1.2, it now follows from Savitch’s theorem [38], see
also [3], [32], that there is a deterministic algorithm that solves the bounded word
problem for presentation (1.4) in space O(max(log |W |, log n)(log |W |)2 ).
To solve the precise word problem for presentation (1.4), assume that we are
given a pair (W, 1n ) and we wish to find out if W is a product of n conjugates
−n2 ±1
of the words (a2 an1 1 a−1
)
and n is minimal with this property. Using the
2 a1
foregoing deterministic algorithm, we check whether the bounded word problem
is solvable for the two pairs (W, 1n−1 ) and (W, 1n ). It is clear that the precise word problem for the pair (W, 1n ) has a positive solution if and only if the
bounded word problem has a negative solution for the pair (W, 1n−1 ) and has a
positive solution for the pair (W, 1n ). Since these facts can be verified in space
O(max(log |W |, log n)(log |W |)2 ), we obtain a solution for the precise word problem in deterministic space O(max(log |W |, log n)(log |W |)2 ), as desired.
Now we describe an algorithm that solves the precise word problem for presentation (1.4) in polynomial time.
Recall that if ai ∈ A and U is a word over A±1 , then |U |ai denotes the number
of all occurrences of letters ai , a−1
in U and |U |āi := |U | − |U |ai .
i
Lemma 7.1. Let ∆ be a reduced disk diagram over presentation (1.4), W ≡
ϕ(∂|0 ∆) and |W |ā1 > 0. Then at least one of the following claims (a)–(b) holds
true.
(a) There is a vertex v of the path PW such that if PW (fact, v) = p1 p2 then
|ϕ(p1 )|ā1 > 0, |ϕ(p2 )|ā1 > 0 and there is a special arc s in ∆ such that s− = α(0)
and s+ = α(v).
(b) Suppose PW (fact, v1 , v2 ) = p1 p2 p3 is a factorization of PW such that |ϕ(p1 )|ā1 =
0, |ϕ(p2 )|ā1 = 0, and if qi = α(pi ), i = 1, 2, 3, e1 , e2 are the first, last, resp., edges of
q2 , then neither of e1 , e2 is an a1 -edge. Then there exists an ak -band Γ in ∆ whose
standard boundary is ∂Γ = ei ri e3−i r3−i , where i = 1 if ϕ(e1 ) ∈ A and i = 2 if
ϕ(e1 )−1 ∈ A. In particular, the subdiagram ∆1 of ∆ defined by ∂∆1 = q1 r2−1 q3 , see
Fig. 7.1, contains no faces and if ∆2 is the subdiagram of ∆ defined by ∂∆2 = q2 r1−1 ,
G
see Fig. 7.1, then |∆(2)| = |Γ(2)| + |∆2 (2)| and ϕ(q2 ) =3 aℓ1 , where ℓ = 0 if k > 2,
ℓ = n1 ℓ0 for some integer ℓ0 if ϕ(e1 ) = a2 , and ℓ = n2 ℓ′0 for some integer ℓ′0 if
ϕ(e1 ) = a−1
2 .
60
S. V. IVANOV
q2
∆2
r1
e1
e2
Γ
r2
∆1
q1
q3
α(0)
Fig. 7.1
Proof. Since |W |ā1 > 0, it follows that there is a unique factorization PW = p1 p2 p3
such that |ϕ(p1 )|ā1 = 0, |ϕ(p2 )|ā1 = 0, and the first and the last edges of p2 are not
a1 -edges. Denote qi = α(pi ), i = 1, 2, 3, and let e1 , f be the first, last, resp., edges
of q2 .
Let ϕ(e1 ) = aδk , δ = ±1. Since k > 1, there is an ak -band Γ whose standard
boundary is ∂Γ = ei ri e3−i r3−i , where i = 1 if ϕ(e1 ) = ak and i = 2 if ϕ(e1 ) = a−1
k .
First assume that e2 6= f . By Lemma 6.1, r2 is a special arc in ∆. Since
|ϕ(q1 )|ā1 = 0, there is a special arc r in ∆ such that r− = α(0), r+ = (e1 )− .
Applying Lemma 6.2 to special arcs r, r2−1 , we obtain a special arc s such that
s− = α(0), s+ = (r2 )− . Now we can see from e2 6= f that claim (a) of Lemma 7.1
holds for v = (p2 )+ .
Suppose e2 = f . In view of the definition of an ak -band and Lemma 6.1, we
conclude that claim (b) of Lemma 7.1 holds true.
G
If U =3 aℓ1 for some integer ℓ, we let µ3 (U ) denote the integer such that the
precise word problem for presentation (1.4) holds for the pair (U a−ℓ
1 , µ3 (U )). If
G3
U 6= aℓ1 for any ℓ, we set µ3 (U ) := ∞.
G
Lemma 7.2. Let U be a word such that U =3 aℓ1 for some integer ℓ and |U |ā1 > 0.
Then at least one of the following claims (a)–(b) holds true.
G
(a) There is a factorization U ≡ U1 U2 such that |U1 |ā1 , |U2 |ā1 > 0, U1 =3 aℓ11 ,
G
U2 =3 aℓ12 for some integers ℓ1 , ℓ2 , ℓ = ℓ1 + ℓ2 , and µ3 (U ) = µ3 (U1 ) + µ3 (U2 ).
(b) There is a factorization U ≡ U1 aδk U2 a−δ
k U3 , where |U1 |ā1 = |U2 |ā1 = 0,
G
k ≥ 2, δ = ±1, such that U2 =3 aℓ12 for some integer ℓ2 and the following is true.
0
If k > 2, then ℓ2 = 0 and U1 U3 = aℓ1 . If k = 2 and δ = 1, then ℓ2 /n1 is an
0
ℓ n /n
integer and U1 a12 2 1 U3 = aℓ1 . If k = 2 and δ = −1, then ℓ2 /n2 is an integer and
0
ℓ n /n
U1 a12 1 2 U3 = aℓ1 .
Then, in case k > 2, we have µ3 (U ) = µ3 (U2 ); in case k = 2 and δ = 1, we
have µ3 (U ) = µ3 (U2 ) + |ℓ2 /n1 |; in case k = 2 and δ = −1, we have µ3 (U ) =
µ3 (U2 ) + |ℓ2 /n2 |.
Proof. Consider a disk diagram over (1.4) such that ϕ(∂|0 ∆) ≡ U a−ℓ
1 and |∆(2)| =
µ3 (U ). Clearly, ∆ is reduced and Lemma 7.1 can be applied to ∆.
Assume that Lemma 7.1(a) holds for ∆. Then there is a special arc s in ∆ such
that s− = α(0) and s+ = (q1 )+ , ∂|0 ∆ = q1 q2 , and |ϕ(q1 )|ā1 , |ϕ(q2 )|ā1 > 0. Cutting
BOUNDED AND PRECISE WORD PROBLEMS FOR GROUP PRESENTATIONS
61
∆ along the simple path s, we obtain two diagrams ∆1 , ∆2 such that ∂∆1 = q1 s−1 ,
0
∂∆2 = q2 s. Since |∆(2)| = µ3 (U ) and ϕ(s) = aℓ10 for some ℓ0 , it follows that if
U1 := ϕ(q1 ), U2 a−ℓ
1 := ϕ(q2 ), U ≡ U1 U2 , then |∆i (2)| = µ3 (Ui ), i = 1, 2. Since
|∆(2)| = |∆1 (2)| + |∆2 (2)|, we obtain
µ3 (U ) = |∆(2)| = |∆1 (2)| + |∆2 (2)| = µ3 (U1 ) + µ3 (U2 ),
as required.
Assuming that Lemma 7.1(b) holds for ∆, we derive claim (b) from the definitions
and Lemma 7.1(b).
Let W be a word over A±1 and let |W |ā1 > 0. As in the proof of Theorem 1.2,
let W (i, j), where 1 ≤ i ≤ |W | and 0 ≤ j ≤ |W |, denote the subword of W that
G
ℓ
starts with the ith letter of W and has length j. If W (i, j) =3 a1ij for some integer
ℓij , we set λ(W (i, j)) := ℓij . Otherwise, it is convenient to define λ(W (i, j)) := ∞.
We now compute the numbers λ(W (i, j)), µ3 (W (i, j)) for all i, j by the method
of dynamic programming in which the parameter is |W (i, j)|ā1 ≥ 0. In other
words, we compute the numbers λ(W (i, j)), µ3 (W (i, j)) by induction on parameter
|W (i, j)|ā1 ≥ 0.
To initialize, or to make the base step, we note that if |W (i, j)|ā1 = 0 then
λ(W (i, j)) = ℓij ,
0
µ3 (W (i, j)) = 0
ℓ
a1ij .
where W (i, j) =
To make the induction step, assume that the numbers λ(W (i′ , j ′ )), µ3 (W (i′ , j ′ ))
are already computed for all W (i′ , j ′ ) such that |W (i′ , j ′ )|ā1 < |W (i, j)|ā1 , where
W (i, j) is fixed and |W (i, j)|ā1 > 0.
Applying Lemma 7.2 to the word U ≡ W (i, j), we consider all factorizations of
the form
U ≡ U1 U2 ,
where |U1 |ā1 , |U2 |ā1 > 0. Since U1 = W (i, j1 ), where j1 = |U1 | and U2 = W (i +
j1 , j − j1 ), it follows from |U1 |ā1 , |U2 |ā1 < |U |ā1 that the numbers λ(U1 ), µ3 (U1 ),
λ(U2 ), µ3 (U2 ) are available by the induction hypothesis. Hence, we can compute
the minimum
Ma = min(µ3 (U1 ) + µ3 (U2 ))
(7.3)
µ3 (U1′ )
over all such factorizations U ≡ U1 U2 . If Ma < ∞ and Ma =
some factorization U ≡ U1′ U2′ , then we set La := λ(U1′ ) + λ(U2′ ).
We also consider a factorization for U = W (i, j) of the form
+
µ3 (U2′ )
U ≡ U1 aδk U2 a−δ
k U3 ,
(7.4)
where |U1 |ā1 = |U2 |ā1 = 0, and k ≥ 2, δ = ±1.
If such a factorization (7.4) is impossible, we set Mb = Lb = ∞.
Assume that a factorization (7.4) exists (its uniqueness is obvious). Since
U2 = W (i + |U1 | + 1, |U2 |),
|U2 |ā1 < |U |ā1 ,
0
it follows that the numbers λ(U2 ), µ3 (U2 ) are available. Denote U1 U3 = aℓ1 .
If k > 2 and λ(U2 ) = 0, we set
Lb := ℓ,
Mb := µ3 (U2 ).
If k > 2 and λ(U2 ) 6= 0, we set Lb = Mb := ∞.
for
62
S. V. IVANOV
If k = 2 and δ = 1, we check whether λ(U2 ) is divisible by n1 . If so, we set
Lb := ℓ + λ(U2 )n2 /n1 ,
Mb := µ3 (U2 ) + |λ(U2 )/n1 |.
If k = 2, δ = 1, and λ(U2 ) is not divisible by n1 , we set Lb = Mb := ∞.
If k = 2 and δ = −1, we check whether λ(U2 ) is divisible by n2 . If so, we set
Lb := ℓ + λ(U2 )n1 /n2 ,
Mb := µ3 (U2 ) + |λ(U2 )/n2 |.
If k = 2, δ = −1, and λ(U2 ) is not divisible by n2 , we set Lb = Mb := ∞.
G
′
It follows from Lemma 7.2 applied to the word U = W (i, j) that if U =3 aℓ1 for
some integer ℓ′ , then λ(U ) = min(La , Lb ) and µ3 (U ) = min(Ma , Mb ). On the other
hand, it is immediate that if λ(U ) = ∞, then La = Lb = ∞ and Ma = Mb = ∞.
Therefore, in all cases, we obtain that
λ(U ) = min(La , Lb ),
µ3 (U ) = min(Ma , Mb ).
This completes our computation of the numbers λ(U ), µ3 (U ) for U = W (i, j).
It follows by the above inductive argument that, for every word W (i, j) with
finite λ(W (i, j)), we have
max(|λ(W (i, j))|, µ3 (W (i, j))) ≤ |W (i, j)| max(|n1 |, |n2 |)|W (i,j)|ā1 ≤ 2O(|W |) . (7.5)
Hence, using binary representation for numbers λ(W (i′ , j ′ )), µ3 (W (i′ , j ′ )), we can
run the foregoing computation of numbers λ(U ), µ3 (U ) for U = W (i, j) in time
O(|W (i, j)|2 ), including division of λ(U2 ) by n1 and by n2 to decide whether
λ(U2 )/n1 , λ(U2 )/n2 are integers and including additions to compute the minimum
(7.3).
Since the number of subwords W (i, j) of W is bounded by O(|W |2 ), it follows that
running time of the described algorithm that computes the numbers λ(W ), µ3 (W )
is O(|W |4 ). This implies that both the bounded word problem and the precise word
problem for presentation (1.4) can be solved in deterministic time O(|W |4 ).
Theorem 1.4 is proved.
8. Minimizing Diagrams over (1.2) and Proofs of Theorem 1.5 and
Corollary 1.6
Theorem 1.5. Both the diagram problem and the minimal diagram problem for
group presentation (1.2) can be solved in deterministic space O((log |W |)3 ) or in
deterministic time O(|W |4 log |W |).
G
Furthermore, let W be a word such that W =2 1 and let
τ (∆) = (τ1 (∆), . . . , τsτ (∆))
be a tuple of integers, where the absolute value |τi (∆)| of each τi (∆) represents
the number of certain vertices or faces in a disk diagram ∆ over (1.2) such that
ϕ(∂∆) ≡ W . Then, in deterministic space O((log |W |)3 ), one can algorithmically
construct such a minimal diagram ∆ which is also smallest relative to the tuple
τ (∆) (the tuples are ordered lexicographically).
Proof. We start by proving the space part of Theorem 1.5.
Let W be a nonempty word over A±1 and let Ω be a finite sequence of elementary
operations that transforms the empty pseudobracket system for W into a final
G
pseudobracket system. By Lemma 4.7, W =2 1 and there is a diagram ∆ over
(1.2) such that ϕ(∂|0 ∆) ≡ W . We now describe an explicit construction of such
BOUNDED AND PRECISE WORD PROBLEMS FOR GROUP PRESENTATIONS
63
a diagram ∆ = ∆(Ω) with property (A) based on the sequence Ω. Note that this
construction is based on the proof of Lemma 4.7 and essentially follows that proof.
The question on how to construct the sequence Ω will be addressed later.
Let B0 , B1 , . . . , Bℓ be the sequence of pseudobracket systems associated with Ω,
where B0 is empty and Bℓ is final. We will construct ∆(Ω) by induction on i. Let
Bi be obtained from Bi−1 , i ≥ 1, by an elementary operation σ and let b′ ∈ Bi be
the pseudobracket obtained from b, c ∈ Bi−1 by σ. Here we assume that one of b, c
(or both) could be missing.
If σ is an addition, then we do not do anything.
Assume that σ is an extension of type 1 or type 2 on the left and a(b′ ) = e1 a(b),
1
where e1 is an edge of PW and ϕ(e1 ) = aεb(3)
. Here we use the notation of the
definition of an extension of type 1 or type 2 on the left. Then we construct ∆(Ω)
so that the edge α(e1 ) = e belongs to the boundary ∂∆(Ω) of ∆(Ω) and e−1 belongs
−ε b(4)
to ∂Π, where Π is a face of ∆(Ω) such that ϕ(∂Π) ≡ ab(3)1 . Note that the label
ϕ(∂Π) is uniquely determined by either of the pseudobrackets b, b′ .
The case of an extension of type 1 or type 2 on the right is analogous.
Suppose that σ is an extension of type 3 and a(b′ ) = e1 a(b)e2 , where e1 , e2 are
edges of PW with ϕ(e1 ) = ϕ(e2 )−1 . Here we use the notation of the definition of
an extension of type 3. Then we construct ∆(Ω) so that α(e1 ) = α(e2 )−1 and both
α(e1 ), α(e2 ) belong to the boundary ∂∆(Ω) of ∆(Ω).
If σ is a turn or a merger, then we do not do anything to the diagram ∆(Ω)
under construction.
Performing these steps during the process of elementary operations Ω, we will
construct, i.e., effectively describe, a required diagram ∆(Ω) such that |∆(Ω)(2)| =
bF (6), where bF is the final pseudobracket of Bℓ , and ϕ(∂|0 ∆(Ω)) ≡ W . Note that
this particular diagram ∆(Ω), depending on the sequence Ω, may not be minimal or
even reduced, however, as in the proof of Lemma 4.7, our construction guarantees
that ∆(Ω) does have the property (A), i.e., the property that if e is an edge of the
boundary path ∂Π of a face Π of ∆(Ω), then e−1 ∈ ∂∆(Ω).
We also observe that, in the proof of Lemma 4.6, for a given pair (W, ∆W ), where
∆W is a diagram over (1.2) such that ϕ(∂|0 ∆W ) ≡ W , we constructed by induction
an operational sequence Ω = Ω(∆W ) of elementary operations that convert the
empty bracket system into the final bracket system and has size bounded by
(11|W |, C(log |W | + 1))).
(8.1)
Recall that every bracket system is also a pseudobracket system. Looking at
details of the construction of the sequence Ω(∆W ), we can see that if we reconstruct
a diagram ∆(Ω(∆W )), by utilizing our above algorithm based upon the sequence
Ω(∆W ), then the resulting diagram will be identical to the original diagram ∆W ,
hence, in this notation, we can write
∆(Ω(∆W )) = ∆W .
(8.2)
Recall that Savitch’s conversion [38], see also [3], [32], of nondeterministic computations in space S and time T into deterministic computations in space O(S log T )
simulates possible nondeterministic computations by using tuples of configurations,
also called instantaneous descriptions, of cardinality O(log T ). Since a configuration takes space S and the number of configurations that are kept in memory at
every moment is at most O(log T ), the space bound O(S log T ) becomes evident.
64
S. V. IVANOV
Furthermore, utilizing the same space O(S log T ), one can use Savitch’s algorithm
to compute an actual sequence of configurations that transform an initial configuration into a final configuration. To do this, we consider every configuration ψ
together with a number k, 0 ≤ k ≤ T , thus replacing every configuration ψ by
a pair (ψ, k), where k plays the role of counter. We apply Savitch’s algorithm to
such pairs (ψ, k), so that (ψ2 , k + 1) is obtained from (ψ1 , k) by a single operation
whenever ψ2 is obtained from ψ1 by a single command. If ψ is a final configuration,
then we also assume that (ψ, k + 1) can be obtained from (ψ, k), 0 ≤ k ≤ T − 1, by
a single operation.
Now we apply Savitch’s algorithm to two pairs (ψ0 , 0), (ψF , T ), where ψ0 is
an initial configuration and ψF is a final configuration. If Savitch’s algorithm
accepts these two pairs (ψ0 , 0), (ψF , T ), i.e., the algorithm turns the pair (ψ0 , 0)
into (ψF , T ), then there will be a corresponding sequence
(ψ0 , 0), (ψ1 , 1), . . . , (ψi , i), . . . , (ψF , T )
(8.3)
of pairs such that (ψi , i) is obtained from (ψi−1 , i − 1), i ≥ 1, by a single operation.
Observe that we can retrieve any term, say, the kth term of the sequence (8.3), by
rerunning Savitch’s algorithm and memorizing a current pair (ψ, k) in a separate
place in memory. Note that the current value of (ψ, k) may change many times
during computations but in the end it will be the desired pair (ψk , k). Clearly, the
space needed to run this modified Savitch’s algorithm is still O(S log T ) and, consecutively outputting these pairs (ψ0 , 0), (ψ1 , 1), . . . , we will construct the required
sequence (8.3) of configurations.
Coming back to our specific situation, we recall that a pseudobracket system
plays the role of a configuration, the empty pseudobracket system is an initial
configuration and a final pseudobracket system is a final configuration. Therefore,
it follows from Lemmas 4.6–4.7, and the foregoing equality (8.2) that picking the
empty pseudobracket system B0 and a final pseudobracket system {bF }, where
bF = (0, |W |, 0, 0, 0, n′), we can use Savitch’s algorithm to verify in deterministic
G
space O((log |W |)3 ) whether W =2 1 and whether there is a disk diagram ∆ over
(1.2) such that ϕ(∂∆) ≡ W and |∆(2)| = n′ . In addition, if the algorithm accepts,
then the algorithm also constructs an operational sequence B0 , B1 , . . . , Bℓ = {bF } of
pseudobracket systems of size bounded by (8.1) which, as was discussed above, can
be used to construct a disk diagram ∆1 with property (A) such that ϕ(∂∆1 ) ≡ W
and |∆1 (2)| = n′ . It is clear that the entire construction of ∆1 can be done in
deterministic space O((log |W |)3 ), as desired.
To construct a minimal disk diagram for W , we consecutively choose n′ =
0, 1, 2, . . . and run Savitch’s algorithm to determine whether the algorithm can
turn the empty pseudobracket system into a final pseudobracket system {bF },
where bF (6) = n′ . If n0 is the minimal such n′ , then we run Savitch’s algorithm
again to construct an operational sequence of pseudobracket systems and a corresponding disk diagram ∆0 such that ϕ(∂∆0 ) ≡ W and |∆0 (2)| = n0 . Similarly
to the above arguments, it follows from Lemmas 4.6–4.7, and from the equality
∆(Ω(∆W )) = ∆W , see (8.2), that ∆0 is a minimal diagram and the algorithm
uses space O((log |W |)3 ). Thus, the minimal diagram problem can be solved for
presentation (1.2) in deterministic space O((log |W |)3 ).
BOUNDED AND PRECISE WORD PROBLEMS FOR GROUP PRESENTATIONS
65
Now we will discuss the additional statement of Theorem 1.5. For a diagram
∆ over presentation (1.2) such that ϕ(∂|0 ∆) ≡ W and ∆ has property (A), we
consider a parameter
τ (∆) = (τ1 (∆), . . . , τsτ (∆)),
(8.4)
where τi (∆) are integers that satisfy 0 ≤ |τi (∆)| ≤ Cτ |W |, Cτ > 0 is a fixed
constant, and that represent numbers of certain vertices, edges, faces of ∆ (we
may also consider some functions of the numbers of certain vertices, edges, faces
of ∆ that are computable in space O((log |W |)3 ) and whose absolute values do not
exceed Cτ |W |).
For example, if we set τ1 (∆) := |∆(2)| then we would be constructing a minimal
diagram which is also smallest with respect to τ2 (∆), then smallest with respect to
τ3 (∆) and so on. Let us set τ2 (∆) := |∆(2)a±n1 |, where |∆(2)a±n1 | is the number
1
1
of faces Π2 in ∆ such that ϕ(∂Π2 ) ≡ a1±n1 , τ3 (∆) := |∆(2)an2 2 |, where |∆(2)an2 2 | is
2
the number of faces Π3 in ∆ such that ϕ(∂Π3 ) ≡ an2P
.
We may also consider functions such as τ4 (∆) := k4,1 ≤|∂Π|≤k4,2 |∂Π|, where the
summation takes place over all faces Π in ∆ such that k4,1 ≤ |∂Π| ≤ k4,2 , where
k4,1 , k4,2 are fixed integers with 0 ≤ k4,1 ≤ k4,2 ≤ |W |. Note that the latter bound
follows from Lemma 2.2(a).
If v is a vertex of ∆, let deg v denote the degree of v, i.e., the number of oriented
edges e of ∆ such that e+ = v. Also, if v is a vertex of ∆, let degF v denote the
face degree of v, i.e., the number of faces Π in ∆ such that v ∈ ∂Π. Note that it
follows from Lemma 2.2(a) that every face Π of a disk diagram ∆ with property
(A) can contribute at most 1 to this sum.
We may also consider entries in τ (∆) that count the number of vertices in ∆ of
certain degree. Note that doing this would be especially interesting and meaningful
when the presentation (1.2) contains no defining relations, hence, diagrams over
(1.2) are diagrams over the free group F (A) = hA k ∅i with no relations. For
instance, we may set τ5 (∆) := |∆(0)≥3 |, where |∆(0)≥3 | is the number of vertices
F
in ∆ of degree at least 3, and set τ6 (∆) := |∆(0)F
[k6,1 ,k6,2 ] |, where |∆(0)[k6,1 ,k6,2 ] | is
the number of vertices v in ∆ such that k6,1 ≤ degF v ≤ k6,2 , where k6,1 , k6,2 are
fixed integers with 0 ≤ k6,1 ≤ k6,2 ≤ |W |.
Our goal is to make some modification of the foregoing algorithm that would
enable us to construct a disk diagram ∆ over (1.2) such that ϕ(∂|0 ∆W ) ≡ W ,
∆ has property (A), and ∆ is smallest relative to the parameter τ (∆). Recall
that the tuples τ (∆) are ordered in the standard lexicographical way. Note that if
τ1 (∆) = −|∆(2)| then ∆ would be a diagram for W with property (A) which has
the maximal number of faces.
Our basic strategy remains unchanged, although we will need more complicated
bookkeeping (to be introduced below). We start by picking a value for the paramτi | ≤ Cτ |W |. As before,
eter τe = (e
τ1 , . . . , τesτ ), where τei are fixed integers with |e
using Lemmas 4.6, 4.7, and the identity ∆(Ω(∆W )) = ∆W , see (8.2), we can utilize
Savitch’s algorithm which verifies, in deterministic space O((log |W |)3 ), whether
the empty pseudobracket system B0 can be turned by elementary operations into
a final pseudobracket system B = {bF } which also changes the τ -parameter from
(0, 0, . . . , 0) to τe. If this is possible, then the algorithm also computes an operational sequence B0 , B1 , . . . , Bℓ = B of pseudobracket systems and, based on this
sequence, constructs a disk diagram ∆ such that ϕ(∂∆) ≡ W , ∆ has property (A)
66
S. V. IVANOV
and τ (∆) = τe. To make sure that ∆ is smallest relative to the parameter τ (∆), we
can use Savitch’s algorithm for every τe′ < τe to verify that the empty pseudobracket
system B0 may not be turned by elementary operations into a final pseudobracket
system B = {bF } so that the τ -parameter concurrently changes from (0, 0, . . . , 0) to
τe′ . This concludes the description of our algorithm for construction of a diagram ∆
over (1.2) such that ϕ(∂|0 ∆W ) ≡ W , ∆ has property (A) and ∆ is smallest relative
to the parameter τ (∆).
Let us describe necessary modifications in bookkeeping. First, we add six more
integer entries to each pseudobracket b. Hence, we now have
b = (b(1), . . . , b(6), b(7), . . . , b(12)),
where b(1), . . . , b(6) are defined exactly as before, b(7), . . . , b(12) are integers such
that
b(7) + b(8) = b(5),
0 ≤ b(7), b(8) ≤ b(5) if b(5) ≥ 0,
(8.5)
b(7) + b(8) = b(5),
b(5) ≤ b(7), b(8) ≤ 0
(8.6)
if
b(5) ≤ 0.
The numbers b(9), . . . , b(12) satisfy the inequalities
0 ≤ b(9), . . . , b(12) ≤ |W |
and the equalities b(7) = b(8) = b(11) = b(12) = 0 whenever b has type PB1. The
purpose of these new entries, to be specified below, is to keep track of the information on degrees and face degrees of the vertices of the diagram ∆(Ω) over (1.2)
being constructed by means of an operational sequence Ω of elementary operations.
We now inductively describe the changes over the entries b(7), . . . , b(12) and over
the parameter τ (∆) = (τ1 (∆), . . . , τsτ (∆)) that are done in the process of performance of an operational sequence Ω of elementary operations over pseudobracket
systems that change the empty pseudobracket system into a final pseudobracket
G
system for W , where W =2 1 and |W | > 0.
To make the inductive definitions below easier to understand, we will first make
informal remarks about meaning of the new entries b(7), . . . , b(12).
Suppose that b is a pseudobracket of type PB1. Then b(7) = b(8) = 0 and
b(11) = b(12) = 0. The entry b(9) represents the current (or intermediate) degree
deg v of the vertex v = α(b(1)) = α(b(2)) which is subject to change in process
of construction of the diagram ∆(Ω). The entry b(10) represents the current face
degree degF v of the vertex v = α(b(1)) = α(b(2)) which is also subject to change.
For example, if a pseudobracket b′ is obtained from pseudobrackets b, c of type
PB1 by a merger, then b′ (9) := b(9) + c(9) and b′ (10) := b(10) + c(10). Note
that b′ (9), b′ (10) are still intermediate degrees of the vertex α(b′ (1)) = α(b′ (2)),
i.e., b′ (9), b′ (10) are not actual degrees of α(b′ (1)), because there could be more
elementary operations such as extension of type 3, mergers, and turns that could
further increase b′ (9), b′ (10).
Assume that b is a pseudobracket of type PB2. Then the entry b(7) represents
b(7)
the exponent in the power ab(3) ≡ ϕ(u−1
1 ), where u1 is the arc of ∂Π, where Π is a
face such that (u1 )− = α(v), (u1 )+ = α(b(1)), see Fig. 8.1, and v is the vertex of
PW at which a turn operation was performed to get a pseudobracket b̄ of type PB2
BOUNDED AND PRECISE WORD PROBLEMS FOR GROUP PRESENTATIONS
67
which, after a number of extensions of type 1 and mergers with pseudobrackets of
type PB1, becomes b.
b(7)
ϕ(u−1
1 ) ≡ ab(3)
u1
α(v)
u2
u = u2 u1
α(b(1))
b(8)
ϕ(u−1
2 ) ≡ ab(3)
α(b(2))
Π
Fig. 8.1
b(8)
Similarly, the entry b(8) represents the exponent in the power ab(3) ≡ ϕ(u−1
2 ), where
u2 is the arc of ∂Π such that (u2 )− = α(b(2)), (u1 )+ = α(v), and v is the vertex
b(5)
of PW defined as above, see Fig. 8.1. Since ϕ(u−1 ) ≡ ab(3) , where u = u2 u1 is the
arc of ∂Π defined by u− = α(b(2)), u+ = α(b(1)), see Fig. 8.1, it follows from the
definitions that the conditions (8.5)–(8.6) hold true.
As in the above case when b has type PB2, the entry b(9) represents the current
(or intermediate) degree deg v1 of the vertex v1 = α(b(1)) which is subject to change
in the process of construction of ∆(Ω). As above, the entry b(10) represents the
current face degree degF v1 of v1 = α(b(1)) which is also subject to change. The
entry b(11) is the intermediate degree deg v2 of the vertex v2 = α(b(2)) and b(12)
represents the current face degree degF v2 of the vertex v2 = α(b(2)).
Our description for meaning of the new entries b(7), . . . , b(12) in thus augmented
pseudobracket b is complete and we now inductively describe the changes over these
new entries b(7), . . . , b(12) and over the parameter τ (∆) in process of performance
of a sequence Ω of elementary operations over pseudobracket systems that change
the empty pseudobracket system into a final pseudobracket system for a word W
G
such that W =2 1 and |W | > 0.
As above, assume that Ω is a sequence of elementary operations that change
the empty pseudobracket system into a final pseudobracket system for W . Let
B0 , B1 , . . . , Bℓ be the corresponding to Ω sequence of pseudobracket systems, where
B0 is empty and Bℓ is final.
If b ∈ Bi+1 , i ≥ 0, is a starting pseudobracket, then we set
b(7) = b(8) = b(9) = b(10) = b(11) = b(12) = 0.
As was mentioned above, the entries b(1), . . . , b(6) in every pseudobracket b ∈
∪ℓj=0 Bj are defined exactly as before.
Suppose that a pseudobracket b′ ∈ Bi+1 , i ≥ 1, is obtained from b ∈ Bi by an
extension of type 1 on the left and a(b′ ) = e1 a(b), where a(b′ ), a(b) are the arcs
1
, ε1 = ±1. Recall that both
of b′ , b, resp., and e1 is an edge of PW , ϕ(e1 ) = aεb(3)
′
pseudobrackets b , b have type PB2.
First we assume that b(5) 6= 0, i.e., one of b(7), b(8) is different from 0. Then we
set
b′ (7) := b(7) + ε1 ,
b′ (10) := 1,
b′ (8) := b(8),
b′ (11) := b(11),
b′ (9) := 2,
b′ (12) := b(12).
Note that it follows from the definitions that |b′ (7)| = |b(7)| + 1. We also update
those entries in the sequence τ (∆) that are affected by the fact that v1 = α(b(1)) is
68
S. V. IVANOV
now known to be a vertex of the diagram ∆(Ω) (which is still under construction)
such that deg v1 = b(9) and degF v1 = b(10).
For example, if τ5 (∆) = |∆(0)≥3 |, where |∆(0)≥3 | is the number of vertices in
∆ of degree at least 3, and b(9) ≥ 3, then τ5 (∆) is increased by 1. If τ6 (∆) =
F
|∆(0)F
[k6,1 ,k6,2 ] |, where |∆(0)[k6,1 ,k6,2 ] | is the number of vertices u in ∆ such that
k6,1 ≤ degF u ≤ k6,2 , where k6,1 , k6,2 are fixed integers with 0 ≤ k6,1 ≤ k6,2 ≤ |W |,
and if degF α(b(1)) = b(10) satisfies k6,1 ≤ b(10) ≤ k6,2 , then we increase τ6 (∆) by
1.
Suppose that b(5) = 0, i.e., b(7) = b(8) = 0. Then we set
b′ (7) := ε1 ,
b′ (8) := 0,
b′ (9) := 2,
b′ (10) := 1,
b′ (11) := b(9),
b′ (12) := b(10).
These formulas, see also (8.7)–(8.8) and other formulas below, reflect the convention
that, for a pseudobracket b of type PB2, the information about edge and face
degrees of the vertex α(b(2)) is stored in components b(9), b(10), resp., whenever
α(b(1)) = α(b(2)), i.e., b(5) = 0. However, when α(b(1)) 6= α(b(2)), i.e., b(5) 6= 0,
this information for the vertex α(b(2)) is separately kept in components b(11), b(12),
whereas this information for the vertex α(b(1)) is stored in components b(9), b(10).
In the current case b(5) = 0, no change to τ (∆) is done.
The case of an extension of type 1 on the left is complete.
Assume that a pseudobracket b′ ∈ Bi+1 , i ≥ 1, is obtained from b ∈ Bi by an
extension of type 1 on the right. This case is similar to its “left” analogue studied
above but not quite symmetric, because of the way we keep information about
degrees of vertices and we will write down all necessary formulas. As above, both
pseudobrackets b′ , b have type PB2.
Let a(b′ ) = a(b)e2 , where a(b′ ), a(b) are the arcs of b′ , b, resp., and e2 is an edge
2
, ε2 = ±1. If b(5) 6= 0, then we define
of PW , ϕ(e2 ) = aεb(3)
b′ (7) := b(7),
′
b (10) := b(10),
b′ (8) := b(8) + ε2 ,
′
b (11) := 2,
′
b′ (9) := b(9),
b (12) := 1.
(8.7)
(8.8)
Note that it follows from the definitions that |b′ (8)| = |b(8)| + 1. We also update
those entries in the sequence τ (∆) that are affected by the fact that v2 = α(b(2))
is a vertex of ∆(Ω) (which is still under construction) such that deg v2 = b(11) and
degF v2 = b(12).
For example, if τ5 (∆) = |∆(0)≥3 | and b(11) ≥ 3, then τ5 (∆) is increased by
F
1. If τ6 (∆) = |∆(0)F
[k6,1 ,k6,2 ] | and if deg α(b(2)) = b(12) satisfies the inequalities
k6,1 ≤ b(12) ≤ k6,2 , then we increase τ6 (∆) by 1.
If b(5) = 0, then, to define b′ , we again use formulas (8.7)–(8.8) but we do not
make any changes to τ (∆).
The case of an extension of type 1 on the right is complete.
Suppose that a pseudobracket b′ ∈ Bi+1 , i ≥ 1, is obtained from b ∈ Bi by an
extension of type 2 on the left. Denote a(b′ ) = e1 a(b), where a(b′ ), a(b) are the arcs
1
, ε1 = ±1. Recall that b has
of b′ , b, resp., and e1 is an edge of PW , ϕ(e1 ) = aεb(3)
′
type PB2 and b has type PB1.
BOUNDED AND PRECISE WORD PROBLEMS FOR GROUP PRESENTATIONS
69
First we assume that b(5) 6= 0, i.e., one of b(7), b(8) is different from 0. Then we
set
b′ (7) := 0,
b′ (8) := 0,
b′ (10) := b(12),
b′ (9) := b(11),
b′ (11) := 0,
b′ (12) := 0.
We also update the sequence τ (∆) according to the information that v1 = α(b(1))
is a vertex of ∆(Ω) (which is still under construction) such that deg v1 = b(9),
−ε b(4)
degF v1 = b(10), and that ∆(Ω) contains a face Πb such that ϕ(∂Πb ) ≡ ab(3)1 .
For example, if τ5 (∆) = |∆(0)≥3 | and b(9) ≥ 3, then τ5 (∆) is increased by 1.
F
If τ6 (∆) = |∆(0)F
[k6,1 ,k6,2 ] | and if deg α(b(1)) = b(10) satisfies k6,1 ≤ b(10) ≤ k6,2 ,
then we increase τ6 (∆) by 1. If τ2 (∆) = |∆(2)a±n1 |, where |∆(2)a±n1 | is the number
1
1
of faces Π2 in ∆ such that ϕ(∂Π2 ) ≡ a1±n1 , and b(3) = 1, b(4) = n1 , then we increase
τ2 (∆) by 1. If τ3 (∆) = |∆(2)an2 2 |, where |∆(2)an2 2 | is the number of faces Π3 in
n2
∆ such that ϕ(∂ΠP
3 ) ≡ a2 , and b(3) = 2, −ε1 b(4) = n2 , then we increase τ3 (∆)
by 1. If τ4 (∆) = k4,1 ≤|∂Π|≤k4,2 |∂Π|, where the summation takes place over all
faces Π in ∆ such that k4,1 ≤ |∂Π| ≤ k4,2 , where k4,1 , k4,2 are fixed integers with
0 ≤ k4,1 ≤ k4,2 ≤ |W |, and k4,1 ≤ b(4) ≤ k4,2 , then we increase τ4 (∆) by 1.
Now assume that b(5) = 0, i.e., b(7) = b(8) = 0, and so b(4) = 1. Then we set
b′ (7) := 0,
b′ (8) := 0,
b′ (10) := b(10),
b′ (9) := b(9),
b′ (11) := 0,
b′ (12) := 0.
We also update the tuple τ (∆) according to the information that ∆(Ω) contains a
−ε1
.
face Πb such that ϕ(∂Πb ) ≡ ab(3)
The case of an extension of type 2 on the left is complete.
Suppose that a pseudobracket b′ ∈ Bi+1 , i ≥ 1, is obtained from b ∈ Bi by an
extension of type 2 on the right. This case is similar to its “left” analogue discussed
above but is not quite symmetric and we will write down all necessary formulas.
Denote a(b′ ) = a(b)e2 , where a(b′ ), a(b) are the arcs of b′ , b, resp., and e2 is an
2
, ε2 = ±1. Recall that b has type PB2 and b′ has type
edge of PW , ϕ(e2 ) = aεb(3)
PB1.
First we assume that b(5) 6= 0. Then we set
b′ (7) := 0,
′
b′ (8) := 0,
b (10) := b(10),
′
b′ (9) := b(9),
b (11) := 0,
′
b (12) := 0.
(8.9)
(8.10)
We also update the sequence τ (∆) according to the information that v2 = α(b(2))
is a vertex of ∆(Ω) (which is still under construction) such that deg v2 = b(11),
−ε b(4)
degF v2 = b(12), and that ∆(Ω) contains a face Πb such that ϕ(∂Πb ) ≡ ab(3)2 .
For example, if τ5 (∆) = |∆(0)≥3 | and b(11) ≥ 3, then τ5 (∆) is increased by 1.
F
If τ6 (∆) = |∆(0)F
[k6,1 ,k6,2 ] | and if deg α(b(2)) = b(12) satisfies k6,1 ≤ b(12) ≤ k6,2 ,
then we increase τ6 (∆) by 1. If τ2 (∆) = |∆(2)a±n1 |, as above, and b(3) = 1, b(4) =
1
n1 , then we increase τ2 (∆) by 1. If τ3 (∆) = |∆(2)an2 2 |, as above, and b(3) = 2,
P
−ε1 b(4) = n2 , then we increase τ3 (∆) by 1. If τ4 (∆) := k4,1 ≤|∂Π|≤k4,2 |∂Π|, as
above, and k4,1 ≤ b(4) ≤ k4,2 , then we increase τ4 (∆) by 1.
70
S. V. IVANOV
Now we assume that b(5) = 0, i.e., b(7) = b(8) = 0, and so b(4) = 1. Then,
to define b, we again use formulas (8.9)–(8.10). We also update the tuple τ (∆)
−ε2
.
according to the information that ∆(Ω) contains a face Πb with ϕ(∂Πb ) ≡ ab(3)
The case of an extension of type 2 on the right is complete.
Suppose that a pseudobracket b′ ∈ Bi+1 , i ≥ 1, results from b ∈ Bi by an
extension of type 3. Denote a(b′ ) = e1 a(b)e2 , where a(b′ ), a(b) are the arcs of b′ , b,
resp., and e1 , e2 are edges of PW , ϕ(e1 ) = ϕ(e2 )−1 = aεj , ε = ±1. Recall that both
b and b′ have type PB1. Then we set
b′ (7) := 0,
b′ (8) := 0,
b′ (10) := 0,
b′ (9) := 1,
b′ (11) := 0,
b′ (12) := 0.
We also update the tuple τ (∆) according to the information that v = α(b(1)) =
α(b(2)) is a vertex of ∆(Ω) (which is still under construction) such that deg v =
b(9) + 1 and degF v = b(10).
For example, if τ5 (∆) = |∆(0)≥3 | and b(9) ≥ 3, then τ5 (∆) is increased by 1. If
F
τ6 (∆) = |∆(0)F
[k6,1 ,k6,2 ] | and if deg v = b(10) satisfies k6,1 ≤ b(10) ≤ k6,2 , then we
increase τ6 (∆) by 1.
The case of an extension of type 3 is complete.
Assume that a pseudobracket b′ ∈ Bi+1 , i ≥ 1, is obtained from b ∈ Bi by a turn
operation. Recall that b has type PB1 and b′ has type PB2. Then we set
b′ (7) := 0,
b′ (8) := 0,
b′ (10) := b(10) + 1,
b′ (9) := b(9) + 2,
b′ (11) := 0,
b′ (12) := 0.
No change over τ (∆) is necessary under a turn operation. The case of a turn
operation is complete.
Suppose that a pseudobracket b′ ∈ Bi+1 , i ≥ 1, is obtained from b, c ∈ Bi by a
merger operation. Without loss of generality, we assume that the pseudobrackets
b, c which satisfy the condition b(2) = c(1), i.e., b is on the left of c. Recall that
one of b, c must have type PB1 and the other one has type PB1 or PB2. Consider
three cases corresponding to the types of the pseudobrackets b, c.
First assume that both b, c have type PB1. Then we set
b′ (7) := 0,
b′ (8) := 0,
b′ (9) := b(9) + c(9),
b′ (10) := b(10) + c(10),
b′ (11) := 0,
b′ (12) := 0.
No change over τ (∆) is made.
Assume that b has type PB1 and c has type PB2. Then, keeping in mind that
b(2) = c(1), we set
b′ (7) := c(7),
b′ (8) := c(8),
b′ (10) := b(10) + c(10),
b′ (9) := b(9) + c(9),
b′ (11) := c(11),
b′ (12) := c(12).
As above, no change over τ (∆) is necessary.
Assume that b has type PB2 and c has type PB1. If b(5) 6= 0, then we set
b′ (7) := b(7),
b′ (10) := b(10),
b′ (8) := b(8),
b′ (9) := b(9),
b′ (11) := b(11) + c(9),
b′ (12) := b(12) + c(10).
BOUNDED AND PRECISE WORD PROBLEMS FOR GROUP PRESENTATIONS
71
On the other hand, if b(5) = 0, i.e., b(7) = b(8) = 0, then we set
b′ (7) := 0,
b′ (8) := 0,
b′ (10) := b(10) + c(10),
b′ (9) := b(9) + c(9),
b′ (11) := 0,
b′ (12) := 0.
As before, no change over τ (∆) is made under a merger operation.
The case of a merger operation is complete.
Our inductive definitions of extended pseudobrackets and modifications of τ (∆)
are complete. To summarize, we conclude that the described changes to the extended pseudobrackets and to the tuple τ (∆) will guarantee that, if our nondeterministic algorithm, which is based on Lemma 4.8 and which follows a sequence Ω of
elementary operations as above, accepts a pair of configurations ψ0 , ψF , where ψ0
is the empty pseudobracket system and ψF is a final pseudobracket system, then
the final tuple τ (∆), i.e., the tuple associated with ψF , will be equal to the tuple
τ (∆(Ω)) which represents the tuple of actual parameters τ1 (∆(Ω)), . . . , τs (∆(Ω))
of the diagram ∆(Ω).
Consider augmented configurations ψ̄ = (B, τ (∆)), corresponding to a sequence
Ω of elementary operations as above, where B is a system of extended pseudobrackets and τ (∆) is the tuple associated with B. Recall that, by Lemma 4.8, we
may assume that Ω has size bounded by (11|W |, C(log |W | + 1)). As in the proof
of Theorem 1.2, we note that the entries b(1), . . . , |b(5)|, b(6) are nonnegative and
bounded by max(|W |, m). It follows from the definitions and Lemma 2.2(a) that
for the new entries b(7), . . . , b(12), we have that |b(7)|, |b(8)| ≤ |b(5)| ≤ |W | and
0 ≤ b(9), b(10), b(11), b(12) ≤ 2|W |.
By the definition, every entry τi (∆) in τ (∆) satisfies |τi (∆)| ≤ Cτ |W |, i = 1, . . . , sτ .
Since sτ , Cτ are constants, we conclude that the space needed to store an augmented configuration ψ̄ = (B, τ (∆)) is O((log |W |)2 ). Therefore, utilizing Savitch’s
algorithm as before, which now applies to augmented configurations of the form
ψ̄ = (B, τ (∆)), we will be able to find out, in deterministic space O((log |W |)3 ),
whether the algorithm accepts a pair (B0 , τ 0 (∆)), (BF , τ F (∆)), where B0 is empty,
τ 0 (∆) consists of all zeros, BF is a final pseudobracket system, and τ F (∆) is a final
tuple corresponding to BF . Since the space needed to store a final configuration (BF , τ F (∆)) is O(log |W |), we will be able to compute, in deterministic space
O((log |W |)3 ), a lexicographically smallest tuple τbF (∆)) relative to the property
bF , τbF (∆)), where B
bF is
that the pair of augmented configurations (B0 , τ 0 (∆)), (B
some final pseudobracket system, is accepted by Savitch’s algorithm. Now we can
do the same counter trick as in the beginning of this proof, see (8.3), to compute, in
deterministic space O((log |W |)3 ), a sequence Ω∗ of elementary operations and the
corresponding to Ω∗ sequence of augmented configurations which turns (B0 , τ 0 (∆))
bF , τbF (∆)) and which is constructed by Savitch’s algorithm. Finally, as in
into (B
the beginning of the proof of Theorem 1.5, using the sequence Ω∗ , we can construct
a desired diagram ∆∗ = ∆(Ω∗ ) so that τ (∆∗ ) = τbF (∆) and this construction can
be done in deterministic space O((log |W |)3 ). This completes the proof of the space
part of Theorem 1.5.
To prove the time part of Theorem 1.5, we review the proof of P part of Theorem 1.2. We observe that our arguments enable us, concurrently with computation
of the number µ2 (W [i, j, k, l]) for every parameterized word W [i, j, k, l] ∈ S2 (W )
72
S. V. IVANOV
G
such that W (i, j)aℓk =2 1, to inductively construct a minimal diagram ∆[i, j, k, l]
over (1.2) such that ϕ(∂|0 ∆[i, j, k, l]) ≡ W (i, j)aℓk . Details of this construction
are straightforward in each of the subcases considered in the proof of P part of
Theorem 1.2. Note that the time needed to run this extended algorithm is still
O(|W |4 log |W |).
Theorem 1.5 is proven.
Corollary 1.6. There is a deterministic algorithm that, for given word W over the
F(A)
alphabet A±1 such that W = 1, where F(A) = hA k ∅i is the free group over A,
constructs a pattern of cancellations of letters in W that result in the empty word
and the algorithm operates in space O((log |W |)3 ).
Furthermore, let ∆ be a disk diagram over F(A) that corresponds to a pattern of
cancellations of letters in W , i.e., ϕ(∂∆) ≡ W , and let
τ (∆) = (τ1 (∆), . . . , τsτ (∆))
be a tuple of integers, where the absolute value |τi (∆)| of each τi (∆) represents
the number of vertices in ∆ of certain degree. Then, also in deterministic space
O((log |W |)3 ), one can algorithmically construct such a diagram ∆ which is smallest
relative to the tuple τ (∆).
Proof. Corollary 1.6 is immediate from Theorem 1.5 and does not need a separate proof. Nevertheless, it is worth mentioning that, in the case of presentation
F(A) = hA k ∅i of the free group F(A) over A, our definitions of brackets, pseudobrackets, elementary operations and subsequent arguments become significantly
simpler. Since there are no relations, we do not define brackets of type B2, nor
we define pseudobrackets of type PB2. In particular, for every bracket or pseudobracket b, the entries b(3), b(4), b(5) are always zeroes and could be disregarded. In
the extended version of a pseudobracket b, defined for minimization of ∆ relative to
τ (∆), the entries b(7), b(8), b(10), b(12) are also always zeroes and could be disregarded. Furthermore, there is no need to consider extensions of type 1, 2 and turn
operations. Hence, in this case, we only need elementary operations which are additions, extensions of type 3 and mergers over brackets of type B1, pseudobrackets
of type PB1, and over their systems.
9. Construction of Minimal Diagrams over (1.4) and Proof of
Theorem 1.7
Theorem 1.7. Suppose that W is a word over the alphabet A±1 such that the
bounded word problem for presentation (1.4) holds for the pair (W, n). Then a
minimal diagram ∆ over (1.4) such that ϕ(∂∆) ≡ W can be algorithmically constructed in deterministic space O(max(log |W |, log n)(log |W |)2 ) or in deterministic
time O(|W |4 ).
In addition, if |n1 | = |n2 | in (1.4), then the minimal diagram problem for presentation (1.4) can be solved in deterministic space O((log |W |)3 ) or in deterministic
time O(|W |3 log |W |).
Proof. First we prove the space part of Theorem 1.7.
G
Let W be a nonempty word over A±1 such that W =3 1, where G3 is defined by
presentation (1.4), and there is a disk diagram ∆ over (1.4) such that ϕ(∂∆) ≡ W
and |∆(2)| ≤ n, i.e., the bounded word problem has a positive solution for the pair
(W, n).
BOUNDED AND PRECISE WORD PROBLEMS FOR GROUP PRESENTATIONS
73
It follows from Lemma 6.13 that there is a finite sequence Ω of elementary
operations such that Ω converts the empty pseudobracket system for W into a
final one and Ω has other properties stated in Lemma 6.13. As in the proof of
Theorem 1.4, Lemma 6.13 gives us a nondeterministic algorithm which runs in
time O(|W |) and space O((max(log |W |, log n) log |W |) and which accepts a word
W over A±1 if and only if the bounded word problem has a positive solution for
the pair (W, n). Note that here the big-O constants can be written down explicitly
(see the proof of Theorem 1.4).
Furthermore, as in the proof of Theorem 1.4, using Savitch’s theorem [38], see
also [3], [32], we obtain a deterministic algorithm which runs in space
O((max(log |W |, log n)(log |W |)2 ),
(9.1)
and which computes a minimal integer n(W ), 0 ≤ n(W ) ≤ n, such that there is a
disk diagram ∆ over (1.4) so that ϕ(∂∆) ≡ W and |∆(2)| = n(W ). To do this, we
can check by Savitch’s algorithm whether the empty pseudobracket system B0 can
be transformed by elementary operations into a final pseudobracket system {bF },
where bF (4) = n′ for n′ = 0, 1, 2, . . . , n.
Without loss of generality, we may assume that n(W ) > 0 because if n(W ) = 0
then Corollary 1.6 yields the desired result.
Having found this number n(W ) ≥ 1 in deterministic space (9.1), we will run
Savitch’s algorithm again for the pair B0 and {b∗F }, where b∗F = (0, |W |, 0, n(W )),
and use the counter trick, as in the proof of Theorem 1.5, to compute an instance
of a sequence Ω of elementary operations and the corresponding to Ω sequence
B0 , B1 , . . . , Bℓ = {b∗F } of pseudobracket systems. After computing these sequences
Ω and B0 , B1 , . . . , Bℓ = {b∗F }, our algorithm halts. Denote this modification of
Savitch’s algorithm by An .
Denote
Ω = (ω1 , . . . , ωℓ ),
where ω1 , . . . , ωℓ are elementary operations and, as above, let B0 , B1 , . . . , Bℓ be the
corresponding to Ω sequence of pseudobracket systems so that Bj is obtained by
application of ωj to Bj−1 , so Bj = ωj (Bj−1 ). We also let
(χ1 , . . . , χℓ2 )
denote the subsequence of Ω that consists of all extensions of type 2. Also, let
ci ∈ Bji −1 denote the pseudobracket to which the elementary operation χi applies
and let di ∈ Bji denote the pseudobracket obtained from ci by application of χi , so
di = χi (ci ).
According to the proof of Lemma 6.12, the sequence Ω or, equivalently, the
sequence B0 , B1 , . . . , Bℓ , defines a disk diagram ∆(Ω) which can be inductively
constructed as in the proof of Claim (D1). Furthermore, according to the proof
of Claim (D1), all faces of ∆(Ω) are contained in ℓ2 a2 -bands Γ1 , . . . , Γℓ2 which
are in bijective correspondence with elementary operations χ1 , . . . , χℓ2 so that Γi
corresponds to χi , i = 1, . . . , ℓ2 . Denote
∂Γi = fi ti gi ui ,
(9.2)
74
S. V. IVANOV
where fi , gi are edges of ∂∆(Ω), ϕ(fi ) = ϕ(gi )−1 ∈ {a±1
2 }, and ti , ui are simple
paths whose labels are powers of a1 . Hence,
ℓ2
X
|Γi (2)| = |∆(Ω)(2)| = n(W ).
i=1
It follows from the definitions that if a(ci ), a(di ) are the arcs of the pseudobrackets ci , di , resp., and e1,i a(ci )e2,i is a subpath of the path PW , where e1,i , e2,i are
edges of PW , then, renaming fi ⇆ gi , ti ⇆ ui if necessary, we have the following
equalities
α(e1,i ) = fi , α(e2,i ) = gi , ϕ(fi ) = ϕ(gi )−1 = aε2i , εi = ±1,
c (3)
ϕ(ti ) ≡ a1i
,
−di (3)
ϕ(ui ) ≡ a1
,
α(a(ci )) = ti ,
α(a(di )) = u−1
i
for every i = 1, . . . , ℓ2 , see Fig. 9.1.
ti = α(a(ci ))
fi = α(e1,i )
Γi
gi = α(e2,i )
ui = α(a(di ))−1
Fig. 9.1
Note that each of these a2 -bands Γ1 , . . . , Γℓ2 can be constructed, as a part of the
sequence Ω, in deterministic space (9.1) by running the algorithm An . For instance,
if we wish to retrieve information about the diagram Γi , we would be looking for
the ith extension of type 2 in the sequence Ω, denoted above by χi . We also remark
that the parameters (ci , εi ), associated with the elementary operation χi , contain
all the information about the diagram Γi .
Observe that we are not able to keep all these pairs (ci , εi ), i = 1, . . . , ℓ2 , in
our intermediate computations aimed to construct ∆(Ω) in polylogarithmic space
because doing this would take polynomial space. However, we can reuse space, so
we keep one pair (ci , εi ), or a few pairs, in memory at any given time and, when
we need a different pair (cj , εj ), we erase (ci , εi ) and compute the new pair (cj , εj )
by running the algorithm An as discussed above.
We can also output all these pairs (ci , εi ), i = 1, . . . , ℓ2 , as a part of our description of the disk diagram ∆(Ω) still under construction.
Thus, in deterministic space (9.1), we have obtained the information about a2 bands Γ1 , . . . , Γℓ2 of ∆(Ω) that contain all of the faces of ∆(Ω) and it remains
to describe, working in space (9.1), how the edges of (∂∆(Ω))±1 and those of
(∂Γ1 )±1 , . . . , (∂Γℓ2 )±1 are attached to each other.
Observe that by taking the subdiagrams Γ1 , . . . , Γℓ2 out of ∆(Ω), we will produce
ℓ2 + 1 connected components ∆1 , . . . , ∆ℓ2 +1 which are disk diagrams with no faces,
i.e., ∆1 , . . . , ∆ℓ2 +1 are disk diagrams over the free group F (A) = hA k ∅i. Note
that the boundary of each disk diagram ∆i has a natural factorization
∂∆i = q1,i r1,i . . . qki ,i rki ,i ,
(9.3)
where every qj,i is a subpath of the cyclic path ∂∆(Ω), perhaps, |qj,i | = 0, and
−1
−1
−1
every rj,i is one of the paths t−1
1 , u1 , . . . , tℓ2 , uℓ2 , |rj,i | > 0, see Fig. 9.2.
BOUNDED AND PRECISE WORD PROBLEMS FOR GROUP PRESENTATIONS
75
q1,1
q1,4
Γ2
q1,3
∆3 r1,3
Γ1
∆
∆1
q2,4
r1,1
Γ3
r1,4
r3,4 r2,4 r1,2 ∆2
q1,2
∆4
q3,4
Fig. 9.2
It follows from the definitions and Lemma 6.13 that
ℓ2
X
|∂∆i | < |∂∆(Ω)| + |∆(Ω)(2)|(|n1 | + |n2 |) = O(max(|W |, n(W ))),
(9.4)
i=1
as |∆(Ω)(2)| = n(W ) and |∂∆(Ω)| = |W |.
Suppose that a vertex v ∈ PW is given, 0 ≤ v ≤ |PW | = |W |. Then there
is a unique disk diagram ∆iv among ∆1 , . . . , ∆ℓ2 +1 such that α(v) ∈ ∂∆iv . We
now describe an algorithm Bv that, for an arbitrary edge e of the boundary ∂∆iv ,
computes in deterministic space (9.1) the label ϕ(e) of e and the unique location
−1
−1
−1
of e in one of the paths ∂|0 ∆(Ω), t−1
1 , u1 , . . . , tℓ2 , uℓ2 . To do this, we will go
around the boundary path ∂|α(v) ∆iv , starting at the vertex α(v). Initializing the
parameters v ∗ , d∗ , we set
v ∗ := v,
d∗ := 1.
If e is the kth edge of ∂|α(v) ∆iv , 1 ≤ k ≤ |∂∆iv |, then we also say that e is the
edge of ∂|α(v) ∆iv with number k.
Let ec denote the edge of PW such that (ec )− = v ∗ if v ∗ < |W | and (ec )− = 0 if
∗
v = |W |. We now consider three possible Cases 1–3.
Case 1. Assume that α(ec ) = fi for some i = 1, . . . , ℓ2 , see the definition (9.2)
of ∂Γi . Note that such an equality α(ec ) = fi can be verified in space (9.1) by
checking, one by one, all a2 -bands Γ1 , . . . , Γℓ2 .
If v ∗ = v, then the first |ti | edges of the boundary path
∂|α(v) ∆iv = t−1
i ...
are the edges of the path ti−1 , see Fig. 9.3(a).
In the general case when v ∗ is arbitrary, we obtain that the edges of the boundary
path ∂|α(v) ∆iv with numbers d∗ , . . . , d∗ + |ti | − 1 are consecutive edges of the path
ti−1 starting from the first one.
Let v ′ ∈ PW denote the vertex such that α(v ′ ) = (ti )− , see Fig. 9.3(a). Also,
denote d′ := d∗ + |ti |.
Case 2. Assume that α(ec ) = gi for some i = 1, . . . , ℓ2 , see (9.2). As in Case 1,
we can verify this equality in space (9.1).
76
S. V. IVANOV
If v ∗ = v, then the first |ui | edges of the boundary path
∂|α(v) ∆iv = u−1
i ...
are the edges of the path u−1
i , see Fig. 9.3(b).
In the general case when v ∗ is arbitrary, we obtain that the edges of the boundary
path ∂|α(v) ∆iv with numbers d∗ , . . . , d∗ + |ui | − 1 are consecutive edges of the path
ui−1 starting from the first one.
Let v ′ ∈ PW denote the vertex such that α(v ′ ) = (ui )− , see Fig. 9.3(b). Also,
denote d′ := d∗ + |ui |.
∂∆
∂∆
∆iv
ui
fi
α(v ∗ )
Γi
ti
α(v ′ )
fi
gi
α(v ′ )
ui
Γi
α(v ∗ )
gi
ti
∆iv
Fig. 9.3(a)
Fig. 9.3(b)
Case 3. Suppose that α(ec ) is not one of the edges fi , gi of a2 -bands Γ1 , . . . , Γℓ2 .
As above, we can verify this claim in space (9.1).
If v ∗ = v, then the first edge of the boundary path
∂|α(v) ∆iv = α(ec ) . . .
is α(ec ).
In the general case when v ∗ is arbitrary, we have that the edge of the boundary
path ∂|α(v) ∆iv with number d∗ is α(ec ), see Fig. 9.4.
Denote v ′ := (ec )− and let d′ := d∗ + 1, see Fig. 9.4.
∂∆
Γj
α(v ′ )
ec
α(v ∗ )
∆iv
Fig. 9.4
The foregoing mutually exclusive Cases 1–3 describe a cycle, including the first
one when v ∗ = v, of the algorithm Bv which is finished by reassignment v ∗ := v ′
and d∗ := d′ .
We now repeat the above cycle with new values of parameters v ∗ , d∗ , whose
storage along with storage of the original vertex v requires additional
O(max(log n, log |W |))
BOUNDED AND PRECISE WORD PROBLEMS FOR GROUP PRESENTATIONS
77
space, as follows from the definitions and inequality (9.4). We keep on repeating
this cycle until we obtain at certain point that the vertex v ∗ is again v which means
that all the edges of the path ∂|α(v) ∆iv have been considered, we are back at v and
we stop. Note that d∗ = |∂|α(v) ∆iv | when we stop.
We remark that, when running the algorithm Bv we can stop at once and abort
computations whenever we encounter the value of parameter v ∗ less than v, because,
in this case, the disk diagram ∆iv which contains the vertex v is identical to ∆iv∗ and
the information about the edges of ∆iv∗ can be obtained by running the algorithm
Bv∗ instead. Thus, we now have that either the algorithm Bv aborts at some point
or Bv performs several cycles and stops when v = v ∗ and d∗ > 1, in which case we
say that the algorithm Bv accepts the diagram ∆iv .
By running the algorithms Bv consecutively for v = 0, 1, . . . , with possible stops
as described above, we will get the information about the edges of all disk diagrams
∆1 , . . . , ∆ℓ2 +1 so that for every j = 1, . . . , ℓ2 + 1, there will be a unique vertex
v(j) ∈ PW such that ∆j = ∆iv(j) and the algorithm Bv(j) accepts ∆j . Recall that,
in the proof of Theorem 1.2, see also Corollary 1.6, we constructed a deterministic
algorithm C such that, when given a word U so that U = 1 in the free group
F (A), the algorithm C outputs a diagram ∆U over F (A) = hA k ∅i such that
ϕ(∂|0 ∆U ) ≡ U and the algorithm C operates in space O((log |U |)3 ). Next, we
observe that our ability to deterministically compute, in space (9.1), the label
ϕ(e) of every edge e of the boundary path ∂|α(v) ∆iv , as well as the location of
−1
−1
−1
e in one of the paths ∂|0 ∆(Ω), t−1
1 , u1 , . . . , tℓ2 , uℓ2 , combined together with the
e iv over F (A) such that
algorithm C, means that we can construct a disk diagram ∆
e iv ) ≡ ϕ(∆iv ) in deterministic space
ϕ(∆
O(|∂∆iv |3 ) = O((max(log |W |, log n(W ))3 )),
see (9.4). Replacing the disk diagrams ∆1 , . . . , ∆ℓ2 +1 in ∆(Ω) with their substitutes
e 1, . . . , ∆
e ℓ2 +1 , we will obtain a disk diagram ∆(Ω)
e
∆
over (1.4) such that
e
ϕ(∂ ∆(Ω))
≡ ϕ(∂∆(Ω)),
e
|∆(Ω)(2)|
= |∆(Ω)(2)| = n(W ),
e
i.e., ∆(Ω)
is as good as ∆(Ω).
e 1, . . . , ∆
e ℓ2 +1 along with a2 -bands Γ1 , . . . , Γℓ2 constitute
Since the disk diagrams ∆
e
e
the entire diagram ∆(Ω), our construction of ∆(Ω),
performed in space (9.1), is now
complete.
It remains to prove the additional claim of Theorem 1.7. We start with the
following.
Lemma 9.1. Suppose ∆0 is a reduced disk diagram over (1.4), where |n1 | = |n2 |.
Then the number |∆0 (2)| of faces in ∆0 satisfies |∆0 (2)| ≤ 4|n1 1 | |∂∆0 |2 .
Proof. Denote n0 := |n1 | and set n2 = κn1 , where κ = ±1. Consider the presentation
−1
−1
κ
κ
κ
G4 = h A, b1 , . . . , bn0 −1 k a2 a1 b−1
1 = a1 , b1 a1 b2 = a1 , . . . , bn0 −1 a1 a2 = a1 i (9.5)
κn1
that has n0 relations which are obtained by splitting the relation a2 an1 1 a−1
2 = a1
of (1.4) into n0 “square” relations, see Fig. 9.5 where the case n0 = 3 is depicted.
78
S. V. IVANOV
a1
a2
a1
b1
a1
a2
b2
aκ1
aκ1
aκ1
Fig. 9.5
Note that there is a natural isomorphism ψ : G4 → G3 defined by the map
ψ(a) = a for a ∈ A, and ψ(bj ) = a−κj
a2 aj1 for j = 1, . . . , n0 − 1.
1
If ∆0 is a reduced disk diagram over (1.4), then we can split faces of ∆0 into
“square” faces over (9.5), as in Fig. 9.5, and obtain a reduced disk diagram ∆0,b
over (9.5) such that
ϕ(∂∆0,b ) ≡ ϕ(∂∆0 ),
|∆0,b (2)| = n0 |∆0 (2)|.
(9.6)
Let ∆ be an arbitrary reduced disk diagram over presentation (9.5).
We modify the definition of an ai -band given in Sect. 6 for diagrams over presentation (1.4) so that this definition would be suitable for diagrams over (9.5).
Two faces Π1 , Π2 in a disk diagram ∆ over (9.5) are called j-related, where
j = 1, 2, denoted Π1 ↔j Π2 , if there is an edge e such that e ∈ ∂Π1 , e−1 ∈ ∂Π2 ,
±1
±1
and ϕ(e) = a±1
if j = 1 or ϕ(e) ∈ {a±1
1
2 , b1 , . . . , bn0 −1 } if j = 2. As before,
we consider a minimal equivalence relation, denoted ∼j , on the set of faces of ∆
generated by the relation ↔j .
An ai -band, where i ≥ 1 is now arbitrary, is a minimal subcomplex Γ of ∆ that
±1
±1
contains an edge e such that ϕ(e) = a±1
if i 6= 2 or ϕ(e) ∈ {a±1
2 , b1 , . . . , bn0 −1 }
i
if i = 2 and Γ has the following property. If there is a face Π in ∆ such that
e ∈ (∂Π)±1 , then Γ must contain all faces of the equivalence class [Π]∼i of Π.
As before, this definition implies that an ai -band Γ is either a subcomplex consisting of a single nonoriented edge {f, f −1 }, where ϕ(f ) = a±1
if i 6= 2 or
i
±1
±1
−1
ϕ(f ) ∈ {a±1
∈ ∂∆, or Γ consists of all faces of
2 , b1 , . . . , bn0 −1 } if i = 2 and f, f
an equivalence class [Π]∼i , here i = 1, 2.
If an ai -band Γ contains faces, then Γ is called essential.
If an ai -band Γ is essential but Γ contains no face whose boundary contains an
±1 ±1
±1
edge f such that f −1 ∈ ∂∆ and ϕ(f ) = a±1
1 if i = 1 or ϕ(f ) ∈ {a2 , b1 , . . . , bn0 −1 }
if i = 2, then Γ is called a closed ai -band. It follows from the definitions that if Γ
is an essential ai -band, then i = 1, 2.
If Γ is an essential ai -band and Π1 , . . . , Πk are all faces of Γ, we consider a simple
arc or a simple curve c(Γ) in the interior of Γ such that the intersection c(Γ)∩Πj for
every j = 1, . . . , k, is a properly embedded arc in Πj whose boundary points belong
to the interior of different edges of ∂Πj whose ϕ-labels are in {a1±1 } if i = 1 or in
±1
±1
{a±1
2 , b1 , . . . , bn0 −1 } if i = 2. This arc or curve c(Γ) will be called a connecting
line of Γ.
Note that if Γ contains a face Π whose boundary has an edge f such that f −1 ∈
±1
±1
if i = 1 or ϕ(f ) ∈ {a±1
∂∆ and ϕ(f ) = a±1
2 , b1 , . . . , bn0 −1 } if i = 2, then a
1
connecting line c(Γ) of Γ connects two points ∂c(Γ) on the boundary ∂∆. On the
other hand, if Γ contains no face Π as above, then c(Γ) is a closed simple curve, in
which case Γ is called a closed ai -band.
BOUNDED AND PRECISE WORD PROBLEMS FOR GROUP PRESENTATIONS
79
If Π is a face in ∆ over (9.5), then there are exactly two bands, a1 -band and
a2 -band, denoted ΓΠ,1 , ΓΠ,2 , resp., whose connecting lines c(ΓΠ,1 ), c(ΓΠ,2 ) pass
through Π. Without loss of generality, we may assume that the intersection
c(ΓΠ,1 ) ∩ c(ΓΠ,2 ) ∩ Π
consists of a single (piercing) point. The following lemma has similarities with
Lemma 6.1.
Lemma 9.2. Suppose that ∆ is a reduced disk diagram over presentation (9.5).
Then there are no closed ai -bands in ∆ and every ai -band Γ of ∆ is a disk subdiagram of ∆ such that
∂|(f1 )− Γ = f1 s1 f2 s2 ,
where f1 , f2 are edges of ∂∆ such that
ϕ(f1 ), ϕ(f2 ) ∈ {a±1
i } if i 6= 2,
±1
±1
ϕ(f1 ), ϕ(f2 ) ∈ {a±1
2 , b1 , . . . , bn0 −1 } if i = 2,
s1 , s2 are simple paths such that |s1 | = |s2 | = |Γ(2)|, and ∂Γ is a reduced simple
closed path when |Γ(2)| > 0. In addition, if Γ1 and Γ2 are essential a1 -band and
a2 -band, resp., in ∆, then their connecting lines c(Γ1 ), c(Γ2 ) intersect in at most
one point.
Proof. Since ∆ is reduced, it follows that if Π1 ↔j Π2 , j = 1, 2, then Π1 is not a
mirror copy of Π2 , i.e., ϕ(Π1 ) 6≡ ϕ(Π2 )−1 . This remark, together with the definition
of an ai -band and defining relations of presentation (9.5), implies that if Π1 ↔j Π2
then the faces Π1 , Π2 share exactly one nonoriented edge. This, in particular,
implies |s1 | = |s2 | = |Γ(2)| if Γ is an ai -band such that ∂Γ = f1 s1 f2 s2 as in the
statement of Lemma 9.2.
Assume, on the contrary, that there is an essential ai -band Γ0 in ∆ such that
either Γ0 is closed or ∂Γ0 = f1 s1 f2 s2 as above, and one of the paths s1 , s2 is not
simple. Then there is a disk subdiagram ∆2 of ∆ bounded by edges of ∂Γ0 such
that ϕ(∂∆2 ) is a nonempty reduced word over the alphabet {a±1
1 } if i = 2 or over
±1
±1
the alphabet {a±1
2 , b1 , . . . , bn0 −1 } if i = 1. Pick such diagrams Γ0 and ∆2 so
that |∆2 (2)| is minimal. Since ∆2 contains faces and ϕ(∂∆2 ) contains no letters
±1 ±1
±1
either from {a±1
1 } if i = 1 or from {a2 , b1 , . . . , bn0 −1 } if i = 2, it follows that
every a3−i -band in ∆2 is closed and bounds a similar to ∆2 diagram ∆′2 such that
|∆′2 (2)| < |∆2 (2)|. This contradiction proves that an ai -band Γ0 with the above
properties does not exist.
To prove the space part of the additional claim, suppose, on the contrary, that
there are essential a1 - and a2 -bands Γ1 and Γ2 , resp., in ∆ such that the intersection
c(Γ1 ) ∩ c(Γ2 ) of their connecting lines c(Γ1 ), c(Γ2 ) contains at least two points. We
pick two such consecutive along c(Γ1 ), c(Γ2 ) points and let Π1 , Π2 be the faces that
contain these two points. Then there exists a disk subdiagram ∆1 in ∆ such that
is a subpath of ∂Γ1 , |u1 | ≥ 0 and ϕ(u1 ) is a reduced or
∂∆1 = u1 u2 , where u−1
1
±1
±1
−1
empty word over {a±1
is a subpath of ∂Γ2 , |u2 | ≥ 0
2 , b1 , . . . , bn0 −1 }, while u2
},
see Fig. 9.6. The equality
and ϕ(u2 ) is a reduced or empty word over {a±1
1
|u1 |+|u2 | = 0 implies that the faces Π1 , Π2 form a reducible pair. This contradiction
to ∆ being reduced proves that |u1 | + |u2 | > 0, whence |∆1 (2)| > 0.
80
S. V. IVANOV
Γ1
Π1
u1
Π2
∆1
u2
Γ2
Fig. 9.6
We now pick a disk subdiagram ∆′1 in ∆ such that |∆′1 (2)| is minimal,
∂∆′1 = u′1 u′2 ,
|u′1 | + |u′2 | > 0,
±1
±1
ϕ(u′1 ) is a reduced or empty word over the alphabet {a±1
2 , b1 , . . . , bn0 −1 }, and
′
ϕ(u2 ) is a reduced or empty word over {a±1
1 }. Consider an a1 -band Γ3 in ∆1 if
′
′
′
′
|u2 | > 0 or consider an a2 -band Γ4 in ∆1 if |u2 | = 0 and |u1 | > 0 such that a
connecting line of Γj , j = 3, 4, connects points on u′2 if j = 3 or on u′1 if j = 4, see
Fig. 9.7. It is easy to check that, taking Γ3 , or Γ4 , out of ∆′1 , we obtain two disk
subdiagrams ∆′1,1 , ∆′1,2 one of which has the above properties of ∆′1 and a fewer
number of faces, see Fig. 9.7. This contradiction to the minimality of ∆′1 completes
the proof of Lemma 9.2.
∆′1
u′1
∆′1,1
Γ3
∆′1,2
u′2
Fig. 9.7
Coming back to the diagrams ∆0 and ∆0,b , see (9.6), we can see from Lemma 9.2
that the number of a1 -bands in ∆0,b is at most 12 |∂∆0,b |a1 ≤ 21 |∂∆0 |. Similarly,
the number of a2 -bands in ∆0,b is at most 12 |∂∆0,b | = 21 |∂∆0 |. It follows from the
definitions and Lemma 9.2 that the number |∆0,b (2)| is equal to the number of
intersections of connecting lines of a1 -bands and a2 -bands in ∆0,b . Hence, referring
to Lemma 9.2 again, we obtain that |∆0,b (2)| ≤ ( 21 |∂∆0 |)2 . In view of (9.6), we
finally have
|∆0 (2)| = n10 |∆0,b (2)| ≤ 4n1 0 |∂∆0 |2 ,
as desired. The proof of Lemma 9.1 is complete.
Suppose ∆0 is a reduced diagram over presentation (1.4), where |n1 | = |n2 |. It
follows from Lemma 9.2 that |∆0 (2)| ≤ 4n1 0 |∂∆0 |2 . This inequality means that the
space bound (9.1) becomes O((log |W |)3 ) as n(W ) ≤ 4n1 0 |∂∆0 |2 = 4n1 0 |W |2 . The
space part of Theorem 1.7 is proved.
BOUNDED AND PRECISE WORD PROBLEMS FOR GROUP PRESENTATIONS
81
To prove the time part of Theorem 1.7, we review the proof of P part of Theorem 1.4. We observe that our arguments enable us, concurrently with computation
of numbers λ(W (i, j)), µ3 (W (i, j)) for every parameterized subword W (i, j) of W
such that λ(W (i, j)) = ℓ < ∞, to inductively construct a minimal diagram ∆(i, j)
over (1.4) such that ϕ(∂|0 ∆(i, j)) ≡ W (i, j)a−ℓ
1 . Details of this construction are
straightforward in each of the subcases considered in the proof of P part of Theorem 1.4. Note that this modified algorithm can still be run in time O(|W |4 ) for the
general case.
If, in addition, |n1 | = |n2 | then the inequality (7.5) can be improved to
|λ(W (i, j))| ≤ |W (i, j)|,
µ3 (W (i, j)) ≤ |W (i, j)|2 .
Hence, using, as before, binary representation for numbers λ(W (i′ , j ′ )), µ3 (W (i′ , j ′ )),
we can run inductive computation of numbers λ(U ), µ3 (U ) for U = W (i, j) and
construction of a diagram ∆(i, j) whenever it exists for given i, j, in time
O(|W (i, j)| log |W (i, j)|).
This improves the bound for running time of our modified algorithm, that computes
the numbers λ(W ), µ3 (W ) and constructs a minimal diagram ∆(1, |W |) for W , from
O(|W |4 ) to O(|W |3 log |W |), as desired.
Theorem 1.7 is proved.
10. Polygonal Curves in the Plane and Proofs of Theorems 1.8, 1.9
and Corollary 1.10
Let T denote a tessellation of the plane R2 into unit squares whose vertices are
points with integer coordinates. Let c be a finite closed path in T so that edges
of c are edges of T. Recall that we have the following two types of elementary
operations over c. If e is an oriented edge of c, e−1 is the edge with an opposite
to e orientation, and ee−1 is a subpath of c so that c = c1 ee−1 c2 , where c1 , c2 are
subpaths of c, then the operation c → c1 c2 over c is called an elementary homotopy
of type 1. Suppose that c = c1 uc2 , where c1 , u, c2 are subpaths of c, and a boundary
path ∂s of a unit square s of T is ∂s = uv, where u, v are subpaths of ∂s and either
of u, v could be of zero length, i.e., either of u, v could be a single vertex of ∂s.
Then the operation c → c1 v −1 c2 over c is called an elementary homotopy of type 2.
Theorem 1.8. Let c be a finite closed path in a tessellation T of the plane R2
into unit squares so that edges of c are edges of T. Then a minimal number m2 (c)
such that there is a finite sequence of elementary homotopies of type 1–2, which
turns c into a single point and which contains m2 (c) elementary homotopies of type
2, can be computed in deterministic space O((log |c|)3 ) or in deterministic time
O(|c|3 log |c|), where |c| is the length of c.
Furthermore, such a sequence of elementary homotopies of type 1–2, which turns
c into a single point and which contains m2 (c) elementary homotopies of type 2,
can also be computed in deterministic space O((log |c|)3 ) or in deterministic time
O(|c|3 log |c|).
Proof. First we assign ϕ-labels to edges of the tessellation T. If an edge e goes from
a point of R2 with coordinates (i, j) to a point with coordinates (i + 1, j), then
2
we set ϕ(e) := a1 and ϕ(e−1 ) := a−1
1 . If an edge f goes from a point of R with
coordinates (i, j) to a point with coordinates (i, j + 1), then we set ϕ(f ) := a2 and
ϕ(f −1 ) := a−1
2 .
82
S. V. IVANOV
Let c be a finite closed path in T whose edges are those of T. Without loss of
generality, we may assume that c starts at the origin, hence c− = c+ has coordinates
(0, 0) (otherwise, we could apply a logspace reduction to achieve this property).
Denote c = fe1 . . . fe|c| , where fe1 , . . . , fe|c| are edges of c, and let
ϕ(c) := ϕ(fe1 ) . . . ϕ(fe|c| ),
±1
where ϕ(c) is a word over the alphabet A±1 = {a±1
1 , a2 }. Since c is closed, it
G
follows that ϕ(c) =5 1, where
−1
G5 := h a1 , a2 k a2 a1 a−1
2 a1 = 1 i.
(10.1)
Lemma 10.1. Suppose that a closed path c in T can be turned into a point by a
finite sequence Ξ of elementary homotopies and m2 (Ξ) is the number of elementary
homotopies of type 2 in this sequence. Then there is a disk diagram ∆ such that
ϕ(∂∆) ≡ ϕ(c) and |∆(2)| = m2 (Ξ).
Conversely, suppose there is a disk diagram ∆ with ϕ(∂∆) ≡ ϕ(c). Then there
is a finite sequence Ξ of elementary homotopies such that Ξ turns c into a point
and m2 (Ξ) = |∆(2)|.
Proof. Given a finite sequence Ξ of elementary homotopies ξ1 , ξ2 , . . . , we can construct a disk diagram ∆ over (10.1) such that ϕ(∂∆) ≡ ϕ(c) and m2 (Ξ) = |∆(2)|.
Indeed, starting with a simple closed path qc in the plane R2 (without any tessellation) such that ϕ(qc ) ≡ ϕ(c), we can simulate elementary homotopies in the
sequence Ξ so that an elementary homotopy of type 1 is simulated by folding a
suitable pair of edges of the path qc , see Fig. 10.1(a), and an elementary homotopy
of type 2 is simulated by attachment of a face Π over (10.1) to the bounded region
of R2 whose boundary is qc , see Fig. 10.1(b).
e1 = e2−1
e1 e2
u1
∂Π = u2 u−1
1
u1
Π
u2
qc
Fig. 10.1(a)
qc′
qc
Fig. 10.1(b)
qc′
As a result, we will fill out the bounded region of R2 whose boundary path is qc
with faces and obtain a required disk diagram ∆ over (10.1).
The converse can be established by a straightforward induction on |∆(2)|.
It follows from Lemma 10.1 and Theorem 1.7 that a minimal number m2 (c) of
elementary homotopies of type 2 in a sequence of elementary homotopies that turns
the path c into a point can be computed in deterministic space O((log |ϕ(c)|)3 ) =
O((log |c|)3 ) or in deterministic time O(|c|3 log |c|).
It remains to show that a desired sequence of elementary homotopies for c can be
computed in deterministic space O((log |c|)3 ) or in deterministic time O(|c|3 log |c|).
We remark that Lemma 10.1 and its proof reveal a close connection between sequences of elementary homotopies that turn c into a point and disk diagrams ∆ over
(10.1) such that ϕ(∂|0 ∆) ≡ ϕ(c). According to Theorem 1.7, we can construct a disk
diagram ∆c such that ϕ(∂|0 ∆c ) ≡ ϕ(c) and |∆c (2)| = m2 (c) in space O((log |c|)3 )
BOUNDED AND PRECISE WORD PROBLEMS FOR GROUP PRESENTATIONS
83
or in time O(|c|3 log |c|). Since |∆c (2)| ≤ |∂∆c |2 /4 = |c|2 /4 by Lemma 9.1, it follows
that we can use this diagram ∆c together with Lemma 10.1 to construct a desired
sequence of elementary homotopies in deterministic time O(|c|2 ). Therefore, a desired sequence of elementary homotopies can be constructed in deterministic time
O(|c|3 log |c|).
It is also tempting to use this diagram ∆c together with Lemma 10.1 to construct,
in space O((log |c|)3 ), a desired sequence of elementary homotopies which turns c
into a point. However, this approach does not quite work for the space part of the
proof because inductive arguments of the proof of Lemma 10.1 use intermediate
paths whose storage would not be possible in O((log |c|)3 ) space. For this reason,
we have to turn back to calculus of pseudobrackets.
Denote W := ϕ(c). As in Sects. 2, 6, let PW be a path such that ϕ(PW ) ≡ W
and vertices of PW are integers 0, 1, . . . , |W | = |c|. Denote PW = f1 . . . f|c| , where
f1 , . . . , f|c| are edges of PW , and let
γ : PW → c
be a cellular map such that γ(fi ) = fei for i = 1, . . . , |W |. Recall that c = fe1 . . . fe|c| .
Note that γ(i) = (fei )+ , where i = 1, . . . , |W |, and γ(0) = (fe1 )− = (fe|c| )+ .
Suppose B is a pseudobracket system for W and let B = {b1 , . . . , bk }, where
bi (2) ≤ bj (1) if i < j. Recall that the arc a(b) of a pseudobracket b ∈ B is a
subpath of PW such that a(b)− = b(1) and a(b)+ = b(2).
We now define a path Γ(B) for B in T so that Γ(B) := γ(PW ) = c if k = 0, i.e.,
B = ∅, and for k > 0 we set
Γ(B) := c0 δ(a(b1 ))c1 . . . δ(a(bk ))ck ,
(10.2)
where ci = γ(di ) and di is a subpath of PW defined by (di )− = bi (2) and (di )+ =
bi+1 (1), except for the case when i = 0, in which case (d0 )− = 0, and except for
the case when i = k, in which case (dk )− = |W |. For every i = 1, . . . , k, the path
δ(a(bi )) in (10.2) is defined by the equalities
δ(a(bi ))− = γ(bi (1)),
δ(a(bi ))+ = γ(bi (2))
and by the equality ϕ(δ(a(bi ))) ≡ a1k1 ak22 with some integers k1 , k2 .
Suppose that B0 , B1 , . . . , Bℓ is an operational sequence of pseudobracket systems
for W that corresponds to a sequence Ω of elementary operations that turns the
empty pseudobracket system B0 into a final pseudobracket system Bℓ = {bℓ,1 }. We
also assume that bℓ,1 (4) = |∆c (2)| and ∆c is a minimal disk diagram over (10.1)
such that ϕ(∂∆c ) ≡ W ≡ ϕ(c). It follows from Lemmas 9.1, 10.1 that
|∆c (2)| = m2 (c) ≤ |∂∆c |2 /4 = |W |2 /4.
(10.3)
Hence, in view of Lemma 6.13, see also inequalities (7.1)–(7.2), we may assume
that, for every pseudobracket b ∈ ∪ℓi=0 Bi , it is true that
0 ≤ b(1), b(2) ≤ |W |,
(10.4)
0 ≤ |b(3)| ≤ (|W |a1 + (|n1 | + |n2 |)|∆c (2)|)/2
≤ (|W |a1 + |W |2 /2)/2 ≤ |W |2 /2,
2
0 ≤ b(4) ≤ |∆c (2)| ≤ |W | /4.
Thus the space needed to store a pseudobracket b ∈ ∪ℓi=0 Bi is O(log |W |).
(10.5)
(10.6)
84
S. V. IVANOV
For every pseudobracket system Bi , denote Bi = {bi,1 , . . . , bi,ki }, where, as
above, bi,j (2) ≤ bi,j ′ (1) whenever j < j ′ .
For every pseudobracket system Bi , we consider a path c(i) := Γ(Bi ) in T defined
by the formula (10.2), hence, we have
c(i) := Γ(Bi ) = c0 (i)δ(a(bi,1 ))c1 (i) . . . δ(a(bi,ki ))cki (i).
(10.7)
As before, for B0 = ∅, we set c(0) := Γ(B0 ) = c.
Note that the last path c(ℓ) in the sequence c(0), . . . , c(ℓ), corresponding to a
final pseudobracket system Bℓ = {bℓ,1 }, where bℓ,1 = (0, |W |, 0, m2 (c)), consists of
a single vertex and has the form (10.7) equal to c(ℓ) = c0 (ℓ)δ(a(bℓ,1 ))c1 (ℓ), where
c0 (ℓ) = c1 (ℓ) = c− = c+ and δ(a(bℓ,1 )) = c.
b
Lemma 10.2. Let bi,j ∈ Bi be a pseudobracket. Then ϕ(δ(a(bi,j ))) ≡ a1i,j
the points γ(bi,j (1)), γ(bi,j (2)) of T have the same y-coordinate.
(3)
, i.e.,
Proof. Since the path γ(a(bi,j ))δ(a(bi,j )−1 in T is closed, it follows that
G
ϕ(γ(a(bi,j ))) =5 ϕ(δ(a(bi,j )−1 ),
where G5 is given by presentation (10.1). On the other hand, ϕ(γ(a(bi,j )) ≡
ϕ(a(bi,j )). Hence, by Claim (D) of the proof of Lemma 6.12, we get that
G
G
b
ϕ(δ(a(bi,j ))) =5 ϕ(a(bi,j ) =5 a1i,j
(3)
.
Since ϕ(δ(a(bi,j ))) ≡ ak11 ak22 with some integers k1 , k2 , it follows that k1 = bi,j (3)
and k2 = 0, as required.
We now analyze how the path c(i), defined by (10.7), changes in comparison
with the path c(i − 1), i ≥ 1.
Suppose that a pseudobracket system Bi , i ≥ 1, is obtained from Bi−1 by an
addition. Then c(i) = c(i − 1), hence no change over c(i − 1) is done. Note that the
form (10.7) does change by an insertion of a subpath consisting of a single vertex
which is the δ-image of the arc of the added to Bi−1 starting pseudobracket.
Assume that a pseudobracket system Bi , i ≥ 1, is obtained from Bi−1 by an
extension of type 1 on the left and bi,j ∈ Bi is obtained from bi−1,j ∈ Bi−1 by this
elementary operation. Let a(bi,j ), a(bi−1,j ) be the arcs of bi,j , bi−1,j , resp., and let
a(bi,j ) = e1 a(bi−1,j ), where e1 is an edge of PW and ϕ(e1 ) = aε11 , ε1 = ±1.
If ε1 · bi−1,j (3) ≥ 0, then we can see that c(i) = c(i − 1) because
cj−1 (i − 1)δ(a(bi−1,j )) = cj−1 (i)δ(a(bi,j ))
and all other syllables of the paths c(i) and c(i−1), as defined in (10.7), are identical.
On the other hand, if ε1 · bi−1,j (3) < 0, then the subpath cj−1 (i)δ(a(bi,j )) of
c(i) differs from the subpath cj−1 (i − 1)δ(a(bi−1,j )) of c(i − 1) by cancelation of
a subpath ee−1 , where e is the last edge of cj−1 (i − 1), e−1 is the first edge of
−b
(3)/|bi−1,j (3)|
δ(a(bi−1,j )), and ϕ(e) = a1 i−1,j
. Since all other syllables of the paths
c(i) and c(i − 1), as defined in (10.7), are identical, the change of c(i − 1), resulting
in c(i), can be described as an elementary homotopy of type 1 which deletes a
subpath ee−1 , where e is an edge of c defined by the equalities
e+ = γ(bi−1,j (1)),
−bi−1,j (3)/|bi−1,j (3)|
ϕ(e) = a1
.
(10.8)
The case when a pseudobracket system Bi , i ≥ 1, is obtained from Bi−1 by
an extension of type 1 on the right is similar but we need to make some changes.
BOUNDED AND PRECISE WORD PROBLEMS FOR GROUP PRESENTATIONS
85
Keeping most of the foregoing notation unchanged, we let a(bi,j ) = a(bi−1,j )e2 ,
where e2 is an edge of PW and ϕ(e2 ) = aε12 , ε2 = ±1.
If ε2 · bi−1,j (3) ≥ 0, then as above we conclude that c(i) = c(i − 1) because
δ(a(bi−1,j ))cj (i − 1) = δ(a(bi,j ))cj (i)
and all other syllables of the paths c(i) and c(i−1), as defined in (10.7), are identical.
If ε1 · bi−1,j (3) < 0, then we can see from the definitions and from Lemma 10.2
that the path c(i) can be obtained from c(i − 1) by an elementary homotopy of type
1 which deletes a subpath e−1 e, where e is an edge of c defined by the equalities
−bi−1,j (3)/|bi−1,j (3)|
e− = γ(bi−1,j (2)),
ϕ(e) = a1
.
(10.9)
We conclude that, in the case when B(i) is obtained from B(i−1) by an extension
of type 1, it follows from inequalities (10.4)–(10.6) and from equations (10.8)–(10.9)
that the elementary homotopy of type 1 that produces c(i) from c(i − 1) can be
computed in space O(log |W |) when the pseudobrackets bi,j ∈ Bi and bi−1,j ∈ Bi−1
are known.
Suppose that a pseudobracket system Bi , i ≥ 1, is obtained from Bi−1 by an
extension of type 2 and bi,j ∈ Bi is obtained from bi−1,j ∈ Bi−1 by this elementary
operation. Denote a(bi,j ) = e1 a(bi−1,j )e2 , where e1 , e2 are edges of PW and ϕ(e1 ) =
ϕ(e2 )−1 = aε2 , ε = ±1. According to the definition of an extension of type 2,
bi−1,j (3) 6= 0 and, in view of the equalities n1 = n2 = 1, see (1.4) and (10.1), we
can derive from Lemma 10.2 that the path c(i − 1) turns into the path c(i) in the
following fashion. Denote cj−1 (i − 1) = e
cj−1 (i − 1)e
e1 and cj (i − 1) = e
e2 e
cj (i − 1),
where e
cj−1 (i − 1), e
cj (i − 1) are subpaths of cj−1 (i − 1), cj (i − 1), resp., and ee1 , ee2 are
cj−1 (i − 1), cj (i) = e
cj (i − 1),
edges such that γ(ei′ ) = eei′ , i′ = 1, 2. Then cj−1 (i) = e
and the path δ(a(bi,j )) has the properties that
δ(a(bi,j ))− = (e
e1 )− ,
ϕ(δ(a(bi−1,j ))) ≡
δ(a(bi,j ))+ = (e
e2 )+ ,
b
(3)
a1i−1,j
≡ ϕ(δ(a(bi,j ))),
see Fig. 10.2, and the other syllables of the paths c(i) and c(i − 1), as defined in
(10.7), are identical.
cj−1 (i − 1) = e
cj−1 (i − 1)e
e1
ee1 = g1
cj−1 (i − 1) = cj (i)
e
s1
g2
δ(a(bi−1,j ))
s2
g3
gk−1
δ(a(bi−1,j ))
cj (i − 1) = ee2 e
cj (i − 1)
sk
ee2 = gk−1
cj (i − 1) = cj (i)
e
Fig. 10.2
Denote k := |bi−1,j (3)| = |bi,j (3)| > 0 and let
δ(a(bi−1,j )) = di−1,1 . . . di−1,k ,
δ(a(bi,j )) = di,1 . . . di,k ,
where di′ ,j ′ are edges of the paths δ(a(bi−1,j )), δ(a(bi,j )). Also, we let g1 , . . . , gk
e1 ) = aε2 and (gi′ )− = (di,i′ )− for i′ =
be the edges of T such that ϕ(gi′ ) = ϕ(e
1, . . . , k, in particular, g1 = ee1 and gk = ee−1
2 , see Fig. 10.2. Then there are k
elementary homotopies of type 2 which turn c(i − 1) into c(i) and which use k
86
S. V. IVANOV
squares of the region bounded by the closed path ee1 δ(a(bi−1,j ))e
e2 δ(a(bi,j ))−1 , see
Fig. 10.2. For example, the first elementary homotopy of type 2 replaces the subpath
e1 di−1,1 = g1 di−1,1 by di,1 g2 . Note that ∂s1 = g1 di−1,1 (di,1 g2 )−1 is a (negatively
e
orientated) boundary path of a square s1 of T. The second elementary homotopy
of type 2 (when k ≥ 2) replaces the subpath g2 di−1,2 by di,2 g3 , where ∂s2 =
g2 di−1,2 (di,2 g3 )−1 is a (negatively orientated) boundary path of a square s2 of T,
and so on.
In view of inequalities (10.4)–(10.6), these k = |bi−1,j (3)| elementary homotopies
of type 2 that produce c(i) from c(i − 1) can be computed in space O(log |W |) when
the pseudobrackets bi,j ∈ Bi and bi−1,j ∈ Bi−1 are available.
Suppose that a pseudobracket system Bi , i ≥ 1, is obtained from Bi−1 by an
extension of type 3 and bi,j ∈ Bi is obtained from bi−1,j ∈ Bi−1 by this elementary
operation. By the definition of an extension of type 3, bi−1,j (3) = 0 and a(bi,j ) =
e1 a(bi−1,j )e2 , where e1 , e2 are edges of PW and ϕ(e1 ) = ϕ(e2 )−1 6= a±1
1 . Hence,
ϕ(e1 ) = ϕ(e2 )−1 = aε2 , ε = ±1. It follows from Lemma 10.2 that
δ(a(bi−1,j )) = cj−1 (i − 1)+ = cj (i − 1)− ,
δ(a(bi,j )) = cj−1 (i)+ = cj (i)− .
As above, denote cj−1 (i − 1) = e
cj−1 (i − 1)e
e1 and cj (i − 1) = ee2 e
cj (i − 1), where
e
cj−1 (i − 1), e
cj (i − 1) are subpaths of cj−1 (i − 1), cj (i − 1), resp., and ee1 , ee2 are
edges such that γ(ei′ ) = eei′ , i′ = 1, 2. Then it follows from the definitions that
cj−1 (i) = e
cj−1 (i − 1), cj (i) = e
cj (i − 1) and that all other syllables of the paths
c(i) and c(i − 1), as defined in (10.7), are identical. Hence, the change of the path
c(i − 1) into c(i) can be described as an elementary homotopy of type 1 that deletes
the subpath ee1 ee2 of c(i − 1). Since (e1 )+ = bi−1,j (1) and ee1 = γ(e1 ), it is easy to
see from inequalities (10.4)–(10.6) that we can compute this elementary homotopy
of type 1 in space O(log |W |) when the pseudobrackets bi,j ∈ Bi and bi−1,j ∈ Bi−1
are given.
Suppose that a pseudobracket system Bi , i ≥ 1, is obtained from Bi−1 by a
merger operation and bi,j ∈ Bi is obtained from pseudobrackets bi−1,j , bi−1,j+1 ∈
Bi−1 by this merger.
First assume that bi−1,j (3) · bi−1,j+1 (3) ≥ 0. It follows from Lemma 10.2 and
from the definitions that
cj (i − 1) = δ(a(bi−1,j ))+ ,
δ(a(bi−1,j ))cj (i − 1)δ(a(bi−1,j+1 )) = δ(a(bi,j ))
and all other syllables of the paths c(i) and c(i − 1), as defined in (10.7), are
identical. Therefore, we have the equality of paths c(i − 1) = c(i) in this case. Note
that factorizations (10.7) of c(i) and c(i − 1) are different.
Now assume that bi−1,j (3) · bi−1,j+1 (3) < 0. It follows from Lemma 10.2 and
from the definitions that cj (i − 1) = δ(a(bi−1,j ))+ and that the subpath δ(a(bi,j )) of
c(i) can be obtained from the subpath δ(a(bi−1,j ))cj (i − 1)δ(a(bi−1,j+1 )) of c(i − 1)
by cancelation of min(|bi−1,j (3)|, |bi−1,j+1 (3)|) pairs of edges in c(i − 1) so that
the last edge of δ(a(bi−1,j )) is canceled with the first edge of δ(a(bi−1,j+1 )) as a
subpath ee−1 in the definition of an elementary homotopy of type 1 and so on until
a shortest path among δ(a(bi−1,j )), δ(a(bi−1,j+1 )) completely cancels. Note that
all other syllables of the paths c(i) and c(i − 1), as defined in (10.7), are identical.
Thus the path c(i) can be obtained from c(i − 1) by min(|bi−1,j (3)|, |bi−1,j+1 (3)|)
elementary homotopies of type 1 which, in view of (10.4)–(10.6), can be computed
BOUNDED AND PRECISE WORD PROBLEMS FOR GROUP PRESENTATIONS
87
in space O(log |W |) when the pseudobrackets bi,j ∈ Bi and bi−1,j , bi−1,j+1 ∈ Bi−1
are known.
Recall that, in the proof of Theorem 1.7, we devised an algorithm An that, in
deterministic space (9.1), constructs a sequence of elementary operations Ω and a
corresponding sequence of pseudobracket systems B0 , . . . , Bℓ for W such that B0
is empty, Bℓ = {bℓ,1 } is final and bℓ,1 (4) = |∆(2)| ≤ n, where ∆ is a minimal
disk diagram over (10.1) with ϕ(∂∆) ≡ W . In view of inequalities (10.3), we may
assume that n = |W |2 /4, hence, the bound (9.1) becomes O((log |W |)3 ) and, as we
saw in (10.4)–(10.6), every pseudobracket b ∈ ∪ℓi=0 Bi requires space O(log |W |) to
store. As was discussed above, when given pseudobracket systems Bi−1 and Bi , we
can construct, in space O(log |W |), a sequence of elementary homotopies that turns
the path c(i − 1) into c(i). Since the sequence of pseudobracket systems B0 , . . . , Bℓ
for W ≡ ϕ(c) is constructible in space O((log |W |)3 ), it follows that a sequence Ξ
of elementary homotopies of type 1–2 that turns the path c = c(0) into a vertex
c(ℓ) = c− = c+ can also be constructed in deterministic space O((log |W |)3 ). This
completes the proof of Theorem 1.8.
Recall that a polygonal closed curve c in the plane R2 , equipped with a tessellation
T into unit squares, consists of finitely many line segments c1 , . . . , ck , k > 0, whose
endpoints are vertices of T, c = c1 . . . ck , and c is closed, i.e., c− = c+ . If ci ⊂ T then
the T-length |ci |T of ci is the number of edges of ci . If ci 6⊂ T then the T-length |ci |T
of ci is the number of connected components in ci \ T. We assume that |ci |T > 0
Pk
for every i and set |c|T := i=1 |ci |T .
Theorem 1.9. Suppose that n ≥ 1 is a fixed integer and c is a polygonal closed
curve in the plane R2 with given tessellation T into unit squares. Then, in deterministic space O((log |c|T )3 ) or in deterministic time O(|c|n+3
log |c|T ), one can
T
compute a rational number rn such that |A(c) − rn | < |c|1n .
T
In particular, if the area A(c) defined by c is known to be an integer multiple of
1
n
L , where L > 0 is a given integer and L < |c|T /2, then A(c) can be computed in
3
deterministic space O((log |c|T ) ) or in deterministic time O(|c|n+3
log |c|T ).
T
Proof. As above, let c = c1 . . . ck , where each ci is a line segment of c of positive
length that connects vertices of T. For every ci , we define an approximating path
ζ(ci ) such that ζ(ci )− = (ci )− , ζ(ci )+ = (ci )+ and ζ(ci ) ⊂ T. If ci ⊂ T then we set
ζ(ci ) := ci .
Assume that ci 6⊂ T. let Ri be a rectangle consisting of unit squares of T so that
ci is a diagonal of Ri . Consider the set N (ci ) ⊆ Ri of all squares s of Ri such that
the intersection s ∩ ci is not empty and is not a single point. Assuming that the
boundary path ∂N (ci ) is negatively, i.e., clockwise, oriented, we represent ∂N (ci )
in the form
∂N (ci ) = q1 (ci )q2 (ci )−1 ,
where q1 (ci ), q2 (ci )−1 are subpath of ∂N (ci ) defined by the equalities
q1 (ci )− = q2 (ci )− = (ci )− ,
q1 (ci )+ = q2 (ci )+ = (ci )+ .
It is easy to see that these equations uniquely determine the paths q1 (ci ), q2 (ci ),
see Fig. 10.3.
88
S. V. IVANOV
q1 (ci )
Ri
ci
q2 (ci )
Fig. 10.3
Lemma 10.3. Suppose ci , Ri are defined as above, ci 6⊂ T, and |Ri |x , |Ri |y are the
lengths of horizontal, vertical, resp., sides of the rectangle Ri . Then
max(|Ri |x , |Ri |y ) ≤ |ci |T ≤ 2 max(|Ri |x , |Ri |y ),
(10.10)
and the area A(ci q1 (ci )−1 ) bounded by the polygonal closed curve ci q1 (ci )−1 satisfies
A(ci q1 (ci )−1 ) ≤ |ci |T .
(10.11)
In addition, the paths q1 (ci ), q2 (ci ) can be constructed in deterministic space
O(log |c|T ) or in deterministic time O(|ci |T log |c|T ).
Proof. We say that all unit squares of Ri whose points have x-coordinates in the
same range compose a column of Ri . Similarly, we say that all unit squares of Ri
whose points have y-coordinates in the same range compose a row of Ri . Since ci is
a diagonal of Ri , it follows that ci has an intersection, different from a single point,
with a unit square from every row and from every column of Ri . Hence, in view of
the definition of the T-length |ci |T of ci , we have max(|Ri |x , |Ri |y ) ≤ |ci |T .
It follows from the definitions of the region N (Ri ) ⊆ Ri and the T-length |ci |T
that the number of unit squares in N (Ri ) is |ci |T , hence, the area A(N (Ri )) is
equal to |ci |T . Since the closed curve ci q1 (ci )−1 is contained in N (Ri ) and it can
be turned into a simple curve by an arbitrarily small deformation, it follows that
A(ci q1 (ci )−1 ) ≤ A(N (Ri )) = |ci |T , as required in (10.11).
Let sl(ci ) denote the slope of the line that goes through ci . If | sl(ci )| ≤ 1 then
we can see that N (ci ) contains at most 2 squares in every column of Ri . On the
other hand, if | sl(ci )| ≥ 1 then N (ci ) contains at most 2 squares in every row of Ri .
Therefore, A(N (ci )) ≤ 2 max(|Ri |x , |Ri |y ). Since A(N (ci )) = |ci |T , the inequalities
(10.10) are proven.
Note that, moving along columns of Ri if sl(ci ) ≤ 1 or moving along rows of Ri if
sl(ci ) > 1, we can detect all squares of Ri that belong to the region N (ci ) in space
O(log |c|T ) or in time O(|ci |T log |c|T ) by checking which squares in current column
or row are intersected by ci in more than one point. This implies that the paths
q1 (ci ), q2 (ci ) can also be constructed in space O(log |c|T ) or in time O(|ci |T log |c|T ),
as desired.
For ci 6⊂ T, define ζ(ci ) := q1 (ci ). Recall that ζ(ci ) = ci if ci ⊂ T. We can now
define an approximating closed path ζ(c) in T for c by setting
ζ(c) := ζ(c1 ) . . . ζ(ck ).
BOUNDED AND PRECISE WORD PROBLEMS FOR GROUP PRESENTATIONS
89
By Lemma 10.3,
|A(c) − A(ζ(c))| ≤
k
X
A(ci ζ(ci )−1 ) ≤
k
X
|ci |T = |c|T .
i=1
i=1
Consider a refined tessellation TM , where M > 1 is an integer, so that every unit
square s of T is divided into M 2 congruent squares each of area M −2 . We repeat
the foregoing definitions of the lengths |ci |TM , |c|TM , the paths qj,M (ci ), j = 1, 2,
i = 1, . . . , k, rectangles Ri,M , regions NM (ci ) ⊆ Ri,M , and approximating paths
ζM (ci ), ζM (c) with respect to the refined tessellation TM in place of T.
Lemma 10.4. Suppose ci 6⊂ T, and |Ri,M |x , |Ri,M |y denote the path length of
horizontal, vertical, resp., sides of the rectangle Ri,M . Then |Ri,M |x = M |Ri |x ,
|Ri,M |y = M |Ri |y ,
max(|Ri,M |x , |Ri,M |y ) ≤ |ci |TM ≤ 2 max(|Ri,M |x , |Ri,M |y ),
(10.12)
and the area bounded by the polygonal closed curve ci q1,M (ci )−1 satisfies
A(ci q1,M (ci )−1 ) ≤ M −2 |ci |TM .
(10.13)
M |ci |T /2 ≤ |ci |TM ≤ 2M |ci |T ,
(10.14)
In addition, we have
and the paths q1,M (ci ), q2,M (ci ) can be constructed in deterministic space
O(log |c|TM ) = O(log(M |c|T ))
or in deterministic time O(|ci |TM log |c|TM ) = O(M |ci |T log(M |c|T )).
Proof. The equalities |Ri,M |x = M |Ri |x , |Ri,M |y = M |Ri |y are obvious from the
definitions. Proofs of inequalities (10.12)–(10.13) are analogous to the proofs of
inequalities (10.10)–(10.11) of Lemma 10.3 with the correction that the area of a
square of the tessellation TM is now M −2 .
The inequalities (10.14) follow from inequalities (10.10), (10.12) and imply that
the space and time bounds O(log |c|TM ), O(|ci |TM log |c|TM ), that are obtained as
corresponding bounds of Lemma 10.3, can be rewritten in the form O(log(M |c|T )),
O(M |ci |T log(M |c|T )), resp.
It follows from the definitions and Lemma 10.4 that
|A(c) − A(ζM (c))| ≤
k
X
A(ci ζM (ci )−1 ) ≤ M −2
|ci |TM
i=1
i=1
≤ M −2
k
X
k
X
2M |ci |T = 2M −1 |c|T
(10.15)
i=1
and that the path ζM (ci ) ⊂ TM can be computed in space O(log(M |c|T )) or in time
O(M |ci |T log(M |c|T )).
Setting M := |c|n+2
, where n ≥ 1 is a fixed integer, we can see from Lemma 10.4
T
that the closed path ζM (c) ⊂ TM can be constructed in space O(log |c|T ) or
in time O(|c|n+3
log |c|T ). Hence, by Theorem 1.8 and Lemma 10.4, the area
T
A(ζM (c)) can be computed in space O((log |c|TM )3 ) = O((log |c|T )3 ) or in time
90
S. V. IVANOV
O(|c|3TM log |c|TM ) = O(|c|n+3
log |c|T ). The inequality (10.15) together with M =
T
|c|n+2
imply
that
T
|A(c) − A(ζM (c))| ≤ 2M −1 |c|T < |c|−n
T ,
(10.16)
here we may assume |c|T > 2 for otherwise A(c) = 0 and rn = 0.
Finally, suppose that the area A(c) is known to be an integer multiple of L1 , where
L > 0 is an integer with L < |c|nT /2. Applying the foregoing approximation result,
in deterministic space O((log |c|T )3 ) or in deterministic time O(|c|n+3
log |c|T ), we
T
can compute a rational number rn = A(ζM (c)), where M = |c|n+2
,
such
that the
T
inequality (10.16) holds true. It follows from Lemma 9.1, inequalities (10.14) and
the definitions that both numerator and denominator of rn are nonnegative integers
that do not exceed
2(n+3)
max(|c|2TM /4, M 2 ) ≤ max(M 2 |c|2T , M 2 ) ≤ |c|T
.
Hence, a binary representation of rn takes space O(log |c|T ).
Since it is known that L1 > |c|2n , it follows from (10.16) that there is at most one
T
integer multiple of
1
L
at the distance <
1
|c|n
T
from rn . This means that an integer
1
L
multiple of closest to rn is the area A(c).
Since rn and L are available and their binary representations take space O(log |c|T ),
it follows that a closest to rn integer multiple of L1 , can be computed in space
O(log |c|T ) or in time O((log |c|T )2 ) and this will be the desired area A(c).
Thus the area A(c) bounded by c can be computed in space O(log |c|3T ) or in
time O(|c|n+3
log |c|T ). Theorem 1.9 is proved.
T
Corollary 1.10. Let K ≥ 1 be a fixed integer and let c be a polygonal closed curve
in the plane R2 with given tessellation T into unit squares such that c has one of
the following two properties (a)–(b).
(a) If ci , cj are two nonparallel line segments of c then their intersection point,
1
.
if it exists, has coordinates that are integer multiples of K
(b) If ci is a line segment of c and ai,x , ai,y are coprime integers such that the
line given by an equation ai,x x + ai,y y = bi , where bi is an integer, contains ci , then
max(|ai,x |, |ai,y |) ≤ K.
Then the area A(c) defined by c can be computed in deterministic space O((log |c|T )3 )
or in deterministic time O(|c|n+3
log |c|T ), where n depends on K.
T
In particular, if T∗ is a tessellation of the plane R2 into equilateral triangles of
unit area, or into regular hexagons of unit area, and q is a finite closed path in T∗
whose edges are edges of T∗ , then the area A(q) defined by q can be computed in
deterministic space O((log |q|)3 ) or in deterministic time O(|q|5 log |q|).
Proof. Assume that the property (a) holds for c. Let t be a triangle in the plane
whose vertices v1 , v2 , v3 are points of the intersection of some nonparallel segments
of c. Then, using the determinant formula
v−→
v ,−
v−→
v ]|
A(t) = 1 | det[−
2
1 2
1 3
1
for the area A(t) of t, we can see that A(t) is an integer multiple of 2K
2 . Since
A(c) is the sum of areas of triangles such as t discussed above, it follows that A(c)
1
is also an integer multiple of 2K
2 . Taking n so that
|c|nT > 4K 2 ,
(10.17)
BOUNDED AND PRECISE WORD PROBLEMS FOR GROUP PRESENTATIONS
91
we see that Theorem 1.9 applies with L = 2K 2 and yields the desired conclusion.
Suppose that the property (b) holds for c. By the Cramer’s rule applied to the
system of two linear equations
ai,x x + ai,y y = bi
aj,x x + aj,y y = bj
which, as in property (b), define the lines that contain nonparallel segments ci , cj ,
we have that the intersection point of ci and cj has rational coordinates whose
denominators do not exceed
a
ai,y
det i,x
≤ 2K 2 .
aj,x aj,y
This means that coordinates of the intersection point of two nonparallel line
segments ci , cj of c are integer multiples of (2K1 2 )! and the case when the property (b)
holds for c is reduced to the case when the property (a) holds for c with K ′ = (2K 2 )!.
It remains to show the last claim of Corollary 1.10. The case when q is a closed
path in the tessellation T6 of the plane R2 into regular hexagons of unit area can
be obviously reduced to the case when q is a closed path in the tessellation T3 of
the plane R2 into equilateral triangles of unit area. For this reason we discuss the
latter case only.
Consider the standard tessellation T = T4 of the plane R2 into unit squares and
draw a diagonal in each square s of T4 which connects the lower left vertex and
the upper right vertex of s. Let T4,1 denote thus obtained tessellation of R2 into
triangles of area 1/2. Note that T3 and T4,1 are isomorphic as graphs and the
areas of corresponding triangles differ by a coefficient of 2±1 . This isomorphism
enables us to define the image q ′ in T4,1 of a closed path q in T3 . It is clear that
the area A(q ′ ) defined by q ′ relative to T4,1 is half of the area A(q) defined by q
relative to T3 , A(q ′ ) = A(q)/2. We can also consider q ′ as a polygonal closed curve
relative to the standard tessellation T = T4 of R2 into unit squares. Note that
|q ′ |T4 = |q ′ |T4,1 = |q|T3 and that the property (a) of Corollary 1.10 holds for q ′ with
the constant K = 1 relative to T4 . Thus, by proven part (a), the area A(q ′ ) can be
computed in space O((log |q ′ |T4 )3 ) = O((log |q|T3 )3 ) = O((log |q|)3 ) or in time
O(|q ′ |5T4 log |q ′ |T4 ) = O(|q|5T3 log |q|T3 ) = O(|q|5 log |q|)
for the reason that K = 1 and we can use n = 2 in (10.17) unless |q ′ |T4 = 2 in
which case A(q ′ ) = A(q) = 0. Since A(q) = 2A(q ′ ), our proof is complete.
It is tempting to try to lift the restrictions of Corollary 1.10 to be able to compute,
in polylogarithmic space, the area A(c) defined by an arbitrary polygonal closed
curve c in the plane equipped with a tessellation T into unit squares. Approximation approach of Theorem 1.9 together with construction of actual elementary
homotopies for the approximating path ζM (c) of Theorem 1.8 that seem to indicate
the sign of the pieces A(ci ζM (ci )−1 ) of the intermediate area between c and ζM (c),
both done in polylogarithmic space, provide certain credibility to this idea.
However, in the general situation, this idea would not work because the rational
number A(c) might have an exponentially large denominator, hence, A(c) could take
polynomial space just to store (let alone the computations). An example would be
provided by a polygonal closed curve c = c(n), where n ≥ 2 is an integer, such that
92
S. V. IVANOV
|c|T < (n + 1)2 and the denominator of A(c) is greater than 2k1 n−1 , where k1 > 0
is a constant. Below are details of such an example.
Let n ≥ 2 be an integer and let p1 , . . . , pℓ be all primes not exceeding n. Recall
that there is a constant k1 > 0 such that p1 . . . pℓ > 2k1 n , see [14]. We construct
line segments c3i+1 , c3i+2 , c3i+3 for i = 0, . . . , ℓ − 1 and c3ℓ+1 of a polygonal closed
curve c = c(n) by induction as follows. Let the initial vertex (c1 )− of c1 be the
point (with coordinates) (0, 0), (c1 )− := (0, 0). Proceeding by induction on i ≥ 0,
assume that the point (c3i+1 )− = (x0 , y0 ) is already defined. Then the line segment
c3i+1 goes from the point (c3i+1 )− to the point (x0 , y0 + 1) = (c3i+1 )− + (0, 1),
written c3i+1 = [(c3i+1 )− , (c3i+1 )− + (0, 1)]. Next, we set
c3i+2 := [(c3i+1 )+ , (c3i+1 )+ + (−1, 0)],
c3i+3 := [(c3i+2 )+ , (c3i+2 )+ + (pi+1 , −1)]
and, completing the induction step, define (c3(i+1)+1 )− := (c3i+3 )+ , see Fig. 10.4,
where the case ℓ = 3 with p1 = 2, p2 = 3, p3 = 5 is depicted.
c2
c5
c3
c1
c8
c9
c6
c4
c7
c10
(c1 )− = (c10 )+ = (0, 0)
Fig. 10.4
Finally, we define (c3ℓ+1 )− := [(c3ℓ )+ , (c1 )− ] and c = c(n) := c1 . . . c3ℓ c3ℓ+1 , see
Fig. 10.4 where c3ℓ+1 = c10 .
It is not difficult to check that
|c|T = ℓ + 2
ℓ
X
pi < (n + 1)2
i=1
and that the area A(c) defined by c is the following sum of areas of 2ℓ triangles
A(c) =
ℓ
X
1
2
1
pi
+ (pi − 1)(1 −
1
pi )
i=1
=
ℓ
X
i=1
1
pi
+
1
2
ℓ
X
i=1
(pi − 2) =
Pℓ
=
1
2
ℓ
X
+ (pi − 1) − 1 +
1
pi
i=1
p1 ...pℓ
pi
p1 ...pℓ
i=1
1
pi
+
1
2
ℓ
X
=
(pi − 2).
i=1
Since p1 . . . pℓ > 2k1 n , it follows that, after possible cancelation of 2, the denominator of A(c) is greater than 2k1 n−1 .
It would be of interest to study similar problems for tessellation of the hyperbolic
plane into regular congruent 2g-gons, where g ≥ 2, which would be technically close
to the precise word problem and to the minimal diagram problem for the standard
group presentation
−1
−1
−1
h a1 , a2 . . . , a2g−1 , a2g k a1 a2 a−1
1 a2 . . . a2g−1 a2g a2g−1 a2g = 1 i
of the fundamental group of an orientable closed surface of genus g ≥ 2. Hopefully,
there could be developed a version of calculus of brackets for such presentations
that would be suitable for these problems.
BOUNDED AND PRECISE WORD PROBLEMS FOR GROUP PRESENTATIONS
93
It is likely that suitable versions of calculus of brackets could be developed for
problems on efficient planar folding of RNA strands which would provide polylogarithmic space algorithms for such problems. Recall that available polynomial time
algorithms for such problems, see [27], [28], [29], [40], use polynomial space. This
approach would give a chance to do, also in polylogarithmic space, maximization of
foldings relative to various parameters similar to those discussed in Theorem 1.5.
Acknowledgements. The author is grateful to Tim Riley for bringing to the author’s attention folklore arguments that solve the precise word problem for presentation h a, b k a = 1, b = 1 i in polynomial time, the article [20], and the similarity
between the precise word problem for presentation h a, b k a = 1, b = 1 i and the
problem on efficient planar folding of RNA strands. The author thanks the referee
for a number of useful remarks and suggestions.
References
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
[21]
[22]
A. V. Anisimov, The group languages, Kibernetika (Kiev) 4(1971), 18–24.
R. Armoni, A. Ta-Shma, A. Wigderson and S. Zhou, A (log n)4/3 space algorithm for (s, t)
connectivity in undirected graphs, J. Assoc. Comput. Mach. 47(2000), 294–311.
S. Arora and B. Barak, Computational complexity – a modern approach, Cambridge Univ.
Press, 2009.
D. Barrington, P. Kadau, K. Lange and P. McKenzie, On the complexity of some problems
on groups input as multiplication tables, J. Comput. System Sci. 63(2001), 186–200.
J.-C. Birget, A. Yu. Ol’shanskii, E. Rips, and M. V. Sapir, Isoperimetric functions of groups
and computational complexity of the word problem, Ann. Math. 156(2002), 467–518.
W. W. Boone, On certain simple undecidable problems in group theory, V, VI, Indag. Math.
19(1957), 22–27, 227–232.
W. W. Boone, The word problem, Ann. Math. 70(1959), 207–265.
V. V. Borisov, Simple examples of groups with unsolvable word problems, Mat. Zametki
6(1969), 521–532.
D. Cummins, Dehn functions, the word problem, and the bounded word problem for decidable
group presentations, preprint, arXiv:1212.2024 [math.GR].
M. Dehn, Über unendliche diskontinuierliche Gruppen, Math. Ann. 71(1911), 116–144.
R. I. Grigorchuk and P. F. Kurchanov, On the width of elements in free groups, Ukrainian
Math. J. 43(1991), 911–918.
R. I. Grigorchuk and S. V. Ivanov, On Dehn functions of infinite presentations of groups,
Geom. Funct. Anal. 18(2008), 1841–1874.
M. Gromov, Hyperbolic groups, Essays in Group Theory(S. Gersten, ed.), MSRI Publ. 8,
Springer-Verlag, 1987, 75–263.
G. H. Hardy and E. M. Wright, An introductton to the theory of numbers, Oxford Univ.
Press, London, 6th ed., 2008.
S. V. Ivanov, The free Burnside groups of sufficiently large exponents, Internat. J. Algebra
Comp. 4(1994), 1–308.
R. J. Lipton and Y. Zalcstein, Word problems solvable in logspace, J. Assoc. Comput. Mach.
24(1977), 522–526.
R. J. Lipton, L. Snyder, and Y. Zalcstein, The complexity of word and isomorphism problems
for finite groups, Proc. Conf. on Information Sciences and Systems, 1976, pp. 33–35.
R. C. Lyndon and P. E. Schupp, Combinatorial group theory, Springer-Verlag, 1977.
W. Magnus, J. Karras, D. Solitar, Combinatorial group theory, Interscience Publ., 1966.
A. Majumdar, J. M. Robbins, M. Zyskin, Tangent unit-vector fields: nonabelian homotopy
invariants and the Dirichlet energy, C. R. Math. Acad. Sci. Paris 347(2009), 1159–1164.
A. Majumdar, J. M. Robbins, M. Zyskin, Tangent unit-vector fields: nonabelian homotopy
invariants and the Dirichlet energy, Acta Math. Sci. Ser. B Engl. Ed. 30 30(2010), 1357–
1399.
A. A. Markov, The impossibility of certain algorithms in the theory of associative systems
I, Dokl. Akad. Nauk 55(1947), 583–586.
94
[23]
[24]
[25]
[26]
[27]
[28]
[29]
[30]
[31]
[32]
[33]
[34]
[35]
[36]
[37]
[38]
[39]
[40]
S. V. IVANOV
A. A. Markov, The impossibility of certain algorithms in the theory of associative systems
II, Dokl. Akad. Nauk 58(1947), 353–356.
N. Megiddo and U. Vishkin, On finding a minimum dominating set in a tournament, Theoret. Comput. Sci. 61(1988), 307–316.
P. S. Novikov, On algorithmic unsolvability of the problem of identity, Doklady Akad. Nauk
SSSR 85(1952), 709–712.
P. S. Novikov, On the algorithmic unsolvability of the word problem in group theory, Trudy
Mat. Inst. Steklov 44(1955), 1–143.
R. Nussinov, G. Pieczenik, J. Griggs, and D. Kleitman, Algorithms for loop matchings,
SIAM J. Appl. Math. 35(1978), 68–82.
R. Nussinov and A. B. Jacobson, Fast algorithm for predicting the secondary structure of
single stranded RNA, Proc. Natl. Acad. Sci. USA 77(1980), 6309–6313.
R. Nussinov, B. Shapiro, and S. Le, J. V. Maizel, Speeding up the dynamic algorithm for
planar RNA folding. Math. Biosci. 100(1990), 33–47.
A. Yu. Ol’shanskii, Geometry of defining relations in groups, Nauka, Moscow, 1989; English
translation: Math. and Its Applications, Soviet series, vol. 70, Kluwer Acad. Publ., 1991.
A. Yu. Ol’shanskii, On calculation of width in free groups, London Math. Soc. Lecture Note
Ser. 204(1995), 255–258.
C. H. Papadimitriou, Computational complexity, Addison-Wesley Publ., 1994.
C. H. Papadimitriou and M. Yannakakis, On limited nondeterminism and the complexity
of the V-C dimension, J. Comput. System Sci. 53(1996), 161–170.
E. L. Post, Recursive unsolvability of a problem of Thue, J. Symb. Logic 12(1947), 1–11.
T. R. Riley, Private communication.
T. R. Riley, Computing area in presentations of the trivial group, preprint,
arXiv:1606.08833 [math.GR].
J. J. Rotman, An introduction to the theory of groups, Springer-Verlag, 1995.
W. J. Savitch, Relationships between nondeterministic and deterministic tape complexities,
J. Comput. System Sci. 4(1970), 177–192.
A. M. Turing, The word problem in semigroups with cancellation, Ann. Math. 52(1950),
491–505.
M. Zuker and P. Stiegler, Optimal computer folding of large RNA sequences using thermodynamics and auxiliary information, Nucleic Acids Res. 9(1981), 133–148.
Department of Mathematics, University of Illinois, 1409 West Green Street, Urbana, IL 61801, USA
E-mail address: [email protected]
| 4math.GR
|
arXiv:1608.04132v3 [cs.DS] 14 Mar 2017
Instantaneous Reaction-Time in
Dynamic-Consistency Checking of Conditional
Simple Temporal Networks
Massimo Cairo
Carlo Comin
Romeo Rizzi
Department of Mathematics
University of Trento, Trento, Italy
([email protected])
Department of Mathematics
University of Trento, Trento, Italy;
LIGM (UMR 8049)
Université Paris-Est, Marne-la-Vallée, France
([email protected])
Department of Computer Science
University of Verona, Verona, Italy
([email protected])
Abstract—Conditional Simple Temporal Network (CSTN) is
a constraint-based graph-formalism for conditional temporal
planning. Three notions of consistency arise for CSTNs and
CSTPs: weak, strong, and dynamic. Dynamic-Consistency (DC) is
the most interesting notion, but it is also the most challenging. In
order to address the DC-Checking problem, in [5] we introduced
ε-DC (a refined, more realistic, notion of DC), and provided
an algorithmic solution to it. Next, given that DC implies εDC for some sufficiently small ε > 0, and that for every
ε > 0 it holds that ε-DC implies DC, we offered a sharp lower
bounding analysis on the critical value of the reaction-time ε̂
under which the two notions coincide. This delivered the first
(pseudo) singly-exponential time algorithm for the DC-Checking
of CSTNs. However, the ε-DC notion is interesting per se, and
the ε-DC-Checking algorithm in [5] rests on the assumption that
the reaction-time satisfies ε > 0; leaving unsolved the question of
what happens when ε = 0. In this work, we introduce and study
π-DC, a sound notion of DC with an instantaneous reaction-time
(i.e., one in which the planner can react to any observation at the
same instant of time in which the observation is made). Firstly, we
demonstrate by a counter-example that π-DC is not equivalent to
0-DC, and that 0-DC is actually inadequate for modeling DC with
an instantaneous reaction-time. This shows that the main results
obtained in our previous work do not apply directly, as they were
formulated, to the case of ε = 0. Motivated by this observation, as
a second contribution, our previous tools are extended in order
to handle π-DC, and the notion of ps-tree is introduced, also
pointing out a relationship between π-DC and HyTN-Consistency.
Thirdly, a simple reduction from π-DC to DC is identified. This
allows us to design and to analyze the first sound-and-complete
π-DC-Checking procedure. Remarkably, the time complexity of
the proposed algorithm remains (pseudo) singly-exponential in
the number of propositional letters.
Index Terms—Conditional Simple Temporal Networks,
Dynamic-Consistency, Instantaneous Reaction-Time, Hyper Temporal Networks, Singly-Exponential Time.
I. I NTRODUCTION AND M OTIVATION
In temporal planning and temporal scheduling, Simple Temporal Networks (STNs) [6] are directed weighted graphs, where
This work was partially supported by Department of Computer Science,
University of Verona, Verona, Italy, under PhD grant “Computational Mathematics and Biology”; and, for the 2nd author, on a co-tutélle agreement with
LIGM, Université Paris-Est in Marne-la-Vallée, Paris, France.
nodes are events to be scheduled in time and arcs represent
temporal distance constraints between pairs of events.
This work is focused on the Conditional Simple Temporal
Problem (CSTP) [12] and its graph-based counterpart Conditional Simple Temporal Network (CSTN) [10], a constraintbased model for conditional temporal planning. The CSTN
formalism extends STNs in that: (1) some of the nodes are
called observation events and to each of them is associated
a propositional letter, to be disclosed only at execution time;
(2) labels (i.e. conjunctions over the literals) are attached to
all nodes and constraints, to indicate the scenarios in which
each of them is required. The planning agent (planner) must
schedule all the required nodes, while respecting all the required temporal constraints among them. This extended framework allows for the off-line construction of conditional plans
that are guaranteed to satisfy complex temporal constraints.
Importantly, this can be achieved even while allowing the
decisions about the precise timing of actions to be postponed
until execution time, in a least-commitment manner, thereby
adding flexibility and making it possible to adapt the plan
dynamically, during execution, in response to the observations
that are made [12].
Then, three notions of consistency arise for CSTNs: weak,
strong, and dynamic. Dynamic-Consistency (DC) is the most
interesting one, as it requires the existence of conditional
plans where decisions about the precise timing of actions are
postponed until execution time, but anyhow guaranteeing that
all of the relevant constraints will be ultimately satisfied. Still,
it is the most challenging one and it was conjectured to be
hard to assess by Tsamardinos, Vidal and Pollack [12].
In our previous work [5], it was unveiled that HyTNs and
MPGs are a natural underlying combinatorial model for the
DC-Checking of CSTNs. Indeed, STNs have been recently
generalized into Hyper Temporal Networks (HyTNs) [3], [4],
by considering weighted directed hypergraphs, where each hyperarc models a disjunctive temporal constraint called hyperconstraint. In [3], [4], the computational equivalence between
checking the consistency of HyTNs and determining the win-
ning regions in Mean Payoff Games (MPGs) was also pointed
out. The approach was shown to be viable and robust thanks
to some extensive experimental evaluations [4]. MPGs [1],
[7], [13] are a family of two-player infinite games played on
finite graphs, with direct and important applications in modelchecking and formal verification [8], and they are known for
having theoretical interest in computational complexity for
their special place among the few (natural) problems lying
in NP ∩ coNP.
All this combined, in [5] it was provided the first (pseudo)
singly-exponential time algorithm for the DC-Checking problem, also producing a dynamic execution strategy whenever
the input CSTN is DC. For this, it was introduced ε-DC (a
refined, more realistic, notion of DC), and provided the first
algorithmic solution to it. Next, given that DC implies ε-DC
for some sufficiently small ε > 0, and that for every ε > 0
it holds that ε-DC implies DC, it was offered a sharp lower
bounding analysis on the critical value of the reaction-time
ε̂ under which the two notions coincide. This delivered the
first (pseudo) singly-exponential time algorithm for the DCChecking of CSTN. However, the ε-DC notion is interesting
per se, and the ε-DC-Checking algorithm in [5] rests on the
assumption that the reaction-time satisfies ε > 0; leaving
unsolved the question of what happens when ε = 0.
Contribution: In this work we introduce and study πDC, a sound notion of DC with an instantaneous reaction-time
(i.e., one in which the planner can react to any observation at
the same instant of time in which the observation is made).
Firstly, we provide a counter-example showing that π-DC is
not just the ε = 0 special case of ε-DC. This implies that the
algorithmic results obtained in [5] do not apply directly to the
study of those situation where the planner is allowed to react
instantaneously. Motivated by this observation, as a second
contribution, we extend the previous formulation to capture a
sound notion of DC with an instantaneous reaction-time, i.e.,
π-DC. Basically, it turns out that π-DC needs to consider an
additional internal ordering among all the observation nodes
that occur at the same time instant. Next, the notion of pstree is introduced to reflect the ordered structure of π-DC,
also pointing out a relationship between π-DC and HyTNConsistency. Thirdly, a simple reduction from π-DC to DC
is identified. This allows us to design and to analyze the
first sound-and-complete π-DC-Checking procedure. The time
complexity of the proposed algorithm remains (pseudo) singlyexponential in the number |P | of propositional letters.
II. BACKGROUND
This section provides some background and preliminary notations. To begin, if G = (V, A) is a directed weighted graph,
every arc a ∈ A is a triplet (u, v, wa ) where u = t(a) ∈ V
is the tail of a, v = h(a) ∈ V is the head of a, and
wa = w(u, v) ∈ Z is the (integer) weight of a.
Let us now recall Simple Temporal Networks (STNs) [6].
Definition 1 (STNs). An STN [6] is a weighted directed
graph whose nodes are events that must be placed on the real
time line and whose arcs, called standard arcs, express binary
constraints on the allocations of their end-points in time.
An STN G = (V, A) is called consistent if it admits a
feasible schedule, i.e., a schedule φ : V 7→ R such that
φ(v) ≤ φ(u) + w(u, v) for all arcs (u, v, w(u, v)) ∈ A.
A. Conditional Simple Temporal Networks
Let us recall the CSTN model from [5], [10], [12].
Let P be a set of propositional letters (boolean variables),
a label is any (possibly empty) conjunction of letters, or
negations of letters, drawn from P . The empty label is denoted
by λ. The set of all these labels is denoted by P ∗ . Two labels,
`1 and `2 , are called consistent, denoted by Con(`1 , `2 ), when
`1 ∧ `2 is satisfiable. A label `1 subsumes a label `2 , denoted
by Sub(`1 , `2 ), whenever `1 ⇒ `2 holds. We are now in the
position to recall the formal definition of CSTNs.
Definition 2 (CSTNs [10], [12]). A Conditional Simple Temporal Network (CSTN) is a tuple hV, A, L, O, OV, P i where:
• V is a finite set of events; P = {p1 , . . . , p|P | } is a finite
set of propositional letters;
• A is a set of labelled temporal constraints (LTCs) each
having the form hv − u ≤ w(u, v), `i, where u, v ∈ V ,
w(u, v) ∈ Z, and ` ∈ P ∗ is satisfiable;
∗
• L : V → P assigns a label to each event in V ;
• OV ⊆ V is a finite set of observation events; O : P →
OV is a bijection that associates a unique observation
event O(p) = Op to each proposition p ∈ P ;
• The following reasonability assumptions must hold:
(WD1) for any LTC hv − u ≤ w, `i ∈ A the label ` is
satisfiable and subsumes both L(u) and L(v); intuitively,
whenever a constraint hv − u ≤ wi is required, then its
endpoints u and v must be scheduled (sooner or later);
(WD2) for each p ∈ P and each u ∈ V such that either
p or ¬p appears in L(u), we require: Sub(L(u), L(Op )),
and hOp −u ≤ 0, L(u)i ∈ A; intuitively, whenever a label
L(u), for some u ∈ V , contains some p ∈ P , and u is
eventually scheduled, then Op = O(p) must be scheduled
before or at the same time of u.
(WD3) for each labelled constraint hv − u ≤ w, `i and
p ∈ P , for which either p or ¬p appears in `, it holds
that Sub(`, L(Op )); intuitively, assuming that a required
constraint contains some p ∈ P , then Op = O(p) must
be scheduled (sooner or later).
In all of the following definitions we shall implicitly refer
to some CSTN Γ = hV, A, L, O, OV, P i.
Definition 3 (Scenario). A scenario over a subset U ⊆ P of
boolean variables is a truth assignment s : U → {0, 1}, i.e.,
s is a function that assigns a truth value to each proposition
p ∈ U . When U ( P and s : U → {0, 1}, then s is said
to be a partial scenario; otherwise, when U = P , then s is
said to be a (complete) scenario. The set comprising all of the
complete scenarios over P is denoted by ΣP . If s ∈ ΣP is a
scenario and ` ∈ P ∗ is a label, then s(`) ∈ {0, 1} denotes
the truth value of ` induced by s in the natural way.
Notice that any scenario s ∈ ΣP can be described by means
of the label `s , l1 ∧· · ·∧l|P | such that, for every 1 ≤ i ≤ |P |,
the literal li ∈ {pi , ¬pi } satisfies s(li ) = 1.
Example 1. Consider the set of boolean variables P = {p, q}.
The scenario s : P → {0, 1} defined as s(p) = 1 and s(q) = 0
can be compactly described by the label `s = p ∧ ¬q.
Definition 4 (Schedule). A schedule for a subset of events
U ⊆ V is a function φ : U → R that assigns a real number to
each event in U . The set of all schedules over U is denoted ΦU .
Definition 5 (Scenario-Restriction). Let s ∈ ΣP be a scenario.
The restriction of V , OV , and A w.r.t. s are defined as follows:
+
+
+
• Vs , {v ∈ V | s(L(v)) = 1}; OVs , OV ∩ Vs ;
+
• As , {hu, v, wi | ∃` hv − u ≤ w, `i ∈ A, s(`) = 1}.
+
+
The restriction of Γ w.r.t. s ∈ ΣP is the STN Γ+
s , hVs , As i.
+
+
Finally, it is worth to introduce the notation Vs1 ,s2 , Vs1 ∩Vs+2 .
Definition 6 (Execution-Strategy). An Execution-Strategy
(ES) for Γ is a mapping σ : ΣP → ΦVs+ such that, for any
scenario s ∈ ΣP , the domain of the schedule σ(s) is Vs+ . The
set of ESs of Γ is denoted by SΓ . The execution time of an event
v ∈ Vs+ in the schedule σ(s) ∈ ΦVs+ is denoted by [σ(s)]v .
Definition 7 (History). Let σ ∈ SΓ be any ES, let s ∈ ΣP be
any scenario and let τ ∈ R. The history Hst(τ, s, σ) of τ in
the
scenario
s under strategy σ is defined as: Hst(τ, s, σ) ,
p, s(p) ∈ P × {0, 1} | Op ∈ Vs+ , [σ(s)]Op < τ .
The history can be compactly encoded as the conjunction
of the literals corresponding to the observations comprising it,
that is, by means of a label.
Definition 8 (Viable Execution-Strategy). We say that σ ∈ SΓ
is a viable ES if, for each scenario s ∈ ΣP , the schedule
σ(s) ∈ ΦVs+ is feasible for the STN Γ+
s .
Definition 9 (Dynamic-Consistency). An ES σ ∈ SΓ is called
dynamic if, for any s1 , s2 ∈ ΣP and any v ∈ Vs+1 ,s2 , the
following implication holds on τ , [σ(s1 )]v :
Con(Hst(τ, s1 , σ), s2 ) ⇒ [σ(s2 )]v = τ.
We say that Γ is dynamically-consistent (DC) if it admits σ ∈
SΓ which is both viable and dynamic. The problem of checking
whether a given CSTN is DC is named DC-Checking.
We provide next the definition of difference set ∆(s1 ; s2 ).
Definition 10 (Difference-Set). Let s1 , s2 ∈ ΣP be any two
scenarios. The set of observation events in OVs+1 at which s1
and s2 differ is denoted by ∆(s1 ; s2 ). Formally,
∆(s1 ; s2 ) , Op ∈ OVs+1 | s1 (p) 6= s2 (p) .
The various definitions of history and dynamic consistency
that are used by different authors [5], [11], [12] are equivalent.
B. Hyper Temporal Networks
This subsection surveys the Hyper Temporal Network
(HyTN) model, which is a strict generalization of STNs. The
reader is referred to [3], [4] for an in-depth treatise on HyTNs.
Definition 11 (Hypergraph). A hypergraph H is a pair (V, A),
where V is the set of nodes, and A is the set of hyperarcs. Each
hyperarc A = (tA , HA , wA ) ∈ A has a distinguished node
tA called the tail of A, and a nonempty set HA ⊆ V \ {tA }
containing the heads of A; to each head v ∈ HA is associated
a weight wA (v) ∈ Z.
Provided that |A| , |HA ∪ {tA }|,P
the size of a hypergraph
H = (V, A) is defined as mA ,
A∈A |A|; it is used as
a measure for the encoding length of H. If |A| = 2, then
A = (u, v, w) can be regarded as a standard arc. In this way,
hypergraphs generalize graphs.
A HyTN is a weighted hypergraph H = (V, A) where
a node represents an event to be scheduled, and a hyperarc
represents a set of temporal distance constraints between the
tail and the heads.
In the HyTN framework the consistency problem is the
following decision problem.
Definition 12 (HyTN-Consistency). Given some HyTN
H = (V, A), decide whether there is a schedule φ : V → R
such that:
φ(tA ) ≥ min {φ(v) − wA (v)}, ∀ A ∈ A
v∈HA
any such a schedule φ : V → R is called feasible.
A HyTN is called consistent whenever it admits at least one
feasible schedule. The problem of checking whether a given
HyTN is consistent is named HyTN-Consistency.
Theorem 1. [3] There exists an O((|V |+|A|)mA W ) pseudopolynomial time algorithm for checking HyTN-Consistency;
moreover, when the input HyTN H = (V, A) is consistent, the
algorithm returns as output a feasible schedule φ : V → R
of H; Here, W , maxA∈A,v∈HA |wA (v)|.
C. ε-Dynamic-Consistency
In CSTNs, decisions about the precise timing of actions are
postponed until execution time, when informations meanwhile
gathered at the observation nodes can be taken into account.
However, the planner is allowed to factor in an outcome, and
differentiate its strategy according to it, only strictly after the
outcome has been observed (whence the strict inequality in
Definition 7). Notice that this definition does not take into
account the reaction-time, which, in most applications, is nonnegligible. In order to deliver algorithms that can also deal
with the reaction-time ε > 0 of the planner, we introduced
in [5] a refined notion of DC.
Definition 13 (ε-Dynamic-Consistency). Given any CSTN
hV, A, L, O, OV, P i and any real number ε ∈ (0, +∞), an ES
σ ∈ SΓ is ε-dynamic if it satisfies all of the Hε -constraints,
namely, for any two scenarios s1 , s2 ∈ ΣP and any event
u ∈ Vs+1 ,s2 , the ES σ satisfies the following constraint, which
is denoted by Hε (s1 ; s2 ; u):
[σ(s1 )]u ≥ min {[σ(s2 )]u }∪{[σ(s1 )]v +ε | v ∈ ∆(s1 ; s2 )}
We say that a CSTN Γ is ε-dynamically-consistent (ε-DC) if
it admits σ ∈ SΓ which is both viable and ε-dynamic.
v1
A,
A, −ε
v2
>
s1
A, 0
us1
[1]
−ε
us2
Fig. 1: An Hε (s1 ; s2 ; u) constraint, modeled as a hyperarc.
a?
+1
III. DC WITH I NSTANTANEOUS R EACTION -T IME
Theorem 2 points out the equivalence between ε-DC and
DC, that arises for a sufficiently small ε > 0. However,
Definition 13 makes sense even if ε = 0, so a natural
question is what happens to the above mentioned relationship
between DC and ε-DC when ε = 0. In this section we first
show that 0-DC doesn’t imply DC, and, moreover, that 0-DC
is in itself too weak to capture an adequate notion of DC
with an instantaneous reaction-time. In light of this we will
introduce a stronger notion, which is named ordered-DynamicConsistency (π-DC); this will turn out to be a suitable notion
of DC with an instantaneous reaction-time.
c?
B
0, ¬a
0, ¬b
−1
C
0, a
0, c
0
0, ¬c
0
0, b
0
⊥
[0]
Theorem 2. Let ε̂ , |ΣP |−1 |V |−1 . Then, Γ is DC if and only
if Γ is ε̂-DC. Moreover, if Γ is ε-DC for some ε > 0, then Γ
is ε0 -DC for every ε0 ∈ (0, ε].
Theorem
3. There exists an O |ΣP |3 |A|2 |V | +
4
2
|ΣP | |A||V | |P | + |ΣP |5 |V |3 |P | W time algorithm for
checking DC on any input CSTN Γ = hV, A, L, O, OV, P i.
In particular, given any dynamically-consistent CSTN Γ, the
algorithm returns a viable and dynamic ES for Γ.
Here, W , maxa∈A |wa |.
0, ¬a¬b
b?
A
As shown in [5], ε-DC can be modeled in terms of HyTNConsistency. Fig. 1 depicts an illustration of an Hε (s1 ; s2 ; u)
constraint, modeled as an hyperarc.
Also, in [5] we proved that DC coincides with ε̂-DC,
provided that ε̂ , |ΣP |−1 |V |−1 .
Then, the main result offered in [5] is a (pseudo) singlyexponential time DC-checking procedure (based on HyTNs).
0, ac
0, b¬c
s2
Fig. 2: A CSTN Γ2 which is 0-DC but not DC.
– [σ2 (s)]B , s(a ∧ b ∧ c) + s(a ∧ ¬b ∧ c);
– [σ2 (s)]C , s(¬a ∧ ¬b ∧ ¬c) + s(¬a ∧ ¬b ∧ c);
– [σ2 (s)]⊥ , 0 and [σ2 (s)]> , 1, for every s ∈ ΣP .
An illustration of σ2 is offered in Fig 3. Three cubical
graphs are depicted in which every node is labelled as vs for
some (v, s) ∈ V2 ×ΣP2 : an edge connects v1s1 and v2s2 if and
only if: (i) v1 = v2 and (ii) the Hamming distance between
s1 and s2 is unitary; each scenario s ∈ ΣP2 is represented
as s = αβγ for α, β, γ ∈ {0, 1}, where s(a) = α, s(b) = β,
s(c) = γ; moreover, each node vs = (v, s) ∈ V2 × ΣP2 is
filled in black if [σ2 (s)]v = 0, and in white if [σ2 (s)]v = 1. So
all three 3-cubes own both black and white nodes, but each of
them, in its own dimension, decomposes into two identically
colored 2-cubes. Fig. 4 offers another visualization of σ2 in
A101
A001
A111
A011
A100
A000
A010
B101
B001
A110
B000
B111
C101
B011
B100
C001
B110
B010
C000
C111
C011
C100
C110
C010
Fig. 3: The ES σ2 for the CSTN Γ2 .
Example 2 (CSTN Γ2 ). Consider the following CSTN Γ2 =
(V2 , A2 , L2 , O2 , OV2 , P2 ); see Fig. 2 for an illustration.
– V2 = {⊥, >, A, B, C};
– A2 = {(> − ⊥ ≤ 1, λ), (⊥ − > ≤ −1, λ), (> − A ≤
0, b∧¬c), (>−B ≤ 0, a∧c), (>−C ≤ 0, ¬a∧¬b), (⊥−A ≤
0, λ), (A − ⊥ ≤ 0, ¬b), (A − ⊥ ≤ 0, c), (⊥ − B ≤ 0, λ), (B −
⊥ ≤ 0, ¬a), (A − ⊥ ≤ 0, ¬c), (⊥ − C ≤ 0, λ), (C − ⊥ ≤
0, a), (C − ⊥ ≤ 0, b)};
– L2 (A) = L2 (B) = L2 (C) = L2 (⊥) = L2 (>) = λ;
– O2 (a) = A, O2 (b) = B, O2 (c) = C;
– OV2 = {A, B, C};
– P2 = {a, b, c}.
which every component of the depicted graph corresponds to
a restriction STN Γ+
2 s for some s ∈ ΣP2 , where si , sj ∈ ΣP2
+
are grouped together whenever Γ+
2 si = Γ2 sj . It is easy to
see from Fig. 4 that σ2 is viable for Γ2 . In order to check
that σ2 is 0-dynamic, look again at Fig. 4, and notice that for
every si , sj ∈ Σ2 , where si 6= sj , there exists an event node
X ∈ {A, B, C} such that [σ2 (si )]X = 0 = [σ2 (sj )]X and
si (X) 6= sj (X). With this in mind it is easy to check that all
of the H0 constraints are thus satisfied by σ2 . Therefore, the
CSTN Γ2 is 0-DC.
Proposition 1. The CSTN Γ2 (Example 2, Fig. 2) is 0-DC.
Proposition 2. The CSTN Γ2 is not DC.
Proof. Consider the execution strategy σ2 : ΣP2 → ΨV2 :
– [σ2 (s)]A , s(a ∧ b ∧ ¬c) + s(¬a ∧ b ∧ ¬c);
Proof. Let σ be a viable ES for Γ2 . Then, σ must be the ES σ2
depicted in Fig. 4, there is no other choice here. Let ŝ ∈ ΣP2 .
ΣP2
[1]
0
⊥
0
0
A
B
C
0
0
1
0
>
000; 001
1
[1]
0
0
1
0
0
1
A
B
C
>
010; 110
A
B
C
>
011; 100
0
0
0
1
0
0
0
⊥
0
0
[1]
0
0
⊥
0
Definition 16 (π-Dynamic-Consistency). Any σ ∈ SΓ is called
π-dynamic when, for any two scenarios s1 , s2 ∈ ΣP and any
event v ∈ Vs+1 ,s2 , if τ , [σ(s1 )]tv and ψ , [σ(s1 )]πv , then:
Con(π-Hst(τ, ψ, s1 , σ), s2 ) ⇒ [σ(s2 )]tv = τ, [σ(s2 )]πv = ψ.
⊥
0
We are finally in the position to define π-DC.
0
1
0
1
A
B
C
>
101; 111
0
[1]
Fig. 4: The restrictions Γ+
2 s for s ∈ ΣP2 , where the execution
times [σ2 (s)]v ∈ {0, 1} are depicted in bold face.
We say that Γ is π-dynamically-consistent (π-DC) if it admits
σ ∈ SΓ which is both viable and π-dynamic. The problem
of checking whether a given CSTN is π-DC is named π-DCChecking.
Remark 1. It is easy to see that, due to the strict inequality
“[σ(s)]πOp < ψ” in the definition of π-Hst(·) (Definition 15), in
a π-dynamic π-ES, there must be exactly one Op0 ∈ OV , for
some p0 ∈ P , which is executed at first (w.r.t. both execution
time and position) under all possible scenarios s ∈ ΣP .
Proposition 3. The CSTN Γ2 is not π-DC.
Proof. The proof goes almost in the same way as that of
Proposition 2. In particular, no observation event is executed
first (i.e., at time t = 0 and position ψ = 1) in all possible
scenarios. Since there is no first-in-time observation event,
then, the ES σ is not π-dynamic.
We provide next a CSTN which is π-DC but not DC.
Then, it is easy to check from Fig. 4 that: (i) [σ2 (ŝ)]⊥ =
0, [σ2 (ŝ)]> = 1, and it holds [σ2 (ŝ)]X ∈ {0, 1} for every
X ∈ {A, B, C}; (ii) there exists at least two observation events
X ∈ {A, B, C} such that [σ(ŝ)]X = 0; still, (iii) there is no
X ∈ {A, B, C} such that [σ(s)]X = 0 for every s ∈ ΣP2 , i.e.,
no observation event is executed first at all possible scenarios.
Therefore, the ES σ2 is not dynamic.
We now introduce a stronger notion of dynamic consistency;
it is named ordered-Dynamic-Consistency (π-DC), and it takes
explicitly into account an additional ordering between the
observation events scheduled at the same execution time.
Definition 14 (π-Execution-Strategy). An ordered-ExecutionStrategy (π-ES) for Γ is a mapping:
σ : s 7→ ([σ(s)]t , [σ(s)]π ),
where s ∈ ΣP , [σ(s)]t ∈ ΦV , and finally, [σ(s)]π : OVs+
{1, . . . , |OVs+ |} is bijective. The set of π-ES of Γ is denoted
by SΓ . For any s ∈ ΣP , the execution time of an event
v ∈ Vs+ in the schedule [σ(s)]t ∈ ΦVs+ is denoted by
[σ(s)]tv ∈ R; the position of an observation Op ∈ OVs+ in
σ(s) is [σ(s)]πOp . We require positions to be coherent w.r.t. execution times, i.e., ∀(Op , Oq ∈ OVs+ ) if [σ(s)]tOp < [σ(s)]tOq
then [σ(s)]πOp < [σ(s)]πOq . In addition, it is worth to adopt the
notation [σ(s)]πv , |OV | + 1 whenever v ∈ Vs+ \ OV .
Definition 15 (π-History). Let σ ∈ SΓ , s ∈ ΣP , and
let τ ∈ R and ψ ∈ {1, . . . , |V |}. The ordered-history
π-Hst(τ, ψ, s, σ) of τ and ψ in the scenario
s,under the π-ES
σ, is defined as: π-Hst(τ, ψ, s, σ) , p, s(p) ∈ P × {0, 1} |
Op ∈ OVs+ , [σ(s)]tOp ≤ τ, [σ(s)]πOp < ψ .
Example 3. Define Γπ = (Vπ , Aπ , Oπ , OVπ , Pπ ) as follows.
Vπ = {Op , X, >}, Aπ = {(> − Op ≤ 1, λ), (Op − > ≤
−1, λ), (X − Op ≤ 0, p), (> − X ≤ 0, ¬p)}, Oπ (p) = Op ,
OVπ = {Op }, Pπ = {p}. Fig. 5 depicts the CSTN Γπ .
p?
−1
Op
1
>
0, ¬p
0, p
X
Fig. 5: The CSTN Γπ .
Proposition 4. The CSTN Γπ is π-DC, but it is not DC.
Proof. Let s1 , s2 ∈ ΣPπ be two scenarios such that s1 (p) =
1 and s2 (p) = 0. Consider the π-ES σ defined as follows:
[σ(s1 )]tOp = [σ(s1 )]tX = 0, [σ(s1 )]t> = 1; and [σ(s2 )]tOp = 0,
[σ(s2 )]tX = [σ(s2 )]t> = 1; finally, [σ(s)]πOp = 1, [σ(s)]πX =
[σ(s)]π> = 2, for all s ∈ {s1 , s2 }. Then, σ is viable and πdynamic for Γπ . To see that Γπ is not DC, pick any ε > 0.
Notice that any viable ES must schedule X either at t = 0 or
t = 1, depending on the outcome of Op , which in turn happens
at t = 0; however, in any ε-dynamic strategy, the planner can’t
react to the outcome of Op before time t = ε > 0. This implies
that Γπ is not ε-DC. Since ε was chosen arbitrarily (ε > 0),
then Γπ can’t be DC by Theorem 2.
So Γπ is ε-DC for ε = 0 but for no ε > 0. In summary, the
following chain of implications holds on the various DCs:
6⇐
6⇐
[ε-DC, ∀ε ∈ (0, ε̂]] ⇔ DC ⇒ π-DC ⇒ [ε-DC, for ε = 0]
where ε̂ , |ΣP |−1 · |V |−1 as in Theorem 2.
A. The ps-tree: a “skeleton” structure for π-dynamic π-ESs
In this subsection we introduce a labelled tree data structure,
named the ps-tree, which turns out to capture the “skeleton”
ordered structure of π-dynamic π-ESs.
Definition 17 (PS-Tree). Let P be any set of boolean variables. A permutation-scenario tree (ps-tree) πT over P is an
outward (non-empty) rooted binary tree such that:
• Each node u of πT is labelled with a letter pu ∈ P ;
• All the nodes that lie along a path leading from the root
to a leaf are labelled with distinct letters from P .
• Each arc (u, v) of πT is labelled by some b(u,v) ∈ {0, 1};
• The two arcs (u, vl ) and (u, vr ) exiting a same node u
have opposite labels, i.e., b(u,vl ) 6= b(u,vr ) .
Fig. 6 depicts an example of a ps-tree.
1
c
1
0
d
1
c
0
c
a
0
1
d
0
d
1
d
0
d
b
b
algorithm for π-DC, though of doubly exponential complexity
in |P |; a bound to be improved in later subsections, but that can
help sizing the sheer dimensionality and depth of the problem.
Firstly, the notion of Expansion of CSTNs is recalled [5].
Definition 19 (Expansion hVΓEx , ΛEx
Γ i). Consider a CSTN Γ =
(V, A, L, O, OV, P ). Consider the family of all (distinct) STNs
(Vs , As ), one for each scenario s ∈ ΣP , defined as follows:
Vs , {vs | v ∈ Vs+ } and As , {(us , vs , w) | (u, v, w) ∈ A+
s }.
The expansion hVΓEx , ΛEx
Γ i of the CSTN Γ is defined as follows:
[
[
A
hVΓEx , ΛEx
i
,
V
,
s .
s
Γ
s∈ΣP
Ex
Notice, (VΓEx , ΛEx
Γ ) is an STN with at most |VΓ | ≤ |ΣP |·|V |
Ex
nodes and at most |ΛΓ | ≤ |ΣP | · |A| standard arcs.
We now show that the expansion of a CSTN can be enriched
with some standard arcs and some hyperarcs in order to model
the π-DC property, by means of an HyTN denoted H0πT (Γ).
Definition 20 (HyTN H0πT (Γ)). Let Γ = (V, A, L, O, OV, P )
be a given CSTN. Let πT be a given c-ps-tree over P .
Then, the HyTN H0πT (Γ) is defined as follows:
+
• For every scenarios s1 , s2 ∈ ΣP and u ∈ Vs1 ,s2 \ OV ,
define a hyperarc α = α0 (s1 ; s2 ; u) as follows (with the
intention to model H0 (s1 ; s2 ; u), see Definition 13):
α = α0 (s1 ; s2 ; u) , htα , Hα , wα i,
Fig. 6: An example of a ps-tree over P = {a, b, c, d}.
Definition 18 (πs , si , Coherent-PS-Tree). Let πT be a ps-tree
over P , let r be the root and s be any leaf. Let (r, v2 , . . . , s)
be the sequence of the nodes encountered along the path going
from r down to s in πT . Then:
• The sequence of labels πs = (pr , pv2 , . . . , ps ) is a permutation of the subset of letters {pr , pv2 , . . . , ps } ⊆ P .
• Each sequence of bits (b(r,v2 ) , . . . , b(vi ,vi+1 ) ), for each
i ∈ {1, 2, . . . , ks − 1} (where v1 , r and vks , s),
can be seen as a partial scenario si over P ; i.e., define
si (vj ) , b(vj ,vj+1 ) , for every j ∈ {1, . . . , i}.
• πT is coherent (c-ps-tree) with Γ if, for every leaf s of πT ,
{Opr , Opv2 , . . . , Ops } = OVs+0 holds for every complete
scenario s0 ∈ ΣP such that Sub(s0 , sks −1 ).
It is not difficult to see that a π-dynamic π-ES induces one
and only one c-ps-tree πT . So, the existence of a suitable c-pstree is a necessary condition for a π-ES to be π-dynamic. One
may ask whether a π-dynamic π-ES can be reconstructed from
its c-ps-tree; the following subsection answers affirmatively.
B. Verifying a c-ps-tree: on π-DC and HyTN-Consistency.
This subsection builds on the notion of c-ps-tree to work
out the details of the relationship between π-DC and HyTNConsistency. Once this picture is in place, it will be easy to
reduce to HyTN-Consistency the problem of deciding whether
a given CSTN admits a valid π-dynamic π-ES with a given cps-tree. This easy result already provides a first combinatorial
s∈ΣP
•
where:
– tα , us1 is the tail of the hyperarc α;
– Hα , {us2 } ∪ ∆(s1 ; s2 ) is the set of the heads;
– wα (us2 ) , 0; ∀(v ∈ ∆(s1 ; s2 )) wα (v) , 0.
Now,
of the CSTN Γ hVΓEx , ΛEx
Γ i=
S consider
Sthe expansion
V
,
A
(as
in
Definition
19).
Then:
s
s
s∈ΣP
s∈ΣP
For each internal node x of πT , A0x is a set of (additional)
standard arcs defined as follows. Let πx = (r, . . . , x0 )
be the sequence of all and only the nodes along the
path going from the root r to the parent x0 of x in πT
(where we can assume r0 = r). Let Px ⊆ P be the
corresponding literals, px excluded, i.e., Px , {pz ∈ P |
z appears in πx and pz is the label of z in πT } \ {px }.
Let sx be the partial scenario defined as follows:
λ,
if x = r;
sx : Px → {0, 1} :
pz 7→ b(z,z0 ) , if x 6= r.
where z 0 is the unique child of z in πT lying on πx .
Let x0 (x1 ) be the unique child of x in πT such that
bx,x0 = 0 (bx,x1 = 1). For every complete s0x ∈ ΣP such
that Sub(s0x , sx ), we define:
0
(Opx0 )s0x , (Opx )s0x , 0 , if sx0 (x) = 0;
Bs0 0x ,
(Opx1 )s0x , (Opx )s0x , 0 , if sx (x) = 1.
Also, for every complete s0x , s00x ∈ ΣP such that
Sub(s0x , sx) and Sub(s00x , sx ), where s0x =
6 s00x , we define:
0
0
00
Cs0x ,s00x , (Opx )sx , (Opx )sx , 0 .
from the root r down to some leaf s. Notice that the
sequence of labels (pr , pv2 , . . . , ps ) can be seen as a
Cs0 0x ,s00x .
A0x ,
Bs0 0x ∪
bijection, i.e., πs : OVs+0
{1, . . . , |OVs+0 |}. Then, for
0 6= s00 ,
+
s0x ∈ΣP : Sub(s0x ,sx )
0
s0x , s00
∈
Σ
:
s
P
x
x
x
any s ∈ ΣP and v ∈ OVs0 , define [σφ,πT (s0 )]πv , πs (v).
Sub(s0x , sx ), Sub(s00
,
s
)
x
x
It is not difficult to check from the definitions, at this point,
π
π
Ex
that since φ is feasible for H0πT (Γ), then σφ,πT must be viable
• Then, H0 (Γ) is defined as H0 (Γ) , hVΓ , AHπ (Γ) i,
0
and π-dynamic for the CSTN Γ. Hence, the CSTN Γ is π-DC.
where,
[
[
(2) The size bounds for H0πT (Γ) follow from Definition 20.
αε (s1 ; s2 ; u) ∪
AHπ0 (Γ) , ΛEx
A0x .
Γ ∪
s1 ,s2 ∈ΣP
x : internal
u∈Vs+ ,s
node of πT
In Fig. 7, Algorithm 1 presents the pseudocode for con1 2
structing
the HyTN H0πT (Γ), as prescribed by Definition 20.
Notice that the following holds: each α (s ; s ; u) has size
Finally,
[
[
ε
1
2
|αε (s1 ; s2 ; u)| = ∆(s1 ; s2 ) + 1 ≤ |P | + 1.
The following theorem establishes the connection between
the π-DC of CSTNs and the consistency of HyTNs.
Theorem 4. Given any CSTN Γ = hV, A, L, O, OV, P i, it
holds that the CSTN Γ is π-DC if and only if there exists
a c-ps-tree πT such that the HyTN H0πT (Γ) is consistent.
Moreover, H0πT (Γ) has at most |VHπ0 T (Γ) | ≤ |ΣP | |V | nodes,
|AHπ0 T (Γ) | = O(|ΣP | |A| + |ΣP |2 |V |) hyperarcs, and it has
size at most mAHπT (Γ) = O(|ΣP | |A| + |ΣP |2 |V | |P |).
0
Proof. (1) Firstly, we prove that the CSTN Γ is π-DC if and
only if there exists a c-ps-tree πT such that the HyTN H0πT (Γ)
is consistent.
(⇒) Let σ ∈ SΓ be a given viable and π-dynamic execution
strategy for the CSTN Γ. Since σ is π-dynamic, then for any
two s1 , s2 ∈ ΣP and any v ∈ Vs+1 ,s2 the following holds on
the execution time τ , [σ(s1 )]tv and position ψ , [σ(s1 )]πv :
Con(π-Hst(τ, ψ, s1 , σ), s2 ) ⇒ [σ(s2 )]tv = τ, [σ(s2 )]πv = ψ.
It is easy to see that this induces one and only one c-ps-tree πT :
indeed, due to Remark 1, there must be exactly one Op0 ∈ OV ,
for some p0 ∈ P , which is executed at first (w.r.t. to both
execution time and position) under all possible scenarios; then,
depending on the boolean result of p0 , a second observation
p00 can be differentiated, and it can occur at the same or at
a subsequent time instant, but still at a subsequent position;
again, by Remark 1, there is exactly one Op00 ∈ OV which
comes first under all possible scenarios that agree on p0 ; and so
on and so forth, thus forming a tree structure over P , rooted
at p0 , which is captured exactly by our notion of c-ps-tree.
Then, let φσ : VΓEx → R be the schedule of H0πT (Γ) defined
as: φσ (vs ) , [σ(s)]tv for every vs ∈ VΓEx , where s ∈ ΣP and
v ∈ Vs+ . It is not difficult to check from the definitions, at
this point, that all of the standard arc and hyperarc constraints
of H0πT (Γ) are satisfied by φσ , that is to say that φσ must be
feasible for H0πT (Γ). Hence, H0πT (Γ) is consistent.
(⇐) Assume that there exists a c-ps-tree πT such that the
HyTN H0πT (Γ) is consistent, and let φ : VΓEx → R be a
feasible schedule for H0πT (Γ). Then, let σφ,πT (s) ∈ SΓ be the
execution strategy defined as follows:
t
Ex
+
• [σφ,πT (s)]v , φ(vs ), ∀ vs ∈ VΓ , s ∈ ΣP , v ∈ Vs ;
0
0
• Let s ∈ ΣP be any complete scenario. Then, s induces
exactly one path in πT , in a natural way, i.e., by going
Algorithm 1: construct_H(Γ, πT )
1
2
3
4
5
6
7
8
9
10
11
12
Input: a CSTN Γ , hV, A, L, O, OV, P i, a c-ps-tree πT
coherent with Γ.
foreach (s ∈ ΣP ) do
Vs ← {vs | v ∈ Vs+ };
As ← {as | a ∈ A+
s };
VΓEx ← ∪s∈ΣP Vs ;
ΛEx
Γ ← ∪s∈ΣP As ;
foreach (s1 , s2 ∈ ΣP , s1 6= s2 ) do
foreach (u ∈ Vs+1 ,s2 \ OV ) do
t α ← u s1 ;
Hα ← {us2 } ∪ (∆(s1 ; s2 ));
wα (us2 ) ← 0;
foreach v ∈ ∆(s1 ; s2 ) do
wα (vs1 ) ← 0;
α0 (s1 ; s2 ; u) ← htα , Hα , wα i;
13
14
15
16
foreach (x : internal node of πT ) do
A0x ← as defined in Definition 20;
[
AHπT (Γ) ← ΛEx
α0 (s1 ; s2 ; u) ∪
Γ ∪
0
s1 ,s2 ∈ΣP
u∈Vs+ ,s
1
17
18
2
H0πT (Γ) ← hVΓEx , AH πT (Γ) i;
0
return H0πT (Γ);
[
A0x ;
x : internal
node of πT
If Γ is π-DC, there is an integral π-dynamic π-ES, as below.
Proposition 5. Assume Γ = hV, A, L, O, OV, P i to be π-DC.
Then, there is some π-ES σ ∈ SΓ which is viable, π-dynamic,
and integral, namely, for every s ∈ ΣP and every v ∈ Vs+ ,
the following integrality property holds:
[σ(s)]tv ∈ 0, 1, 2, . . . , MΓ ⊆ N,
where MΓ , |ΣP ||V | + |ΣP ||A| + |ΣP |2 |V | W .
Proof. By Theorem 4, since Γ is π-DC, there exists some c-pstree πT such that the HyTN H0πT (Γ) is consistent; moreover,
by Theorem 4 again, H0πT (Γ) has |VHπ0 T (Γ) | ≤ |ΣP | |V |
nodes and |AHπ0 T (Γ) | ≤ |ΣP | |A| + |ΣP |2 |V | hyperarcs. Since
H0πT (Γ) is consistent, it follows from Theorem 4 [also see
Lemma 1 and Theorem 8 in [4]] that H0πT (Γ) admits an
integral and feasible schedule φ such that:
φ : VHπ0 T (Γ) → 0, 1, 2, . . . , MΓ ,
where MΓ ≤ (|VHπ0 T (Γ) | + |AHπ0 T (Γ) |)W . Therefore, it holds
that MΓ ≤ (|ΣP | |V | + |ΣP | |A| + |ΣP |2 |V |)W .
Given a CSTN Γ and some c-ps-tree πT , it is thus easy to
check whether there exists some π-ES for Γ whose ordering
relations are exactly the same as those prescribed by πT .
Indeed, it is sufficient to construct H0πT (Γ) with Algorithm 1,
then checking the consistency of H0πT (Γ) with the algorithm
mentioned in Theorem 1. This results into Algorithm 2. The
corresponding time complexity is also that of Theorem 1.
Algorithm 2: check_π-DC_on_c-ps-tree(Γ, πT )
4
Input: a CSTN Γ , hV, A, L, O, OV, P i, a c-ps-tree πT
coherent with Γ.
H0πT (Γ) ← construct_H(Γ, πT ); // ref. Algorithm 1
φ ← check_HyTN-Consistency(H0πT (Γ)); // ref. Thm 1
if (φ is a feasible schedule of H0πT (Γ)) then
return hYES, φ, πT i;
5
return NO;
1
2
3
Fig. 7: Checking π-DC given a c-ps-tree πT , by reduction to
HyTN-Consistency.
Notice that, in principle, one could generate all of the possible c-ps-trees πT given P , one by one, meanwhile checking for
the consistency state of H0πT (Γ) with Algorithm 2. However,
it is not difficult to see that, in general, the total number f|P |
of possible c-ps-trees over P is not singly-exponential in |P |.
Indeed, a moment’s reflection revelas that for every n > 1 it
2
holds that fn = n · fn−1
, and f1 = 1. So, any algorithm based
on the exhaustive exploration of the whole space comprising
all of the possible c-ps-trees over P would not have a (pseudo)
singly-exponential time complexity in |P |. Nevertheless, we
have identified another solution, that allows us to provide
a sound-and-complete (pseudo) singly-exponential time πDC-Checking procedure: it is a simple and self-contained
reduction from π-DC-Checking to DC-Checking. This allows
us to provide the first sound-and-complete (pseudo) singlyexponential time π-DC-Checking algorithm which employs
our previous DC-Checking algorithm (i.e., that underlying
Theorem 3) in a direct manner, as a black box, thus avoiding
a more fundamental restructuring of it.
C. A Singly-Exponential Time π-DC-Checking Algorithm
This section presents a sound-and-complete (pseudo) singlyexponential time algorithm for solving π-DC, also producing
a viable and π-dynamic π-ES whenever the input CSTN is
really π-DC. The main result of this section goes as follows.
Theorem 5. There exists an algorithm for checking π-DC
on any input given CSTN Γ = (V, A, L, O, OV, P ) with the
following (pseudo) singly-exponential time complexity:
O |ΣP |4 |A|2 |V |3 + |ΣP |5 |A||V |4 |P | + |ΣP |6 |V |5 |P | W.
Moreover, when Γ is π-DC, the algorithm also returns a viable
and π-dynamic π-ES for Γ. Here, W , maxa∈A |wa |.
The algorithm mentioned in Theorem 5 consits of a simple
reduction from π-DC to (classical) DC in CSTNs.
Basically, the idea is to give a small margin γ so that the
planner can actually do before, in the sense of the time value
[σ(s)]v , what he did “before” in the ordering π. Given any ES
in the relaxed network, the planner would then turn it into a
π-ES for the original network (which has some more stringent
constraints), by rounding-down each time value [σ(s)]vto the
largest integer less than or equal to it, i.e., [σ(s)]v . The
problem is that one may (possibly) violate some constraints
when there is a “leap” in the rounding (i.e., a difference of one
unit, in the rounded value, w.r.t. what one would have wanted).
Anyhow, we have identified a technique that allows us to get
around this subtle case, provided that γ is exponentially small.
Definition 21. Relaxed CSTN Γ0 . Let Γ = hV, A, L, O, OV, P i
be any CSTN with integer constraints. Let γ ∈ (0, 1) be a real.
Define Γ0γ , hV, A0γ , L, O, OV, P i to be a CSTN that differs
from Γ only in the numbers appearing in the constraints.
Specifically, each constraint hu − v ≤ δ, `i ∈ A is replaced
in Γ0γ by a slightly relaxed constraint, hu − v ≤ δγ0 , `i ∈ A0γ ,
where:
δγ0 , δ + |V | · γ.
The following two lemmata hold for any CSTN Γ.
Lemma 1. Let γ be any real in (0, |V |−1 ).
If Γ is π-DC, then Γ0γ is DC.
Proof. Since Γ is π-DC, by Proposition 5, there exists an
integral, viable and π-dynamic, π-ES σ for Γ. Let us fix some
real γ ∈ (0, |V |−1 ). Define the ES σγ0 ∈ SΓ0γ as follows, for
every s ∈ ΣP and v ∈ Vs+ :
[σγ0 (s)]v , [σ(s)]tv + [σ(s)]πv · γ.
Since [σ(s)]πv ≤ |V |, then:
[σ(s)]πv · γ < |V | · |V |−1 = 1,
and so the total ordering of the values [σγ0 (s)]v , for a given s ∈
ΣP , coincides with [σ(s)]π . Hence, the fact that σγ0 is dynamic
follows directly from the π-dynamicity of σ. Moreover, no
LTC (u − v ≤ δγ0 , `) of Γ0γ is violated in any scenario s ∈ ΣP
since, if ∆0γ,u,v , [σγ0 (s)]u − [σγ0 (s)]v then:
∆0γ,u,v = [σ(s)]tu + [σ(s)]πu · γ − [σ(s)]tv + [σ(s)]πv · γ
≤ [σ(s)]tu − [σ(s)]tv + |V | · γ
≤ δ + |V | · γ = δ 0 .
So, σγ0 is viable. Since σγ0 is also dynamic, then Γ0γ is DC.
The next lemma shows that the converse direction holds as
well, but for (exponentially) smaller values of γ.
Lemma 2. Let γ be any real in (0, |ΣP |−1 · |V |−2 ).
If Γ0γ is DC, then Γ is π-DC.
Proof. Let σγ0 ∈ SΓ0γ be some viable and dynamic ES for Γ0γ .
Firstly, we aim at showing that, w.l.o.g., the following lower subsequent inequality is just a rewriting; finally, the last two
inequalities hold by (A) and by δγ0 = δ + |V | · γ (respectively).
bound holds:
Now, since we have the strict inequality [σ(s)]tu −[σ(s)]tv < δ+
[σγ0 (s)]v − [σγ0 (s)]v ≥ |V |·γ, for all s ∈ ΣP and v ∈ Vs+ . (LB)
1, and since [σ(s)]tu −[σ(s)]tv ∈ Z, then [σ(s)]tu −[σ(s)]tv ≤ δ,
This will allow us to simplify the rest of the proof. In order as desired. So, σ is viable. Since σ is both viable and πdynamic, then Γ is π-DC.
to prove it, let us pick any η ∈ [0, 1) such that:
[σγ0 (s)]v −η −k ∈ [0, |V |·γ), for no v ∈ V, s ∈ ΣP , k ∈ Z.
Observe that such a value η exists. Indeed, there are only
|ΣP | · |V | choices of pairs (s, v) ∈ ΣP × V and each pair
rules out a (circular) semi-open interval of length |V | · γ in
[0, 1), so the total measure of invalid values for η in the semiopen real interval [0, 1) is at most |ΣP | · |V | · |V | · γ < 1. So
η exists.
See Fig. 8 for an intuitive illustration of this fact.
|V | · γ
then, lγ ≥ |ΣP |−1 · |V |−2 .
In summary, Lemma 1 and Lemma 2 imply Theorem 6.
η
[0
γ
2γ
···
3γ
jγ
(j + 1)γ
···
1 − γ 1)
Fig. 8: An illustration of the proof of Lemma 2.
By subtracting η to all time values {[σγ0 (s)]v }v∈V,s∈ΣP we
can assume w.l.o.g. that η = 0 holds for the rest ofthis proof;
and thus, that (LB) holds. Now, define [σ(s)]tv , [σγ0 (s)]v ,
and let [σ(s)]π be the ordering induced by σγ0 (s). Observe that
σ is a well-defined π-ES (i.e., that [σ(s)]π is coherent w.r.t.
[σ(s)]t ), thanks to the fact that b·c is a monotone operator.
Since the ordering [σ(s)]π is the same as that of σγ0 (s), then
σ is π-dynamic.
It remains to prove that σ is viable. For this, take any
constraint (u − v ≤ δ, `) ∈ A in Γ, and suppose that:
[σγ0 (s)]u
−
[σγ0 (s)]v
≤
δγ0
Fig. 8 illustrates the proof of Lemma 2, in which a family
of (circular) semi-open intervals of length |V | · γ are depicted
as shaded rectangles. Lemma 2 ensures that at least one chunk
on length lγ ≥ 1 − |ΣP | · |V |2 · γ is not covered by the union
of those (circular) semi-open intervals, and it is therefore free
to host η; in Fig. 8, this is represented by the blue interval,
and η = j · γ for some j ∈ [0, γ −1 ). Also notice that γ can
be fixed as follows:
1
;
γ,
|ΣP | · |V |2 + 1
= δ + |V | · γ.
[σγ0 (s)]u − [σγ0 (s)]v
(A)
[σ(s)]tu − [σ(s)]tv
≤ δ.
≤ δ, then clearly
If
So, the interesting case that we really need to check is when:
0 < [σγ0 (s)]u − [σγ0 (s)]v − δ ≤ |V | · γ.
For this, we observe that the following (∗) holds by (LB):
0
[σγ (s)]u ≤ [σγ0 (s)]u − |V | · γ. (∗)
Also, it is clear that:
0
[σγ (s)]v > [σγ0 (s)]v − 1.
Theorem 6. Let Γ be a CSTN and let γ ∈ (0, |ΣP |−1 ·|V |−2 ).
Then, Γ is π-DC if and only if Γ0γ is DC.
This allows us to design a simple algorithm for solving πDC-Checking, by reduction to DC-Checking, which is named
Check-π-DC() (Algorithm 3). Its pseudocode follows below.
Algorithm 3: Check-π-DC(Γ)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
(∗∗)
Input: a CSTN Γ , hV, A, L, O, OV, P i
1
;
γ ← |Σ |·|V
|2 +1
P
0
Aγ ← hu − v ≤ δ + |V | · γ, `i | hu − v ≤ δ, `i ∈ A ;
Γ0γ ← hV, A0γ , L, O, OV, P i;
σγ0 ← check_DC(Γ0γ ); // see Theorem 3
if σγ0 is a viable and dynamic ES for Γ0γ then
η ← pick η ∈ [0, 1) as in the proof of Lemma 2;
foreach (s, v) ∈ ΣP × Vs+ do
[σγ0 (s)]v ← [σγ0 (s)]v − η; // shift by η;
let σ ∈ ΣΓ be constructed as follows;
foreach s ∈ ΣP do
foreach v ∈ Vs+j do
k
[σ(s)]tv ← [σγ0 (s)]v ;
[σ(s)]π ← the ordering on P induced by σγ0 (s);
return hYES, σi;
return NO;
Fig. 9: Checking π-DC by reduction to DC-Checking.
Then,
[σ(s)]tu − [σ(s)]tv = [σγ0 (s)]u − [σγ0 (s)]v
< ([σγ0 (s)]u − |V | · γ) − ([σγ0 (s)]v − 1)
≤ ([σγ0 (s)]u − [σγ0 (s)]v ) − |V | · γ + 1
≤ δγ0 − |V | · γ + 1
≤ δ + 1,
where: the first equality holds by definition of [σ(s)]tx , for
x ∈ {u, v}; the strict inequality holds by (∗) and (∗∗); the
Description of Algorithm 3: It takes in input a CSTN Γ.
When Γ is π-DC, it aims at returning hYES, σi, where σ ∈ SΓ
is a viable and π-dynamic π-ES for Γ. Otherwise, if Γ is
not π-DC, then Check-π-DC() (Algorithm 3) returns NO.
Of course the algorithm implements the reduction described
in Definition 21, whereas the π-ES is computed as prescribed
1
by Lemma 2. At line 1, we set γ ← |ΣP |·|V
|2 +1 . Then, at
0
lines 2-3, Γγ is constructed as in Definition
21, i.e., Γ0γ ←
0
0
hV, Aγ , L, O, OV, P i, where Aγ ← hu − v ≤ δ + |V | · γ, `i |
hu − v ≤ δ, `i ∈ A . At this point, at line 5, the DC-Checking
algorithm of Theorem 3 is invoked on input Γ0γ . Let σγ0 be its
output. If Γ0γ is not DC, then Check-π-DC() (Algorithm 3)
returns NO at line 15. When σγ0 is a viable and dynamic ES
for Γ0γ at line 5, then Check-π-DC() (Algorithm 3) proceeds
as follows. At line 6, some η ∈ [0, 1) is computed as in the
proof of Lemma 2, i.e., such that [σγ0 (s)]v − η − k ∈ [0, |V | ·
γ) holds for no v ∈ V, s ∈ ΣP , k ∈ Z. Notice that it is easy
to find such η in practice. Indeed, one may view the real semiopen interval [0, 1) as if it was partitioned into chunks (i.e.,
smaller semi-open intervals) of length γ; as observed in the
proof of Lemma 2, there are only |ΣP | · |V | choices of pairs
(s, v) ∈ ΣP × V , and each pair rules out a (circular) semiopen interval of length |V | · γ; therefore, there is at least one
chunk of length lγ ≥ |ΣP |−1 · |V |−2 , within [0, 1), where
η can be placed, and we can easily find it just by inspecting
(exhaustively) the pairs (s, v) ∈ ΣP ×V . In fact, the algorithm
underlying Theorem 3 always deliver an earliest-ES (i.e., one
in which the time values are the smallest possible, in the space
of all consistent ESs), so that for each interval of length |V |·γ,
the only time values that we really need to check and rule out
are |V | multiples of γ. Therefore, at line 6, η exists and it can
be easily found in time O(|ΣP | · |V |2 ). So, at line 7, for each
s ∈ ΣP and v ∈ Vs+ , the value [σγ0 (s)]v is shifted to the left
by setting [σγ0 (s)]v ← [σγ0 (s)]v − η. Then, the following π-ES
σ ∈ SΓ is constructed at lines 9-13: for each
s ∈ ΣP and
v ∈ Vs+ , the execution-time is set [σ(s)]tv ← [σγ0 (s)]v , and
the ordering [σ(s)]π follows the ordering on P that is induced
by σγ0 (s). Finally, hYES, σi is returned to output at line 14.
To conclude, we can prove the main result of this section.
Proof of Theorem 5. The correctness of Algorithm 3 follows
directly from Theorems 6 and 3, plus the fact that η ∈ [0, 1)
can be computed easily, at line 6, as we have already mentioned above. The (pseudo) singly-exponential time complexity of Algorithm 3 follows from that of Theorem 3 plus the
fact that all the integer weights in Γ are scaled-up by a factor
1/γ = |ΣP | · |V |2 + 1 in Γ0γ ; also notice that η ∈ [0, 1)
can be computed in time O(|ΣP | · |V |2 ), as we have already
mentioned. Therefore, all in, the time complexity stated in
Theorem 3 increases by a factor 1/γ = |ΣP | · |V |2 + 1.
IV. R ELATED W ORKS
This section discusses of some related approaches offered
in the current literature. The article of Tsamardinos, et al. [12]
introduced DC for CSTNs. Subsequently, this notion has been
analyzed and further formalized in [10], finally leading to a
sound notion of DC for CSTNs. However, neither of these
two works takes into account an instantaneous reaction-time.
Cimatti, et al. [2] provided the first sound-and-complete procedure for checking the Dynamic-Controllability of CSTNs with
Uncertainty (CSTNUs) and this algorithm can be employed
for checking DC on CSTNs as a special case. Their approach
is based on reducing the problem to solving Timed Game
Automata (TGA). However, solving TGAs is a problem of
much higher complexity than solving MPGs. Indeed, no upper
bound is given in [2] on the time complexity of their solution.
Moreover, neither ε-DC nor any other notion of DC with
an instantaneous reaction-time are dealt with in that work.
The first work to approach a notion of DC with an instantaneous reaction-time is [11]; its aim was to offer a soundand-complete propagation-based DC-checking algorithm for
CSTNs. The subsequent work [9] extended and amended [11]
so that to check ε-DC, both for ε > 0 and for ε = 0. However,
to the best of our knowledge, the worst-case complexity of
those algorithms is currently unsettled. Moreover, it is not
clear to us how one variant of the algorithm offered in [9], [11]
(i.e., the one that aims at checking DC with an instantaneous
reaction-time) can adequately handle cases like the CSTN
counter-example Γ2 that we have provided in Example 2.
In summary, we believe that the present work can possibly
help in clarifying DC with an instantaneous reaction-time
also when the perspective had to be that of providing soundand-complete algorithms based on the propagation of labelled
temporal constraints.
V. C ONCLUSION
The notion of ε-DC has been introduced and analysed
in [5] where an algorithm was also given to check whether
a CSTN is ε-DC. By the interplay between ε-DC and the
standard notion of DC, also disclosed in [5], this delivered
the first (pseudo) singly-exponential time algorithm checking
whether a CSTN is DC (essentially, DC-Checking reduces
to ε-DC-Checking for a suitable value of ε). In this paper,
we proposed and formally defined π-DC, a natural and sound
notion of DC for CSTNs in which the planner is allowed to
react instantaneously to the observations that are made during
the execution. A neat counter-example shows that π-DC with
instantaneous reaction-time is not just the special case of εDC with ε = 0. Therefore, to conclude, we offer the first
sound-and-complete π-DC-Checking algorithm for CSTNs.
The time complexity of the procedure is still (pseudo) singlyexponential in |P |. The solution is based on a simple reduction
from π-DC-Checking to DC-Checking of CSTNs.
R EFERENCES
[1] L. Brim, J. Chaloupka, L. Doyen, R. Gentilini, and J.F. Raskin. Faster
algorithms for mean-payoff games. Formal Methods in System Design,
38(2):97–118, 2011.
[2] A. Cimatti, L. Hunsberger, A. Micheli, R. Posenato, and M. Roveri.
Sound and complete algorithms for checking the dynamic controllability
of temporal networks with uncertainty, disjunction and observation. In
21st Intern. Symp. on Temp. Repres. and Reasoning, TIME 2014, Verona,
Italy, pages 27–36, 2014.
[3] C. Comin, R. Posenato, and R. Rizzi. A tractable generalization of
simple temporal networks and its relation to mean payoff games. In 21th
International Symposium on Temporal Representation and Reasoning
(TIME 2014), Verona, Italy, Sept 2014.
[4] C. Comin, R. Posenato, and R. Rizzi. Hyper temporal networks.
Constraints, pages 1–39, 2016.
[5] C. Comin and R. Rizzi. Dynamic consistency of conditional simple
temporal networks via mean payoff games: a singly-exponential time
DC-Checking. In 22nd International Symposium on Temporal Representation and Reasoning (TIME 2015), pages 19–28. IEEE CPS, September
2015.
[6] R. Dechter, I. Meiri, and J. Pearl. Temporal constraint networks.
Artificial Intelligence, 49(1–3):61–95, 1991.
[7] A. Ehrenfeucht and J. Mycielski. Positional strategies for mean payoff
games. International Journal of Game Theory, 8(2):109–113, 1979.
[8] E. Grädel, W. Thomas, and T. Wilke, editors. Automata, Logics, and
Infinite Games: A Guide to Current Research. Springer-Verlag New
York, Inc., New York, NY, USA, 2002.
[9] L. Hunsberger and R. Posenato. Checking the dynamic consistency
of conditional temporal networks with bounded reaction times. In
Proceedings of the Twenty-Sixth International Conference on Automated
Planning and Scheduling, ICAPS 2016, London, UK, June 12-17, 2016.,
pages 175–183, 2016.
[10] L. Hunsberger, R. Posenato, and C. Combi. The dynamic controllability
of conditional stns with uncertainty. In Proc. of the Plan. and Plan
Exec. for Real-World Syst.: Princip. and Pract. (PlanEx), ICAPS-2012,
page 121–128, Atibaia, Sao Paulo, Brazil, 2012.
[11] L. Hunsberger, R. Posenato, and C. Combi. A sound-and-complete
propagation-based algorithm for checking the dynamic consistency of
conditional simple temporal networks. In 22nd International Symposium on Temporal Representation and Reasoning, TIME 2015, Kassel,
Germany, September 23-25, 2015, pages 4–18, 2015.
[12] I. Tsamardinos, T. Vidal, and M. Pollack. Ctp: A new constraint-based
formalism for conditional, temporal planning. Constraints, 8(4):365–
388, 2003.
[13] U. Zwick and M. Paterson. The complexity of mean payoff games on
graphs. Theoretical Computer Science, 158:343–359, 1996.
| 8cs.DS
|
Coalescent-based species tree estimation:
a stochastic Farris transform ∗
arXiv:1707.04300v1 [cs.LG] 13 Jul 2017
Gautam Dasarathy†
Elchanan Mossel‡
Sebastien Roch¶
Robert Nowak§
Abstract
The reconstruction of a species phylogeny from genomic data faces two significant hurdles:
1) the trees describing the evolution of each individual gene—i.e., the gene trees—may differ
from the species phylogeny and 2) the molecular sequences corresponding to each gene often
provide limited information about the gene trees themselves. In this paper we consider an
approach to species tree reconstruction that addresses both these hurdles. Specifically, we
propose an algorithm for phylogeny reconstruction under the multispecies coalescent model
with a standard model of site substitution. The multispecies coalescent is commonly used
to model gene tree discordance due to incomplete lineage sorting, a well-studied populationgenetic effect.
In previous work, an information-theoretic trade-off was derived in this context between the
number of loci, m, needed for an accurate reconstruction and the length of the locus sequences,
k. It was shown
√ that to reconstruct an internal branch of length f , one needs m to be of the
2
order of 1/[f k]. That previous result was obtained under the molecular clock assumption,
i.e., under the assumption that mutation rates (as well as population sizes) are constant across
the species phylogeny.
Here we generalize this result beyond the restrictive molecular clock assumption, and obtain a new reconstruction algorithm that has the same data requirement (up to log factors).
Our main contribution is a novel reduction to the molecular clock case under the multispecies
coalescent. As a corollary, we also obtain a new identifiability result of independent interest:
for any species tree with n ≥ 3 species, the rooted species tree can be identified from the
distribution of its unrooted weighted gene trees even in the absence of a molecular clock.
∗
Keywords: Phylogenetic Reconstruction, Coalescent, Gene Tree/Species Tree, Distance Methods, Data Requirement.
†
Department of Electrical and Computer Engineering at Rice University. [email protected]
‡
Department of Mathematics and IDS at the Massachusetts Institute of Technology. [email protected]
§
Department of Electrical and Computer Engineering at the University of Wisconsin–Madison.
[email protected]
¶
Department of Mathematics at the University of Wisconsin–Madison. Supported by NSF grants DMS-1149312
(CAREER) and DMS-1614242. [email protected]
1
Introduction
Modern molecular sequencing technology has provided a wealth of data to assist biologists in the
inference of evolutionary relationships between species. Not only is it now possible to quickly
sequence a single gene across a wide range of species, but in fact thousands of genes—or entire
genomes—can be sequenced simultaneously. With this abundance of data comes new algorithmic and statistical challenges. One such challenge arises because phylogenomic inference entails
dealing with the interplay of two processes, as we now explain.
While the tree of life (also referred to as a species phylogeny) depicts graphically the history
of speciation of living organisms, each gene within the genomes of these organisms has its own
history. That history is captured by a gene tree. In practice, by contrasting the DNA sequences
of a common gene across many current species, one can reconstruct the corresponding gene tree.
Indeed the accumulation of mutations along the gene tree reflects, if imperfectly, the underlying
history. Much is known about the reconstruction of single-gene trees, a subject with a long history; see [SS03a, Fel04, Yan14, Ste16, War] for an overview. The theoretical computer science
literature, in particular, has contributed a deep understanding of the computational complexity
and data requirements of the problem, under standard stochastic models of sequence evolution
on a tree. See, e.g., [GF82, ABF+ 99, FK99, ESSW99a, ESSW99b, Att99, CGG02, SS02, KZZ03,
Mos03, Mos04, CT06, Roc06, MR06, BCMR06, MLP09, DMR11a, DMR11b, ADHR12, GMS12,
MHR13, DR13, RS].
But a gene tree is only an approximation to the species phylogeny. Indeed various evolutionary
mechanisms lead to discordance between gene trees and species phylogenies. These include the
transfer of genetic material between unrelated species, hybrid speciation events and a populationgenetic effect known as incomplete lineage sorting [Mad97]. The wide availability of genomic
datasets has brought to the fore the major impact these discordances have on phylogenomic inference [DBP05, DR09]. As a result, in addition to the stochastic process governing the evolution
of DNA sequences on a fixed gene tree, one is led to model the structure of the gene tree itself, in relation to the species phylogeny, through a separate stochastic process. The inference of
these complex, two-level evolutionary models is an active area of research. See the recent monographs [HRS10, Ste16, War] for an introduction.
In this paper, we focus on incomplete lineage sorting (from hereon ILS) and consider the reconstruction of a species phylogeny from multiple genes (or loci) under a standard populationgenetic model known as the multispecies coalescent [RY03a]. The problem is of great practical
interest in computational evolutionary biology and is currently the subject of intense study; see
e.g. [LYK+ 09, DR09, ALPE12, Nak13] for a survey. There is in particular a growing body of theoretical results in this area [DR06, DDBR09, DD10, MR10, LYP10, ADR11b, ADR11a, Roc13,
DNR14, RS15, DD14, CK15, RW15, MR15, ADR17, SRM], although much remains to be understood. This inference problem is also closely related to another active area of research, the
reconstruction of demographic history in population genetics. See e.g. [MFP08, BS14, KMRR15]
for some recent theoretical results.
A significant fraction of prior rigorous work on species phylogeny estimation in the presence
of ILS has been aimed at the case where “true” gene trees are assumed to be available. However, in reality, one needs to estimate gene trees from DNA sequences, leading to reconstruction
1
errors, and indeed there has been a recent thrust towards understanding the effect of this important
source of error in phylogenomic inference, both from empirical [KD07, MBW16] and theoretical [MR10, DD14, RS15, RW15, SRM] standpoints. Another option, which we adopt here, is
to bypass the reconstruction of gene trees altogether and infer the species history directly from
sequence data [DNR15, MR15, CK15].
In previous work on this latter approach [MR15], an optimal information-theoretic trade-off
was derived between the number of genes m needed to accurately reconstruct a species phylogeny
and the length of the genes k (which is linked to the quality of the phylogenetic signal that can
√ be
extracted from each separate gene). Specifically, it was shown that m needs to scale like 1/[f 2 k],
where f is the length of the shortest branch in the tree (which controls the extent of the ILS). This
result was obtained under a restrictive molecular clock assumption, where the leaves are equidistant
from the root; in essence, it was assumed that the mutation rates and population sizes do not vary
across the species phylogeny, which is rarely the case in practice.
In the current work, we design and analyze a new reconstruction algorithm that achieves the
same optimal data requirement (up to log factors) beyond the molecular clock assumption. Our
key contribution is of independent interest: we show how to transform sequence data to appear as
though it was generated under the multispecies coalescent with a molecular clock. We achieve this
through a novel reduction which we call a stochastic Farris transform. Our construction relies on
a new identifiability result: for any species phylogeny with n ≥ 3 species, the rooted species tree
can be identified from the distribution of the unrooted weighted gene trees even in the absence of
a molecular clock.
We state our main results formally in Section 2 and describe our new reduction in Section 3.
The proofs are in Sections B, C, D and E.
2
Background and main results
In this section, we state formally our main results and provide a high-level view of the proof.
2.1
Basic definitions
We begin with a brief description of our modeling assumptions. More details on the models, which
are standard in the phylogenetic literature (see e.g. [Ste16]), are provided in Section A.
Species phylogeny v. gene trees A species phylogeny is a graphical depiction of the evolutionary history of a set of species. The leaves of the tree correspond to extant species while internal
vertices indicate a speciation event. Each edge (or branch) corresponds to an ancestral population
and will be described here by two numbers: one that indicates the amount of time that the corresponding population lived, and a second one that specifies the rate of genetic mutation in that
population. Formally, we define the species phylogeny (or tree) as follows.
Definition 1 (Species phylogeny). A species phylogeny S = (Vs , Es ; r, ~τ , ~µ) is a directed tree
rooted at r ∈ Vs with vertex set Vs , edge set Es , and n labelled leaves L = {1, 2, . . . , n} such that
2
(a) the degree of all internal vertices is 3 except for the root r which has degree 2, and (b) each
edge e ∈ Es is associated with a length τe ∈ (0, ∞) and a mutation rate µe ∈ (0, ∞).
To be more precise, as is standard in coalescent theory (see, e.g., [Ste16]), the length τe of a branch
e ∈ Es is expressed in coalescent time units, which is the duration of the branch te divided by
its population size Ne . That is, τe = te /Ne . We pictorially represent species phylogenies as thick
shaded trees; see Fig. 1 for an example with n = 3 leaves.
r
µrs
r
s
⌧rs
s
⌧r3
⌧s1
µr3
µs2
µs1
⌧s2
2
3
1
2
3
1
(a)
(b)
Figure 1: (a) A species phylogeny with n = 3 leaves {1, 2, 3} and two internal vertices r and s.
The branch lengths are denoted by {τs1 , τs2 , τrs , τr3 }. Depicted above is the special case where
all population sizes are the same, in which case the species phylogeny is “ultrametric,” that is,
all leaves are the same distance from the root under ~τ . (b) The same phylogeny with branches
stretched by the corresponding mutation rates. The mutation rate-weighted branch lengths are
denoted by {µs1 , µs2 , µrs , µr3 }, with respect to which ultrametricity is in general lost—our focus
here.
While a species phylogeny describes the history of speciation, each gene has its own history
which is captured by a gene tree.
Definition 2 (Gene trees). A gene tree G(i) = (V (i) , E (i) ; r, ~δ(i) ) corresponding to gene i is a
directed tree rooted at r with vertex set V (i) and edge set E (i) , and the same labeled leaf set
L = {1, 2, . . . , n} as S such that (a) the degree of each internal vertex is 3, except the root r whose
(i)
degree is 2, and (b) each branch e ∈ E (i) is associated with a branch length δe ∈ (0, ∞).
In essence, these gene trees “evolve” on the species phylogeny. More specifically, following
[RY03b], we assume that a multispecies coalescent (MSC) process produces m independent random gene trees G(1) , G(2) , . . . , G(m) . This process is parametrized by the species phylogeny S.
In words, proceeding backwards in time, in each population, every pair of lineages entering from
descendant populations merge at a unit exponential rate. We describe this process formally in Algorithm 4 in Section A. For the present discussion, it suffices to think of the MSC as a random
3
process generating “noisy versions” of the species phylogeny. We highlight one key feature of the
gene trees: their topology may be distinct from that of the species phylogeny. This discordance,
which in this context is referred to as incomplete lineage sorting (see e.g. [DR09]), is a major challenge for species tree estimation from multiple genes. See Figure 2 for an illustration in the case
n = 3.
Gene 2
t4
t3
Gene 1
r
t2
s
t1
1
2
3
Figure 2: Two sample draws from the MSC on a species phylogeny with n = 3 leaves. The (rooted)
topology of Gene 1 (red gene) agrees with the topology of the underlying species phylogeny (i.e.,
species 1 and 2 are closest), while the topology of Gene 2 (blue gene) does not (here species 2 and
3 are closest instead).
Sequence data and inference problem The gene trees are not observed directly. Rather, they
are typically inferred from sequence data “evolving” on the gene trees. We model this sample
generation process according to the standard Jukes-Cantor (JC) model (see, e.g., [Ste16]). That is,
given a gene tree G(i) = (V (i) , E (i) ; r, ~δ(i) ), we associate to each e ∈ E (i) , a probability
(i)
3
− 43 δe
1
−
e
,
p(i)
=
e
4
(i)
where δe is the mutation rate-weighted edge length. In words, the corresponding gene i is a
sequence of length k in {A, T, G, C}k . Each position in the sequence evolves independently, starting
from a uniform state in {A, T, G, C} at the root. Moving away from the root, a substitution occurs
(i)
on edge e with probability pe , in which case the state changes to state chosen uniformly among
the remaining states. After repeating this process for all positions, one obtains a sequence of length
k for each leaf of G(i) , for each i ∈ [m]—that is our input. A full algorithmic description of the
Jukes-Cantor process is provided as Algorithm 3 in Section A.
For gene i, we will let {ξxij : j ∈ [k], x ∈ L} denote the data generated at the leaves L of
the tree G(i) per the Jukes-Cantor process, the superscript j runs across the positions of the gene
sequence. To simplify the notation, we denote ξ ij = (ξxij )x∈L . The species phylogeny estimation
problem can then be stated as follows:
4
The n × m × k data array {ξ ij }i∈[m],j∈[k] is generated according to the Jukes-Cantor
process on the m gene trees, each of which in turn is generated according to the multispecies coalescent on S. The goal is to recover the topology of the species phylogeny
S from {ξ ij }i∈[m],j∈[k] .
We abbreviate this two-step data generation process by saying that {ξ ij }i∈[m],j∈[k] is generated
according to the MSC-JC(m, k) process on S.
2.2
Main result
ij
Require: Sequence output by Algorithm 2 {ξx,N
: x ∈ X = {1, 2, 3} , i ∈ MQ , j ∈ [k]}. A
partition of the set of genes [m] = MR t MQ , where MR = MR1 t MR2 and MQ =
MQ1 t MQ2 satisfy the conditions of Proposition 3.
P
ij
ij
i
1: For each x, y ∈ X and i ∈ MQ , let qbxy
= kj=1 1{ξx,N
6= ξy,N
}.
−1
√
−0.5
log k , and partition MQ = MQ1 t MQ2 such that
2: Set α , max m log m, k
|MQ1 | , |MQ2 | satisfy the conditions in Proposition 3.
i
: x, y ∈ X , i ∈ MQ1
Ultrametric Quantile Test on qbxy
3:
(c α)
For each pair of leaves x, y ∈ X , compute qbxy3 , the c3 α-th quantile with respect to the data
i
: i ∈ MQ1 . The constant c3 is as in Proposition 7. Define
qbxy
qb∗ ,
4:
max
x,y∈{1,2,3}
(c3 α)
qbxy
.
Next, for x, y ∈ X , define a similarity measure
sbxy ,
1
i
i ∈ MQ2 : qbxy
≤ qb∗
|MQ2 |
.
Return Declare that the topology is xy|z if sbxy > max {b
sxz , sbyz } .
Algorithm 1: Quantile-based triplet test
We now state our main result for the species phylogeny estimation problem. For any 3 leaves
x, y, z ∈ L, the species phylogeny S restricted to these three leaves has one of three possible rooted
topologies: xy|z, xz|y, or yz|x. For instance, 12|3 is depicted in Figure 1 (a) and indicates that
1 and 2 are closest. It is a classical phylogenetic result that if one is able to correctly reconstruct
the topology of all triples of leaves in L, then the topology of the full species phylogeny can
be correctly reconstructed as well (see e.g., [Ste16]). Therefore, to simplify the presentation, in
what follows our algorithms and theoretical guarantees are stated for a fixed triple X = {1, 2, 3}
(without loss of generality) among the set of leaves L.
Our main contribution is a novel polynomial-time reconstruction algorithm for the species phylogeny estimation problem, along with a rigorous data requirement which is optimal (up to log factors) by the work of [MR15]. Moreover, unlike [MR15], our results hold when mutation rates and
5
Require: Sequences {ξxij : x ∈ X = {1, 2, 3} , i ∈ [m], j ∈ [k]}. A partition of the set of genes
[m] = MR t MQ , where MR = MR1 t MR2 and MQ = MQ1 t MQ2 satisfy the conditions
of Proposition 3.
P
Pk/2
2
ij
1: For each x, y ∈ X and i ∈ MR , define p
bixy = k1 kj=1 1{ξxij 6= ξyij }, pbi↓
xy = k
j=1 1{ξx 6=
P
k
2
ij
ij
ξyij }, and pbi↑
xy = k
j=k/2+1 1{ξx 6= ξy }.
2: Let {xi , yi }, i = 1, 2, 3, be the three distinct (unordered) pairs of distinct leaves in X .
3: for i = 1,2 do
4:
5:
6:
7:
8:
9:
10:
11:
12:
Fixing gene tree topologies
Let x = xi and y = yi .
Let z be the unique element in X − {x, y}.
(1/3)
(2/3)
(5/6)
(2/3)
(5/6)
Compute the empirical quantiles pbxy , pbxz , pbxz , pbyz , and pbyz from the loci in MR1 .
(1/3)
For instance, to compute pbxy , sort the set pbixy : i ∈ MR1 in ascending order and pick the
j
k
|MR1 |
-th element, breaking ties arbitrarily.
3
o
n
(5/6)
(5/6) (2/3)
(1/3) (2/3)
i↓
i↓
.
≤
p
b
≤
p
b
,
p
b
≤
p
b
≤
p
b
,
p
b
≤
p
b
Set I := i ∈ MR2 : pbi↓
yz
xz
yz
xy
xz
yz
xz
xy
Estimation of differences ∆xy
P
1
bIyz .
bi↑
Set pbIxz := |I|
xz , and similarly for p
i∈I p
4 I
b xy := −∆
b yx := − 3 log 1− 43 pbIyz
Set ∆
4
1− 3 pbxz
end for
Let z3 be the unique element in X − {x3 , y3 }.
b x3 y3 := ∆
b x3 z 3 − ∆
b y3 z3 .
Set ∆
Stochastic Farris transform
b zx , ∆
b zy } ≥ 0.
13: Find a permutation {x, y, z} of X such that min{∆
ij
ij
14: For each gene i ∈ MQ and j ∈ [k], set ξz,N = ξzij . Also set ξx,N = ξxij with probability
b zx ) and otherwise choose ξ ij uniformly from {A, T, G, C} \ ξ ij . Do the same to ξ ij
1 − p(∆
y
x
x,N
ij
b
b
(with ∆yz instead of ∆xz ) to obtain ξy,N .
ij
: i ∈ MQ , j ∈ [k], x ∈ X
Return “noisy” sequence data ξx,N
Algorithm 2: Reduction step
populations sizes are allowed to vary across the species phylogeny. Our reconstruction algorithm
comprises two steps, which are detailed as Algorithm 1 and Algorithm 2. Our data requirement
applies to an unknown species phylogeny in the following class. We assume that: mutation rates
are in the interval (µL , µU ); leaf-edge lengths are in (f 0 , g 0 ); and internal-edge lengths are in (f, g).
We suppress the dependence on µL , µU , f 0 , g 0 , g, which we think of as constants, and focus here
on the role of f . The latter indeed plays a critical role in both the random processes described
above. Short internal branches are known to be hard to reconstruct from sequence data even when
6
dealing with a single gene tree [SS02] and a smaller f also leads to more discordance between
gene trees [RY03b]. We also suppress the dependence on the number of leaves n = |L|, which we
also consider here to be a constant (see the concluding remarks in Section 4 for more on this).
We state here a simplified version of our results (the more general statement appearing as
Proposition 3 in Section C). Specifically, we answer the following question: as f → 0, how many
genes m of length k are needed for a correct reconstruction with high probability? For technical
reasons, our results apply only when k grows at least polynomially with f (with an arbitrarily
small exponent). Throughout, we use the notation & (similarly, .) to indicate that constants and
poly(log f −1 ) factors are suppressed in a lower bound. Recall that x ∨ y = max{x, y}.
Theorem 1 (Data requirement). Suppose that we have sequence data {ξ ij }i∈[m],j∈[k] generated according to the MSC-JC(m, k) process on a species phylogeny S = (Vs , Es ; r, ~τ , ~µ) . The mutation
rates, leaf-edge lengths and internal-edge lengths are respectively in (µL , µU ), (f 0 , g 0 ) and (f, g).
We assume further that there is C > 0 such that k & f −C . Then Algorithm 1 correctly identifies
the topology of S restricted to X = {1, 2, 3} with probability at least 95% provided that
m&
1
1
∨√
.
f
kf 2
(1)
Two regimes are implicit in Theorem 1:
• “Long” sequences: When k & f −2 , we require m & f −1 . As first observed by [MR10],
this condition is always required for high-probability reconstruction under this setting.
• “Short” sequences: When k . f −2 , we require the stronger condition that m & k −1/2 f −2 .
This is known to be optimal (up to the log factor) by the information-theoretic lower bound
in [MR15]. As mentioned above, the matching algorithmic upper bound of [MR15] only
applies when all mutation rates and population sizes are identical. Our main contribution
here is to relax this assumption.
On the other hand, our results do not apply to the regime of “very short” sequences of constant
length. In that regime, the reconstruction algorithm of [DNR15], which applies under the same
setting we are considering here, achieves the optimal bound of m & f −2 .
2.3
Proof idea and further results
We give a brief overview of the proof. The full details are given in Section 3 as well as Sections C, D and E. Again, fix a triple of leaves X = {1, 2, 3}.
Tree metrics Phylogenies are naturally equipped with a notion of distance between leaves, and
in general any pair of vertices, which is known as a tree metric (see e.g. [Ste16] for more details).
Our species phylogeny reconstruction method rests on such tree metrics.
7
Definition 3 (Weighted species metric). A species phylogeny S = (Vs , Es ; r, ~τ , ~µ) induces the
following metric on the leaf set L. For any pair of leaves a, b ∈ L, we let
X
µab =
τe µe ,
e∈π(a,b;S)
where π(a, b; S) is the unique path connecting a and b in S interpreted as a set of edges. We will
refer to {µab }a,b∈L as the weighted species metric induced by S.
The above definition is valid for any pair of vertices in Vs . That is, the metric µ can be extended
to the entire set Vs . In the species phylogeny estimation problem, the sequence data only carries
information about the rate-weighted distances {µab }a,b∈L . The algorithm in [MR15] is guaranteed
to recover the topology of S only in the case that {µab }a,b∈L is an ultrametric on the leaf set L, in
which case we refer to S as an ultrametric species phylogeny. The metric {µab }a,b∈L is ultrametric
when µra = µrb for all a, b ∈ L, that is, when the distance from the root to every leaf is the same.
Recall from Definition 2 that each each random gene tree has an associated set of branch
lengths. From the description of the multispecies coalescent (see Section A), it follows that a
single branch of a gene tree may span across multiple branches of the species phylogeny; this can
also be seen in Fig. 2. Let tẽ denote the (random) length of the branch ẽ ∈ E (i) . For any species
phylogeny branch e ∈ Es , let tẽ∩e denote the length of the branch ẽ that overlaps with e. Then, δẽ
and tẽ satisfy the following relationship
X
µe tẽ∩e .
δẽ =
e∈Es
This set of weights again defines a different metric on the leaves L of the species tree.
Definition 4 (Gene metric). A gene tree G(i) = (V (i) , E (i) ; r, ~δ(i) ) induces the following metric on
the leaf set L. For any pair of leaves a, b ∈ L, we (overload the notation δ) and let
X
(i)
δab =
δe(i)
e∈π(a,b;G(i) )
(i)
where, again, π(a, b; Gn(i) ) is
o the unique path connecting a and b in G interpreted as a set of
(i)
edges. We will refer to δab
as the gene metric induced by G(i) .
a,b∈L
Note that, when the species phylogeny S is ultrametric, so are the gene trees.
Ultrametric reduction At a high level, our reconstruction algorithm relies on a quantile triplet
test developed in [MR15]. Roughly speaking this test, which is detailed in n
Algorithm
1, compares
o
(i)
a well-chosen quantile of the sequence-based estimates of gene metrics δab
in order to
a,b∈X
determine which pair of leaves is closest. The algorithm of [MR15], however, only works when all
mutation rates and population sizes are equal. In that case, the species phylogeny and gene trees
8
are ultrametric, as defined above. That property leads to symmetries that play a crucial role in the
algorithm. Our first main contribution here is a reduction to the this ultrametric case.
That is, in order to apply the quantile triplet test, we first transform the sequence data to appear
as though it was was generated by an ultrametric species phylogeny. This ultrametric reduction,
inspired by a classical technique known as the Farris transform (see e.g. [SS03b]), may be of
independent interest as it could be used to generalize other reconstruction algorithms. Formally,
we prove the following theorem. Again, we state a simplified version of our result which gives a
lower bound on the number of genes m of length k needed to achieve a desired accuracy (the more
general statement appearing as Proposition 4 in Section D). More specifically, Algorithm 2 takes
as input two sets of genes, MR and MQ . The set MR is used to estimate parameters needed for
the reduction. The reduction is subsequently performed on MQ . We let m0 = |MR |. Here we give
a lower bound on m0 (while, for the purposes of this theorem, |MQ | can be arbitrarily large). For
θ > 0, we say that two metrics µ0 and µ00 over X are θ-close if µ0xy − µ00xy ≤ θ, for all x, y ∈ X .
Theorem 2 (Ultrametric reduction). Suppose that we have sequence data {ξ ij }i∈[m],j∈[k] generated
according to the MSC-JC(m, k) process on a three-species phylogeny S = (Vs , Es ; r, ~τ , ~µ). The
mutation rates, leaf-edge lengths and internal-edge lengths are respectively in (µL , µU ), (f 0 , g 0 )
and (f, g). We assume further that there is C > 0 such that k & f −C . Then, with probability at
least 95%, the output of Algorithm 2 is distributed according to the MSC-JC process on a species
tree S 0 that is φ-close to an ultrametric species phylogeny with rooted topology identical to that of
S restricted to X = {1, 2, 3}, provided that
m0 & 1 ∨
1
,
kf 2
(2)
where φ = Θ(f / log f −1 ).
The log factor in φ is needed in our analysis of the quantile test below. The key to the proof of
Theorem 2 is the establishment of a new identifiability result of independent interest.
Theorem 3 (Identifiability of rooted species tree from unrooted weighted gene trees). Let S =
(Vs , Es ; r, ~τ , ~µ) be a species tree with n ≥ 3 leaves and root r and let G = (V, E; r, ~δ) be a
sampled gene tree from the MSC with branch lengths δe , e ∈ E. Then the rooted topology of the
species tree S is identifiable from the distribution of the unrooted weighted gene tree G.
The case n ≥ 5 is not new. Indeed, it follows from [ADR11b, Theorem 9], where it is shown that in
fact the distribution of the unrooted gene tree topologies (without any branch length information)
suffices to identify the rooted species phylogeny when the number of leaves exceeds 4. On the other
hand, it was also shown in [ADR11b, Proposition 3] that, when n = 4, the gene tree topologies
are not enough to locate the root of the species phylogeny (and the case n = 3 is trivial). Here we
show that, already with three species (and therefore also when n > 3), the extra information in the
gene tree branch lengths allows to recover the root. We give a constructive proof of Theorem 3,
which we then adapt to obtain Algorithm 2. More details on this key step are given in Section 3.
9
ij
Robustness of quantile test Algorithm 2 produces a new sequence dataset ξx,N
: x ∈ X that
appears close to being distributed according to an ultrametric species phylogeny. The next step is
to perform a triplet test of [MR15], detailed in Algorithm 1. Roughly speaking, this test is based
on comparing an appropriately chosen quantile of the gene metrics. In fact, because we do not
have direct access to the latter, we use a sequence-based surrogate, the empirical p-distances
k
i
qbxy
1 X ij
ij
1 ξx,N 6= ξy,N
,
=
k j=1
for each gene i ∈ MQ in the output of the reduction, whose expectation is a monotone transformation of the corresponding gene metrics. The idea of Algorithm 1 is to use the above p-distances
to define a “similarity measure” sbxy between each pair of leaves x, y ∈ X to reveal the underlying
species tree topology on X . It works as follows. The set of genes MQ is divided into two disjoint
(c α)
i
subsets MQ1 , MQ2 . The set MQ1 is used to compute the c3 α-quantile qbxy3 of {b
qxy
: i ∈ MQ1 },
where c3 > 0 is a constant determined in the proofs and
)
(
r
log m
log k
,
.
α = max
m
k
n
o
(c α)
Let qb∗ denote the maximum among qbxy3 : x, y ∈ X . We then use the genes in MQ2 to define
the similarity measure
1
i
sbxy =
i ∈ MQ2 : qbxy
≤ qb∗ .
|MQ2 |
Whichever pair x, y ∈ X produces the largest value of sbxy is declared the closest, i.e., the output
is xy|z where z is the remaining leaf in X .
Why does it work? Intuitively, the closest pair of species x, y will tend to produce a larger
number of genes with few differences between their sequences at x and y, as measured by the
p-distance. In fact it was shown in [MR10] that, under the MSC-JC process on an ultrametric
phylogeny when sequences are long enough (namely k & f −2 ), choosing the pair of species
achieving the smallest p-distance across genes succeeds with high probability under optimal data
requirements. When k is short on the other hand (namely k . f −2 ), the randomness from the JC
process produces outliers that confound this approach. To make the test more robust, it is natural
to turn to quantiles, i.e., to remove a small, fixed √
fraction of outliers. On a fixed gene tree, the
k. It was shown in [MR15] that, as a result,
standard
deviation
of
the
p-distance
is
of
order
1/
√
1/ k is in a sense the smallest quantile that can be meaningfully controlled and that it leads to a
successful test under optimal data requirements. Our choice of quantile α is meant to cover both
regimes above simultaneously. See [MR15], as well as [MR10, DNR15], for more details.
As stated in Theorem 2, the output to the ultrametric reduction is almost—but not perfectly—
ultrametric. In our second main contribution, to account for this extra error, we perform a delicate
robustness analysis of the quantile-based triplet test. This step is detailed in Section E. At a high
level, the proof follows [MR15]. After 1) controlling the deviation of the quantiles, we establish
that 2) the test works in expectation and then 3) finish off with concentration inequalities. All
10
these steps must be updated to account for the error introduced in the reduction step. Step 2) is
particularly involved and requires the delicate analysis of the CDF of a mixture of binomials.
3
Key ideas in the ultrametric reduction
The goal of the ultrametric reduction step, Algorithm 2, is to transform the sequence data to appear
statistically as though it is the output of an MSC-JC process on an ultrametric species phylogeny
with the same topology as S restricted to X .
3.1
Preliminary step: a new identifiability result
Before diving into the description of Algorithm 2, we provide some insights into the algebra of our
reduction by first deriving a new identifiability result, Theorem 3. That is, we show that, under the
multispecies coalescent, the rooted topology of the species phylogeny can be recovered from the
distribution of the unrooted weighted gene trees.
Our reduction is inspired by the Farris transform (also related to the Gromov product; see
e.g. [SS03b]), a classical technique to transform a general metric into an ultrametric. In a typical
application of the Farris transform, one “roots” the species phylogeny S at an “outgroup” o (i.e., a
species that is “far away” from the leaves of S) and then uses the quantities µox , x ∈ L to implicitly
stretch the leaf edges appropriately, so as to make all inter-species distances to o equal, without
changing the underlying topology. More specifically, let S be a species phylogeny. Suppose X =
{1, 2, 3} and let o ∈ L − X be any leaf of S outside X . Assume that µo1 ≥ max{µo2 , µo3 } (the
other cases being similar) and define the Farris transform
µ̇xy , µxy + 2µo1 − µox − µoy ,
∀x, y ∈ X .
(3)
A classical phylogenetic result (proved for instance in [SS03a, Lemma 7.2.2]) states that {µ̇xy }x,y∈X
is an ultrametric on X consistent with the topology of S re-rooted at o and, then, restricted to X .
In the multi-gene context, however, we cannot apply a Farris transform in this manner. For one,
we do not have direct access to the species phylogeny distances {µxy }; rather, we only estimate
(i)
the gene tree distances {δxy }. Moreover the latter vary across genes according to the MSC. In
particular, distance differences (such as those appearing in (3)) are affected by the topology of the
gene tree (see Figure 3 for an illustration).
Key idea 1: To get around this problem, we artificially fix gene tree topologies
through conditioning. We also take advantage of the effect of the rooting on the
MSC process to avoid using an outgroup.
We give more details on our approach next.
We turn to the proof of Theorem 3. We prove the claim for n = 3. As we discussed, it is
straightforward to extend the proof to n > 4. Let S be a species phylogeny with three leaves
and recall that r is the root of S. Unlike the classical Farris transform above, we do not use an
outgroup. Instead, we show how to achieve the same outcome by using only the distribution of
11
G and, in particular, of the random distances {δe }e∈Es . Notice from (3) that we only need the
differences of distances between pairs of species in X ∪ {r}
∆xy , µrx − µry .
It is these quantities that we derive from the distribution of weighted gene trees.
The idea is to:
1. Condition on an event such that the rooted topology of a gene tree is guaranteed to be equal
to a fixed, chosen topology. Intuitively, we achieve this by considering an event where one
pair of leaves is “somewhat close” while the other two pairs are “somewhat far.”
2. Conditioning on this event, we recover the species-based difference ∆xy = µrx − µry from
the distribution of gene-based difference δxz − δyz . Intuitively, letting w be the most recent
common ancestor of x and y on G, when the topology is xy|z then the difference δxz − δyz
is equal to ∆xy irrespective of when w occurred. See Figure 3 for an illustration.
Gene 1
Gene 2
2
2
3
13
23
=
3
12
13
1
23
6=
12
1
(b)
(a)
Figure 3: (a) Gene 1 (red gene) has the topology 12|3. Therefore, the gene distance on this gene
satisfies the condition that δ13 − δ23 = ∆12 . (b) In this case, Gene 2 (blue gene) has the topology
1|23. Observe that therefore, δ13 − δ23 6= ∆12 .
More formally, we establish the following two propositions, whose proofs are in Section B. For
(β)
(β)
x, y ∈ L and β ∈ [0, 1], let δxy be the β-th quantile of δxy . That is, δxy is the smallest number
α ∈ [0, 1] such that
P [δxy ≤ α] ≥ β.
Note that this quantile is a function of the distribution of G (and of the ~δs). Our event of interest is
defined next.
12
Proposition 1 (Fixing the rooted topology of the gene tree). Let (x, y, z) be an arbitrary permutation of (1, 2, 3). The event
(1/2)
(1/2)
(1/2)
EI = δxy ≤ δxy
, δxz > δxz
, δyz > δyz
,
(4)
has positive probability and implies that the rooted topology of G is xy|z.
Conditioning on the event EI , we then show how to recover the difference ∆xy from the distribution
of δxz − δyz .
Proposition 2 (A formula for the height difference). Using the notation of Proposition 1, we have
E[δxz − δyz | EI ] = ∆xy ,
(5)
almost surely.
Note that the quantity on the l.h.s. of (5) is a function of the distribution of G. From the values
of ∆xy , x, y ∈ X , we can solve for µrx , x ∈ X . Hence, combining the properties of the Farris
transform with Propositions 1 and 2, we arrive at Theorem 3.
3.2
Algorithm 2: the reduction step
We are now ready to describe the reduction algorithm (Algorithm 2) and provide guarantees about
its behavior. Recall that we are restricting our attention to three leaves X = {1, 2, 3} whose species
tree topology is 12|3. The main idea underlying the reduction algorithm is based on the proof of
the identifiability result (Theorem 3). That is, we find a set of genes whose topology is highly
likely to be a fixed triplet, we estimate the height differences on this set using the “sample version”
of (5), and we perform what could be thought of as a “sequence-based” Farris transform.
Given that we do not have access to the actual gene tree distribution, but only sequence data,
there are several differences with the identifiability proof that make the analysis and the algorithm
more involved. A primary challenge is that, in the regime where sequence length is “short,” i.e.,
when k f −2 , the sequence-based estimates of the gene tree distances are very inaccurate—much
less accurate then what is needed for our reduction step to be useful.
Key idea 2: To get around this issue, we show how to combine genes satisfying a
condition similar to (4) to produce a much better estimate of distance differences.
We detail the main steps of Algorithm 2 next.
Fixing gene tree topologies. Here we only have access to sequence data. In particular the δs are
unknown. So, we work instead with the p-distances
pbixy =
1 X ij
1 ξx 6= ξyij ,
k
j∈[k]
13
(β)
for gene i and x, y ∈ X , and their empirical quantiles pbxy .1 Similar to Proposition 1, we then
consider those genes for which the event
i
(5/6) (2/3)
(5/6)
pbxy ≤ pb(1/3)
b(2/3)
≤ pbixz ≤ pbxz
, pbyz ≤ pbiyz ≤ pbyz
,
(6)
xy , p
xz
holds for some chosen permutation (x, y, z) of (1, 2, 3). We will call this set of genes I. We show
that this set has a “non-trivial” size and that, with high probability, the genes satisfying (6) have
topology xy|z (see Proposition 5).2 In particular, the analysis of this construction accounts for the
“sequence noise” around the expected values
3
i
i /3
−4δxy
i
i
(i)
1−e
, p(δxy
),
(7)
=
pxy , E pbxy G
4
where p(x) = 34 1 − e−4x/3 .
Estimating distance differences. Because we work with p-distances, we adapt formula (5) for
the difference ∆xy as follows3 . Using
1 X i
1 X i
pbIxz =
pbxz
and
pbIyz =
pb ,
|I| i∈I
|I| i∈I yz
our estimate of the distance differences is given by
3
4 I
4 I
3
b
∆xy = − log 1 − pbxz
− − log 1 − pbyz
.
4
3
4
3
Recall that, for this formula to work, we need to ensure that the topologies of the gene trees used are
fixed to be xy|z; see Fig. 3, for instance. The logarithmic transforms in the curly brackets are the
usual distance corrections in the Jukes-Cantor sequence model (see e.g. [Ste16]). Note, however,
that we perform an average over I before the correction; this is important to obtain the correct
statistical power of our estimator. A similar phenomenon was leveraged in the METAL algorithm
of [DNR15]. The non-trivial part of the analysis of this step is to bound the estimation error.
Indeed, unlike the identifiability result, we have a finite amount of gene data and, moreover, we
must account for the sequence noise. This is done using concentration inequalities in Proposition 6.
Stochastic Farris transform. The quantile test of Section E below is not a distance-based method in the traditional sense of the term. That is, we do not define a pairwise distance matrix on the
leaves and use it to deduce the species phylogeny. Instead, our method uses the empirical distribution of the p-distances across genes. It is for this reason that we do not simply apply the classical
Farris transform of (3) to the estimated distances. Rather, we perform what we call a “stochastic”
Farris transform. That is, we transform the sequence data itself to mimic the distribution under an
ultrametric species phylogeny.
1
Actually, the quantiles are estimated from part of the gene set (MR1 ) to avoid unwanted correlations. The rest of
the analysis is done on the other part.
2
In fact, the p-distances in (6) are estimated over half the gene length to avoid unwanted correlations. That is, we
use pbi↓
xy to compute I (see Step 5 of Algorithm 2).
3
Again, here we use the other half of the sites to avoid correlations with Step 5.
14
Key idea 3: This is done by adding the right amount of noise to the sequence data
at each gene, as detailed next. It ensures that we properly mimic the contributions from both the multispecies coalescent and the Jukes-Cantor model to the
distribution of p-distances.
See Algorithm 2 for the full details.
For the sake of notational convenience, we will let ⊕ denote addition mod-4 and identify
A, T, G, C with {0, 1, 2, 3} in that order when doing this addition. For instance, this means that
A ⊕ 1 = T and G ⊕ 2 = A.
Definition 5 (Stochastic Farris transform). For a gene i, let {ξxi }x∈X be a sequence dataset over the
species X = {1, 2, 3} and let ∆xy = µrx − µry , x, y ∈ X . Assume without loss of generality that
i
min{∆12 , ∆13 } ≥ 04 . The stochastic Farris transform defines a new set of sequences {ξx,N
}x∈X
k
i
i
i
i
such that ξx,N = ξx ⊕ x , where x ∈ {0, 1, 2, 3} is an independent random sequence whose j-th
coordinate is drawn according to
0, w.p. 1 − p(∆1x ),
1, w.p. p(∆ )/3,
1x
ij
x =
2, w.p. p(∆1x )/3,
3, w.p. p(∆1x )/3.
i
}x∈X = F({ξxi }x∈X ; {∆xy }x,y∈X ).
We write this as {ξx,N
By the Markov property, for x, y ∈ X , the “noisy” sequence data above satisfy
i
i
i
i
P ξx,N
6= ξy,N
= p δxy
+ ∆1x + ∆1y , rxy
.
i
, the random gene tree distance between x and y under gene i, can be decomposed
Notice that δxy
i
as µxy + Γxy , where Γixy is the random component contributed by the multispecies coalescent. On
the other hand, the set of distances µxy + ∆1x + ∆1y is ultrametric by the properties of the classical
Farris transform. As a result, the stochastic Farris transform modifies the sequence data so that it
appears as though it was generated from an ultrametric MSC-JC process. We show this pictorially
in Fig. 4.
In reality, we do not have access to the true differences ∆xy , x, y ∈ X . Instead, we employ our
b xy for all x, y ∈ X in the previous step to obtain the following approximate stochastic
estimates ∆
Farris transform:
i
b xy }x,y∈X ).
(8)
{ξx,N
}x∈X = F({ξxi }x∈X ; {∆
This is the output of the reduction. See Algorithm 2 for details. We prove Theorem 2 in Section D.
In what follows, we will condition on the implications of Theorem 2 holding.
4
This is equivalent to assuming that µr1 ≥ max{µr2 , µr3 }.
15
Gene 2
Gene 1
2
3
12
13
1
Figure 4: After the stochastic Farris transform, the leaf edges corresponding to leaves 2 and 3, in
both Gene 1 (red gene) and Gene 2 (blue gene), are “stretched” by ∆12 and ∆13 respectively. As a
result, the sequence data appears as though it is drawn from an ultrametric species phylogeny.
4
Concluding remarks
We have extended the optimal tradeoff (up to log factors) of [MR15] beyond the case of equal
mutation rates and population sizes. Several open problems remain:
1. Our results assume that the number of leaves n is constant (as k, m → ∞). As n gets
larger, the depth of the species phylogeny typically increases. In fact, in the single gene tree
reconstruction context, the depth is known to play a critical and intricate role in the data
requirement [ESSW99a, Mos04, Mos07, DMR11a, DMR11b]. Understanding the role of
the depth under the MSC-JC is an interesting avenue for future work.
2. We have assumed here that the mutation rates are the same across genes. This assumption
is not realistic and relaxing it is important for the practical relevance of this line of work.
Identifiability issues may arise however [MS07, Ste09]. In a related issue, we have assumed,
to simplify, that all genes have the same length. (Gene lengths and mutation rates together
control the amount of phylogenetic signal in a gene.) We leave for future work how best to
take advantage of differing gene lengths (beyond simply truncating to the shortest gene).
3. A more technical point left open here is to remove the assumption that k grows polynomially
with f . This may require new ideas.
16
References
[ABF+ 99]
Richa Agarwala, Vineet Bafna, Martin Farach, Mike Paterson, and Mikkel Thorup.
On the approximability of numerical taxonomy (fitting distances by tree metrics).
SIAM J. Comput., 28(3):1073–1085 (electronic), 1999.
[ADHR12] Alexandr Andoni, Constantinos Daskalakis, Avinatan Hassidim, and Sebastien Roch.
Global alignment of molecular sequences via ancestral state reconstruction. Stochastic Processes and their Applications, 122(12):3852 – 3874, 2012.
[ADR11a]
Elizabeth S. Allman, James H. Degnan, and John A. Rhodes. Determining species
tree topologies from clade probabilities under the coalescent. Journal of Theoretical
Biology, 289:96 – 106, 2011.
[ADR11b]
Elizabeth S. Allman, James H. Degnan, and John A. Rhodes. Identifying the rooted
species tree from the distribution of unrooted gene trees under the coalescent. Journal
of Mathematical Biology, 62(6):833–862, 2011.
[ADR17]
E. Allman, J. Degnan, and J. Rhodes. Species tree inference from gene splits by unrooted star methods. IEEE/ACM Transactions on Computational Biology and Bioinformatics, PP(99):1–1, 2017.
[ALPE12]
Christian N.K. Anderson, Liang Liu, Dennis Pearl, and Scott V. Edwards. Tangled
trees: The challenge of inferring species trees from coalescent and noncoalescent
genes. In Maria Anisimova, editor, Evolutionary Genomics, volume 856 of Methods
in Molecular Biology, pages 3–28. Humana Press, 2012.
[Att99]
K. Atteson. The performance of neighbor-joining methods of phylogenetic reconstruction. Algorithmica, 25(2-3):251–278, 1999.
[BCMR06] Christian Borgs, Jennifer T. Chayes, Elchanan Mossel, and Sébastien Roch. The
Kesten-Stigum reconstruction bound is tight for roughly symmetric binary channels.
In FOCS, pages 518–530, 2006.
[BLM13]
S. Boucheron, G. Lugosi, and P. Massart. Concentration Inequalities: A Nonasymptotic Theory of Independence. OUP Oxford, 2013.
[BS14]
Anand Bhaskar and Yun S. Song. Descartes’ rule of signs and the identifiability of population demographic models from genomic variation data. Ann. Statist.,
42(6):2469–2493, 2014.
[CGG02]
M. Cryan, L. A. Goldberg, and P. W. Goldberg. Evolutionary trees can be learned in
polynomial time. SIAM J. Comput., 31(2):375–397, 2002. short version, Proceedings
of the 39th Annual Symposium on Foundations of Computer Science (FOCS 98),
pages 436-445, 1998.
17
[CK15]
Julia Chifman and Laura Kubatko. Identifiability of the unrooted species tree topology under the coalescent model with time-reversible substitution processes, sitespecific rate variation, and invariable sites. Journal of Theoretical Biology, 374:35 –
47, 2015.
[CT06]
Benny Chor and Tamir Tuller. Finding a maximum likelihood tree is hard. J. ACM,
53(5):722–744, 2006.
[DBP05]
Frederic Delsuc, Henner Brinkmann, and Herve Philippe. Phylogenomics and the
reconstruction of the tree of life. Nat Rev Genet, 6(5):361–375, 05 2005.
[DD10]
Michael DeGiorgio and James H Degnan. Fast and consistent estimation of species
trees using supermatrix rooted triples. Molecular Biology and Evolution, 27(3):552–
69, March 2010.
[DD14]
Michael DeGiorgio and James H. Degnan. Robustness to divergence time underestimation when inferring species trees from estimated gene trees. Systematic Biology,
63(1):66, 2014.
[DDBR09] James H. Degnan, Michael DeGiorgio, David Bryant, and Noah A. Rosenberg. Properties of consensus methods for inferring species trees from gene trees. Systematic
Biology, 58(1):35–54, 2009.
[DMR11a] Constantinos Daskalakis, Elchanan Mossel, and Sébastien Roch. Evolutionary trees
and the ising model on the bethe lattice: a proof of steel’s conjecture. Probability
Theory and Related Fields, 149:149–189, 2011. 10.1007/s00440-009-0246-2.
[DMR11b] Constantinos Daskalakis, Elchanan Mossel, and Sébastien Roch. Phylogenies without branch bounds: Contracting the short, pruning the deep. SIAM J. Discrete Math.,
25(2):872–893, 2011.
[DNR14]
Gautam Dasarathy, Robert D. Nowak, and Sébastien Roch. New sample complexity
bounds for phylogenetic inference from multiple loci. In 2014 IEEE International
Symposium on Information Theory, Honolulu, HI, USA, June 29 - July 4, 2014, pages
2037–2041, 2014.
[DNR15]
G. Dasarathy, R. Nowak, and S. Roch. Data requirement for phylogenetic inference
from multiple loci: A new distance method. Computational Biology and Bioinformatics, IEEE/ACM Transactions on, 12(2):422–432, March 2015.
[DR06]
J. H. Degnan and N. A. Rosenberg. Discordance of species trees with their most
likely gene trees. PLoS Genetics, 2(5), May 2006.
[DR09]
James H. Degnan and Noah A. Rosenberg. Gene tree discordance, phylogenetic inference and the multispecies coalescent. Trends in Ecology and Evolution, 24(6):332
– 340, 2009.
18
[DR13]
Constantinos Daskalakis and Sebastien Roch. Alignment-free phylogenetic reconstruction: sample complexity via a branching process analysis. Ann. Appl. Probab.,
23(2):693–721, 2013.
[Dur96]
Richard Durrett. Probability: theory and examples. Duxbury Press, Belmont, CA,
second edition, 1996.
[ESSW99a] P. L. Erdös, M. A. Steel, L. A. Székely, and T. A. Warnow. A few logs suffice to build
(almost) all trees (part 1). Random Struct. Algor., 14(2):153–184, 1999.
[ESSW99b] P. L. Erdös, M. A. Steel, L. A. Székely, and T. A. Warnow. A few logs suffice to build
(almost) all trees (part 2). Theor. Comput. Sci., 221:77–118, 1999.
[Fel04]
J. Felsenstein. Inferring Phylogenies. Sinauer, Sunderland, MA, 2004.
[FK99]
Martin Farach and Sampath Kannan. Efficient algorithms for inverting evolution. J.
ACM, 46(4):437–449, 1999.
[GF82]
R. L. Graham. and L. R. Foulds. Unlikelihood that minimal phylogenies for a realistic
biological study can be constructed in reasonable computational time. Math. Biosci.,
60:133–142, 1982.
[GMS12]
Ilan Gronau, Shlomo Moran, and Sagi Snir. Fast and reliable reconstruction of phylogenetic trees with indistinguishable edges. Random Struct. Algorithms, 40(3):350–
384, 2012.
[Hoe63]
Wassily Hoeffding. Probability inequalities for sums of bounded random variables.
Journal of the American Statistical Association, 58(301):13–30, 1963.
[HRS10]
D.H. Huson, R. Rupp, and C. Scornavacca. Phylogenetic Networks: Concepts, Algorithms and Applications. Cambridge University Press, 2010.
[KD07]
L. S. Kubatko and J. H. Degnan. Inconsistency of phylogenetic estimates from concatenated data under coalescence. Systematic Biology, 56(1):17–24, February 2007.
[KMRR15] Junhyong Kim, Elchanan Mossel, Mikls Z. Rcz, and Nathan Ross. Can one hear the
shape of a population history? Theoretical Population Biology, 100(0):26 – 38, 2015.
[KZZ03]
Valerie King, Li Zhang, and Yunhong Zhou. On the complexity of distance-based
evolutionary tree reconstruction. In SODA ’03: Proceedings of the fourteenth annual
ACM-SIAM symposium on Discrete algorithms, pages 444–453, Philadelphia, PA,
USA, 2003. Society for Industrial and Applied Mathematics.
[LYK+ 09]
Liang Liu, Lili Yu, Laura Kubatko, Dennis K. Pearl, and Scott V. Edwards. Coalescent methods for estimating phylogenetic trees. Molecular Phylogenetics and
Evolution, 53(1):320 – 328, 2009.
19
[LYP10]
Liang Liu, Lili Yu, and DennisK. Pearl. Maximum tree: a consistent estimator of the
species tree. Journal of Mathematical Biology, 60(1):95–106, 2010.
[Mad97]
Wayne P. Maddison. Gene trees in species trees. Systematic Biology, 46(3):523–536,
1997.
[MBW16]
Siavash Mirarab, Md Shamsuzzoha Bayzid, and Tandy Warnow. Evaluating summary
methods for multilocus species tree estimation in the presence of incomplete lineage
sorting. Systematic Biology, 65(3):366, 2016.
[MFP08]
Simon Myers, Charles Fefferman, and Nick Patterson. Can one learn history from
the allelic spectrum? Theoretical Population Biology, 73(3):342 – 348, 2008.
[MHR13]
Radu Mihaescu, Cameron Hill, and Satish Rao. Fast phylogeny reconstruction
through learning of ancestral sequences. Algorithmica, 66(2):419–449, 2013.
[MLP09]
Radu Mihaescu, Dan Levy, and Lior Pachter. Why neighbor-joining works. Algorithmica, 54(1):1–24, May 2009.
[Mos03]
E. Mossel. On the impossibility of reconstructing ancestral data and phylogenies. J.
Comput. Biol., 10(5):669–678, 2003.
[Mos04]
E. Mossel. Phase transitions in phylogeny. Trans. Amer. Math. Soc., 356(6):2379–
2404, 2004.
[Mos07]
E. Mossel. Distorted metrics on trees and phylogenetic forests. IEEE/ACM Trans.
Comput. Bio. Bioinform., 4(1):108–116, 2007.
[MR95]
Rajeev Motwani and Prabhakar Raghavan. Randomized algorithms. Cambridge University Press, Cambridge, 1995.
[MR06]
Elchanan Mossel and Sébastien Roch. Learning nonsingular phylogenies and hidden
Markov models. Ann. Appl. Probab., 16(2):583–614, 2006.
[MR10]
Elchanan Mossel and Sébastien Roch. Incomplete lineage sorting: Consistent phylogeny estimation from multiple loci. IEEE/ACM Trans. Comput. Biology Bioinform.,
7(1):166–171, 2010.
[MR15]
Elchanan Mossel and Sébastien Roch. Distance-based species tree estimation:
Information-theoretic trade-off between number of loci and sequence length under
the coalescent. In Approximation, Randomization, and Combinatorial Optimization.
Algorithms and Techniques, APPROX/RANDOM 2015, August 24-26, 2015, Princeton, NJ, USA, pages 931–942, 2015.
[MS07]
Frederick A. Matsen and Mike Steel. Phylogenetic mixtures on a single tree can
mimic a tree of another topology. Systematic Biology, 56(5):767–775, 2007.
20
[Nak13]
Luay Nakhleh.
Computational approaches to species phylogeny inference and gene tree reconciliation.
Trends in ecology & evolution,
28(12):10.1016/j.tree.2013.09.004, 12 2013.
[Roc06]
Sébastien Roch. A short proof that phylogenetic tree reconstruction by maximum
likelihood is hard. IEEE/ACM Trans. Comput. Biology Bioinform., 3(1):92–94, 2006.
[Roc13]
Sébastien Roch. An analytical comparison of multilocus methods under the multispecies coalescent: The three-taxon case. In Biocomputing 2013: Proceedings of the
Pacific Symposium, Kohala Coast, Hawaii, USA, January 3-7, 2013, pages 297–306,
2013.
[Roo01]
Bero Roos. Binomial approximation to the poisson binomial distribution: The
krawtchouk expansion. Theory of Probability & Its Applications, 45(2):258–272,
2001.
[RS]
Sebastien Roch and Allan Sly. Phase transition in the sample complexity of
likelihood-based phylogeny inference. Submitted, arXiv:1508.01964.
[RS15]
Sebastien Roch and Mike Steel. Likelihood-based tree reconstruction on a concatenation of alignments can be positively misleading. Theoretical Population Biology,
2015. To appear.
[RW15]
Sebastien Roch and Tandy Warnow. On the robustness to gene tree estimation error
(or lack thereof) of coalescent-based species tree methods. Systematic Biology, 2015.
In press.
[RY03a]
Bruce Rannala and Ziheng Yang. Bayes estimation of species divergence times
and ancestral population sizes using DNA sequences from multiple loci. Genetics,
164(4):1645–1656, 2003.
[RY03b]
Bruce Rannala and Ziheng Yang. Bayes estimation of species divergence times
and ancestral population sizes using dna sequences from multiple loci. Genetics,
164(4):1645–1656, 2003.
[SRM]
Shubhanshu Shekhar, Sebastien Roch, and Siavash Mirarab. Species tree estimation
using ASTRAL: how many genes are enough? Preprint, 2017. To appear in the Proceedings of the 21st Annual International Conference on Research in Computational
Molecular Biology (RECOMB), 2017.
[SS02]
M. A. Steel and L. A. Székely. Inverting random functions. II. Explicit bounds for
discrete maximum likelihood estimation, with applications. SIAM J. Discrete Math.,
15(4):562–575 (electronic), 2002.
[SS03a]
C. Semple and M. Steel. Phylogenetics, volume 22 of Mathematics and its Applications series. Oxford University Press, 2003.
21
[SS03b]
Charles Semple and Mike A Steel. Phylogenetics, volume 24. Oxford University
Press, 2003.
[Ste09]
Mike Steel. A basic limitation on inferring phylogenies by pairwise sequence comparisons. Journal of Theoretical Biology, 256(3):467 – 472, 2009.
[Ste16]
Mike Steel. Phylogeny—discrete and random processes in evolution, volume 89 of
CBMS-NSF Regional Conference Series in Applied Mathematics. Society for Industrial and Applied Mathematics (SIAM), Philadelphia, PA, 2016.
[War]
Tandy Warnow. Computational phylogenetics: An introduction to designing methods
for phylogeny estimation. To be published by Cambridge University Press , 2017.
[Yan14]
Z. Yang. Molecular Evolution: A Statistical Approach. Oxford University Press,
2014.
22
A
Models: full definitions
In this section, we provide full definitions of the multispecies coalescent and Jukes-Cantor model.
Jukes-Cantor model of sequence evolution
detailed in Algorithm 3.
The Jukes-Cantor model of sequence evolution is
1. Associate to the root r a sequence ξr = ξr1 , . . . , ξrk ∈ {A, T, G, C}k of length k, where
each character ξrj is drawn independently and uniformly at random from {A, T, G, C}.
2. Initialize the set U with the children of the root r of G(i) .
3. Repeat until U = ∅.
(a) Pick u ∈ U , and let u− be the parent of u in G(i) .
(b) Associate a sequence ξu ∈ {A, T, G, C}k as follows. ξu is obtained from ξu− by
mutating each site independently with probability p(u,u− ) . If a mutation occurs at
a site j, it gets assigned a uniformly random character from {A, T, G, C}, else the
corresponding character from ξu− simply gets copied.
(c) Remove u from U and add any children of u to U .
Algorithm 3: The Jukes-Cantor process
Multispecies coalescent Let S = (Vs , Es ; r, ~τ , ~µ) be a fixed species phylogeny. For the sake
of this algorithmic description, we will work with what we call root-extended trees and forests.
Given a weighted rooted tree, the corresponding root-extended tree simply has a new vertex r0 that
is connected to r with a (potentially zero-) weighted edge. We will call r0 the root of such a tree,
and the edge connecting r and r0 as the root edge. A root-extended forest is simply a union of
root-extended trees. Only for the description that follows, when we write tree and forest, we mean
the root-extended versions unless otherwise specified.
We first describe a function MSC that takes as input a vertex vc ∈ Vs , and returns a forest. We
will obtain a random gene tree from the multispecies coalescent as follows: (1) invoke the function
MSC with r, the root of S as input, and (2) contract the root-edge of the tree returned (thus making
it a gene tree per Definition 2). That is, for i ∈ [m], G(i) = MSC(r), with the root edge contracted.
This function, as we can see in Algorithm 4, recursively descends the species phylogeny S and it
calls the function coalesce() at every stage of this recursion.
The coalesce() function works with rooted forests. This function operates at each branch
of the species tree, and (potentially) merges the genealogies of its two descendant populations. It
takes as input two forests F1 and F2 corresponding to the descendants of the current branch (or
23
Function MSC(vc ).
1. If vc ∈ L, i.e., vc is a leaf of S:
(a) Return the following single edge (root-extended) tree: (vc , r0 ). One vertex of this
edge corresponds to the current leaf and the other is an ancestor to this leaf, and the
root of the tree. The length of the edge created is µe × τe , where e ∈ Es is the edge
incident upon vc in S.
2. Else
(a) Let d1 and d2 be the descendants of vc .
(b) If vc is r, the root of S, then set τ, µ = ∞. Otherwise, let τ and µ respectively be
the length and mutation rate of the branch connecting vc to its immediate ancestor
in S.
(c) Return the following forest: coalesce(MSC(d1 ), MSC(d2 ), τ , µ)
Function coalesce(F1 , F2 , τ , µ).
1. Create a new forest F that is a union of F1 and F2 . Set k = number of roots (or lineages)
in F .
2. While k > 1:
(a) Choose a random
pair of (distinct) roots r1 and r2 from F . Also draw a random
time t ∼ Exp( k2 ).
(b) If t ≥ τ
i. Increase the length of all the k root edges in F by µ × τ . Return F .
(c) Else
i. Set τ := τ − t. Create new vertices r3 , r30 .
ii. Make r1 and r2 descendants of r3 , where the lengths of the branches (r1 , r3 )
and (r2 , r3 ) are both set equal to µ × t. Make r30 the root of this newly created
tree, connecting r30 to r3 with a length 0 branch.
iii. Also, add µ × t to the root edges of the other trees in F . Now, F has one fewer
root, so set k := k − 1.
3. If k is 1, then return F , adding µ × τ to the unique root edge of F .
Algorithm 4: The multispecies coalescent process
population) that it is invoked at. It also takes the mutation rate µ and length τ associated with the
current branch. It then returns a single forest F after performing (potentially) multiple coalescence
operations. The details are in Algorithm 4.
24
Gene 1
Time
Gene 2
Figure 5: A species phylogeny (the thick, shaded tree) and two samples from the multispecies
coalescent.
Fig. 2 shows two sample draws from the multispecies coalescent process. Notice that while the
topology of Gene 1 (red gene) agrees with the topology of the underlying species tree, the topology
of Gene 2 (blue gene) does not. This happens since in Step 2(b) of coalesce, if the randomly
drawn time t is larger than τ , the chosen pair of lineages do not coalesce in the current population.
This sort of discordance in the topologies of gene trees is called incomplete lineage sorting (ILS).
For more details, we refer the reader e.g. to [DR09].
The density of the likelihood of a gene tree G(i) = V (i) , E (i) can now be written down as
follows. We will focus our attention on the branch e ∈ E of the species tree and for the gene tree
(i)
(i)
G(i) , let Ie and Oe be the number of lineages entering and leaving the branch e respectively.
For instance, consider Gene 1 in Figure 5. Here, two lineages enter the branch e4 and one lineage
leaves it. On the other hand, in the case of n
Gene 2 in Figure 5, two olineages enter the branch
(i)
(i)
(i)
e4 and two lineages leave it. Let te,s , s = 1, 2, . . . , Ie − Oe + 1 be the s−th coalescent
time corresponding to G(i) in the branch e. From the algorithm described above, each pair of
lineages in a population can coalesce at a random time drawn according to the Exp(1) distribution
(i)
independently of each other. Therefore, after the (s − 1)-th coalescent event at time te,s−1 , there
(i)
are Ie − s + 1 surviving lineages in branch e and the likelihood that the s−th coalescence time in
(i)
(i)
branch e is te,s corresponds to the event that the minimum of Ie −s+1
random variables distributed
2
(i)
(i)
according to Exp(1) has the value te,s − te,s−1 . It follows that the density of the likelihood of G(i)
can be written as
)
(
(i)
(i)
(i)
i
Y Ie −O
Ye +1
Ie − s + 1 h (i)
(i)
exp −
te,s − te,s−1 ,
(9)
2
s=1
e∈E
(i)
(i)
where, for convenience, we let te,0 and t
(i)
(i)
e,Ie −Oe +1
25
be respectively the divergence times of the
population in e and of its parent population.
We will also need the density and quantiles of gene tree pairwise distances. For a pair of leaves
a, b ∈ L, δab is the branch length induced by the random gene tree which is drawn according to
the MSC. Notice that, by definition, δab ≥ µab . We will let fab (·) and Fab (·) denote respectively
ab
. Because
the density and the cumulative density function of the random variable Zab , δab −µ
2
of the memoryless property of the exponential, it is natural to think of the distribution of Zab as a
mixture of distributions whose supports are disjoint, corresponding to the different branches that
the lineages go through before coalescing. We state this more generally as follows. Suppose that
U1 , U2 , . . . , Ur are subsets of R such that they satisfy
sup(Ui ) ≤ inf(Ui+1 ),
i = 1, 2, . . . , r.
Suppose that f is a probability density function such that
f (x) =
r
X
ωi fi (x),
i=1
P
where ω1 , ω2 , . . . , ωr ∈ (0, 1) are such that ri=1 ωi = 1, and the density fi is supported on Si for
i = 1, . . . , r. Then, the quantile function of f is given as follows
(
" i−1
!)
Pi−1 !
r
i
X
X
X
ωi
α − j=0
,
QF (α) , inf {x ∈ R : α ≤ F (x)} =
1 α∈
QFi
ωj ,
ωj
ωi
i=1
j=0
j=0
where, we set ω0 = 0. Specializing to fab under the multispecies coalescent, it follows that there
exists a finite sequence of constants µ1 , . . . , µr ∈ [µL , µU ] and h0 , . . . , hr−1 ∈ [f 0 , g 0 + ng] such
that
Pi−1 −1
Pi
−1
ωi = e− j=1 µj (hj −hj−1 ) − e− j=1 µj (hj −hj−1 ) ,
and
−1
fi (x) =
Cancellations lead to
fab (x) =
r
X
e−
−µi
µ−1
i e
−1
1 − e−µi
Pi−1
j=1
(x−hi−1 )
(hi −hi−1 )
.
−1
µ−1
j (hj −hj−1 ) µ−1 e−µi (x−hi−1 ) .
i
i=1
This formula implies that the density is bounded between positive constants. We will need the
(α)
(α)
following implication. For any α ∈ [0, 1), we let δab and pab denote the α-quantile of the δab
(α)
(α)
and pab respectively.
Since by definition pab = p(δab ), we have that pab = p(δab ), where p(x) =
3
1 − e−4x/3 . Then, for any 0 < β 0 < β < 1, there are constants 0 < c0 < c00 < +∞ (depending
4
on µL , µU , g, g 0 , n, β 0 , β) such that for any ξ ∈ (0, 1 − β 0 ), we have
(β+ξ)
− δab ≤ c00 ξ,
(β+ξ)
− pab ≤ c00 ξ.
c0 ξ ≤ δab
(β)
(10)
(β)
(11)
and, hence,
c0 ξ ≤ pab
26
B
Identifiability result: proofs
The key steps in the proof of Theorem 3 follow.
Proof (Proposition 1): Recall the definition of the event
(1/2)
(1/2)
(1/2)
.
, δyz > δyz
, δxz > δxz
EI = δxy ≤ δxy
Our goal is to show that it has positive probability and that it implies that the rooted topology of G
is xy|z. This makes sense on an intuitive level because the event EI requires that δxy is somewhat
small and that δxz , δyz are somewhat large. To make this rigorous we use the fact that, conditioned
on coalescence occurring in the top population, the time to coalescence inside that population is
identically distributed for all pairs of lineages. That observation facilitates the comparison the
δ-quantiles, as we show now.
Let Γxy be twice the weighted height of the MSC process between the lineages of x and y in the
common ancestral population. Notice that with this definition of Γxy , we can write δxy = µxy +Γxy .
We let Γxy = 0 if the coalescence between the lineages of x and y occurs below the common
ancestral population (which in this case is only possible for the two closest populations in the
(β)
species tree), an event we denote by Bxy . For β ∈ [0, 1], let Γxy be the β-th quantile of Γxy . We
define the quantities above similary for the other pairs. We make three observations:
(1/2)
(1/2)
a) By definition of Γxy , the event {δxy ≤ δxy } implies the event {Γxy ≤ Γxy }. (Note that
(1/2)
the two events are not in fact equivalent, though, because of the possibility that Γxy = 0.)
(1/2)
(1/2)
(1/2)
(1/2)
Similarly, {δxz > δxz } implies {Γxz > Γxz } and {δyz > δyz } implies {Γyz > Γyz }.
b) Irrespective of the species tree topology, the lineages of at least of one the pairs (x, z) or
(y, z) can only coalesce in the common ancestral population. See Figures 2 and 3, for instance.
c) By symmetry, conditioned on coalescence in the common ancestral population, all Γs are
equal in distribution, i.e.,
d
d
c
c
c
Γxy | Bxy
= Γxz | Bxz
= Γyz | Byz
.
As a consequence of b) and c), we have
(1/2)
Γ(1/2)
≤ max{Γxz
, Γ(1/2)
yz },
xy
(12)
where we used that, conditioned on Bxy , it holds that Γxy = 0. Combining this with a), we get
that EI implies the event
(1/2)
(1/2)
,
Γxy ≤ Γ(1/2)
xy , Γxz > Γxz , Γyz > Γyz
which together with (12) implies the event
Γxy < max{Γxz , Γyz }.
27
This last event can only happen when the gene tree topology is xy|z.
It remains to prove that EI has positive probability. Let F be the event that G has rooted
topology xy|z. Note that
P[EI ] ≥ P[EI , F ]
(1/2)
(1/2)
(1/2)
(1/2)
(1/2)
= P[δxy ≤ δxy
] P[F | δxy ≤ δxy
] P[δxz > δxz
, δyz > δyz
| F , δxy ≤ δxy
],
where each term on the last line is clearly positive under the MSC.
Proof (Proposition 2): Conditioned on EI , we know from Lemma 1 that the coalescence between
the lineages of x and z happens in the common ancestral population of x, y and z, irrespective of
the species tree topology. The same holds for y and z. This implies that
δxz = µrx + µrz + Γxz ,
(13)
δyz = µry + µrz + Γyz ,
(14)
and
where the Γs are defined in the proof of Lemma 1. Observe further that in fact, conditioned on EI ,
Γxz = Γyz ,
(15)
almost surely. Hence, combining (13), (14), and (15),
E[δxz − δyz | EI ] = µrx − µry = ∆xy ,
as claimed.
C
Main theorem: proof
Theorem 1 follows from Proposition 3 below.
Proposition 3 (Data requirement: general version). Suppose we have data {ξ ij }i∈[m],j∈[k] generated according to the MSC-JC(m, k) process on a species phylogeny S = (Vs , Es ; r, ~τ , ~µ). The
mutation rates, leaf-edge lengths and internal-edge lengths are respectively in (µL , µU ), (f 0 , g 0 )
and (f, g). For any ε > 0 and C > 0 there is a constant c1 > 0 such that Algorithm 1 correctly
identifies the topology of S restricted to X = {1, 2, 3} with probability at least 1 − ε, provided
there is a partitioning of the set of genes [m] = MR1 tMR2 tMQ1 tMQ2 such that the following
conditions hold:
log k
−1
log ε−1
|MR1 | ≥ c1 log ε
|MR2 | ≥ c1 1 ∨
kf 2
|MQ2 | ≥ c1 f −2 (α + f ) log ε−1 ,
|MQ1 | ≥ c1 α−1 log ε−1
√
where α = m−1 log m ∨ k −0.5 log k. And, the sequence length k satisfies:
C
p
−1
−1
k ≥ c1 log |MR2 | ε
∨ c1 f
log k .
28
Under condition (1) of Theorem 1, to satisfy the inequalities above, we can choose |MR1 | & 1,
1
|MR2 | & 1 ∨ kf1 2 , |MQ1 | & 1 and |MQ2 | & f1 ∨ √kf
2.
Proof (Proposition 3): Without loss of generality we assume that the topology of the true species
tree restricted to this triple is 12|3.
Main steps Let SX be the species tree S restricted to X = {1,2, 3} and let r0 denote
its root,
√
i.e., the most recent common ancestor of X . We let α = max m−1 log m, k −0.5 log k . We
partition the loci [m] = MR1 t MR2 t MQ1 t MQ2 , such that the size of each partition satisfies
the conditions specified in Proposition 3. The reconstruction algorithm on X has two steps: 1) a
reduction to the ultrametric case by the addition of noise and 2) the ultrametric quantile test. We
divide the proof accordingly:
1) Ultrametric reduction: In this step, we invoke Algorithm 2 with sequence data {ξxij : x ∈
ij
X , i ∈ [m], j ∈ [k]}. The algorithm outputs new sequences {ξx,N
: x ∈ X , i ∈ MQ1 t
MQ2 , j ∈ [k]}, and Proposition 4 (proved in Section D) guarantees that these output sequences appear as though they were drawn from an almost-ultrametric species phylogeny
SX0 that has the same topology as SX . In particular, using Proposition 4, we know that there
is a constant c1 > 0 such that if
|MR1 |, |MR2 | ≥ c1 log(4ε−1 ),
k ≥ c1 log |MR2 | + c1 log(4ε−1 ),
k|MR2 | ≥ c1 φ−2 log(4ε−1 ).
ij
then, {ξx,N
: x ∈ X , i ∈ MQ1 t MQ2 , j ∈ [k]} has the same distribution as a multispecies
ˆxy ), where SX0 and S have the same rooted
coalescent process on SX0 with branch lengths
(µ̇
√
ˆxy ) and (µ̇xy ) are O(f / log k)-close with probability at least 1 − ε. Notice
topology, and (µ̇
√
that we have set the value of φ in Proposition 4 to O(f / log k), which will turn out to be
what we need in Step 2 below.
ij
2) Quantile test: Now, we invoke Algorithm 1 with the sequence data {ξx,N
: i ∈ MQ1 t
MQ2 , j ∈ [k]} output by Step 1. By Propositions 7 and 9 (proved in Section E), it follows
that if
|MQ1 | ≥ c1 α−1 log ε−1
|MQ2 | ≥ c1 f −2 (α + f ) log ε−1 ,
then, with probability at least 1 − ε, Algorithm 1 returns the right topology .
This concludes the proof.
29
D
Ultrametric reduction: proofs
Theorem 2 follows from Proposition 4 below. In Proposition 4, we show that the approximate
stochastic Farris transform defined in Section 3.2 outputs sequence data that looks statistically as
though it was generated from an ultrametric species phylogeny.
b xy for all x, y ∈ X , and supposing that min{∆
b 12 , ∆
b 13 } ≥ 0 (the other cases
Given estimates ∆
follow similarly), we let
b 1x + ∆
b 1y ,
ḃxy = µxy + ∆
µ
x, y ∈ X .
~ˆ be a
Compare this to the definition of µ̇xy in (3). Recall Definition 1, and let S 0 = (Vs , Es , r, ~τ , µ̇)
species phylogeny with the same topology and branch lengths as S restricted to X , and mutation
ḃ e∈Es that are chosen such that: (a) µ
ḃe = µe if e ∈ Es is an internal branch, and (b) for
rates {µ}
ḃe be chosen (uniquely) such that mutation rate
all e ∈ Es that are incident on the leaves of S, let µ
ḃxy . The sets MR1 , MR2
weighted distance on S 0 between any pair of leaves x, y ∈ X is given by µ
referred to below are defined in Algorithm 2.
Proposition 4 (Ultrametric reduction: general version). Suppose that we have sequence data
{ξ ij }i∈[m],j∈[k] generated according to the MSC-JC(m, k) process on a three-species phylogeny
S = (Vs , Es ; r, ~τ , ~µ). The mutation rates, leaf-edge lengths and internal-edge lengths are respectively in (µL , µU ), (f 0 , g 0 ) and (f, g). Then, the output of Algorithm 2 is distributed according to
the MSC-JC process on the species tree S 0 defined above. Furthermore there is a constant c2 > 0
ḃxy is φ-close to the ultrametric
such that, for any ε, φ ∈ (0, 1), with probability at least 1 − ε, µ
(µ̇xy ), provided
|MR1 |, |MR2 | ≥ c2 log(4ε−1 ),
k ≥ c2 log |MR2 | + c2 log(4ε−1 ),
k|MR2 | ≥ c2 φ−2 log(4ε−1 ).
Under condition (2) of Theorem 2, to satisfy the inequalities above, we can choose |MR1 | & 1 and
|MR2 | & 1 ∨ kf1 2 .
D.1
Proof of Proposition 4
Proof (Proposition 4): As explained in Section 3.2, there are three main steps to this proof, which
we summarize in a series of propositions.
For a gene i and leaves x, y ∈ X , let
pbixy =
1 X ij
1 ξx 6= ξyij .
k
j∈[k]
Furthermore, we need to split the above average into two halves to avoid unwanted correlations as
30
we explain below. We denote these as5
k/2
pbi↓
xy
2X
=
1{ξxij 6= ξyij }
k j=1
2
=
k
pbi↓
xy
and
k
X
j=k/2+1
1{ξxij 6= ξyij }.
(β)
And, for β ∈ [0, 1], let pbxy be the corresponding empirical quantiles computed based on the set
i
: i ∈ MR1 }. Fix a permutation (x, y, z) of (1, 2, 3). Consider the following subset of genes
{b
qxy
in MR2 :
(1/3) (2/3)
(2/3)
I = i ∈ MR2 : pbi↓
bxy
, pbxz ≤ pbi↓
b(5/6)
byz
≤ pbi↓
b(5/6)
.
xy ≤ p
xz ≤ p
xz , p
yz ≤ p
yz
We first show that the rooted topologies in I are highly likely to be xy|z. We also prove some
technical claims that will be useful in the proof of Proposition 6 below.
Proposition 5 (Fixing gene tree topologies). There are constants c9 , c10 , c010 , ε0 > 0 such that, with
probability at least
1 − 10 exp(−2c29 |MR1 |) − 6|MR2 | exp −kε20 − 2 exp −2c210 |MR2 | ,
the following hold:
(a) the rooted topology of all gene trees in I is xy|z,
(7/24)
(b) for all i ∈ I, pixy ≤ pxy
(17/24)
, pxz
(19/24)
≤ pixz ≤ pxz
(17/24)
, pyz
(19/24)
≤ piyz ≤ pyz
,
(c) the size of I is greater than c010 |MR2 |.
The proof is given in Section D.1.1. Notice that Proposition 5 guarantees that all the genes in
I satisfy conditions (a) and (b). We can weaken our requirements on how big k needs to be by
relaxing this and performing a more careful analysis.
Using
1 X i↑
1 X i↑
pbIxz =
pbxz
and
pbIyz =
pb ,
|I| i∈I
|I| i∈I yz
let
b xy =
∆
3
3
4 I
4 I
− log 1 − pbyz
− − log 1 − pbxz
.
4
3
4
3
Recall that
∆xy = µrx − µry .
b xy is a good approximation to ∆xy . Let I be the event that the conclusion of
We next show that ∆
e and E
e to denote the
Proposition 5 holds. To simplify the notation, throughout this proof, we use P
probability and expectation operators conditioned on I .
5
For simplicity, we assume that k is even. This is not a critical requirement and can be easily relaxed.
31
Proposition 6 (Estimating distance differences). There is a constant c11 ∈ (0, 1) such that with
e
P-probability
at least
1 − 4 exp −c11 k|MR2 |φ2 ,
the following holds:
b xy − ∆xy ≤ φ/2.
∆
The proof is in Section D.1.2.
We repeat the height difference estimation above for all pairs in X . Therefore, by a union
bound, we get the above guarantee for all pairs with probability at least 1−12 exp (−c11 k|MR2 |φ2 ).
Without loss of generality, assume that
µr1 ≥ max{µr2 , µr3 },
and recall the Farris transform
µ̇xy = µxy + 2µr1 − µrx − µry = µxy + ∆1x + ∆1y ,
x, y ∈ X ,
which defines an ultrametric, and consider the approximation
b 1x + ∆
b 1y ,
ḃxy = µxy + ∆
µ
x, y ∈ X .
ḃxy )
Assuming that the conclusion of Proposition 6 holds for all x, y ∈ X , we have shown that (µ
is φ-close to the ultrametric (µ̇xy ). As we explained in Section 3.2, we produce a new sequence
dataset using an approximate stochastic Farris transform
i
{ξx,N
} = F({ξxi }; (b
µrx )).
By the Markov property, the transformation F has the effect of stretching the leaf edges of the
gene trees by the appropriate amount.
Hence, again by a union bound, we get the claim of Theorem 2 except with probability
10 exp(−2c29 |MR1 |) + 6|MR2 | exp −2kε20 + 2 exp −2c210 |MR2 | + 12 exp −c11 k|MR2 |φ2 .
(16)
We can get the data requirement result by asking for the conditions under which the above quantity
is less than ε.
D.1.1
Proof of Proposition 5
Proof (Proposition 5): Let (x, y, z) be an arbitrary permutation of the leaves (1, 2, 3). The idea of
the proof is to rely on Proposition 1, which we rephrase in terms of p-distances. For a gene Gi Let
3
i
1 − e−4δxy /3 .
pixy =
4
And, for β ∈ [0, 1], the corresponding β-th quantile is given by
(β)
3
(β)
pxy
=
1 − e−4δxy /3 ;
4
32
similarly for the other pairs. Then, by Proposition 1, the event
(1/2) (1/2)
(1/2)
, pxz < pixz , pyz
< piyz ,
EIi = pixy ≤ pxy
implies that the rooted topology of Gi is xy|z. Our goal is to show that
Qi = pbi↓
b(1/3)
b(2/3)
≤ pbi↓
b(5/6)
b(2/3)
≤ pbi↓
b(5/6)
,
xy ≤ p
xy , p
xz
xz ≤ p
xz , p
yz
yz ≤ p
yz
(17)
(β)
implies EIi with high probability. We do this by controlling the deviations of pbuw and pbi↓
uw . We state
i↓
in
(17)
are
included
and
p
b
the necessary claims as a series of lemmas. (The upper bounds on pbi↓
yz
xz
for technical reasons that will be explained in the proof of Proposition 6. This requirement may
not be needed, but it makes the analysis simpler.)
Recall that we use only the genes in MR for the reduction step and this in turn is divided
into disjoint subsets MR1 and MR2 . The quantiles are estimated using MR1 , while MR2 is used
(β)
to compute I. We do not argue about the deviation of pbuw from the true β-th quantile of the
(β)
(β)
distribution of pbiuw . Instead we show that pbuw is close to the β-th quantile puw of the disagreement
probability under the MSC, that is, the quantile without the sequence noise. We argue this way
because the events that we are ultimately interested in (whether a certain coalescence event has
occured in a particular population) are expressed in terms of the MSC. Note that, in order to obtain
a useful bound of this type, we must assume that the sequence length is sufficiently long, that is,
that the sequence noise is reasonably small. Hence this is one of the steps of our argument where
we require a lower bound on k.
(β)
Lemma 1 (Deviation of pbuw ). Fix a pair u, w ∈ X and a constant β ∈ (0, 1). For all ε0 > 0 and
0 < ε1 < min{β, 1 − β}, there is a constant c0 > 0 such that
(β+ε1 )
1)
P p(β−ε
− ε0 ≤ pb(β)
+ ε0 ≥ 1 − 2 exp −2c29 |MR1 | ,
uw
uw ≤ puw
provided that k is greater than a constant depending on ε0 and ε1 .
Proof: We prove one side of the first equation. The other inequalities follow similarly. Define the
random variable
(β+ε1 )
M = i ∈ MR1 : pbi↓
+ ε0 ,
uw ≤ puw
and observe that
(β+ε1 )
P pb(β)
+ ε0 ≤ P [M < β|MR1 |] .
uw > puw
To bound the probability on the r.h.s., we note that
i
(β+ε1 )
(β+ε1 )
1)
1)
P pbi↓
+ ε0 ≥ P pbi↓
+ ε0 piuw ≤ p(β+ε
P puw ≤ p(β+ε
uw ≤ puw
uw ≤ puw
uw
uw
≥ 1 − exp −kε20 (β + ε1 ) ,
(β+ε )
pi↓
by Hoeffding’s inequality [Hoe63] and the definition of puw 1 . We also used that E[b
uw | puw ] =
puw . By Hoeffding’s inequality applied to M , we have that
(β+ε1 )
P pb(β)
+ ε0 ≤ P [M < β|MR1 |] ≤ exp −2c29 |MR1 | ,
uw > puw
33
where
c0 = (β + ε1 ) 1 − exp −kε20 − β,
which is strictly positive, provided that k is greater than a constant depending on ε0 and ε1 .
On the other hand, standard concentration inequalities allow us to control the deviation of pbiuw .
Observe that, piuw being itself random, the deviation holds conditionally on the value of piuw .
Lemma 2 (Deviation of pbiuw ). Fix a pair u, w ∈ X . For all i and ε0 > 0,
i↓
P |b
puw − piuw | ≥ ε0 | piuw ≤ 2 exp −kε20 ,
almost surely.
i
Proof: Note that, conditioned on piuw , k/2b
pi↓
uw is distributed as Bin(k, puw ). The result then follows
from Hoeffding’s inequality.
Fix 0 < ε1 < 1/24 and pick ε0 > 0 small enough that
(1/3−ε1 )
(9/24)
1)
p(7/24)
≤ pxy
− 2ε0 ≤ p(1/3+ε
+ 2ε0 ≤ pxy
xy
xy
1)
1)
p(15/24)
≤ p(2/3−ε
− 2ε0 ≤ p(2/3+ε
+ 2ε0 ≤ p(17/24)
xz
xz
xz
xz
1)
1)
p(19/24)
≤ p(5/6−ε
− 2ε0 ≤ p(5/6+ε
+ 2ε0 ≤ p(21/24)
xz
xz
xz
xz
(18)
1)
1)
p(15/24)
≤ p(2/3−ε
− 2ε0 ≤ p(2/3+ε
+ 2ε0 ≤ p(17/24)
yz
yz
yz
yz
1)
1)
p(19/24)
≤ p(5/6−ε
− 2ε0 ≤ p(5/6+ε
+ 2ε0 ≤ p(21/24)
.
yz
yz
yz
yz
Notice that the fact that these inequalities hold is guaranteed by (11) (in Section A) which characterizes the behavior of the quantile functions of the random variables associated with the MSC.
(β−ε )
(β)
(β+ε )
Let Equ be the event that the inequality in Lemma 1, i.e., puw 1 − ε0 ≤ pbuw ≤ puw 1 +
(1/3)
(2/3)
(5/6)
(2/3)
(5/6)
ε0 , holds for pbxy , pbxz , pbxz , pbyz and pbyz , which occurs with probability at least 1 −
10 exp(−2c20 |MR1 |) by a union bound. Let Di be the event that the inequality in Lemma 2, i.e.,
i
|b
pi↓
uw − puw | ≥ ε0 , holds for all pairs (u, w) in X , an event which occurs with probability at least
1 − 6 exp(−kε20 ) by a union bound. Given Equ , Di and Qi , we have
1)
pixy ≤ pbixy + ε0 ≤ pb(1/3)
+ ε0 ≤ p(1/3+ε
+ 2ε0 ≤ p(1/2)
xy
xy
xy ,
and similarly for the other pairs. That is, EIi holds. Finally, we bound the probability that all i in I
satisfy Di with the probability that all i in MR2 satisfy Di . (In fact we show below that with high
probablity |I| = Θ(|MR2 |).) That is, the probability that all i ∈ I satisfy EIi simultaneously is at
least
P[EIi , ∀i ∈ I] ≥ 1 − 10 exp(−2c20 |MR1 |) − 6|MR2 | exp −kε20 .
(19)
It remains to bound the size of I.
Lemma 3 (Size of I). There are constants c10 , c010 > 0 such that
P[|I| ≥ c010 |MR2 | | Equ ] ≥ 1 − 2 exp −2c210 |MR2 | ,
provided k is greater than a constant depending on ε0 .
34
(20)
Proof: We show that, under Equ , the event Qi has constant probability and we apply Hoeffding’s
inequality.
(7/24)
Observe that, by (18), the events {pixy ≤ pxy } and Di imply
(7/24)
1)
− ε0 ≤ pb(1/3)
+ ε0 ≤ p(1/3−ε
pbixy ≤ pixy + ε0 ≤ pxy
xy .
xy
Hence, a similar argument shows that
, p(17/24)
≤ pixz ≤ p(19/24)
, p(17/24)
≤ piyz ≤ p(19/24)
},
Di ∩ {pixy ≤ p(7/24)
xy
xz
xz
yz
yz
implies Qi . This leads to the following lower bound
P[Qi | Equ ]
≥ P[Di ∩ {pixy ≤ p(7/24)
, p(17/24)
≤ pixz ≤ p(19/24)
, p(17/24)
≤ piyz ≤ p(19/24)
} | Equ ]
xy
xz
xz
yz
yz
≥ P[pixy ≤ p(7/24)
, p(17/24)
≤ pixz ≤ p(19/24)
, p(17/24)
≤ piyz ≤ p(19/24)
]
xy
xz
xz
yz
yz
×P[Di | {pixy ≤ p(7/24)
, p(17/24)
≤ pixz ≤ p(19/24)
, p(17/24)
≤ piyz ≤ p(19/24)
} ∩ Equ ]
xy
xz
xz
yz
yz
00
2
≥ c10 1 − 6 exp −kε0 ,
for some constant c0010 > 0. This existence of the latter constant follows from an argument similar
to that leading up to (11) (but is somewhat complicated by the fact that pixy , pixz and piyz are not
independent). The expression on the last line is a strictly positive constant provided k is greater
than a constant depending on ε0 . Finally, applying Hoeffding’s inequality to |I|, we get the result.
Combining (19) and (20) concludes the proof.
D.1.2
Proof of Proposition 6
Proof (Proposition 6): Fix x, y ∈ X and let z be the unique element in X − {x, y}. The proof
idea is based on Proposition 2. Recall that I be the event that the conclusion of Proposition 5
holds. Let also GI be the event that the weighted gene trees in I are {Gi }i∈I . Similarly to the proof
of Proposition 2 we note that, conditioned on I , in all genes in I the coalescences between the
lineages of x and z happen in the common ancestral population of x, y and z, irrespective of the
species tree topology. The same holds for y and z. That implies that, for i ∈ I,
i
δxz
= µrx + µrz + Γixz ,
and
i
δyz
= µry + µrz + Γiyz ,
where the Γi s are defined as in the proof of Lemma 1. Observe further that in fact, conditioned on
I , for i ∈ I
Γixz = Γiyz ,
(21)
35
almost surely. Hence
"
1 X i
pb GI
|I| i∈I xz
1 X i
=
p
|I| i∈I xz
#
e pbI GI = E
e
E
xz
3
=
4
3
=
4
1 X −4δxz
i /3
1−
e
|I| i∈I
!
1 − 2e−4µrx /3−4µrz /3
1 X −4Γixz /3
e
|I| i∈I
!!
,
and similarly for the pair (y, z). Letting `(x) = − 43 log 1 − 43 x , we get
!
e pbI GI
1 − 4/3E
xz
e
1 − 4/3E pbIyz GI
P
1
−4µrx /3−4µrz /3
−4Γixz /3
e
e
i∈I
|I|
3
P
= − log
1
−4Γiyz /3
4
−4µ
/3−4µ
/3
ry
rz
e
i∈I e
|I|
3
= − log e−4µrx /3+4µry /3
4
= ∆xy ,
e pbI GI − ` E
e pbI GI = − 3 log
` E
xz
yz
4
where we used (21) on the third line. Observe that the computation above relies crucially on the
conditioning on GI .
It remains to bound the deviation of
b xy = ` pbIxz − ` pbIyz ,
∆
around
∆xy
I
I
e
e
= ` E pbxz GI − ` E pbyz GI ,
and take expectations with respect to GI . We do this by controlling the error on pbIxz and pbIyz ,
conditionally on GI . Indeed, observe that the function ` satisfies the following Lipschitz property:
for 0 ≤ x ≤ y ≤ M < 1/2,
Z y
|x − y|
1
dt ≤
.
(22)
|`(x) − `(y)| =
1 − 4M/3
x 1 − 4t/3
n
o
e pbI GI and max pbI , E
e pbI GI
b xy −∆xy |, it suffices to bound pbIuz − E
Hence, to control |∆
uz
uz
uz
for u = x, y.
e pbI GI , we use the upper bounds on pbi and pbi in the definition of the set I.
To bound E
uz
xz
yz
36
Lemma 4 (Conditional expectation of pbIuz ). Fix u = x or y. There is a constant c012 ∈ (0, 1/2)
small enough,
e pbI GI ≤ 1 − c0 ,
E
uz
12
2
e
P-almost
surely.
e [b
Proof: Using piuz = E
piuz | GI ] for i ∈ I, by Proposition 5 (b), we have that
1
i
4
i /3
i
−4δuz
e
E pbuz GI = puz =
1−e
≤ − c012 ,
3
2
for some constant c02 ∈ (0, 1/2). Again, this constant depends on bounds on the mutation rate and
the depth of the tree. Hence,
X
e pbI GI = 1
e pbi GI ≤ 1 − c0 .
E
E
uz
uz
2
|I| i∈I
2
e pbI GI , we use Hoeffding’s inequality.
To bound pbIuz − E
uz
Lemma 5 (Conditional deviation of pbIuz ). Fix u = x or y. For all φ0 > 0,
i
h
e pbI − E
e pbI GI ≥ φ0 GI ≤ 2 exp −k|I|(φ0 )2 ,
P
uz
uz
almost surely.
Proof: Observe first that, conditioned on GI , the k |I| sites that are averaged over in the computation of
k
2 X X
I
pbuz =
1 ξuij 6= ξzij ,
(23)
k|I| i∈I
j=k/2+1
are independent. Secondly, each random variable in (23) is bounded by 1. Therefore, from Hoeffding’s inequality, we have that
h
i
e pbI − E
e pbI GI ≥ φ0 GI ≤ 2 exp −k|I|(φ0 )2 ,
P
uz
uz
almost surely.
We set φ0 = 21 c012 (φ/2), which is < c012 since φ ≤ 1. Combining (22) and Lemmas 4 and 5, we
37
get that conditioned on GI
I
e pbI GI
e pbI GI
b xy − ∆xy ≤ ` pbI − ` E
∆
+
`
p
b
−
`
E
xz
xz
yz
yz
e pbI GI
e pbI GI
pbIxz − E
pbIyz − E
xz
yz
n
o+
n
≤
o
I
I
I
I
e
e
1 − 4/3 max pbxz , E [b
pxz | GI ]
1 − 4/3 max pbyz , E pbyz GI
φ0
2(c012 − φ0 )
1 0
c (φ/2)
≤ 2 0 2 121 0
2(c12 − 2 c12 (φ/2))
≤2
≤ φ/2,
e
where we used that φ/2 ≤ 1, except with P-probability
1 0 2 0
2
0 2
4 exp −k|I|(φ ) ≤ 4 exp − (c12 ) c10 k|MR2 |φ = 4 exp −c12 k|MR2 |φ2 ,
8
by setting c2 = 81 (c02 )2 c01 . Taking expectations with respect to GI gives the result.
E
Quantile test: robustness analysis
In this section, we analyze Algorithm 1.
Control of empirical quantiles To perform the quantile test, Algorithm 1 has access to a set of
genes MQ that were not used in the reduction step above; this is to avoid unwanted correlations.
This set is in turn partitioned as MQ = MQ1 t MQ2 so that |MQ1 | , |MQ2 | satisfy the conditions
of Proposition 3. The first step in Algorithm 1 is to compute
ai well-chosen empirical quantile
: i ∈ MQ1 . The quantile we
of qbxy for each pair of leaves x, y ∈ X based on the dataset qbxy
compute is (a constant multiple of)
n
o
p
−1
−0.5
α = max m log m, k
log k .
In the following proposition, we show that these empirical quantiles are well-behaved, and provide
a good estimate of the α-quantile of the underlying MSC random variables. We define the random
i
i
variables qxy
and rxy
associated to a gene tree i:
i
i
b 1x + ∆
b 1y ),
qxy
= p(δxy
+∆
i
i
rxy
= p(δxy
+ ∆1x + ∆1y ).
Also, we need the 0-th quantile of these random variables. Notice that
(0)
(0)
b
b
b 1x + ∆
b 1y ).
qxy = p δxy + ∆1x + ∆1y = p(µxy + ∆
38
And similarly,
(0)
rxy
= p(µxy + ∆1x + ∆1y ).
Finally, φ > 0 is the closeness parameter from Proposition 4.
√
Proposition 7 (Quantile behaviour). Let α = max m−1 log m, k −0.5 log k . Then, there exist
constants c3 , c4 , c5 > 0 such that, for each pair of leaves x, y ∈ X , that the c3 α-quantile satisfies
the following
(0) (0)
(0)
(c3 α)
(0)
qbxy
∈ qxy
, qxy + c5 α ⊂ rxy
− c5 φ, rxy
+ c5 φ + c5 α
with probability at least 1 − 6 exp (−c4 |MQ1 | α), provided we condition on the implications of
Proposition 4 holding.
We prove this proposition in Section E.1. In what follows, we will let P̄ and Ē denote the probability and expectation measures conditioned on the event that implications of Propositions 4 and 7.
n
o
(c3 α)
Expected version of quantile test Let qb∗ denote the maximum among qbxy : x, y ∈ X . We
use the genes in MQ2 (which are not affected by the conditioning under P̄) to define a similarity
measure among pairs of leaves in X :
sbxy =
1
i
i ∈ MQ2 : qbxy
≤ qb∗
|MQ2 |
.
(24)
We next show that this similarity measure has the right behavior in expectation. That is, defining
sxy , Ē [b
sxy ] ,
which is the expected version of our similarity measure, we show that s12 > max{s13 , s23 }. This
means that the sxy s expose the topology of the tree SX .
Proposition 8 (Expected version of quantile test). Let sxy be as defined above. Then for any
C2 > 0, there exist constants c6 , c7 > 0 such that s12 − max{s13 , s23 } ≥ c6 p(3f /4) > 0 provided
1
1
log
m ≥ c7
p(3f /4)
p(3f /4)
√
1/C2
log k
k ≥ c7
,
p(3f /4)
√
and the closeness parameter φ ∈ O(p(3f /4)/ log k).
The proof is given in Section E.2.
39
Sample version of quantile test Finally, we conclude the proof by demonstrating that the empirical versions of the similarity measures defined above are also consistent with the underlying
species tree topology with high probability. This follows from a concentration argument detailed
in Section E.3.
Proposition 9 (Sample version of quantile test). There exists a constant c8 > 0 such that the P̄probability that Algorithm 1 fails to identify the correct topology of the triple X is bounded from
above by
|MQ2 | p(3f /4)2
,
4 exp −
c8 (p(3f /4) + α)
provided the conditions of Proposition 8 hold.
E.1
Proof of Proposition 7
In this section we prove Proposition 7, which provides us a control over the behavior of the empirical quantiles computed in the first part of Algorithm 1.
b xy and ∆xy are close. Therefore, using the
Proof (Proposition 7): Proposition 6 guarantees that ∆
fact that p(·) is a Lipschitz function, we know that there exists a constant c05 > 0 such that
(0)
(0)
rxy
− qxy
≤ c05 φ.
The second containment in the statement of the lemma follows from this (after adjusting the constant appropriately).
We will prove the first part following along the lines of [MR15]. Let W be the number of
(0)
i
f be the number of genes i ∈ MQ1 such
≤ qxy , and let W
genes i ∈ MQ1 that are such that qbxy
(0)
i
≤ qxy + c5 α; we will choose c5 below. Notice that the conclusion of Proposition 7 follows
that qbxy
f ≥ |MQ1 | c3 α. So, we
if we show that there is a const c3 > 0 such that W ≤ |MQ1 | c3 α and W
bound the probability that each of these events fail. First, we restate the following lemma about
the cumulative distribution function from [MR15].
Lemma 6 (CDF behavior [MR15]). There exists a constant c03 > 0 such that
c0
(0)
P qbxy ≤ qxy
≤ √3
k
Therefore, the above lemma implies that
c0
(0)
P qbxy ≤ qxy
≤ √3 ≤ c03 α.
k
On the other hand, for every constant c005 > 0, there is a constant c003 > 0 such that
(0) (0)
(0) (0)
(0)
(0)
P qbxy ≤ qxy
+ c005 α ≥ P qbxy ≤ qxy
+ c005 α qxy ∈ qxy
, qxy + c005 α P qxy ∈ qxy
, qxy + c005 α
≥ c003 α,
40
where the last inequality follows from the Berry-Esséen theorem (see e.g. [Dur96]) and (11). We
choose c005 large enough so that we can take c003 > c03 . Then, we choose c5 to be the maximum
among c05 and c005 ; observe that the above inequality holds when c005 is replaced by c5 . Finally, we set
c0 +c00
c3 = 3 2 3 .
There is a constant c4 > 0 such that
(0)
(0)
≥ |MQ1 | c3 α − P qbxy ≤ qxy
P [W ≥ |MQ1 | c3 α] = P W − |MQ1 | P qbxy ≤ qxy
(c3 − c03 )
(0)
≤ P W − |MQ1 | P qbxy ≤ qxy ≥ |MQ1 |
α
2
≤ exp (−c4 |MQ1 | α) ,
(25)
where the last step follows from Bernstein’s inequality (see e.g. [BLM13]). Similarly, it can be
shown that
h
i
f ≤ |MQ1 | c3 α ≤ exp (−c4 |MQ1 | α) .
(26)
P W
Now, a union bound over these two probabilities for each of the three pairs of leaves in X gives us
the stated result.
E.2
Proof of Proposition 8
In this section, we show that the expected version of the quantile test succeeds, and hence prove
Proposition 8 .
Proof (Proposition 8): Recall that we fix a triple of leaves X = {1, 2, 3} such that their topology
with respect to S is given by 12|3 without loss of generality.
Let E12|3 be the event that there is a coalescence in the internal branch and observe that s12 can
be decomposed as follows
s12 = Ē[b
s12 ] = P̄ [b
q12 ≤ qb∗ ]
c
c
= P̄ E12|3 P̄ qb12 ≤ qb∗ E12|3 + P̄ E12|3
P̄ qb12 ≤ qb∗ E12|3
(27)
In the proof in [MR15], instead of qb12 , one deals with rb12 ; indeed, in the ultrametric setting, there is
no difference correction. And it follows from the
h symmetries iof the MSC (namely, the exchangec
ability of the lineages in a population) that P̄ rb12 ≤ qb∗ E12|3
= P̄ [b
r13 ≤ qb∗ ]. This turns out to
suffice to establish the expected version of the quantile test in [MR15]. In our setting, however, we
must control quantitatively the difference between these two probabilities due to the slack added
by the reduction step (Algorithm 2).
Lemma 7 (Closeness to symmetry). For qb∗ as defined in Algorithm 1 and any constant C2 > 0,
there exist constants c07 , c007 > 0 such that
p
c
P̄ [b
q13 ≤ qb∗ ] − P̄ qb12 ≤ qb∗ E12|3
≤ φ2 , c07 φ log k,
41
provided
1
√
log
m≥
φ log k
1/C2
1
k≥
.
φ
c007
1
√
φ log k
We prove this in Section E.2.1.
Using the above lemma in (27), we can now bound s12 from below as follows
c
P̄ [b
r13 ≤ qb∗ ] − φ2
s12 ≥ P̄ E12|3 P̄ qb12 ≤ qb∗ E12|3 + P̄ E12|3
= P̄ E12|3 P̄ qb12 ≤ qb∗ E12|3 − s13 + s13 − φ2 .
(28)
This implies that
s12 − s13 > P̄ E12|3 P̄ qb12 ≤ qb∗ E12|3 − s13 − φ2 .
The expected version of the quantile test succeeds provided the latter quantity is bounded from
below by 0. We establish a better lower bound, which will be useful in the analysis of the sample
version of the quantile test. Towards this end, we will state the following lemma, which is proved
in Section E.2.2.
Lemma 8 (Bounds on tails). There exist positive constants c06 and c006 such that the following hold
P̄ qb12 ≤ qb∗ E12|3 ≥ c06
s13 = P̄ [b
q13 ≤ qb∗ ] ≤ c006 α.
The first inequality captures the intuition that, conditioned on the coalescence, the probability of
qb12 being small is high. The second inequality captures the intuition that since qb∗ behaves roughly
(α)
(α)
b 13 ), the event that qb13 ≤ qb∗ is dominated by the event that the underlying
like q13 = p(δ13 + ∆
MSC random variable satisfies the same inequality (the deviations of the JC random variable on
top of this being of order k −0.5 ).
Notice that, if we use Lemma 8 in (28), there is a constant c6 > 0
s12 − s13 ≥ c6 P̄ E12|3
provided φ2 ≤ c6 p(3f /4) for a large enough c6 > 0, where we used that P̄ E12|3 is lower bounded
by p(3f /4). This, along with a similar argument for s23 , concludes the proof of Proposition 8.
E.2.1
Proof of Lemma 7
In this section, we prove Lemma 7 which is key to accounting for the slack added in the reduction
phase of Algorithm 2.
h
i
∗ c
Proof (Lemma 7): First observe that P̄ rb13 ≤ qb E12|3 = P̄ [b
r13 ≤ qb∗ ]. We prove Lemma 7
h
i
h
i
c
c
by arguing that P̄ qb12 ≤ qb∗ E12|3
is close to P̄ rb13 ≤ qb∗ E12|3
, and that P̄ [b
q13 ≤ qb∗ ] is close to
P̄ [b
r13 ≤ qb∗ ]. Both these statements follow from Lemma 9 below.
42
For a pair of leaves x, y ∈ X , let δxy is the distance between x and y on a random gene
tree drawn according to the MSC, and let pxy = p(δxy ) denote the corresponding expected pdistance. Conditioned on the value of δxy , suppose that we have two Bernoulli random variables
J1 ∼ Bin(k, pxy ) and J2 ∼ Bin(k, pxy + β), for some fixed β ∈ (0, 1 − pxy ). Then we have the
following.
Lemma 9 (Mixture of binomials: CDF perturbation). Suppose that we are given constants c14 , γ >
(c α)
0 such that γ < pxy14 . Then, for any constant C2 > 0 there exist constants c13 , c013 > 0 such that
the following holds
p
P̄ [J1 ≤ kγ] − P̄ [J2 ≤ kγ] ≤ c013 β log k,
provided
1
m ≥ c13 √
log
β log k
1/C2
1
k≥
.
β
1
√
β log k
Observe that, although J1 and J2 above do not depend on m, γ—through α—does. While we
stated the lemma in terms of pxy , this lemma applies to Farris-transformed variables qxy and rxy as
well. We prove this lemma at the end of this section. Notice, first, that this result implies that there
exists a constant c07 > 0 such that
c07 p
φ log k
2
c0 p
c
c
≤ 7 φ log k.
− P̄ rb13 ≤ qb∗ E12|3
P̄ qb12 ≤ qb∗ E12|3
2
P̄ [b
r13 ≤ qb∗ ] − P̄ [b
q13 ≤ qb∗ ] ≤
(0)
(c0 α)
To see why this is true, first observe that qb∗ ≤ q13 + c5 α ≤ q135 ; the first inequality follows from
Proposition 7, and the second inequality follows from (11). This is also true (up to a factor of φ) if
we replace the r.h.s. random variables by rs. So we can take γ = qb∗ . Using this, and taking β to
be O(φ)), we get the above two inequalities. This concludes the proof of Lemma 7.
Proof (Lemma 9): All that remains is to prove Lemma 9. Before doing this, we prove an auxiliary lemma which characterizes the difference between two binomial distributions in terms of the
difference of the underlying probabilities, i.e., we condition on pxy . This follows from the work of
Roos [Roo01]. For J1 and J2 as defined above:
Lemma 10 (Binomial: CDF perturbation). For any γ ∈ (0, 1), we have
√ √
2 2e k + 2
β.
P̄ [J1 ≤ kγ|pxy ] − P̄ [J2 ≤ kγ|pxy ] ≤ p
(pxy + β)(1 − pxy − β)
43
Proof: It holds that
P̄ [J1 ≤ kγ|pxy ] − P̄ [J2 ≤ kγ|pxy ] ≤ kBin(k, pxy ) − Bin(k, pxy + β)k1
p
√
θ(pxy + β)
≤ e
2 , if θ(pxy + β) < 1
p
1 − θ(pxy + β)
2
β (k+2)
, and the above inequality comes from [Roo01, (15)], by
where θ(pxy + β) = 2(pxy +β)(1−p
xy −β)
setting s = 0 there, and choosing
to simply be the binomial
q the Poisson-Binomial distribution
p
(pxy +β)(1−pxy −β)
distribution Bin(k, pxy ). If β ≤
, then 1 − θ(pxy + β) ≥ 0.5. In this case, we
2(k+2)
have
√ √
2 2e k + 2
P̄ [J1 ≤ kγ|pxy ] − P̄ [J2 ≤ kγ|pxy ] ≤ p
β.
(pxy + β)(1 − pxy − β)
q
xy −β)
, then since the difference between two probabilities
On the other hand, if β > (pxy +β)(1−p
2(k+2)
is upper bounded by 2, the following upper bound holds trivially
√ √
2 2 k+2
P̄ [J1 ≤ kγ|pxy ] − P̄ [J2 ≤ kγ|pxy ] ≤ p
β.
(pxy + β)(1 − pxy + β)
This concludes the proof.
√
We cannot directly apply Lemma 10 to prove Lemma 9 since the k + 2 factor on the upper
bound is too loose for our purposes. Instead, we employ a more careful argument that splits the
domain of the underlying MSC
random variables.
q
log k log m
, m . Now, consider the following partition of the domain
First, recall that α = max
k
of pxy ; we will choose Cf below.
√ log k
)
(0) (2c14
k
, low substitution regime for small k
I1 = pxy , pxy
√ k
q
(2c14 log
)
(2c
α)
k
14
, pxy
I2 = pxy
, low substitution regime for large k; (empty if α = logk k )
(2c14 α)
I3 = pxy
, 0.5 , high substitution regime
Now, we proceed by bounding the above difference in each of these intervals.
Low substitution regime, small k (p ∈ I1 )
√
In this case, we use the fact that Lemma 10 guarantees that the binomial distributions are O(β k)
apart. That is, there exists a constant c015 > 0 such that
√
P̄ [pxy ∈ I1 ] Ē P̄ [J1 ≤ kγ] − P̄ [J2 ≤ kγ] pxy ∈ I1 ≤ P̄ [pxy ∈ I1 ] c015 β k + 2
p
≤ c015 β log k,
44
where the last step follows after appropriately increasing the constant c015 . This follows from the
definition of a quantile.
Low substitution regime, large k (p ∈ I2 )
q
log k
Notice that if α =
, then this interval is empty. In the case that it is not, there exists a
k
00
constant c15 > 0 such that
P̄ [pxy ∈ I2 ] E P̄ [J1 ≤ kγ] − P̄ [J2 ≤ kγ] pxy ∈ I2 ≤ P̄[pxy ∈ I2 ]
(a)
(b)
≤ c0015 α ≤ c0015
log m
.
m
(a) follows from (11), and (b) follows from the definition of α.
High substitution regime (p ∈ I3 )
(c α)
In this case observe that, since γ < pxy14 (i.e., we are looking at a left tail below the mean), we can
apply Chernoff’s bound (see e.g. [MR95]) on each of the two terms in the difference individually.
In fact, depending on how large we want C2 to be, we can choose c14 > 0 so that the following
inequality holds:
−C2
P̄ [pxy ∈ I3 ] E P̄ [J1 ≤ kγ] − P̄ [J2 ≤ kγ] pxy ∈ I3 ≤ c000
,
15 k
for some c000
15 > 0.
Putting the bounds in the above three regimes together, we see that there is a constant c07 > 0
(that does not depend on f, m, k) such that
p
c07
log m
−C2
β log k +
+k
.
P̄ [J1 ≤ kγ] − P̄ [J2 ≤ kγ] ≤
3
m
that the following inThe lemma follows by observing that there exists
a constant c007 > 0 such √
√
−1
−C2
equalities respectively imply that m log m ≤ β log k, and that k
≤ β log k.
1
1
√
m ≥ c7 √
log
,
β log k
β log k
1/C2
1
k≥
.
β
This concludes the proof of Lemma 9.
E.2.2
Proof of Lemma 8
In this section, we prove Lemma 8, which is the final piece needed to complete the proof of
Proposition 8.
45
Proof (Lemma 8): Notice that, from Proposition 7 (on which P̄ is conditioning), we know that
(0)
qb∗ ≤ r13 + c5 φ + c5 α. This implies the second inequality of the lemma, i.e., there exists a constant
r13 ≤ qb∗ ] ≤ c006 α by (11).
c006 > 0 such that P̄ [b
To see the first implication of the lemma, we reason as follows. First we make a few observations:
(c α)
1. Again, from Proposition 7, qb∗ ≥ qb133
definitions.
(0)
≥ r13 − c5 φ; the last equality follows from the
b 12 ) and, conditioned on δ12 , qb12 is distributed as Bin(k, q12 ).
2. By definition q12 = p(δ12 + ∆
b 12 ), r12 = p(δ12 + ∆12 ) and by Proposition 6, it follows that there
3. Because q12 = p(δ12 + ∆
(0)
(0)
is c16 > 0 such that the event {r12 ≤ r13 − c5 φ − c16 φ} implies the event {q12 ≤ r13 − c5 φ}
under P̄.
(0)
4. The event E12|3 is equivalent to the condition that r12 = p(δ12 + ∆12 ) ≤ p(µ13 + ∆13 ) = r13 .
We use these facts in the following chain of inequalities, which will lead us to a lower bound on
the desired quantity:
i
(a) h
(0)
P̄ qb12 ≤ qb∗ E12|3 ≥ P̄ qb12 ≤ r13 − c5 φ E12|3
i
h
(b)
(0)
(0)
≥ P̄ Bin(k, q12 ) ≤ r13 − c5 φ q12 ≤ r13 − c5 φ, E12|3
i
h
(0)
× P̄ q12 ≤ r13 − c5 φ E12|3
h
i
(c)
(0)
(0)
≥ P̄ Bin(k, q12 ) ≤ r13 − c5 φ q12 ≤ r13 − c5 φ, E12|3
i
h
(0)
(0)
× P̄ r12 ≤ r13 − c5 φ − c16 φ r12 ≤ r13
(d)
≥ Cj0 ,
for some constant Cj0 > 0, where (a) follows from Observation 1 above. Inequality (b) follows
(0)
after conditioning on the event that q12 ≤ r13 − c5 φ, and (c) follows from Observations 3 and
4 above. Finally, the inequality (d) follows from the Berry-Esséen theorem (see e.g. [Dur96]),
which gives a constant lower bound on the first line, and (11) together with the assumption that
φ p(3f /4) and the fact that the probability of E12|3 is lower bounded by p(3f /4), which gives a
constant lower bound on the second line.
E.3
Proof of Proposition 9
We first prove the following lemma:
46
Lemma 11 (Variance upper bound). There is a constant c08 > 0 such that for x, y ∈ X ,
Var(b
sxy ) ≤
c08
P̄ E12|3 + φ + α
|MQ2 |
where recall that E12|3 is the event that there is a coalescence in the internal branch.
Proof (Lemma 11): We begin by observing that the variance of sb12 is bounded from above by
1
P̄ [b
q12 ≤ qb∗ ]. So we devote the rest of the proof to controlling this probability.
|MQ2 |
(0)
First, observe that by Proposition 7, qb∗ ≤ r13 + c5 φ + c5 α. Also, observe that conditioned
on
b
the random distance δ12 , qb12 is distributed according to Bin(k, q12 ), where q12 = p δ12 + ∆12 .
Finally, we observe that from Lemma 6, we can conclude that
h
i
(0)
(0)
P̄ qb12 ≤ r13 + c5 φ + c5 α q12 > r13 + c5 φ + c5 α ≤ c03 α.
Note that while this looks different from what Lemma 6 guarantees, it follows given the memoryless property of the exponential distribution. Therefore, we can bound P̄ [b
q12 ≤ qb∗ ] as follows
h
i
(0)
P̄ [b
q12 ≤ qb∗ ] ≤ P̄ qb12 ≤ r13 + c5 φ + c5 α
h
i
(0)
≤ P̄ q12 ≤ r13 + c5 φ + c5 α +
i
h
(0)
(0)
P̄ qb12 ≤ r13 + c5 φ + c5 α q12 > r13 + c5 φ + c5 α
h
i
(0)
≤ P̄ q12 ≤ r13 + c5 φ + c5 α + c03 α.
(29)
From (11), it follows that there is a constant c008 > 0 such that
h
i
(0)
(0)
(0)
00
P̄ q12 ≤ r13 + c5 φ + c5 α ≤ c8 r13 + c5 φ + c5 α − q12 .
(30)
Moreover,
(0)
r13
(0)
q12
(0)
δ13
(0)
δ12
+ ∆13 − p
(0)
(0)
≤ p δ13 + ∆13 − p δ12 + ∆12 + c000
8φ
≤ c000
8 P̄ E12|3 + φ .
The last inequality follows from the fact that P̄ E12|3 is of the order of the length of the internal
branch; and so is the difference on the second line. Notice that this along with (29) and (30)
imply that there is a constant c08 > 0 (after changing it appropriately) such that P̄ [b
q12 ≤ qb∗ ] ≤
0
c8 P̄ E12|3 + φ + α , and concludes the proof.
−
=p
b 12
+∆
Proof (Proposition 9): Recall that we are restricting our attention to a particular triple of leaves
X = {1, 2, 3} which has a topology 12|3 with respect to the true species tree S. In this case, we
47
know that an error in the quantile test implies that either sb13 > sb12 or sb23 > sb12 . Therefore, we can
control the probability that the algorithm makes an error in correctly identifying the topology of
the triple at hand as follows.
P [error] ≤ P̄ [b
s13 ≥ sb12 ] + P̄ [b
s23 ≥ sb12 ]
s12 − s13
s12 − s13
+ P̄ s12 − sb12 ≥
≤ P̄ sb13 − s13 ≥
2
2
s12 − s23
s12 − s23
+ P̄ sb23 − s23 ≥
+ P̄ s12 − sb12 ≥
2
2
, I1 + I2 + I3 + I4
(31)
We will now use concentration inequalities to control each of the above terms.
Consider I2 first. We need two ingredients for invoking Bernstein’s inequality: (1) an effective
13
lower bound on the “gap” s12 −s
, and (2) an effective upper bound on the variance of the random
2
variable sb12 . We will use Proposition 8 for (1), and Lemma 11 for (2). That is,
!
2
13
0.5 s12 −s
s12 − s13 (a)
2
≤ exp −
I2 = P̄ s12 − sb12 ≥
2
Var(b
s12 ) + 16 (s12 − s13 )
!
2
(b)
|MQ2 | c6 P̄ E12|3
≤ exp − 0
c8 α + φ + P̄ E12|3 + c66 P̄ E12|3
(c)
|MQ2 | p(3f /4)2
≤ exp −
,
(32)
c8 (p(3f /4) + α)
where (a) follows from Bernstein’s inequality (see e.g. [BLM13]); (b) follows from the lower
bound on s12 − s13 provided by Proposition 8 and the
s12 ) provided by
upper bound on Var(b
Lemma 11; and (c) follows from the fact that P̄ E12|3 is bounded from below by p(3f /4). We
have absorbed all constants into c8 .
We can similarly control I1 , I3 ,and I4 . Putting all this back in (31) concludes the proof.
48
| 10math.ST
|
1
Spectral Efficiency Optimization For Millimeter
Wave Multi-User MIMO Systems
arXiv:1801.07560v1 [cs.IT] 23 Jan 2018
Qingjiang Shi and Mingyi Hong
Abstract—As a key enabling technology for 5G wireless, millimeter wave (mmWave) communication motivates the utilization
of large-scale antenna arrays for achieving highly directional
beamforming. However, the high cost and power consumption of
RF chains stands in the way of adoption of the optimal fullydigital precoding in large-array systems. To reduce the number
of RF chains while still maintaining the spatial multiplexing gain
of large-array, hybrid precoding architecture has been proposed
for mmWave systems and received considerable interest in both
industry and academia. However, the optimal hybrid precoding
design has not been fully understood, especially for the multi-user
MIMO case. This paper is the first work that directly addresses
the nonconvex hybrid precoding problem of mmWave multi-user
MIMO systems (without any approximation) by using penalty
dual decomposition (PDD) method. The proposed PDD method
have a guaranteed convergence to KKT solutions of the hybrid
precoding problem under a mild assumption. Simulation results
show that, even when both the transmitter and the receivers are
equipped with the fewest RF chains that are required to support
multi-stream transmission, hybrid precoding can still approach
the performance of fully-digital precoding in both the infinite
resolution phase shifter case and the finite resolution phase shifter
case with several bits quantization.
I. I NTRODUCTION
The frequency bandwidth scarcity has motivated the exploration of the underutilized millimeter wave (mmWave) frequency spectrum for future broadband cellular communication
networks [1]–[5]. The shorter wavelength of the mmWave
frequencies enables more antenna components to be packed
in the same physical dimension, which allows for largescale spatial multiplexing and highly directional beamforming.
For massive multiple-input-multiple-output (MIMO) mmWave
systems, the conventional fully-digital (FD) precoding scheme,
which requires one radio frequency (RF) chain per antenna
element, is not viable due to the high cost and the high power
consumption of RF chain components in high frequencies.
To address the challenge of this hardware limitation while
exploiting the multiplexing gain of MIMO, hybrid precoding
architectures have been proposed for mmWave systems and
widely investigated in the literature [6]–[19].
This work was supported in part by NSFC under Grant 61671411,
61631020, 61374020, 61611130211, and by Zhejiang Provincial NSF of China
under Grant LR15F010002. The work of M. Hong is supported in part by NSF
under Grant CCF-1526078 and in part by AFOSR under Grant 15RT0767.
Q. Shi is with the College of EIE, Nanjing University of Aeronautics and
Astronautics, Nanjing 210016, China. Email: [email protected].
M. Hong is with the Dept. of ECE, University of Minnesota, Minneapolis,
MN 55455, USA. Email: [email protected].
This work was done when both authors worked at the Dept. of Industrial
and Manufacturing Systems Engineering, Iowa State University, IA 50011,
USA.
The key idea of hybrid precoding is using a linear network
of variable phase shifters in the RF domain in addition to
the baseband digital precoding. Such a scheme was firstly
known as soft antenna selection (SAS) which was proposed
in [20] to improve the performance of the traditional antenna
selection scheme. It was shown in [20] that the SAS method
can even achieve the optimal performance of a fully-digitally
precoded single-stream MIMO system when each end is
equipped with two or more RF chains. Recently, SAS is reintroduced to mmWave system design under the name of
hybrid precoding/beamforming [2], [6], which has received
significant interest in recent literature on large-scale antenna
array systems.
The pioneering work on hybrid precoding for mmWave
systems [6] investigated hybrid precoding methods for a pointto-point mmWave MIMO system. It was shown in [6] that
the spectral efficiency maximization problem for mmWave
MIMO systems can be approximately solved by addressing
a matrix approximation or reconstruction problem, i.e., minimizing the Frobenius norm of the difference between the
optimal FD precoder and the hybrid precoder. By exploiting
the sparse nature of mmWave channels, the matrix approximation problem is reformulated as a compressive-sensing-like
problem which is addressed by using a modified orthogonal
matching pursuit (OMP) algorithm. The OMP method in [6]
can achieve good performance when hundreds of antennas are
used at both ends of transceiver and/or the number of RF
chains is strictly greater than the number of data streams.
But there is still a significant performance gap between the
hybrid precoding method proposed in [6] and the optimal FD
precoding method especially when the number of RF chains
is equal to the number of data streams. Hence, several works
have devoted to reducing this performance gap. The authors
of [7] proposed using alternating optimization to approximate
the solution of the matrix approximation problem. To deal
with the difficulty arising from the unit-modulus constraints,
[7] used manifold optimization to solve the analog (or RF)
precoder design problem with fixed digital precoder. Other
works that are based on matrix approximation can be found
in [8]–[10]. Together with [6], [7], all the above methods
don’t directly take into consideration the power constraint in
the optimization, they inevitably incur some performance loss
as compared to the optimal hybrid precoding performance.
Differently from the matrix approximation methods [6]–[10],
the work [15] proposed another hybrid precoding algorithm
that can approximately solve the spectral efficiency maximization problem of mmWave MIMO systems. First, by removing
the coupling among the hybrid precoder and decoders (or
2
combiners) using some approximations under large-antenna
array, the analog precoder is designed independently. Then,
given the analog precoder, the digital precoder, the analog
combiner, and the digital combiner are sequentially designed.
Simulations confirm that the hybrid precoding method in
[15] can achieve a performance that is very close to the FD
precoding performance, though it is still a heuristic algorithm.
While the above works considered hybrid precoding for
single-user MIMO systems, there are a few works on hybrid
precoding for spectral efficiency optimization of multi-user
MIMO/MISO systems [11], [12], [15]–[18]. In [11], the authors developed a low-complexity two-stage hybrid precoding
algorithm for downlink mmWave single-stream multi-user
MIMO systems with limited feedback channel. In the first
stage, the analog precoder at the base station (BS) and the
analog combiners at the users are jointly designed to maximize the desired signal power of each user while neglecting
the resulting interference among users. In the second stage,
the BS digital precoder is designed to manage the multiuser interference. The authors of [12] proposed equal-gaintransmission-based analog beamforming combined with block
diagonalization (BD) based digital precoding for generic channel setup. By exploiting the knowledge of angle of departure,
the work [13] proposed iterative matrix decomposition based
BD for mmWave multiuser MIMO systems. In [14], the
authors first proposed a low complexity mmWave channel
estimation algorithm for multiuser mmWave systems with
single RF chain at each user and then designed a hybrid analog
precoding and digital zero-forcing precoding scheme based
on the channel estimations. In [15], assuming perfect channel
state information, the authors proposed an iterative hybrid
precoding algorithm for multi-user MISO systems, which
iterates between the design of zero-forcing (ZF) precoder
and the analog precoder. The works [16], [17] proposed low
complexity hybrid precoding schemes for multi-user MISO
systems by performing phase-only maximum ratio combining
(MRC) or channel matched filter in the analog domain and
ZF beamforming in the digital domain based on the effective
channel. Extending the hybrid precoding methods in [16],
[17], the work [18] proposed similar low complexity hybrid
precoding algorithms for multi-user MIMO systems by designing the analog and digital precoders sequentially, where
digital precoders are separated to pre-filters and post-filters.
First, analog precoders and decoders are designed based on
a per-user channel matching criterion. Then digital pre-filters
are applied at both the transmitter side and the receiver side.
Finally, optimal digital post-filters are derived based on four
linear transmit strategies, including ZF [21], MMSE [22],
block diagonalization (BD) [23], regularized BD (RBD) [24],
followed by optimal power allocation via the water-filling
algorithm. The simulation results in [18] show that the RBD
method performs the best among various transmit strategies.
Differently from the above works focusing on the downlink
systems, the authors of [19] investigated the hybrid beamforming design problems for a mmWave massive multicell
MIMO uplink transmission system, tackling both the intracell
and inter-cell interference under the hardware constraints.
For tractability, most of the above works typically assume
infinite resolution phase shifters used for analog precoders,
although it is very expensive to realize accurate phase shifters
[25], [26]. Usually, low resolution but less expensive phase
shifters are used in practice, resulting in spectral efficiency
optimization problems with discrete unit modulus constraints
[6], [15], [17]. Since it is impossible to directly solve the
discrete optimization problem, the most straightforward way
to design analog precoder with finite resolution phase shifters
is to address the spectral efficiency problem assuming infinite
resolution first and then to quantize analog precoder using
a finite set of available phases [17]. However, this approach
could be ineffective for the case of very low resolution phase
shifters. An alternative way to approximately addressing the
discrete spectral efficiency problem was proposed in [15],
where the discrete unit modulus constraints are directly taken
into consideration in the optimization of analog precoder.
To the best of our knowledge, all the existing hybrid
precoding methods are heuristic and there is no work that tries
to directly solve the spectral efficiency optimization problems
of various MIMO systems under hybrid precoding setups from
the perspective of mathematical optimization. While some
of the existing hybrid precoding methods can achieve very
high spectral efficiency when the number of RF chains is
strictly greater than the number of symbols and even the same
performance as the fully-digital precoding method when the
number of RF chains is no less than twice the number of
streams, there is in some cases a significant gap between
the achievable rate of the existing methods and the theoretical maximum rate. Hence, an interesting and fundamental
question is: if the spectral efficiency optimization problem
is directly addressed from a perspective of mathematical
optimization, is it possible for the hybrid precoding method
to obtain a better system rate and how close it is to the fullydigital precoding performance when the minimum number of
RF chains are used. To answer this fundamental question,
we propose an iterative hybrid precoding algorithm for multistream multi-user MIMO systems, aiming to directly solve the
spectral efficiency optimization problem. Apparently, the main
difficulty of the spectral efficiency optimization problem arises
from the unit modulus constraints and the coupling among
the analog precoder and the digital precoders. To address
these difficulties, we apply penalty dual decomposition (PDD)
method [27], [28] to the spectral efficiency optimization problem by first penalizing and dualizing the coupling constraint
into the objective, and then iteratively solving the augmented
Lagrangian problem using block successive upper-bound minimization (BSUM) method [29] which can naturally exploit
the special structure of the unit modulus constraints. The PDD
method has guaranteed convergence to KKT solutions of the
spectral efficiency optimization problem. Moreover, it can be
straightforwardly extended to the finite resolution phase shifter
case. Our simulation results show that, even when both the
transmitter and the receivers are equipped with the fewest RF
chains that are required to support multi-stream transmission,
hybrid precoding can still approach the performance of fullydigital precoding in both the infinite resolution phase shifter
case and the finite resolution phase shifter case with several
bits quantization.
3
Notations: scalars are denoted by lower-case letters, boldface lower-case letters are used for vectors, and bold-face
upper-case letters for matrices. For a matrix A, AT , AH , A† ,
Tr(A) and det(A) denote its transpose, conjugate transpose,
pseudo-inverse, trace, and determinant, respectively. I denotes
an identity matrix whose dimension will be clear from the
context. |x|, ℜe {x} and x∗ are the absolute value, real part
and conjugate of a complex scalar x, respectively, while ∥x∥
and ∥X∥ denote the Euclidean norm and the Frobenius norm
of a complex vector x and a complex matrix X, respectively.
∥x∥∞ denotes the infinity norm. For a matrix X, we use
[X]ij or X(i, j) to denote its (i, j)-th entry. Particularly,
we use X ∈ M to denote that each entry of X has a
unit modulus constraint, i.e., |Xij | = 1. The distribution of
a circularly symmetric complex Gaussian (CSCG) random
vector variable with mean µ and covariance matrix C is
denoted by CN (µ, C), and ‘∼’ stands for ‘ distributed as’.
Cm×n denotes the space of m × n complex matrices and Rn
denotes the n-dimensional real vector space.
the final processed signal, given by
H
ŝk = UH
BBk URFk yk
H
= UH
BBk URFk Hk VRF VBBk sk
+
K
∑
Consider a narrow band single-cell mmWave downlink
multi-user multi-stream MIMO system, where a base station
(BS) equipped with N antennas and NRF (≤ N ) transmit
RF chains sends signals to K > 1 users, each equipped with
M > 1 antennas and MRF (≤ M ) receive RF chains. Let
d > 1 denote the number of streams intended for each receiver.
For successful symbol detection at each user, it is assumed that
d ≤ min(NRF /K, MRF ). Furthermore, to reduce hardware
complexity, we consider a hybrid digital and analog precoding
architecture for both BS and users (see Fig. 1 of [15] for an
illustration of the hybrid precoding architecture).
In the multi-user hybrid precoding scheme, the BS first
processes the data streams digitally at the baseband using
digital precoders, and then up-converts the digitally processed
signals to the carrier frequency through RF chains, followed
by an analog precoder which is implemented by analog phase
shifters. Let VRF ∈ CN ×NRF denote the BS analog precoder,
and VBBk ∈ CNRF ×d denote the digital precoder for user
k’s data stream sk ∈ Cd×1 . Mathematically, the BS transmit
signal after hybrid precoding is expressed as
x = VRF
K
∑
VBBk sk
(1)
k=1
and the received signal at user k is given by
yk = Hk x + nk
(2)
where Hk denotes the channel between the BS and user k,
and nk denotes the additive white Gaussian noise (AWGN)
with zero mean and variance σ 2 .
At the receiver side, user k first processes the received signal
by using an analog combiner URFk ∈ CM ×MRF , and then
down-converts the signals to the baseband through RF chains,
followed by a digital combiner UBBk ∈ CMRF ×d to obtain
H
H
H
UH
BBk URFk Hk VRF VBBj sj + UBBk URFk nk .
j̸=k
Assuming Gaussian signaling for the data streams each
with zero mean and unit variance, and that the noises nk ’s
and the data streams sk ’s are independent of each other, the
overall system spectrum efficiency maximization problem can
be formulated as1
(
K
∑
max
log det I + UH
RFk Hk VRF VBBk
V,U
k=1
×
s.t.
K
∑
k=1
II. S YSTEM M ODEL A ND P ROBLEM STATEMENT
(3)
−1
H
H
VBB
VRF
HH
k URFk Υ̃k
k
)
2
∥VRF VBBk ∥ ≤ P,
|VRF (i, j)|=1, ∀i=1, 2, . . . , N, j=1, 2, . . . , NRF ,
|URFk (i, j)|=1, ∀i=1, 2, . . . , M, j=1, 2, . . . , MRF , ∀k.
(4)
where the first constraint is the BS power constraint with
power budget P ; the last two sets of unit modulus constraints
are due to the fact that both the analog precoder and the analog
combiners are implemented using low-cost phase shifters; and
the matrix
(
∑
H
Υ̃k , URFk σ 2 I +
Hk VRF VBBj
j̸=k
)
H
H
H
× VBB
V
H
URFk .
RF
k
j
(5)
is the so-called interference-plus-noise covariance matrix
associated with user k.
It was shown in [15], [18] that, when the number of transmit
RF chains NRF are greater than or equal to twice the total
number of streams, i.e., NRF ≥ 2Kd, the performance of the
fully-digital precoding scheme can be perfectly achieved by
the hybrid precoding scheme. This implies that problem (4)
can be addressed by first solving the fully-digital precoding
problem and then constructing the hybrid precoders VBBk
and VRF based on the fully-digital precoder. It is noted
that, the fully-digital precoding problem can be addressed
using the well-known WMMSE algorithm [30], [31] and the
corresponding hybrid precoders can be found in closed-form;
see [15, Prop. 2]. However, for the general case, the hybrid
precoding problem (4) is extremely hard to solve due to not
only the coupling of the digital/analog precoders but also
the unit modulus constraints. To the best of our knowledge,
all the existing works that deal with these difficulties are
1 The system rate expression does not include the digital combiners explicitly since it is well-known [30] that the optimal digital combiners (i.e., MMSE
receivers) can achieve the maximum system rate. In addition, for convenience,
we use V to denote the set of variables VBBk ’s and VRF ; similarly for U
and other variables later.
4
heuristic algorithms and thus inevitably incurs performance
loss. Although most of heuristic algorithms for single-user
MIMO scenarios perform well, it is still not known how
close their performance is to the optimal one. Hence, it is
important to devise an algorithm that can solve problem
(4) with some theoretical guarantees (e.g., achieve at least
stationary solutions to problem (4)). In this paper, we propose
using penalty dual decomposition method [28] to address
problem (4). Before proceeding to solving problem (4), let
us first give a brief introduction to the PDD method in what
follows.
III. A BRIEF INTRODUCTION TO PDD METHOD
The PDD method is a general algorithmic framework that
can be applied to the minimization of a nonconvex nonsmooth
function subject to nonconvex coupling constraints. Considering that problem (4) has a differentiable objective function, we
are here only concern about the differentiable case for ease of
understanding of the PDD framework.
Consider the following problem
(P )
min f (x)
x
s.t. h(x) = 0,
(6)
x ∈ X.
where f (x) is a scalar continuously differentiable function and
h(x) ∈ Rp×1 is a vector of p continuously differentiable functions; the feasible set X is the Cartesian product of n closed
sets: X , ∑
X1 × X2 × . . . × Xn with Xi , {xi | gi (xi )≤0} ⊆
n
Rmi and
i=1 mi = m, and accordingly the optimization
variable x ∈ Rm can be decomposed as x = (x1 , x2 , . . . , xn )
with xi ∈ Xi i = 1, 2, . . . , n; gi (xi ) ∈ Rqi is a vector of
differentiable functions.
If no coupling constraints h(x) = 0 exist, the classical
block coordination descent (BCD)-type algorithms [32] can be
applied to decompose problem (P ) into a sequence of smallscale problems. This observation motivates us to dualize the
difficult coupling constraints with appropriate penalty, and use
coordinate-decomposition to perform fast computation, hence
the name penalty dual decomposition method. Specifically,
the PDD method applied to problem (P ) is a double-loop
iterative algorithm. It employs inner iterations to solve to
some accuracy a nonconvex augmented Lagrangian problem
via an inexact or exact block coordinate descent method,
while updating dual variables and a penalty parameter in outer
iterations.
The PDD method is summarized in TABLE I, where the oracle ‘BSUM(Pϱk ,λk , L̃k , z k−1 , ϵk )’ means that, starting from
z k−1 , the BSUM algorithm [29] or its variant randomized
BSUM proposed in [28] for nonconvex constraint cases is
invoked to iteratively solve problem (Pϱk ,λk ), given below
{
(Pϱk ,λk ) min Lk (x) , f (x)
xi ∈Xi
}
1
+ λTk h(x) +
∥h(x)∥2
(7)
2ϱk
where Lk (x) is the augmented Lagrange function with dual
variable λk and penalty parameter ϱk . Further, the BSUM
algorithm utilizes a locally tight upper bound of Lk (x), denoted as L̃k , and it terminates when certain solution accuracy
ϵk is reached. A basic implementation of the BSUM oracle is
presented in Appendix A.
It is worthy noting that, when the penalty parameter ϱk in
(7) is sufficiently small, the term ∥h(x)∥2 or ∥h(x)∥∞ will
be driven to vanish after solving problem (7). That is, in this
case, solving problem (7) yields a solution to problem (P )
satisfying the constraint ∥h(x)∥∞ = 0. However, we have
generally ∥h(x)∥∞ ̸= 0 during iterations. To measure the
violation of the constraint h(x) = 0, we refer to the value of
∥h(x)∥∞ as constraint violation.
Define g(x) , (gi (xi ))i . Then we have the following
theory regarding the convergence the PDD method, which is
a direct result of [28, Theorem 3.1].
Corollary 3.1: Let {xk , ν k } be the sequence generated
by Algorithm 1 for problem (P ), where ν k =(νik )i denotes the Lagrange multipliers associated with the constraints
gi (xi )≤0, ∀i. The stop criterion for the BSUM algorithm
involved in Algorithm 1 is
∥∇x Lk (xk ) + ∇g(xk )T ν k ∥∞ ≤ ϵk ,
∀k
(8)
with ϵk , ηk → 0 as k → ∞. Suppose that x⋆ is a limit
point of the sequence {xk } and Robinson’s condition holds
for problem (P ) at x⋆ . Then x⋆ is a KKT point of problem
(P ), i.e., it satisfies the KKT condition of problem (P ).
A general proof of this theorem can be found in [28].
Note that Robinson’s condition is a constraint qualification
condition which is often used to characterize the first order optimality condition of nonconvex problems [33]–[37].
For the problem at hand, it is equivalent to the commonly
used Mangasarian-Fromovitz constraint qualification (MFCQ)
condition. Further, a simple way to set ηk is to make it
explicitly related to the constraint violation of the last iteration
or the current minimum constraint violation. For example,
we set ηk = 0.9∥h(z k−1 )∥∞ in our simulations later. Furthermore, it is reasonable to terminate the BSUM algorithm
based on the progress of the objective value Lk (xr ), i.e.,
|Lk (xr )−Lk (xr−1 )|
≤ ϵk . Here, the superscript ‘r’ denotes the
|Lk (xr−1 )|
BSUM iterations. In addition, since the penalty term ∥h(x)∥∞
vanishes eventually, a practical choice of the termination
condition for the PDD method is ∥h(xk )∥∞ ≤ ϵO . Here, ϵO
is some prescribed small constant, e.g., ϵO = 1e−6. See more
details in [28].
IV. PDD
METHOD FOR HYBRID PRECODING
This section proposes a PDD-based hybrid precoding
method for spectral efficiency optimization. To tackle the
difficulty arising from the rate function, i.e., the log det(·)
function, we apply the PDD method to an equivalent problem
of (4) instead of directly to (4).
A. Equivalent formulations of (4)
Let us first rewrite problem (4) in the form of (P ). To make
the power constraint easy to handle and remove the coupling
of VRF and VBBk later through penalty [cf. Eq. (13)], we
5
TABLE I
A LGORITHM 1: PDD METHOD FOR PROBLEM (6)
Proof: The proposition can be proven by following [30,
Theorem 3] and [38, Lemma 4.1]. We omit the proof for
brevity.
Problem (12) is now in the form of problem (P ) is more
tractable than problem (10) because the optimization with
respect to Wk ’s is easy (see (17) below). Moreover, according
to the result of Proposition 4.1, it is known that, the KKT
solutions of problem (10) can be obtained by solving problem
(12). Hence, in the following we solve problem (12) by using
the PDD method [27].
0. initialize x0 , ϱ0 > 0, λ0 , and set 0 < c < 1, k = 1
1. repeat
2.
xk = BSUM(Pϱk ,λk , L̃k , xk−1 , ϵk )
3.
4.
if ∥h(xk )∥∞ ≤ ηk
λk+1 = λk +
1
h(xk )
ϱk
5.
ϱk+1 = ϱk
6.
else
7.
λk+1 = λk
8.
ϱk+1 = cϱk
9.
end
10.
k =k+1
11. until some termination criterion is met
B. PDD method for (12)
introduce an auxiliary variable Xk such Xk = VRF VBBk
with the definition of
∑
H
σ 2 I +
Υk , UH
Hk Xj XH
URFk . (9)
RFk
j Hk
j̸=k
Now we turn our attention to solving problem (12). It is
readily known that the key to the PDD method is the inner
iterations for solving augmented Lagrangian problems. Hence,
our main efforts are devoted to developing BSUM/BCD algorithm for the augmented Lagrangian problem associated with
(12). Specifically, by introducing dual variables Yk ’s for the
coupling constraints Xk = VRF VBBk , k = 1, 2, . . . , K, we
define the augmented Lagrangian problem of (12) as follows
Then we can recast problem (4) as
max
V,U,X
s.t.
K
∑
k=1
K
∑
k=1
max
(
)
−1
H
log det I + UH
RFk Hk Xk Xk Hk URFk Υk
U,V,W,X
2
Xk = VRF VBBk , ∀k = 1, 2, . . . , K,
|VRF (i, j)| = 1, ∀i = 1, 2, . . . , N, j = 1, 2, . . . , NRF ,
|URFk (i, j)|=1, ∀i=1, 2, . . . , M, j=1, 2, . . . , MRF , ∀k.
(10)
Furthermore, to address the difficulty arising from the
log det(·) function, based on the theory of the well-known
WMMSE method [30], we write (10) in an equivalent
WMMSE form, which is shown in Proposition 4.1 with the
definition of mean-square-error (MSE) matrix
H
H
H
H
Ek (U, X) , (I−UH
BBk URFk Hk Xk )(I−UBB URF Hk Xk )
+ UH
BBk Υk UBBk .
(11)
K
∑
s.t.
max
s.t.
K
∑
k=1
k=1
k=1
∥Xk ∥ ≤ P,
Xk = VRF VBBk , ∀k,
|VRF (i, j)| = 1, ∀i, j,
|URFk (i, j)| = 1, ∀i, j, k.
K
∑
1
2
∥Xk −VRF VBBk + ρYk ∥
2ρ
k=1
2
|URFk (i, j)| = 1, ∀i, j, k.
It is seen that the constraints of the above problem are
separable. Hence, we use BCD-type algorithm to address (13)
with the block variables {UBB , W}, VBB , URF , VRF , and
X, i.e., each time we optimize one block among them while
fixing the others. This leads to the following six kinds of
subproblems in each BCD iteration.
1) The subproblem w.r.t UBBk : The variable UBBk is updated by minimizing the term Tr(Wk Ek (U, X)), i.e., solving
where
(12)
Moreover, if {U, V, W, X} is a KKT point of problem (12),
{U, V, X} is a KKT point of problem (10), and further
{U, V} is a KKT point of problem (4).
(13)
∥Xk ∥ ≤ P,
H
min Tr(Wk UH
BBk URFk Ak URFk UBBk )
{
}
H
− 2ℜe Tr(Wk UH
BBk URFk Hk Xk )
UBBk
(log det(Wk ) − Tr(Wk Ek (U, X)) + d)
2
(log det(Wk ) − Tr(Wk Ek (U, X)) + d)
|VRF (i, j)| = 1, ∀i, j,
Proposition 4.1: Problem (10) is equivalent to
U,V,W,X
k=1
−
∥Xk ∥ ≤ P,
K
∑
K
∑
Ak , σk2 I+
K
∑
H
Hk Xj XH
j Hk .
(14)
(15)
j=1
Since each matrix Wk is positive definite, the optimal solution
UBBk to problem (14) takes the form
(
)† ( H
)
UBBk = UH
URFk Hk Xk .
RFk Ak URFk
(16)
The reason why we use pseudo-inverse here is that the matrix
UH
RFk Ak URFk may be rank-deficient during iterations.
6
2) The subproblem w.r.t Wk : Consequently, given the
optimal UBBk in (16), the optimal Wk can be expressed as
Wk = arg min log det(W̄k ) − Tr(W̄k Ek (U, X))
equivalently,
K ∑
N
∑
k=1 i=1
W̄
= Ek (U, X)−1
(
( H
))−1
= I − UH
.
BBk URFk Hk Xk
(17)
where the last equality is obtained by plugging (16) into
Ek (U, X). Moreover, it can be verified that the bracketed
matrix in the last equality is positive definite and so is the
matrix Wk .
3) The subproblem w.r.t VBBk : The variable VBBk appears only in the summand of the second term of the objective
2
of (13), i.e., ∥Xk − VRF VBBk + ρYk ∥ . Thus, the subproblem with respect to VBBk is given by
2
min ∥VRF VBBk − Zk ∥
VBBk
(18)
where Zk , Xk + ρYk . Solving the above unconstrained
quadratic optimization problem, we obtain the optimal VBBk
as follows
VBBk = (VRF )† Zk .
(19)
4) The subproblem w.r.t X: The variables Xk ’s appear in
both the first and the second terms of the objective of (13).
Thus, the subproblem with respect to X is given by
min
X
s.t.
K
∑
K
∑
1
2
Tr (Wk Ek (U, X))+
∥Xk+ρYk −VRF VBBk ∥
2ρ
k=1
K
∑
k=1
k=1
2
∥Xk ∥ ≤ P.
For notational convenience, we define
)
1
H
H
U
U
W
U
U
H
+ I,
HH
RF
BB
j
j
BBj RFj
j
j
j
2ρ
j=1
(
)
1 1
H
, Hk URFk UBBk Wk +
VRFk VBBk − Yk .
2 ρ
Aρ ,
Bρ,k
K (
∑
Then, with some appropriate rearrange, problem (20) can be
equivalently written as
min
X
s.t.
K
∑
(
{
})
H
Tr(XH
k Aρ Xk ) − 2ℜe Tr(Xk Bρ,k )
k=1
K
∑
k=1
(21)
2
∥Xk ∥ ≤ P.
Further, by introducing a Lagrange multiplier µ ≥ 0 to the
power constraint, we can express the optimal Xk as
−1
Xk = (Aρ + µI)
Bρ,k
(22)
where the optimal multiplier µ can be easily found by using
Bisection method such that
K
(
)
∑
−2
Tr BH
Bρ,k = P,
ρ,k (Aρ + µI)
k=1
H
where bk,i = [Uρ Bρ,k BH
ρ,k Uρ ]ii , Uρ is a unitary matrix
consisting of the eigenvectors of Aρ and ai ’s are the corresponding eigenvalues, i.e., Uρ diag {a1 , a2 , . . . , aN } UH
ρ is
the eigenvalue decomposition of Aρ .
5) The subproblem w.r.t VRF : The variable VRF only
appears in the summand of the second term of the objective
of (13). Thus, the subproblem with respect to VRF is given
by
min
VRF ∈M
K
∑
k=1
2
∥Xk − VRF VBBk + ρYk ∥
(23)
By appropriate rearrange, the above problem can be equivalently formulated as follows
{
}
H
H
min Tr(VRF
VRF CVRF ) − 2ℜe Tr(VRF
BVRF )
VRF ∈M
(24)
∑K
H
BVRF
,
CVRF
,
k=1 VBBk VBBk ,
H
k=1 (Xk + ρYk ) VBBk . Since the unit modulus constraints
are separable, we can use one-iteration BCD-type algorithm
to recursively solve problem (24). See the details in Appendix
B.
6) The subproblem w.r.t URFk : The variable URFk appears
only in the term Ek (U, X). Thus, the subproblem with respect
to URFk is given by
where
∑K
min
(20)
bk,i
=P
(ai + µ)2
URFk ∈M
Tr(Wk Ek (U, X))
(25)
By appropriate rearrange, the above problem can be equivalently formulated as follows
(
)
min Tr UH
RFk Ak URFk CURFk
URFk ∈M
{
}
− 2ℜe Tr(UH
(26)
RFk BURFk )
where Ak is defined in (15), CURFk , UBBk Wk UH
BBk ,
BURFk , Hk Xk Wk UH
.
Again,
we
can
use
one-iteration
BBk
BCD-type algorithm to address problem (26). See the details
in Appendix B.
The BCD-type algorithm for (13) is summarized in TABLE
II. Note that, all updates can be implemented in parallel for
different users. For instance, once URFk ’s and Xk ’s are given,
each UBBk can be updated independently. By using parallel
implementation, the proposed BCD-type algorithm is scalable
to the number of users. In particular, when N >> M , it can
be shown that the complexity of Algorithm 2 is dominated by
Step 7, which is O(N 3 KI2 ). Here, I2 denotes the maximum
number of iterations required by Algorithm 2.
In each iteration of the PDD method, after running Algorithm 2, we update the dual variable Yk or the penalty
parameter ρ according to the constraint violation condition
maxk ∥Xk − VRF VBBk ∥∞ ≤ ηk (see Step 3 in TABLE I).
That is, if the constraint violation condition is satisfied, we update the dual variables Yk ’s according to Yk ← Yk + ρ1 (Xk −
7
VRF VBBk ), ∀k; otherwise, we increase the penalty parameter
by updating ρ = cρ where 0 < c < 1. The PDD method is
terminated when the feasibility of the penalized constraints
(i.e., Xk = VRF VBBk , ∀k) is approximately achieved; see
the discussion in the end of Section III. In addition, it can
be shown2 that the MFCQ condition is satisfied for problem
(12) at any feasible solution ({DXk }, DVRF , DURF ). Hence,
we can conclude that the PDD method converges to the set
of KKT solutions of problem (4) according to the results of
Proposition 4.1 and Corollary 3.1.
TABLE II
A LGORITHM 2: BCD- TYPE ALGORITHM FOR PROBLEM (13)
0. initialize VRF , VBBk , URFk , UBBk , ∀k, such the constraints
and set Xk = VRF VBBk , ∀k
1. repeat
2.
update UBBk ’s according to (16)
3.
update Wk ’s according to (17)
4.
update VRF by solving (24) using Algorithm 4
5.
update URFk ’s by solving (26) using Algorithm 4
6.
update VBBk ’s according to (19)
7.
update Xk ’s according to (22) using Bisection method
8. until some termination criterion is met
V. M ATRIX -A PPROXIMATION -BASED H YBRID P RECODING
METHOD
The PDD method requires running Algorithm 2 repeatedly.
In this section, we provide a lightweight solution based on
matrix approximation (MAP).
The intuition behind MAP is that [6], the performance
achieved by hybrid precoding is upper bounded by the one
achieved by the fully-digital precoding. Hence, if the optimal
fully-digital precoder/decoder can be approximated well using hybrid precoder/decoder, we can obtain a good spectral
efficiency performance that is close to that of the fullydigital precoding. Mathematically, given the optimal fullydigital precoder Voptk , ∀k, MAP for precoder design3 is to
find a structured matrix Vk = VRF VBBk to approximate
Vopt,k , ∀k, i.e., solving the following matrix approximation
problem
min
VRF ,VBB
∥Vopt − VRF VBB ∥2
s.t. VRF ∈ M
(27)
where Vopt , [Vopt1 Vopt2 . . . VoptK ], VBB ,
[VBB1 VBB2 . . . VBBK ] and M , {VRF | |VRF (i, j)| =
1, ∀i, j}. Note that, for tractability, we here have neglected
the coupling power constraint which will be taken into consideration in the end by scaling VBBk ’s to satisfy the power
constraint. However, even after removing this difficulty, the
problem is still difficult to solve due to the unit modulus
constraint.
2 The MFCQ for this problem is equivalent to verify that 1) the equality constraint gradients are linearly independent, and 2) there exists
({DXk }, DVRF , DURF ) such that DXk − VRF DURF − DRF URF =
∑
H
0, ∀k and K
k=1 Tr(Xk DXk ) < 0. Note that both are easy to be verified.
3 The decoder can be designed in a similar way.
The quadratic nature of the objective function motivates us
to use BCD-type algorithm to address problem (27). Specifically, we divide the variable (VRF , VBB ) into NRF d + 1
blocks, i.e., VBB , VRF (i, j), i = 1, 2, . . . , NRF , j =
1, 2, . . . , d, and successively update each block by solving
(27) for the selected block while fixing the others. Clearly,
given VRF , VBB can be updated in closed-form: VBB =
†
VRF
Vopt . Further, according to Appendix B, each entry of
VRF , VRF (i, j), can be also updated in closed-form with
the other variables fixed. Note that the proposed algorithm for
problem (27) differs significantly from the alternating optimization method proposed in [7], where the analog precoder
VRF is updated as a whole by using manifold optimization
(MO) method [39]. Since the MO method is a gradientdescent-like iterative algorithm, our algorithm is more efficient
than the alternating optimization method in [7] by using
closed-form update of VRF . This will be verified later using
numerical examples (see Fig. 3).
To summarize, the MAP method is a two-phase hybrid
precoding method. In the first phase, we obtain the fullydigital precoder/decoder using the well-known WMMSE algorithm. In the second phase, we run BCD method to find the
hybrid precoders/decoders. Assuming the number of transmit
antennas N is much larger than the number of user receive
antennas M , it can be shown that the complexity of the first
phase is O(N 3 KIwmmse ) where Iwmmse denotes the number
of iterations required by the WMMSE method. Furthermore, as
shown in Appendix B, updating all the entries of VRF once re2
quires complexity of O(N 2 NRF
). Similarly, updating URFk ’s
2
requires complexity of O(KM 2 MRF
). Hence, the complexity
2 2
2
of the second phase is O(Ibcd (N NRF + KM 2 MRF
) where
Ibcd denotes the maximum number of iterations required for
updating precoders/decoders. Therefore, the total complexity
of the MAP method is cubic in the number of BS antennas.
It is lower than that of the PDD method because the latter
requires repeatedly running Algorithm 2 which also has cubic
complexity in the number of BS antennas.
VI. H YBRID P RECODING W ITH F INITE R ESOLUTION
P HASE S HIFTERS
So far, we have assumed that arbitrary resolution phase
shifters are available for realizing analog precoders. However,
the hardware for accurate phase control could be very expensive. Furthermore, infinite resolution phase shifter is not
always practical for large-array systems. Hence, we need to
consider hybrid precoding with finite resolution phase shifters,
i.e., the candidate phases for each phase shifter is finite.
Let F denote the set of finite phases, with |F| = 2b , where
b is the number of bits used to quantize the phases. The
corresponding spectral efficiency optimization problem turns
8
out to be
max
V,U
K
∑
k=1
(
H
log det I + UH
BBk URFk Hk VRF VBBk
×
s.t.
K
∑
k=1
−1
H
H
VBB
VRF
HH
k URFk UBBk Υk
k
)
(28)
2
∥VRF VBBk ∥ ≤ P,
VRF (i, j) ∈ F , ∀i, j
URF (i, j) ∈ F , ∀i, j, k
Due to the combinatorial nature of the phases available for
the phase shifters, problem (28) generally requires exhaustive search which however is computationally prohibitive in
practice. A heuristic way to deal with the constraints of finite
phases is to first addressing the spectral efficiency optimization
problem under the assumption of infinite resolution phase and
then quantize each component of the analog precoders to
the nearest point in the set F. This heuristic method could
work effectively when the number of available phases is large
(i.e., relatively high resolution phase). However, for the low
resolution case (e.g., b ≤ 4), the effect of phase quantization
could be significant. Hence, it is important to devise algorithms
that can incorporate the constraints of finite resolution phases
directly into the optimization procedure.
Fortunately, our PDD algorithm4 proposed above can be
easily adapted to the finite resolution phase case. The modification lies only in Step 4 of Algorithm 4, i.e., modify Step 4
as follows
x = arg
min
X(i,j)∈F
ℜe {b∗ X(i, j)}
The above problem can be globally solved via one-dimensional
exhaustive search. As a result, the modified Algorithm 4
requires complexity of O(I3 mn(mn + 2b )).
Finally, we make a remark on iterations of our algorithms.
All the algorithms proposed for both the infinite resolution
phase case and the finite resolution phase case are iterative algorithms. While the algorithms requires a number of iterations
for achieving convergence, we can terminate them early at the
price of system spectral efficiency performance. It is worth
mentioning that, when the algorithms are terminated early,
we need to scale VBBk ’s such the BS power constraint and
eventually obtain a feasible solution to the spectral efficiency
optimization problem.
VII. S IMULATION RESULTS
This section presents simulation results to illustrate the
performance of the proposed hybrid precoding algorithms.
Since we focus on hybrid precoding for multi-stream multiuser MIMO cases without exploiting the angle information
4 It is worthy mentioning that, as F is a discrete set, the convergence result
in Corollary 3.1 does not apply to the finite resolution case. While numerical
results still show good convergence performance for the PDD method in
the finite resolution case, the convergence issue remains open. However, our
method can serve as a reference point for studying the performance of hybrid
precoding architecture with finite resolution phase shifters in comparison with
fully digital precoding.
of channel knowledge, we first present the simulation results
for MU-MIMO systems and then for SU-MIMO and multiuser MISO systems as special cases. The recent work [18] has
shown that the regularized block diagonalization (RBD)-based
hybrid precoding method performs better than ZF-based, BDbased, and MMSE-based hybrid precoding methods for multiuser MIMO cases. Hence, in the simulations, we compare our
hybrid precoding methods with the RBD method in addition
to the benchmark performance—the fully digital precoding
schemes in terms of the achieved spectral efficiency5 .
As in [15], we use a geometric channel model of L = 15
paths with uniform linear array antenna configurations. Specifically, the channel matrix between the BS and each user is
expressed as
√
L
NM ∑ ℓ
Hk =
αk ar (ϕℓk )at (φℓk )H , ∀k.
(29)
L
ℓ=1
where
∼ CN (0, 1) is the complex gain of the ℓ-th path,
ar (ϕℓk ) and at (φℓk ) are respectively the normalized receive
and transmit array response vector at the azimuth angle of
ϕℓk ∈ [0, 2π) and φℓk ∈ [0, 2π), which are given by
]T
1 [ jπsin(θ)
a(θ) =
1, e
, . . . , ej(N −1)πsin(θ) .
(30)
N
In our simulations, it is assumed that the base station is
equipped with N = 64 antennas and each user with M = 16
antennas. Unless otherwise specified, we set NRF = 8 and
MRF = 4. For the PDD method, the initial penalty parameter
ρ is set to 100/N and the control parameter c is set to 0.8.
Furthermore, we set η0 = ϵ0 = 1e − 3 and ϵk = cϵk−1 .
Moreover, in practical implementation, we set the maximum
number of inner BSUM/BCD iterations of the PDD method6
to 30. The simulation results versus SN R are averaged over
100 channel realizations, where SN R is defined by SN R =
10 log10 ( σP2 ).
αkℓ
A. Convergence performance of the PDD method
In the first set of simulations, we examine the convergence
performance of the proposed methods. To obtain a benchmark
performance, we set K = d = 2 so that we have NRF = 2Kd
and MRF = 2d. For this setup, it is known that the fullydigital precoding performance can be perfectly achieved by
the hybrid precoding in the infinite resolution phase shifter
case [15].
First, we examine the convergence performance of the PDD
method. For 100 randomly generated multi-user channels, we
run the PDD method for each channel with different random
initialization. Note that, different problem instances yield
different spectral efficiency value. To remove this effect and
5 The most costly step of the RBD method is the SVD operations performed
on the user channels (each with size of N ×M ) and the analog precoder (with
size of N × NRF ). Hence, the RBD method has quadratic complexity in the
number of BS antennas, which is lower than the cubic complexity of the PDD
method.
6 When the constraint violation is satisfactory, the maximum iteration
strategy can be used to avoid the possibly slow convergence of the BSUM/BCD algorithm due to large penalty (i.e., 1/ρ is large). From our
numerical experience, this strategy works effectively without sacrificing any
performance.
9
Constraint violation
Objective function
−3
10
−4
−5
5
10
0
2000
4000
6000
Number of iterations
8000
10000
Fig. 3. The convergence performance of BCD (for (27) with some fixed
VBB ) versus MO.
2
10
−2
10
−4
10
−8
10
0
50
100
150
200
250
Number of iterations
300
350
400
−6
0
10
20
40
Number of iterations
0
20
40
Number of iterations
The average convergence behavior of PDD with b = ∞.
−1
0.81
10
0.8
−2
10
Constraint violation
0.79
Objective function
0
0.6
10
10
0.78
0.77
0.76
−3
10
−4
10
−5
10
0.75
−6
10
0.74
0.73
Fig. 2.
0
−6
10
0.965
Fig. 1.
0.5
0.8
0.2
Approximation error
−2
10
0.97
1
1
0
0.99
0.975
1.5
10
0.995
0.98
1.2
0.4
10
0.985
MO
BCD
1.4
−1
1
0.96
1.6
Approximation error
also clearly demonstrate how much percentage of fully-digital
precoding performance the hybrid precoding can achieve,
we normalize the objective value of problem (12) by the
spectral efficiency value of the fully-digital precoding, and
plot them in the left subfigure of Figs. 1 and 2, where the
average convergence behavior of the PDD method is shown
for the case of infinite resolution phase shifters (denoted by
b = ∞ for short) and finite resolution phase shifters with
b = 1, respectively. In terms of the objective value and the
constraint violation, it is observed from the plots that the
PDD method can converge well within 50 iterations for both
cases. Furthermore, it is seen that the hybrid precoding can
achieve the same performance as the fully-digital precoding in
the example of infinite resolution phase shifter case, implying
its excellent convergence performance in achieving possibly
optimal solutions. While for the finite resolution phase shifter
case with b = 1, the hybrid precoding can achieve about
73.7 percent of the fully-digital precoding performance in this
example.
−7
0
20
40
Number of iterations
10
0
20
40
Number of iterations
The average convergence behavior of PDD with b = 1.
Second, we examine the convergence performance of the
MAP method as compared to the alternating optimization
Fig. 4. Ten examples of convergence behavior of BCD (for (27)) with random
initialization. Each curve corresponds to an example of convergence.
method in [7]. Since the two methods have the same way
in the update of VBB but totally different ways in the update
of VRF , we here focus on the comparison of the efficiency
of the update of VRF and thus first compare the convergence
performance of the BCD method and the MO method when
they are applied to problem (27) with fixed VBB . Figure
3 illustrates an average convergence behavior of the two
methods over ten problem instances, where ‘Approximation
error’ denotes the objective value of problem (27). It is seen
that the BCD method requires significantly less iterations
(generally serveral iterations) for convergence than the MO
method, while both can achieve the same convergence result.
Furthermore, Figure 4 shows that, with different initializations,
the MAP method can always achieve global optimality of
problem (27). This will be further verified by the simulation
results later.
B. Hybrid precoding with infinite resolution phase shiters
In the second set of simulations, we demonstrate the spectral efficiency performance of the proposed hybrid precoding
10
Average spectral efficiency (bps/Hz)
35
FD
PDD
MAP
RBD
30
25
20
15
10
−10
−8
−6
−4
−2
0
SNR (dB)
2
4
6
Fig. 5. Spectral efficiency achieved by different methods versus SNR when
NRF = 8, MRF = 4, and K = d = 2.
C. Hybrid precoding with finite resolution phase shiters
In the third set of simulations, we demonstrate the spectral
efficiency performance achieved by the PDD-based hybrid
60
FD
PDD
MAP
RBD
55
Average spectral efficiency (bps/Hz)
methods as compared to the benchmark fully-digital precoding
performance and the existing multi-user hybrid precoding
method—RBD [18] in the infinite resolution phase shifter case.
Figures 5 shows the spectral efficiency performance of various
precoding methods for the case of K = d = 2. In this case,
it is known that the fully-digital precoding performance is
achievable by the hybrid precoding. Figure 5 shows that both
the PDD method and the MAP method can achieve the same
performance as the fully-digital precoding. This again verifies
the excellent convergence performance of the PDD method
and the MAP method (possibly achieve global optimality in
this case). Moreover, it is observed from Fig. 5 that both the
PDD method and the MAP method has better performance
than the RBD method with about 1dB gain.
Figures 6 shows the spectral efficiency performance of
various hybrid precoding methods for the cases when NRF =
Kd < 2Kd. It is seen that, even for the case when
NRF < 2Kd, the PDD method can achieve a performance
that is extremely close to the FD precoding performance.
Furthermore, one can see that, unlike the case of NRF ≥ 2Kd
(where the MAP method has the same performance as the
PDD method), the PDD method has a better performance than
the MAP method in the cases of NRF < 2Kd. Moreover, it
is observed again that both the PDD method and the MAP
method outperform the RBD method, and the performance
gaps increase with the SNR. The reason for this observation
is explained as follows. Recall that both the MAP method
and the RBD method first neglects the power constraint and
then scale the digital precoder VBBk ’s to satisfy the power
constraint. Such a heuristic scaling approach inevitably incurs
performance degradation especially when P is large (i.e., the
SNR is large). Particularly, when SN R = 6, the PDD method
improves the spectral efficiency of the MAP method and the
RBD method by 5 bps/Hz and 8 bps/Hz, respectively.
50
NRF=8, MRF=4, K=4, d=2
45
40
35
30
25
NRF=MRF=6, K=3, d=2
20
15
10
−10
−8
−6
−4
−2
0
SNR (dB)
2
4
6
Fig. 6. Spectral efficiency achieved by different methods versus SNR when
K = 4 and d = 2.
precoding method in the finite resolution phase shifter case.
Since it is not possible to get good matrix approximations
of the fully-digital precoder/decoders in this case, the MAP
method does not work well and thus its simulation result is
not shown here. Furthermore, since the RBD method was
proposed only for the infinite resolution phase shifter case
in [18], we adapt it to the finite resolution phase shifter case
by quantizing the analog precoder/decoders7 followed by the
design of the digital precoder/decorders. For convenience, we
still refer to the modified RBD method in the finite resolution
phase shifter case as ‘RBD’ in the plot. In addition, since
the hybrid precoding problem has much more local maxima
in the finite resolution phase shifter case than in the infinite
resolution phase shifter case, the PDD method could get
easily trapped in some bad point and thus may result in bad
performance. To deal with this issue, we run 20 iterations of
the PDD method assuming infinite resolution phase shifters
to get a good initialization, followed by the PDD method of
the finite resolution phase shifter case8 . Such a strategy makes
the PDD method work very well in the finite resolution phase
shifter case from our numerical experience.
Figure 7 illustrates the spectral efficiency of the PDD
method and the modified RBD method when three bits quantization is used, i.e., b = 3. It is observed that the PDD
method can still achieve most of the spectral efficiency of
the FD precoding in this example of finite resolution phase
shifter case. Moreover, one can see again that the PDD method
outperforms the RBD method and their gap increases with
the SNR. For example, the gap between the PDD method
and the modified RBD method is about 5 bps/Hz when
SN R = 6 dB, while it is much smaller (almost negligible)
when SN R = −10 dB.
Figure 8 illustrates the spectral efficiency of the PDD
method versus SNR when different quantization levels are
used. It is observed that, the spectral efficiency of the PDD
7 Specifically, we quantize each element of the analog precoder/decoder to
the nearest point in the set F.
8 Note that the penalty parameter ρ needs not to be re-initialized. Thus, it
is in essence running PDD once.
11
60
60
FD
PDD
RBD
50
NRF=8, MRF=4, K=4, d=2
Average spectral efficiency (bps/Hz)
Average spectral efficiency (bps/Hz)
55
45
40
35
30
25
NRF=MRF=6, K=3, d=2
20
FD
PDD, b=4
PDD, b=3
PDD, b=2
PDD, b=1
50
40
30
20
10
15
Fig. 7.
b = 3.
−8
−6
−4
−2
0
SNR (dB)
2
4
0
−10
6
Spectral efficiency achieved by different methods versus SNR when
method improves when more bits are used in the phase
quantization, and the improvement shrinks as the quantization
level increases. Particularly, one can see that three bits quantization is enough for achieving about 95% of the fully-digital
precoding performance in this example.
Figure 9 shows the spectral efficiency of the PDD method
when different number of RF chains are used with the lowest
resolution phase shifters, i.e., b = 1. It is seen that the
performance gap between the PDD-based hybrid precoding
and the FD precoding can be reduced by increasing the number
of RF chains. Therefore, the number of RF chains can be
used to trade off the resolution of the phase shifters in hybrid
precoding design.
Lastly, we consider the cases when the numbers of transmit/receive RF chains fullfil the minimum requirement, i.e.,
NRF = Kd and MRF = d, and examine how much percentage of the FD precoding performance the hybrid precoding
can achieve. In the simulations, we test five system setups with
different combination of K and d which satisfies NRF = Kd
and MRF = d. For each setup, we run the PDD method
for 100 randomly generated channels under three quantization
levels b = ∞, b = 4, and b = 2. The minimum, average,
maximum spectral efficiency of the PDD method relative to
the spectral efficiency of the FD precoding are respectively
listed in TABLE I for SN R=0 dB. It can be observed
that, in the infinite resolution phase shifter case, the hybrid
precoding can achieve more than 95 percentage of the FD
precoding performance, implying the excellent efficiency of
the hybrid precoding while enjoying the benefit of significantly
reducing the number of RF chains in Massive MIMO systems.
Furthermore, even with low resolution phase shifters, the
hybrid precoding can still achieve most percentage of the FD
precoding performance, e.g., it is on average about 95% when
b = 4 and about 80% when b = 2.
D. Hybrid precoding for special cases: SU-MIMO & MUMISO
In this set of simulations, we show the hybrid precoding
performance achieved by the PDD method for SU-MIMO
−8
−6
−4
−2
0
SNR (dB)
2
4
6
Fig. 8.
Spectral efficiency achieved by PDD versus SNR for different
quantization levels when K = 4 and d = 2.
45
Average spectral efficiency (bps/Hz)
10
−10
FD
PDD, NRF=12
40
PDD, NRF=10
35
PDD, NRF=8
PDD, NRF=6
30
25
20
15
10
5
−10
−8
−6
−4
−2
0
SNR (dB)
2
4
6
Fig. 9. Spectral efficiency achieved by PDD versus SNR for different NRF
when MRF = 4, K = 3, d = 2 and b = 1.
TABLE III
T HE ACHIEVABLE RELATIVE PERFORMANCE OF HYBRID PRECODING AS
COMPARED TO FULLY- DIGITAL PRECODING (NRF = Kd, MRF = d)
Value
b=∞
b=4
b=2
(K, d)
min.
avg.
max.
min
avg.
max.
min.
avg.
max.
(2, 2)
(4, 2)
(2, 4)
(2, 3)
(3, 2)
94.35%
96.84%
98.78%
92.10%
95.21%
97.30%
78.81%
84.11%
88.73%
94.74%
96.86%
98.10%
85.00%
94.80%
97.11%
74.19%
80.84%
86.05%
94.54%
97.31%
98.23%
91.20%
94.57%
96.64%
70.92%
77.83%
82.76%
95.00%
97.12%
99.19%
91.70%
94.97%
96.72%
71.63%
80.47%
86.72%
93.46%
96.85%
98.50%
92.67%
95.10%
97.02%
76.09%
82.06%
87.24%
systems and MU-MISO systems as compared to the proposed
method in [15] named as ‘Sohrabi-Yu’.
First, we consider a MIMO system with N = 64, NRF =
16, MRF = 8 and d = 4. For this system, it is known
that the hybrid precoding can achieve the FD performance
in the inifinite resolution phase shifter case. The simulation
results are presented in Fig. 10 with b = ∞ and Fig. 11
12
30
FD
PDD
MAP
Sohrabi−YU
25
Average spectral efficiency (bps/Hz)
30
20
15
FD
PDD
Sohrabi−YU
25
20
15
10
5
−10
−8
−6
−4
−2
0
SNR (dB)
2
4
6
Fig. 11. Spectral efficiency achieved by PDD versus SNR for MIMO system
when N = 64, NRF = 16, MRF = 8, d = 4, b = 3.
24
FD
PDD
MAP
YU
22
Average spectral efficiency (bps/Hz)
with b = 3. It is observed from Fig. 10 that, in the inifinite
resolution phase shifter case, both the PDD method and the
MAP method can achieve the FD performance while there
exists a performance gap between the Sohrabi-Yu method and
the FD scheme. Furthermore, we find from Fig. 10 that, in
the finite resolution phase shifter case with b = 3, the PDD
method is slightly better than the Sohrabi-Yu method in the
low SNR case and almost has the same performance as the
Sohrabi-Yu method.However, the MAP method perform worst
in this case.
Second, we consider a 4-user MISO system with N = 64
and NRF = 8. In this case, there is no analog combiner.
Hence, the performance loss due to using the finite resolution
phase shifter is less than that in the MU-MIMO and MIMO
case. The simulation results are presented in Fig. 12 with b =
∞ and Fig. 13 with b = 3. It is observed from Fig. 12 that the
PDD method provides the same performance as the FD scheme
in the infinite resolution phase shifter case. Furthermore, as
shown in Fig. 13, the PDD method exhibits performance that is
very close to the FD performance in the finite resolution phase
shifter case, while significantly outperforming the Sohrabi-Yu
method.
20
18
16
14
12
10
8
6
4
−10
10
5
−10
−8
−6
−4
−2
SNR (dB)
0
2
4
6
−4
−2
0
SNR (dB)
2
4
6
25
Average spectral efficiency (bps/Hz)
By applying penalty dual decomposition method, this paper
has proposed an iterative algorithm to address the difficult
hybrid precoding problem for mmWave multi-user MIMO
systems. Different from the existing hybrid precoding algorithms which are all heuristic, the proposed algorithm has
guaranteed convergence to KKT solutions of the hybrid precoding problem. Although the optimality of the proposed
hybrid precoding method is not proven, simulation results
verify that the proposed hybrid precoding method performs
very well and its achievable spectral efficiency is very close
to the performance of the fully-digital precoding, implying
the capability of hybrid precoding in achieving near-optimal
spectral efficiency while greatly reducing the number of RF
chains. It is worth mentioning that, all the results are obtained
−6
Fig. 12. Spectral efficiency achieved by PDD versus SNR for 4-user MISO
system when N = 64, NRF = 8, and b = ∞.
Fig. 10. Spectral efficiency achieved by PDD versus SNR for MIMO system
when N = 64, NRF = 16, MRF = 8, d = 4, b = ∞.
VIII. C ONCLUSIONS
−8
FD
PDD
Sohrabi−YU
20
15
10
5
0
−10
−8
−6
−4
−2
0
SNR (dB)
2
4
6
Fig. 13. Spectral efficiency achieved by PDD versus SNR for 4-user MISO
system when N = 64, NRF = 8, and b = 3.
13
TABLE IV
A LGORITHM 3: BSUM A LGORITHM
•
•
0.
1.
2.
3.
4.
Input: xk−1
output: xk
initialize y 0 = xk−1 and set j = 0
repeat
w = yj
for each i ∈ {1, . . . , n}
Aji = arg minyi ∈Xi L̃k (yi ; w)
5.
set wi to be an arbitrary element in Aji
6.
end
7.
y j+1 = w
8.
j =j+1
9. until some termination criterion is met with accuracy ϵk
10. xk = y j
under the assumption of perfect channel state information. In
the future, we will consider hybrid precoding with imperfect
channel state information.
A PPENDIX A
A BASIC IMPLEMENTATION OF THE BSUM
ORACLE
We outline the basic BSUM algorithm [29] in TABLE
IV which can be used to implement the oracle xk =
BSUM(Pϱk ,λk , L̃k , xk−1 , ϵk ) in Step 2 of the PDD method.
Here, L̃k (yi ; w) denotes a locally tight upper bound of Lk (x)
w.r.t xi (i.e., the i-th block of x) at the point w.
In the BSUM algorithm [29] applied to a minimization
problem with multiple block variables, each time one block
variable is cyclicly picked to be optimized while fixing the
others by minimizing a locally tight upper bound of the
objective. In particular, when the upper bound is simply chosen
as the objective function itself, the BSUM algorithm reduces to
the BCD method [32]. Hence, the former includes the latter
as a special case. On the other hand, note that the BSUM
algorithm [29] was proposed for convex constraint cases.
When the problem has separable nonconvex constraints, we
need to resort to a randomized BSUM algorithm proposed in
[28] instead of BSUM for theoretical convergence guarantee.
For convenience, we sill refer to it as BSUM algorithm,
but keep in mind that the BCD or BSUM algorithm used
throughout this paper refers to randomized BSUM algorithm
in nonconvex constraint cases, where each time one block
variable is randomly picked to be optimized while fixing the
others.
where the matrices A ∈ Cm×m and C ∈ Cn×n are positive semidefinite and M denotes the set of unit modulus
constraints, i.e., |X(i, j)| = 1, ∀i, j.
We use BCD-type algorithm to address problem (31) with
guaranteed convergence to stationary solutions [32], i.e., in
each step we update one entry of X while fixing the others.
Without loss of generality, let us consider the problem of
minimizing ϕ(X) with respect to X(i, j) subject to the unit
modulus constraint |X(i, j)| = 1, i.e.,
min
|X(i,j)|=1
ϕ(X)
(32)
It is easily known that, the function ϕ(X) with respect to
X(i, j) can be expressed as a quadratic function of X(i, j)
in the form of ϕ̃(X(i, j)) , a|X(i, j)|2 − 2ℜe {b∗ X(i, j)} for
some real number a and some complex number b. Considering
that |X(i, j)| = 1, problem (32) reduces to
max
|X(i,j)|=1
ℜe {b∗ X(i, j)} .
(33)
It follows that the optimal X(i, j) is equal to b/|b|. Hence, in
order to update X(i, j), we only need to know the value of b.
In what follows, we show how the complex number b can
be easily obtained. First, we have [40]
∂ ϕ̃(X(i, j))
∂X∗ (i, j)
=
X(i,j)=X̃(i,j)
1
(aX̃(i, j) − b).
2
On the other hand, we have [40]
∂ϕ(X)
∂X∗
=
X=X̃
1
(AX̃C − B).
2
Combining the above equations, we obtain [AX̃C − B]ij =
aX̃(i, j) − b. By expanding [AX̃C]ij and checking the coefficient of X̃(i, j), we have aX̃(i, j) = A(i, i)X̃(i, j)C(j, j).
It follows that
b = A(i, i)X̃(i, j)C(j, j) − [AX̃C]ij + B(i, j).
According to the above analysis, the entries of X can be
recursively updated. The corresponding algorithm for updating
X is summarized in Table III, where the recursion step 5 is due
to the fact that Q should be updated accordingly once X(i, j)
is updated (which is done in step 6). It is easily known that
step 5 is the most costly step requiring complexity O(mn).
Hence, it can be shown that Algorithm 4 has complexity of
O(I3 m2 n2 ) where I3 denotes the total number of iterations
required by Algorithm 4.
R EFERENCES
A PPENDIX B
Q UADRATIC OPTIMIZATION WITH UNIT MODULUS
CONSTRAINTS
This appendix provides an iterative algorithm to address the
following quadratic optimization problem with unit modulus
constraints:
{
}
min ϕ(X) , Tr(XH AXC) − 2ℜe Tr(XH B)
(31)
X∈M
[1] T. S. Rappaport, S. Sun, R. Mayzus, H. Zhao, Y. Azar, K. Wang, G. N.
Wong, J. K. Schulz, M. Samimi, and F. Gutierrez, “Millimeter wave
mobile communications for 5G cellular: It will work!” IEEE Access,
vol. 1, pp. 335–349, 2013.
[2] R. W. Heath, N. Gonzlez-Prelcic, S. Rangan, W. Roh, and A. M. Sayeed,
“An overview of signal processing techniques for millimeter wave
MIMO systems,” IEEE Journal of Selected Topics in Signal Processing,
vol. 10, no. 3, pp. 436–453, April 2016.
[3] Z. Pi and F. Khan, “An introduction to millimeter-wave mobile broadband systems,” IEEE Communications Magazine, vol. 49, no. 6, pp.
101–107, June 2011.
14
TABLE V
A LGORITHM 4: BCD- TYPE ALGORITHM FOR PROBLEM (31)
[20]
0. set k = 1 and Q = AXC
1. repeat
2.
for (i, j) ∈ {1, 2, . . . , m} × {1, 2, . . . , n}
3.
b = A(i, i)X(i, j)C(j, j) − Q(i, j) + B(i, j)
b
4.
x = |b|
5.
Q = Q + (x − X(i, j))A(:, i)C(j, :)
6.
X(i, j) = x
7.
end
8.
k =k+1
9. until some termination criterion is met
[21]
[22]
[23]
[24]
[4] Z. Gao, L. Dai, D. Mi, Z. Wang, M. A. Imran, and M. Z. Shakir,
“MmWave massive-MIMO-based wireless backhaul for the 5G ultradense network,” IEEE Wireless Communications, vol. 22, no. 5, pp.
13–21, May 2015.
[5] S. K. Yong and C.-C. Chong, “An overview of multigigabit
wireless through millimeter wave technology: Potentials and technical
challenges,” EURASIP Journal on Wireless Communications and
Networking, vol. 2007, no. 1, p. 078907, 2006. [Online]. Available:
http://dx.doi.org/10.1155/2007/78907
[6] O. E. Ayach, S. Rajagopal, S. Abu-Surra, Z. Pi, and R. W. Heath,
“Spatially sparse precoding in millimeter wave MIMO systems,” IEEE
Transactions on Wireless Communications, vol. 13, no. 3, pp. 1499–
1513, March 2014.
[7] X. Yu, J. C. Shen, J. Zhang, and K. B. Letaief, “Alternating minimization
algorithms for hybrid precoding in millimeter wave MIMO systems,”
IEEE Journal of Selected Topics in Signal Processing, vol. 10, no. 3,
pp. 485–500, April 2016.
[8] W. Ni, X. Dong, and W.-S. Lu, “Near-optimal hybrid processing
for massive MIMO systems via matrix decomposition,” arXiv, 2015.
[Online]. Available: http://arxiv.org/abs/1504.03777
[9] C. E. Chen, “An iterative hybrid transceiver design algorithm for millimeter wave MIMO systems,” IEEE Wireless Communications Letters,
vol. 4, no. 3, pp. 285–288, June 2015.
[10] R. Mndez-Rial, C. Rusu, N. Gonzlez-Prelcic, and R. W. Heath,
“Dictionary-free hybrid precoders and combiners for mmwave MIMO
systems,” in 2015 IEEE 16th International Workshop on Signal Processing Advances in Wireless Communications (SPAWC), June 2015, pp.
151–155.
[11] A. Alkhateeb, G. Leus, and R. W. Heath, “Limited feedback hybrid
precoding for multi-user millimeter wave systems,” IEEE Transactions
on Wireless Communications, vol. 14, no. 11, pp. 6481–6494, Nov 2015.
[12] W. Ni and X. Dong, “Hybrid block diagonalization for massive multiuser
MIMO systems,” IEEE Trans. Commun., vol. 64, no. 1, pp. 201-211,
Jan. 2016.
[13] R. Rajashekar and L. Hanzo, “Iterative matrix decomposition aided
block diagonalization for mm-Wave multiuser MIMO systems,” IEEE
Trans. Wireless Commun., vol. 16, no. 3, pp. 1372-1384, Mar. 2017.
[14] L. Zhao, D. W. K. Ng, and J. Yuan, “Multi-User Precoding and Channel
Estimation for Hybrid Millimeter Wave Systems,” IEEE Journal of
Selected Topics in Signal Processing, vol. 35, no. 7, pp. 1576-1590,
July 2017.
[15] F. Sohrabi and W. Yu, “Hybrid digital and analog beamforming design
for large-scale antenna arrays,” IEEE Journal of Selected Topics in
Signal Processing, vol. 10, no. 3, pp. 501–513, April 2016.
[16] D. Ying, F. W. Vook, T. A. Thomas, and D. J. Love, “Hybrid structure
in massive MIMO: Achieving large sum rate with fewer RF chains,” in
2015 IEEE International Conference on Communications (ICC), June
2015, pp. 2344–2349.
[17] L. Liang, W. Xu, and X. Dong, “Low-complexity hybrid precoding
in massive multiuser MIMO systems,” IEEE Wireless Communications
Letters, vol. 3, no. 6, pp. 653–656, Dec 2014.
[18] J. Zhang, M. Haardt, I. Soloveychik, and A. Wiesel, “A channel matching
based hybrid analog-digital strategy for massive multi-user MIMO
downlink systems,” in 2016 IEEE Sensor Array and Multichannel Signal
Processing Workshop (SAM), July 2016, pp. 1–5.
[19] G. Zhu, K. Huang, V. K. N. Lau, B. Xia, X. Li, and S. Zhang, “Hybrid
beamforming via the Kronecker decomposition for the millimeter-wave
[25]
[26]
[27]
[28]
[29]
[30]
[31]
[32]
[33]
[34]
[35]
[36]
[37]
[38]
[39]
[40]
massive MIMO systems,” IEEE Journal of Selected Topics in Signal
Processing, vol. 35, no. 9, pp. 2097-2114, Sept. 2017.
X. Zhang, A. F. Molisch, and S.-Y. Kung, “Variable-phase-shift-based
RF-baseband codesign for MIMO antenna selection,” IEEE Transactions
on Signal Processing, vol. 53, no. 11, pp. 4091–4103, Nov 2005.
C. B. Peel, B. M. Hochwald, and A. L. Swindlehurst, “A
vector-perturbation technique for near-capacity multiantenna multiuser communication-part i: channel inversion and regularization,” IEEE
Transactions on Communications, vol. 53, no. 1, pp. 195–202, Jan 2005.
S. Shi, M. Schubert, and H. Boche, “Downlink mmse transceiver
optimization for multiuser MIMO systems: Duality and sum-mse minimization,” IEEE Transactions on Signal Processing, vol. 55, no. 11, pp.
5436–5446, Nov 2007.
Q. H. Spencer, A. L. Swindlehurst, and M. Haardt, “Zero-forcing methods for downlink spatial multiplexing in multiuser MIMO channels,”
IEEE Transactions on Signal Processing, vol. 52, no. 2, pp. 461–471,
Feb 2004.
V. Stankovic and M. Haardt, “Generalized design of multi-user MIMO
precoding matrices,” IEEE Transactions on Wireless Communications,
vol. 7, no. 3, pp. 953–961, March 2008.
S. Montori, C. Fritzsch, L. Marcaccioli, R. V. Gatti, R. Jakoby, and
R. Sorrentino, “Design and measurements of a 1-bit reconfigurable
elementary cell for large electronic steerable reflectarrays,” in Microwave
Conference (EuMC), 2010 European, Sept 2010, pp. 918–921.
J. D. Krieger, C. P. Yeang, and G. W. Wornell, “Dense delta-sigma
phased arrays,” IEEE Transactions on Antennas and Propagation,
vol. 61, no. 4, pp. 1825–1837, April 2013.
Q. Shi and M. Hong, “Penalty dual decomposition method with application in signal processing”, in Proc. ICASSP, Mar. 2017, pp. 1-5.
Q. Shi and M. Hong, “Penalty dual decomposition method for
nonconvex nonsmooth optimization–part I: Theory,” Tech. Report,
2017. [Online]. Available: http://arxiv.org/
M. Razaviyayn, M. Hong, and Z.-Q. Luo, “A unified convergence
analysis of block successive minimization methods for nonsmooth
optimization,” SIAM Journal on Optimization, vol. 23, no. 2, pp. 1126–
1153, 2013.
Q. Shi, M. Razaviyayn, Z. Q. Luo, and C. He, “An iteratively weighted
mmse approach to distributed sum-utility maximization for a MIMO
interfering broadcast channel,” IEEE Transactions on Signal Processing,
vol. 59, no. 9, pp. 4331–4340, Sept 2011.
S. S. Christensen, R. Agarwal, E. D. Carvalho, and J. M. Cioffi,
“Weighted sum-rate maximization using weighted MMSE for MIMOBC beamforming design,” IEEE Transactions on Wireless Communications, vol. 7, no. 12, pp. 4792–4799, December 2008.
D. Bertsekas, Nonlinear Programming, 2nd ed. Belmont, MA: Athena
Scientific, 1999.
Z. Lu and Y. Zhang, “Sparse approximation via penalty decomposition
methods,” SIAM Journal on Optimization, vol. 23, no. 4, pp. 2448–2478,
2013.
Z. Lu, Y. Zhang, and X. Li, “Penalty decomposition methods for rank
minimization,” Optimization Methods and Software, vol. 30, no. 3, pp.
531–558, 2015.
A. Ruszczynski, Nonlinear optimization. Princeton University Press,
2011.
A. F. Izmailov and M. V. Solodov, “Optimality conditions for irregular
inequality-constrained problems,” SIAM Journal on Control and Optimization, vol. 40, no. 4, pp. 1280–1295, 2002.
Q. Shi, M. Hong, X. Gao, E. Song, Y. Cai, and W. Xu, “Joint
source-relay design for full–duplex MIMO AF relay systems,” IEEE
Transactions on Signal Processing, vol. 64, no. 23, pp. 6118-6131,
Dec. 2016.
Q. Shi, W. Xu, J. Wu, E. Song, and Y. Wang, “Secure beamforming
for MIMO broadcasting with wireless information and power transfer,”
IEEE Transactions on Wireless Communications, vol. 14, no. 5, pp.
2841–2853, May 2015.
P.-A. Absil, R. Mahony, and R. Sepulchre, Optimization Algorithms on
Matrix Manifolds. Princeton, NJ, USA: Princeton Univ. Press, 2009.
K.
B.
Petersen
and
M.
S.
Pedersen,
“The
matrix
cookbook,”
Nov
2012.
[Online].
Available:
https://www.math.uwaterloo.ca/∼hwolkowi/matrixcookbook.pdf
| 7cs.IT
|
Distributed Semi-Stochastic Optimization with Quantization Refinement
arXiv:1603.06306v1 [math.OC] 21 Mar 2016
Neil McGlohon and Stacy Patterson
Abstract— We consider the problem of regularized regression
in a network of communication-constrained devices. Each
node has local data and objectives, and the goal is for the
nodes to optimize a global objective. We develop a distributed
optimization algorithm that is based on recent work on semistochastic proximal gradient methods. Our algorithm employs
iteratively refined quantization to limit message size. We present
theoretical analysis and conditions for the algorithm to achieve
a linear convergence rate. Finally, we demonstrate the performance of our algorithm through numerical simulations.
I. I NTRODUCTION
We consider the problem of distributed optimization in a
network where communication is constrained, for example a
wireless sensor network. In particular, we focus on problems
where each node has local data and objectives, and the goal
is for the nodes to learn a global objective that includes this
local information. Such problems arise in networked systems
problems such as estimation, prediction, resource allocation,
and control.
Recent works have proposed distributed optimization
methods that reduce communication by using quantization.
For example, in [1], the authors propose a distributed algorithm to solve unconstrained problems based on a centralized
inexact proximal gradient method [2]. In [3], the authors
extend their work to constrained optimization problems. In
these algorithms, the nodes compute a full gradient step in
each iteration, requiring quantized communication between
every pair of neighboring nodes. Quantization has been
applied in distributed consensus algorithms [4], [5], [6] and
distributed subgradient methods [7].
In this work, we address the specific problem of distributed
regression with regularization over the variables across all
nodes. Applications of our approach include distributed
compressed sensing, LASSO, group LASSO, and regression
with Elastic Net regularization, among others. Our approach
is inspired by [1], [3]. We seek to further reduce per-iteration
communication by using an approach based on a stochastic
proximal gradient algorithm. This approach only requires
communication between a small subset of nodes in each
iteration. In general, stochastic gradients may suffer from
slow convergence. Thus any per-iteration communication
savings could be counter-acted by an extended number of
iterations. Recently, however, several works have proposed
semi-stochastic gradient methods [8], [9], [10]. To reduce the
variance of the iterates generated by a stochastic approach,
these algorithms periodically incorporate a full gradient
*This work was funded in part by NSF grants 1553340 and 1527287.
N. McGlohon and S. Patterson are with the Department of Computer Science, Rensselaer Polytechnic Institute, Troy, NY 12180, USA
[email protected], [email protected]
computation. It has been shown that these algorithms achieve
a linear rate of convergence to the optimal solution.
We propose a distributed algorithm for regularized regression based on the centralized semi-stochastic proximal gradient of [10]. In most iterations, only a subset of nodes need
communicate. We further reduce communication overhead
by employing quantized messaging. Our approach reduces
both the length of messages sent between nodes as well
as the number of messages sent in total to converge to the
optimal solution. The detailed contributions of our work are
as follows:
• We extend the centralized semi-stochastic proximal
gradient algorithm to include errors in the gradient
computations and show the convergence rate of this
inexact algorithm.
• We propose a distributed optimization algorithm based
on this centralized algorithm that uses iteratively refined
quantization to limit message size.
• We show that our distributed algorithm is equivalent to
the centralized algorithm, where the errors introduced
by quantization can be interpreted as inexact gradient
computations. We further design quantizers that guarantees a linear convergence rate to the optimal solution.
• We demonstrate the performance of the proposed algorithm in numerical simulations.
The remainder of this paper is organized as follows.
In Section II, we present the centralized inexact proximal
gradient algorithm and give background on quantization.
In Section III, we give the system model and problem
formulation. Section IV details our distributed algorithm.
Section V provides theoretical analysis of our proposed
algorithm. Section VI presents our simulation results, and
we conclude in Section VII.
II. P RELIMINARIES
A. Inexact Semi-Stochastic Proximal Gradient Algorithm
We consider an optimization problem over the form:
minimize G(x) = F (x) + R(x),
x∈RP
(1)
PN
where F (x) = N1 i=1 fi (x), and the following assumptions are satisfied.
Assumption 1: Each fi (x) is differentiable, and its gradient ∇fi (x) is Lipschitz continuous with constant Li , i.e., for
all x, y ∈ RP ,
k∇fi (x) − ∇fi (y)k ≤ Li kx − yk.
(2)
Algorithm 1 Inexact Prox-SVRG.
their gradients ∇fi (x(st ) ). Let x? = arg minx G(x), and let
T be such that,
Initialize: x̃(s) = 0
for s = 0, 1, 2, . . . do
g̃ (s) = ∇F (x̃(s) )
x(s0 ) = x̃(s)
for t = 0, 1, 2, . . . , T − 1 do
Choose ` uniformly at random from {1, . . . , N }.
v (st ) = ∇f` (x(st ) ) − ∇f` (x̃(s) ) + g̃ (s) + e(st )
x(st+1 ) = proxηR (x̃(st ) − ηv (st ) )
end for
PT
(st )
x̃(s+1) = T1
t=1 x̃
end for
α=
1
4Lη(T + 1)
+
< 1.
µη(1 − 4Lη)T
(1 − 4Lη)T
Then,
h
i
E G(x̃(s) ) − G(x? )
≤ αs
G(x̃(0) ) − G(x? ) + β
s
X
!
α−i Γ(i)
i=1
Assumption 2: The function R(x) is lower semicontinuous, convex, and its effective domain, dom(R) := {x ∈
RP | R(x) < +∞}, is closed.
Assumption 3: The function G(x) is strongly convex with
parameter µ > 0, i.e., for all x, y ∈ dom(R) and for all
ξ ∈ ∂G(x),
G(x) − G(y) − 12 µkx − yk2 ≥ ξ T (x − y),
(3)
where ∂G(x) is the subdifferential of G at x. This strong
convexity may come from either F (x) or R(x) (or both).
Problem (1) can be solved using a stochastic proximal
gradient algorithm [11] where, in each iteration, a single ∇f`
is computed for a randomly chosen ` ∈ {1, . . . , N }, and the
iterate is updated accordingly as,
x(t+1) = proxηR (x(t) − η (t) ∇f` (x(t) )).
Here, proxηR (·) is the proximal operator
1
proxηR (v) = arg min ky − vk2 + ηR(y).
y∈Rp 2
While stochastic methods offer the benefit of reduced periteration computation over standard gradient methods, the
iterates may have high variance. These methods typically use
a decreasing step-size η (t) to compensate for this variance,
resulting in slow convergence. Recently, Xiao and Zhang proposed a semi-stochastic proximal gradient algorithm, ProxSVRG that reduces the variance by periodically incorporating a full gradient computation [10]. This modification allows
Prox-SVRG to use a constant step size, and thus, Prox-SVRG
achieves a linear convergence rate.
We extend Prox-SVRG to include a zero-mean error in the
gradient computation. Our resulting algorithm, Inexact ProxSVRG, is given in Algorithm 1. The algorithm consists of
an outer loop where the full gradient is computed and an
inner loop where the iterate is updated based on both the
stochastic and full gradients.
The following theorem states the convergence behavior of
Algorithm 1.
Theorem 1: Let {x̃(s) }s≥0 be the sequence generated by
1
Algorithm 1, with 0 < η < 4L
, where L = maxi Li .
Assume that the functions R, G, and fi , i = 1, . . . , N ,
satisfy Assumptions 1, 2, and 3, and that the errors e(st )
are zero-mean and uncorrelated with the iterates x(st ) and
η
T (1−4Lη)
(i)
PT −1
and Γ = t=0 Eke(it ) k2 .
where β =
The proof is given in the appendix.
From this theorem, we can derive conditions for the
algorithm to converge to the optimal x? . Let the sequence
{Γ(s) }s≥0 decrease linearly at a rate κ. Then
1) If κ < α, then E G(x̃(s) ) − G(x? ) converges linearly
with a rate of α.
2) If α < κ < 1, then E G(x̃(s) ) − G(x? ) converges
linearly with a rate
of κ.
3) If κ = α, then E G(x̃(s) ) − G(x? ) converges linearly
with a rate in O(sαs ).
B. Subtractively Dithered Quantization
We employ a subtractively dithered quantizer to quantize
values before transmission. We use a substractively dithered
quantizer rather than non-subtractively dithered quantizer
because the quantization error of the subtractively dithered
quantizer is not correlated with its input. We briefly summarize the quantizer and its key properties below.
Let z be real number to be quantized into n bits. The
quantizer is parameterized by an interval size U and a
midpoint value z ∈ R. Thus the quantization interval
is [z − U/2, z + U/2], and the quantization step-size is
∆ = 2nU−1 . We first define the uniform quantizer,
|z − z| 1
q(z) , z + sgn(z − z) · ∆ ·
+
.
(4)
∆
2
In subtractively dithered quantization, a dither ν is added to
z, the resulting value is quantized using a uniform quantizer,
and then transmitted. The recipient then subtracts ν from
this value. The subtractively dithered quantized value of z,
denoted ẑ, is thus
ẑ = Q(z) , q(z + ν) − ν.
(5)
Note that this quantizer requires both the sender and recipient
to use the same value for ν, for example, by using the same
pseudorandom number generator.
The following theorem describes the statistical properties
of the quantization error.
Theorem 2 (See [12]): Let z ∈ [z − U/2, z + U/2] and
ẑ = Q(z), for Q(·) in (5). Further, let ν is a real number
drawn uniformly at random from the interval (−∆/2, ∆/2).
The quantization error ε(z) , z − ẑ satisfies the following:
1) E [ε(z)]
=
E [ν] =0. 2
2) E ε(z)2 = E ν 2 = ∆
12
3) E [zε(z)] = E [z] E [ε(z)] = 0
4) For z1 and z2 in the interval [z − U/2, z + U/2],
E [ε(z1 )ε(z2 )] = E [ε(z1 )] E [ε(z2 )] = 0.
With some abuse of notation, we also write Q(v) where v
is a vector. In this case, the quantization operator is applied
to each component of v independently, using a vector-valued
midpoint and the same scalar-valued interval bounds.
III. P ROBLEM F ORMULATION
We consider a similar system model to that in [1]. The
network is a connected graph of N nodes where inter-node
communication is limited to the local neighborhood of each
node. The neighbor set Ni consists of node i’s neighbors and
itself. The neighborhoods exist corresponding to the fixed
undirected graph G = (V, E). We denote D as the maximum
degree of the graph G.
Each node i has a state vector xi with dimension mi . The
T
T T
state of the system is x = [xT
1 x2 . . . xN ] . We let xNi be the
vector consisting of the concatenation of states of all nodes in
Ni . For ease of exposition, we define the selecting matrices
Ai , i = 1, . . . , N , where xNi = Ai x and the matrices Bij ,
i, j = 1, . . . , N where xj = Bij xNi . These matrices each
have `2 -norm of 1.
Every node i has a local objective function over the states
in Ni . The distributed optimization problem is thus,
minimize G(x) = F (x) + R(x),
x∈RP
(6)
PN
where F (x) = N1 i=1 fi (xNi ). We assume that Assumptions 1 and 3 are satisfied. Further, we require the following
assumptions hold.
Assumption 4: For all i, ∇fi (xNi ) is linear or constant.
This implies that, for a zero-mean random variable ν,
E [∇fi (xNi + ν)] = ∇fi (xNi ).
Assumption 5: The proximal operation proxR (x) can be
performed by each node locally, i.e.,
proxR (x) = [proxR (x1 )T proxR (x2 )T . . . proxR (xN )T ]T .
We note that Assumption 5 holds for standard regularization
functions used in LASSO (kxk1 ), group LASSO where each
xi its own group, and Elastic Net regularization (λ1 kxk1 +
λ2
2
2 kxk2 ).
In the next section, we present our distributed implementation of Prox-SVRG to solve Problem (6).
IV. A LGORITHM
Our distributed algorithm is given in Algorithm 2. In
(s)
each outer iteration s, node i quantizes its iterate x̃i and
(s)
the gradient ∇fi and sends it to all of its neighbors.
These values are quantized using two subtractively dithered
(s)
(s)
quantizers, Qa,i and Qb,i , whereby the sender (node i) sends
an n bit representation and the recipient reconstructs the
value from this representation and subtracts the dither. The
(s)
(s)
midpoints for Qa,i and Qb,i are set to be the quantized values
from the previous iteration. Thus, the recipients already
know these midpoints. The quantized values (after the dither
Algorithm 2 Inexact Semi-stochastic Gradient Descent as
executed by node i
1: Parameters: inner loop size T , step size η
(0)
ˆ (−1) = 0
ˆ(−1) = 0, ∇f
2: Initialize: x̃i = 0, x̃
i
i
3: for s = 0, 1, . . . do
4: Update quantizer parameters:
(s)
(s)
ˆ(s−1) ,
Ua,i = Ca κ(s+1)/2 , xa,i = x̃
i
5:
(s)
(s)
(s+1)/2
ˆ (s−1)
Ub,i = Cb κ
, ∇fb,i = ∇f
i
6: Quantize local variable and send to all j ∈ Ni :
(s)
(s)
(s)
(s)
(s)
ˆ
x̃
= Qa,i (x̃i ) = x̃i + ai
i
(s)
(s)
ˆ )
7: Compute: ∇fi = ∇fi (x̃
Ni
8: Quantize gradient and send to all j ∈ Ni :
(s)
(s)
(s)
(s)
(s)
ˆ
∇f
= Qb,i (∇fi ) = ∇fi + bi
i
P
(s)
(s)
ˆ
9: Compute: h̃
= 1
Bij ∇f
i
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
22:
23:
24:
25:
26:
N
j∈Ni
j
(s)
ˆ (s) + h̃(s) for all j ∈ Ni
Compute: vij = −Bij ∇f
j
i
Update quantizer parameters:
(s)
(s)
(s)
ˆ ,
Uc,i = Cc κ(s+1)/2 , xc,i = x̃
i
(s)
Ud,i
(s0 )
xi
=
(s)
(s)
ˆ
= Cd κ(s+1)/2 , ∇fd,i = ∇f
i
(s)
x̃i
for t = 0, 1, . . . , T − 1 do
Randomly pick ` ∈ {1, 2, 3, . . . , N }
if i ∈ N` then
Quantize local variable and send to `:
(s )
(s ) (s )
(s )
(s )
x̂i t = Qc,it (xi t ) = xi t + ci t
if i = ` then
(s )
(s )
Compute: ∇fi t = ∇fi (x̂N t )
i
Quantize gradient and send to all j ∈ Ni :
(s
)
(s
)
(s
)
(s )
(s )
t
t
t
ˆ
∇f
= Qd,i (∇fi ) = ∇fi t + di t
i
end if
Update local variable:
(st+1 )
xi
(st )
= proxηR (xi
else
Update local variable:
(st+1 )
xi
27:
end if
28: end for
(s+1)
29: x̃i
=
30: end for
1
T
PT
(st )
ˆ
− η(Bi` ∇f
`
(s)
+ vi` ))
(s)
= proxηR (x(st ) − η h̃i )
t=1
x(st )
ˆ (s) , and the
ˆ(s) and ∇f
is subtracted) are denoted by x̃
i
i
(s)
(s)
quantization errors are ai and bi , respectively.
For every iteration s of the outer loop of the algorithm,
there is an inner loop of T iterations. In each inner iteration,
a single node `, chosen at random, computes its gradient.
To do this, node ` and its neighbors exchange their states
(s )
(s )
xi t and gradients ∇fi t . These values are quantized using
(s )
(s )
two subtractively dithered quantizers, Qc,it and Qd,it . The
ˆ (s) . Each
ˆ(s) and ∇f
midpoints for these quantizers are x̃
i
i
node sends these values to their neighbors before the inner
loop, so all nodes are aware of the midpoints. The quantized
values (after the dither is subtracted) are denoted by x̂(st )
ˆ (st ) , and their quantization errors are c(st ) and d(st ) ,
and ∇f
i
i
i
(s)
(s)
respectively. The quantization interval bounds Ua,i , Ub,i ,
(s)
(s)
Uc,i , and Ud,i , are initialized to Ca , Cb , Cc , and Cd ,
respectively, and each iteration, the bounds are multiplied
by κ1/2 . Thus the quantizers are refined in each iteration.
The quantizers limit the length of a single variable transmission to n bits. In the outer loop of the algorithm, each
node i sends its local variable, consisting of mi quantized
components, to every neighbor. It also sends its gradient,
consisting of |Ni |mi quantized components to every neighbor.
PNThus the number of bits exchanged by all nodes is
n i=1 |Ni |mi + |Ni |2 mi bits. In each inner iteration, only
nodes j ∈ N` exchange messages. Each node j quantizes
mj state variables P
and sends them to node `. This yields a
transmission of n j∈N` mj bits in total. In turn, node `
quantizes its gradient and sends it to all of its neighbors,
2
which is n|N` |P
m` total bits. Thus, in each inner iteration
2
n(|N` | m` + j∈N` mj ) bits are transmitted. The total
number of bits transmitted in a single outer iteration is
therefore,
N
T
−1
X
X
X
|N` |2 m` +
n
(|Ni |mi (1 + |Ni |)) +
mj .
t=0
i=1
j∈N`
Let D = maxi |Ni | and m = maxi mi . An upper bound
on the number bits transmitted by the algorithm in each outer
iteration is nm(N + T )(D + D2 ).
V. A LGORITHM A NALYSIS
We now present our analysis of Algorithm 2. First we
show that the algorithm is equivalent to Algorithm 1, where
the quantization errors are encapsulated in the error term
e(st ) . We also give an explicit expression for this error term.
We now show that e(st ) is is uncorrelated with x(st ) and
(s )
the gradients ∇f` (xN`t ), ` = 1, . . . , N . Clearly, x(st ) and
(st )
∇f` (xN` ) are uncorrelated with the terms of e(st ) contain(s )
(s)
(s)
ing d` t , b` , and bi . In accordance with Assumption 5,
the gradients ∇f` and ∇fi are either linear or constant.
(s )
(s )
If they are constant, then ∇f` (x̂N`t ) − ∇f` (xN`t ) = 0
ˆ(s) ) − ∇fi (x̃(s) ) = 0. Thus, the terms in e(st )
and ∇fi (x̃
Ni
Ni
containing these differences are also 0. If they are linear,
e.g., ∇f` (z) = Hz + h, for an appropriately sized, matrix
H and vector h (possibly 0). Then,
(s )
(s )
∇f` (x̂N`t ) − ∇f` (xN`t )
(s )
(st )
= (H(xN`t + ci
(st )
) + h) − (Hx(st ) + h) = Hci
.
(s )
ci t
By Theorem 2,
is uncorrelated with x(st ) . It is clearly
(s )
also uncorrelated with ∇f` (xN`t ). Similar arguments can be
(s )
used to show that x(st ) and ∇f` (xN`t ) are uncorrelated with
the remaining terms in e(st ) .
With respect to Eke(st ) k2 , we have
(st )
(st )
Eke(st ) k2 = EkAT
` ∇f` (x̂N` ) − ∇f` (xN` )
(s)
ˆ(s)
− AT
` ∇f` (x̃N` ) − ∇f` (x̃N` )
PN
(s)
2
ˆ(s)
+ N1 i=1 AT
i ∇fi (x̃Ni ) − ∇fi (x̃Ni ) k
PN
(s)
(st )
T (s) 2
+ EkAT
+ N1 i=1 AT
` d`
i bi − A` b` k .
The first term on the right hand side can be bounded using
Lemma 1: Algorithm 2 is equivalent to the Inexact Proxthe fact that ka + bk2 ≤ 2kak2 + 2kbk2 , as
SVG method in Algorithm 1, with
(st )
(st )
T
≤
2EkA
∇f
(x̂
)
−
∇f
(x
)
k2
`
`
`
(s
)
(s
)
(s
)
N
N
t
t
t
T
`
`
e(st ) = AT
` ∇f` (x̂N` ) − ∇f` (xN` ) + A` d`
(s)
(s)
T
ˆ
+
2EkA
∇f
(
x̃
)
−
∇f
(x̃
)
`
`
`
(s)
N`
N`
T (s)
ˆ(s)
− AT
` ∇f` (x̃N` ) − ∇f` (x̃N` ) − A` b`
PN
(s)
(s)
1
T
ˆ
A
∇f
(
x̃
)
−
∇f
(x̃
)
k2 .
+
P
P
i Ni
i Ni
(s)
(s)
i=1 i
N
N
N
1
T (s)
ˆ
+ N1 i=1 AT
A
b
.
∇f
(
x̃
)
−
∇f
(x̃
)
+
i Ni
i Ni
i
i=1 i i
N
Further, Eke(st ) k2 is upper-bounded by,
2 X
2 X
(s )
(s)
Ekcj t k2 + 2L
Ekaj k2
Eke(st ) k2 ≤ 2L
j∈N`
(st ) 2
+ Ekd`
j∈N`
(s)
k + 2Ekb` k2 +
N
2 X
(s)
Ekbi k2 .
N 2 i=1
Proof: The error e(st ) is:
ˆ (st ) − AT ∇f
ˆ (s) + 1 PN Ai T ∇f
ˆ (s)
e(st ) = A` T ∇f
i
`
i=1
`
`
N
(st )
(s)
T
T
− A` ∇f` (xN` ) − A` ∇f` (x̃N` )
PN
(s)
+ N1 i=1 Ai T ∇fi (x̃Ni )
(st )
(st )
T (st )
= AT
` ∇f` (x̂N` ) − ∇f` (xN` ) + A` d`
(s)
(s)
(s)
ˆ
x̃
− AT
∇f
(
)
−
∇f
(x̃
)
− AT
`
`
`
` b`
N`
N`
PN
PN
(s)
1
T (s)
ˆ(s)
+ N1 i=1 AT
i ∇fi (x̃Ni ) − ∇fi (x̃Ni ) + N
i=1 Ai bi .
We note that all quantization errors are zero-mean. Further,
by Assumption 5, E [∇fi (x + δ)]= ∇fi (x), for a zero-mean
random variable δ. Therefore, E e(st ) = 0.
We now bound the first term in this expression,
(st )
(st )
2
2EkAT
` ∇f` (x̂N` ) − ∇f` (xN` ) k
2 X
(s )
(s )
(s )
≤ 2E(L2i kx̂N`t − xN`t k2 ) ≤ 2L
Ekcj t k2 ,
j∈N`
where the first inequality follows from Assumptions 1 and 5
and the fact that kA` k = 1. The second inequality follows
from the independence of quantization errors (Theorem 2).
Next we bound the second term,
(s)
(s)
ˆ
)
2EkAT
∇f
(
x̃
)
−
∇f
(x̃
`
`
`
N`
N`
PN
(s)
1
T
ˆ ) − ∇fi (x̃(s) ) k2
+ N i=1 Ai ∇fi (, x̃
Ni
Ni
(s)
(s)
T
ˆ ) − ∇f` (x̃ )
= 2EkA` ∇f` (x̃
N`
N`
i
h
(s)
(s)
ˆ
− E AT
k2
` ∇f` (x̃N` ) − ∇f` (x̃N` )
(s)
2
ˆ(s)
≤ 2EkAT
` ∇f` (x̃N` ) − ∇f` (x̃N` ) k
ˆ(s) − x̃(s) k2 )
≤ 2E(L2i k(x̃
N`
N`
P
2
(s) 2
≤ 2L
Eka
j k ,
j∈N`
where the first inequality uses the fact that for a random
variable υ, Ekυ − Eυk2 = Ekυk2 − kEυk2 ≤ Ekυk2 . The
remaining inequalities follow from Assumptions 1 and 5, the
fact that kA` k = 1, and the independence of the quantization
errors.
Finally, again from the independence of the quantization
errors, we have,
PN
(s)
(st )
T (s) 2
EkAT
+ N1 i=1 AT
` d`
i bi − A ` b` k
PN
(st ) 2
1
T (s) 2
T (s) 2
≤ EkAT
` d` k + Ek N
i=1 Ai bi k − A` b` k
PN
(s )
(s)
(s)
≤ Ekd` t k2 + 2Ekb` k2 + N22 i=1 Ekbi k2 .
Combining these bounds, we obtain the desired result,
2 X
2 X
(s )
(s)
Eke(st ) k2 ≤ 2L
Ekcj t k2 + 2L
Ekaj k2
j∈N`
j∈N`
N
2 X
(s)
(s )
(s)
Ekbi k2 .
+ Ekd` t k2 + 2Ekb` k2 + 2
N i=1
We next show that, if all of the values fall within their
respective quantization intervals, then the error term Γ(s) decreases linearly with rate κ, and thus the algorithm converges
to the optimal solution linearly with rate κ.
Theorem 3: Given p, if for all 1 ≤ s ≤ (p − 1), the
(s)
(s)
(s )
values of x̃i , ∇fi , x(st ) , and ∇fi t fall inside of the
(s)
(s)
(s )
(s )
respective quantization intervals Qa,i , Qb,i , Qc,it , and Qd,it ,
then Γ(k) ≤ Cκk , where,
DT m 2
N +1
C=
(C
+
C
)
+
2(
2L
)C
+
C
,
a
b
b
d
N
12(2` − 1)2
with D = maxi |Ni | and m = maxi mi .
It follows that, for α < κ < 1,
h
i
E G(x̃(s) ) − G(x? )
≤ κs G(x̃(0) ) − G(x? ) + βC
1
1 − ακ
Proof: First we note that, by Theorem 2 and the update
rule for the quantization intervals, we have:
(s) 2
Ua,i
(s)
m
s
Ekai k2 ≤ 12
≤ 12(2m
` −1)2 Ca κ
2` −1
(s) 2
U
(s) 2
Dm
s
Ekbi k ≤ 12 2`b,i
≤ 12(2Dm
` −1)2 Cb κ
−1
(s) 2
U
(st ) 2
m
s
Ekci k ≤ 12 2`c,i
≤ 12(2m
` −1)2 Cc κ
−1
(s) 2
Ud,i
(s )
s
Ekdi t k2 ≤ Dm
≤ 12(2Dm
` −1)2 Cd κ .
12
2` −1
We use these inequalities to bound ke(st ) k2 ,
2
s
Eke(st ) k2 ≤ 2L D 12(2m
` −1)2 Cc κ
2
s
s
+ 2 L D 12(2m
+ 12(2Dm
` −1)2 Ca κ
` −1)2 Cd κ
s
s
+ 2 12(2Dm
+ N2 12(2Dm
` −1)2 Cb κ
` −1)2 Cb κ
2
= 12(2Dm
2L (Ca + Cc ) + 2( NN+1 )Cb + Cd κs .
` −1)2
Summing over t = 0, . . . , T − 1, we obtain,
Γ(s) =
T
−1
X
Eke(st ) k2 ≤ Cκs ,
t=0
where
C=
DT m
12(2` −1)2
2
2L (Ca + Cc ) + 2( NN+1 )Cb + Cd
Applying Theorem 1, with κ > α, we have
h
i
E G(x̃(s) ) − G(x? )
s
X
≤ αs G(x̃(0) ) − G(x? ) + β
αs−i Cκi
s
≤κ
G(x̃
(0)
?
i=1
s
X
) − G(x ) + Cβ
!
−(s−i) s−i
κ
α
i=1
)s
1−( α
Cβ 1−κα
κ
≤ κs G(x̃(0) ) − G(x? ) +
.
≤ κs G(x̃(0) ) − G(x? ) + Cβ 1−1 α
κ
While we do not yet have theoretical guarantees that
all values will fall within their quantization intervals, our
simulations indicate that is always possible to find parameters
Ca , Cb , Cc , and Cd , for which all values lie within their
quantization intervals for all iterations. Thus, in practice, our
algorithm achieves a linear convergence rate. We anticipate
that it is possible to develop a programmatic approach,
similar to that in [1], to identify values for Ca , Cb , Cc , and
Cd that guarantee linear convergence. This is a subject of
current work.
VI. N UMERICAL E XAMPLE
This section illustrates the performance of Algorithm 2 by
solving a distributed linear regression problem with elastic
net regularization.
We randomly generate a d-regular graph with N = 40
and uniform degree of 8, i.e., ∀i |Ni | = 9. We set each
subsystem size, mi , to be 10. Each node has a local function
fi (xNi ) = kHi xNi − hi k2 where Hi is a 80 × 90 random
matrix. We generate hi by first generating a random vector x
and then computing hi = Hi x. The global objective function
is:
N
λ2
1 X
fi (xNi ) + λ1 kxk2 + kxk1 .
G(x) =
N i
2
This simulation was implemented in Matlab and the optimal value x? was computed using CVX. We set the total
number of inner iterations to be T = 2N and use the step
size η = 0.1/L. With these values, α < 1, as required by
Theorem 1. We set κ = 0.97, which ensures that κ > α. We
use the quantization parameters Ca = 50, Cb = 300, Cc =
50, Cd = 400. With these parameters, the algorithms values
always fell within their quantization intervals.
Fig. 1 shows the performance of the algorithm where the
number of bits n is 11, 13, and 15, as well as the performance
of the algorithm without quantization. In these results, x(s)
convex, then µF (µR ) is its strong convexity parameter; if
F (x) (R(x)) is only convex, then µF (µR ) is 0. For any
x ∈ dom(R) and any v ∈ RP , define,
105
n = 11
n = 13
n = 15
No Error
100
x+ = proxηR (x − ηv)
jjx(s) ! x? jj
h = η1 (x − x+ )
∆ = v − ∇F (x),
10!5
where 0 < η <
1
L.
Then, for any y ∈ RP ,
G(y) ≥ G(x+ ) + hT (y − x) + η2 khk2 +
10!10
µR
2 ky
10!15
0
50
100
150
200
250
300
Iteration s
Fig. 1: Comparison of the performance of Algorithm 2
with differing quantized message lengths and that with no
quantization applied.
+ 2
T
µF
2
ky − xk2
+
+
− x k + ∆ (x − y).
We now proceed to prove Theorem 1. For brevity, we
omit some details that are identical to those in the proof
of Theorem 3.1 in [10]. We have indicated these omissions
below.
Proof: First, we define
h(st ) = η1 (x(st−1 ) − x(st ) )
= η1 (x(st−1 ) − proxηR (x(st−1 ) − ηv (st−1 ) )),
(s)
is the concatenation of the x̃i vectors, for i = 1 . . . N . It
is important to note the rate of convergence of the algorithm
in all four cases is linear, and, performance improves as the
number of bits increases.
where v (st−1 ) is as defined in Algorithm 1.
We analyze the change in the distance between x(st ) and
?
x in a single inner iteration,
kx(st ) − x? k2 = kx(st−1 ) − ηh(st ) − x? k2
VII. C ONCLUSION
T
We have presented a distributed algorithm for regularized regression in communication-constrained networks. This
algorithm is based on recently proposed semi-stochastic
proximal gradient methods. Our algorithm reduces communication requirements by (1) using a stochastic approach
where only a subset of nodes communicate in each iteration
and (2) quantizing all messages. We have shown that this
distributed algorithm is equivalent to a centralized version
with inexact gradient computations, and we have used this
equivalence to analyze the convergence rate of the distributed
method. Finally, we have demonstrated the performance of
our algorithm in numerical simulations.
In future work, we plan to extend our theoretical analysis
to develop a programmatic way to identify initial quantization intervals. We also plan to explore the integration of more
complex regularization functions.
A PPENDIX
Proof of Theorem 1
= kx(st−1 ) x? k2 − 2ηh(st ) (x(st−1 ) − x? ) + η 2 kh(st ) k2 .
We next apply Lemma 3, with x = x(st−1 ) , x+ = x(st ) ,
h = h(st ) , v = v (st−1 ) , and y = x? , to obtain,
T
− h(st ) (x(st−1 ) − x? ) + η2 kh(st ) k2
≤ G(x? ) − G(x(st ) ) −
−
(s)
T
− x? k2 − ∆(st ) (x(st ) − x? ),
kx(st ) − x? k2 ≤ kx(st−1 ) − x? k2 − 2η(G(x(st ) ) − G(x? ))
T
− 2η∆(st ) (x(st ) − x? ).
We follow the same reasoning as in the proof of Theorem
3.1 in [10] to obtain the following expression, which is
conditioned on x(st−1 ) and takes expectation with respect
to `,
Ekx(st ) − x? k2 ≤ kx(st−1 ) − x? k2
− 2ηE(G(x(st ) ) − G(x? )) + 2η 2 Ek∆(st ) k2
h
i
T
− 2ηE ∆(st ) (x(st ) − x? ) ,
(s)
) − ∇f` (x̃ ) + ∇F (x̃ ).
Then, conditioned on x(st−1 ) , E w(st ) = ∇F (x(st−1 ) ) and
w
= ∇f` (x
(st−1 )
kx(st−1 ) − x? k2
where ∆(st ) = v (st−1 ) − ∇F (x(st−1 ) ) = w(st−1 ) + e(st−1 ) −
∇F (x(st−1 ) ). This implies,
We first restate some useful results from [10].
Lemma 2: Let
(st )
µR
(st )
2 kx
µF
2
Ekw(st ) − ∇F (x(st−1 ) )k2
≤ 4L(G(x(st−1 ) ) − G(x? ) + G(x̃(s) ) − G(x? )),
Lemma 3: Let G(x) = F (x) + R(x), where G is strongly
convex, ∇F (x) is Lipschitz continuous with parameter L.
Further let F (x) and R(x) have convexity parameters µF and
µR , respectively. In other words, if F (x) (R(x)) is strongly
where
x(st ) = proxηR (x(st−1 ) − η∇F (x(st−1 ) )).
Since ` and e(st−1 ) are independent of x(st ) and x? , and
since e(st−1 ) is zero-mean,
h
i
T
E ∆(st ) (x(st ) − x? ) = (E∆(st ) )T (x(st ) − x? ) = 0.
Further, since e(st−1 ) is independent of w(st−1 ) and
∇F (x(st−1 ) ),
Ek∆
(st ) 2
k = Ekw
(st−1 )
(st−1 )
− ∇F (x
2
)k + Eke
(st−1 ) 2
k
Applying Lemma 2, we obtain,
Ekx(st ) − x? k2 ≤ kx(st−1 ) − x? k2 − 2ηE(G(x(st ) ) − G(x? ))
+ 8Lη 2 (G(x(st−1 ) ) − G(x? ) + G(x̃(s) ) − G(x? ))
+ 2η 2 Eke(st−1 ) k2
We consider a single execution of the inner P
iteration of
T
the algorithm, so x(s0 ) = x̃(s) and x̃(s+1) = T1 t=1 x(st ) .
Summing over t = 1, . . . , T on both sides gives and taking
expectation over `, for t = 1, . . . , T gives us,
Ekx(sT ) − x? k2 + 2ηE(G(x(sT ) ) − G(x? ))
+ 2η(1 − 4Lη)
≤ kx
(s0 )
T
−1
X
E(G(x(st ) ) − G(x? ))
t=1
? 2
− x k + 8Lη 2 (G(x(s0 ) ) − G(x? )
+ T (G(x̃(s) ) − G(x? )) + 2η 2 Γ(s) .
Following the same reasoning as in [10], we obtain,
h
i
E(G(x̃(s+1) ) − G(x? )) ≤ αE G(x̃(s) − G(x? ) + βΓ(s) .
Applying this bound recursively, we obtain the expression in
our theorem.
R EFERENCES
[1] Y. Pu, M. N. Zeilinger, and C. N. Jones, “Quantization design for
unconstrained distributed optimization,” in American Control Conference, July 2015, pp. 1229–1234.
[2] M. Schmidt, N. L. Roux, and F. R. Bach, “Convergence rates of inexact
proximal-gradient methods for convex optimization,” in Advances in
Neural Information Processing Systems, J. Shawe-Taylor, R. Zemel,
P. Bartlett, F. Pereira, and K. Weinberger, Eds., 2011, pp. 1458–1466.
[3] Y. Pu, M. N. Zeilinger, and C. N. Jones, “Quantization design for
distributed optimization,” arXiv preprint arXiv:1504.02317, 2015.
[4] A. Kashyap, T. Başar, and R. Srikant, “Quantized consensus,” Automatica, vol. 43, no. 7, pp. 1192–1203, 2007.
[5] D. Thanou, E. Kokiopoulou, Y. Pu, and P. Frossard, “Distributed
average consensus with quantization refinement,” IEEE Transactions
on Signal Processing, vol. 61, no. 1, pp. 194–205, 2013.
[6] R. Carli, F. Fagnani, P. Frasca, T. Taylor, and S. Zampieri, “Average
consensus on networks with transmission noise or quantization,” in
Proceedings of European Control Conference, 2007, pp. 1852–1857.
[7] A. Nedić, A. Olshevsky, A. Ozdaglar, and J. N. Tsitsiklis, “Distributed
subgradient methods and quantization effects,” in 47th IEEE Conference on Decision and Control, 2008, pp. 4177–4184.
[8] A. Nitanda, “Stochastic proximal gradient descent with acceleration
techniques,” in Advances in Neural Information Processing Systems,
2014, pp. 1574–1582.
[9] R. Johnson and T. Zhang, “Accelerating stochastic gradient descent using predictive variance reduction,” in Advances in Neural Information
Processing Systems, 2013, pp. 315–323.
[10] L. Xiao and T. Zhang, “A proximal stochastic gradient method
with progressive variance reduction,” SIAM Journal on Optimization,
vol. 24, no. 4, pp. 2057–2075, 2014.
[11] J. Duchi and Y. Singer, “Efficient online and batch learning using
forward backward splitting,” J. Mach. Learn. Res., vol. 10, pp. 2873–
2898, 2009.
[12] S. P. Lipshitz, R. A. Wannamaker, and J. Vanderkooy, “Quantization
and dither: A theoretical survey,” Journal of the Audio Engineering
Society, vol. 40, no. 5, pp. 355–375, 1992.
| 3cs.SY
|
1
Evaluating the Performance of eMTC and
NB-IoT for Smart City Applications
Mohieddine El Soussi, Pouria Zand, Frank Pasveer and Guido Dolmans
Holst Centre/imec, Eindhoven, The Netherlands
arXiv:1711.07268v1 [cs.IT] 20 Nov 2017
e-mail:{mohieddine.elsoussi, pouria.zand, frank.pasveer and
guido.dolmans}@imec-nl.nl
Abstract
Low power wide area network (LPWAN) is a wireless telecommunication network that is designed
for interconnecting devices with low bitrate focusing on long range and power efficiency. In this
paper, we study two recent technologies built from existing Long-Term Evolution (LTE) functionalities:
Enhanced machine type communications (eMTC) and Narrow band internet of things (NB-IoT). These
technologies are designed to coexist with existing LTE infrastructure, spectrum, and devices. We first
briefly introduce both systems and then compare their performance in terms of energy consumption,
latency and scalability. We introduce a model for calculating the energy consumption and study the
effect of clock drift and propose a method to overcome it. We also propose a model for analytically
evaluating the latency and the maximum number of devices in a network. Furthermore, we implement
the main functionality of both technologies and simulate the end-to-end latency and maximum number
of devices in a discrete-event network simulator NS-3. Numerical results show that 8 years battery
life time can be achieved by both technologies in a poor coverage scenario and that depending on the
coverage conditions and data length, one technology consumes less energy than the other. The results
also show that eMTC can serve more devices in a network than NB-IoT, while providing a latency that
is 10 times lower.
I. I NTRODUCTION
The internet of things (IoT) interconnects a massive number of devices, e.g., machines, vehicles
and sensors, to the internet and exploits the data that is generated. The IoT world is growing
fast, from 2 billion objects in 2006 to a projected 200 billion by 2020 [1]. IoT is expected to
create steady growth in the economy since these smart objects give vital data to track inventory,
manage machines, increase efficiency, save costs, and even save lives.
2
While many IoT devices will be served by short-range radio technologies that operate on
an unlicensed spectrum such as WiFi, ZigBee and Bluetooth, a significant proportion will be
enabled by wide area networks (WANs) [2]. Low power wide area network (LPWAN) is a type
of wireless telecommunication intended for wireless battery operated things and designed to
allow long range communications at a low bit rate. Currently, there are two connectivity tracks
for LPWAN, one operating on an unlicensed spectrum such as SigFox and LoRa [3] and one
operating on a licensed spectrum such as Cellular IoT.
The third generation partnership project (3GPP) has introduced two Cellular IoT technologies
based on Long Term Evolution (LTE), namely: eMTC (enhanced Machine Type Communications) or CAT-M1 [4], and Narrowband Internet of Things (NB-IoT) or CAT-NB1 [5]. These
systems are designed to coexist with existing LTE infrastructure, spectrum, and devices. eMTC
targets applications such as VoLTE (Voice over Long-Term Evolution), tracking devices and
objects that require mobility, high data rate and low power consumption with wide area coverage.
NB-IoT targets applications such as control equipment, sensors, and meters that require low
complexity and low power consumption with wide area coverage.
In smart cities, a massive number of IoT devices will be deployed for several use cases. Key
challenges to enabling a large-scale uptake of massive IoT include: device costs, battery life,
scalability, latency and coverage. In this paper, we consider some of these challenges while
studying the energy consumption, the latency and scalability for both technologies.
In order to ensure long battery lifetime and to reduce energy consumption, these technologies
are enabled with two power saving features: extended discontinuous reception (eDRX) and power
saving mode (PSM) [6]. eDRX is a mechanism that enables the device to switch off part of its
circuitry to save power. An eDRX cycle consists of an “On Duration” during which the device
checks for paging and an “eDRX period” during which the device is in sleep mode. This feature
is useful for device-terminated applications, e.g., smart grid. PSM is a low-power mode that
allows the device to skip the periodic page monitoring cycles, allowing the device to sleep for
longer. However, as a result the device becomes unreachable. It is therefore best utilized by
device-originated or scheduled applications, e.g., smart metering.
The energy consumption of both eMTC and NB-IoT devices has been studied in, for example,
[7], [8], where a highly accurate expensive clock has been considered for deep sleep periods.
However, since another key requirement of these technologies is low cost, it is beneficial to
consider a low cost, low power clock for deep sleep periods at the expense of having a clock
3
drift. Clock drift causes the device to loose its time and frequency synchronization. Furthermore,
depending on the length of the sleep period and the amount of time and frequency drift, more
time is needed to synchronize. Hence, the device must wake up earlier to synchronize and to
not miss its scheduled transmission or reception. This increases the energy consumption of the
device since the accurate clock is turned on for additional duration. In this paper, we study
this problem and reduce the energy consumption by allowing the device to wake up several
times during the sleep period to synchronize. We formulate an optimization problem that aims
to evaluate the sleep duration and the number of waking up times in order to minimize the total
consumed energy during the sleep period.
Other key requirements for these technologies are scalability and latency. In general, IoT
devices are deployed on a large scale and in order to a priori study the network performance,
expensive test-beds are required. Network simulators such as NS-3 offer an alternative solution
to those expensive test-beds to test and evaluate the performance of large-scale networks. Based
on the existing LTE module in NS-3, we implement eMTC and NB-IoT modules in NS-3,
where we adapt both the physical (PHY) layer and the media access control (MAC) layer
according to eMTC and NB-IoT standards. In the PHY layer, we modify the PHY error model
by introducing the link-level results of the convolution code for different modulation and coding
schemes (MCSs). In addition, we limit the number of resource blocks (RBs) to 6 RBs and 1 RB
for eMTC and NB-IoT, respectively. In the MAC layer, we include switching and cross subframe
delays for the control and data channels according to 3GPP specifications. The latency based on
system level has been studied and evaluated in [8] and [9]. In this paper, in addition to system
level evaluation, the latency and scalability are also evaluated at the network level using NS-3.
Thus, we estimate the end-to-end latency from sensors/actuators to the cloud as well as the
maximum number of devices that can be served by a cell. In addition to the simulation results,
we formulate the latency and the maximum number of sensors based on the channel quality, the
maximum coupling loss (MCL) and the transmission period.
Through simulation results, we show that at least 8 years battery life time can be achieved
by both technologies in a poor coverage scenario taking into account the clock drift. We also
show that depending on the coverage conditions and data length, one technology consumes less
energy than the other and that eMTC can serve more devices in a network than NB-IoT, while
providing a latency that is 10 times lower. We also formulate the data transmission delay and
the maximum number of UEs per cell or network and demonstrate that the results derived from
4
analytical formulations correspond to the results obtained from the network simulator.
II. S YSTEM D ESIGN OVERVIEW
In this section, we briefly provide an overview of eMTC and NB-IoT with a focus on the key
aspects where they deviate from LTE.
eMTC is an evolution of LTE optimized for IoT. It has introduced a set of physical layer
features that aim to reduce cost and power consumption while extending coverage [10]. eMTC
devices operate with a bandwidth of 1.08MHz (6 LTE Physical Resource Blocks (PRBs)) for the
transmission and reception of physical channels and signals. The downlink of eMTC is based
on orthogonal frequency division multiplexing (OFDM) scheme with 15kHz subcarrier spacing
as in LTE. In the uplink, eMTC also uses the same numerology as LTE. The transmission is
based on single carrier frequency division multiple access (SC-FDMA) with 15kHz subcarrier
spacing. The reader can refer to [10] for more details.
NB-IoT is a new 3GPP radio-access technology built from LTE functionalities. It is not
backward compatible with existing 3GPP devices. It is, nevertheless, designed to coexist with
LTE. NB-IoT occupies a frequency band of 180kHz, which is further decreased compared to
eMTC, and corresponds to one PRB in LTE transmission. This narrowband bandwidth allows
the device complexity to be further reduced at the expense of lower data rate. The downlink
of NB-IoT is based on OFDM with 15kHz subcarrier spacing as in LTE (the PRB contains
12 subcarriers). In the uplink, NB-IoT also uses the same numerology as LTE, however NBIoT supports not only multi-tone but also single-tone transmissions. A single-tone transmission
supports 3.75kHz and 15kHz subcarrier spacing. The 3.75kHz numerology uses 2ms slot duration
instead of 0.5ms, and the PRB contains 48 subcarriers instead of 12 subcarriers to remain
compatible with the LTE numerology. The 15kHz numerology is identical to LTE. In the
multi-tone transmissions, 3, 6, or 12 tones with 15kHz subcarrier spacing can be used and
the transmission is based on SC-FDMA scheme [11]. The reader can refer to [12] for more
details.
III. P OWER C ONSUMPTION A NALYSIS
Low power consumption is a key requirement for these technologies to ensure years of battery
lifetime. In Figure 1, we present a model for calculating the energy consumption of an LTE device
5
as described in [5]. The model shows the different events performed by the device to transmit
a packet and the corresponding power consumption of each event.
The model considers four different operating states: transmission state, receiving state, idle
state and sleeping state. In this model, we assume that the device transmits with a reporting
interval equal to ttot . To transmit a report, the device needs to obtain downlink synchronization
(Synch period), acquire the system information (PBCH period) and achieve the uplink synchronization (random access period). Once the random access procedure is completed, the device
waits for an uplink grant on the physical downlink control channel (PDCCH), transmits the
report and then waits for the network and application acknowledgement. The device monitors
the control channel according to its DRX cycle until the ready timer expires, and then it either
continues to monitor the control channel according to its idle eDRX cycle or sleeps for a longer
period according to its PSM (Standby period).
During the eDRX cycle, the receiver is switched on for a duration of tPDCCH to check for
paging and then goes back to sleep. During the sleep period, most of the device circuitries
are switched off, including the high accurate crystal clock and only the very low power clock
is enabled. Unfortunately, this clock is inaccurate and causes the device to loose its time and
frequency synchronization. As a result, the device might wake up late and miss the scheduled
paging, transmission or reception. Hereafter, we assume that the error due to clock drift only
depends on the length of the sleep period. Hence, the device’s circuitry must be switched on
m ∗ tsleep earlier to synchronize with the network where m is the fractional error. This increases
the power consumption and thus decreases the lifetime of the battery. We should note that the
accurate clock is switched on during the synchronization period and it is switched off only
during the sleep period. In order to reduce the power consumption, the receiver, during the off
period of the eDRX cycle, wakes up and then goes back to sleep several times to obtain the
time and frequency synchronization. Following this, the high accurate clock stays switched on
for a duration of tact = m ∗ tsleep after the last synchronization and before receiving the control
channel (PDCCH) as shown in Figure 1. In order to minimize the total energy consumed during
an eDRX period, the sleeping duration should be large, the number of waking up should be small
and the active time should be also small. However, this is not achievable since the active time
is proportional to the sleeping time. To overcome this problem, the device will have different
sleeping duration values. The first sleeping duration is the largest to reduce the number of waking
up for synchronizing and the following ones are smaller so that tact is minimized. Hence, the
6
optimization problem to reduce the energy consumption during an eDRX period can be stated
as:
min
k
K, tK
act , tsleep
KPRx tsynch +
K
X
Psleep tksleep + PIdle tK
act
(1a)
k=1
s.t. teDRX − tPDCCH − Ktsynch −
K
X
tksleep = tK
act
(1b)
k=1
K
tK
act = mtsleep
(1c)
tksleep > 0,
(1d)
∀k,
where tsynch is the synchronization time, tPDCCH is the time to check for paging, K is the number
of cycles or iterations, tksleep is the sleeping time at cycle k, tK
act is the active time during the last
cycle and PRx , Psleep and PIdle are the power value for the receiving, sleeping and idle states,
respectively.
The optimization problem is a mixed integer optimization problem and it is hard to find K,
K
{tksleep }K
k=1 and tact jointly in a reasonable time. Therefore, we propose an iterative optimization,
K
where we find appropriate {tksleep }K
k=1 and tact , and K, alternately. More specifically, for a given
K
value of K, we can compute {tksleep }K
k=1 and tact using (1). It can be easily seen from (1a) that the
energy consumption is minimized for small K and tK
act . Thus, the optimization starts by setting
K
K = 1 and optimizing over {tksleep }K
k=1 and tact . Next, we increment K by one and optimize again
K−1
k
K
over tK
sleep and tact given the values of {tsleep }k=1 from the previous iterations. The iterative process
K
K
terminates if either of these conditions hold: tK
act < tsynch or PRx tsynch + Psleep tsleep + PIdle tact >
k
K
PIdle tK−1
act . For the first condition, the minimum energy consumption is attained using {tsleep }k=1
K−1
k
and tK
act and for the latter case the minimum energy consumption is attained using {tsleep }k=1
and tK−1
act .
The above optimization can also be used for PSM by simply replacing teDRX with tPSM and
equating tPDCCH to 0. The main difference is that during the standby period there exist multiple
eDRX and only one PSM.
IV. L ATENCY AND S CALABILITY M ODEL
In this section, we introduce a simple model to formulate the data transmission delay and the
maximum number of users per cell.
7
Power
Reporting period
DRX
Random access
eDRX
HARQ ACK
Ack
PDCCH
Data
HARQ
msg4
PDCCH
PDCCH
msg3
msg2
PRACH
PDCCH
PBCH
Synch
. . .
PDCCH
Synch
Active period
. . .
PDCCH
Synch
Synch
PDCCH
Sleep
PDCCH
Idle
Sleep period
Ready timer
PDCCH
Rx
HARQ ACK
Tx
time
Standby period
Fig. 1: Different states of the power model and the resulting power levels.
Latency consists of synchronization delay, random access channel (RACH) procedure delay
and data transmission delay. Herein, we only focus on formulating the data transmission delay.
The synchronization and RACH procedure delays for different MCLs are provided in [9]. The
data transmission delay per user equipment (UE) consists of the reception of downlink control
information (DCI), transmission of data and transmission or reception of the acknowledgment.
The data transmission delay per UE for the downlink (DL) and the uplink (UL) transmissions
are given as follows
DelayiUE
DataLen
, i ∈ {DL, UL}.
= TLi ∗
TBS(MCS, RBU)
(2)
It can be easily seen that the transmission delay per user depends on the total number of transport
l
m
DataLen
blocks needed to transmit the user data, i.e., TBS(MCS,RBU)
and the transmission latency per
transport block, TLi , where DataLen is the data size per user, TBS is the transport block size
that depends on MCS and the allocated RB per user (RBU), and TLi on the UL and DL are
given by
TLDL = RLDC ∗ tPDCCH + tD + RLDS ∗ tPDSCH
+ tDUS + RLUC ∗ tULACK
(3)
TLUL = RLDC ∗ tPDCCH + tDUS + RLUS ∗ tPUSCH
+ tUDS + RLDC ∗ tDLACK ,
(4)
where RLDC is the number of repetitions of the downlink control channel (PDCCH). RLDS
and RLUS are the number of repetitions of the data on the physical downlink shared channel
(PDSCH) and the physical uplink shared channel (PUSCH), respectively. tPDCCH is the transmission time needed to transmit the control information on the PDCCH, tPDSCH and tPUSCH are the
8
transmission times needed to transmit one transport block on PDSCH and PUSCH, respectively.
tD is the cross subframe delay, tDUS and tUDS are the radio frequency (RF) tuning delay for
switching from DL to UL and UL to DL channels, respectively. We should note that the values
of RLDC, RLDS and RLUS depend on MCL. The data transmission latency for uplink and
downlink are illustrated in Figures (2a) and (2b) for eMTC and NB-IoT, respectively. In these
figures, the vertical axis represents the spectrum while the horizontal one represents the time.
The total data transmission delay in a cell can be formulated as
'
&
NUE
i
i
DelayT = DelayUE ∗ NRB ,
(5)
RBU
where NUE is the total number of users, NRB is the total number of RBs and RBU is the number
of allocated RB per user.
NRB
We should note that
RBU represents the number of users in a group that can concurrently
NUE
represents the total number of groups.
transmit and
RB
c
b NRBU
eMTC TLDL
1 RB
NPDSCH to NPUSCH
NPUSCH
NPDCCH to NPDSCH
tDUS ≥ 12ms
NB‐IoT TLDL
NPDCCH to NPUSCH
eMTC TLUL
NB‐IoT TLUL
(a)
(b)
tUDS ≥ 3ms
NPUSCH to
NPDCCH
1 RB
tDUS ≥ 8ms
NPDCCH
(Ack.)
(Data)
NPUSCH
MPUSCH to
MPDCCH
(DCI)
NPDCCH
tUDS ≥ 3ms
MPDCCH
MPUSCH
MPDCCH
tDUS =
3ms
tD ≥ 4ms
(Ack.)
(Data)
to MPUSCH
MPDCCH
1‐6 RBs
(DCI)
(Ack.)
(Data)
NPDSCH
MPDSCH to
MPUCCH
(DCI)
NPDCCH
tDUS ≥ 3ms
MPUCCH
tD =
1ms
MPDSCH
MPDCCH
(Ack.)
(Data)
MPDCCH
to MPDSCH
1‐6 RBs
(DCI)
Fig. 2: Transmission latency for (a) eMTC and (b) NB-IoT on the downlink and uplink.
In order to evaluate the scalability of the network, we calculate the maximum number of users
per cell, which can be formulated as follows
Reporting Period
NRB
max NUE =
∗
,
RBU
DelayiUE
(6)
9
where
j
Reporting Period
DelayiUE
k
represents the total number of groups of transmission that can occur during
a reporting period. In (6), we assume that the users have the same reporting period.
We should note that the scheduler is able to adapt the number of repetitions and the MCS,
based on the channel quality of the link or MCL, in such a way that the block error rate is
minimized at the expense of increasing the transmission air-time of the packets. We should
also note that the RF tuning delay between control and data channels is needed to reduce the
complexity and cost of the radio module and to support longer decoding time at the modules.
Hence, the above equations can be used to roughly estimate the data transmission delays and
the maximum number of users that can be served by a cell with a very low complexity.
V.
E MTC AND
NB-I OT IMPLEMENTATIONS IN NS-3
In this section, we provide some insights into the implementation of eMTC and NB-IoT in
NS-3.
NS-3 has a complete LTE module that includes EPC control and data plane, an end-to-end
data-control plane protocol stack, a physical-layer data error model, radio propagation models
and different MAC schedulers. Hence, based on the LTE module in NS-3, we implement eMTC
and NB-IoT modules where we adapt the LTE module according to eMTC and NB-IoT standards.
We modify the PHY layer, MAC layer, and the scheduler according to 3GPP specifications.
In NS-3, a link abstraction model is used to provide an accurate link performance metric
at a low computational cost [13]. This model is based on link-level results obtained under the
configuration of the PHY-layer turbo encoder in terms of code block length and MCS. In order
to evaluate the PHY layer performance of downlink data transmission in NB-IoT, we adapt the
model and include the link-level results of convolution codes in terms of MCS as shown in
Figure 3. In addition, we adapt the number of RBs that is supported by the users for eMTC
(1-6 RBs) and NB-IoT (1 RB). We also limit the DL and UL modulation schemes for eMTC to
16QAM and NB-IoT to QPSK.
Moreover, we dedicate separate subframes for control and data channels. In LTE, both control
and data channels are transmitted in the same subframe, while in eMTC and NB-IoT separate
subframes are required. We also introduce cross subframe delays and RF switching delays
according to 3GPP specifications. In future work, we intend to establish a complete network
simulator, implement single-tone NB-IoT and publish the source code with a public license
GNU in a public repository.
10
100
10-1
MCS-12
BLER
MCS-0
10-2
10-3
10-4
-15
-10
-5
0
5
10
15
SNR (dB)
Fig. 3: NB-IoT BLER curves used in NS-3
VI. S IMULATION RESULTS
In this section, we provide some numerical examples where we evaluate the battery life time
of the devices, the latency and the maximum number of devices that can be served in a network.
A. Power consumption
The power parameters used to evaluate the energy consumption of the devices are based on
the data sheet of the ublox SARA-N2 module, where PTx = 792 mW (at transmitted power of
23dBm), PRx = 72 mW, PIdle = 22 mW and PSleep = 18 µW. The duration parameters are based
on the results of [9]. We assume that the battery capacity is 18000J (5Wh). Hereafter, we do
not consider battery leakage impact nor temperature effect nor the peak currents situation.
Figure 4 depicts the battery life time obtained for different clocks: perfect crystal clock “Perfect
XC”, low power crystal clock “Low power XC” that has an error of m = 0.01% and optimized
using (1), optimized low cost clock “Opti low cost clock” that has an error of m = 0.1% and
optimized using (1) and unoptimized low cost clock “Unopti low cost clock” that has an error
11
of m = 0.1%. For the unoptimized low cost clock, the device simply wakes up m ∗ tsleep earlier.
The battery life time is evaluated for NB-IoT at 164dB MCL.
In Figure 4, the battery life time is taken as function of eDRX in which we assume that the
device is going to transmit a report of 200 bytes with a reporting interval ttot = 24hrs. We observe
that the “Opti low cost clock” yields more battery life time than the “Unopti low cost clock” at
an eDRX higher than 12min. We also observe that “Opti low cost clock” yields similar battery
life time as “Low power XC” at low and high eDRX and maximum one year less than the “Low
power XC” at eDRX of 12min. However, the “Opti low cost clock” is, in general, much cheaper
than the “XC” making the module cheaper and more suitable for IoT. Moreover, it can be easily
seen that with a larger sleeping period, i.e. using PSM, the “Opti low cost clock” yields a battery
life time similar to that of “Low power XC”. We should note that in general a low cost clock
consumes less energy, however, in this example, we assume that all clocks consume the same
energy.
Figure 5 depicts the battery life time obtained using “Opti low cost clock” under different data
lengths and coverage scenarios for eMTC and NB-IoT. The good coverage scenario corresponds
to the situation where the node is outdoor (close to the base station) and has the highest MCS
value and zero repetition whereas the poor coverage scenario corresponds to the situation where
the node is deep indoor (far from the base station) and has the lowest MCS value and the highest
repetition value. The battery life time is taken as function of the reporting number per day. We
should note that in this figure the PSM technique is being used. We should also note that a
bandwidth of 6 RBs and 1 RB is assumed for eMTC and NB-IoT, respectively. We also assume
that PTx and PRx of eMTC are 1.25 times larger than the ones of NB-IoT.
In Figure 5, we observe that for small data length NB-IoT has a larger battery life time. We
also observe that in a good coverage scenario NB-IoT is slightly higher and in a poor coverage
scenario NB-IoT yields a maximum of 2.5 additional years over eMTC. For large data length,
we observe that in a good coverage scenario eMTC yields more life time whereas in a poor
coverage scenario NB-IoT yields more life time.
In general, the energy consumption or battery life time during a reporting period depends on
MCL, data length, bandwidth, RF module and latency. Thus, depending on the values of these
parameters, the energy consumption of one technology will be higher or lower than the other.
For instance, in a good coverage scenario, even though the RF module of eMTC consumes more
energy than the NB-IoT one, eMTC energy consumption is lower if the data length is large. This
12
Battery life factor in yrs with ttot = 24 hours
8
7
6
5
4
3
2
Perfect XC
Low power XC
Opti low cost clock
Unopti low cost clock
1
0
0
20
40
60
80
100
120
140
160
180
eDRX cycle (min)
Fig. 4: Battery life time under different eDRX values for NB-IoT at 164dB MCL.
is due to the larger bandwidth of eMTC compared to NB-IoT, which reduces the transmission
time and allows the device to go into sleep mode earlier. However, if the data length is small, the
energy consumption of NB-IoT is lower. In a poor coverage scenario, NB-IoT is more efficient,
especially in a single-tone transmission. Consequently, its energy consumption is lower than the
one of eMTC, which needs a very high number of repetitions. Hence, NB-IoT is a good fit for
simple sensors and low-rate applications in medium to poor coverage scenarios, while eMTC is
best fit for applications transmitting large amount of data in good to medium coverage scenarios.
We should note that a battery life time of almost 10 years can be achieved with the current
technology in a poor coverage scenario and with a reporting period of 24 hrs. However, for more
frequent transmission, e.g., 10 reports per day, the battery life time is decreased to almost one
year, which is not practical for IoT modules. Hence, more efforts should be taken to reduce the
power consumption of the RF modules, e.g., power amplifier.
13
25
eMTC : 10 Bytes, Good Coverage
eMTC : 10 3 Bytes, Good Coverage
eMTC : 10 Bytes, Poor Coverage
Battery life factor in yrs
20
eMTC : 10 3 Bytes, Poor Coverage
NB-IoT: 10 Bytes, Good Coverage
NB-IoT: 10 3 Bytes, Good Coverage
NB-IoT: 10 Bytes, Poor Coverage
15
NB-IoT: 10 3 Bytes, Poor Coverage
10
5
0
0
5
10
15
20
25
30
Reporting Number per day
Fig. 5: Battery life time under different reporting number, data lengths and coverage scenarios
for eMTC and NB-IoT.
B. Latency and scalability
In order to evaluate the latency and scalability of a network in a smart city, we consider a
simple city model that has various types of environments, e.g., urban, suburban and open area,
and a varied number of buildings with different inter-building distances as shown in Figure
6. In this city, seven evolved NodeB (eNB) are deployed to serve the UEs, which are placed
randomly over the 2D grid and shown by a white point in Figure 6. The figure also shows the
signal-to-interference-plus-noise ratio (SINR) values at the UEs.
Based on this network setup and the NS-3 simulation parameters shown in Table I, we evaluate
the latency of the received packets at the cloud and the maximum number of devices that can be
served in the different areas. We assume that 92% of the eNBs bandwidth is allocated to legacy
LTE, 6% to eMTC and 2% to NB-IoT.
We assume that the number of deployed UEs or devices in each cell can scale up to 1800.
In the simulation, we set the propagation model to Hybrid Buildings for indoor and outdoor
14
5000
70
6 buildings
(w=100m)
4000
60
3000
50
Sub Urban
1000
Urban
40
0
30
-1000
20
6 buildings
(w=25m)
-2000
-3000
10
Open Area
0
-4000
-5000
-5000
SINR (dB)
Y (meter)
2000
-10
-4000
-3000
-2000
-1000
0
1000
2000
3000
4000
5000
X (meter)
Fig. 6: Radio environment map (REM) of a simple city with urban, suburban and open-area.
communications that includes the Hata model, COST231, ITU-R P.1411 (short range communications) and ITU-R P.1238 (indoor communications). All these propagation models are combined
in order to be able to evaluate the path loss in different environments (urban, suburban and open
area). In addition, we consider an uplink traffic model and we configure each UE to randomly
send a packet to the cloud every couple of minutes.
Figure 7a depicts the average end-to-end latency per user on the UL for both eMTC and NBIoT for a given data length. The latency is taken as function of the ratio of the number of indoor
UEs to the total number of available UEs. We assume that all UEs choose a random instant,
within the given reporting period, to initiate a transmission. We observe that, for all areas, the
delay of transmitting a packet (12 or 160 Bytes) in eMTC is lower than the delay of transmitting
a packet (12 Bytes) in NB-IoT. This occurs because eMTC is using a larger bandwidth (6 RBs)
than NB-IoT (1 RB), and the cross subframe delays and RF tuning of eMTC are lower than the
ones of NB-IoT.
15
TABLE I: NS-3 simulation parameters
Parameters
Number of UEs and eNBs
Propagation model
External Wall Loss
Frequency band
System Bandwidth (NRB )
Uplink traffic period
Sector per cell
Antenna model
Fractional frequency reuse
Scheduler
App. packet size
Allocation size
Resource Unit (RU)
eNB and UE Tx power
OpenArea (good coverage)
SubUrban (med. coverage)
Urban (poor coverage)
Values
300-1800 per cell and seven eNBs
Hybrid Buildings
Concrete wall no windows (15dB)
DL: 925MHz, UL: 880MHz
6PRBs (eMTC), 2PRBs (NB-IoT)
< 1 minute
Three
Parabolic
Hard (FFR)
Round Robin
12 Byte and 160 Byte (IP packet)
12 tones @ 15 KHz
1 ms
46 dBm and 20 dBm
0-33% indoor,100-67% outdoor UEs
33-66% indoor,66-33% outdoor UEs
66-100% indoor,33-0% outdoor UEs
We also observe that, for all areas, as the number of UEs increases the end-to-end latency per
user increases. This occurs because as the number of UEs increases, the probability of waking
up and asking for the uplink resources at the same time is higher. The situation aggravates in
urban areas (more UEs installed indoor with poor coverage) since the UEs’ transmission time
increases due to larger MCL.
Figure 7b depicts the total airtime of all users on the UL for both eMTC and NB-IoT for a
given data length. It also depicts the total transmission delay obtained in (5). We assume that
all users wake up at the same instant and simultaneously request uplink resources from eNB.
Herein, we use the round robin (RR) scheduler, i.e., the resources are allocated to each user
in equal portions and in circular order. The total airtime is calculated from the instant the first
packet of the first user is transmitted until the last packet of the last user is delivered. The total
airtime is taken as function of the ratio of the number of indoor UEs to the total number of
available UEs. We observe that the total airtime of all users in eMTC is less than the one in
NB-IoT. We also observe that the total data transmission time obtained in (5), is similar to the
results obtained in NS-3. We should note that in indoor scenarios more repetitions and smaller
TBS are required, which explains the higher total air-time.
Figure 7c depicts the maximum number of UEs that can be supported in different environments
of a network. It also depicts the maximum number of UEs obtained in (6). We assume that all UEs
choose a random instant, within the given reporting period, to initiate a transmission. We should
eMTC 160B
NB-IoT 12B
0.8
0.6
0.4
0.2
0.0
Total number of UEs
OpenArea OpenArea SubUrban SubUrban
(0%)
(25%)
(38%)
(63%)
Urban
(75%)
14
12
eMTC 12B
eMTC 160B
NB-IoT 12B
Analytical model
10
8
6
4
2
0
200
500
600
200
500
600
200
500
600
200
500
600
200
500
600
200
500
600
eMTC 12B
Total airtime of all users (Sec.)
1.0
200
500
600
200
500
600
200
500
600
200
500
600
200
500
600
200
500
600
Avg. E2E latency per user (Sec.)
16
Total number of UEs
OpenArea OpenArea SubUrban SubUrban
(0%)
(25%)
(38%)
(63%)
Urban
(100%)
Urban
(75%)
Urban
(100%)
Areas (Ratio of indoor-to-total number of UEs)
Areas (Ratio of indoor-to-total number of UEs)
(a)
(b)
>
(c)
Fig. 7: (a) Average E2E UL latency per user, (b) Total UL air-time of all users, (c) Maximum
number of supported users
note that this figure only displays the number of UEs who have more than 90% successfully
delivered packets at the eNB/cloud in the given reporting period. Due to memory and process
limitation in the simulation, the maximum number of UEs that can be simulated is limited to
600. We should also note that the maximum number of UEs is constrained by the reporting
period and the percentage of indoor UEs. We can observe in Figure 7c that eMTC can support
more users than NB-IoT. For instance, eMTC at 100% of indoor UEs, can serve 500 UEs (with
a 160 Bytes data length) while for NB-IoT this is not the case. In the latter, when the number
of indoor UEs is above 400, the airtime is high and the eNB/scheduler, that has only 2 RBs,
cannot guarantee anymore the delivery of UEs’ reports in a timely manner. Thus, the request of
some UEs is discarded. However, in eMTC, since the airtime is low, the eNB/scheduler, that has
6 RBs, can deliver all 500 UEs’ reports in a timely manner. Moreover, we can observe that the
maximum number of users obtained in (6) also decreases whenever there are more indoor users
than outdoor. We should note that the results obtained using the analytical model are slightly
larger than the simulated ones. This is because in the analytical model, unlike in NS-3, we
assume that there are no collisions among the users requests and that the reporting period is
optimally divided among the UEs.
17
VII. C ONCLUSION
In this paper, we study two recent LPWAN technologies, eMTC and NB-IoT, that are based
on LTE and we evaluate their performance for smart city applications. We evaluate the battery
life time of both technologies considering the clock drift. We show that a battery life time of
8 years can be achieved by both technologies in a poor coverage scenario with a reporting
interval of one day. We also formulate the data transmission delay and the maximum number of
UEs per cell or network and show that these metrics depend on MCL, channel bandwidth, data
length, RF switching delay, cross subframe delay as well as the reporting period. In addition,
we introduce a network simulator tool that can evaluate the end-to-end latency and scalability
of eMTC and NB-IoT in a large-scale network in various environments. We show that eMTC
can serve more devices in a network than NB-IoT, while providing a latency that is 10 times
lower. We also demonstrate that the results obtained from the network simulator correspond to
the results derived from the analytical formulations. In future works, we intend to compare the
simulator results of eMTC and NB-IoT in NS-3 with field-test results.
R EFERENCES
[1] “A guide to the internet of things,” Available at https://www.intel.com/content/www/us/en/internet-of-things/infographics/
guide-to-iot.html.
[2] “LTE evolution for IoT connectivity,” Available at http://resources.alcatel-lucent.com/asset/200178 (2017/01/01).
[3] U. Raza, P. Kulkarni, and M. Sooriyabandara, “Low power wide area networks: An overview,” IEEE Communications
Surveys Tutorials, vol. 19, no. 2, pp. 855–873, Second quarter 2017.
[4] “Evolved Universal Terrestrial Radio Access (E-UTRA) and Evolved Universal Terrestrial Radio Access Network (EUTRAN); Overall description,” 3GPP TS 36.300, 2016.
[5] “Cellular System Support for ultra-low complexity and low throughput Internet of Things (CIoT),” TR 45.820, 2015.
[6] “Architecture enhancements to facilitate communications with packet data networks and applications,” 3GPP TS 23.682,
2016.
[7] M. Lauridsen, I. Z. Kovacs, P. Mogensen, M. Sorensen, and S. Holst, “Coverage and capacity analysis of lte-m and nb-iot
in a rural area,” in 2016 IEEE 84th Vehicular Technology Conference (VTC-Fall), Sept 2016, pp. 1–5.
[8] R. Ratasuk, B. Vejlgaard, N. Mangalvedhe, and A. Ghosh, “Nb-iot system for m2m communication,” in 2016 IEEE Wireless
Communications and Networking Conference, April 2016, pp. 1–5.
[9] R1-156006, “NB-IoT - battery lifetime evaluation,” 3GPP TSG RAN1#82bis, 2015.
[10] A. Rico-Alvarino, M. Vajapeyam, H. Xu, X. Wang, Y. Blankenship, J. Bergman, T. Tirronen, and E. Yavuz, “An overview
of 3gpp enhancements on machine to machine communications,” IEEE Communications Magazine, vol. 54, no. 6, pp.
14–21, June 2016.
[11] R. Ratasuk, N. Mangalvedhe, Y. Zhang, M. Robert, and J. P. Koskinen, “Overview of narrowband iot in lte rel-13,” in
2016 IEEE Conference on Standards for Communications and Networking (CSCN), Oct 2016, pp. 1–7.
18
[12] Y. P. E. Wang, X. Lin, A. Adhikary, A. Grovlen, Y. Sui, Y. Blankenship, J. Bergman, and H. S. Razaghi, “A primer on
3gpp narrowband internet of things,” IEEE Communications Magazine, vol. 55, no. 3, pp. 117–123, March 2017.
[13] M. Mezzavilla, M. Miozzo, M. Rossi, N. Baldo, and M. Zorzi, “A lightweight and accurate link abstraction model for the
simulation of LTE networks in ns-3,” in Proceedings of the 15th ACM International Conference on Modeling, Analysis
and Simulation of Wireless and Mobile Systems.
New York, NY, USA: ACM, 2012, pp. 55–60.
| 7cs.IT
|
A Basic Recurrent Neural Network Model
Fathi M. Salem
arXiv:1612.09022v1 [cs.NE] 29 Dec 2016
Circuits, Systems, and Neural Networks (CSANN) Lab
Department of Electrical and Computer Engineering
Michigan State University
East Lansing, Micigan 48824–1226
Email:[email protected]
Abstract—We present a model of a basic recurrent neural
network (or bRNN) that includes a separate linear term with
a slightly ”stable” fixed matrix to guarantee bounded solutions
and fast dynamic response. We formulate a state space viewpoint
and adapt the constrained optimization Lagrange Multiplier
(CLM) technique and the vector Calculus of Variations (CoV)
to derive the (stochastic) gradient descent. In this process, one
avoids the commonly used re-application of the circular chainrule and identifies the error backpropagation with the co-state
backward dynamic equations. We assert that this bRNN can
successfully perform regression tracking of time-series. Moreover,
the “vanishing and exploding” gradients are explicitly quantified
and explained through the co-state dynamics and the update
laws. The adapted CoV framework, in addition, can correctly
and principally integrate new loss functions in the network on
any variable and for varied goals, e.g., for supervised learning on
the outputs and unsupervised learning on the internal (hidden)
states.
I. I NTRODUCTION
The so-called simple Recurrent Neural Networks (sRNN)
have been reported to be difficult to train by the stochastic
gradient descent, see [1], [2] and the references therein.
This has spurred developments of new training approaches,
modifications, and more complex architectures to enable
convergent training and learning in state-of-the-art applications
[2], [3], [4]. Recent trends have focused on reducing the
computational load and complexity while preserving their
demonstrated learning capabilities of sequences and timeseries [3] and [5].
In this work, we define a basic recurrent neural network
(bRNN) as a model that can achieve regression (i.e., output
tracking) of real (i.e. analog) output values. Using the classical
Lagrange multiplier approach for non-convex constrained
optimization, we obtain a bounded input-bounded-output
(BIBO) stable flexible recurrent neural network that can
be trained to learn (discrete) categories or (continuous)
regression profiles. For ease of presentation, we adopt a
tutorial style and relegate most derivations to the appendices.
This bRNN has the following unique attributes:
•
Stable behavior without a need of additional “gating”
networks. Gating networks would at least double the
number of the overall adaptive parameters in the overall
RNN. In addition, no need for special caps on the growth
of the gradient signals.
•
•
•
Specify a predictive state space formulation. Specifically,
the network takes input data and state values at one
index (or time) value, say k and produces a state value
and/or output at the next index, say k + 1.
Update law derivations, including the backpropagation
network, are shown to follow easily using the Lagrange
Multiplier Constrained Optimization method. One
identifies the forward and backward propagation
networks and the split boundary conditions in a
principled way. The split boundary conditions are
the initial state and final co-state, and they are the
fundamental reason behind the state-forward and
costate-backward propagating dynamic processing for
the stochastic gradient descent.
In this framework, extensions and definitions of the
output or state loss functions can easily be incorporated
without complex re-derivations. E.g., each loss function
can be independently specified to be a supervised cost
(with a given reference/target), or an unsupervised loss
function, e.g., minimizing the entropy or maximizing
sparsity in an internal representation (state) vector.
II.
SIMPLE
R ECURRENT N EURAL N ETWORKS
The simple recurrent neural network (sRNN) model is
expressed as:
hk = σk (U hk−1 + W sk + b),
k = 0, ..., N
(1)
where k is the discrete (time) index, and N is the final finitehorizon time, sk is the m-d input vector sequence, and hk
is the n-d output via the nonlinear function σk . Here, σk is
a general nonlinear function which may be specified to be
the logistic function sigm or the hyperbolic tangent tanh as
is common in the literature (see [1]- [4], and the references
therein), or even as the rectified linear unit reLU as in [5].
The non-indexed parameters, to be determined by training, are
the n × n-matrix U , the n × m-matrix W , and the n × 1-vector
bias b. This model is a discrete nonlinear dynamic system, see,
e.g., [2].
Using variants of stochastic gradient descent to train the
sRNN for long-term dependencies or sequences, it has been
widely reported that the sRNN may exhibit the so-called
“vanishing gradient” or the “exploding gradient,” see [2], [5]
and the references therein.
The Long Short-Term Memory (LSTM) network was
developed to overcome difficulties in training sRNNs. They
have demonstrated their state-of-the-art performance in many
applications, e.g., in mapping input sequence to an output
sequence as in speech recognition, machine translations, and
language modeling, see [1-3]. In a nutshell, the three gating
control signals in the LSTM network, each roughly replicates
a sRNN and bring a four-fold increase in parameters in
comparison to parameters of the same state-size sRNN. More
recently, there has been interest in (i) determining the optimal
RNNs with decreased computational load and parameters.
e.g., the GRU architecture [2], reduces the gating signals to
two and thus parameters become three folds in comparison
to the parameters in the same state-size sRNN. Moreover,
the recent MGU architecture [4] further reduces the number
of independent control gates to one and thus the parameters
become two-folds in comparison to the parameters in the
same state-size sRNN. There is also interest in reviving
simple RNNs with modifications in nonlinearity-type, and an
approach to careful training which has introduced the IRRN
design, i.e., sRNN with rectified Linear Units (ReLUs) for
the nonlinearity, an initialization of the U matrix in eqn (1) at
the identity, and setting a cap limit on the resulting gradient
computation [5].
and is adaptively changing over the training iterations. That
′
is, U hk−1 may become unstable and may lead to unstable
behavior of the dynamic network during the training phase! If
′
U hk−1 is an unstable matrix (meaning that at least one of its
eigenvalues has a modulus greater than unity), the nonlinear
dynamic system may become unstable which may lead to
xk growing without bound. Even though hk is bounded
when using the compressive nonlinearity σk , the state xk can
become unbounded and thus the dynamic system may in fact
become unstable in the sense that xk grows unbounded. Of
course, Liapunov theory and Liapunov functions constitue
one approach to definitively confirm the stability of such a
nonlinear system.
Remark II.1: From a system and signal processing
viewpoints, the model in eqn (1), without the nonlinearity
(σk ), represents a filtering that can be expressed in the
(frequency) Z-transform as follows:
−1
(W S(z) + b)
H(z) = I − U z −1
III. BASIC R ECURRENT N EURAL N ETWORKS : B RNN–
= z [zI − U ]
−1
It would be beneficial to add a (stable) linear term in
the state eqn (2) that bounds the growth of xk to within a
bounded region. In Appendix I, we include the analysis based
on Liapunov theory to show that such dynamic system would
indeed be bounded to a region in the (xk -) state space. In
addition, one includes a linear output layer in the modeling
to allow for flexible filtering which is intended to render the
recurrent network with its output layer a general bandpass
filter. Such a filter would have the equivalent of poles and
zeros that produce, after training, a linearizable system with
an equivalent proper stable transfer function. These notions
are incorporated into the basic RNN model defined next.
THE GENERAL CASE
In contrast to the sRNN, see eqns (2)-(3), now consider this
extended recurrent neural network model:
xk+1 = Axk + U hk + W sk + b,
k = 0, ..., N − 1
(W S(z) + b)
In general, this represents a high-pass filtering, and thus
all high frequency input components, including noise, will
pass through to the output. This motivates one to revise
the indexing at the input and to include an output layer as
part of the model in order to realize a flexible filtering, and
specifically to be able to realize a bandpass filtering with
stable poles after training. This description of course is valid
only for linear systems, and it serves only as a motivation of
the design of the nonlinear system.
To motivate the proposed model in the next section, we
can now re-write this sRNN model as:
xk = U hk−1 + W sk + b,
hk = σk (xk ),
k = 0, ..., N
k = 0, ..., N
(2)
(3)
where we explicitly identified the state vector xk . Eqn (2)
captures the stability behavior. The (local) stability properties
′
of the network is determined by the matrix U hk−1 , where
′
hk−1 is the diagonal matrix derivative of the vector hk−1 .
However, the matrix U is typically initialized randomly small
hk = σk (xk ),
k = 0, ..., N
(4)
(5)
yk = V hk + Dsk + c,
k = 0, ..., N
(6)
where, as before, k = 0, ..., N , is the discrete (time) index, sk
is the m-d input vector sequence, xk is the n-d state vector,
hk is the n-d nonlinear function of the state vector which is
labeled as the hidden unit, and yk is the r-d corresponding output vector sequence. For this neural system, eqn (4) includes
the dynamic transition from the index step k (all variables
on the right-hand side have the same index k) to the state at
the next step k + 1. Eqn (5) represents the static nonlinear
state transformation which may include any of the common
nonlinearity, e.g., the logistic function, hyperbolic tangent,
reLU, etc.. Eqn (6) represents the static output equation.
This neural model extends the simple recurrent neural
networks (sRNN) by adding a linear state term to the dynamic
equation, eqn ((4), via the matrix A, and also by adding the
direct input term in the output equation, eqn (6), via the matrix
D. The dimensions of the parameters are obvious to achieve
compatibility of the equations. Specifically, A and U are n×n,
W is n × m, b is n × 1, V is r × n, D is r × m, and c is
r × 1.
In eqn (4), the state matrix A is set to be constant with
eigenvalues having moduli within (or on) the unit circle. As
a special case, one may choose the eigenvalues to be distinct,
real, or random, with amplitude ≤ 1. For large scale models, it
may be computationally easier to choose A = αI, 0 < α ≤ 1.
The state matrix A must be a stable matrix. In eqn (6), the
direct input term enriches the model since the state equation
can only produce transformations of delayed versions of the
input signal but not an instantaneous input.
The parameters in the dynamic equation, eqn (4), enumerated in the matrices U, W , and the bias vector b, can
be represented by a single vector parameter θ, whereas the
parameters of the output equation eqn (6), enumerated in
the matrices V, D, and the bias vector c, can be represented
by the single vector parameter ν. We shall also index these
parameters by the index k in order to facilitate the upcoming
analysis. Thus we shall represent the dynamic equations as
follows:
xk+1 = f k (xk , hk , sk , θk )
= Axk + Uk hk + Wk sk + bk ,
hk = σk (xk ),
k = 0, ..., N − 1
k = 0, ..., N
k
yk = g (hk , sk , νk )
= Vk hk + Dk sk + ck ,
k = 0, ..., N
where sk , k = 0, ..., N is the sequence of input signal, and
k is the time index of the dynamic system. In the case of
supervised adaptive learning, we associate with each sample
sk the desired sequence label dk .
The cost or loss function is thus given in general as a
function on the final time output yN as well as the variables
at the intermediate times within o ≤ k < N :
Jo (θo , ..., θN −1 , νo , ..., νN ) = φ(yN ) + LN (νN )
N
−1
X
Lk (yk , xk , hk , θk , νk )
k=o
where the general loss function includes all variables and
parameters.
Following the calculus of variations in constrained optimization and the Lagrange multiplier technique, see e.g., [6] and
[7], one can define the Hamiltonian at each time step k as
H k = Lk (yk , xk , hk , θk , νk )+(λk+1 )T f k (xk , hk , sk , θk ),
i = 1, ..., N − 1
where the sequence λk , 0 ≤ k ≤ N of the Lagrange multipliers would become the co-state with dynamics generated
as in constrained optimization and optimal control, see [6]
and [7]. The state equations and the co-state equations are
reproduced from optimization, similar to optimal estimation
and control procedures with minor adjustments. We apply
the (stochastic) gradient descent to all the parameters. Recall
that the parameters θk represent the elements of the matrices
Uk , Wk and bias vector bk . Similarly, the parameters νk
represent the elements of the matrices Vk , Dk and the bias
vector ck .
Thus, the state equations representing the network is reproduced as:
∂H k
= fk
xk+1 =
∂λk+1
= Axk + Uk hk + Wk sk + bk , i = 0, ..., N − 1
(7)
where the output layer is
yk = Vk hk + Dk sk + ck ,
k = 0, ..., N
(8)
0<k ≤N −1
(9)
The co-state dynamics are generated as:
λk =
∂H k
∂xk
∂Lk
=
+
∂xk
∂f k
∂xk
T
λk+1 ,
The co-state dynamics are linearized (sensitivity) equations
along the state trajectory.
The gradient change in the parameters (within the state
equation) at each time step k are:
∆θk = −η
∂H k
∂θk
= −η
∂Lk
+
∂θk
∂f k
∂θk
T
k+1
λ
!
(10)
where η is a general (sufficiently small) learning rate. Similarly, the gradient change in the parameters (within the output
equation) at the time step k are:
k
∂H k
∂L
∆νk = −η
(11)
= −η
∂νk
∂νk
Note that eqns (10)-(11) are written as deterministic
expressions for clarity; however, they should be viewed with
an expectation operator applied to the righ-hand side. These
changes are expressed at each time step k, thus there is a
“sequence” of gradient changes over the whole time horizon
duration 0 to N . The goal is to use these time-step changes
to compute the changes to the parameters over the whole
time horizon of the sequence trajectory, known as an epoch.
We note that the parameters of the network are “shared” and
must remain constant during the trajectory/epoch and can
only be incremented from an epoch to the next epoch till
they converge to a constant set of parameters.
Finally, the split boundary conditions are:
∂φ
− λN
∂xN
T
T
(λo ) dxo = 0
dxN = 0
(12)
(13)
derivatives. This provides the final co-state at the final time N .
It is usually assumed that the initial state stage xo is given or
specified as constant, and thus dxo = 0. Therefore eqn (13)
is trivially satisfied. Also, for the boundary condition eqn
(12) we take the quantity in parenthesis to equal zero. The
boundary conditions thus become:
Remark IV.1: Note that for computational expediency, the
′
diagonal matrix σN can be expressed as a vector of derivatives
multiplied point-wise (i.e., a Hadamard multiplication) to
the vector yN − dN . This is more commonly adopted in
computational/coding implementations.
(i) Initial boundary condition:
xo
(14)
Lk (yk , xk , hk , θk , νk ) = Lk (yk )
(ii) Final boundary condition:
∂φ
=
∂xN
The general loss function may be simplified to be a sum of
separate loss functions in each vector variable, e.g., as
+ βLk (xk ) + β0 Lk (hk )
(15)
+ γ1 Lk (θk ) + γ2 Lk (νk ) (18)
Thus, given a set of parameters values, the iteration processing
is to use the initial state xo , the input signal sequence sk and
the corresponding desired sequence dk to generate the state
and output sequences (forward in time) up to the final timeN .
Then, compute the final boundary of the co-state using eqn
(15) and subsequently generate the co-state sequence λk of
eqn (9) backward in-time. For each iteration, or a batch of
iterations, all parameters need to be updated, and continue
doing so according to some stopping criterion. In principle,
this is the core parameters iteration procedure in training
neural networks. After each iteration is completed, and/or
after the stopping criterion is met, one may use the achieved
fixed parameters in the recurrent network, eqns (7)-(8), for
evaluation, testing, or processing.
where the first term is a loss function on the output at
index k and is chosen to be a supervised loss function with
corresponding reference or target dk . One may use the scaled
L2 norm of the error to be consistent with the final-time loss
function in eqn (16), specifically,
1
Lk (yk ) = ||yk − dk ||22
2
1
= (yk − dk )T (yk − dk )
(19)
2
All other terms have the tuning (or hyper-) parameters
β, βo , γ1 and γ2 as scaling penalty factors between 0 and 1 to
emphasize the importance one places on these individual costs.
The second two loss terms (usually, one needs to use only one
or the other), are on the internal state or its hidden function.
One may use either one with an unsupervised loss function,
e.g., to optimize the entropy, cross entropy, or sparsity of this
internal (state) representation. Here, we set β0 to zero and
choose the loss function to be the entropy defined as
λN
IV. BASIC R ECURRENT N EURAL N ETWORKS :
B RNN– SPECIAL CASE
We now provide specific details of the bRNN model and
its parameter updates. We basically specify the components
of the loss/cost function and the bounday conditions, namely
the initial condition for the state forward dynmaics and final
condition for the co-state backward dynamics.
The final-time component of the loss function φ(yN )
can be specified for supervised learning as, e.g., an L2 norm
of an error quantity at the final time N . Specifically,
1
||yN − dN ||22
2
1
= (yN − dN )T (yN − dN )
2
φ(yN ) =
(16)
And thus its derivative is calculated from vector-calculus to
produce the final condition of the co-state as:
λN =
′
∂φ
= (σN )T VNT (yN − dN )
∂xN
(17)
where one uses the output eqn (8) for the definition of
′
yN . The matrix σN is the derivative of the nonlinearity
vector σN expressed as a diagonal matrix of element-wise
Lk (xk ) = − ln |pxk (xk )|
≈ ||xk ||1 ,
(20)
where one may use the L1 norm to approximate the entropy
(i.e., one imposes the sup-Gaussian Laplacian density function
on the state vector xk ). Another candidate is to use the general
approximating function for sup-Gaussian densities, namely,
tanh(α ∗ xk ), 1 < α ≤ 3, for further details see, e.g. [8]
and [9].
Finally, the loss function terms on the parameters θk and νk
may be used for regularization. In that case, one provides
a scaled quadratic expression for every scalar parameter. An
example is to use
1
||θk ||22 0 ≤ k ≤ N − 1
2
and for the parameters in the output layer,
1
Lk (νk ) = ||νk ||22 0 ≤ k ≤ N
2
These specific loss function terms represent common
choices in RNNs.
Lk (θk ) =
V.
BASIC
R ECURRENT N EURAL N ETWORKS ( B RNN):
SUMMARY EQUATIONS
We now summarize the set of equations to be used in
coding computations as follows:
The basic RNN (bRNN) model is
xk+1 = Axk + Ui hk + Wi sk + bi , xo ,
hk = σk (xk ),
yk = Vi hk + Di sk + ci ,
k = 0, ..., N − 1
k = 0, ..., N
k = 0, ..., N
(21)
where the adaptive parameters are indexed with the iteration
(i.e., epoch) index i to denote update of the parameters during
training. These parameters are expected to converge to a
constant at the end of the training process. The time index k
denotes the sample trajectory over the finite horizon training
sequence, i.e., input-label pairs (sk , dk ), 0 ≤ k ≤ N . We
thus reserve the time index k along each step of the sequence
trajectory over the finite-horizon, o ≤ k ≤ N , while the
parameter update iteration index i per full sequence trajectory
as illustrated below:
− |−−−−−− →k
|
i ↓
Now the back propagating co-state dynamics are (see Appendix II)
′
From Appendix II, the state equations parameter gradient
changes at each time k, 0 ≤ k ≤ N −1 (without regularization)
are:
∆Uk = − ηλk+1 (hk )T
(25)
∆Wk = − ηλk+1 (sk )T
(26)
∆bk = − ηλk+1
(27)
Similarly, from Appendix II, the parameter (weight and bias)
updates in the output equation at each time k, 0 ≤ k ≤ N
(without regularization) are:
k
∂L
= − ηek (hk )T
(28)
∆Vk = − η
∂Vk
∆Dk = − η
For each iteration i, this model propagates forward for a
given (or chosen) initial condition xo , and a sequence pair
(sk , dk ) to generate the state xk , the corresponding hidden
nonlinearity hk and the output yk over the sequence duration
k = 0, ..., N . Thus the output error sequence is computed.
The error backpropagation co-state equations need the final
co-state and the output error sequence to generate the costate dynamics backward in time k. Here, the final co-state
is expressed as (see, 17):
′
λN = (σN )T ViT (yN − dN )
′
where σN is the derivative of the nonlinear vector σN which
is represented as a diagonal matrix of element-wise derivatives of the corresponding scalar nonlinearity. It may also
be represented as a vector of derivatives which point-wise
multiplies to the error vector as a Hadamard vector-vector
multiplication. Usually, in codes, the latter representation is
followed. However, for analysis, it is more convenient to view
it as a diagonal matrix (transpose) as it follows from the vector
calculus. Let us compactly express the output error signal over
the duration of k as:
ek = (yk − dk ), 0 ≤ k ≤ N
(22)
Then the final time boundary condition of the co-state is
expressed as:
′
λN = (σN )T ViT eN
(23)
′
λk =(A + Ui σk )T λk+1 + (σk )T ViT ek +
k
k
′
∂L
∂L
T
+ β2 (σk )
,
β1
∂xk
∂hk
0 < k ≤ N − 1 (24)
∆ck = − η
∂Lk
∂Dk
∂Lk
∂ck
= − ηek (sk )T
= − ηek
(29)
(30)
Usually, these parameter change contributions at each k are
summed over k over their time horizon to provide one change
per trajectory duration or epoch. Then, one may updates all
the parameters at each single iteration (i.e. epoch) or over a
patch or several iterations.
Thus, per sequene trajectory, one can accumulate the parameter change contributions to obtain their updates as:
∆Ui = − η
N
−1
X
λk+1 (hk )T
(31)
k=o
∆Wi = − η
N
−1
X
λk+1 (sk )T
(32)
k=o
∆bi = − η
N
−1
X
λk+1
(33)
k=o
Similarly, the parameter (weight and bias) updates in the
output equation at one iteration (epoch) are:
∆Vi = − η
N
X
∂Lk
k=o
∂Vk
= −η
N
X
k=o
ek (hk )T
(34)
∆Di = − η
N
X
k=o
∆ci = − η
k
∂L
∂Dk
N
X
∂Lk
k=o
∂ck
= −η
N
X
ek (sk )T
(35)
k=o
= −η
N
X
ek
(36)
k=o
Remark V.1: The above parameter update expressions
are commonly used forms, which are basically a sum
of the contributions over the time index k over the full
sequence trajectory. We note that this is only one choice.
Observe also that it is equivalent to using the mean of the
gradient changes over the sequences, eqns (25)-(30), where
η absorbs any scaling due to the number of elements of
the sequence. The mean, however, may be small or even
zero while some gradient changes at some time indices k
may be relatively large. In fact, this is one explanation of
the “vanishing gradient” phenomena. That is, even though
the co-state and output error sequences are not vanishing,
the gradient changes using the ”mean” as in eqns (31)-(36) are.
Remark V.2: Another choice to explore is to use the
median of the changes, or the minimum of the changes, over
the gradient change sequences, eqns (25)-(30). Moreover,
the change in power or strength in the sequence of changes
is best captured by the variance which should influence the
learning rate in a multitude of ways. Thus there are alternate
possibilities of the iteration update laws for RNN besides the
ones in eqns (31)-(36).
Remark V.3: The “vanishing and exploding” gradients
can be explained easily by investigating the update eqns
(31)-(36), and the backward co-state dynamics eqns (23)-(24).
By their nature, the state and co-state equations form an
unstable saddle point dynamics, i.e. if the forward network
dynamics are (locally) stable, the backward sensitivity
dynamics are also stable but only backward in time. This is
another justification for the error back-propagation in addition
to the fact that we use the final co-state value. In contrast,
if the forward network dynamics are (locally) unstable, the
backward sensitivity dynamics are also unstable backward
in time. In that event, the state will grow as it propagates
forward, thus, the final costate will be large, and in turn the
co-state will be further growing. As the co-state and the error
signals are used in the expression of the gradient updates.
the gradient will continue to grow unbounded. This process
will make the state and co-state grow unbounded. As the
time horizon increases (i.e., N increases), this will amplify
the growth of the state and co-state leading to the so-called
“exploding gradient” phonemena.
In the update laws and the iterations, one must use sufficiently small η to avoid the occurence of numerical instability
while allow the iteration process of the parameter update to be
sufficiently small. The presence of the dominant stable linear
term in bRNN ensures that the network for each bounded
parameters is BIBO stable and thus the states, costates, and
consequently the gradient would not grow unbounded. The
bRRN can steer the error and the co-state sequences towards
zero during training. This is possible of course for all co-states
except for the co-state λ0 which can be non-zero and does not
play any role in the (state equations) parameter updates, eqn
(31)-(33).
VI. C ONCLUSION
This work introduces a design for a basic recurrent neural
network (bRNN) which has a basic recurrent structure and
output layer with sufficient parameters to enable a flexible
filtering suitable for RNN applications of sequence to sequence
mappings. The framework adopts the classical constrained
optimization and calculus of variations to derive the backpropagation and (stochastic) gradient parameter update. It enables
an ease in incorporating general loss function components
on any variable in the networks, namely outputs, states or
hidden units (nonlinear functions of the states). It shows the
correspondence of the backpropagations through time (BTT)
approach with applying the classical Lagrange multiplier
method in constrained optimization. It further shows that
the usual sum of contributions of changes to the parameters
inherent in the BTT approach is only one form of update, and it
could be a source of the “vanishing gradient” phenomena. The
“exploding gradient” phenomena is explained to emanate from
either(i) numerical instability due to summing contributions
over a long time horizon, or (ii) due to the instability of
the networks and the costate dynamics manivesting in the
gradient update laws. Finaly, we state that we have conducted
simulations using this bRRN model to verify and demostrate
its perfromance and will be reported on the details of these
results in another publication. [10].
ACKNOWLEDGMENT
The work is supported in part by the National Science
foundation grant No. ECCS-1549517. The author thankfully
acknowledges this support.
VII. APPENDIX I: BIBO
STABILITY OF B RNN
We show that the presence of the linear term of the
constant stable matrix in the basic recurrent neural network
(bRNN) model as in eqn(4) is crucial for the boundedness of
all trajectories, and that all trajectories would remain within,
or converge to, a bounded region. This guarantees that, for
bounded input signals and bounded parameters, trajectories
would not go to infinity but rather remain confined to, or
converge to, a bounded region.
The bRNN stability and dynamic behavior are governed
by the dynamic eqn(4). To pursue the Bounded-InputBounded-Output (BIBO) stability analysis, we define the
compact assumedly bounded vector:
Mk := U hk + W sk + b
(37)
We choose the constant matrix A to be a stable matrix with
eigenvalues having moduli less than unity.
Then, eqn(4) is re-written as
xk+1 = Axk + Mk ,
k = 0, ..., N − 1
(38)
For stability, we choose the following general quadratic
form as a candidate Liapunov function:
Vk = (xk + x∗1 )T S (xk + x∗1 )
(39)
where S is a symmetric positive definite matrix, and x∗1 is
a center point which may be different from the origin. In
Liapunov theory, typically, x∗1 is chosen to be an equilibrium
point of interest. In the present case, however, x∗1 is simply
a point in (the state) space as a center to be determined. We
note that this Liapunov candidate is positive definite for all
points with reference to the center point.
Now, we calculate the “difference equation” of this Liapunov
function, i.e.,
∆Vk = Vk+1 − Vk
= (xk+1 + x∗1 )T S(xk+1 + x∗1 ) −
(xk + x∗1 )T S(xk + x∗1 )
(40)
We use the dynamic equations, eqn(38) into the equality (40),
and expand terms to obtain a quadratic equation in xk . We
then complete the square and match the expanded equation
terms to a quadratic expression term-by-term. We thus obtain
the following general quadratic form:
Then to match eqn (41) to eqn (40), term by term, the
following equalities must be satisfied:
GT x∗2 = AT Mk
(x∗2 )T x∗2
(46)
D =
MkT Mk
+
(x∗2 )T x∗2
(47)
(48)
Thus all terms are well-defined in the difference eqn (41).
This completes the proof. Thus, we indeed found a Liapunov
function candidate centered around x∗1 = 0 and is described
by a spherical (or more generally, an ellipsoidal) region. Such
region can be made to conservatively include the point x∗2
where outside this bounded region the difference equation
along trajectories is negative definite.
We remark that our purpose here is to simply determine a
bounded region where BIBO stability is ensured. In general,
and for a tighter bound constraints, the goal is to solve the
general constraints in order to determine a point x∗1 as near
as possible to a point x∗2 where both satisfy the constraints.
Finally, we remark that, in the training process of RNNs,
the parameters are frozen over the finite-horizon forward state
dynamics and backward co-state dynamics. This constitutes
one iteration (or epoch). The gradient descent update of
the parameters occur per one or many iterations over the
forzen training sequences. This constitutes a decoupling of the
dynamics of the bRNN and the parameter gradient updates.
Thus the BIBO stability of the overall system is maintained.
VIII. APPENDIX II: D ERIVATIONS
(41)
where the matrix G satisfies the equality
(42)
Thus the difference Liapunov function will be negative
definite outside the ellipsoid defined by eqn (41) with the
left-hand side set to zero. This summarizes the general
approach.
To be specific, and to simplify the calculations, we choose
S = I (the identity matrix), the vector x∗1 = 0, and proceed
as follows. Eqn (42) becomes
(43)
Note that GT G is positive definite and thus invertible. Λ is a
diagonal matrix of the square roots of the singular values of
A with moduli less than 1, and Q is the matrix formed of the
corresponding orthogonal right-singular vectors. Note that the
singular values of A result in
AT A = QΛT ΛQT
+D
x∗2 = (GGT )−1 GAT Mk
T
GT G = I − AT A = QT [I − ΛT Λ]Q
(45)
The above results in the following solution equalities
∆Vk = − [Gxk − x∗2 ] [Gxk − x∗2 ] + D
GT G = S − AT SA
= −
MkT Mk
(44)
We now provide specific details of the model and its
parameter updates. We assume the state iniital condition xo
is chosen or given as a constant, and thus dxo = 0. The split
boundary conditions then are:
xo ,
λN =
∂φ
∂xN
(49)
Thus, the processing is to use the initial state xo , the input
signal sequence sk and the coressponding desired sequence
dk to generate the internal state and output sequences. This
consequently generates the output error sequnece expressed
compacted as:
ek = (yk − dk ), 0 ≤ k ≤ N
(50)
The final-time loss function φ(yN ) is defined as an L2 norm
in the output error as:
1
||yN − dN ||22
2
1
= eTN eN
2
φ(yN ) =
(51)
And thus its derivative in eqn (49) provides the final co-state
as
′
∂φ
λN =
= (σN )T VNT eN
(52)
∂xN
The general loss function is simplified to be a sum of
separate loss functions as
Lk (yk , xk , hk , θk , νk ) = Lk (yk )
k
k
+ βL (xk ) + β0 L (hk )
+ γ1 Lk (θk ) + γ2 Lk (νk ) (53)
We now calculate the derivatives needed in the co-state
equation (9) and in the gradient descent parameter update
equations (10) and (11), specilized to each parameter.
∂Lk (hk )
∂Lk (yk )
∂Lk (xk )
∂Lk
=
+β
+ β0
∂xk
∂xk
∂xk
∂xk
k
k
′
′
∂L
∂L
= (σk )T VkT ek + β
+ β0 (σk )T
,
∂xk
∂hk
0<k ≤N −1
(54)
And the derivative (The Jacobian):
k
′
∂f
= (A + Uk (σk ) ), 0 < k ≤ N − 1
∂xk
(55)
Also, one derives the equalities:
∂Lk
= γ1 θk
∂θk
(56)
The parameter (weight and bias) updates in the output equation
are now calculated. The regularization terms for all parameters
contribute
∂Lk (ν)
= γ2 νk
∂νk
(63)
Finally, the gradient changes at each time step k are
k
∂H k
∂L
(64)
∆νk = −η
= −η
∂νk
∂νk
where
∆Vk = −η
∆Dk = −η
∆ck = −η
∂Lk
∂Vk
∂Lk
∂Dk
∂Lk
∂ck
= −η γ2 Vk + ek (hk )T
= −η γ2 Dk + ek (sk )T
= −η (γ2 ck + ek )
(65)
(66)
(67)
Thus, the (training) processing is to use the initial state
xo , and the input signal sequence sk and its corresponding
desired sequence dk to generate the gradient descent parameter
changes. Then one updates the parameter online, over the
full sequence (epoch), or over several epoches, to steer the
parameters along a stochastic gradient descent towards a good
local optimal.
R EFERENCES
T
λk+1 = λk+1 (hk )T
(57)
T
λk+1 = λk+1 (sk )T
(58)
∂f k
∂Uk
∂f k
∂Wk
∂f k
∂bk
T
λk+1 = λk+1
(59)
Thus, this gives:
∆Uk = − η γ1 Uk + λk+1 (hk )T
(60)
∆Wk = − η γ1 Wk + λk+1 (sk )T
(61)
∆bk = − η (γ1 bk + λk+1 )
(62)
[1] Y. LeCun, Y. Bengio, and G. Hinton, Deep learning, Nature, vol. 521,
pp. 436-444, May 2015.
[2] Chung, Junyoung and Gulcehre, Caglar and Cho, KyungHyun and Bengio, Yoshua, Empirical Evaluation of Gated Recurrent Neural Networks
on Sequence Modeling, https://arxiv.org/abs/1412.3555, 2014.
[3] R. Jozefowicz, W. Zaremba, and I. Sutskever, An empirical exploration
of recurrent network architectures, 2015.
[4] G-B. Zhou, J. Wu, Jianxin, C-L. Zhang, and Z-H. Zhou, Minimal Gated Unit for Recurrent Neural Networks,
ArXiv:
https://arxiv.org/abs/1603.09420, 2016.
[5] Q. V. Le, N. Jaitly, and G. E. Hinton, A simple way to initialize recurrent
networks of rectified linear units,
arXiv preprint arXiv:1504.00941,
[cs.NE] 7 April 2015.
[6] A. E. Bryson, Jr and Y-C. Ho, Applied Optimal Control: Optimization,
Estimation and Control, CRC Press, 1975
[7] F. Lewis, D. Vrabie, and V. Syrmos, Optimal Control, 3rd ed Wiley
2012
[8] K. Waheed, F. M. Salem, Blind source recovery: A framework in the state
space, Journal of Machine Learning Research 4 (Dec), pp. 1411-1446.
[9] Z. Albataineh and F. M. Salem, Adaptive Blind CDMA Receivers Based
on ICA Filtered Structures,
Circuits Syst Signal Process (2016).
doi:10.1007/s00034-016-0459-4.
[10] Z. Zhou and F. M. Salem, Perfomance Evaluation of the basic RNN
model, Memorandum.10.12.2016. MSU, Dec. 2016
| 9cs.NE
|
arXiv:1606.00539v2 [math.GR] 5 Jan 2017
Geometric embedding properties
of Bestvina-Brady subgroups
HUNG CONG TRAN
We compute the relative divergence of right-angled Artin groups with respect to
their Bestvina-Brady subgroups and the subgroup distortion of Bestvina-Brady
subgroups. We also show that for each integer n ≥ 3, there is a free subgroup of
rank n of some right-angled Artin group whose inclusion is not a quasi-isometric
embedding. The corollary answers the question of Carr about the minimum rank
n such that some right-angled Artin group has a free subgroup of rank n whose
inclusion is not a quasi-isometric embedding. It is well-known that a right-angled
Artin group AΓ is the fundamental group of a graph manifold whenever the defining
graph Γ is a tree with at least 3 vertices. We show that the Bestvina-Brady subgroup
HΓ in this case is a horizontal surface subgroup.
20F65; 20F36, 20F67
1 Introduction
For each Γ a finite simplicial graph the associated right-angled Artin group AΓ has
generating set S the vertices of Γ, and relations st = ts whenever s and t are adjacent
vertices. If Γ is non-empty, there is a homomorphism from AΓ onto the integers, that
takes every generator to 1. The Bestvina-Brady subgroup HΓ is defined to be the
kernel of this homomorphism.
Bestvina-Brady subgroups were introduced by Bestvina-Brady in [BB97] to study the
finiteness properties of subgroups of right-angled Artin groups. One result in [BB97] is
that the Bestvina-Brady subgroup HΓ is finitely generated iff the graph Γ is connected.
This fact is a motivation to study the geometric connection between a right-angled
Artin group and its Bestvina-Brady subgroup. More precisely, we examine the relative
divergence of right-angled Artin groups with respect to their Bestvina-Brady subgroups
and the subgroup distortion of Bestvina-Brady subgroups (see the following theorem).
Theorem 1.1 Let Γ be a connected, finite, simplicial graph with at least 2 vertices. Let
AΓ be the associated right-angled Artin group and HΓ the Bestvina-Brady subgroup.
2
Hung Cong Tran
Then the relative divergence Div(AΓ , HΓ ) and the subgroup distortion DistAHΓΓ are both
linear if Γ is a join graph. Otherwise, the relative divergence Div(AΓ , HΓ ) and the
subgroup distortion DistAHΓΓ are both quadratic.
In the above theorem, we can see that the relative divergence Div(AΓ , HΓ ) and the
subgroup distortion DistAHΓΓ are equivalent. In general, we showed that the relative
divergence is always dominated by the subgroup distortion for any pair of finitely
generated groups (G, H), where H is a normal subgroup of G such that the quotient
group G/H is an infinite cyclic group (see Proposition 4.3).
Carr [Car] proved that non-abelian two-generator subgroups of right-angled Artin
groups are quasi-isometrically embedded free groups. In his paper, he also showed
an example of a distorted free subgroup of a right-angled Artin group. However, the
minimum rank n such that some right-angled Artin group has a free subgroup of rank
n whose inclusion is not a quasi-isometric embedding was still unknown (see [Car]).
A corollary of Theorem 1.1 answered this question (see the following corollary).
Corollary 1.2 For each integer n ≥ 3 , there is a right-angled Artin group containing
a free subgroup of rank n whose inclusion is not a quasi-isometric embedding.
We remark that a special case of Theorem 1.1 can also be derived as a consequence
of previous work by Hruska–Nguyen ([HN]) on distortion of surfaces in graph manifolds. Hruska–Nguyen showed that every virtually embedded horizontal surface in
a 3–dimensional graph manifold has quadratic distortion. After learning about this
result, the we proved the following theorem, which implies that many Bestvina–Brady
subgroups are also horizontal surface subgroups.
Theorem 1.3 If Γ is a finite tree with at least 3 vertices, then the associated rightangled Artin group AΓ is a fundamental group of a graph manifold and the BestvinaBrady subgroup HΓ is a horizontal surface subgroup.
It is well-known that a right-angled Artin group AΓ is the fundamental group of a graph
manifold whenever the defining graph Γ is a tree with at least 3 vertices. However, the
fact that the Bestvina-Brady subgroup HΓ is a horizontal subgroup does not seem to be
recorded in the literature. With the use of Theorem 1.3, we see that Theorem 1.1 can
be viewed as a generalization of a special case of the quadratic distortion theorem of
Hruska–Nguyen. Moreover, Theorem 1.3 combined with the Hruska–Nguyen theorem
gives an alternative proof of Corollary 1.2.
Geometric embedding properties of Bestvina-Brady subgroups
3
Acknowledgments
I would like to thank Prof. Christopher Hruska and Hoang Nguyen for very helpful
comments and suggestions, especially their help on the proof of Theorem 1.3. Hoang
Nguyen helped me a lot with the background of graph manifolds and horizontal surfaces. He also showed me his joint work with Hruska on subgroup distortion of
horizontal surface subgroups and their work is a motivation for this article. I also thank
the referee for advice that improved the exposition of the paper.
2 Right-angled Artin groups and Bestvina-Brady subgroups
Definition 2.1 Given a finite simplicial graph Γ, the associated right-angled Artin
group AΓ has generating set S the vertices of Γ, and relations st = ts whenever s and
t are adjacent vertices.
Let S1 be a subset of S. The subgroup of AΓ generated by S1 is a right-angled Artin
group AΓ1 , where Γ1 is the induced subgraph of Γ with vertex set S1 (i.e. Γ1 is the
union of all edges of Γ with both endpoints in S1 ). The subgroup AΓ1 is called a
special subgroup of AΓ .
Definition 2.2 Let Γ be a finite simplicial graph with the set S of vertices. Let T be
a torus of dimension |S| with edges labeled by the elements of S. Let XΓ denote the
subcomplex of T consisting of all faces whose edge labels span a complete subgraph
in Γ (or equivalently, mutually commute in AΓ ). XΓ is called the Salvetti complex.
Remark 2.3 The fundamental group of XΓ is AΓ . The universal cover X̃Γ of XΓ is a
CAT(0) cube complex with a free, cocompact action of AΓ . Obviously, the 1–skeleton
of X̃Γ is the Cayley graph of AΓ with respect the generating set S.
Definition 2.4 Let Γ be a finite simplicial graph. Let Φ : AΓ → Z be an epimorphism
which sends all the generators of AΓ to 1 in Z. The kernel HΓ of Φ is called the
Bestvina-Brady subgroup.
Remark 2.5 There is a natural continuous map f : XΓ → S1 which induces the
homomorphism Φ : AΓ → Z. Moreover, it is not hard to see that the lifting map
f̃ : X̃Γ → R is an extension of Φ.
4
Hung Cong Tran
Theorem 2.6 (Bestvina-Brady [BB97] and Dicks-Leary [DL99]) Let Γ be a finite
simplicial graph. The Bestvina-Brady subgroup HΓ is finitely generated iff Γ is
connected. Moreover, the set T of all elements of the form st−1 whenever s and t are
adjacent vertices form a finite generating set for HΓ . Moreover, if Γ is a tree with n
edges, then the Bestvina-Brady subgroup HΓ is a free group of rank n .
Definition 2.7 Let Γ1 and Γ2 be two graphs, the join of Γ1 and Γ2 is a graph obtained
by connecting every vertex of Γ1 to every vertex of Γ2 by an edge.
Let J be a complete subgraph of Γ which decomposes as a nontrivial join. We call AJ
a join subgroup of AΓ .
Let Γ be a finite simplicial graph with the vertex set S and let g an element of AΓ . A
reduced word for g is a minimal length word in the free group F(S) representing g.
Given an arbitrary word representing g, one can obtain a reduced word by a process
of “shuffling” (i.e. interchanging commuting elements) and canceling inverse pairs.
Any two reduced words for g differ only by shuffling. For an element g ∈ AΓ , a
cyclic reduction of g is a minimal length element of the conjugacy class of g. If w
is a reduced word representing g, then we can find a cyclic reduction g by shuffling
commuting generators in w to get a maximal length word u such that w = uwu−1 . In
particular, g itself is cyclically reduced if and only if every shuffle of w is cyclically
reduced as a word in the free group F(S).
3 Relative divergence, geodesic divergence, and subgroup
distortion
Before we define the concepts of relative divergence, geodesic divergence, and subgroup distortion, we need to build the notions of domination and equivalence. These
notions are the tools to measure the relative divergence, geodesic divergence, and
subgroup distortion.
Definition 3.1 Let M be the collection of all functions from [0, ∞) to [0, ∞]. Let
f and g be arbitrary elements of M. The function f is dominated by the function
g, denoted f g, if there are positive constants A, B, C and D such that f (x) ≤
Ag(Bx) + Cx for all x > D. Two function f and g are equivalent, denoted f ∼ g, if
f g and g f .
Geometric embedding properties of Bestvina-Brady subgroups
5
Remark 3.2 A function f in M is linear, quadratic or exponential... if f is respectively equivalent to any polynomial with degree one, two or any function of the form
abx+c , where a > 1, b > 0.
Definition 3.3 Let {δρn } and {δρ′n } be two families of functions of M, indexed over
ρ ∈ (0, 1] and positive integers n ≥ 2. The family {δρn } is dominated by the family
{δρ′n }, denoted {δρn } {δρ′n }, if there exists constant L ∈ (0, 1] and a positive integer
n δ ′Mn . Two families {δ n } and {δ ′n } are equivalent, denoted
M such that δLρ
ρ
ρ
ρ
′n
n
{δρ } ∼ {δρ }, if {δρn } {δρ′n } and {δρ′n } {δρn }.
Remark 3.4 A family {δρn } is dominated by (or dominates) a function f in M if {δρn }
is dominated by (or dominates) the family {δρ′n } where δρ′n = f for all ρ and n. The
equivalence between a family {δρn } and a function f in M can be defined similarly.
Thus, a family {δρn } is linear, quadratic, exponential, etc if {δρn } is equivalent to the
function f where f is linear, quadratic, exponential, etc.
Definition 3.5 Let X be a geodesic space and A a subspace of X . Let r be any positive
number.
(1) Nr (A) = { x ∈ X | dX (x, A) < r }
(2) ∂Nr (A) = { x ∈ X | dX (x, A) = r }
(3) Cr (A) = X − Nr (A).
(4) Let dr,A be the induced length metric on the complement of the r–neighborhood
of A in X . If the subspace A is clear from context, we can use the notation dr
instead of using dr,A .
Definition 3.6 Let (X, A) be a pair of metric spaces. For each ρ ∈ (0, 1] and positive
integer n ≥ 2, we define a function δρn : [0, ∞) → [0, ∞] as follows:
For each r, let δρn (r) = sup dρr (x1 , x2 ) where the supremum is taken over all x1 , x2 ∈
∂Nr (A) such that dr (x1 , x2 ) < ∞ and d(x1 , x2 ) ≤ nr.
The family of functions {δρn } is the relative divergence of X with respect A, denoted
Div(X, A).
We now define the concept of relative divergence of a finitely generated group with
respect to a subgroup.
6
Hung Cong Tran
Definition 3.7 Let G be a finitely generated group and H its subgroup. We define
the relative divergence of G with respect to H , denoted Div(G, H), to be the relative
divergence of the Cayley graph Γ(G, S) with respect to H for some finite generating
set S.
Remark 3.8 The concept of relative divergence was introduced by the author in
[Tra15] with the name upper relative divergence. The relative divergence of geodesic
spaces is a pair quasi-isometry invariant concept. This implies that the relative divergence on a finitely generated group does not depend on the choice of finite generating
sets.
Definition 3.9 The divergence of a bi-infinite geodesic α, denoted Divα , is a function
g : (0, ∞) → (0, ∞) which for each positive number r the value g(r) is the infimum
on the lengths of all paths outside the open ball with radius r about α(0) connecting
α(−r) and α(r).
The following lemma is deduced from the proof of Corollary 4.8 in [BC12].
Lemma 3.10 Let Γ be a connected, finite, simplicial graph with at least 2 vertices.
Assume that Γ is not a join. Let g be a cyclically reduced element in AΓ that does not
lie in any join subgroup. Then the divergence of bi-infinite geodesic · · · ggggg · · · is
at least quadratic.
Definition 3.11 Let G be a group with a finite generating set S and H a subgroup of
G with a finite generating set T . The subgroup distortion of H in G is the function
H : (0, ∞) → (0, ∞) defined as follows:
DistG
H
DistG
(r) = max{ |h|T | h ∈ H, |h|S ≤ r }.
Remark 3.12 It is well-known that the concept of distortion does not depend on the
choice of finite generating sets.
4 Connection between subgroup distortion and relative divergence
Lemma 4.1 Let H be a finitely generated group with finite generating set T and φ in
Aut(H) . Let G = hH, t/tht−1 = φ(h)i and S = T ∪ {t} . Then:
Geometric embedding properties of Bestvina-Brady subgroups
7
(1) All element in G can be written uniquely in the from htn where h is a group
element in H .
(2) The set S is a finite generating set of G and dS (htm , h′ tn ) ≥ |m − n| , and
dS (htm , Htn ) = |m − n| .
Proof The statement (1) is a well-known and we only need to prove statement (2).
Let ψ be the map from G to Z by sending element t to 1 and each generator in T to 0.
It is not hard to see that ψ is a group homomorphism. We first show that the absolute
value of ψ(g) is at most the length of g with respect to S for each group element g in
G. In fact, let w1 tn1 w2 tn2 · · · wk tnk be the shortest word in S that represents g, where
each wi is a word in T . Therefore,
ψ(g) = n1 + n2 + · · · + nk
and
|g|S = ℓ(w1 ) + ℓ(w2 ) + · · · + ℓ(wk ) + |n1 | + |n2 | + · · · + |nk | .
This implies that the absolute value of ψ(g) is at most the length of g with respect to
S. The distance between two elements htm and h′ tn is the length of the group element
g = (htm )−1 h′ tn . Obviously, ψ(g) = n − m. Therefore, the distance between two
elements htm and h′ tn is at least |m − n|. This fact directly implies that the distance
between htm and any element in Htn is at least |m − n|. Also, htn is an element in Htn
and the distance between htm , htn is at most |m − n|. Therefore, the distance between
htm and Htn is at exactly |m − n|.
Lemma 4.2 Let H be a finitely generated group with finite generating set T and φ
in Aut(H) . Let G = hH, t/tht−1 = φ(h)i and S = T ∪ {t} . Let n be an arbitrary
positive integer and x , y be two points in ∂Nn (H) . Then there is path outside Nn (H)
connecting x and y iff the pair (x, y) is either of the form (h1 tn , h2 tn ) or (h1 t−n , h2 t−n )
where h1 and h2 are elements in H .
Proof By Lemma 4.1, the pair (x, y) must be of the form (h1 tm1 , h2 tm2 ) where |m1 | =
|m2 | = n. We first assume that m1 m2 < 0. Let γ be an arbitrary path connecting x and
′
y. By Lemma 4.1, we observe that if two vertices htm and h′ tm of γ are consecutive,
then |m − m′ | ≤ 1. Therefore, there exists a vertex of γ that belongs to H . Thus,
there is no path outside Nn (H) connecting x and y.
If m1 = m2 , then x and y both lie in the same coset tm1 H . Therefore, there is a path α
with all vertices in tm1 H connecting x and y. By Lemma 4.1 again, α must lie outside
Nn (H). Therefore, the pair (x, y) is either of the form (h1 tn , h2 tn ) or (h1 t−n , h2 t−n ).
8
Hung Cong Tran
Proposition 4.3 Let H be a finitely generated group and G = hH, t/tht−1 = φ(h)i
where φ in Aut(H) . Then, Div(G, H) DistGH .
Proof Let T be a finite generating set of H and let S = T ∪ {t}. Then, S is a
finite generating set of G. Suppose that Div(G, H) = {δρn }. We will show that
H (nr) for all positive integer r.
δρn (r) ≤ DistG
Indeed, let x, y be arbitrary points in ∂Nr (H) such that dr,H (x, y) < ∞ and dS (x, y) ≤
nr. By Lemma 4.2, x, y both lie in the same coset tm H where |m| = r. Therefore,
there is a path γ with all vertices in tm H connecting x and y and the length of γ
H (nr). By Lemma 4.1 again, the path γ must lie outside N (H).
is at most DistG
r
H (nr). Thus, δ n (r) ≤ DistH (nr). This implies that
Therefore, dρr,H (x, y) ≤ DistG
ρ
G
H.
Div(G, H) DistG
5 Relative divergence of right-angled Artin groups with respect to Bestvina-Brady subgroups and subgroup distortion of Bestvina-Brady subgroups
From now, we let Γ be a finite, connected, simplicial graph with at least 2 vertices. Let
AΓ be the associated right-angled Artin group and HΓ be its Bestvina-Brady subgroup.
Let XΓ be the associated Salvetti complex and X̃Γ its universal covering. We consider
the 1–skeleton of X̃Γ as a Cayley graph of AΓ and the vertex set S of Γ as a finite
generating set of AΓ . By Theorem 2.6, we can choose the set T of all elements of the
form st−1 whenever s and t are adjacent vertices as a finite generating set for HΓ . Let
Φ and f̃ be group homomorphism and continuous map as in Remark 2.5.
Lemma 5.1 Let M be the diameter of Γ . Let a and b be arbitrary vertices in S . For
each integer m , the length of am b−m with respect to T is at most M|m| .
Proof Since the diameter of Γ is M , we can choose positive integer n ≤ M and n + 1
generators s0 , s1 , · · · , sn in S such that the following conditions hold:
(1) s0 = a and sn = b.
(2) si and si+1 commutes where i ∈ {0, 1, 2, · · · , n − 1}.
Obviously,
−m
−m m −m m −m
m
−m
m
−m
am b−m = sm
= (sm
0 sn
0 s1 )(s1 s2 )(s2 s3 ) · · · (sn−2 sn−1 )(sn−1 sn )
−1 m
−1 m
−1 m
−1 m
m
= (s0 s−1
1 ) (s1 s2 ) (s2 s3 ) · · · (sn−2 sn−1 ) (sn−1 sn ) .
Geometric embedding properties of Bestvina-Brady subgroups
9
Also, si−1 s−1
belongs to T . Therefore, the length of am b−m with respect to T is
i
at most n|m|. This implies that the length of am b−m with respect to T is at most
M|m|.
Proposition 5.2 The subgroup distortion DistAHΓΓ is dominated by a quadratic function.
Moreover, DistAHΓΓ is linear when Γ is a join.
Proof We first show that DistAHΓΓ is dominated by a quadratic function. Let n be an
arbitrary positive integer and h be an arbitrary element in HΓ such that |h|S ≤ n. We
mk
1 m2 m3
can write h = sm
1 s2 s3 · · · sk such that:
(1) Each si lies in S, |mi | ≥ 1 and |m1 | + |m2 | + |m3 | + · · · + |mk | ≤ n.
(2) m1 + m2 + m3 + · · · + mk = 0
Obviously, we can rewrite h as follows:
(m +m2 +···+mk−1 ) −(m1 +m2 +···+mk−1 )
).
sk
1
1 +m2 ) −(m1 +m2 )
1 −m1
) · · · (sk−1
s3
h = (sm
)(s(m
2
1 s2
Let M be the diameter of Γ. By Lemma 5.1, we have
|h|T ≤ M|m1 | + M|m1 + m2 | + · · · + M|m1 + m2 + · · · + mk−1 |
≤ M|m1 | + M |m1 | + |m2 | + · · · + M |m1 | + |m2 | + · · · + |mk−1 |
≤ M(k − 1)n ≤ Mn2 .
Therefore, the distortion function DistAHΓΓ is bounded above by Mn2 .
We now assume that Γ is a join of Γ1 and Γ2 . We need to prove that the distortion
DistAHΓΓ is linear. Let n be an arbitrary positive integer and h be an arbitrary element in
HΓ such that |h|S ≤ n. Since AΓ is the direct product of AΓ1 and AΓ2 , we can write
nℓ
mk
n1 n2
1 m2
h = (am
1 a2 · · · ak )(b1 b2 · · · bℓ ) such that:
(1) Each ai is a vertex of Γ1 and each bj is a vertex of Γ2 .
(2) |m1 | + |m2 | + · · · + |mk | + |n1 | + |n2 | + · · · + |nℓ | ≤ n.
(3) (m1 + m2 + · · · + mk ) + (n1 + n2 + · · · + nℓ ) = 0
Let m = m1 + m2 + · · · + mk . Then, n1 + n2 + · · · + nℓ = −m and |m| ≤ n. Let a
be a vertex in Γ1 and b a vertex in Γ2 . Since a commutes with each bj , b commutes
with each ai and a, b commute, we can rewrite h as follows:
mk −m
1 m2
h = (am
)(bm a−m )(am bn11 bn22 · · · bnℓ ℓ )
1 a2 · · · ak b
−nℓ
−n2
−n1
.
· · · (ab−1
(ab−1
= (a1 b−1 )m1 (a2 b−1 )m2 · · · (ak b−1 )mk (ba−1 )m (ab−1
2 )
1 )
ℓ )
10
Hung Cong Tran
−1 and ba−1 all belong to T . Therefore,
Also, ab−1
j , ai b
|h|T ≤ |m1 | + |m2 | + · · · + |mk | + |n1 | + |n2 | + · · · + |nℓ | + |m| ≤ 2n.
Therefore, the distortion function DistAHΓΓ is bounded above by 2n.
Proposition 5.3 If Γ is not a join graph, then the relative divergence Div(AΓ , HΓ ) is
at least quadratic.
Proof Let J be a maximal join in Γ and let v be a vertex not in J . Let g in AJ be the
product of all vertices in J . Let n = Φ(g) and let h = gv−n . Then h is an element in
HΓ . Since J is a maximal join in Γ and let v be a vertex not in J , then h does not lie in
any join subgroup. Also, h is a cyclically reduced element. Therefore, the divergence
of the bi-infinite geodesic α = · · · hhhhh · · · is at least quadratic by Lemma 3.10.
Let t be an arbitrary generator in S and k = |h|S . We can assume that α(0) =
e, α(km) = hm , and α(−km) = h−m . In order to prove the relative divergence
Div(AΓ , HΓ ) is at least quadratic, it is sufficient to prove each function δρn dominates
the divergence function of α for each n ≥ 2k + 2.
Indeed, let r be an arbitrary positive integer. Let x = h−r tr and y = hr tr . By the
similar argument as in Lemma 4.1 and Lemma 4.2, two points x and y both lie in
∂Nr (HΓ ) and dr,HΓ (x, y) < ∞. Moreover,
dS (x, y) ≤ dS (x, h−r ) + dS (h−r , hr ) + dS (hr , y) ≤ r + 2kr + r ≤ (2k + 2)r ≤ nr.
Let γ be an arbitrary path outside Nρr (H) connecting
x and y. Obviously, the path
γ must lie outside the open ball B α(0), ρr . It is obvious that
we can connect x
and h−r by a path γ1 of length r which lies outside B α(0), ρr . Similarly,
we can
r
connect y and h by a path γ2 of length r which lies outside B α(0), ρr . Let γ3 be
the subsegment of α connecting α(−ρr) and h−r . Let γ4 be the subsegment of α
connecting α(ρr) and hr . It is not hard to see the length of γ3 and γ4 are both (k − ρ)r.
Let γ = γ3 ∪γ1 ∪γ ∪γ2 ∪γ4 . Then, γ is a path that lies outside B α(0), ρr connecting
α(−ρr) and α(ρr). Therefore, the length of γ is at least Divα (ρr). Also,
ℓ(γ) = ℓ(γ3 ) + ℓ(γ1 ) + ℓ(γ) + ℓ(γ2 ) + ℓ(γ5 ) = ℓ(γ) + 2(k − ρ + 1)r.
Thus,
ℓ(γ) ≥ Divα (ρr) − 2(k − ρ + 1)r.
This implies that
dρr,HΓ (x, y) ≥ Divα (ρr) − 2(k − ρ + 1)r.
Geometric embedding properties of Bestvina-Brady subgroups
11
Therefore,
δρn (r) ≥ Divα (ρr) − 2(k − ρ + 1)r.
Thus, the relative divergence Div(AΓ , HΓ ) is at least quadratic.
The following theorem is deduced from Proposition 4.3, Proposition 5.2, and Proposition 5.3.
Theorem 5.4 Let Γ be a connected, finite, simplicial graph with at least 2 vertices. Let
AΓ be the associated right-angled Artin group and HΓ the Bestvina-Brady subgroup.
Then the relative divergence Div(AΓ , HΓ ) and the subgroup distortion DistAHΓΓ are both
linear if Γ is a join graph. Otherwise, the relative divergence Div(AΓ , HΓ ) and the
subgroup distortion DistAHΓΓ are both quadratic.
Corollary 5.5 For each integer n ≥ 3 , there is a right-angled Artin group containing
a free subgroup of rank n whose inclusion is not a quasi-isometric embedding.
Proof For each positive integer n ≥ 3, let Γ be a tree with n edges such that Γ is
not a join graph. By the above theorem, the distortion of HΓ in the right-angled Artin
group AΓ is quadratic. Also, HΓ is the free group of rank n by Theorem 2.6.
6 Connection to horizontal surface subgroups
Definition 6.1 A graph manifold is a compact, irreducible, connected orientable 3–
manifold M that can be decomposed along T into finitely many Seifert manifolds,
where T is the canonical decomposition tori of Johannson and of Jaco-Shalen. We
call the collection T is JSJ-decomposition in M , and each element in T is JSJ-torus.
Definition 6.2 If M is a Seifert manifold, a properly immersed surface g : S # M
is horizontal, if g(S) is transverse to the Seifert fibers everywhere. In the case M is
a graph manifold, a properly immersed surface g : S # M horizontal if g(S) ∩ Pv is
horizontal for every Seifert component Pv .
Theorem 6.3 If Γ is a finite tree with at least 3 vertices, then the associated rightangled Artin group AΓ is a fundamental group of a graph manifold and the BestvinaBrady subgroup HΓ is a horizontal surface subgroup.
12
Hung Cong Tran
bv
u2
u3
u1
u4
u6
u5
u7
Figure 1: A punctured disk Σv when the degree of v in Γ is 7.
Proof First, we construct the graph manifold M whose fundamental group is AΓ as
follows:
Let v be a vertex of Γ of degree k ≥ 2. Let u1 , u2 , · · · , uk be all elements in ℓk(v).
Let Σv be a punctured disk with k inside holes in which their boundaries are labeled
by elements in ℓk(v). We also label the outside boundary component of Σv by bv (see
Figure 1). Obviously, π1 (Σv ) is the free group generated by u1 , u2 , · · · uk .
Let Pv = Σv × Sv1 , here we label the circle factor in Pv by v. Obviously, each Pv is
a Seifert manifold. Moreover, for each ui in ℓk(v), the Seifert manifold Pv contains
torus Su1i × Sv1 as a component of its boundary.
We construct the graph manifold by gluing pair of Seifert manifolds (Pv1 , Pv2 ) along
their tori Sv11 × Sv12 whenever v1 and v2 are adjacent vertices in Γ. We observe that the
pair of such regions are glued together by switching fiber and base directions. It is not
hard to see that the fundamental group of M is the right-angled Artin group AΓ .
We now construct the horizontal surface S in M with the Bestvina-Brady subgroup HΓ
as its fundamental group. We first construct the horizontal surface Sv on each Seifert
piece Pv = Σv × Sv1 , where v is a vertex of Γ of degree k ≥ 2.
We remind the reader that Σv is a punctured disk with k inside holes in which their
boundaries are labeled by all elements u1 , u2 , · · · , uk in ℓk(v). We also label the outside
boundary component of Σv by bv (see Figure 1). We label the circle factor in Pv by v.
Geometric embedding properties of Bestvina-Brady subgroups
13
Let Sv be a copy of the punctured disk Σv . However, we relabel all inside circles by
c1 , c2 , · · · , ck and the outside circle by cv . We will construct a map (g, h) : Sv → Σv ×Sv1
as follows:
(1) The map g is the identity map that maps each ci to ui and cv to bv .
(2) The map h has degree −1 on boundary component ci and degree k on cv .
We now construct the map h with the above properties. We observe that the fundamental
group of Sv is generated by c1 , c2 , · · · , ck , and cv with a unique relator c1 c2 c3 · · · ck cv =
e. Here we abused notation for the presentation of π1 (Sv ). By that presentation of
π1 (Sv ), we can see that there is a group homomorphism φ from π1 (Sv ) to Z that maps
each ci to −1 and cv to k. By [Hat02, Proposition 1B.9], the group homomorphism φ
is induced by a map h from Sv to Sv1 . Therefore, we constructed a desired map h.
Finally, we identify the surface Sv with its image via the map (g, h). By construction,
π1 (Sv ) is the subgroup of π1 (Pv ) generated by elements u1 v−1 , u2 v−1 , · · · , uk v−1 . We
observe that if we glue pair of Seifert manifolds (Pv1 , Pv2 ) along their tori Sv11 ×Sv12 , pair
of horizontal surfaces (Sv1 , Sv2 ) will be matched up along their boundaries in Sv11 × Sv12 .
Therefore, we constructed a horizontal surface S in M . By Vankampen theorem, the
fundamental group of S is generated by all elements of the form st−1 whenever s and
t are adjacent vertices in Γ. In other words, π1 (S) is the Bestvina-Brady subgroup by
Theorem 2.6.
References
[BB97] Mladen Bestvina and Noel Brady. Morse theory and finiteness properties of groups.
Invent. Math., 129(3):445–470, 1997.
[BC12] Jason Behrstock and Ruth Charney. Divergence and quasimorphisms of right-angled
Artin groups. Math. Ann., 352(2):339–356, 2012.
[Car]
Mike Carr. Two-generator subgroups of right-angled artin groups are quasiisometrically embedded. Preprint. arXiv:1412.0642.
[DL99] Warren Dicks and Ian J. Leary. Presentations for subgroups of Artin groups. Proc.
Amer. Math. Soc., 127(2):343–348, 1999.
[Hat02] Allen Hatcher. Algebraic topology. Cambridge University Press, Cambridge, 2002.
[HN]
G. Christopher Hruska and Hoang Thanh Nguyen. Distortion of surfaces in graph
manifolds. In Preparation.
[Tra15] Hung Cong Tran. Relative divergence of finitely generated groups. Algebr. Geom.
Topol., 15(3):1717–1769, 2015.
14
Hung Cong Tran
Department of Mathematics, The University of Georgia, 1023 D. W. Brooks Drive, Athens,
GA 30605, United States
[email protected]
This figure "i1.jpg" is available in "jpg" format from:
http://arxiv.org/ps/1606.00539v2
| 4math.GR
|
Dynamic Weight Alignment for Convolutional Neural Networks
Brian Kenji Iwana and Seiichi Uchida
Department of Advanced Information Technology
Kyushu University, Fukuoka, Japan
Abstract
Input
In this paper, we propose a method of improving Convolutional Neural Networks (CNN) by determining the optimal alignment of weights and inputs using dynamic programming. Conventional CNNs convolve learnable shared
weights, or filters, across the input data. The filters use a
linear matching of weights to inputs using an inner product between the filter and a window of the input. However,
it is possible that there exists a more optimal alignment of
weights compared to linear matching. Thus, we propose
the use of Dynamic Time Warping (DTW) to dynamically
align the weights to optimized input elements. This dynamic
alignment is useful for time series recognition due to the
complexities with temporal relations and temporal distortions. We demonstrate the effectiveness of the proposed architecture on the Unipen online handwritten digit and character datasets, the UCI Spoken Arabic Digit dataset, and
the UCI Activities of Daily Life dataset.
Convolution Window
arXiv:1712.06530v2 [cs.CV] 25 Jan 2018
brian, [email protected]
Convolutional Layer
with Dynamic
Weight Alignment
Fully-Connected
Layer
Output
𝑤4
𝑤3
𝑤2
𝑤1
Figure 1. A 3-layer CNN with the first hidden layer as a convolutional layer with dynamically aligned weights. In each stride, the
shared weights are individually aligned to the optimal input.
Memory (LSTM) networks [20].
The key feature of a CNN is using sparsely connected
shared weights as a convolution. In this way, the shared
weights, or filter, act as a feature extractor and maintains the
structural aspects from the input. Furthermore, the shared
weights are linearly aligned to each corresponding window
value of the input. This interaction is much like a linear
inner product or linear kernel in that the result is large when
the filter is similar to the receptive field and small when it is
different.
Difficulties arise with time series recognition due to issues with temporal distortions such as varying rates. The
linear alignment assumes that each element of the receptive field correspond directly to each weight of the filter in a
one-to-one fashion. Conversely, time series patterns do not
have linear relationships between elements. RNNs address
these issues with the use of recurrent connections to store
persistent information across time without strict time step
alignment. In addition, LSTMs are capable of overcoming
local rate inconsistencies with the use of gates to control
the learning and develop long-term dependencies. However, most feedforward network models, including CNNs,
1. Introduction
Neural networks and perceptron learning models have
become a powerful tool in machine learning and pattern recognition. Early models were introduced in the
1970s, but recently have achieved state-of-the-art results
due to improvements in data availability and computational
power [46]. Convolutional Neural Networks (CNN) [34] in
particular have achieved the state-of-the-art results in many
areas of image recognition, such as offline handwritten digit
recognition [52], text digit recognition [35, 50], and object
recognition [6, 16, 17].
In addition to the image domain, CNN models have been
used for time series patterns. A predecessor to CNNs, Time
Delay Neural Networks (TDNN) [51, 33] used time-delay
windows similar to the filters of CNNs. CNNs were also
used to classify time series by embedding the sequences into
vectors [59] and matrices [43, 55]. However, most recent
successes have been through the use of Recurrent Neural
Networks (RNN) [27] and in particular, Long Short-Term
1
do not have the same considerations for time series and do
not protect against temporal distortions.
To overcome this, we propose a method of finding the
optimal alignment between the weights and the inputs using dynamic programming, namely Dynamic Time Warping (DTW) [44]. DTW determines an optimal distance between two time series patterns by elastically matching elements using dynamic programming along a constrained
path on a cost matrix. It combines the local distances between matched elements to determine the optimized global
distance. We exploit the elastic matching utility of DTW
to align the weights of the filter to the elements of the corresponding receptive field to create more efficient feature
extractors for CNNs.
The contribution of this paper is twofold. First, we propose a novel method of aligning weights within the convolutional filters of CNNs. This alignment of weights allows CNNs to effectively tackle time series patterns and
be used for time series recognition. Second, we demonstrate the effectiveness of the proposed method on multiple
time series datasets including: Unipen online handwritten
digit and character datasets, the UCI Spoken Arabic Digit
dataset, and the UCI Activities of Daily Life dataset. We
perform a comparative study using a traditional CNN to reveal the benefits of the proposed weight alignment. In addition, we compare the proposed method to other state-ofthe-art methods for time series classification.
The rest of this paper is organized as follows. In Section 2, we review other neural methods of time series recognition. Section 3 details the proposed dynamic weight alignment method for CNNs. Section 4 reports the experiments
with results and Section 5 discusses the advantages of dynamic weight alignment. Finally, in Section 6, we provide
a conclusion and lay out future work.
2. Related Work
Neural networks and perceptron learning methods for
time series recognition have been extensively explored in
literature. Many classic models were used for time series
forecasting [57, 23, 58] where artificial neural networks
were provided a sliding window of a discrete signal and
were tasked with predicting future time steps. However,
these methods learn rigid relationships to temporal features
and are susceptible to temporal distortions and shift variance.
TDNNs are classical feed-forward neural networks designed to address shift variance [51] by introducing wider
sliding window receptive fields to represent multiple points
in time. These windows, or time delays, are passed through
learned filters to create subsequent feature map time series.
TDNNs have been successful in speech and phoneme recognition [48]. CNNs continue the idea of learned filters permeated across a structural input to the domain of images.
However, CNNs have been used for time series recognition by considering the time steps as a dimension for the
convolutions. Zheng et al. [59] use CNNs with 1D subsequences created from multivariate time series. There also
have been attempts [31, 43, 55, 53] to classify time series
patterns by embedding them into 2D matrices for classification by CNN.
WaveNet [40] employs dilated causal convolutional layers to generate audio. The causal convolutions have relatively small receptive fields, however, the convolutions are
dilated meaning that higher layers skip input values and
increase the effective range of the receptive field. The
WaveNet architecture stacks multiple dilated causal convolutions to a very large receptive field which is considerate
to order.
There are other notable feedforward network architectures that have been proposed to address the characteristics
or features of time series patterns. For instance, MLPs have
been applied to wavelet transforms [1] and ARMA(p, q)
time series data [23] for forecasting and classification. Spiking neurons [36, 3] create sparse representations using brief
spikes time series data.
Other than feedforward networks, recurrent models like
RNNs have been successful in time series recognition. Similar to other models, RNNs are given a sliding window of
the input time series. However, unlike feedforward networks, RNNs use recurrent connections and stored hidden
states to maintain a dependence on time. LSTMs are an improvement on RNNs which use gates to control the amount
of learned information to solve the vanishing gradient problem [19]. LSTMs have been successful in many time series and sequence domains such as handwriting [10, 12],
speech [11], and natural language processing [49].
Dynamic neural networks is an emerging field in neural model learning There are generally two approaches,
dynamically embedding trained parameters into a network
and using dynamic connections. Dynamic Filter Networks (DFN) [7] use filter-generating networks to produce
filters that are dynamically used depending on the input.
Klein et al. [32] similarly uses filters that vary depending on the input. Dynamic Convolutional Neural Networks (DCNN) [29] use dynamic k-Max Pooling to simplify CNNs for sentence modeling. The distinction between
these models and the proposed method is that we use dynamic programming to optimize the weight alignment inside nonlinear convolutions.
DTW-NNs [26] similarly use DTW as a nonlinear inner product for neural networks. The difference between
a DTW-NN and the proposed model is that DTW-NN uses
a sum of Euclidean distances between a learned prototype
sequence and the input. This means the weights are only
pseudo weights whereas the CNN described in this paper
are used as true weights. In addition, we use a shared weight
CNN model instead of sparsely aligned network. The advantage of a convolution is that the output of the node retains temporal qualities instead of a scalar output, which allows for the ability to stack multiple nonlinear inner product
convolutional layers.
3. Dynamic Weight Alignment for CNNs
The goal of the proposed method is to exploit dynamic programming to determine the optimal alignment of
weights for convolutional layers in CNNs. By using DTW,
we can adapt the conventional linear inner product of a convolution to optimize the connections between convolutional
filter weights and inputs. Figure 2 demonstrates the difference between a conventional convolutional layer with linear weight alignment and the proposed CNN with dynamic
weight alignment.
3.1. Convolutional Neural Networks
A key feature of CNNs is that they maintain sparse connections that promote local connectivity. Each element of
a convolutional node only has a receptive field of a small
number of locally neighboring inputs. Besides enforcing
local relations, the sparsity of a CNN reduces the number
of parameters needed to be calculated and allows for the
use of larger input sizes.
Parameter sharing further reduces the number of parameters required training a CNN. The idea is that the weights
of a convolutional layer are shared for each corresponding
output element’s local receptive field. In this way, a forward
calculation of a convolutional layer is identical to a convolution operation where the shared weights are the filter and
the output is a feature map.
Formally, the feature map zxl of a convolutional layer is
defined as:
F
X
l
wfl al−1
(1)
zxl =
x+f + bx
f =0
for each element x, where l is the convolutional layer, l − 1
is the previous layer, f is the index of the filter, and F + 1
l
is the window size. We denote wfl , al−1
x+f , and bx as the
shared weights, the previous layer activations, and the bias
respectively. In other words, zxl is the inner product of the
shared weights wl and each window of the previous layer
l−1
al−1
x , . . . , ax+F .
This inner product linearly matches the weights to the
inputs within the window. However, it is plausible that there
exist instances where particular weights should be matched
with more optimal inputs, for example noisy elements or
feature translation and scale variance within the filter.
3.2. Dynamic Weight Alignment
To optimize the alignment of weights wl and the window
l−1
of the previous layer al−1
x , . . . , ax+F , we adopt a dynamic
programming solution, namely DTW. Giving the nonlinear
matching of weights and the values from the window, we
consider the result of DTW as a nonlinear inner product.
Dynamic Time Warping DTW is an asymmetric positive
semi-definite similarity function that traditionally used as a
distance measure between sequences. It is calculated using
dynamic programming to determine the optimal match of
elements between two sequences. By matching elements,
the sequences are warped in the time dimension to align
similar features of the time series. DTW distance has shown
to be a successful distance measure for time series recognition [54, 8, 42].
DTW finds the total cost over an optimal warping path of
a local cost matrix using dynamic programming. Given two
discrete time series, sequence p = p1 , . . . , pi , . . . , pI of
length I and sequence s = s1 , . . . , sj , . . . , sJ , where i and
j are the index of each time step and pi and sj are elements
at each time step, the DTW-distance is the global summation of local distances between pairwise element matches.
Namely, the DTW-distance is denoted as:
X
DTW(p, s) =
||pi0 − sj 0 || ,
(2)
(i0 ,j 0 )∈M
where (i0 , j 0 ) is a pair of matched indices i0 and j 0 corresponding to the original indices i of p and j of s, respectively. The set M contains all matched pairs of i0 and j 0 .
Additionally, the set of matched pairs M can contain repeated and skipped indices of i and j from the original sequences, therefore, M has a nonlinear correspondence to
1, . . . , i, . . . , I and 1, . . . , j, . . . , J. || · || is a local distance
function between elements.
The use of DTW as a distance measure or similarity
function is not unlike a kernel function and has been integrated into the kernel functions of Support Vector Machines (SVM) [5, 47, 2] and as a replacement for the inner product of fully-connected neural network layers in Dynamic Time Warping Neural Networks (DTW-NN) [26].
Dynamic weight alignment with shared weights. If we
define the distance function || · || of Eq. 2 as the product
of pi0 and sj 0 , then we can consider DTW as a nonlinear
inner product. Unlike the conventional convolution defined
in Eq. 1, the result of DTW would become the nonlinear
product of weights and the window of the previous layer.
Namely, we propose using DTW to determine zxl :
l−1
l
zxl = DTW(wl , {al−1
x , . . . , ax+F }) + bx
X
l
=
wfl 0 al−1
x 0 + bx ,
(3)
(4)
(f 0 ,x0 )∈M
where M is the set of optimally matched indices f 0 and
x0 corresponding to the index f of wl and the index x
Layer 𝑙 − 1
Layer 𝑙
Layer 𝑙 − 1
𝑎7
𝑧4
𝑎7
𝑎6
𝑧3
𝑎6
Product Matrix
Layer 𝑙
Layer 𝑙 − 1
Layer 𝑙
Layer 𝑙 − 1
𝑎7
𝑧4
𝑎7
𝑎6
𝑧3
𝑎6
𝑎5
𝑧2
𝑎5
𝑎4
𝑧1
𝑎4
Product Matrix
Layer 𝑙
𝑧4
𝑎5
𝑧2
𝑎5
𝑎4
𝑧1
𝑎4
𝑧4
𝑎2
𝑎1
𝑧2
𝑎3
𝑎2
𝑤4
𝑤3
𝑤2
𝑧1
𝑤1
𝑤2
𝑤3
𝑎3
𝑎2
𝑧2
𝑎3
𝑤4
𝑤2
𝑧1
𝑤1
𝑎1
𝑤1
(a) Convolution with linear weight alignment
𝑤4
𝑤3
𝑎2
𝑎1
𝑎1
𝑤1
Convolution Window
𝑎3
𝑧3
Convolution Window
Convolution Window
Convolution Window
𝑧3
𝑤2
𝑤3
𝑤4
(b) Convolution with dynamic weight alignment
Figure 2. The comparison between a conventional linear convolution (a) and the proposed convolution with dynamic weight alignment (b).
Both illustrate 1D convolutions with four weights w1 , . . . , w4 at stride 1. The layer l − 1 is the previous layer with elements a1 , . . . , a7
and layer l is the resulting feature map from the convolution with elements z1 , . . . , z4 . Each black dot is the product of the corresponding
weight and input and the blue circle is the sum of the products.
in {x, . . . , x + F } of al−1 , respectively. When used in
this manner, we create a nonlinear convolutional filter that
acts as a feature extractor similar to using shapelets with
DTW [56].
To create the nonlinearity of the weight alignment of the
filters, DTW allows for a relaxed matching as determined
by the warping path within constraints. One constraint, a
slope constraint should be applied to the warping path to
prevent overfitting. We propose using an asymmetric slope
constraint which is a modification of the slope constraint
proposed by Itakura [25]. In the traditional implementation
of DTW, a the local distance metric is a dissimilarity function and the goal is to minimize the global distance. However, when using inner product space as the cost matrix, the
local distance metric represents similarity, therefore we aim
to find the maximal product. Thus, we use a slope constraint
defined by the recurrent function:
D(f, x) = wfl al−1
+
x
max
x0 ∈{x,x−1,x−2}
D(f − 1, x0 ),
(5)
where D(f, x) is the cumulative sum at the f -th element of
wfl and the x-th element of al−1
x . The global DTW distance,
or the result of the nonlinear inner product, is defined as the
value at D(F, X).
In addition are many other possible slope and local constraints that have been researched [44, 38, 28]. However,
the proposed method uses a slope constraint that ensures
that the warping path is monotonic and that the number of
matches is always equal to the number of elements in the
filter.
3.3. Backpropagation of Convolutions with Dynamic Weight Alignment
In order to train the network, we use Stochastic Gradient
Decent (SGD) to determine the gradients of the weights in
respect to the error in order update the weights to minimize
the loss. For a CNN, the gradient of the shared weights with
respect to the error is the partial derivative:
X ∂C ∂z l
∂C
x
=
,
l ∂w l
∂z
∂wfl
x
f
x
(6)
where C is the loss function. In a conventional CNN, wfl
has a linear relationship to zxl , thus
l
∂zx
∂wfl
can be calcu-
lated simply. However, given the nonlinearity of the weight
alignment, the calculation of the gradient is reliant on the
matched elements determined by the forward pass in:
P
l l−1
l
X ∂C ∂
(f 0 ,x0 )∈M wf 0 ax0 + bx
∂C
=
. (7)
∂zxl
∂wfl
∂wfl
x
Unlike a conventional CNN, the set of matched indicies M in Eq. 4 allows for duplicate and skipped values of wfl 0 and al−1
Thus, the relationship between
x0 .
P
l l−1
l
l
(f 0 ,x0 )∈M wf 0 ax0 +bx and wf is the sum of the windows
l−1
of the matched activations ax0 from the previous layer for
each matched weight wfl 0 . Finally, the Eq. 7 can be solved,
giving us:
X
∂C
l+1
=
δ
al−1
(8)
x0 ,
∂wfl
0 0
(f ,x )∈M
l+1
where δ
is the backpropagated error from the previous
layer as determined by the chain rule.
4. Experiments
We demonstrate the effectiveness of the proposed
method by quantitatively evaluating the architecture and
compare it baseline methods.
4.1. Architecture Settings
For the experiment, we implement a four layer dynamically aligned CNN. The first hidden layer is a convolutional
layer with 50 nodes of the proposed dynamically aligned
filters. In addition, we use batch normalization on the results of the convolutional layer. The second and third layers
are fully-connected layers with a hyperbolic tangent tanh
activation and have 400 and 100 nodes respectively. The
final output layer uses softmax with the number of outputs
corresponding to the number of classes.
Batch normalization. Batch normalization helps in the
generalization ability of a neural network [24]. It aims to
reduce the effects of the internal covariate shift which forces
higher layers of a network to drift which causes whitening
of the inputs. To solve this problem, batch normalization
shifts the values so that the mean is zero and the variance
is one. The proposed method uses batch normalization as
described by [24] after the convolutional layer.
Learning Rate. By using a learning rate with a progressive decay, gradient decent of the network will quickly learn
the weight parameters in the beginning and slowly refine
them later in the training. The learning rate ηt at iteration t
is defined as:
η0
,
(9)
ηt =
1 + αt
where η0 is the initial learning rate and α is the decay parameter. For all of the experiments, we use the 1/t progressive learning rate with a η0 = 0.001 and α = 0.001.
4.2. Comparison Methods
To evaluate the proposed method, we compare the accuracy to comparison methods. We report classification results literature as well as evaluate the datasets on established
state-of-the-art neural network methods.
Reported baselines. We compare the results of the proposed method to reported results from literature.
For the online handwritten digits and character evaluation, we compare results from two neural network models,
a Regional-Fuzzy Representation (RFR) with an MLP [41]
and a DTW-NN [26], and two other methods, an SVM
using a Gaussian DTW kernel (DAG-SVM-GDTW) [2]
and a Hidden Markov Model (HMM) [21]. We also
compare our results to a Fuzzy Adaptive System ARTbased (FasArt) [45] model and and a Kohonen-Perceptron
network using fuzzy vector representation (Fuzzy Rep.
KP) [18].
For the spoken Arabic digits, there is one reported
neural network solution using a Wavelet Neural Network (WNN) [22] as well as other models using a Tree
Distribution model [14] and a Continuous Hidden Markov
Model of the second-order derivative MFCC (CHMM w/
∆(∆MFCC)) [15].
For the activities of daily life (ADL) dataset, we compare our results to the original dataset proposal by Bruno
et al. [4] who modeled the ADL accelerometer data using
Gaussian Mixture Modeling and Gaussian Mixture Regression (GMM + GMR). We also report the best results of
Kanna et al. [30] who used a Decision Tree.
Evaluated baselines. The evaluated baselines were designed to be direct comparisons for the proposed method.
The LSTM is used as the established state-of-the-art neural
network method for sequence and time series recognition
and a traditional CNN is used as a direct comparison using
standard convolutional layers. Both comparative models are
provided with the same exact training, test, and validation
sets as the proposed method. Furthermore, the evaluated
methods use the same batch size and number of iterations
as the proposed method for the respective trials.
For the LSTM evaluation, an LSTM with two recursive
hidden layers and a softmax output layer was used. The
LSTM was set to a forget gate bias of 1.0 and had a static
learning rate of 0.001.
The second comparative evaluation was using a CNN
with the same exact hyperparameters as the proposed
method, but with standard convolutional nodes. It was constructed of one 50-node convolutional layer with a window
size and stride equal each proposed method evaluation with
batch normalization, two fully-connected tanh layers size
400 and 100 respectively, and a softmax output layer. The
convolutional layer weights had the same progressive 1/t
decay learning rate as evaluation of the proposed method
with η0 = 0.001 and α = 0.001 and a static learning rate of
0.0001 between the fully-connected layers.
4.3. Classification of Online Handwritten Digits and
Characters
The isolated online handwritten digits and characters offer an ideal source of data for time series recognition due to
having distinct classes but having numerous variations.
4.3.1
Dataset
The Unipen multi-writer 1a, 1b, and 1c datasets [13] is constructed from pen tip trajectories of isolated numerical digits, uppercase alphabet characters, and lowercase alphabet
Method
DAG-SVM-GDTW [2]
HMM [21]
RFR [41]
DTW-NN [26]
FasArt [45]
Fuzzy Rep. KP [18]
LSTM
CNN
Proposed Method
1a
96.2
96.8
96.9
98.2
85.66
96.1
95.92
97.92
98.08
1b
92.4
93.6
94.9
89.39
94.50
95.67
1c
87.9
85.9
85.6
94.5
86.88
94.75
94.83
Table 1. Accuracy (%) on the Unipen online handwritten digit (1a),
uppercase character (1b), and lowercase character (1c) datasets.
The highest accuracy for each dataset is in bold.
characters respectively. The datasets are provided by the
International Unipen Foundation (iUF).1
We used a set of 13,000 randomly selected digits of
10 classes for the 1a experiment and a set of 12,350 randomly selected characters of 26 classes for the 1b and 1c
experiments. For the experiments, we preprocessed the patterns with re-sampling to 50 sequence elements and scaled
to fit in a square defined by the corner coordinates of
(−0.5, −0.5) and (0.5, 0.5). The datasets were divided into
three sets for training, a test of 10% of the data, a training
set of 90% of the data, and 50 patterns set a side from the
training set for a validation set.
4.3.2
Training
Given that the experimental dataset is made of sequences of
two dimensional coordinates, the filters should correspond
accordingly. Thus, the convolutional filters were of size
7 × 2 and were applied to the input sequences at stride 2. A
stride 2 was used to reduce redundant information and decrease computation time. The filters were initialized using
a Gaussian distribution with a minimum value of -0.5 and a
maximum of 0.5, within a similar range as the dataset. The
experiment uses batch gradient decent with a batch size of
100 for 30,000 iterations.
4.3.3
Results
The results of the experiments on the Unipen 1a, 1b, and
1c datasets are shown in Table 1. The proposed method
achieved a test accuracy of 98.08% for digits, 95.67% for
uppercase characters, and 94.83% for lowercase characters after 30,000 iterations of training, surpassing the conventional CNN and the LSTM. Furthermore, the proposed
method surpassed all of the comparative methods except for
DTW-NN with 98.2% for the Unipen 1a trial.
1 URL:
http://www.unipen.org/home.html
4.4. Classification of Spoken Arabic Digits
Audio is a prime example of a time series and speech
and phoneme classification is essential for speech recognition. One popular method of audio recognition is the
use of Mel-Frequency Cepstrum Coefficients (MFCC) [9].
MFCCs transforms the audio signals into discrete time series by using mel-frequency cepstrum representations with
the number of dimensions equaling the number of coefficients.
4.4.1
Dataset
We used the UCI Spoken Arabic Digit Data Set2 for the
experiments. The dataset consists of 8,800 13-frequency
MFCCs of spoken Arabic numerals in 10 classes. The
samples were taken at 11,025 Hz, 16 bit sampling rate
with a Hamming window, and with filters pre-emphasized
1 − 0.97z −1 .
The UCI Spoken Arabic Digit Data Set has a pre-defined
division of the data with a speaker-independent training set
and test set of 6,600 patterns and 2,200 patterns respectively. 50 patterns from the training set were separated as
validation during training. For the experiments, we preprocessed the patterns with re-sampling to 40 sequence ele1
.
ments and scaled by 15
4.4.2
Training
The UCI Spoken Arabic Digit Data Set has 13-frequency
MFCCs, therefore we use the data as 40 time steps of 13
dimensional data. The convolutional filters are set to 5 × 13
at stride 2 with a minimum value of -0.5 and a maximum of
0.5 to fit the data. We trained the proposed method with a
batch size of 50 for 30,000 iterations.
4.4.3
Results
The proposed method had an accuracy of 96.95% for the
predefined test set with 2,200 speaker independent MFCC
spoken Arabic digits. When comparing the results in Table 2, it is noted that the proposed method was outperformed by CHMM w/ ∆(∆MFCC). Nonetheless, the proposed CNN with dynamically aligned weights performed
better than the conventional CNN, the LSTM, and the other
reported results. In addition, CHMM w/ ∆(∆MFCC) is a
method tailored to MFCC speech recognition whereas the
proposed method works well with other domains.
4.5. Classification of Activities of Daily Life
Study of ADL data is important for understanding the
mechanics of daily life and helpful for for applications such
2 URL:
https://archive.ics.uci.edu/ml/datasets/
Spoken+Arabic+Digit
Accuracy (%)
93.1
98.4
96.7
96.0
94.77
96.95
Table 2. Accuracy (%) on the UCI Spoken Arabic Digit Data Set.
The highest accuracy for each dataset is in bold.
Method
GMM + GMR [4]
Decision Tree [30]
LSTM
CNN
Proposed Method
Table 3. Accuracy (%) on the UCI ADL Recognition with Wristworn Accelerometer Data Set. The highest accuracy for each
dataset is in bold.
4.5.1
Dataset
One common method of studying ADL is the use of wearable sensors. An example of this is the UCI ADL Recognition with Wrist-worn Accelerometer Data Set.3 The dataset
contains 705 recorded accelerometer data split into 7 classes
of about 100 patterns each. The actions include “Climbing
stairs”, “Drinking from a glass”, “Getting up from bed”,
“Pouring water in a glass”, “Sitting in a chair”, “Standing
from a chair”, and “Walking.” They were recorded at 32 Hz
in the three spacial dimensions.
1
For the experiment, we scaled the data to 1.5g
where g
is the acceleration due to gravity. The time series were also
sampled to 200 frames. The original experiment [4] uses
a median filter, however we found it unnecessary and used
the raw data for the experiment. For training, the dataset
was divided at random into a training set of 600 patterns, a
validation set of 5 patterns, and a test set of 70 patterns.
4.5.2
Training
The dataset is made of tri-axial accelerometer data. Due
to the length of the samples, 200 frames, we used a larger
convolutional filter of 9 × 3 at stride 4. They were initiated similarly with a Gaussian distribution with a minimum
value of -0.5 and a maximum of 0.5 in all three dimensions.
Finally, only a batch size of 5 was used due to the very small
dataset.
4.5.3
Results
Similar to the other datasets, the results in Table 3 show that
the proposed method improved the accuracy over the conventional CNN with linearly aligned weights, with 85.7%
and 84.3% respectively. The result of the proposed CNN
with dynamically aligned weights far exceeds the accuracy
of the original dataset proposal using GMM + GMR.
3 URL:
https://archive.ics.uci.edu/ml/datasets/
Dataset+for+ADL+Recognition+with+Wrist-worn+
100
95
90
85
80
0
Proposed Method
CNN
5000
10000
15000
Iteration
20000
25000
Unipen 1a
Test Accuracy (%)
Test Accuracy (%)
100
as robotics, assisted ADL, and the monitoring of disabled
or elderly patients [39].
Accuracy (%)
63.1
80.9
80.0
84.3
85.7
30000
100
95
90
85
80
0
Proposed Method
CNN
5000
10000
15000
Iteration
20000
Unipen 1b
25000
Test Accuracy (%)
Method
Tree Distribution [14]
CHMM w/ ∆(∆MFCC) [15]
WNN [22]
LSTM
CNN
Proposed Method
30000
95
90
85
80
0
Proposed Method
CNN
5000
10000
15000
Iteration
20000
25000
30000
Unipen 1c
Figure 3. Comparison of test accuracies of the Unipen online handwritten datasets between a conventional CNN and the proposed
CNN with dynamic weight alignment over the number of training
iterations.
5. Discussion
In this section, we discuss comparisons between the proposed method and the evaluated comparisons. We also offer
explanations for the improved accuracy. Additional results
such as the confusion matrices and accuracy comparisons
are provided in the supplementary materials.
CNN with dynamically aligned weights versus LSTM.
LSTMs are the state-of-the-art sequence and time series
neural network model. They are RNNs designed to maintain long-term dependencies through gating.
In the online handwriting and ADL experiments, the
LSTM performed poorly compared to both CNNs. One reason for the limited performance of the LSTM is that each
individual element of those datasets do not contain a significant amount of information and the model needs to know
how all elements work together to form spacial structures.
For example, for large tri-axial accelerometer data, individual long-term dependencies are not as important as the local
and global structures whereas CNNs excels. Another reason
for the poor performance of the ADL dataset could be the
low amount of training data (600 training samples), high
amounts of noise, and a high variation of patterns within
each class. However, the LSTM did comparatively well on
the spoken Arabic digits.
CNN with dynamically aligned weights versus CNN with
linearly aligned weights. It is important to analyze the
Accelerometer
differences between the conventional CNN with linearly
aligned weights and the proposed method with dynamically
aligned weights. Table 1 reveals a 7.69% reduction in error
for Unipen 1a, 21.21% reduction for 1b, and a 1.59% reduction for 1c. In addition to the increased accuracy, we can observe from Fig. 3 that compared to the conventional CNN,
the proposed method achieves a higher accuracy during all
parts of training but especially during the early stages. This
indicates that the nonlinear alignment optimizes the use of
the weights before the generalization of the CNN converges.
0
1
2
3
4
5
6
7
8
9
(a) CNN with linear weight alignment
Efficient feature extraction. One explanation of the improved accuracy is that aligning the weights to their similar
corresponding inputs is more efficient than conventional linear matching. The weights of a convolutional layer learned
learned by a CNN act like filter for feature extraction [34].
And, the purpose of using dynamically aligned weights is to
warp the assignment of weights to their most similar corresponding inputs. The DTW algorithm skips redundant and
noisy inputs while ensuring that the target features are detected. This way, the features that are extracted by the filter
are enhanced.
0
1
2
3
4
5
6
7
8
9
(b) CNN with dynamic weight alignment
Improved discriminability at the layer level. Figure 4
is a visualization of the outputs of the convolutional layers from the test set using t-Distributed Stochastic Neighbor
Embedding (t-SNE) [37]. t-SNE is an algorithm to visualize high-dimensional data which allows the visualization of
the response from the convolutional node for each sample in
low-dimensional space. The figure shows that the classes,
represented by different colors, are more discriminable with
the dynamic weight alignment. Figure 4 (a) is from the test
samples using a conventional CNN and it shows that many
of the classes are made of multiple small clusters whereas
Fig. 4 (b) generally contains larger unified clusters. For example, classes “0”, “6”, and “9” are made of multiple clusters in Fig 4 (a), but are unified in (b). The t-SNE figures for
the other evaluations can be found in the supplementary materials. The higher discriminability means that there would
be less confusion for the higher layers.
6. Conclusion
In this paper, we proposed a novel method of optimizing
the weights within a convolutional filter of a CNN through
the use of dynamic programming. We implemented DTW
as a method of sequence element alignment between the
weights of a filter and the inputs of the corresponding receptive field. By defining the local distance function of DTW
as a product, we show that it is possible to use DTW as a
less rigid nonlinear inner product. In this way, the weights
of the convolutional layer are aligned to maximize their relationship to the data from the previous layer.
Figure 4. t-SNE visualization of the output of the convolutional
node for the Unipen 1a (digits) experiment. Each point is a test
sample colored to indicate the class membership. (a) is a conventional CNN and (b) is a CNN with the proposed weight alignment.
We show that the proposed model is able to tackle
time series pattern recognition. We evaluated the proposed
model on the Unipen online handwritten digit and character
datasets. The results show n-class classification test accuracies of 98.08% for online handwritten digits, 95.67% for
uppercase characters, and 94.83% for lowercase characters.
In addition to handwriting, we demonstrate the effectiveness of the proposed model on other types of data including
13-coefficient MFCC speech of Arabic digits and tri-axial
accelerometer ADL data. The spoken Arabic digit evaluate
achieved a 96.95% accuracy and the accelerometer-based
ADL evaluation achieved a 85.7% accuracy. This shows
that not only is the proposed method a viable feedforward
neural network model for time series recognition, but it is
an effective method of optimizing the convolutional filter
for CNNs.
Dynamic weight alignment for convolutions of CNNs is
a promising area of research. The proposed method can
be applied to other CNN based models such as deep CNNs
or fully convolutional networks. In the future, we plan on
extending this method to other time series datasets as well
as other domains, such as images.
References
[1] A. K. Alexandridis and A. D. Zapranis. Wavelet neural networks: A practical guide. Neural Networks, 42:1–27, 2013.
2
[2] C. Bahlmann, B. Haasdonk, and H. Burkhardt. Online handwriting recognition with support vector machines-a kernel
approach. In Proc. 8th Int. Workshop Frontiers in Handwriting Recognition, pages 49–54. IEEE, 2002. 3, 5, 6
[3] R. Brette, M. Rudolph, T. Carnevale, M. Hines, D. Beeman,
J. M. Bower, M. Diesmann, A. Morrison, P. H. Goodman,
F. C. Harris, et al. Simulation of networks of spiking neurons: a review of tools and strategies. J. Computational Neuroscience, 23(3):349–398, 2007. 2
[4] B. Bruno, F. Mastrogiovanni, A. Sgorbissa, T. Vernazza, and
R. Zaccaria. Analysis of human behavior recognition algorithms based on acceleration data. In 2013 IEEE Int. Conf.
Robotics and Automation, pages 1602–1607. IEEE, 2013. 5,
7
[5] W. Chaovalitwongse and P. Pardalos. On the time series
support vector machine using dynamic time warping kernel for brain activity classification. Cybern. and Syst. Anal.,
44(1):125–138, 2008. 3
[6] D.-A. Clevert, T. Unterthiner, and S. Hochreiter. Fast and
accurate deep network learning by exponential linear units
(elus). arXiv preprint arXiv:1511.07289, 2015. 1
[7] B. De Brabandere and X. Jia. Dynamic filter networks. In
Conf. Neural Inform. Process. Systems, 2016. 2
[8] H. Ding, G. Trajcevski, P. Scheuermann, X. Wang, and
E. Keogh. Querying and mining of time series data: experimental comparison of representations and distance measures.
Proc. Very Large Data Base Endowment, 1(2):1542–1552,
2008. 3
[9] T. Ganchev, N. Fakotakis, and G. Kokkinakis. Comparative
evaluation of various mfcc implementations on the speaker
verification task. In 2005 Proc. Int. Conf. Speech and Comput., pages 191–194, 2005. 6
[10] A. Graves, M. Liwicki, S. Fernandez, R. Bertolami,
H. Bunke, and J. Schmidhuber. A novel connectionist system for unconstrained handwriting recognition. IEEE Trans.
Pattern Anal. Mach. Intell., 31(5):855–868, may 2009. 2
[11] A. Graves, A.-R. Mohamed, and G. Hinton. Speech recognition with deep recurrent neural networks. In 2013 IEEE
Int. Conf. Acoustics, Speech and Signal Process. IEEE, May
2013. 2
[12] A. Graves and J. Schmidhuber. Offline handwriting recognition with multidimensional recurrent neural networks. In
Advances in Neural Inform. Process. Syst., pages 545–552,
2009. 2
[13] I. Guyon, L. Schomaker, R. Plamondon, M. Liberman, and
S. Janet. Unipen project of on-line data exchange and recognizer benchmarks. In Proc. 12th IAPR Int. Conf. Pattern
Recognition, volume 2, pages 29–33. IEEE, 1994. 5
[14] N. Hammami and M. Bedda. Improved tree model for arabic
speech recognition. In 3rd IEEE Int. Conf. Comp. Sci. and
Inform. Technology, volume 5, pages 521–526. IEEE, 2010.
5, 7
[15] N. Hammami, M. Bedda, and F. Nadir. The secondorder derivatives of mfcc for improving spoken arabic digits
recognition using tree distributions approximation model and
hmms. In 2012 Int. Conf. Commun. and Inform. Technology,
pages 1–5. IEEE, 2012. 5, 7
[16] K. He, X. Zhang, S. Ren, and J. Sun. Delving deep into
rectifiers: Surpassing human-level performance on imagenet
classification. In Proc. IEEE Int. Conf. on Comput. Vision,
pages 1026–1034, 2015. 1
[17] K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning
for image recognition. In Proc. IEEE Conf. Comp. Vision
and Pattern Recognition, pages 770–778, 2016. 1
[18] J.-F. Hébert, M. Parizeau, and N. Ghazzali. A new fuzzy geometric representation for online isolated character recognition. In Proc. 14th Int. Conf. Pattern Recognition, volume 2,
pages 1121–1123. IEEE, 1998. 5, 6
[19] S. Hochreiter, Y. Bengio, P. Frasconi, and J. Schmidhuber.
Gradient flow in recurrent nets: the difficulty of learning
long-term dependencies. Wiley-IEEE Press, 2001. 2
[20] S. Hochreiter and J. Schmidhuber. Long short-term memory.
Neural Computation, 9(8):1735–1780, nov 1997. 1
[21] J. Hu, S. G. Lim, and M. K. Brown. Writer independent online handwriting recognition using an hmm approach. Pattern Recognition, 33(1):133–147, 2000. 5, 6
[22] X. Hu, L. Zhan, Y. Xue, W. Zhou, and L. Zhang. Spoken arabic digits recognition based on wavelet neural networks. In
2011 IEEE Int. Conf. Syst., Man, and Cybern., pages 1481–
1485. IEEE, 2011. 5, 7
[23] H. B. Hwarng and H. Ang. A simple neural network for arma
(p, q) time series. Omega, 29(4):319–333, 2001. 2
[24] S. Ioffe and C. Szegedy. Batch normalization: Accelerating
deep network training by reducing internal covariate shift.
arXiv preprint arXiv:1502.03167, 2015. 5
[25] F. Itakura. Minimum prediction residual principle applied to
speech recognition. IEEE Trans. Acoustics, Speech, and Sig.
Proc., 23(1):67–72, 1975. 4
[26] B. K. Iwana, V. Frinken, and S. Uchida.
A robust
dissimilarity-based neural network for temporal pattern
recognition. In 15th Int. Conf. Frontiers in Handwriting
Recognition, pages 265–270. IEEE, 2016. 2, 3, 5, 6
[27] H. Jaeger. Tutorial on training recurrent neural networks,
covering BPPT, RTRL, EKF and the” echo state network”
approach, volume 5. 2002. 1
[28] R. Kakisako, S. Uchida, and V. Frinken. Learning nonmarkovian constraints for handwriting recognition. In 13th
Int. Conf. on Document Anal. and Recognition, pages 446–
450. IEEE, 2015. 4
[29] N. Kalchbrenner, E. Grefenstette, and P. Blunsom. A convolutional neural network for modelling sentences. In Proc.
52nd Annu. Meeting Assoc. for Comput. Linguistics, pages
655–665, 2014. 2
[30] K. R. Kanna, V. Sugumaran, T. Vijayaram, and
C. Karthikeyan. Activities of daily life (adl) recognition using wrist-worn accelerometer. Int. J. of Eng. and
Technology, 4(3):1406–1413, 6 2016. 5, 7
[31] Y. Kim. Convolutional neural networks for sentence classification. In Proc. 2014 Conf. Empirical Methods in Natural
Language Process., pages 1746–1751, 2014. 2
[32] B. Klein, L. Wolf, and Y. Afek. A dynamic convolutional
layer for short range weather prediction. In Proc. IEEE Conf.
Comput. Vision and Pattern Recognition, pages 4840–4848,
2015. 2
[33] K. J. Lang, A. H. Waibel, and G. E. Hinton. A time-delay
neural network architecture for isolated word recognition.
Neural Networks, 3(1):23–43, 1990. 1
[34] Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner. Gradientbased learning applied to document recognition. Proc. IEEE,
86(11):2278–2324, 1998. 1, 8
[35] C.-Y. Lee, P. W. Gallagher, and Z. Tu. Generalizing pooling
functions in convolutional neural networks: Mixed, gated,
and tree. In Int. Conf. Artificial Intell. and Stat., pages 464–
472, 2016. 1
[36] W. Maass. Networks of spiking neurons: the third generation
of neural network models. Neural Networks, 10(9):1659–
1671, 1997. 2
[37] L. v. d. Maaten and G. Hinton. Visualizing data using t-sne.
Journal of Machine Learning Research, 9(Nov):2579–2605,
2008. 8
[38] C. Myers, L. Rabiner, and A. Rosenberg. Performance tradeoffs in dynamic time warping algorithms for isolated word
recognition. IEEE Trans. Acoustics, Speech, and Sig. Proc.,
28(6):623–635, 1980. 4
[39] T. S. of the Benjamin Rose Hospital. Multidisciplinary studies of illness in aged persons. J. of Chronic Diseases, 9(1):55
– 62, 1959. 7
[40] A. v. d. Oord, S. Dieleman, H. Zen, K. Simonyan, O. Vinyals,
A. Graves, N. Kalchbrenner, A. Senior, and K. Kavukcuoglu.
Wavenet: A generative model for raw audio. arXiv preprint
arXiv:1609.03499, 2016. 2
[41] M. Parizeau, A. Lemieux, and C. Gagné. Character recognition experiments using unipen data. In Proc. 6th Int. Conf.
Document Anal. and Recognition, pages 481–485. IEEE,
2001. 5, 6
[42] T. Rakthanmanon, B. Campana, A. Mueen, G. Batista,
B. Westover, Q. Zhu, J. Zakaria, and E. Keogh. Searching
and mining trillions of time series subsequences under dynamic time warping. In Proc. 18th Int. Conf. on Knowledge
Discovery and Data Mining. ACM, 2012. 3
[43] N. Razavian and D. Sontag. Temporal convolutional neural networks for diagnosis from lab tests. arXiv preprint
arXiv:1511.07938, 2015. 1, 2
[44] H. Sakoe and S. Chiba. Dynamic programming algorithm
optimization for spoken word recognition. IEEE Trans.
Acoustics, Speech, and Sig. Proc., 26(1):43–49, 1978. 2, 4
[45] E. G. Sánchez, J. G. González, Y. A. Dimitriadis, J. C.
Izquierdo, and J. L. Coronado. Experimental study of a
novel neuro-fuzzy system for on-line handwritten unipen
digit recognition. Pattern Recognition Letters, 19(3):357–
364, 1998. 5, 6
[46] J. Schmidhuber. Deep learning in neural networks: An
overview. Neural Networks, 61:85–117, 2015. 1
[47] H. Shimodaira, K.-i. Noma, M. Nakai, and S. Sagayama.
Dynamic time-alignment kernel in support vector machine.
2002. 3
[48] M. Sugiyama, H. Sawai, and A. H. Waibel. Review of tdnn
(time delay neural network) architectures for speech recognition. In IEEE Int. Symp. on Circuits and Syst., pages 582–
585. IEEE, 1991. 2
[49] M. Sundermeyer, R. Schlüter, and H. Ney. Lstm neural networks for language modeling. In Interspeech, 2010. 2
[50] S. Uchida, S. Ide, B. K. Iwana, and A. Zhu. A further step to
perfect accuracy by training cnn with larger data. In 15th Int.
Conf. Frontiers in Handwriting Recognition. IEEE, 2016. 1
[51] A. Waibel, T. Hanazawa, G. Hinton, K. Shikano, and K. J.
Lang. Phoneme recognition using time-delay neural networks. IEEE Trans. Acoust., Speech, Signal Process.,
37(3):328–339, 1989. 1, 2
[52] L. Wan, M. Zeiler, S. Zhang, Y. L. Cun, and R. Fergus. Regularization of neural networks using dropconnect. In Proc.
30th Int. Conf. Mach. Learning, pages 1058–1066, 2013. 1
[53] Z. Wang and T. Oates. Encoding time series as images for
visual inspection and classification using tiled convolutional
neural networks. In Assoc. Advancement Artificial Intell.
Workshop, pages 40–46, 2015. 2
[54] X. Xi, E. Keogh, C. Shelton, L. Wei, and C. A. Ratanamahatana. Fast time series classification using numerosity reduction. In Proc. 23rd Int. Conf. Machine learning, pages
1033–1040. ACM, 2006. 3
[55] J. Yang, M. N. Nguyen, P. P. San, X. Li, and S. Krishnaswamy. Deep convolutional neural networks on multichannel time series for human activity recognition. In 2015
Int. Joint Conf. Artificial Intell., pages 3995–4001, 2015. 1,
2
[56] L. Ye and E. Keogh. Time series shapelets: a new primitive
for data mining. In Proc. 15th ACM Int. Conf. Knowledge
Discovery and Data Mining, pages 947–956. ACM, 2009. 4
[57] G. Zhang, B. E. Patuwo, and M. Y. Hu. Forecasting with
artificial neural networks:: The state of the art. Int. J. Forecasting, 14(1):35–62, 1998. 2
[58] G. P. Zhang. Time series forecasting using a hybrid arima
and neural network model. Neurocomputing, 50:159–175,
2003. 2
[59] Y. Zheng, Q. Liu, E. Chen, Y. Ge, and J. L. Zhao. Time series
classification using multi-channels deep convolutional neural
networks. In Web-Age Information Management, pages 298–
310. Springer Science + Business Media, 2014. 1, 2
| 9cs.NE
|
Submitted to the Annals of Statistics
arXiv: arXiv:1701.06088
DISTRIBUTED INFERENCE FOR QUANTILE
REGRESSION PROCESSES
arXiv:1701.06088v3 [math.ST] 10 Apr 2018
By Stanislav Volgushev¶,§ , Shih-Kang Chaok,† and Guang
Chengk,‡,†
University of Toronto¶ and Purdue Universityk
The increased availability of massive data sets provides a unique
opportunity to discover subtle patterns in their distributions, but
also imposes overwhelming computational challenges. To fully utilize
the information contained in big data, we propose a two-step procedure: (i) estimate conditional quantile functions at different levels in
a parallel computing environment; (ii) construct a conditional quantile regression process through projection based on these estimated
quantile curves. Our general quantile regression framework covers
both linear models with fixed or growing dimension and series approximation models. We prove that the proposed procedure does not
sacrifice any statistical inferential accuracy provided that the number of distributed computing units and quantile levels are chosen
properly. In particular, a sharp upper bound for the former and a
sharp lower bound for the latter are derived to capture the minimal
computational cost from a statistical perspective. As an important
application, the statistical inference on conditional distribution functions is considered. Moreover, we propose computationally efficient
approaches to conducting inference in the distributed estimation setting described above. Those approaches directly utilize the availability of estimators from sub-samples and can be carried out at almost
no additional computational cost. Simulations confirm our statistical
inferential theory.
1. Introduction. The advance of technology in applications such as
meteorological and environmental surveillance or e-commerce has led to extremely large data sets which cannot be processed with standalone computers due to processor, memory, or disk bottlenecks. Dividing data and
∗
The authors would like to thank the Editors and two anonymous Referees for helpful
comments that helped to considerably improve an earlier version of this manuscript.
†
Partially supported by Office of Naval Research (ONR N00014-15-1-2331).
‡
Partially supported by NSF CAREER Award DMS-1151692, DMS-1418042.
§
Partially supported by discovery grant from Natural Sciences and Engineering Research Council of Canada.
MSC 2010 subject classifications: Primary 62F12, 62G15, 62G20
Keywords and phrases: B-spline estimation, conditional distribution function, distributed computing, divide-and-conquer, quantile regression process
1
2
computing among many machines is a common way to alleviate such bottlenecks, and can be implemented by parallel computing systems like Hadoop
(White, 2012) with the aid of communication-efficient algorithms.
In statistics, this approach to estimation has recently been adopted under
the name divide-and-conquer, see Li et al. (2012); Jordan (2013). Pioneering contributions on theoretical analysis of divide and conquer algorithms
focus on mean squared error rates Zhang et al. (2013, 2015). The analysis
therein does not take into account a core question in statistics – inferential procedures. In the last two years, such procedures have been developed
for various non- and semi-parametric estimation approaches that focus on
the mean or other notions of central tendency, this line of work includes
Shang and Cheng (2017); Zhao et al. (2016); Shi et al. (2017); Banerjee
et al. (2017), among others. Focusing on the mean tendency illuminates
one important aspect of the dependence between predictors and response,
but ignores all other aspects of the conditional distribution of the response
which can be equally important. Additionally, most of the work cited above
assumes homoskedastic errors or sub-Gaussian tails. Both assumptions are
easily violated in modern messy and large-scale data sets, and this limits
the applicability of approaches that are available to date.
We propose to use quantile regression in order to extract features of the
conditional distribution of the response while avoiding tail conditions and
taking heteroskedasticity into account. This approach focuses on estimating
the conditional quantile function x 7→ Q(x; τ ) := FY−1
|X (τ |x) where FY |X
denotes the conditional distribution of response given predictor. A flexible
class of models for conditional quantile functions can be obtained by basis
expansions of the form Q(x; τ ) ≈ Z(x)> β(τ ). This includes linear models
of fixed or increasing dimension where the approximation is replaced by an
equality, as well as a rich set of non- and semi-parametric procedures such
as tensor product B-splines or additive models.
Given data {(Xi , Yi )}N
i=1 , quantile regression for such models is classically
formulated through the following minimization problem:
(1.1)
βbor (τ ) := arg minm
b∈R
N
X
ρτ {Yi − b> Z(Xi )},
i=1
where ρτ (u) := {τ − 1(u ≤ 0)}u and 1(·) is the indicator function. However,
solving this minimization problem by classical algorithms requires that the
complete data set can be processsed on a single computer. For large samples,
this might not be feasible. This motivates us to utilize the divide and conquer
approach where the full sample is randomly split across several computers
into S smaller sub-samples of size n, the minimization in (1.1) is solved on
DISTRIBUTED INFERENCE FOR QUANTILE REGRESSION PROCESSES
3
each sub-sample, and results are averaged in the end to obtain the final
estimator β(τ ) (see (2.3) in Section 2 for a formal definition). While this
approach is easy to implement, the resulting estimator β(τ ) is typically
not a solution to the original minimization problem (1.1). As illustrated in
Figure 1.1, this can be problematic for inference procedures. More precisely,
the left panel of Figure 1.1 depicts the coverage probabilities (on the y-axis)
against number of sub-samples (on the x-axis) for asymptotic confidence
intervals that are based on asymptotic distributions of βbor (τ ) but computed
using β(τ ) for three data generating processes (linear models with different
dimension) and a fixed quantile τ = 0.1. This indicates that choosing S too
large results in invalid inference, as reflected by a rapid drop in coverage
after a certain threshold is reached. For different models this threshold is
different and it intrinsically depends on various properties of the underlying
model such as dimension of predictors. These observations indicate that
developing valid statistical inference procedures based on β(τ ) requires a
careful theoretical analysis of divide and conquer procedures. The first major
contribution of the present paper is to provide statistical inferential theory
for this framework.
1.0
Normal(0, 1) , τ = 0.1
●
●
●
●
●
K
●
●
Coverage Probability
0.4
0.6
0.8
●
●
0.2
N 1/(2ητ )
N
m
0.0
N = 214 , m = 4
N = 214 , m = 16
N = 214 , m = 32
0.0
0.2
1/(2ητ )
●
0.4
logN(S)
0.6
●
●
0.8
N 1/2
m1/2 log2 N
1/2
N
m
N 1/2
log2 N
N 1/2
S
Fig 1.1: Left penal: coverage probabilities (y-axis) of confidence intervals
for estimators computed from divide and conquer; x-axis: number of subsamples S in log-scale. Different colors correspond to linear models with
different m = dim(Z(X)). Right panel: necessary and sufficient conditions on
(S, K) for the oracle rule in linear models with fixed dimension (Blue) and Bspline nonparametric models (Green). The dotted region is the discrepancy
between the sufficient and necessary conditions.
The approach described so far provides a way to estimate the conditional
quantile function for a single quantile. To obtain a picture of the entire conditional distribution of response given predictors, estimation of the conditional
quantile function at several quantile levels is required. Performing such an
estimation for a dense grid of quantiles can be computationally costly as
4
estimation at each new quantile requires running the above algorithm anew.
To relax this computational burden, we introduce the two-step quantile projection algorithm. In the first step, the divide and conquer algorithm is used
to compute β(τ1 ), ..., β(τK ) for a grid of quantile values τ1 , ..., τK . In the secb is computed from β(τ1 ), ..., β(τK ) (see (2.7) for a formal
ond step, a matrix Ξ
b ) can
b the quantile projection estimator β(τ
definition). Given this matrix Ξ,
be computed for any τ by performing a multiplication of this matrix with a
known vector (depending on τ ) without access to the original dataset. Based
b ), we can estimate the conditional distribution function; see (2.8) and
on β(τ
(2.9).
The computational cost of our procedure is determined by the number
of sub-samples S and the number of quantile levels K. To minimize the
computational burden, K should be chosen as small as possible. Choosing S
large will allow us to maintain a small sample size n on each local machine,
thus enabling us to process larger data sets. At the same time, classical
inference procedures for βbor (e.g. procedures based on asymptotic normality
b ).
of βbor ) should remain valid for β(τ ), β(τ
A detailed analysis of conditions which ensure this ”oracle rule” is conducted in Section 3. There, we derive sufficient conditions for both S and
K, which are also proved to be necessary in some cases up
to log N terms.
Sufficient conditions of S take the form S = o N 1/2 aN,m where aN,m → 0.
The specific form of aN,m depends on the precise form of the model, with
more complex models leading to aN,m converging to zero faster, thus placing
more restrictions on S. Necessary and sufficient conditions on K take the
form K N 1/(2η) kZ(x)k−1/η , where η > 0 is the degree of Hölder continuity (see (1.2)) of τ 7→ Q(x; τ ). In particular, this lower bound is completely
independent of S. An interesting insight which will be explained in more
detail in Section 3.2 is that for large values of kZk, which typically corresponds to more complex models, the restriction on K becomes weaker; see
the paragraph right after Corollary 3.10 for a more detailed explanation.
A graphical summary of the necessary and sufficient conditions on K and
S derived in Sections 3.1 and 3.2 is provided in the right panel of Figure 1.1.
Deriving such necessary conditions on S, K is a crucial step in understanding the limitations of divide and conquer procedures, as it shows that
the interplay between S, K and model complexity is indeed a feature of the
underlying problem and not an artifact of the proof; see also Shang and
Cheng (2017).
While the oracle rules described above provide justification for using the
asymptotic distribution of βbor for inference, this distribution is typically not
pivotal. We discuss several ways to overcome this problem. First, we propose
DISTRIBUTED INFERENCE FOR QUANTILE REGRESSION PROCESSES
5
simple and computationally efficient approaches which directly make use of
the fact that β̄ is formed by averaging results from independent sub-samples.
Those approaches are based on normal and t-approximations as well as a
novel bootstrap procedure. Second, we comment on extensions of traditional
approaches to inference which rely on estimating the asymptotic variance
of βbor by using kernel smoothing techniques. Simulations demonstrate finite
sample properties of the proposed inference strategies.
The rest of this paper is organized as follows. In Section 2, we provide a
formal description of all algorithms and estimation procedures discussed in
this paper. Section 3 contains the main theoretical results, with additional
results presented in Appendix A. In Section 4, we propose several practical
approaches to inference. In Section 5, we validate the finite sample relevance
of our theoretical results through extensive Monte Carlo experiments. Proofs
of all theoretical results are deferred to the online supplement. The following
notation will be used throughout the paper.
Notation. Let X := supp(X). Let Z = Z(X) and Zi = Z(Xi ) and assume
T = [τL , τU ] for some 0 < τL < τU < 1. Define the true underlying measure
of (Yi , Xi ) by P and denote the corresponding expectation as E. Denote by
kbk the L2 -norm of a vector b. λmin (A) and λmax (A) are the smallest and
largest eigenvalue of a matrix A. Let bηc denote the integer part of a real
number η, and |j| = j1 + ... + jd for d-tuple j = (j1 , ..., jd ). S m−1 ⊂ Rm
is the unit sphere. an bn means that (|an /bn |)n∈N and (|bn /an |)n∈N are
bounded, and an bn means an /bn → ∞. Define the class of functions
n
Ληc (T ) := f ∈ C bηc (T ) : sup sup |Dj f (τ )| ≤ c,
j≤bηc τ ∈T
(1.2)
sup sup
j=bηc τ 6=τ 0
o
− Dj f (τ 0 )|
≤
c
,
kτ − τ 0 kη−bηc
|Dj f (τ )
where η is called the ”degree of Hölder continuity”, and C α (X ) denotes the
class of α-continuously differentiable functions on a set X .
2. A two-step procedure for computing quantile process. In this
section, we formally introduce two algorithms studied throughout the paper
- divide and conquer and quantile projection. The former is used to estimate
quantile functions at fixed quantile levels, based on which the latter constructs an estimator for the quantile process. We also note that algorithms
presented below can be applied for processing data that are locally stored
for any reason and not necessarily large. As an important application, the
estimation of conditional distribution functions will be presented.
6
We consider a general approximately linear model:
Q(x; τ ) ≈ Z(x)> βN (τ ),
(2.1)
where Q(x; τ ) is the τ -th quantile of the distribution of Y conditional on
X = x ∈ Rd , and Z(x) ∈ Rm is a vector of transformation of x. This
framework (2.1) incorporates various estimation procedures, for instance,
series approximation models; see Belloni et al. (2017) for more discussion.
In this paper, we will focus on three classes of transformation Z(x) ∈ Rm
which include many models as special cases: fixed and finite m, diverging m
with local support structure (e.g. B-splines) and diverging m without local
support structure.
The divide and conquer algorithm for estimating βN (τ ) at a fixed τ ∈
(0, 1) is described below:
1
1. Divide the data {(Xi , Yi )}N
i=1 into S sub-samples of size n . Denote
the s-th sub-sample as {(Xis , Yis )}ni=1 where s = 1, ..., S.
2. For each s and fixed τ , estimate the sub-sample based quantile regression coefficient as follows
(2.2)
βbs (τ ) := arg minm
β∈R
n
X
i=1
ρτ Yis − β > Z(Xis ) .
3. Each local machine sends βbs (τ ) ∈ Rm to the master that outputs a
pooled estimator
(2.3)
β(τ ) := S −1
S
X
s=1
βbs (τ ).
Remark 2.1. The minimization problem in (2.2) is in general not strictly
convex. Thus, several minimizers could exist. In the rest of this paper, we
will only focus on the minimizer with the smallest `2 norm. This is assumed
for presentational convenience, and a close look at the proofs reveals that
all statements remain unchanged if any other minimizer is chosen instead.
While β(τ ) defined in (2.3) gives an estimator at a fixed τ ∈ T , a complete
picture of the conditional distribution is often desirable. To achieve this, we
propose a two-step procedure. First compute β(τk ) ∈ Rm for each τk ∈ TK ,
where TK ⊂ T = [τL , τU ] is grid of quantile values in T with |TK | = K ∈ N.
1
The equal sub-sample size is assumed for ease of presentation, and can be certainly
relaxed.
DISTRIBUTED INFERENCE FOR QUANTILE REGRESSION PROCESSES
7
Second project each component of the vectors {β(τ1 ), ..., β(τK )} on a space
of spline functions in τ . More precisely, let
(2.4)
b j := arg minq
α
α∈R
K
X
k=1
2
β j (τk ) − α> B(τk ) ,
j = 1, ..., m.
where B := (B1 , ..., Bq )> is a B-spline basis defined on G equidistant knots
τL = t1 < ... < tG = τU with degree rτ ∈ N (we use the normalized version
of B-splines as given in Definition 4.19 on page 124 in Schumaker (1981)).
Here, uniformity of knots is assumed for simplicity, all theoretical results
b j , we define
can be extended to knots with quasi uniform partitions. Using α
b : T → Rm :
the quantile projection estimator β(·)
b ) := Ξ
b > B(τ ),
β(τ
(2.5)
b is defined below. The jth element βbj (τ ) = α
b>
where Ξ
j B(τ ) can be viewed
PK
2
as projection, with respect to kf kK := ( k=1 f (τk ))1/2 , of β j onto the
polynomial spline space with basis B1 , ..., Bq . In what follows, this projection
is denoted by ΠK . Although we focus on B-splines in this paper, other basis
functions are certainly possible.
b is summaThe algorithm for computing the quantile projection matrix Ξ
rized below, here the divide and conquer algorithm is applied as a subroutine:
1. Define a grid of quantile levels τk = τL +(k/K)(τU −τL ) for k = 1, ..., K.
For each τk , compute β(τk ) as (2.3).
2. For each j = 1, ..., m, compute
(2.6)
bj =
α
X
K
B(τk )B(τk )
k=1
>
−1 X
K
k=1
B(τk )β j (τk ) ,
which is a closed form solution of (2.4).
3. Set the matrix
(2.7)
b := [α
b1 α
b 2 ... α
b m ].
Ξ
A direct application of the above algorithm is to estimate the quantile
b ) is the estifunction for any τ ∈ T . A more important application of β(τ
mation of the conditional distribution function FY |X (y|x) for any fixed x.
More precisely, we consider
Z τU
b ) < y}dτ,
b
(2.8)
FY |X (y|x) := τL +
1{Z(x)> β(τ
τL
8
where τL and τU are chosen close to 0 and 1. The intuition behind this
approach is the observation that
(2.9)
τL +
Z
τU
τL
if FY |X (y|x) < τL ;
τL
FY |X (y|x) if τL ≤ FY |X (y|x) ≤ τU ;
1{Q(x; τ ) < y}dτ =
τU
if FY |X (y|x) > τU ,
see Chernozhukov et al. (2010); Volgushev (2013). The estimator FbY |X is a
b ) (Chernozhukov et al., 2010).
smooth functional of the map τ 7→ Z(x)> β(τ
b ) as a process
Hence, properties of FbY |X depend on the behavior of Z(x)> β(τ
in τ , which we will study in detail in Section 3.
3. Theoretical analysis. The following regularity conditions are needed
throughout this paper.
(A1) Assume that kZi k ≤ ξm < ∞, where ξm is allowed to diverge, and that
1/M ≤ λmin (E[ZZ> ]) ≤ λmax (E[ZZ> ]) ≤ M holds uniformly in N for
some fixed constant M .
(A2) The conditional distribution FY |X (y|x) is twice differentiable w.r.t. y,
with the corresponding derivatives fY |X (y|x) and fY0 |X (y|x). Assume
|f 0 (y|x)| < ∞
f¯ := sup
|fY |X (y|x)| < ∞, f 0 := sup
y∈R,x∈X
y∈R,x∈X
Y |X
uniformly in N .
(A3) Assume that uniformly in N , there exists a constant fmin < f such
that
0 < fmin ≤ inf inf fY |X (Q(x; τ )|x).
τ ∈T x∈X
In these assumptions, we explicitly work with triangular array asymptotics for {(Xi , Yi )}N
i=1 , where d = dim(Xi ) is allowed to grow as well.
Assumptions (A1) – (A3) are fairly standard in the quantile regression
literature, see Belloni et al. (2017). The eigenvalue condition in (A1) is imposed for the purpose of normalization, which in turn determines the value
of ξm . For basis expansions with local support such as B-splines, the right
√
scaling to ensure the eigenvalue condition results is ξm m.
3.1. Fixed dimensional linear models. In this section, we assume for all
τ ∈ T and x ∈ X
(3.1)
Q(x; τ ) = Z(x)> β(τ ),
where Z(x) has fixed dimension m. This simple model setup allows us to
derive a simple and clean bound on the difference between β, βb and the
DISTRIBUTED INFERENCE FOR QUANTILE REGRESSION PROCESSES
9
oracle estimator βbor and, leading to a better understanding of resulting
oracle rules. Our first main result is as follows.
Theorem 3.1. Suppose that (3.1) and assumptions (A1) – (A3) hold
and that K N 2 , S = o(N (log N )−1 ). Then
S log N S 1/4 (log N )7/4
b
(3.2) sup kβ(τ )− βor (τ )k = OP
+oP (N −1/2 ).
+
3/4
N
N
τ ∈TK
If additionally K G 1 we also have
1/2
2
b ) − βbor (τ ))| ≤ OP S log N + S (log N ) + oP (N −1/2 )
sup |Z(x0 )> (β(τ
N
N
τ ∈T
+ sup |(ΠK Q(x0 ; ·))(τ ) − Q(x0 ; τ )|
τ ∈T
where the projection operator ΠK was defined right after (2.5).
The proof of Theorem 3.1 is given in Section S.3.1. To obtain this result, we develop a new Bahadur representation for each local estimator
βbs (τ ), see Section S.2.1. The main novelty compared to existing results
is a sharp bound on the expectation of the remainder term. This is crucial
to obtain the bound in (3.2). In contrast, relying on previously available
versions of the Bahadur representation would result in a bound of the form
(S/N )3/4 (log N )3/4 , which is somewhat more restrictive. See Remark S.2.3
for additional details. Theorem 3.1 can be generalized to potentially missspecified linear models with dimension that depends on the sample size.
Technical details are provided in the Appendix, Section A.
The bound in (3.2) quantifies the difference between β and βbor in terms of
the number of sub-samples S and can be equivalently formulated in terms of
the sub-sample size n = N/S provided that log N ∼ log n. When considering
the projection estimator, an additional bias term ΠK Q(x0 ; ·) − Q(x0 ; ·) is
introduced. Provided that for a given x0 ∈ X one has τ 7→ Q(x0 ; τ ) ∈
Ληc (T ), this bias term can be further bounded by O(G−η ). Note that in
the setting of Theorem 3.1 the oracle estimator converges to β(τ ) at rate
OP (N −1/2 ), uniformly in τ ∈ T . By combining the results in Theorem 3.1
with this bound, upper bounds on convergence rates for β(τ ) − β(τ ) and
b ) − Q(x0 ; τ )| follow as direct Corollaries.
supτ ∈T |Z(x0 )> (β(τ
Theorem 3.1 only provides upper bounds on the differences between β, βb
and βbor . While a more detailed expression for this difference is derived in
the proof of Theorem S.2.1, this expression is very unwieldy and does not
lead to useful explicit formulas. However, it is possible to prove that the
10
bounds given in Theorem 3.1 are sharp up to log N factors, which is the main
result of the next Theorem. Before stating this result, we need to introduce
some additional notation. Denote by P1 (ξm , M, f¯, f 0 , fmin ) all pairs (P, Z) of
distributions P and transformations Z satisfying (3.1) and (A1)-(A3) with
constants 0 < ξm , M, f¯, f 0 < ∞, fmin > 0. Since m, ξm are constant in this
section, we use the shortened notation P1 (ξ, M, f¯, f 0 , fmin ).
Theorem 3.2. For any τ in T there exists (P, Z) ∈ P1 (ξ, M, f¯, f 0 , fmin )
and a C > 0 such that
CS
(3.3)
lim sup P kβ(τ ) − βbor (τ )k ≥
> 0.
N
N →∞
Moreover for any c, η > 0 there exist (P, Z) ∈ P1 (ξ, M, f¯, f 0 , fmin ) such that
τ 7→ βj (τ ) ∈ Ληc (T ), j = 1, ..., d and
(3.4)
b ) − βbor (τ )k ≥ CS + CG−η > 0.
lim sup P sup kβ(τ
N
N →∞
τ ∈T
The proof of Theorem 3.2 is given in Section S.3.1. The result provided
in (3.3) has an interesting implication: the best possible precision of estimating Q in a divide-and-conquer framework is restricted by n = N/S, the
sample size that can be processed on a single machine, regardless of the
total sample size N . A related observation was made in Example 1 of Zhang
et al. (2013) who construct a data-generating process where the MSE rate
of a divide and conquer estimator is limited by the sample size that can be
processed on a single computer.
As corollaries to the above results, we derive sufficient and necessary conditions on S under which β and Z(x)> βb satisfy the oracle rule. Note that
the asymptotic distribution of the oracle estimator βbor (τ ) under various
conditions has been known for a long time, see for instance Theorem 4.1 of
Koenker (2005). Under (A1)-(A3) it was developed in Belloni et al. (2017)
and Chao et al. (2017) who show that
√
(3.5)
N βbor (·) − β(·)
G(·) in (`∞ (T ))d ,
where G is a centered Gaussian process with covariance structure
(3.6) H(τ, τ 0 ) := E G(τ )G(τ 0 )>
= Jm (τ )−1 E Z(X)Z(X)> Jm (τ 0 )−1 (τ ∧ τ 0 − τ τ 0 ).
where Jm (τ ) := E[ZZ> fY |X (Q(X; τ )|X)].
DISTRIBUTED INFERENCE FOR QUANTILE REGRESSION PROCESSES 11
√
Corollary 3.3 (Oracle rules for β). A sufficient condition for N (β̄(τ )−
β(τ ))
N (0, H(τ, τ )) for any (P, Z) ∈ P1 (ξ, M, f¯, f 0 , fmin ) is that S =
1/2
o(N / log N ). A necessary condition for the same result is that S = o(N 1/2 ).
The necessary and sufficient conditions above match up to a factor of
log N . In order to guarantee the oracle rule for the process, we need to
impose additional conditions on the smoothness of τ 7→ β(τ ) and on the
number of grid points G.
b
Corollary 3.4 (Oracle rules for β).
Assume that τ 7→ βj (τ ) ∈ Ληc (T )
for j = 1, ..., d and√given c, η > 0, that N 2 K G and rτ ≥ η. A suffib
cient condition for N (β(·)−β(·))
G(·) for any (P, Z) ∈ P1 (ξ, M, f¯, f 0 , fmin )
satisfying above conditions is S = o(N 1/2 (log N )−1 ) and G N 1/(2η) . A
necessary condition for the same result is S = o(N 1/2 ) and G N 1/(2η) .
Corollary 3.4 characterizes the choice for parameters (S, K) which determine computational cost. The conditions on S remain the same as in
Corollary 3.3. The process oracle rule requires restrictions on K based on
the smoothness of τ 7→ Q(x0 ; τ ), denoted as η. Note that, compared to the
results in Belloni et al. (2017), smoothness of τ 7→ Q(x0 ; τ ) is the only additional
condition of the data that is needed to ensure process convergence of
√
b − β(·)). Specifically, the lower bound on K in terms of N becomes
N (β(·)
smaller as η increases, which implies that smoother τ 7→ Q(x0 ; τ ) allow for
larger computational savings. Corollary 3.4 implies that the condition on K
is necessary for the oracle rule, no matter how S is chosen.
Next we apply Corollary 3.4 to the estimation of conditional distribution
functions. Define
Z τU
or
b
(3.7)
1{Z(x)> βbor (τ ) < y}dτ.
FY |X (·|x0 ) := τL +
τL
The asymptotic distribution of FbYor|X (·|x0 ) was derived in Chao et al. (2017).
Corollary 3.5. Under the same conditions as Corollary 3.4, we have
for any x0 ∈ X ,
√
N FbY |X (·|x0 ) − FY |X (·|x0 )
−fY |X (·|x0 )Z(x0 )> G FY |X (·|x0 )
in `∞ (Q(x0 ; τL ), Q(x0 ; τU )) ,
where FbY |X (·|x0 ) is defined in (2.8) and G is a centered Gaussian process
with covariance given in (3.6). The same process convergence result holds
with FbYor|X (·|x0 ) replacing FbY |X (·|x0 ).
12
The proof of Corollary 3.5 uses the fact that y 7→ FbY |X (y|x) is a Hadamard
b ) for any fixed x. The result of this
differentiable functional of τ 7→ Z(x)> β(τ
corollary can be extended to other functionals with this property.
3.2. Approximate linear models with local basis structure. In this section,
we consider models with Q(x; τ ) ≈ Z(x)> β(τ ) with m = dim(Z) → ∞ as
N → ∞ where the transformation Z corresponds to a basis expansion. The
analysis in this section focuses on transformations Z with a specific local
support structure, which will be defined more formally in Condition (L).
Examples of such transformations include polynomial splines or wavelets.
Since the model Q(x; τ ) ≈ Z(x)> β(τ ) holds only approximately, there
is no unique ”true” value for β(τ ). Theoretical results for such models are
often stated in terms of the following vector
(3.8)
γN (τ ) := arg minm E (Z> γ − Q(X; τ ))2 f (Q(X; τ )|X) ,
γ∈R
see Chao et al. (2017) and Remark A.6. Note that Z> γN (τ ) can be viewed
as the (weighted L2 ) projection of Q(X; τ ) onto the approximation space.
The resulting L∞ approximation error is defined as
(3.9)
cm (γN ) :=
sup
x∈X ,τ ∈T
Q(x; τ ) − γN (τ )> Z(x) .
For any v ∈ Rm define the matrix Jem (v) := E[ZZ> f (Z> v|X)] . For
−1 (b(τ ))Z|].
e b) := supτ ∈T E[|a> Jem
any a ∈ Rm , b(·) : T → Rm , define E(a,
Throughout the rest of this subsection, we assume the following condition:
(L) For each x ∈ X , the vector Z(x) has zeroes in all but at most r consecue
tive entries, where r is fixed. Furthermore, supx∈X E(Z(x),
γN ) = O(1).
Condition (L) ensures that the matrix Jem (v) has a band structure for any
−1 (v) decay exponentially fast
v ∈ Rm such that the off-diagonal entries of Jem
(Lemma 6.3 in Zhou et al. (1998)). Next, we discuss an example of Z which
satisfies (L).
Example 3.6 (Univariate polynomial spline). Suppose that (A2)-(A3)
hold and that X has a density on X = [0, 1] uniformly bounded away from
e
e1 (x), ..., B
eJ−p−1 (x))> be a polynomial spline
zero and infinity. Let B(x)
= (B
basis of degree p defined on J uniformly spaced knots 0 = t1 < ... < tJ = 1,
ej is contained in the interval [tj , tj+p+1 ) and
such that the support of each B
normalization is as given in Definition 4.19 on page 124 in Schumaker (1981).
e1 (x), ..., B
eJ−p−1 (x))> , then there exists a constant M >
Let Z(x) := m1/2 (B
DISTRIBUTED INFERENCE FOR QUANTILE REGRESSION PROCESSES 13
1 such that M −1 < E[ZZ> ] < M (by Lemma 6.2 of Zhou et al. (1998)). With
√
this scaling we have ξm m. Moreover, the first part of assumption (L)
e
holds with r = p+1, while the second part, i.e., supx∈X E(Z(x),
γN ) = O(1),
is verified in Lemma S.2.6.
Theorem 3.7. Suppose that assumptions (A1) – (A3) and (L) hold, that
4 log N = o(N ), c (γ ) = o(ξ −1 ∧ (log N )−2 ). Then
K N 2 and Sξm
m N
m
(3.10)
sup Z(x0 )> (β(τ ) − βbor (τ )) = oP (kZ(x0 )kN −1/2 )
τ ∈TK
2 (log N )2
log N 2
Sξm
c
(γ
)
+
N
m
N
S 1/2
kZ(x )kξ S log N
Sξ 2 (log N )10 1/4
kZ(x
)k
0
0
m
m
.
+
+ OP
N
N
N 1/2
+ OP
1+
If additionally K G 1 and c2m (γN ) = o(N −1/2 ) we also have
b )−βbor (τ ))| ≤ kZ(x0 )k sup kβ(τ )−βbor (τ )k+oP (kZ(x0 )kN −1/2 )
sup |Z(x0 )> (β(τ
τ ∈T
τ ∈TK
+ sup |(ΠK Q(x0 ; ·))(τ ) − Q(x0 ; τ )| + |Z(x0 )> γN (τ ) − Q(x0 ; τ )|
τ ∈T
where the projection operator ΠK was defined right after (2.5).
The strategy for proving Theorem 3.7 is similar to that for Theorem 3.1,
the difference is that we now apply an aggregated Bahadur representation
which makes explicit use of the local basis structure of Z (Section S.2.2).
Similarly to Theorem 3.1, Theorem 3.7 only provides upper bounds on
the differences between β, βb and βbor . As in the setting of fixed-dimensional
linear models, this result can be complemented by a corresponding lower
bound which we state next.
Theorem 3.8. For any τ in T there exists a sequence of distributions
of (Y, X) and sequence of transformations Z such that assumptions (A1)
4 (log N )10 = o(N ), c2 (γ ) = o(N −1/2 ) and
– (A3) and (L) hold, that Sξm
m N
there exists a C > 0 with
CSξm
(3.11)
lim sup P |Z(x0 )> β(τ ) − Z(x0 )> βbor (τ )| ≥
> 0.
N
N →∞
Moreover for any c, η > 0 there exists a sequence of distributions of (Y, X)
and sequence of transformations Z satisfying the above conditions and an
x0 ∈ X with τ 7→ Q(x0 ; τ ) ∈ Ληc (T ) such that
(3.12)
b ) − Z(x0 )> βbor (τ )| ≥ CSξm + CG−η > 0.
lim sup P sup |Z(x0 )> β(τ
N
N →∞
τ ∈T
14
Compared to Section 3.1, we obtain an interesting insight. The sufficient
and necessary conditions on S explicitly depend on m (note that under
(A1) – (A3) we have ξm & m) and become more restrictive as m increases.
This shows that there is a fundamental limitation on possible computational
savings that depends on the model complexity. In other words more complex
models (as measured by m) allow for less splitting and require larger subsamples, resulting in a larger computational burden.
We conclude this section by providing sufficient and necessary conditions for oracle rules in the local basis expansion setting. To state those
results, denote by PL (M, f¯, f 0 , fmin , R) the collection of all sequences PN
of distributions of (X, Y ) on Rd+1 and fixed Z with the following properties: (A1)-(A3) hold with constants M, f¯, f 0 < ∞, fmin > 0, (L) holds for
4 (log N )6 = o(N ), c2 (γ ) = o(N −1/2 ). The conditions in
some r < R, ξm
m N
PL (M, f¯, f 0 , fmin , R) ensure the weak convergence of the oracle estimator
βbor (τ ), see Chao et al. (2017).
The following condition characterizes the upper bound on S which is
sufficient to ensure the oracle rule for β(τ ).
(L1) Assume that
N
N 1/2
N 1/2 kZ(x0 )k
N
∧ 2
∧
∧ 2
.
S=o
2 log N
mξm
ξm (log N )10 ξm log N
ξm (log N )2
For specific examples, Condition (L1) can be simplified. For instance, in
the setting of Example 3.6, we can reduce (L1) to the form
S = o(N 1/2 (log N )−2 m−1/2 ∧ N (log N )−10 m−2 ).
We now present the sufficient and necessary conditions for the oracle rule
of β(τ ) under the Condition (L) for Z.
Corollary 3.9 (Oracle Rule for β(τ )). Assume that (L1) holds and
data are generated from PN with (PN , Z) ∈ PL (M, f¯, f 0 , fmin , R). Then the
pooled estimator β(τ ) defined in (2.3) satisfies for any fixed τ ∈ T , x0 ∈ X ,
√
N Z(x0 )> (β(τ ) − γN (τ ))
(3.13)
N
0,
τ
(1
−
τ
)
,
(Z(x0 )> Jm (τ )−1 E[ZZ> ]Jm (τ )−1 Z(x0 ))1/2
where Jm (τ ) is defined in Corollary 3.3. This matches the limiting behaviour
−1 the weak convergence result (3.13) fails for some
of βbor . If S & N 1/2 ξm
0
¯
(PN , Z) ∈ PL (1, f , f , fmin , R), x0 ∈ X .
DISTRIBUTED INFERENCE FOR QUANTILE REGRESSION PROCESSES 15
To compare the necessary and sufficient conditions in Corollary 3.9, assume for simplicity that m N α for some α > 0 and kZ(x0 )k ξm (this
is, for instance, the case for univariate splines as discussed in Example 3.6).
Since under (A1)-(A3) we have m1/2 . ξm , it follows that (L1) holds pro−1 (log N )−2 ) and the necessary and sufficient convided that S = o(N 1/2 ξm
ditions match up to a factor of (log N )2 .
Next we discuss sufficient conditions for the process oracle rule.
Corollary 3.10. Consider an arbitrary vector x0 ∈ X . Assume that
data are generated from PN with (PN , Z) ∈ PL (M, f¯, f 0 , fmin , R), that (L1)
holds, that τ 7→ Q(x0 ; τ ) ∈ Ληc (T ), rτ ≥ η, supτ ∈T |Z(x0 )> γN (τ )−Q(x0 ; τ )| =
o(kZ(x0 )kN −1/2 ), that N 2 K G N 1/(2η) kZ(x0 )k−1/η , c2m (γN ) =
o(N −1/2 ) and that the limit
(3.14)
−1 (τ )E[ZZ> ]J −1 (τ )Z(x )(τ ∧ τ − τ τ )
Z(x0 )> Jm
1
2
0
1
2
1 2
m
N →∞
kZ(x0 )k2
Hx0 (τ1 , τ2 ) := lim
exists and is non-zero for any τ1 , τ2 ∈ T , where Jm (τ ) is defined in the
statement of Corollary 3.3.
b ) defined in (2.5) satisfies
1. The projection estimator β(τ
√
N
b − Q(x0 ; ·)
(3.15)
Z(x0 )> β(·)
Gx0 (·) in `∞ (T ),
kZ(x0 )k
where Gx0 is a centered Gaussian process with E[Gx0 (τ )Gx0 (τ 0 )] = Hx0 (τ, τ 0 ).
The same holds for the oracle estimator βbor (τ ).
−1 the weak convergence in
If G . N 1/(2η) kZ(x0 )k−1/η or S & N 1/2 ξm
(3.15) fails for some (PN , Z) which satisfies the above conditions.
2. For FbY |X (·|x0 ) defined in (2.8) and G defined above we have
√
N
FbY |X (·|x0 ) − FY |X (·|x0 )
−fY |X (·|x0 )Gx0 FY |X (·|x0 )
kZ(x0 )k
in `∞ (Q(x0 ; τL ), Q(x0 ; τU )) .
This matches the process convergence for FbYor|X (·|x0 ).
The proof of the sufficient conditions in Corollary 3.10 is presented in
Section S.3.2.3, and the collapse of the weak convergence (3.15) is shown
in Section S.3.3. Similar to the discussion after Corollary 3.4, the process
oracle rule does not place additional restrictions on the number of subsamples S besides (L1). However, the process oracle rule requires additional
16
assumptions on the quantile grid TK . An interesting observation is that G
(N/kZ(x0 )k2 )1/(2η) in Theorem 3.14 can be weaker than G N 1/(2η) from
Corollary 3.4. For instance, this is true in the setting of Example 3.6 where
kZ(x0 )k m1/2 . Consequently, the computational burden is reduced since
K can be chosen smaller. The intuition behind this surprising phenomenon is
that the convergence rate for the estimator Z(x0 )> βbor (τ ) in nonparametric
models is typically slower. Thus, less stringent assumptions are needed to
ensure that the bias induced by quantile projection is negligible compared
to the convergence rate of βbor (τ ).
The sufficient conditions in this section can be extended to cover approximately linear models with increasing dimension that do not satisfy condition
(L). Technical details are provided in the Appendix, Section A.
Remark 3.11. To the best of our knowledge, the only paper that also
studies the sharpness of upper bounds for S that guarantee valid inference
in a divide and conquer setting with nonparametric regression is Shang and
Cheng (2017). However, Shang and Cheng (2017) only consider nonparametric mean regression in the smoothing spline setup.
4. Practical aspects of inference. In the previous section, we derived conditions which guarantee that the divide and conquer estimator β
and the quantile projection estimator βb share the same asymptotic distribution as the ’oracle’ estimator βbor , so that inference based on the asymptotic
distribution of βbor remains valid. In practice, this result can not be directly utilized since the asymptotic variance of the oracle estimator βbor is in
general not pivotal. Classical approaches to inference for this estimator are
typically based on estimating its asymptotic variance from the data directly,
or conducting bootstrap to approximate the asymptotic distribution.
Estimating the limiting variance requires the choice of a bandwidth parameter, and existing research indicates that classical rules for bandwidth
selection need to be adjusted in a divide and conquer setting (see e.g. Banerjee et al. (2017); Xu et al. (2016)). We discuss related issues for variance
estimation in Section 4.2.
Conducting bootstrap substantially increases the computational burden
of any procedure, which is problematic in a massive data setting we are
considering here. While recent proposals by Kleiner et al. (2014); Sengupta
et al. (2016) provide a way to reduce the computational cost of bootstrap in
a big data setting, the approaches described in those papers are not easily
combined with the divide and conquer setting which we consider here.
As an alternative to variance estimation or classical bootstrap approaches,
we propose several simple inference procedures which directly utilize the fact
DISTRIBUTED INFERENCE FOR QUANTILE REGRESSION PROCESSES 17
that in a divide and conquer setting estimators from sub-samples are available. Those procedures are very easy to implement, and require only a very
small amount of computation on the central computer without additional
communication costs. Details are provided in section 4.1.
4.1. Inference utilizing results from subsamples. We begin by discussing
inference at a fixed quantile level τ . The key idea in this section is to make
direct use of the fact that S estimators βbs (τ ) from subsamples are available.
Observe that the estimator β(τ ) is simply an average of βb1 (τ ), ..., βbS (τ )
which can be seen as iid realizations (provided groups are homogeneous) of
a random variable with approximately normal distribution. This suggests
two simple options
b D , of βb1 (τ ), ..., βbS (τ ) in order
1. Use the sample covariance matrix, say Σ
to conduct inference on β(τ ) or linear functionals thereof such as Q(x; τ ) :=
Z(x)> β(τ ). For example, a simple asymptotic level α confidence interval for
Q(x; τ ) is given by
b D Z(x0 ))1/2 Φ−1 (1 − α/2) ,
(4.1)
Z(x0 )> β(τ ) ± S −1/2 (Z(x0 )> Σ
2. A refined version of the previous approach is to additionally exploit
b
the fact that a suitably scaled version of u>
N βs (τ ) should be approximately
normal since each βbs (τ ) is itself an estimator based on sample of iid data.
Hence for small S (say S ≤ 30) more precise confidence intervals can be
obtained by using quantiles of the student t distribution (if uN is a vector) or F distribution (if uN is a fixed-dimensional matrix). For example, a
modification of the confidence interval in (4.1) would take the form
b D Z(x0 ))1/2 tS−1,1−α/2
(4.2)
Z(x0 )> β(τ ) ± S −1/2 (Z(x0 )> Σ
where tS−1,1−α/2 denotes the 1−α/2-quantile of the t-distribution with S −1
degrees of freedom. The asymptotic validity of both intervals discussed above
is provided in the following theorem.
Theorem 4.1. Assume that the conditions of either Corollary 3.3, Corollary 3.9 or Corollary A.2 hold, that cm (γN ) = o(kZ(x0 )kN −1/2 ) and that
S ≥ 2 (S can be fixed). Then the confidence interval (4.2) has asymptotic
(for N → ∞) coverage probability 1 − α.
If additionally S → ∞, the confidence interval given in (4.1) also has
asymptotic coverage probability 1 − α.
See Section S.4.1 for a proof of Theorem 4.1. The main advantage of the
two approaches discussed above lies in their simplicity as they do not require
18
any costly computation or communication between machines. There are two
main limitations.
First, for small values of S (say S ≤ 30) the confidence intervals in (4.1)
will not have the correct coverage while the interval in (4.2) can be substantially wider than the one based on the exact asymptotic distribution since
quantiles of the t-distribution with few degrees of freedom can be substantially larger than corresponding normal quantiles. Moreover, the approach is
not applicable if S is smaller than the dimension of the parameter of interest.
Second the approaches are not straightforward to generalize to inference on
non-linear functionals of β(τ ) such as FbY |X (y|x) or inference which is uniform over a continuum of quantiles τ .
The first limitation is not likely to become relevant in a big data setting
since here we typically expect that S is large due to computational bottlenecks and memory constraints. For the sake of completeness we discuss this
case in the next section. To deal with the second limitation, we propose to
use a bootstrap procedure that can be conducted by solely using the subsample estimators {βbs (τk )}s=1,..,S,k=1,...,K which are stored on the central
machine. Details are provided below.
1. Sample
iid random weights {ωs,b }s=1,...,S,b=1,...,B
from taking value 1 −
√
√
1/ 2 with probability 2/3 and 1+ 2 with probability 1/3 (i.e. weights
are chosen such that ωs,b ≥ 0, E[ωs,b ] = V ar(ωs,b ) = 1) and let ω̄·,b :=
P
S −1 Ss=1 ωs,b .
2. For b = 1, ..., B, k = 1, ..., K compute the bootstrap estimators
(4.3)
β
(b)
(τk ) :=
S
1 X ωs,b bs
β (τk )
S
ω̄·,b
s=1
b (b) from (2.7) with β (b) (τk ) replacing β(τk ).
and define the matrix Ξ
b (b) defined
3. Similarly as in (2.5) compute βb(b) (·) from the matrix Ξ
above. For a functional of interest Φ approximate quantiles of the disb
tribution of Φ(β(·))−Φ(β(·))
by the empirical quantiles of {Φ(βb(b) (·))−
b
Φ(β(·))}b=1,...,B .
A formal justification for this bootstrap approach is provided by the following result.
Theorem 4.2. Let the assumptions of either Corollary 3.4, Corollary 3.10
or Corollary A.5 hold and assume that additionally S → ∞. Then we have
conditionally on the data (Xi , Yi )i=1,...,N
√
N
b
(Z(x0 )> βb(1) (·) − Z(x0 )> β(·))
Gx0 (·) in `∞ (T ).
kZ(x0 )k
DISTRIBUTED INFERENCE FOR QUANTILE REGRESSION PROCESSES 19
where the limit Gx0 denotes the centered Gaussian process from Corollary 3.10
or Corollary A.5 under their respective assumptions and Gx0 = Z(x0 )> G
under the assumptions of Corollary 3.4.
The proof of Theorem 4.2 is given in Section S.4.2. We conclude this section by remarking that the bootstrap proposed differ from the cluster-robust
bootstrap of Hagemann (2017). The main difference is that we propose to directly utilize the sub-sample estimators βbs while the approach of Hagemann
(2017) requires repeated estimation on the complete sample.
4.2. Inference based on estimating the asymptotic covariance matrix. As
suggested by a referee, an alternative way to conduct inference is to compute,
for each sub-sample, not only the estimator βbs but also a variance estimator
and pool those estimators. Here, we provide some details on this approach.
For the sake of simplicity, we only discuss the case where m is fixed and the
model is specified correctly, i.e. the setting of Section 3.1.
It is well known that the asymptotic variance-covariance matrix of the
√
difference n(βbor (τ ) − β(τ )) (for fixed m) takes the ’sandwich’ form
Σ(τ ) = τ (1−τ )Jm (τ )−1 E[ZZ> ]Jm (τ )−1 ,
Jm (τ ) = E[ZZ> fY |X (Q(X; τ )|X)].
1 P P
>
The middle part E[ZZ> ] is easily estimated by nS
i
s Zis Zis . Since this is
1 P
>
a simple mean of the subsample-based quantities n i Zis Zis , implementing
this in a distributed computing setting is straightforward.
The matrix Jm (τ ) involves the conditional density fY |X (Q(X; τ )|X) and
is more difficult to estimate. A popular approach is based on Powell’s estimator (Powell (1986))
(4.4)
P
Jbms
(τ ) :=
n
n
o
1 X
> bs
Zis Z>
is 1 |Yis − Zis β (τ )| ≤ hn .
2nhn
i=1
Here, hn denotes a bandwidth parameter that needs to be chosen carefully
in order to balance the resulting bias and variance.
P (τ ) in a parallel comThere are several possible approaches to estimate Jm
puting environment. If an additional round of communication is acceptable,
it is possible to construct estimators with the same convergence rate and
asymptotic distribution as the estimator based on the full sample. Details
are provided in section S.1 in the appendix. If only a single round of communication is allowed, the following algorithm can be used instead.
1. For s = 1, ..., S, in the same round
as computing βbs (τ ), compute
1 P
P
b
b
Jms (τ ) from (4.4) and Σ1s := n i Zis Z>
is .
20
P (τ ), Σ
b 1s )s=1,...,S to the master ma2. Along with (βbs (τ ))s=1,...,S send (Jbms
P
PS b
P
S
P (τ ), Σ := 1
chine and compute J m (τ ) := S1 s=1 Jbms
1
s=1 Σ1s .
S
P
P
3. The final variance estimator is given by Σ̄(τ ) = τ (1−τ )J m (τ )−1 Σ1 J m (τ )−1 .
Remark 4.3. Note that in the above algorithm we first take averages
P (τ ) and only invert the aggregated maover the subsampled estimators Jbms
P
trix J m (τ ). An alternative approach would have been to compute the esP (τ )−1 Σ
P (τ )−1 for each subsample and average in the end.
b s Jbms
timator Jbms
However, given that A 7→ A−1 is non-linear, this might result in additional
bias since in general for random matrices (E[A])−1 6= E[A−1 ].
An important question for implementing the above algorithm is the choice
of the bandwidth parameter hn . To gain some intuition about the optimal
choice of hn , we will formally discuss the case of a linear model fixed dimension. First, observe that by a Taylor expansion we have almost surely
P
P
J m (τ )−1 Σ1 J m (τ )−1 − Jm (τ )−1 Σ1 Jm (τ )−1
P
= −Jm (τ )−1 (J m (τ ) − Jm (τ ))Jm (τ )−1 Σ1 Jm (τ )−1
P
− Jm (τ )−1 Σ1 Jm (τ )−1 (J m (τ ) − Jm (τ ))Jm (τ )−1
+ Jm (τ )−1 (Σ1 − Σ1 )Jm (τ )−1
P
P
+ O(kJ m (τ ) − Jm (τ )k2 + kΣ1 − Σ1 kkJ m (τ ) − Jm (τ )k).
Ignoring higher-order terms and terms that do not depend on hn it suffices
P
to analyze the properties of J m (τ ) − Jm (τ ).
Theorem 4.4. Under assumptions (A1)-(A3), assume that additionally y 7→ fY0 |X (y|x) is continuously differentiable with first derivative being
jointly continuous and uniformly bounded as a function of x, y and that
nhn (log n)−1 → ∞. Then
log n
P
J m (τ ) − Jm (τ ) = An (τ ) + Op
nhn
where the exact form of An (τ ) is given in the proof. Moreover
E[An (τ )] =
i
log n
h2n h > 00
E ZZ fY |X (Z> β(τ )|X) + O
+ o(h2n )
6
n
and for AN,(j,k) denoting the entry of AN in row j and column k
Cov(AN,(j,k) (τ ), AN,(u,v) (τ )) =
1
1
E[fY |X (Z> β(τ )|X)Zj Zk Zu Zv ]+o
.
N hn
N hn
DISTRIBUTED INFERENCE FOR QUANTILE REGRESSION PROCESSES 21
The proof of Theorem 4.4 is given in Section S.4.3. Theorem 4.4 has
several interesting implications. First, note that the asymptotic MSE of
AN (τ ) is of the order h4n + (N hn )−1 , which is minimized for hN ∼ N −1/5
(note that the term log n/n is negligible). Under the additional condition
S = o(N 2/5 (log N )−1 ) we have log n/(nhn ) = o(N −2/5 ), and in this setting
P
the MSE of the first order expansion of J m (τ ) matches that of the Powell
’oracle’ estimator as derived in Kato (2012). This shows that, despite using
P (τ ) as from the
estimators from subsamples, the same rate for estimating Jm
full sample can be achieved. That requires a stronger condition on S than
b It is not clear if the latter condition is
the oracle rate for the estimator β.
sharp, and we leave an investigation of this issue to future research.
Second, the choice hn ∼ n−1/5 which would have been optimal for estimating Jm (τ ) based on a subsample of size n does not lead to the optimal
P
error rate for the averaged estimator J m (τ ). In fact the optimal bandwidth
P
for J m (τ ) is always smaller, which corresponds to undersmoothing. Similar
effects were observed in various settings by Zhang et al. (2015), Banerjee
et al. (2017) (see their Appendix A.13) and Shang and Cheng (2017).
5. Monte Carlo experiments. In this section, we demonstrate our
theory with simulation experiments. Due to space limitations, we restrict
our attention to correctly specified linear models with different dimensions
of predictors. More precisely, we consider data generated from
(5.1)
>
Yi = 0.21 + βm−1
Xi + εi ,
i = 1, ..., N.
where εi ∼ N (0, 0.01) iid and m ∈ {4, 16, 32}. For each m, the covariate Xi
follows a multivariate uniform distribution U([0, 1]m−1 ) with Cov(Xij , Xik ) =
0.12 0.7|j−k| for j, k = 1, ..., m − 1, and the vector βm−1 takes the form
β3 = (0.21, −0.89, 0.38)> ;
(5.2)
β15 = (β3> , 0.63, 0.11, 1.01, −1.79, −1.39, 0.52, −1.62,
1.26, −0.72, 0.43, −0.41, −0.02)> ;
>
> >
β31 = (β15
, 0.21, β15
) .
Throughout this section, we fix T = [0.05, 0.95]. Section 5.1 contains
results for the estimator β(τ ), while results for FbY |X (y|x) are collected in
Section 5.2. Additional simulations (including models with heteroskedastic
errors) are presented in Section S.9 of the online supplement.
5.1. Results for the divide and conquer estimator β(τ ). We fix the subsample size n and consider the impact of the number of sub-samples S on the
22
coverage probabilities of different 95% confidence intervals. To benchmark
our results, we use the infeasible asymptotic confidence interval
>
(5.3)
x0 β(τ ) ± N −1/2 σ(τ )Φ−1 (1 − α/2) ,
where σ 2 (τ ) denotes the theoretical asymptotic variance of β or (τ ); this CI
is valid by the oracle rule but contains unknown quantities. The coverage
properties of this interval also indicate whether we are in a regime where
the oracle rule holds.
In a first step, we consider the properties of confidence intervals discussed
in Section 4.1 which directly utilize the availability of results from subsamples. We consider the following three types of confidence intervals.
1. The normal confidence interval (4.1).
2. The confidence interval (4.2) based on quantiles of the t distribution.
3. The bootstrap confidence interval based on sample quantiles of (β̄ (1) (τ )−
β̄(τ )), ..., (β̄ (B) (τ ) − β̄(τ )), where β̄ (1) (τ ), ..., β̄ (B) (τ ) are computed as
in (4.3); in this section we set B = 500.
The coverage probabilities of corresponding confidence intervals are summarized in Figure 5.1 where we fix two sub-sample sizes n = 512, 2048 and
present two types of plots: coverage probabilities for ’small’ values of S ranging from S = 2 to S = 50 and coverage probabilities for ’large’ values of
S = 2k , k = 1, ..., 10. Only m = 4, 32 is considered here, m = 16 can be
found in the online supplement.
Plots for small S help to access how different procedures behave if the
number of sub-samples is small. As expected from the theory, oracle confidence intervals and simple confidence intervals based on the t distribution
maintain nominal coverage for small values of S. Confidence intervals based
on normal approximation and the bootstrap under-cover for smaller values
of S. Starting from about S = 20, coverage probabilities of all four types
of intervals are very close. We also see that for m = 32, the oracle rule
does not apply for a sub-sample size of n = 512 with any number of subsamples S > 10; the situation improves for n = 2048. This is in line with our
asymptotic oracle theory. For ’larger’ values of S, there is no difference in
the coverage probabilities of different intervals. As predicted by the oracle
theory, coverage starts to drop earlier for models with larger m.
Next, we analyze the properties of asymptotic confidence intervals which
are based on estimating the asymptotic variance of x>
0 β(τ ) from data. We
compare three different ways of estimating the asymptotic variance
1. A simple pooled estimator which uses the default settings in the package quantreg to obtain estimated variances in each sub-sample and
DISTRIBUTED INFERENCE FOR QUANTILE REGRESSION PROCESSES 23
takes the average over all sub-samples (additional details are provided
in the online Supplement, Section S.9.1).
2. The estimator Σ̄ based on the bandwidth c∗ (τ )n−1/5 which minimizes
the asymptotic MSE of the estimator Jm within sub-samples of size n
(additional details on the choice of optimal constant c∗ (τ ) are provided
in the online Supplement, Section S.9.1).
3. The estimator Σ̄ based on the bandwidth c∗ (τ )N −1/5 which is motivated by the theory developed in Theorem 4.4.
The results are reported in Table 5.1. Since there is no notable difference between all approaches when S is large, only results for S ≤ 50 are
displayed for the sake of brevity. Interestingly, we do not observe a big difference between the naive bandwidth choice hn ∼ n−1/5 and the optimal
undersmoothing choice hn ∼ N −1/5 . This finding is quite intuitive since,
once the asymptotic variance is estimated with at most 5 − 10% relative
error, a further increase in estimation accuracy does not lead to substantial
improvements in coverage probabilities. The completely automatic choice
implemented in the quantreg package also performs reasonably well.
Finally, note that the pattern of coverage probabilities varies at different
τ . For example, in the linear models with normal errors, the coverage probabilities at tail quantiles (τ = 0.1, 0.9) drop to zero much faster than those at
τ = 0.5. These empirical observations are not inconsistent with our theory
where only the orders of the upper bound for S are shown to be the same
irrespective of the value of τ . Rather, this phenomenon might be explained
by different derivatives of the error density that appear in the estimation
bias, and is left for future study.
5.2. Results for the estimator FbY |X (y|x). In this section, we consider
inference on FY |X (y|x). We compare the coverage probability of the oracle
asymptotic confidence interval
(5.4)
FbY |X (Q(x0 ; τ )|x0 ) ± N −1/2 σF2 (τ )Φ−1 (1 − α/2) .
(here σF2 (τ ) is the asymptotic variance of the oracle estimator) and the
bootstrap confidence interval described above Theorem 4.2. Note that the
other approaches described in Section 4 are not directly applicable here since
b
FbY |X (y|x) is a functional of the whole process β(·).
Since we focus on bootstrap reliability, the number of quantile levels K = 65 and knots for spline
interpolation G = 32 are chosen sufficiently large to ensure nominal coverage
of oracle intervals. A detailed study of the impact of K, G on coverage of oracle intervals is provided in Section S.9.3.2 of the online supplement. Due to
space limitations, we only show the results for small values of S, results for
24
S
or
def
nai
adj
or
def
nai
adj
or
def
nai
adj
or
def
nai
adj
or
def
nai
adj
or
def
nai
adj
1
10
30
50
n = 512, m = 4, τ = 0.1
94.9
94.8
94.9
94.3
92.6
93.7
93.6
92.8
94.2
93.9
93.6
93.2
94.2
94.5
94.3
94.2
n = 512, m = 4, τ = 0.5
94.7
96
95.9
95
97.8
98.2
98.2
98.2
95.9
96.8
96.4
96.3
95.9
96.4
96
95.2
n = 512, m = 4, τ = 0.9
95.4
94.6
94.2
93.6
94
93.6
92.8
92.4
94.8
93.8
93.1
92.6
94.8
94.3
93.8
93.7
1
10
30
50
n = 512, m = 16, τ = 0.1
94.7
94.3
92.2
90.7
92
92.3
90.1
88.4
96.6
93.3
91
89.4
96.6
94.3
92.1
90.8
n = 512, m = 16, τ = 0.5
94.2
95.7
94.8
95.4
96.7
98.2
98
98
96.7
97
96.4
97
96.7
96.4
95.7
96.2
n = 512, m = 16, τ = 0.9
94.6
94
91.7
90.1
92.6
92.2
90
88.1
96.6
93.2
90.8
88.6
96.6
94
91.9
90.4
1
10
30
50
n = 512, m = 32, τ = 0.1
94.4
92.5
88.6
85.5
92.5
90.6
86.2
82.9
98.9
92.7
88.4
85
98.9
94.4
89.5
87.2
n = 512, m = 32, τ = 0.5
95.4
95.5
94.8
95
97.6
97.5
97.3
97.7
99
97
96.2
96.7
99
96.6
95.7
95.9
n = 512, m = 32, τ = 0.9
95.2
92.1
90.6
87.1
92.2
90.2
88.2
84.3
99
92.5
90.1
86.6
99
93.8
91.4
88.4
n = 2048, m = 4, τ = 0.1
95.6
94.7
94.4
94.3
94.7
94.1
93.8
94
95.1
94.1
93.7
94
95.1
94.6
94.4
94.5
n = 2048, m = 4, τ = 0.5
94.6
94.9
95.2
95.3
96
96.2
96.7
96.2
95.1
95.4
95.5
95.6
95.1
95.1
95.1
95.3
n = 2048, m = 4, τ = 0.9
94.6
95.6
95.8
94.6
94.2
94.9
94.9
94.1
94.4
94.9
94.8
94.1
94.4
95.3
95.6
94.6
n = 2048, m = 16, τ = 0.1
95
94.2
94.4
94.4
93.8
93.2
93.5
93.7
95.3
93.5
93.5
93.7
95.3
94
94.1
94.4
n = 2048, m = 16, τ = 0.5
95.3
94.2
94.2
95.5
96.7
96.2
95.4
96.7
96.8
95.2
95
96
96.8
95
94.7
95.6
n = 2048, m = 16, τ = 0.9
95.1
95
94.5
94.7
94.3
94.5
94
94.3
95.5
94.7
94.2
94.4
95.5
95.1
94.6
95
n = 2048, m = 32, τ = 0.1
95
95
94
94
93.6
94.2
92.7
93.1
95.7
94.6
93.1
93.2
95.7
94.9
93.8
93.8
n = 2048, m = 32, τ = 0.5
95.3
95.5
95.1
95.5
96
96.4
96.3
96.6
96.5
95.8
95.7
95.9
96.5
95.6
95.2
95.5
n = 2048, m = 32, τ = 0.9
95.2
94.4
92.9
92.9
94.4
93.8
92.6
92.2
96.7
94.4
92.7
92.5
96.7
94.9
93.2
93.3
Table 5.1
Coverage probabilities based on estimating the asymptotic variance. Different rows
correspond to different methods for obtaining covariance matrix. or: using true
asymptotic variance matrix, def: default choice implemented in quantreg package, nai:
asymptotically optimal constant with scaling hn ∼ n−1/5 , adj: asymptotically optimal
constant with scaling hn ∼ N −1/5 as suggested by Theorem 4.4.
large values of S do not give crucial additional insights and are deferred to
the online supplement. Coverage probabilities for m = 4, 32 are reported in
Figure 5.2. For m = 4, the bootstrap and oracle confidence interval show a
very similar performance as soon as S ≥ 20; this is in line with the coverage
properties for the bootstrap for β̄. For m = 32, coverage probabilities of the
oracle confidence interval indicate that the sub-sample size n is too small
and the oracle rule does not apply, even for S = 2. Interestingly, coverage
probabilities of bootstrap and asymptotic confidence intervals differ in this
setting. This does not contradict our theory for the bootstrap since that
was only developed under the assumption that we are in a regime where the
oracle rule holds.
DISTRIBUTED INFERENCE FOR QUANTILE REGRESSION PROCESSES 25
(a) m = 4
30
40
50
1.0
0.8
10
20
30
40
50
0
10
20
30
τ = 0.1
τ = 0.5
τ = 0.9
8
50
8
10
40
50
8
10
0.8
0.6
0.0
0.2
0.4
coverage
0.8
0.6
0.0
0.2
0.4
coverage
0.8
0.6
0.4
6
log2(S)
40
1.0
S
1.0
S
0.2
4
0.6
coverage
0.2
0.0
0
S
0.0
2
0.4
0.6
0.0
0.2
coverage
0.8
20
1.0
10
0.4
0.8
0.6
coverage
0.0
0.2
0.4
oracle asymp. CI n=512
oracle asymp. CI n=2048
CI from (4.2) t quant. n=512
CI from (4.2) t quant. n=2048
CI from (4.2) norm quant. n=512
CI from (4.2) norm quant. n=2048
bootstrap n=512, B=500
bootstrap n=2048, B=500
0
coverage
τ = 0.9
1.0
τ = 0.5
1.0
τ = 0.1
10
2
4
6
log2(S)
8
10
2
4
6
log2(S)
(b) m = 32
30
40
50
1.0
0.8
10
20
30
40
50
0
10
20
30
τ = 0.1
τ = 0.5
τ = 0.9
6
log2(S)
8
10
0.8
0.6
0.0
0.2
0.4
coverage
0.8
0.6
0.4
0.0
0.2
0.4
0.6
coverage
0.8
1.0
S
1.0
S
0.2
4
0.6
coverage
0.2
0.0
0
S
0.0
2
0.4
0.6
0.0
0.2
coverage
0.8
20
1.0
10
0.4
0.8
0.6
coverage
0.0
0.2
0.4
oracle asymp. CI n=512
oracle asymp. CI n=2048
CI from (4.2) t quant. n=512
CI from (4.2) t quant. n=2048
CI from (4.2) norm quant. n=512
CI from (4.2) norm quant. n=2048
bootstrap n=512, B=500
bootstrap n=2048, B=500
0
coverage
τ = 0.9
1.0
τ = 0.5
1.0
τ = 0.1
2
4
6
log2(S)
8
10
2
4
6
log2(S)
Fig 5.1: Coverage probabilities for x>
0 β(τ ) for different values of S and
τ = 0.1, 0.5, 0.9 (left, middle, right row). Solid lines: n = 512, dashed lines:
n = 2048. Black: asymptotic oracle CI, blue: CI from (4.2) based on t distribution, red: CI from (4.1) based on normal distribution, green: bootstrap
CI. Throughout x0 = (1, ..., 1)/m1/2 , nominal coverage 0.95.
26
Summarizing all results obtained so far, we can conclude that inference
based on results from sub-samples is reliable for S ≥ 20. Since this does
not require additional computation, we recommend using the normal approximation for S > 20 for point-wise inference and the bootstrap if process
level results are required. For S < 20, estimating the asymptotic variance
within sub-samples and aggregating is recommendable. The simplest approach which is based on averaging variance estimators from the quanteg
package works well and does not require additional implementation, so we
recommend to use this for S < 20.
30
40
1.0
0.8
0
10
20
30
40
50
0
20
30
50
50
40
50
0.8
0.6
0.4
0.0
0.2
0.4
0.6
coverage
0.8
40
0.2
coverage
30
S
0.0
0.8
0.6
0.4
0.2
20
40
1.0
S
τ = 0.9 , m=32
1.0
S
τ = 0.5 , m=32
0.0
10
10
S
oracle asymp. CI n=512
bootstrap n=512, B=500
0
0.6
0.2
0.0
0.0
50
0.4
0.6
coverage
0.8
20
τ = 0.1 , m=32
1.0
10
0.4
0.2
coverage
0.8
0.6
0.4
coverage
0.2
0.0
oracle asymp. CI n=512
bootstrap n=512, B=500
0
coverage
τ = 0.9 , m=4
1.0
τ = 0.5 , m=4
1.0
τ = 0.1 , m=4
0
10
20
30
S
40
50
0
10
20
30
S
Fig 5.2: Coverage probabilities for oracle confidence intervals (red) and bootstrap confidence intervals (black) for FY |X (y|x0 ) for x0 = (1, ..., 1)/m1/2 and
y = Q(x0 ; τ ), τ = 0.1, 0.5, 0.9. n = 512 and nominal coverage 0.95.
References.
Banerjee, M., Durot, C., and Sen, B. (2017). Divide and conquer in non-standard problems
and the super-efficiency phenomenon. The Annals of Statistics, page to appear.
Belloni, A., Chernozhukov, V., Chetverikov, D., and Fernández-Val, I. (2017). Conditional
quantile processes based on series or many regressors. arXiv preprint arXiv:1105.6154.
Billingsley, P. (1995). Probability and Measure. Wiley.
Bücher, A. and Kojadinovic, I. (2017). A note on conditional versus joint unconditional
weak convergence in bootstrap consistency results. arXiv preprint arXiv:1706.01031.
Chao, S., Volgushev, S., and Cheng, G. (2017). Quantile process for semi and nonparametric regression. Electronic Journal of Statistics, 11(2):3272–3331.
Chernozhukov, V., Fernández-Val, I., and Galichon, A. (2010). Quantile and probability
curves without crossing. Econometrica, 78(3):1093–1125.
DISTRIBUTED INFERENCE FOR QUANTILE REGRESSION PROCESSES 27
Hagemann, A. (2017). Cluster-robust bootstrap inference in quantile regression models.
Journal of the American Statistical Association, 112(517):446–456.
Huang, J. Z. (2003). Local asymptotics for polynomial spline regression. Annals of Statistics, 31(5):1600–1635.
Jordan, M. I. (2013). On statistics, computation and scalability. Bernoulli, 19(4):1378–
1390.
Kato, K. (2012). Asymptotic normality of Powell’s kernel estimator. Annals of the Institute
of Statistical Mathematics, 64(2):255–273.
Kleiner, A., Talwalkar, A., Sarkar, P., and Jordan, M. I. (2014). A scalable bootstrap for
massive data. Journal of the Royal Statistical Society: Series B (Statistical Methodology), 76(4):795–816.
Kley, T., Volgushev, S., Dette, H., and Hallin, M. (2016). Quantile spectral processes:
Asymptotic analysis and inference. Bernoulli, 22(3):1770–1807.
Koenker, R. (2005). Quantile Regression. Cambridge University Press, New York.
Koenker, R. (2016). quantreg: Quantile Regression. R package version 5.29.
Koltchinskii, V. (2006). Local Rademacher complexities and oracle inequalities in risk
minimization. Annals of Statistics, 34(6):2593–2656.
Ledoux, M. and Talagrand, M. (1991). Probability in banach spaces, volume 23 of ergebnisse der mathematik und ihrer grenzgebiete (3)[results in mathematics and related
areas (3)].
Li, R., Lin, D. K. J., and Li, B. (2012). Statistical inference in massive data sets. Applied
Stochastic Models in Business and Industry, 29:399–409.
Massart, P. (2000). About the constants in Talagrand’s concentration inequalities for
empirical processes. Annals of Probability, pages 863–884.
McDiarmid, C. (1989). On the method of bounded differences. In Siemons, J., editor,
Surveys in Combinatorics, London Mathematical Society Lecture Note Series 141, pages
148–188. Cambridge University Press.
Powell, J. L. (1986). Censored regression quantiles. Journal of econometrics, 32(1):143–
155.
Schumaker, L. (1981). Spline Functions: Basic Theory. Wiley, New York.
Sengupta, S., Volgushev, S., and Shao, X. (2016). A subsampled double bootstrap for
massive data. Journal of the American Statistical Association, 111(515):1222–1232.
Shang, Z. and Cheng, G. (2017). Computational limits of a distributed algorithm for
smoothing spline. Journal of Machine Learning Research, 18(108):1–37.
Shi, C., Lu, W., and Song, R. (2017). A massive data framework for M-estimators with
cubic-rate. Journal of the American Statistical Association, page to appear.
Volgushev, S. (2013). Smoothed quantile regression processes for binary response models.
arXiv preprint arXiv:1302.5644.
White, T. (2012). Hadoop: The Definitive Guide. O’Reilly Media / Yahoo Press.
Xu, G., Shang, Z., and Cheng, G. (2016). Optimal tuning for divide-and-conquer kernel
ridge regression with massive data. arXiv preprint arXiv:1612.05907.
Zhang, Y., Duchi, J., and Wainwright, M. J. (2013). Communication-efficient algorithms
for statistical optimization. Journal of Machine Learning Research, 14:3321–3363.
Zhang, Y., Duchi, J., and Wainwright, M. J. (2015). Divide and conquer kernel ridge
regression: A distributed algorithm with minimax optimal rates. Journal of Machine
Learning Research, 16:3299–3340.
Zhao, T., Cheng, G., and Liu, H. (2016). A partially linear framework for massive heterogeneous data. Annals of Statistics, 44(4):1400–1437.
Zhou, S., Shen, X., and Wolfe, D. (1998). Local asymptotics for regression splines and
confidence regions. Annals of Statistics, 26(5):1760–1782.
28
APPENDIX A: APPROXIMATE LINEAR MODELS WITHOUT
LOCAL BASIS STRUCTURE
In this section, we consider models with transformation Z of increasing
dimension that do not have the special local structure considered in Section 3.2. The price for this generality is that we need to assume a more
stringent upper bound for S and put additional growth restrictions on m in
order to prove the oracle rule. The conditions on K remain the same.
Theorem A.1. Assume that conditions (A1)-(A3) hold and that addi2 log N = o(N/S), c (γ ) = o(ξ −1 ), K N 2 . Then
tionally mξm
m N
m
(A.1)
sup kβ(τ ) − βbor (τ )k
τ ∈TK
mc2 log N 1/2
Smξ 2 log N
log N
m
m
4
1 + 1/2
+ c2m ξm +
+ c4m ξm
N
N
S
1/2 Sm3 ξ 2 (log N )7 1/4
1
m
2 2
cm (log N )3
+
+ oP (N −1/2 ).
+ 1/2 OP mξm
N
N
= OP
2 (log N )3 = o(N ), c2 (γ )ξ = o(N −1/2 )
If additionally K G 1, m3 ξm
m N m
we also have for any x0 ∈ X
(A.2)
b ) − Z(x0 )> βbor (τ )|
sup |Z(x0 )> β(τ
τ ∈T
≤ kZ(x0 )k sup kβ(τ ) − βbor (τ )k + sup |(ΠK Q(x0 ; ·))(τ ) − Q(x0 ; τ )|
τ ∈TK
τ ∈T
>
+ sup |Z(x0 ) βN (τ ) − Q(x0 ; τ )| + oP (kZ(x0 )kN −1/2 )
τ ∈T
where the projection operator ΠK was defined right after (2.5).
The proof of Theorem A.1 is given in Section S.3.4.2. The general upper
bound provided in (A.1) takes a rather complicated form. Under additional
assumptions on cm , major simplifications are possible. Due to space considerations we do not provide additional details here, rather we implicitly carry
out the simplifications when proving the following result.
Corollary A.2 (Oracle Rule for β(τ )). Assume that conditions (A1)(A3) hold and that additionally m4 (log N )10 = o(N ), c2m (γN )ξm = o(N −1/2 ).
2 (log N )2 )) the estimator β(τ )
Provided that additionally S = o(N 1/2 /(mξm
defined in (2.3) satisfies
√
N u>
N (β(τ ) − γN (τ ))
(A.3)
N 0, τ (1 − τ ) ,
−1
−1
>
1/2
>
(uN Jm (τ )E[ZZ ]Jm (τ )uN )
DISTRIBUTED INFERENCE FOR QUANTILE REGRESSION PROCESSES 29
for any τ ∈ T , uN ∈ Rm , where Jm (τ ) is defined in the statement of Corollary 3.3. The same holds for the oracle estimator βbor (τ ).
We note that m4 (log N )10 = o(N ) imposes restriction on model complexity. In particular, this requires m = o(N 1/4 (log N )−5/2 ). An immediate
consequence of Corollary A.2 is the oracle rule for correctly specified models,
including linear quantile regression with increasing dimension as a special
case. In this case cm (γN ) = 0.
Corollary A.3 (Oracle rule for correctly specified models). Assume
that conditions (A1)-(A3) hold and that the quantile function satisfies Q(x; τ ) =
Z(x)> γN (τ ) with a transformation Z(x) of possibly increasing dimension m
with each entry bounded almost surely. Then u>
N β(τ ) satisfies the oracle rule
4
10
provided that m (log N ) = o(N ) and S = o(N 1/2 m−2 (log N )−2 ).
This corollary reduces to Corollary 3.3 in Section 3.1 when m is fixed. It
describes the effect of allowing m to increase on the sufficient upper bound
for S. We note that cm (γN ) = 0 and ξm m1/2 under the settings of
Corollary A.3, whose proof follows directly from Corollary A.2.
Both Corollary 3.9 and Corollary A.2 can be applied to local polynomial
spline models, but Corollary 3.9 puts less assumptions on S and m than
Corollary A.2, because Corollary 3.9 exploits the local support property of
splines. This is illustrated in the following Remark A.4 for the specific setting
of Example 3.6.
Remark A.4 (Comparing Corollary 3.9 and Corollary A.2 with univariate splines). Let Z denote the univariate splines from Example 3.6 and
let uN := Z(x0 ) for a fixed x0 ∈ X . We assume that (A2) and (A3)
hold and that X has a density on X = [0, 1] that is uniformly bounded
away from zero and infinity. We will verify in Section S.3.4.4 that (A1)
holds with ξm m1/2 . For simplicity, assume that the bias cm (γN ) satisfies
ξm cm (γN )2 = o(N −1/2 ). Corollary 3.9 shows that sufficient conditions for
the oracle rule are m2 (log N )6 = o(N ) and S = o((N m−1 (log N )−4 )1/2 ∧
(N m−2 (log N )−10 )). On the other hand, Corollary A.2 requires the more
restrictive conditions m4 (log N )10 = o(N ) and S = o(N 1/2 m−2 (log N )−2 ).
Remark A.4 indicates that Corollary 3.9 gives sharper bounds than Corollary A.2 when both results are applicable. Note however that Corollary A.2
applies to more general settings since, in contrast to Corollary 3.9, it does
not require Condition (L). For instance, linear models as discussed in Corollary A.3 can in general not be handled by Corollary 3.9. Finally, we discuss
sufficient conditions for the process oracle rule.
30
b )). Let x0 ∈ X , let the conditions
Corollary A.5 (Oracle Rule for β(τ
of Corollary A.2 hold. Suppose that τ 7→ Q(x0 ; τ ) ∈ Ληc (T ), rτ ≥ η, N 2
K G N 1/(2η) kZ(x0 )k−1/η , and supτ ∈T |Z(x0 )> γN (τ ) − Q(x0 ; τ )| =
o(kZ(x0 )kN −1/2 ). Let the limit Hx0 (τ1 , τ2 ) defined in (3.14) exist and be
non-zero for any τ1 , τ2 ∈ T
b ) defined in (2.5) satisfies
1. The projection estimator β(τ
√
N
b − Q(x0 ; ·)
(A.4)
Z(x0 )> β(·)
Gx0 (·) in `∞ (T ),
kZ(x0 )k
where Gx0 is a centered Gaussian process with E[Gx0 (τ )Gx0 (τ 0 )] =
Hx0 (τ, τ 0 ). This matches the process convergence of Z(x0 )> βbor (τ ).
2. The estimator FbY |X (·|x0 ) defined in (2.8) satisfies,
√
N
FbY |X (·|x0 ) − FY |X (·|x0 )
−fY |X (·|x0 )Gx0 FY |X (·|x0 )
kZ(x0 )k
in `∞ (Q(x0 ; τL ), Q(x0 ; τU )) ,
where and Gx0 is the centered Gaussian process from (A.4). The same
is true for FbYor|X .
The proof of Corollary A.5 is given in Section S.3.4.5. Note that the
condition on K is the same as in Corollary 3.10. Results along the lines
b
of Corollary A.5 can be obtained for any estimator of the form u>
N β(·), as
long as uN satisfies certain technical conditions. For example, the partial
derivative, the average derivative and the conditional average derivative of
Q(x; τ ) in x fall into this framework. For brevity, we refer the interested
reader to Section 2.3 of Belloni et al. (2017) for examples of vectors uN and
omit the technical details.
Remark A.6. The weak convergence in (A.3) was derived with the centering γN for notational convenience. As pointed out in Chao et al. (2017),
other choices of centering sequences can be advantageous (for instance, this
is the case in the setting of partial linear models, see also Belloni et al. (2017)
for an alternative centering sequence). The conclusions of Theorem A.2 can
be generalized by replacing γN in there by other vectors βN (τ ) as long as
certain technical conditions are satisfied. However, this requires additional
notation and technicalities. To keep the presentation simple we provide those
details in Section S.3.4.1 in the appendix. Similarly, the results in Section 3.2
could be stated in terms of more general sequences βN instead of the one
considered in (3.8) at the cost of additional notation and conditions.
DISTRIBUTED INFERENCE FOR QUANTILE REGRESSION PROCESSES
1
SUPPLEMENTARY MATERIAL
ONLINE SUPPLEMENTARY MATERIAL: Distributed inference for quantile regression processes
(doi: COMPLETED BY THE TYPESETTER; .pdf). The supplementary
materials contain all the proofs, additional technical details, and additional
simulation results.
Additional Notation. This notation will be used throughout the supplement. Let S m−1 := {u ∈ Rm : kuk = 1} denote the unit sphere in Rm . For
a set I ⊂ {1, ..., m}, define
(S.0.5)
>
m
Rm
I := {u = (u1 , ..., um ) ∈ R : uj 6= 0 if and only if j ∈ I}
(S.0.6) SIm−1 := {u = (u1 , ..., um )> ∈ S m−1 : uj 6= 0 if and only if j ∈ I}
For simplicity, we sometimes write supτ (inf τ ) and supx (inf x ) instead of
supτ ∈T (inf τ ∈T ) and supx∈X (inf x∈X ) throughout the Appendix. Ik is the
k-dimensional identity matrix for k ∈ N.
P (τ )
APPENDIX S.1: AN ALTERNATIVE ESTIMATOR FOR JM
Here, we provide a simple way to estimate the matrix Jm (τ ) in a divide
and conquer setup provided that an additional round of communication is
acceptable.
1. Compute β(τ ) on the master computer and send to each distributed
unit.
2. On each machine, compute
P
Jems
(τ ) :=
n
n
o
1 X
>
Zis Z>
is 1 |Yis − Zis β(τ )| ≤ hn
2nhn
i=1
and send results back to master machine.
PS eP
P (τ ) := 1
3. On the master, compute Jem
s=1 Jms (τ ). The final variance
S
P
−1
P (τ )−1 .
estimator is given by τ (1 − τ )Jem (τ ) Σ1 Jem
Provided that β(τ ) satisfies the oracle rule, the theory developed in Kato
(2012) for Powell’s estimator
applies (note that the results developed in Kato
√
(2012) only require N consistency of the estimator β, which is guaranteed
by the oracle rule). In particular, following Kato (2012), the optimal rate for
the bandwidth hn is given by hN ∼ N −1/5 , and an explicit formula for the
optimal bandwidth can be found on page 263 of Kato (2012). In practice,
the ’rule of thumb’ provided on page 264 in Kato (2012) can be used to
determine the bandwidth hn .
2
The algorithm described above provides a simple way to estimate the
asymptotic covariance matrix of β(τ ) at oracle rate N −2/5 under the same
assumptions that guarantee β to satisfy the oracle rule. The additional communication and computational costs are also fairly low since only the vectors
P (τ ) need to be communicated and computation of
β and the matrices Jems
P (τ ) on each sub-sample does not involve any optimization.
Jems
APPENDIX S.2: AGGREGATED BAHADUR REPRESENTATIONS
S.2.1. Aggregated Bahadur representation for general series approximation model. Note that Assumptions (A1)-(A3) imply that for
any sequence of Rm -valued functions βN (τ ) satisfying supτ ∈T supx |βN (τ )> Z(x)−
Q(x; τ )| = o(1), the smallest eigenvalues of the matrices
(S.2.1)
Jem (βN (τ )) := E[ZZ> fY |X (βN (τ )> Z|X)],
Jm (τ ) := E[ZZ> fY |X (Q(X; τ )|X)]
are bounded away from zero uniformly in τ for sufficiently large n. Define
for any vector of functions bN : T → Rm ,
(S.2.2)
gN (bN ) := sup E[Zi (1{Yi ≤ Z>
i bN (τ )} − τ )] .
τ ∈T
Theorem S.2.1 (ABR for General Series Model). Assume that the un2 log n =
derlying distribution of (Xi , Yi ) satisfies conditions (A1)-(A3) and that mξm
−1
o(n). Consider any sequence βN (τ ) such that gN := gN (βN ) = o(ξm ), cm :=
cm (βN ) = o(1). Then we have
N
β(τ )−βN (τ ) = −
X
1
(1)
(2)
Jm (τ )−1
Zi (1{Yi ≤ Z>
i βN (τ )}−τ )+rN (τ )+rN (τ )
N
i=1
2 , any sufficiently large n, and a constant C indewhere, for any κn n/ξm
pendent of n
(1)
(S.2.3)
P sup krN (τ )k ≥ CR1,n (κn ) ≤ Se−κn
τ ∈T
with
(S.2.4)
R1,n (κn ) := ξm
m
n
log n
1/2 κ 1/2 2
m
1/2 κ 1/2
n
n
+
+
+cm
log n
+gN .
n
n
n
DISTRIBUTED INFERENCE FOR QUANTILE REGRESSION PROCESSES
3
Moreover there exists a constant C independent of n, S such that for n sufficiently large and A ≥ 1
2
(2)
(S.2.5)
sup P rN (τ ) > CR2,n (A, κn ) ≤ 2Sn−A + 2e−κn
τ ∈T
where
(S.2.6)
R2,n (A, κn ) :=
mξ 2 (log n)3 1/4
Aκn m 1/2
m
1/2
(ξ
g
log
n)
+
m N
n
S 1/2 n
2 log n
mξm
κn
2 2
ξm +
+ 1 + 1/2 gN
.
n
S
See Section S.5.1 of the online Appendix for a proof for Theorem S.2.1. We
give a simpler expression for the remainder terms in the following remark.
Remark S.2.2. For subsequent applications of the Bahadur representation Theorem S.2.1, we will often set κn = A log n for a sufficiently large
(1)
(2)
constant A > 0, and assume the remainder terms rN (τ ) + rN (τ ), defined
in Theorem S.2.1, are bounded with certain order.
Under the conditions κn = A log n, (log n)2 S −1 = o(1) and m ≥ 1,
R1,n (κn ) + R2,n (κn ) can be bounded by (neglecting a positive constant)
(S.2.7)
1/2
mξ 2 (log n)3 1/4
m
log n m 1/2
m
log n
(ξm gN log n)1/2 +
cm
+gN + 1/2
n
n
n
S
2
mξ
m log n
2 2
+ gN
ξm +
.
n
We note that without the condition (log n)2 S −1 = o(1), we have an additional log n factor to the last two terms in (S.2.7).
Remark S.2.3. A result along the lines of Theorem S.2.1 can also be
obtained by a direct application of the classical Bahadur representation (e.g.
Theorem 5.1 in Chao et al. (2017)), but that leads to bounds on remainder
terms that are less sharp compared to the ones provided in (S.2.3)-(S.2.6).
To illustrate this point, assume that Q(x; τ ) = x> β(τ ) with x denoting a
covariate of fixed dimension. In this case gN (βN ) = cm (βN ) = 0 and ξm , m
are constant. The classical Bahadur representation in Theorem 5.1 of Chao
et al. (2017) yields an expansion of the form
1
βbs (τ ) − βN (τ ) = − Jm (τ )−1
n
n
X
i=1
Xis (1{Yis ≤ Q(Xis ; τ )} − τ ) + rns
4
where Jm (τ ) := E[XX > fY |X (Q(X; τ )|X)], rns = OP (n−3/4 log n) and the
exponent of n in the bound on rns can in general not be improved. Applying
this expansion to the estimator in each group we obtain
(S.2.8)
β(τ ) − βN (τ ) = −
S
N
X
1
1X s
Jm (τ )−1
rn .
Xi (1{Yi ≤ Q(Xi ; τ )} − τ ) +
N
S
s=1
i=1
rns ,
Without additional information on the moments of
the bound
rns =
P
S
OP (n−3/4 log n) would at best provide a bound of the form S1 s=1 rns =
OP (n−3/4 log n). Thus by the representation (S.2.8) the oracle rule holds provided that n−3/4 log n = o(N −1/2 ), which is equivalent to S = o(N 1/3 (log N )−4/3 )
for an arbitrarily small ε > 0. In contrast, for cm (βN ) = gN (βN ) = 0 the
bound in (S.2.7) reduces to n−1 log n + n−3/4 S −1/2 (log n)7/4 . Since N = Sn
this is of order o(N −1/2 ) if S = o(N 1/2 (log N )−2 ). This is a considerably
less restrictive condition than S = o(N 1/3 (log N )−4/3 ) which was derived
by the ’direct’ approach above. Similar arguments apply in a more general
setting of a linear model with increasing dimension or series expansions.
S.2.2. Aggregated Bahadur representation for local basis functions.
Theorem S.2.4 (ABR for local basis functions). Assume that the underlying distribution satisfies conditions (A1)-(A3) and let Condition (L) hold.
2 log n = o(n), c (γ ) = o(1) and that c (γ )2 = o(ξ −1 ).
Assume that mξm
m N
m N
m
Then we have for uN ∈ SIm−1 (defined in (S.0.6)) with I consisting of
L = O(1) consecutive integers (I is allowed to depend on n)
(S.2.9)
N
>
u>
N β(τ ) − uN γN (τ ) = −
X
1 >e
uN Jm (γN (τ ))−1
Zi (1{Yi ≤ Z>
i γN (τ )} − τ )
N
i=1
(1)
(2)
+ rN (τ, uN ) + rN (τ, uN )
2 , any sufficiently large n, and a constant C indewhere, for any κn n/ξm
pendent of n
(1)
(L)
(S.2.10)
P sup sup krN (τ, uN )k ≥ CR1,n (κn ) ≤ Se−κn .
τ ∈T uN ∈S m−1
I
where
(S.2.11)
(L)
R1,n (κn )
2 (log n)2 + ξ 2 κ
ξm
ξm log n
m n
2
e
+ sup E(uN , γN ) cm (γN ) +
.
:=
n
n
m−1
uN ∈S
I
DISTRIBUTED INFERENCE FOR QUANTILE REGRESSION PROCESSES
5
Moreover there exists a constant C independent of n, S such that for n sufficiently large and A ≥ 1
2
(2)
(L)
rN (τ, uN ) > CRn,2 (A, κn ) ≤ 2Sn−A + 2e−κn
(S.2.12) sup P
sup
τ
uN ∈SIm−1
where
(L)
Rn,2 (A, κn ) := A
2 (log n)2
ξm
4
e N , γN ) 1 + κn
E(u
c
(γ
)
+
m
N
n
S 1/2
uN ∈S m−1
sup
I
κn
ξm (κn ∨ log n)3/2
1/2
c
(γ
)(κ
∨
log
n)
+
.
m N
n
n1/2 S 1/2
n1/4
1/2
(S.2.13)
+A
1/2
See Section S.5.2 of the online Appendix for a proof of Theorem S.2.4.
The following corollary gives an ABR representation which centered at
the true function.
Corollary S.2.5. If the conditions in Theorem S.2.4 hold, and additionally for a fixed x0 ∈ X , cm (γN ) = o(kZ(x0 )kN −1/2 ) then for any τ ∈ T ,
(S.2.9) in Theorem S.2.4 holds with uN and u>
N γN (τ ) being replaced by
Z(x0 ) and Q(x0 ; τ ).
Lemma S.2.6.
In setting of Example 3.6 we have
sup
τ ∈T ,x∈[0,1]
−1
E|Z(x)Jem
(γN (τ ))Z| = O(1)
sup
τ ∈T ,x∈[0,1]
−1
E|Z(x)Jm
(τ )Z| = O(1)
The proof of Lemma S.2.6 follows by similar arguments as the proof of
Example 2.3 on page 3279 in Chao et al. (2017) and is omitted for brevity.
APPENDIX S.3: PROOFS FOR SECTION 3 AND SECTION A
S.3.1. Proofs for Section 3.1. Here we provide proofs for Theorem 3.1
and the sufficient conditions for the oracle rule in Corollary 3.3 and Corollary 3.4. Theorem 3.2 and the necessary conditions of Corollary 3.3 and
Corollary 3.4 will be established in Section S.3.3.
Theorem 3.1 follows from Theorem A.1 by noting that for any P ∈ P1 ,
the quantities m, ξm are constants that do not depend on N . Moreover, since
Jm (τ ) is invertible by assumptions (A1)-(A3), γN ≡ β, and additionally we
have cm (γN ) = gN (γN ) = 0.
6
The sufficiency of S = o(N 1/2 (log N )−1 ) for weak convergence of β̄ in
Corollary 3.3 follows from the corresponding weak convergence of βbor and (3.2)
which implies that under S = o(N 1/2 (log N )−1 ) we have β̄−βbor = oP (N −1/2 ).
It remains to prove the ’sufficient’ part of Corollary 3.4. As a direct consequence of the second part of Theorem 3.1 we obtain
sup |βbj (τ ) − βbor,j (τ ))| ≤ kΠK βj (·) − βj (·)k∞ + oP (N −1/2 ).
τ ∈T
Next, note that by assumption, βj ∈ Ληc (T ). Applying Theorem 2.62 and
Theorem 6.25 of Schumaker (1981), we find that
inf kβj (·) − gk∞ . G−η ,
(S.3.1)
g∈ΘG
where ΘG denotes the space of splines of degree rτ over the grid t1 , ..., tG on
T = [τL , τU ] ⊂ (0, 1). This together with Lemma 5.1 of Huang (2003) shows
that
kΠK βj (·) − βj (·)k∞ = O(G−η ).
(S.3.2)
Thus the ’sufficient’ part of the statement of Corollary 3.4 follows.
2
S.3.2. Proofs for Section 3.2. This section contains the proofs of
Theorem 3.7 and the sufficient conditions for weak convergence in Corollary 3.9 and Corollary 3.10. The proof of Theorem 3.8 and the necessary
conditions in Corollary 3.9 and Corollary 3.10 are given in section S.3.3.
S.3.2.1. Proof of Theorem 3.7. We begin with a proof of (3.10). Define
log N 2
Sξ 2 (log N )2
(S.3.3) Rn (S) := 1 + 1/2
cm (γN ) + m
N
S
2 (log N )10 1/4
kZ(x0 )k Sξm
kZ(x0 )kξm S log N
+
+
.
N
N
N 1/2
From Theorem S.6.2 in the supplementary material we obtain the following
representation for the oracle estimator
(S.3.4)
sup
τ ∈T
where
Z(x0 )> b
e 1,N (τ ) = OP (RN (1))
(βor (τ ) − γN (τ )) − U
kZ(x0 )k
(S.3.5)
N
>
X
e 1,N (τ ) := −N −1 Z(x0 ) Jem (γN (τ ))−1
U
Zi (1{Yi ≤ Z>
i γN (τ )} − τ ).
kZ(x0 )k
i=1
DISTRIBUTED INFERENCE FOR QUANTILE REGRESSION PROCESSES
Let uN :=
Z(x0 )
kZ(x0 )k
7
and apply Theorem S.2.4 to obtain
(1)
(2)
e
u>
N (β(τ ) − γN (τ )) = U1,N (τ ) + rN (τ, uN ) + rN (τ, uN )
(S.3.6)
with
C (L)
C (L)
(1)
(2)
P sup |rN (τ, uN ) + rN (τ, uN )| ≥ R1,n (A log n) + R2,n (A, A log n)
2
2
τ ∈TK
≤2N 2 (3S + 2)N −A
(L)
(L)
and R1,n (A log n), R2,n (A, A log n) defined in Theorem S.2.4. Note that un−1 ). Now a straightforward
e N , γN ) = O(ξm
der assumption (L) we have E(u
but tedious calculation shows that under the assumptions of Theorem 3.7
we have, for any fixed constant A > 0,
(L)
(L)
R1,n (A log n) + R2,n (A, A log n) = O(Rn (S)).
Choosing A = 3, a combination of (S.3.4) and the bound above yields (3.10).
To prove the second claim we first shall prove that
e 1,N (τ ) − UN (τ ) = oP N −1/2 ,
(S.3.7)
sup U
τ ∈T
where
N
(S.3.8)
UN (τ ) := −N −1
X
Z(x0 )>
Jm (τ )−1
Zi (1{Yi ≤ Q(Xi ; τ )} − τ ).
kZ(x0 )k
i=1
This follows by similar arguments as Step 1 in the proof of Theorem 2.4 in
Section A.2 on pages 3294-3296 of Chao et al. (2017), by noting that the
conditions required there follow exactly by our (A1)-(A3), (L), (L1) where
4 (log N )6 = o(N ), c (γ )2 = o(N −1/2 ) by (L1), and I and L
we use that ξm
m N
in the statement of Theorem 2.4 of Chao et al. (2017) being the coordinates
of the non-zero components of Z(x0 ) and r, respectively.
The proof of the second claim is very similar to the proof of (A.2).
Use (S.3.7) instead of (S.3.34) and (S.3.4) instead of (S.3.30) and similar
arguments as in the proof of (A.2) to obtain
sup (ΠK Z(x0 )> βbor (·))(τ ) − Z(x0 )> βbor (τ ) ≤ oP (kZ(x0 )kN −1/2 )
τ ∈T
+ sup [ΠK Q(x0 ; ·)](τ ) − Q(x0 ; τ ) + sup |Z(x0 )> γN (τ ) − Q(x0 ; τ )|
τ ∈T
τ ∈T
+ sup [ΠK Z(x0 )> UN ](τ ) − Z(x0 )> UN (τ ) ,
τ ∈T
and bound the last term exactly as in the proof of (A.2).
8
S.3.2.2. Proof of (3.13) and the weak convergence of βbor (τ ) in Corollary 3.9. Combining (S.3.4) and (S.3.7) we obtain the representation
Z(x0 )> (βbor (τ ) − γN (τ ))
= −N −1 Z(x0 )> Jm (τ )−1
N
X
Zi (1{Yi ≤ Q(Xi ; τ )} − τ ) + oP
i=1
kZ(x )k
0
.
N 1/2
Moreover, by straightforward calculations bounding the terms in (3.10) we
b )) =
obtain under assumption (L1) the representation Z(x0 )> (βbor (τ ) − β(τ
−1/2
oP (kZ(x0 )kN
).
To prove the weak convergence result for both βbor and β, it thus suffices
to prove that
N
1 X Z(x0 )> Jm (τ )−1 Zi (1{Yi ≤ Q(Xi ; τ )} − τ )
√
N i=1 (Z(x0 )> Jm (τ )−1 E[ZZ> ]Jm (τ )−1 Z(x0 ))1/2
N (0, τ (1 − τ )).
This follows by an application of the Lindeberg CLT. The verification of the
Lindeberg condition here is a simple modification from finite dimensional
joint convergence in the proof of (2.4) in Section A.1.2 in Chao et al. (2017)
to pointwise convergence.
S.3.2.3. Proof of (3.15) and part 2 of Corollary 3.10. The proof of the
process convergence of βbor (τ ) follows directly by Theorem 2.4 of Chao et al.
(2017). Moreover, the bound
b ) − βbor (τ ))| = oP (N −1/2 kZ(x0 )k)
sup |Z(x0 )(β(τ
τ ∈T
follows from Theorem 3.7 after some simple computations to bound the term
supτ ∈TK |Z(x0 )(β(τ ) − βbor (τ ))| and using similar arguments as in the proof
of (S.3.2) to obtain the bound
sup [ΠK Q(x0 ; ·)](τ ) − Q(x0 ; τ ) = O(G−η ).
τ ∈T
This implies (3.15). The proof of the second part follows by similar arguments as the proof of Corollary 4.1 in Chao et al. (2017) and is omitted for
the sake of brevity.
S.3.3. Proof of Theorem 3.2, Theorem 3.8 and necessary conditions for S, G in Section 3.1 and Section 3.2.
DISTRIBUTED INFERENCE FOR QUANTILE REGRESSION PROCESSES
9
First, we will prove in Section S.3.3.1 that the oracle rule in Corollary 3.4
fails whenever G . N 1/(2η) and that the oracle rule in Corollary 3.10 fail
whenever G . N 1/(2η) kZ(x0 )k−1/η , no matter how S is chosen.
Second, in Section S.3.3.2 we will construct a special class of data generation processes. This construction will be utilized in Section S.3.3.3 to prove
that the oracle rules in Corollary 3.3 and Corollary 3.9 fail if S & N 1/2 or
−1 , respectively.
S & N 1/2 ξm
Third, in Section S.3.3.4 we will prove that, provided G N 1/(2ητ )
and S & N 1/2 , the oracle rule in Corollary 3.4 fails and that, for G
−1 , the oracle rule in Corollary 3.10
N 1/(2η) kZ(x0 )k−1/ητ and S & N 1/2 ξm
fails.
Fourth, in Section S.3.3.5 we will derive the lower bounds in Theorem 3.2
and Theorem 3.8 from the necessity of the conditions in Corollary 3.9 and
3.10.
S.3.3.1. Sharpness of the conditions on G in Corollary 3.4 and 3.10. Fix
an integer ητ ≥ 2 and assume that G = GN . N 1/(2ητ ) . We will now prove
that weak convergence in (3.5) fails for some (P, Z) ∈ P1 (d1/2 , 1, f , f 0 , fmin ),
independently of the choice of S.
We begin by noting that it is possible to construct a function g on R
with the following properties: g has support [−1, 1], g is ητ times continuously differentiable on R with absolute values of all the function and
derivatives bounded uniformly by 1, g does not vanish on R. From the
properties listed here it is clear that g is not a polynomial and hence,
drτ := inf p∈Prτ supx∈[−1,1] |g(x) − p(x)| > 0 for any fixed rτ ≥ ητ and rτ ∈ N,
where Prτ is the space of polynomials of degree rτ .
We now construct a function Q on T = [τL , τU ] that is a proper quantile
function with the following properties: the corresponding distribution function F is twice continuously differentiable, its density f satisfies (A2)-(A3)
and there exists a subsequence Nk and ε > 0 such that for all k
(S.3.9)
GηNτk
inf
sup |Q(τ ) − p(τ )| ≥ ε,
p∈PPrτ (GNk ) τ ∈T
where PPrτ (GN ) is the set of piecewise polynomials of degree rτ defined on a
the equally spaced grid τL = t1,N < ... < tGN ,N = τU . Begin by considering
the case lim supN →∞ GN = ∞. Pick a subsequence Nk of N and a sequence
10
of integers νk ≤ GNk such that
h
νk (τU − τL )
(νk + 1)(τU − τL ) i
Ik := τL +
, τL +
GN k
GNk
k−1
k
X
X
−j
⊂ τL + (τU − τL )
2 , τL + (τU − τL )
2−j .
j=1
j=1
By construction, the intervals Ik are mutually disjoint and each interval is
contained between two adjacent knots in the grid t1,Nk , ..., tGNk ,Nk .
Define for some a, c > 0 (specified later):
(S.3.10)
Q(τ ) := aτ + c
∞
X
(2GNk )−ητ g(2GNk (τ − sk )/(τU − τL ))
k=2
where sk denotes the midpoint of Ik . Note that by construction the supports
of τ 7→ g(2GNk (τ − sk )/(τU − τL )) do not overlap for different k, so for each
fixed τ only one summand in the sum above does not vanish. Moreover, again
by construction, the support of τ 7→ g(2GNk (τ − sk )/(τU − τL )) is contained
between two knots of the form (τU − τL )νk /GNk , (τU − τL )(νk + 1)/GNk for
some νk ∈ N. Hence, (S.3.10) implies that
(S.3.11)
inf
sup |Q(τ ) − p(τ )| ≥ c(2GNk )−ητ drτ .
p∈PPrτ (GNk ) τ ∈T
Suppose now that lim supN →∞ GN < ∞. Since in this case GN is a
bounded sequence of natural numbers, it must take at least one value, say
N0 , infinitely often. Let s := τL + (τU − τL )/(2GN0 ). Define for a, c > 0,
(S.3.12)
Q(τ ) := aτ + c(2GN0 )−ητ g(2GN0 (τ − s)/(τU − τL )).
Similar to (S.3.11), we have for Nk ≡ N0
(S.3.13)
inf
sup |Q(τ ) − p(τ )| ≥ c(2GNk )−ητ drτ .
p∈PPrτ (GNk ) τ ∈T
Simple computations show that for the Q defined in (S.3.10) or (S.3.12)
supτ |Q0 (τ )| ≤ a + c, supτ |Q00 (τ )| ≤ c, inf τ Q0 (τ ) ≥ a − c. Hence a, c >
0 can be chosen such that (A2)-(A3) hold for F = Q−1 . Thus we have
established (S.3.9).
Now let P (X1 = m1/2 ej ) = 1/m for j = 1, ..., m and let Z(x) ≡ x and
assume X1 , ..., XN are i.i.d. Let Y1 , ..., YN be independent of {Xi }1=1,...,N
and have quantile function Q for which (S.3.9) holds and (A2)-(A3) hold
DISTRIBUTED INFERENCE FOR QUANTILE REGRESSION PROCESSES 11
for F = Q−1 . To simplify notation, assume that (S.3.9) holds with N ≡ Nk .
Observe that by definition τ 7→ βbj (τ ) ∈ PPrτ (GN ) for any j = 1, ..., d, where
we recall that B(τ ) is a vector of B-spline basis with degree rτ . Thus we have
almost surely for all N
(S.3.14)
GηNτ sup |βbj (τ ) − βj (τ )| ≥ GηNτ
τ ∈T
inf
sup |p(τ ) − Q(τ )| ≥ ε > 0.
p∈PPrτ (GNk ) τ ∈T
Note that (S.3.14) holds regardless of the number of groups S. Now, since
GN . N 1/(2ητ ) , we have C := supN ≥1 GηNτ /N 1/2 < ∞ and thus (S.3.14)
implies that for all N
N 1/2 sup |βbj (τ ) − βj (τ )| ≥ ε/C > 0
(S.3.15)
a.s.
τ ∈T
Now assume that the weak convergence (3.5) holds. We have by the continuous mapping theorem
√
(S.3.16)
N sup |βbj (τ ) − βj (τ )|
sup |G(τ )|,
τ ∈T
τ ∈T
where G is a tight, centered Gaussian process with continuous sample paths.
By the arguments given on pages 60-61 of Ledoux and Talagrand (1991), it
follows that supτ ∈T |G(τ )| has a continuous distribution with left support
point zero, and thus P (supτ ∈T |G(τ )| < δ) > 0 for all δ > 0. This also implies
that [ε/(2C), ∞) is a continuity set of the distribution of supτ ∈T |G(τ )|, and
thus the weak convergence together with the Portmanteau Theorem implies
that
√
lim P
N sup |βbj (τ ) − βj (τ )| ≥ ε/(2C) = P sup |G(τ )| ≥ ε/(2C) < 1.
N →∞
τ ∈T
τ ∈T
This contradicts (S.3.15), and thus the weak convergence (3.5) can not hold.
Arguments for the case in Corollary 3.10 with increasing m√are similar as
above, except that the rate of weak convergence in (S.3.16) is N kZ(x0 )k−1 ,
and GN . N 1/(2ητ ) kZ(x0 )k−1/ητ .
S.3.3.2. A special class of data generation processes. Consider a family
of quantile functions Qa,b (τ ) = aτ 2 + bτ , τ ∈ (0, 1) indexed by a ≥ 0, b > 0.
The corresponding distribution functions Fa,b have support [0, a + b] and
take the form
(
1
(−b + (b2 + 4ay)1/2 ), a > 0, b > 0;
(S.3.17)
Fa,b (y) = 2a
y
a = 0, b > 0.
b,
12
Note that the first two derivatives of the function Fa,b defined in (S.3.17)
are
0
00
Fa,b
(y) = (b2 + 4ay)−1/2 , Fa,b
(y) = −2a(b2 + 4ay)−3/2
and in particular
inf
y∈[0,a+b]
0
Fa,b
(y) ≥ (b+2a)−1
sup
y∈[0,a+b]
0
|Fa,b
(y)| ≤ b−1 ,
sup
y∈[0,a+b]
00
|Fa,b
(y)| ≤ 2ab−3 .
Choosing b = 1/f and a such that
(S.3.18)
0 ≤ 2a < 1/fmin − 1/f ∧ f 0 (f )−3
ensures that
(S.3.19)
0
(y) ≥ fmin ,
inf Fa,b
y∈[0,a+b]
sup
y∈[0,a+b]
0
(y)| ≤ f ,
|Fa,b
sup
y∈[0,a+b]
00
(y)| ≤ f 0 .
|Fa,b
Let Y1 , ..., Yn be i.i.d. with distribution function Fa,b and X1 , ..., Xn i.i.d.,
√
independent of Y1 , ..., Yn , and such that P (X1 = ej m) = 1/m, j = 1, ..., m
where ej ∈ Rm denotes the jth unit vector. In this model, we have
n
X
i=1
ρτ (Yi − β > Xi ) =
m
X
X
ρτ (Yi −
√
mβj ).
j=1 i:Xi =ej
Define Aj := {i : Xi = ej }, nj := #Aj . Letting
(S.3.20)
e ) = argmin
β(τ
n
X
β∈Rm i=1
ρτ (Yi − β > Xi )
P
√
we find that βej (τ ) = argminb∈R i:Xi =ej ρτ (Yi − mb), and by p.9 in
Koenker (2005) the solution of this minimization problem with smallest
j
j
absolute value is βej (τ ) = m−1/2 Y(dn
1{nj > 0} where Y(k)
is the k-th
j τ e)
order statistic of the sample {Yi }i∈Aj and due denotes the smallest integer
greater or equal u.
Now assume that m, n, K, G are sequences of positive integers indexed by
N → ∞ such that n ≥ 2, m ≤ n, n = N/S for an integer S.
In Section S.7 of the online Appendix we shall prove the following facts
about βej (τ ) defined above
1. For any fixed τ ∈ T , b > 0 and any amax > 0 there exists a ∈ [0, amax ]
such that
n
(S.3.21)
E[m1/2 βej (τ )] − Qa,b (τ ) > 0
lim sup
N →∞ m
n
(S.3.22)
Var(m1/2 βej (τ )) . 1
m
DISTRIBUTED INFERENCE FOR QUANTILE REGRESSION PROCESSES 13
2. Assume that G (N/m)α for some α > 0 and N/m ≥ 1. Let β̌j (τ ) :=
PK
e
k=1 Ak (τk )βj (τk ) where
(S.3.23)
>
Ak (τ ) := B(τ )
X
K
k=1
B(τk )B(τk )
>
−1
B(τk ),
and we recall that B = (B1 , ..., Bq )> where {B1 , ..., Bq } B-spline basis
with equidistant knots τL = t1 < ... < tG = τU and degree rτ ≥ ητ . For
any closed interval T0 ⊂ T with non-empty interior and any amax , b > 0
there exists a ∈ [0, amax ] and a sequence τN in T0 such that
(S.3.24)
lim sup
N →∞
(S.3.25)
n
E[m1/2 β̌j (τN )] − Qa,b (τN ) > 0
m
n
Var(m1/2 β̌j (τN )) . 1.
m
S.3.3.3. Necessity of conditions of oracle rules in Corollary 3.3 and Corollary 3.9. Fix an arbitrary τ ∈ T . Let Y1 , ..., YN be i.i.d. with distribution
function Fa,b defined in (S.3.17) and independent of X1 , ..., XN where Xi
√
i.i.d. with P (X1 = ej m) = 1/m, j = 1, ..., m. Let Z(x) := x. Define
b = 1/f and pick a satisfying (S.3.18) such that (S.3.21)-(S.3.22) hold.
We begin by considering the case where m = d is fixed. Denote by P the
measure corresponding to X1 , Y1 . Due to the choice of a, b and the distribution of X the pair (Z, P ) is an element of P1 (d1/2 , 1, f , f 0 , fmin ). Moreover,
by construction, {βbjs (τ )}s=1,...,S are i.i.d. and βbj1 (τ ) has the same distribution
as βej (τ ) defined in (S.3.20).
Next, we shall prove that weak convergence of β in Corollary 3.3 fails for
u = e1 . Given (S.3.22), a simple computation shows that
√
N
(S.3.26) E[( N (β 1 (τ ) − E[β 1 (τ )]))2 ] = Var(βe1 (τ )) = nVar(βe1 (τ )) . 1
S
√
for any fixed τ ∈ T . Thus the sequence N (β 1 (τ ) − E[β 1 (τ )]) is uniformly
integrable.
Now assume that the that weak convergence of β in Corollary 3.3 holds.
By (S.3.21) we know that for S & N 1/2 and any fixed τ ∈ T
E[β 1 (τ )] − β1 (τ ) & 1/n = S/N & N −1/2 .
√
If E[β 1 (τ )]−β1 (τ ) N −1/2 , uniform integrability of N (β 1 (τ )−E[β 1 (τ )])
√
implies uniform integrability
of N (β 1 (τ ) − β1 (τ )). In that case weak con√
vergence of β implies E[ N (β 1 (τ ) − β1 (τ ))] → 0 by Theorem 25.12 on p.338
of Billingsley (1995), which contradicts (S.3.27).
(S.3.27)
14
If E[β 1 (τ )]−β1 (τ ) N −1/2 , (S.3.26) implies that N E[(β 1 (τ )−β1 (τ ))2 ] →
√
∞, and thus N β 1 (τ ) − β1 (τ ) diverges to infinity in probability, which
contradicts weak convergence of β.
Now consider the setting of Section 3.2 where m → ∞. With the choice of
X described in Section S.3.3.2 we have γN (τ ) = m−1/2 Qa,b (τ )(1, ..., 1)> ∈
Rm , ξm = m1/2 , cm (γN ) = 0. The matrix E[ZZ> ] is the m × m identity matrix while Jm (τ ), Jem (γN (τ )) are diagonal with entries bounded away from
zero and infinity. A simple computation shows that (L) holds with r = 1. Denote by PN the sequence of measures corresponding to Xi , Yi . Picking m such
that m2 (log N )6 = o(N ) ensures that (Z, PN ) lies in PL (1, f , f 0 , fmin , R) for
any R ≥ 1. Letting x0 = Z(x0 ) = e1 ∈ Rm the weak convergence in (3.13)
takes the form
r
N 1/2
(m β 1 (τ ) − Qa,b (τ ))
N (0, σ(τ ))
m
for some constant σ(τ ). Moreover, (S.3.22) yields
√
E[( N m−1/2 (m1/2 β 1 (τ ) − m1/2 E[β 1 (τ )]))2 ]
N
n
=
Var(m1/2 βe1 (τ )) = Var(m1/2 βe1 (τ )) . 1
Sm
m
−1
while (S.3.21) implies that for S & N 1/2 /m1/2 = N 1/2 ξm
E[m1/2 β 1 (τ )] − Qa,b (τ ) & m/n = mS/N & m1/2 N −1/2 .
Now failure of weak convergence in (3.13) can be established for x0 =
Z(x0 ) = e1 ∈ Rm by exactly the same arguments as given in the first part
of the proof for the fixed m case, details are omitted for the sake of brevity.
S.3.3.4. Necessity of conditions for oracle rules in Corollary 3.4 and Corollary 3.10. Fix a subset T0 ⊂ T which is closed and has non-empty interior.
Let Y1 , ..., YN be i.i.d. with distribution function Fa,b defined in (S.3.17) and
√
independent of X1 , ..., XN where Xi i.i.d. with P (X1 = ej m) = 1/m, j =
1, ..., m. Let b = 1/f and pick a satisfying (S.3.18) and a sequence τN in T0
such that (S.3.24)-(S.3.25) hold. Since τN is a sequence in a compact set, it
must have a convergent subsequence τNk → τ0 ∈ T0 . From now on, we will
without loss of generality assume that N = Nk .
Assume that (3.5) holds. Process convergence in (3.5), along with the
continuity of the sample paths√of the limiting processes, implies that the
sequence of random variables N (m1/2 βb1 (τN ) − Qa,b (τN )) converges to a
centered normal random variable. This can be led to a contradiction with
DISTRIBUTED INFERENCE FOR QUANTILE REGRESSION PROCESSES 15
(S.3.24)-(S.3.25) by similar arguments
as in Section S.3.3.3 after observing
P
that by definition βb1 (τ ) = S −1 Ss=1 β̌1 (τ ).
The proof that (3.15) fails can be done similarly, and details are omitted
for the sake of brevity.
2
S.3.3.5. Proof of Theorem 3.8 and Theorem 3.2. The proofs of both results follow by the same type of arguments, and thus we only give details
for the proof of Theorem 3.2. This proof will be based on the necessary
conditions in Corollary 3.3 and Corollary 3.4.
First we derive (3.3) by contradiction. Assume that for some τ ∈ T we
have for any (P, Z) ∈ P1 and C > 0 that √
P (kβ̄(τ ) − βbor (τ )k > CS/N ) → 0.
Then β̄(τ ) − βbor (τ ) = oP (S/N ) and thus N (β̄(τ ) − β(τ ))
N (0, H(τ, τ ))
1/2
even if S = N . However, this contradicts the necessity statement in
Corollary 3.3 (in the proof of Corollary 3.3 we show that for any τ ∈ T
if S/N 1/2 & 1 then there exists (P, Z) ∈ P1 such that this weak convergence
fails).
The claim in (3.4) will be derived from the ’necessary’ part of Corollary 3.4
by similar arguments. Assume that (3.4) does not hold, i.e. there exist c, η >
0 such that for any (P, Z) ∈ P1 and any x0 ∈ X with τ 7→ Q(x0 ; τ ) ∈ Ληc we
have for any C > 0
b ) − βbor (τ )k ≥ CS + CG−η → 0.
P sup kβ(τ
N
τ ∈T
This implies
b ) − βbor (τ )k = oP S + G−η ,
sup kβ(τ
N
τ ∈T
√
√
G implies N (βb − β)
G.
and the process convergence N (βbor − β)
This contradicts the necessity of S = o(N 1/2 ) and G N 1/(2η) in Corollary 3.4. Hence (3.4) follows.
2
S.3.4. Proofs for Section A. Recall the definition of gN (bN ) in (S.2.2)
and cm (bN ) in (3.9). As mentioned in Remark A.6, the results in Section A
hold for general ’centerings’ βN (which includes γN as a special case) provided that certain technical conditions are satisfied. The precise form of
those conditions is discussed in section S.3.4.1. There, we show that γN satisfies those general conditions. In sections S.3.4.2 and S.3.4.5, all theoretical
results will be proved for βN instead of γN under the general conditions
described in section S.3.4.1. The results with γN in Section A follow as a
special case.
16
S.3.4.1. Details of Remark A.6. Assume that (A1)-(A3) hold. Corollary A.2 holds for any sequence of centering coefficients βN under the fol2 (log N )2 )), Sc2 (β )m log N = o(1),
lowing conditions: S = o(N 1/2 /(mξm
m N
and
4 = o(N ).
(B1) gN (βN ) = o(N −1/2 ) and m4 (log N )10 + m2 ξm
(B2) mcm (βN ) log N = o(1) and the sequence uN (in Corollary A.2) satisfies
−1
>
χN (uN , βN ) := sup u>
N Jm (τ ) E Zi 1{Yi ≤ Q(Xi ; τ )} − 1{Yi ≤ Zi βN (τ )}
τ ∈T
= o(kuN kN −1/2 ).
Furthermore, Corollary A.5 holds under the additional restriction cm (βN ) =
o(kZ(x0 )kN −1/2 ) and the other conditions on Q, G, K stated in Corollary A.5.
In the sequel, we verify that, assuming the conditions stated in Corollary A.2 and Corollary A.5 respectively, γN satisfies the conditions given
above. Note that some conditions in these theorems overlap with the conditions stated above, so we only need to verify that the conditions c2m (γN )ξm =
2 (log N )2 )) in both theorems imply g (γ ) =
o(N −1/2 ) and S = o(N 1/2 /(mξm
N
N
o(N −1/2 ), χN (uN , γN ) = o(kuN kN −1/2 ), Sc2m (γN )m log N = o(1) and moreover mcm (γN ) log N = o(1).
Under the definition of γN in (3.8) and (A1)-(A3), we have gN (γN ) =
O(ξm c2m (γN )). To see this, note that under (A2),
Z(FY |X (Z> γN (τ )|X) − τ ) = ZfY |X (Q(X; τ )|X)(Z> γN (τ ) − Q(X; τ ))
1
(S.3.28)
+ ZfY0 |X (ζn (X, τ )|X)(Z> γN (τ ) − Q(X; τ ))2 ,
2
where ζn (X, τ ) is a value between Z(X)> γN (τ ) and Q(X; τ ). Thus
(S.3.29) gN (γN ) = sup E[Z(FY |X (Z> γN (τ )|X) − τ )]
τ ∈T
≤
f 0 cm (γN )2
kZk = O(ξm c2m (γN )),
2
where the second inequality follows from the first order condition
E[ZfY |X (Q(X; τ )|X)(Z> γN (τ ) − Q(X; τ ))] = 0
by the definition of γN in (3.8) and (S.3.28). Moreover,
−1
χN (uN , γN ) = sup u>
N Jm (τ ) µ(γN (τ ), τ ) . kuN k2 gN (γN ).
τ ∈T
DISTRIBUTED INFERENCE FOR QUANTILE REGRESSION PROCESSES 17
Hence, c2m (γN )ξm = o(N −1/2 ) implies gN (γN ) = o(N −1/2 ) and χN (uN , γN ) =
o(kuN kN −1/2 ). In addition,
−1
Sc2m (γN )m log N = o(SN −1/2 ξm
m log N ) = o(1),
where the first equality is from c2m (γN )ξm = o(N −1/2 ) and the second from
2 (log N )2 )). Finally
S = o(N 1/2 /(mξm
mcm (γN ) log N = o(mN −1/4 log N ) = o((m4 N −1 (log N )4 )1/4 ) = o(1)
where the first bound follows from c2m (γN )ξm = o(N −1/2 ), ξm & 1 and the
third bound from m4 (log N )10 = o(N ).
S.3.4.2. Proof of Theorem A.1. We begin by proving of (A.1). From Theorem S.6.1 in the supplementary material, we obtain the following representation for the oracle estimator
(S.3.30)
N
1 X
or
βbor (τ ) − βN (τ ) = − Jm (τ )−1
Zi (1{Yi ≤ Z>
i βN (τ )} − τ ) + rn (τ )
N
i=1
or ) and
where supτ ∈T krnor (τ )k = oP (RN
or
RN
:= cm
m log N 1/2
mξ 2 log N
m
4
+ c2m ξm +
+ c4m ξm
N
N
1/2 m3 ξ 2 (log N )7 1/4
1 2 2
m
.
+
+ 1/2 mξm cm (log N )3
N
N
Next we will prove that under the assumptions of Theorem A.1 we also have
(S.3.31)
sup β(τ ) − βN (τ ) + Jm (τ )−1
τ ∈TK
N
1 X
Zi (1{Yi ≤ Z>
i βN (τ )} − τ ) = OP (RN )
N
i=1
where
Smξ 2 log N
log N
m
4
+ c2m ξm +
+ c4m ξm
1 + 1/2
N
N
S
1/2 Sm3 ξ 2 (log N )7 1/4
1 2 2
m
+ 1/2 mξm cm (log N )3
+
.
N
N
RN := cm
m log N 1/2
18
To this end, apply Theorem S.2.1. A simple computation shows that the
quantities R1,n , R2,n defined in (S.2.4) and (S.2.6) of Theorem S.2.1 satisfy
R1,n (A log n) + R2,n (A, A log n) ≤ CRN
for any fixed A > 0. Thus we obtain from (S.2.3) and (S.2.5) that there
exists N0 such that for all N ≥ N0 and any τ ∈ T
(1)
(2)
e N ≤ 2n−A (3S + 1),
(S.3.32)
P rN (τ ) + rN (τ ) > CR
where
β(τ ) − βN (τ ) + Jm (τ )−1
N
1 X
(1)
(2)
Zi (1{Yi ≤ Z>
i βN (τ )} − τ ) = rN (τ ) + rN (τ ).
N
i=1
Choosing A large enough ensures that
(1)
(S.3.33)
(2)
sup rN (τ ) + rN (τ ) = oP (RN ),
τ ∈TK
and we obtain (S.3.31).
Next we prove (A.2). Recall the definition of the projection operator ΠK
right after (2.5).
Step 1: kΠK k∞ := supkf k∞ =1 kΠK f k∞ < C for a constant C > 0 independent of K.
We will apply Theorem A.1 of Huang (2003), and for that we need to validate conditions A.1-A.3 of Huang (2003). Note that none of these conditions
is associated with the functions being projected on the spline space – all conditions are associated with the design and the selection of basis functions.
We first show that Condition A.1 of Huang (2003) holds in our context. By
the definition of grids t1 , ..., tG and τ1 , ..., τK , each interval [tk , tk+1 ] contains
ck,K spaced grid points of the form τU + l(τL − τU )/K. Hence, applying
Lemma S.8.2, we obtain for j = 0, ..., G − 1 and any f ∈ ΘG
Z
X
Crτ tj+1 2
−1
2
K
f (τk ) ≤ 1 +
f (x)dx,
cj,K − 1 tj
k:τk ∈(tj ,tj+1 ]
and
K
−1
X
k:τk ∈(tj ,tj+1 ]
Crτ
f (τk ) ≥ 1 −
cj,K − 1
2
Z
tj+1
tj
f 2 (x)dx,
DISTRIBUTED INFERENCE FOR QUANTILE REGRESSION PROCESSES 19
for a constant Crτ that depends only on rτ . Since mink=1,...,K cj,K → ∞ it
follows that for sufficiently large K Condition A.1 of Huang (2003) holds
with γ1 = 1/2 and γ2 = 2.
Conditions A.2 (i) and (ii) of Huang (2003) easily follow from the fact
that T is bounded. Finally, the definition of the knots ensures that A.2 (iii)
in Huang (2003) holds. Condition A.3 holds since B-spline basis is used in
the projection procedure, cf. the discussion following Condition A.3 on page
1630 of Huang (2003).
Thus conditions A.1-A.3 in Huang (2003) are satisfied, and Theorem A.1
of Huang (2003) shows the existence of a constant C such that kΠK f k∞ ≤
Ckf k∞ for all functions f ∈ `∞ (T ). This completes the proof of step 1.
From the definition of βb and linearity of ΠK we obtain
and thus
b ) = (ΠK Z(x0 )> β(·))(τ )
Z(x0 )> β(τ
b ) − βbor (τ )) = (ΠK {Z(x0 )> [β(·) − βbor (·)]})(τ )
Z(x0 )> (β(τ
+ (ΠK {Z(x0 )> βbor (·)})(τ ) − Z(x0 )> βbor (τ ).
Observe that
sup (ΠK {Z(x0 )> [β(·) − βbor (·)]})(τ ) . sup Z(x0 )> [β(τ ) − βbor (τ )] .
τ ∈T
τ ∈TK
Hence it remains to prove that
sup (ΠK Z(x0 )> βbor (·))(τ ) − Z(x0 )> βbor (τ )
τ ∈T
≤ sup (ΠK Q(x0 ; ·))(τ ) − Q(x0 ; τ ) + oP (kZ(x0 )kN −1/2 ).
τ ∈T
From Theorem 2.1 of Chao et al. (2017) we obtain
(S.3.34)
where
(S.3.35)
Z(x0 )> βbor (τ ) = Q(x0 ; τ ) + Z(x0 )> UN (τ ) + rN (τ )
UN (τ ) := −N
−1
−1
Jm (τ )
N
X
i=1
Zi (1{Yi ≤ Q(Xi ; τ )} − τ )
20
and supτ ∈T |rN (τ )| ≤ supτ ∈T |Z(x0 )> γN (τ )−Q(x0 ; τ )|+oP (kZ(x0 )kN −1/2 ).
This implies
kZ(x )k
0
sup (ΠK Z(x0 )> βbor (·))(τ ) − Z(x0 )> βbor (τ ) ≤ oP
1/2
N
τ ∈T
+ sup [ΠK Q(x0 ; ·)](τ ) − Q(x0 ; τ ) + sup |Z(x0 )> γN (τ ) − Q(x0 ; τ )|
τ ∈T
τ ∈T
+ sup [ΠK Z(x0 )> UN ](τ ) − Z(x0 )> UN (τ ) .
τ ∈T
and it remains to bound the last term. From Lemma 5.1 of Huang (2003)
we obtain that
(S.3.36)
sup (ΠK Z(x0 )> UN )(τ ) − Z(x0 )> UN (τ )| ≤ C inf sup |Z(x0 )> UN (τ ) − fe(τ )|,
τ ∈T
fe∈ΘG τ ∈T
where C is an absolute constant and ΘG is the space of splines of degree rτ
with knots t1 , ..., tG on T = [τL , τU ] ⊂ (0, 1). The right hand side of (S.3.36)
can be related to the modulus of continuity. Indeed, using Theorem 6.27 of
Schumaker (1981) we have for an absolute constant C > 0,
(S.3.37)
inf sup Z(x0 )> UN (τ ) − fe(τ )
fe∈ΘG τ ∈T
≤C
sup
τ,τ 0 ∈T ,|τ −τ 0 |≤δ
|Z(x0 )> UN (τ ) − Z(x0 )> UN (τ 0 )| = oP
kZ(x )k
0
,
N 1/2
where δ = max0≤j≤K−1 (tj+1 − tj ) = o(1), and the last equality follows from
Lemma S.8.1. Thus the proof of (A.2) is complete.
2
S.3.4.3. Proof of Corollary A.2. From Theorem 2.1 of Chao et al. (2017)
we obtain the representation
Z(x0 )> (βbor (τ ) − γN (τ ))
= −Z(x0 )> Jm (τ )−1
N
kZ(x )k
1 X
0
Zi (1{Yi ≤ Q(Xi ; τ )} − τ ) + oP
.
N
N 1/2
i=1
Moreover, by straightforward calculations bounding the terms in (A.1) we
obtain the representation Z(x0 )> (βbor (τ )−β(τ )) = oP (kZ(x0 )kN −1/2 ). Hence
to prove the weak convergence result for both βbor and β, it suffices to prove
that
N
1 X Z(x0 )> Jm (τ )−1 Zi (1{Yi ≤ Q(Xi ; τ )} − τ )
√
N i=1 (Z(x0 )> Jm (τ )−1 E[ZZ> ]Jm (τ )−1 Z(x0 ))1/2
N (0, τ (1 − τ )).
DISTRIBUTED INFERENCE FOR QUANTILE REGRESSION PROCESSES 21
This follows by an application of the Lindeberg CLT. The verification of the
Lindeberg condition here is a simple modification from finite dimensional
joint convergence in the proof of (2.4) on page 3292 in Chao et al. (2017) to
pointwise convergence.
S.3.4.4. Details for Remark A.4. In the proof below, we will use the
ej , j’th knot tj and degree q.
notations in Example 3.6: polynomial spline B
We first verify (A1). The bounded spectrum E[ZZ> ] follows by the argument in Example 3.6. We note that every univariate polynomial spline
ej | ≤ C almost surely for some constant C > 0 for all
function satisfies |B
ej and B
ek is disjoint unless |j − k| ≤ q + 1.
j ≤ m. Moreover, the support of B
Hence, we have almost surely
1/2
(S.3.38) kZi k = m
k−q−1
X
j=1
ej (Xi )2
B
1/2
.m
1/2
k−q−1
X
j=1
1{tj ≤ Xi ≤ tj+q+1 }
1/2
. m1/2 .
ej (Xi ) is nonzero almost surely, kZi k & m1/2
On the other hand, since each B
1/2
and ξm m .
We now verify the sufficient conditions of Corollary 3.9 and Corollary A.2
under the conditions made in this remark. Condition (L) holds given the dis4 (log N )6 = o(N ), c2 (γ ) =
cussion in Example 3.6. (L1) and the conditions ξm
m N
o(N −1/2 ) hold under the assumptions of this remark. On the other hand,
the sufficient conditions of Corollary A.2 can be easily verified by the assumptions of this remark since ξm m1/2 .
S.3.4.5. Proof of Corollary A.5. Similar to Section S.3.4.2, we will prove
b ) in models with centering coefficient βN (τ )
the weak convergence of β(τ
from a class specified in Section S.3.4.1.
√
The process convergence of kZ(xN0 )k Z(x0 )> βbor (·) − Q(x0 ; ·) is exactly
Theorem 2.1 of Chao et al. (2017). A straightforward computation shows
that under the assumptions of Corollary A.5 the results in (A.1) and (A.2)
yield
√
N
b ) = oP (1).
Z(x0 )> βbor (τ ) − Z(x0 )> β(τ
sup
τ ∈T kZ(x0 )k
This yields the claim in the first part. The proof of the second part follows
by similar arguments as the proof of Corollary 4.1 in Chao et al. (2017), and
is therefore omitted for brevity.
22
APPENDIX S.4: PROOFS FOR SECTION 4
S.4.1. Proof of Theorem 4.1. Since the proofs under the assumptions
of Corollary 3.4, Corollary 3.10 and Corollary A.5 are similar, we will only
give a proof under the conditions of Corollary A.5. Define
√
b D Z(x0 )
WN := n(Z(x0 )> β̄(τ ) − Q(x0 ; τ )), σ
b2 := nZ(x0 )> Σ
and
pN := P S 1/2 |WN | ≤ σ
btS−1,1−α/2 .
Throughout this proof, we will write SN instead of S in order to emphasize
that S depends on N . First, observe that along any sub-sequence Nk such
that SNk only takes the constant value S0 > 2 we have
√
n
(Z(x0 )> βb1 (τ ) − Q(x0 ; τ ), ..., Z(x0 )> βbS0 (τ ) − Q(x0 ; τ ))
N (0, IS0 )
σn
where IS0 denotes the S0 × S0 identity matrix and
σn2 = τ (1 − τ )
Z(x0 )> Jm (τ )−1 E[ZZ> ]Jm (τ )−1 Z(x0 )
.
kZ(x0 )k2
b D Z(x0 ) is the sample variance of {Z(x0 )> βbs }s=1,...,S .
Observe that Z(x0 )> Σ
As a direct consequence of the continuous mapping theorem, we obtain
pNk → (1−α). By the definition of convergence of a sequence of real numbers
to a limit, this also implies that for any fixed S0 > 2
lim
sup
N0 →∞ N ≥N0 :SN =S0
|pN − (1 − α)| = 0,
which in turn yields that for any fixed S̄
lim
sup
N0 →∞ N ≥N :2<S ≤S̄
0
N
|pN − (1 − α)| = 0.
Next observe that for any fixed ε > 0
σ
btS−1,1−α/2
−
1
≥
ε
σN Φ−1 (1 − α/2)
−1
|WN | − Φ−1 (1 − α/2) ≤ εΦ−1 (1 − α/2)
+ P S 1/2 σN
−1
+ P S 1/2 σN
|WN | ≤ Φ−1 (1 − α/2) − (1 − α) .
pN − (1 − α) ≤ P
DISTRIBUTED INFERENCE FOR QUANTILE REGRESSION PROCESSES 23
Under the assumptions of Corollary A.5, we have the decomposition
Z(x0 )> βbs (τ ) = Q(x0 ; τ ) + Lsn (τ ) + rns (τ )
where sups=1,...,S |rns (τ )| = oP (kZ(x0 )kn−1/2 ) and
n
Lsn (τ )
1X
:=
Z(x0 )> Jm (τ )−1 Z(Xis )(1{Yis ≤ Q(Xis ; τ )} − τ ).
n
i=1
This implies
N
2
X
n
Z(x0 )> βbs (τ ) − Z(x0 )> β̄(τ )
SN − 1
S
σ
b2 =
=
s=1
n
SN − 1
1
=
SN − 1
=:
2
σ
bL,N
SN
X
s=1
SN
X
−1
Lsn (τ ) − SN
S
X
s=1
√
( nLsn (τ ))2 −
s=1
2
Lsn (τ ) + oP (1)
SN
X
√ s 2
1
nLn (τ ) + oP (1)
SN (SN − 1)
s=1
2
+ oP (kZ(x0 )k )
where the oP (1) term is for N → ∞ for arbitrary sequences SN . Noting
√
4 ] is bounded uniformly in n, s and that E[Ls (τ )] = 0,
that E[( nLsn (τ ))4 /σN
n
√ s
2
2
E[( nLn (τ )) ] = σN , an application of the Markov inequality shows that
there exists a constant C such that for any fixed ε > 0 and any SN > 2
P
2
σ
bL,N
2
σN
−1 −2
ε .
− 1 ≥ ε ≤ CSN
2 /σ 2 − 1| ≥ |b
σL,N /σN − 1| and thus
Next, observe that |b
σL,N
N
σ
btS−1,1−α/2
t1,1−α/2
tS−1,1−α/2
σ
b
−1 ≤
− 1 −1
+ −1
−1
−1
σN Φ (1 − α/2)
σN
Φ (1 − α/2)
Φ (1 − α/2)
t1,1−α/2
tS−1,1−α/2
σ
b2
≤ 2 − 1 −1
+ −1
−1 .
Φ (1 − α/2)
Φ (1 − α/2)
σN
It follows that for any fixed ε > 0 and S̄ with |tS̄−1,1−α/2 /Φ−1 (1−α/2)−1| <
e
ε/2 we have for some constant C
lim sup P
N →∞:SN ≥S̄
t
2
σ
btS−1,1−α/2
1,1−α/2
e S̄ −1
−1
≥
ε
≤
C
ε−2
σN Φ−1 (1 − α/2)
Φ−1 (1 − α/2)
24
−1
Finally, noting that S 1/2 σN
|WN |
lim sup P
N →∞
and
|N (0, 1)|, we have
−1
S 1/2 σN
|WN | − Φ−1 (1 − α/2) ≤ εΦ−1 (1 − α/2)
= P |N (0, 1)| − Φ−1 (1 − α/2) ≤ εΦ−1 (1 − α/2)
−1
P S 1/2 σN
|WN | ≤ Φ−1 (1 − α/2) − (1 − α) → 0.
Combining all results so far, we obtain for any fixed ε > 0 and S̄ with
|tS̄−1,1−α/2 /Φ−1 (1 − α/2) − 1| < ε/2
2
t
1,1−α/2
e S̄ −1
ε−2
lim sup pN − (1 − α) ≤C
−1 (1 − α/2)
Φ
N →∞
+ P |N (0, 1)| − Φ−1 (1 − α/2) ≤ εΦ−1 (1 − α/2) .
The right-hand side can be made arbitrarily small by the choice of fixed
values for ε, S̄, and thus pN → (1 − α). This completes the proof of the first
part of Theorem 4.1. The second part follows easily since tSN −1,1−α/2 →
Φ−1 (1 − α/2) if S → ∞.
2
S.4.2. Proof of Theorem 4.2. Since the arguments under the assumptions of Corollary 3.10 are very similar to the arguments under the
assumptions of Corollary A.5, we will only give a proof in the latter case.
Note also that Corollary 3.4 follows from Corollary A.5. Define
(1)
GN (τ ) := −
S
n
X
1 X
−1
(ωs,1 −1)Z(x0 )> Jm
(τ )
Z(Xis )(1{Yis ≤ Q(Xis ; τ )}−τ )
N
s=1
i=1
and
S
n
X
1 X
> −1
Z(x0 ) Jm (τ )
Z(Xis )(1{Yis ≤ Q(Xis ; τ )} − τ ).
GN (τ ) := −
N
s=1
i=1
√
Throughout the proof, let Cω := 1 + 2 and note that ωs,b ≤ Cω almost
surely. The proof will rely on the following two key steps:
Step 1: prove that for any fixed B ∈ N
√
N
(1)
(B)
(B)
(S.4.1)
(GN (·), GN (·), . . . , GN (·))
(Gx0 (·), G(1)
x0 (·), ..., Gx0 (·))
kZ(x0 )k
DISTRIBUTED INFERENCE FOR QUANTILE REGRESSION PROCESSES 25
(1)
(B)
as elements in (`∞ (T ))B+1 where the processes Gx0 , ..., Gx0 iid copies of
(1)
the process Gx0 defined in (A.4) and that GN (·) is asymptotically uniformly
equicontinuous in probability.
Step 2: prove that uniformly in τ ∈ Tk we have for j = 1, ..., B
Z(x0 )> (β
(S.4.2)
(j)
(j)
(τ ) − β(τ )) = GN (τ ) + oP
kZ(x )k
0
.
N 1/2
Given the results in step 1 and step 2, exactly the same arguments as those
in the proof of Corollary A.5 (replacing (S.3.34) by (S.4.2) and (S.3.37) by
(1)
(1)
uniform asymptotic equicontinuity of Gx0 ; note that Gx0 in the present
>
setting corresponds to Z(x0 ) UN (τ ) in the proof of Corollary A.5) show
that for j = 1, ..., B
(j)
b )) − G (τ ) = oP
sup Z(x0 )> (βb(j) (τ ) − β(τ
N
τ ∈T
kZ(x )k
0
,
N 1/2
and together with Step 2, (S.4.1) and Lemma 3.1 from Bücher and Kojadinovic (2017) the assertion follows.
It remains to establish the claims in step 1 and step 2.
Proof of step 1 Finite-dimensional convergence follows by a simple application
array Lyapunov CLT with δ = 2 after observing
√ of the triangular
P
(1)
(1)
that kZ(xN0 )k GN (τ ) = Ss=1 VN,s (τ ) where
n
X
(ωs,1 − 1)
(1)
−1
VN,s (τ ) := √
Z(x0 )> Jm
(τ )
Z(Xis )(1{Yis ≤ Q(Xis ; τ )} − τ )
N kZ(x0 )k
i=1
are independent summands. Hence it suffices to establish asymptotic uniform
equicontinuity, which entails tightness. The proof is based on similar ideas
as the proof of Lemma A.3 in Chao et al. (2017), and we will only stress the
main differences. Following the latter paper, define
UN (τ ) :=
S
S
1X
1X
(ωs,1 − 1)Jm (τ )−1
Z(Xis )(1{Yis ≤ Q(Xis ; τ )} − τ )
S
n
s=1
s=1
∆is (τ, η) := 1{Yis ≤ Q(Xis ; τ )} − 1{Yis ≤ Q(Xis ; η)} − (τ − η).
Similarly to the proof of equations (A.34) and (A.35) in the latter paper we
find that for any vector uN ∈ Rm with kuN k = 1
h
i
−1
4
2
E |(ωs,1 − 1)u>
. ξm
|τ − η|
N Jm (τ ) Z(Xis )∆is (τ, η)|
26
and
h
i
−1
−1
4
E |(ωs,1 − 1)u>
(J
(τ
)
−
J
(η))Z(X
)(1{Y
≤
Q(X
;
τ
)}
−
τ
)|
is
is
is
N m
m
2
. ξm
|τ − η|2
where the constants on the right-hand side are independent of τ, η, n. Next,
similar computations as on page 3307 (above the derivation of equation
(A.37)) in Chao et al. (2017) yield for any τ, η ∈ T
(S.4.3)
h
i
ξ2
4
E |N 1/2 u>
. |τ − η|2 + m |τ − η|.
N (UN (τ ) − UN (η))|
N
6 /N 3
In particular we have for |τ − η| ≥ ξm
i
h
4
. |τ − η|4/3 .
(S.4.4)
E |N 1/2 u>
N (UN (τ ) − UN (η))|
Apply Lemma A.1 of Kley et al. (2016) with T = T , d(s, t) = |t − s|1/3 ,
2 /N , Ψ(x) = x4 . Note that with the notation introduced in the latter
η̄n = ξm
2 /N
reference D(ε, d) . ε−3 and thus for any v ≥ ξm
sup |N 1/2 u>
N (UN (τ ) − UN (η))|
|τ −η|≤δ
≤ S1 (v) + 2
sup
|N 1/2 u>
N (UN (τ ) − UN (η))|
2 /N,τ ∈T
e
|τ −η|1/3 ≤ξm
−6 N 3 ) points and the random variable
where the set Te contains at most O(ξm
S1 (v) satisfies
Z v
2
kS1 (v)k4 .
ε−3/4 dε + (δ 1/3 + 2ξm
/N )v −6/4 .
0
Later we will prove that
(S.4.5)
sup
2 /N,τ ∈T
e
|τ −η|1/3 ≤ξm
|N 1/2 u>
N (UN (τ ) − UN (η))| = oP (1).
DISTRIBUTED INFERENCE FOR QUANTILE REGRESSION PROCESSES 27
Given (S.4.5), it follows that for any c, v > 0
−1/2
lim lim sup P
sup |N 1/2 u>
(U
(τ
)
−
U
(η))|
≥
cN
N
N
N
δ↓0 N →∞
≤ lim sup P
N →∞
|τ −η|≤δ
−1/2
|N 1/2 u>
/2
N (UN (τ ) − UN (η))| ≥ cN
sup
2 /N,τ ∈T
e
|τ −η|1/3 ≤ξm
+ lim lim sup P (S1 (v) ≥ c/2)
δ↓0 N →∞
Z
i4
16 h v −3/4
2
ε
dε + (δ 1/3 + 2ξm
/N )v −6/4
. lim lim sup 4
δ↓0 N →∞ c
0
Z
16 h v −3/4 i4
= 4
ε
dε .
c
0
Since the right-hand side can be made arbitrarily small by choosing v small,
it follows that
lim lim sup P
sup |N 1/2 u>
(U
(τ
)
−
U
(η))|
≥
c
= 0,
N
N
N
δ↓0 N →∞
|τ −η|≤δ
and hence it remains to prove (S.4.5). To do so, consider the decomposition
u>
N (UN (τ ) − UN (η))
X
1
−1
−1
= u>
(ωs,1 − 1)Z(Xis )(1{Yis ≤ Q(Xis ; τ )} − τ )
N (Jm (τ ) − Jm (η))
N
s,i
X
1
−1
+ u>
J
(η)
(ωs,1 − 1)Z(Xis )∆is (τ, η).
N N m
s,i
Similarly as in the proof of Lemma A.3 in Chao et al. (2017) (in the calculation below (A.34) on page 3306) we obtain
−1
−1
kJm
(τ ) − Jm
(η)kop . |τ − η|
6 N −3
where k · kop denotes the operator norm, and thus for |τ − η| ≤ ξm
X
1 > −1
−1
uN (Jm (τ ) − Jm
(η))
(ωs,1 − 1)Z(Xis )(1{Yis ≤ Q(Xis ; τ )} − τ )
N
s,i
(S.4.6)
7
.|τ − η|ξm = O(ξm
N −3 ) = o(N −1 )
a.s.
Next observe that for τ ≤ η
|∆is (τ, η)| ≤ 1{Q(Xis ; τ ) ≤ Yis ≤ Q(Xis ; η)} + |τ − η|.
28
This implies that for any fixed τ we have a.s. for a constant C independent
of τ, η, N
X
−1
−1
sup u>
(ωs,1 − 1)Z(Xis )∆is (τ, η)
N Jm (η)N
|τ −η|≤εN
≤C
ξm
N
Xh
i,s
s,i
i
εN + 1{Q(Xis ; τ − εN ) ≤ Yis ≤ Q(Xis ; τ + εN )} .
Thus
sup
sup
6 N −3
τ ∈Te |τ −η|≤ξm
−1
−1/2
u>
N Jm (η)N
X
(ωs,1 − 1)Z(Xis )∆is (τ, η)
s,i
εN ξ m
+ ξm sup BN (τ, εN )
.
N
τ ∈Te
where
BN (τ, εN ) :=
X
1{Q(Xis ; τ − εN ) ≤ Yis ≤ Q(Xis ; τ + εN )}.
i,s
Since BN (τ, εN ) ∼ Bin(N, 2εN ) distribution and by applying the multiplicative Chernoff bound for Binomial random variables we find that for
any t ≥ 1
P BN (τ, εN ) > N (1 + t)εN ≤ exp(−2tN εN /3).
6 N −3 and t = 3AN 2 log N/(2ξ 6 ) 1 we obtain
Plugging in εN = ξm
m
6
6
P BN (τ, 2ξm
N −3 ) > 2ξm
N −2 + 3A log N ≤ N −A .
By choosing A = 3 and applying the union bound for probabilities we finally
obtain for N sufficiently large and a constant 0 < c < ∞
X
−1
P sup
sup
u>
J
(η)
(ω
−
1)Z(X
)∆
(τ,
η)
≥
cξ
log
N
s,1
is
is
m
N m
6 N −3
τ ∈Te |τ −η|≤ξm
≤|Te|N −3 = o(1).
s,i
Combining this with (S.4.6), (S.4.5) follows.
DISTRIBUTED INFERENCE FOR QUANTILE REGRESSION PROCESSES 29
Proof of step 2 Observe that
Z(x0 )> (β
(1)
(τ ) − β(τ ))
S
1 X ωs,1
= Z(x0 )>
− 1 βbs (τ )
S
ω̄·,1
s=1
=
S
1
1 X
Z(x0 )>
ωs,1 − ω̄·,1 βbs (τ ) − βN (τ )
ω̄·,1
S
1
1
=
Z(x0 )>
ω̄·,1
S
−
s=1
S
X
s=1
s
b
ωs,1 − 1 β (τ ) − βN (τ )
S
ω̄·,1 − 1
1 X bs
Z(x0 )>
β (τ ) − βN (τ )
ω̄·,1
S
s=1
=
1
1
Z(x0 )>
ω̄·,1
S
S
X
s=1
kZ(x )k
0
ωs,1 − 1 βbs (τ ) − βN (τ ) + oP
.
1/2
N
Note that under the assumptions on the weights ω̄·,1 = 1+oP (1). Thus (S.4.2)
will follow from step 1 once we prove that, uniformly in τ ∈ TK ,
(S.4.7) Z(x0 )>
S
kZ(x )k
1 X
(1)
0
ωs,1 −1 βbs (τ )−βN (τ ) = GN (τ )+oP
.
S
N 1/2
s=1
To this end apply Theorem S.6.1, which implies that
S
1X
e (1) (τ )
(ωs,1 − 1) βbs (τ ) − βN (τ ) − G
Z(x0 )
N
S
>
s=1
= Z(x0 )>
S
1X
s
s
s
s
(ωs,1 − 1) rn,1
(τ ) + rn,2
(τ ) + rn,3
(τ ) + rn,4
(τ )
S
s=1
where
S
n
X
1 X
(1)
> −1
e
(ωs,1 − 1)Z(x0 ) Jm (τ )
ψ(Yis , Z(Xis ); βN (τ ), τ )
GN (τ ) := −
N
s=1
i=1
s (τ ) correspond to r (τ ) defined in Theoand the remainder terms rn,j
n,j
30
rem S.6.1. We will prove that
(S.4.8)
S
1X
s
s
s
s
sup Z(x0 )
(ωs,1 − 1) rn,1
(τ ) + rn,2
(τ ) + rn,3
(τ ) + rn,4
(τ )
S
τ ∈TK
>
s=1
−1/2
= oP (kZ(x0 )kN
),
(S.4.9)
e (1) (τ ) − G(1) (τ ) = oP (kZ(x0 )kN −1/2 ).
sup G
N
N
τ ∈TK
We begin by a proof of (S.4.8). Applying Theorem S.6.1 with κn = A log n
and A sufficiently large we find that
S
1X
s
s
s
Z(x0 )
(ωs,1 − 1) rn,1
(τ ) + rn,2
(τ ) + rn,4
(τ )
S
>
s=1
s
s
s
≤ sup sup Z(x0 )> (rn,1
(τ ) + rn,2
(τ ) + rn,4
(τ ))
s
τ ∈T
S
1X
|ωj,1 − 1|
S
j=1
kZ(x )k
0
.
=oP
N 1/2
Next, note that by independence between ωs,1 and the sample we have
s (τ )] = 0 for all s, τ . Let
E[(ωs,1 − 1)rn,3
Dn (A) := C<3 (A log n)
where <3 (·) is defined in (S.6.5), and C is the constant appearing in (S.6.3).
Consider the decomposition
s
s
s
(ωs,1 − 1)Z(x0 )> rn,3
(τ ) =Z(x0 )> (ωs,1 − 1)rn,3
(τ )1{|rn,3
(τ )| ≤ Dn (A)Cω }
s
s
+ Z(x0 )> (ωs,1 − 1)rn,3
(τ )1{|rn,3
(τ )| > Dn (A)Cω }
(s)
(s)
=:Rn,1 (τ ) + Rn,2 (τ )
By (S.6.3), (S.6.8) in Theorem S.6.1, we see that
(s)
(S.4.10)
P
sup sup kRn,2 (τ )k =
6 0 ≤ 2Sn−A .
s=1,...,S
τ
(s)
Moreover, by independence of the weights ωs,1 from the sample E[Rn,1 (τ )] =
(s)
0 and by construction |Rn,1 (τ )| ≤ kZ(x0 )kCω DN (A) a.s. Apply Hoeffding’s
DISTRIBUTED INFERENCE FOR QUANTILE REGRESSION PROCESSES 31
inequality to obtain
S
1X
2St2
(s)
P
Rn,1 (τ ) ≥ tkZ(x0 )k ≤ 2 exp −
2 (A)
S
4Cω2 DN
s=1
√
Letting t = 4 log nS −1/2 DN (6)Cω and combining this with (S.4.10) shows
that
S
p
1X
s
(ωs,1 − 1)rn,3
(τ ) ≥ 4kZ(x0 )k log nS −1/2 DN (6)Cω
S
τ ∈TK
s=1
≤N 2 2n−8 + 2Sn−6 = o(1).
P
sup Z(x0 )>
Straightforward but tedious computations show that
p
log nS −1/2 DN (6) = o(N −1/2 ),
and thus the proof of (S.4.8) is complete. Next we prove (S.4.9). Note that
e (1) (τ ) − G(1) (τ )
G
N
N
1 X
−1
≤ Cω
|Z(x0 )> Jm
(τ )Z(Xis )|1{|Yis − Q(Xis ; τ )| ≤ cm (γN )}
N
i,s
1 X
Wis (τ ).
=: Cω kZ(x0 )k
N
i,s
We have
sup E[Wis (τ )] = O(cm (γN )) = o(N −1/2 kZ(x0 )k)
τ ∈T
and supτ ∈T |Wis (τ )| ≤ Cξm almost surely for some constant C < ∞ . Moreover, the Wis (τ ) are iid across i, s and
sup Var(W11 (τ )) = O(cN (γN )) = O(kZ(x0 )kN −1/2 ) = o(N −1/4 ).
τ ∈T
Hence Bernstein’s inequality and the union bound for probabilities yields
X
Wis (τ ) − E[Wis (τ )] ≥ N 1/2 (log N )−1
P sup
τ ∈TK
i,s
≤2|TK | exp −
N supτ ∈T
N (log N )−2 /2
= o(1).
Var(W11 (τ )) + Cξm N 1/2 (log N )−1 /3
Thus the proof of (S.4.9) is complete.
2
32
S.4.3. Proof of Theorem 4.4. Define the class of functions
n
n
o
o
G3 (δ) := (z, y) 7→ kzz > k1 |y − z > b + t| ≤ δ 1{kzk ≤ ξ} t ∈ R, b ∈ Rm .
Let
βes,−i (τ ) := β(τ ) − (n − 1)−1/2 Jm (τ )−1 Gns,−i ψ(·; βbs,−i (τ ), τ )
where Gs,−i
is the empirical process corresponding to sub-sample s with i’th
n
observation removed and
βbs,−j (τ ) := arg minm
b∈R
n
X
ρτ (Yis − Z>
is b).
i=1,i6=j
By Lemma S.6.3 we obtain for A ≥ 1 and some constant C
log n
P sup kβbs,−j (τ ) − βbs (τ )k ≥ CA
≤ 2Sn1−A .
n
s,τ,j
Following the arguments in the proof of Lemma S.6.3 we also find that for
some constant C1 we have almost surely
1
βes,−j (τ ) − βbs,−j (τ ) ≤ C1
+ kβbs,−j (τ ) − β(τ )k2 ,
n
so that by (S.6.18) there exists a constant C2 such that
(S.4.11)
log n
P sup kβes,−j (τ ) − βbs (τ )k ≥ C2 A
≤ Sn1−A .
n
s,τ,j
Moreover, the arguments in the proof of Theorem S.6.1 show that
sup E[βes,−i (τ )] − β(τ )
s,i,τ
(S.4.12)
h
i
log n
= sup E n−1/2 Jm (τ )−1 Gns,−i ψ(·; βbs,−i (τ ), τ ) = O
.
n
s,i,τ
Define the events
n
o
es,−j (τ ) − βbs (τ )k ≤ C2 4 log n ,
Ωs,j
:=
k
β
1,n
n
Ω1,n := ∩s,j Ωs,j
1,n .
DISTRIBUTED INFERENCE FOR QUANTILE REGRESSION PROCESSES 33
and observe that for δn := 4C2 logn n
n
h n
o
n
oi
X
1
> bs
> es,−i
Zis Z>
1
|Y
−
Z
β
(τ
)|
≤
h
−
1
|Y
−
Z
β
(τ
)|
≤
h
is
n
is
n
is
is
is
2nhn
i=1
n
n
o ξ2
1X
>
1 C
kZis Z>
k1
|Y
−
Z
b
−
t|
≤
ξδ
is
n +
is
is
hn Ω1,n
b∈Rm ,t∈R n i=1
h
n
oi
1
>
≤
sup E kZ1s Z>
k1
|Y
−
Z
b
−
t|
≤
ξδ
1s
n
1s
1s
hn b∈Rm ,t∈R
≤
1
hn
+
=
sup
1
ξ2
kPn,s − P kG3 (δn ) +
1 C
hn
hn Ω1,n
1
ξ2
kPn,s − P kG3 (ξδn ) + O(h−1
δ
)
+
1 C .
n
n
hn
hn Ω1,n
Standard arguments show that G3 (ξδn ) satisfies (S.8.1) with fixed constant
V and thus by (S.8.2) we have
sup EkPn,s − P kG3 (ξδn ) = O
s
log n
n
.
This implies
P
sup J m (τ )−
τ
S
n
n
o
log n
1X 1 X
> es,−i
.
Zis Z>
1
|Y
−Z
β
(τ
)|
≤
h
=
O
is
n
P
is
is
S
2nhn
nhn
s=1
i=1
Next note that by a Taylor expansion we have
FY |X (a + hn |x) − FY |X (a − hn |x) = 2hn fY |X (a|x) +
h3n 00
f
(a|x) + o(h3n )
3 Y |X
34
where the remainder term is uniform in a ∈ R. Thus we have by (S.4.12)
h 1
n
oi
> es,−1
E
Z1s Z>
1
|Y
−
Z
β
(τ
)|
≤
h
1s
n
1s
1s
2hn
Z
h
1
=
Z(x)Z(x)> E FY |X (Z(x)> βes,−1 (τ ) + hn |x)
2hn
i
− FY |X (Z(x)> βes,−1 (τ ) − hn |x) dP X (x)
Z
h
=
Z(x)Z(x)> E fY |X (Z(x)> βes,−1 (τ )|x)
i
h2n 00
fY |X (Z(x)> βes,−1 (τ )|x) dP X (x) + o(h2n )
6
Z
i
h
h2
Z(x)Z(x)> fY |X (Z(x)> β(τ )|x) + n fY00 |X (Z(x)> β(τ )|x) dP X (x)
=
6
log n
+O
+ o(h2n ).
n
+
Similar but simpler computations show that for
1
n
o
>
Ws (b) := vec
Z1s Z>
1
|Y
−
Z
b|
≤
h
1s
n
1s
1s
2hn
we have uniformly in s
Var(Ws (βes,−1 (τ ))) = Var(Ws (β(τ ))) + o(h−1
n ).
This completes the proof.
2
APPENDIX S.5: PROOFS FOR APPENDIX S.2
S.5.1. Proof of Theorem S.2.1. From Theorem S.6.1 we obtain the
representation
βbs (τ ) − βN (τ ) = −n−1/2 Jm (τ )−1 Gsn (ψ(·; βN (τ ), τ ))
s
s
s
s
+ rn,1
(τ ) + rn,2
(τ ) + rn,3
(τ ) + rn,4
(τ ).
1 , ..., r S are i.i.d. for each j = 1, ..., 4 and
where the quantities rn,j
n,j
Gsn (ψ(·; βN (τ ), τ )) = n−1/2
n
X
i=1
ψ(Yis , Z(Xis ); βN (τ ), τ ).
DISTRIBUTED INFERENCE FOR QUANTILE REGRESSION PROCESSES 35
Letting
(1)
rN (τ )
S
1 X s
s
s
:=
rn,1 (τ ) + rn,2 (τ ) + rn,4 (τ )
S
s=1
we find that the bound in (S.2.3) is a direct consequence of Theorem S.6.1,
2 = o(g ) by the
the union probability bound and the observation that ξm gN
N
−1 ).
assumption gN = o(ξm
Next, let
S
1X s
(2)
rn,3 (τ ).
rN (τ ) :=
S
s=1
Define
s (τ )
ren,3
:=
s (τ )
rn,3
−
s (τ )]
E[rn,3
and consider the decomposition
s
s
s
rn,3
(τ ) = ren,3
1 ke
rn,3 (τ )k ≤ Dn (A)
+
s
ren,3
(τ )1
3
X
s
(s)
s
ke
rn,3 (τ )k > Dn (A) + E[rn,3 (τ )] =:
Rn,j (τ )
j=1
where, for a sufficiently large constant C,
mξm log n
2
Dn (A) := C<3 (A log n) + C3,2 ξm gN
ξm +
,
n
<3 (·) is defined in (S.6.5) and C3,2 denotes the constant appearing in (S.6.8).
By (S.6.3), (S.6.8) in Theorem S.6.1, we see that
(s)
(S.5.1)
P
sup sup kRn,2 (τ )k =
6 0 ≤ 2Sn−A ;
s=1,...,S
τ
and by (S.6.8) in Theorem S.6.1,
(S.5.2)
mξm log n
(s)
2
sup kRn,3 (τ )k ≤ C3,2 ξm gN
ξm +
.
n
τ
(s)
Next we deal with Rn,1 (τ ). Consider the function g : (Rm )S → R,
S
1X
g(x1 , ..., xS ) :=
xs .
S
s=1
Direct computations show that
|g(x1 , ..., xS ) − g(x1 , ..., x0i , ..., xS )| ≤
kx0i − xi k
.
S
36
(s)
Note that by construction kRn,1 (τ )k ≤ Dn (A) almost surely. Apply McDiarmid’s inequality (Lemma 1.2 on page 149 in McDiarmid (1989)) to
(1)
(S)
g(Rn,1 (τ ), ..., Rn,1 (τ )) to obtain for all t > 0
sup P
τ
√
S
1X
−St2
(s)
Rn,1 (τ ) ≥ t ≤ 2 exp
S
2Dn2 (A)
s=1
Letting t = 2κn S −1/2 Dn (A) and combining this with (S.5.1) and (S.5.2)
shows that
√
mξm log n
2
(2)
2
sup P rN (τ ) > 2κn S −1/2 Dn (A)+C3,2 ξm gN
ξm +
≤ 2Sn−A +2e−κn .
n
τ
Finally, from an elementary but tedious calculation and the definition of
<3 (A log n) in (S.6.5), we obtain
κn S −1/2 <3 (A log n) . A
mξ 2 (log n)3 1/4
κn m 1/2
m
1/2
(ξ
g
log
n)
+
.
m
N
n
S 1/2 n
Thus the proof of Theorem S.2.1 is complete.
S.5.2. Proof of Theorem S.2.4. The proof proceeds similarly to the
proof of Theorem S.2.1. From Theorem S.6.2 we obtain the representation
bs
u>
N (β (τ ) − γN (τ ))
=−
+
−1
n−1 u>
N Jm (γN (τ ))
n
X
Zsi (1{Yis
≤
(Zsi )> γN (τ )}
i=1
>
−1
uN Jm (γN (τ )) E[Z(FY |X (Z> γN (τ )|X)
− τ) +
u>
N
4
X
s
rn,k
(τ )
k=1
− τ )].
1 (τ ), ..., r S (τ ) are i.i.d. for each j = 1, ..., 4 and
where the quantities rn,j
n,j
Gsn (ψ(·; γN (τ ), τ ))
=n
−1/2
n
X
ψ(Yis , Z(Xis ); γN (τ ), τ ).
i=1
Next, note that
Z(FY |X (Z> γN (τ )|X) − τ ) = ZfY |X (Q(X; τ )|X)(Z> γN (τ ) − Q(X; τ ))
1
+ ZfY0 |X (ζn (X, τ )|X)(Z> γN (τ ) − Q(X; τ ))2
2
DISTRIBUTED INFERENCE FOR QUANTILE REGRESSION PROCESSES 37
where ζn (X, τ ) is a value between Z(X)> γN (τ ) and Q(X; τ ). From the definition of γN as (3.8), we obtain E[ZfY |X (Q(X; τ )|X)(Z> γN (τ )−Q(X; τ ))] =
0 so that
−1
>
2
e
e
sup u>
N Jm (γN (τ )) E[Z(FY |X (Z γN (τ )|X) − τ )] . E(uN , γN )cn (γN ).
τ ∈T
Letting
(1)
rN (τ, uN )
S
1 > X s
s
s
:=
u
rn,1 (τ ) + rn,2 (τ ) + rn,4 (τ )
S N
s=1
−1
>
+ u>
N Jm (γN (τ )) E[Z(FY |X (Z γN (τ )|X) − τ )]
(1)
we find that the bound on rN (τ, uN ) is a direct consequence of Theorem S.6.2. Next, let
(2)
rN (τ, uN )
S
1X > s
:=
uN rn,3 (τ ).
S
s=1
s (τ ) := r s (τ ) − E[r s (τ )] and consider the decomposition
Define ren,3
n,3
n,3
n
s
s
rn,3
(τ ) = ren,3
1
+
=:
sup
uN ∈SIm−1
s
ren,3
(τ )1
3
X
n
s
|u>
en,3
(τ )| ≤ Dn(L) (A)
Nr
sup
uN ∈SIm−1
(s)
Rn,j (τ )
s
|u>
en,3
(τ )|
Nr
>
o
Dn(L) (A)
o
1
+ E[rn,3
(τ )]
j=1
where for a constant A > 0, and a sufficiently large constant C
(L)
Dn(L) (A) := C<3 (A log n)+C3,2
2
2
e N , γN ) c4 (γN )+ ξm (log n) ,
E(u
m
n
uN ∈S m−1
sup
I
(L)
<3 is defined in (S.6.13), C3,2 denotes the constant appearing in (S.6.15).
By (S.6.11) in Theorem S.6.2,
(s)
(S.5.3)
P
sup sup sup |u>
6 0 ≤ 2Sn−A .
N Rn,2 (τ )| =
s=1,...,S
τ
uN ∈SIm−1
38
Applying (S.6.15) in Theorem S.6.2 yields
(S.5.4)
(s)
sup
sup
τ
uN ∈SIm−1
|u>
N Rn,3 (τ )| ≤ C3,2
2
2
e N , γN ) c4 (γN ) + ξm (log n) .
E(u
m
n
uN ∈S m−1
sup
I
Next consider he function g : (Rm )S → R,
g(x1 , ..., xS ) :=
sup
uN ∈SIm−1
u>
N
S
S
1 X
(I)
1X
xs =
xs
.
S
S
s=1
s=1
The reverse triangle inequality shows that
|g(x1 , ..., xS ) − g(x1 , ..., x0i , ..., xS )| ≤
sup
uN ∈SIm−1
(s)
0
|u>
N (xi − xi )|
.
S
(L)
Note that by construction supuN ∈S m−1 |u>
N Rn,1 (τ )| ≤ Dn (A) almost surely.
I
Apply McDiarmid’s inequality (Lemma 1.2 on page 149 in McDiarmid (1989))
(1)
(S)
to g(Rn,1 (τ ), ..., Rn,1 (τ )) to obtain for all t > 0
sup P
τ
1 > X (s)
Rn,1 (τ ) ≥ t ≤ 2 exp
uN
S
S
sup
uN ∈SIm−1
s=1
−St2
(L)
2(Dn (A))2
.
√
(L)
Letting t = 2κn S −1/2 Dn (A) and combining this with (S.5.3)-(S.5.4)
shows that
√
(2)
sup P
sup
rN (τ, uN ) > 2κn S −1/2 Dn(L) (A)
τ
uN ∈SIm−1
+ C3,2
2
2
e N , γN ) c4 (γN ) + ξm (log n)
E(u
m
n
uN ∈S m−1
sup
I
2
≤ 2Sn−A + 2e−κn .
Thus the proof of Theorem S.2.4 is complete.
APPENDIX S.6: REFINED BAHADUR REPRESENTATIONS FOR
SUB-SAMPLES AND THEIR PROOFS
In this section, we consider a triangular array {(Xis , Yis )}ni=1 from the
s-th group. In order to keep the notation simple, the sample will be denoted by {(Xi , Yi )}ni=1 . In addition to the notation introduced in the main
DISTRIBUTED INFERENCE FOR QUANTILE REGRESSION PROCESSES 39
part of the manuscript, we will make use of the following notation throughout the appendix. Denote the empirical measure of
R n samples (Yi , Zi ) by
Pn . For a functionR x 7→ f (x) define Gn (f ) := n1/2 f (x)(dPn (x) − dP (x))
and kf kLp (P ) = ( |f (x)|p dP (x))1/p for 0 < p < ∞. For a class of functions G, let kPn − P kG := supf ∈G |Pn f − P f |. For any > 0, the covering
number N (, G, Lp ) is the minimal number of balls of radius (under Lp norm) that is needed to cover G. The bracketing number N[ ] (, G, Lp ) is
the minimal number of -brackets that is needed to cover G. An -bracket
refers to a pair of functions within an distance: ku − lkLp < . Throughout the proofs, C, C1 , Cj etc. will denote constants which do not depend
on n but may have different values in different lines. For
τ ∈ T , define
> b} − τ ) and µ(b, τ ) := E ψ(Y , Z ; b, τ ) =
ψ(Y
,
Z
;
b,
τ
)
:=
Z
(1{Y
≤
Z
i
i
i
i
i
i
i
E Zi FY |X (Z>
.
i b|X) − τ
The aim of this section is to provide precise bounds on the remainder
terms in the Bahadur representation of the estimator β̆(τ ) which is defined
as
n
X
β̆(τ ) := arg minm
ρτ (Yi − b> Z(Xi )), τ ∈ T .
b∈R
i=1
Theorem S.6.1. Suppose Conditions (A1)-(A3) hold. Assume that ad2 log n = o(n). Then, for any β (·) satisfying g (β ) =
ditionally mξm
N
N
N
−1
o(ξm ) and cm (βN ) = o(1), we have
(S.6.1) β̆(τ ) − βN (τ ) = −n−1/2 Jm (τ )−1 Gn (ψ(·; βN (τ ), τ ))
+ rn,1 (τ ) + rn,2 (τ ) + rn,3 (τ ) + rn,4 (τ ).
The remainder terms rn,j ’s can be bounded as follows:
sup krn,1 (τ )k ≤ C1
(S.6.2)
τ
mξm
n
a.s.
2 ,
for a constant C1 independent of n. Moreover, we have for any κn n/ξm
all sufficiently large n, and a constant C independent of n
(S.6.3)
P sup krn,j (τ )k ≤ C<j (κn ) ≥ 1 − 2e−κn , j = 2, 3, 4,
τ
40
where
(S.6.4)
<2 (κn ) := ξm
(S.6.5)
<3 (κn ) :=
(S.6.6)
m
n
log n
1/2
m log n 1/2
n
+
+ n−1/2 κn1/2 + gN
κ 1/2
n
n
+ gN
2
,
1/2 mξ log n 1/2 ξ κ 1/2
m
m n
+
,
n
n
m
1/2 κ1/2
n
<4 (κn ) := cm (βN )
+ 1/2 + gN .
log n
n
n
Additionally,
sup krn,3 (τ )k ≤ C3,1 ξm ,
(S.6.7)
a.s.,
τ
and it holds that for sufficiently large n,
(S.6.8)
sup kE[rn,3 (τ )]k ≤
τ
2
C3,2 ξm
2
gN
m log n
+
n
where C3,1 , C3,2 are constants that are independent of n.
The statements in (S.6.2)-(S.6.6) are from Theorem 5.1 of Chao et al.
(2017) and reproduced here for the sake of completeness. The bound in
(S.6.8) is new and crucial for the ABR in Theorem S.2.1 in the main manuscript.
See Section S.6.1 in the supplementary material for a proof.
The next result provides a refined version of the previous theorem for
local basis functions that satisfy Condition (L).
Theorem S.6.2. Suppose Conditions (A1)-(A3) and (L) hold. Assume
2 log n = o(n). Assume that the set I consists of at
that additionally mξm
most L ≥ 1 consecutive integers. Then, for γN defined in (3.8) we have
β̆(τ ) − γN (τ ) = −n
−1/2
Jem (γN (τ ))−1 Gn (ψ(·; γN (τ ), τ )) +
4
X
(L)
rn,k (τ ).
k=1
where remainder terms rn,j ’s can be bounded as follows (C1 , C4 are constants
that do not depend on n):
(S.6.9)
(L)
sup sup |a> rn,1 (τ )| ≤ C1
a∈SIm−1
(S.6.10)
τ
(L)
sup sup |a> rn,4 (τ )| ≤
a∈SIm−1
τ
ξm log n
n
a.s.,
1
e γN )
+ C4 c2m (γN ) sup E(a,
n
m−1
a∈S
I
a.s.
DISTRIBUTED INFERENCE FOR QUANTILE REGRESSION PROCESSES 41
2 , all sufficiently large n, and a conMoreover, we have for any κn n/ξm
stant C independent of n
(L)
(L)
(S.6.11) P
sup sup |a> rn,j (τ )| ≤ C<j (κn ) ≥ 1 − n2 e−κn , j = 2, 3
τ
a∈SIm−1
where
(S.6.12)
(L)
<2 (κn ) :=
1/2
2
e γN ) ξm (log n + κn ) + c2 (γN ) ,
sup E(a,
m
n1/2
a∈S m−1
I
(S.6.13)
(L)
<3 (κn )
κn ∨ log n ξm (κn ∨ log n)3/2
:= cm (γN )
+
,
n1/2
n3/4
1/2
1/2
1/2
e γN ) := supτ ∈T E[a> Jem (γN (τ ))−1 Z]. Additionally,
and E(a,
(S.6.14)
(L)
sup sup |a> rn,3 (τ )| ≤ C3,1 ξm
τ
a.s.
a∈SIm−1
and for sufficiently large n we have
(S.6.15)
2
2
(L)
e γN ) c4 (γN ) + ξm (log n) .
sup sup |a> E[rn,3 (τ )]| ≤ C3,2 sup E(a,
m
n
a∈S m−1 τ
a∈S m−1
I
I
where C3,1 , C3,2 are constants that do not depend on n.
Similarly to the setting in Theorem S.6.1, the statements in (S.6.9)(S.6.13) are proved in Theorem 5.2 of Chao et al. (2017) and reproduced
here for the sake of completeness. The bound in (S.6.15) is new and crucial
for the ABR in Theorem S.2.4.
In both proofs, we will repeatedly use the following leave-one-out estimator
X
(S.6.16) β̆ (−j) (τ ) := arg minm
ρτ (Yi − b> Z(Xi )), τ ∈ T , j = 1, ..., n
b∈R
i6=j
S.6.1. Proof of Theorem S.6.1. Observe the representation
β̆(τ ) − βN (τ ) = −n−1/2 Jm (τ )−1 Gn (ψ(·; βN (τ ), τ ))
+ rn,1 (τ ) + rn,2 (τ ) + rn,3 (τ ) + rn,4 (τ )
42
where
rn,1 (τ ) := Jem (βN (τ ))−1 Pn ψ(·; β̆(τ ), τ ),
rn,2 (τ ) := −Jem (βN (τ ))−1 µ(β̆(τ ), τ ) − µ(βN (τ ), τ )
−Jem (βN (τ ))(β̆(τ ) − βN (τ )) ,
rn,3 (τ ) := −n−1/2 Jem (βN (τ ))−1 Gn (ψ(·; β̆(τ ), τ )) − Gn (ψ(·; βN (τ ), τ )) ,
rn,4 (τ ) := −n−1/2 (Jm (τ )−1 − Jem (βN (τ ))−1 )Gn (ψ(·; βN (τ ), τ ))
−Jem (βN (τ ))−1 µ(βN (τ ), τ ).
The bounds in (S.6.2)-(S.6.6) follow similarly as in the proof of Theorem 5.1
in Chao et al. (2017). The bound in (S.6.7) follows from the definition of
rn,3 . It remains to prove (S.6.8). Note that the expectation of rn,3 (τ ) is not
zero in general. Write
rn,3 (τ )
h
= − Jem (βN (τ ))−1 − n−1/2 Gn ψ(·; βN (τ ), τ )
+ n−1
n
X
ψ(Yj , Zj ; β̆ (−j) (τ ), τ ) − µ(β̆ (−j) (τ ), τ )
j=1
− n−1
n
X
ψ(Yj , Zj ; β̆ (−j) (τ ), τ ) − ψ(Yj , Zj ; β̆(τ ), τ )
j=1
− µ(β̆ (−j) (τ ), τ ) + µ(β̆(τ ), τ )
i
.
By definition of β̆ (−j) (τ ) in (S.6.16), the expectation of the first line in the
above representation is zero. To bound the expectation of the second line,
observe that for
m log n
2
e m gN
δn := Cξ
(βN ) +
n
DISTRIBUTED INFERENCE FOR QUANTILE REGRESSION PROCESSES 43
e sufficiently large we have
with C
n
h
i
X
−1
sup E n
ψ(Yj , Zj ; β̆ (−j) (τ ), τ ) − ψ(Yj , Zj ; β̆(τ ), τ )
τ ∈T
j=1
n
h
X
i
(−j)
>
= sup sup E n−1
a> Zj 1{Yj ≤ Z>
β̆
(τ
)}
−
1{Y
≤
Z
β̆(τ
)}
j
j
j
τ ∈T kak=1
j=1
≤ sup sup n−1
τ ∈T kak=1
n
X
h
(−j)
E |a> Zj |1 |Yj − Z>
(τ )| ≤ ξm δn
j β̆
j
n
X
h
(−j)
(τ ) + ξm δn |Xj )
E |a> Zj | FY |X (Z>
j β̆
j=1
i
> (−j)
(τ ) ≤ ξm δn
× 1 sup Z>
j β̆(τ ) − Zj β̆
j
> (−j)
>
ξ
δ
+ ξm sup P sup Z>
β̆(τ
)
−
Z
β̆
(τ
)
m
n
j
j
τ ∈T
≤ sup sup n−1
τ ∈T kak=1
j=1
(−j)
− FY |X (Z>
(τ ) − ξm δn |Xj )
j β̆
> (−j)
>
ξ
δ
+ ξm P sup sup Z>
β̆(τ
)
−
Z
β̆
(τ
)
m n
j
j
τ ∈T j
> (−j)
≤ξm P sup sup Z>
β̆(τ
)
−
Z
β̆
(τ
)
>
ξ
δ
m n
j
j
τ ∈T
i
j
+ 2ξm δn f¯ sup n−1
kak=1
n
X
E |a> Zj |
j=1
> (−j)
≤ 2ξm δn f¯λmax (E[ZZ> ])1/2 + ξm P sup sup Z>
>
ξ
δ
β̆(τ
)
−
Z
β̆
(τ
)
m
n
j
j
2
2
≤C3,2 ξm
gN
(βN ) +
τ ∈T
m log n
j
n
for all sufficiently large n and a constant C3,2 independent of τ, n. Here, the
last line follows from Lemma S.6.3 and the definition of δn provided that
e in the definition of δn is sufficiently large. Thus it remains to bound the
C
expectation of µ(β̆ (−j) (τ ), τ ) − µ(β̆(τ ), τ ). A Taylor expansion shows that
sup kµ(b1 , τ ) − µ(b2 , τ )k ≤ sup E |a> Z|1 |Y − Z> b1 | ≤ |Z> (b1 − b2 )|
τ
kak=1
≤ 2f¯ sup E |a> Z||Z> (b1 − b2 )|
kak=1
≤ 2f¯λmax (E[ZZ> ])kb1 − b2 k,
44
where the last inequality follows by Cauchy-Schwarz inequality. On the other
hand kµ(b1 , τ ) − µ(b2 , τ )k ≤ 2ξm . Thus we have for any δn > 0,
sup E µ(β̆ (−j) (τ ), τ ) − µ(β̆(τ ), τ )
τ,j
≤ 2f¯λmax (E[ZZ> ])δn + 2ξm P sup kβ̆ (−j) (τ ) − β̆(τ )k > δn .
τ,j
Choosing δn as above completes the proof.
Lemma S.6.3. Under the assumptions of Theorem S.6.1 we have for any
2 , all sufficiently large n, and a constant C independent of n we
κn n/ξm
have for β̆ (−j) defined in (S.6.16)
ξm κn mξm log n
2
P max sup kβ̆ (−j) (τ )−β̆(τ )k ≥ C gN
(βN )ξm +
+
≤ 2ne−κn .
j=1,...,n τ
n
n
Proof of Lemma S.6.3 P
√
(−j)
(−j)
(−j)
Let Pn (f ) := (n − 1)−1 i6=j f (Xi , Yi ), Gn (f ) := n − 1(Pn (f ) −
Ef (X1 , Y1 )). Similarly to the decomposition considered previously (S.6.1),
we have
(S.6.17)
4
X
(−j)
β̆ (−j) (τ ) − βN (τ ) = −(n − 1)−1/2 Jm (τ )−1 G(−j)
(ψ(·;
β
(τ
),
τ
))
+
rn,k (τ )
N
n
k=1
where
(−j)
rn,1 (τ ) := Jem (βN (τ ))−1 P(−j)
ψ(·; β̆ (−j) (τ ), τ ),
n
(−j)
rn,2 (τ ) := −Jem (βN (τ ))−1 µ(β̆ (−j) (τ ), τ ) − µ(βN (τ ), τ )
−Jem (βN (τ ))(β̆ (−j) (τ ) − βN (τ )) ,
−Jem (βN (τ ))−1 (−j)
√
Gn (ψ(·; β̆ (−j) (τ ), τ )) − G(−j)
(ψ(·;
β
(τ
),
τ
))
,
N
n
n−1
(−j)
rn,4 (τ ) := −(n − 1)−1/2 (Jm (τ )−1 − Jem (βN (τ ))−1 )G(−j)
(ψ(·; βN (τ ), τ ))
n
(−j)
rn,3 (τ ) :=
−Jem (βN (τ ))−1 µ(βN (τ ), τ ).
By similar arguments as in the proof for rn,2 in Theorem 5.1 in Chao et al.
(j)
(2017) there exists C independent of n such that P (Ω1,n (C)) ≤ e−κn , j =
1, ..., n where we defined the event
(S.6.18)
(j)
Ω1,n (C)
:=
n
sup kβ̆
τ
(−j)
m log n 1/2 o
κn
(τ ) − βN (τ )k ≥ C gN (βN ) + 1/2 +
.
n
n
1/2
DISTRIBUTED INFERENCE FOR QUANTILE REGRESSION PROCESSES 45
Combining (S.6.1) and (S.6.17) we find that, almost surely,
sup sup β̆
(−j)
(τ ) − β̆(τ ) −
τ
j
4
X
(−j)
rn,k (τ ) − rn,k (τ )
k=1
.
ξm
.
n
As in the proof of Lemma 34 on page 106 in Belloni et al. (2017), arguments
from linear optimization lead to the bound
(−j)
sup sup krn,1 (τ ) − rn,1 (τ )k ≤
j
τ
Direct calculations show that
mξm
2
inf τ λmin (Jem (βN (τ ))) n − 1
(−j)
sup sup krn,4 (τ ) − rn,4 (τ )k ≤ C1
j
τ
a.s.
ξm
,
n
and from Lemma S.6.5 we obtain for a constant C independent of j, τ, n
(−j)
krn,2 (τ ) − rn,2 (τ )k ≤ Cξm kβ̆ (−j) (τ ) − βN (τ )k2 + kβ̆(τ ) − βN (τ )k2 .
The probability of the event
n
Ω1,n (C) := sup kβ̆ (−j) (τ ) − βN (τ )k2 + kβ̆(τ ) − βN (τ )k2
τ,j
κn m log n o
2
≥ C gN
(βN ) +
+
n
n
can be bounded by (n + 1)e−κn if C is chosen suitably, this follows from the
(j)
bound P (Ω1,n (C)) ≤ e−κn in (S.6.18). Finally, define
(S.6.19)
n
G2 (δ) := (z, y) 7→ a> z(1{y ≤ z > b1 } − 1{y ≤ z > b2 })1{kzk ≤ ξm }
o
b1 , b2 ∈ Rm , kb1 − b2 k ≤ δ, a ∈ S m−1 .
Letting δn := supj supτ kβ̆(τ )(−j) − β̆(τ )k, we have
(−j)
sup sup krn,3 (τ ) − rn,3 (τ )k ≤ C3
j
τ
ξ
m
n
+ kPn − P kG2 (δn ) .
Summarizing the bounds obtained so far we find for a constant C4 independent of n
mξ
m
δ n ≤ C4
+kPn −P kG2 (δn ) +sup ξm kβ̆ (−j) (τ )−βN (τ )k2 +kβ̆(τ )−βN (τ )k2 .
n
τ,j
46
Consider the event
Ω2,n (C) :=
n
o
kPn − P kG2 (t)
≥C ,
χn (t, κn )
0≤t≤1
sup
where
χn (t, κn ) =
1/2 1/2
ξm
t
1/2 mξ
ξm tκn 1/2 ξm κn
m
log(ξm ∨n)
+
log(ξm ∨n)+
+
n
n
n
n
m
defined as (S.6.22) in Lemma S.6.4. It follows from (S.6.21) that P (Ω2,n (C)) ≤
e−κn log2 n if we choose C suitably. Thus, δn satisfies the inequality
κn m log n
2
δn ≤ C8 ξm gN
(βN ) +
+
+ C8 χn (δn , κn )
n
n
κn m log n
mξm
ξ m κn
2
≤ C8 ξm gN
(βN ) +
+
+ C8
log(ξm ∨ n) + C8
n
n
n
n
ξm κn 1/2
1/2 ξm m
+ C9 δn
log(ξm ∨ n) +
n
n
1/2
:= an + δn bn
with probability at least 1 − 2ne−κn (note that (n + 1)e−κn + log2 ne−κn ≤
2ne−κn ). A simple calculation shows that the inequality 0 ≤ δn ≤ an +
1/2
bn δn implies 0 ≤ δn ≤ 4 max(an , b2n ). Straightforward calculations show
that under the assumption of the theorem we have
2
an + b2n . gN
(βN )ξm +
ξm κn mξm log n
+
n
n
provided that κn → ∞. Thus we have for some constant C independent of
n
ξm κn mξm log n
2
P δn ≥ C gN
(βN )ξm +
+
≤ 2ne−κn .
n
n
This completes the proof.
Lemma S.6.4. Consider the class of functions G2 (δn ) defined in (S.6.19).
Under assumptions (A1)-(A3) and ξm δn n−1 , ξm = O(nb ) for some b <
∞, we have for some constant C independent of n, sufficiently large n and
arbitrary κn > 0,
(S.6.20)
P kPn − P kG2 (δn ) ≥ Cχn (δn , κn ) ≤ e−κn
and for κn ≥ 1
(S.6.21)
P
kPn − P kG2 (t)
≥ 2C ≤ e−κn log2 n
χn (t, κn )
0≤t≤1
sup
DISTRIBUTED INFERENCE FOR QUANTILE REGRESSION PROCESSES 47
where
(S.6.22)
χn (t, κn ) :=
mξ
1/2 mξ
ξ κ t 1/2 ξ κ
m
m n
m n
log(ξm ∨ n)t
+
log(ξm ∨ n) +
.
+
n
n
n
m
n
Proof of Lemma S.6.4. For a proof of (S.6.20) see Lemma C.3 in the
Appendix of Chao et al. (2017). For a proof of (S.6.21), observe that for
κn ≥ 1 we have χn (t/2, κn ) ≥ χn (t, κn )/2, χn (0, κn ) ≥ χn (n−1 , κn )/2, and
thus
kPn − P kG2 (t)
χn (t, κn )
0≤t≤1
kPn − P kG2 (t)
kPn − P kG2 (t)
≤
sup
∨
max
sup
χn (t, κn )
χn (t, κn )
k:n−1 ≤2−k ≤1 2−k−1 ≤t≤2−k
0≤t≤n−1
2kPn − P kG2 (2−k ) 2kPn − P kG2 (n−1 )
∨
≤
max
χn (n−1 , κn )
χn (2−k , κn )
k:n−1 ≤2−k ≤1
sup
Now the set {k : n−1 ≤ 2−k ≤ 1} contains not more than log2 n elements,
and (S.6.21) thus follows from (S.6.20).
Lemma S.6.5.
sup
Under assumptions (A1)-(A3) we have
sup
τ ∈T kb−βN (τ )k≤ω
kµ(b, τ ) − µ(βN (τ ), τ ) − Jem (βN (τ ))(b − βN (τ ))k
≤ λmax (E[ZZ> ])f 0 ω 2 ξm ,
Proof of Lemma S.6.5 see the proof of Lemma C.1 in the Appendix of
Chao et al. (2017).
Lemma S.6.6.
n
Let assumptions (A1)-(A3) hold. Then, for any t > 1
sup kβ̆(τ ) − βN (τ )k ≤
τ ∈T
2t(sn,1 + gN (βN )) o
inf τ ∈T λmin (Jem (βN (τ )))
n
inf τ ∈T λ2min (Jem (βN (τ ))) o
⊇ (sn,1 + gN (βN )) <
.
4tξm f 0 λ2max (E[ZZ> ])
where sn,1 := kPn − P kG1 and
n
o
G1 := (z, y) 7→ a> z(1{y ≤ z > b}−τ )1{kzk ≤ ξm } τ ∈ T , b ∈ Rm , a ∈ S m
Proof of Lemma S.6.6. See the proof of Lemma C.2 in the Appendix of
Chao et al. (2017).
48
S.6.2. Proof of Theorem S.6.2. We begin by introducing some notation and useful preliminary results which will be used throughout this
section. For a vector v ∈ Rm and a set I ⊂ {1, ..., m} let v(I) ∈ Rm denote
the vector that has entries vi for i ∈ I and zero otherwise. For a vector
a ∈ Rm let kak0 denote the number of non-zero entries in this vector, let ka
denote the position of the first non-zero entry and define
(S.6.23)
Ia (D) := i ∈ {1, ..., m} : |i − ka | ≤ kak0 + D .
Under (A1)-(A3) and (L) we obtain by similar arguments as in Lemma
A.1 in Chao et al. (2017)
(S.6.24)
−1
−1
(γN (τ )))(Ia (D)) k ≤ Ckak∞ kak0 λD
(γN (τ )) − (a> Jem
ka> Jem
for constants λ ∈ (0, 1), C > 0 independent of n, τ .
In the sequel, we will make use of the following decomposition which holds
for any b ∈ Rm such that Jem (b) is invertible
(S.6.25) β̆(τ ) − b =
where
−1
−n−1/2 Jem
(b)Gn (ψ(·; b, τ ))
−1
+ Jem
(b)
4
X
Rn,k (τ, b)
k=1
(S.6.26) Rn,1 (τ, b) := Pn ψ(·; β̆(τ ), τ ),
(S.6.27) Rn,2 (τ, b) := − µ(β̆(τ ), τ ) − µ(b, τ ) − Jem (b)(β̆(τ ) − b) ,
(S.6.28) Rn,3 (τ, b) := −n−1/2 Gn (ψ(·; β̆(τ ), τ )) − Gn (ψ(·; b, τ )) ,
(S.6.29) Rn,4 (τ, b) := −µ(b, τ ).
(L)
−1 (γ (τ ))R
Let rn,k (τ ) := Jem
N
n,k (τ, γN (τ )), k = 1, ..., 4. The bounds in
(S.6.9)-(S.6.10) follow similarly to the bounds (5.6)-(5.9) in Theorem 5.2
in Chao et al. (2017). The bound in (S.6.14) follows from the definition of
rn,3 . Thus it remains to establish (S.6.15). Write
h
Rn,3 (τ, γN (τ )) = − − n−1/2 Gn ψ(·; γN (τ ), τ )
+n
−1
n
X
j=1
− n−1
n
X
j=1
ψ(Yj , Zj ; β̆ (−j) (τ ), τ ) − µ(β̆ (−j) (τ ), τ )
ψ(Yj , Zj ; β̆ (−j) (τ ), τ ) − ψ(Yj , Zj ; β̆(τ ), τ )
i
− µ(β̆ (−j) (τ ), τ ) + µ(β̆(τ ), τ ) .
DISTRIBUTED INFERENCE FOR QUANTILE REGRESSION PROCESSES 49
Define
2
2
e cm (γN )4 + ξm (log n) .
δen := C
n
and observe that for any a ∈ SIm−1 , τ ∈ T ,
n
h
i
X
−1
−1
e
(a Jm (γN (τ )) )E n
ψ(Yj , Zj ; β̆ (−j) (τ ), τ ) − ψ(Yj , Zj ; β̆(τ ), τ )
>
j=1
n
h
i
X
(−j)
>
= E n−1
(a> Jem (γN (τ ))−1 )Zj (1{Yi ≤ Z>
β̆
(τ
)}
−
1{Y
≤
Z
β̆(τ
)})
i
j
j
. n−1
j=1
n
X h
j=1
n
o
(−j)
(τ )| ≤ δen
E |(a> Jem (γN (τ ))−1 )Zj |1 |Yj − Z>
j β̆
oi
n
> (−j)
× 1 sup Z>
(τ ) ≤ δen
j β̆(τ ) − Zj β̆
j
> (−j)
+ ξm P sup Z>
(τ ) > δen
j β̆(τ ) − Zj β̆
j
≤ n−1
n
X
j=1
h
(−j)
E |(a> Jem (γN (τ ))−1 )Zj | FY |X (Z>
(τ ) + δen |Xj )
j β̆
(−j)
− FY |X (Z>
(τ ) − δen |Xj )
j β̆
> (−j)
+ ξm P sup Z>
(τ ) > δen
j β̆(τ ) − Zj β̆
≤ 2δen f¯n−1
j
n
X
j=1
i
h
i
> (−j)
E |(a> Jem (γN (τ ))−1 )Zj | + ξm P sup Z>
(τ ) > δen
j β̆(τ ) − Zj β̆
j
2
2
e γN ) c4 (γN ) + ξm (log n)
≤C3,2 E(a,
m
n
for all sufficiently large n and a constant C3,2 independent of τ, n, a. Here, the
e in the definition of δen chosen suffilast line follows from Lemma S.6.7 with C
ciently large. Thus it remains to bound (a> Jem (γN (τ ))−1 )E[µ(β̆ (−j) (τ ), τ ) −
µ(β̆(τ ), τ )]. A Taylor expansion shows that for b1 , b2 ∈ Rm
sup |(a> Jem (γN (τ ))−1 )µ(b1 , τ ) − µ(b2 , τ )|
τ
h
n
oi
>
≤ E |(a> Jem (γN (τ ))−1 )Zj |1 |Yj − Z>
j b1 | ≤ |Zj (b1 − b2 )|
e γN ) sup |Z(x)> (b1 − b2 )|.
≤ 2f¯E(a,
x∈X
50
On the other hand kµ(b1 , τ ) − µ(b2 , τ )k ≤ 2ξm . Thus we have for any
δen > 0, a ∈ SIm−1
sup (a> Jem (γN (τ ))−1 )E[µ(β̆ (−j) (τ ), τ ) − µ(β̆(τ ), τ )]
τ,j
> (−j)
en .
e γN )δen + ξm P sup kZ>
. E(a,
β̆(τ
)
−
Z
β̆
(τ
)k
>
δ
j
j
τ,j
Choosing δen as above completes the proof.
Lemma S.6.7. Under the assumptions of Theorem S.6.2 we have for any
2 , all sufficiently large n, and a constant C independent of n
κn n/ξm
ξ 2 (κn + (log n)2 )
P sup kZ(x)> β̆ (−j) (τ ) − Z(x)> β̆(τ )k ≥ C c4m (γN ) + m
n
j,τ,x
≤3m(n + 1)e−κn .
Proof of Lemma S.6.7
Recall that for a vector a ∈ Rm kak0 denotes the number of non-zero entries
in a and ka denotes the position of the first non-zero entry. Consider the
following additional notation which is used exclusively in this proof:
(S.6.30)
I(x) := {i ∈ {1, ..., m} : |i − kZ(x) | ≤ r + D log n},
(S.6.31)
I 0 (x) := {i ∈ {1, ..., m} : ∃ j ∈ I(x) such that |i − j| ≤ r},
where r is defined in (L), and we suppress the dependence of I(x) and I 0 (x)
on D for the sake of a simpler notation.
√
P
(−j)
(−j)
(−j)
Let Pn f := (n−1)−1 i6=j f (Xi , Yi ), Gn f := n − 1(Pn f −Ef (X1 , Y1 )).
We have
(S.6.32) Z(x)> (β̆ (−j) (τ ) − γN (τ ))
5
X
(−j)
= −(n−1)−1/2 (Z(x)> Jem (γN (τ ))−1 )(I(x)) G(−j)
(ψ(·;
γ
(τ
),
τ
))+
rn,k (τ, x)
N
n
k=1
(−j)
(−j)
where rn,k (τ, x) := (Z(x)> Jem (γN (τ ))−1 )(I(x)) Rn,k (τ, γN (τ )), k = 1, ..., 4,
(−j)
Rn,k is defined exactly as Rn,k in (S.6.26)-(S.6.29) with n, β̆, Pn , Gn replaced
DISTRIBUTED INFERENCE FOR QUANTILE REGRESSION PROCESSES 51
(−j)
by n − 1, β̆ (−j) , Pn
(−j)
, Gn
, respectively, and
(−j)
rn,5 (τ, x)
h
i
:= − (Z(x)> Jem (γN (τ ))−1 )(I(x)) − (Z(x)> Jem (γN (τ ))−1 )
h
−1/2
× − − (n − 1)
Gn(−j) (ψ(·; γN (τ ), τ ))
+
4
X
k=1
i
(−j)
Rn,k (τ, γN (τ ))
i
h
= (Z(x)> Jem (γN (τ ))−1 )(I(x)) − (Z(x)> Jem (γN (τ ))−1 ) (β̆ (−j) (τ ) − γN (τ )),
where the last expression follows from (S.6.25).
Similarly we have
Z(x)> (β̆(τ ) − γN (τ ))
= −n
where
−1/2
(Z(x) Jem (γN (τ ))−1 )(I(x)) Gn (ψ(·; γN (τ ), τ )) +
>
5
X
rn,k (τ )
k=1
rn,5 (τ, x)
h
i
:= − (Z(x)> Jem (γN (τ ))−1 )(I(x)) − (Z(x)> Jem (γN (τ ))−1 )
4
h
i
X
× − − n−1/2 Gn (ψ(·; γN (τ ), τ )) +
Rn,k (τ, γN (τ )) .
k=1
h
i
= (Z(x)> Jem (γN (τ ))−1 )(I(x)) − (Z(x)> Jem (γN (τ ))−1 ) (β̆(τ ) − γN (τ )),
where the last expression follows from (S.6.25).
We obtain the representation
(−j)
sup Z(x)> β̆ (−j) (τ ) − Z(x)> β̆(τ ) − (rn,5 (τ, x) − rn,5 (τ, x))
j,τ,x
−
3
X
k=1
ξ2
(−j)
rn,k (τ, x) − rn,k (τ, x) . m
n
(−j)
a.s.
Observe that rn,4 (τ ) − rn,4 (τ ) = 0 for all τ because rn,4 does not depend
on β̆(τ ) or β̆ (−j) (τ ).
The results in Lemma S.6.11 show that
(−j)
sup krn,1 (τ, x) − rn,1 (τ, x)k .
j,τ,x
2 log n
ξm
n
a.s.
52
From Lemma S.6.10 and condition (L) we obtain for a constant C independent of j, τ, n, x
(−j)
krn,2 (τ ) − rn,2 (τ )k
≤ C |Z(x)> β̆ (−j) (τ ) − Z(x)> γN (τ )|2 + |Z(x)> β̆(τ ) − Z(x)> γN (τ )|2 .
Let wn := c4m (γN ) +
e 1,n (C) :=
Ω
n
2 (log n)2 +ξ 2 κ
ξm
m n
n
and define the event
o
sup |Z(x)> (βb(−j) (τ )−γN (τ ))|2 +|Z(x)> (β̆(τ )−γN (τ ))|2 ≥ Cwn .
τ,j,x
e 1,n (C)) can be bounded by m(n +
By an application of Lemma S.6.8, P (Ω
−κ
1)e n if C is chosen suitably.
Next, let δen := supj,τ,x kZ(x)> β̆(τ )(−j) − Z(x)> β̆(τ )k. We obtain for the
class of functions G2 defined in (S.6.34) and using (L)
(−j)
sup |rn,3 (τ, x) − rn,3 (τ, x)| ≤ C3
j,τ,x
ξ2
m
n
+ ξm sup kPn − P kG2 (δen ,I(x),I 0 (x)) .
x
Consider the event [here χ
en (t, I1 , Ij , κn ) is defined as (S.6.38) in Lemma
S.6.9]
n
o
kPn − P kG2 (t,I(x),I 0 (x))
e 2,n (C) := sup sup
Ω
≥
C
.
en (t, I(x), I 0 (x), κn )
x 0≤t≤1 χ
Since for x ∈ X the pair of sets (I(x), I 0 (x)) takes at most m distinct
e 2,n (C)) ≤ me−κn log2 n if we choose
values, it follows from (S.6.37) that P (Ω
C suitably.
Finally, apply (S.6.24) to find that
(−j)
sup |rn,5 (τ, x)|+|rn,5 (τ, x)| . ξm λD log n kβ̆(τ )−γN (τ )k+kβ̆ (−j) (τ )−γN (τ )k ,
j,τ,x
almost surely. Applying Theorem S.6.1 with βN = γN allows to bound
the probability of the event Ω3,n (C) := {kβ̆(τ ) − γN (τ )k + kβ̆ (−j) (τ ) −
γN (τ )k > C} by e−κn , and by choosing D sufficiently large it follows that
(−j)
supj,τ,x |rn,5 (τ, x)| + |rn,5 (τ, x)| ≤ n−1 with probability at least 1 − e−κn .
Summarizing the bounds obtained so far, there is a constant Cj indepene 1,n (C2 ) ∩ Ω
e 2,n (C2 ) ∩ Ω3,n (C2 )) ≤ m(n + 1)e−κn +
dent of n such that P (Ω
DISTRIBUTED INFERENCE FOR QUANTILE REGRESSION PROCESSES 53
e 1,n (C2 ) ∩ Ω
e 2,n (C2 ) ∩ Ω3,n (C2 ),
m log2 ne−κn + e−κn ≤ 3m(n + 1)e−κn . On Ω
ξ 2 κn ξ 2 (log n)2
δn ≤ C2 c4m (γN ) + m + m
+ C2 sup ξm χn (δn , I(x), I 0 (x), κn )
n
n
x
ξ 2 log2 (ξ ∨ n) ξ 2 κ
2 κ
2 (log n)2
ξ
ξ
n
m
n
≤ C2 c4m (γN ) + m + m
+ C2 m
+ m
n
n
n
n
ξ 2 log2 (ξ ∨ n) ξ 2 κ 1/2
m
n
+ C2 δn1/2 m
+ m
n
n
:= an + δn1/2 bn .
1/2
A simple calculation shows that the inequality 0 ≤ δn ≤ an + bn δn implies
0 ≤ δn ≤ 4 max(an , b2n ). Straightforward calculations show that under the
assumption of the theorem we have
an + b2n . c4m (γN ) +
2 log2 (ξ ∨ n)
ξm
ξ 2 κn
m
+ m .
n
n
Thus we have for some constant C3 independent of n
2 κ
ξ 2 log2 (ξm ∨ n) ξm
n
+
≤ 3m(n + 1)e−κn .
P δn ≥ C3 c4m (γN ) + m
n
n
This completes the proof.
Lemma S.6.8. Under the assumptions of Theorem S.6.2 we have for
2
sufficiently large n and any κn n/ξm
ξ log n + ξ κ1/2
m
m n
2
P sup |Z(x)> (β̆(τ )−γN (τ ))| ≥ C
+c
(γ
)
≤ (m+1)e−κn .
N
m
n1/2
τ,x
where the constant C does not depend on n.
Proof of Lemma S.6.8. See the proof for Lemma C.4 in the Appendix of
Chao et al. (2017).
Lemma S.6.9. Let Z := {Z(x)|x ∈ X } where X is the support of X. For
arbitrary index sets I, I 0 ⊂ {1, ..., m}, define the classes of functions
n
0
(S.6.33) G1 (I, I 0 ) := (z, y) 7→ a> z (I) (1{y ≤ z > b(I ) } − τ )1{kzk ≤ ξm }
o
τ ∈ T , b ∈ Rm , a ∈ S m−1 ,
54
(S.6.34)
n
(I 0 )
(I 0 )
G2 (δ, I, I 0 ) := (z, y) 7→ a> z (I) (1{y ≤ z > b1 }−1{y ≤ z > bj })1{z ∈ Z}
o
b1 , bj ∈ Rm , sup kv > b1 − v > bj k ≤ δ, a ∈ S m−1 .
v∈Z
Under assumptions (A1)-(A3) we have
h max(|I|, |I 0 |)
1/2
(S.6.35) P kPn − P kG1 ≥ C
log ξm
n
1/2
0
ξm κn i
max(|I|, |I |)ξm
κn
+
log ξm + 1/2 +
≤ e−κn
n
n
n
and for ξm δn n−1 we have for sufficiently large n and arbitrary κn > 0
(S.6.36)
P kPn − P kG2 (δ,I,I 0 ) ≥ C χ
en (δn , I, I 0 , κn ) ≤ e−κn
and for κn ≥ 1
(S.6.37)
P
where
kPn − P kG2 (t,I,I 0 )
≥
2C
≤ e−κn log2 n
en (t, I, I 0 , κn )
0≤t≤1 χ
sup
max(|I|, |I 0 |)
1/2
log(ξm ∨ n)
n
0
max(|I|, |I |)ξm
tκn 1/2 ξm κn
+
log(ξm ∨ n) +
+
.
n
n
n
(S.6.38) χ
en (t, I, I 0 , κn ) := t1/2
Proof of Lemma S.6.9. See the proof for Lemma C.5 in the Appendix of
Chao et al. (2017). The proof of (S.6.37) is similar to the proof for (S.6.21).
Lemma S.6.10.
Under assumptions (A1)-(A3) we have for any a ∈ Rm
a> Jem (γN (τ ))−1 µ(b, τ ) − a> Jem (γN (τ ))−1 µ(γN (τ ), τ ) − a> (b − γN (τ ))
≤ f 0 sup |Z(x)> b − Z(x)> γN (τ )|2 E[|a> Jem (γN (τ ))−1 Z|].
x
Proof of Lemma S.6.10 See the proof for Lemma C.6 in the Appendix of
Chao et al. (2017).
DISTRIBUTED INFERENCE FOR QUANTILE REGRESSION PROCESSES 55
Lemma S.6.11. Under assumptions (A1)-(A3) and (L) we have for any
a ∈ Rm having zero entries everywhere except at L consecutive positions:
|a> Pn ψ(·; β̆(τ ), τ )| ≤
(L + 2r)kakξm
.
n
Proof of Lemma S.6.11. See the proof for Lemma C.7 in the Appendix
of Chao et al. (2017).
APPENDIX S.7: PROOF OF (S.3.21)-(S.3.22) AND (S.3.24)-(S.3.25)
Let Ui := Fa,b (Yi ). By strict monotonicity of Fa,b on its support we have
j
j
j
Ui ∼ U [0, 1]. Moreover Y(k)
= Qa,b (U(k)
) where U(k)
denotes the kth order
j
statistic of the sample {Ui }i∈Aj if Aj is non-empty and U(k)
:= 0 if Aj is
empty. By independence of {Y1 , ..., Yn } and {X1 , ..., Xn } we have conditionally on {X1 , ..., Xn },
(S.7.1)
j
U(k)
∼ 1{nj > 0}Beta(k, nj + 1 − k).
Throughout this section we use the notation introduced in section S.3.3.2.
S.7.1. Proof of (S.3.21) and (S.3.24) (bias). Proof of (S.3.21): note
that
j
j
j
m1/2 βej (τ ) = Y(dn
1{nj > 0} = Qa,b (U(dn
)1{nj > 0} = Qa,b (U(dn
)
j τ e)
j τ e)
j τ e)
j
where nj ∼ Bin(n, 1/m) and the last equality uses the definition of U(dn
j τ e)
and the fact that Qa,b (0) = 0. Now we distinguish two cases.
j
− τ ] > 0, using the definition of Qa,b :
If lim supN →∞ (n/m) E[U(dn
j τ e)
(S.7.2)
j
j
j
E[Qa,b (U(dn
) − Qa,b (τ )] = a(E[(U(dn
)2 − τ 2 ]) + b(E[U(dn
− τ ]) ,
j τ e)
j τ e)
j τ e)
so (S.3.21) holds for any b > 0 and a = 0.
j
If lim supN →∞ (n/m) E[U(dn
− τ ] = 0 apply (S.7.1) and (S.7.13) in
j τ e)
Lemma S.7.1 to obtain
(S.7.3)
lim sup
N →∞
n
aτ (1 − τ )
j
E[Qa,b (U(dn
)] − Qa,b (τ ) ≥
>0
τ
e)
j
m
200
since T = [τL , τU ] with 0 < τL < τU < 1. Thus (S.3.21) holds for any b > 0
and 0 < a < amax .
56
Proof of (S.3.22): note that Qa,b is a polynomial of degree 2 and
K
X
Ak (·)Qa,b (τk ) − Qa,b (·)
= ΠK Qa,b − Qa,b
∞
k=1
∞
where
ΠK f denotes the projection, with respect to the inner product hf, gi :=
P
f
(τ
k )g(τk ), of a function f onto the space of functions spanned by
k
B1 , ..., Bq . Since B1 , ..., Bq are splines of degree rτ ≥ ητ ≥ 2, Qa,b lies in
the space spanned by B1 , ..., Bq (see p.111 of Schumaker (1981)). In other
words, Qa,b ∈ ΘG where ΘG is defined in the beginning of Section S.3.4.5.
Therefore, by similar arguments as in Section S.3.4.5
ΠK Qa,b − Qa,b
(S.7.4)
∞
. inf kp − Qa,b k∞ = 0.
p∈ΘG
We have from (S.7.4),
E[m1/2 β̌j (τ ) − Qa,b (τ )]
=
K
X
k=1
(S.7.5)
=
Ak (τ )E[m1/2 βej (τk ) − Qa,b (τk )]
X
j
Ak (τ )E[Qa,b (U(dn
) − Qa,b (τk )] + o(m/n),
j τk e)
k:|τk −τ |≤(log G)2 /G
where the o(m/n) is uniform in τ ∈ T . The first equality above is from
(S.7.4), and the second follows from the fact that supτ E[m1/2 βej (τ )−Qa,b (τ )] ≤
2(a + b) and
X
(S.7.6)
k:|τk −τ |>(log G)2 /G
≤ 2(a + b)
Ak (τ )E m1/2 βej (τk ) − Qa,b (τk )
k:|τk
X
2 /2
|Ak (τ )| . γ (log G)
= o(m/n),
−τ |>(log G)2 /G
where 0 < γ < 1; in the second above inequality we apply the bound (S.7.33)
in Lemma S.7.3.
Hence, by (S.7.5), it suffices to prove that for b > 0 and any amax > 0,
there exists a ∈ [0, amax ] and a sequence τN such that
(S.7.7)
lim sup
N →∞
n
m
X
k:|τk −τN |≤(log G)2 /G
j
> 0.
Ak (τN )E Qa,b (U(dn
)
−
Q
(τ
)
a,b
k
j τk e)
DISTRIBUTED INFERENCE FOR QUANTILE REGRESSION PROCESSES 57
In the following, we will distinguish the cases lim supN →∞ (n/m) < ∞ and
lim supN →∞ (n/m) = ∞.
First consider the case lim supN →∞ (n/m) < ∞. Invoke (S.7.17) in Lemma
S.7.2 to see that there exist a sequence τN in T such that
X
n
j
Ak (τN )E Qa,b (U(dn
) − Qa,b (τk )
lim sup
τ
e)
j
k
N →∞ m
2
k:|τk −τN |≤(log G) /G
≥ lim sup
N →∞
n
j
E Qa,b (U(dn
) − Qa,b (τN )
τ
e)
j
N
m
j
If lim supN →∞ (n/m) E[U(dn
− τN ] 6= 0, use (S.7.2), to see that the
j τN e)
right-hand side in the equation above is non-zero for any b > 0 and a = 0.
j
If lim supN →∞ (n/m) E[U(dn
− τN ] = 0 use (S.7.13) in Lemma S.7.1
j τN e)
to obtain
aτN (1 − τN )
n
j
>0
E Qa,b (U(dn
)
−
Q
(τ
)
≥
lim
sup
lim sup
N
a,b
τ
e)
j
N
200
N →∞
N →∞ m
where the last bound follows since 0 < τL ≤ τN ≤ τU < 1 for all N . Thus
we have established (S.7.7) for the case lim supN →∞ (n/m) > 0.
Next consider the case lim supN →∞ (n/m) = ∞. By (S.7.18) in Lemma
S.7.2 there exist a sequence τN ∈ T such that
lim sup
N →∞
≥ lim sup
N →∞
n
m
X
k:|τk −τN |≤(log G)2 /G
j
Ak (τN )E Qa,b (U(dn
) − Qa,b (τk )
j τk e)
n τN (1 − τN )
a
+(b+2aτN )
m (n/m) + 1
X
k:|τk −τN |≤(log G)2 /G
j
Ak (τN ) E[U(dn
]−τ
N .
j τk e)
We distinguish two cases.
P
j
If lim supN →∞ (n/m)
>
k:|τk −τN |≤(log G)2 /G Ak (τN ) E[U(dnj τk e) ] − τN
0, (S.7.7) holds for τN with any b > 0 and a = 0.
P
j
If lim supN →∞ (n/m)
A
(τ
)
E[U
]
−
τ
=
2
N
N
k
k:|τk −τN |≤(log G) /G
(dnj τk e)
0, we can pick an arbitrary 0 < a ≤ amax since τN ∈ T = [τL , τU ] with
0 < τL < τU < 1 and thus
1
n τN (1 − τN )
a
≥ lim sup aτN (1 − τN )
(n/m) + 1
2 N →∞
N →∞ m
1
≥ a min τL (1 − τL ), τU (1 − τU ) > 0.
2
Thus, (S.7.7) holds when lim supN →∞ (n/m) = ∞. Combining the two cases
above completes the proof of (S.7.7).
2
lim sup
58
S.7.2. Proof of (S.3.22) and (S.3.25). We begin by proving a slightly
stronger version of (S.3.22) for any j:
sup Var(βej (τ )) . 1/n.
(S.7.8)
τ ∈T
Observe that, as discussed in te beginning of Section S.7.1, m1/2 βej (τ ) =
j
Qa,b (U(dn
), and
j τ e)
Var m1/2 βej (τ )
j
= Var Qa,b (U(dn
)
j τ e)
2
j
j
j
j
2
2
= E a (U(dn
)
−
E[(U
)
]
+
b
U
−
E[U
]
τ
e)
(dn
τ
e)
(dn
τ
e)
(dn
τ
e)
j
j
j
j
j
j
j
j
j
2
2
= E a (U(dnj τ e) ) − E[U(dnj τ e) ] − Var(U(dnj τ e) ) + b U(dnj τ e) − E[U(dn
]
j τ e)
2
2
j
j
2
2
≤ 9 (4a + b )Var U(dnj τ e) + a Var U(dnj τ e)
.
2
Therefore, it is enough to show that
j
sup Var U(dn
. m/n.
j τ e)
(S.7.9)
τ ∈T
Distinguish two cases. If m/n is bounded
away from zero, this follows since
j
j
|U(dn
|
≤
1
and
thus
Var
U
is
bounded
by a constant independent
(dnj τ e)
j τ e)
of τ . We will thus without loss of generality assume that m/n → 0. Observe
the decomposition
h
2 i
j
j
j
Var(U(dn
)
=
E
U
−
E[U
]
(dnj τ e)
(dnj τ e)
j τ e)
h
2 i
h
2 i
j
j
j
j
= E U(dn
−
E[U
|n
]
+
E
E[U
|n
]
−
E[U
]
.
j
j
(dnj τ e)
(dnj τ e)
(dnj τ e)
j τ e)
First observe that
2 i
h
j
j
−
E[U
|n
]
E U(dn
(dnj τ e) j
j τ e)
r
n oi
h
2 n
n
n
j
j
≤6
log
+ O (n/m)−2
= E U(dnj τ e) − E[U(dnj τ e) |nj ] 1 nj −
m
m
m
h
2 i
j
j
≤
max
E U(dlτ e) − E[U(dlτ e) ]
+ O (n/m)−2
√
l:|l−(n/m)|≤6
=
max
√
l:|l−(n/m)|≤6
(S.7.10)
m/n,
(n/m) log(n/m)
(n/m) log(n/m)
Var U(dlτ e) + O (n/m)−2 ,
DISTRIBUTED INFERENCE FOR QUANTILE REGRESSION PROCESSES 59
j
uniformly in τ ∈ T . Here, the first equality follows since, U(dn
−
j τ e)
2
j
E[U(dn
|nj ]
≤ 1 a.s., nj ∼ Bin(n, 1/m) and the following result which
j τ e)
is a consequence of Bernstein’s inequality
r
n
n −C/3
n
n
(S.7.11) P nj −
≥C
log
≤2
for all C ≥ 1
m
m
m
m
The final inequality is from (S.7.1).
Next, observe that, almost surely,
j
E[U(dn
|nj ] − τ ≤
j τ e)
1
nj + 1
j
by (S.7.1). This implies |E[U(dn
] − τ | ≤ E[(1 + nj )−1 ], and thus
j τ e)
(S.7.12)
h
2 i
j
j
E E[U(dn
|n
]−E[U
]
. E (1+nj )−2 +(E[(1+nj )−1 ])2 = O((m/n)2 )
j
(dnj τ e)
j τ e)
by an application of (S.7.11). Combining (S.7.10) and (S.7.12) finishes the
proof of (S.7.9).
To prove (S.3.25), note that by Hölder’s inequality and Lemma S.7.3
K
h X
2 i
E[|β̌j (τ ) − E[β̌j (τ )]|2 ] = E
Ak (τ ){βej (τk ) − E[βej (τk )]}
k=1
K
K
h X
X
i
≤E
|Ak (τ )|
|Ak (τ )||βej (τk ) − E[βej (τk )]|2
k=1
.
sup
k=1,...,K
. 1/n,
k=1
Var(βej (τk ))
where the last inequality is an application of (S.7.8).
2
S.7.3. Auxiliary Lemmas.
Lemma S.7.1. Assume that Wn ∼ Bin(n, 1/m) and that Uτ |Wn = w ∼
Beta(dwτ e, w + 1 − dwτ e) if w > 0 and Uτ = 0 a.s. conditional on Wn = 0.
Let Qa,b (x) := ax2 + bx. Then for any a, b ≥ 0 and n, m such that n ≥ 2,
n/m ≥ 1, we have for all τ ∈ (0, 1) that
(S.7.13)
aτ (1 − τ )
n
n
n
E[Qa,b (Uτ ) − Qa,b (τ )] ≥
− b E[Uτ − τ ] − 2a E[Uτ − τ ] .
m
200
m
m
60
Proof of Lemma S.7.1. For a proof of (S.7.13), note that
|(E[Uτ ])2 − τ 2 | ≤ 2|E[Uτ ] − τ |.
Note also that
Var(Uτ ) = E[Var(Uτ |W )] + Var(E[Uτ |W ]) ≥ E[Var(Uτ |W )].
Thus with an application of triangle inequality, for a, b ≥ 0,
E[Qa,b (Uτ ) − Qa,b (τ )] = a(E[Uτ2 − τ 2 ]) + b(E[Uτ − τ ])
≥ aE[Var(Uτ |W )] − b E[Uτ − τ ] − 2a E[Uτ − τ ] .
(S.7.14)
Next we will prove a lower bound on E[Var(Uτ |W )]. The following tail bound
on binomial probabilities follows from the Bernstein inequality
n
x2 n
P Wn ≥ (1 + x)
≤ exp −
m
2+xm
∀x ≥ 0
In particular for n/m ≥ 1 setting C = x − 1 yields for C ≥ 1
(C − 1)2
n
P Wn ≥ C
≤ exp −
m
1+C
(S.7.15)
Next note that
E[Var(Uτ |W )] =
∞
X
w=1
P (Wn = w)
dwτ e(w + 1 − dwτ e)
,
(w + 1)2 (w + 2)
and moreover for any w > 0
τ (1 − τ )
dwτ e(w + 1 − dwτ e)
1
≤
≤
.
2
4(w + 2)
(w + 1) (w + 2)
w+2
Thus from (S.7.15) we obtain for n/m ≥ 1 by setting C = 3
m τ (1 − τ ) 1
τ (1 − τ )
P (Wn > 0) − e−1 ≥
− e−1
4(3(n/m) + 2)
n
20
2
m τ (1 − τ )
≥
.
n 200
E[Var(Uτ |W )] ≥
(S.7.16)
where we used the fact that for n ≥ 2, n/m ≥ 1 we have P (Wn > 0) ≥ 1/2.
The result follows by combining this with (S.7.14).
DISTRIBUTED INFERENCE FOR QUANTILE REGRESSION PROCESSES 61
Lemma S.7.2 (Localization). Let all conditions in Lemma S.7.1 hold
and assume that additionally N ≥ m, G (N/m)α for some α > 0 and
S (N/m)1/2 . Recall the definitions of Ak , Qa,b given in Section S.3.3.2
and assume that rτ ≥ 2.
1. If lim supN →∞ n/m < ∞ there exists a sequence τN in T such that
(S.7.17)
lim sup
N →∞
X
n
m
Ak (τN )E[Qa,b (Uτk ) − Qa,b (τk )]
k:|τk −τN |≤(log G)2 /G
≥ lim sup (n/m) E[Qa,b (UτN ) − Qa,b (τN )]
N →∞
2. If lim supN →∞ n/m = ∞ there exists a sequence τN in T such that
(S.7.18)lim sup
N →∞
≥ lim sup
N →∞
n
m
k:|τk −τN
X
Ak (τN )E[Qa,b (Uτk ) − Qa,b (τk )]
|≤(log G)2 /G
n τN (1 − τN )
a
+(b+2aτN )
m (n/m) + 1
X
k:|τk −τN |≤(log G)2 /G
Ak (τN ) E[Uτk ]−τN
Proof of Lemma S.7.2. For a proof of (S.7.17), for arbitrary w0 ∈ N,
observe that the sum
w0
X
(S.7.19) Sw0 (τ ) :=
P (Wn = w) aE[Uτ2 |Wn = w] + bE[Uτ |Wn = w]
w=1
is a piecewise constant function of τ . This function is constant on intervals of
the form [l1 /w1 , l2 /w2 ] for some positive integers l1 , l2 , w1 , w2 ≤ w0 satisfying
τL ≤ l1 /w1 < l2 /w2 ≤ τU . Thus, the length of each interval where Sw0 is
constant can be bounded from below as follows
l1
l2
l1 w2 − l2 w1
(S.7.20)
−
=
≥ w0−2 ,
w1 w2
w1 w2
since l1 w2 − l2 w1 ≥ 1 because both l1 w2 and l2 w1 are positive integers, and
w1 , w2 ≤ w0 . Let wN = log G. Then
(S.7.21)
a(E[Uτ2 − τ 2 ]) + b(E[Uτ − τ ]) = SwN (τ ) − P (Wn ≤ wN )(bτ + aτ 2 ) + rG (τ )
where the remainder satisfies supτ ∈T |rG (τ )| ≤ 2(a + b)P (Wn > wN ), and
by (S.7.15)
(S.7.22)
P (Wn > wN ) = o(1),
sup |rG (τ )| = o(m/n)
τ ∈T
62
since wN → ∞ and we are in the case n/m bounded. For each N , select τN
as the midpoint of any open interval JN ⊂ T on which SwN (τ ) is a constant.
From the choice of τN and wN , we have when N is sufficiently large,
2 )
SwN (τk ) = SwN (τN ) for all k : |τk − τN | ≤ 1/(3wN
(S.7.23)
and in particular SwN (τk ) = SwN (τN ) for k : |τk − τN | ≤ (log G)2 /G since
G → ∞. Hence, for each N , we get from (S.7.21) that
X
Ak (τN )E[Qa,b (Uτk ) − Qa,b (τk )]
k:|τk −τN |≤(log G)2 /G
=
X
Ak (τN ) a(E[Uτ2k − τk2 ]) + b(E[Uτk − τk ])
X
o
n
Ak (τN ) SwN (τk ) + − P (Wn ≤ wN )(bτk + aτk2 ) + rG (τk ) .
k:|τk −τN |≤(log G)2 /G
(S.7.24)
=
k:|τk −τN |≤(log G)2 /G
Note that by similar argument as (S.7.4) and (S.7.6), we have
X
k:|τk −τN |≤(log G)2 /G
o
n
2
)
Ak (τN )(bτk + aτk2 ) − (bτN + aτN
(S.7.25)
(S.7.6)
=
K n
o
X
(S.7.4)
2
) + o(m/n) = o(m/n),
Ak (τN )(bτk + aτk2 ) − (bτN + aτN
k=1
and
(S.7.26)
X
k:|τk −τN |≤(log G)2 /G
Ak (τN )rG (τk )
(S.7.6)
=
K
X
Ak (τN )rG (τk ) + o(m/n) = o(m/n).
k=1
where the last identity follows from (S.7.21) and Lemma S.7.3. Hence, we
DISTRIBUTED INFERENCE FOR QUANTILE REGRESSION PROCESSES 63
obtain from (S.7.24) that
lim sup
N →∞
(S.7.25)
(S.7.26)
=
n
m
k:|τk −τN
lim sup
N →∞
(S.7.23)
=
lim sup
N →∞
(S.7.6)
n
m
X
Ak (τN )E[Qa,b (Uτk ) − Qa,b (τk )]
|≤(log G)2 /G
X
k:|τk −τN |≤(log G)2 /G
n
SwN (τN )
m
n
o
2
Ak (τN )SwN (τk ) − P (Wn ≤ wN )(bτN + aτN
)
X
k:|τk −τN |≤(log G)2 /G
n
o
2
Ak (τN ) − P (Wn ≤ wN )(bτN + aτN
)
K n
o
X
n
2
SwN (τN )
Ak (τN ) − P (Wn ≤ wN )(bτN + aτN
)
N →∞ m
k=1
n
2
SwN (τN ) − P (Wn ≤ wN )(bτN + aτN
) + rG (τN ) .
lim sup
m
N →∞
= lim sup
=
Here the last line follows from Lemma S.7.3. Apply (S.7.21) to complete the
proof of (S.7.17).
For a proof of (S.7.18), note that by the law of iterated expectation,
(S.7.27)
2
E Qa,b (Uτ ) − Qa,b (τ ) = a(E[v2,n (τ )] + E[v1,n
(τ )] − τ 2 ) + b(E[v1,n (τ )] − τ )
where
dWn τ e
;
Wn + 1
dWn τ e(Wn + 1 − dWn τ e)
v2,n (τ ) := Var(Uτ |Wn ) = 1{Wn > 0}
.
(Wn + 1)2 (Wn + 2)
v1,n (τ ) := E[Uτ |Wn ] = 1{Wn > 0}
(S.7.28)
The following bound for binomial random variables can be obtained by Bernstein’s inequality for n/m ≥ 1:
p
P |Wn − n/m| ≥ C (n/m) log(n/m) ≤ 2(n/m)−C/3 for all C ≥ 1
Using this bound, straightforward calculations show that
E[v1,n (τ )] = τ + O(m/n)
τ (1 − τ )
E[v2,n (τ )] =
+ o(m/n)
(n/m) + 1
2
E[v1,n
(τ ) − τ 2 ] = 2τ E[v1,n (τ )] − τ + o(m/n)
64
where all remainder terms are uniform in τ ∈ T . Here, the first equation is
a consequence of the fact that |v1,n (τ ) − τ | ≤ (1 + Wn )−1 while the third
equation follows since
2
E[v1,n
(τ ) − τ 2 ] − 2τ E[v1,n (τ )] − τ = E[(v1,n (τ ) − τ )2 ].
Thus, from (S.7.27),
(S.7.29)
τ (1 − τ )
E Qa,b (Uτ ) − Qa,b (τ ) = a
+ (b + 2aτ ) E[v1,n (τ )] − τ + r1,N (τ )
(n/m) + 1
where r1,N (τ ) = o((n/m)−1 ) uniformly in τ ∈ T . Pick τN ≡ τ for an arbitrary τ ∈ T . From Lemma S.7.3 we obtain
(S.7.30)
k:|τk −τN
X
Ak (τN )r1,N (τk ) ≤
|≤(log G)2 /G
sup
k=1,...,K
|r1,N (τk )|
K
X
|Ak (τ )| = o(m/n).
k=1
Hence, combine (S.7.30) and (S.7.29) to get
lim sup
N →∞
(S.7.29)
(S.7.30)
=
n
m
X
k:|τk −τN |≤(log G)2 /G
lim sup
N →∞
(S.7.6)
(S.7.4)
= lim sup
N →∞
n
m
X
Ak (τN )E Qa,b (Uτk ) − Qa,b (τk )
k:|τk −τN |≤(log G)2 /G
n τ (1 − τ )
o
k
k
+ (b + 2aτk ) E[v1,n (τk )] − τk
Ak (τN ) a
(n/m) + 1
n τN (1 − τN )
a
+ (b + 2aτN )
m (n/m) + 1
X
Ak (τN ) E[v1,n (τk )] − τN
k:|τk −τN |≤(log G)2 /G
where the last inequality follows since for all k in the sum |τk − τN | =
O((log G)2 /G) = o(1) and |E[v1,n (τk )]| = τk + O(m/n) uniformly in τk . This
completes the proof of (S.7.18).
Lemma S.7.3. Assume that K ≥ G. Recall the definition of Ak (τ ) in
(S.3.23). There exist constants c, C > 0 independent of K, G such that
sup |Ak (τ )| ≤ c,
k,τ ∈T
Moreover, for any τ ∈ T ,
PK
sup
K
X
τ ∈T k=1
k=1 Ak (τ )
= 1.
|Ak (τ )| ≤ C.
DISTRIBUTED INFERENCE FOR QUANTILE REGRESSION PROCESSES 65
P
Proof of Lemma S.7.3. We start with proving K
k=1 Ak (τ ) = 1. ObPK
serve that k=1 Ak (τ ) = ΠK g1 , where ΠK is defined in Section S.3.4.5 and
g1 (x) ≡ 1. Since the degree of the piecewise polynomials in B is greater or
equal to 2 and g1 is an order 0 polynomial, similar arguments
as used in the
PK
proof of (S.7.4) show that kΠK g1 − 1k∞ = 0. This implies k=1 Ak (τ ) = 1
for all τ ∈ T .
To obtain a bound for supk,τ ∈T |Ak (τ )|, note that
sup Ak (τ ) = sup kΠK ek k∞ ≤ kΠK k∞ sup kek k∞ ,
k
k,τ ∈T
k
where ek (·) : T → [0, 1] is the function interpolating (τ1 , 0),..., (τk−1 , 0),
(τk , 1), (τk+1 , 0)..., (τK , 0) , so supk kek k∞ = 1. In the first step in Section S.3.4.5 we have shown that kΠK k∞ < C 0 for some constant C 0 > 0
independent of K. Hence, taking
PK c = kΠK k∞ finishes the first claim.
Next we show that supτ ∈T k=1 |Ak (τ )| ≤ C. For any τ ∈ T and τk ∈ TK ,
define the index sets
I(τ ) := j ∈ {1, ..., q} : Bj (τ ) 6= 0 ;
Ik := j ∈ {1, ..., q} : Bj (τk ) 6= 0 .
Note that |I(τ )| = |I(τk )| = rτ + 1 for all τ ∈ T and k, and the elements in
both sets are consecutive positive integers by the connective support of Bj .
By Lemma 6.3 of Zhou et al. (1998) (note that their N, n and Gk,n are
P
>
our B, K and K −1 K
k=1 B(τk )B(τk ) , while their h is of order 1/G see page
1761, 1762 and 1765 of Zhou et al. (1998)), there exist constants 0 < γ < 1
and c1 > 0 independent of K, G, τ such that the ij’th element of the matrix
satisfies
(S.7.31)
K
X
B(τk )B(τk )>
k=1
−1
ij
≤ c1 K −1 Gγ |j−i| ≤ c1 γ |j−i|
where the second inequality follows from our assumption that K ≥ G.
Recall that our Bj is defined as given in Definition 4.19 on page 124 of
Schumaker (1981). Thus, by equation (4.31) in the same reference,
sup max |Bj (τ )|2 ≤ 1,
τ ∈T j≤q
66
and by (S.7.31),
(S.7.32)
Ak (τ ) = B(τ )>
K
X
B(τk )B(τk )>
k=1
(S.7.33)
≤ c1
∞
X
l=0
−1
B(τk ) ≤ c1
X
0
γ |j−j |
j∈I(τ ),j 0 ∈Ik
(I(τ ) + l) ∩ Ik + (I(τ ) − l) ∩ Ik γ l ,
where I(τ ) − l = {j − l : j ∈ I(τ )}, I(τ ) + l = {j + l : j ∈ I(τ )}. Now,
we make a key observation that (I(τ ) + l) ∩ Ik + (I(τ ) − l) ∩ Ik 6= 0
for at most 6(rτ + 1) many k at each fixed l ∈ N. This is due to the fact
that both I(τ ) and Ik consist of rτ + 1 consecutive positive integers, and
the same is true for both I(τ ) − l and I(τ ) + l. Hence, from (S.7.33) and
Fubini’s theorem,
K
X
k=1
Ak (τ ) ≤ c1
∞ X
K
X
l=0 k=1
≤ 6c1 (rτ + 1)
(I(τ ) + l) ∩ Ik + (I(τ ) − l) ∩ Ik γ l
∞
X
γ l < ∞.
l=0
This completes the proof.
APPENDIX S.8: TECHNICAL RESULTS FROM EMPIRICAL
PROCESS THEORY
In this section, we collect some basic results from empirical process theory
that we use throughout the proofs. Denote by G a class of functions that
satisfies |f (x)| ≤ F (x) ≤ U for every f ∈ G and let σ 2 ≥ supf ∈G P f 2 .
Additionally, let for some A > 0, V > 0 and all ε > 0
AkF k 2
L (Pn ) V
(S.8.1)
N (ε, G, L2 (Pn )) ≤
.
ε
In that case, the symmetrization inequality and inequality (2.2) from Koltchinskii (2006) yield
h V
AkF kL2 (P ) 1/2 V U
AkF kL2 (P ) i
(S.8.2) EkPn − P kG ≤ c0 σ
log
+
log
n
σ
n
σ
for a universal constant c0 > 0 provided that 1 ≥ σ 2 > constn−1 [in fact,
the inequality in Koltchinskii (2006) is for σ 2 = supf ∈G P f 2 . However, this
DISTRIBUTED INFERENCE FOR QUANTILE REGRESSION PROCESSES 67
is not a problem since we can replace G by Gσ/(supf ∈G P f 2 )1/2 ]. The second inequality (a refined version of Talagrand’s concentration inequality)
states that for any countable class of measurable functions F with elements
mapping into [−M, M ]
(S.8.3)
n
1/2 √
o
P kPn −P kF ≥ 2EkPn −P kF +c1 n−1/2 sup P f 2
t+n−1 cj M t ≤ e−t ,
f ∈F
for all t > 0 and universal constants c1 , cj > 0. This is a special case of
Theorem 3 in Massart (2000) [in the notation of that paper, set ε = 1].
2 (log n)2 = o(N ) we have for any
Lemma S.8.1. Under (A1)-(A3) and ξm
δ > 0 and uN ∈ Rm ,
(S.8.4)
1/2
>
>
N
sup
u
U
(τ
)−u
U
(τ
)
>
ε
= 0,
lim lim sup P kuN k−1
N N 1
N N 2
2
δ→0 N →∞
τ1 ,τ2 ∈T ,|τ1 −τ2 |≤δ
where UN (τ ) is defined in (S.3.35).
Proof of Lemma S.8.1. See the proof for Lemma A.3 in Chao et al.
(2017).
Lemma S.8.2. Let p be a polynomial of degree q. Then for arbitrary
intervals [a, b]and α ∈ [0, 1) we have
Z
a
b
b−a
p (x)dx −
J
2
X
j:j+α≤J
p
2
Z b
Cq
j+α
a+
p2 (x)dx
(b − a) ≤
J
J −1 a
for a constant Cq > 1 that depends only on q.
Proof of Lemma S.8.2. We begin by observing that it suffices to prove
that for any polynomial p of degree q we have
Z 1
Z
Cq 1 2
1 X
2
2 j+α
p (x)dx −
(S.8.5)
p
≤
p (x)dx.
J
J
J 0
0
j:j+α≤J
To see this, note that for any polynomial p of degree q the function pea,b (x) :=
p(a + x(b − a))(b − a)1/2 is also a polynomial of degree q and we have
R1 2
Rb 2
p
e
(x)dx
=
a,b
0
a p (x)dx by a change of variables and
b−a
J
X
j:j+α≤J
j+α
1
p2 a +
(b − a) =
J
J
X
j:j+α≤J
pe2a,b
j + α
J
.
68
To prove (S.8.5), consider the bound
Z 1
j + α
1 X
1
dp2 (x) 2
p2 (x)dx−
≤
+ sup p2 (x),
p2
sup
J
J
J − 1 x∈[0,1] dx
J x∈[0,1]
0
j:j+α≤J
where the first term on the right is from the difference of the two integrals on
the interior partitions, and the second term handles the boundary partitions
which arennot 2included in the
o summation
R 1 on the left. Thus we need to bound
supx∈[0,1] dpdx(x) + p2 (x) by Cq 0 p2 (x)dx. To this end, denote by Pq
the space of polynomials on [0, 1] with degree less or equal to q. This is a
finite-dimensional vector space. Note that the operator Φ : f 7→ f 0 + f is a
linear operator between the finite-dimensional, normed spaces (P2q , k · k1 )
and (P2q , k · k∞ ). Such an operator must be bounded, and hence there exists
a constant Cq with kΦ(p2 )k∞ ≤ Cq kp2 k1 . This completes the proof.
APPENDIX S.9: TECHNICAL DETAILS FOR SIMULATION AND
ADDITIONAL SIMULATION RESULTS
This part of the supplement contains additional results and some technical details for the simulation study in Section 5. Section S.9.1 contains
some technical details related to the simulation settings in Section 5, and
additional figures (the case m = 16 and coverage probabilities of confidence
intervals for FY |X (y|x0 ) and large values of S). In Section S.9.3, using the
same linear model as in Section 5, we provide additional simulation results
for the oracle properties of the divide and conquer estimator β(τ ) and the
distribution function estimator FbY |X (y|x); the emphasis of this section is
on illustrating our oracle theory. In Section S.9.4, we consider a non-linear
model and illustrate oracle properties of the divide and conquer estimator β(τ ) and the distribution function estimator FbY |X (y|x). Some practical
conclusions from the simulations are summarized in Section S.9.5
S.9.1. Details for Section 5 and additional simulation results.
The simple pooled estimator for the asymptotic covariance matrix of βor (τ )−
β(τ ) was computed by first using the options se=’ker’, covariance=TRUE
in the rq function in quantreg (Koenker, 2016) for each sub-sample, and
then averaging the results over sub-samples.
The values of the optimal constant c∗ (τ ) were determined from the theory
developed in Kato (2012). More precisely, we use the following formula given
on its page 264 of the latter reference,
P
1/5
2 2
4.5 m
j,k=1 E[Zj Zk ]
∗
P
(S.9.1)
,
c (τ ) = σ
2
α(τ ) m
j,k=1 (E[Zj Zk ])
DISTRIBUTED INFERENCE FOR QUANTILE REGRESSION PROCESSES 69
where α(τ ) = (1 − Φ−1 (τ )2 )2 φ(Φ−1 (τ )) (note that this is a corrected version
of Kato’s formula). Note that c∗ (τ ) is symmetric around τ = 0.5. In order
to save computation time in the simulation, we simulated an independent
data set to compute c∗ (τ ), and used the same c∗ (τ ) (based on m = 4) for
different m in all the simulations, because the exact value of c∗ (τ ) does not
change much with m under our design of Z; see Table S.9.1.
m = 4 m = 16 m = 32
0.242
0.252
0.254
0.173
0.179
0.180
Table S.9.1
Values of c∗ (τ ) based on (S.9.1).
τ = 0.1
0.5
In the following, we show additional simulation results. Figure S.9.1 shows
the coverage probability of the confidence interval for Q(x0 ; τ ) based on
x>
0 β̄(τ ) for dimension m = 16. For additional simulations for FY |X (y|x0 )
30
40
50
1.0
0.8
10
20
30
40
50
0
10
20
30
8
10
50
8
10
0.8
0.6
0.0
0.2
0.4
coverage
0.8
0.6
0.0
0.2
0.4
coverage
0.8
0.6
0.4
6
log2(S)
40
1.0
S
τ = 0.9
1.0
S
τ = 0.5
0.2
4
0.6
coverage
0.2
0.0
0
S
τ = 0.1
0.0
2
0.4
0.6
0.0
0.2
coverage
0.8
20
1.0
10
0.4
0.8
0.6
coverage
0.0
0.2
0.4
oracle asymp. CI n=512
oracle asymp. CI n=2048
CI from (4.2) t quant. n=512
CI from (4.2) t quant. n=2048
CI from (4.2) norm quant. n=512
CI from (4.2) norm quant. n=2048
bootstrap n=512, B=500
bootstrap n=2048, B=500
0
coverage
τ = 0.9
1.0
τ = 0.5
1.0
τ = 0.1
2
4
6
log2(S)
8
10
2
4
6
log2(S)
Fig S.9.1: Coverage probabilities for Q(x0 ; τ ) for m = 16 in homoskedastic
model (5.1). Coverage probabilities for x>
0 β(τ ) for different values of S and
τ = 0.1, 0.5, 0.9 (left, middle, right row). Solid lines: n = 512, dashed lines:
n = 2048. Black: asymptotic oracle CI, blue: CI from (4.2) based on t distribution, red: CI from (4.1) based on normal distribution, green: bootstrap
CI. Throughout x0 = (1, ..., 1)/m1/2 , nominal coverage 0.95.
under the same settings as that of Figure 5.2 in Section 5, Figure S.9.2 shows
70
the coverage probability of the confidence interval with large S for m = 4
and 32 and Figure S.9.3 shows the coverage probability of the confidence
interval for m = 16.
8
1.0
0.8
4
6
log2(S)
8
10
2
10
8
10
8
10
1.0
coverage
0.8
1.0
0.6
0.2
0.0
0.2
0.4
coverage
0.0
0.0
8
6
log2(S)
τ = 0.9 , m=32
0.8
1.0
0.8
0.6
0.2
0.4
6
log2(S)
4
τ = 0.5 , m=32
oracle asymp. CI n=512
bootstrap n=512, B=500
4
0.6
0.2
0.0
2
τ = 0.1 , m=32
2
0.4
0.6
0.0
10
0.6
log2(S)
0.4
6
coverage
0.8
4
0.4
0.2
coverage
0.8
0.6
0.4
coverage
0.2
0.0
oracle asymp. CI n=512
bootstrap n=512, B=500
2
coverage
τ = 0.9 , m=4
1.0
τ = 0.5 , m=4
1.0
τ = 0.1 , m=4
2
4
6
log2(S)
8
10
2
4
6
log2(S)
Fig S.9.2: Coverage probabilities for FY |X (y|x0 ) in homoskedastic
model (5.1) with m = 4, 32 with large number of subsamples S, n = 512
and nominal coverage 0.95. Red: oracle asymptotic CI, black: bootstrap CI
with B = 500.
S.9.2. Heteroskedastic model. We consider a linear location-scale
shift model,
(S.9.2)
>
Yi = 0.21 + βm−1
Xi + (1 + z>
m−1 Xi )εi ,
i = 1, ..., N.
where m ∈ {4, 16, 32}, βm−1 is chosen as (S.9.3), Xi and εi are chosen as
model (5.1), and the vector zm−1 takes the form
z3 = (0.69, 0.56, 0.35)> ;
(S.9.3)
> > > > >
z15 = (z>
3 , z3 , z3 , z3 , z3 ) ;
>
>
z31 = (β15
, β15
, 0.69)> .
S.9.2.1. Results for the divide and conquer estimator β̄(τ ). We fix the
sub-sample size n and consider the impact of S on the coverage probabilities
of various 95% confidence intervals as in Section 5.1. We use the infeasible
DISTRIBUTED INFERENCE FOR QUANTILE REGRESSION PROCESSES 71
30
40
1.0
0.8
0
10
20
30
40
50
0
30
τ = 0.5 , m=16
τ = 0.9 , m=16
10
0.8
0.6
0.2
0.0
0.0
8
50
0.4
coverage
0.8
0.6
0.2
0.4
coverage
0.8
6
log2(S)
40
1.0
τ = 0.1 , m=16
1.0
S
0.6
0.2
0.0
20
S
0.4
4
10
S
oracle asymp. CI n=512
bootstrap n=512, B=500
2
0.6
0.2
0.0
0.0
50
0.4
0.6
coverage
0.8
20
1.0
10
0.4
0.2
coverage
0.8
0.6
0.4
coverage
0.2
0.0
oracle asymp. CI n=512
bootstrap n=512, B=500
0
coverage
τ = 0.9 , m=16
1.0
τ = 0.5 , m=16
1.0
τ = 0.1 , m=16
2
4
6
log2(S)
8
10
2
4
6
log2(S)
8
10
Fig S.9.3: Coverage probabilities in homoskedastic model (5.1) for
FY |X (y|x0 ) for m = 16, n = 512 and nominal coverage 0.95. Red: oracle
asymptotic CI, black: bootstrap CI with B = 500.
asymptotic confidence interval (5.3) as benchmark (note that the asymptotic
variance is different in the homoskedastic and heteroskedastic model).
In the first step, as in Section 5.1, we consider the three types of confidence intervals in Section 4.1. The coverage probabilities of the corresponding confidence intervals are presented in Figure S.9.4. The general patterns
are similar to those in Figure 5.1, but coverage starts to drop earlier for
both m = 4, 32. The same conclusion holds for the coverage probabilities
for dimension m = 16 when comparing Figure S.9.5 to Figure S.9.1 for the
homoskedastic model.
Next, we analyse the asymptotic confidence intervals using the empirical
asymptotic variance of x>
0 β(τ ) from data, which is estimated with the three
ways described on p.22 in Section 5.1. The constants c∗ (τ ) there are replaced
by c∗g (τ ), which is computed by adapting the formula on page 263 of Kato
(2012) to the N (0, σ 2 ) location-scale shift model (S.9.2). This yields the
formula:
P
1/5
>
−1 2 2
4.5 m
j,k=1 E[(1 + zm−1 X) Zj Zk ]
∗
P
(S.9.4)
cg (τ ) = σ
,
>
−3
2
α(τ ) m
j,k=1 (E[(1 + zm−1 X) Zj Zk ])
where α(τ ) = (1 − Φ−1 (τ )2 )2 φ(Φ−1 (τ )). The values are listed in Table S.9.2.
The results for the coverage probabilities are reported in Table S.9.3. The
72
m = 4 m = 16 m = 32
0.4206
1.2118
2.3222
0.2990
0.8614
1.6508
Table S.9.2
Values of c∗g (τ ) based on (S.9.4) of the linear Gaussian location-scale shift model (5.1) .
τ = 0.1
0.5
overall patterns are similar to those in Table 5.1, but the coverage probabilities start to drop earlier, this effect is similar to what we observed in
Figure S.9.4.
S
or
def
nai
adj
or
def
nai
adj
or
def
nai
adj
or
def
nai
adj
or
def
nai
adj
or
def
nai
adj
1
10
30
50
n = 512, m = 4, τ = 0.1
94.9
94.1
95.0
93.9
92.8
92.7
93.8
92.6
93.7
93.1
94.4
93.0
93.7
93.6
94.9
93.8
n = 512, m = 4, τ = 0.5
95.4
94.7
95.5
94.9
97.4
97.0
97.8
97.4
96.0
95.6
96.1
95.8
96.0
95.2
95.6
95.1
n = 512, m = 4, τ = 0.9
95.0
95.2
93.9
94.5
92.9
94.5
92.5
93.6
93.8
95.0
93.0
93.7
93.8
95.2
93.6
94.5
1
10
30
50
n = 512, m = 16, τ = 0.1
95.1
94.0
92.7
90.5
94.9
93.5
91.9
89.8
97.7
93.2
91.2
88.5
97.7
94.1
92.9
90.9
n = 512, m = 16, τ = 0.5
95.4
96.2
95.3
94.3
99.3
99.3
98.8
99.0
98.6
98.0
97.3
97.1
98.6
97.4
96.1
95.6
n = 512, m = 16, τ = 0.9
95.3
94.2
92.3
91.6
94.0
93.8
91.6
90.8
97.1
93.3
90.8
89.8
97.1
94.6
92.4
91.8
1
10
30
50
n = 512, m = 32, τ = 0.1
94.3
91.6
83.0
73.8
94.6
90.4
80.2
71.3
99.4
91.6
80.9
72.0
99.4
93.4
83.7
75.6
n = 512, m = 32, τ = 0.5
95.3
94.3
95.7
94.7
98.4
98.6
98.6
98.2
99.0
97.5
97.2
97.4
99.0
96.7
96.4
96.1
n = 512, m = 32, τ = 0.9
94.7
90.9
80.1
72.7
94.0
88.5
79.2
69.8
99.4
90.0
80.1
70.6
99.4
91.6
83.2
73.9
n = 2048, m = 4, τ = 0.1
95.1
96.3
94.9
95.1
94.0
95.8
94.2
94.8
94.6
95.8
94.4
94.8
94.6
96.0
94.6
95.0
n = 2048, m = 4, τ = 0.5
95.1
94.4
95.6
94.6
95.9
95.5
96.6
95.7
95.4
94.9
96.2
95.0
95.4
94.6
95.8
94.8
n = 2048, m = 4, τ = 0.9
95.1
94.9
95.6
94.6
94.2
94.4
95.1
94.1
94.6
94.5
95.1
94.1
94.6
94.7
95.6
94.4
n = 2048, m = 16, τ = 0.1
95.6
94.1
94.8
93.8
94.6
93.5
93.7
92.8
95.7
93.1
93.2
92.4
95.7
93.8
94.4
93.4
n = 2048, m = 16, τ = 0.5
94.8
95.0
95.4
95.1
97.6
97.8
98.0
98.0
96.6
96.8
96.6
96.4
96.6
95.9
95.8
95.6
n = 2048, m = 16, τ = 0.9
95.0
94.8
94.2
93.2
94.0
94.2
93.4
92.1
95.4
93.9
93.0
91.8
95.4
94.4
94.0
93.2
n = 2048, m = 32, τ = 0.1
95.2
95.0
90.5
88.2
94.7
94.3
89.3
87.0
95.6
94.2
89.1
86.8
95.6
94.9
90.2
87.9
n = 2048, m = 32, τ = 0.5
95.2
95.1
95.2
95.0
97.5
97.1
97.6
97.4
97.0
96.4
96.4
96.4
97.0
96.0
95.4
95.6
n = 2048, m = 32, τ = 0.9
95.3
94.5
91.0
89.4
94.8
93.8
90.1
88.2
95.6
93.8
90.0
87.6
95.6
94.8
91.1
89.4
Table S.9.3
Coverage probabilities under the linear Gaussian location-scale shift mode (S.9.2) based
on estimating the asymptotic variance. Different rows correspond to different methods for
obtaining covariance matrix. or: using true asymptotic variance matrix, def: default
choice implemented in quantreg package, nai: asymptotically optimal constant with
scaling hn ∼ n−1/5 , adj: asymptotically optimal constant with scaling hn ∼ N −1/5 as
suggested by Theorem 4.4.
DISTRIBUTED INFERENCE FOR QUANTILE REGRESSION PROCESSES 73
(a) m = 4
30
40
50
1.0
0.8
10
20
30
40
50
0
10
20
30
τ = 0.1
τ = 0.5
τ = 0.9
8
50
8
10
40
50
8
10
0.8
0.6
0.0
0.2
0.4
coverage
0.8
0.6
0.0
0.2
0.4
coverage
0.8
0.6
0.4
6
log2(S)
40
1.0
S
1.0
S
0.2
4
0.6
coverage
0.2
0.0
0
S
0.0
2
0.4
0.6
0.0
0.2
coverage
0.8
20
1.0
10
0.4
0.8
0.6
coverage
0.0
0.2
0.4
oracle asymp. CI n=512
oracle asymp. CI n=2048
CI from (4.2) t quant. n=512
CI from (4.2) t quant. n=2048
CI from (4.2) norm quant. n=512
CI from (4.2) norm quant. n=2048
bootstrap n=512, B=500
bootstrap n=2048, B=500
0
coverage
τ = 0.9
1.0
τ = 0.5
1.0
τ = 0.1
10
2
4
6
log2(S)
8
10
2
4
6
log2(S)
(b) m = 32
30
40
50
1.0
0.8
10
20
30
40
50
0
10
20
30
τ = 0.1
τ = 0.5
τ = 0.9
6
log2(S)
8
10
0.8
0.6
0.0
0.2
0.4
coverage
0.8
0.6
0.4
0.0
0.2
0.4
0.6
coverage
0.8
1.0
S
1.0
S
0.2
4
0.6
coverage
0.2
0.0
0
S
0.0
2
0.4
0.6
0.0
0.2
coverage
0.8
20
1.0
10
0.4
0.8
0.6
coverage
0.0
0.2
0.4
oracle asymp. CI n=512
oracle asymp. CI n=2048
CI from (4.2) t quant. n=512
CI from (4.2) t quant. n=2048
CI from (4.2) norm quant. n=512
CI from (4.2) norm quant. n=2048
bootstrap n=512, B=500
bootstrap n=2048, B=500
0
coverage
τ = 0.9
1.0
τ = 0.5
1.0
τ = 0.1
2
4
6
log2(S)
8
10
2
4
6
log2(S)
Fig S.9.4: Coverage probabilities under the linear Gaussian location-scale
shift model (S.9.2) for x>
0 β(τ ) for different values of S and τ = 0.1, 0.5, 0.9
(left, middle, right row). Solid lines: n = 512, dashed lines: n = 2048. Black:
asymptotic oracle CI, blue: CI from (4.2) based on t distribution, red: CI
from (4.1) based on normal distribution, green: bootstrap CI. Throughout
x0 = (1, ..., 1)/m1/2 , nominal coverage 0.95.
74
30
40
50
1.0
0.8
10
20
30
40
50
0
10
20
30
8
10
50
8
10
0.8
0.6
0.0
0.2
0.4
coverage
0.8
0.6
0.0
0.2
0.4
coverage
0.8
0.6
0.4
6
log2(S)
40
1.0
S
τ = 0.9
1.0
S
τ = 0.5
0.2
4
0.6
coverage
0.2
0.0
0
S
τ = 0.1
0.0
2
0.4
0.6
0.0
0.2
coverage
0.8
20
1.0
10
0.4
0.8
0.6
coverage
0.0
0.2
0.4
oracle asymp. CI n=512
oracle asymp. CI n=2048
CI from (4.2) t quant. n=512
CI from (4.2) t quant. n=2048
CI from (4.2) norm quant. n=512
CI from (4.2) norm quant. n=2048
bootstrap n=512, B=500
bootstrap n=2048, B=500
0
coverage
τ = 0.9
1.0
τ = 0.5
1.0
τ = 0.1
2
4
6
log2(S)
8
10
2
4
6
log2(S)
Fig S.9.5: Coverage probabilities for Q(x0 ; τ ) for m = 16 under the linear Gaussian location-scale shift model (S.9.2). Coverage probabilities for
x>
0 β(τ ) for different values of S and τ = 0.1, 0.5, 0.9 (left, middle, right
row). Solid lines: n = 512, dashed lines: n = 2048. Black: asymptotic oracle
CI, blue: CI from (4.2) based on t distribution, red: CI from (4.1) based on
normal distribution, green: bootstrap CI. Throughout x0 = (1, ..., 1)/m1/2 ,
nominal coverage 0.95.
S.9.2.2. Results for the estimator FbY |X (y|x). In this section, we compare the coverage probabilities of the bootstrap confidence intervals for
FY |X (Q(x0 ; τ )|x0 ) = τ with the oracle asymptotic confidence interval (5.4).
To estimate FbY |X , we use the same number of quantile grid K = 65 and
the same number of equidistant knots G = 32 for spline interpolation as
in Section 5.2, which are sufficiently large to ensure the nominal coverage
of the oracle confidence intervals. Coverage probabilities for m = 4, 32 are
reported in Figure S.9.6. Results for large values of S and dimension m = 16
are presented in Figure S.9.7 and S.9.8. The general patterns are similar to
the homoskedastic model with overall coverage starting to drop earlier.
DISTRIBUTED INFERENCE FOR QUANTILE REGRESSION PROCESSES 75
30
40
1.0
0.8
0
10
20
30
40
50
0
20
30
τ = 0.1 , m=32
τ = 0.5 , m=32
τ = 0.9 , m=32
50
50
40
50
0.8
0.6
0.4
0.0
0.2
0.4
0.6
coverage
0.8
40
0.2
coverage
30
S
0.0
0.8
0.6
0.4
0.2
20
40
1.0
S
1.0
S
0.0
10
10
S
oracle asymp. CI n=512
bootstrap n=512, B=500
0
0.6
0.2
0.0
0.0
50
0.4
0.6
coverage
0.8
20
1.0
10
0.4
0.2
coverage
0.8
0.6
0.4
coverage
0.2
0.0
oracle asymp. CI n=512
bootstrap n=512, B=500
0
coverage
τ = 0.9 , m=4
1.0
τ = 0.5 , m=4
1.0
τ = 0.1 , m=4
0
10
20
30
S
40
50
0
10
20
30
S
Fig S.9.6: Coverage probabilities for oracle confidence intervals (red) and
bootstrap confidence intervals (black) for FY |X (Q(x0 ; τ )|x0 ) = τ under the
linear Gaussian location-scale shift model (S.9.2) for x0 = (1, ..., 1)/m1/2
and y = Q(x0 ; τ ), τ = 0.1, 0.5, 0.9. n = 512 and nominal coverage 0.95.
76
8
1.0
0.8
4
6
log2(S)
8
10
2
10
8
10
8
10
1.0
0.6
coverage
0.8
1.0
0.6
0.2
0.0
0.2
0.4
coverage
0.0
0.0
8
6
log2(S)
τ = 0.9 , m=32
0.8
1.0
0.8
0.6
0.2
0.4
6
log2(S)
4
τ = 0.5 , m=32
oracle asymp. CI n=512
bootstrap n=512, B=500
4
0.6
0.2
0.0
2
τ = 0.1 , m=32
2
0.4
0.6
0.0
10
0.4
6
log2(S)
coverage
0.8
4
0.4
0.2
coverage
0.8
0.6
0.4
coverage
0.2
0.0
oracle asymp. CI n=512
bootstrap n=512, B=500
2
coverage
τ = 0.9 , m=4
1.0
τ = 0.5 , m=4
1.0
τ = 0.1 , m=4
2
4
6
log2(S)
8
10
2
4
6
log2(S)
Fig S.9.7: Coverage probabilities for FY |X (Q(x0 ; τ )|x0 ) = τ under the linear
Gaussian location-scale shift model (S.9.2) for m = 4, 32 with large number
of subsamples S, n = 512 and nominal coverage 0.95. Red: oracle asymptotic
CI, black: bootstrap CI with B = 500.
DISTRIBUTED INFERENCE FOR QUANTILE REGRESSION PROCESSES 77
30
40
1.0
0.8
0
10
20
30
40
50
0
30
τ = 0.5 , m=16
τ = 0.9 , m=16
10
0.8
0.6
0.2
0.0
0.0
8
50
0.4
coverage
0.8
0.6
0.2
0.4
coverage
0.8
6
log2(S)
40
1.0
τ = 0.1 , m=16
1.0
S
0.6
0.2
0.0
20
S
0.4
4
10
S
oracle asymp. CI n=512
bootstrap n=512, B=500
2
0.6
0.2
0.0
0.0
50
0.4
0.6
coverage
0.8
20
1.0
10
0.4
0.2
coverage
0.8
0.6
0.4
coverage
0.2
0.0
oracle asymp. CI n=512
bootstrap n=512, B=500
0
coverage
τ = 0.9 , m=16
1.0
τ = 0.5 , m=16
1.0
τ = 0.1 , m=16
2
4
6
log2(S)
8
10
2
4
6
log2(S)
8
10
Fig S.9.8: Coverage probabilities for FY |X (Q(x0 ; τ )|x0 ) = τ under the linear
Gaussian location-scale shift model (S.9.2) for m = 16, n = 512 and nominal
coverage 0.95. Red: oracle asymptotic CI, black: bootstrap CI with B = 500.
S.9.3. Additional simulations for the linear model. The purpose
of the simulations in this section is to show the thresholds for the number
of sub-samples S ∗ and the number of quantile grids K ∗ that guarantee the
b ) predicted by the sufficient and necessary
oracle properties of β(τ ) and β(τ
conditions in the theorems in Section 3. Besides normal location shift models,
we also consider models with exponential error distributions. Using the linear
model (5.1) in Section 5, we show the coverage probabilities of the oracle
asymptotic confidence intervals, i.e. the confidence intervals based on the
true asymptotic variances. Similar to Section 5, we select τ = 0.1, 0.5, 0.9,
m = 4, 16, 32 and T = [0.05, 0.95]. Section S.9.3.1 contains results for β(τ )
defined in (2.3), while Section S.9.3.2 discusses FbY |X (y|x) defined in (2.8).
S.9.3.1. Oracle rule for β(τ ). Recall that from Corollary 3.3, an asymptotic 1 − α confidence interval for x>
0 β(τ ) is
q
>
−1
−1
−1
(S.9.5)
x0 β(τ ) ± N −1/2 fε,τ
τ (1 − τ )x>
0 ΣX x0 Φ (1 − α/2) ,
where fε,τ = f (F−1 (τ )), f is the error density and ΣX = E[(1, Xi> )> (1, Xi> )].
>
Here, we set x0 = (1, (m − 1)−1/2 l>
m−1 ) , where lm−1 is an (m − 1) vector
with each entry equal to 1.
78
We verify the oracle rule by checking whether the empirical coverage
probability of (5.3) equals 1 − α = 95%. Figure S.9.9 shows the coverage
probabilities for ε ∼ N (0, σ 2 ). In all plots, the coverage probabilities reach
the nominal level 95% for S < S ∗ , and then drop quickly to 0 for S > S ∗ for
some S ∗ . When N increases, S ∗ shifts toward N 1/2 , which is the sharp upper
bound in Corollary 3.3. Also, when N is fixed, a larger dimensionality m
leads to a smaller S ∗ . For ε ∼ Exp(λ), Figure S.9.10 shows that the coverage
probabilities are no longer symmetric in τ . When τ is small, N needs to be
large enough to attain the same magnitude of S ∗ due to the skewness of
Exp(λ).
S.9.3.2. Oracle rule for FbY |X (y|x). We compute FbY |X (y0 |x0 ) defined in
(2.8) discretely with an equidistant partition of size 1000 on [τL , τU ]. The
b ) is computed as in (2.5) with B being cubic B-spline with
involved β(τ
dimension q defined on G = 4+q knots. The knots form a partition on [τL , τU ]
with repetitions on the boundary; see Corollary 4.10 of Schumaker (1981).
For simplicity, we set y0 = Q(x0 ; τ ) so that FY |X (y0 |x0 ) = τ , where x0 is
chosen as in Section S.9.3.1. Recall from Corollary 3.5 that an asymptotic
1 − α confidence interval for FY |X (Q(x0 ; τ )|x0 ) = τ is
(S.9.6)
FbY |X (Q(x0 ; τ )|x0 ) ± N −1/2
q
−1
−1
τ (1 − τ )x>
0 ΣX x0 Φ (1 − α/2) .
We fix N = 214 in this section.
We demonstrate here the impact of number of subsamples S, model dimension m, number of basis q and number of quantile grid points K on
the coverage probability of (S.9.6). We note that q and K have similar impact on coverage probabilities given that other parameters are held fixed. In
Figure S.9.11 with ε ∼ N (0, 0.12 ), we note that an increase in q = dim(B)
improves the coverage probabilities given that m, S are held fixed. Also, at a
fixed S, q can be chosen smaller if m is larger. This is consistent with Corollary 3.10 that requires q N 1/(2η) kZ(x0 )k−η . Similar to Section S.9.3.1, an
increase of the dimensionality m leads to a smaller S ∗ for any fixed q. For
ε ∼ Exp(0.8) shown in Figure S.9.12, a difference to the normal case is that
the performance of the coverage probability is better when τ is small.
S.9.4. Nonparametric model. In this section, we consider a nonlinear model
(S.9.7)
Yi = 2.5 + sin(2Xi ) + 2 exp(−16Xi2 ) + 0.7i ,
Xi ∼ U(−1, 1),
i ∼ N (0, 1),
DISTRIBUTED INFERENCE FOR QUANTILE REGRESSION PROCESSES 79
where the function x 7→ 2.5 + sin(2x) + 2 exp(−16x2 ) is plotted in Figure
S.9.13. The basis Z is set as cubic B-spline defined at m+4 knots. The knots
form a partition on [−1, 1] with repetitions on the boundary; see Corollary
4.10 of Schumaker (1981). The model (S.9.7) implies the quantile of Y is
Q(x; τ ) = 2.5 + sin(2x) + 2 exp(−16x2 ) + 0.7Φ−1 (τ ) for 0 < τ < 1.
Section S.9.4.1 concerns the coverage probabilities of the confidence intervals for Q(x; τ ) for fixed τ . Section S.9.4.2 deals with the coverage probabilities of the confidence intervals for FY |X (y|x). In both sections, we fix
N = 216 .
S.9.4.1. Oracle rule for β(τ ). According to Corollary 3.9, an asymptotic
1 − α confidence interval for Q(x0 ; τ ) is
p
−1
−1
(S.9.8) Z(x0 )> β(τ ) ± N −1/2 φ−1
σ (Φσ (τ )) τ (1 − τ )σ0 (Z)Φ (1 − α/2) ,
where σ02 (Z) = Z(x0 )> E[Z(X)Z> (X)]−1 Z(x0 ) and φσ is the density of
N (0, σ 2 ).
Our simulation results show how S and m (the number of basis functions
m = dim(Z)) influence the coverage probabilities of (S.9.8). For all the
plots in Figure S.9.14, the coverage probabilities corresponding to m = 13
at x0 = 0 (the solid green curves) performs the worst. This is partly caused
by the large bias from the large curvature of Q(x0 ; τ ) at x0 = 0 (see Figure
S.9.13). This bias can be reduced by setting greater m, but the increase of
m results in a smaller S ∗ .
S.9.4.2. Oracle rule for FbY |X (y|x). We compute FbY |X (y|x) in a similar
way as Section S.9.3.2 by taking y0 = Q(x0 ; τ ), which gives FY |X (y0 |x0 ) = τ .
By Corollary 3.10, an 1 − α confidence interval for τ is
p
FbY |X (Q(x0 ; τ )|x0 ) ± N −1/2 τ (1 − τ )σ0 (Z)Φ−1 (1 − α/2) ,
(S.9.9)
where σ02 (Z) is defined as in (S.9.8).
We show the impact of number of subsamples S, model dimension m,
number of basis q and number of quantile grid points K on the coverage
probability of (S.9.9) in Figure S.9.15. We only show τ = 0.1 and omit the
results for τ = 0.5, 0.9, as they do not show additional insights to Figure
S.9.11. It can be seen from Figure S.9.15 that larger m lead to a smaller
estimation bias, and thus improves performance under fixed S, q and K.
However, such an increase of m also results in smaller S ∗ . Under a fixed
m, the influence from S, q and K on the coverage probabilities is similar to
the linear models with normal errors in Section S.9.3.2. These findings are
consistent with Corollary 3.10.
80
S.9.5. Practical conclusions from simulations. In this section, we
briefly comment on some practical recommendations regarding the choice
of S, K. Our general recommendation would be to choose S as small as
possible and K as large as possible since the only advantage of choosing
large S or small K are computational savings. Both theory and simulations
suggest that for models with more predictors, particular care in choosing S
is necessary. Giving very specific default recommendations is difficult since,
as seen from the simulations, the exact value of logN (S) when coverage
starts to drop below nominal values depends on the quantile at hand and
(unknown) conditional density of response given predictors.
As a rule of thumb, we would recommend to ensure that the ’effective subsample size’ n/m is at least 50 and that S does not exceed n/m. Additionally,
if the spread of conditional quantiles as a function of τ changes rapidly, extra
caution needs to be taken. This recommendation is based on Figure S.9.10
where the conditional density of response given predictors is exponential and
coverage starts to drop earlier for smaller quantiles, which corresponds to
regions where the second derivative of the conditional quantile function is
larger. This is also in line with our derivation of lower bonds for the oracle
rule, where the second derivative of the conditional quantile function enters
as constant a (see Section S.3.3.2)
Regarding K, K > N 1/4 worked well in all simulations we considered.
Again, the precise constant will depend on the roughness of the function
τ 7→ Q(x; τ ). The choice of G should depend on K. Although the formal
statement of our theorem requires K/G → ∞, a look at the proof reveals
that for cubic splines K/G > c for some constant c is sufficient. Simulations
indicate that G = K/2 combined with cubic splines is a reasonable choice.
DISTRIBUTED INFERENCE FOR QUANTILE REGRESSION PROCESSES 81
●
●
●
●
●
●
●
●
●
●
Normal(0, 0.12) , τ = 0.9
1.0
Normal(0, 0.12) , τ = 0.5
1.0
1.0
Normal(0, 0.12) , τ = 0.1
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
0.4
logN(S)
●
0.6
●
0.8
0.8
0.0
●
●
●
●
●
●
●
●
●
●
0.2
0.4
logN(S)
●
0.6
●
0.8
Normal(0, 1) , τ = 0.9
●
●
●
●
●
0.8
0.8
Coverage Probability
0.4
0.6
N = 214 , m = 4
N = 222 , m = 4
N = 214 , m = 16
N = 222 , m = 16
N = 214 , m = 32
N = 222 , m = 32
●
0.4
logN(S)
●
0.6
●
0.0
0.2
N = 214 , m = 4
N = 222 , m = 4
N = 214 , m = 16
N = 222 , m = 16
N = 214 , m = 32
N = 222 , m = 32
0.2
0.0
Coverage Probability
0.4
0.6
0.8
Coverage Probability
0.4
0.6
●
0.2
0.8
●
●
●
●
●
●
●
●
0.0
●
●
0.0
0.6
1.0
●
0.4
logN(S)
Normal(0, 1) , τ = 0.5
1.0
1.0
●
0.2
0.2
Normal(0, 1) , τ = 0.1
●
0.0
0.2
●
0.2
●
N = 214 , m = 4
N = 222 , m = 4
N = 214 , m = 16
N = 222 , m = 16
N = 214 , m = 32
N = 222 , m = 32
0.8
0.0
●
●
N = 214 , m = 4
N = 222 , m = 4
N = 214 , m = 16
N = 222 , m = 16
N = 214 , m = 32
N = 222 , m = 32
0.2
0.0
N = 214 , m = 4
N = 222 , m = 4
N = 214 , m = 16
N = 222 , m = 16
N = 214 , m = 32
N = 222 , m = 32
0.0
0.0
0.2
N = 214 , m = 4
N = 222 , m = 4
N = 214 , m = 16
N = 222 , m = 16
N = 214 , m = 32
N = 222 , m = 32
●
Coverage Probability
0.4
0.6
0.8
Coverage Probability
0.4
0.6
●
0.0
Coverage Probability
0.4
0.6
0.8
●
0.2
0.4
logN(S)
0.6
0.8
0.0
●
0.2
0.4
logN(S)
●
0.6
●
0.8
Fig S.9.9: Coverage
of the 1 − α = 95% confidence interval
for
>probability
−1
>
1/2
−1
−1
Q(x0 ; τ ) in (5.3): x0 β(τ )±fε,τ (τ (1−τ )x0 ΣX x0 /N ) Φ (1−α/2) under
> X + ε with β
2
the model Yi = 0.21 + βm−1
i
i
m−1 as (S.9.3) and ε ∼ N (0, σ ),
where N is the total sample size, S is the number of subsamples and m =
1 + dim(βm−1 ).
●
1.0
●
●
●
●
●
●
●
●
●
●
●
0.4
logN(S)
●
●
0.6
●
0.8
●
●
●
●
0.2
0.4
logN(S)
●
0.6
●
0.8
0.0
●
●
●
●
0.4
logN(S)
●
●
●
●
●
●
●
●
0.2
●
0.6
●
●
0.8
0.0
0.2
0.8
Coverage Probability
0.4
0.6
0.2
●
●
0.4
logN(S)
●
N = 214 , m = 4
N = 222 , m = 4
N = 214 , m = 16
N = 222 , m = 16
N = 214 , m = 32
N = 222 , m = 32
0.6
●
●
0.8
0.0
Coverage Probability
0.4
0.6
0.2
●
0.0
●
0.4
logN(S)
●
N = 214 , m = 4
N = 222 , m = 4
N = 214 , m = 16
N = 222 , m = 16
N = 214 , m = 32
N = 222 , m = 32
0.8
●
●
●
0.8
0.8
Coverage Probability
0.4
0.6
0.2
0.0
0.0
●
0.6
Exp (8) , τ = 0.9
1.0
●
●
0.2
●
●
●
●
Exp (8) , τ = 0.5
●
N = 214 , m = 4
N = 222 , m = 4
N = 214 , m = 16
N = 222 , m = 16
N = 214 , m = 32
N = 222 , m = 32
●
0.8
0.0
1.0
●
●
N = 214 , m = 4
N = 222 , m = 4
N = 214 , m = 16
N = 222 , m = 16
N = 214 , m = 32
N = 222 , m = 32
0.2
0.2
N = 214 , m = 4
N = 222 , m = 4
N = 214 , m = 16
N = 222 , m = 16
N = 214 , m = 32
N = 222 , m = 32
0.0
●
Coverage Probability
0.4
0.6
Coverage Probability
0.4
0.6
●
Exp (8) , τ = 0.1
●
●
●
0.8
0.8
Coverage Probability
0.4
0.6
0.0
0.2
N = 214 , m = 4
N = 222 , m = 4
N = 214 , m = 16
N = 222 , m = 16
N = 214 , m = 32
N = 222 , m = 32
0.2
●
●
●
0.0
●
●
0.0
●
1.0
Exp (0.8) , τ = 0.9
1.0
Exp (0.8) , τ = 0.5
1.0
Exp (0.8) , τ = 0.1
●
0.0
0.2
●
●
0.4
logN(S)
0.6
0.8
Fig S.9.10: Coverage probability under exactly the same setting as Figure
S.9.9, only except for ε ∼ Exp(λ).
82
(a) m = 4
1.0
Normal(0, 0.12) , m= 4 , y0 = Q( x0 ; τ = 0.9 )
1.0
Normal(0, 0.12) , m= 4 , y0 = Q( x0 ; τ = 0.5 )
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
0.8
0.8
●
●
●
0.8
1.0
Normal(0, 0.12) , m= 4 , y0 = Q( x0 ; τ = 0.1 )
●
●
●
0.0
●
●
0.0
0.2
0.4
logN(S)
0.6
●
0.0
●
0.2
0.2
●
0.8
Coverage Probability
0.4
0.6
q = 7 , K = 20
q = 7 , K = 30
q = 10 , K = 20
q = 10 , K = 30
q = 15 , K = 20
q = 15 , K = 30
●
●
0.0
●
q = 7 , K = 20
q = 7 , K = 30
q = 10 , K = 20
q = 10 , K = 30
q = 15 , K = 20
q = 15 , K = 30
0.2
●
0.0
q = 7 , K = 20
q = 7 , K = 30
q = 10 , K = 20
q = 10 , K = 30
q = 15 , K = 20
q = 15 , K = 30
Coverage Probability
0.4
0.6
Coverage Probability
0.4
0.6
●
0.2
0.4
logN(S)
0.6
0.8
0.0
●
0.2
0.4
logN(S)
0.6
●
●
0.8
(b) m = 16
●
●
●
●
●
Normal(0, 0.12) , m= 16 , y0 = Q( x0 ; τ = 0.9 )
1.0
1.0
Normal(0, 0.12) , m= 16 , y0 = Q( x0 ; τ = 0.5 )
1.0
Normal(0, 0.12) , m= 16 , y0 = Q( x0 ; τ = 0.1 )
●
●
●
●
●
●
●
●
●
●
●
●
0.8
0.8
0.8
●
●
0.2
0.4
logN(S)
●
0.6
0.8
0.0
q = 7 , K = 20
q = 7 , K = 30
q = 10 , K = 20
q = 10 , K = 30
q = 15 , K = 20
q = 15 , K = 30
0.2
0.2
●
0.0
0.0
0.0
q = 7 , K = 20
q = 7 , K = 30
q = 10 , K = 20
q = 10 , K = 30
q = 15 , K = 20
q = 15 , K = 30
0.0
●
Coverage Probability
0.4
0.6
Coverage Probability
0.4
0.6
●
●
q = 7 , K = 20
q = 7 , K = 30
q = 10 , K = 20
q = 10 , K = 30
q = 15 , K = 20
q = 15 , K = 30
0.2
Coverage Probability
0.4
0.6
●
●
●
0.2
0.4
logN(S)
0.6
0.8
0.0
●
●
0.2
0.4
logN(S)
●
0.6
0.8
(c) m = 32
●
●
●
●
●
●
●
●
●
●
●
●
●
●
0.8
●
0.8
0.8
●
●
●
●
●
Normal(0, 0.12) , m= 32 , y0 = Q( x0 ; τ = 0.9 )
1.0
1.0
Normal(0, 0.12) , m= 32 , y0 = Q( x0 ; τ = 0.5 )
1.0
Normal(0, 0.12) , m= 32 , y0 = Q( x0 ; τ = 0.1 )
0.0
0.2
0.2
●
0.4
logN(S)
●
●
0.6
0.8
Coverage Probability
0.4
0.6
q = 7 , K = 20
q = 7 , K = 30
q = 10 , K = 20
q = 10 , K = 30
q = 15 , K = 20
q = 15 , K = 30
0.0
0.0
0.2
q = 7 , K = 20
q = 7 , K = 30
q = 10 , K = 20
q = 10 , K = 30
q = 15 , K = 20
q = 15 , K = 30
0.0
0.2
●
q = 7 , K = 20
q = 7 , K = 30
q = 10 , K = 20
q = 10 , K = 30
q = 15 , K = 20
q = 15 , K = 30
0.2
●
0.0
Coverage Probability
0.4
0.6
Coverage Probability
0.4
0.6
●
●
0.4
logN(S)
0.6
0.8
0.0
0.2
●
0.4
logN(S)
●
●
0.6
0.8
Fig S.9.11: Coverage probability of the 1 − α = 95% confidence interval for
1/2 −1
−1
Φ (1 −
F (y0 |x0 ) in (S.9.6): FbY |X (Q(x0 ; τ )|x0 ) ± τ (1 − τ )x>
0 ΣX x0 /N
>
α/2) for Yi = 0.21 + βm−1 Xi + εi with βm−1 as (S.9.3) and ε ∼ N (0, 0.12 ),
where the total number of samples is N = 214 , S is the number of subsamples, m = 1 + dim(βm−1 ), q = dim(B) (corresponds to projection in τ
direction) and K is the number of the quantile grid points.
DISTRIBUTED INFERENCE FOR QUANTILE REGRESSION PROCESSES 83
(a) m = 4
●
Exp (0.8) , m= 4 , y0 = Q( x0 ; τ = 0.9 )
●
●
●
●
0.6
●
●
0.8
●
●
●
●
●
●
0.0
●
Coverage Probability
0.4
0.6
Coverage Probability
0.4
0.6
●
0.4
logN(S)
●
●
q = 7 , K = 20
q = 7 , K = 30
q = 10 , K = 20
q = 10 , K = 30
q = 15 , K = 20
q = 15 , K = 30
0.2
●
0.2
●
●
0.0
0.2
0.0
0.0
●
●
●
0.2
0.4
logN(S)
0.6
●
●
0.8
●
q = 7 , K = 20
q = 7 , K = 30
q = 10 , K = 20
q = 10 , K = 30
q = 15 , K = 20
q = 15 , K = 30
0.2
0.8
Coverage Probability
0.4
0.6
●
q = 7 , K = 20
q = 7 , K = 30
q = 10 , K = 20
q = 10 , K = 30
q = 15 , K = 20
q = 15 , K = 30
●
0.8
●
●
0.0
●
●
1.0
1.0
●
●
Exp (0.8) , m= 4 , y0 = Q( x0 ; τ = 0.5 )
0.8
1.0
Exp (0.8) , m= 4 , y0 = Q( x0 ; τ = 0.1 )
0.0
●
0.2
0.4
logN(S)
0.6
●
0.8
(b) m = 16
Exp (0.8) , m= 16 , y0 = Q( x0 ; τ = 0.9 )
1.0
●
●
●
●
●
●
●
●
●
●
●
●
●
●
0.8
0.8
●
0.8
●
Exp (0.8) , m= 16 , y0 = Q( x0 ; τ = 0.5 )
1.0
1.0
Exp (0.8) , m= 16 , y0 = Q( x0 ; τ = 0.1 )
●
●
0.0
●
●
●
0.4
logN(S)
●
0.6
0.8
0.0
q = 7 , K = 20
q = 7 , K = 30
q = 10 , K = 20
q = 10 , K = 30
q = 15 , K = 20
q = 15 , K = 30
0.2
0.2
●
0.2
Coverage Probability
0.4
0.6
Coverage Probability
0.4
0.6
●
q = 7 , K = 20
q = 7 , K = 30
q = 10 , K = 20
q = 10 , K = 30
q = 15 , K = 20
q = 15 , K = 30
●
0.2
●
0.4
logN(S)
0.0
0.0
0.2
q = 7 , K = 20
q = 7 , K = 30
q = 10 , K = 20
q = 10 , K = 30
q = 15 , K = 20
q = 15 , K = 30
0.0
Coverage Probability
0.4
0.6
●
●
0.6
0.8
0.0
●
●
0.2
0.4
logN(S)
0.6
0.8
(c) m = 32
●
●
●
●
Exp (0.8) , m= 32 , y0 = Q( x0 ; τ = 0.9 )
1.0
Exp (0.8) , m= 32 , y0 = Q( x0 ; τ = 0.5 )
1.0
1.0
Exp (0.8) , m= 32 , y0 = Q( x0 ; τ = 0.1 )
●
●
●
●
●
●
●
●
●
●
●
0.8
0.8
0.8
●
0.2
●
●
0.4
logN(S)
●
0.0
●
0.6
0.8
0.0
0.2
Coverage Probability
0.4
0.6
●
●
0.4
logN(S)
q = 7 , K = 20
q = 7 , K = 30
q = 10 , K = 20
q = 10 , K = 30
q = 15 , K = 20
q = 15 , K = 30
0.2
0.2
0.0
0.0
q = 7 , K = 20
q = 7 , K = 30
q = 10 , K = 20
q = 10 , K = 30
q = 15 , K = 20
q = 15 , K = 30
0.2
●
q = 7 , K = 20
q = 7 , K = 30
q = 10 , K = 20
q = 10 , K = 30
q = 15 , K = 20
q = 15 , K = 30
0.0
Coverage Probability
0.4
0.6
Coverage Probability
0.4
0.6
●
●
0.6
0.8
0.0
0.2
●
●
0.4
logN(S)
0.6
0.8
1.5
2.0
2.5
3.0
3.5
4.0
4.5
Fig S.9.12: Coverage probability under exactly the same setting as Figure
S.9.11, only except for ε ∼ Exp(λ).
-1.0
-0.5
0.0
0.5
1.0
x
Fig S.9.13: The plot for the function x 7→ 2.5+sin(2x)+2 exp(−16x2 ), which
is used in model (S.9.7).
84
●
●
●
●
1.0
τ = 0.9
1.0
τ = 0.5
1.0
τ = 0.1
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
0.1
0.2
●
0.3
logN(S)
0.4
0.5
0.0
0.1
0.2
●
0.2
x0 = −0.9 , m = 13
x0 = −0.9 , m = 17
x0 = −0.9 , m = 21
x0 = 0 , m = 13
x0 = 0 , m = 17
x0 = 0 , m = 21
x0 = 0.9 , m = 13
x0 = 0.9 , m = 17
x0 = 0.9 , m = 21
0.0
0.2
x0 = −0.9 , m = 13
x0 = −0.9 , m = 17
x0 = −0.9 , m = 21
x0 = 0 , m = 13
x0 = 0 , m = 17
x0 = 0 , m = 21
x0 = 0.9 , m = 13
x0 = 0.9 , m = 17
x0 = 0.9 , m = 21
0.0
0.0
0.2
x0 = −0.9 , m = 13
x0 = −0.9 , m = 17
x0 = −0.9 , m = 21
x0 = 0 , m = 13
x0 = 0 , m = 17
x0 = 0 , m = 21
x0 = 0.9 , m = 13
x0 = 0.9 , m = 17
x0 = 0.9 , m = 21
0.0
0.8
Coverage Probability
0.4
0.6
Coverage Probability
0.4
0.6
●
Coverage Probability
0.4
0.6
0.8
0.8
●
0.3
logN(S)
0.4
0.5
0.0
0.1
0.2
●
0.3
logN(S)
0.4
0.5
Fig S.9.14: Coverage probability of the 1 − α
=
95%
>
confidence interval for Q(x0 ; τ ) in (S.9.8):
Z(x0 ) β(τ ) ±
p
−1
−1/2
−1
−1
N
φσ (Φσ (τ )) τ (1 − τ )σ0 (Z)Φ (1 − α/2) under the nonlinear
model (S.9.7), where m = dim(Z) (corresponds to basis expansion in x)
and N = 216 .
(a) m = 17
1.0
m=17, y0 = Q( x0 = 0.9 ; τ = 0.1 )
1.0
m=17, y0 = Q( x0 = 0 ; τ = 0.1 )
1.0
m=17, y0 = Q( x0 = −0.9 ; τ = 0.1 )
●
●
0.1
0.2
●
0.2
0.3
logN(S)
0.4
0.5
●
●
●
●
●
●
q = 7 , K = 20
q = 7 , K = 30
q = 10 , K = 20
q = 10 , K = 30
q = 15 , K = 20
q = 15 , K = 30
0.0
0.0
0.0
●
●
0.0
0.1
0.2
0.3
logN(S)
0.4
0.5
●
●
●
●
●
q = 7 , K = 20
q = 7 , K = 30
q = 10 , K = 20
q = 10 , K = 30
q = 15 , K = 20
q = 15 , K = 30
0.2
●
0.0
Coverage Probability
0.4
0.6
●
●
Coverage Probability
0.4
0.6
●
q = 7 , K = 20
q = 7 , K = 30
q = 10 , K = 20
q = 10 , K = 30
q = 15 , K = 20
q = 15 , K = 30
0.2
Coverage Probability
0.4
0.6
●
●
●
0.8
0.8
0.8
●
●
●
●
●
●
0.0
0.1
●
0.2
0.3
logN(S)
0.4
0.5
(b) m = 23
●
0.0
0.0
0.1
●
0.2
0.3
logN(S)
0.4
0.5
●
●
0.8
●
0.0
0.1
●
●
0.2
0.3
logN(S)
0.4
0.5
●
●
●
●
●
●
q = 7 , K = 20
q = 7 , K = 30
q = 10 , K = 20
q = 10 , K = 30
q = 15 , K = 20
q = 15 , K = 30
0.2
q = 7 , K = 20
q = 7 , K = 30
q = 10 , K = 20
q = 10 , K = 30
q = 15 , K = 20
q = 15 , K = 30
0.2
0.2
q = 7 , K = 20
q = 7 , K = 30
q = 10 , K = 20
q = 10 , K = 30
q = 15 , K = 20
q = 15 , K = 30
●
●
Coverage Probability
0.4
0.6
Coverage Probability
0.4
0.6
●
●
0.0
●
●
●
●
0.0
Coverage Probability
0.4
0.6
●
●
0.8
0.8
●
●
●
●
●
●
1.0
m=23, y0 = Q( x0 = 0.9 ; τ = 0.1 )
1.0
m=23, y0 = Q( x0 = 0 ; τ = 0.1 )
1.0
m=23, y0 = Q( x0 = −0.9 ; τ = 0.1 )
0.0
0.1
●
0.2
0.3
logN(S)
0.4
0.5
Fig S.9.15: Coverage probability
of the 1 − α = p
95% confidence interval
−1/2
b
for F(y0 |x0 ) in (S.9.9): FY |X (Q(x0 ; τ )|x0 ) ± N
τ (1 − τ )σ0 (Z)Φ−1 (1 −
α/2) under the nonlinear model (S.9.7), where m = dim(Z) (corresponds
to basis expansion in x), S is the number of subsamples, N = 216 , q =
dim(B) (corresponds to projection in τ direction) and K is the number of
the quantile grid points.
| 10math.ST
|
APPROXIMATE SUBLOOPS AND FREIMAN’S THEOREM IN
FINITELY GENERATED COMMUTATIVE MOUFANG LOOPS
arXiv:1802.10471v1 [math.GR] 28 Feb 2018
ARINDAM BISWAS
Abstract. Fix a parameter K > 1. A K-approximate subgroup is a finite set A in a
group G which contains the identity, is symmetric and such that A.A can covered by K left
translates of A. This article deals with the generalisation of the concept of approximate
groups in the case of loops which we call approximate loops and the description of Kapproximate subloops when the ambient loop is a finitely generated commutative moufang
loop. Specifically we have a Freiman type theorem where such approximate subloops are
controlled by arithmetic progressions defined in the commutative moufang loops.
1. Introduction
This article deals with the study of sets of small tripling in non-associative, commutative
Moufang loops. We study the connections between groups, quasigroups, loops and combinatorics. We shall make the relevant ideas more precise in the following sections but first let
us recall the known literature in the case of groups.
1.1. History and Background. The formal study of the structure of approximate subgroups began with the celebrated theorem of Freiman
Theorem 1.1 (Freiman [Fre64]). Let A ⊆ Z be a finite set of integers with small sumset
|A + A| 6 α|A|
Then A ⊆ P where P is a d−dimensional arithmetic progression with d 6 d(α) and |P | 6
C(α)|A| (i.e., the length and dimension of the progression is a constant depending only on
α).
The above was a qualitative statement about the subsets of integers. The quantitative
bounds on the dimension and size of the progression P were established by Chang in [Cha02].
A natural question was the extension of this result for arbitrary abelian groups. However
Z is a torison free abelian group and it was clear that Freiman’s theorem cannot hold exactly
as stated for Z, since if G is a finite abelian group of high rank then G itself satisfies the
small doubling condition for every α > 1, but it cannot be contained in any set P of the
form {x0 + l1 x1 + ... + lr xr : 0 6 li 6 Li } and r 6 C(α).
Green and Ruzsa showed that this is essentially the only hindrance to Freiman’s theorem
in an arbitrary abelian group in the sense that if G is such a group and A ⊂ G has doubling
constant α then A ⊂ H + P where H is a subgroup and P is an arithmetic progression with
|H + P | 6 C(α)|A|. H + P is called a coset progression.
Key words and phrases. approximate groups, approximate loops, commutative Moufang loops.
1
2
ARINDAM BISWAS
Theorem 1.2 (Green-Ruzsa [GR07]). Suppose that X is a symmetric subset of an abelian
group G with doubling constant at most K. Then there is a subgroup H ⊂ 4X and an
arithmetic progression P = {l1 x1 + ... + lr xr : 1 6 li 6 Li , xi ∈ G ∀i = 1, ..., r} of rank r at
most K O(1) such that
xi ∈ 4X (i = 1, ..., r)
and
X ⊂ H + P ⊂ K O(1) X
Extension of Freiman’s theorem for arbitrary groups and also establishing better quantitative bounds for particular classes of groups rapidly followed. Some of the notable works
in this direction are by Tao - [Tao10], Sanders - [San12], Breuillard,Green,Tao - [BGT12],
Tointon - [Toi14] etc to name a few. As our goal here is different so we shall not go into
further details in those directions.
The formal definition of an approximate group was introduced by Tao in [Tao08] and
was in part motivated by its use in the work of Bourgain-Gamburd [BG08] on super-strong
approximation for Zariski dense subgroups of SL2 (Z). Approximate groups were also used
extensively in Helfgott’s seminal paper [Hel08].
Definition 1.3 (Approximate subgroup). Let G be some group and K > 1 be some parameter. A finite set A ⊆ G is called a K-approximate subgroup if
(1) Identity of G, e ∈ A.
(2) It is symmetric, i.e. if a ∈ A then a−1 ∈ A.
(3) There is a symmetric subset X lying in A.A with |X| 6 K such that A.A ⊆ X.A
The notion of groups can be generalised to loops. Briefly speaking, a loop is a quasi-group
with an identity element but not necessarily being associative. However loops are considered
to have very little in the way of algebraic structure and it is for that reason we sometimes limit
our investigations to loops which satisfy a weak form of associativity. Common examples are
the Moufang loops. We leave the general defintion of a Moufang loop for the next section
and state the definition of a commutative moufang loop directly.
Definition 1.4 (commutative moufang loop). A loop L is called a commutative moufang
loop if
(1.1)
xy = yx,
2
x (yz) = (xy)(xz)
are satisfied for every x, y, z ∈ L.
In this article we generalize the notion of approximate subgroups in the case of loops. We
call them approximate subloops (see 2.14) and show a structure theorem for approximate
subloops of finitely generated commutative moufang loops. We shall formally define all the
terms in the next section.
1.2. Main Result. (Freiman’s Theorem in finitely generated CML) Let M be a n generated
commutative moufang loop. Let A be a K-approximate subloop (2.14) of M. Then A2
is contained in a coset progression in M of dimension (rank) (K|M ′ |)O(1) and of size of
progression at most exp((K|M ′ |)O(1) )|A2 |.
APPROXIMATE SUBLOOPS AND FREIMAN’S THEOREM IN FINITELY GENERATED COMMUTATIVE MOUFANG LOOPS
3
1.3. Outline of the paper. The paper is divided into the following sections.
(1) Introduction
(2) Preliminaries
(3) Properties of associator subloops
(4) Progressions in loops - Here we define the effective notion of arithmetic progressions
in CMLs.
(5) Freiman’s theorem in CMLs - This section contains the lemmas and propositions for
the main result of Freiman’s theorem for finitely generated CMLs
(6) Concluding remarks
1.4. Acknowledgements. I wish to thank Emmanuel Breuillard for a number of helpful
discussions and advice on this subject.
2. Preliminaries
To begin with we state the definitions and properties of groupoids, quasigroups, loops and
related structures.
2.1. Groupoids, Quasigroups, Loops, Moufang loops.
Definition 2.1 (Groupoids). Fix a positive integer n. An n-ary groupoid (G, T ) is a nonempty set G together with an n-ary operation T defined on G.
The order of an n-ary groupoid (G, T ) is the cardinality of G. An n-ary groupoid is said
to be finite whenever its order is finite.We shall be dealing with binary groupoids and then
we denote the operation T as ∗.
Definition 2.2 (Translation maps). Let (G, ∗) be a groupoid and let a be a fixed element in
G. Then the translation maps L(a) : G → G and R(a) : G → G for each a ∈ G are defined
by
L(a)x = a ∗ x, R(a)x = x ∗ a ∀x ∈ G
Definition 2.3 (Cancellation groupoids). A groupoid (G, ∗) is called left cancellation (resp.
right cancellation) if the left (resp. right) translation map L(a) (resp. R(a)) is injective for
any a ∈ G : a ∗ x = a ∗ y ⇒ x = y for all a, x, y ∈ G (resp. x ∗ a = y ∗ a ⇒ x = y for all
a, x, y ∈ G).
A groupoid (G, ∗) is called cancellation if it is both left and right cancellation.
Definition 2.4 (Division groupoids). A groupoid (G, ∗) is called left division (resp. right
division) if the left (resp. right) translation map L(a) (resp. R(a)) is surjective for any
a ∈ G : a ∗ x = b has solutions for any ordered pair of elements a, b ∈ Q (y ∗ a = b has
solutions for any ordered pair of elements a, b ∈ Q).
A groupoid (G, ∗) is called division groupoid if it is both left and right division.
Definition 2.5 (Idem groupoids). An element x such that x ∗ x = x is called an idempotent
element of the binary groupoid (G, ∗). A groupoid which has idempotent elements is called
an idem groupoid.
Definition 2.6 (Quasigroup). An n-ary groupoid (G, T ) with n-ary operation T is called
a quasigroup if in the equality T (x1 , x2 , ...xn ) = xn+1 , knowledge of any n elements among
x1 , x2 , ..., xn , xn+1 ∈ G uniquely specifies the remaining one.
4
ARINDAM BISWAS
In binary case, this is equivalent to the following:
Definition 2.7 (Binary quasigroup). Binary groupoid (Q, ∗) is called a binary quasigroup
if for all ordered pairs (a, b) ∈ Q2 there exist unique solutions x, y ∈ Q to the equations
x ∗ a = b and a ∗ y = b.
From here onwards we shall be only concerned with binary operations.
It is easy to see the following equivalent criteria for a quasigroup.
Remark 2.8. A groupoid (G, ∗) is a quasigroup iff the maps L(a) : G → G, R(a) : G → G
are bijections for all a ∈ G.
From now on we shall drop the reference to the binary operation ∗ and denote algebraic
objects (Q, ∗) as Q and operations a ∗ x = y as ax = y.
Definition 2.9 (Loop). A quasigroup Q is a loop if Q possesses a neutral element e i.e., if
there exists e such that
ae = a = ea
for every a ∈ Q.
Clearly the neutral element if it exists is idempotent so loops are idem groupoids. Also
the neutral element is unique (but there might exist other idempotent elements).
Neither quasigroups nor loops are necessarily associative and so care needs to be taken
when writing down complex expressions. We employ the following evolution rules, juxtaposition has the highest priority followed by . and then parentheses.
One usually studies loops satisfying some weak form of associativity. This leads us to the
notion of moufang loops
Definition 2.10 (Moufang loop). A loop L is called a Moufang loop if the Moufang identities
xy.zx = x(yz.x),
(2.1)
x(y.xz) = (xy.x)z,
x(y.zy) = (x.yz)y
are satisfied for every x, y, z ∈ L.
It must be noted that any one of the three identities implies the other two. Our main
object of interest will be commutative moufang loops. For commutative moufang loops the
above set of identities reduce to a single one
x2 (yz) = (xy)(xz)
Let us give some examples.
Example 2.11. To begin with any group is trivially a Moufang loop. For the non-trivial
case, we first give examples of non-commutative moufang loops and then commutative ones.
(1) Octonions over the real numbers form a non-commutative, non-associative moufang
loop.
APPROXIMATE SUBLOOPS AND FREIMAN’S THEOREM IN FINITELY GENERATED COMMUTATIVE MOUFANG LOOPS
5
α a
(2) (Zorn’s vector matrix algebra). Let T be the set of matrices
with real scalars
b β
α, β and real 3-vectors a, b. Define the product of two elements of T by
α a
γ c
α.γ + a.d
αc + δa − b × d
.
=
.
b β
d δ
γb + βd + a × c
b.c + βδ
using the scalar product x.y and cross-product x × y of 3-vectors. Let us define the
determinant of a matrix by
α a
det
= αβ − a.b
b β
Let Q be the subset of T consisting of matrices whose determinant is 1. Let us set
1 0
I=
.
0 1
Then (Q, ., I) is a non-commutative and non-associative Moufang loop.
(3) (Zassenhaus’s Commutative Moufang loop) Let F3 denote the finite field with 3 elements. Let Q be the set F43 . Define a new multiplication ◦ on F43 for x = (x1 , x2 , x3 , x4 )
and y = (y1 , y2 , y3, y4 ), by
x ◦ y = x + y + (0, 0, 0, (x3 − y3 )(x1 y2 − x2 y1 )).
Then (Q, ◦) is a commutative moufang loop that is not associative.
(4) This construction is due to J.D.H. Smith [Smi78a] and [Smi78b].
Let G be a group of exponent 3 with involution operator σ, M(G) = {x ∈ G : xσ =
x−1 }. Then M(G) is a commutative moufang loop of exponent 3 (i.e, x3 = 1∀ x ∈
M(G)) with respect to multiplication defined by x ∗ y = x−1 yx−1 . It is also clear that
every 3 exponent CML can be obtained in this way.
Proof. It is clear to see that M(G) is a loop with identity element 1 ∈ G. Also the
Moufang identity is satisfied,
x2 ∗ (y ∗ z) = x−1 ∗ (y −1 zy −1 ) = x(y −1 zy −1 )x
= (xy −1 x)(x−1 zx−1 )(xy −1 x) = (x−1 yx−1 ) ∗ (x−1 zx−1 ) = (x ∗ y) ∗ (x ∗ z)
(5) This is Bruck [Bru46] and has also been mentioned by L.Bénétau in [Bén84]. A group
G is said to be 3-abelian if for all x, y ∈ G we have x3 y 3 = (xy)3 identically. In such
a group, any cube x3 lies in the centre Z(G) of the group.
(Generic CML) If G is a non-trivial 3-abelian group ,
(a) the binary law x ∗ y = x−1 yx2 makes the set G a commutative moufang loop
(denoted by G∗ ).
(b) the set product (Z/3Z) × G = {(p, x) : p ∈ Z/3Z, x ∈ G} together with the law
of composition :
(p, x)˜∗(q, y) = (p + q, zq−p (x, y))
where z−1 (x, y) = yx, z0 (x, y) = x−1 yx2 , z1 (x, y) = xy becomes a CML G̃∗ which
contains {0} × G ≃ G∗ as a maximal subloop of index 3.
If we suppose that G is non-abelian then the associative centre of G̃∗ consists of
elements of the form (0, z) where z ∈ Z(G) while if G is abelian then G̃∗ = Z×G
(c) When G is of exponent 3t , the loops G∗ and G̃∗ also have the same form of
exponent.
6
ARINDAM BISWAS
2.2. Approximate groups and approximate loops. To motivate the discussion on approximate loops, let us first recall Tao’s definition of approximate groups already stated in
the introduction 1.3.
A finite set A ⊆ G is called a K-approximate subgroup if
(1) Identity of G, e ∈ A.
(2) It is symmetric, i.e. if a ∈ A then a−1 ∈ A.
(3) There is a symmetric subset X lying in A.A with |X| 6 K such that A.A ⊆ X.A
Remark 2.12. From the above definition we see that the property of associativity is not
needed. But in the study of approximate groups we are mostly interested in the growth of
iterated powers of sets. For example A2 ⊂ XA in G automatically guarantees the fact
that An ⊂ X n−1 A in G which implies that |An | 6 K n−1 |A|. In a non-associative loop it
is not clear how to define the powers of sets. The fact that AA ⊂ XA doesn’t guarantee
A(AA) ⊂ (X 2 )A and also the balls A3 , A4 , ..., An are not well defined, let alone the growth of
them to be bounded in terms of a constant times |A| (an essential condition for the study of
approximate sets). To fix this problem we define the powers of a set A in a loop L as follows.
2n
1
Notation 1. Let L be a loop, n ∈ N and A ∈ L be a finite set. Let Cn = n+1
denote
n
the nth Catalan number, which represents the number of ways of inserting parentheses in
the expression A.A...A
| {z }. Let each Bi for i = 1, ..., Cn represent one such set. Then we define
n+1 times
n+1
n
n
:= ∩C
Ahn+1i := ∪C
k=1 Bk .
k=1 Bk and A
For example if n = 3 we have C3 = 5, B1 = (A2 ).(A2 ), B2 = ((A2 )A)A, B3 = (A(A2 ))A, B4 =
A((A2 )A), B5 = A(A(A2 )) and
Ah4i = ((A2 ).(A2 )) ∪ (((A2 )A)A) ∪ ((A(A2 ))A) ∪ (A((A2 )A)) ∪ (A(A(A2 )))
A4 = ((A2 ).(A2 )) ∩ (((A2 )A)A) ∩ ((A(A2 ))A) ∩ (A((A2 )A)) ∩ (A(A(A2 )))
Remark 2.13.
(1) If 1 ∈ A then A2 ∈ An and hence An is non-empty.
(2) In the case of groups we have Ahni = An .
(3) In a commutative moufang loop we have the identification B2 = B3 = B4 = B5 and
hence the number reduces.
We are now in a position to define approximate loops.
Definition 2.14 (Approximate subloops). Let L be a loop with e the neutral element. Let
K > 1 be a parameter and A ⊆ L be a finite set. We say that A is a K approximate subloop
of L if
(i) 1 ∈ A,
(ii) Symmetricity - A is symmetric: A = A−1 ,
(iii) Bounded growth condition - |Ah3i | < K|A|.
The definition is consistent as the usual operation of taking inverses is valid in loops in
general. So for all subsets S ⊆ L we know that S −1 exists and |S −1 | = |S|. In this chapter we
shall be mainly concerned with the growth of approximate subloops of commutative moufang
loops (CMLs in short).
APPROXIMATE SUBLOOPS AND FREIMAN’S THEOREM IN FINITELY GENERATED COMMUTATIVE MOUFANG LOOPS
7
Remark 2.15. In the case of approximate groups bounded growth condition and subset criterion are equivalent. We have seen that the subset criterion implies bounded growth of sets
An ∀n ∈ N. The other direction can be seen from - If |A3 | 6 K|A|, then B := (A∪A−1 ∪{1})2
is a O(K O(1) )-approximate group. Later we shall see how this result can be extended in the
setting of finitely generated commutative moufang loops.
2.3. Properties of moufang loops.
2.4. Moufang loops. As a starting point we give some definitions and recall some properties
of Moufang loops.
Definition 2.16 (Associator). Let L be a loop and x, y, z ∈ L. Then the associator of x, y, z
denoted by (x, y, z) is defined by
xy.z = (x.yz).(x, y, z)
or equivalently
(x, y, z) := (x.yz)−1 (xy.z)
The associator measures the deviation of a loop from associativity.
Definition 2.17 (Commutator). In a loop L, the commutator of x, y ∈ L , denoted by [x, y]
is defined as
[x, y] := x−1 y −1 xy
The commutator measures the deviation of a loop from commutativity.
Definition 2.18 (Nucleus or centre). The nucleus Z(M) of a moufang loop M (called centre
in case of CMLs) is the set of elements in M which commute and associate with all other
elements in M,
Z(M) := {z ∈ M : [z, x] = 1, (z, x, y) = 1∀ x, y ∈ M}
Remark 2.19. The centre Z(M(G)) of the moufang loop 4 is the set
{p ∈ M(G) : (p, x, y) = 1∀ x, y ∈ G} = Z(G)
Definition 2.20 (power associativity). A groupoid is said to be power associative if every
element generates an associative subgroupoid.
Due to power associativity, the expression xn has a unique interpretation for every nonnegative integer n and every x ∈ G. In a power associative loop the identity is the unique
idempotent element.
Definition 2.21 (diassociativity). A groupoid is said to be diassociative if every two elements
generate an associative subgroupoid.
Due to diassociativity, we may omit parentheses in expressions involving only powers of
two elements.
Proposition 2.22. In a diassociative quasigroup if there exists an idempotent element then
it is unique.
Proof. Let us assume that there exists two idempotent elements x and y. Then x2 = x
and y 2 = y. Consider the well defined product xy 2 . We get that xy 2 = xy which implies
xy = x. Similarly considering x2 y we get that xy = y (we use the property of cancellation
in quasigroups). Hence x = y.
8
ARINDAM BISWAS
Definition 2.23 (Homomorphisms and kernels). Let M1 and M2 be moufang loops. A single
valued mapping θ of M1 into M2 is said to be a homomorphism of moufang loops if
θ(ab) = θ(a)θ(b) ∀a, b ∈ M1
The kernel K of θ is the set {k ∈ M1 : θ(k) = 1 ∈ M2 }
Definition 2.24 (semi-endomorphism). A single valued mapping θ of the Moufang loop G
into itself is called a semi-endomorphism of G provided that
θ(xyx) = (θx)(θy)(θx), θ1 = 1.
for all x, y ∈ G.
Definition 2.25 (subloops and normal subloops). Let M be a loop. A subset H ⊆ M is a
subloop of M iff H ⊆ M and a.b ∈ H for all a, b ∈ H. A subloop N of M which satisfies
xN = Nx, (Nx)y = N(xy), y(xN) = (yx)N
is called a normal subloop of M. We shall denote N normal in M by N ⋖ M
Remark 2.26. We note that
(1) The intersection of subloops of a loop M is also a subloop of M.
(2) In the loop M for A ⊆ M, the subloop generated by A is denoted by hAi and defined
to be the intersection of all subloops of M containing A. It is the smallest subloop
containing A.
(3) Any subgroup of the nucleus Z(M) of a loop M is a normal subloop of M.
(4) Let N be a normal subloop of M. The right cosets Nx of M modulo N partition M.
That is
y ∈ Nx ⇔ Ny = Nx
Definition 2.27 (quotient loops). Let M be a CML and N ⋖ M. Then M/N has the
structure of a loop and is called the quotient loop of M modulo N. The right cosets Nx of
M mod N partitions M and x ≡ y mod N iff xy −1 ∈ N.
The fundamental theorem for Moufang loops is the following which ensures that we don’t
need to bother to give parentheses when evaluating expressions in moufang loops involving
only two elements and their powers. This was proved by Moufang in [Mou35] simultaneously
for Moufang loops and alternative division rings.
Theorem 2.28 (Moufang’s theorem). Moufang loops are diassociative.
Proof. See [Mou35].
2.5. Properties of commutative moufang loops. The following lemma forms the basis
for calculation in commutative moufang loops. It can be found in any standard text on
commutative moufang loops (for example [Bru58]) but since we shall be using the identities
repetitively so we state it here.
Lemma 2.29 (Identities involving associators in case of commutative moufang loops). In a
commutative moufang loop M with x, y, z, w ∈ M the associator has the following properties
(1) (x, y, z) = (y, z, x) = (y, x, z)−1
(2) (x−1 , y, z) = (x, y, z)−1 and ∀ m, n, p ∈ Z, (xm , y n , z p ) = (x, y, z)mnp
(3) (xy, z, w) = {(x, z, w)(y, z, w)}{((x, z, w), x, y).((y, z, w), y, x)}.
(4) (x, y, z)3 = 1
APPROXIMATE SUBLOOPS AND FREIMAN’S THEOREM IN FINITELY GENERATED COMMUTATIVE MOUFANG LOOPS
9
Proof. The proof depends on several associator results. Let L(x, y) and R(x, y) denote the
mappings
L(x, y) = L(yx)−1 L(y)L(x), R(x, y) = R(yx)−1 R(y)R(x)
We have L(z, y)(x) = x(x, y, z)−1 .
Proof :- By definition, (xy)z = [x(yz)](x, y, z). Hence (x, y, z)−1 = [z −1 (y −1 x−1 )][x(yz)] and
x(x, y, z)−1 = {x[z −1 (y −1x−1 )]x}(yz) = [(xz −1 )y −1](yz) = R(z −1 , y −1)(x) = L(z, y)(x) .
Since Moufang loops are power associative for every semi-endomorphism (2.24 ) θ and for
all x ∈ M we have θ(xn ) = (θ(x))n . Now L(z, y) is a semi-endomorphism of M and hence we
have xn (xn , y, z)−1 = [x(x, y, z)−1 ]n for all integers n. Taking n = −1 we have the first result
of (2) and (xn , y, z)−1 = (x, y, z)−n from which we get that (xn , y, z) = (x, y, z)n ∀n ∈ Z.
The rest of (2) follows.
For (1) we first prove the fact that (x, y, z) = (x, yz, z) = (x, y, zy). This is a direct
consequence of L(x, y) = L(xy, y) = L(x, yx) and L(z, y)x = x(x, y, z)−1 . Using this we
get that (x, y, z) = (xy, z, y)−1. By (2) we have (x, y, z) = (x−1 , y, z)−1 = (x−1 y, z, y) =
(y −1x, z, y)−1 = (y −1x, z, x)−1 = (y −1, x, z) = (y, x, z)−1 ; this, applied to (x, y, z) = (x, z, y −1 )−1 ,
gives (x, y, z) = (x, z, y −1 ) = (z, x, y −1 )−1 = (z, y, x)−1 = (y, z, x).This shows (1).
To show (3), we first prove the following lemma :
Lemma 2.30. If w, x, a, b are elements of a commutative Moufang loop M, then (wa)(xb) =
(wx)c where c = pq −1 and
p = [R(w, x)a][R(x, w)b] = [a(a, w, x)][b(b, x, w)], q = (w −1 x, R(w, x)a, R(x, w)b)
Proof. We multiply the equation (wx)c = (wa)(xb) by w −2 to get x(w −1 c) = a[w −1 (xb)]. If
θ = R(w, x) then R(w −1 , x) = θ−1 , R(x, w −1 ) = θ. Hence x(w −1 c) = (xw −1 )(θ−1 (c)) and
w −1 (xb) = (w −1 x)(θ(b)), so the equation becomes (w −1 x)(θ−1 (c)) = a[(w −1 x)(θ(b))]. Also
θ3 = I gives us θ2 = θ−1 . Moreover, θ leaves w and x fixed. Hence
(w −1 x)c = θ[(w −1 x)(θ−1 (c))] = θ{a[(w −1 x)(θ(b))]} = (θ(a))[(w −1 x)(θ−1 (b))] = [(w −1 x)(θ−1 (b))](θ(a)).
From here we get
(w −1 x)c = (w −1 x){[(θ−1 (b))(θ(a))](w −1 x, θ−1 (b), θ(a))},
whence
c = [(θ(a))(θ−1 (b))] × (w −1 x, θ(a), θ−1 (b))−1 .
Since θ = R(w, x), θ−1 = R(x, w), the proof is complete
Now we deduce the expansion formula for (wx, y, z). Since φ = R(y, z) is an automorphism,
(wx)(wx, y, z) = φ(wx) = (φ(w))(φ(x)) = [w(w, y, z)][x(x, y, z)]. We apply the above lemma
with a = (w, y, z), b = (x, y, z), c = (wx, y, z). In this case, replacement of y by y −1 replaces
a, b, c, p by their inverses but leaves q fixed. Hence we have both c = pq −1 and c−1 = p−1 q −1 .
By multiplication, 1 = q −2 . However, q 3 = 1 and so q = 1. Therefore c = p, 1 = q that is,
(wx, y, z) = [R(w, x)(w, y, z)][R(x, w)(x, y, z)] = [(w, y, z)((w, y, z), w, x)][(x, y, z)((x, y, z), x, w)]
To show (4), we use the fact that (xy)3 = x3 y 3 by Moufang’s theorem and commutativity
for all x, y ∈ M. For each x in M, the mapping T (x) = R(x)L(x)−1 = I and also T (x) is a
10
ARINDAM BISWAS
pseudo-automorphism 1 with companion x−3 . So x3 lies in the centre of G. The kernel of the
endomorphism x → x3 must contain G′ which shows that (x, y, z)3 = 1 for all x, y, z ∈ M.
Theorem 2.31 (Bruck-Slaby). If M is a commutative moufang loop generated by n elements
then M is centrally nilpotent and if n > 2 then the nilpotency class (size of the lower central
series) is bounded by n − 1.
Proof. See [Bru58].
3. Associator subloops
In this section we collect some of the important definitions and results concerning associator subloops following Bruck [Bru58] which we shall need later. Let us first recall the
definition of a normal subloop of a CML M. N is said to be normal in M (denoted by
N ⋖ M) if it’s a subloop and it satisfies
xN = Nx, (Nx)y = N(xy), y(xN) = (yx)N
for all x, y ∈ M.
Definition 3.1 (associator subloops, central series and derived series). If A, B, C are normal
subloops of a commutative moufang loop M one defines the associator subloop of A, B, C
denoted by h(A, B, C)i [or sometimes just as (A, B, C) when it’s clear from the context] to
be the subloop generated by all associators (a, b, c) with a ∈ A, b ∈ B, c ∈ C.
The lower central series {Mα } of a commutative moufang loop M is defined by
M0 = M, Mi+1 = h(Mi , M, M)i, i = 0, 1, ... ;
and the upper central series {Zα },
Z0 = 1, Zi+1 /Zi = Z(M/Zi ) ∀i = 0, 1, ... ;
while the derived series is defined as
M (0) = M, M (i+1) = (M (i) )′ = h(M (i) , M (i) , M (i) )i, i = 0, 1, ... ;
The first lemma is the normality of the associator subloop h(A, B, C)i when A, B, C are
normal in M.
Lemma 3.2. Let M be a CML and A ⋖ M, B ⋖ M, C ⋖ M. If A, B, C are generated by
the self-conjugate subsets 1 U, V, W respectively, then their associator subloop h(A, B, C)i is
generated by the set of all (u, v, w) with u ∈ U, v ∈ V, w ∈ W . In particluar h(A, B, C)i ⋖ M.
Proof. We shall briefly state the steps in the proof (for details see [Bru58]) : Let H be the
subloop generated by the set P consisting of all associators (u, v, w), u ∈ U, v ∈ V, w ∈
W . Then the fact that U, V, W are self conjugate and every inner mapping of M is an
automorphism of M implies P is also self conjugate and H is normal in M. So we can
consider the quotient loop G/H which is a commutative moufang loop. The set X of all
x ∈ M such that (x, V, W ) = 1 mod H is a subloop of M. Since X contains U so X contains
1A
permutation S of a loop M is called a pseudo-automorphism of M provided there exists at least one
element c of M , called a companion of S, such that (xS)(yS.c) = (xy)S.c holds for all x, y of M .
1A non-empty subset S of a loop M is called self-conjugate in M if SG ⊂ S where G = G(M ) is the inner
mapping group. Each element x ∈ M determines a self-conjugate subset xG called the conjugate class of x
in M .
APPROXIMATE SUBLOOPS AND FREIMAN’S THEOREM IN FINITELY GENERATED COMMUTATIVE MOUFANG LOOPS 11
A which implies (A, V, W ) = 1 mod H. Similarly, (A, B, W ) = 1 mod H and (A, B, C) = 1
mod H. This implies
h(A, B, C)i ⊆ H ⊆ h(A, B, C)i ⇒ h(A, B, C)i = H
The next lemma gives us an inclusion between associator subloops which is required to
deduce relations between lower and upper central series in CMLs.
Lemma 3.3. If A, B, C, X, Y are normal subloops of the commutative moufang loop M, then
h((A, B, C), X, Y )i ⊆ h((A, X, Y, )B, C)ih((B, X, Y ), C, A)ih((C, X, Y ), A, B)i
Proof. By the previous lemma, h((A, B, C), X, Y )i is generated by the set of all elements
((a, b, c), x, y) with a, b, c, x, y in A, B, C, X, Y respectively. If a is in A and p, q are in M,
then (a, p, q) is in A. Using the product rule for associators
(wx, y, z) = [(w, y, z)(x, y, z)][((w, y, z)w, x)((x, y, z), x, w)]
and the fact that every CML satisfies (p, (p, w, x), (p, y, z)) = 1 we have, (a(a, x, y), b, c) =
[(a, b, c)((a, x, y), b, c)]p where p = (((a, x, y), b, c), (a, x, y), a). Therefore (a(a, x, y), b, c) ≡
(a, b, c) mod h((A, X, Y ), B, C)i. If P denotes
h((A, X, Y, )B, C)ih((B, X, Y ), C, A)ih((C, X, Y ), A, B)i continuing as above we have
(3.1)
(a(a, x, y), b, c) ≡ (a, b, c) mod P
and similarly for any permutation of a, b, c. If θ = R(x, y), then
(a, b, c)((a, b, c), x, y) = θ(a, b, c) = (θ(a), θ(b), θ(c)) = (a(a, x, y), b(b, x, y), c(c, x, y)).
Therefore, using (3.1) thrice, (a, b, c)((a, b, c), x, y) ≡ (a, b(b, x, y), c(c, x, y)) ≡ (a, b, c(c, x, y)) ≡
(a, b, c) mod P which implies ((a, b, c), x, y) ≡ 1 mod P. We are done.
We come to one of the main lemmas in this section which relates elements of the lower
central series, the upper central series and the derived series.
Lemma 3.4. If M is a commutative moufang loop,
h(Mi , Mj , Mk )i ⊆ Mi+j+k+1
h(Mi , Mj , Zk )i ⊆ Zk−i−j−1
M (i) ⊆ M(3i −1)/2
for all i, j, k > 0.
Proof. We proceed by induction on j. Start with j = 0 so that M0 = M. The identity
h(Mi , M0 , M)i ⊆ Mi+1
holds for all non-negative integers i. Let us assume that it holds for all i > 0 and some j.
Then for j + 1 we have
h(Mi , Mj+1 , M)i = h(Mi , (Mj , M, M), M)i = h((Mj , M, M), Mi , M)i
⊆ h((Mj , Mi , M), M, M)ih((M, Mi , M), Mj , M)i
⊆ h(Mi+j+1, M, M)ih(Mi+1 , Mj , M)i ⊆ Mi+j+2 ,
This implies that the statement holds for all i, j > 0, k = 0. We now proceed by induction
on k to get the first result.
12
ARINDAM BISWAS
Next we show that M (i) ⊆ M(3i −1)/2 . The result holds for i = 0, 1. Let us assume that it
holds for some fixed i. Then using the first result we have that
M (i+1) = h(M i , M i , M i )i ⊆ h(M(3i −1)/2 , M(3i −1)/2 , M(3i −1)/2 )i ⊆ M3i + 3i −1 = M(3i+1 −1)/2
2
The result follows by induction.
Finally, from the definition of the upper central series, for i = j = 0 and for all k we
have h(M, M, Zk )i ⊆ Zk−1 . We use the previous lemma and induction to show this for all
non-negative i, j.
Proposition 3.5. Let M be a CML generated by n elements. Then the derived subloop
M ′ = h(M, M, M)i of M is finite.
Proof. Let M be generated by a set S with |S| = n. By Theorem 2.31, we know that M
is centrally nilpotent of class t where t 6 n − 1. For t = 0, 1, we have M ′ is either trivial
or a finitely generated group of exponent 3 and hence is finite (by the restricted Burnside
theorem). The rest of the proof is based on induction. Let t = k + 1 with k > 1 and let us
assume that the statement is true for commutative moufang loops of class at most k. The
quotient loop H = M/Mk has nilpotency class k and is finitely generated. This implies by
the induction hypothesis that H ′ = M ′ /Mk is finite. All that now remains to be shown is
that Mk is finite.
If k > 1 then Mk−1 ⊆ M ′ and hence Mk−1 /Mk is a subloop of H ′ , implying that Mk−1 /Mk
is finite. If k = 1 then Mk−1 /Mk = M/Mk is finitely generated. We can thus find a finite
non-empty subset T of M such that T and Mk generate Mk−1 . We also have the fact that S
generates M.
Claim 1. h(T, S, S)i is finite and Mk = h(T, S, S)i.
Proof of claim. Clearly J = h(T, S, S)i is a finitely generated subloop of Mk . Mk ⊂ Z(M)
where Z(M) denotes the centre of M and hence J is a normal subloop of M. J is infact
a finitely generated abelian group of exponent 3 and so J is finite. Using the fact that S
generates M, we have h(T, M, M)i = h(T, S, S)i = J. Now h(Mk , M, M) = Mk+1 = 1. But T
and Mk generate Mk−1 which means that h(Mk−1 , M, M)i ⊂ h(T, M, M)i∪h(Mk , M, M)i ⊂ J.
Thus we have
Mk = h(Mk−1 , M, M) ⊂ J ⊂ Mk .
Thus Mk is finite which implies M ′ is finite.
4. Progressions in commutative moufang loops
We follow the notation of group theorists with respect to mappings. Let us define a notion
of arithmetic progressions (APs) in commutative moufang loops. We recall the notion of
arithmetic progressions in case of arbitrary abelian groups.
Definition 4.1 (arithmetic progression in groups). Let G be an abelian group. An arithmetic
progression of dimension d and size L is a set of the form
P = {v0 + l1 v1 + ... + ld vd : 0 6 lj < Lj }
where l1 l2 ...ld = L. P is said to be proper if all of the sums in the above set are distinct, in
which case |P | = L.
APPROXIMATE SUBLOOPS AND FREIMAN’S THEOREM IN FINITELY GENERATED COMMUTATIVE MOUFANG LOOPS 13
The dimension of an arithmetic progression is the measure of its linear independence.
An equivalent notion of arithmetic progressions of dimension d can de defined in case of
commutative moufang loops but care must be taken because the lack of associativity implies
that even if the dimension is fixed the same “formal” expression for APs can have multiple
values depending on the position of the parentheses.
A naive way to define an AP in case of a commutative moufang loop is to look at the inverse
images of the projection onto the quotient loop. Formally,
Definition 4.2 (Generalised arithmetic progression in a CML). Let M be a commutative
moufang loop. Let M ′ = (M, M, M) denote its associator subloop which is a normal subloop
of M. Let π : M → M/M ′ denote the projection map onto the abelian group M/M ′ . Let
P be an arithmetic progression of dimension d in M/M ′ . Then π −1 (P ) is defined to be the
generalised arithmetic progression of dimension d in M.
Remark 4.3. Let P = {l1 v1 + l2 v2 + l3 v3 : |lj | < 5}. A typical element of a generalised AP
in a CML looks like ((x21 x22 )x3 )m where m ∈ M ′ and xi = π −1 (vi ), i = 1, 2, 3.
Definition 4.4 (Equality of generalised AP’s). Two elements R1 and R2 of a generalised
AP are said to be equal if π(R1 ) = π(R2 ). Alternatively if ∃m ∈ M ′ such that R1 = R2 m.
This an equivalence relation which partitions the set of all generalised AP’s into equivalence
classes.
We shall state a special form of APs called canonical APs
Definition 4.5 (Canonical form of generalised arithmetic progressions). Let x1 , x2 , ..., xr be
fixed elements in a commutative moufang loop M and L = (L1 , ..., Lr ) be a vector of positive
integers. Then the set of all products in the xi and their inverses which are of the form
{(...((xl11 xl22 )xl33 )...)xlrr : |li | < Li ∀i = 1, 2, ..., r}
is a canonical form of an AP of rank r and side lengths L.
Intuitively the canonical form of a generalised AP is easier to handle given that the position
of the parentheses are fixed and so two APs in canonical form can be easily compared. But
we need a method to convert each generalised AP into its canonical form. This process is
called associator collection.
Lemma 4.6 (associator collection). Let x1 , ..., xr be elements in a commutative moufang
loop M and let S be a finite string defined on M containing the elements x1 , ..., xr which
occur l1 , ..., lr times respectively. Then S can be written as
S = ((...((xl11 xl22 )xl33 )...)xlrr )m
with m ∈ M ′ = h(M, M, M)i
Proof. Use lemma 3.4 to see that h(Mi , Mj , Mk )i ⊆ M ′ for all i, j, k > 0. Consider two
strings S1 , S2 having the same elements, the same number of times and differing only in the
placement of the parentheses. Then it is direct to see that S1 = S2 m where m ∈ M ′ . For
example take S1 = (x21 x22 )x3 , S2 = (x21 (x2 x3 ))x2 . Then
S2 = (x21 (x2 (x2 x3 )))(x21 , x2 , x2 x3 )
= (((x21 x22 )x3 )(x21 , x22 , x3 ))(x21 , x2 , x2 x3 )
= S1 ((x21 x22 )x3 , (x21 , x22 , x3 ), (x21 , x2 , x2 x3 ))
14
ARINDAM BISWAS
= S1 m1
where m1 ∈ M ′ . Taking m = m−1
1 we get the result.
Lemma 4.7. Each generalised AP of dimension d can be written as Rm where R is an AP
in canonical form of dimension (rank) d and m ∈ M ′ .
Proof. It is sufficient to collect the terms in the formal expression of a generalised AP of
dimension d.
We also have an implicit notion of an AP inside a CML (without using the abelian quotient). We call this a “usual” arithmetic progression.
Definition 4.8 (Usual arithmetic progression in a CML). Let x1 , x2 , ..., xr be fixed elements
in a commutative moufang loop M and L = (L1 , ..., Lr ) be a vector of positive integers. Then
the set of all products in the xi and their inverses, in which each xi and x−1
i appear at most Li
times between them and also the position of the parentheses vary is called a “usual” arithmetic
∗
progression of rank r and side lengths L1 , ..., Lr . It is denoted by Pcml
(x1 , ..., xr ; L).
Lemma 4.9. Let M be a CML. Each usual AP in M can be reduced to an element of the
canonical form of a generalised AP and two usual APs are said to be equivalent (denoted by
≡) if they have the same canonical portion after reduction.
Proof. Start with any usual AP of rank (dimension) r and apply the process of collection.
∗
(x1 , x2 , x3 ; 4, 4, 4) of rank 3.Two typical elements of this usual
Remark 4.10. Take Pcml
arithmetic progression look like
−1 −1
2 2
((x1 x22 )(x−1
1 (x3 x2 )))(x3 (x1 x2 )) and
−1 −1
2
2
((x1 x22 )(x−1
1 (x3 x2 ))x3 )(x1 x2 )
Note that they are the same in case of groups but not in case of commutative moufang loops.
The above two elements have the same canonical portion after reduction (which is equal to
((x21 x22 )x3 )).
∗
Lemma 4.11. Let M be an n generated CML.If we have two usual APs, Pcml
and Q∗cml with
∗
∗
∗
′
∗
∗
′
∗
Pcml ≡ Qcml then |Pcml | 6 |M ||Qcml | and |Qcml | 6 |M ||Pcml |.
Proof. M is an n generated commutative moufang loop. Hence |M ′ | is finite. Take two
equivalent usual arithmetic progressions and reduce it to the canonical generalised form. Let
them be R1 m1 and R2 m2 where R1 = R2 (since they are equivalent) and m1 , m2 ∈ M ′ . As
′
M ′ is a subloop, m1 m−1
2 ∈ M and the result follows.
5. Structure of approximate subloops of finitely generated commutative
moufang loops
We start with generalisation of some well known results in groups adapted in the case of
commutative moufang loops.
Lemma 5.1 (Ruzsa’s covering lemma for loops). Let S, T be subsets of a loop L such that
|S.T | 6 K|S|. Then there is a set X ⊂ T , |X| 6 K, such that T ⊆ S −1 (SX).
APPROXIMATE SUBLOOPS AND FREIMAN’S THEOREM IN FINITELY GENERATED COMMUTATIVE MOUFANG LOOPS 15
Proof. We choose a maximal set X ⊆ T such that the sets S.x, x ∈ X, are pairwise disjoint.
Since ∪x∈X (S.x) ⊆ S.T we have |S||X| 6 K|S|, which implies that |X| 6 K. Now suppose
that t ∈ T . By maximality there must be some x ∈ X such that (S.t) ∩ (S.x) 6= φ, which
means that t ∈ S −1 (Sx) ⇒ T ⊆ S −1 (SX).
Lemma 5.2 (Ruzsa inequality for CML). Let A, Y, Z be finite sets in a commutative moufang
loop M. Let [A, Y, Z] = {(a, y, z) : a ∈ A, y ∈ Y, z ∈ Z}. Then we have
|A||Y Z −1 | 6 |A−1 Y | × |AZ −1 | × |[A, Y, Z]|
Proof. If yz −1 is an element of Y Z −1 with y ∈ Y and z ∈ Z then use 2.30 to see that
(ya−1 )(az −1 ) = (yz −1 )(y, a, z)
Hence we have that yz −1 can be written as a product of an element of A−1 Y and an element
of AZ −1 in at least |A|/|[A, Y, Z]| ways.
Remark 5.3. A special case is when one of A, Y, Z is in the centre Z(M) of the Moufang
loop M. For example in a CML one has ∀x ∈ M x3 ∈ Z(M). So if A(3) = {a3 : a ∈ A}
has size |A(3) | > |A|/C where C is an absolute constant and if A satisfies a small tripling
condition (A is an approximate loop) then a large portion of A has bounded growth in terms
of |A|.
An interesting result in the growth of finite sets in group theory is Petridis’s lemma [Pet12].
It is a direct consequence to see that in the setting of finitely generated commutative moufang
loops the same result holds with an extra term of the size of a finite set which depends on
the loop.
Lemma 5.4 (Petridis type lemma for Loops). Let X and B be finite sets in a Moufang loop.
Suppose that
|XB|
|ZB|
K :=
6
, ∀ Z ⊆ X.
|X|
|Z|
Then for all finite sets C
|(CX)B| 6 K|[C, X, B]||CX|
where [A, B, C] := {(a, b, c) : a ∈ A, b ∈ B, c ∈ C} and (a, b, c) := (ab.c)(a.bc)−1
Proof. In the proof of [Pet12] we need to add the associator subset [C, X, B].
Let C = {c1 , ..., cr } be ordered arbitrarily.
CX = ∪i=r
i=1 (ci Xi )
where X1 = X and for i > 1
Xi = {x ∈ X : ci x ∈
/ {c1 , ..., ci−1 }X}
Then for all j:
{c1 , ..., cj }X = ∪ji=1 (ci X) = ∪ji=1 (ci Xi )
The sets ci Xi are all disjoint and for all j, we have
|{c1 , ..., cj }X| = Σji=1 |ci Xi | = Σji=1 |Xi |
We now apply induction on r.
Step 1 : For r = 1 we have
|(c1 X)B| 6 |c1 (XB)||[c1 , X, B]| = |XB||[c1 , X, B]| = K|[c1 , X, B]||c1 X|
16
ARINDAM BISWAS
Step 2 : For r > 1 let us write Xrc = X\Xr , for the complement of Xr in X. By definition
of Xr , cr Xrc ⊆ {c1 , c2 , ..., cr−1 }X and (cr Xrc )B ⊆ ({c1 , c2 , ..., cr−1 }X)B
(CX)B = ({c1 , c2 , ..., cr }X)B = [({c1 , c2 , ..., cr−1 }X)B] ∪ [(cr X)B\(cr Xrc )B]
We note that |(cr X)B\(cr Xrc )B| 6 |cr (XB)\cr (Xrc B)||[cr , X, B]| = (|XB|−|Xrc B|)(|[cr , X, B]|)
This implies that
|(CX)B| 6 |({c1 , c2 , ..., cr−1 }X)B| + (|XB| − |Xrc B|)|[cr , X, B]|
The first summand is bounded by the inductive hypothesis on r namely
|({c1 , c2 , ..., cr−1 }X)B| 6 K|{c1 , ..., cr−1 }X||[C, X, B]| =
and the second summand is at most K|Xr ||[cr , X, B]|
Combining the above we get that
|(CX)B| 6 K|[C, X, B]||CX|
Remark 5.5. The proof of the above inequality shows that the following statements are also
true |(CX)B| 6 K max |[c, X, B]||CX|
c∈C
and
|C(XB)| 6 K max |[c, X, B]||CX|
c∈C
The above two lemmas suggest that if we have a K approximate subloop A inside a
commutative moufang loop such that the size of the associator set |[A, A, A]| is independent
of the size of |A| then we have the following |Ahni | 6 C(n, K)|A| where C is a constant
independant of the size of A.
Proposition 5.6 (Lemmas on approximate loops). Let M be a finitely generated commutative moufang loop having derived subloop M ′ . Fix K > 0.
(1) Let there exist X, A ⊂ M with |X| 6 K, A finite and A2 ⊆ X.A. Then |Ah3i | =
|A3 | 6 K 2 |M ′ ||A|
n−1
(2) Suppose A ∈ M be a symmetric set, |A3 | 6 K|A|. Then |Ahni | 6 ⌊ C2n−2
⌋K n−2 |M ′ |2n−6 |A|
(where Cn denotes the nth Catalan number)
(3) Let π : M → M ′ be a homomorphism of loops having a finite kernel (normal subloop)
N. Let A ∈ M ′ be such that A2 ⊆ X.A with |X| 6 K. Then (π −1 (A))2 ⊆ Y.π −1 (A)
where Y = Y −1 ∈ M and |Y | 6 2K.
(4) Let the notations be as in (2), then ∀n > 3 the following holds : if T ⊂ Ahni and
|T A| 6 K|A| then there exists X ⊂ T with |X| 6 K and T ⊆ A−1 (AX)
(1) It is clear to see that A2 ⊆ X.A implies A3 ⊆ (XA)A ⊆ (X(A2 ))(X, A, A) ⊆
(X(XA))M ′ ⊆ ((X 2 )A)M ′ and we have the cardinality inequality.
(2) We recall the Ruzsa inequality for CML,
Proof.
|A||Y Z −1 | 6 |A−1 Y | × |AZ −1 | × |[A, Y, Z]| 6 |M ′ ||A−1 Y | × |AZ −1 |.
APPROXIMATE SUBLOOPS AND FREIMAN’S THEOREM IN FINITELY GENERATED COMMUTATIVE MOUFANG LOOPS 17
Let us proceed by induction.
For n = 4 we have Ah4i = (A3 )A ∪ (A2 ).(A2 ).
Let T = (A2 )2 , choose Y = Z = A2 and we have
|A||T | 6 |M ′ ||A3 |2 6 K 2 |M ′ ||A|2 ⇒ |T | 6 K 2 |M ′ ||A|
Noting that ((A3 )A) ⊆ ((A2 )2 )M ′ we get that |Ah4i | 6 2(K|M ′ |)2 |A|.
For n = 5 we have Ah5i = (A3 )A2 ∪ ((A2 )2 )A ∪ (A3 .A)A Applying Ruzsa’s lemma
for CML to (A3 )A2 = A3 .A2 we get that
|A||A3.A2 | 6 |M ′ ||A3 .A||A3 | 6 K 3 |M ′ |3 |A|2 ,
hence |A3 .A2 | 6 (K|M ′ |)3 |A|. Considering the fact that the other elements in the
C4
union are subsets of (A3 .A2 )M ′ and there are at most
⌊ 22 ⌋ sets in the union we
2n
1
get that |Ah5i | 6 ⌊ C224 ⌋K 3 |M ′ |4 |A| (where Cn = n+1
denotes the nth Catalan
n
number.)
Fix k ∈ N. Let us assume that the result is true for all n 6 k − 1. For n 6 k − 1 by
k−2
the induction hypothesis, |Ahk−1i | 6 ⌊ C2k−3
⌋K k−3 |M ′ |2(k−4) |A|. This implies that
Ck−2 k−2 ′ 2(k−4)+1 2
⌋K |M |
|A|
2k−3
i=k−2
Now writing Ahki in the union form and noting that Ck−1 = Σi=0
Ci Ck−i we have
k−1
⌋K k−2 |M ′ |2(k−3) |A′ |.
|Ahki | 6 ⌊ C2k−2
(3) The fact that ker π is finite implies that π −1 (A) is finite. It remains to check the
existence of a set Y ∈ M of cardinality at most 2K such that (π −1 (A))2 ⊆ Y.π −1 (A).
By hypothesis there is a set X ∈ M ′ of cardinality at most K such that A2 ⊆ XA.
For each x ∈ X select an element ω(x) ∈ π −1 (x). Set Y = {ω(x) : x ∈ X} ∪
{ω −1 (x) : x ∈ X}, so that Y is symmetric and of cardinality at most 2K. Now given
a1 , a2 ∈ π −1 (A) , note that by definition of X there exist x ∈ X and a ∈ π −1 (A)
such that π(a1 )π(a2 ) = xπ(a). This implies that there exists b ∈ kerπ such that
a1 a2 = ω(x)ab. However, ab ∈ π −1 (A), and so a1 a2 ∈ Y π −1 (A), as desired.
(4) The last one is a direct consequence of Ruzsa covering lemma.
|A||Ahk−2i .A2 | 6 |M ′ ||Ahk−2i .A||A3 | 6 ⌊
Proposition 5.7. Let M be an n generated CML. There exists a correspondance between
approximate subloops of M and approximate subgroups of M/M ′ where M ′ denotes the derived subloop of M. If A is an approximate subloop in M then the canonical homomorphism
π : M → M/M ′ turns π(A2 ) into an approximate subgroup of M/M ′ and if B is an approximate subgroup inside the abelian group M/M ′ then π −1 (B) is an approximate subloop inside
M.
Proof. Let A ⊆ M be a K-approximate subloop in the CML M. We have |A3 | 6 K|A|. We
use Ruzsa’s inequality for loops to conclude that |A.(A3 )| ≤ K|M ′ ||A| and Ruzsa’s covering
lemma for loops to conclude that A3 ⊆ (XA)A where X is of size at most K|M ′ |. Taking the
canonical homomorphism on both sides we have π(A3 ) ⊂ π(X)π(A)2 in M/M ′ . As M/M ′
is an abelian group we see that π(A2 )2 = π(A3 )π(A) ⊆ π(X)π(A2 )π(A) ⊆ π(X)π(A3 ) ⊆
π(X)π(A2 ). Hence π(A2 ) is a π(X) approximate subgroup in M/M ′ where |X| 6 K|M ′ |.
18
ARINDAM BISWAS
In the other direction, consider the inverse image, π −1 : M/M ′ → M. We know that M is
a n generated CML, hence M ′ is a finite loop (also it is the kernel of the homomorphism π) .
If we have a K-approximatif subloop (subgroup) B in M/M ′ (abelian group) then B 2 ⊂ Y.B
(with |Y | 6 K). We can look at the inverse images π −1 (B) and conclude that π −1 (B) is a
subloop such that (π −1 (B))2 ⊆ Z.π −1 (B) (with |Z| 6 2K and Z ∈ M) and π −1 (B) ∈ M,
|π −1 (B)| < ∞ (we use the finiteness of the kernel M ′ here). Now from (1) of the previous
proposition we can conclude that π −1 (B) is an approximate subloop of M.
Remark 5.8. The essential point in the above proposition is the finiteness of M ′ which
is guaranteed in a finitely generated commutative moufang loop. For moufang loops having
approximate subloops A such that |[A, A, A]| 6 C where C is a constant independent of A,
we also have a same sort of structure theorem in the sense that these approximate subloops
are essentially controlled by approximate subgroups in M/M ′ . Note that commutativity is
not required in this case.
Definition 5.9 (coset progression in a CML). By a coset progression in a CML M, we mean
sets of the form π −1 (H)π −1(P ) ⊂ M where π is the projection map from M to M/M ′ , H is
a subgroup of M/M ′ and P is an arithmetic progression in M/M ′ . The dimension or rank
of the coset progression is the dimension of P .
We finally state a sort of structure theorem for finitely generated commutative moufang
loops.
Theorem 5.10 (Freiman’s theorem for finitely generated commutative moufang loops). Let
M be a n generated commutative moufang loop. Let A be a K-approximate subloop of M.
Then A2 is contained in a coset progression in M of dimension (rank) depending on K|M ′ |
and of size of progression at most f (K, |M ′ |)|A2 |.
Proof. Let M be a n generated CML. Let A ⊆ M be a K-approximate subloop. Let M ′ =
h(M, M, M)i and
π : M → M/M ′
be the canonical homomorphism. M ′ is finite and by the proposition 5.7, we have π(A2 ) is
a π(X) approximate group in the abelian group M/M ′ with |π(X)| 6 |X| 6 K|M ′ |.
Applying Freiman’s theorem in case of general abelian groups, Theorem 1.2 we have π(A2 )
is contained inside a coset progression H + P where P is of rank at most d(K|M ′ |) =
(K|M ′ |)O(1) and size at most |P | 6 exp((K|M ′ |)O(1) )|π(A2 )|. Thus A2 is contained inside
π −1 (H + P ) ⊆ π −1 (H)π −1 (P ) which has rank at most (K|M ′ |)O(1) and size of π −1 (P ) at
most exp((K|M ′ |)O(1) )|A2 |.
6. Concluding remarks
The question can be posed in the general setting of infinitely generated CMLs. A loop
L is said to have finite rank if there exists an integer r such that every finitely generated
subloop of L can be generated by at most r elements.
Proposition 6.1. Let M be a commutative Moufang loop having finite rank r associator
subloop h(M, M, M)i. Let A be a K-approximate subloop of M. Then A2 is contained in a
coset progression in M of dimension depending on CM K and of size of progression at most
f (K, CM )|A2 | where CM is an absolute constant depending on M.
APPROXIMATE SUBLOOPS AND FREIMAN’S THEOREM IN FINITELY GENERATED COMMUTATIVE MOUFANG LOOPS 19
Proof. The proof of this proposition follows from the following lemmas.
Lemma 6.2. Every commutative moufang loop without elements of infinite order is locally
finite.
Proof. Let M be a commutative moufang loop without elements of infinite order and N be
a finitely generated subloop of M. Then N/N ′ is a finitely generated abelian group without
elements of infinite order. So N/N ′ is finite. Using the same arguments as in Prop 3.5 we
can show that N ′ is also finite. This implies that N is finite and we are done.
Lemma 6.3. Let M be a commutative moufang loop with associator subloop M ′ and centre
Z(M), then M ′ and M/Z(M) are locally finite loops of exponent 3.
Proof. The mapping x → x3 is an endomorphism of M into Z(M). The loops M ′ and
M/Z(M) therefore have exponent 3. Using Lemma 6.2 we have M ′ and M/Z(M) are locally
finite.
Lemma 6.4. Let M be commutative moufang loop having finite rank associator subloop
h(M, M, M)i. Let A ∈ M be a finite set. Then |h(A, A, A)i| is finite and independent of |A|.
Proof. We recall that a loop M is said to have finite rank if there exists an integer r such
that every finitely generated subloop of M can be generated by at most r elements. The
subloop h(A, A, A)i is finitely generated and hence there exists an r (uniform) such that it
can be generated by at most r elements. Also h(A, A, A)i is a CML of exponent 3 and hence
is locally finite. Thus h(A, A, A)i is finite and its size is independent of |A|.
6.1. Manin’s problem and effective bounds on the progressions. An interesting note
to add is the dependency of the size and rank of the progression to the size of the associator
subloop. Manin’s problem of specifying the 3 rank of a commutative moufang loop deals
with the size of the associator subloops with respect to the number of generators [Man74].
Progress in this area has been pretty slow even for small values of n.An important approach
to a solution to this question was done by J.D.H.Smith in 1982. He gave a hypothetical
formula for calculating the 3 rank of a finitely generated CML provided the triple argument
hypothesis holds 2. Under this assumption he was able to compute the orders of the associator subloops. The details can be found in [Smi82].
But it was shown by N. Sandu in [San87] that triple argument hypothesis doesn’t work for
n > 9. Recent works by Grishkov and Sheshtakov [GS11] showed that the triple argument
hypothesis actually fails for n > 7. Smith’s formula still gives correct results for n 6 6.
References
[Bén84] Lucien Bénéteau, 3-abelian groups and commutative Moufang loops, European J. Combin. 5 (1984),
no. 3, 193–196. MR 765625
[BG08] Jean Bourgain and Alex Gamburd, Uniform expansion bounds for Cayley graphs of SL2 (Fp ), Ann.
of Math. (2) 167 (2008), no. 2, 625–642. MR 2415383
[BGT12] Emmanuel Breuillard, Ben Green, and Terence Tao, The structure of approximate groups, Publ.
Math. Inst. Hautes Études Sci. 116 (2012), 115–221. MR 3090256
[Bru46] R. H. Bruck, Contributions to the theory of loops, Trans. Amer. Math. Soc. 60 (1946), 245–354.
MR 0017288
2The
triple argument hypothesis says that the normal subloop of the loop generated by all associators in
with an argument appearing thrice vanishes.
20
[Bru58]
[Cha02]
[Fre64]
[GR07]
[GS11]
[Hel08]
[Man74]
[Mou35]
[Pet12]
[San87]
[San12]
[Smi78a]
[Smi78b]
[Smi82]
[Tao08]
[Tao10]
[Toi14]
ARINDAM BISWAS
Richard Hubert Bruck, A survey of binary systems, Ergebnisse der Mathematik und ihrer Grenzgebiete. Neue Folge, Heft 20. Reihe: Gruppentheorie, Springer Verlag, Berlin-Göttingen-Heidelberg,
1958. MR 0093552
Mei-Chu Chang, A polynomial bound in Freiman’s theorem, Duke Math. J. 113 (2002), no. 3,
399–419. MR 1909605 (2003d:11151)
G. A. Freı̆man, On the addition of finite sets, Dokl. Akad. Nauk SSSR 158 (1964), 1038–1041.
MR 0168529 (29 #5791)
Ben Green and Imre Z. Ruzsa, Freiman’s theorem in an arbitrary abelian group, J. Lond. Math.
Soc. (2) 75 (2007), no. 1, 163–175. MR 2302736 (2007m:20087)
Alexander N. Grishkov and Ivan P. Shestakov, Commutative Moufang loops and alternative algebras, J. Algebra 333 (2011), 1–13. MR 2785933
H. A. Helfgott, Growth and generation in SL2 (Z/pZ), Ann. of Math. (2) 167 (2008), no. 2, 601–623.
MR 2415382
Yu. I. Manin, Cubic forms: algebra, geometry, arithmetic, North-Holland Publishing Co.,
Amsterdam-London; American Elsevier Publishing Co., New York, 1974, Translated from the
Russian by M. Hazewinkel, North-Holland Mathematical Library, Vol. 4. MR 0460349
Ruth Moufang, Zur Struktur von Alternativkörpern, Math. Ann. 110 (1935), no. 1, 416–430.
MR 1512948
Giorgis Petridis, New proofs of Plünnecke-type estimates for product sets in groups, Combinatorica
32 (2012), no. 6, 721–733. MR 3063158
N. I. Sandu, Infinite irreducible systems of identities of commutative Moufang loops and of distributive Steiner quasigroups, Izv. Akad. Nauk SSSR Ser. Mat. 51 (1987), no. 1, 171–188, 208.
MR 887606
Tom Sanders, On the Bogolyubov-Ruzsa lemma, Anal. PDE 5 (2012), no. 3, 627–655. MR 2994508
J. D. H. Smith, On the nilpotence class of commutative Moufang loops, Math. Proc. Cambridge
Philos. Soc. 84 (1978), no. 3, 387–404. MR 0498931
, A second grammar of associators, Math. Proc. Cambridge Philos. Soc. 84 (1978), no. 3,
405–415. MR 0498933
Jonathan D. H. Smith, Commutative Moufang loops and Bessel functions, Invent. Math. 67 (1982),
no. 1, 173–187. MR 664331
Terence Tao, Product set estimates for non-commutative groups, Combinatorica 28 (2008), no. 5,
547–594. MR 2501249
, Freiman’s theorem for solvable groups, Contrib. Discrete Math. 5 (2010), no. 2, 137–184.
MR 2791295
Matthew C. H. Tointon, Freiman’s theorem in an arbitrary nilpotent group, Proc. Lond. Math. Soc.
(3) 109 (2014), no. 2, 318–352. MR 3254927
Universität Wien, Fakultät für Mathematik, Oskar-Morgenstern-Platz 1, 1090 Wien,
Austria.
E-mail address: [email protected]
| 4math.GR
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.